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" |
87760e5e | 37 | #include "blk-wbt.h" |
bd166ef1 | 38 | #include "blk-mq-sched.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 { | |
288 | if (blk_mq_tag_busy(data->hctx)) { | |
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 CH |
334 | data->ctx->rq_dispatched[op_is_sync(op)]++; |
335 | return rq; | |
5dee8577 CH |
336 | } |
337 | ||
d2c0d383 CH |
338 | static struct request *blk_mq_get_request(struct request_queue *q, |
339 | struct bio *bio, unsigned int op, | |
340 | struct blk_mq_alloc_data *data) | |
341 | { | |
342 | struct elevator_queue *e = q->elevator; | |
343 | struct request *rq; | |
e4cdf1a1 | 344 | unsigned int tag; |
21e768b4 | 345 | bool put_ctx_on_error = false; |
d2c0d383 CH |
346 | |
347 | blk_queue_enter_live(q); | |
348 | data->q = q; | |
21e768b4 BVA |
349 | if (likely(!data->ctx)) { |
350 | data->ctx = blk_mq_get_ctx(q); | |
351 | put_ctx_on_error = true; | |
352 | } | |
d2c0d383 CH |
353 | if (likely(!data->hctx)) |
354 | data->hctx = blk_mq_map_queue(q, data->ctx->cpu); | |
03a07c92 GR |
355 | if (op & REQ_NOWAIT) |
356 | data->flags |= BLK_MQ_REQ_NOWAIT; | |
d2c0d383 CH |
357 | |
358 | if (e) { | |
359 | data->flags |= BLK_MQ_REQ_INTERNAL; | |
360 | ||
361 | /* | |
362 | * Flush requests are special and go directly to the | |
17a51199 JA |
363 | * dispatch list. Don't include reserved tags in the |
364 | * limiting, as it isn't useful. | |
d2c0d383 | 365 | */ |
17a51199 JA |
366 | if (!op_is_flush(op) && e->type->ops.mq.limit_depth && |
367 | !(data->flags & BLK_MQ_REQ_RESERVED)) | |
5bbf4e5a | 368 | e->type->ops.mq.limit_depth(op, data); |
d2c0d383 CH |
369 | } |
370 | ||
e4cdf1a1 CH |
371 | tag = blk_mq_get_tag(data); |
372 | if (tag == BLK_MQ_TAG_FAIL) { | |
21e768b4 BVA |
373 | if (put_ctx_on_error) { |
374 | blk_mq_put_ctx(data->ctx); | |
1ad43c00 ML |
375 | data->ctx = NULL; |
376 | } | |
037cebb8 CH |
377 | blk_queue_exit(q); |
378 | return NULL; | |
d2c0d383 CH |
379 | } |
380 | ||
e4cdf1a1 | 381 | rq = blk_mq_rq_ctx_init(data, tag, op); |
037cebb8 CH |
382 | if (!op_is_flush(op)) { |
383 | rq->elv.icq = NULL; | |
5bbf4e5a | 384 | if (e && e->type->ops.mq.prepare_request) { |
44e8c2bf CH |
385 | if (e->type->icq_cache && rq_ioc(bio)) |
386 | blk_mq_sched_assign_ioc(rq, bio); | |
387 | ||
5bbf4e5a CH |
388 | e->type->ops.mq.prepare_request(rq, bio); |
389 | rq->rq_flags |= RQF_ELVPRIV; | |
44e8c2bf | 390 | } |
037cebb8 CH |
391 | } |
392 | data->hctx->queued++; | |
393 | return rq; | |
d2c0d383 CH |
394 | } |
395 | ||
cd6ce148 | 396 | struct request *blk_mq_alloc_request(struct request_queue *q, unsigned int op, |
9a95e4ef | 397 | blk_mq_req_flags_t flags) |
320ae51f | 398 | { |
5a797e00 | 399 | struct blk_mq_alloc_data alloc_data = { .flags = flags }; |
bd166ef1 | 400 | struct request *rq; |
a492f075 | 401 | int ret; |
320ae51f | 402 | |
3a0a5299 | 403 | ret = blk_queue_enter(q, flags); |
a492f075 JL |
404 | if (ret) |
405 | return ERR_PTR(ret); | |
320ae51f | 406 | |
cd6ce148 | 407 | rq = blk_mq_get_request(q, NULL, op, &alloc_data); |
3280d66a | 408 | blk_queue_exit(q); |
841bac2c | 409 | |
bd166ef1 | 410 | if (!rq) |
a492f075 | 411 | return ERR_PTR(-EWOULDBLOCK); |
0c4de0f3 | 412 | |
1ad43c00 | 413 | blk_mq_put_ctx(alloc_data.ctx); |
1ad43c00 | 414 | |
0c4de0f3 CH |
415 | rq->__data_len = 0; |
416 | rq->__sector = (sector_t) -1; | |
417 | rq->bio = rq->biotail = NULL; | |
320ae51f JA |
418 | return rq; |
419 | } | |
4bb659b1 | 420 | EXPORT_SYMBOL(blk_mq_alloc_request); |
320ae51f | 421 | |
cd6ce148 | 422 | struct request *blk_mq_alloc_request_hctx(struct request_queue *q, |
9a95e4ef | 423 | unsigned int op, blk_mq_req_flags_t flags, unsigned int hctx_idx) |
1f5bd336 | 424 | { |
6d2809d5 | 425 | struct blk_mq_alloc_data alloc_data = { .flags = flags }; |
1f5bd336 | 426 | struct request *rq; |
6d2809d5 | 427 | unsigned int cpu; |
1f5bd336 ML |
428 | int ret; |
429 | ||
430 | /* | |
431 | * If the tag allocator sleeps we could get an allocation for a | |
432 | * different hardware context. No need to complicate the low level | |
433 | * allocator for this for the rare use case of a command tied to | |
434 | * a specific queue. | |
435 | */ | |
436 | if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT))) | |
437 | return ERR_PTR(-EINVAL); | |
438 | ||
439 | if (hctx_idx >= q->nr_hw_queues) | |
440 | return ERR_PTR(-EIO); | |
441 | ||
3a0a5299 | 442 | ret = blk_queue_enter(q, flags); |
1f5bd336 ML |
443 | if (ret) |
444 | return ERR_PTR(ret); | |
445 | ||
c8712c6a CH |
446 | /* |
447 | * Check if the hardware context is actually mapped to anything. | |
448 | * If not tell the caller that it should skip this queue. | |
449 | */ | |
6d2809d5 OS |
450 | alloc_data.hctx = q->queue_hw_ctx[hctx_idx]; |
451 | if (!blk_mq_hw_queue_mapped(alloc_data.hctx)) { | |
452 | blk_queue_exit(q); | |
453 | return ERR_PTR(-EXDEV); | |
c8712c6a | 454 | } |
20e4d813 | 455 | cpu = cpumask_first_and(alloc_data.hctx->cpumask, cpu_online_mask); |
6d2809d5 | 456 | alloc_data.ctx = __blk_mq_get_ctx(q, cpu); |
1f5bd336 | 457 | |
cd6ce148 | 458 | rq = blk_mq_get_request(q, NULL, op, &alloc_data); |
3280d66a | 459 | blk_queue_exit(q); |
c8712c6a | 460 | |
6d2809d5 OS |
461 | if (!rq) |
462 | return ERR_PTR(-EWOULDBLOCK); | |
463 | ||
464 | return rq; | |
1f5bd336 ML |
465 | } |
466 | EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx); | |
467 | ||
6af54051 | 468 | void blk_mq_free_request(struct request *rq) |
320ae51f | 469 | { |
320ae51f | 470 | struct request_queue *q = rq->q; |
6af54051 CH |
471 | struct elevator_queue *e = q->elevator; |
472 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
473 | struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, ctx->cpu); | |
474 | const int sched_tag = rq->internal_tag; | |
475 | ||
5bbf4e5a | 476 | if (rq->rq_flags & RQF_ELVPRIV) { |
6af54051 CH |
477 | if (e && e->type->ops.mq.finish_request) |
478 | e->type->ops.mq.finish_request(rq); | |
479 | if (rq->elv.icq) { | |
480 | put_io_context(rq->elv.icq->ioc); | |
481 | rq->elv.icq = NULL; | |
482 | } | |
483 | } | |
320ae51f | 484 | |
6af54051 | 485 | ctx->rq_completed[rq_is_sync(rq)]++; |
e8064021 | 486 | if (rq->rq_flags & RQF_MQ_INFLIGHT) |
0d2602ca | 487 | atomic_dec(&hctx->nr_active); |
87760e5e | 488 | |
7beb2f84 JA |
489 | if (unlikely(laptop_mode && !blk_rq_is_passthrough(rq))) |
490 | laptop_io_completion(q->backing_dev_info); | |
491 | ||
a8a45941 | 492 | wbt_done(q->rq_wb, rq); |
0d2602ca | 493 | |
85acb3ba SL |
494 | if (blk_rq_rl(rq)) |
495 | blk_put_rl(blk_rq_rl(rq)); | |
496 | ||
1d9bd516 | 497 | blk_mq_rq_update_state(rq, MQ_RQ_IDLE); |
bd166ef1 JA |
498 | if (rq->tag != -1) |
499 | blk_mq_put_tag(hctx, hctx->tags, ctx, rq->tag); | |
500 | if (sched_tag != -1) | |
c05f8525 | 501 | blk_mq_put_tag(hctx, hctx->sched_tags, ctx, sched_tag); |
6d8c6c0f | 502 | blk_mq_sched_restart(hctx); |
