Commit | Line | Data |
---|---|---|
75bb4625 JA |
1 | /* |
2 | * Block multiqueue core code | |
3 | * | |
4 | * Copyright (C) 2013-2014 Jens Axboe | |
5 | * Copyright (C) 2013-2014 Christoph Hellwig | |
6 | */ | |
320ae51f JA |
7 | #include <linux/kernel.h> |
8 | #include <linux/module.h> | |
9 | #include <linux/backing-dev.h> | |
10 | #include <linux/bio.h> | |
11 | #include <linux/blkdev.h> | |
f75782e4 | 12 | #include <linux/kmemleak.h> |
320ae51f JA |
13 | #include <linux/mm.h> |
14 | #include <linux/init.h> | |
15 | #include <linux/slab.h> | |
16 | #include <linux/workqueue.h> | |
17 | #include <linux/smp.h> | |
18 | #include <linux/llist.h> | |
19 | #include <linux/list_sort.h> | |
20 | #include <linux/cpu.h> | |
21 | #include <linux/cache.h> | |
22 | #include <linux/sched/sysctl.h> | |
105ab3d8 | 23 | #include <linux/sched/topology.h> |
174cd4b1 | 24 | #include <linux/sched/signal.h> |
320ae51f | 25 | #include <linux/delay.h> |
aedcd72f | 26 | #include <linux/crash_dump.h> |
88c7b2b7 | 27 | #include <linux/prefetch.h> |
320ae51f JA |
28 | |
29 | #include <trace/events/block.h> | |
30 | ||
31 | #include <linux/blk-mq.h> | |
32 | #include "blk.h" | |
33 | #include "blk-mq.h" | |
9c1051aa | 34 | #include "blk-mq-debugfs.h" |
320ae51f | 35 | #include "blk-mq-tag.h" |
cf43e6be | 36 | #include "blk-stat.h" |
bd166ef1 | 37 | #include "blk-mq-sched.h" |
c1c80384 | 38 | #include "blk-rq-qos.h" |
320ae51f | 39 | |
ea435e1b | 40 | static bool blk_mq_poll(struct request_queue *q, blk_qc_t cookie); |
34dbad5d OS |
41 | static void blk_mq_poll_stats_start(struct request_queue *q); |
42 | static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb); | |
43 | ||
720b8ccc SB |
44 | static int blk_mq_poll_stats_bkt(const struct request *rq) |
45 | { | |
46 | int ddir, bytes, bucket; | |
47 | ||
99c749a4 | 48 | ddir = rq_data_dir(rq); |
720b8ccc SB |
49 | bytes = blk_rq_bytes(rq); |
50 | ||
51 | bucket = ddir + 2*(ilog2(bytes) - 9); | |
52 | ||
53 | if (bucket < 0) | |
54 | return -1; | |
55 | else if (bucket >= BLK_MQ_POLL_STATS_BKTS) | |
56 | return ddir + BLK_MQ_POLL_STATS_BKTS - 2; | |
57 | ||
58 | return bucket; | |
59 | } | |
60 | ||
320ae51f JA |
61 | /* |
62 | * Check if any of the ctx's have pending work in this hardware queue | |
63 | */ | |
79f720a7 | 64 | static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx) |
320ae51f | 65 | { |
79f720a7 JA |
66 | return !list_empty_careful(&hctx->dispatch) || |
67 | sbitmap_any_bit_set(&hctx->ctx_map) || | |
bd166ef1 | 68 | blk_mq_sched_has_work(hctx); |
1429d7c9 JA |
69 | } |
70 | ||
320ae51f JA |
71 | /* |
72 | * Mark this ctx as having pending work in this hardware queue | |
73 | */ | |
74 | static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx, | |
75 | struct blk_mq_ctx *ctx) | |
76 | { | |
88459642 OS |
77 | if (!sbitmap_test_bit(&hctx->ctx_map, ctx->index_hw)) |
78 | sbitmap_set_bit(&hctx->ctx_map, ctx->index_hw); | |
1429d7c9 JA |
79 | } |
80 | ||
81 | static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx, | |
82 | struct blk_mq_ctx *ctx) | |
83 | { | |
88459642 | 84 | sbitmap_clear_bit(&hctx->ctx_map, ctx->index_hw); |
320ae51f JA |
85 | } |
86 | ||
f299b7c7 JA |
87 | struct mq_inflight { |
88 | struct hd_struct *part; | |
89 | unsigned int *inflight; | |
90 | }; | |
91 | ||
92 | static void blk_mq_check_inflight(struct blk_mq_hw_ctx *hctx, | |
93 | struct request *rq, void *priv, | |
94 | bool reserved) | |
95 | { | |
96 | struct mq_inflight *mi = priv; | |
97 | ||
6131837b OS |
98 | /* |
99 | * index[0] counts the specific partition that was asked for. index[1] | |
100 | * counts the ones that are active on the whole device, so increment | |
101 | * that if mi->part is indeed a partition, and not a whole device. | |
102 | */ | |
103 | if (rq->part == mi->part) | |
104 | mi->inflight[0]++; | |
105 | if (mi->part->partno) | |
106 | mi->inflight[1]++; | |
f299b7c7 JA |
107 | } |
108 | ||
109 | void blk_mq_in_flight(struct request_queue *q, struct hd_struct *part, | |
110 | unsigned int inflight[2]) | |
111 | { | |
112 | struct mq_inflight mi = { .part = part, .inflight = inflight, }; | |
113 | ||
b8d62b3a | 114 | inflight[0] = inflight[1] = 0; |
f299b7c7 JA |
115 | blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi); |
116 | } | |
117 | ||
bf0ddaba OS |
118 | static void blk_mq_check_inflight_rw(struct blk_mq_hw_ctx *hctx, |
119 | struct request *rq, void *priv, | |
120 | bool reserved) | |
121 | { | |
122 | struct mq_inflight *mi = priv; | |
123 | ||
124 | if (rq->part == mi->part) | |
125 | mi->inflight[rq_data_dir(rq)]++; | |
126 | } | |
127 | ||
128 | void blk_mq_in_flight_rw(struct request_queue *q, struct hd_struct *part, | |
129 | unsigned int inflight[2]) | |
130 | { | |
131 | struct mq_inflight mi = { .part = part, .inflight = inflight, }; | |
132 | ||
133 | inflight[0] = inflight[1] = 0; | |
134 | blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight_rw, &mi); | |
135 | } | |
136 | ||
1671d522 | 137 | void blk_freeze_queue_start(struct request_queue *q) |
43a5e4e2 | 138 | { |
4ecd4fef | 139 | int freeze_depth; |
cddd5d17 | 140 | |
4ecd4fef CH |
141 | freeze_depth = atomic_inc_return(&q->mq_freeze_depth); |
142 | if (freeze_depth == 1) { | |
3ef28e83 | 143 | percpu_ref_kill(&q->q_usage_counter); |
055f6e18 ML |
144 | if (q->mq_ops) |
145 | blk_mq_run_hw_queues(q, false); | |
cddd5d17 | 146 | } |
f3af020b | 147 | } |
1671d522 | 148 | EXPORT_SYMBOL_GPL(blk_freeze_queue_start); |
f3af020b | 149 | |
6bae363e | 150 | void blk_mq_freeze_queue_wait(struct request_queue *q) |
f3af020b | 151 | { |
3ef28e83 | 152 | wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter)); |
43a5e4e2 | 153 | } |
6bae363e | 154 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait); |
43a5e4e2 | 155 | |
f91328c4 KB |
156 | int blk_mq_freeze_queue_wait_timeout(struct request_queue *q, |
157 | unsigned long timeout) | |
158 | { | |
159 | return wait_event_timeout(q->mq_freeze_wq, | |
160 | percpu_ref_is_zero(&q->q_usage_counter), | |
161 | timeout); | |
162 | } | |
163 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout); | |
43a5e4e2 | 164 | |
f3af020b TH |
165 | /* |
166 | * Guarantee no request is in use, so we can change any data structure of | |
167 | * the queue afterward. | |
168 | */ | |
3ef28e83 | 169 | void blk_freeze_queue(struct request_queue *q) |
f3af020b | 170 | { |
3ef28e83 DW |
171 | /* |
172 | * In the !blk_mq case we are only calling this to kill the | |
173 | * q_usage_counter, otherwise this increases the freeze depth | |
174 | * and waits for it to return to zero. For this reason there is | |
175 | * no blk_unfreeze_queue(), and blk_freeze_queue() is not | |
176 | * exported to drivers as the only user for unfreeze is blk_mq. | |
177 | */ | |
1671d522 | 178 | blk_freeze_queue_start(q); |
454be724 ML |
179 | if (!q->mq_ops) |
180 | blk_drain_queue(q); | |
f3af020b TH |
181 | blk_mq_freeze_queue_wait(q); |
182 | } | |
3ef28e83 DW |
183 | |
184 | void blk_mq_freeze_queue(struct request_queue *q) | |
185 | { | |
186 | /* | |
187 | * ...just an alias to keep freeze and unfreeze actions balanced | |
188 | * in the blk_mq_* namespace | |
189 | */ | |
190 | blk_freeze_queue(q); | |
191 | } | |
c761d96b | 192 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue); |
f3af020b | 193 | |
b4c6a028 | 194 | void blk_mq_unfreeze_queue(struct request_queue *q) |
320ae51f | 195 | { |
4ecd4fef | 196 | int freeze_depth; |
320ae51f | 197 | |
4ecd4fef CH |
198 | freeze_depth = atomic_dec_return(&q->mq_freeze_depth); |
199 | WARN_ON_ONCE(freeze_depth < 0); | |
200 | if (!freeze_depth) { | |
3ef28e83 | 201 | percpu_ref_reinit(&q->q_usage_counter); |
320ae51f | 202 | wake_up_all(&q->mq_freeze_wq); |
add703fd | 203 | } |
320ae51f | 204 | } |
b4c6a028 | 205 | EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue); |
320ae51f | 206 | |
852ec809 BVA |
207 | /* |
208 | * FIXME: replace the scsi_internal_device_*block_nowait() calls in the | |
209 | * mpt3sas driver such that this function can be removed. | |
210 | */ | |
211 | void blk_mq_quiesce_queue_nowait(struct request_queue *q) | |
212 | { | |
8814ce8a | 213 | blk_queue_flag_set(QUEUE_FLAG_QUIESCED, q); |
852ec809 BVA |
214 | } |
215 | EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait); | |
216 | ||
6a83e74d | 217 | /** |
69e07c4a | 218 | * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished |
6a83e74d BVA |
219 | * @q: request queue. |
220 | * | |
221 | * Note: this function does not prevent that the struct request end_io() | |
69e07c4a ML |
222 | * callback function is invoked. Once this function is returned, we make |
223 | * sure no dispatch can happen until the queue is unquiesced via | |
224 | * blk_mq_unquiesce_queue(). | |
6a83e74d BVA |
225 | */ |
226 | void blk_mq_quiesce_queue(struct request_queue *q) | |
227 | { | |
228 | struct blk_mq_hw_ctx *hctx; | |
229 | unsigned int i; | |
230 | bool rcu = false; | |
231 | ||
1d9e9bc6 | 232 | blk_mq_quiesce_queue_nowait(q); |
f4560ffe | 233 | |
6a83e74d BVA |
234 | queue_for_each_hw_ctx(q, hctx, i) { |
235 | if (hctx->flags & BLK_MQ_F_BLOCKING) | |
05707b64 | 236 | synchronize_srcu(hctx->srcu); |
6a83e74d BVA |
237 | else |
238 | rcu = true; | |
239 | } | |
240 | if (rcu) | |
241 | synchronize_rcu(); | |
242 | } | |
243 | EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue); | |
244 | ||
e4e73913 ML |
245 | /* |
246 | * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue() | |
247 | * @q: request queue. | |
248 | * | |
249 | * This function recovers queue into the state before quiescing | |
250 | * which is done by blk_mq_quiesce_queue. | |
251 | */ | |
252 | void blk_mq_unquiesce_queue(struct request_queue *q) | |
253 | { | |
8814ce8a | 254 | blk_queue_flag_clear(QUEUE_FLAG_QUIESCED, q); |
f4560ffe | 255 | |
1d9e9bc6 ML |
256 | /* dispatch requests which are inserted during quiescing */ |
257 | blk_mq_run_hw_queues(q, true); | |
e4e73913 ML |
258 | } |
259 | EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue); | |
260 | ||
aed3ea94 JA |
261 | void blk_mq_wake_waiters(struct request_queue *q) |
262 | { | |
263 | struct blk_mq_hw_ctx *hctx; | |
264 | unsigned int i; | |
265 | ||
266 | queue_for_each_hw_ctx(q, hctx, i) | |
267 | if (blk_mq_hw_queue_mapped(hctx)) | |
268 | blk_mq_tag_wakeup_all(hctx->tags, true); | |
269 | } | |
270 | ||
320ae51f JA |
271 | bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx) |
272 | { | |
273 | return blk_mq_has_free_tags(hctx->tags); | |
274 | } | |
275 | EXPORT_SYMBOL(blk_mq_can_queue); | |
276 | ||
e4cdf1a1 CH |
277 | static struct request *blk_mq_rq_ctx_init(struct blk_mq_alloc_data *data, |
278 | unsigned int tag, unsigned int op) | |
320ae51f | 279 | { |
e4cdf1a1 CH |
280 | struct blk_mq_tags *tags = blk_mq_tags_from_data(data); |
281 | struct request *rq = tags->static_rqs[tag]; | |
bf9ae8c5 | 282 | req_flags_t rq_flags = 0; |
c3a148d2 | 283 | |
e4cdf1a1 CH |
284 | if (data->flags & BLK_MQ_REQ_INTERNAL) { |
285 | rq->tag = -1; | |
286 | rq->internal_tag = tag; | |
287 | } else { | |
d263ed99 | 288 | if (data->hctx->flags & BLK_MQ_F_TAG_SHARED) { |
bf9ae8c5 | 289 | rq_flags = RQF_MQ_INFLIGHT; |
e4cdf1a1 CH |
290 | atomic_inc(&data->hctx->nr_active); |
291 | } | |
292 | rq->tag = tag; | |
293 | rq->internal_tag = -1; | |
294 | data->hctx->tags->rqs[rq->tag] = rq; | |
295 | } | |
296 | ||
af76e555 | 297 | /* csd/requeue_work/fifo_time is initialized before use */ |
e4cdf1a1 CH |
298 | rq->q = data->q; |
299 | rq->mq_ctx = data->ctx; | |
bf9ae8c5 | 300 | rq->rq_flags = rq_flags; |
7c3fb70f | 301 | rq->cpu = -1; |
ef295ecf | 302 | rq->cmd_flags = op; |
1b6d65a0 BVA |
303 | if (data->flags & BLK_MQ_REQ_PREEMPT) |
304 | rq->rq_flags |= RQF_PREEMPT; | |
e4cdf1a1 | 305 | if (blk_queue_io_stat(data->q)) |
e8064021 | 306 | rq->rq_flags |= RQF_IO_STAT; |
7c3fb70f | 307 | INIT_LIST_HEAD(&rq->queuelist); |
af76e555 CH |
308 | INIT_HLIST_NODE(&rq->hash); |
309 | RB_CLEAR_NODE(&rq->rb_node); | |
af76e555 CH |
310 | rq->rq_disk = NULL; |
311 | rq->part = NULL; | |
522a7775 | 312 | rq->start_time_ns = ktime_get_ns(); |
544ccc8d | 313 | rq->io_start_time_ns = 0; |
af76e555 CH |
314 | rq->nr_phys_segments = 0; |
315 | #if defined(CONFIG_BLK_DEV_INTEGRITY) | |
316 | rq->nr_integrity_segments = 0; | |
317 | #endif | |
af76e555 CH |
318 | rq->special = NULL; |
319 | /* tag was already set */ | |
af76e555 | 320 | rq->extra_len = 0; |
e14575b3 | 321 | rq->__deadline = 0; |
af76e555 | 322 | |
af76e555 | 323 | INIT_LIST_HEAD(&rq->timeout_list); |
f6be4fb4 JA |
324 | rq->timeout = 0; |
325 | ||
af76e555 CH |
326 | rq->end_io = NULL; |
327 | rq->end_io_data = NULL; | |
328 | rq->next_rq = NULL; | |
329 | ||
7c3fb70f JA |
330 | #ifdef CONFIG_BLK_CGROUP |
331 | rq->rl = NULL; | |
7c3fb70f JA |
332 | #endif |
333 | ||
e4cdf1a1 | 334 | data->ctx->rq_dispatched[op_is_sync(op)]++; |
12f5b931 | 335 | refcount_set(&rq->ref, 1); |
e4cdf1a1 | 336 | return rq; |
5dee8577 CH |
337 | } |
338 | ||
d2c0d383 CH |
339 | static struct request *blk_mq_get_request(struct request_queue *q, |
340 | struct bio *bio, unsigned int op, | |
341 | struct blk_mq_alloc_data *data) | |
342 | { | |
343 | struct elevator_queue *e = q->elevator; | |
344 | struct request *rq; | |
e4cdf1a1 | 345 | unsigned int tag; |
21e768b4 | 346 | bool put_ctx_on_error = false; |
d2c0d383 CH |
347 | |
348 | blk_queue_enter_live(q); | |
349 | data->q = q; | |
21e768b4 BVA |
350 | if (likely(!data->ctx)) { |
351 | data->ctx = blk_mq_get_ctx(q); | |
352 | put_ctx_on_error = true; | |
353 | } | |
d2c0d383 CH |
354 | if (likely(!data->hctx)) |
355 | data->hctx = blk_mq_map_queue(q, data->ctx->cpu); | |
03a07c92 GR |
356 | if (op & REQ_NOWAIT) |
357 | data->flags |= BLK_MQ_REQ_NOWAIT; | |
d2c0d383 CH |
358 | |
359 | if (e) { | |
360 | data->flags |= BLK_MQ_REQ_INTERNAL; | |
361 | ||
362 | /* | |
363 | * Flush requests are special and go directly to the | |
17a51199 JA |
364 | * dispatch list. Don't include reserved tags in the |
365 | * limiting, as it isn't useful. | |
d2c0d383 | 366 | */ |
17a51199 JA |
367 | if (!op_is_flush(op) && e->type->ops.mq.limit_depth && |
368 | !(data->flags & BLK_MQ_REQ_RESERVED)) | |
5bbf4e5a | 369 | e->type->ops.mq.limit_depth(op, data); |
d263ed99 JW |
370 | } else { |
371 | blk_mq_tag_busy(data->hctx); | |
d2c0d383 CH |
372 | } |
373 | ||
e4cdf1a1 CH |
374 | tag = blk_mq_get_tag(data); |
375 | if (tag == BLK_MQ_TAG_FAIL) { | |
21e768b4 BVA |
376 | if (put_ctx_on_error) { |
377 | blk_mq_put_ctx(data->ctx); | |
1ad43c00 ML |
378 | data->ctx = NULL; |
379 | } | |
037cebb8 CH |
380 | blk_queue_exit(q); |
381 | return NULL; | |
d2c0d383 CH |
382 | } |
383 | ||
e4cdf1a1 | 384 | rq = blk_mq_rq_ctx_init(data, tag, op); |
037cebb8 CH |
385 | if (!op_is_flush(op)) { |
386 | rq->elv.icq = NULL; | |
5bbf4e5a | 387 | if (e && e->type->ops.mq.prepare_request) { |
44e8c2bf CH |
388 | if (e->type->icq_cache && rq_ioc(bio)) |
389 | blk_mq_sched_assign_ioc(rq, bio); | |
390 | ||
5bbf4e5a CH |
391 | e->type->ops.mq.prepare_request(rq, bio); |
