Commit | Line | Data |
---|---|---|
3dcf60bc | 1 | // SPDX-License-Identifier: GPL-2.0 |
75bb4625 JA |
2 | /* |
3 | * Block multiqueue core code | |
4 | * | |
5 | * Copyright (C) 2013-2014 Jens Axboe | |
6 | * Copyright (C) 2013-2014 Christoph Hellwig | |
7 | */ | |
320ae51f JA |
8 | #include <linux/kernel.h> |
9 | #include <linux/module.h> | |
10 | #include <linux/backing-dev.h> | |
11 | #include <linux/bio.h> | |
12 | #include <linux/blkdev.h> | |
fe45e630 | 13 | #include <linux/blk-integrity.h> |
f75782e4 | 14 | #include <linux/kmemleak.h> |
320ae51f JA |
15 | #include <linux/mm.h> |
16 | #include <linux/init.h> | |
17 | #include <linux/slab.h> | |
18 | #include <linux/workqueue.h> | |
19 | #include <linux/smp.h> | |
e41d12f5 | 20 | #include <linux/interrupt.h> |
320ae51f | 21 | #include <linux/llist.h> |
320ae51f JA |
22 | #include <linux/cpu.h> |
23 | #include <linux/cache.h> | |
24 | #include <linux/sched/sysctl.h> | |
105ab3d8 | 25 | #include <linux/sched/topology.h> |
174cd4b1 | 26 | #include <linux/sched/signal.h> |
320ae51f | 27 | #include <linux/delay.h> |
aedcd72f | 28 | #include <linux/crash_dump.h> |
88c7b2b7 | 29 | #include <linux/prefetch.h> |
a892c8d5 | 30 | #include <linux/blk-crypto.h> |
82d981d4 | 31 | #include <linux/part_stat.h> |
320ae51f JA |
32 | |
33 | #include <trace/events/block.h> | |
34 | ||
54d4e6ab | 35 | #include <linux/t10-pi.h> |
320ae51f JA |
36 | #include "blk.h" |
37 | #include "blk-mq.h" | |
9c1051aa | 38 | #include "blk-mq-debugfs.h" |
986d413b | 39 | #include "blk-pm.h" |
cf43e6be | 40 | #include "blk-stat.h" |
bd166ef1 | 41 | #include "blk-mq-sched.h" |
c1c80384 | 42 | #include "blk-rq-qos.h" |
82b74cac | 43 | #include "blk-ioprio.h" |
320ae51f | 44 | |
f9ab4918 | 45 | static DEFINE_PER_CPU(struct llist_head, blk_cpu_done); |
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); | |
3e08773c | 52 | |
f70299f0 CH |
53 | static inline struct blk_mq_hw_ctx *blk_qc_to_hctx(struct request_queue *q, |
54 | blk_qc_t qc) | |
55 | { | |
54bdd67d | 56 | return xa_load(&q->hctx_table, qc); |
c6699d6f CH |
57 | } |
58 | ||
3e08773c CH |
59 | static inline blk_qc_t blk_rq_to_qc(struct request *rq) |
60 | { | |
54bdd67d | 61 | return rq->mq_hctx->queue_num; |
3e08773c CH |
62 | } |
63 | ||
320ae51f | 64 | /* |
85fae294 YY |
65 | * Check if any of the ctx, dispatch list or elevator |
66 | * have pending work in this hardware queue. | |
320ae51f | 67 | */ |
79f720a7 | 68 | static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx) |
320ae51f | 69 | { |
79f720a7 JA |
70 | return !list_empty_careful(&hctx->dispatch) || |
71 | sbitmap_any_bit_set(&hctx->ctx_map) || | |
bd166ef1 | 72 | blk_mq_sched_has_work(hctx); |
1429d7c9 JA |
73 | } |
74 | ||
320ae51f JA |
75 | /* |
76 | * Mark this ctx as having pending work in this hardware queue | |
77 | */ | |
78 | static void blk_mq_hctx_mark_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 | if (!sbitmap_test_bit(&hctx->ctx_map, bit)) | |
84 | sbitmap_set_bit(&hctx->ctx_map, bit); | |
1429d7c9 JA |
85 | } |
86 | ||
87 | static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx, | |
88 | struct blk_mq_ctx *ctx) | |
89 | { | |
f31967f0 JA |
90 | const int bit = ctx->index_hw[hctx->type]; |
91 | ||
92 | sbitmap_clear_bit(&hctx->ctx_map, bit); | |
320ae51f JA |
93 | } |
94 | ||
f299b7c7 | 95 | struct mq_inflight { |
8446fe92 | 96 | struct block_device *part; |
a2e80f6f | 97 | unsigned int inflight[2]; |
f299b7c7 JA |
98 | }; |
99 | ||
2dd6532e | 100 | static bool blk_mq_check_inflight(struct request *rq, void *priv) |
f299b7c7 JA |
101 | { |
102 | struct mq_inflight *mi = priv; | |
103 | ||
b81c14ca HW |
104 | if (rq->part && blk_do_io_stat(rq) && |
105 | (!mi->part->bd_partno || rq->part == mi->part) && | |
b0d97557 | 106 | blk_mq_rq_state(rq) == MQ_RQ_IN_FLIGHT) |
bb4e6b14 | 107 | mi->inflight[rq_data_dir(rq)]++; |
7baa8572 JA |
108 | |
109 | return true; | |
f299b7c7 JA |
110 | } |
111 | ||
8446fe92 CH |
112 | unsigned int blk_mq_in_flight(struct request_queue *q, |
113 | struct block_device *part) | |
f299b7c7 | 114 | { |
a2e80f6f | 115 | struct mq_inflight mi = { .part = part }; |
f299b7c7 | 116 | |
f299b7c7 | 117 | blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi); |
e016b782 | 118 | |
a2e80f6f | 119 | return mi.inflight[0] + mi.inflight[1]; |
bf0ddaba OS |
120 | } |
121 | ||
8446fe92 CH |
122 | void blk_mq_in_flight_rw(struct request_queue *q, struct block_device *part, |
123 | unsigned int inflight[2]) | |
bf0ddaba | 124 | { |
a2e80f6f | 125 | struct mq_inflight mi = { .part = part }; |
bf0ddaba | 126 | |
bb4e6b14 | 127 | blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi); |
a2e80f6f PB |
128 | inflight[0] = mi.inflight[0]; |
129 | inflight[1] = mi.inflight[1]; | |
bf0ddaba OS |
130 | } |
131 | ||
1671d522 | 132 | void blk_freeze_queue_start(struct request_queue *q) |
43a5e4e2 | 133 | { |
7996a8b5 BL |
134 | mutex_lock(&q->mq_freeze_lock); |
135 | if (++q->mq_freeze_depth == 1) { | |
3ef28e83 | 136 | percpu_ref_kill(&q->q_usage_counter); |
7996a8b5 | 137 | mutex_unlock(&q->mq_freeze_lock); |
344e9ffc | 138 | if (queue_is_mq(q)) |
055f6e18 | 139 | blk_mq_run_hw_queues(q, false); |
7996a8b5 BL |
140 | } else { |
141 | mutex_unlock(&q->mq_freeze_lock); | |
cddd5d17 | 142 | } |
f3af020b | 143 | } |
1671d522 | 144 | EXPORT_SYMBOL_GPL(blk_freeze_queue_start); |
f3af020b | 145 | |
6bae363e | 146 | void blk_mq_freeze_queue_wait(struct request_queue *q) |
f3af020b | 147 | { |
3ef28e83 | 148 | wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter)); |
43a5e4e2 | 149 | } |
6bae363e | 150 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait); |
43a5e4e2 | 151 | |
f91328c4 KB |
152 | int blk_mq_freeze_queue_wait_timeout(struct request_queue *q, |
153 | unsigned long timeout) | |
154 | { | |
155 | return wait_event_timeout(q->mq_freeze_wq, | |
156 | percpu_ref_is_zero(&q->q_usage_counter), | |
157 | timeout); | |
158 | } | |
159 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout); | |
43a5e4e2 | 160 | |
f3af020b TH |
161 | /* |
162 | * Guarantee no request is in use, so we can change any data structure of | |
163 | * the queue afterward. | |
164 | */ | |
3ef28e83 | 165 | void blk_freeze_queue(struct request_queue *q) |
f3af020b | 166 | { |
3ef28e83 DW |
167 | /* |
168 | * In the !blk_mq case we are only calling this to kill the | |
169 | * q_usage_counter, otherwise this increases the freeze depth | |
170 | * and waits for it to return to zero. For this reason there is | |
171 | * no blk_unfreeze_queue(), and blk_freeze_queue() is not | |
172 | * exported to drivers as the only user for unfreeze is blk_mq. | |
173 | */ | |
1671d522 | 174 | blk_freeze_queue_start(q); |
f3af020b TH |
175 | blk_mq_freeze_queue_wait(q); |
176 | } | |
3ef28e83 DW |
177 | |
178 | void blk_mq_freeze_queue(struct request_queue *q) | |
179 | { | |
180 | /* | |
181 | * ...just an alias to keep freeze and unfreeze actions balanced | |
182 | * in the blk_mq_* namespace | |
183 | */ | |
184 | blk_freeze_queue(q); | |
185 | } | |
c761d96b | 186 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue); |
f3af020b | 187 | |
aec89dc5 | 188 | void __blk_mq_unfreeze_queue(struct request_queue *q, bool force_atomic) |
320ae51f | 189 | { |
7996a8b5 | 190 | mutex_lock(&q->mq_freeze_lock); |
aec89dc5 CH |
191 | if (force_atomic) |
192 | q->q_usage_counter.data->force_atomic = true; | |
7996a8b5 BL |
193 | q->mq_freeze_depth--; |
194 | WARN_ON_ONCE(q->mq_freeze_depth < 0); | |
195 | if (!q->mq_freeze_depth) { | |
bdd63160 | 196 | percpu_ref_resurrect(&q->q_usage_counter); |
320ae51f | 197 | wake_up_all(&q->mq_freeze_wq); |
add703fd | 198 | } |
7996a8b5 | 199 | mutex_unlock(&q->mq_freeze_lock); |
320ae51f | 200 | } |
aec89dc5 CH |
201 | |
202 | void blk_mq_unfreeze_queue(struct request_queue *q) | |
203 | { | |
204 | __blk_mq_unfreeze_queue(q, false); | |
205 | } | |
b4c6a028 | 206 | EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue); |
320ae51f | 207 | |
852ec809 BVA |
208 | /* |
209 | * FIXME: replace the scsi_internal_device_*block_nowait() calls in the | |
210 | * mpt3sas driver such that this function can be removed. | |
211 | */ | |
212 | void blk_mq_quiesce_queue_nowait(struct request_queue *q) | |
213 | { | |
e70feb8b ML |
214 | unsigned long flags; |
215 | ||
216 | spin_lock_irqsave(&q->queue_lock, flags); | |
217 | if (!q->quiesce_depth++) | |
218 | blk_queue_flag_set(QUEUE_FLAG_QUIESCED, q); | |
219 | spin_unlock_irqrestore(&q->queue_lock, flags); | |
852ec809 BVA |
220 | } |
221 | EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait); | |
222 | ||
6a83e74d | 223 | /** |
9ef4d020 | 224 | * blk_mq_wait_quiesce_done() - wait until in-progress quiesce is done |
483239c7 | 225 | * @set: tag_set to wait on |
6a83e74d | 226 | * |
9ef4d020 | 227 | * Note: it is driver's responsibility for making sure that quiesce has |
483239c7 CH |
228 | * been started on or more of the request_queues of the tag_set. This |
229 | * function only waits for the quiesce on those request_queues that had | |
230 | * the quiesce flag set using blk_mq_quiesce_queue_nowait. | |
6a83e74d | 231 | */ |
483239c7 | 232 | void blk_mq_wait_quiesce_done(struct blk_mq_tag_set *set) |
6a83e74d | 233 | { |
483239c7 CH |
234 | if (set->flags & BLK_MQ_F_BLOCKING) |
235 | synchronize_srcu(set->srcu); | |
704b914f | 236 | else |
6a83e74d BVA |
237 | synchronize_rcu(); |
238 | } | |
9ef4d020 ML |
239 | EXPORT_SYMBOL_GPL(blk_mq_wait_quiesce_done); |
240 | ||
241 | /** | |
242 | * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished | |
243 | * @q: request queue. | |
244 | * | |
245 | * Note: this function does not prevent that the struct request end_io() | |
246 | * callback function is invoked. Once this function is returned, we make | |
247 | * sure no dispatch can happen until the queue is unquiesced via | |
248 | * blk_mq_unquiesce_queue(). | |
249 | */ | |
250 | void blk_mq_quiesce_queue(struct request_queue *q) | |
251 | { | |
252 | blk_mq_quiesce_queue_nowait(q); | |
8537380b CH |
253 | /* nothing to wait for non-mq queues */ |
254 | if (queue_is_mq(q)) | |
483239c7 | 255 | blk_mq_wait_quiesce_done(q->tag_set); |
9ef4d020 | 256 | } |
6a83e74d BVA |
257 | EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue); |
258 | ||
e4e73913 ML |
259 | /* |
260 | * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue() | |
261 | * @q: request queue. | |
262 | * | |
263 | * This function recovers queue into the state before quiescing | |
264 | * which is done by blk_mq_quiesce_queue. | |
265 | */ | |
266 | void blk_mq_unquiesce_queue(struct request_queue *q) | |
267 | { | |
e70feb8b ML |
268 | unsigned long flags; |
269 | bool run_queue = false; | |
270 | ||
271 | spin_lock_irqsave(&q->queue_lock, flags); | |
272 | if (WARN_ON_ONCE(q->quiesce_depth <= 0)) { | |
273 | ; | |
274 | } else if (!--q->quiesce_depth) { | |
275 | blk_queue_flag_clear(QUEUE_FLAG_QUIESCED, q); | |
276 | run_queue = true; | |
277 | } | |
278 | spin_unlock_irqrestore(&q->queue_lock, flags); | |
f4560ffe | 279 | |
1d9e9bc6 | 280 | /* dispatch requests which are inserted during quiescing */ |
e70feb8b ML |
281 | if (run_queue) |
282 | blk_mq_run_hw_queues(q, true); | |
e4e73913 ML |
283 | } |
284 | EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue); | |
285 | ||
414dd48e CL |
286 | void blk_mq_quiesce_tagset(struct blk_mq_tag_set *set) |
287 | { | |
288 | struct request_queue *q; | |
289 | ||
290 | mutex_lock(&set->tag_list_lock); | |
291 | list_for_each_entry(q, &set->tag_list, tag_set_list) { | |
292 | if (!blk_queue_skip_tagset_quiesce(q)) | |
293 | blk_mq_quiesce_queue_nowait(q); | |
294 | } | |
295 | blk_mq_wait_quiesce_done(set); | |
296 | mutex_unlock(&set->tag_list_lock); | |
297 | } | |
298 | EXPORT_SYMBOL_GPL(blk_mq_quiesce_tagset); | |
299 | ||
300 | void blk_mq_unquiesce_tagset(struct blk_mq_tag_set *set) | |
301 | { | |
302 | struct request_queue *q; | |
303 | ||
304 | mutex_lock(&set->tag_list_lock); | |
305 | list_for_each_entry(q, &set->tag_list, tag_set_list) { | |
306 | if (!blk_queue_skip_tagset_quiesce(q)) | |
307 | blk_mq_unquiesce_queue(q); | |
308 | } | |
309 | mutex_unlock(&set->tag_list_lock); | |
310 | } | |
311 | EXPORT_SYMBOL_GPL(blk_mq_unquiesce_tagset); | |
312 | ||
aed3ea94 JA |
313 | void blk_mq_wake_waiters(struct request_queue *q) |
314 | { | |
315 | struct blk_mq_hw_ctx *hctx; | |
4f481208 | 316 | unsigned long i; |
aed3ea94 JA |
317 | |
318 | queue_for_each_hw_ctx(q, hctx, i) | |
319 | if (blk_mq_hw_queue_mapped(hctx)) | |
320 | blk_mq_tag_wakeup_all(hctx->tags, true); | |
321 | } | |
322 | ||
52fdbbcc CH |
323 | void blk_rq_init(struct request_queue *q, struct request *rq) |
324 | { | |
325 | memset(rq, 0, sizeof(*rq)); | |
326 | ||
327 | INIT_LIST_HEAD(&rq->queuelist); | |
328 | rq->q = q; | |
329 | rq->__sector = (sector_t) -1; | |
330 | INIT_HLIST_NODE(&rq->hash); | |
331 | RB_CLEAR_NODE(&rq->rb_node); | |
332 | rq->tag = BLK_MQ_NO_TAG; | |
333 | rq->internal_tag = BLK_MQ_NO_TAG; | |
334 | rq->start_time_ns = ktime_get_ns(); | |
335 | rq->part = NULL; | |
336 | blk_crypto_rq_set_defaults(rq); | |
337 | } | |
338 | EXPORT_SYMBOL(blk_rq_init); | |
339 | ||
e4cdf1a1 | 340 | static struct request *blk_mq_rq_ctx_init(struct blk_mq_alloc_data *data, |
fe6134f6 | 341 | struct blk_mq_tags *tags, unsigned int tag, u64 alloc_time_ns) |
320ae51f | 342 | { |
605f784e PB |
343 | struct blk_mq_ctx *ctx = data->ctx; |
344 | struct blk_mq_hw_ctx *hctx = data->hctx; | |
345 | struct request_queue *q = data->q; | |
e4cdf1a1 | 346 | struct request *rq = tags->static_rqs[tag]; |
c3a148d2 | 347 | |
c7b84d42 JA |
348 | rq->q = q; |
349 | rq->mq_ctx = ctx; | |
350 | rq->mq_hctx = hctx; | |
351 | rq->cmd_flags = data->cmd_flags; | |
352 | ||
353 | if (data->flags & BLK_MQ_REQ_PM) | |
354 | data->rq_flags |= RQF_PM; | |
355 | if (blk_queue_io_stat(q)) | |
356 | data->rq_flags |= RQF_IO_STAT; | |
357 | rq->rq_flags = data->rq_flags; | |
358 | ||
dd6216bb | 359 | if (data->rq_flags & RQF_SCHED_TAGS) { |
56f8da64 JA |
360 | rq->tag = BLK_MQ_NO_TAG; |
361 | rq->internal_tag = tag; | |
dd6216bb CH |
362 | } else { |
363 | rq->tag = tag; | |
364 | rq->internal_tag = BLK_MQ_NO_TAG; | |
e4cdf1a1 | 365 | } |
c7b84d42 | 366 | rq->timeout = 0; |
e4cdf1a1 | 367 | |
4f266f2b PB |
368 | if (blk_mq_need_time_stamp(rq)) |
369 | rq->start_time_ns = ktime_get_ns(); | |
370 | else | |
371 | rq->start_time_ns = 0; | |
af76e555 | 372 | rq->part = NULL; |
6f816b4b TH |
373 | #ifdef CONFIG_BLK_RQ_ALLOC_TIME |
374 | rq->alloc_time_ns = alloc_time_ns; | |
375 | #endif | |
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 JA |
404 | static inline struct request * |
405 | __blk_mq_alloc_requests_batch(struct blk_mq_alloc_data *data, | |
406 | u64 alloc_time_ns) | |
407 | { | |
408 | unsigned int tag, tag_offset; | |
fe6134f6 | 409 | struct blk_mq_tags *tags; |
349302da | 410 | struct request *rq; |
fe6134f6 | 411 | unsigned long tag_mask; |
349302da JA |
412 | int i, nr = 0; |
413 | ||
fe6134f6 JA |
414 | tag_mask = blk_mq_get_tags(data, data->nr_tags, &tag_offset); |
415 | if (unlikely(!tag_mask)) | |
349302da JA |
416 | return NULL; |
417 | ||
fe6134f6 JA |
418 | tags = blk_mq_tags_from_data(data); |
419 | for (i = 0; tag_mask; i++) { | |
420 | if (!(tag_mask & (1UL << i))) | |
349302da JA |
421 | continue; |
422 | tag = tag_offset + i; | |
a22c00be | 423 | prefetch(tags->static_rqs[tag]); |
fe6134f6 JA |
424 | tag_mask &= ~(1UL << i); |
425 | rq = blk_mq_rq_ctx_init(data, tags, tag, alloc_time_ns); | |
013a7f95 | 426 | rq_list_add(data->cached_rq, rq); |
c5fc7b93 | 427 | nr++; |
349302da | 428 | } |
c5fc7b93 JA |
429 | /* caller already holds a reference, add for remainder */ |
430 | percpu_ref_get_many(&data->q->q_usage_counter, nr - 1); | |
349302da JA |
431 | data->nr_tags -= nr; |
432 | ||
013a7f95 | 433 | return rq_list_pop(data->cached_rq); |
349302da JA |
434 | } |
435 | ||
b90cfaed | 436 | static struct request *__blk_mq_alloc_requests(struct blk_mq_alloc_data *data) |
d2c0d383 | 437 | { |
e6e7abff | 438 | struct request_queue *q = data->q; |
6f816b4b | 439 | u64 alloc_time_ns = 0; |
47c122e3 | 440 | struct request *rq; |
600c3b0c | 441 | unsigned int tag; |
d2c0d383 | 442 | |
6f816b4b TH |
443 | /* alloc_time includes depth and tag waits */ |
444 | if (blk_queue_rq_alloc_time(q)) | |
445 | alloc_time_ns = ktime_get_ns(); | |
446 | ||
f9afca4d | 447 | if (data->cmd_flags & REQ_NOWAIT) |
03a07c92 | 448 | data->flags |= BLK_MQ_REQ_NOWAIT; |
d2c0d383 | 449 | |
781dd830 | 450 | if (q->elevator) { |
dd6216bb CH |
451 | /* |
452 | * All requests use scheduler tags when an I/O scheduler is | |
453 | * enabled for the queue. | |
454 | */ | |
455 | data->rq_flags |= RQF_SCHED_TAGS; | |
781dd830 | 456 | |
d2c0d383 | 457 | /* |
8d663f34 | 458 | * Flush/passthrough requests are special and go directly to the |
dd6216bb | 459 | * dispatch list. |
d2c0d383 | 460 | */ |
be4c4278 | 461 | if ((data->cmd_flags & REQ_OP_MASK) != REQ_OP_FLUSH && |
dd6216bb CH |
462 | !blk_op_is_passthrough(data->cmd_flags)) { |
463 | struct elevator_mq_ops *ops = &q->elevator->type->ops; | |
464 | ||
465 | WARN_ON_ONCE(data->flags & BLK_MQ_REQ_RESERVED); | |
466 | ||
467 | data->rq_flags |= RQF_USE_SCHED; | |
468 | if (ops->limit_depth) | |
469 | ops->limit_depth(data->cmd_flags, data); | |
470 | } | |
d2c0d383 CH |
471 | } |
472 | ||
bf0beec0 | 473 | retry: |
600c3b0c CH |
474 | data->ctx = blk_mq_get_ctx(q); |
475 | data->hctx = blk_mq_map_queue(q, data->cmd_flags, data->ctx); | |
dd6216bb | 476 | if (!(data->rq_flags & RQF_SCHED_TAGS)) |
600c3b0c CH |
477 | blk_mq_tag_busy(data->hctx); |
478 | ||
99e48cd6 JG |
479 | if (data->flags & BLK_MQ_REQ_RESERVED) |
480 | data->rq_flags |= RQF_RESV; | |
481 | ||
349302da JA |
482 | /* |
483 | * Try batched alloc if we want more than 1 tag. | |
484 | */ | |
485 | if (data->nr_tags > 1) { | |
486 | rq = __blk_mq_alloc_requests_batch(data, alloc_time_ns); | |
487 | if (rq) | |
488 | return rq; | |
489 | data->nr_tags = 1; | |
490 | } | |
491 | ||
bf0beec0 ML |
492 | /* |
493 | * Waiting allocations only fail because of an inactive hctx. In that | |
494 | * case just retry the hctx assignment and tag allocation as CPU hotplug | |
495 | * should have migrated us to an online CPU by now. | |
496 | */ | |
e4cdf1a1 | 497 | tag = blk_mq_get_tag(data); |
bf0beec0 ML |
498 | if (tag == BLK_MQ_NO_TAG) { |
499 | if (data->flags & BLK_MQ_REQ_NOWAIT) | |
500 | return NULL; | |
bf0beec0 | 501 | /* |
349302da JA |
502 | * Give up the CPU and sleep for a random short time to |
503 | * ensure that thread using a realtime scheduling class | |
504 | * are migrated off the CPU, and thus off the hctx that | |
505 | * is going away. | |
bf0beec0 ML |
506 | */ |
507 | msleep(3); | |
508 | goto retry; | |
509 | } | |
47c122e3 | 510 | |
fe6134f6 JA |
511 | return blk_mq_rq_ctx_init(data, blk_mq_tags_from_data(data), tag, |
512 | alloc_time_ns); | |
d2c0d383 CH |
513 | } |
514 | ||
4b6a5d9c JA |
515 | static struct request *blk_mq_rq_cache_fill(struct request_queue *q, |
516 | struct blk_plug *plug, | |
517 | blk_opf_t opf, | |
518 | blk_mq_req_flags_t flags) | |
320ae51f | 519 | { |
e6e7abff CH |
520 | struct blk_mq_alloc_data data = { |
521 | .q = q, | |
522 | .flags = flags, | |
16458cf3 | 523 | .cmd_flags = opf, |
4b6a5d9c JA |
524 | .nr_tags = plug->nr_ios, |
525 | .cached_rq = &plug->cached_rq, | |
e6e7abff | 526 | }; |
bd166ef1 | 527 | struct request *rq; |
320ae51f | 528 | |
4b6a5d9c JA |
529 | if (blk_queue_enter(q, flags)) |
530 | return NULL; | |
531 | ||
532 | plug->nr_ios = 1; | |
320ae51f | 533 | |
b90cfaed | 534 | rq = __blk_mq_alloc_requests(&data); |
4b6a5d9c JA |
535 | if (unlikely(!rq)) |
536 | blk_queue_exit(q); | |
537 | return rq; | |
538 | } | |
539 | ||
540 | static struct request *blk_mq_alloc_cached_request(struct request_queue *q, | |
541 | blk_opf_t opf, | |
542 | blk_mq_req_flags_t flags) | |
543 | { | |
544 | struct blk_plug *plug = current->plug; | |
545 | struct request *rq; | |
546 | ||
547 | if (!plug) | |
548 | return NULL; | |
40467282 | 549 | |
4b6a5d9c JA |
550 | if (rq_list_empty(plug->cached_rq)) { |
551 | if (plug->nr_ios == 1) | |
552 | return NULL; | |
553 | rq = blk_mq_rq_cache_fill(q, plug, opf, flags); | |
40467282 JC |
554 | if (!rq) |
555 | return NULL; | |
556 | } else { | |
557 | rq = rq_list_peek(&plug->cached_rq); | |
558 | if (!rq || rq->q != q) | |
559 | return NULL; | |
4b6a5d9c | 560 | |
40467282 JC |
561 | if (blk_mq_get_hctx_type(opf) != rq->mq_hctx->type) |
562 | return NULL; | |
563 | if (op_is_flush(rq->cmd_flags) != op_is_flush(opf)) | |
564 | return NULL; | |
565 | ||
566 | plug->cached_rq = rq_list_next(rq); | |
567 | } | |
4b6a5d9c | 568 | |
4b6a5d9c JA |
569 | rq->cmd_flags = opf; |
570 | INIT_LIST_HEAD(&rq->queuelist); | |
571 | return rq; | |
572 | } | |
573 | ||
574 | struct request *blk_mq_alloc_request(struct request_queue *q, blk_opf_t opf, | |
575 | blk_mq_req_flags_t flags) | |
576 | { | |
577 | struct request *rq; | |
578 | ||
579 | rq = blk_mq_alloc_cached_request(q, opf, flags); | |
580 | if (!rq) { | |
581 | struct blk_mq_alloc_data data = { | |
582 | .q = q, | |
583 | .flags = flags, | |
584 | .cmd_flags = opf, | |
585 | .nr_tags = 1, | |
586 | }; | |
587 | int ret; | |
588 | ||
589 | ret = blk_queue_enter(q, flags); | |
590 | if (ret) | |
591 | return ERR_PTR(ret); | |
592 | ||
593 | rq = __blk_mq_alloc_requests(&data); | |
594 | if (!rq) | |
595 | goto out_queue_exit; | |
596 | } | |
0c4de0f3 CH |
597 | rq->__data_len = 0; |
598 | rq->__sector = (sector_t) -1; | |
599 | rq->bio = rq->biotail = NULL; | |
320ae51f | 600 | return rq; |
a5ea5811 CH |
601 | out_queue_exit: |
602 | blk_queue_exit(q); | |
603 | return ERR_PTR(-EWOULDBLOCK); | |
320ae51f | 604 | } |
4bb659b1 | 605 | EXPORT_SYMBOL(blk_mq_alloc_request); |
320ae51f | 606 | |
cd6ce148 | 607 | struct request *blk_mq_alloc_request_hctx(struct request_queue *q, |
16458cf3 | 608 | blk_opf_t opf, blk_mq_req_flags_t flags, unsigned int hctx_idx) |
1f5bd336 | 609 | { |
e6e7abff CH |
610 | struct blk_mq_alloc_data data = { |
611 | .q = q, | |
612 | .flags = flags, | |
16458cf3 | 613 | .cmd_flags = opf, |
47c122e3 | 614 | .nr_tags = 1, |
e6e7abff | 615 | }; |
600c3b0c | 616 | u64 alloc_time_ns = 0; |
e3c5a78c | 617 | struct request *rq; |
6d2809d5 | 618 | unsigned int cpu; |
600c3b0c | 619 | unsigned int tag; |
1f5bd336 ML |
620 | int ret; |
621 | ||
600c3b0c CH |
622 | /* alloc_time includes depth and tag waits */ |
623 | if (blk_queue_rq_alloc_time(q)) | |
624 | alloc_time_ns = ktime_get_ns(); | |
625 | ||
1f5bd336 ML |
626 | /* |
627 | * If the tag allocator sleeps we could get an allocation for a | |
628 | * different hardware context. No need to complicate the low level | |
629 | * allocator for this for the rare use case of a command tied to | |
630 | * a specific queue. | |
631 | */ | |
6ee858a3 KS |
632 | if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT)) || |
633 | WARN_ON_ONCE(!(flags & BLK_MQ_REQ_RESERVED))) | |
1f5bd336 ML |
634 | return ERR_PTR(-EINVAL); |
635 | ||
636 | if (hctx_idx >= q->nr_hw_queues) | |
637 | return ERR_PTR(-EIO); | |
638 | ||
3a0a5299 | 639 | ret = blk_queue_enter(q, flags); |
1f5bd336 ML |
640 | if (ret) |
641 | return ERR_PTR(ret); | |
642 | ||
c8712c6a CH |
643 | /* |
644 | * Check if the hardware context is actually mapped to anything. | |
645 | * If not tell the caller that it should skip this queue. | |