3ef28e83 | 503 | blk_queue_exit(q); |
320ae51f | 504 | } |
1a3b595a | 505 | EXPORT_SYMBOL_GPL(blk_mq_free_request); |
320ae51f | 506 | |
2a842aca | 507 | inline void __blk_mq_end_request(struct request *rq, blk_status_t error) |
320ae51f | 508 | { |
522a7775 OS |
509 | u64 now = ktime_get_ns(); |
510 | ||
4bc6339a OS |
511 | if (rq->rq_flags & RQF_STATS) { |
512 | blk_mq_poll_stats_start(rq->q); | |
522a7775 | 513 | blk_stat_add(rq, now); |
4bc6339a OS |
514 | } |
515 | ||
522a7775 | 516 | blk_account_io_done(rq, now); |
0d11e6ac | 517 | |
91b63639 | 518 | if (rq->end_io) { |
a8a45941 | 519 | wbt_done(rq->q->rq_wb, rq); |
320ae51f | 520 | rq->end_io(rq, error); |
91b63639 CH |
521 | } else { |
522 | if (unlikely(blk_bidi_rq(rq))) | |
523 | blk_mq_free_request(rq->next_rq); | |
320ae51f | 524 | blk_mq_free_request(rq); |
91b63639 | 525 | } |
320ae51f | 526 | } |
c8a446ad | 527 | EXPORT_SYMBOL(__blk_mq_end_request); |
63151a44 | 528 | |
2a842aca | 529 | void blk_mq_end_request(struct request *rq, blk_status_t error) |
63151a44 CH |
530 | { |
531 | if (blk_update_request(rq, error, blk_rq_bytes(rq))) | |
532 | BUG(); | |
c8a446ad | 533 | __blk_mq_end_request(rq, error); |
63151a44 | 534 | } |
c8a446ad | 535 | EXPORT_SYMBOL(blk_mq_end_request); |
320ae51f | 536 | |
30a91cb4 | 537 | static void __blk_mq_complete_request_remote(void *data) |
320ae51f | 538 | { |
3d6efbf6 | 539 | struct request *rq = data; |
320ae51f | 540 | |
30a91cb4 | 541 | rq->q->softirq_done_fn(rq); |
320ae51f | 542 | } |
320ae51f | 543 | |
453f8341 | 544 | static void __blk_mq_complete_request(struct request *rq) |
320ae51f JA |
545 | { |
546 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
38535201 | 547 | bool shared = false; |
320ae51f JA |
548 | int cpu; |
549 | ||
1d9bd516 | 550 | WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT); |
5a61c363 | 551 | blk_mq_rq_update_state(rq, MQ_RQ_COMPLETE); |
1d9bd516 | 552 | |
453f8341 CH |
553 | if (rq->internal_tag != -1) |
554 | blk_mq_sched_completed_request(rq); | |
453f8341 | 555 | |
38535201 | 556 | if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) { |
30a91cb4 CH |
557 | rq->q->softirq_done_fn(rq); |
558 | return; | |
559 | } | |
320ae51f JA |
560 | |
561 | cpu = get_cpu(); | |
38535201 CH |
562 | if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags)) |
563 | shared = cpus_share_cache(cpu, ctx->cpu); | |
564 | ||
565 | if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) { | |
30a91cb4 | 566 | rq->csd.func = __blk_mq_complete_request_remote; |
3d6efbf6 CH |
567 | rq->csd.info = rq; |
568 | rq->csd.flags = 0; | |
c46fff2a | 569 | smp_call_function_single_async(ctx->cpu, &rq->csd); |
3d6efbf6 | 570 | } else { |
30a91cb4 | 571 | rq->q->softirq_done_fn(rq); |
3d6efbf6 | 572 | } |
320ae51f JA |
573 | put_cpu(); |
574 | } | |
30a91cb4 | 575 | |
04ced159 | 576 | static void hctx_unlock(struct blk_mq_hw_ctx *hctx, int srcu_idx) |
b7435db8 | 577 | __releases(hctx->srcu) |
04ced159 JA |
578 | { |
579 | if (!(hctx->flags & BLK_MQ_F_BLOCKING)) | |
580 | rcu_read_unlock(); | |
581 | else | |
05707b64 | 582 | srcu_read_unlock(hctx->srcu, srcu_idx); |
04ced159 JA |
583 | } |
584 | ||
585 | static void hctx_lock(struct blk_mq_hw_ctx *hctx, int *srcu_idx) | |
b7435db8 | 586 | __acquires(hctx->srcu) |
04ced159 | 587 | { |
08b5a6e2 JA |
588 | if (!(hctx->flags & BLK_MQ_F_BLOCKING)) { |
589 | /* shut up gcc false positive */ | |
590 | *srcu_idx = 0; | |
04ced159 | 591 | rcu_read_lock(); |
08b5a6e2 | 592 | } else |
05707b64 | 593 | *srcu_idx = srcu_read_lock(hctx->srcu); |
04ced159 JA |
594 | } |
595 | ||
1d9bd516 TH |
596 | static void blk_mq_rq_update_aborted_gstate(struct request *rq, u64 gstate) |
597 | { | |
598 | unsigned long flags; | |
599 | ||
600 | /* | |
601 | * blk_mq_rq_aborted_gstate() is used from the completion path and | |
602 | * can thus be called from irq context. u64_stats_fetch in the | |
603 | * middle of update on the same CPU leads to lockup. Disable irq | |
604 | * while updating. | |
605 | */ | |
606 | local_irq_save(flags); | |
607 | u64_stats_update_begin(&rq->aborted_gstate_sync); | |
608 | rq->aborted_gstate = gstate; | |
609 | u64_stats_update_end(&rq->aborted_gstate_sync); | |
610 | local_irq_restore(flags); | |
611 | } | |
612 | ||
613 | static u64 blk_mq_rq_aborted_gstate(struct request *rq) | |
614 | { | |
615 | unsigned int start; | |
616 | u64 aborted_gstate; | |
617 | ||
618 | do { | |
619 | start = u64_stats_fetch_begin(&rq->aborted_gstate_sync); | |
620 | aborted_gstate = rq->aborted_gstate; | |
621 | } while (u64_stats_fetch_retry(&rq->aborted_gstate_sync, start)); | |
622 | ||
623 | return aborted_gstate; | |
624 | } | |
625 | ||
30a91cb4 CH |
626 | /** |
627 | * blk_mq_complete_request - end I/O on a request | |
628 | * @rq: the request being processed | |
629 | * | |
630 | * Description: | |
631 | * Ends all I/O on a request. It does not handle partial completions. | |
632 | * The actual completion happens out-of-order, through a IPI handler. | |
633 | **/ | |
08e0029a | 634 | void blk_mq_complete_request(struct request *rq) |
30a91cb4 | 635 | { |
95f09684 | 636 | struct request_queue *q = rq->q; |
5197c05e TH |
637 | struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, rq->mq_ctx->cpu); |
638 | int srcu_idx; | |
95f09684 JA |
639 | |
640 | if (unlikely(blk_should_fake_timeout(q))) | |
30a91cb4 | 641 | return; |
5197c05e | 642 | |
1d9bd516 TH |
643 | /* |
644 | * If @rq->aborted_gstate equals the current instance, timeout is | |
645 | * claiming @rq and we lost. This is synchronized through | |
646 | * hctx_lock(). See blk_mq_timeout_work() for details. | |
647 | * | |
648 | * Completion path never blocks and we can directly use RCU here | |
649 | * instead of hctx_lock() which can be either RCU or SRCU. | |
650 | * However, that would complicate paths which want to synchronize | |
651 | * against us. Let stay in sync with the issue path so that | |
652 | * hctx_lock() covers both issue and completion paths. | |
653 | */ | |
5197c05e | 654 | hctx_lock(hctx, &srcu_idx); |
634f9e46 | 655 | if (blk_mq_rq_aborted_gstate(rq) != rq->gstate) |
ed851860 | 656 | __blk_mq_complete_request(rq); |
5197c05e | 657 | hctx_unlock(hctx, srcu_idx); |
30a91cb4 CH |
658 | } |
659 | EXPORT_SYMBOL(blk_mq_complete_request); | |
320ae51f | 660 | |
973c0191 KB |
661 | int blk_mq_request_started(struct request *rq) |
662 | { | |
5a61c363 | 663 | return blk_mq_rq_state(rq) != MQ_RQ_IDLE; |
973c0191 KB |
664 | } |
665 | EXPORT_SYMBOL_GPL(blk_mq_request_started); | |
666 | ||
e2490073 | 667 | void blk_mq_start_request(struct request *rq) |
320ae51f JA |
668 | { |
669 | struct request_queue *q = rq->q; | |
670 | ||
bd166ef1 JA |
671 | blk_mq_sched_started_request(rq); |
672 | ||
320ae51f JA |
673 | trace_block_rq_issue(q, rq); |
674 | ||
cf43e6be | 675 | if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) { |
544ccc8d OS |
676 | rq->io_start_time_ns = ktime_get_ns(); |