392 | rq->rq_flags |= RQF_ELVPRIV; | |
44e8c2bf | 393 | } |
037cebb8 CH |
394 | } |
395 | data->hctx->queued++; | |
396 | return rq; | |
d2c0d383 CH |
397 | } |
398 | ||
cd6ce148 | 399 | struct request *blk_mq_alloc_request(struct request_queue *q, unsigned int op, |
9a95e4ef | 400 | blk_mq_req_flags_t flags) |
320ae51f | 401 | { |
5a797e00 | 402 | struct blk_mq_alloc_data alloc_data = { .flags = flags }; |
bd166ef1 | 403 | struct request *rq; |
a492f075 | 404 | int ret; |
320ae51f | 405 | |
3a0a5299 | 406 | ret = blk_queue_enter(q, flags); |
a492f075 JL |
407 | if (ret) |
408 | return ERR_PTR(ret); | |
320ae51f | 409 | |
cd6ce148 | 410 | rq = blk_mq_get_request(q, NULL, op, &alloc_data); |
3280d66a | 411 | blk_queue_exit(q); |
841bac2c | 412 | |
bd166ef1 | 413 | if (!rq) |
a492f075 | 414 | return ERR_PTR(-EWOULDBLOCK); |
0c4de0f3 | 415 | |
1ad43c00 | 416 | blk_mq_put_ctx(alloc_data.ctx); |
1ad43c00 | 417 | |
0c4de0f3 CH |
418 | rq->__data_len = 0; |
419 | rq->__sector = (sector_t) -1; | |
420 | rq->bio = rq->biotail = NULL; | |
320ae51f JA |
421 | return rq; |
422 | } | |
4bb659b1 | 423 | EXPORT_SYMBOL(blk_mq_alloc_request); |
320ae51f | 424 | |
cd6ce148 | 425 | struct request *blk_mq_alloc_request_hctx(struct request_queue *q, |
9a95e4ef | 426 | unsigned int op, blk_mq_req_flags_t flags, unsigned int hctx_idx) |
1f5bd336 | 427 | { |
6d2809d5 | 428 | struct blk_mq_alloc_data alloc_data = { .flags = flags }; |
1f5bd336 | 429 | struct request *rq; |
6d2809d5 | 430 | unsigned int cpu; |
1f5bd336 ML |
431 | int ret; |
432 | ||
433 | /* | |
434 | * If the tag allocator sleeps we could get an allocation for a | |
435 | * different hardware context. No need to complicate the low level | |
436 | * allocator for this for the rare use case of a command tied to | |
437 | * a specific queue. | |
438 | */ | |
439 | if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT))) | |
440 | return ERR_PTR(-EINVAL); | |
441 | ||
442 | if (hctx_idx >= q->nr_hw_queues) | |
443 | return ERR_PTR(-EIO); | |
444 | ||
3a0a5299 | 445 | ret = blk_queue_enter(q, flags); |
1f5bd336 ML |
446 | if (ret) |
447 | return ERR_PTR(ret); | |
448 | ||
c8712c6a CH |
449 | /* |
450 | * Check if the hardware context is actually mapped to anything. | |
451 | * If not tell the caller that it should skip this queue. | |
452 | */ | |
6d2809d5 OS |
453 | alloc_data.hctx = q->queue_hw_ctx[hctx_idx]; |
454 | if (!blk_mq_hw_queue_mapped(alloc_data.hctx)) { | |
455 | blk_queue_exit(q); | |
456 | return ERR_PTR(-EXDEV); | |
c8712c6a | 457 | } |
20e4d813 | 458 | cpu = cpumask_first_and(alloc_data.hctx->cpumask, cpu_online_mask); |
6d2809d5 | 459 | alloc_data.ctx = __blk_mq_get_ctx(q, cpu); |
1f5bd336 | 460 | |
cd6ce148 | 461 | rq = blk_mq_get_request(q, NULL, op, &alloc_data); |
3280d66a | 462 | blk_queue_exit(q); |
c8712c6a | 463 | |
6d2809d5 OS |
464 | if (!rq) |
465 | return ERR_PTR(-EWOULDBLOCK); | |
466 | ||
467 | return rq; | |
1f5bd336 ML |
468 | } |
469 | EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx); | |
470 | ||
12f5b931 KB |
471 | static void __blk_mq_free_request(struct request *rq) |
472 | { | |
473 | struct request_queue *q = rq->q; | |
474 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
475 | struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, ctx->cpu); | |
476 | const int sched_tag = rq->internal_tag; | |
477 | ||
478 | if (rq->tag != -1) | |
479 | blk_mq_put_tag(hctx, hctx->tags, ctx, rq->tag); | |
480 | if (sched_tag != -1) | |
481 | blk_mq_put_tag(hctx, hctx->sched_tags, ctx, sched_tag); | |
482 | blk_mq_sched_restart(hctx); | |
483 | blk_queue_exit(q); | |
484 | } | |
485 | ||
6af54051 | 486 | void blk_mq_free_request(struct request *rq) |
320ae51f | 487 | { |
320ae51f | 488 | struct request_queue *q = rq->q; |
6af54051 CH |
489 | struct elevator_queue *e = q->elevator; |
490 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
491 | struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, ctx->cpu); | |
6af54051 | 492 | |
5bbf4e5a | 493 | if (rq->rq_flags & RQF_ELVPRIV) { |
6af54051 CH |
494 | if (e && e->type->ops.mq.finish_request) |
495 | e->type->ops.mq.finish_request(rq); | |
496 | if (rq->elv.icq) { | |
497 | put_io_context(rq->elv.icq->ioc); | |
498 | rq->elv.icq = NULL; | |
499 | } | |
500 | } | |
320ae51f | 501 | |
6af54051 | 502 | ctx->rq_completed[rq_is_sync(rq)]++; |
e8064021 | 503 | if (rq->rq_flags & RQF_MQ_INFLIGHT) |
0d2602ca | 504 | atomic_dec(&hctx->nr_active); |
87760e5e | 505 | |
7beb2f84 JA |
506 | if (unlikely(laptop_mode && !blk_rq_is_passthrough(rq))) |
507 | laptop_io_completion(q->backing_dev_info); | |
508 | ||
a7905043 | 509 | rq_qos_done(q, rq); |
0d2602ca | 510 | |
85acb3ba SL |
511 | if (blk_rq_rl(rq)) |
512 | blk_put_rl(blk_rq_rl(rq)); | |
513 | ||
12f5b931 KB |
514 | WRITE_ONCE(rq->state, MQ_RQ_IDLE); |
515 | if (refcount_dec_and_test(&rq->ref)) | |
516 | __blk_mq_free_request(rq); | |
320ae51f | 517 | } |
1a3b595a | 518 | EXPORT_SYMBOL_GPL(blk_mq_free_request); |
320ae51f | 519 | |
2a842aca | 520 | inline void __blk_mq_end_request(struct request *rq, blk_status_t error) |
320ae51f | 521 | { |
522a7775 OS |
522 | u64 now = ktime_get_ns(); |
523 | ||
4bc6339a OS |
524 | if (rq->rq_flags & RQF_STATS) { |
525 | blk_mq_poll_stats_start(rq->q); | |
522a7775 | 526 | blk_stat_add(rq, now); |
4bc6339a OS |
527 | } |
528 | ||
522a7775 | 529 | blk_account_io_done(rq, now); |
0d11e6ac | 530 | |
91b63639 | 531 | if (rq->end_io) { |
a7905043 | 532 | rq_qos_done(rq->q, rq); |
320ae51f | 533 | rq->end_io(rq, error); |
91b63639 CH |
534 | } else { |
535 | if (unlikely(blk_bidi_rq(rq))) | |
536 | blk_mq_free_request(rq->next_rq); | |
320ae51f | 537 | blk_mq_free_request(rq); |
91b63639 | 538 | } |
320ae51f | 539 | } |
c8a446ad | 540 | EXPORT_SYMBOL(__blk_mq_end_request); |
63151a44 | 541 | |
2a842aca | 542 | void blk_mq_end_request(struct request *rq, blk_status_t error) |
63151a44 CH |
543 | { |
544 | if (blk_update_request(rq, error, blk_rq_bytes(rq))) | |
545 | BUG(); | |
c8a446ad | 546 | __blk_mq_end_request(rq, error); |
63151a44 | 547 | } |
c8a446ad | 548 | EXPORT_SYMBOL(blk_mq_end_request); |
320ae51f | 549 | |
30a91cb4 | 550 | static void __blk_mq_complete_request_remote(void *data) |
320ae51f | 551 | { |
3d6efbf6 | 552 | struct request *rq = data; |
320ae51f | 553 | |
30a91cb4 | 554 | rq->q->softirq_done_fn(rq); |
320ae51f | 555 | } |
320ae51f | 556 | |
453f8341 | 557 | static void __blk_mq_complete_request(struct request *rq) |
320ae51f JA |
558 | { |
559 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
38535201 | 560 | bool shared = false; |
320ae51f JA |
561 | int cpu; |
562 | ||
0fc09f92 | 563 | if (!blk_mq_mark_complete(rq)) |
12f5b931 | 564 | return; |
453f8341 CH |
565 | if (rq->internal_tag != -1) |
566 | blk_mq_sched_completed_request(rq); | |
453f8341 | 567 | |
38535201 | 568 | if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) { |
30a91cb4 CH |
569 | rq->q->softirq_done_fn(rq); |
570 | return; | |
571 | } | |
320ae51f JA |
572 | |
573 | cpu = get_cpu(); | |
38535201 CH |
574 | if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags)) |
575 | shared = cpus_share_cache(cpu, ctx->cpu); | |
576 | ||
577 | if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) { | |
30a91cb4 | 578 | rq->csd.func = __blk_mq_complete_request_remote; |
3d6efbf6 CH |
579 | rq->csd.info = rq; |
580 | rq->csd.flags = 0; | |
c46fff2a | 581 | smp_call_function_single_async(ctx->cpu, &rq->csd); |
3d6efbf6 | 582 | } else { |
30a91cb4 | 583 | rq->q->softirq_done_fn(rq); |
3d6efbf6 | 584 | } |
320ae51f JA |
585 | put_cpu(); |
586 | } | |
30a91cb4 | 587 | |
04ced159 | 588 | static void hctx_unlock(struct blk_mq_hw_ctx *hctx, int srcu_idx) |
b7435db8 | 589 | __releases(hctx->srcu) |
04ced159 JA |
590 | { |
591 | if (!(hctx->flags & BLK_MQ_F_BLOCKING)) | |
592 | rcu_read_unlock(); | |
593 | else | |
05707b64 | 594 | srcu_read_unlock(hctx->srcu, srcu_idx); |
04ced159 JA |
595 | } |
596 | ||
597 | static void hctx_lock(struct blk_mq_hw_ctx *hctx, int *srcu_idx) | |
b7435db8 | 598 | __acquires(hctx->srcu) |
04ced159 | 599 | { |
08b5a6e2 JA |
600 | if (!(hctx->flags & BLK_MQ_F_BLOCKING)) { |
601 | /* shut up gcc false positive */ | |
602 | *srcu_idx = 0; | |
04ced159 | 603 | rcu_read_lock(); |
08b5a6e2 | 604 | } else |
05707b64 | 605 | *srcu_idx = srcu_read_lock(hctx->srcu); |
04ced159 JA |
606 | } |
607 | ||
30a91cb4 CH |
608 | /** |
609 | * blk_mq_complete_request - end I/O on a request | |
610 | * @rq: the request being processed | |
611 | * | |
612 | * Description: | |
613 | * Ends all I/O on a request. It does not handle partial completions. | |
614 | * The actual completion happens out-of-order, through a IPI handler. | |
615 | **/ | |
08e0029a | 616 | void blk_mq_complete_request(struct request *rq) |
30a91cb4 | 617 | { |
12f5b931 | 618 | if (unlikely(blk_should_fake_timeout(rq->q))) |
30a91cb4 | 619 | return; |
12f5b931 | 620 | __blk_mq_complete_request(rq); |
30a91cb4 CH |
621 | } |
622 | EXPORT_SYMBOL(blk_mq_complete_request); | |
320ae51f | 623 | |
973c0191 KB |
624 | int blk_mq_request_started(struct request *rq) |
625 | { | |
5a61c363 | 626 | return blk_mq_rq_state(rq) != MQ_RQ_IDLE; |
973c0191 KB |
627 | } |
628 | EXPORT_SYMBOL_GPL(blk_mq_request_started); | |
629 | ||
e2490073 | 630 | void blk_mq_start_request(struct request *rq) |
320ae51f JA |
631 | { |
632 | struct request_queue *q = rq->q; | |
633 | ||
bd166ef1 JA |
634 | blk_mq_sched_started_request(rq); |
635 | ||
320ae51f JA |
636 | trace_block_rq_issue(q, rq); |
637 | ||
cf43e6be | 638 | if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) { |
544ccc8d OS |
639 | rq->io_start_time_ns = ktime_get_ns(); |
640 | #ifdef CONFIG_BLK_DEV_THROTTLING_LOW | |
641 | rq->throtl_size = blk_rq_sectors(rq); | |
642 | #endif | |
cf43e6be | 643 | rq->rq_flags |= RQF_STATS; |
a7905043 | 644 | rq_qos_issue(q, rq); |
cf43e6be JA |
645 | } |
646 | ||
1d9bd516 | 647 | WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE); |
538b7534 | 648 | |
1d9bd516 | 649 | blk_add_timer(rq); |
12f5b931 | 650 | WRITE_ONCE(rq->state, MQ_RQ_IN_FLIGHT); |
49f5baa5 CH |
651 | |
652 | if (q->dma_drain_size && blk_rq_bytes(rq)) { | |
653 | /* | |
654 | * Make sure space for the drain appears. We know we can do | |
655 | * this because max_hw_segments has been adjusted to be one | |
656 | * fewer than the device can handle. | |
657 | */ | |
658 | rq->nr_phys_segments++; | |
659 | } | |
320ae51f | 660 | } |
e2490073 | 661 | EXPORT_SYMBOL(blk_mq_start_request); |
320ae51f | 662 | |
ed0791b2 | 663 | static void __blk_mq_requeue_request(struct request *rq) |
320ae51f JA |
664 | { |
665 | struct request_queue *q = rq->q; | |
666 | ||
923218f6 ML |
667 | blk_mq_put_driver_tag(rq); |
668 | ||
320ae51f | 669 | trace_block_rq_requeue(q, rq); |
a7905043 | 670 | rq_qos_requeue(q, rq); |
49f5baa5 | 671 | |
12f5b931 KB |
672 | if (blk_mq_request_started(rq)) { |
673 | WRITE_ONCE(rq->state, MQ_RQ_IDLE); | |
da661267 | 674 | rq->rq_flags &= ~RQF_TIMED_OUT; |
e2490073 CH |
675 | if (q->dma_drain_size && blk_rq_bytes(rq)) |
676 | rq->nr_phys_segments--; | |
677 | } | |
320ae51f JA |
678 | } |
679 | ||
2b053aca | 680 | void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list) |
ed0791b2 | 681 | { |
ed0791b2 | 682 | __blk_mq_requeue_request(rq); |
ed0791b2 | 683 | |
105976f5 ML |
684 | /* this request will be re-inserted to io scheduler queue */ |
685 | blk_mq_sched_requeue_request(rq); | |
686 | ||
ed0791b2 | 687 | BUG_ON(blk_queued_rq(rq)); |
2b053aca | 688 | blk_mq_add_to_requeue_list(rq, true, kick_requeue_list); |
ed0791b2 CH |
689 | } |
690 | EXPORT_SYMBOL(blk_mq_requeue_request); | |
691 | ||
6fca6a61 CH |
692 | static void blk_mq_requeue_work(struct work_struct *work) |
693 | { | |
694 | struct request_queue *q = | |
2849450a | 695 | container_of(work, struct request_queue, requeue_work.work); |
6fca6a61 CH |
696 | LIST_HEAD(rq_list); |
697 | struct request *rq, *next; | |
6fca6a61 | 698 | |
18e9781d | 699 | spin_lock_irq(&q->requeue_lock); |
6fca6a61 | 700 | list_splice_init(&q->requeue_list, &rq_list); |
18e9781d | 701 | spin_unlock_irq(&q->requeue_lock); |
6fca6a61 CH |
702 | |
703 | list_for_each_entry_safe(rq, next, &rq_list, queuelist) { | |
e8064021 | 704 | if (!(rq->rq_flags & RQF_SOFTBARRIER)) |
6fca6a61 CH |
705 | continue; |
706 | ||
e8064021 | 707 | rq->rq_flags &= ~RQF_SOFTBARRIER; |
6fca6a61 | 708 | list_del_init(&rq->queuelist); |
9e97d295 | 709 | blk_mq_sched_insert_request(rq, true, false, false); |
6fca6a61 CH |
710 | } |
711 | ||
712 | while (!list_empty(&rq_list)) { | |
713 | rq = list_entry(rq_list.next, struct request, queuelist); | |
714 | list_del_init(&rq->queuelist); | |
9e97d295 | 715 | blk_mq_sched_insert_request(rq, false, false, false); |
6fca6a61 CH |
716 | } |
717 | ||
52d7f1b5 | 718 | blk_mq_run_hw_queues(q, false); |
6fca6a61 CH |
719 | } |
720 | ||
2b053aca BVA |
721 | void blk_mq_add_to_requeue_list(struct request *rq, bool at_head, |
722 | bool kick_requeue_list) | |
6fca6a61 CH |
723 | { |
724 | struct request_queue *q = rq->q; | |
725 | unsigned long flags; | |
726 | ||
727 | /* | |
728 | * We abuse this flag that is otherwise used by the I/O scheduler to | |
ff821d27 | 729 | * request head insertion from the workqueue. |
6fca6a61 | 730 | */ |
e8064021 | 731 | BUG_ON(rq->rq_flags & RQF_SOFTBARRIER); |
6fca6a61 CH |
732 | |
733 | spin_lock_irqsave(&q->requeue_lock, flags); | |
734 | if (at_head) { | |
e8064021 | 735 | rq->rq_flags |= RQF_SOFTBARRIER; |
6fca6a61 CH |
736 | list_add(&rq->queuelist, &q->requeue_list); |
737 | } else { | |
738 | list_add_tail(&rq->queuelist, &q->requeue_list); | |
739 | } | |
740 | spin_unlock_irqrestore(&q->requeue_lock, flags); | |
2b053aca BVA |
741 | |
742 | if (kick_requeue_list) | |
743 | blk_mq_kick_requeue_list(q); | |
6fca6a61 CH |
744 | } |
745 | EXPORT_SYMBOL(blk_mq_add_to_requeue_list); | |
746 | ||
747 | void blk_mq_kick_requeue_list(struct request_queue *q) | |
748 | { | |
ae943d20 | 749 | kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0); |
6fca6a61 CH |
750 | } |
751 | EXPORT_SYMBOL(blk_mq_kick_requeue_list); | |
752 | ||
2849450a MS |
753 | void blk_mq_delay_kick_requeue_list(struct request_queue *q, |
754 | unsigned long msecs) | |
755 | { | |
d4acf365 BVA |
756 | kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, |
757 | msecs_to_jiffies(msecs)); | |
2849450a MS |
758 | } |
759 | EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list); | |
760 | ||
0e62f51f JA |
761 | struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag) |
762 | { | |
88c7b2b7 JA |
763 | if (tag < tags->nr_tags) { |
764 | prefetch(tags->rqs[tag]); | |
4ee86bab | 765 | return tags->rqs[tag]; |
88c7b2b7 | 766 | } |
4ee86bab HR |
767 | |
768 | return NULL; | |
24d2f903 CH |
769 | } |
770 | EXPORT_SYMBOL(blk_mq_tag_to_rq); | |
771 | ||
358f70da | 772 | static void blk_mq_rq_timed_out(struct request *req, bool reserved) |
320ae51f | 773 | { |
da661267 | 774 | req->rq_flags |= RQF_TIMED_OUT; |
d1210d5a CH |