646 | */ | |
a5ea5811 | 647 | ret = -EXDEV; |
4e5cc99e | 648 | data.hctx = xa_load(&q->hctx_table, hctx_idx); |
e6e7abff | 649 | if (!blk_mq_hw_queue_mapped(data.hctx)) |
a5ea5811 | 650 | goto out_queue_exit; |
e6e7abff | 651 | cpu = cpumask_first_and(data.hctx->cpumask, cpu_online_mask); |
14dc7a18 BVA |
652 | if (cpu >= nr_cpu_ids) |
653 | goto out_queue_exit; | |
e6e7abff | 654 | data.ctx = __blk_mq_get_ctx(q, cpu); |
1f5bd336 | 655 | |
dd6216bb CH |
656 | if (q->elevator) |
657 | data.rq_flags |= RQF_SCHED_TAGS; | |
781dd830 | 658 | else |
dd6216bb | 659 | blk_mq_tag_busy(data.hctx); |
600c3b0c | 660 | |
99e48cd6 JG |
661 | if (flags & BLK_MQ_REQ_RESERVED) |
662 | data.rq_flags |= RQF_RESV; | |
663 | ||
a5ea5811 | 664 | ret = -EWOULDBLOCK; |
600c3b0c CH |
665 | tag = blk_mq_get_tag(&data); |
666 | if (tag == BLK_MQ_NO_TAG) | |
a5ea5811 | 667 | goto out_queue_exit; |
e3c5a78c | 668 | rq = blk_mq_rq_ctx_init(&data, blk_mq_tags_from_data(&data), tag, |
fe6134f6 | 669 | alloc_time_ns); |
e3c5a78c JG |
670 | rq->__data_len = 0; |
671 | rq->__sector = (sector_t) -1; | |
672 | rq->bio = rq->biotail = NULL; | |
673 | return rq; | |
600c3b0c | 674 | |
a5ea5811 CH |
675 | out_queue_exit: |
676 | blk_queue_exit(q); | |
677 | return ERR_PTR(ret); | |
1f5bd336 ML |
678 | } |
679 | EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx); | |
680 | ||
12f5b931 KB |
681 | static void __blk_mq_free_request(struct request *rq) |
682 | { | |
683 | struct request_queue *q = rq->q; | |
684 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
ea4f995e | 685 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; |
12f5b931 KB |
686 | const int sched_tag = rq->internal_tag; |
687 | ||
a892c8d5 | 688 | blk_crypto_free_request(rq); |
986d413b | 689 | blk_pm_mark_last_busy(rq); |
ea4f995e | 690 | rq->mq_hctx = NULL; |
76647368 | 691 | if (rq->tag != BLK_MQ_NO_TAG) |
cae740a0 | 692 | blk_mq_put_tag(hctx->tags, ctx, rq->tag); |
76647368 | 693 | if (sched_tag != BLK_MQ_NO_TAG) |
cae740a0 | 694 | blk_mq_put_tag(hctx->sched_tags, ctx, sched_tag); |
12f5b931 KB |
695 | blk_mq_sched_restart(hctx); |
696 | blk_queue_exit(q); | |
697 | } | |
698 | ||
6af54051 | 699 | void blk_mq_free_request(struct request *rq) |
320ae51f | 700 | { |
320ae51f | 701 | struct request_queue *q = rq->q; |
ea4f995e | 702 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; |
6af54051 | 703 | |
dd6216bb | 704 | if ((rq->rq_flags & RQF_USE_SCHED) && |
222ee581 CH |
705 | q->elevator->type->ops.finish_request) |
706 | q->elevator->type->ops.finish_request(rq); | |
320ae51f | 707 | |
e8064021 | 708 | if (rq->rq_flags & RQF_MQ_INFLIGHT) |
bccf5e26 | 709 | __blk_mq_dec_active_requests(hctx); |
87760e5e | 710 | |
7beb2f84 | 711 | if (unlikely(laptop_mode && !blk_rq_is_passthrough(rq))) |
d152c682 | 712 | laptop_io_completion(q->disk->bdi); |
7beb2f84 | 713 | |
a7905043 | 714 | rq_qos_done(q, rq); |
0d2602ca | 715 | |
12f5b931 | 716 | WRITE_ONCE(rq->state, MQ_RQ_IDLE); |
0a467d0f | 717 | if (req_ref_put_and_test(rq)) |
12f5b931 | 718 | __blk_mq_free_request(rq); |
320ae51f | 719 | } |
1a3b595a | 720 | EXPORT_SYMBOL_GPL(blk_mq_free_request); |
320ae51f | 721 | |
47c122e3 | 722 | void blk_mq_free_plug_rqs(struct blk_plug *plug) |
320ae51f | 723 | { |
013a7f95 | 724 | struct request *rq; |
fe1f4526 | 725 | |
c5fc7b93 | 726 | while ((rq = rq_list_pop(&plug->cached_rq)) != NULL) |
47c122e3 | 727 | blk_mq_free_request(rq); |
47c122e3 | 728 | } |
522a7775 | 729 | |
22350ad7 CH |
730 | void blk_dump_rq_flags(struct request *rq, char *msg) |
731 | { | |
732 | printk(KERN_INFO "%s: dev %s: flags=%llx\n", msg, | |
f3fa33ac | 733 | rq->q->disk ? rq->q->disk->disk_name : "?", |
16458cf3 | 734 | (__force unsigned long long) rq->cmd_flags); |
22350ad7 CH |
735 | |
736 | printk(KERN_INFO " sector %llu, nr/cnr %u/%u\n", | |
737 | (unsigned long long)blk_rq_pos(rq), | |
738 | blk_rq_sectors(rq), blk_rq_cur_sectors(rq)); | |
739 | printk(KERN_INFO " bio %p, biotail %p, len %u\n", | |
740 | rq->bio, rq->biotail, blk_rq_bytes(rq)); | |
741 | } | |
742 | EXPORT_SYMBOL(blk_dump_rq_flags); | |
743 | ||
9be3e06f JA |
744 | static void req_bio_endio(struct request *rq, struct bio *bio, |
745 | unsigned int nbytes, blk_status_t error) | |
746 | { | |
478eb72b | 747 | if (unlikely(error)) { |
9be3e06f | 748 | bio->bi_status = error; |
478eb72b | 749 | } else if (req_op(rq) == REQ_OP_ZONE_APPEND) { |
9be3e06f JA |
750 | /* |
751 | * Partial zone append completions cannot be supported as the | |
752 | * BIO fragments may end up not being written sequentially. | |
753 | */ | |
297db731 | 754 | if (bio->bi_iter.bi_size != nbytes) |
9be3e06f JA |
755 | bio->bi_status = BLK_STS_IOERR; |
756 | else | |
757 | bio->bi_iter.bi_sector = rq->__sector; | |
758 | } | |
759 | ||
478eb72b PB |
760 | bio_advance(bio, nbytes); |
761 | ||
762 | if (unlikely(rq->rq_flags & RQF_QUIET)) | |
763 | bio_set_flag(bio, BIO_QUIET); | |
9be3e06f JA |
764 | /* don't actually finish bio if it's part of flush sequence */ |
765 | if (bio->bi_iter.bi_size == 0 && !(rq->rq_flags & RQF_FLUSH_SEQ)) | |
766 | bio_endio(bio); | |
767 | } | |
768 | ||
769 | static void blk_account_io_completion(struct request *req, unsigned int bytes) | |
770 | { | |
771 | if (req->part && blk_do_io_stat(req)) { | |
772 | const int sgrp = op_stat_group(req_op(req)); | |
773 | ||
774 | part_stat_lock(); | |
775 | part_stat_add(req->part, sectors[sgrp], bytes >> 9); | |
776 | part_stat_unlock(); | |
777 | } | |
778 | } | |
779 | ||
0d7a29a2 CH |
780 | static void blk_print_req_error(struct request *req, blk_status_t status) |
781 | { | |
782 | printk_ratelimited(KERN_ERR | |
783 | "%s error, dev %s, sector %llu op 0x%x:(%s) flags 0x%x " | |
784 | "phys_seg %u prio class %u\n", | |
785 | blk_status_to_str(status), | |
f3fa33ac | 786 | req->q->disk ? req->q->disk->disk_name : "?", |
16458cf3 BVA |
787 | blk_rq_pos(req), (__force u32)req_op(req), |
788 | blk_op_str(req_op(req)), | |
789 | (__force u32)(req->cmd_flags & ~REQ_OP_MASK), | |
0d7a29a2 CH |
790 | req->nr_phys_segments, |
791 | IOPRIO_PRIO_CLASS(req->ioprio)); | |
792 | } | |
793 | ||
5581a5dd JA |
794 | /* |
795 | * Fully end IO on a request. Does not support partial completions, or | |
796 | * errors. | |
797 | */ | |
798 | static void blk_complete_request(struct request *req) | |
799 | { | |
800 | const bool is_flush = (req->rq_flags & RQF_FLUSH_SEQ) != 0; | |
801 | int total_bytes = blk_rq_bytes(req); | |
802 | struct bio *bio = req->bio; | |
803 | ||
804 | trace_block_rq_complete(req, BLK_STS_OK, total_bytes); | |
805 | ||
806 | if (!bio) | |
807 | return; | |
808 | ||
809 | #ifdef CONFIG_BLK_DEV_INTEGRITY | |
810 | if (blk_integrity_rq(req) && req_op(req) == REQ_OP_READ) | |
811 | req->q->integrity.profile->complete_fn(req, total_bytes); | |
812 | #endif | |
813 | ||
9cd1e566 EB |
814 | /* |
815 | * Upper layers may call blk_crypto_evict_key() anytime after the last | |
816 | * bio_endio(). Therefore, the keyslot must be released before that. | |
817 | */ | |
818 | blk_crypto_rq_put_keyslot(req); | |
819 | ||
5581a5dd JA |
820 | blk_account_io_completion(req, total_bytes); |
821 | ||
822 | do { | |
823 | struct bio *next = bio->bi_next; | |
824 | ||
825 | /* Completion has already been traced */ | |
826 | bio_clear_flag(bio, BIO_TRACE_COMPLETION); | |
a12821d5 PR |
827 | |
828 | if (req_op(req) == REQ_OP_ZONE_APPEND) | |
829 | bio->bi_iter.bi_sector = req->__sector; | |
830 | ||
5581a5dd JA |
831 | if (!is_flush) |
832 | bio_endio(bio); | |
833 | bio = next; | |
834 | } while (bio); | |
835 | ||
836 | /* | |
837 | * Reset counters so that the request stacking driver | |
838 | * can find how many bytes remain in the request | |
839 | * later. | |
840 | */ | |
ab3e1d3b JA |
841 | if (!req->end_io) { |
842 | req->bio = NULL; | |
843 | req->__data_len = 0; | |
844 | } | |
5581a5dd JA |
845 | } |
846 | ||
9be3e06f JA |
847 | /** |
848 | * blk_update_request - Complete multiple bytes without completing the request | |
849 | * @req: the request being processed | |
850 | * @error: block status code | |
851 | * @nr_bytes: number of bytes to complete for @req | |
852 | * | |
853 | * Description: | |
854 | * Ends I/O on a number of bytes attached to @req, but doesn't complete | |
855 | * the request structure even if @req doesn't have leftover. | |
856 | * If @req has leftover, sets it up for the next range of segments. | |
857 | * | |
858 | * Passing the result of blk_rq_bytes() as @nr_bytes guarantees | |
859 | * %false return from this function. | |
860 | * | |
861 | * Note: | |
862 | * The RQF_SPECIAL_PAYLOAD flag is ignored on purpose in this function | |
863 | * except in the consistency check at the end of this function. | |
864 | * | |
865 | * Return: | |
866 | * %false - this request doesn't have any more data | |
867 | * %true - this request has more data | |
868 | **/ | |
869 | bool blk_update_request(struct request *req, blk_status_t error, | |
870 | unsigned int nr_bytes) | |
871 | { | |
872 | int total_bytes; | |
873 | ||
8a7d267b | 874 | trace_block_rq_complete(req, error, nr_bytes); |
9be3e06f JA |
875 | |
876 | if (!req->bio) | |
877 | return false; | |
878 | ||
879 | #ifdef CONFIG_BLK_DEV_INTEGRITY | |
880 | if (blk_integrity_rq(req) && req_op(req) == REQ_OP_READ && | |
881 | error == BLK_STS_OK) | |
882 | req->q->integrity.profile->complete_fn(req, nr_bytes); | |
883 | #endif | |
884 | ||
9cd1e566 EB |
885 | /* |
886 | * Upper layers may call blk_crypto_evict_key() anytime after the last | |
887 | * bio_endio(). Therefore, the keyslot must be released before that. | |
888 | */ | |
889 | if (blk_crypto_rq_has_keyslot(req) && nr_bytes >= blk_rq_bytes(req)) | |
890 | __blk_crypto_rq_put_keyslot(req); | |
891 | ||
9be3e06f | 892 | if (unlikely(error && !blk_rq_is_passthrough(req) && |
3d973a76 CH |
893 | !(req->rq_flags & RQF_QUIET)) && |
894 | !test_bit(GD_DEAD, &req->q->disk->state)) { | |
9be3e06f | 895 | blk_print_req_error(req, error); |
d5869fdc YS |
896 | trace_block_rq_error(req, error, nr_bytes); |
897 | } | |
9be3e06f JA |
898 | |
899 | blk_account_io_completion(req, nr_bytes); | |
900 | ||
901 | total_bytes = 0; | |
902 | while (req->bio) { | |
903 | struct bio *bio = req->bio; | |
904 | unsigned bio_bytes = min(bio->bi_iter.bi_size, nr_bytes); | |
905 | ||
906 | if (bio_bytes == bio->bi_iter.bi_size) | |
907 | req->bio = bio->bi_next; | |
908 | ||
909 | /* Completion has already been traced */ | |
910 | bio_clear_flag(bio, BIO_TRACE_COMPLETION); | |
911 | req_bio_endio(req, bio, bio_bytes, error); | |
912 | ||
913 | total_bytes += bio_bytes; | |
914 | nr_bytes -= bio_bytes; | |
915 | ||
916 | if (!nr_bytes) | |
917 | break; | |
918 | } | |
919 | ||
920 | /* | |
921 | * completely done | |
922 | */ | |
923 | if (!req->bio) { | |
924 | /* | |
925 | * Reset counters so that the request stacking driver | |
926 | * can find how many bytes remain in the request | |
927 | * later. | |
928 | */ | |
929 | req->__data_len = 0; | |
930 | return false; | |
931 | } | |
932 | ||
933 | req->__data_len -= total_bytes; | |
934 | ||
935 | /* update sector only for requests with clear definition of sector */ | |
936 | if (!blk_rq_is_passthrough(req)) | |
937 | req->__sector += total_bytes >> 9; | |
938 | ||
939 | /* mixed attributes always follow the first bio */ | |
940 | if (req->rq_flags & RQF_MIXED_MERGE) { | |
941 | req->cmd_flags &= ~REQ_FAILFAST_MASK; | |
942 | req->cmd_flags |= req->bio->bi_opf & REQ_FAILFAST_MASK; | |
943 | } | |
944 | ||
945 | if (!(req->rq_flags & RQF_SPECIAL_PAYLOAD)) { | |
946 | /* | |
947 | * If total number of sectors is less than the first segment | |
948 | * size, something has gone terribly wrong. | |
949 | */ | |
950 | if (blk_rq_bytes(req) < blk_rq_cur_bytes(req)) { | |
951 | blk_dump_rq_flags(req, "request botched"); | |
952 | req->__data_len = blk_rq_cur_bytes(req); | |
953 | } | |
954 | ||
955 | /* recalculate the number of segments */ | |
956 | req->nr_phys_segments = blk_recalc_rq_segments(req); | |
957 | } | |
958 | ||
959 | return true; | |
960 | } | |
961 | EXPORT_SYMBOL_GPL(blk_update_request); | |
962 | ||
450b7879 CH |
963 | static inline void blk_account_io_done(struct request *req, u64 now) |
964 | { | |
5a80bd07 HC |
965 | trace_block_io_done(req); |
966 | ||
450b7879 CH |
967 | /* |
968 | * Account IO completion. flush_rq isn't accounted as a | |
969 | * normal IO on queueing nor completion. Accounting the | |
970 | * containing request is enough. | |
971 | */ | |
972 | if (blk_do_io_stat(req) && req->part && | |
06965037 CK |
973 | !(req->rq_flags & RQF_FLUSH_SEQ)) { |
974 | const int sgrp = op_stat_group(req_op(req)); | |
450b7879 | 975 | |
06965037 CK |
976 | part_stat_lock(); |
977 | update_io_ticks(req->part, jiffies, true); | |
978 | part_stat_inc(req->part, ios[sgrp]); | |
979 | part_stat_add(req->part, nsecs[sgrp], now - req->start_time_ns); | |
980 | part_stat_unlock(); | |
981 | } | |
450b7879 CH |
982 | } |
983 | ||
984 | static inline void blk_account_io_start(struct request *req) | |
985 | { | |
5a80bd07 HC |
986 | trace_block_io_start(req); |
987 | ||
e165fb4d CK |
988 | if (blk_do_io_stat(req)) { |
989 | /* | |
990 | * All non-passthrough requests are created from a bio with one | |
991 | * exception: when a flush command that is part of a flush sequence | |
992 | * generated by the state machine in blk-flush.c is cloned onto the | |
993 | * lower device by dm-multipath we can get here without a bio. | |
994 | */ | |
995 | if (req->bio) | |
996 | req->part = req->bio->bi_bdev; | |
997 | else | |
998 | req->part = req->q->disk->part0; | |
999 | ||
1000 | part_stat_lock(); | |
1001 | update_io_ticks(req->part, jiffies, false); | |
1002 | part_stat_unlock(); | |
1003 | } | |
450b7879 CH |
1004 | } |
1005 | ||
f794f335 | 1006 | static inline void __blk_mq_end_request_acct(struct request *rq, u64 now) |
320ae51f | 1007 | { |
54bdd67d | 1008 | if (rq->rq_flags & RQF_STATS) |
522a7775 | 1009 | blk_stat_add(rq, now); |
4bc6339a | 1010 | |
87890092 | 1011 | blk_mq_sched_completed_request(rq, now); |
522a7775 | 1012 | blk_account_io_done(rq, now); |
f794f335 | 1013 | } |
522a7775 | 1014 | |
f794f335 JA |
1015 | inline void __blk_mq_end_request(struct request *rq, blk_status_t error) |
1016 | { | |
1017 | if (blk_mq_need_time_stamp(rq)) | |
1018 | __blk_mq_end_request_acct(rq, ktime_get_ns()); | |
0d11e6ac | 1019 | |
91b63639 | 1020 | if (rq->end_io) { |
a7905043 | 1021 | rq_qos_done(rq->q, rq); |
de671d61 JA |
1022 | if (rq->end_io(rq, error) == RQ_END_IO_FREE) |
1023 | blk_mq_free_request(rq); | |
91b63639 | 1024 | } else { |
320ae51f | 1025 | blk_mq_free_request(rq); |
91b63639 | 1026 | } |
320ae51f | 1027 | } |
c8a446ad | 1028 | EXPORT_SYMBOL(__blk_mq_end_request); |
63151a44 | 1029 | |
2a842aca | 1030 | void blk_mq_end_request(struct request *rq, blk_status_t error) |
63151a44 CH |
1031 | { |
1032 | if (blk_update_request(rq, error, blk_rq_bytes(rq))) | |
1033 | BUG(); | |
c8a446ad | 1034 | __blk_mq_end_request(rq, error); |
63151a44 | 1035 | } |
c8a446ad | 1036 | EXPORT_SYMBOL(blk_mq_end_request); |
320ae51f | 1037 | |
f794f335 JA |
1038 | #define TAG_COMP_BATCH 32 |
1039 | ||
1040 | static inline void blk_mq_flush_tag_batch(struct blk_mq_hw_ctx *hctx, | |
1041 | int *tag_array, int nr_tags) | |
1042 | { | |
1043 | struct request_queue *q = hctx->queue; | |
1044 | ||
3b87c6ea ML |
1045 | /* |
1046 | * All requests should have been marked as RQF_MQ_INFLIGHT, so | |
1047 | * update hctx->nr_active in batch | |
1048 | */ | |
1049 | if (hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) | |
1050 | __blk_mq_sub_active_requests(hctx, nr_tags); | |
1051 | ||
f794f335 JA |
1052 | blk_mq_put_tags(hctx->tags, tag_array, nr_tags); |
1053 | percpu_ref_put_many(&q->q_usage_counter, nr_tags); | |
1054 | } | |
1055 | ||
1056 | void blk_mq_end_request_batch(struct io_comp_batch *iob) | |
1057 | { | |
1058 | int tags[TAG_COMP_BATCH], nr_tags = 0; | |
02f7eab0 | 1059 | struct blk_mq_hw_ctx *cur_hctx = NULL; |
f794f335 JA |
1060 | struct request *rq; |
1061 | u64 now = 0; | |
1062 | ||
1063 | if (iob->need_ts) | |
1064 | now = ktime_get_ns(); | |
1065 | ||
1066 | while ((rq = rq_list_pop(&iob->req_list)) != NULL) { | |
1067 | prefetch(rq->bio); | |
1068 | prefetch(rq->rq_next); | |
1069 | ||
5581a5dd | 1070 | blk_complete_request(rq); |
f794f335 JA |
1071 | if (iob->need_ts) |
1072 | __blk_mq_end_request_acct(rq, now); | |
1073 | ||
98b26a0e JA |
1074 | rq_qos_done(rq->q, rq); |
1075 | ||
ab3e1d3b JA |
1076 | /* |
1077 | * If end_io handler returns NONE, then it still has | |
1078 | * ownership of the request. | |
1079 | */ | |
1080 | if (rq->end_io && rq->end_io(rq, 0) == RQ_END_IO_NONE) | |
1081 | continue; | |
1082 | ||
f794f335 | 1083 | WRITE_ONCE(rq->state, MQ_RQ_IDLE); |
0a467d0f | 1084 | if (!req_ref_put_and_test(rq)) |
f794f335 JA |
1085 | continue; |
1086 | ||
1087 | blk_crypto_free_request(rq); | |
1088 | blk_pm_mark_last_busy(rq); | |
f794f335 | 1089 | |
02f7eab0 JA |
1090 | if (nr_tags == TAG_COMP_BATCH || cur_hctx != rq->mq_hctx) { |
1091 | if (cur_hctx) | |
1092 | blk_mq_flush_tag_batch(cur_hctx, tags, nr_tags); | |
f794f335 | 1093 | nr_tags = 0; |
02f7eab0 | 1094 | cur_hctx = rq->mq_hctx; |
f794f335 JA |
1095 | } |
1096 | tags[nr_tags++] = rq->tag; | |
f794f335 JA |
1097 | } |
1098 | ||
1099 | if (nr_tags) | |
02f7eab0 | 1100 | blk_mq_flush_tag_batch(cur_hctx, tags, nr_tags); |
f794f335 JA |
1101 | } |
1102 | EXPORT_SYMBOL_GPL(blk_mq_end_request_batch); | |
1103 | ||
f9ab4918 | 1104 | static void blk_complete_reqs(struct llist_head *list) |
320ae51f | 1105 | { |
f9ab4918 SAS |
1106 | struct llist_node *entry = llist_reverse_order(llist_del_all(list)); |
1107 | struct request *rq, *next; | |
c3077b5d | 1108 | |
f9ab4918 | 1109 | llist_for_each_entry_safe(rq, next, entry, ipi_list) |
c3077b5d | 1110 | rq->q->mq_ops->complete(rq); |
320ae51f | 1111 | } |
320ae51f | 1112 | |
f9ab4918 | 1113 | static __latent_entropy void blk_done_softirq(struct softirq_action *h) |
320ae51f | 1114 | { |
f9ab4918 | 1115 | blk_complete_reqs(this_cpu_ptr(&blk_cpu_done)); |
115243f5 CH |
1116 | } |
1117 | ||
c3077b5d CH |
1118 | static int blk_softirq_cpu_dead(unsigned int cpu) |
1119 | { | |
f9ab4918 | 1120 | blk_complete_reqs(&per_cpu(blk_cpu_done, cpu)); |
c3077b5d CH |
1121 | return 0; |
1122 | } | |
1123 | ||
40d09b53 | 1124 | static void __blk_mq_complete_request_remote(void *data) |
c3077b5d | 1125 | { |
f9ab4918 | 1126 | __raise_softirq_irqoff(BLOCK_SOFTIRQ); |
c3077b5d CH |
1127 | } |
1128 | ||
96339526 CH |
1129 | static inline bool blk_mq_complete_need_ipi(struct request *rq) |
1130 | { | |
1131 | int cpu = raw_smp_processor_id(); | |
1132 | ||
1133 | if (!IS_ENABLED(CONFIG_SMP) || | |
1134 | !test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) | |
1135 | return false; | |
71425189 SAS |
1136 | /* |
1137 | * With force threaded interrupts enabled, raising softirq from an SMP | |
1138 | * function call will always result in waking the ksoftirqd thread. | |
1139 | * This is probably worse than completing the request on a different | |
1140 | * cache domain. | |
1141 | */ | |
91cc470e | 1142 | if (force_irqthreads()) |
71425189 | 1143 | return false; |
96339526 CH |
1144 | |
1145 | /* same CPU or cache domain? Complete locally */ | |
1146 | if (cpu == rq->mq_ctx->cpu || | |
1147 | (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags) && | |
1148 | cpus_share_cache(cpu, rq->mq_ctx->cpu))) | |
1149 | return false; | |
1150 | ||
1151 | /* don't try to IPI to an offline CPU */ | |
1152 | return cpu_online(rq->mq_ctx->cpu); | |
1153 | } | |
1154 | ||
f9ab4918 SAS |
1155 | static void blk_mq_complete_send_ipi(struct request *rq) |
1156 | { | |
1157 | struct llist_head *list; | |
1158 | unsigned int cpu; | |
1159 | ||
1160 | cpu = rq->mq_ctx->cpu; | |
1161 | list = &per_cpu(blk_cpu_done, cpu); | |
1162 | if (llist_add(&rq->ipi_list, list)) { | |
1163 | INIT_CSD(&rq->csd, __blk_mq_complete_request_remote, rq); | |
1164 | smp_call_function_single_async(cpu, &rq->csd); | |
1165 | } | |
1166 | } | |
1167 | ||
1168 | static void blk_mq_raise_softirq(struct request *rq) | |
1169 | { | |
1170 | struct llist_head *list; | |
1171 | ||
1172 | preempt_disable(); | |
1173 | list = this_cpu_ptr(&blk_cpu_done); | |
1174 | if (llist_add(&rq->ipi_list, list)) | |
1175 | raise_softirq(BLOCK_SOFTIRQ); | |
1176 | preempt_enable(); | |
1177 | } | |
1178 | ||
40d09b53 | 1179 | bool blk_mq_complete_request_remote(struct request *rq) |
320ae51f | 1180 | { |
af78ff7c | 1181 | WRITE_ONCE(rq->state, MQ_RQ_COMPLETE); |
36e76539 | 1182 | |
4ab32bf3 | 1183 | /* |
f168420c LS |
1184 | * For request which hctx has only one ctx mapping, |
1185 | * or a polled request, always complete locally, | |
1186 | * it's pointless to redirect the completion. | |
4ab32bf3 | 1187 | */ |
f168420c LS |
1188 | if (rq->mq_hctx->nr_ctx == 1 || |
1189 | rq->cmd_flags & REQ_POLLED) | |
40d09b53 | 1190 | return false; |
38535201 | 1191 | |
96339526 | 1192 | if (blk_mq_complete_need_ipi(rq)) { |
f9ab4918 SAS |
1193 | blk_mq_complete_send_ipi(rq); |
1194 | return true; | |
3d6efbf6 | 1195 | } |
40d09b53 | 1196 | |
f9ab4918 SAS |
1197 | if (rq->q->nr_hw_queues == 1) { |
1198 | blk_mq_raise_softirq(rq); | |
1199 | return true; | |
1200 | } | |
1201 | return false; | |
40d09b53 CH |
1202 | } |
1203 | EXPORT_SYMBOL_GPL(blk_mq_complete_request_remote); | |
1204 | ||
1205 | /** | |
1206 | * blk_mq_complete_request - end I/O on a request | |
1207 | * @rq: the request being processed | |
1208 | * | |
1209 | * Description: | |
1210 | * Complete a request by scheduling the ->complete_rq operation. | |
1211 | **/ | |
1212 | void blk_mq_complete_request(struct request *rq) | |
1213 | { | |
1214 | if (!blk_mq_complete_request_remote(rq)) | |
1215 | rq->q->mq_ops->complete(rq); | |
320ae51f | 1216 | } |
15f73f5b | 1217 | EXPORT_SYMBOL(blk_mq_complete_request); |
30a91cb4 | 1218 | |
105663f7 AA |
1219 | /** |
1220 | * blk_mq_start_request - Start processing a request | |
1221 | * @rq: Pointer to request to be started | |
1222 | * | |
1223 | * Function used by device drivers to notify the block layer that a request | |
1224 | * is going to be processed now, so blk layer can do proper initializations | |
1225 | * such as starting the timeout timer. | |
1226 | */ | |
e2490073 | 1227 | void blk_mq_start_request(struct request *rq) |
320ae51f JA |
1228 | { |
1229 | struct request_queue *q = rq->q; | |
1230 | ||
a54895fa | 1231 | trace_block_rq_issue(rq); |
320ae51f | 1232 | |
cf43e6be | 1233 | if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) { |
4cddeaca | 1234 | rq->io_start_time_ns = ktime_get_ns(); |
3d244306 | 1235 | rq->stats_sectors = blk_rq_sectors(rq); |
cf43e6be | 1236 | rq->rq_flags |= RQF_STATS; |
a7905043 | 1237 | rq_qos_issue(q, rq); |
cf43e6be JA |
1238 | } |
1239 | ||
1d9bd516 | 1240 | WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE); |
538b7534 | 1241 | |
1d9bd516 | 1242 | blk_add_timer(rq); |
12f5b931 | 1243 | WRITE_ONCE(rq->state, MQ_RQ_IN_FLIGHT); |
49f5baa5 | 1244 | |
54d4e6ab MG |
1245 | #ifdef CONFIG_BLK_DEV_INTEGRITY |
1246 | if (blk_integrity_rq(rq) && req_op(rq) == REQ_OP_WRITE) | |
1247 | q->integrity.profile->prepare_fn(rq); | |
1248 | #endif | |
3e08773c CH |
1249 | if (rq->bio && rq->bio->bi_opf & REQ_POLLED) |
1250 | WRITE_ONCE(rq->bio->bi_cookie, blk_rq_to_qc(rq)); | |
320ae51f | 1251 | } |