677 | #ifdef CONFIG_BLK_DEV_THROTTLING_LOW | |
678 | rq->throtl_size = blk_rq_sectors(rq); | |
679 | #endif | |
cf43e6be | 680 | rq->rq_flags |= RQF_STATS; |
a8a45941 | 681 | wbt_issue(q->rq_wb, rq); |
cf43e6be JA |
682 | } |
683 | ||
1d9bd516 | 684 | WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE); |
538b7534 | 685 | |
87ee7b11 | 686 | /* |
1d9bd516 TH |
687 | * Mark @rq in-flight which also advances the generation number, |
688 | * and register for timeout. Protect with a seqcount to allow the | |
689 | * timeout path to read both @rq->gstate and @rq->deadline | |
690 | * coherently. | |
a7af0af3 | 691 | * |
1d9bd516 TH |
692 | * This is the only place where a request is marked in-flight. If |
693 | * the timeout path reads an in-flight @rq->gstate, the | |
694 | * @rq->deadline it reads together under @rq->gstate_seq is | |
695 | * guaranteed to be the matching one. | |
87ee7b11 | 696 | */ |
1d9bd516 TH |
697 | preempt_disable(); |
698 | write_seqcount_begin(&rq->gstate_seq); | |
699 | ||
700 | blk_mq_rq_update_state(rq, MQ_RQ_IN_FLIGHT); | |
701 | blk_add_timer(rq); | |
702 | ||
703 | write_seqcount_end(&rq->gstate_seq); | |
704 | preempt_enable(); | |
49f5baa5 CH |
705 | |
706 | if (q->dma_drain_size && blk_rq_bytes(rq)) { | |
707 | /* | |
708 | * Make sure space for the drain appears. We know we can do | |
709 | * this because max_hw_segments has been adjusted to be one | |
710 | * fewer than the device can handle. | |
711 | */ | |
712 | rq->nr_phys_segments++; | |
713 | } | |
320ae51f | 714 | } |
e2490073 | 715 | EXPORT_SYMBOL(blk_mq_start_request); |
320ae51f | 716 | |
d9d149a3 | 717 | /* |
5a61c363 TH |
718 | * When we reach here because queue is busy, it's safe to change the state |
719 | * to IDLE without checking @rq->aborted_gstate because we should still be | |
720 | * holding the RCU read lock and thus protected against timeout. | |
d9d149a3 | 721 | */ |
ed0791b2 | 722 | static void __blk_mq_requeue_request(struct request *rq) |
320ae51f JA |
723 | { |
724 | struct request_queue *q = rq->q; | |
725 | ||
923218f6 ML |
726 | blk_mq_put_driver_tag(rq); |
727 | ||
320ae51f | 728 | trace_block_rq_requeue(q, rq); |
a8a45941 | 729 | wbt_requeue(q->rq_wb, rq); |
49f5baa5 | 730 | |
5a61c363 | 731 | if (blk_mq_rq_state(rq) != MQ_RQ_IDLE) { |
1d9bd516 | 732 | blk_mq_rq_update_state(rq, MQ_RQ_IDLE); |
e2490073 CH |
733 | if (q->dma_drain_size && blk_rq_bytes(rq)) |
734 | rq->nr_phys_segments--; | |
735 | } | |
320ae51f JA |
736 | } |
737 | ||
2b053aca | 738 | void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list) |
ed0791b2 | 739 | { |
ed0791b2 | 740 | __blk_mq_requeue_request(rq); |
ed0791b2 | 741 | |
105976f5 ML |
742 | /* this request will be re-inserted to io scheduler queue */ |
743 | blk_mq_sched_requeue_request(rq); | |
744 | ||
ed0791b2 | 745 | BUG_ON(blk_queued_rq(rq)); |
2b053aca | 746 | blk_mq_add_to_requeue_list(rq, true, kick_requeue_list); |
ed0791b2 CH |
747 | } |
748 | EXPORT_SYMBOL(blk_mq_requeue_request); | |
749 | ||
6fca6a61 CH |
750 | static void blk_mq_requeue_work(struct work_struct *work) |
751 | { | |
752 | struct request_queue *q = | |
2849450a | 753 | container_of(work, struct request_queue, requeue_work.work); |
6fca6a61 CH |
754 | LIST_HEAD(rq_list); |
755 | struct request *rq, *next; | |
6fca6a61 | 756 | |
18e9781d | 757 | spin_lock_irq(&q->requeue_lock); |
6fca6a61 | 758 | list_splice_init(&q->requeue_list, &rq_list); |
18e9781d | 759 | spin_unlock_irq(&q->requeue_lock); |
6fca6a61 CH |
760 | |
761 | list_for_each_entry_safe(rq, next, &rq_list, queuelist) { | |
e8064021 | 762 | if (!(rq->rq_flags & RQF_SOFTBARRIER)) |
6fca6a61 CH |
763 | continue; |
764 | ||
e8064021 | 765 | rq->rq_flags &= ~RQF_SOFTBARRIER; |
6fca6a61 | 766 | list_del_init(&rq->queuelist); |
9e97d295 | 767 | blk_mq_sched_insert_request(rq, true, false, false); |
6fca6a61 CH |
768 | } |
769 | ||
770 | while (!list_empty(&rq_list)) { | |
771 | rq = list_entry(rq_list.next, struct request, queuelist); | |
772 | list_del_init(&rq->queuelist); | |
9e97d295 | 773 | blk_mq_sched_insert_request(rq, false, false, false); |
6fca6a61 CH |
774 | } |
775 | ||
52d7f1b5 | 776 | blk_mq_run_hw_queues(q, false); |
6fca6a61 CH |
777 | } |
778 | ||
2b053aca BVA |
779 | void blk_mq_add_to_requeue_list(struct request *rq, bool at_head, |
780 | bool kick_requeue_list) | |
6fca6a61 CH |
781 | { |
782 | struct request_queue *q = rq->q; | |
783 | unsigned long flags; | |
784 | ||
785 | /* | |
786 | * We abuse this flag that is otherwise used by the I/O scheduler to | |
ff821d27 | 787 | * request head insertion from the workqueue. |
6fca6a61 | 788 | */ |
e8064021 | 789 | BUG_ON(rq->rq_flags & RQF_SOFTBARRIER); |
6fca6a61 CH |
790 | |
791 | spin_lock_irqsave(&q->requeue_lock, flags); | |
792 | if (at_head) { | |
e8064021 | 793 | rq->rq_flags |= RQF_SOFTBARRIER; |
6fca6a61 CH |
794 | list_add(&rq->queuelist, &q->requeue_list); |
795 | } else { | |
796 | list_add_tail(&rq->queuelist, &q->requeue_list); | |
797 | } | |
798 | spin_unlock_irqrestore(&q->requeue_lock, flags); | |
2b053aca BVA |
799 | |
800 | if (kick_requeue_list) | |
801 | blk_mq_kick_requeue_list(q); | |
6fca6a61 CH |
802 | } |
803 | EXPORT_SYMBOL(blk_mq_add_to_requeue_list); | |
804 | ||
805 | void blk_mq_kick_requeue_list(struct request_queue *q) | |
806 | { | |
ae943d20 | 807 | kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0); |
6fca6a61 CH |
808 | } |
809 | EXPORT_SYMBOL(blk_mq_kick_requeue_list); | |
810 | ||
2849450a MS |
811 | void blk_mq_delay_kick_requeue_list(struct request_queue *q, |
812 | unsigned long msecs) | |
813 | { | |
d4acf365 BVA |
814 | kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, |
815 | msecs_to_jiffies(msecs)); | |
2849450a MS |
816 | } |
817 | EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list); | |
818 | ||
0e62f51f JA |
819 | struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag) |
820 | { | |
88c7b2b7 JA |
821 | if (tag < tags->nr_tags) { |
822 | prefetch(tags->rqs[tag]); | |
4ee86bab | 823 | return tags->rqs[tag]; |
88c7b2b7 | 824 | } |
4ee86bab HR |
825 | |
826 | return NULL; | |
24d2f903 CH |
827 | } |
828 | EXPORT_SYMBOL(blk_mq_tag_to_rq); | |
829 | ||
320ae51f | 830 | struct blk_mq_timeout_data { |
46f92d42 CH |
831 | unsigned long next; |
832 | unsigned int next_set; | |
1d9bd516 | 833 | unsigned int nr_expired; |
320ae51f JA |
834 | }; |
835 | ||
358f70da | 836 | static void blk_mq_rq_timed_out(struct request *req, bool reserved) |
320ae51f | 837 | { |
f8a5b122 | 838 | const struct blk_mq_ops *ops = req->q->mq_ops; |
46f92d42 | 839 | enum blk_eh_timer_return ret = BLK_EH_RESET_TIMER; |
87ee7b11 | 840 | |
634f9e46 | 841 | req->rq_flags |= RQF_MQ_TIMEOUT_EXPIRED; |
87ee7b11 | 842 | |
46f92d42 | 843 | if (ops->timeout) |
0152fb6b | 844 | ret = ops->timeout(req, reserved); |
46f92d42 CH |
845 | |
846 | switch (ret) { | |
847 | case BLK_EH_HANDLED: | |