775 | if (req->q->mq_ops->timeout) { |
776 | enum blk_eh_timer_return ret; | |
777 | ||
778 | ret = req->q->mq_ops->timeout(req, reserved); | |
779 | if (ret == BLK_EH_DONE) | |
780 | return; | |
781 | WARN_ON_ONCE(ret != BLK_EH_RESET_TIMER); | |
46f92d42 | 782 | } |
d1210d5a CH |
783 | |
784 | blk_add_timer(req); | |
87ee7b11 | 785 | } |
5b3f25fc | 786 | |
12f5b931 | 787 | static bool blk_mq_req_expired(struct request *rq, unsigned long *next) |
81481eb4 | 788 | { |
12f5b931 | 789 | unsigned long deadline; |
87ee7b11 | 790 | |
12f5b931 KB |
791 | if (blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT) |
792 | return false; | |
da661267 CH |
793 | if (rq->rq_flags & RQF_TIMED_OUT) |
794 | return false; | |
a7af0af3 | 795 | |
12f5b931 KB |
796 | deadline = blk_rq_deadline(rq); |
797 | if (time_after_eq(jiffies, deadline)) | |
798 | return true; | |
a7af0af3 | 799 | |
12f5b931 KB |
800 | if (*next == 0) |
801 | *next = deadline; | |
802 | else if (time_after(*next, deadline)) | |
803 | *next = deadline; | |
804 | return false; | |
87ee7b11 JA |
805 | } |
806 | ||
12f5b931 | 807 | static void blk_mq_check_expired(struct blk_mq_hw_ctx *hctx, |
1d9bd516 TH |
808 | struct request *rq, void *priv, bool reserved) |
809 | { | |
12f5b931 KB |
810 | unsigned long *next = priv; |
811 | ||
812 | /* | |
813 | * Just do a quick check if it is expired before locking the request in | |
814 | * so we're not unnecessarilly synchronizing across CPUs. | |
815 | */ | |
816 | if (!blk_mq_req_expired(rq, next)) | |
817 | return; | |
818 | ||
819 | /* | |
820 | * We have reason to believe the request may be expired. Take a | |
821 | * reference on the request to lock this request lifetime into its | |
822 | * currently allocated context to prevent it from being reallocated in | |
823 | * the event the completion by-passes this timeout handler. | |
824 | * | |
825 | * If the reference was already released, then the driver beat the | |
826 | * timeout handler to posting a natural completion. | |
827 | */ | |
828 | if (!refcount_inc_not_zero(&rq->ref)) | |
829 | return; | |
830 | ||
1d9bd516 | 831 | /* |
12f5b931 KB |
832 | * The request is now locked and cannot be reallocated underneath the |
833 | * timeout handler's processing. Re-verify this exact request is truly | |
834 | * expired; if it is not expired, then the request was completed and | |
835 | * reallocated as a new request. | |
1d9bd516 | 836 | */ |
12f5b931 | 837 | if (blk_mq_req_expired(rq, next)) |
1d9bd516 | 838 | blk_mq_rq_timed_out(rq, reserved); |
12f5b931 KB |
839 | if (refcount_dec_and_test(&rq->ref)) |
840 | __blk_mq_free_request(rq); | |
1d9bd516 TH |
841 | } |
842 | ||
287922eb | 843 | static void blk_mq_timeout_work(struct work_struct *work) |
320ae51f | 844 | { |
287922eb CH |
845 | struct request_queue *q = |
846 | container_of(work, struct request_queue, timeout_work); | |
12f5b931 | 847 | unsigned long next = 0; |
1d9bd516 | 848 | struct blk_mq_hw_ctx *hctx; |
81481eb4 | 849 | int i; |
320ae51f | 850 | |
71f79fb3 GKB |
851 | /* A deadlock might occur if a request is stuck requiring a |
852 | * timeout at the same time a queue freeze is waiting | |
853 | * completion, since the timeout code would not be able to | |
854 | * acquire the queue reference here. | |
855 | * | |
856 | * That's why we don't use blk_queue_enter here; instead, we use | |
857 | * percpu_ref_tryget directly, because we need to be able to | |
858 | * obtain a reference even in the short window between the queue | |
859 | * starting to freeze, by dropping the first reference in | |
1671d522 | 860 | * blk_freeze_queue_start, and the moment the last request is |
71f79fb3 GKB |
861 | * consumed, marked by the instant q_usage_counter reaches |
862 | * zero. | |
863 | */ | |
864 | if (!percpu_ref_tryget(&q->q_usage_counter)) | |
287922eb CH |
865 | return; |
866 | ||
12f5b931 | 867 | blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &next); |
320ae51f | 868 | |
12f5b931 KB |
869 | if (next != 0) { |
870 | mod_timer(&q->timeout, next); | |
0d2602ca | 871 | } else { |
fcd36c36 BVA |
872 | /* |
873 | * Request timeouts are handled as a forward rolling timer. If | |
874 | * we end up here it means that no requests are pending and | |
875 | * also that no request has been pending for a while. Mark | |
876 | * each hctx as idle. | |
877 | */ | |
f054b56c ML |
878 | queue_for_each_hw_ctx(q, hctx, i) { |
879 | /* the hctx may be unmapped, so check it here */ | |
880 | if (blk_mq_hw_queue_mapped(hctx)) | |
881 | blk_mq_tag_idle(hctx); | |
882 | } | |
0d2602ca | 883 | } |
287922eb | 884 | blk_queue_exit(q); |
320ae51f JA |
885 | } |
886 | ||
88459642 OS |
887 | struct flush_busy_ctx_data { |
888 | struct blk_mq_hw_ctx *hctx; | |
889 | struct list_head *list; | |
890 | }; | |
891 | ||
892 | static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data) | |
893 | { | |
894 | struct flush_busy_ctx_data *flush_data = data; | |
895 | struct blk_mq_hw_ctx *hctx = flush_data->hctx; | |
896 | struct blk_mq_ctx *ctx = hctx->ctxs[bitnr]; | |
897 | ||
88459642 OS |
898 | spin_lock(&ctx->lock); |
899 | list_splice_tail_init(&ctx->rq_list, flush_data->list); | |
e9a99a63 | 900 | sbitmap_clear_bit(sb, bitnr); |
88459642 OS |
901 | spin_unlock(&ctx->lock); |
902 | return true; | |
903 | } | |
904 | ||
1429d7c9 JA |
905 | /* |
906 | * Process software queues that have been marked busy, splicing them | |
907 | * to the for-dispatch | |
908 | */ | |
2c3ad667 | 909 | void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list) |
1429d7c9 | 910 | { |
88459642 OS |
911 | struct flush_busy_ctx_data data = { |
912 | .hctx = hctx, | |
913 | .list = list, | |
914 | }; | |
1429d7c9 | 915 | |
88459642 | 916 | sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data); |
1429d7c9 | 917 | } |
2c3ad667 | 918 | EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs); |
1429d7c9 | 919 | |
b347689f ML |
920 | struct dispatch_rq_data { |
921 | struct blk_mq_hw_ctx *hctx; | |
922 | struct request *rq; | |
923 | }; | |
924 | ||
925 | static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr, | |
926 | void *data) | |
927 | { | |
928 | struct dispatch_rq_data *dispatch_data = data; | |
929 | struct blk_mq_hw_ctx *hctx = dispatch_data->hctx; | |
930 | struct blk_mq_ctx *ctx = hctx->ctxs[bitnr]; | |
931 | ||
932 | spin_lock(&ctx->lock); | |
b4f6f38d | 933 | if (!list_empty(&ctx->rq_list)) { |
b347689f ML |
934 | dispatch_data->rq = list_entry_rq(ctx->rq_list.next); |
935 | list_del_init(&dispatch_data->rq->queuelist); | |
936 | if (list_empty(&ctx->rq_list)) | |
937 | sbitmap_clear_bit(sb, bitnr); | |
938 | } | |
939 | spin_unlock(&ctx->lock); | |
940 | ||
941 | return !dispatch_data->rq; | |
942 | } | |
943 | ||
944 | struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx, | |
945 | struct blk_mq_ctx *start) | |
946 | { | |
947 | unsigned off = start ? start->index_hw : 0; | |
948 | struct dispatch_rq_data data = { | |
949 | .hctx = hctx, | |
950 | .rq = NULL, | |
951 | }; | |
952 | ||
953 | __sbitmap_for_each_set(&hctx->ctx_map, off, | |
954 | dispatch_rq_from_ctx, &data); | |
955 | ||
956 | return data.rq; | |
957 | } | |
958 | ||
703fd1c0 JA |
959 | static inline unsigned int queued_to_index(unsigned int queued) |
960 | { | |
961 | if (!queued) | |
962 | return 0; | |
1429d7c9 | 963 | |
703fd1c0 | 964 | return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1); |
1429d7c9 JA |
965 | } |
966 | ||
8ab6bb9e | 967 | bool blk_mq_get_driver_tag(struct request *rq) |
bd166ef1 JA |
968 | { |
969 | struct blk_mq_alloc_data data = { | |
970 | .q = rq->q, | |
bd166ef1 | 971 | .hctx = blk_mq_map_queue(rq->q, rq->mq_ctx->cpu), |
8ab6bb9e | 972 | .flags = BLK_MQ_REQ_NOWAIT, |
bd166ef1 | 973 | }; |
d263ed99 | 974 | bool shared; |
5feeacdd | 975 | |
81380ca1 OS |
976 | if (rq->tag != -1) |
977 | goto done; | |
bd166ef1 | 978 | |
415b806d SG |
979 | if (blk_mq_tag_is_reserved(data.hctx->sched_tags, rq->internal_tag)) |
980 | data.flags |= BLK_MQ_REQ_RESERVED; | |
981 | ||
d263ed99 | 982 | shared = blk_mq_tag_busy(data.hctx); |
bd166ef1 JA |
983 | rq->tag = blk_mq_get_tag(&data); |
984 | if (rq->tag >= 0) { | |
d263ed99 | 985 | if (shared) { |
200e86b3 JA |
986 | rq->rq_flags |= RQF_MQ_INFLIGHT; |
987 | atomic_inc(&data.hctx->nr_active); | |
988 | } | |
bd166ef1 | 989 | data.hctx->tags->rqs[rq->tag] = rq; |
bd166ef1 JA |
990 | } |
991 | ||
81380ca1 | 992 | done: |
81380ca1 | 993 | return rq->tag != -1; |
bd166ef1 JA |
994 | } |
995 | ||
eb619fdb JA |
996 | static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode, |
997 | int flags, void *key) | |
da55f2cc OS |
998 | { |
999 | struct blk_mq_hw_ctx *hctx; | |
1000 | ||
1001 | hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait); | |
1002 | ||
5815839b | 1003 | spin_lock(&hctx->dispatch_wait_lock); |
eb619fdb | 1004 | list_del_init(&wait->entry); |
5815839b ML |
1005 | spin_unlock(&hctx->dispatch_wait_lock); |
1006 | ||
da55f2cc OS |
1007 | blk_mq_run_hw_queue(hctx, true); |
1008 | return 1; | |
1009 | } | |
1010 | ||
f906a6a0 JA |
1011 | /* |
1012 | * Mark us waiting for a tag. For shared tags, this involves hooking us into | |
ee3e4de5 BVA |
1013 | * the tag wakeups. For non-shared tags, we can simply mark us needing a |
1014 | * restart. For both cases, take care to check the condition again after | |
f906a6a0 JA |
1015 | * marking us as waiting. |
1016 | */ | |
2278d69f | 1017 | static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx *hctx, |
f906a6a0 | 1018 | struct request *rq) |
da55f2cc | 1019 | { |
5815839b | 1020 | struct wait_queue_head *wq; |
f906a6a0 JA |
1021 | wait_queue_entry_t *wait; |
1022 | bool ret; | |
da55f2cc | 1023 | |
2278d69f ML |
1024 | if (!(hctx->flags & BLK_MQ_F_TAG_SHARED)) { |
1025 | if (!test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state)) | |
1026 | set_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state); | |
f906a6a0 | 1027 | |
c27d53fb BVA |
1028 | /* |
1029 | * It's possible that a tag was freed in the window between the | |
1030 | * allocation failure and adding the hardware queue to the wait | |
1031 | * queue. | |
1032 | * | |
1033 | * Don't clear RESTART here, someone else could have set it. | |
1034 | * At most this will cost an extra queue run. | |
1035 | */ | |
8ab6bb9e | 1036 | return blk_mq_get_driver_tag(rq); |
eb619fdb | 1037 | } |
eb619fdb | 1038 | |
2278d69f | 1039 | wait = &hctx->dispatch_wait; |
c27d53fb BVA |
1040 | if (!list_empty_careful(&wait->entry)) |
1041 | return false; | |
1042 | ||
5815839b ML |
1043 | wq = &bt_wait_ptr(&hctx->tags->bitmap_tags, hctx)->wait; |
1044 | ||
1045 | spin_lock_irq(&wq->lock); | |
1046 | spin_lock(&hctx->dispatch_wait_lock); | |
c27d53fb | 1047 | if (!list_empty(&wait->entry)) { |
5815839b ML |
1048 | spin_unlock(&hctx->dispatch_wait_lock); |
1049 | spin_unlock_irq(&wq->lock); | |
c27d53fb | 1050 | return false; |
eb619fdb JA |
1051 | } |
1052 | ||
5815839b ML |
1053 | wait->flags &= ~WQ_FLAG_EXCLUSIVE; |
1054 | __add_wait_queue(wq, wait); | |
c27d53fb | 1055 | |
da55f2cc | 1056 | /* |
eb619fdb JA |
1057 | * It's possible that a tag was freed in the window between the |
1058 | * allocation failure and adding the hardware queue to the wait | |
1059 | * queue. | |
da55f2cc | 1060 | */ |
8ab6bb9e | 1061 | ret = blk_mq_get_driver_tag(rq); |
c27d53fb | 1062 | if (!ret) { |
5815839b ML |
1063 | spin_unlock(&hctx->dispatch_wait_lock); |
1064 | spin_unlock_irq(&wq->lock); | |
c27d53fb | 1065 | return false; |
eb619fdb | 1066 | } |
c27d53fb BVA |
1067 | |
1068 | /* | |
1069 | * We got a tag, remove ourselves from the wait queue to ensure | |
1070 | * someone else gets the wakeup. | |
1071 | */ | |
c27d53fb | 1072 | list_del_init(&wait->entry); |
5815839b ML |
1073 | spin_unlock(&hctx->dispatch_wait_lock); |
1074 | spin_unlock_irq(&wq->lock); | |
c27d53fb BVA |
1075 | |
1076 | return true; | |
da55f2cc OS |
1077 | } |
1078 | ||
6e768717 ML |
1079 | #define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT 8 |
1080 | #define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR 4 | |
1081 | /* | |
1082 | * Update dispatch busy with the Exponential Weighted Moving Average(EWMA): | |
1083 | * - EWMA is one simple way to compute running average value | |
1084 | * - weight(7/8 and 1/8) is applied so that it can decrease exponentially | |
1085 | * - take 4 as factor for avoiding to get too small(0) result, and this | |
1086 | * factor doesn't matter because EWMA decreases exponentially | |
1087 | */ | |
1088 | static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx *hctx, bool busy) | |
1089 | { | |
1090 | unsigned int ewma; | |
1091 | ||
1092 | if (hctx->queue->elevator) | |
1093 | return; | |
1094 | ||
1095 | ewma = hctx->dispatch_busy; | |
1096 | ||
1097 | if (!ewma && !busy) | |
1098 | return; | |
1099 | ||
1100 | ewma *= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT - 1; | |
1101 | if (busy) | |
1102 | ewma += 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR; | |
1103 | ewma /= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT; | |
1104 | ||
1105 | hctx->dispatch_busy = ewma; | |
1106 | } | |
1107 | ||
86ff7c2a ML |
1108 | #define BLK_MQ_RESOURCE_DELAY 3 /* ms units */ |
1109 | ||
1f57f8d4 JA |
1110 | /* |
1111 | * Returns true if we did some work AND can potentially do more. | |
1112 | */ | |
de148297 | 1113 | bool blk_mq_dispatch_rq_list(struct request_queue *q, struct list_head *list, |
eb619fdb | 1114 | bool got_budget) |
320ae51f | 1115 | { |
81380ca1 | 1116 | struct blk_mq_hw_ctx *hctx; |
6d6f167c | 1117 | struct request *rq, *nxt; |
eb619fdb | 1118 | bool no_tag = false; |
fc17b653 | 1119 | int errors, queued; |
86ff7c2a | 1120 | blk_status_t ret = BLK_STS_OK; |
320ae51f | 1121 | |
81380ca1 OS |
1122 | if (list_empty(list)) |
1123 | return false; | |
1124 | ||
de148297 ML |
1125 | WARN_ON(!list_is_singular(list) && got_budget); |
1126 | ||
320ae51f JA |
1127 | /* |
1128 | * Now process all the entries, sending them to the driver. | |
1129 | */ | |
93efe981 | 1130 | errors = queued = 0; |
81380ca1 | 1131 | do { |
74c45052 | 1132 | struct blk_mq_queue_data bd; |
320ae51f | 1133 | |
f04c3df3 | 1134 | rq = list_first_entry(list, struct request, queuelist); |
0bca799b ML |
1135 | |
1136 | hctx = blk_mq_map_queue(rq->q, rq->mq_ctx->cpu); | |
1137 | if (!got_budget && !blk_mq_get_dispatch_budget(hctx)) | |
1138 | break; | |
1139 | ||
8ab6bb9e | 1140 | if (!blk_mq_get_driver_tag(rq)) { |
3c782d67 | 1141 | /* |
da55f2cc | 1142 | * The initial allocation attempt failed, so we need to |
eb619fdb JA |
1143 | * rerun the hardware queue when a tag is freed. The |
1144 | * waitqueue takes care of that. If the queue is run | |
1145 | * before we add this entry back on the dispatch list, | |
1146 | * we'll re-run it below. | |
3c782d67 | 1147 | */ |
2278d69f | 1148 | if (!blk_mq_mark_tag_wait(hctx, rq)) { |
0bca799b | 1149 | blk_mq_put_dispatch_budget(hctx); |
f906a6a0 JA |
1150 | /* |
1151 | * For non-shared tags, the RESTART check | |
1152 | * will suffice. | |
1153 | */ | |