e2490073 | 1252 | EXPORT_SYMBOL(blk_mq_start_request); |
320ae51f | 1253 | |
a327c341 ML |
1254 | /* |
1255 | * Allow 2x BLK_MAX_REQUEST_COUNT requests on plug queue for multiple | |
1256 | * queues. This is important for md arrays to benefit from merging | |
1257 | * requests. | |
1258 | */ | |
1259 | static inline unsigned short blk_plug_max_rq_count(struct blk_plug *plug) | |
1260 | { | |
1261 | if (plug->multiple_queues) | |
1262 | return BLK_MAX_REQUEST_COUNT * 2; | |
1263 | return BLK_MAX_REQUEST_COUNT; | |
1264 | } | |
1265 | ||
1266 | static void blk_add_rq_to_plug(struct blk_plug *plug, struct request *rq) | |
1267 | { | |
1268 | struct request *last = rq_list_peek(&plug->mq_list); | |
1269 | ||
1270 | if (!plug->rq_count) { | |
1271 | trace_block_plug(rq->q); | |
1272 | } else if (plug->rq_count >= blk_plug_max_rq_count(plug) || | |
1273 | (!blk_queue_nomerges(rq->q) && | |
1274 | blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) { | |
1275 | blk_mq_flush_plug_list(plug, false); | |
878eb6e4 | 1276 | last = NULL; |
a327c341 ML |
1277 | trace_block_plug(rq->q); |
1278 | } | |
1279 | ||
1280 | if (!plug->multiple_queues && last && last->q != rq->q) | |
1281 | plug->multiple_queues = true; | |
dd6216bb | 1282 | if (!plug->has_elevator && (rq->rq_flags & RQF_USE_SCHED)) |
a327c341 ML |
1283 | plug->has_elevator = true; |
1284 | rq->rq_next = NULL; | |
1285 | rq_list_add(&plug->mq_list, rq); | |
1286 | plug->rq_count++; | |
1287 | } | |
1288 | ||
4054cff9 CH |
1289 | /** |
1290 | * blk_execute_rq_nowait - insert a request to I/O scheduler for execution | |
4054cff9 CH |
1291 | * @rq: request to insert |
1292 | * @at_head: insert request at head or tail of queue | |
4054cff9 CH |
1293 | * |
1294 | * Description: | |
1295 | * Insert a fully prepared request at the back of the I/O scheduler queue | |
1296 | * for execution. Don't wait for completion. | |
1297 | * | |
1298 | * Note: | |
1299 | * This function will invoke @done directly if the queue is dead. | |
1300 | */ | |
e2e53086 | 1301 | void blk_execute_rq_nowait(struct request *rq, bool at_head) |
4054cff9 | 1302 | { |
f0dbe6e8 CH |
1303 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; |
1304 | ||
ae948fd6 CH |
1305 | WARN_ON(irqs_disabled()); |
1306 | WARN_ON(!blk_rq_is_passthrough(rq)); | |
4054cff9 | 1307 | |
ae948fd6 | 1308 | blk_account_io_start(rq); |
110fdb44 PR |
1309 | |
1310 | /* | |
1311 | * As plugging can be enabled for passthrough requests on a zoned | |
1312 | * device, directly accessing the plug instead of using blk_mq_plug() | |
1313 | * should not have any consequences. | |
1314 | */ | |
f0dbe6e8 | 1315 | if (current->plug && !at_head) { |
ae948fd6 | 1316 | blk_add_rq_to_plug(current->plug, rq); |
f0dbe6e8 CH |
1317 | return; |
1318 | } | |
1319 | ||
710fa378 | 1320 | blk_mq_insert_request(rq, at_head ? BLK_MQ_INSERT_AT_HEAD : 0); |
f0dbe6e8 | 1321 | blk_mq_run_hw_queue(hctx, false); |
4054cff9 CH |
1322 | } |
1323 | EXPORT_SYMBOL_GPL(blk_execute_rq_nowait); | |
1324 | ||
32ac5a9b CH |
1325 | struct blk_rq_wait { |
1326 | struct completion done; | |
1327 | blk_status_t ret; | |
1328 | }; | |
1329 | ||
de671d61 | 1330 | static enum rq_end_io_ret blk_end_sync_rq(struct request *rq, blk_status_t ret) |
32ac5a9b CH |
1331 | { |
1332 | struct blk_rq_wait *wait = rq->end_io_data; | |
1333 | ||
1334 | wait->ret = ret; | |
1335 | complete(&wait->done); | |
de671d61 | 1336 | return RQ_END_IO_NONE; |
32ac5a9b CH |
1337 | } |
1338 | ||
c6e99ea4 | 1339 | bool blk_rq_is_poll(struct request *rq) |
4054cff9 CH |
1340 | { |
1341 | if (!rq->mq_hctx) | |
1342 | return false; | |
1343 | if (rq->mq_hctx->type != HCTX_TYPE_POLL) | |
1344 | return false; | |
4054cff9 CH |
1345 | return true; |
1346 | } | |
c6e99ea4 | 1347 | EXPORT_SYMBOL_GPL(blk_rq_is_poll); |
4054cff9 CH |
1348 | |
1349 | static void blk_rq_poll_completion(struct request *rq, struct completion *wait) | |
1350 | { | |
1351 | do { | |
38a8c4d1 | 1352 | blk_mq_poll(rq->q, blk_rq_to_qc(rq), NULL, 0); |
4054cff9 CH |
1353 | cond_resched(); |
1354 | } while (!completion_done(wait)); | |
1355 | } | |
1356 | ||
1357 | /** | |
1358 | * blk_execute_rq - insert a request into queue for execution | |
4054cff9 CH |
1359 | * @rq: request to insert |
1360 | * @at_head: insert request at head or tail of queue | |
1361 | * | |
1362 | * Description: | |
1363 | * Insert a fully prepared request at the back of the I/O scheduler queue | |
1364 | * for execution and wait for completion. | |
1365 | * Return: The blk_status_t result provided to blk_mq_end_request(). | |
1366 | */ | |
b84ba30b | 1367 | blk_status_t blk_execute_rq(struct request *rq, bool at_head) |
4054cff9 | 1368 | { |
f0dbe6e8 | 1369 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; |
32ac5a9b CH |
1370 | struct blk_rq_wait wait = { |
1371 | .done = COMPLETION_INITIALIZER_ONSTACK(wait.done), | |
1372 | }; | |
4054cff9 | 1373 | |
ae948fd6 CH |
1374 | WARN_ON(irqs_disabled()); |
1375 | WARN_ON(!blk_rq_is_passthrough(rq)); | |
4054cff9 CH |
1376 | |
1377 | rq->end_io_data = &wait; | |
ae948fd6 | 1378 | rq->end_io = blk_end_sync_rq; |
4054cff9 | 1379 | |
ae948fd6 | 1380 | blk_account_io_start(rq); |
710fa378 | 1381 | blk_mq_insert_request(rq, at_head ? BLK_MQ_INSERT_AT_HEAD : 0); |
f0dbe6e8 | 1382 | blk_mq_run_hw_queue(hctx, false); |
4054cff9 | 1383 | |
ae948fd6 | 1384 | if (blk_rq_is_poll(rq)) { |
32ac5a9b | 1385 | blk_rq_poll_completion(rq, &wait.done); |
ae948fd6 CH |
1386 | } else { |
1387 | /* | |
1388 | * Prevent hang_check timer from firing at us during very long | |
1389 | * I/O | |
1390 | */ | |
1391 | unsigned long hang_check = sysctl_hung_task_timeout_secs; | |
1392 | ||
1393 | if (hang_check) | |
32ac5a9b | 1394 | while (!wait_for_completion_io_timeout(&wait.done, |
ae948fd6 CH |
1395 | hang_check * (HZ/2))) |
1396 | ; | |
1397 | else | |
32ac5a9b | 1398 | wait_for_completion_io(&wait.done); |
ae948fd6 | 1399 | } |
4054cff9 | 1400 | |
32ac5a9b | 1401 | return wait.ret; |
4054cff9 CH |
1402 | } |
1403 | EXPORT_SYMBOL(blk_execute_rq); | |
1404 | ||
ed0791b2 | 1405 | static void __blk_mq_requeue_request(struct request *rq) |
320ae51f JA |
1406 | { |
1407 | struct request_queue *q = rq->q; | |
1408 | ||
923218f6 ML |
1409 | blk_mq_put_driver_tag(rq); |
1410 | ||
a54895fa | 1411 | trace_block_rq_requeue(rq); |
a7905043 | 1412 | rq_qos_requeue(q, rq); |
49f5baa5 | 1413 | |
12f5b931 KB |
1414 | if (blk_mq_request_started(rq)) { |
1415 | WRITE_ONCE(rq->state, MQ_RQ_IDLE); | |
da661267 | 1416 | rq->rq_flags &= ~RQF_TIMED_OUT; |
e2490073 | 1417 | } |
320ae51f JA |
1418 | } |
1419 | ||
2b053aca | 1420 | void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list) |
ed0791b2 | 1421 | { |
214a4418 | 1422 | struct request_queue *q = rq->q; |
9a67aa52 | 1423 | unsigned long flags; |
214a4418 | 1424 | |
ed0791b2 | 1425 | __blk_mq_requeue_request(rq); |
ed0791b2 | 1426 | |
105976f5 ML |
1427 | /* this request will be re-inserted to io scheduler queue */ |
1428 | blk_mq_sched_requeue_request(rq); | |
1429 | ||
9a67aa52 CH |
1430 | spin_lock_irqsave(&q->requeue_lock, flags); |
1431 | list_add_tail(&rq->queuelist, &q->requeue_list); | |
1432 | spin_unlock_irqrestore(&q->requeue_lock, flags); | |
214a4418 CH |
1433 | |
1434 | if (kick_requeue_list) | |
1435 | blk_mq_kick_requeue_list(q); | |
ed0791b2 CH |
1436 | } |
1437 | EXPORT_SYMBOL(blk_mq_requeue_request); | |
1438 | ||
6fca6a61 CH |
1439 | static void blk_mq_requeue_work(struct work_struct *work) |
1440 | { | |
1441 | struct request_queue *q = | |
2849450a | 1442 | container_of(work, struct request_queue, requeue_work.work); |
6fca6a61 | 1443 | LIST_HEAD(rq_list); |
9a67aa52 CH |
1444 | LIST_HEAD(flush_list); |
1445 | struct request *rq; | |
6fca6a61 | 1446 | |
18e9781d | 1447 | spin_lock_irq(&q->requeue_lock); |
6fca6a61 | 1448 | list_splice_init(&q->requeue_list, &rq_list); |
9a67aa52 | 1449 | list_splice_init(&q->flush_list, &flush_list); |
18e9781d | 1450 | spin_unlock_irq(&q->requeue_lock); |
6fca6a61 | 1451 | |
9a67aa52 CH |
1452 | while (!list_empty(&rq_list)) { |
1453 | rq = list_entry(rq_list.next, struct request, queuelist); | |
aef1897c | 1454 | /* |
a1e948b8 CH |
1455 | * If RQF_DONTPREP ist set, the request has been started by the |
1456 | * driver already and might have driver-specific data allocated | |
1457 | * already. Insert it into the hctx dispatch list to avoid | |
1458 | * block layer merges for the request. | |
aef1897c | 1459 | */ |
a1e948b8 | 1460 | if (rq->rq_flags & RQF_DONTPREP) { |
a1e948b8 | 1461 | list_del_init(&rq->queuelist); |
2b597613 | 1462 | blk_mq_request_bypass_insert(rq, 0); |
9a67aa52 | 1463 | } else { |
a1e948b8 | 1464 | list_del_init(&rq->queuelist); |
710fa378 | 1465 | blk_mq_insert_request(rq, BLK_MQ_INSERT_AT_HEAD); |
a1e948b8 | 1466 | } |
6fca6a61 CH |
1467 | } |
1468 | ||
9a67aa52 CH |
1469 | while (!list_empty(&flush_list)) { |
1470 | rq = list_entry(flush_list.next, struct request, queuelist); | |
6fca6a61 | 1471 | list_del_init(&rq->queuelist); |
710fa378 | 1472 | blk_mq_insert_request(rq, 0); |
6fca6a61 CH |
1473 | } |
1474 | ||
52d7f1b5 | 1475 | blk_mq_run_hw_queues(q, false); |
6fca6a61 CH |
1476 | } |
1477 | ||
6fca6a61 CH |
1478 | void blk_mq_kick_requeue_list(struct request_queue *q) |
1479 | { | |
ae943d20 | 1480 | kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0); |
6fca6a61 CH |
1481 | } |
1482 | EXPORT_SYMBOL(blk_mq_kick_requeue_list); | |
1483 | ||
2849450a MS |
1484 | void blk_mq_delay_kick_requeue_list(struct request_queue *q, |
1485 | unsigned long msecs) | |
1486 | { | |
d4acf365 BVA |
1487 | kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, |
1488 | msecs_to_jiffies(msecs)); | |
2849450a MS |
1489 | } |
1490 | EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list); | |
1491 | ||
2dd6532e | 1492 | static bool blk_mq_rq_inflight(struct request *rq, void *priv) |
ae879912 JA |
1493 | { |
1494 | /* | |
8ab30a33 JG |
1495 | * If we find a request that isn't idle we know the queue is busy |
1496 | * as it's checked in the iter. | |
1497 | * Return false to stop the iteration. | |
ae879912 | 1498 | */ |
8ab30a33 | 1499 | if (blk_mq_request_started(rq)) { |
ae879912 JA |
1500 | bool *busy = priv; |
1501 | ||
1502 | *busy = true; | |
1503 | return false; | |
1504 | } | |
1505 | ||
1506 | return true; | |
1507 | } | |
1508 | ||
3c94d83c | 1509 | bool blk_mq_queue_inflight(struct request_queue *q) |
ae879912 JA |
1510 | { |
1511 | bool busy = false; | |
1512 | ||
3c94d83c | 1513 | blk_mq_queue_tag_busy_iter(q, blk_mq_rq_inflight, &busy); |
ae879912 JA |
1514 | return busy; |
1515 | } | |
3c94d83c | 1516 | EXPORT_SYMBOL_GPL(blk_mq_queue_inflight); |
ae879912 | 1517 | |
9bdb4833 | 1518 | static void blk_mq_rq_timed_out(struct request *req) |
320ae51f | 1519 | { |
da661267 | 1520 | req->rq_flags |= RQF_TIMED_OUT; |
d1210d5a CH |
1521 | if (req->q->mq_ops->timeout) { |
1522 | enum blk_eh_timer_return ret; | |
1523 | ||
9bdb4833 | 1524 | ret = req->q->mq_ops->timeout(req); |
d1210d5a CH |
1525 | if (ret == BLK_EH_DONE) |
1526 | return; | |
1527 | WARN_ON_ONCE(ret != BLK_EH_RESET_TIMER); | |
46f92d42 | 1528 | } |
d1210d5a CH |
1529 | |
1530 | blk_add_timer(req); | |
87ee7b11 | 1531 | } |
5b3f25fc | 1532 | |
82c22947 DJ |
1533 | struct blk_expired_data { |
1534 | bool has_timedout_rq; | |
1535 | unsigned long next; | |
1536 | unsigned long timeout_start; | |
1537 | }; | |
1538 | ||
1539 | static bool blk_mq_req_expired(struct request *rq, struct blk_expired_data *expired) | |
81481eb4 | 1540 | { |
12f5b931 | 1541 | unsigned long deadline; |
87ee7b11 | 1542 | |
12f5b931 KB |
1543 | if (blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT) |
1544 | return false; | |
da661267 CH |
1545 | if (rq->rq_flags & RQF_TIMED_OUT) |
1546 | return false; | |
a7af0af3 | 1547 | |
079076b3 | 1548 | deadline = READ_ONCE(rq->deadline); |
82c22947 | 1549 | if (time_after_eq(expired->timeout_start, deadline)) |
12f5b931 | 1550 | return true; |
a7af0af3 | 1551 | |
82c22947 DJ |
1552 | if (expired->next == 0) |
1553 | expired->next = deadline; | |
1554 | else if (time_after(expired->next, deadline)) | |
1555 | expired->next = deadline; | |
12f5b931 | 1556 | return false; |
87ee7b11 JA |
1557 | } |
1558 | ||
2e315dc0 ML |
1559 | void blk_mq_put_rq_ref(struct request *rq) |
1560 | { | |
de671d61 JA |
1561 | if (is_flush_rq(rq)) { |
1562 | if (rq->end_io(rq, 0) == RQ_END_IO_FREE) | |
1563 | blk_mq_free_request(rq); | |
1564 | } else if (req_ref_put_and_test(rq)) { | |
2e315dc0 | 1565 | __blk_mq_free_request(rq); |
de671d61 | 1566 | } |
2e315dc0 ML |
1567 | } |
1568 | ||
2dd6532e | 1569 | static bool blk_mq_check_expired(struct request *rq, void *priv) |
1d9bd516 | 1570 | { |
82c22947 | 1571 | struct blk_expired_data *expired = priv; |
12f5b931 KB |
1572 | |
1573 | /* | |
c797b40c ML |
1574 | * blk_mq_queue_tag_busy_iter() has locked the request, so it cannot |
1575 | * be reallocated underneath the timeout handler's processing, then | |
1576 | * the expire check is reliable. If the request is not expired, then | |
1577 | * it was completed and reallocated as a new request after returning | |
1578 | * from blk_mq_check_expired(). | |
1d9bd516 | 1579 | */ |
82c22947 DJ |
1580 | if (blk_mq_req_expired(rq, expired)) { |
1581 | expired->has_timedout_rq = true; | |
1582 | return false; | |
1583 | } | |
1584 | return true; | |
1585 | } | |
1586 | ||
1587 | static bool blk_mq_handle_expired(struct request *rq, void *priv) | |
1588 | { | |
1589 | struct blk_expired_data *expired = priv; | |
1590 | ||
1591 | if (blk_mq_req_expired(rq, expired)) | |
9bdb4833 | 1592 | blk_mq_rq_timed_out(rq); |
7baa8572 | 1593 | return true; |
1d9bd516 TH |
1594 | } |
1595 | ||
287922eb | 1596 | static void blk_mq_timeout_work(struct work_struct *work) |
320ae51f | 1597 | { |
287922eb CH |
1598 | struct request_queue *q = |
1599 | container_of(work, struct request_queue, timeout_work); | |
82c22947 DJ |
1600 | struct blk_expired_data expired = { |
1601 | .timeout_start = jiffies, | |
1602 | }; | |
1d9bd516 | 1603 | struct blk_mq_hw_ctx *hctx; |
4f481208 | 1604 | unsigned long i; |
320ae51f | 1605 | |
71f79fb3 GKB |
1606 | /* A deadlock might occur if a request is stuck requiring a |
1607 | * timeout at the same time a queue freeze is waiting | |
1608 | * completion, since the timeout code would not be able to | |
1609 | * acquire the queue reference here. | |
1610 | * | |
1611 | * That's why we don't use blk_queue_enter here; instead, we use | |
1612 | * percpu_ref_tryget directly, because we need to be able to | |
1613 | * obtain a reference even in the short window between the queue | |
1614 | * starting to freeze, by dropping the first reference in | |
1671d522 | 1615 | * blk_freeze_queue_start, and the moment the last request is |
71f79fb3 GKB |
1616 | * consumed, marked by the instant q_usage_counter reaches |
1617 | * zero. | |
1618 | */ | |
1619 | if (!percpu_ref_tryget(&q->q_usage_counter)) | |
287922eb CH |
1620 | return; |
1621 | ||
82c22947 DJ |
1622 | /* check if there is any timed-out request */ |
1623 | blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &expired); | |
1624 | if (expired.has_timedout_rq) { | |
1625 | /* | |
1626 | * Before walking tags, we must ensure any submit started | |
1627 | * before the current time has finished. Since the submit | |
1628 | * uses srcu or rcu, wait for a synchronization point to | |
1629 | * ensure all running submits have finished | |
1630 | */ | |
483239c7 | 1631 | blk_mq_wait_quiesce_done(q->tag_set); |
82c22947 DJ |
1632 | |
1633 | expired.next = 0; | |
1634 | blk_mq_queue_tag_busy_iter(q, blk_mq_handle_expired, &expired); | |
1635 | } | |
320ae51f | 1636 | |
82c22947 DJ |
1637 | if (expired.next != 0) { |
1638 | mod_timer(&q->timeout, expired.next); | |
0d2602ca | 1639 | } else { |
fcd36c36 BVA |
1640 | /* |
1641 | * Request timeouts are handled as a forward rolling timer. If | |
1642 | * we end up here it means that no requests are pending and | |
1643 | * also that no request has been pending for a while. Mark | |
1644 | * each hctx as idle. | |
1645 | */ | |
f054b56c ML |
1646 | queue_for_each_hw_ctx(q, hctx, i) { |
1647 | /* the hctx may be unmapped, so check it here */ | |
1648 | if (blk_mq_hw_queue_mapped(hctx)) | |
1649 | blk_mq_tag_idle(hctx); | |
1650 | } | |
0d2602ca | 1651 | } |
287922eb | 1652 | blk_queue_exit(q); |
320ae51f JA |
1653 | } |
1654 | ||
88459642 OS |
1655 | struct flush_busy_ctx_data { |
1656 | struct blk_mq_hw_ctx *hctx; | |
1657 | struct list_head *list; | |
1658 | }; | |
1659 | ||
1660 | static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data) | |
1661 | { | |
1662 | struct flush_busy_ctx_data *flush_data = data; | |
1663 | struct blk_mq_hw_ctx *hctx = flush_data->hctx; | |
1664 | struct blk_mq_ctx *ctx = hctx->ctxs[bitnr]; | |
c16d6b5a | 1665 | enum hctx_type type = hctx->type; |
88459642 | 1666 | |
88459642 | 1667 | spin_lock(&ctx->lock); |
c16d6b5a | 1668 | list_splice_tail_init(&ctx->rq_lists[type], flush_data->list); |
e9a99a63 | 1669 | sbitmap_clear_bit(sb, bitnr); |
88459642 OS |
1670 | spin_unlock(&ctx->lock); |
1671 | return true; | |
1672 | } | |
1673 | ||
1429d7c9 JA |
1674 | /* |
1675 | * Process software queues that have been marked busy, splicing them | |
1676 | * to the for-dispatch | |
1677 | */ | |
2c3ad667 | 1678 | void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list) |
1429d7c9 | 1679 | { |
88459642 OS |
1680 | struct flush_busy_ctx_data data = { |
1681 | .hctx = hctx, | |
1682 | .list = list, | |
1683 | }; | |
1429d7c9 | 1684 | |
88459642 | 1685 | sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data); |
1429d7c9 | 1686 | } |
2c3ad667 | 1687 | EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs); |
1429d7c9 | 1688 | |
b347689f ML |
1689 | struct dispatch_rq_data { |
1690 | struct blk_mq_hw_ctx *hctx; | |
1691 | struct request *rq; | |
1692 | }; | |
1693 | ||
1694 | static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr, | |
1695 | void *data) | |
1696 | { | |
1697 | struct dispatch_rq_data *dispatch_data = data; | |
1698 | struct blk_mq_hw_ctx *hctx = dispatch_data->hctx; | |
1699 | struct blk_mq_ctx *ctx = hctx->ctxs[bitnr]; | |
c16d6b5a | 1700 | enum hctx_type type = hctx->type; |
b347689f ML |
1701 | |
1702 | spin_lock(&ctx->lock); | |
c16d6b5a ML |
1703 | if (!list_empty(&ctx->rq_lists[type])) { |
1704 | dispatch_data->rq = list_entry_rq(ctx->rq_lists[type].next); | |
b347689f | 1705 | list_del_init(&dispatch_data->rq->queuelist); |
c16d6b5a | 1706 | if (list_empty(&ctx->rq_lists[type])) |
b347689f ML |
1707 | sbitmap_clear_bit(sb, bitnr); |
1708 | } | |
1709 | spin_unlock(&ctx->lock); | |
1710 | ||
1711 | return !dispatch_data->rq; | |
1712 | } | |
1713 | ||
1714 | struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx, | |
1715 | struct blk_mq_ctx *start) | |
1716 | { | |
f31967f0 | 1717 | unsigned off = start ? start->index_hw[hctx->type] : 0; |
b347689f ML |
1718 | struct dispatch_rq_data data = { |
1719 | .hctx = hctx, | |
1720 | .rq = NULL, | |
1721 | }; | |
1722 | ||
1723 | __sbitmap_for_each_set(&hctx->ctx_map, off, | |
1724 | dispatch_rq_from_ctx, &data); | |
1725 | ||
1726 | return data.rq; | |
1727 | } | |
1728 | ||
a808a9d5 | 1729 | static bool __blk_mq_alloc_driver_tag(struct request *rq) |
570e9b73 | 1730 | { |
ae0f1a73 | 1731 | struct sbitmap_queue *bt = &rq->mq_hctx->tags->bitmap_tags; |
570e9b73 | 1732 | unsigned int tag_offset = rq->mq_hctx->tags->nr_reserved_tags; |
570e9b73 ML |
1733 | int tag; |
1734 | ||
568f2700 ML |
1735 | blk_mq_tag_busy(rq->mq_hctx); |
1736 | ||
570e9b73 | 1737 | if (blk_mq_tag_is_reserved(rq->mq_hctx->sched_tags, rq->internal_tag)) { |
ae0f1a73 | 1738 | bt = &rq->mq_hctx->tags->breserved_tags; |
570e9b73 | 1739 | tag_offset = 0; |
28500850 ML |
1740 | } else { |
1741 | if (!hctx_may_queue(rq->mq_hctx, bt)) | |
1742 | return false; | |
570e9b73 ML |
1743 | } |
1744 | ||
570e9b73 ML |
1745 | tag = __sbitmap_queue_get(bt); |
1746 | if (tag == BLK_MQ_NO_TAG) | |
1747 | return false; | |
1748 | ||
1749 | rq->tag = tag + tag_offset; | |
570e9b73 ML |
1750 | return true; |
1751 | } | |
1752 | ||
a808a9d5 | 1753 | bool __blk_mq_get_driver_tag(struct blk_mq_hw_ctx *hctx, struct request *rq) |
570e9b73 | 1754 | { |
a808a9d5 | 1755 | if (rq->tag == BLK_MQ_NO_TAG && !__blk_mq_alloc_driver_tag(rq)) |
568f2700 ML |
1756 | return false; |
1757 | ||
51db1c37 | 1758 | if ((hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) && |
568f2700 ML |
1759 | !(rq->rq_flags & RQF_MQ_INFLIGHT)) { |
1760 | rq->rq_flags |= RQF_MQ_INFLIGHT; | |
bccf5e26 | 1761 | __blk_mq_inc_active_requests(hctx); |
568f2700 ML |
1762 | } |
1763 | hctx->tags->rqs[rq->tag] = rq; | |
1764 | return true; | |
570e9b73 ML |
1765 | } |
1766 | ||
eb619fdb JA |
1767 | static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode, |
1768 | int flags, void *key) | |
da55f2cc OS |
1769 | { |
1770 | struct blk_mq_hw_ctx *hctx; | |
1771 | ||
1772 | hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait); | |
1773 | ||
5815839b | 1774 | spin_lock(&hctx->dispatch_wait_lock); |
e8618575 JA |
1775 | if (!list_empty(&wait->entry)) { |
1776 | struct sbitmap_queue *sbq; | |
1777 | ||
1778 | list_del_init(&wait->entry); | |
ae0f1a73 | 1779 | sbq = &hctx->tags->bitmap_tags; |
e8618575 JA |
1780 | atomic_dec(&sbq->ws_active); |
1781 | } | |
5815839b ML |
1782 | spin_unlock(&hctx->dispatch_wait_lock); |
1783 | ||
da55f2cc OS |
1784 | blk_mq_run_hw_queue(hctx, true); |
1785 | return 1; | |
1786 | } | |
1787 | ||
f906a6a0 JA |
1788 | /* |
1789 | * Mark us waiting for a tag. For shared tags, this involves hooking us into | |
ee3e4de5 BVA |
1790 | * the tag wakeups. For non-shared tags, we can simply mark us needing a |
1791 | * restart. For both cases, take care to check the condition again after | |
f906a6a0 JA |
1792 | * marking us as waiting. |
1793 | */ | |
2278d69f | 1794 | static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx *hctx, |
f906a6a0 | 1795 | struct request *rq) |
da55f2cc | 1796 | { |
98b99e94 | 1797 | struct sbitmap_queue *sbq; |
5815839b | 1798 | struct wait_queue_head *wq; |
f906a6a0 JA |
1799 | wait_queue_entry_t *wait; |
1800 | bool ret; | |
da55f2cc | 1801 | |
47df9ce9 KS |
1802 | if (!(hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) && |
1803 | !(blk_mq_is_shared_tags(hctx->flags))) { | |
684b7324 | 1804 | blk_mq_sched_mark_restart_hctx(hctx); |
f906a6a0 | 1805 | |
c27d53fb BVA |
1806 | /* |
1807 | * It's possible that a tag was freed in the window between the | |
1808 | * allocation failure and adding the hardware queue to the wait | |
1809 | * queue. | |
1810 | * | |
1811 | * Don't clear RESTART here, someone else could have set it. | |
1812 | * At most this will cost an extra queue run. | |
1813 | */ | |
8ab6bb9e | 1814 | return blk_mq_get_driver_tag(rq); |
eb619fdb | 1815 | } |
eb619fdb | 1816 | |
2278d69f | 1817 | wait = &hctx->dispatch_wait; |
c27d53fb BVA |
1818 | if (!list_empty_careful(&wait->entry)) |
1819 | return false; | |
1820 | ||
98b99e94 KS |
1821 | if (blk_mq_tag_is_reserved(rq->mq_hctx->sched_tags, rq->internal_tag)) |
1822 | sbq = &hctx->tags->breserved_tags; | |
1823 | else | |
1824 | sbq = &hctx->tags->bitmap_tags; | |
e8618575 | 1825 | wq = &bt_wait_ptr(sbq, hctx)->wait; |
5815839b ML |
1826 | |
1827 | spin_lock_irq(&wq->lock); | |
1828 | spin_lock(&hctx->dispatch_wait_lock); | |
c27d53fb | 1829 | if (!list_empty(&wait->entry)) { |