848 | __blk_mq_complete_request(req); | |
849 | break; | |
850 | case BLK_EH_RESET_TIMER: | |
1d9bd516 TH |
851 | /* |
852 | * As nothing prevents from completion happening while | |
853 | * ->aborted_gstate is set, this may lead to ignored | |
854 | * completions and further spurious timeouts. | |
855 | */ | |
856 | blk_mq_rq_update_aborted_gstate(req, 0); | |
46f92d42 | 857 | blk_add_timer(req); |
46f92d42 CH |
858 | break; |
859 | case BLK_EH_NOT_HANDLED: | |
860 | break; | |
861 | default: | |
862 | printk(KERN_ERR "block: bad eh return: %d\n", ret); | |
863 | break; | |
864 | } | |
87ee7b11 | 865 | } |
5b3f25fc | 866 | |
81481eb4 CH |
867 | static void blk_mq_check_expired(struct blk_mq_hw_ctx *hctx, |
868 | struct request *rq, void *priv, bool reserved) | |
869 | { | |
870 | struct blk_mq_timeout_data *data = priv; | |
1d9bd516 TH |
871 | unsigned long gstate, deadline; |
872 | int start; | |
87ee7b11 | 873 | |
1d9bd516 | 874 | might_sleep(); |
87ee7b11 | 875 | |
5a61c363 | 876 | if (rq->rq_flags & RQF_MQ_TIMEOUT_EXPIRED) |
46f92d42 | 877 | return; |
a7af0af3 | 878 | |
1d9bd516 TH |
879 | /* read coherent snapshots of @rq->state_gen and @rq->deadline */ |
880 | while (true) { | |
881 | start = read_seqcount_begin(&rq->gstate_seq); | |
882 | gstate = READ_ONCE(rq->gstate); | |
0a72e7f4 | 883 | deadline = blk_rq_deadline(rq); |
1d9bd516 TH |
884 | if (!read_seqcount_retry(&rq->gstate_seq, start)) |
885 | break; | |
886 | cond_resched(); | |
887 | } | |
a7af0af3 | 888 | |
1d9bd516 TH |
889 | /* if in-flight && overdue, mark for abortion */ |
890 | if ((gstate & MQ_RQ_STATE_MASK) == MQ_RQ_IN_FLIGHT && | |
891 | time_after_eq(jiffies, deadline)) { | |
892 | blk_mq_rq_update_aborted_gstate(rq, gstate); | |
893 | data->nr_expired++; | |
894 | hctx->nr_expired++; | |
a7af0af3 PZ |
895 | } else if (!data->next_set || time_after(data->next, deadline)) { |
896 | data->next = deadline; | |
46f92d42 CH |
897 | data->next_set = 1; |
898 | } | |
87ee7b11 JA |
899 | } |
900 | ||
1d9bd516 TH |
901 | static void blk_mq_terminate_expired(struct blk_mq_hw_ctx *hctx, |
902 | struct request *rq, void *priv, bool reserved) | |
903 | { | |
904 | /* | |
905 | * We marked @rq->aborted_gstate and waited for RCU. If there were | |
906 | * completions that we lost to, they would have finished and | |
907 | * updated @rq->gstate by now; otherwise, the completion path is | |
908 | * now guaranteed to see @rq->aborted_gstate and yield. If | |
909 | * @rq->aborted_gstate still matches @rq->gstate, @rq is ours. | |
910 | */ | |
634f9e46 TH |
911 | if (!(rq->rq_flags & RQF_MQ_TIMEOUT_EXPIRED) && |
912 | READ_ONCE(rq->gstate) == rq->aborted_gstate) | |
1d9bd516 TH |
913 | blk_mq_rq_timed_out(rq, reserved); |
914 | } | |
915 | ||
287922eb | 916 | static void blk_mq_timeout_work(struct work_struct *work) |
320ae51f | 917 | { |
287922eb CH |
918 | struct request_queue *q = |
919 | container_of(work, struct request_queue, timeout_work); | |
81481eb4 CH |
920 | struct blk_mq_timeout_data data = { |
921 | .next = 0, | |
922 | .next_set = 0, | |
1d9bd516 | 923 | .nr_expired = 0, |
81481eb4 | 924 | }; |
1d9bd516 | 925 | struct blk_mq_hw_ctx *hctx; |
81481eb4 | 926 | int i; |
320ae51f | 927 | |
71f79fb3 GKB |
928 | /* A deadlock might occur if a request is stuck requiring a |
929 | * timeout at the same time a queue freeze is waiting | |
930 | * completion, since the timeout code would not be able to | |
931 | * acquire the queue reference here. | |
932 | * | |
933 | * That's why we don't use blk_queue_enter here; instead, we use | |
934 | * percpu_ref_tryget directly, because we need to be able to | |
935 | * obtain a reference even in the short window between the queue | |
936 | * starting to freeze, by dropping the first reference in | |
1671d522 | 937 | * blk_freeze_queue_start, and the moment the last request is |
71f79fb3 GKB |
938 | * consumed, marked by the instant q_usage_counter reaches |
939 | * zero. | |
940 | */ | |
941 | if (!percpu_ref_tryget(&q->q_usage_counter)) | |
287922eb CH |
942 | return; |
943 | ||
1d9bd516 | 944 | /* scan for the expired ones and set their ->aborted_gstate */ |
0bf6cd5b | 945 | blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &data); |
320ae51f | 946 | |
1d9bd516 TH |
947 | if (data.nr_expired) { |
948 | bool has_rcu = false; | |
949 | ||
950 | /* | |
951 | * Wait till everyone sees ->aborted_gstate. The | |
952 | * sequential waits for SRCUs aren't ideal. If this ever | |
953 | * becomes a problem, we can add per-hw_ctx rcu_head and | |
954 | * wait in parallel. | |
955 | */ | |
956 | queue_for_each_hw_ctx(q, hctx, i) { | |
957 | if (!hctx->nr_expired) | |
958 | continue; | |
959 | ||
960 | if (!(hctx->flags & BLK_MQ_F_BLOCKING)) | |
961 | has_rcu = true; | |
962 | else | |
05707b64 | 963 | synchronize_srcu(hctx->srcu); |
1d9bd516 TH |
964 | |
965 | hctx->nr_expired = 0; | |
966 | } | |
967 | if (has_rcu) | |
968 | synchronize_rcu(); | |
969 | ||
970 | /* terminate the ones we won */ | |
971 | blk_mq_queue_tag_busy_iter(q, blk_mq_terminate_expired, NULL); | |
972 | } | |
973 | ||
81481eb4 CH |
974 | if (data.next_set) { |
975 | data.next = blk_rq_timeout(round_jiffies_up(data.next)); | |
976 | mod_timer(&q->timeout, data.next); | |
0d2602ca | 977 | } else { |
fcd36c36 BVA |
978 | /* |
979 | * Request timeouts are handled as a forward rolling timer. If | |
980 | * we end up here it means that no requests are pending and | |
981 | * also that no request has been pending for a while. Mark | |
982 | * each hctx as idle. | |
983 | */ | |
f054b56c ML |
984 | queue_for_each_hw_ctx(q, hctx, i) { |
985 | /* the hctx may be unmapped, so check it here */ | |
986 | if (blk_mq_hw_queue_mapped(hctx)) | |
987 | blk_mq_tag_idle(hctx); | |
988 | } | |
0d2602ca | 989 | } |
287922eb | 990 | blk_queue_exit(q); |
320ae51f JA |
991 | } |
992 | ||
88459642 OS |
993 | struct flush_busy_ctx_data { |
994 | struct blk_mq_hw_ctx *hctx; | |
995 | struct list_head *list; | |
996 | }; | |
997 | ||
998 | static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data) | |
999 | { | |
1000 | struct flush_busy_ctx_data *flush_data = data; | |
1001 | struct blk_mq_hw_ctx *hctx = flush_data->hctx; | |
1002 | struct blk_mq_ctx *ctx = hctx->ctxs[bitnr]; | |
1003 | ||
88459642 OS |
1004 | spin_lock(&ctx->lock); |
1005 | list_splice_tail_init(&ctx->rq_list, flush_data->list); | |
e9a99a63 | 1006 | sbitmap_clear_bit(sb, bitnr); |
88459642 OS |
1007 | spin_unlock(&ctx->lock); |
1008 | return true; | |
1009 | } | |
1010 | ||
1429d7c9 JA |
1011 | /* |
1012 | * Process software queues that have been marked busy, splicing them | |
1013 | * to the for-dispatch | |
1014 | */ | |
2c3ad667 | 1015 | void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list) |
1429d7c9 | 1016 | { |
88459642 OS |
1017 | struct flush_busy_ctx_data data = { |
1018 | .hctx = hctx, | |
1019 | .list = list, | |
1020 | }; | |
1429d7c9 | 1021 | |