1154 | if (hctx->flags & BLK_MQ_F_TAG_SHARED) | |
1155 | no_tag = true; | |
de148297 ML |
1156 | break; |
1157 | } | |
1158 | } | |
1159 | ||
320ae51f | 1160 | list_del_init(&rq->queuelist); |
320ae51f | 1161 | |
74c45052 | 1162 | bd.rq = rq; |
113285b4 JA |
1163 | |
1164 | /* | |
1165 | * Flag last if we have no more requests, or if we have more | |
1166 | * but can't assign a driver tag to it. | |
1167 | */ | |
1168 | if (list_empty(list)) | |
1169 | bd.last = true; | |
1170 | else { | |
113285b4 | 1171 | nxt = list_first_entry(list, struct request, queuelist); |
8ab6bb9e | 1172 | bd.last = !blk_mq_get_driver_tag(nxt); |
113285b4 | 1173 | } |
74c45052 JA |
1174 | |
1175 | ret = q->mq_ops->queue_rq(hctx, &bd); | |
86ff7c2a | 1176 | if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE) { |
6d6f167c JW |
1177 | /* |
1178 | * If an I/O scheduler has been configured and we got a | |
ff821d27 JA |
1179 | * driver tag for the next request already, free it |
1180 | * again. | |
6d6f167c JW |
1181 | */ |
1182 | if (!list_empty(list)) { | |
1183 | nxt = list_first_entry(list, struct request, queuelist); | |
1184 | blk_mq_put_driver_tag(nxt); | |
1185 | } | |
f04c3df3 | 1186 | list_add(&rq->queuelist, list); |
ed0791b2 | 1187 | __blk_mq_requeue_request(rq); |
320ae51f | 1188 | break; |
fc17b653 CH |
1189 | } |
1190 | ||
1191 | if (unlikely(ret != BLK_STS_OK)) { | |
93efe981 | 1192 | errors++; |
2a842aca | 1193 | blk_mq_end_request(rq, BLK_STS_IOERR); |
fc17b653 | 1194 | continue; |
320ae51f JA |
1195 | } |
1196 | ||
fc17b653 | 1197 | queued++; |
81380ca1 | 1198 | } while (!list_empty(list)); |
320ae51f | 1199 | |
703fd1c0 | 1200 | hctx->dispatched[queued_to_index(queued)]++; |
320ae51f JA |
1201 | |
1202 | /* | |
1203 | * Any items that need requeuing? Stuff them into hctx->dispatch, | |
1204 | * that is where we will continue on next queue run. | |
1205 | */ | |
f04c3df3 | 1206 | if (!list_empty(list)) { |
86ff7c2a ML |
1207 | bool needs_restart; |
1208 | ||
320ae51f | 1209 | spin_lock(&hctx->lock); |
c13660a0 | 1210 | list_splice_init(list, &hctx->dispatch); |
320ae51f | 1211 | spin_unlock(&hctx->lock); |
f04c3df3 | 1212 | |
9ba52e58 | 1213 | /* |
710c785f BVA |
1214 | * If SCHED_RESTART was set by the caller of this function and |
1215 | * it is no longer set that means that it was cleared by another | |
1216 | * thread and hence that a queue rerun is needed. | |
9ba52e58 | 1217 | * |
eb619fdb JA |
1218 | * If 'no_tag' is set, that means that we failed getting |
1219 | * a driver tag with an I/O scheduler attached. If our dispatch | |
1220 | * waitqueue is no longer active, ensure that we run the queue | |
1221 | * AFTER adding our entries back to the list. | |
bd166ef1 | 1222 | * |
710c785f BVA |
1223 | * If no I/O scheduler has been configured it is possible that |
1224 | * the hardware queue got stopped and restarted before requests | |
1225 | * were pushed back onto the dispatch list. Rerun the queue to | |
1226 | * avoid starvation. Notes: | |
1227 | * - blk_mq_run_hw_queue() checks whether or not a queue has | |
1228 | * been stopped before rerunning a queue. | |
1229 | * - Some but not all block drivers stop a queue before | |
fc17b653 | 1230 | * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq |
710c785f | 1231 | * and dm-rq. |
86ff7c2a ML |
1232 | * |
1233 | * If driver returns BLK_STS_RESOURCE and SCHED_RESTART | |
1234 | * bit is set, run queue after a delay to avoid IO stalls | |
1235 | * that could otherwise occur if the queue is idle. | |
bd166ef1 | 1236 | */ |
86ff7c2a ML |
1237 | needs_restart = blk_mq_sched_needs_restart(hctx); |
1238 | if (!needs_restart || | |
eb619fdb | 1239 | (no_tag && list_empty_careful(&hctx->dispatch_wait.entry))) |
bd166ef1 | 1240 | blk_mq_run_hw_queue(hctx, true); |
86ff7c2a ML |
1241 | else if (needs_restart && (ret == BLK_STS_RESOURCE)) |
1242 | blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY); | |
1f57f8d4 | 1243 | |
6e768717 | 1244 | blk_mq_update_dispatch_busy(hctx, true); |
1f57f8d4 | 1245 | return false; |
6e768717 ML |
1246 | } else |
1247 | blk_mq_update_dispatch_busy(hctx, false); | |
f04c3df3 | 1248 | |
1f57f8d4 JA |
1249 | /* |
1250 | * If the host/device is unable to accept more work, inform the | |
1251 | * caller of that. | |
1252 | */ | |
1253 | if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE) | |
1254 | return false; | |
1255 | ||
93efe981 | 1256 | return (queued + errors) != 0; |
f04c3df3 JA |
1257 | } |
1258 | ||
6a83e74d BVA |
1259 | static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx) |
1260 | { | |
1261 | int srcu_idx; | |
1262 | ||
b7a71e66 JA |
1263 | /* |
1264 | * We should be running this queue from one of the CPUs that | |
1265 | * are mapped to it. | |
7df938fb ML |
1266 | * |
1267 | * There are at least two related races now between setting | |
1268 | * hctx->next_cpu from blk_mq_hctx_next_cpu() and running | |
1269 | * __blk_mq_run_hw_queue(): | |
1270 | * | |
1271 | * - hctx->next_cpu is found offline in blk_mq_hctx_next_cpu(), | |
1272 | * but later it becomes online, then this warning is harmless | |
1273 | * at all | |
1274 | * | |
1275 | * - hctx->next_cpu is found online in blk_mq_hctx_next_cpu(), | |
1276 | * but later it becomes offline, then the warning can't be | |
1277 | * triggered, and we depend on blk-mq timeout handler to | |
1278 | * handle dispatched requests to this hctx | |
b7a71e66 | 1279 | */ |
7df938fb ML |
1280 | if (!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) && |
1281 | cpu_online(hctx->next_cpu)) { | |
1282 | printk(KERN_WARNING "run queue from wrong CPU %d, hctx %s\n", | |
1283 | raw_smp_processor_id(), | |
1284 | cpumask_empty(hctx->cpumask) ? "inactive": "active"); | |
1285 | dump_stack(); | |
1286 | } | |
6a83e74d | 1287 | |
b7a71e66 JA |
1288 | /* |
1289 | * We can't run the queue inline with ints disabled. Ensure that | |
1290 | * we catch bad users of this early. | |
1291 | */ | |
1292 | WARN_ON_ONCE(in_interrupt()); | |
1293 | ||
04ced159 | 1294 | might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING); |
bf4907c0 | 1295 | |
04ced159 JA |
1296 | hctx_lock(hctx, &srcu_idx); |
1297 | blk_mq_sched_dispatch_requests(hctx); | |
1298 | hctx_unlock(hctx, srcu_idx); | |
6a83e74d BVA |
1299 | } |
1300 | ||
f82ddf19 ML |
1301 | static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx) |
1302 | { | |
1303 | int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask); | |
1304 | ||
1305 | if (cpu >= nr_cpu_ids) | |
1306 | cpu = cpumask_first(hctx->cpumask); | |
1307 | return cpu; | |
1308 | } | |
1309 | ||
506e931f JA |
1310 | /* |
1311 | * It'd be great if the workqueue API had a way to pass | |
1312 | * in a mask and had some smarts for more clever placement. | |
1313 | * For now we just round-robin here, switching for every | |
1314 | * BLK_MQ_CPU_WORK_BATCH queued items. | |
1315 | */ | |
1316 | static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx) | |
1317 | { | |
7bed4595 | 1318 | bool tried = false; |
476f8c98 | 1319 | int next_cpu = hctx->next_cpu; |
7bed4595 | 1320 | |
b657d7e6 CH |
1321 | if (hctx->queue->nr_hw_queues == 1) |
1322 | return WORK_CPU_UNBOUND; | |
506e931f JA |
1323 | |
1324 | if (--hctx->next_cpu_batch <= 0) { | |
7bed4595 | 1325 | select_cpu: |
476f8c98 | 1326 | next_cpu = cpumask_next_and(next_cpu, hctx->cpumask, |
20e4d813 | 1327 | cpu_online_mask); |
506e931f | 1328 | if (next_cpu >= nr_cpu_ids) |
f82ddf19 | 1329 | next_cpu = blk_mq_first_mapped_cpu(hctx); |
506e931f JA |
1330 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; |
1331 | } | |
1332 | ||
7bed4595 ML |
1333 | /* |
1334 | * Do unbound schedule if we can't find a online CPU for this hctx, | |
1335 | * and it should only happen in the path of handling CPU DEAD. | |
1336 | */ | |
476f8c98 | 1337 | if (!cpu_online(next_cpu)) { |
7bed4595 ML |
1338 | if (!tried) { |
1339 | tried = true; | |
1340 | goto select_cpu; | |
1341 | } | |
1342 | ||
1343 | /* | |
1344 | * Make sure to re-select CPU next time once after CPUs | |
1345 | * in hctx->cpumask become online again. | |
1346 | */ | |
476f8c98 | 1347 | hctx->next_cpu = next_cpu; |
7bed4595 ML |
1348 | hctx->next_cpu_batch = 1; |
1349 | return WORK_CPU_UNBOUND; | |
1350 | } | |
476f8c98 ML |
1351 | |
1352 | hctx->next_cpu = next_cpu; | |
1353 | return next_cpu; | |
506e931f JA |
1354 | } |
1355 | ||
7587a5ae BVA |
1356 | static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async, |
1357 | unsigned long msecs) | |
320ae51f | 1358 | { |
5435c023 | 1359 | if (unlikely(blk_mq_hctx_stopped(hctx))) |
320ae51f JA |
1360 | return; |
1361 | ||
1b792f2f | 1362 | if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) { |
2a90d4aa PB |
1363 | int cpu = get_cpu(); |
1364 | if (cpumask_test_cpu(cpu, hctx->cpumask)) { | |
398205b8 | 1365 | __blk_mq_run_hw_queue(hctx); |
2a90d4aa | 1366 | put_cpu(); |
398205b8 PB |
1367 | return; |
1368 | } | |
e4043dcf | 1369 | |
2a90d4aa | 1370 | put_cpu(); |
e4043dcf | 1371 | } |
398205b8 | 1372 | |
ae943d20 BVA |
1373 | kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work, |
1374 | msecs_to_jiffies(msecs)); | |
7587a5ae BVA |
1375 | } |
1376 | ||
1377 | void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs) | |
1378 | { | |
1379 | __blk_mq_delay_run_hw_queue(hctx, true, msecs); | |
1380 | } | |
1381 | EXPORT_SYMBOL(blk_mq_delay_run_hw_queue); | |
1382 | ||
79f720a7 | 1383 | bool blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) |
7587a5ae | 1384 | { |
24f5a90f ML |
1385 | int srcu_idx; |
1386 | bool need_run; | |
1387 | ||
1388 | /* | |
1389 | * When queue is quiesced, we may be switching io scheduler, or | |
1390 | * updating nr_hw_queues, or other things, and we can't run queue | |
1391 | * any more, even __blk_mq_hctx_has_pending() can't be called safely. | |
1392 | * | |
1393 | * And queue will be rerun in blk_mq_unquiesce_queue() if it is | |
1394 | * quiesced. | |
1395 | */ | |
04ced159 JA |
1396 | hctx_lock(hctx, &srcu_idx); |
1397 | need_run = !blk_queue_quiesced(hctx->queue) && | |
1398 | blk_mq_hctx_has_pending(hctx); | |
1399 | hctx_unlock(hctx, srcu_idx); | |
24f5a90f ML |
1400 | |
1401 | if (need_run) { | |
79f720a7 JA |
1402 | __blk_mq_delay_run_hw_queue(hctx, async, 0); |
1403 | return true; | |
1404 | } | |
1405 | ||
1406 | return false; | |
320ae51f | 1407 | } |
5b727272 | 1408 | EXPORT_SYMBOL(blk_mq_run_hw_queue); |
320ae51f | 1409 | |
b94ec296 | 1410 | void blk_mq_run_hw_queues(struct request_queue *q, bool async) |
320ae51f JA |
1411 | { |
1412 | struct blk_mq_hw_ctx *hctx; | |
1413 | int i; | |
1414 | ||
1415 | queue_for_each_hw_ctx(q, hctx, i) { | |
79f720a7 | 1416 | if (blk_mq_hctx_stopped(hctx)) |
320ae51f JA |
1417 | continue; |
1418 | ||
b94ec296 | 1419 | blk_mq_run_hw_queue(hctx, async); |
320ae51f JA |
1420 | } |
1421 | } | |
b94ec296 | 1422 | EXPORT_SYMBOL(blk_mq_run_hw_queues); |
320ae51f | 1423 | |
fd001443 BVA |
1424 | /** |
1425 | * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped | |
1426 | * @q: request queue. | |
1427 | * | |
1428 | * The caller is responsible for serializing this function against | |
1429 | * blk_mq_{start,stop}_hw_queue(). | |
1430 | */ | |
1431 | bool blk_mq_queue_stopped(struct request_queue *q) | |
1432 | { | |
1433 | struct blk_mq_hw_ctx *hctx; | |
1434 | int i; | |
1435 | ||
1436 | queue_for_each_hw_ctx(q, hctx, i) | |
1437 | if (blk_mq_hctx_stopped(hctx)) | |
1438 | return true; | |
1439 | ||
1440 | return false; | |
1441 | } | |
1442 | EXPORT_SYMBOL(blk_mq_queue_stopped); | |
1443 | ||
39a70c76 ML |
1444 | /* |
1445 | * This function is often used for pausing .queue_rq() by driver when | |
1446 | * there isn't enough resource or some conditions aren't satisfied, and | |
4d606219 | 1447 | * BLK_STS_RESOURCE is usually returned. |
39a70c76 ML |
1448 | * |
1449 | * We do not guarantee that dispatch can be drained or blocked | |
1450 | * after blk_mq_stop_hw_queue() returns. Please use | |
1451 | * blk_mq_quiesce_queue() for that requirement. | |
1452 | */ | |
2719aa21 JA |
1453 | void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx) |
1454 | { | |
641a9ed6 | 1455 | cancel_delayed_work(&hctx->run_work); |
280d45f6 | 1456 | |
641a9ed6 | 1457 | set_bit(BLK_MQ_S_STOPPED, &hctx->state); |
2719aa21 | 1458 | } |
641a9ed6 | 1459 | EXPORT_SYMBOL(blk_mq_stop_hw_queue); |
2719aa21 | 1460 | |
39a70c76 ML |
1461 | /* |
1462 | * This function is often used for pausing .queue_rq() by driver when | |
1463 | * there isn't enough resource or some conditions aren't satisfied, and | |
4d606219 | 1464 | * BLK_STS_RESOURCE is usually returned. |
39a70c76 ML |
1465 | * |
1466 | * We do not guarantee that dispatch can be drained or blocked | |
1467 | * after blk_mq_stop_hw_queues() returns. Please use | |
1468 | * blk_mq_quiesce_queue() for that requirement. | |
1469 | */ | |
2719aa21 JA |
1470 | void blk_mq_stop_hw_queues(struct request_queue *q) |
1471 | { | |
641a9ed6 ML |
1472 | struct blk_mq_hw_ctx *hctx; |
1473 | int i; | |
1474 | ||
1475 | queue_for_each_hw_ctx(q, hctx, i) | |
1476 | blk_mq_stop_hw_queue(hctx); | |
280d45f6 CH |
1477 | } |
1478 | EXPORT_SYMBOL(blk_mq_stop_hw_queues); | |
1479 | ||
320ae51f JA |
1480 | void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx) |
1481 | { | |
1482 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
e4043dcf | 1483 | |
0ffbce80 | 1484 | blk_mq_run_hw_queue(hctx, false); |
320ae51f JA |
1485 | } |
1486 | EXPORT_SYMBOL(blk_mq_start_hw_queue); | |
1487 | ||
2f268556 CH |
1488 | void blk_mq_start_hw_queues(struct request_queue *q) |
1489 | { | |
1490 | struct blk_mq_hw_ctx *hctx; | |
1491 | int i; | |
1492 | ||
1493 | queue_for_each_hw_ctx(q, hctx, i) | |
1494 | blk_mq_start_hw_queue(hctx); | |
1495 | } | |
1496 | EXPORT_SYMBOL(blk_mq_start_hw_queues); | |
1497 | ||
ae911c5e JA |
1498 | void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) |
1499 | { | |
1500 | if (!blk_mq_hctx_stopped(hctx)) | |
1501 | return; | |
1502 | ||
1503 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
1504 | blk_mq_run_hw_queue(hctx, async); | |
1505 | } | |
1506 | EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue); | |
1507 | ||
1b4a3258 | 1508 | void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async) |
320ae51f JA |
1509 | { |
1510 | struct blk_mq_hw_ctx *hctx; | |
1511 | int i; | |
1512 | ||
ae911c5e JA |
1513 | queue_for_each_hw_ctx(q, hctx, i) |
1514 | blk_mq_start_stopped_hw_queue(hctx, async); | |
320ae51f JA |
1515 | } |
1516 | EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues); | |
1517 | ||
70f4db63 | 1518 | static void blk_mq_run_work_fn(struct work_struct *work) |
320ae51f JA |
1519 | { |
1520 | struct blk_mq_hw_ctx *hctx; | |
1521 | ||
9f993737 | 1522 | hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work); |
320ae51f | 1523 | |
21c6e939 | 1524 | /* |
15fe8a90 | 1525 | * If we are stopped, don't run the queue. |
21c6e939 | 1526 | */ |
15fe8a90 | 1527 | if (test_bit(BLK_MQ_S_STOPPED, &hctx->state)) |
0196d6b4 | 1528 | return; |
7587a5ae BVA |
1529 | |