5815839b ML |
1830 | spin_unlock(&hctx->dispatch_wait_lock); |
1831 | spin_unlock_irq(&wq->lock); | |
c27d53fb | 1832 | return false; |
eb619fdb JA |
1833 | } |
1834 | ||
e8618575 | 1835 | atomic_inc(&sbq->ws_active); |
5815839b ML |
1836 | wait->flags &= ~WQ_FLAG_EXCLUSIVE; |
1837 | __add_wait_queue(wq, wait); | |
c27d53fb | 1838 | |
da55f2cc | 1839 | /* |
eb619fdb JA |
1840 | * It's possible that a tag was freed in the window between the |
1841 | * allocation failure and adding the hardware queue to the wait | |
1842 | * queue. | |
da55f2cc | 1843 | */ |
8ab6bb9e | 1844 | ret = blk_mq_get_driver_tag(rq); |
c27d53fb | 1845 | if (!ret) { |
5815839b ML |
1846 | spin_unlock(&hctx->dispatch_wait_lock); |
1847 | spin_unlock_irq(&wq->lock); | |
c27d53fb | 1848 | return false; |
eb619fdb | 1849 | } |
c27d53fb BVA |
1850 | |
1851 | /* | |
1852 | * We got a tag, remove ourselves from the wait queue to ensure | |
1853 | * someone else gets the wakeup. | |
1854 | */ | |
c27d53fb | 1855 | list_del_init(&wait->entry); |
e8618575 | 1856 | atomic_dec(&sbq->ws_active); |
5815839b ML |
1857 | spin_unlock(&hctx->dispatch_wait_lock); |
1858 | spin_unlock_irq(&wq->lock); | |
c27d53fb BVA |
1859 | |
1860 | return true; | |
da55f2cc OS |
1861 | } |
1862 | ||
6e768717 ML |
1863 | #define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT 8 |
1864 | #define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR 4 | |
1865 | /* | |
1866 | * Update dispatch busy with the Exponential Weighted Moving Average(EWMA): | |
1867 | * - EWMA is one simple way to compute running average value | |
1868 | * - weight(7/8 and 1/8) is applied so that it can decrease exponentially | |
1869 | * - take 4 as factor for avoiding to get too small(0) result, and this | |
1870 | * factor doesn't matter because EWMA decreases exponentially | |
1871 | */ | |
1872 | static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx *hctx, bool busy) | |
1873 | { | |
1874 | unsigned int ewma; | |
1875 | ||
6e768717 ML |
1876 | ewma = hctx->dispatch_busy; |
1877 | ||
1878 | if (!ewma && !busy) | |
1879 | return; | |
1880 | ||
1881 | ewma *= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT - 1; | |
1882 | if (busy) | |
1883 | ewma += 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR; | |
1884 | ewma /= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT; | |
1885 | ||
1886 | hctx->dispatch_busy = ewma; | |
1887 | } | |
1888 | ||
86ff7c2a ML |
1889 | #define BLK_MQ_RESOURCE_DELAY 3 /* ms units */ |
1890 | ||
c92a4103 JT |
1891 | static void blk_mq_handle_dev_resource(struct request *rq, |
1892 | struct list_head *list) | |
1893 | { | |
c92a4103 JT |
1894 | list_add(&rq->queuelist, list); |
1895 | __blk_mq_requeue_request(rq); | |
1896 | } | |
1897 | ||
0512a75b KB |
1898 | static void blk_mq_handle_zone_resource(struct request *rq, |
1899 | struct list_head *zone_list) | |
1900 | { | |
1901 | /* | |
1902 | * If we end up here it is because we cannot dispatch a request to a | |
1903 | * specific zone due to LLD level zone-write locking or other zone | |
1904 | * related resource not being available. In this case, set the request | |
1905 | * aside in zone_list for retrying it later. | |
1906 | */ | |
1907 | list_add(&rq->queuelist, zone_list); | |
1908 | __blk_mq_requeue_request(rq); | |
1909 | } | |
1910 | ||
75383524 ML |
1911 | enum prep_dispatch { |
1912 | PREP_DISPATCH_OK, | |
1913 | PREP_DISPATCH_NO_TAG, | |
1914 | PREP_DISPATCH_NO_BUDGET, | |
1915 | }; | |
1916 | ||
1917 | static enum prep_dispatch blk_mq_prep_dispatch_rq(struct request *rq, | |
1918 | bool need_budget) | |
1919 | { | |
1920 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; | |
2a5a24aa | 1921 | int budget_token = -1; |
75383524 | 1922 | |
2a5a24aa ML |
1923 | if (need_budget) { |
1924 | budget_token = blk_mq_get_dispatch_budget(rq->q); | |
1925 | if (budget_token < 0) { | |
1926 | blk_mq_put_driver_tag(rq); | |
1927 | return PREP_DISPATCH_NO_BUDGET; | |
1928 | } | |
1929 | blk_mq_set_rq_budget_token(rq, budget_token); | |
75383524 ML |
1930 | } |
1931 | ||
1932 | if (!blk_mq_get_driver_tag(rq)) { | |
1933 | /* | |
1934 | * The initial allocation attempt failed, so we need to | |
1935 | * rerun the hardware queue when a tag is freed. The | |
1936 | * waitqueue takes care of that. If the queue is run | |
1937 | * before we add this entry back on the dispatch list, | |
1938 | * we'll re-run it below. | |
1939 | */ | |
1940 | if (!blk_mq_mark_tag_wait(hctx, rq)) { | |
1fd40b5e ML |
1941 | /* |
1942 | * All budgets not got from this function will be put | |
1943 | * together during handling partial dispatch | |
1944 | */ | |
1945 | if (need_budget) | |
2a5a24aa | 1946 | blk_mq_put_dispatch_budget(rq->q, budget_token); |
75383524 ML |
1947 | return PREP_DISPATCH_NO_TAG; |
1948 | } | |
1949 | } | |
1950 | ||
1951 | return PREP_DISPATCH_OK; | |
1952 | } | |
1953 | ||
1fd40b5e ML |
1954 | /* release all allocated budgets before calling to blk_mq_dispatch_rq_list */ |
1955 | static void blk_mq_release_budgets(struct request_queue *q, | |
2a5a24aa | 1956 | struct list_head *list) |
1fd40b5e | 1957 | { |
2a5a24aa | 1958 | struct request *rq; |
1fd40b5e | 1959 | |
2a5a24aa ML |
1960 | list_for_each_entry(rq, list, queuelist) { |
1961 | int budget_token = blk_mq_get_rq_budget_token(rq); | |
1fd40b5e | 1962 | |
2a5a24aa ML |
1963 | if (budget_token >= 0) |
1964 | blk_mq_put_dispatch_budget(q, budget_token); | |
1965 | } | |
1fd40b5e ML |
1966 | } |
1967 | ||
34c9f547 KS |
1968 | /* |
1969 | * blk_mq_commit_rqs will notify driver using bd->last that there is no | |
1970 | * more requests. (See comment in struct blk_mq_ops for commit_rqs for | |
1971 | * details) | |
1972 | * Attention, we should explicitly call this in unusual cases: | |
1973 | * 1) did not queue everything initially scheduled to queue | |
1974 | * 2) the last attempt to queue a request failed | |
1975 | */ | |
1976 | static void blk_mq_commit_rqs(struct blk_mq_hw_ctx *hctx, int queued, | |
1977 | bool from_schedule) | |
1978 | { | |
1979 | if (hctx->queue->mq_ops->commit_rqs && queued) { | |
1980 | trace_block_unplug(hctx->queue, queued, !from_schedule); | |
1981 | hctx->queue->mq_ops->commit_rqs(hctx); | |
1982 | } | |
1983 | } | |
1984 | ||
1f57f8d4 JA |
1985 | /* |
1986 | * Returns true if we did some work AND can potentially do more. | |
1987 | */ | |
445874e8 | 1988 | bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx *hctx, struct list_head *list, |
1fd40b5e | 1989 | unsigned int nr_budgets) |
320ae51f | 1990 | { |
75383524 | 1991 | enum prep_dispatch prep; |
445874e8 | 1992 | struct request_queue *q = hctx->queue; |
f1ce99f7 | 1993 | struct request *rq; |
4ea58fe4 | 1994 | int queued; |
86ff7c2a | 1995 | blk_status_t ret = BLK_STS_OK; |
0512a75b | 1996 | LIST_HEAD(zone_list); |
9586e67b | 1997 | bool needs_resource = false; |
320ae51f | 1998 | |
81380ca1 OS |
1999 | if (list_empty(list)) |
2000 | return false; | |
2001 | ||
320ae51f JA |
2002 | /* |
2003 | * Now process all the entries, sending them to the driver. | |
2004 | */ | |
4ea58fe4 | 2005 | queued = 0; |
81380ca1 | 2006 | do { |
74c45052 | 2007 | struct blk_mq_queue_data bd; |
320ae51f | 2008 | |
f04c3df3 | 2009 | rq = list_first_entry(list, struct request, queuelist); |
0bca799b | 2010 | |
445874e8 | 2011 | WARN_ON_ONCE(hctx != rq->mq_hctx); |
1fd40b5e | 2012 | prep = blk_mq_prep_dispatch_rq(rq, !nr_budgets); |
75383524 | 2013 | if (prep != PREP_DISPATCH_OK) |
0bca799b | 2014 | break; |
de148297 | 2015 | |
320ae51f | 2016 | list_del_init(&rq->queuelist); |
320ae51f | 2017 | |
74c45052 | 2018 | bd.rq = rq; |
f1ce99f7 | 2019 | bd.last = list_empty(list); |
74c45052 | 2020 | |
1fd40b5e ML |
2021 | /* |
2022 | * once the request is queued to lld, no need to cover the | |
2023 | * budget any more | |
2024 | */ | |
2025 | if (nr_budgets) | |
2026 | nr_budgets--; | |
74c45052 | 2027 | ret = q->mq_ops->queue_rq(hctx, &bd); |
7bf13729 ML |
2028 | switch (ret) { |
2029 | case BLK_STS_OK: | |
2030 | queued++; | |
320ae51f | 2031 | break; |
7bf13729 | 2032 | case BLK_STS_RESOURCE: |
9586e67b NA |
2033 | needs_resource = true; |
2034 | fallthrough; | |
7bf13729 ML |
2035 | case BLK_STS_DEV_RESOURCE: |
2036 | blk_mq_handle_dev_resource(rq, list); | |
2037 | goto out; | |
2038 | case BLK_STS_ZONE_RESOURCE: | |
0512a75b KB |
2039 | /* |
2040 | * Move the request to zone_list and keep going through | |
2041 | * the dispatch list to find more requests the drive can | |
2042 | * accept. | |
2043 | */ | |
2044 | blk_mq_handle_zone_resource(rq, &zone_list); | |
9586e67b | 2045 | needs_resource = true; |
7bf13729 ML |
2046 | break; |
2047 | default: | |
e21ee5a6 | 2048 | blk_mq_end_request(rq, ret); |
320ae51f | 2049 | } |
81380ca1 | 2050 | } while (!list_empty(list)); |
7bf13729 | 2051 | out: |
0512a75b KB |
2052 | if (!list_empty(&zone_list)) |
2053 | list_splice_tail_init(&zone_list, list); | |
2054 | ||
632bfb63 | 2055 | /* If we didn't flush the entire list, we could have told the driver |
2056 | * there was more coming, but that turned out to be a lie. | |
2057 | */ | |
e4ef2e05 KS |
2058 | if (!list_empty(list) || ret != BLK_STS_OK) |
2059 | blk_mq_commit_rqs(hctx, queued, false); | |
2060 | ||
320ae51f JA |
2061 | /* |
2062 | * Any items that need requeuing? Stuff them into hctx->dispatch, | |
2063 | * that is where we will continue on next queue run. | |
2064 | */ | |
f04c3df3 | 2065 | if (!list_empty(list)) { |
86ff7c2a | 2066 | bool needs_restart; |
75383524 ML |
2067 | /* For non-shared tags, the RESTART check will suffice */ |
2068 | bool no_tag = prep == PREP_DISPATCH_NO_TAG && | |
47df9ce9 KS |
2069 | ((hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) || |
2070 | blk_mq_is_shared_tags(hctx->flags)); | |
86ff7c2a | 2071 | |
2a5a24aa ML |
2072 | if (nr_budgets) |
2073 | blk_mq_release_budgets(q, list); | |
86ff7c2a | 2074 | |
320ae51f | 2075 | spin_lock(&hctx->lock); |
01e99aec | 2076 | list_splice_tail_init(list, &hctx->dispatch); |
320ae51f | 2077 | spin_unlock(&hctx->lock); |
f04c3df3 | 2078 | |
d7d8535f ML |
2079 | /* |
2080 | * Order adding requests to hctx->dispatch and checking | |
2081 | * SCHED_RESTART flag. The pair of this smp_mb() is the one | |
2082 | * in blk_mq_sched_restart(). Avoid restart code path to | |
2083 | * miss the new added requests to hctx->dispatch, meantime | |
2084 | * SCHED_RESTART is observed here. | |
2085 | */ | |
2086 | smp_mb(); | |
2087 | ||
9ba52e58 | 2088 | /* |
710c785f BVA |
2089 | * If SCHED_RESTART was set by the caller of this function and |
2090 | * it is no longer set that means that it was cleared by another | |
2091 | * thread and hence that a queue rerun is needed. | |
9ba52e58 | 2092 | * |
eb619fdb JA |
2093 | * If 'no_tag' is set, that means that we failed getting |
2094 | * a driver tag with an I/O scheduler attached. If our dispatch | |
2095 | * waitqueue is no longer active, ensure that we run the queue | |
2096 | * AFTER adding our entries back to the list. | |
bd166ef1 | 2097 | * |
710c785f BVA |
2098 | * If no I/O scheduler has been configured it is possible that |
2099 | * the hardware queue got stopped and restarted before requests | |
2100 | * were pushed back onto the dispatch list. Rerun the queue to | |
2101 | * avoid starvation. Notes: | |
2102 | * - blk_mq_run_hw_queue() checks whether or not a queue has | |
2103 | * been stopped before rerunning a queue. | |
2104 | * - Some but not all block drivers stop a queue before | |
fc17b653 | 2105 | * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq |
710c785f | 2106 | * and dm-rq. |
86ff7c2a ML |
2107 | * |
2108 | * If driver returns BLK_STS_RESOURCE and SCHED_RESTART | |
2109 | * bit is set, run queue after a delay to avoid IO stalls | |
ab3cee37 | 2110 | * that could otherwise occur if the queue is idle. We'll do |
9586e67b NA |
2111 | * similar if we couldn't get budget or couldn't lock a zone |
2112 | * and SCHED_RESTART is set. | |
bd166ef1 | 2113 | */ |
86ff7c2a | 2114 | needs_restart = blk_mq_sched_needs_restart(hctx); |
9586e67b NA |
2115 | if (prep == PREP_DISPATCH_NO_BUDGET) |
2116 | needs_resource = true; | |
86ff7c2a | 2117 | if (!needs_restart || |
eb619fdb | 2118 | (no_tag && list_empty_careful(&hctx->dispatch_wait.entry))) |
bd166ef1 | 2119 | blk_mq_run_hw_queue(hctx, true); |
6d5e8d21 | 2120 | else if (needs_resource) |
86ff7c2a | 2121 | blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY); |
1f57f8d4 | 2122 | |
6e768717 | 2123 | blk_mq_update_dispatch_busy(hctx, true); |
1f57f8d4 | 2124 | return false; |
4ea58fe4 | 2125 | } |
f04c3df3 | 2126 | |
4ea58fe4 KS |
2127 | blk_mq_update_dispatch_busy(hctx, false); |
2128 | return true; | |
f04c3df3 JA |
2129 | } |
2130 | ||
f82ddf19 ML |
2131 | static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx) |
2132 | { | |
2133 | int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask); | |
2134 | ||
2135 | if (cpu >= nr_cpu_ids) | |
2136 | cpu = cpumask_first(hctx->cpumask); | |
2137 | return cpu; | |
2138 | } | |
2139 | ||
506e931f JA |
2140 | /* |
2141 | * It'd be great if the workqueue API had a way to pass | |
2142 | * in a mask and had some smarts for more clever placement. | |
2143 | * For now we just round-robin here, switching for every | |
2144 | * BLK_MQ_CPU_WORK_BATCH queued items. | |
2145 | */ | |
2146 | static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx) | |
2147 | { | |
7bed4595 | 2148 | bool tried = false; |
476f8c98 | 2149 | int next_cpu = hctx->next_cpu; |
7bed4595 | 2150 | |
b657d7e6 CH |
2151 | if (hctx->queue->nr_hw_queues == 1) |
2152 | return WORK_CPU_UNBOUND; | |
506e931f JA |
2153 | |
2154 | if (--hctx->next_cpu_batch <= 0) { | |
7bed4595 | 2155 | select_cpu: |
476f8c98 | 2156 | next_cpu = cpumask_next_and(next_cpu, hctx->cpumask, |
20e4d813 | 2157 | cpu_online_mask); |
506e931f | 2158 | if (next_cpu >= nr_cpu_ids) |
f82ddf19 | 2159 | next_cpu = blk_mq_first_mapped_cpu(hctx); |
506e931f JA |
2160 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; |
2161 | } | |
2162 | ||
7bed4595 ML |
2163 | /* |
2164 | * Do unbound schedule if we can't find a online CPU for this hctx, | |
2165 | * and it should only happen in the path of handling CPU DEAD. | |
2166 | */ | |
476f8c98 | 2167 | if (!cpu_online(next_cpu)) { |
7bed4595 ML |
2168 | if (!tried) { |
2169 | tried = true; | |
2170 | goto select_cpu; | |
2171 | } | |
2172 | ||
2173 | /* | |
2174 | * Make sure to re-select CPU next time once after CPUs | |
2175 | * in hctx->cpumask become online again. | |
2176 | */ | |
476f8c98 | 2177 | hctx->next_cpu = next_cpu; |
7bed4595 ML |
2178 | hctx->next_cpu_batch = 1; |
2179 | return WORK_CPU_UNBOUND; | |
2180 | } | |
476f8c98 ML |
2181 | |
2182 | hctx->next_cpu = next_cpu; | |
2183 | return next_cpu; | |
506e931f JA |
2184 | } |
2185 | ||
105663f7 | 2186 | /** |
1aa8d875 | 2187 | * blk_mq_delay_run_hw_queue - Run a hardware queue asynchronously. |
105663f7 | 2188 | * @hctx: Pointer to the hardware queue to run. |
fa94ba8a | 2189 | * @msecs: Milliseconds of delay to wait before running the queue. |
105663f7 | 2190 | * |
1aa8d875 | 2191 | * Run a hardware queue asynchronously with a delay of @msecs. |
105663f7 | 2192 | */ |
1aa8d875 | 2193 | void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs) |
320ae51f | 2194 | { |
5435c023 | 2195 | if (unlikely(blk_mq_hctx_stopped(hctx))) |
320ae51f | 2196 | return; |
ae943d20 BVA |
2197 | kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work, |
2198 | msecs_to_jiffies(msecs)); | |
7587a5ae | 2199 | } |
7587a5ae BVA |
2200 | EXPORT_SYMBOL(blk_mq_delay_run_hw_queue); |
2201 | ||
105663f7 AA |
2202 | /** |
2203 | * blk_mq_run_hw_queue - Start to run a hardware queue. | |
2204 | * @hctx: Pointer to the hardware queue to run. | |
2205 | * @async: If we want to run the queue asynchronously. | |
2206 | * | |
2207 | * Check if the request queue is not in a quiesced state and if there are | |
2208 | * pending requests to be sent. If this is true, run the queue to send requests | |
2209 | * to hardware. | |
2210 | */ | |
626fb735 | 2211 | void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) |
7587a5ae | 2212 | { |
24f5a90f ML |
2213 | bool need_run; |
2214 | ||
4d5bba5b CH |
2215 | /* |
2216 | * We can't run the queue inline with interrupts disabled. | |
2217 | */ | |
2218 | WARN_ON_ONCE(!async && in_interrupt()); | |
2219 | ||
24f5a90f ML |
2220 | /* |
2221 | * When queue is quiesced, we may be switching io scheduler, or | |
2222 | * updating nr_hw_queues, or other things, and we can't run queue | |
2223 | * any more, even __blk_mq_hctx_has_pending() can't be called safely. | |
2224 | * | |
2225 | * And queue will be rerun in blk_mq_unquiesce_queue() if it is | |
2226 | * quiesced. | |
2227 | */ | |
41adf531 | 2228 | __blk_mq_run_dispatch_ops(hctx->queue, false, |
2a904d00 ML |
2229 | need_run = !blk_queue_quiesced(hctx->queue) && |
2230 | blk_mq_hctx_has_pending(hctx)); | |
24f5a90f | 2231 | |
1aa8d875 CH |
2232 | if (!need_run) |
2233 | return; | |
2234 | ||
2235 | if (async || (hctx->flags & BLK_MQ_F_BLOCKING) || | |
2236 | !cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask)) { | |
2237 | blk_mq_delay_run_hw_queue(hctx, 0); | |
2238 | return; | |
2239 | } | |
2240 | ||
4d5bba5b CH |
2241 | blk_mq_run_dispatch_ops(hctx->queue, |
2242 | blk_mq_sched_dispatch_requests(hctx)); | |
320ae51f | 2243 | } |
5b727272 | 2244 | EXPORT_SYMBOL(blk_mq_run_hw_queue); |
320ae51f | 2245 | |
b6e68ee8 JK |
2246 | /* |
2247 | * Return prefered queue to dispatch from (if any) for non-mq aware IO | |
2248 | * scheduler. | |
2249 | */ | |
2250 | static struct blk_mq_hw_ctx *blk_mq_get_sq_hctx(struct request_queue *q) | |
2251 | { | |
5d05426e | 2252 | struct blk_mq_ctx *ctx = blk_mq_get_ctx(q); |
b6e68ee8 JK |
2253 | /* |
2254 | * If the IO scheduler does not respect hardware queues when | |
2255 | * dispatching, we just don't bother with multiple HW queues and | |
2256 | * dispatch from hctx for the current CPU since running multiple queues | |
2257 | * just causes lock contention inside the scheduler and pointless cache | |
2258 | * bouncing. | |
2259 | */ | |
51ab80f0 | 2260 | struct blk_mq_hw_ctx *hctx = ctx->hctxs[HCTX_TYPE_DEFAULT]; |
5d05426e | 2261 | |
b6e68ee8 JK |
2262 | if (!blk_mq_hctx_stopped(hctx)) |
2263 | return hctx; | |
2264 | return NULL; | |
2265 | } | |
2266 | ||
105663f7 | 2267 | /** |
24f7bb88 | 2268 | * blk_mq_run_hw_queues - Run all hardware queues in a request queue. |
105663f7 AA |
2269 | * @q: Pointer to the request queue to run. |
2270 | * @async: If we want to run the queue asynchronously. | |
2271 | */ | |
b94ec296 | 2272 | void blk_mq_run_hw_queues(struct request_queue *q, bool async) |
320ae51f | 2273 | { |
b6e68ee8 | 2274 | struct blk_mq_hw_ctx *hctx, *sq_hctx; |
4f481208 | 2275 | unsigned long i; |
320ae51f | 2276 | |
b6e68ee8 | 2277 | sq_hctx = NULL; |
4d337ceb | 2278 | if (blk_queue_sq_sched(q)) |
b6e68ee8 | 2279 | sq_hctx = blk_mq_get_sq_hctx(q); |
320ae51f | 2280 | queue_for_each_hw_ctx(q, hctx, i) { |
79f720a7 | 2281 | if (blk_mq_hctx_stopped(hctx)) |
320ae51f | 2282 | continue; |
b6e68ee8 JK |
2283 | /* |
2284 | * Dispatch from this hctx either if there's no hctx preferred | |
2285 | * by IO scheduler or if it has requests that bypass the | |
2286 | * scheduler. | |
2287 | */ | |
2288 | if (!sq_hctx || sq_hctx == hctx || | |
2289 | !list_empty_careful(&hctx->dispatch)) | |
2290 | blk_mq_run_hw_queue(hctx, async); | |
320ae51f JA |
2291 | } |
2292 | } | |
b94ec296 | 2293 | EXPORT_SYMBOL(blk_mq_run_hw_queues); |
320ae51f | 2294 | |
b9151e7b DA |
2295 | /** |
2296 | * blk_mq_delay_run_hw_queues - Run all hardware queues asynchronously. | |
2297 | * @q: Pointer to the request queue to run. | |
fa94ba8a | 2298 | * @msecs: Milliseconds of delay to wait before running the queues. |
b9151e7b DA |
2299 | */ |
2300 | void blk_mq_delay_run_hw_queues(struct request_queue *q, unsigned long msecs) | |
2301 | { | |
b6e68ee8 | 2302 | struct blk_mq_hw_ctx *hctx, *sq_hctx; |
4f481208 | 2303 | unsigned long i; |
b9151e7b | 2304 | |
b6e68ee8 | 2305 | sq_hctx = NULL; |
4d337ceb | 2306 | if (blk_queue_sq_sched(q)) |
b6e68ee8 | 2307 | sq_hctx = blk_mq_get_sq_hctx(q); |
b9151e7b DA |
2308 | queue_for_each_hw_ctx(q, hctx, i) { |
2309 | if (blk_mq_hctx_stopped(hctx)) | |
2310 | continue; | |
8f5fea65 DJ |
2311 | /* |
2312 | * If there is already a run_work pending, leave the | |
2313 | * pending delay untouched. Otherwise, a hctx can stall | |
2314 | * if another hctx is re-delaying the other's work | |
2315 | * before the work executes. | |
2316 | */ | |
2317 | if (delayed_work_pending(&hctx->run_work)) | |
2318 | continue; | |
b6e68ee8 JK |
2319 | /* |
2320 | * Dispatch from this hctx either if there's no hctx preferred | |
2321 | * by IO scheduler or if it has requests that bypass the | |
2322 | * scheduler. | |
2323 | */ | |
2324 | if (!sq_hctx || sq_hctx == hctx || | |
2325 | !list_empty_careful(&hctx->dispatch)) | |
2326 | blk_mq_delay_run_hw_queue(hctx, msecs); | |
b9151e7b DA |
2327 | } |
2328 | } | |
2329 | EXPORT_SYMBOL(blk_mq_delay_run_hw_queues); | |
2330 | ||
39a70c76 ML |
2331 | /* |
2332 | * This function is often used for pausing .queue_rq() by driver when | |
2333 | * there isn't enough resource or some conditions aren't satisfied, and | |
4d606219 | 2334 | * BLK_STS_RESOURCE is usually returned. |
39a70c76 ML |
2335 | * |
2336 | * We do not guarantee that dispatch can be drained or blocked | |
2337 | * after blk_mq_stop_hw_queue() returns. Please use | |
2338 | * blk_mq_quiesce_queue() for that requirement. | |
2339 | */ | |
2719aa21 JA |
2340 | void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx) |
2341 | { | |
641a9ed6 | 2342 | cancel_delayed_work(&hctx->run_work); |
280d45f6 | 2343 | |
641a9ed6 | 2344 | set_bit(BLK_MQ_S_STOPPED, &hctx->state); |
2719aa21 | 2345 | } |
641a9ed6 | 2346 | EXPORT_SYMBOL(blk_mq_stop_hw_queue); |
2719aa21 | 2347 | |
39a70c76 ML |
2348 | /* |
2349 | * This function is often used for pausing .queue_rq() by driver when | |
2350 | * there isn't enough resource or some conditions aren't satisfied, and | |
4d606219 | 2351 | * BLK_STS_RESOURCE is usually returned. |
39a70c76 ML |
2352 | * |
2353 | * We do not guarantee that dispatch can be drained or blocked | |
2354 | * after blk_mq_stop_hw_queues() returns. Please use | |
2355 | * blk_mq_quiesce_queue() for that requirement. | |
2356 | */ | |
2719aa21 JA |
2357 | void blk_mq_stop_hw_queues(struct request_queue *q) |
2358 | { | |
641a9ed6 | 2359 | struct blk_mq_hw_ctx *hctx; |
4f481208 | 2360 | unsigned long i; |
641a9ed6 ML |
2361 | |
2362 | queue_for_each_hw_ctx(q, hctx, i) | |
2363 | blk_mq_stop_hw_queue(hctx); | |
280d45f6 CH |
2364 | } |
2365 | EXPORT_SYMBOL(blk_mq_stop_hw_queues); | |
2366 | ||
320ae51f JA |
2367 | void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx) |
2368 | { | |
2369 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
e4043dcf | 2370 | |
0ffbce80 | 2371 | blk_mq_run_hw_queue(hctx, false); |
320ae51f JA |
2372 | } |
2373 | EXPORT_SYMBOL(blk_mq_start_hw_queue); | |
2374 | ||
2f268556 CH |
2375 | void blk_mq_start_hw_queues(struct request_queue *q) |
2376 | { | |
2377 | struct blk_mq_hw_ctx *hctx; | |
4f481208 | 2378 | unsigned long i; |
2f268556 CH |
2379 | |
2380 | queue_for_each_hw_ctx(q, hctx, i) | |
2381 | blk_mq_start_hw_queue(hctx); | |
2382 | } | |
2383 | EXPORT_SYMBOL(blk_mq_start_hw_queues); | |