88459642 | 1022 | sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data); |
1429d7c9 | 1023 | } |
2c3ad667 | 1024 | EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs); |
1429d7c9 | 1025 | |
b347689f ML |
1026 | struct dispatch_rq_data { |
1027 | struct blk_mq_hw_ctx *hctx; | |
1028 | struct request *rq; | |
1029 | }; | |
1030 | ||
1031 | static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr, | |
1032 | void *data) | |
1033 | { | |
1034 | struct dispatch_rq_data *dispatch_data = data; | |
1035 | struct blk_mq_hw_ctx *hctx = dispatch_data->hctx; | |
1036 | struct blk_mq_ctx *ctx = hctx->ctxs[bitnr]; | |
1037 | ||
1038 | spin_lock(&ctx->lock); | |
1039 | if (unlikely(!list_empty(&ctx->rq_list))) { | |
1040 | dispatch_data->rq = list_entry_rq(ctx->rq_list.next); | |
1041 | list_del_init(&dispatch_data->rq->queuelist); | |
1042 | if (list_empty(&ctx->rq_list)) | |
1043 | sbitmap_clear_bit(sb, bitnr); | |
1044 | } | |
1045 | spin_unlock(&ctx->lock); | |
1046 | ||
1047 | return !dispatch_data->rq; | |
1048 | } | |
1049 | ||
1050 | struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx, | |
1051 | struct blk_mq_ctx *start) | |
1052 | { | |
1053 | unsigned off = start ? start->index_hw : 0; | |
1054 | struct dispatch_rq_data data = { | |
1055 | .hctx = hctx, | |
1056 | .rq = NULL, | |
1057 | }; | |
1058 | ||
1059 | __sbitmap_for_each_set(&hctx->ctx_map, off, | |
1060 | dispatch_rq_from_ctx, &data); | |
1061 | ||
1062 | return data.rq; | |
1063 | } | |
1064 | ||
703fd1c0 JA |
1065 | static inline unsigned int queued_to_index(unsigned int queued) |
1066 | { | |
1067 | if (!queued) | |
1068 | return 0; | |
1429d7c9 | 1069 | |
703fd1c0 | 1070 | return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1); |
1429d7c9 JA |
1071 | } |
1072 | ||
bd6737f1 JA |
1073 | bool blk_mq_get_driver_tag(struct request *rq, struct blk_mq_hw_ctx **hctx, |
1074 | bool wait) | |
bd166ef1 JA |
1075 | { |
1076 | struct blk_mq_alloc_data data = { | |
1077 | .q = rq->q, | |
bd166ef1 JA |
1078 | .hctx = blk_mq_map_queue(rq->q, rq->mq_ctx->cpu), |
1079 | .flags = wait ? 0 : BLK_MQ_REQ_NOWAIT, | |
1080 | }; | |
1081 | ||
5feeacdd JA |
1082 | might_sleep_if(wait); |
1083 | ||
81380ca1 OS |
1084 | if (rq->tag != -1) |
1085 | goto done; | |
bd166ef1 | 1086 | |
415b806d SG |
1087 | if (blk_mq_tag_is_reserved(data.hctx->sched_tags, rq->internal_tag)) |
1088 | data.flags |= BLK_MQ_REQ_RESERVED; | |
1089 | ||
bd166ef1 JA |
1090 | rq->tag = blk_mq_get_tag(&data); |
1091 | if (rq->tag >= 0) { | |
200e86b3 JA |
1092 | if (blk_mq_tag_busy(data.hctx)) { |
1093 | rq->rq_flags |= RQF_MQ_INFLIGHT; | |
1094 | atomic_inc(&data.hctx->nr_active); | |
1095 | } | |
bd166ef1 | 1096 | data.hctx->tags->rqs[rq->tag] = rq; |
bd166ef1 JA |
1097 | } |
1098 | ||
81380ca1 OS |
1099 | done: |
1100 | if (hctx) | |
1101 | *hctx = data.hctx; | |
1102 | return rq->tag != -1; | |
bd166ef1 JA |
1103 | } |
1104 | ||
eb619fdb JA |
1105 | static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode, |
1106 | int flags, void *key) | |
da55f2cc OS |
1107 | { |
1108 | struct blk_mq_hw_ctx *hctx; | |
1109 | ||
1110 | hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait); | |
1111 | ||
eb619fdb | 1112 | list_del_init(&wait->entry); |
da55f2cc OS |
1113 | blk_mq_run_hw_queue(hctx, true); |
1114 | return 1; | |
1115 | } | |
1116 | ||
f906a6a0 JA |
1117 | /* |
1118 | * Mark us waiting for a tag. For shared tags, this involves hooking us into | |
ee3e4de5 BVA |
1119 | * the tag wakeups. For non-shared tags, we can simply mark us needing a |
1120 | * restart. For both cases, take care to check the condition again after | |
f906a6a0 JA |
1121 | * marking us as waiting. |
1122 | */ | |
1123 | static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx **hctx, | |
1124 | struct request *rq) | |
da55f2cc | 1125 | { |
eb619fdb | 1126 | struct blk_mq_hw_ctx *this_hctx = *hctx; |
da55f2cc | 1127 | struct sbq_wait_state *ws; |
f906a6a0 JA |
1128 | wait_queue_entry_t *wait; |
1129 | bool ret; | |
da55f2cc | 1130 | |
c27d53fb | 1131 | if (!(this_hctx->flags & BLK_MQ_F_TAG_SHARED)) { |
f906a6a0 JA |
1132 | if (!test_bit(BLK_MQ_S_SCHED_RESTART, &this_hctx->state)) |
1133 | set_bit(BLK_MQ_S_SCHED_RESTART, &this_hctx->state); | |
f906a6a0 | 1134 | |
c27d53fb BVA |
1135 | /* |
1136 | * It's possible that a tag was freed in the window between the | |
1137 | * allocation failure and adding the hardware queue to the wait | |
1138 | * queue. | |
1139 | * | |
1140 | * Don't clear RESTART here, someone else could have set it. | |
1141 | * At most this will cost an extra queue run. | |
1142 | */ | |
1143 | return blk_mq_get_driver_tag(rq, hctx, false); | |
eb619fdb | 1144 | } |
eb619fdb | 1145 | |
c27d53fb BVA |
1146 | wait = &this_hctx->dispatch_wait; |
1147 | if (!list_empty_careful(&wait->entry)) | |
1148 | return false; | |
1149 | ||
1150 | spin_lock(&this_hctx->lock); | |
1151 | if (!list_empty(&wait->entry)) { | |
1152 | spin_unlock(&this_hctx->lock); | |
1153 | return false; | |
eb619fdb JA |
1154 | } |
1155 | ||
c27d53fb BVA |
1156 | ws = bt_wait_ptr(&this_hctx->tags->bitmap_tags, this_hctx); |
1157 | add_wait_queue(&ws->wait, wait); | |
1158 | ||
da55f2cc | 1159 | /* |
eb619fdb JA |
1160 | * It's possible that a tag was freed in the window between the |
1161 | * allocation failure and adding the hardware queue to the wait | |
1162 | * queue. | |
da55f2cc | 1163 | */ |
f906a6a0 | 1164 | ret = blk_mq_get_driver_tag(rq, hctx, false); |
c27d53fb | 1165 | if (!ret) { |
eb619fdb | 1166 | spin_unlock(&this_hctx->lock); |
c27d53fb | 1167 | return false; |
eb619fdb | 1168 | } |
c27d53fb BVA |
1169 | |
1170 | /* | |
1171 | * We got a tag, remove ourselves from the wait queue to ensure | |
1172 | * someone else gets the wakeup. | |
1173 | */ | |
1174 | spin_lock_irq(&ws->wait.lock); | |
1175 | list_del_init(&wait->entry); | |
1176 | spin_unlock_irq(&ws->wait.lock); | |
1177 | spin_unlock(&this_hctx->lock); | |
1178 | ||
1179 | return true; | |
da55f2cc OS |
1180 | } |
1181 | ||
86ff7c2a ML |
1182 | #define BLK_MQ_RESOURCE_DELAY 3 /* ms units */ |
1183 | ||
de148297 | 1184 | bool blk_mq_dispatch_rq_list(struct request_queue *q, struct list_head *list, |
eb619fdb | 1185 | bool got_budget) |
320ae51f | 1186 | { |
81380ca1 | 1187 | struct blk_mq_hw_ctx *hctx; |
6d6f167c | 1188 | struct request *rq, *nxt; |
eb619fdb | 1189 | bool no_tag = false; |
fc17b653 | 1190 | int errors, queued; |
86ff7c2a | 1191 | blk_status_t ret = BLK_STS_OK; |
320ae51f | 1192 | |
81380ca1 OS |
1193 | if (list_empty(list)) |
1194 | return false; | |
1195 | ||
de148297 ML |
1196 | WARN_ON(!list_is_singular(list) && got_budget); |
1197 | ||
320ae51f JA |
1198 | /* |
1199 | * Now process all the entries, sending them to the driver. | |
1200 | */ | |
93efe981 | 1201 | errors = queued = 0; |
81380ca1 | 1202 | do { |
74c45052 | 1203 | struct blk_mq_queue_data bd; |
320ae51f | 1204 | |
f04c3df3 | 1205 | rq = list_first_entry(list, struct request, queuelist); |