1530 | __blk_mq_run_hw_queue(hctx); | |
1531 | } | |
1532 | ||
cfd0c552 | 1533 | static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx, |
cfd0c552 ML |
1534 | struct request *rq, |
1535 | bool at_head) | |
320ae51f | 1536 | { |
e57690fe JA |
1537 | struct blk_mq_ctx *ctx = rq->mq_ctx; |
1538 | ||
7b607814 BVA |
1539 | lockdep_assert_held(&ctx->lock); |
1540 | ||
01b983c9 JA |
1541 | trace_block_rq_insert(hctx->queue, rq); |
1542 | ||
72a0a36e CH |
1543 | if (at_head) |
1544 | list_add(&rq->queuelist, &ctx->rq_list); | |
1545 | else | |
1546 | list_add_tail(&rq->queuelist, &ctx->rq_list); | |
cfd0c552 | 1547 | } |
4bb659b1 | 1548 | |
2c3ad667 JA |
1549 | void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq, |
1550 | bool at_head) | |
cfd0c552 ML |
1551 | { |
1552 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
1553 | ||
7b607814 BVA |
1554 | lockdep_assert_held(&ctx->lock); |
1555 | ||
e57690fe | 1556 | __blk_mq_insert_req_list(hctx, rq, at_head); |
320ae51f | 1557 | blk_mq_hctx_mark_pending(hctx, ctx); |
320ae51f JA |
1558 | } |
1559 | ||
157f377b JA |
1560 | /* |
1561 | * Should only be used carefully, when the caller knows we want to | |
1562 | * bypass a potential IO scheduler on the target device. | |
1563 | */ | |
b0850297 | 1564 | void blk_mq_request_bypass_insert(struct request *rq, bool run_queue) |
157f377b JA |
1565 | { |
1566 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
1567 | struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(rq->q, ctx->cpu); | |
1568 | ||
1569 | spin_lock(&hctx->lock); | |
1570 | list_add_tail(&rq->queuelist, &hctx->dispatch); | |
1571 | spin_unlock(&hctx->lock); | |
1572 | ||
b0850297 ML |
1573 | if (run_queue) |
1574 | blk_mq_run_hw_queue(hctx, false); | |
157f377b JA |
1575 | } |
1576 | ||
bd166ef1 JA |
1577 | void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx, |
1578 | struct list_head *list) | |
320ae51f JA |
1579 | |
1580 | { | |
3f0cedc7 ML |
1581 | struct request *rq; |
1582 | ||
320ae51f JA |
1583 | /* |
1584 | * preemption doesn't flush plug list, so it's possible ctx->cpu is | |
1585 | * offline now | |
1586 | */ | |
3f0cedc7 | 1587 | list_for_each_entry(rq, list, queuelist) { |
e57690fe | 1588 | BUG_ON(rq->mq_ctx != ctx); |
3f0cedc7 | 1589 | trace_block_rq_insert(hctx->queue, rq); |
320ae51f | 1590 | } |
3f0cedc7 ML |
1591 | |
1592 | spin_lock(&ctx->lock); | |
1593 | list_splice_tail_init(list, &ctx->rq_list); | |
cfd0c552 | 1594 | blk_mq_hctx_mark_pending(hctx, ctx); |
320ae51f | 1595 | spin_unlock(&ctx->lock); |
320ae51f JA |
1596 | } |
1597 | ||
1598 | static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b) | |
1599 | { | |
1600 | struct request *rqa = container_of(a, struct request, queuelist); | |
1601 | struct request *rqb = container_of(b, struct request, queuelist); | |
1602 | ||
1603 | return !(rqa->mq_ctx < rqb->mq_ctx || | |
1604 | (rqa->mq_ctx == rqb->mq_ctx && | |
1605 | blk_rq_pos(rqa) < blk_rq_pos(rqb))); | |
1606 | } | |
1607 | ||
1608 | void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule) | |
1609 | { | |
1610 | struct blk_mq_ctx *this_ctx; | |
1611 | struct request_queue *this_q; | |
1612 | struct request *rq; | |
1613 | LIST_HEAD(list); | |
1614 | LIST_HEAD(ctx_list); | |
1615 | unsigned int depth; | |
1616 | ||
1617 | list_splice_init(&plug->mq_list, &list); | |
1618 | ||
1619 | list_sort(NULL, &list, plug_ctx_cmp); | |
1620 | ||
1621 | this_q = NULL; | |
1622 | this_ctx = NULL; | |
1623 | depth = 0; | |
1624 | ||
1625 | while (!list_empty(&list)) { | |
1626 | rq = list_entry_rq(list.next); | |
1627 | list_del_init(&rq->queuelist); | |
1628 | BUG_ON(!rq->q); | |
1629 | if (rq->mq_ctx != this_ctx) { | |
1630 | if (this_ctx) { | |
bd166ef1 JA |
1631 | trace_block_unplug(this_q, depth, from_schedule); |
1632 | blk_mq_sched_insert_requests(this_q, this_ctx, | |
1633 | &ctx_list, | |
1634 | from_schedule); | |
320ae51f JA |
1635 | } |
1636 | ||
1637 | this_ctx = rq->mq_ctx; | |
1638 | this_q = rq->q; | |
1639 | depth = 0; | |
1640 | } | |
1641 | ||
1642 | depth++; | |
1643 | list_add_tail(&rq->queuelist, &ctx_list); | |
1644 | } | |
1645 | ||
1646 | /* | |
1647 | * If 'this_ctx' is set, we know we have entries to complete | |
1648 | * on 'ctx_list'. Do those. | |
1649 | */ | |
1650 | if (this_ctx) { | |
bd166ef1 JA |
1651 | trace_block_unplug(this_q, depth, from_schedule); |
1652 | blk_mq_sched_insert_requests(this_q, this_ctx, &ctx_list, | |
1653 | from_schedule); | |
320ae51f JA |
1654 | } |
1655 | } | |
1656 | ||
1657 | static void blk_mq_bio_to_request(struct request *rq, struct bio *bio) | |
1658 | { | |
da8d7f07 | 1659 | blk_init_request_from_bio(rq, bio); |
4b570521 | 1660 | |
85acb3ba SL |
1661 | blk_rq_set_rl(rq, blk_get_rl(rq->q, bio)); |
1662 | ||
6e85eaf3 | 1663 | blk_account_io_start(rq, true); |
320ae51f JA |
1664 | } |
1665 | ||
fd2d3326 JA |
1666 | static blk_qc_t request_to_qc_t(struct blk_mq_hw_ctx *hctx, struct request *rq) |
1667 | { | |
bd166ef1 JA |
1668 | if (rq->tag != -1) |
1669 | return blk_tag_to_qc_t(rq->tag, hctx->queue_num, false); | |
1670 | ||
1671 | return blk_tag_to_qc_t(rq->internal_tag, hctx->queue_num, true); | |
fd2d3326 JA |
1672 | } |
1673 | ||
0f95549c MS |
1674 | static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx, |
1675 | struct request *rq, | |
1676 | blk_qc_t *cookie) | |
f984df1f | 1677 | { |
f984df1f | 1678 | struct request_queue *q = rq->q; |
f984df1f SL |
1679 | struct blk_mq_queue_data bd = { |
1680 | .rq = rq, | |
d945a365 | 1681 | .last = true, |
f984df1f | 1682 | }; |
bd166ef1 | 1683 | blk_qc_t new_cookie; |
f06345ad | 1684 | blk_status_t ret; |
0f95549c MS |
1685 | |
1686 | new_cookie = request_to_qc_t(hctx, rq); | |
1687 | ||
1688 | /* | |
1689 | * For OK queue, we are done. For error, caller may kill it. | |
1690 | * Any other error (busy), just add it to our list as we | |
1691 | * previously would have done. | |
1692 | */ | |
1693 | ret = q->mq_ops->queue_rq(hctx, &bd); | |
1694 | switch (ret) { | |
1695 | case BLK_STS_OK: | |
6ce3dd6e | 1696 | blk_mq_update_dispatch_busy(hctx, false); |
0f95549c MS |
1697 | *cookie = new_cookie; |
1698 | break; | |
1699 | case BLK_STS_RESOURCE: | |
86ff7c2a | 1700 | case BLK_STS_DEV_RESOURCE: |
6ce3dd6e | 1701 | blk_mq_update_dispatch_busy(hctx, true); |
0f95549c MS |
1702 | __blk_mq_requeue_request(rq); |
1703 | break; | |
1704 | default: | |
6ce3dd6e | 1705 | blk_mq_update_dispatch_busy(hctx, false); |
0f95549c MS |
1706 | *cookie = BLK_QC_T_NONE; |
1707 | break; | |
1708 | } | |
1709 | ||
1710 | return ret; | |
1711 | } | |
1712 | ||
0f95549c MS |
1713 | static blk_status_t __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx, |
1714 | struct request *rq, | |
396eaf21 ML |
1715 | blk_qc_t *cookie, |
1716 | bool bypass_insert) | |
0f95549c MS |
1717 | { |
1718 | struct request_queue *q = rq->q; | |
d964f04a ML |
1719 | bool run_queue = true; |
1720 | ||
23d4ee19 ML |
1721 | /* |
1722 | * RCU or SRCU read lock is needed before checking quiesced flag. | |
1723 | * | |
1724 | * When queue is stopped or quiesced, ignore 'bypass_insert' from | |
c77ff7fd | 1725 | * blk_mq_request_issue_directly(), and return BLK_STS_OK to caller, |
23d4ee19 ML |
1726 | * and avoid driver to try to dispatch again. |
1727 | */ | |
f4560ffe | 1728 | if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)) { |
d964f04a | 1729 | run_queue = false; |
23d4ee19 | 1730 | bypass_insert = false; |
d964f04a ML |
1731 | goto insert; |
1732 | } | |
f984df1f | 1733 | |
396eaf21 | 1734 | if (q->elevator && !bypass_insert) |
2253efc8 BVA |
1735 | goto insert; |
1736 | ||
0bca799b | 1737 | if (!blk_mq_get_dispatch_budget(hctx)) |
bd166ef1 JA |
1738 | goto insert; |
1739 | ||
8ab6bb9e | 1740 | if (!blk_mq_get_driver_tag(rq)) { |
0bca799b | 1741 | blk_mq_put_dispatch_budget(hctx); |
de148297 | 1742 | goto insert; |
88022d72 | 1743 | } |
de148297 | 1744 | |
0f95549c | 1745 | return __blk_mq_issue_directly(hctx, rq, cookie); |
2253efc8 | 1746 | insert: |
396eaf21 ML |
1747 | if (bypass_insert) |
1748 | return BLK_STS_RESOURCE; | |
0f95549c | 1749 | |
23d4ee19 | 1750 | blk_mq_sched_insert_request(rq, false, run_queue, false); |
0f95549c | 1751 | return BLK_STS_OK; |
f984df1f SL |
1752 | } |
1753 | ||
5eb6126e CH |
1754 | static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx, |
1755 | struct request *rq, blk_qc_t *cookie) | |
1756 | { | |
0f95549c | 1757 | blk_status_t ret; |
04ced159 | 1758 | int srcu_idx; |
bf4907c0 | 1759 | |
04ced159 | 1760 | might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING); |
bf4907c0 | 1761 | |
04ced159 | 1762 | hctx_lock(hctx, &srcu_idx); |
0f95549c | 1763 | |
396eaf21 | 1764 | ret = __blk_mq_try_issue_directly(hctx, rq, cookie, false); |
86ff7c2a | 1765 | if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE) |
23d4ee19 | 1766 | blk_mq_sched_insert_request(rq, false, true, false); |
0f95549c MS |
1767 | else if (ret != BLK_STS_OK) |
1768 | blk_mq_end_request(rq, ret); | |
1769 | ||
04ced159 | 1770 | hctx_unlock(hctx, srcu_idx); |
5eb6126e CH |
1771 | } |
1772 | ||
c77ff7fd | 1773 | blk_status_t blk_mq_request_issue_directly(struct request *rq) |
396eaf21 ML |
1774 | { |
1775 | blk_status_t ret; | |
1776 | int srcu_idx; | |
1777 | blk_qc_t unused_cookie; | |
1778 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
1779 | struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(rq->q, ctx->cpu); | |
1780 | ||
1781 | hctx_lock(hctx, &srcu_idx); | |
1782 | ret = __blk_mq_try_issue_directly(hctx, rq, &unused_cookie, true); | |
1783 | hctx_unlock(hctx, srcu_idx); | |
1784 | ||
1785 | return ret; | |
5eb6126e CH |
1786 | } |
1787 | ||
6ce3dd6e ML |
1788 | void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx, |
1789 | struct list_head *list) | |
1790 | { | |
1791 | while (!list_empty(list)) { | |
1792 | blk_status_t ret; | |
1793 | struct request *rq = list_first_entry(list, struct request, | |
1794 | queuelist); | |
1795 | ||
1796 | list_del_init(&rq->queuelist); | |
1797 | ret = blk_mq_request_issue_directly(rq); | |
1798 | if (ret != BLK_STS_OK) { | |
8824f622 ML |
1799 | if (ret == BLK_STS_RESOURCE || |
1800 | ret == BLK_STS_DEV_RESOURCE) { | |
1801 | list_add(&rq->queuelist, list); | |
1802 | break; | |
1803 | } | |
1804 | blk_mq_end_request(rq, ret); | |
6ce3dd6e ML |
1805 | } |
1806 | } | |
1807 | } | |
1808 | ||
dece1635 | 1809 | static blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio) |
07068d5b | 1810 | { |
ef295ecf | 1811 | const int is_sync = op_is_sync(bio->bi_opf); |
f73f44eb | 1812 | const int is_flush_fua = op_is_flush(bio->bi_opf); |
5a797e00 | 1813 | struct blk_mq_alloc_data data = { .flags = 0 }; |
07068d5b | 1814 | struct request *rq; |
5eb6126e | 1815 | unsigned int request_count = 0; |
f984df1f | 1816 | struct blk_plug *plug; |
5b3f341f | 1817 | struct request *same_queue_rq = NULL; |
7b371636 | 1818 | blk_qc_t cookie; |
07068d5b JA |
1819 | |
1820 | blk_queue_bounce(q, &bio); | |
1821 | ||
af67c31f | 1822 | blk_queue_split(q, &bio); |
f36ea50c | 1823 | |
e23947bd | 1824 | if (!bio_integrity_prep(bio)) |
dece1635 | 1825 | return BLK_QC_T_NONE; |
07068d5b | 1826 | |
87c279e6 OS |
1827 | if (!is_flush_fua && !blk_queue_nomerges(q) && |
1828 | blk_attempt_plug_merge(q, bio, &request_count, &same_queue_rq)) | |
1829 | return BLK_QC_T_NONE; | |
f984df1f | 1830 | |
bd166ef1 JA |
1831 | if (blk_mq_sched_bio_merge(q, bio)) |
1832 | return BLK_QC_T_NONE; | |
1833 | ||
c1c80384 | 1834 | rq_qos_throttle(q, bio, NULL); |
87760e5e | 1835 | |
bd166ef1 JA |
1836 | trace_block_getrq(q, bio, bio->bi_opf); |
1837 | ||
d2c0d383 | 1838 | rq = blk_mq_get_request(q, bio, bio->bi_opf, &data); |
87760e5e | 1839 | if (unlikely(!rq)) { |
c1c80384 | 1840 | rq_qos_cleanup(q, bio); |
03a07c92 GR |
1841 | if (bio->bi_opf & REQ_NOWAIT) |
1842 | bio_wouldblock_error(bio); | |
dece1635 | 1843 | return BLK_QC_T_NONE; |
87760e5e JA |
1844 | } |
1845 | ||
c1c80384 | 1846 | rq_qos_track(q, rq, bio); |
07068d5b | 1847 | |
fd2d3326 | 1848 | cookie = request_to_qc_t(data.hctx, rq); |
07068d5b | 1849 | |
f984df1f | 1850 | plug = current->plug; |
07068d5b | 1851 | if (unlikely(is_flush_fua)) { |
f984df1f | 1852 | blk_mq_put_ctx(data.ctx); |
07068d5b | 1853 | blk_mq_bio_to_request(rq, bio); |
923218f6 ML |
1854 | |
1855 | /* bypass scheduler for flush rq */ | |
1856 | blk_insert_flush(rq); | |
1857 | blk_mq_run_hw_queue(data.hctx, true); | |
a4d907b6 | 1858 | } else if (plug && q->nr_hw_queues == 1) { |
600271d9 SL |
1859 | struct request *last = NULL; |
1860 | ||
b00c53e8 | 1861 | blk_mq_put_ctx(data.ctx); |
e6c4438b | 1862 | blk_mq_bio_to_request(rq, bio); |
0a6219a9 ML |
1863 | |
1864 | /* | |
1865 | * @request_count may become stale because of schedule | |
1866 | * out, so check the list again. | |
1867 | */ | |
1868 | if (list_empty(&plug->mq_list)) | |
1869 | request_count = 0; | |
254d259d CH |
1870 | else if (blk_queue_nomerges(q)) |
1871 | request_count = blk_plug_queued_count(q); | |
1872 | ||
676d0607 | 1873 | if (!request_count) |
e6c4438b | 1874 | trace_block_plug(q); |
600271d9 SL |
1875 | else |
1876 | last = list_entry_rq(plug->mq_list.prev); | |
b094f89c | 1877 | |
600271d9 SL |
1878 | if (request_count >= BLK_MAX_REQUEST_COUNT || (last && |
1879 | blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) { | |
e6c4438b JM |
1880 | blk_flush_plug_list(plug, false); |
1881 | trace_block_plug(q); | |
320ae51f | 1882 | } |
b094f89c | 1883 | |
e6c4438b | 1884 | list_add_tail(&rq->queuelist, &plug->mq_list); |
2299722c | 1885 | } else if (plug && !blk_queue_nomerges(q)) { |
bd166ef1 | 1886 | blk_mq_bio_to_request(rq, bio); |
07068d5b | 1887 | |
07068d5b | 1888 | /* |
6a83e74d | 1889 | * We do limited plugging. If the bio can be merged, do that. |
f984df1f SL |
1890 | * Otherwise the existing request in the plug list will be |
1891 | * issued. So the plug list will have one request at most | |
2299722c CH |
1892 | * The plug list might get flushed before this. If that happens, |
1893 | * the plug list is empty, and same_queue_rq is invalid. | |
07068d5b | 1894 | */ |
2299722c CH |
1895 | if (list_empty(&plug->mq_list)) |
1896 | same_queue_rq = NULL; | |
1897 | if (same_queue_rq) | |
1898 | list_del_init(&same_queue_rq->queuelist); | |
1899 | list_add_tail(&rq->queuelist, &plug->mq_list); | |
1900 | ||
bf4907c0 JA |
1901 | blk_mq_put_ctx(data.ctx); |
1902 | ||
dad7a3be ML |
1903 | if (same_queue_rq) { |
1904 | data.hctx = blk_mq_map_queue(q, | |
1905 | same_queue_rq->mq_ctx->cpu); | |
2299722c CH |
1906 | blk_mq_try_issue_directly(data.hctx, same_queue_rq, |
1907 | &cookie); | |
dad7a3be | 1908 | } |
6ce3dd6e ML |
1909 | } else if ((q->nr_hw_queues > 1 && is_sync) || (!q->elevator && |
1910 | !data.hctx->dispatch_busy)) { | |
bf4907c0 | 1911 | blk_mq_put_ctx(data.ctx); |
2299722c | 1912 | blk_mq_bio_to_request(rq, bio); |
2299722c | 1913 | blk_mq_try_issue_directly(data.hctx, rq, &cookie); |
ab42f35d | 1914 | } else { |
b00c53e8 | 1915 | blk_mq_put_ctx(data.ctx); |
ab42f35d | 1916 | blk_mq_bio_to_request(rq, bio); |