2384 | ||
ae911c5e JA |
2385 | void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) |
2386 | { | |
2387 | if (!blk_mq_hctx_stopped(hctx)) | |
2388 | return; | |
2389 | ||
2390 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
2391 | blk_mq_run_hw_queue(hctx, async); | |
2392 | } | |
2393 | EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue); | |
2394 | ||
1b4a3258 | 2395 | void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async) |
320ae51f JA |
2396 | { |
2397 | struct blk_mq_hw_ctx *hctx; | |
4f481208 | 2398 | unsigned long i; |
320ae51f | 2399 | |
ae911c5e JA |
2400 | queue_for_each_hw_ctx(q, hctx, i) |
2401 | blk_mq_start_stopped_hw_queue(hctx, async); | |
320ae51f JA |
2402 | } |
2403 | EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues); | |
2404 | ||
70f4db63 | 2405 | static void blk_mq_run_work_fn(struct work_struct *work) |
320ae51f | 2406 | { |
c20a1a2c CH |
2407 | struct blk_mq_hw_ctx *hctx = |
2408 | container_of(work, struct blk_mq_hw_ctx, run_work.work); | |
7b607814 | 2409 | |
4d5bba5b CH |
2410 | blk_mq_run_dispatch_ops(hctx->queue, |
2411 | blk_mq_sched_dispatch_requests(hctx)); | |
320ae51f JA |
2412 | } |
2413 | ||
105663f7 AA |
2414 | /** |
2415 | * blk_mq_request_bypass_insert - Insert a request at dispatch list. | |
2416 | * @rq: Pointer to request to be inserted. | |
2b597613 | 2417 | * @flags: BLK_MQ_INSERT_* |
105663f7 | 2418 | * |
157f377b JA |
2419 | * Should only be used carefully, when the caller knows we want to |
2420 | * bypass a potential IO scheduler on the target device. | |
2421 | */ | |
360f2648 | 2422 | static void blk_mq_request_bypass_insert(struct request *rq, blk_insert_t flags) |
157f377b | 2423 | { |
ea4f995e | 2424 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; |
157f377b JA |
2425 | |
2426 | spin_lock(&hctx->lock); | |
2b597613 | 2427 | if (flags & BLK_MQ_INSERT_AT_HEAD) |
01e99aec ML |
2428 | list_add(&rq->queuelist, &hctx->dispatch); |
2429 | else | |
2430 | list_add_tail(&rq->queuelist, &hctx->dispatch); | |
157f377b | 2431 | spin_unlock(&hctx->lock); |
157f377b JA |
2432 | } |
2433 | ||
05a93117 CH |
2434 | static void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, |
2435 | struct blk_mq_ctx *ctx, struct list_head *list, | |
2436 | bool run_queue_async) | |
320ae51f | 2437 | { |
3f0cedc7 | 2438 | struct request *rq; |
c16d6b5a | 2439 | enum hctx_type type = hctx->type; |
3f0cedc7 | 2440 | |
94aa228c CH |
2441 | /* |
2442 | * Try to issue requests directly if the hw queue isn't busy to save an | |
2443 | * extra enqueue & dequeue to the sw queue. | |
2444 | */ | |
2445 | if (!hctx->dispatch_busy && !run_queue_async) { | |
2446 | blk_mq_run_dispatch_ops(hctx->queue, | |
2447 | blk_mq_try_issue_list_directly(hctx, list)); | |
2448 | if (list_empty(list)) | |
2449 | goto out; | |
2450 | } | |
2451 | ||
320ae51f JA |
2452 | /* |
2453 | * preemption doesn't flush plug list, so it's possible ctx->cpu is | |
2454 | * offline now | |
2455 | */ | |
3f0cedc7 | 2456 | list_for_each_entry(rq, list, queuelist) { |
e57690fe | 2457 | BUG_ON(rq->mq_ctx != ctx); |
a54895fa | 2458 | trace_block_rq_insert(rq); |
320ae51f | 2459 | } |
3f0cedc7 ML |
2460 | |
2461 | spin_lock(&ctx->lock); | |
c16d6b5a | 2462 | list_splice_tail_init(list, &ctx->rq_lists[type]); |
cfd0c552 | 2463 | blk_mq_hctx_mark_pending(hctx, ctx); |
320ae51f | 2464 | spin_unlock(&ctx->lock); |
94aa228c CH |
2465 | out: |
2466 | blk_mq_run_hw_queue(hctx, run_queue_async); | |
320ae51f JA |
2467 | } |
2468 | ||
710fa378 | 2469 | static void blk_mq_insert_request(struct request *rq, blk_insert_t flags) |
2bd215df CH |
2470 | { |
2471 | struct request_queue *q = rq->q; | |
2bd215df CH |
2472 | struct blk_mq_ctx *ctx = rq->mq_ctx; |
2473 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; | |
2474 | ||
53548d2a CH |
2475 | if (blk_rq_is_passthrough(rq)) { |
2476 | /* | |
2477 | * Passthrough request have to be added to hctx->dispatch | |
2478 | * directly. The device may be in a situation where it can't | |
2479 | * handle FS request, and always returns BLK_STS_RESOURCE for | |
2480 | * them, which gets them added to hctx->dispatch. | |
2481 | * | |
2482 | * If a passthrough request is required to unblock the queues, | |
2483 | * and it is added to the scheduler queue, there is no chance to | |
2484 | * dispatch it given we prioritize requests in hctx->dispatch. | |
2485 | */ | |
2b597613 | 2486 | blk_mq_request_bypass_insert(rq, flags); |
be4c4278 | 2487 | } else if (req_op(rq) == REQ_OP_FLUSH) { |
2bd215df CH |
2488 | /* |
2489 | * Firstly normal IO request is inserted to scheduler queue or | |
2490 | * sw queue, meantime we add flush request to dispatch queue( | |
2491 | * hctx->dispatch) directly and there is at most one in-flight | |
2492 | * flush request for each hw queue, so it doesn't matter to add | |
2493 | * flush request to tail or front of the dispatch queue. | |
2494 | * | |
2495 | * Secondly in case of NCQ, flush request belongs to non-NCQ | |
2496 | * command, and queueing it will fail when there is any | |
2497 | * in-flight normal IO request(NCQ command). When adding flush | |
2498 | * rq to the front of hctx->dispatch, it is easier to introduce | |
2499 | * extra time to flush rq's latency because of S_SCHED_RESTART | |
2500 | * compared with adding to the tail of dispatch queue, then | |
2501 | * chance of flush merge is increased, and less flush requests | |
2502 | * will be issued to controller. It is observed that ~10% time | |
2503 | * is saved in blktests block/004 on disk attached to AHCI/NCQ | |
2504 | * drive when adding flush rq to the front of hctx->dispatch. | |
2505 | * | |
2506 | * Simply queue flush rq to the front of hctx->dispatch so that | |
2507 | * intensive flush workloads can benefit in case of NCQ HW. | |
2508 | */ | |
2b597613 | 2509 | blk_mq_request_bypass_insert(rq, BLK_MQ_INSERT_AT_HEAD); |
53548d2a | 2510 | } else if (q->elevator) { |
2bd215df CH |
2511 | LIST_HEAD(list); |
2512 | ||
53548d2a CH |
2513 | WARN_ON_ONCE(rq->tag != BLK_MQ_NO_TAG); |
2514 | ||
2bd215df | 2515 | list_add(&rq->queuelist, &list); |
93fffe16 | 2516 | q->elevator->type->ops.insert_requests(hctx, &list, flags); |
2bd215df | 2517 | } else { |
4ec5c055 CH |
2518 | trace_block_rq_insert(rq); |
2519 | ||
2bd215df | 2520 | spin_lock(&ctx->lock); |
710fa378 | 2521 | if (flags & BLK_MQ_INSERT_AT_HEAD) |
4ec5c055 CH |
2522 | list_add(&rq->queuelist, &ctx->rq_lists[hctx->type]); |
2523 | else | |
2524 | list_add_tail(&rq->queuelist, | |
2525 | &ctx->rq_lists[hctx->type]); | |
a88db1e0 | 2526 | blk_mq_hctx_mark_pending(hctx, ctx); |
2bd215df CH |
2527 | spin_unlock(&ctx->lock); |
2528 | } | |
320ae51f JA |
2529 | } |
2530 | ||
14ccb66b CH |
2531 | static void blk_mq_bio_to_request(struct request *rq, struct bio *bio, |
2532 | unsigned int nr_segs) | |
320ae51f | 2533 | { |
93f221ae EB |
2534 | int err; |
2535 | ||
f924cdde CH |
2536 | if (bio->bi_opf & REQ_RAHEAD) |
2537 | rq->cmd_flags |= REQ_FAILFAST_MASK; | |
2538 | ||
2539 | rq->__sector = bio->bi_iter.bi_sector; | |
14ccb66b | 2540 | blk_rq_bio_prep(rq, bio, nr_segs); |
93f221ae EB |
2541 | |
2542 | /* This can't fail, since GFP_NOIO includes __GFP_DIRECT_RECLAIM. */ | |
2543 | err = blk_crypto_rq_bio_prep(rq, bio, GFP_NOIO); | |
2544 | WARN_ON_ONCE(err); | |
4b570521 | 2545 | |
b5af37ab | 2546 | blk_account_io_start(rq); |
320ae51f JA |
2547 | } |
2548 | ||
0f95549c | 2549 | static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx, |
3e08773c | 2550 | struct request *rq, bool last) |
f984df1f | 2551 | { |
f984df1f | 2552 | struct request_queue *q = rq->q; |
f984df1f SL |
2553 | struct blk_mq_queue_data bd = { |
2554 | .rq = rq, | |
be94f058 | 2555 | .last = last, |
f984df1f | 2556 | }; |
f06345ad | 2557 | blk_status_t ret; |
0f95549c | 2558 | |
0f95549c MS |
2559 | /* |
2560 | * For OK queue, we are done. For error, caller may kill it. | |
2561 | * Any other error (busy), just add it to our list as we | |
2562 | * previously would have done. | |
2563 | */ | |
2564 | ret = q->mq_ops->queue_rq(hctx, &bd); | |
2565 | switch (ret) { | |
2566 | case BLK_STS_OK: | |
6ce3dd6e | 2567 | blk_mq_update_dispatch_busy(hctx, false); |
0f95549c MS |
2568 | break; |
2569 | case BLK_STS_RESOURCE: | |
86ff7c2a | 2570 | case BLK_STS_DEV_RESOURCE: |
6ce3dd6e | 2571 | blk_mq_update_dispatch_busy(hctx, true); |
0f95549c MS |
2572 | __blk_mq_requeue_request(rq); |
2573 | break; | |
2574 | default: | |
6ce3dd6e | 2575 | blk_mq_update_dispatch_busy(hctx, false); |
0f95549c MS |
2576 | break; |
2577 | } | |
2578 | ||
2579 | return ret; | |
2580 | } | |
2581 | ||
2b71b877 | 2582 | static bool blk_mq_get_budget_and_tag(struct request *rq) |
0f95549c | 2583 | { |
2a5a24aa | 2584 | int budget_token; |
d964f04a | 2585 | |
2b71b877 | 2586 | budget_token = blk_mq_get_dispatch_budget(rq->q); |
2a5a24aa | 2587 | if (budget_token < 0) |
2b71b877 | 2588 | return false; |
2a5a24aa | 2589 | blk_mq_set_rq_budget_token(rq, budget_token); |
8ab6bb9e | 2590 | if (!blk_mq_get_driver_tag(rq)) { |
2b71b877 CH |
2591 | blk_mq_put_dispatch_budget(rq->q, budget_token); |
2592 | return false; | |
88022d72 | 2593 | } |
2b71b877 | 2594 | return true; |
fd9c40f6 BVA |
2595 | } |
2596 | ||
105663f7 AA |
2597 | /** |
2598 | * blk_mq_try_issue_directly - Try to send a request directly to device driver. | |
2599 | * @hctx: Pointer of the associated hardware queue. | |
2600 | * @rq: Pointer to request to be sent. | |
105663f7 AA |
2601 | * |
2602 | * If the device has enough resources to accept a new request now, send the | |
2603 | * request directly to device driver. Else, insert at hctx->dispatch queue, so | |
2604 | * we can try send it another time in the future. Requests inserted at this | |
2605 | * queue have higher priority. | |
2606 | */ | |
fd9c40f6 | 2607 | static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx, |
3e08773c | 2608 | struct request *rq) |
fd9c40f6 | 2609 | { |
e1f44ac0 CH |
2610 | blk_status_t ret; |
2611 | ||
2612 | if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(rq->q)) { | |
710fa378 | 2613 | blk_mq_insert_request(rq, 0); |
e1f44ac0 CH |
2614 | return; |
2615 | } | |
2616 | ||
dd6216bb | 2617 | if ((rq->rq_flags & RQF_USE_SCHED) || !blk_mq_get_budget_and_tag(rq)) { |
710fa378 | 2618 | blk_mq_insert_request(rq, 0); |
f0dbe6e8 | 2619 | blk_mq_run_hw_queue(hctx, false); |
e1f44ac0 CH |
2620 | return; |
2621 | } | |
fd9c40f6 | 2622 | |
e1f44ac0 CH |
2623 | ret = __blk_mq_issue_directly(hctx, rq, true); |
2624 | switch (ret) { | |
2625 | case BLK_STS_OK: | |
2626 | break; | |
2627 | case BLK_STS_RESOURCE: | |
2628 | case BLK_STS_DEV_RESOURCE: | |
2b597613 | 2629 | blk_mq_request_bypass_insert(rq, 0); |
2394395c | 2630 | blk_mq_run_hw_queue(hctx, false); |
e1f44ac0 CH |
2631 | break; |
2632 | default: | |
fd9c40f6 | 2633 | blk_mq_end_request(rq, ret); |
e1f44ac0 CH |
2634 | break; |
2635 | } | |
fd9c40f6 BVA |
2636 | } |
2637 | ||
06c8c691 | 2638 | static blk_status_t blk_mq_request_issue_directly(struct request *rq, bool last) |
fd9c40f6 | 2639 | { |
e1f44ac0 CH |
2640 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; |
2641 | ||
2642 | if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(rq->q)) { | |
710fa378 | 2643 | blk_mq_insert_request(rq, 0); |
e1f44ac0 CH |
2644 | return BLK_STS_OK; |
2645 | } | |
2646 | ||
2647 | if (!blk_mq_get_budget_and_tag(rq)) | |
2648 | return BLK_STS_RESOURCE; | |
2649 | return __blk_mq_issue_directly(hctx, rq, last); | |
5eb6126e CH |
2650 | } |
2651 | ||
3e368fb0 | 2652 | static void blk_mq_plug_issue_direct(struct blk_plug *plug) |
b84c5b50 CH |
2653 | { |
2654 | struct blk_mq_hw_ctx *hctx = NULL; | |
2655 | struct request *rq; | |
2656 | int queued = 0; | |
0d617a83 | 2657 | blk_status_t ret = BLK_STS_OK; |
b84c5b50 CH |
2658 | |
2659 | while ((rq = rq_list_pop(&plug->mq_list))) { | |
2660 | bool last = rq_list_empty(plug->mq_list); | |
b84c5b50 CH |
2661 | |
2662 | if (hctx != rq->mq_hctx) { | |
34c9f547 KS |
2663 | if (hctx) { |
2664 | blk_mq_commit_rqs(hctx, queued, false); | |
2665 | queued = 0; | |
2666 | } | |
b84c5b50 CH |
2667 | hctx = rq->mq_hctx; |
2668 | } | |
2669 | ||
2670 | ret = blk_mq_request_issue_directly(rq, last); | |
2671 | switch (ret) { | |
2672 | case BLK_STS_OK: | |
2673 | queued++; | |
2674 | break; | |
2675 | case BLK_STS_RESOURCE: | |
2676 | case BLK_STS_DEV_RESOURCE: | |
2b597613 | 2677 | blk_mq_request_bypass_insert(rq, 0); |
2394395c | 2678 | blk_mq_run_hw_queue(hctx, false); |
0d617a83 | 2679 | goto out; |
b84c5b50 CH |
2680 | default: |
2681 | blk_mq_end_request(rq, ret); | |
b84c5b50 CH |
2682 | break; |
2683 | } | |
2684 | } | |
2685 | ||
0d617a83 KS |
2686 | out: |
2687 | if (ret != BLK_STS_OK) | |
34c9f547 | 2688 | blk_mq_commit_rqs(hctx, queued, false); |
b84c5b50 CH |
2689 | } |
2690 | ||
518579a9 KB |
2691 | static void __blk_mq_flush_plug_list(struct request_queue *q, |
2692 | struct blk_plug *plug) | |
2693 | { | |
2694 | if (blk_queue_quiesced(q)) | |
2695 | return; | |
2696 | q->mq_ops->queue_rqs(&plug->mq_list); | |
2697 | } | |
2698 | ||
26fed4ac JA |
2699 | static void blk_mq_dispatch_plug_list(struct blk_plug *plug, bool from_sched) |
2700 | { | |
2701 | struct blk_mq_hw_ctx *this_hctx = NULL; | |
2702 | struct blk_mq_ctx *this_ctx = NULL; | |
2703 | struct request *requeue_list = NULL; | |
34e0a279 | 2704 | struct request **requeue_lastp = &requeue_list; |
26fed4ac | 2705 | unsigned int depth = 0; |
d97217e7 | 2706 | bool is_passthrough = false; |
26fed4ac JA |
2707 | LIST_HEAD(list); |
2708 | ||
2709 | do { | |
2710 | struct request *rq = rq_list_pop(&plug->mq_list); | |
2711 | ||
2712 | if (!this_hctx) { | |
2713 | this_hctx = rq->mq_hctx; | |
2714 | this_ctx = rq->mq_ctx; | |
d97217e7 ML |
2715 | is_passthrough = blk_rq_is_passthrough(rq); |
2716 | } else if (this_hctx != rq->mq_hctx || this_ctx != rq->mq_ctx || | |
2717 | is_passthrough != blk_rq_is_passthrough(rq)) { | |
34e0a279 | 2718 | rq_list_add_tail(&requeue_lastp, rq); |
26fed4ac JA |
2719 | continue; |
2720 | } | |
34e0a279 | 2721 | list_add(&rq->queuelist, &list); |
26fed4ac JA |
2722 | depth++; |
2723 | } while (!rq_list_empty(plug->mq_list)); | |
2724 | ||
2725 | plug->mq_list = requeue_list; | |
2726 | trace_block_unplug(this_hctx->queue, depth, !from_sched); | |
05a93117 CH |
2727 | |
2728 | percpu_ref_get(&this_hctx->queue->q_usage_counter); | |
d97217e7 ML |
2729 | /* passthrough requests should never be issued to the I/O scheduler */ |
2730 | if (this_hctx->queue->elevator && !is_passthrough) { | |
05a93117 | 2731 | this_hctx->queue->elevator->type->ops.insert_requests(this_hctx, |
93fffe16 | 2732 | &list, 0); |
05a93117 CH |
2733 | blk_mq_run_hw_queue(this_hctx, from_sched); |
2734 | } else { | |
2735 | blk_mq_insert_requests(this_hctx, this_ctx, &list, from_sched); | |
2736 | } | |
2737 | percpu_ref_put(&this_hctx->queue->q_usage_counter); | |
26fed4ac JA |
2738 | } |
2739 | ||
b84c5b50 CH |
2740 | void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule) |
2741 | { | |
3c67d44d | 2742 | struct request *rq; |
b84c5b50 CH |
2743 | |
2744 | if (rq_list_empty(plug->mq_list)) | |
2745 | return; | |
2746 | plug->rq_count = 0; | |
2747 | ||
2748 | if (!plug->multiple_queues && !plug->has_elevator && !from_schedule) { | |
3c67d44d JA |
2749 | struct request_queue *q; |
2750 | ||
2751 | rq = rq_list_peek(&plug->mq_list); | |
2752 | q = rq->q; | |
2753 | ||
2754 | /* | |
2755 | * Peek first request and see if we have a ->queue_rqs() hook. | |
2756 | * If we do, we can dispatch the whole plug list in one go. We | |
2757 | * already know at this point that all requests belong to the | |
2758 | * same queue, caller must ensure that's the case. | |
2759 | * | |
2760 | * Since we pass off the full list to the driver at this point, | |
2761 | * we do not increment the active request count for the queue. | |
2762 | * Bypass shared tags for now because of that. | |
2763 | */ | |
2764 | if (q->mq_ops->queue_rqs && | |
2765 | !(rq->mq_hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) { | |
2766 | blk_mq_run_dispatch_ops(q, | |
518579a9 | 2767 | __blk_mq_flush_plug_list(q, plug)); |
3c67d44d JA |
2768 | if (rq_list_empty(plug->mq_list)) |
2769 | return; | |
2770 | } | |
73f3760e ML |
2771 | |
2772 | blk_mq_run_dispatch_ops(q, | |
3e368fb0 | 2773 | blk_mq_plug_issue_direct(plug)); |
b84c5b50 CH |
2774 | if (rq_list_empty(plug->mq_list)) |
2775 | return; | |
2776 | } | |
2777 | ||
b84c5b50 | 2778 | do { |
26fed4ac | 2779 | blk_mq_dispatch_plug_list(plug, from_schedule); |
b84c5b50 | 2780 | } while (!rq_list_empty(plug->mq_list)); |
b84c5b50 CH |
2781 | } |
2782 | ||
94aa228c | 2783 | static void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx, |
6ce3dd6e ML |
2784 | struct list_head *list) |
2785 | { | |
536167d4 | 2786 | int queued = 0; |
984ce0a7 | 2787 | blk_status_t ret = BLK_STS_OK; |
536167d4 | 2788 | |
6ce3dd6e | 2789 | while (!list_empty(list)) { |
6ce3dd6e ML |
2790 | struct request *rq = list_first_entry(list, struct request, |
2791 | queuelist); | |
2792 | ||
2793 | list_del_init(&rq->queuelist); | |
fd9c40f6 | 2794 | ret = blk_mq_request_issue_directly(rq, list_empty(list)); |
27e8b2bb KS |
2795 | switch (ret) { |
2796 | case BLK_STS_OK: | |
536167d4 | 2797 | queued++; |
27e8b2bb KS |
2798 | break; |
2799 | case BLK_STS_RESOURCE: | |
2800 | case BLK_STS_DEV_RESOURCE: | |
2b597613 | 2801 | blk_mq_request_bypass_insert(rq, 0); |
2394395c CH |
2802 | if (list_empty(list)) |
2803 | blk_mq_run_hw_queue(hctx, false); | |
27e8b2bb KS |
2804 | goto out; |
2805 | default: | |
2806 | blk_mq_end_request(rq, ret); | |
2807 | break; | |
2808 | } | |
6ce3dd6e | 2809 | } |
d666ba98 | 2810 | |
27e8b2bb | 2811 | out: |
984ce0a7 KS |
2812 | if (ret != BLK_STS_OK) |
2813 | blk_mq_commit_rqs(hctx, queued, false); | |
6ce3dd6e ML |
2814 | } |
2815 | ||
b131f201 | 2816 | static bool blk_mq_attempt_bio_merge(struct request_queue *q, |
0c5bcc92 | 2817 | struct bio *bio, unsigned int nr_segs) |
900e0807 JA |
2818 | { |
2819 | if (!blk_queue_nomerges(q) && bio_mergeable(bio)) { | |
0c5bcc92 | 2820 | if (blk_attempt_plug_merge(q, bio, nr_segs)) |
900e0807 JA |
2821 | return true; |
2822 | if (blk_mq_sched_bio_merge(q, bio, nr_segs)) | |
2823 | return true; | |
2824 | } | |
2825 | return false; | |
2826 | } | |
2827 | ||
71539717 JA |
2828 | static struct request *blk_mq_get_new_requests(struct request_queue *q, |
2829 | struct blk_plug *plug, | |
0a5aa8d1 SK |
2830 | struct bio *bio, |
2831 | unsigned int nsegs) | |
71539717 JA |
2832 | { |
2833 | struct blk_mq_alloc_data data = { | |
2834 | .q = q, | |
2835 | .nr_tags = 1, | |
9d497e29 | 2836 | .cmd_flags = bio->bi_opf, |
71539717 JA |
2837 | }; |
2838 | struct request *rq; | |
2839 | ||
5b13bc8a | 2840 | if (unlikely(bio_queue_enter(bio))) |
b637108a | 2841 | return NULL; |
900e0807 | 2842 | |
0a5aa8d1 SK |
2843 | if (blk_mq_attempt_bio_merge(q, bio, nsegs)) |
2844 | goto queue_exit; | |
2845 | ||
2846 | rq_qos_throttle(q, bio); | |
2847 | ||
71539717 JA |
2848 | if (plug) { |
2849 | data.nr_tags = plug->nr_ios; | |
2850 | plug->nr_ios = 1; | |
2851 | data.cached_rq = &plug->cached_rq; | |
2852 | } | |
2853 | ||
2854 | rq = __blk_mq_alloc_requests(&data); | |
373b5416 JA |
2855 | if (rq) |
2856 | return rq; | |
71539717 JA |
2857 | rq_qos_cleanup(q, bio); |
2858 | if (bio->bi_opf & REQ_NOWAIT) | |
2859 | bio_wouldblock_error(bio); | |
0a5aa8d1 | 2860 | queue_exit: |
5b13bc8a | 2861 | blk_queue_exit(q); |
71539717 JA |
2862 | return NULL; |
2863 | } | |
2864 | ||
5b13bc8a | 2865 | static inline struct request *blk_mq_get_cached_request(struct request_queue *q, |
0a5aa8d1 | 2866 | struct blk_plug *plug, struct bio **bio, unsigned int nsegs) |
71539717 | 2867 | { |
b637108a | 2868 | struct request *rq; |
77465647 | 2869 | enum hctx_type type, hctx_type; |
b637108a | 2870 | |
5b13bc8a CH |
2871 | if (!plug) |
2872 | return NULL; | |
81ea1222 ML |
2873 | rq = rq_list_peek(&plug->cached_rq); |
2874 | if (!rq || rq->q != q) | |
2875 | return NULL; | |
71539717 | 2876 | |
0a5aa8d1 SK |
2877 | if (blk_mq_attempt_bio_merge(q, *bio, nsegs)) { |
2878 | *bio = NULL; | |
2879 | return NULL; | |
2880 | } | |
2881 | ||
77465647 PB |
2882 | type = blk_mq_get_hctx_type((*bio)->bi_opf); |
2883 | hctx_type = rq->mq_hctx->type; | |
2884 | if (type != hctx_type && | |
2885 | !(type == HCTX_TYPE_READ && hctx_type == HCTX_TYPE_DEFAULT)) | |
5b13bc8a | 2886 | return NULL; |
0a5aa8d1 | 2887 | if (op_is_flush(rq->cmd_flags) != op_is_flush((*bio)->bi_opf)) |
5b13bc8a CH |
2888 | return NULL; |
2889 | ||
2645672f JA |
2890 | /* |
2891 | * If any qos ->throttle() end up blocking, we will have flushed the | |
2892 | * plug and hence killed the cached_rq list as well. Pop this entry | |
2893 | * before we throttle. | |
2894 | */ | |
5b13bc8a | 2895 | plug->cached_rq = rq_list_next(rq); |
2645672f JA |
2896 | rq_qos_throttle(q, *bio); |
2897 | ||
2898 | rq->cmd_flags = (*bio)->bi_opf; | |
5b13bc8a | 2899 | INIT_LIST_HEAD(&rq->queuelist); |
5b13bc8a | 2900 | return rq; |
71539717 JA |
2901 | } |
2902 | ||
82b74cac JK |
2903 | static void bio_set_ioprio(struct bio *bio) |
2904 | { | |
a78418e6 JK |
2905 | /* Nobody set ioprio so far? Initialize it based on task's nice value */ |
2906 | if (IOPRIO_PRIO_CLASS(bio->bi_ioprio) == IOPRIO_CLASS_NONE) | |
2907 | bio->bi_ioprio = get_current_ioprio(); | |
82b74cac JK |
2908 | blkcg_set_ioprio(bio); |
2909 | } | |
2910 | ||
105663f7 | 2911 | /** |
c62b37d9 | 2912 | * blk_mq_submit_bio - Create and send a request to block device. |
105663f7 AA |
2913 | * @bio: Bio pointer. |
2914 | * | |
2915 | * Builds up a request structure from @q and @bio and send to the device. The | |
2916 | * request may not be queued directly to hardware if: | |
2917 | * * This request can be merged with another one | |
2918 | * * We want to place request at plug queue for possible future merging | |
2919 | * * There is an IO scheduler active at this queue | |
2920 | * | |
2921 | * It will not queue the request if there is an error with the bio, or at the | |
2922 | * request creation. | |
105663f7 | 2923 | */ |
3e08773c | 2924 | void blk_mq_submit_bio(struct bio *bio) |
07068d5b | 2925 | { |
ed6cddef | 2926 | struct request_queue *q = bdev_get_queue(bio->bi_bdev); |
6deacb3b | 2927 | struct blk_plug *plug = blk_mq_plug(bio); |
ef295ecf | 2928 | const int is_sync = op_is_sync(bio->bi_opf); |
f0dbe6e8 | 2929 | struct blk_mq_hw_ctx *hctx; |
07068d5b | 2930 | struct request *rq; |
abd45c15 | 2931 | unsigned int nr_segs = 1; |
a892c8d5 | 2932 | blk_status_t ret; |
07068d5b | 2933 | |
51d798cd | 2934 | bio = blk_queue_bounce(bio, q); |
613b1488 | 2935 | if (bio_may_exceed_limits(bio, &q->limits)) { |
c55ddd90 | 2936 | bio = __bio_split_to_limits(bio, &q->limits, &nr_segs); |
613b1488 JA |
2937 | if (!bio) |
2938 | return; | |
2939 | } | |
f36ea50c | 2940 | |
e23947bd | 2941 | if (!bio_integrity_prep(bio)) |
900e0807 | 2942 | return; |
87760e5e | 2943 | |
9c6227e0 JK |
2944 | bio_set_ioprio(bio); |
2945 | ||
0a5aa8d1 | 2946 | rq = blk_mq_get_cached_request(q, plug, &bio, nr_segs); |
5b13bc8a | 2947 | if (!rq) { |
0a5aa8d1 SK |
2948 | if (!bio) |
2949 | return; | |
2950 | rq = blk_mq_get_new_requests(q, plug, bio, nr_segs); | |
5b13bc8a CH |
2951 | if (unlikely(!rq)) |
2952 | return; | |
2953 | } | |
87760e5e | 2954 | |
e8a676d6 | 2955 | trace_block_getrq(bio); |
d6f1dda2 | 2956 | |