0bca799b ML |
1206 | |
1207 | hctx = blk_mq_map_queue(rq->q, rq->mq_ctx->cpu); | |
1208 | if (!got_budget && !blk_mq_get_dispatch_budget(hctx)) | |
1209 | break; | |
1210 | ||
1211 | if (!blk_mq_get_driver_tag(rq, NULL, false)) { | |
3c782d67 | 1212 | /* |
da55f2cc | 1213 | * The initial allocation attempt failed, so we need to |
eb619fdb JA |
1214 | * rerun the hardware queue when a tag is freed. The |
1215 | * waitqueue takes care of that. If the queue is run | |
1216 | * before we add this entry back on the dispatch list, | |
1217 | * we'll re-run it below. | |
3c782d67 | 1218 | */ |
f906a6a0 | 1219 | if (!blk_mq_mark_tag_wait(&hctx, rq)) { |
0bca799b | 1220 | blk_mq_put_dispatch_budget(hctx); |
f906a6a0 JA |
1221 | /* |
1222 | * For non-shared tags, the RESTART check | |
1223 | * will suffice. | |
1224 | */ | |
1225 | if (hctx->flags & BLK_MQ_F_TAG_SHARED) | |
1226 | no_tag = true; | |
de148297 ML |
1227 | break; |
1228 | } | |
1229 | } | |
1230 | ||
320ae51f | 1231 | list_del_init(&rq->queuelist); |
320ae51f | 1232 | |
74c45052 | 1233 | bd.rq = rq; |
113285b4 JA |
1234 | |
1235 | /* | |
1236 | * Flag last if we have no more requests, or if we have more | |
1237 | * but can't assign a driver tag to it. | |
1238 | */ | |
1239 | if (list_empty(list)) | |
1240 | bd.last = true; | |
1241 | else { | |
113285b4 JA |
1242 | nxt = list_first_entry(list, struct request, queuelist); |
1243 | bd.last = !blk_mq_get_driver_tag(nxt, NULL, false); | |
1244 | } | |
74c45052 JA |
1245 | |
1246 | ret = q->mq_ops->queue_rq(hctx, &bd); | |
86ff7c2a | 1247 | if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE) { |
6d6f167c JW |
1248 | /* |
1249 | * If an I/O scheduler has been configured and we got a | |
ff821d27 JA |
1250 | * driver tag for the next request already, free it |
1251 | * again. | |
6d6f167c JW |
1252 | */ |
1253 | if (!list_empty(list)) { | |
1254 | nxt = list_first_entry(list, struct request, queuelist); | |
1255 | blk_mq_put_driver_tag(nxt); | |
1256 | } | |
f04c3df3 | 1257 | list_add(&rq->queuelist, list); |
ed0791b2 | 1258 | __blk_mq_requeue_request(rq); |
320ae51f | 1259 | break; |
fc17b653 CH |
1260 | } |
1261 | ||
1262 | if (unlikely(ret != BLK_STS_OK)) { | |
93efe981 | 1263 | errors++; |
2a842aca | 1264 | blk_mq_end_request(rq, BLK_STS_IOERR); |
fc17b653 | 1265 | continue; |
320ae51f JA |
1266 | } |
1267 | ||
fc17b653 | 1268 | queued++; |
81380ca1 | 1269 | } while (!list_empty(list)); |
320ae51f | 1270 | |
703fd1c0 | 1271 | hctx->dispatched[queued_to_index(queued)]++; |
320ae51f JA |
1272 | |
1273 | /* | |
1274 | * Any items that need requeuing? Stuff them into hctx->dispatch, | |
1275 | * that is where we will continue on next queue run. | |
1276 | */ | |
f04c3df3 | 1277 | if (!list_empty(list)) { |
86ff7c2a ML |
1278 | bool needs_restart; |
1279 | ||
320ae51f | 1280 | spin_lock(&hctx->lock); |
c13660a0 | 1281 | list_splice_init(list, &hctx->dispatch); |
320ae51f | 1282 | spin_unlock(&hctx->lock); |
f04c3df3 | 1283 | |
9ba52e58 | 1284 | /* |
710c785f BVA |
1285 | * If SCHED_RESTART was set by the caller of this function and |
1286 | * it is no longer set that means that it was cleared by another | |
1287 | * thread and hence that a queue rerun is needed. | |
9ba52e58 | 1288 | * |
eb619fdb JA |
1289 | * If 'no_tag' is set, that means that we failed getting |
1290 | * a driver tag with an I/O scheduler attached. If our dispatch | |
1291 | * waitqueue is no longer active, ensure that we run the queue | |
1292 | * AFTER adding our entries back to the list. | |
bd166ef1 | 1293 | * |
710c785f BVA |
1294 | * If no I/O scheduler has been configured it is possible that |
1295 | * the hardware queue got stopped and restarted before requests | |
1296 | * were pushed back onto the dispatch list. Rerun the queue to | |
1297 | * avoid starvation. Notes: | |
1298 | * - blk_mq_run_hw_queue() checks whether or not a queue has | |
1299 | * been stopped before rerunning a queue. | |
1300 | * - Some but not all block drivers stop a queue before | |
fc17b653 | 1301 | * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq |
710c785f | 1302 | * and dm-rq. |
86ff7c2a ML |
1303 | * |
1304 | * If driver returns BLK_STS_RESOURCE and SCHED_RESTART | |
1305 | * bit is set, run queue after a delay to avoid IO stalls | |
1306 | * that could otherwise occur if the queue is idle. | |
bd166ef1 | 1307 | */ |
86ff7c2a ML |
1308 | needs_restart = blk_mq_sched_needs_restart(hctx); |
1309 | if (!needs_restart || | |
eb619fdb | 1310 | (no_tag && list_empty_careful(&hctx->dispatch_wait.entry))) |
bd166ef1 | 1311 | blk_mq_run_hw_queue(hctx, true); |
86ff7c2a ML |
1312 | else if (needs_restart && (ret == BLK_STS_RESOURCE)) |
1313 | blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY); | |
320ae51f | 1314 | } |
f04c3df3 | 1315 | |
93efe981 | 1316 | return (queued + errors) != 0; |
f04c3df3 JA |
1317 | } |
1318 | ||
6a83e74d BVA |
1319 | static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx) |
1320 | { | |
1321 | int srcu_idx; | |
1322 | ||
b7a71e66 JA |
1323 | /* |
1324 | * We should be running this queue from one of the CPUs that | |
1325 | * are mapped to it. | |
7df938fb ML |
1326 | * |
1327 | * There are at least two related races now between setting | |
1328 | * hctx->next_cpu from blk_mq_hctx_next_cpu() and running | |
1329 | * __blk_mq_run_hw_queue(): | |
1330 | * | |
1331 | * - hctx->next_cpu is found offline in blk_mq_hctx_next_cpu(), | |
1332 | * but later it becomes online, then this warning is harmless | |
1333 | * at all | |
1334 | * | |
1335 | * - hctx->next_cpu is found online in blk_mq_hctx_next_cpu(), | |
1336 | * but later it becomes offline, then the warning can't be | |
1337 | * triggered, and we depend on blk-mq timeout handler to | |
1338 | * handle dispatched requests to this hctx | |
b7a71e66 | 1339 | */ |
7df938fb ML |
1340 | if (!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) && |
1341 | cpu_online(hctx->next_cpu)) { | |
1342 | printk(KERN_WARNING "run queue from wrong CPU %d, hctx %s\n", | |
1343 | raw_smp_processor_id(), | |
1344 | cpumask_empty(hctx->cpumask) ? "inactive": "active"); | |
1345 | dump_stack(); | |
1346 | } | |
6a83e74d | 1347 | |
b7a71e66 JA |
1348 | /* |
1349 | * We can't run the queue inline with ints disabled. Ensure that | |
1350 | * we catch bad users of this early. | |
1351 | */ | |
1352 | WARN_ON_ONCE(in_interrupt()); | |
1353 | ||
04ced159 | 1354 | might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING); |
bf4907c0 | 1355 | |
04ced159 JA |
1356 | hctx_lock(hctx, &srcu_idx); |
1357 | blk_mq_sched_dispatch_requests(hctx); | |
1358 | hctx_unlock(hctx, srcu_idx); | |
6a83e74d BVA |
1359 | } |
1360 | ||
f82ddf19 ML |
1361 | static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx) |
1362 | { | |
1363 | int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask); | |
1364 | ||
1365 | if (cpu >= nr_cpu_ids) | |