8fa9f556 | 1917 | blk_mq_sched_insert_request(rq, false, true, true); |
ab42f35d | 1918 | } |
320ae51f | 1919 | |
7b371636 | 1920 | return cookie; |
320ae51f JA |
1921 | } |
1922 | ||
cc71a6f4 JA |
1923 | void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags, |
1924 | unsigned int hctx_idx) | |
95363efd | 1925 | { |
e9b267d9 | 1926 | struct page *page; |
320ae51f | 1927 | |
24d2f903 | 1928 | if (tags->rqs && set->ops->exit_request) { |
e9b267d9 | 1929 | int i; |
320ae51f | 1930 | |
24d2f903 | 1931 | for (i = 0; i < tags->nr_tags; i++) { |
2af8cbe3 JA |
1932 | struct request *rq = tags->static_rqs[i]; |
1933 | ||
1934 | if (!rq) | |
e9b267d9 | 1935 | continue; |
d6296d39 | 1936 | set->ops->exit_request(set, rq, hctx_idx); |
2af8cbe3 | 1937 | tags->static_rqs[i] = NULL; |
e9b267d9 | 1938 | } |
320ae51f | 1939 | } |
320ae51f | 1940 | |
24d2f903 CH |
1941 | while (!list_empty(&tags->page_list)) { |
1942 | page = list_first_entry(&tags->page_list, struct page, lru); | |
6753471c | 1943 | list_del_init(&page->lru); |
f75782e4 CM |
1944 | /* |
1945 | * Remove kmemleak object previously allocated in | |
1946 | * blk_mq_init_rq_map(). | |
1947 | */ | |
1948 | kmemleak_free(page_address(page)); | |
320ae51f JA |
1949 | __free_pages(page, page->private); |
1950 | } | |
cc71a6f4 | 1951 | } |
320ae51f | 1952 | |
cc71a6f4 JA |
1953 | void blk_mq_free_rq_map(struct blk_mq_tags *tags) |
1954 | { | |
24d2f903 | 1955 | kfree(tags->rqs); |
cc71a6f4 | 1956 | tags->rqs = NULL; |
2af8cbe3 JA |
1957 | kfree(tags->static_rqs); |
1958 | tags->static_rqs = NULL; | |
320ae51f | 1959 | |
24d2f903 | 1960 | blk_mq_free_tags(tags); |
320ae51f JA |
1961 | } |
1962 | ||
cc71a6f4 JA |
1963 | struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, |
1964 | unsigned int hctx_idx, | |
1965 | unsigned int nr_tags, | |
1966 | unsigned int reserved_tags) | |
320ae51f | 1967 | { |
24d2f903 | 1968 | struct blk_mq_tags *tags; |
59f082e4 | 1969 | int node; |
320ae51f | 1970 | |
59f082e4 SL |
1971 | node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx); |
1972 | if (node == NUMA_NO_NODE) | |
1973 | node = set->numa_node; | |
1974 | ||
1975 | tags = blk_mq_init_tags(nr_tags, reserved_tags, node, | |
24391c0d | 1976 | BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags)); |
24d2f903 CH |
1977 | if (!tags) |
1978 | return NULL; | |
320ae51f | 1979 | |
590b5b7d | 1980 | tags->rqs = kcalloc_node(nr_tags, sizeof(struct request *), |
36e1f3d1 | 1981 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, |
59f082e4 | 1982 | node); |
24d2f903 CH |
1983 | if (!tags->rqs) { |
1984 | blk_mq_free_tags(tags); | |
1985 | return NULL; | |
1986 | } | |
320ae51f | 1987 | |
590b5b7d KC |
1988 | tags->static_rqs = kcalloc_node(nr_tags, sizeof(struct request *), |
1989 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, | |
1990 | node); | |
2af8cbe3 JA |
1991 | if (!tags->static_rqs) { |
1992 | kfree(tags->rqs); | |
1993 | blk_mq_free_tags(tags); | |
1994 | return NULL; | |
1995 | } | |
1996 | ||
cc71a6f4 JA |
1997 | return tags; |
1998 | } | |
1999 | ||
2000 | static size_t order_to_size(unsigned int order) | |
2001 | { | |
2002 | return (size_t)PAGE_SIZE << order; | |
2003 | } | |
2004 | ||
1d9bd516 TH |
2005 | static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq, |
2006 | unsigned int hctx_idx, int node) | |
2007 | { | |
2008 | int ret; | |
2009 | ||
2010 | if (set->ops->init_request) { | |
2011 | ret = set->ops->init_request(set, rq, hctx_idx, node); | |
2012 | if (ret) | |
2013 | return ret; | |
2014 | } | |
2015 | ||
12f5b931 | 2016 | WRITE_ONCE(rq->state, MQ_RQ_IDLE); |
1d9bd516 TH |
2017 | return 0; |
2018 | } | |
2019 | ||
cc71a6f4 JA |
2020 | int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags, |
2021 | unsigned int hctx_idx, unsigned int depth) | |
2022 | { | |
2023 | unsigned int i, j, entries_per_page, max_order = 4; | |
2024 | size_t rq_size, left; | |
59f082e4 SL |
2025 | int node; |
2026 | ||
2027 | node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx); | |
2028 | if (node == NUMA_NO_NODE) | |
2029 | node = set->numa_node; | |
cc71a6f4 JA |
2030 | |
2031 | INIT_LIST_HEAD(&tags->page_list); | |
2032 | ||
320ae51f JA |
2033 | /* |
2034 | * rq_size is the size of the request plus driver payload, rounded | |
2035 | * to the cacheline size | |
2036 | */ | |
24d2f903 | 2037 | rq_size = round_up(sizeof(struct request) + set->cmd_size, |
320ae51f | 2038 | cache_line_size()); |
cc71a6f4 | 2039 | left = rq_size * depth; |
320ae51f | 2040 | |
cc71a6f4 | 2041 | for (i = 0; i < depth; ) { |
320ae51f JA |
2042 | int this_order = max_order; |
2043 | struct page *page; | |
2044 | int to_do; | |
2045 | void *p; | |
2046 | ||
b3a834b1 | 2047 | while (this_order && left < order_to_size(this_order - 1)) |
320ae51f JA |
2048 | this_order--; |
2049 | ||
2050 | do { | |
59f082e4 | 2051 | page = alloc_pages_node(node, |
36e1f3d1 | 2052 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO, |
a5164405 | 2053 | this_order); |
320ae51f JA |
2054 | if (page) |
2055 | break; | |
2056 | if (!this_order--) | |
2057 | break; | |
2058 | if (order_to_size(this_order) < rq_size) | |
2059 | break; | |
2060 | } while (1); | |
2061 | ||
2062 | if (!page) | |
24d2f903 | 2063 | goto fail; |
320ae51f JA |
2064 | |
2065 | page->private = this_order; | |
24d2f903 | 2066 | list_add_tail(&page->lru, &tags->page_list); |
320ae51f JA |
2067 | |
2068 | p = page_address(page); | |
f75782e4 CM |
2069 | /* |
2070 | * Allow kmemleak to scan these pages as they contain pointers | |
2071 | * to additional allocations like via ops->init_request(). | |
2072 | */ | |
36e1f3d1 | 2073 | kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO); |
320ae51f | 2074 | entries_per_page = order_to_size(this_order) / rq_size; |
cc71a6f4 | 2075 | to_do = min(entries_per_page, depth - i); |
320ae51f JA |
2076 | left -= to_do * rq_size; |
2077 | for (j = 0; j < to_do; j++) { | |
2af8cbe3 JA |
2078 | struct request *rq = p; |
2079 | ||
2080 | tags->static_rqs[i] = rq; | |
1d9bd516 TH |
2081 | if (blk_mq_init_request(set, rq, hctx_idx, node)) { |
2082 | tags->static_rqs[i] = NULL; | |
2083 | goto fail; | |
e9b267d9 CH |
2084 | } |
2085 | ||
320ae51f JA |
2086 | p += rq_size; |
2087 | i++; | |
2088 | } | |
2089 | } | |
cc71a6f4 | 2090 | return 0; |
320ae51f | 2091 | |
24d2f903 | 2092 | fail: |
cc71a6f4 JA |
2093 | blk_mq_free_rqs(set, tags, hctx_idx); |
2094 | return -ENOMEM; | |
320ae51f JA |
2095 | } |
2096 | ||
e57690fe JA |
2097 | /* |
2098 | * 'cpu' is going away. splice any existing rq_list entries from this | |
2099 | * software queue to the hw queue dispatch list, and ensure that it | |
2100 | * gets run. | |
2101 | */ | |
9467f859 | 2102 | static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node) |
484b4061 | 2103 | { |
9467f859 | 2104 | struct blk_mq_hw_ctx *hctx; |
484b4061 JA |
2105 | struct blk_mq_ctx *ctx; |
2106 | LIST_HEAD(tmp); | |
2107 | ||
9467f859 | 2108 | hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead); |
e57690fe | 2109 | ctx = __blk_mq_get_ctx(hctx->queue, cpu); |
484b4061 JA |
2110 | |
2111 | spin_lock(&ctx->lock); | |
2112 | if (!list_empty(&ctx->rq_list)) { | |
2113 | list_splice_init(&ctx->rq_list, &tmp); | |
2114 | blk_mq_hctx_clear_pending(hctx, ctx); | |
2115 | } | |
2116 | spin_unlock(&ctx->lock); | |
2117 | ||
2118 | if (list_empty(&tmp)) | |
9467f859 | 2119 | return 0; |
484b4061 | 2120 | |
e57690fe JA |
2121 | spin_lock(&hctx->lock); |
2122 | list_splice_tail_init(&tmp, &hctx->dispatch); | |
2123 | spin_unlock(&hctx->lock); | |
484b4061 JA |
2124 | |
2125 | blk_mq_run_hw_queue(hctx, true); | |
9467f859 | 2126 | return 0; |
484b4061 JA |
2127 | } |
2128 | ||
9467f859 | 2129 | static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx) |
484b4061 | 2130 | { |
9467f859 TG |
2131 | cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD, |
2132 | &hctx->cpuhp_dead); | |
484b4061 JA |
2133 | } |
2134 | ||
c3b4afca | 2135 | /* hctx->ctxs will be freed in queue's release handler */ |
08e98fc6 ML |
2136 | static void blk_mq_exit_hctx(struct request_queue *q, |
2137 | struct blk_mq_tag_set *set, | |
2138 | struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx) | |
2139 | { | |
9c1051aa OS |
2140 | blk_mq_debugfs_unregister_hctx(hctx); |
2141 | ||
8ab0b7dc ML |
2142 | if (blk_mq_hw_queue_mapped(hctx)) |
2143 | blk_mq_tag_idle(hctx); | |
08e98fc6 | 2144 | |
f70ced09 | 2145 | if (set->ops->exit_request) |
d6296d39 | 2146 | set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx); |
f70ced09 | 2147 | |
93252632 OS |
2148 | blk_mq_sched_exit_hctx(q, hctx, hctx_idx); |
2149 | ||
08e98fc6 ML |
2150 | if (set->ops->exit_hctx) |
2151 | set->ops->exit_hctx(hctx, hctx_idx); | |
2152 | ||
6a83e74d | 2153 | if (hctx->flags & BLK_MQ_F_BLOCKING) |
05707b64 | 2154 | cleanup_srcu_struct(hctx->srcu); |
6a83e74d | 2155 | |
9467f859 | 2156 | blk_mq_remove_cpuhp(hctx); |
f70ced09 | 2157 | blk_free_flush_queue(hctx->fq); |
88459642 | 2158 | sbitmap_free(&hctx->ctx_map); |
08e98fc6 ML |
2159 | } |
2160 | ||
624dbe47 ML |
2161 | static void blk_mq_exit_hw_queues(struct request_queue *q, |
2162 | struct blk_mq_tag_set *set, int nr_queue) | |
2163 | { | |
2164 | struct blk_mq_hw_ctx *hctx; | |
2165 | unsigned int i; | |
2166 | ||
2167 | queue_for_each_hw_ctx(q, hctx, i) { | |
2168 | if (i == nr_queue) | |
2169 | break; | |
08e98fc6 | 2170 | blk_mq_exit_hctx(q, set, hctx, i); |
624dbe47 | 2171 | } |
624dbe47 ML |
2172 | } |
2173 | ||
08e98fc6 ML |
2174 | static int blk_mq_init_hctx(struct request_queue *q, |
2175 | struct blk_mq_tag_set *set, | |
2176 | struct blk_mq_hw_ctx *hctx, unsigned hctx_idx) | |
320ae51f | 2177 | { |
08e98fc6 ML |
2178 | int node; |
2179 | ||
2180 | node = hctx->numa_node; | |
2181 | if (node == NUMA_NO_NODE) | |
2182 | node = hctx->numa_node = set->numa_node; | |
2183 | ||
9f993737 | 2184 | INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn); |
08e98fc6 ML |
2185 | spin_lock_init(&hctx->lock); |
2186 | INIT_LIST_HEAD(&hctx->dispatch); | |
2187 | hctx->queue = q; | |
2404e607 | 2188 | hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED; |
08e98fc6 | 2189 | |
9467f859 | 2190 | cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead); |
08e98fc6 ML |
2191 | |
2192 | hctx->tags = set->tags[hctx_idx]; | |
320ae51f JA |
2193 | |
2194 | /* | |
08e98fc6 ML |
2195 | * Allocate space for all possible cpus to avoid allocation at |
2196 | * runtime | |
320ae51f | 2197 | */ |
d904bfa7 | 2198 | hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *), |
08e98fc6 ML |
2199 | GFP_KERNEL, node); |
2200 | if (!hctx->ctxs) | |
2201 | goto unregister_cpu_notifier; | |
320ae51f | 2202 | |
88459642 OS |
2203 | if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8), GFP_KERNEL, |
2204 | node)) | |
08e98fc6 | 2205 | goto free_ctxs; |
320ae51f | 2206 | |
08e98fc6 | 2207 | hctx->nr_ctx = 0; |
320ae51f | 2208 | |
5815839b | 2209 | spin_lock_init(&hctx->dispatch_wait_lock); |
eb619fdb JA |
2210 | init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake); |
2211 | INIT_LIST_HEAD(&hctx->dispatch_wait.entry); | |
2212 | ||
08e98fc6 ML |
2213 | if (set->ops->init_hctx && |
2214 | set->ops->init_hctx(hctx, set->driver_data, hctx_idx)) | |
2215 | goto free_bitmap; | |
320ae51f | 2216 | |
93252632 OS |
2217 | if (blk_mq_sched_init_hctx(q, hctx, hctx_idx)) |
2218 | goto exit_hctx; | |
2219 | ||
f70ced09 ML |
2220 | hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size); |
2221 | if (!hctx->fq) | |
93252632 | 2222 | goto sched_exit_hctx; |
320ae51f | 2223 | |
1d9bd516 | 2224 | if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx, node)) |
f70ced09 | 2225 | goto free_fq; |
320ae51f | 2226 | |
6a83e74d | 2227 | if (hctx->flags & BLK_MQ_F_BLOCKING) |
05707b64 | 2228 | init_srcu_struct(hctx->srcu); |
6a83e74d | 2229 | |
9c1051aa OS |
2230 | blk_mq_debugfs_register_hctx(q, hctx); |
2231 | ||
08e98fc6 | 2232 | return 0; |
320ae51f | 2233 | |
f70ced09 ML |
2234 | free_fq: |
2235 | kfree(hctx->fq); | |
93252632 OS |
2236 | sched_exit_hctx: |
2237 | blk_mq_sched_exit_hctx(q, hctx, hctx_idx); | |
f70ced09 ML |
2238 | exit_hctx: |
2239 | if (set->ops->exit_hctx) | |
2240 | set->ops->exit_hctx(hctx, hctx_idx); | |
08e98fc6 | 2241 | free_bitmap: |
88459642 | 2242 | sbitmap_free(&hctx->ctx_map); |
08e98fc6 ML |
2243 | free_ctxs: |
2244 | kfree(hctx->ctxs); | |
2245 | unregister_cpu_notifier: | |
9467f859 | 2246 | blk_mq_remove_cpuhp(hctx); |
08e98fc6 ML |
2247 | return -1; |
2248 | } | |
320ae51f | 2249 | |
320ae51f JA |
2250 | static void blk_mq_init_cpu_queues(struct request_queue *q, |
2251 | unsigned int nr_hw_queues) | |
2252 | { | |
2253 | unsigned int i; | |
2254 | ||
2255 | for_each_possible_cpu(i) { | |
2256 | struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i); | |
2257 | struct blk_mq_hw_ctx *hctx; | |
2258 | ||
320ae51f JA |
2259 | __ctx->cpu = i; |
2260 | spin_lock_init(&__ctx->lock); | |
2261 | INIT_LIST_HEAD(&__ctx->rq_list); | |
2262 | __ctx->queue = q; | |
2263 | ||
320ae51f JA |
2264 | /* |
2265 | * Set local node, IFF we have more than one hw queue. If | |
2266 | * not, we remain on the home node of the device | |
2267 | */ | |
20e4d813 | 2268 | hctx = blk_mq_map_queue(q, i); |
320ae51f | 2269 | if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE) |
bffed457 | 2270 | hctx->numa_node = local_memory_node(cpu_to_node(i)); |
320ae51f JA |
2271 | } |
2272 | } | |
2273 | ||
cc71a6f4 JA |
2274 | static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, int hctx_idx) |
2275 | { | |
2276 | int ret = 0; | |
2277 | ||
2278 | set->tags[hctx_idx] = blk_mq_alloc_rq_map(set, hctx_idx, | |
2279 | set->queue_depth, set->reserved_tags); | |
2280 | if (!set->tags[hctx_idx]) | |
2281 | return false; | |
2282 | ||
2283 | ret = blk_mq_alloc_rqs(set, set->tags[hctx_idx], hctx_idx, | |
2284 | set->queue_depth); | |
2285 | if (!ret) | |
2286 | return true; | |
2287 | ||
2288 | blk_mq_free_rq_map(set->tags[hctx_idx]); | |
2289 | set->tags[hctx_idx] = NULL; | |
2290 | return false; | |
2291 | } | |
2292 | ||
2293 | static void blk_mq_free_map_and_requests(struct blk_mq_tag_set *set, | |
2294 | unsigned int hctx_idx) | |
2295 | { | |
bd166ef1 JA |
2296 | if (set->tags[hctx_idx]) { |
2297 | blk_mq_free_rqs(set, set->tags[hctx_idx], hctx_idx); | |
2298 | blk_mq_free_rq_map(set->tags[hctx_idx]); | |
2299 | set->tags[hctx_idx] = NULL; | |
2300 | } | |
cc71a6f4 JA |
2301 | } |
2302 | ||
4b855ad3 | 2303 | static void blk_mq_map_swqueue(struct request_queue *q) |
320ae51f | 2304 | { |
4412efec | 2305 | unsigned int i, hctx_idx; |
320ae51f JA |
2306 | struct blk_mq_hw_ctx *hctx; |
2307 | struct blk_mq_ctx *ctx; | |
2a34c087 | 2308 | struct blk_mq_tag_set *set = q->tag_set; |
320ae51f | 2309 | |
60de074b AM |
2310 | /* |
2311 | * Avoid others reading imcomplete hctx->cpumask through sysfs | |
2312 | */ | |
2313 | mutex_lock(&q->sysfs_lock); | |
2314 | ||
320ae51f | 2315 | queue_for_each_hw_ctx(q, hctx, i) { |
e4043dcf | 2316 | cpumask_clear(hctx->cpumask); |