c1c80384 | 2957 | rq_qos_track(q, rq, bio); |
07068d5b | 2958 | |
970d168d BVA |
2959 | blk_mq_bio_to_request(rq, bio, nr_segs); |
2960 | ||
9cd1e566 | 2961 | ret = blk_crypto_rq_get_keyslot(rq); |
a892c8d5 ST |
2962 | if (ret != BLK_STS_OK) { |
2963 | bio->bi_status = ret; | |
2964 | bio_endio(bio); | |
2965 | blk_mq_free_request(rq); | |
3e08773c | 2966 | return; |
a892c8d5 ST |
2967 | } |
2968 | ||
360f2648 | 2969 | if (op_is_flush(bio->bi_opf) && blk_insert_flush(rq)) |
d92ca9d8 CH |
2970 | return; |
2971 | ||
f0dbe6e8 | 2972 | if (plug) { |
ce5b009c | 2973 | blk_add_rq_to_plug(plug, rq); |
f0dbe6e8 CH |
2974 | return; |
2975 | } | |
2976 | ||
2977 | hctx = rq->mq_hctx; | |
dd6216bb | 2978 | if ((rq->rq_flags & RQF_USE_SCHED) || |
f0dbe6e8 | 2979 | (hctx->dispatch_busy && (q->nr_hw_queues == 1 || !is_sync))) { |
710fa378 | 2980 | blk_mq_insert_request(rq, 0); |
f0dbe6e8 CH |
2981 | blk_mq_run_hw_queue(hctx, true); |
2982 | } else { | |
2983 | blk_mq_run_dispatch_ops(q, blk_mq_try_issue_directly(hctx, rq)); | |
2984 | } | |
320ae51f JA |
2985 | } |
2986 | ||
248c7933 | 2987 | #ifdef CONFIG_BLK_MQ_STACKING |
06c8c691 | 2988 | /** |
a5efda3c | 2989 | * blk_insert_cloned_request - Helper for stacking drivers to submit a request |
a5efda3c | 2990 | * @rq: the request being queued |
06c8c691 | 2991 | */ |
28db4711 | 2992 | blk_status_t blk_insert_cloned_request(struct request *rq) |
06c8c691 | 2993 | { |
28db4711 | 2994 | struct request_queue *q = rq->q; |
06c8c691 | 2995 | unsigned int max_sectors = blk_queue_get_max_sectors(q, req_op(rq)); |
49d24398 | 2996 | unsigned int max_segments = blk_rq_get_max_segments(rq); |
a5efda3c | 2997 | blk_status_t ret; |
06c8c691 CH |
2998 | |
2999 | if (blk_rq_sectors(rq) > max_sectors) { | |
3000 | /* | |
3001 | * SCSI device does not have a good way to return if | |
3002 | * Write Same/Zero is actually supported. If a device rejects | |
3003 | * a non-read/write command (discard, write same,etc.) the | |
3004 | * low-level device driver will set the relevant queue limit to | |
3005 | * 0 to prevent blk-lib from issuing more of the offending | |
3006 | * operations. Commands queued prior to the queue limit being | |
3007 | * reset need to be completed with BLK_STS_NOTSUPP to avoid I/O | |
3008 | * errors being propagated to upper layers. | |
3009 | */ | |
3010 | if (max_sectors == 0) | |
3011 | return BLK_STS_NOTSUPP; | |
3012 | ||
3013 | printk(KERN_ERR "%s: over max size limit. (%u > %u)\n", | |
3014 | __func__, blk_rq_sectors(rq), max_sectors); | |
3015 | return BLK_STS_IOERR; | |
3016 | } | |
3017 | ||
3018 | /* | |
3019 | * The queue settings related to segment counting may differ from the | |
3020 | * original queue. | |
3021 | */ | |
3022 | rq->nr_phys_segments = blk_recalc_rq_segments(rq); | |
49d24398 US |
3023 | if (rq->nr_phys_segments > max_segments) { |
3024 | printk(KERN_ERR "%s: over max segments limit. (%u > %u)\n", | |
3025 | __func__, rq->nr_phys_segments, max_segments); | |
06c8c691 CH |
3026 | return BLK_STS_IOERR; |
3027 | } | |
3028 | ||
28db4711 | 3029 | if (q->disk && should_fail_request(q->disk->part0, blk_rq_bytes(rq))) |
06c8c691 CH |
3030 | return BLK_STS_IOERR; |
3031 | ||
5b8562f0 EB |
3032 | ret = blk_crypto_rq_get_keyslot(rq); |
3033 | if (ret != BLK_STS_OK) | |
3034 | return ret; | |
06c8c691 CH |
3035 | |
3036 | blk_account_io_start(rq); | |
3037 | ||
3038 | /* | |
3039 | * Since we have a scheduler attached on the top device, | |
3040 | * bypass a potential scheduler on the bottom device for | |
3041 | * insert. | |
3042 | */ | |
28db4711 | 3043 | blk_mq_run_dispatch_ops(q, |
4cafe86c | 3044 | ret = blk_mq_request_issue_directly(rq, true)); |
592ee119 YK |
3045 | if (ret) |
3046 | blk_account_io_done(rq, ktime_get_ns()); | |
4cafe86c | 3047 | return ret; |
06c8c691 CH |
3048 | } |
3049 | EXPORT_SYMBOL_GPL(blk_insert_cloned_request); | |
3050 | ||
3051 | /** | |
3052 | * blk_rq_unprep_clone - Helper function to free all bios in a cloned request | |
3053 | * @rq: the clone request to be cleaned up | |
3054 | * | |
3055 | * Description: | |
3056 | * Free all bios in @rq for a cloned request. | |
3057 | */ | |
3058 | void blk_rq_unprep_clone(struct request *rq) | |
3059 | { | |
3060 | struct bio *bio; | |
3061 | ||
3062 | while ((bio = rq->bio) != NULL) { | |
3063 | rq->bio = bio->bi_next; | |
3064 | ||
3065 | bio_put(bio); | |
3066 | } | |
3067 | } | |
3068 | EXPORT_SYMBOL_GPL(blk_rq_unprep_clone); | |
3069 | ||
3070 | /** | |
3071 | * blk_rq_prep_clone - Helper function to setup clone request | |
3072 | * @rq: the request to be setup | |
3073 | * @rq_src: original request to be cloned | |
3074 | * @bs: bio_set that bios for clone are allocated from | |
3075 | * @gfp_mask: memory allocation mask for bio | |
3076 | * @bio_ctr: setup function to be called for each clone bio. | |
3077 | * Returns %0 for success, non %0 for failure. | |
3078 | * @data: private data to be passed to @bio_ctr | |
3079 | * | |
3080 | * Description: | |
3081 | * Clones bios in @rq_src to @rq, and copies attributes of @rq_src to @rq. | |
3082 | * Also, pages which the original bios are pointing to are not copied | |
3083 | * and the cloned bios just point same pages. | |
3084 | * So cloned bios must be completed before original bios, which means | |
3085 | * the caller must complete @rq before @rq_src. | |
3086 | */ | |
3087 | int blk_rq_prep_clone(struct request *rq, struct request *rq_src, | |
3088 | struct bio_set *bs, gfp_t gfp_mask, | |
3089 | int (*bio_ctr)(struct bio *, struct bio *, void *), | |
3090 | void *data) | |
3091 | { | |
3092 | struct bio *bio, *bio_src; | |
3093 | ||
3094 | if (!bs) | |
3095 | bs = &fs_bio_set; | |
3096 | ||
3097 | __rq_for_each_bio(bio_src, rq_src) { | |
abfc426d CH |
3098 | bio = bio_alloc_clone(rq->q->disk->part0, bio_src, gfp_mask, |
3099 | bs); | |
06c8c691 CH |
3100 | if (!bio) |
3101 | goto free_and_out; | |
3102 | ||
3103 | if (bio_ctr && bio_ctr(bio, bio_src, data)) | |
3104 | goto free_and_out; | |
3105 | ||
3106 | if (rq->bio) { | |
3107 | rq->biotail->bi_next = bio; | |
3108 | rq->biotail = bio; | |
3109 | } else { | |
3110 | rq->bio = rq->biotail = bio; | |
3111 | } | |
3112 | bio = NULL; | |
3113 | } | |
3114 | ||
3115 | /* Copy attributes of the original request to the clone request. */ | |
3116 | rq->__sector = blk_rq_pos(rq_src); | |
3117 | rq->__data_len = blk_rq_bytes(rq_src); | |
3118 | if (rq_src->rq_flags & RQF_SPECIAL_PAYLOAD) { | |
3119 | rq->rq_flags |= RQF_SPECIAL_PAYLOAD; | |
3120 | rq->special_vec = rq_src->special_vec; | |
3121 | } | |
3122 | rq->nr_phys_segments = rq_src->nr_phys_segments; | |
3123 | rq->ioprio = rq_src->ioprio; | |
3124 | ||
3125 | if (rq->bio && blk_crypto_rq_bio_prep(rq, rq->bio, gfp_mask) < 0) | |
3126 | goto free_and_out; | |
3127 | ||
3128 | return 0; | |
3129 | ||
3130 | free_and_out: | |
3131 | if (bio) | |
3132 | bio_put(bio); | |
3133 | blk_rq_unprep_clone(rq); | |
3134 | ||
3135 | return -ENOMEM; | |
3136 | } | |
3137 | EXPORT_SYMBOL_GPL(blk_rq_prep_clone); | |
248c7933 | 3138 | #endif /* CONFIG_BLK_MQ_STACKING */ |
06c8c691 | 3139 | |
f2b8f3ce CH |
3140 | /* |
3141 | * Steal bios from a request and add them to a bio list. | |
3142 | * The request must not have been partially completed before. | |
3143 | */ | |
3144 | void blk_steal_bios(struct bio_list *list, struct request *rq) | |
3145 | { | |
3146 | if (rq->bio) { | |
3147 | if (list->tail) | |
3148 | list->tail->bi_next = rq->bio; | |
3149 | else | |
3150 | list->head = rq->bio; | |
3151 | list->tail = rq->biotail; | |
3152 | ||
3153 | rq->bio = NULL; | |
3154 | rq->biotail = NULL; | |
3155 | } | |
3156 | ||
3157 | rq->__data_len = 0; | |
3158 | } | |
3159 | EXPORT_SYMBOL_GPL(blk_steal_bios); | |
3160 | ||
bd63141d ML |
3161 | static size_t order_to_size(unsigned int order) |
3162 | { | |
3163 | return (size_t)PAGE_SIZE << order; | |
3164 | } | |
3165 | ||
3166 | /* called before freeing request pool in @tags */ | |
f32e4eaf JG |
3167 | static void blk_mq_clear_rq_mapping(struct blk_mq_tags *drv_tags, |
3168 | struct blk_mq_tags *tags) | |
bd63141d | 3169 | { |
bd63141d ML |
3170 | struct page *page; |
3171 | unsigned long flags; | |
3172 | ||
76dd2980 YK |
3173 | /* |
3174 | * There is no need to clear mapping if driver tags is not initialized | |
3175 | * or the mapping belongs to the driver tags. | |
3176 | */ | |
3177 | if (!drv_tags || drv_tags == tags) | |
4f245d5b JG |
3178 | return; |
3179 | ||
bd63141d ML |
3180 | list_for_each_entry(page, &tags->page_list, lru) { |
3181 | unsigned long start = (unsigned long)page_address(page); | |
3182 | unsigned long end = start + order_to_size(page->private); | |
3183 | int i; | |
3184 | ||
f32e4eaf | 3185 | for (i = 0; i < drv_tags->nr_tags; i++) { |
bd63141d ML |
3186 | struct request *rq = drv_tags->rqs[i]; |
3187 | unsigned long rq_addr = (unsigned long)rq; | |
3188 | ||
3189 | if (rq_addr >= start && rq_addr < end) { | |
0a467d0f | 3190 | WARN_ON_ONCE(req_ref_read(rq) != 0); |
bd63141d ML |
3191 | cmpxchg(&drv_tags->rqs[i], rq, NULL); |
3192 | } | |
3193 | } | |
3194 | } | |
3195 | ||
3196 | /* | |
3197 | * Wait until all pending iteration is done. | |
3198 | * | |
3199 | * Request reference is cleared and it is guaranteed to be observed | |
3200 | * after the ->lock is released. | |
3201 | */ | |
3202 | spin_lock_irqsave(&drv_tags->lock, flags); | |
3203 | spin_unlock_irqrestore(&drv_tags->lock, flags); | |
3204 | } | |
3205 | ||
cc71a6f4 JA |
3206 | void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags, |
3207 | unsigned int hctx_idx) | |
95363efd | 3208 | { |
f32e4eaf | 3209 | struct blk_mq_tags *drv_tags; |
e9b267d9 | 3210 | struct page *page; |
320ae51f | 3211 | |
e02657ea ML |
3212 | if (list_empty(&tags->page_list)) |
3213 | return; | |
3214 | ||
079a2e3e JG |
3215 | if (blk_mq_is_shared_tags(set->flags)) |
3216 | drv_tags = set->shared_tags; | |
e155b0c2 JG |
3217 | else |
3218 | drv_tags = set->tags[hctx_idx]; | |
f32e4eaf | 3219 | |
65de57bb | 3220 | if (tags->static_rqs && set->ops->exit_request) { |
e9b267d9 | 3221 | int i; |
320ae51f | 3222 | |
24d2f903 | 3223 | for (i = 0; i < tags->nr_tags; i++) { |
2af8cbe3 JA |
3224 | struct request *rq = tags->static_rqs[i]; |
3225 | ||
3226 | if (!rq) | |
e9b267d9 | 3227 | continue; |
d6296d39 | 3228 | set->ops->exit_request(set, rq, hctx_idx); |
2af8cbe3 | 3229 | tags->static_rqs[i] = NULL; |
e9b267d9 | 3230 | } |
320ae51f | 3231 | } |
320ae51f | 3232 | |
f32e4eaf | 3233 | blk_mq_clear_rq_mapping(drv_tags, tags); |
bd63141d | 3234 | |
24d2f903 CH |
3235 | while (!list_empty(&tags->page_list)) { |
3236 | page = list_first_entry(&tags->page_list, struct page, lru); | |
6753471c | 3237 | list_del_init(&page->lru); |
f75782e4 CM |
3238 | /* |
3239 | * Remove kmemleak object previously allocated in | |
273938bf | 3240 | * blk_mq_alloc_rqs(). |
f75782e4 CM |
3241 | */ |
3242 | kmemleak_free(page_address(page)); | |
320ae51f JA |
3243 | __free_pages(page, page->private); |
3244 | } | |
cc71a6f4 | 3245 | } |
320ae51f | 3246 | |
e155b0c2 | 3247 | void blk_mq_free_rq_map(struct blk_mq_tags *tags) |
cc71a6f4 | 3248 | { |
24d2f903 | 3249 | kfree(tags->rqs); |
cc71a6f4 | 3250 | tags->rqs = NULL; |
2af8cbe3 JA |
3251 | kfree(tags->static_rqs); |
3252 | tags->static_rqs = NULL; | |
320ae51f | 3253 | |
e155b0c2 | 3254 | blk_mq_free_tags(tags); |
320ae51f JA |
3255 | } |
3256 | ||
4d805131 ML |
3257 | static enum hctx_type hctx_idx_to_type(struct blk_mq_tag_set *set, |
3258 | unsigned int hctx_idx) | |
3259 | { | |
3260 | int i; | |
3261 | ||
3262 | for (i = 0; i < set->nr_maps; i++) { | |
3263 | unsigned int start = set->map[i].queue_offset; | |
3264 | unsigned int end = start + set->map[i].nr_queues; | |
3265 | ||
3266 | if (hctx_idx >= start && hctx_idx < end) | |
3267 | break; | |
3268 | } | |
3269 | ||
3270 | if (i >= set->nr_maps) | |
3271 | i = HCTX_TYPE_DEFAULT; | |
3272 | ||
3273 | return i; | |
3274 | } | |
3275 | ||
3276 | static int blk_mq_get_hctx_node(struct blk_mq_tag_set *set, | |
3277 | unsigned int hctx_idx) | |
3278 | { | |
3279 | enum hctx_type type = hctx_idx_to_type(set, hctx_idx); | |
3280 | ||
3281 | return blk_mq_hw_queue_to_node(&set->map[type], hctx_idx); | |
3282 | } | |
3283 | ||
63064be1 JG |
3284 | static struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, |
3285 | unsigned int hctx_idx, | |
3286 | unsigned int nr_tags, | |
e155b0c2 | 3287 | unsigned int reserved_tags) |
320ae51f | 3288 | { |
4d805131 | 3289 | int node = blk_mq_get_hctx_node(set, hctx_idx); |
24d2f903 | 3290 | struct blk_mq_tags *tags; |
320ae51f | 3291 | |
59f082e4 SL |
3292 | if (node == NUMA_NO_NODE) |
3293 | node = set->numa_node; | |
3294 | ||
e155b0c2 JG |
3295 | tags = blk_mq_init_tags(nr_tags, reserved_tags, node, |
3296 | BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags)); | |
24d2f903 CH |
3297 | if (!tags) |
3298 | return NULL; | |
320ae51f | 3299 | |
590b5b7d | 3300 | tags->rqs = kcalloc_node(nr_tags, sizeof(struct request *), |
36e1f3d1 | 3301 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, |
59f082e4 | 3302 | node); |
7edfd681 JC |
3303 | if (!tags->rqs) |
3304 | goto err_free_tags; | |
320ae51f | 3305 | |
590b5b7d KC |
3306 | tags->static_rqs = kcalloc_node(nr_tags, sizeof(struct request *), |
3307 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, | |
3308 | node); | |
7edfd681 JC |
3309 | if (!tags->static_rqs) |
3310 | goto err_free_rqs; | |
2af8cbe3 | 3311 | |
cc71a6f4 | 3312 | return tags; |
7edfd681 JC |
3313 | |
3314 | err_free_rqs: | |
3315 | kfree(tags->rqs); | |
3316 | err_free_tags: | |
3317 | blk_mq_free_tags(tags); | |
3318 | return NULL; | |
cc71a6f4 JA |
3319 | } |
3320 | ||
1d9bd516 TH |
3321 | static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq, |
3322 | unsigned int hctx_idx, int node) | |
3323 | { | |
3324 | int ret; | |
3325 | ||
3326 | if (set->ops->init_request) { | |
3327 | ret = set->ops->init_request(set, rq, hctx_idx, node); | |
3328 | if (ret) | |
3329 | return ret; | |
3330 | } | |
3331 | ||
12f5b931 | 3332 | WRITE_ONCE(rq->state, MQ_RQ_IDLE); |
1d9bd516 TH |
3333 | return 0; |
3334 | } | |
3335 | ||
63064be1 JG |
3336 | static int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, |
3337 | struct blk_mq_tags *tags, | |
3338 | unsigned int hctx_idx, unsigned int depth) | |
cc71a6f4 JA |
3339 | { |
3340 | unsigned int i, j, entries_per_page, max_order = 4; | |
4d805131 | 3341 | int node = blk_mq_get_hctx_node(set, hctx_idx); |
cc71a6f4 | 3342 | size_t rq_size, left; |
59f082e4 | 3343 | |
59f082e4 SL |
3344 | if (node == NUMA_NO_NODE) |
3345 | node = set->numa_node; | |
cc71a6f4 JA |
3346 | |
3347 | INIT_LIST_HEAD(&tags->page_list); | |
3348 | ||
320ae51f JA |
3349 | /* |
3350 | * rq_size is the size of the request plus driver payload, rounded | |
3351 | * to the cacheline size | |
3352 | */ | |
24d2f903 | 3353 | rq_size = round_up(sizeof(struct request) + set->cmd_size, |
320ae51f | 3354 | cache_line_size()); |
cc71a6f4 | 3355 | left = rq_size * depth; |
320ae51f | 3356 | |
cc71a6f4 | 3357 | for (i = 0; i < depth; ) { |
320ae51f JA |
3358 | int this_order = max_order; |
3359 | struct page *page; | |
3360 | int to_do; | |
3361 | void *p; | |
3362 | ||
b3a834b1 | 3363 | while (this_order && left < order_to_size(this_order - 1)) |
320ae51f JA |
3364 | this_order--; |
3365 | ||
3366 | do { | |
59f082e4 | 3367 | page = alloc_pages_node(node, |
36e1f3d1 | 3368 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO, |
a5164405 | 3369 | this_order); |
320ae51f JA |
3370 | if (page) |
3371 | break; | |
3372 | if (!this_order--) | |
3373 | break; | |
3374 | if (order_to_size(this_order) < rq_size) | |
3375 | break; | |
3376 | } while (1); | |
3377 | ||
3378 | if (!page) | |
24d2f903 | 3379 | goto fail; |
320ae51f JA |
3380 | |
3381 | page->private = this_order; | |
24d2f903 | 3382 | list_add_tail(&page->lru, &tags->page_list); |
320ae51f JA |
3383 | |
3384 | p = page_address(page); | |
f75782e4 CM |
3385 | /* |
3386 | * Allow kmemleak to scan these pages as they contain pointers | |
3387 | * to additional allocations like via ops->init_request(). | |
3388 | */ | |
36e1f3d1 | 3389 | kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO); |
320ae51f | 3390 | entries_per_page = order_to_size(this_order) / rq_size; |
cc71a6f4 | 3391 | to_do = min(entries_per_page, depth - i); |
320ae51f JA |
3392 | left -= to_do * rq_size; |
3393 | for (j = 0; j < to_do; j++) { | |
2af8cbe3 JA |
3394 | struct request *rq = p; |
3395 | ||
3396 | tags->static_rqs[i] = rq; | |
1d9bd516 TH |
3397 | if (blk_mq_init_request(set, rq, hctx_idx, node)) { |
3398 | tags->static_rqs[i] = NULL; | |
3399 | goto fail; | |
e9b267d9 CH |
3400 | } |
3401 | ||
320ae51f JA |
3402 | p += rq_size; |
3403 | i++; | |
3404 | } | |
3405 | } | |
cc71a6f4 | 3406 | return 0; |
320ae51f | 3407 | |
24d2f903 | 3408 | fail: |
cc71a6f4 JA |
3409 | blk_mq_free_rqs(set, tags, hctx_idx); |
3410 | return -ENOMEM; | |
320ae51f JA |
3411 | } |
3412 | ||
bf0beec0 ML |
3413 | struct rq_iter_data { |
3414 | struct blk_mq_hw_ctx *hctx; | |
3415 | bool has_rq; | |
3416 | }; | |
3417 | ||
2dd6532e | 3418 | static bool blk_mq_has_request(struct request *rq, void *data) |
bf0beec0 ML |
3419 | { |
3420 | struct rq_iter_data *iter_data = data; | |
3421 | ||
3422 | if (rq->mq_hctx != iter_data->hctx) | |
3423 | return true; | |
3424 | iter_data->has_rq = true; | |
3425 | return false; | |
3426 | } | |
3427 | ||
3428 | static bool blk_mq_hctx_has_requests(struct blk_mq_hw_ctx *hctx) | |
3429 | { | |
3430 | struct blk_mq_tags *tags = hctx->sched_tags ? | |
3431 | hctx->sched_tags : hctx->tags; | |
3432 | struct rq_iter_data data = { | |
3433 | .hctx = hctx, | |
3434 | }; | |
3435 | ||
3436 | blk_mq_all_tag_iter(tags, blk_mq_has_request, &data); | |
3437 | return data.has_rq; | |
3438 | } | |
3439 | ||
3440 | static inline bool blk_mq_last_cpu_in_hctx(unsigned int cpu, | |
3441 | struct blk_mq_hw_ctx *hctx) | |
3442 | { | |
9b51d9d8 | 3443 | if (cpumask_first_and(hctx->cpumask, cpu_online_mask) != cpu) |
bf0beec0 ML |
3444 | return false; |
3445 | if (cpumask_next_and(cpu, hctx->cpumask, cpu_online_mask) < nr_cpu_ids) | |
3446 | return false; | |
3447 | return true; | |
3448 | } | |
3449 | ||
3450 | static int blk_mq_hctx_notify_offline(unsigned int cpu, struct hlist_node *node) | |
3451 | { | |
3452 | struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node, | |
3453 | struct blk_mq_hw_ctx, cpuhp_online); | |
3454 | ||
3455 | if (!cpumask_test_cpu(cpu, hctx->cpumask) || | |
3456 | !blk_mq_last_cpu_in_hctx(cpu, hctx)) | |
3457 | return 0; | |
3458 | ||
3459 | /* | |
3460 | * Prevent new request from being allocated on the current hctx. | |
3461 | * | |
3462 | * The smp_mb__after_atomic() Pairs with the implied barrier in | |
3463 | * test_and_set_bit_lock in sbitmap_get(). Ensures the inactive flag is | |
3464 | * seen once we return from the tag allocator. | |
3465 | */ | |
3466 | set_bit(BLK_MQ_S_INACTIVE, &hctx->state); | |
3467 | smp_mb__after_atomic(); | |
3468 | ||
3469 | /* | |
3470 | * Try to grab a reference to the queue and wait for any outstanding | |
3471 | * requests. If we could not grab a reference the queue has been | |
3472 | * frozen and there are no requests. | |
3473 | */ | |
3474 | if (percpu_ref_tryget(&hctx->queue->q_usage_counter)) { | |
3475 | while (blk_mq_hctx_has_requests(hctx)) | |
3476 | msleep(5); | |
3477 | percpu_ref_put(&hctx->queue->q_usage_counter); | |
3478 | } | |
3479 | ||
3480 | return 0; | |
3481 | } | |
3482 | ||
3483 | static int blk_mq_hctx_notify_online(unsigned int cpu, struct hlist_node *node) | |
3484 | { | |
3485 | struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node, | |
3486 | struct blk_mq_hw_ctx, cpuhp_online); | |
3487 | ||
3488 | if (cpumask_test_cpu(cpu, hctx->cpumask)) | |
3489 | clear_bit(BLK_MQ_S_INACTIVE, &hctx->state); | |
3490 | return 0; | |
3491 | } | |
3492 | ||
e57690fe JA |
3493 | /* |
3494 | * 'cpu' is going away. splice any existing rq_list entries from this | |
3495 | * software queue to the hw queue dispatch list, and ensure that it | |
3496 | * gets run. | |
3497 | */ | |
9467f859 | 3498 | static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node) |
484b4061 | 3499 | { |
9467f859 | 3500 | struct blk_mq_hw_ctx *hctx; |
484b4061 JA |
3501 | struct blk_mq_ctx *ctx; |
3502 | LIST_HEAD(tmp); | |
c16d6b5a | 3503 | enum hctx_type type; |
484b4061 | 3504 | |
9467f859 | 3505 | hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead); |
bf0beec0 ML |
3506 | if (!cpumask_test_cpu(cpu, hctx->cpumask)) |
3507 | return 0; | |
3508 | ||
e57690fe | 3509 | ctx = __blk_mq_get_ctx(hctx->queue, cpu); |
c16d6b5a | 3510 | type = hctx->type; |
484b4061 JA |
3511 | |
3512 | spin_lock(&ctx->lock); | |
c16d6b5a ML |
3513 | if (!list_empty(&ctx->rq_lists[type])) { |
3514 | list_splice_init(&ctx->rq_lists[type], &tmp); | |
484b4061 JA |
3515 | blk_mq_hctx_clear_pending(hctx, ctx); |
3516 | } | |
3517 | spin_unlock(&ctx->lock); | |
3518 | ||
3519 | if (list_empty(&tmp)) | |
9467f859 | 3520 | return 0; |
484b4061 | 3521 | |
e57690fe JA |
3522 | spin_lock(&hctx->lock); |
3523 | list_splice_tail_init(&tmp, &hctx->dispatch); | |
3524 | spin_unlock(&hctx->lock); | |
484b4061 JA |
3525 | |
3526 | blk_mq_run_hw_queue(hctx, true); | |
9467f859 | 3527 | return 0; |
484b4061 JA |
3528 | } |
3529 | ||
9467f859 | 3530 | static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx) |
484b4061 | 3531 | { |
bf0beec0 ML |
3532 | if (!(hctx->flags & BLK_MQ_F_STACKING)) |
3533 | cpuhp_state_remove_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE, | |
3534 | &hctx->cpuhp_online); | |
9467f859 TG |
3535 | cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD, |
3536 | &hctx->cpuhp_dead); | |
484b4061 JA |
3537 | } |
3538 | ||
364b6181 ML |
3539 | /* |
3540 | * Before freeing hw queue, clearing the flush request reference in | |
3541 | * tags->rqs[] for avoiding potential UAF. | |
3542 | */ | |
3543 | static void blk_mq_clear_flush_rq_mapping(struct blk_mq_tags *tags, | |
3544 | unsigned int queue_depth, struct request *flush_rq) | |
3545 | { | |
3546 | int i; | |
3547 | unsigned long flags; | |
3548 | ||
3549 | /* The hw queue may not be mapped yet */ | |
3550 | if (!tags) | |
3551 | return; | |
3552 | ||
0a467d0f | 3553 | WARN_ON_ONCE(req_ref_read(flush_rq) != 0); |
364b6181 ML |
3554 | |
3555 | for (i = 0; i < queue_depth; i++) | |
3556 | cmpxchg(&tags->rqs[i], flush_rq, NULL); | |
3557 | ||
3558 | /* | |
3559 | * Wait until all pending iteration is done. | |
3560 | * | |
3561 | * Request reference is cleared and it is guaranteed to be observed | |
3562 | * after the ->lock is released. | |
3563 | */ | |
3564 | spin_lock_irqsave(&tags->lock, flags); | |
3565 | spin_unlock_irqrestore(&tags->lock, flags); | |
3566 | } | |
3567 | ||
c3b4afca | 3568 | /* hctx->ctxs will be freed in queue's release handler */ |
08e98fc6 ML |
3569 | static void blk_mq_exit_hctx(struct request_queue *q, |
3570 | struct blk_mq_tag_set *set, | |
3571 | struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx) | |
3572 | { | |
364b6181 ML |
3573 | struct request *flush_rq = hctx->fq->flush_rq; |
3574 | ||
8ab0b7dc ML |
3575 | if (blk_mq_hw_queue_mapped(hctx)) |
3576 | blk_mq_tag_idle(hctx); | |
08e98fc6 | 3577 | |
6cfeadbf ML |
3578 | if (blk_queue_init_done(q)) |
3579 | blk_mq_clear_flush_rq_mapping(set->tags[hctx_idx], | |