1366 | cpu = cpumask_first(hctx->cpumask); | |
1367 | return cpu; | |
1368 | } | |
1369 | ||
506e931f JA |
1370 | /* |
1371 | * It'd be great if the workqueue API had a way to pass | |
1372 | * in a mask and had some smarts for more clever placement. | |
1373 | * For now we just round-robin here, switching for every | |
1374 | * BLK_MQ_CPU_WORK_BATCH queued items. | |
1375 | */ | |
1376 | static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx) | |
1377 | { | |
7bed4595 | 1378 | bool tried = false; |
476f8c98 | 1379 | int next_cpu = hctx->next_cpu; |
7bed4595 | 1380 | |
b657d7e6 CH |
1381 | if (hctx->queue->nr_hw_queues == 1) |
1382 | return WORK_CPU_UNBOUND; | |
506e931f JA |
1383 | |
1384 | if (--hctx->next_cpu_batch <= 0) { | |
7bed4595 | 1385 | select_cpu: |
476f8c98 | 1386 | next_cpu = cpumask_next_and(next_cpu, hctx->cpumask, |
20e4d813 | 1387 | cpu_online_mask); |
506e931f | 1388 | if (next_cpu >= nr_cpu_ids) |
f82ddf19 | 1389 | next_cpu = blk_mq_first_mapped_cpu(hctx); |
506e931f JA |
1390 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; |
1391 | } | |
1392 | ||
7bed4595 ML |
1393 | /* |
1394 | * Do unbound schedule if we can't find a online CPU for this hctx, | |
1395 | * and it should only happen in the path of handling CPU DEAD. | |
1396 | */ | |
476f8c98 | 1397 | if (!cpu_online(next_cpu)) { |
7bed4595 ML |
1398 | if (!tried) { |
1399 | tried = true; | |
1400 | goto select_cpu; | |
1401 | } | |
1402 | ||
1403 | /* | |
1404 | * Make sure to re-select CPU next time once after CPUs | |
1405 | * in hctx->cpumask become online again. | |
1406 | */ | |
476f8c98 | 1407 | hctx->next_cpu = next_cpu; |
7bed4595 ML |
1408 | hctx->next_cpu_batch = 1; |
1409 | return WORK_CPU_UNBOUND; | |
1410 | } | |
476f8c98 ML |
1411 | |
1412 | hctx->next_cpu = next_cpu; | |
1413 | return next_cpu; | |
506e931f JA |
1414 | } |
1415 | ||
7587a5ae BVA |
1416 | static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async, |
1417 | unsigned long msecs) | |
320ae51f | 1418 | { |
5435c023 | 1419 | if (unlikely(blk_mq_hctx_stopped(hctx))) |
320ae51f JA |
1420 | return; |
1421 | ||
1b792f2f | 1422 | if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) { |
2a90d4aa PB |
1423 | int cpu = get_cpu(); |
1424 | if (cpumask_test_cpu(cpu, hctx->cpumask)) { | |
398205b8 | 1425 | __blk_mq_run_hw_queue(hctx); |
2a90d4aa | 1426 | put_cpu(); |
398205b8 PB |
1427 | return; |
1428 | } | |
e4043dcf | 1429 | |
2a90d4aa | 1430 | put_cpu(); |
e4043dcf | 1431 | } |
398205b8 | 1432 | |
ae943d20 BVA |
1433 | kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work, |
1434 | msecs_to_jiffies(msecs)); | |
7587a5ae BVA |
1435 | } |
1436 | ||
1437 | void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs) | |
1438 | { | |
1439 | __blk_mq_delay_run_hw_queue(hctx, true, msecs); | |
1440 | } | |
1441 | EXPORT_SYMBOL(blk_mq_delay_run_hw_queue); | |
1442 | ||
79f720a7 | 1443 | bool blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) |
7587a5ae | 1444 | { |
24f5a90f ML |
1445 | int srcu_idx; |
1446 | bool need_run; | |
1447 | ||
1448 | /* | |
1449 | * When queue is quiesced, we may be switching io scheduler, or | |
1450 | * updating nr_hw_queues, or other things, and we can't run queue | |
1451 | * any more, even __blk_mq_hctx_has_pending() can't be called safely. | |
1452 | * | |
1453 | * And queue will be rerun in blk_mq_unquiesce_queue() if it is | |
1454 | * quiesced. | |
1455 | */ | |
04ced159 JA |
1456 | hctx_lock(hctx, &srcu_idx); |
1457 | need_run = !blk_queue_quiesced(hctx->queue) && | |
1458 | blk_mq_hctx_has_pending(hctx); | |
1459 | hctx_unlock(hctx, srcu_idx); | |
24f5a90f ML |
1460 | |
1461 | if (need_run) { | |
79f720a7 JA |
1462 | __blk_mq_delay_run_hw_queue(hctx, async, 0); |
1463 | return true; | |
1464 | } | |
1465 | ||
1466 | return false; | |
320ae51f | 1467 | } |
5b727272 | 1468 | EXPORT_SYMBOL(blk_mq_run_hw_queue); |
320ae51f | 1469 | |
b94ec296 | 1470 | void blk_mq_run_hw_queues(struct request_queue *q, bool async) |
320ae51f JA |
1471 | { |
1472 | struct blk_mq_hw_ctx *hctx; | |
1473 | int i; | |
1474 | ||
1475 | queue_for_each_hw_ctx(q, hctx, i) { | |
79f720a7 | 1476 | if (blk_mq_hctx_stopped(hctx)) |
320ae51f JA |
1477 | continue; |
1478 | ||
b94ec296 | 1479 | blk_mq_run_hw_queue(hctx, async); |
320ae51f JA |
1480 | } |
1481 | } | |
b94ec296 | 1482 | EXPORT_SYMBOL(blk_mq_run_hw_queues); |
320ae51f | 1483 | |
fd001443 BVA |
1484 | /** |
1485 | * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped | |
1486 | * @q: request queue. | |
1487 | * | |
1488 | * The caller is responsible for serializing this function against | |
1489 | * blk_mq_{start,stop}_hw_queue(). | |
1490 | */ | |
1491 | bool blk_mq_queue_stopped(struct request_queue *q) | |
1492 | { | |
1493 | struct blk_mq_hw_ctx *hctx; | |
1494 | int i; | |
1495 | ||
1496 | queue_for_each_hw_ctx(q, hctx, i) | |
1497 | if (blk_mq_hctx_stopped(hctx)) | |
1498 | return true; | |
1499 | ||
1500 | return false; | |
1501 | } | |
1502 | EXPORT_SYMBOL(blk_mq_queue_stopped); | |
1503 | ||
39a70c76 ML |
1504 | /* |
1505 | * This function is often used for pausing .queue_rq() by driver when | |
1506 | * there isn't enough resource or some conditions aren't satisfied, and | |
4d606219 | 1507 | * BLK_STS_RESOURCE is usually returned. |
39a70c76 ML |
1508 | * |
1509 | * We do not guarantee that dispatch can be drained or blocked | |
1510 | * after blk_mq_stop_hw_queue() returns. Please use | |
1511 | * blk_mq_quiesce_queue() for that requirement. | |
1512 | */ | |
2719aa21 JA |
1513 | void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx) |
1514 | { | |
641a9ed6 | 1515 | cancel_delayed_work(&hctx->run_work); |
280d45f6 | 1516 | |
641a9ed6 | 1517 | set_bit(BLK_MQ_S_STOPPED, &hctx->state); |
2719aa21 | 1518 | } |
641a9ed6 | 1519 | EXPORT_SYMBOL(blk_mq_stop_hw_queue); |
2719aa21 | 1520 | |
39a70c76 ML |
1521 | /* |
1522 | * This function is often used for pausing .queue_rq() by driver when | |
1523 | * there isn't enough resource or some conditions aren't satisfied, and | |
4d606219 | 1524 | * BLK_STS_RESOURCE is usually returned. |
39a70c76 ML |
1525 | * |
1526 | * We do not guarantee that dispatch can be drained or blocked | |
1527 | * after blk_mq_stop_hw_queues() returns. Please use | |
1528 | * blk_mq_quiesce_queue() for that requirement. | |
1529 | */ | |
2719aa21 JA |
1530 | void blk_mq_stop_hw_queues(struct request_queue *q) |
1531 | { | |
641a9ed6 ML |
1532 | struct blk_mq_hw_ctx *hctx; |
1533 | int i; | |
1534 | ||
1535 | queue_for_each_hw_ctx(q, hctx, i) | |
1536 | blk_mq_stop_hw_queue(hctx); | |
280d45f6 CH |
1537 | } |
1538 | EXPORT_SYMBOL(blk_mq_stop_hw_queues); | |
1539 | ||
320ae51f JA |
1540 | void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx) |
1541 | { | |
1542 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
e4043dcf | 1543 | |
0ffbce80 | 1544 | blk_mq_run_hw_queue(hctx, false); |
320ae51f JA |
1545 | } |
1546 | EXPORT_SYMBOL(blk_mq_start_hw_queue); | |