320ae51f | 2317 | hctx->nr_ctx = 0; |
d416c92c | 2318 | hctx->dispatch_from = NULL; |
320ae51f JA |
2319 | } |
2320 | ||
2321 | /* | |
4b855ad3 | 2322 | * Map software to hardware queues. |
4412efec ML |
2323 | * |
2324 | * If the cpu isn't present, the cpu is mapped to first hctx. | |
320ae51f | 2325 | */ |
20e4d813 | 2326 | for_each_possible_cpu(i) { |
4412efec ML |
2327 | hctx_idx = q->mq_map[i]; |
2328 | /* unmapped hw queue can be remapped after CPU topo changed */ | |
2329 | if (!set->tags[hctx_idx] && | |
2330 | !__blk_mq_alloc_rq_map(set, hctx_idx)) { | |
2331 | /* | |
2332 | * If tags initialization fail for some hctx, | |
2333 | * that hctx won't be brought online. In this | |
2334 | * case, remap the current ctx to hctx[0] which | |
2335 | * is guaranteed to always have tags allocated | |
2336 | */ | |
2337 | q->mq_map[i] = 0; | |
2338 | } | |
2339 | ||
897bb0c7 | 2340 | ctx = per_cpu_ptr(q->queue_ctx, i); |
7d7e0f90 | 2341 | hctx = blk_mq_map_queue(q, i); |
868f2f0b | 2342 | |
e4043dcf | 2343 | cpumask_set_cpu(i, hctx->cpumask); |
320ae51f JA |
2344 | ctx->index_hw = hctx->nr_ctx; |
2345 | hctx->ctxs[hctx->nr_ctx++] = ctx; | |
2346 | } | |
506e931f | 2347 | |
60de074b AM |
2348 | mutex_unlock(&q->sysfs_lock); |
2349 | ||
506e931f | 2350 | queue_for_each_hw_ctx(q, hctx, i) { |
4412efec ML |
2351 | /* |
2352 | * If no software queues are mapped to this hardware queue, | |
2353 | * disable it and free the request entries. | |
2354 | */ | |
2355 | if (!hctx->nr_ctx) { | |
2356 | /* Never unmap queue 0. We need it as a | |
2357 | * fallback in case of a new remap fails | |
2358 | * allocation | |
2359 | */ | |
2360 | if (i && set->tags[i]) | |
2361 | blk_mq_free_map_and_requests(set, i); | |
2362 | ||
2363 | hctx->tags = NULL; | |
2364 | continue; | |
2365 | } | |
484b4061 | 2366 | |
2a34c087 ML |
2367 | hctx->tags = set->tags[i]; |
2368 | WARN_ON(!hctx->tags); | |
2369 | ||
889fa31f CY |
2370 | /* |
2371 | * Set the map size to the number of mapped software queues. | |
2372 | * This is more accurate and more efficient than looping | |
2373 | * over all possibly mapped software queues. | |
2374 | */ | |
88459642 | 2375 | sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx); |
889fa31f | 2376 | |
484b4061 JA |
2377 | /* |
2378 | * Initialize batch roundrobin counts | |
2379 | */ | |
f82ddf19 | 2380 | hctx->next_cpu = blk_mq_first_mapped_cpu(hctx); |
506e931f JA |
2381 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; |
2382 | } | |
320ae51f JA |
2383 | } |
2384 | ||
8e8320c9 JA |
2385 | /* |
2386 | * Caller needs to ensure that we're either frozen/quiesced, or that | |
2387 | * the queue isn't live yet. | |
2388 | */ | |
2404e607 | 2389 | static void queue_set_hctx_shared(struct request_queue *q, bool shared) |
0d2602ca JA |
2390 | { |
2391 | struct blk_mq_hw_ctx *hctx; | |
0d2602ca JA |
2392 | int i; |
2393 | ||
2404e607 | 2394 | queue_for_each_hw_ctx(q, hctx, i) { |
97889f9a | 2395 | if (shared) |
2404e607 | 2396 | hctx->flags |= BLK_MQ_F_TAG_SHARED; |
97889f9a | 2397 | else |
2404e607 JM |
2398 | hctx->flags &= ~BLK_MQ_F_TAG_SHARED; |
2399 | } | |
2400 | } | |
2401 | ||
8e8320c9 JA |
2402 | static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set, |
2403 | bool shared) | |
2404e607 JM |
2404 | { |
2405 | struct request_queue *q; | |
0d2602ca | 2406 | |
705cda97 BVA |
2407 | lockdep_assert_held(&set->tag_list_lock); |
2408 | ||
0d2602ca JA |
2409 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
2410 | blk_mq_freeze_queue(q); | |
2404e607 | 2411 | queue_set_hctx_shared(q, shared); |
0d2602ca JA |
2412 | blk_mq_unfreeze_queue(q); |
2413 | } | |
2414 | } | |
2415 | ||
2416 | static void blk_mq_del_queue_tag_set(struct request_queue *q) | |
2417 | { | |
2418 | struct blk_mq_tag_set *set = q->tag_set; | |
2419 | ||
0d2602ca | 2420 | mutex_lock(&set->tag_list_lock); |
705cda97 | 2421 | list_del_rcu(&q->tag_set_list); |
2404e607 JM |
2422 | if (list_is_singular(&set->tag_list)) { |
2423 | /* just transitioned to unshared */ | |
2424 | set->flags &= ~BLK_MQ_F_TAG_SHARED; | |
2425 | /* update existing queue */ | |
2426 | blk_mq_update_tag_set_depth(set, false); | |
2427 | } | |
0d2602ca | 2428 | mutex_unlock(&set->tag_list_lock); |
a347c7ad | 2429 | INIT_LIST_HEAD(&q->tag_set_list); |
0d2602ca JA |
2430 | } |
2431 | ||
2432 | static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set, | |
2433 | struct request_queue *q) | |
2434 | { | |
2435 | q->tag_set = set; | |
2436 | ||
2437 | mutex_lock(&set->tag_list_lock); | |
2404e607 | 2438 | |
ff821d27 JA |
2439 | /* |
2440 | * Check to see if we're transitioning to shared (from 1 to 2 queues). | |
2441 | */ | |
2442 | if (!list_empty(&set->tag_list) && | |
2443 | !(set->flags & BLK_MQ_F_TAG_SHARED)) { | |
2404e607 JM |
2444 | set->flags |= BLK_MQ_F_TAG_SHARED; |
2445 | /* update existing queue */ | |
2446 | blk_mq_update_tag_set_depth(set, true); | |
2447 | } | |
2448 | if (set->flags & BLK_MQ_F_TAG_SHARED) | |
2449 | queue_set_hctx_shared(q, true); | |
705cda97 | 2450 | list_add_tail_rcu(&q->tag_set_list, &set->tag_list); |
2404e607 | 2451 | |
0d2602ca JA |
2452 | mutex_unlock(&set->tag_list_lock); |
2453 | } | |
2454 | ||
e09aae7e ML |
2455 | /* |
2456 | * It is the actual release handler for mq, but we do it from | |
2457 | * request queue's release handler for avoiding use-after-free | |
2458 | * and headache because q->mq_kobj shouldn't have been introduced, | |
2459 | * but we can't group ctx/kctx kobj without it. | |
2460 | */ | |
2461 | void blk_mq_release(struct request_queue *q) | |
2462 | { | |
2463 | struct blk_mq_hw_ctx *hctx; | |
2464 | unsigned int i; | |
2465 | ||
2466 | /* hctx kobj stays in hctx */ | |
c3b4afca ML |
2467 | queue_for_each_hw_ctx(q, hctx, i) { |
2468 | if (!hctx) | |
2469 | continue; | |
6c8b232e | 2470 | kobject_put(&hctx->kobj); |
c3b4afca | 2471 | } |
e09aae7e | 2472 | |
a723bab3 AM |
2473 | q->mq_map = NULL; |
2474 | ||
e09aae7e ML |
2475 | kfree(q->queue_hw_ctx); |
2476 | ||
7ea5fe31 ML |
2477 | /* |
2478 | * release .mq_kobj and sw queue's kobject now because | |
2479 | * both share lifetime with request queue. | |
2480 | */ | |
2481 | blk_mq_sysfs_deinit(q); | |
2482 | ||
e09aae7e ML |
2483 | free_percpu(q->queue_ctx); |
2484 | } | |
2485 | ||
24d2f903 | 2486 | struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set) |
b62c21b7 MS |
2487 | { |
2488 | struct request_queue *uninit_q, *q; | |
2489 | ||
5ee0524b | 2490 | uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node, NULL); |
b62c21b7 MS |
2491 | if (!uninit_q) |
2492 | return ERR_PTR(-ENOMEM); | |
2493 | ||
2494 | q = blk_mq_init_allocated_queue(set, uninit_q); | |
2495 | if (IS_ERR(q)) | |
2496 | blk_cleanup_queue(uninit_q); | |
2497 | ||
2498 | return q; | |
2499 | } | |
2500 | EXPORT_SYMBOL(blk_mq_init_queue); | |
2501 | ||
07319678 BVA |
2502 | static int blk_mq_hw_ctx_size(struct blk_mq_tag_set *tag_set) |
2503 | { | |
2504 | int hw_ctx_size = sizeof(struct blk_mq_hw_ctx); | |
2505 | ||
05707b64 | 2506 | BUILD_BUG_ON(ALIGN(offsetof(struct blk_mq_hw_ctx, srcu), |
07319678 BVA |
2507 | __alignof__(struct blk_mq_hw_ctx)) != |
2508 | sizeof(struct blk_mq_hw_ctx)); | |
2509 | ||
2510 | if (tag_set->flags & BLK_MQ_F_BLOCKING) | |
2511 | hw_ctx_size += sizeof(struct srcu_struct); | |
2512 | ||
2513 | return hw_ctx_size; | |
2514 | } | |
2515 | ||
868f2f0b KB |
2516 | static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set, |
2517 | struct request_queue *q) | |
320ae51f | 2518 | { |
868f2f0b KB |
2519 | int i, j; |
2520 | struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx; | |
f14bbe77 | 2521 | |
868f2f0b | 2522 | blk_mq_sysfs_unregister(q); |
fb350e0a ML |
2523 | |
2524 | /* protect against switching io scheduler */ | |
2525 | mutex_lock(&q->sysfs_lock); | |
24d2f903 | 2526 | for (i = 0; i < set->nr_hw_queues; i++) { |
868f2f0b | 2527 | int node; |
f14bbe77 | 2528 | |
868f2f0b KB |
2529 | if (hctxs[i]) |
2530 | continue; | |
2531 | ||
2532 | node = blk_mq_hw_queue_to_node(q->mq_map, i); | |
07319678 | 2533 | hctxs[i] = kzalloc_node(blk_mq_hw_ctx_size(set), |
cdef54dd | 2534 | GFP_KERNEL, node); |
320ae51f | 2535 | if (!hctxs[i]) |
868f2f0b | 2536 | break; |
320ae51f | 2537 | |
a86073e4 | 2538 | if (!zalloc_cpumask_var_node(&hctxs[i]->cpumask, GFP_KERNEL, |
868f2f0b KB |
2539 | node)) { |
2540 | kfree(hctxs[i]); | |
2541 | hctxs[i] = NULL; | |
2542 | break; | |
2543 | } | |
e4043dcf | 2544 | |
0d2602ca | 2545 | atomic_set(&hctxs[i]->nr_active, 0); |
f14bbe77 | 2546 | hctxs[i]->numa_node = node; |
320ae51f | 2547 | hctxs[i]->queue_num = i; |
868f2f0b KB |
2548 | |
2549 | if (blk_mq_init_hctx(q, set, hctxs[i], i)) { | |
2550 | free_cpumask_var(hctxs[i]->cpumask); | |
2551 | kfree(hctxs[i]); | |
2552 | hctxs[i] = NULL; | |
2553 | break; | |
2554 | } | |
2555 | blk_mq_hctx_kobj_init(hctxs[i]); | |
320ae51f | 2556 | } |
868f2f0b KB |
2557 | for (j = i; j < q->nr_hw_queues; j++) { |
2558 | struct blk_mq_hw_ctx *hctx = hctxs[j]; | |
2559 | ||
2560 | if (hctx) { | |
cc71a6f4 JA |
2561 | if (hctx->tags) |
2562 | blk_mq_free_map_and_requests(set, j); | |
868f2f0b | 2563 | blk_mq_exit_hctx(q, set, hctx, j); |
868f2f0b | 2564 | kobject_put(&hctx->kobj); |
868f2f0b KB |
2565 | hctxs[j] = NULL; |
2566 | ||
2567 | } | |
2568 | } | |
2569 | q->nr_hw_queues = i; | |
fb350e0a | 2570 | mutex_unlock(&q->sysfs_lock); |
868f2f0b KB |
2571 | blk_mq_sysfs_register(q); |
2572 | } | |
2573 | ||
2574 | struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set, | |
2575 | struct request_queue *q) | |
2576 | { | |
66841672 ML |
2577 | /* mark the queue as mq asap */ |
2578 | q->mq_ops = set->ops; | |
2579 | ||
34dbad5d | 2580 | q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn, |
720b8ccc SB |
2581 | blk_mq_poll_stats_bkt, |
2582 | BLK_MQ_POLL_STATS_BKTS, q); | |
34dbad5d OS |
2583 | if (!q->poll_cb) |
2584 | goto err_exit; | |
2585 | ||
868f2f0b KB |
2586 | q->queue_ctx = alloc_percpu(struct blk_mq_ctx); |
2587 | if (!q->queue_ctx) | |
c7de5726 | 2588 | goto err_exit; |
868f2f0b | 2589 | |
737f98cf ML |
2590 | /* init q->mq_kobj and sw queues' kobjects */ |
2591 | blk_mq_sysfs_init(q); | |
2592 | ||
590b5b7d | 2593 | q->queue_hw_ctx = kcalloc_node(nr_cpu_ids, sizeof(*(q->queue_hw_ctx)), |
868f2f0b KB |
2594 | GFP_KERNEL, set->numa_node); |
2595 | if (!q->queue_hw_ctx) | |
2596 | goto err_percpu; | |
2597 | ||
bdd17e75 | 2598 | q->mq_map = set->mq_map; |
868f2f0b KB |
2599 | |
2600 | blk_mq_realloc_hw_ctxs(set, q); | |
2601 | if (!q->nr_hw_queues) | |
2602 | goto err_hctxs; | |
320ae51f | 2603 | |
287922eb | 2604 | INIT_WORK(&q->timeout_work, blk_mq_timeout_work); |
e56f698b | 2605 | blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ); |
320ae51f JA |
2606 | |
2607 | q->nr_queues = nr_cpu_ids; | |
320ae51f | 2608 | |
94eddfbe | 2609 | q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT; |
320ae51f | 2610 | |
05f1dd53 | 2611 | if (!(set->flags & BLK_MQ_F_SG_MERGE)) |
f78bac2c | 2612 | queue_flag_set_unlocked(QUEUE_FLAG_NO_SG_MERGE, q); |
05f1dd53 | 2613 | |
1be036e9 CH |
2614 | q->sg_reserved_size = INT_MAX; |
2615 | ||
2849450a | 2616 | INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work); |
6fca6a61 CH |
2617 | INIT_LIST_HEAD(&q->requeue_list); |
2618 | spin_lock_init(&q->requeue_lock); | |
2619 | ||
254d259d | 2620 | blk_queue_make_request(q, blk_mq_make_request); |
ea435e1b CH |
2621 | if (q->mq_ops->poll) |
2622 | q->poll_fn = blk_mq_poll; | |
07068d5b | 2623 | |
eba71768 JA |
2624 | /* |
2625 | * Do this after blk_queue_make_request() overrides it... | |
2626 | */ | |
2627 | q->nr_requests = set->queue_depth; | |
2628 | ||
64f1c21e JA |
2629 | /* |
2630 | * Default to classic polling | |
2631 | */ | |
2632 | q->poll_nsec = -1; | |
2633 | ||
24d2f903 CH |
2634 | if (set->ops->complete) |
2635 | blk_queue_softirq_done(q, set->ops->complete); | |
30a91cb4 | 2636 | |
24d2f903 | 2637 | blk_mq_init_cpu_queues(q, set->nr_hw_queues); |
0d2602ca | 2638 | blk_mq_add_queue_tag_set(set, q); |
4b855ad3 | 2639 | blk_mq_map_swqueue(q); |
4593fdbe | 2640 | |
d3484991 JA |
2641 | if (!(set->flags & BLK_MQ_F_NO_SCHED)) { |
2642 | int ret; | |
2643 | ||
131d08e1 | 2644 | ret = elevator_init_mq(q); |
d3484991 JA |
2645 | if (ret) |
2646 | return ERR_PTR(ret); | |
2647 | } | |
2648 | ||
320ae51f | 2649 | return q; |
18741986 | 2650 | |
320ae51f | 2651 | err_hctxs: |
868f2f0b | 2652 | kfree(q->queue_hw_ctx); |
320ae51f | 2653 | err_percpu: |
868f2f0b | 2654 | free_percpu(q->queue_ctx); |
c7de5726 ML |
2655 | err_exit: |
2656 | q->mq_ops = NULL; | |
320ae51f JA |
2657 | return ERR_PTR(-ENOMEM); |
2658 | } | |
b62c21b7 | 2659 | EXPORT_SYMBOL(blk_mq_init_allocated_queue); |
320ae51f JA |
2660 | |
2661 | void blk_mq_free_queue(struct request_queue *q) | |
2662 | { | |
624dbe47 | 2663 | struct blk_mq_tag_set *set = q->tag_set; |
320ae51f | 2664 | |
0d2602ca | 2665 | blk_mq_del_queue_tag_set(q); |
624dbe47 | 2666 | blk_mq_exit_hw_queues(q, set, set->nr_hw_queues); |
320ae51f | 2667 | } |
320ae51f JA |
2668 | |
2669 | /* Basically redo blk_mq_init_queue with queue frozen */ | |
4b855ad3 | 2670 | static void blk_mq_queue_reinit(struct request_queue *q) |
320ae51f | 2671 | { |
4ecd4fef | 2672 | WARN_ON_ONCE(!atomic_read(&q->mq_freeze_depth)); |
320ae51f | 2673 | |
9c1051aa | 2674 | blk_mq_debugfs_unregister_hctxs(q); |
67aec14c JA |
2675 | blk_mq_sysfs_unregister(q); |
2676 | ||
320ae51f JA |
2677 | /* |
2678 | * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe | |
ff821d27 JA |
2679 | * we should change hctx numa_node according to the new topology (this |
2680 | * involves freeing and re-allocating memory, worth doing?) | |
320ae51f | 2681 | */ |
4b855ad3 | 2682 | blk_mq_map_swqueue(q); |
320ae51f | 2683 | |
67aec14c | 2684 | blk_mq_sysfs_register(q); |
9c1051aa | 2685 | blk_mq_debugfs_register_hctxs(q); |
320ae51f JA |
2686 | } |
2687 | ||
a5164405 JA |
2688 | static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set) |
2689 | { | |
2690 | int i; | |
2691 | ||
cc71a6f4 JA |
2692 | for (i = 0; i < set->nr_hw_queues; i++) |
2693 | if (!__blk_mq_alloc_rq_map(set, i)) | |
a5164405 | 2694 | goto out_unwind; |
a5164405 JA |
2695 | |
2696 | return 0; | |
2697 | ||
2698 | out_unwind: | |
2699 | while (--i >= 0) | |
cc71a6f4 | 2700 | blk_mq_free_rq_map(set->tags[i]); |
a5164405 | 2701 | |
a5164405 JA |
2702 | return -ENOMEM; |
2703 | } | |
2704 | ||
2705 | /* | |
2706 | * Allocate the request maps associated with this tag_set. Note that this | |
2707 | * may reduce the depth asked for, if memory is tight. set->queue_depth | |
2708 | * will be updated to reflect the allocated depth. | |
2709 | */ | |
2710 | static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set) | |
2711 | { | |
2712 | unsigned int depth; | |
2713 | int err; | |
2714 | ||
2715 | depth = set->queue_depth; | |
2716 | do { | |
2717 | err = __blk_mq_alloc_rq_maps(set); | |
2718 | if (!err) | |
2719 | break; | |
2720 | ||
2721 | set->queue_depth >>= 1; | |