3580 | set->queue_depth, flush_rq); | |
f70ced09 | 3581 | if (set->ops->exit_request) |
364b6181 | 3582 | set->ops->exit_request(set, flush_rq, hctx_idx); |
f70ced09 | 3583 | |
08e98fc6 ML |
3584 | if (set->ops->exit_hctx) |
3585 | set->ops->exit_hctx(hctx, hctx_idx); | |
3586 | ||
9467f859 | 3587 | blk_mq_remove_cpuhp(hctx); |
2f8f1336 | 3588 | |
4e5cc99e ML |
3589 | xa_erase(&q->hctx_table, hctx_idx); |
3590 | ||
2f8f1336 ML |
3591 | spin_lock(&q->unused_hctx_lock); |
3592 | list_add(&hctx->hctx_list, &q->unused_hctx_list); | |
3593 | spin_unlock(&q->unused_hctx_lock); | |
08e98fc6 ML |
3594 | } |
3595 | ||
624dbe47 ML |
3596 | static void blk_mq_exit_hw_queues(struct request_queue *q, |
3597 | struct blk_mq_tag_set *set, int nr_queue) | |
3598 | { | |
3599 | struct blk_mq_hw_ctx *hctx; | |
4f481208 | 3600 | unsigned long i; |
624dbe47 ML |
3601 | |
3602 | queue_for_each_hw_ctx(q, hctx, i) { | |
3603 | if (i == nr_queue) | |
3604 | break; | |
08e98fc6 | 3605 | blk_mq_exit_hctx(q, set, hctx, i); |
624dbe47 | 3606 | } |
624dbe47 ML |
3607 | } |
3608 | ||
08e98fc6 ML |
3609 | static int blk_mq_init_hctx(struct request_queue *q, |
3610 | struct blk_mq_tag_set *set, | |
3611 | struct blk_mq_hw_ctx *hctx, unsigned hctx_idx) | |
320ae51f | 3612 | { |
7c6c5b7c ML |
3613 | hctx->queue_num = hctx_idx; |
3614 | ||
bf0beec0 ML |
3615 | if (!(hctx->flags & BLK_MQ_F_STACKING)) |
3616 | cpuhp_state_add_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE, | |
3617 | &hctx->cpuhp_online); | |
7c6c5b7c ML |
3618 | cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead); |
3619 | ||
3620 | hctx->tags = set->tags[hctx_idx]; | |
3621 | ||
3622 | if (set->ops->init_hctx && | |
3623 | set->ops->init_hctx(hctx, set->driver_data, hctx_idx)) | |
3624 | goto unregister_cpu_notifier; | |
08e98fc6 | 3625 | |
7c6c5b7c ML |
3626 | if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx, |
3627 | hctx->numa_node)) | |
3628 | goto exit_hctx; | |
4e5cc99e ML |
3629 | |
3630 | if (xa_insert(&q->hctx_table, hctx_idx, hctx, GFP_KERNEL)) | |
3631 | goto exit_flush_rq; | |
3632 | ||
7c6c5b7c ML |
3633 | return 0; |
3634 | ||
4e5cc99e ML |
3635 | exit_flush_rq: |
3636 | if (set->ops->exit_request) | |
3637 | set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx); | |
7c6c5b7c ML |
3638 | exit_hctx: |
3639 | if (set->ops->exit_hctx) | |
3640 | set->ops->exit_hctx(hctx, hctx_idx); | |
3641 | unregister_cpu_notifier: | |
3642 | blk_mq_remove_cpuhp(hctx); | |
3643 | return -1; | |
3644 | } | |
3645 | ||
3646 | static struct blk_mq_hw_ctx * | |
3647 | blk_mq_alloc_hctx(struct request_queue *q, struct blk_mq_tag_set *set, | |
3648 | int node) | |
3649 | { | |
3650 | struct blk_mq_hw_ctx *hctx; | |
3651 | gfp_t gfp = GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY; | |
3652 | ||
704b914f | 3653 | hctx = kzalloc_node(sizeof(struct blk_mq_hw_ctx), gfp, node); |
7c6c5b7c ML |
3654 | if (!hctx) |
3655 | goto fail_alloc_hctx; | |
3656 | ||
3657 | if (!zalloc_cpumask_var_node(&hctx->cpumask, gfp, node)) | |
3658 | goto free_hctx; | |
3659 | ||
3660 | atomic_set(&hctx->nr_active, 0); | |
08e98fc6 | 3661 | if (node == NUMA_NO_NODE) |
7c6c5b7c ML |
3662 | node = set->numa_node; |
3663 | hctx->numa_node = node; | |
08e98fc6 | 3664 | |
9f993737 | 3665 | INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn); |
08e98fc6 ML |
3666 | spin_lock_init(&hctx->lock); |
3667 | INIT_LIST_HEAD(&hctx->dispatch); | |
3668 | hctx->queue = q; | |
51db1c37 | 3669 | hctx->flags = set->flags & ~BLK_MQ_F_TAG_QUEUE_SHARED; |
08e98fc6 | 3670 | |
2f8f1336 ML |
3671 | INIT_LIST_HEAD(&hctx->hctx_list); |
3672 | ||
320ae51f | 3673 | /* |
08e98fc6 ML |
3674 | * Allocate space for all possible cpus to avoid allocation at |
3675 | * runtime | |
320ae51f | 3676 | */ |
d904bfa7 | 3677 | hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *), |
7c6c5b7c | 3678 | gfp, node); |
08e98fc6 | 3679 | if (!hctx->ctxs) |
7c6c5b7c | 3680 | goto free_cpumask; |
320ae51f | 3681 | |
5b202853 | 3682 | if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8), |
c548e62b | 3683 | gfp, node, false, false)) |
08e98fc6 | 3684 | goto free_ctxs; |
08e98fc6 | 3685 | hctx->nr_ctx = 0; |
320ae51f | 3686 | |
5815839b | 3687 | spin_lock_init(&hctx->dispatch_wait_lock); |
eb619fdb JA |
3688 | init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake); |
3689 | INIT_LIST_HEAD(&hctx->dispatch_wait.entry); | |
3690 | ||
754a1572 | 3691 | hctx->fq = blk_alloc_flush_queue(hctx->numa_node, set->cmd_size, gfp); |
f70ced09 | 3692 | if (!hctx->fq) |
7c6c5b7c | 3693 | goto free_bitmap; |
320ae51f | 3694 | |
7c6c5b7c | 3695 | blk_mq_hctx_kobj_init(hctx); |
6a83e74d | 3696 | |
7c6c5b7c | 3697 | return hctx; |
320ae51f | 3698 | |
08e98fc6 | 3699 | free_bitmap: |
88459642 | 3700 | sbitmap_free(&hctx->ctx_map); |
08e98fc6 ML |
3701 | free_ctxs: |
3702 | kfree(hctx->ctxs); | |
7c6c5b7c ML |
3703 | free_cpumask: |
3704 | free_cpumask_var(hctx->cpumask); | |
3705 | free_hctx: | |
3706 | kfree(hctx); | |
3707 | fail_alloc_hctx: | |
3708 | return NULL; | |
08e98fc6 | 3709 | } |
320ae51f | 3710 | |
320ae51f JA |
3711 | static void blk_mq_init_cpu_queues(struct request_queue *q, |
3712 | unsigned int nr_hw_queues) | |
3713 | { | |
b3c661b1 JA |
3714 | struct blk_mq_tag_set *set = q->tag_set; |
3715 | unsigned int i, j; | |
320ae51f JA |
3716 | |
3717 | for_each_possible_cpu(i) { | |
3718 | struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i); | |
3719 | struct blk_mq_hw_ctx *hctx; | |
c16d6b5a | 3720 | int k; |
320ae51f | 3721 | |
320ae51f JA |
3722 | __ctx->cpu = i; |
3723 | spin_lock_init(&__ctx->lock); | |
c16d6b5a ML |
3724 | for (k = HCTX_TYPE_DEFAULT; k < HCTX_MAX_TYPES; k++) |
3725 | INIT_LIST_HEAD(&__ctx->rq_lists[k]); | |
3726 | ||
320ae51f JA |
3727 | __ctx->queue = q; |
3728 | ||
320ae51f JA |
3729 | /* |
3730 | * Set local node, IFF we have more than one hw queue. If | |
3731 | * not, we remain on the home node of the device | |
3732 | */ | |
b3c661b1 JA |
3733 | for (j = 0; j < set->nr_maps; j++) { |
3734 | hctx = blk_mq_map_queue_type(q, j, i); | |
3735 | if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE) | |
576e85c5 | 3736 | hctx->numa_node = cpu_to_node(i); |
b3c661b1 | 3737 | } |
320ae51f JA |
3738 | } |
3739 | } | |
3740 | ||
63064be1 JG |
3741 | struct blk_mq_tags *blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set *set, |
3742 | unsigned int hctx_idx, | |
3743 | unsigned int depth) | |
cc71a6f4 | 3744 | { |
63064be1 JG |
3745 | struct blk_mq_tags *tags; |
3746 | int ret; | |
cc71a6f4 | 3747 | |
e155b0c2 | 3748 | tags = blk_mq_alloc_rq_map(set, hctx_idx, depth, set->reserved_tags); |
63064be1 JG |
3749 | if (!tags) |
3750 | return NULL; | |
cc71a6f4 | 3751 | |
63064be1 JG |
3752 | ret = blk_mq_alloc_rqs(set, tags, hctx_idx, depth); |
3753 | if (ret) { | |
e155b0c2 | 3754 | blk_mq_free_rq_map(tags); |
63064be1 JG |
3755 | return NULL; |
3756 | } | |
cc71a6f4 | 3757 | |
63064be1 | 3758 | return tags; |
cc71a6f4 JA |
3759 | } |
3760 | ||
63064be1 JG |
3761 | static bool __blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set *set, |
3762 | int hctx_idx) | |
cc71a6f4 | 3763 | { |
079a2e3e JG |
3764 | if (blk_mq_is_shared_tags(set->flags)) { |
3765 | set->tags[hctx_idx] = set->shared_tags; | |
1c0706a7 | 3766 | |
e155b0c2 | 3767 | return true; |
bd166ef1 | 3768 | } |
e155b0c2 | 3769 | |
63064be1 JG |
3770 | set->tags[hctx_idx] = blk_mq_alloc_map_and_rqs(set, hctx_idx, |
3771 | set->queue_depth); | |
3772 | ||
3773 | return set->tags[hctx_idx]; | |
cc71a6f4 JA |
3774 | } |
3775 | ||
645db34e JG |
3776 | void blk_mq_free_map_and_rqs(struct blk_mq_tag_set *set, |
3777 | struct blk_mq_tags *tags, | |
3778 | unsigned int hctx_idx) | |
cc71a6f4 | 3779 | { |
645db34e JG |
3780 | if (tags) { |
3781 | blk_mq_free_rqs(set, tags, hctx_idx); | |
e155b0c2 | 3782 | blk_mq_free_rq_map(tags); |
bd166ef1 | 3783 | } |
cc71a6f4 JA |
3784 | } |
3785 | ||
e155b0c2 JG |
3786 | static void __blk_mq_free_map_and_rqs(struct blk_mq_tag_set *set, |
3787 | unsigned int hctx_idx) | |
3788 | { | |
079a2e3e | 3789 | if (!blk_mq_is_shared_tags(set->flags)) |
e155b0c2 JG |
3790 | blk_mq_free_map_and_rqs(set, set->tags[hctx_idx], hctx_idx); |
3791 | ||
3792 | set->tags[hctx_idx] = NULL; | |
cc71a6f4 JA |
3793 | } |
3794 | ||
4b855ad3 | 3795 | static void blk_mq_map_swqueue(struct request_queue *q) |
320ae51f | 3796 | { |
4f481208 ML |
3797 | unsigned int j, hctx_idx; |
3798 | unsigned long i; | |
320ae51f JA |
3799 | struct blk_mq_hw_ctx *hctx; |
3800 | struct blk_mq_ctx *ctx; | |
2a34c087 | 3801 | struct blk_mq_tag_set *set = q->tag_set; |
320ae51f JA |
3802 | |
3803 | queue_for_each_hw_ctx(q, hctx, i) { | |
e4043dcf | 3804 | cpumask_clear(hctx->cpumask); |
320ae51f | 3805 | hctx->nr_ctx = 0; |
d416c92c | 3806 | hctx->dispatch_from = NULL; |
320ae51f JA |
3807 | } |
3808 | ||
3809 | /* | |
4b855ad3 | 3810 | * Map software to hardware queues. |
4412efec ML |
3811 | * |
3812 | * If the cpu isn't present, the cpu is mapped to first hctx. | |
320ae51f | 3813 | */ |
20e4d813 | 3814 | for_each_possible_cpu(i) { |
4412efec | 3815 | |
897bb0c7 | 3816 | ctx = per_cpu_ptr(q->queue_ctx, i); |
b3c661b1 | 3817 | for (j = 0; j < set->nr_maps; j++) { |
bb94aea1 JW |
3818 | if (!set->map[j].nr_queues) { |
3819 | ctx->hctxs[j] = blk_mq_map_queue_type(q, | |
3820 | HCTX_TYPE_DEFAULT, i); | |
e5edd5f2 | 3821 | continue; |
bb94aea1 | 3822 | } |
fd689871 ML |
3823 | hctx_idx = set->map[j].mq_map[i]; |
3824 | /* unmapped hw queue can be remapped after CPU topo changed */ | |
3825 | if (!set->tags[hctx_idx] && | |
63064be1 | 3826 | !__blk_mq_alloc_map_and_rqs(set, hctx_idx)) { |
fd689871 ML |
3827 | /* |
3828 | * If tags initialization fail for some hctx, | |
3829 | * that hctx won't be brought online. In this | |
3830 | * case, remap the current ctx to hctx[0] which | |
3831 | * is guaranteed to always have tags allocated | |
3832 | */ | |
3833 | set->map[j].mq_map[i] = 0; | |
3834 | } | |
e5edd5f2 | 3835 | |
b3c661b1 | 3836 | hctx = blk_mq_map_queue_type(q, j, i); |
8ccdf4a3 | 3837 | ctx->hctxs[j] = hctx; |
b3c661b1 JA |
3838 | /* |
3839 | * If the CPU is already set in the mask, then we've | |
3840 | * mapped this one already. This can happen if | |
3841 | * devices share queues across queue maps. | |
3842 | */ | |
3843 | if (cpumask_test_cpu(i, hctx->cpumask)) | |
3844 | continue; | |
3845 | ||
3846 | cpumask_set_cpu(i, hctx->cpumask); | |
3847 | hctx->type = j; | |
3848 | ctx->index_hw[hctx->type] = hctx->nr_ctx; | |
3849 | hctx->ctxs[hctx->nr_ctx++] = ctx; | |
3850 | ||
3851 | /* | |
3852 | * If the nr_ctx type overflows, we have exceeded the | |
3853 | * amount of sw queues we can support. | |
3854 | */ | |
3855 | BUG_ON(!hctx->nr_ctx); | |
3856 | } | |
bb94aea1 JW |
3857 | |
3858 | for (; j < HCTX_MAX_TYPES; j++) | |
3859 | ctx->hctxs[j] = blk_mq_map_queue_type(q, | |
3860 | HCTX_TYPE_DEFAULT, i); | |
320ae51f | 3861 | } |
506e931f JA |
3862 | |
3863 | queue_for_each_hw_ctx(q, hctx, i) { | |
4412efec ML |
3864 | /* |
3865 | * If no software queues are mapped to this hardware queue, | |
3866 | * disable it and free the request entries. | |
3867 | */ | |
3868 | if (!hctx->nr_ctx) { | |
3869 | /* Never unmap queue 0. We need it as a | |
3870 | * fallback in case of a new remap fails | |
3871 | * allocation | |
3872 | */ | |
e155b0c2 JG |
3873 | if (i) |
3874 | __blk_mq_free_map_and_rqs(set, i); | |
4412efec ML |
3875 | |
3876 | hctx->tags = NULL; | |
3877 | continue; | |
3878 | } | |
484b4061 | 3879 | |
2a34c087 ML |
3880 | hctx->tags = set->tags[i]; |
3881 | WARN_ON(!hctx->tags); | |
3882 | ||
889fa31f CY |
3883 | /* |
3884 | * Set the map size to the number of mapped software queues. | |
3885 | * This is more accurate and more efficient than looping | |
3886 | * over all possibly mapped software queues. | |
3887 | */ | |
88459642 | 3888 | sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx); |
889fa31f | 3889 | |
484b4061 JA |
3890 | /* |
3891 | * Initialize batch roundrobin counts | |
3892 | */ | |
f82ddf19 | 3893 | hctx->next_cpu = blk_mq_first_mapped_cpu(hctx); |
506e931f JA |
3894 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; |
3895 | } | |
320ae51f JA |
3896 | } |
3897 | ||
8e8320c9 JA |
3898 | /* |
3899 | * Caller needs to ensure that we're either frozen/quiesced, or that | |
3900 | * the queue isn't live yet. | |
3901 | */ | |
2404e607 | 3902 | static void queue_set_hctx_shared(struct request_queue *q, bool shared) |
0d2602ca JA |
3903 | { |
3904 | struct blk_mq_hw_ctx *hctx; | |
4f481208 | 3905 | unsigned long i; |
0d2602ca | 3906 | |
2404e607 | 3907 | queue_for_each_hw_ctx(q, hctx, i) { |
454bb677 | 3908 | if (shared) { |
51db1c37 | 3909 | hctx->flags |= BLK_MQ_F_TAG_QUEUE_SHARED; |
454bb677 YK |
3910 | } else { |
3911 | blk_mq_tag_idle(hctx); | |
51db1c37 | 3912 | hctx->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED; |
454bb677 | 3913 | } |
2404e607 JM |
3914 | } |
3915 | } | |
3916 | ||
655ac300 HR |
3917 | static void blk_mq_update_tag_set_shared(struct blk_mq_tag_set *set, |
3918 | bool shared) | |
2404e607 JM |
3919 | { |
3920 | struct request_queue *q; | |
0d2602ca | 3921 | |
705cda97 BVA |
3922 | lockdep_assert_held(&set->tag_list_lock); |
3923 | ||
0d2602ca JA |
3924 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
3925 | blk_mq_freeze_queue(q); | |
2404e607 | 3926 | queue_set_hctx_shared(q, shared); |
0d2602ca JA |
3927 | blk_mq_unfreeze_queue(q); |
3928 | } | |
3929 | } | |
3930 | ||
3931 | static void blk_mq_del_queue_tag_set(struct request_queue *q) | |
3932 | { | |
3933 | struct blk_mq_tag_set *set = q->tag_set; | |
3934 | ||
0d2602ca | 3935 | mutex_lock(&set->tag_list_lock); |
08c875cb | 3936 | list_del(&q->tag_set_list); |
2404e607 JM |
3937 | if (list_is_singular(&set->tag_list)) { |
3938 | /* just transitioned to unshared */ | |
51db1c37 | 3939 | set->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED; |
2404e607 | 3940 | /* update existing queue */ |
655ac300 | 3941 | blk_mq_update_tag_set_shared(set, false); |
2404e607 | 3942 | } |
0d2602ca | 3943 | mutex_unlock(&set->tag_list_lock); |
a347c7ad | 3944 | INIT_LIST_HEAD(&q->tag_set_list); |
0d2602ca JA |
3945 | } |
3946 | ||
3947 | static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set, | |
3948 | struct request_queue *q) | |
3949 | { | |
0d2602ca | 3950 | mutex_lock(&set->tag_list_lock); |
2404e607 | 3951 | |
ff821d27 JA |
3952 | /* |
3953 | * Check to see if we're transitioning to shared (from 1 to 2 queues). | |
3954 | */ | |
3955 | if (!list_empty(&set->tag_list) && | |
51db1c37 ML |
3956 | !(set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) { |
3957 | set->flags |= BLK_MQ_F_TAG_QUEUE_SHARED; | |
2404e607 | 3958 | /* update existing queue */ |
655ac300 | 3959 | blk_mq_update_tag_set_shared(set, true); |
2404e607 | 3960 | } |
51db1c37 | 3961 | if (set->flags & BLK_MQ_F_TAG_QUEUE_SHARED) |
2404e607 | 3962 | queue_set_hctx_shared(q, true); |
08c875cb | 3963 | list_add_tail(&q->tag_set_list, &set->tag_list); |
2404e607 | 3964 | |
0d2602ca JA |
3965 | mutex_unlock(&set->tag_list_lock); |
3966 | } | |
3967 | ||
1db4909e ML |
3968 | /* All allocations will be freed in release handler of q->mq_kobj */ |
3969 | static int blk_mq_alloc_ctxs(struct request_queue *q) | |
3970 | { | |
3971 | struct blk_mq_ctxs *ctxs; | |
3972 | int cpu; | |
3973 | ||
3974 | ctxs = kzalloc(sizeof(*ctxs), GFP_KERNEL); | |
3975 | if (!ctxs) | |
3976 | return -ENOMEM; | |
3977 | ||
3978 | ctxs->queue_ctx = alloc_percpu(struct blk_mq_ctx); | |
3979 | if (!ctxs->queue_ctx) | |
3980 | goto fail; | |
3981 | ||
3982 | for_each_possible_cpu(cpu) { | |
3983 | struct blk_mq_ctx *ctx = per_cpu_ptr(ctxs->queue_ctx, cpu); | |
3984 | ctx->ctxs = ctxs; | |
3985 | } | |
3986 | ||
3987 | q->mq_kobj = &ctxs->kobj; | |
3988 | q->queue_ctx = ctxs->queue_ctx; | |
3989 | ||
3990 | return 0; | |
3991 | fail: | |
3992 | kfree(ctxs); | |
3993 | return -ENOMEM; | |
3994 | } | |
3995 | ||
e09aae7e ML |
3996 | /* |
3997 | * It is the actual release handler for mq, but we do it from | |
3998 | * request queue's release handler for avoiding use-after-free | |
3999 | * and headache because q->mq_kobj shouldn't have been introduced, | |
4000 | * but we can't group ctx/kctx kobj without it. | |
4001 | */ | |
4002 | void blk_mq_release(struct request_queue *q) | |
4003 | { | |
2f8f1336 | 4004 | struct blk_mq_hw_ctx *hctx, *next; |
4f481208 | 4005 | unsigned long i; |
e09aae7e | 4006 | |
2f8f1336 ML |
4007 | queue_for_each_hw_ctx(q, hctx, i) |
4008 | WARN_ON_ONCE(hctx && list_empty(&hctx->hctx_list)); | |
4009 | ||
4010 | /* all hctx are in .unused_hctx_list now */ | |
4011 | list_for_each_entry_safe(hctx, next, &q->unused_hctx_list, hctx_list) { | |
4012 | list_del_init(&hctx->hctx_list); | |
6c8b232e | 4013 | kobject_put(&hctx->kobj); |
c3b4afca | 4014 | } |
e09aae7e | 4015 | |
4e5cc99e | 4016 | xa_destroy(&q->hctx_table); |
e09aae7e | 4017 | |
7ea5fe31 ML |
4018 | /* |
4019 | * release .mq_kobj and sw queue's kobject now because | |
4020 | * both share lifetime with request queue. | |
4021 | */ | |
4022 | blk_mq_sysfs_deinit(q); | |
e09aae7e ML |
4023 | } |
4024 | ||
5ec780a6 | 4025 | static struct request_queue *blk_mq_init_queue_data(struct blk_mq_tag_set *set, |
2f227bb9 | 4026 | void *queuedata) |
b62c21b7 | 4027 | { |
26a9750a CH |
4028 | struct request_queue *q; |
4029 | int ret; | |
b62c21b7 | 4030 | |
80bd4a7a | 4031 | q = blk_alloc_queue(set->numa_node); |
26a9750a | 4032 | if (!q) |
b62c21b7 | 4033 | return ERR_PTR(-ENOMEM); |
26a9750a CH |
4034 | q->queuedata = queuedata; |
4035 | ret = blk_mq_init_allocated_queue(set, q); | |
4036 | if (ret) { | |
6f8191fd | 4037 | blk_put_queue(q); |
26a9750a CH |
4038 | return ERR_PTR(ret); |
4039 | } | |
b62c21b7 MS |
4040 | return q; |
4041 | } | |
2f227bb9 CH |
4042 | |
4043 | struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set) | |
4044 | { | |
4045 | return blk_mq_init_queue_data(set, NULL); | |
4046 | } | |
b62c21b7 MS |
4047 | EXPORT_SYMBOL(blk_mq_init_queue); |
4048 | ||
6f8191fd CH |
4049 | /** |
4050 | * blk_mq_destroy_queue - shutdown a request queue | |
4051 | * @q: request queue to shutdown | |
4052 | * | |
81ea42b9 BVA |
4053 | * This shuts down a request queue allocated by blk_mq_init_queue(). All future |
4054 | * requests will be failed with -ENODEV. The caller is responsible for dropping | |
4055 | * the reference from blk_mq_init_queue() by calling blk_put_queue(). | |
6f8191fd CH |
4056 | * |
4057 | * Context: can sleep | |
4058 | */ | |
4059 | void blk_mq_destroy_queue(struct request_queue *q) | |
4060 | { | |
4061 | WARN_ON_ONCE(!queue_is_mq(q)); | |
4062 | WARN_ON_ONCE(blk_queue_registered(q)); | |
4063 | ||
4064 | might_sleep(); | |
4065 | ||
4066 | blk_queue_flag_set(QUEUE_FLAG_DYING, q); | |
4067 | blk_queue_start_drain(q); | |
56c1ee92 | 4068 | blk_mq_freeze_queue_wait(q); |
6f8191fd CH |
4069 | |
4070 | blk_sync_queue(q); | |
4071 | blk_mq_cancel_work_sync(q); | |
4072 | blk_mq_exit_queue(q); | |
6f8191fd CH |
4073 | } |
4074 | EXPORT_SYMBOL(blk_mq_destroy_queue); | |
4075 | ||
4dcc4874 CH |
4076 | struct gendisk *__blk_mq_alloc_disk(struct blk_mq_tag_set *set, void *queuedata, |
4077 | struct lock_class_key *lkclass) | |
9316a9ed JA |
4078 | { |
4079 | struct request_queue *q; | |
b461dfc4 | 4080 | struct gendisk *disk; |
9316a9ed | 4081 | |
b461dfc4 CH |
4082 | q = blk_mq_init_queue_data(set, queuedata); |
4083 | if (IS_ERR(q)) | |
4084 | return ERR_CAST(q); | |
9316a9ed | 4085 | |
4a1fa41d | 4086 | disk = __alloc_disk_node(q, set->numa_node, lkclass); |
b461dfc4 | 4087 | if (!disk) { |
0a3e5cc7 | 4088 | blk_mq_destroy_queue(q); |
2b3f056f | 4089 | blk_put_queue(q); |
b461dfc4 | 4090 | return ERR_PTR(-ENOMEM); |
9316a9ed | 4091 | } |
6f8191fd | 4092 | set_bit(GD_OWNS_QUEUE, &disk->state); |
b461dfc4 | 4093 | return disk; |
9316a9ed | 4094 | } |
b461dfc4 | 4095 | EXPORT_SYMBOL(__blk_mq_alloc_disk); |
9316a9ed | 4096 | |
6f8191fd CH |
4097 | struct gendisk *blk_mq_alloc_disk_for_queue(struct request_queue *q, |
4098 | struct lock_class_key *lkclass) | |
4099 | { | |
22c17e27 CH |
4100 | struct gendisk *disk; |
4101 | ||
6f8191fd CH |
4102 | if (!blk_get_queue(q)) |
4103 | return NULL; | |
22c17e27 CH |
4104 | disk = __alloc_disk_node(q, NUMA_NO_NODE, lkclass); |
4105 | if (!disk) | |
4106 | blk_put_queue(q); | |
4107 | return disk; | |
6f8191fd CH |
4108 | } |
4109 | EXPORT_SYMBOL(blk_mq_alloc_disk_for_queue); | |
4110 | ||
34d11ffa JW |
4111 | static struct blk_mq_hw_ctx *blk_mq_alloc_and_init_hctx( |
4112 | struct blk_mq_tag_set *set, struct request_queue *q, | |
4113 | int hctx_idx, int node) | |
4114 | { | |
2f8f1336 | 4115 | struct blk_mq_hw_ctx *hctx = NULL, *tmp; |
34d11ffa | 4116 | |
2f8f1336 ML |
4117 | /* reuse dead hctx first */ |
4118 | spin_lock(&q->unused_hctx_lock); | |
4119 | list_for_each_entry(tmp, &q->unused_hctx_list, hctx_list) { | |
4120 | if (tmp->numa_node == node) { | |
4121 | hctx = tmp; | |
4122 | break; | |
4123 | } | |
4124 | } | |
4125 | if (hctx) | |
4126 | list_del_init(&hctx->hctx_list); | |
4127 | spin_unlock(&q->unused_hctx_lock); | |
4128 | ||
4129 | if (!hctx) | |
4130 | hctx = blk_mq_alloc_hctx(q, set, node); | |
34d11ffa | 4131 | if (!hctx) |
7c6c5b7c | 4132 | goto fail; |
34d11ffa | 4133 | |
7c6c5b7c ML |
4134 | if (blk_mq_init_hctx(q, set, hctx, hctx_idx)) |
4135 | goto free_hctx; | |
34d11ffa JW |
4136 | |
4137 | return hctx; | |
7c6c5b7c ML |
4138 | |
4139 | free_hctx: | |
4140 | kobject_put(&hctx->kobj); | |
4141 | fail: | |
4142 | return NULL; | |
34d11ffa JW |
4143 | } |
4144 | ||
868f2f0b KB |
4145 | static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set, |
4146 | struct request_queue *q) | |
320ae51f | 4147 | { |
4e5cc99e ML |
4148 | struct blk_mq_hw_ctx *hctx; |
4149 | unsigned long i, j; | |
ac0d6b92 | 4150 | |
fb350e0a ML |
4151 | /* protect against switching io scheduler */ |
4152 | mutex_lock(&q->sysfs_lock); | |
24d2f903 | 4153 | for (i = 0; i < set->nr_hw_queues; i++) { |
306f13ee | 4154 | int old_node; |
4d805131 | 4155 | int node = blk_mq_get_hctx_node(set, i); |
4e5cc99e | 4156 | struct blk_mq_hw_ctx *old_hctx = xa_load(&q->hctx_table, i); |
868f2f0b | 4157 | |
306f13ee ML |
4158 | if (old_hctx) { |
4159 | old_node = old_hctx->numa_node; | |
4160 | blk_mq_exit_hctx(q, set, old_hctx, i); | |
4161 | } | |
868f2f0b | 4162 | |
4e5cc99e | 4163 | if (!blk_mq_alloc_and_init_hctx(set, q, i, node)) { |
306f13ee | 4164 | if (!old_hctx) |
34d11ffa | 4165 | break; |
306f13ee ML |
4166 | pr_warn("Allocate new hctx on node %d fails, fallback to previous one on node %d\n", |
4167 | node, old_node); | |
4e5cc99e ML |
4168 | hctx = blk_mq_alloc_and_init_hctx(set, q, i, old_node); |
4169 | WARN_ON_ONCE(!hctx); | |
868f2f0b | 4170 | } |
320ae51f | 4171 | } |
e01ad46d JW |
4172 | /* |
4173 | * Increasing nr_hw_queues fails. Free the newly allocated | |
4174 | * hctxs and keep the previous q->nr_hw_queues. | |
4175 | */ | |
4176 | if (i != set->nr_hw_queues) { | |
4177 | j = q->nr_hw_queues; | |
e01ad46d JW |
4178 | } else { |
4179 | j = i; | |
e01ad46d JW |
4180 | q->nr_hw_queues = set->nr_hw_queues; |
4181 | } | |
34d11ffa | 4182 | |
4e5cc99e ML |
4183 | xa_for_each_start(&q->hctx_table, j, hctx, j) |
4184 | blk_mq_exit_hctx(q, set, hctx, j); | |
fb350e0a | 4185 | mutex_unlock(&q->sysfs_lock); |
868f2f0b KB |
4186 | } |
4187 | ||
42ee3061 ML |
4188 | static void blk_mq_update_poll_flag(struct request_queue *q) |
4189 | { | |
4190 | struct blk_mq_tag_set *set = q->tag_set; | |
4191 | ||
4192 | if (set->nr_maps > HCTX_TYPE_POLL && | |
4193 | set->map[HCTX_TYPE_POLL].nr_queues) | |
4194 | blk_queue_flag_set(QUEUE_FLAG_POLL, q); | |
4195 | else | |
4196 | blk_queue_flag_clear(QUEUE_FLAG_POLL, q); | |
4197 | } | |
4198 | ||
26a9750a CH |
4199 | int blk_mq_init_allocated_queue(struct blk_mq_tag_set *set, |
4200 | struct request_queue *q) | |
868f2f0b | 4201 | { |
66841672 ML |
4202 | /* mark the queue as mq asap */ |
4203 | q->mq_ops = set->ops; | |
4204 | ||
1db4909e | 4205 | if (blk_mq_alloc_ctxs(q)) |
54bdd67d | 4206 | goto err_exit; |
868f2f0b | 4207 | |
737f98cf ML |
4208 | /* init q->mq_kobj and sw queues' kobjects */ |
4209 | blk_mq_sysfs_init(q); | |
4210 | ||
2f8f1336 ML |
4211 | INIT_LIST_HEAD(&q->unused_hctx_list); |
4212 | spin_lock_init(&q->unused_hctx_lock); | |
4213 | ||
4e5cc99e ML |
4214 | xa_init(&q->hctx_table); |
4215 | ||
868f2f0b KB |
4216 | blk_mq_realloc_hw_ctxs(set, q); |
4217 | if (!q->nr_hw_queues) | |
4218 | goto err_hctxs; | |
320ae51f | 4219 | |
287922eb | 4220 | INIT_WORK(&q->timeout_work, blk_mq_timeout_work); |
e56f698b | 4221 | blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ); |
320ae51f | 4222 | |
a8908939 | 4223 | q->tag_set = set; |
320ae51f | 4224 | |
94eddfbe | 4225 | q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT; |
42ee3061 | 4226 | blk_mq_update_poll_flag(q); |
320ae51f | 4227 | |
2849450a | 4228 | INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work); |
9a67aa52 | 4229 | INIT_LIST_HEAD(&q->flush_list); |
6fca6a61 CH |
4230 | INIT_LIST_HEAD(&q->requeue_list); |
4231 | spin_lock_init(&q->requeue_lock); | |
4232 | ||
eba71768 JA |
4233 | q->nr_requests = set->queue_depth; |
4234 | ||
24d2f903 | 4235 | blk_mq_init_cpu_queues(q, set->nr_hw_queues); |
0d2602ca | 4236 | blk_mq_add_queue_tag_set(set, q); |
4b855ad3 | 4237 | blk_mq_map_swqueue(q); |
26a9750a | 4238 | return 0; |
18741986 | 4239 | |
320ae51f | 4240 | err_hctxs: |
943f45b9 | 4241 | blk_mq_release(q); |
c7de5726 ML |
4242 | err_exit: |
4243 | q->mq_ops = NULL; | |
26a9750a | 4244 | return -ENOMEM; |
320ae51f | 4245 | } |
b62c21b7 | 4246 | EXPORT_SYMBOL(blk_mq_init_allocated_queue); |
320ae51f | 4247 | |
c7e2d94b ML |
4248 | /* tags can _not_ be used after returning from blk_mq_exit_queue */ |
4249 | void blk_mq_exit_queue(struct request_queue *q) | |
320ae51f | 4250 | { |
630ef623 | 4251 | struct blk_mq_tag_set *set = q->tag_set; |
320ae51f | 4252 | |
630ef623 | 4253 | /* Checks hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED. */ |
624dbe47 | 4254 | blk_mq_exit_hw_queues(q, set, set->nr_hw_queues); |
630ef623 BVA |
4255 | /* May clear BLK_MQ_F_TAG_QUEUE_SHARED in hctx->flags. */ |
4256 | blk_mq_del_queue_tag_set(q); | |
320ae51f | 4257 | } |
320ae51f | 4258 | |
a5164405 JA |
4259 | static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set) |
4260 | { | |
4261 | int i; | |
4262 | ||
079a2e3e JG |
4263 | if (blk_mq_is_shared_tags(set->flags)) { |
4264 | set->shared_tags = blk_mq_alloc_map_and_rqs(set, | |
e155b0c2 JG |
4265 | BLK_MQ_NO_HCTX_IDX, |
4266 | set->queue_depth); | |
079a2e3e | 4267 | if (!set->shared_tags) |
e155b0c2 JG |
4268 | return -ENOMEM; |
4269 | } | |
4270 | ||
8229cca8 | 4271 | for (i = 0; i < set->nr_hw_queues; i++) { |
63064be1 | 4272 | if (!__blk_mq_alloc_map_and_rqs(set, i)) |
a5164405 | 4273 | goto out_unwind; |
8229cca8 XT |
4274 | cond_resched(); |
4275 | } | |
a5164405 JA |
4276 | |
4277 | return 0; | |
4278 | ||
4279 | out_unwind: | |
4280 | while (--i >= 0) | |
e155b0c2 JG |
4281 | __blk_mq_free_map_and_rqs(set, i); |
4282 | ||
079a2e3e JG |
4283 | if (blk_mq_is_shared_tags(set->flags)) { |
4284 | blk_mq_free_map_and_rqs(set, set->shared_tags, | |
e155b0c2 | 4285 | BLK_MQ_NO_HCTX_IDX); |
645db34e | 4286 | } |
a5164405 | 4287 | |
a5164405 JA |
4288 | return -ENOMEM; |
4289 | } | |
4290 | ||
4291 | /* | |
4292 | * Allocate the request maps associated with this tag_set. Note that this | |
4293 | * may reduce the depth asked for, if memory is tight. set->queue_depth | |
4294 | * will be updated to reflect the allocated depth. | |
4295 | */ | |
63064be1 | 4296 | static int blk_mq_alloc_set_map_and_rqs(struct blk_mq_tag_set *set) |
a5164405 JA |
4297 | { |
4298 | unsigned int depth; | |
4299 | int err; | |
4300 | ||
4301 | depth = set->queue_depth; | |
4302 | do { | |
4303 | err = __blk_mq_alloc_rq_maps(set); | |
4304 | if (!err) | |
4305 | break; | |
4306 | ||
4307 | set->queue_depth >>= 1; | |
4308 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) { | |
4309 | err = -ENOMEM; | |
4310 | break; | |
4311 | } | |
4312 | } while (set->queue_depth); | |
4313 | ||
4314 | if (!set->queue_depth || err) { | |
4315 | pr_err("blk-mq: failed to allocate request map\n"); | |
4316 | return -ENOMEM; | |
4317 | } | |
4318 | ||
4319 | if (depth != set->queue_depth) | |
4320 | pr_info("blk-mq: reduced tag depth (%u -> %u)\n", | |
4321 | depth, set->queue_depth); | |
4322 | ||
4323 | return 0; | |
4324 | } | |
4325 | ||
a4e1d0b7 | 4326 | static void blk_mq_update_queue_map(struct blk_mq_tag_set *set) |
ebe8bddb | 4327 | { |
6e66b493 BVA |
4328 | /* |
4329 | * blk_mq_map_queues() and multiple .map_queues() implementations | |
4330 | * expect that set->map[HCTX_TYPE_DEFAULT].nr_queues is set to the | |
4331 | * number of hardware queues. | |
4332 | */ | |
4333 | if (set->nr_maps == 1) | |
4334 | set->map[HCTX_TYPE_DEFAULT].nr_queues = set->nr_hw_queues; | |
4335 | ||
59388702 | 4336 | if (set->ops->map_queues && !is_kdump_kernel()) { |
b3c661b1 JA |
4337 | int i; |
4338 | ||
7d4901a9 ML |
4339 | /* |
4340 | * transport .map_queues is usually done in the following | |
4341 | * way: | |
4342 | * | |
4343 | * for (queue = 0; queue < set->nr_hw_queues; queue++) { | |
4344 | * mask = get_cpu_mask(queue) | |
4345 | * for_each_cpu(cpu, mask) | |
b3c661b1 | 4346 | * set->map[x].mq_map[cpu] = queue; |
7d4901a9 ML |
4347 | * } |
4348 | * | |
4349 | * When we need to remap, the table has to be cleared for | |
4350 | * killing stale mapping since one CPU may not be mapped | |
4351 | * to any hw queue. | |
4352 | */ | |
b3c661b1 JA |
4353 | for (i = 0; i < set->nr_maps; i++) |
4354 | blk_mq_clear_mq_map(&set->map[i]); | |
7d4901a9 | 4355 | |
a4e1d0b7 | 4356 | set->ops->map_queues(set); |
b3c661b1 JA |
4357 | } else { |
4358 | BUG_ON(set->nr_maps > 1); | |
a4e1d0b7 | 4359 | blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]); |
b3c661b1 | 4360 | } |
ebe8bddb OS |
4361 | } |
4362 | ||
f7e76dbc | 4363 | static int blk_mq_realloc_tag_set_tags(struct blk_mq_tag_set *set, |
ee9d5521 | 4364 | int new_nr_hw_queues) |
f7e76dbc BVA |
4365 | { |
4366 | struct blk_mq_tags **new_tags; | |
4367 | ||
ee9d5521 | 4368 | if (set->nr_hw_queues >= new_nr_hw_queues) |
d4b2e0d4 | 4369 | goto done; |
f7e76dbc BVA |
4370 | |
4371 | new_tags = kcalloc_node(new_nr_hw_queues, sizeof(struct blk_mq_tags *), | |
4372 | GFP_KERNEL, set->numa_node); | |
4373 | if (!new_tags) | |
4374 | return -ENOMEM; | |
4375 | ||
4376 | if (set->tags) | |
ee9d5521 | 4377 | memcpy(new_tags, set->tags, set->nr_hw_queues * |
f7e76dbc BVA |
4378 | sizeof(*set->tags)); |
4379 | kfree(set->tags); | |
4380 | set->tags = new_tags; | |
d4b2e0d4 | 4381 | done: |
f7e76dbc | 4382 | set->nr_hw_queues = new_nr_hw_queues; |
f7e76dbc BVA |
4383 | return 0; |
4384 | } | |
4385 | ||
a4391c64 JA |
4386 | /* |
4387 | * Alloc a tag set to be associated with one or more request queues. | |
4388 | * May fail with EINVAL for various error conditions. May adjust the | |
c018c84f | 4389 | * requested depth down, if it's too large. In that case, the set |
a4391c64 JA |
4390 | * value will be stored in set->queue_depth. |
4391 | */ | |
24d2f903 CH |
4392 | int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set) |
4393 | { | |
b3c661b1 | 4394 | int i, ret; |
da695ba2 | 4395 | |
205fb5f5 BVA |
4396 | BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS); |
4397 | ||
24d2f903 CH |
4398 | if (!set->nr_hw_queues) |
4399 | return -EINVAL; | |
a4391c64 | 4400 | if (!set->queue_depth) |
24d2f903 CH |
4401 | return -EINVAL; |
4402 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) | |
4403 | return -EINVAL; | |
4404 | ||
7d7e0f90 | 4405 | if (!set->ops->queue_rq) |
24d2f903 CH |
4406 | return -EINVAL; |
4407 | ||
de148297 ML |
4408 | if (!set->ops->get_budget ^ !set->ops->put_budget) |
4409 | return -EINVAL; | |
4410 | ||
a4391c64 JA |
4411 | if (set->queue_depth > BLK_MQ_MAX_DEPTH) { |
4412 | pr_info("blk-mq: reduced tag depth to %u\n", | |
4413 | BLK_MQ_MAX_DEPTH); | |
4414 | set->queue_depth = BLK_MQ_MAX_DEPTH; | |
4415 | } | |
24d2f903 | 4416 | |
b3c661b1 JA |
4417 | if (!set->nr_maps) |
4418 | set->nr_maps = 1; | |
4419 | else if (set->nr_maps > HCTX_MAX_TYPES) | |
4420 | return -EINVAL; | |
4421 | ||
6637fadf SL |
4422 | /* |
4423 | * If a crashdump is active, then we are potentially in a very | |
4424 | * memory constrained environment. Limit us to 1 queue and | |
4425 | * 64 tags to prevent using too much memory. | |
4426 | */ | |
4427 | if (is_kdump_kernel()) { | |
4428 | set->nr_hw_queues = 1; | |
59388702 | 4429 | set->nr_maps = 1; |
6637fadf SL |
4430 | set->queue_depth = min(64U, set->queue_depth); |
4431 | } | |
868f2f0b | 4432 | /* |
392546ae JA |
4433 | * There is no use for more h/w queues than cpus if we just have |
4434 | * a single map | |
868f2f0b | 4435 | */ |
392546ae | 4436 | if (set->nr_maps == 1 && set->nr_hw_queues > nr_cpu_ids) |
868f2f0b | 4437 | set->nr_hw_queues = nr_cpu_ids; |
6637fadf | 4438 | |
80bd4a7a CH |
4439 | if (set->flags & BLK_MQ_F_BLOCKING) { |
4440 | set->srcu = kmalloc(sizeof(*set->srcu), GFP_KERNEL); | |
4441 | if (!set->srcu) | |
4442 | return -ENOMEM; | |
4443 | ret = init_srcu_struct(set->srcu); | |
4444 | if (ret) | |
4445 | goto out_free_srcu; | |
4446 | } | |
24d2f903 | 4447 | |
da695ba2 | 4448 | ret = -ENOMEM; |
5ee20298 CH |
4449 | set->tags = kcalloc_node(set->nr_hw_queues, |
4450 | sizeof(struct blk_mq_tags *), GFP_KERNEL, | |
4451 | set->numa_node); | |
4452 | if (!set->tags) | |
80bd4a7a | 4453 | goto out_cleanup_srcu; |
24d2f903 | 4454 | |
b3c661b1 JA |
4455 | for (i = 0; i < set->nr_maps; i++) { |
4456 | set->map[i].mq_map = kcalloc_node(nr_cpu_ids, | |
07b35eb5 | 4457 | sizeof(set->map[i].mq_map[0]), |
b3c661b1 JA |
4458 | GFP_KERNEL, set->numa_node); |
4459 | if (!set->map[i].mq_map) | |
4460 | goto out_free_mq_map; | |
59388702 | 4461 | set->map[i].nr_queues = is_kdump_kernel() ? 1 : set->nr_hw_queues; |
b3c661b1 | 4462 | } |
bdd17e75 | 4463 | |
a4e1d0b7 | 4464 | blk_mq_update_queue_map(set); |
da695ba2 | 4465 | |
63064be1 | 4466 | ret = blk_mq_alloc_set_map_and_rqs(set); |
da695ba2 | 4467 | if (ret) |
bdd17e75 | 4468 | goto out_free_mq_map; |
24d2f903 | 4469 | |
0d2602ca JA |
4470 | mutex_init(&set->tag_list_lock); |
4471 | INIT_LIST_HEAD(&set->tag_list); | |
4472 | ||
24d2f903 | 4473 | return 0; |
bdd17e75 CH |
4474 | |
4475 | out_free_mq_map: | |
b3c661b1 JA |
4476 | for (i = 0; i < set->nr_maps; i++) { |
4477 | kfree(set->map[i].mq_map); | |
4478 | set->map[i].mq_map = NULL; | |
4479 | } | |
5676e7b6 RE |
4480 | kfree(set->tags); |
4481 | set->tags = NULL; | |
80bd4a7a CH |
4482 | out_cleanup_srcu: |
4483 | if (set->flags & BLK_MQ_F_BLOCKING) | |
4484 | cleanup_srcu_struct(set->srcu); | |
4485 | out_free_srcu: | |
4486 | if (set->flags & BLK_MQ_F_BLOCKING) | |
4487 | kfree(set->srcu); | |
da695ba2 | 4488 | return ret; |
24d2f903 CH |
4489 | } |
4490 | EXPORT_SYMBOL(blk_mq_alloc_tag_set); | |
4491 | ||
cdb14e0f CH |
4492 | /* allocate and initialize a tagset for a simple single-queue device */ |
4493 | int blk_mq_alloc_sq_tag_set(struct blk_mq_tag_set *set, | |
4494 | const struct blk_mq_ops *ops, unsigned int queue_depth, | |
4495 | unsigned int set_flags) | |
4496 | { | |
4497 | memset(set, 0, sizeof(*set)); | |
4498 | set->ops = ops; | |
4499 | set->nr_hw_queues = 1; | |
4500 | set->nr_maps = 1; | |
4501 | set->queue_depth = queue_depth; | |
4502 | set->numa_node = NUMA_NO_NODE; | |
4503 | set->flags = set_flags; | |
4504 | return blk_mq_alloc_tag_set(set); | |
4505 | } | |
4506 | EXPORT_SYMBOL_GPL(blk_mq_alloc_sq_tag_set); | |
4507 | ||
24d2f903 CH |
4508 | void blk_mq_free_tag_set(struct blk_mq_tag_set *set) |
4509 | { | |
b3c661b1 | 4510 | int i, j; |
24d2f903 | 4511 | |
f7e76dbc | 4512 | for (i = 0; i < set->nr_hw_queues; i++) |
e155b0c2 | 4513 | __blk_mq_free_map_and_rqs(set, i); |
484b4061 | 4514 | |
079a2e3e JG |
4515 | if (blk_mq_is_shared_tags(set->flags)) { |
4516 | blk_mq_free_map_and_rqs(set, set->shared_tags, | |
e155b0c2 JG |
4517 | BLK_MQ_NO_HCTX_IDX); |
4518 | } | |
32bc15af | 4519 | |
b3c661b1 JA |
4520 | for (j = 0; j < set->nr_maps; j++) { |
4521 | kfree(set->map[j].mq_map); | |
4522 | set->map[j].mq_map = NULL; | |
4523 | } | |
bdd17e75 | 4524 | |
981bd189 | 4525 | kfree(set->tags); |
5676e7b6 | 4526 | set->tags = NULL; |
80bd4a7a CH |
4527 | if (set->flags & BLK_MQ_F_BLOCKING) { |
4528 | cleanup_srcu_struct(set->srcu); | |
4529 | kfree(set->srcu); | |
4530 | } | |
24d2f903 CH |
4531 | } |
4532 | EXPORT_SYMBOL(blk_mq_free_tag_set); | |
4533 | ||
e3a2b3f9 JA |
4534 | int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr) |
4535 | { | |
4536 | struct blk_mq_tag_set *set = q->tag_set; | |
4537 | struct blk_mq_hw_ctx *hctx; | |
4f481208 ML |
4538 | int ret; |
4539 | unsigned long i; | |
e3a2b3f9 | 4540 | |
bd166ef1 | 4541 | if (!set) |
e3a2b3f9 JA |
4542 | return -EINVAL; |
4543 | ||
e5fa8140 AZ |
4544 | if (q->nr_requests == nr) |
4545 | return 0; | |
4546 | ||
70f36b60 | 4547 | blk_mq_freeze_queue(q); |
24f5a90f | 4548 | blk_mq_quiesce_queue(q); |
70f36b60 | 4549 | |
e3a2b3f9 JA |
4550 | ret = 0; |
4551 | queue_for_each_hw_ctx(q, hctx, i) { | |
e9137d4b KB |
4552 | if (!hctx->tags) |
4553 | continue; | |
bd166ef1 JA |
4554 | /* |
4555 | * If we're using an MQ scheduler, just update the scheduler | |
4556 | * queue depth. This is similar to what the old code would do. | |
4557 | */ | |
f6adcef5 | 4558 | if (hctx->sched_tags) { |
70f36b60 | 4559 | ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags, |
f6adcef5 | 4560 | nr, true); |
f6adcef5 JG |
4561 | } else { |
4562 | ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr, | |
4563 | false); | |
70f36b60 | 4564 | } |
e3a2b3f9 JA |
4565 | if (ret) |
4566 | break; | |
77f1e0a5 JA |
4567 | if (q->elevator && q->elevator->type->ops.depth_updated) |
4568 | q->elevator->type->ops.depth_updated(hctx); | |
e3a2b3f9 | 4569 | } |
d97e594c | 4570 | if (!ret) { |
e3a2b3f9 | 4571 | q->nr_requests = nr; |
079a2e3e | 4572 | if (blk_mq_is_shared_tags(set->flags)) { |
8fa04464 | 4573 | if (q->elevator) |
079a2e3e | 4574 | blk_mq_tag_update_sched_shared_tags(q); |
8fa04464 | 4575 | else |
079a2e3e | 4576 | blk_mq_tag_resize_shared_tags(set, nr); |
8fa04464 | 4577 | } |
d97e594c | 4578 | } |
e3a2b3f9 | 4579 | |
24f5a90f | 4580 | blk_mq_unquiesce_queue(q); |
70f36b60 | 4581 | blk_mq_unfreeze_queue(q); |
70f36b60 | 4582 | |
e3a2b3f9 JA |
4583 | return ret; |
4584 | } | |
4585 | ||
d48ece20 JW |
4586 | /* |
4587 | * request_queue and elevator_type pair. | |
4588 | * It is just used by __blk_mq_update_nr_hw_queues to cache | |
4589 | * the elevator_type associated with a request_queue. | |
4590 | */ | |
4591 | struct blk_mq_qe_pair { | |
4592 | struct list_head node; | |
4593 | struct request_queue *q; | |
4594 | struct elevator_type *type; | |
4595 | }; | |
4596 | ||
4597 | /* | |
4598 | * Cache the elevator_type in qe pair list and switch the | |
4599 | * io scheduler to 'none' | |
4600 | */ | |
4601 | static bool blk_mq_elv_switch_none(struct list_head *head, | |
4602 | struct request_queue *q) | |
4603 | { | |
4604 | struct blk_mq_qe_pair *qe; | |
4605 | ||
4606 | if (!q->elevator) | |
4607 | return true; | |
4608 | ||
4609 | qe = kmalloc(sizeof(*qe), GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY); | |
4610 | if (!qe) | |
4611 | return false; | |
4612 | ||
5fd7a84a ML |
4613 | /* q->elevator needs protection from ->sysfs_lock */ |
4614 | mutex_lock(&q->sysfs_lock); | |
4615 | ||
d48ece20 JW |
4616 | INIT_LIST_HEAD(&qe->node); |
4617 | qe->q = q; | |
4618 | qe->type = q->elevator->type; | |
dd6f7f17 CH |
4619 | /* keep a reference to the elevator module as we'll switch back */ |
4620 | __elevator_get(qe->type); | |
d48ece20 | 4621 | list_add(&qe->node, head); |
64b36075 | 4622 | elevator_disable(q); |
d48ece20 JW |
4623 | mutex_unlock(&q->sysfs_lock); |
4624 | ||
4625 | return true; | |
4626 | } | |
4627 | ||
4a3b666e JK |
4628 | static struct blk_mq_qe_pair *blk_lookup_qe_pair(struct list_head *head, |
4629 | struct request_queue *q) | |
d48ece20 JW |
4630 | { |
4631 | struct blk_mq_qe_pair *qe; | |
d48ece20 JW |
4632 | |
4633 | list_for_each_entry(qe, head, node) | |
4a3b666e JK |
4634 | if (qe->q == q) |
4635 | return qe; | |
d48ece20 | 4636 | |
4a3b666e JK |
4637 | return NULL; |
4638 | } | |
d48ece20 | 4639 | |
4a3b666e JK |
4640 | static void blk_mq_elv_switch_back(struct list_head *head, |
4641 | struct request_queue *q) | |
4642 | { | |
4643 | struct blk_mq_qe_pair *qe; | |
4644 | struct elevator_type *t; | |
4645 | ||
4646 | qe = blk_lookup_qe_pair(head, q); | |
4647 | if (!qe) | |
4648 | return; | |
4649 | t = qe->type; | |
d48ece20 JW |
4650 | list_del(&qe->node); |
4651 | kfree(qe); | |
4652 | ||
4653 | mutex_lock(&q->sysfs_lock); | |
8237c01f | 4654 | elevator_switch(q, t); |
8ed40ee3 JC |
4655 | /* drop the reference acquired in blk_mq_elv_switch_none */ |
4656 | elevator_put(t); | |
d48ece20 JW |
4657 | mutex_unlock(&q->sysfs_lock); |
4658 | } | |
4659 | ||
e4dc2b32 KB |
4660 | static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, |
4661 | int nr_hw_queues) | |
868f2f0b KB |
4662 | { |
4663 | struct request_queue *q; | |
d48ece20 | 4664 | LIST_HEAD(head); |
e01ad46d | 4665 | int prev_nr_hw_queues; |
868f2f0b | 4666 | |
705cda97 BVA |
4667 | lockdep_assert_held(&set->tag_list_lock); |
4668 | ||
392546ae | 4669 | if (set->nr_maps == 1 && nr_hw_queues > nr_cpu_ids) |
868f2f0b | 4670 | nr_hw_queues = nr_cpu_ids; |
fe35ec58 WZ |
4671 | if (nr_hw_queues < 1) |
4672 | return; | |
4673 | if (set->nr_maps == 1 && nr_hw_queues == set->nr_hw_queues) | |
868f2f0b KB |
4674 | return; |
4675 | ||
4676 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
4677 | blk_mq_freeze_queue(q); | |
d48ece20 JW |
4678 | /* |
4679 | * Switch IO scheduler to 'none', cleaning up the data associated | |
4680 | * with the previous scheduler. We will switch back once we are done | |
4681 | * updating the new sw to hw queue mappings. | |
4682 | */ | |
4683 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
4684 | if (!blk_mq_elv_switch_none(&head, q)) | |
4685 | goto switch_back; | |
868f2f0b | 4686 | |
477e19de JW |
4687 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
4688 | blk_mq_debugfs_unregister_hctxs(q); | |
eaa870f9 | 4689 | blk_mq_sysfs_unregister_hctxs(q); |
477e19de JW |
4690 | } |
4691 | ||
a2584e43 | 4692 | prev_nr_hw_queues = set->nr_hw_queues; |
ee9d5521 | 4693 | if (blk_mq_realloc_tag_set_tags(set, nr_hw_queues) < 0) |
f7e76dbc BVA |
4694 | goto reregister; |
4695 | ||
e01ad46d | 4696 | fallback: |
aa880ad6 | 4697 | blk_mq_update_queue_map(set); |
868f2f0b KB |
4698 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
4699 | blk_mq_realloc_hw_ctxs(set, q); | |
42ee3061 | 4700 | blk_mq_update_poll_flag(q); |
e01ad46d | 4701 | if (q->nr_hw_queues != set->nr_hw_queues) { |
a846a8e6 YB |
4702 | int i = prev_nr_hw_queues; |
4703 | ||
e01ad46d JW |
4704 | pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n", |
4705 | nr_hw_queues, prev_nr_hw_queues); | |
a846a8e6 YB |
4706 | for (; i < set->nr_hw_queues; i++) |
4707 | __blk_mq_free_map_and_rqs(set, i); | |
4708 | ||
e01ad46d | 4709 | set->nr_hw_queues = prev_nr_hw_queues; |
7d76f856 | 4710 | blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]); |
e01ad46d JW |
4711 | goto fallback; |
4712 | } | |
477e19de JW |
4713 | blk_mq_map_swqueue(q); |
4714 | } | |
4715 | ||
f7e76dbc | 4716 | reregister: |
477e19de | 4717 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
eaa870f9 | 4718 | blk_mq_sysfs_register_hctxs(q); |
477e19de | 4719 | blk_mq_debugfs_register_hctxs(q); |
868f2f0b KB |
4720 | } |
4721 | ||
d48ece20 JW |
4722 | switch_back: |
4723 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
4724 | blk_mq_elv_switch_back(&head, q); | |
4725 | ||
868f2f0b KB |
4726 | list_for_each_entry(q, &set->tag_list, tag_set_list) |
4727 | blk_mq_unfreeze_queue(q); | |
4728 | } | |
e4dc2b32 KB |
4729 | |
4730 | void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues) | |
4731 | { | |
4732 | mutex_lock(&set->tag_list_lock); | |
4733 | __blk_mq_update_nr_hw_queues(set, nr_hw_queues); | |
4734 | mutex_unlock(&set->tag_list_lock); | |
4735 | } | |
868f2f0b KB |
4736 | EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues); |
4737 | ||
54bdd67d KB |
4738 | int blk_mq_poll(struct request_queue *q, blk_qc_t cookie, struct io_comp_batch *iob, |
4739 | unsigned int flags) | |
bbd7bb70 | 4740 | { |
c6699d6f CH |
4741 | struct blk_mq_hw_ctx *hctx = blk_qc_to_hctx(q, cookie); |
4742 | long state = get_current_state(); | |
4743 | int ret; | |
bbd7bb70 | 4744 | |
aa61bec3 | 4745 | do { |
5a72e899 | 4746 | ret = q->mq_ops->poll(hctx, iob); |
bbd7bb70 | 4747 | if (ret > 0) { |
849a3700 | 4748 | __set_current_state(TASK_RUNNING); |
85f4d4b6 | 4749 | return ret; |
bbd7bb70 JA |
4750 | } |
4751 | ||
4752 | if (signal_pending_state(state, current)) | |
849a3700 | 4753 | __set_current_state(TASK_RUNNING); |
b03fbd4f | 4754 | if (task_is_running(current)) |
85f4d4b6 | 4755 | return 1; |
c6699d6f | 4756 | |
ef99b2d3 | 4757 | if (ret < 0 || (flags & BLK_POLL_ONESHOT)) |
bbd7bb70 JA |
4758 | break; |
4759 | cpu_relax(); | |
aa61bec3 | 4760 | } while (!need_resched()); |
bbd7bb70 | 4761 | |
67b4110f | 4762 | __set_current_state(TASK_RUNNING); |
85f4d4b6 | 4763 | return 0; |
bbd7bb70 | 4764 | } |
1052b8ac | 4765 | |
9cf2bab6 JA |
4766 | unsigned int blk_mq_rq_cpu(struct request *rq) |
4767 | { | |
4768 | return rq->mq_ctx->cpu; | |
4769 | } | |
4770 | EXPORT_SYMBOL(blk_mq_rq_cpu); | |
4771 | ||
2a19b28f ML |
4772 | void blk_mq_cancel_work_sync(struct request_queue *q) |
4773 | { | |
219cf43c JC |
4774 | struct blk_mq_hw_ctx *hctx; |
4775 | unsigned long i; | |
2a19b28f | 4776 | |
219cf43c | 4777 | cancel_delayed_work_sync(&q->requeue_work); |
2a19b28f | 4778 | |
219cf43c JC |
4779 | queue_for_each_hw_ctx(q, hctx, i) |
4780 | cancel_delayed_work_sync(&hctx->run_work); | |
2a19b28f ML |
4781 | } |
4782 | ||
320ae51f JA |
4783 | static int __init blk_mq_init(void) |
4784 | { | |
c3077b5d CH |
4785 | int i; |
4786 | ||
4787 | for_each_possible_cpu(i) | |
f9ab4918 | 4788 | init_llist_head(&per_cpu(blk_cpu_done, i)); |
c3077b5d CH |
4789 | open_softirq(BLOCK_SOFTIRQ, blk_done_softirq); |
4790 | ||
4791 | cpuhp_setup_state_nocalls(CPUHP_BLOCK_SOFTIRQ_DEAD, | |
4792 | "block/softirq:dead", NULL, | |
4793 | blk_softirq_cpu_dead); | |
9467f859 TG |
4794 | cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL, |
4795 | blk_mq_hctx_notify_dead); | |
bf0beec0 ML |
4796 | cpuhp_setup_state_multi(CPUHP_AP_BLK_MQ_ONLINE, "block/mq:online", |
4797 | blk_mq_hctx_notify_online, | |
4798 | blk_mq_hctx_notify_offline); | |
320ae51f JA |
4799 | return 0; |
4800 | } | |
4801 | subsys_initcall(blk_mq_init); |