2722 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) { | |
2723 | err = -ENOMEM; | |
2724 | break; | |
2725 | } | |
2726 | } while (set->queue_depth); | |
2727 | ||
2728 | if (!set->queue_depth || err) { | |
2729 | pr_err("blk-mq: failed to allocate request map\n"); | |
2730 | return -ENOMEM; | |
2731 | } | |
2732 | ||
2733 | if (depth != set->queue_depth) | |
2734 | pr_info("blk-mq: reduced tag depth (%u -> %u)\n", | |
2735 | depth, set->queue_depth); | |
2736 | ||
2737 | return 0; | |
2738 | } | |
2739 | ||
ebe8bddb OS |
2740 | static int blk_mq_update_queue_map(struct blk_mq_tag_set *set) |
2741 | { | |
7d4901a9 | 2742 | if (set->ops->map_queues) { |
7d4901a9 ML |
2743 | /* |
2744 | * transport .map_queues is usually done in the following | |
2745 | * way: | |
2746 | * | |
2747 | * for (queue = 0; queue < set->nr_hw_queues; queue++) { | |
2748 | * mask = get_cpu_mask(queue) | |
2749 | * for_each_cpu(cpu, mask) | |
2750 | * set->mq_map[cpu] = queue; | |
2751 | * } | |
2752 | * | |
2753 | * When we need to remap, the table has to be cleared for | |
2754 | * killing stale mapping since one CPU may not be mapped | |
2755 | * to any hw queue. | |
2756 | */ | |
0da73d00 | 2757 | blk_mq_clear_mq_map(set); |
7d4901a9 | 2758 | |
ebe8bddb | 2759 | return set->ops->map_queues(set); |
7d4901a9 | 2760 | } else |
ebe8bddb OS |
2761 | return blk_mq_map_queues(set); |
2762 | } | |
2763 | ||
a4391c64 JA |
2764 | /* |
2765 | * Alloc a tag set to be associated with one or more request queues. | |
2766 | * May fail with EINVAL for various error conditions. May adjust the | |
c018c84f | 2767 | * requested depth down, if it's too large. In that case, the set |
a4391c64 JA |
2768 | * value will be stored in set->queue_depth. |
2769 | */ | |
24d2f903 CH |
2770 | int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set) |
2771 | { | |
da695ba2 CH |
2772 | int ret; |
2773 | ||
205fb5f5 BVA |
2774 | BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS); |
2775 | ||
24d2f903 CH |
2776 | if (!set->nr_hw_queues) |
2777 | return -EINVAL; | |
a4391c64 | 2778 | if (!set->queue_depth) |
24d2f903 CH |
2779 | return -EINVAL; |
2780 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) | |
2781 | return -EINVAL; | |
2782 | ||
7d7e0f90 | 2783 | if (!set->ops->queue_rq) |
24d2f903 CH |
2784 | return -EINVAL; |
2785 | ||
de148297 ML |
2786 | if (!set->ops->get_budget ^ !set->ops->put_budget) |
2787 | return -EINVAL; | |
2788 | ||
a4391c64 JA |
2789 | if (set->queue_depth > BLK_MQ_MAX_DEPTH) { |
2790 | pr_info("blk-mq: reduced tag depth to %u\n", | |
2791 | BLK_MQ_MAX_DEPTH); | |
2792 | set->queue_depth = BLK_MQ_MAX_DEPTH; | |
2793 | } | |
24d2f903 | 2794 | |
6637fadf SL |
2795 | /* |
2796 | * If a crashdump is active, then we are potentially in a very | |
2797 | * memory constrained environment. Limit us to 1 queue and | |
2798 | * 64 tags to prevent using too much memory. | |
2799 | */ | |
2800 | if (is_kdump_kernel()) { | |
2801 | set->nr_hw_queues = 1; | |
2802 | set->queue_depth = min(64U, set->queue_depth); | |
2803 | } | |
868f2f0b KB |
2804 | /* |
2805 | * There is no use for more h/w queues than cpus. | |
2806 | */ | |
2807 | if (set->nr_hw_queues > nr_cpu_ids) | |
2808 | set->nr_hw_queues = nr_cpu_ids; | |
6637fadf | 2809 | |
590b5b7d | 2810 | set->tags = kcalloc_node(nr_cpu_ids, sizeof(struct blk_mq_tags *), |
24d2f903 CH |
2811 | GFP_KERNEL, set->numa_node); |
2812 | if (!set->tags) | |
a5164405 | 2813 | return -ENOMEM; |
24d2f903 | 2814 | |
da695ba2 | 2815 | ret = -ENOMEM; |
590b5b7d KC |
2816 | set->mq_map = kcalloc_node(nr_cpu_ids, sizeof(*set->mq_map), |
2817 | GFP_KERNEL, set->numa_node); | |
bdd17e75 CH |
2818 | if (!set->mq_map) |
2819 | goto out_free_tags; | |
2820 | ||
ebe8bddb | 2821 | ret = blk_mq_update_queue_map(set); |
da695ba2 CH |
2822 | if (ret) |
2823 | goto out_free_mq_map; | |
2824 | ||
2825 | ret = blk_mq_alloc_rq_maps(set); | |
2826 | if (ret) | |
bdd17e75 | 2827 | goto out_free_mq_map; |
24d2f903 | 2828 | |
0d2602ca JA |
2829 | mutex_init(&set->tag_list_lock); |
2830 | INIT_LIST_HEAD(&set->tag_list); | |
2831 | ||
24d2f903 | 2832 | return 0; |
bdd17e75 CH |
2833 | |
2834 | out_free_mq_map: | |
2835 | kfree(set->mq_map); | |
2836 | set->mq_map = NULL; | |
2837 | out_free_tags: | |
5676e7b6 RE |
2838 | kfree(set->tags); |
2839 | set->tags = NULL; | |
da695ba2 | 2840 | return ret; |
24d2f903 CH |
2841 | } |
2842 | EXPORT_SYMBOL(blk_mq_alloc_tag_set); | |
2843 | ||
2844 | void blk_mq_free_tag_set(struct blk_mq_tag_set *set) | |
2845 | { | |
2846 | int i; | |
2847 | ||
cc71a6f4 JA |
2848 | for (i = 0; i < nr_cpu_ids; i++) |
2849 | blk_mq_free_map_and_requests(set, i); | |
484b4061 | 2850 | |
bdd17e75 CH |
2851 | kfree(set->mq_map); |
2852 | set->mq_map = NULL; | |
2853 | ||
981bd189 | 2854 | kfree(set->tags); |
5676e7b6 | 2855 | set->tags = NULL; |
24d2f903 CH |
2856 | } |
2857 | EXPORT_SYMBOL(blk_mq_free_tag_set); | |
2858 | ||
e3a2b3f9 JA |
2859 | int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr) |
2860 | { | |
2861 | struct blk_mq_tag_set *set = q->tag_set; | |
2862 | struct blk_mq_hw_ctx *hctx; | |
2863 | int i, ret; | |
2864 | ||
bd166ef1 | 2865 | if (!set) |
e3a2b3f9 JA |
2866 | return -EINVAL; |
2867 | ||
70f36b60 | 2868 | blk_mq_freeze_queue(q); |
24f5a90f | 2869 | blk_mq_quiesce_queue(q); |
70f36b60 | 2870 | |
e3a2b3f9 JA |
2871 | ret = 0; |
2872 | queue_for_each_hw_ctx(q, hctx, i) { | |
e9137d4b KB |
2873 | if (!hctx->tags) |
2874 | continue; | |
bd166ef1 JA |
2875 | /* |
2876 | * If we're using an MQ scheduler, just update the scheduler | |
2877 | * queue depth. This is similar to what the old code would do. | |
2878 | */ | |
70f36b60 | 2879 | if (!hctx->sched_tags) { |
c2e82a23 | 2880 | ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr, |
70f36b60 JA |
2881 | false); |
2882 | } else { | |
2883 | ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags, | |
2884 | nr, true); | |
2885 | } | |
e3a2b3f9 JA |
2886 | if (ret) |
2887 | break; | |
2888 | } | |
2889 | ||
2890 | if (!ret) | |
2891 | q->nr_requests = nr; | |
2892 | ||
24f5a90f | 2893 | blk_mq_unquiesce_queue(q); |
70f36b60 | 2894 | blk_mq_unfreeze_queue(q); |
70f36b60 | 2895 | |
e3a2b3f9 JA |
2896 | return ret; |
2897 | } | |
2898 | ||
e4dc2b32 KB |
2899 | static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, |
2900 | int nr_hw_queues) | |
868f2f0b KB |
2901 | { |
2902 | struct request_queue *q; | |
2903 | ||
705cda97 BVA |
2904 | lockdep_assert_held(&set->tag_list_lock); |
2905 | ||
868f2f0b KB |
2906 | if (nr_hw_queues > nr_cpu_ids) |
2907 | nr_hw_queues = nr_cpu_ids; | |
2908 | if (nr_hw_queues < 1 || nr_hw_queues == set->nr_hw_queues) | |
2909 | return; | |
2910 | ||
2911 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
2912 | blk_mq_freeze_queue(q); | |
2913 | ||
2914 | set->nr_hw_queues = nr_hw_queues; | |
ebe8bddb | 2915 | blk_mq_update_queue_map(set); |
868f2f0b KB |
2916 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
2917 | blk_mq_realloc_hw_ctxs(set, q); | |
4b855ad3 | 2918 | blk_mq_queue_reinit(q); |
868f2f0b KB |
2919 | } |
2920 | ||
2921 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
2922 | blk_mq_unfreeze_queue(q); | |
2923 | } | |
e4dc2b32 KB |
2924 | |
2925 | void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues) | |
2926 | { | |
2927 | mutex_lock(&set->tag_list_lock); | |
2928 | __blk_mq_update_nr_hw_queues(set, nr_hw_queues); | |
2929 | mutex_unlock(&set->tag_list_lock); | |
2930 | } | |
868f2f0b KB |
2931 | EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues); |
2932 | ||
34dbad5d OS |
2933 | /* Enable polling stats and return whether they were already enabled. */ |
2934 | static bool blk_poll_stats_enable(struct request_queue *q) | |
2935 | { | |
2936 | if (test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) || | |
7dfdbc73 | 2937 | blk_queue_flag_test_and_set(QUEUE_FLAG_POLL_STATS, q)) |
34dbad5d OS |
2938 | return true; |
2939 | blk_stat_add_callback(q, q->poll_cb); | |
2940 | return false; | |
2941 | } | |
2942 | ||
2943 | static void blk_mq_poll_stats_start(struct request_queue *q) | |
2944 | { | |
2945 | /* | |
2946 | * We don't arm the callback if polling stats are not enabled or the | |
2947 | * callback is already active. | |
2948 | */ | |
2949 | if (!test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) || | |
2950 | blk_stat_is_active(q->poll_cb)) | |
2951 | return; | |
2952 | ||
2953 | blk_stat_activate_msecs(q->poll_cb, 100); | |
2954 | } | |
2955 | ||
2956 | static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb) | |
2957 | { | |
2958 | struct request_queue *q = cb->data; | |
720b8ccc | 2959 | int bucket; |
34dbad5d | 2960 | |
720b8ccc SB |
2961 | for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) { |
2962 | if (cb->stat[bucket].nr_samples) | |
2963 | q->poll_stat[bucket] = cb->stat[bucket]; | |
2964 | } | |
34dbad5d OS |
2965 | } |
2966 | ||
64f1c21e JA |
2967 | static unsigned long blk_mq_poll_nsecs(struct request_queue *q, |
2968 | struct blk_mq_hw_ctx *hctx, | |
2969 | struct request *rq) | |
2970 | { | |
64f1c21e | 2971 | unsigned long ret = 0; |
720b8ccc | 2972 | int bucket; |
64f1c21e JA |
2973 | |
2974 | /* | |
2975 | * If stats collection isn't on, don't sleep but turn it on for | |
2976 | * future users | |
2977 | */ | |
34dbad5d | 2978 | if (!blk_poll_stats_enable(q)) |
64f1c21e JA |
2979 | return 0; |
2980 | ||
64f1c21e JA |
2981 | /* |
2982 | * As an optimistic guess, use half of the mean service time | |
2983 | * for this type of request. We can (and should) make this smarter. | |
2984 | * For instance, if the completion latencies are tight, we can | |
2985 | * get closer than just half the mean. This is especially | |
2986 | * important on devices where the completion latencies are longer | |
720b8ccc SB |
2987 | * than ~10 usec. We do use the stats for the relevant IO size |
2988 | * if available which does lead to better estimates. | |
64f1c21e | 2989 | */ |
720b8ccc SB |
2990 | bucket = blk_mq_poll_stats_bkt(rq); |
2991 | if (bucket < 0) | |
2992 | return ret; | |
2993 | ||
2994 | if (q->poll_stat[bucket].nr_samples) | |
2995 | ret = (q->poll_stat[bucket].mean + 1) / 2; | |
64f1c21e JA |
2996 | |
2997 | return ret; | |
2998 | } | |
2999 | ||
06426adf | 3000 | static bool blk_mq_poll_hybrid_sleep(struct request_queue *q, |
64f1c21e | 3001 | struct blk_mq_hw_ctx *hctx, |
06426adf JA |
3002 | struct request *rq) |
3003 | { | |
3004 | struct hrtimer_sleeper hs; | |
3005 | enum hrtimer_mode mode; | |
64f1c21e | 3006 | unsigned int nsecs; |
06426adf JA |
3007 | ktime_t kt; |
3008 | ||
76a86f9d | 3009 | if (rq->rq_flags & RQF_MQ_POLL_SLEPT) |
64f1c21e JA |
3010 | return false; |
3011 | ||
3012 | /* | |
3013 | * poll_nsec can be: | |
3014 | * | |
3015 | * -1: don't ever hybrid sleep | |
3016 | * 0: use half of prev avg | |
3017 | * >0: use this specific value | |
3018 | */ | |
3019 | if (q->poll_nsec == -1) | |
3020 | return false; | |
3021 | else if (q->poll_nsec > 0) | |
3022 | nsecs = q->poll_nsec; | |
3023 | else | |
3024 | nsecs = blk_mq_poll_nsecs(q, hctx, rq); | |
3025 | ||
3026 | if (!nsecs) | |
06426adf JA |
3027 | return false; |
3028 | ||
76a86f9d | 3029 | rq->rq_flags |= RQF_MQ_POLL_SLEPT; |
06426adf JA |
3030 | |
3031 | /* | |
3032 | * This will be replaced with the stats tracking code, using | |
3033 | * 'avg_completion_time / 2' as the pre-sleep target. | |
3034 | */ | |
8b0e1953 | 3035 | kt = nsecs; |
06426adf JA |
3036 | |
3037 | mode = HRTIMER_MODE_REL; | |
3038 | hrtimer_init_on_stack(&hs.timer, CLOCK_MONOTONIC, mode); | |
3039 | hrtimer_set_expires(&hs.timer, kt); | |
3040 | ||
3041 | hrtimer_init_sleeper(&hs, current); | |
3042 | do { | |
5a61c363 | 3043 | if (blk_mq_rq_state(rq) == MQ_RQ_COMPLETE) |
06426adf JA |
3044 | break; |
3045 | set_current_state(TASK_UNINTERRUPTIBLE); | |
3046 | hrtimer_start_expires(&hs.timer, mode); | |
3047 | if (hs.task) | |
3048 | io_schedule(); | |
3049 | hrtimer_cancel(&hs.timer); | |
3050 | mode = HRTIMER_MODE_ABS; | |
3051 | } while (hs.task && !signal_pending(current)); | |
3052 | ||
3053 | __set_current_state(TASK_RUNNING); | |
3054 | destroy_hrtimer_on_stack(&hs.timer); | |
3055 | return true; | |
3056 | } | |
3057 | ||
bbd7bb70 JA |
3058 | static bool __blk_mq_poll(struct blk_mq_hw_ctx *hctx, struct request *rq) |
3059 | { | |
3060 | struct request_queue *q = hctx->queue; | |
3061 | long state; | |
3062 | ||
06426adf JA |
3063 | /* |
3064 | * If we sleep, have the caller restart the poll loop to reset | |
3065 | * the state. Like for the other success return cases, the | |
3066 | * caller is responsible for checking if the IO completed. If | |
3067 | * the IO isn't complete, we'll get called again and will go | |
3068 | * straight to the busy poll loop. | |
3069 | */ | |
64f1c21e | 3070 | if (blk_mq_poll_hybrid_sleep(q, hctx, rq)) |
06426adf JA |
3071 | return true; |
3072 | ||
bbd7bb70 JA |
3073 | hctx->poll_considered++; |
3074 | ||
3075 | state = current->state; | |
3076 | while (!need_resched()) { | |
3077 | int ret; | |
3078 | ||
3079 | hctx->poll_invoked++; | |
3080 | ||
3081 | ret = q->mq_ops->poll(hctx, rq->tag); | |
3082 | if (ret > 0) { | |
3083 | hctx->poll_success++; | |
3084 | set_current_state(TASK_RUNNING); | |
3085 | return true; | |
3086 | } | |
3087 | ||
3088 | if (signal_pending_state(state, current)) | |
3089 | set_current_state(TASK_RUNNING); | |
3090 | ||
3091 | if (current->state == TASK_RUNNING) | |
3092 | return true; | |
3093 | if (ret < 0) | |
3094 | break; | |
3095 | cpu_relax(); | |
3096 | } | |
3097 | ||
67b4110f | 3098 | __set_current_state(TASK_RUNNING); |
bbd7bb70 JA |
3099 | return false; |
3100 | } | |
3101 | ||
ea435e1b | 3102 | static bool blk_mq_poll(struct request_queue *q, blk_qc_t cookie) |
bbd7bb70 JA |
3103 | { |
3104 | struct blk_mq_hw_ctx *hctx; | |
bbd7bb70 JA |
3105 | struct request *rq; |
3106 | ||
ea435e1b | 3107 | if (!test_bit(QUEUE_FLAG_POLL, &q->queue_flags)) |
bbd7bb70 JA |
3108 | return false; |
3109 | ||
bbd7bb70 | 3110 | hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)]; |
bd166ef1 JA |
3111 | if (!blk_qc_t_is_internal(cookie)) |
3112 | rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie)); | |
3a07bb1d | 3113 | else { |
bd166ef1 | 3114 | rq = blk_mq_tag_to_rq(hctx->sched_tags, blk_qc_t_to_tag(cookie)); |
3a07bb1d JA |
3115 | /* |
3116 | * With scheduling, if the request has completed, we'll | |
3117 | * get a NULL return here, as we clear the sched tag when | |
3118 | * that happens. The request still remains valid, like always, | |
3119 | * so we should be safe with just the NULL check. | |
3120 | */ | |
3121 | if (!rq) | |
3122 | return false; | |
3123 | } | |
bbd7bb70 JA |
3124 | |
3125 | return __blk_mq_poll(hctx, rq); | |
3126 | } | |
bbd7bb70 | 3127 | |
320ae51f JA |
3128 | static int __init blk_mq_init(void) |
3129 | { | |
9467f859 TG |
3130 | cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL, |
3131 | blk_mq_hctx_notify_dead); | |
320ae51f JA |
3132 | return 0; |
3133 | } | |
3134 | subsys_initcall(blk_mq_init); |