1547 | ||
2f268556 CH |
1548 | void blk_mq_start_hw_queues(struct request_queue *q) |
1549 | { | |
1550 | struct blk_mq_hw_ctx *hctx; | |
1551 | int i; | |
1552 | ||
1553 | queue_for_each_hw_ctx(q, hctx, i) | |
1554 | blk_mq_start_hw_queue(hctx); | |
1555 | } | |
1556 | EXPORT_SYMBOL(blk_mq_start_hw_queues); | |
1557 | ||
ae911c5e JA |
1558 | void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) |
1559 | { | |
1560 | if (!blk_mq_hctx_stopped(hctx)) | |
1561 | return; | |
1562 | ||
1563 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
1564 | blk_mq_run_hw_queue(hctx, async); | |
1565 | } | |
1566 | EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue); | |
1567 | ||
1b4a3258 | 1568 | void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async) |
320ae51f JA |
1569 | { |
1570 | struct blk_mq_hw_ctx *hctx; | |
1571 | int i; | |
1572 | ||
ae911c5e JA |
1573 | queue_for_each_hw_ctx(q, hctx, i) |
1574 | blk_mq_start_stopped_hw_queue(hctx, async); | |
320ae51f JA |
1575 | } |
1576 | EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues); | |
1577 | ||
70f4db63 | 1578 | static void blk_mq_run_work_fn(struct work_struct *work) |
320ae51f JA |
1579 | { |
1580 | struct blk_mq_hw_ctx *hctx; | |
1581 | ||
9f993737 | 1582 | hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work); |
320ae51f | 1583 | |
21c6e939 | 1584 | /* |
15fe8a90 | 1585 | * If we are stopped, don't run the queue. |
21c6e939 | 1586 | */ |
15fe8a90 | 1587 | if (test_bit(BLK_MQ_S_STOPPED, &hctx->state)) |
21c6e939 | 1588 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); |
7587a5ae BVA |
1589 | |
1590 | __blk_mq_run_hw_queue(hctx); | |
1591 | } | |
1592 | ||
cfd0c552 | 1593 | static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx, |
cfd0c552 ML |
1594 | struct request *rq, |
1595 | bool at_head) | |
320ae51f | 1596 | { |
e57690fe JA |
1597 | struct blk_mq_ctx *ctx = rq->mq_ctx; |
1598 | ||
7b607814 BVA |
1599 | lockdep_assert_held(&ctx->lock); |
1600 | ||
01b983c9 JA |
1601 | trace_block_rq_insert(hctx->queue, rq); |
1602 | ||
72a0a36e CH |
1603 | if (at_head) |
1604 | list_add(&rq->queuelist, &ctx->rq_list); | |
1605 | else | |
1606 | list_add_tail(&rq->queuelist, &ctx->rq_list); | |
cfd0c552 | 1607 | } |
4bb659b1 | 1608 | |
2c3ad667 JA |
1609 | void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq, |
1610 | bool at_head) | |
cfd0c552 ML |
1611 | { |
1612 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
1613 | ||
7b607814 BVA |
1614 | lockdep_assert_held(&ctx->lock); |
1615 | ||
e57690fe | 1616 | __blk_mq_insert_req_list(hctx, rq, at_head); |
320ae51f | 1617 | blk_mq_hctx_mark_pending(hctx, ctx); |
320ae51f JA |
1618 | } |
1619 | ||
157f377b JA |
1620 | /* |
1621 | * Should only be used carefully, when the caller knows we want to | |
1622 | * bypass a potential IO scheduler on the target device. | |
1623 | */ | |
b0850297 | 1624 | void blk_mq_request_bypass_insert(struct request *rq, bool run_queue) |
157f377b JA |
1625 | { |
1626 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
1627 | struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(rq->q, ctx->cpu); | |
1628 | ||
1629 | spin_lock(&hctx->lock); | |
1630 | list_add_tail(&rq->queuelist, &hctx->dispatch); | |
1631 | spin_unlock(&hctx->lock); | |
1632 | ||
b0850297 ML |
1633 | if (run_queue) |
1634 | blk_mq_run_hw_queue(hctx, false); | |
157f377b JA |
1635 | } |
1636 | ||
bd166ef1 JA |
1637 | void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx, |
1638 | struct list_head *list) | |
320ae51f JA |
1639 | |
1640 | { | |
320ae51f JA |
1641 | /* |
1642 | * preemption doesn't flush plug list, so it's possible ctx->cpu is | |
1643 | * offline now | |
1644 | */ | |
1645 | spin_lock(&ctx->lock); | |
1646 | while (!list_empty(list)) { | |
1647 | struct request *rq; | |
1648 | ||
1649 | rq = list_first_entry(list, struct request, queuelist); | |
e57690fe | 1650 | BUG_ON(rq->mq_ctx != ctx); |
320ae51f | 1651 | list_del_init(&rq->queuelist); |
e57690fe | 1652 | __blk_mq_insert_req_list(hctx, rq, false); |
320ae51f | 1653 | } |
cfd0c552 | 1654 | blk_mq_hctx_mark_pending(hctx, ctx); |
320ae51f | 1655 | spin_unlock(&ctx->lock); |
320ae51f JA |
1656 | } |
1657 | ||
1658 | static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b) | |
1659 | { | |
1660 | struct request *rqa = container_of(a, struct request, queuelist); | |
1661 | struct request *rqb = container_of(b, struct request, queuelist); | |
1662 | ||
1663 | return !(rqa->mq_ctx < rqb->mq_ctx || | |
1664 | (rqa->mq_ctx == rqb->mq_ctx && | |
1665 | blk_rq_pos(rqa) < blk_rq_pos(rqb))); | |
1666 | } | |
1667 | ||
1668 | void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule) | |
1669 | { | |
1670 | struct blk_mq_ctx *this_ctx; | |
1671 | struct request_queue *this_q; | |
1672 | struct request *rq; | |
1673 | LIST_HEAD(list); | |
1674 | LIST_HEAD(ctx_list); | |
1675 | unsigned int depth; | |
1676 | ||
1677 | list_splice_init(&plug->mq_list, &list); | |
1678 | ||
1679 | list_sort(NULL, &list, plug_ctx_cmp); | |
1680 | ||
1681 | this_q = NULL; | |
1682 | this_ctx = NULL; | |
1683 | depth = 0; | |
1684 | ||
1685 | while (!list_empty(&list)) { | |
1686 | rq = list_entry_rq(list.next); | |
1687 | list_del_init(&rq->queuelist); | |
1688 | BUG_ON(!rq->q); | |
1689 | if (rq->mq_ctx != this_ctx) { | |
1690 | if (this_ctx) { | |
bd166ef1 JA |
1691 | trace_block_unplug(this_q, depth, from_schedule); |
1692 | blk_mq_sched_insert_requests(this_q, this_ctx, | |
1693 | &ctx_list, | |
1694 | from_schedule); | |
320ae51f JA |
1695 | } |
1696 | ||
1697 | this_ctx = rq->mq_ctx; | |
1698 | this_q = rq->q; | |
1699 | depth = 0; | |
1700 | } | |
1701 | ||
1702 | depth++; | |
1703 | list_add_tail(&rq->queuelist, &ctx_list); | |
1704 | } | |
1705 | ||
1706 | /* | |
1707 | * If 'this_ctx' is set, we know we have entries to complete | |
1708 | * on 'ctx_list'. Do those. | |
1709 | */ | |
1710 | if (this_ctx) { | |
bd166ef1 JA |
1711 | trace_block_unplug(this_q, depth, from_schedule); |
1712 | blk_mq_sched_insert_requests(this_q, this_ctx, &ctx_list, | |
1713 | from_schedule); | |
320ae51f JA |
1714 | } |
1715 | } | |
1716 | ||
1717 | static void blk_mq_bio_to_request(struct request *rq, struct bio *bio) | |
1718 | { | |
da8d7f07 | 1719 | blk_init_request_from_bio(rq, bio); |
4b570521 | 1720 | |
85acb3ba SL |
1721 | blk_rq_set_rl(rq, blk_get_rl(rq->q, bio)); |
1722 | ||
6e85eaf3 | 1723 | blk_account_io_start(rq, true); |
320ae51f JA |
1724 | } |
1725 | ||
ab42f35d ML |
1726 | static inline void blk_mq_queue_io(struct blk_mq_hw_ctx *hctx, |
1727 | struct blk_mq_ctx *ctx, | |
1728 | struct request *rq) | |
1729 | { | |
1730 | spin_lock(&ctx->lock); | |
1731 | __blk_mq_insert_request(hctx, rq, false); | |
1732 | spin_unlock(&ctx->lock); | |
07068d5b | 1733 | } |
14ec77f3 | 1734 | |
fd2d3326 JA |
1735 | static blk_qc_t request_to_qc_t(struct blk_mq_hw_ctx *hctx, struct request *rq) |
1736 | { | |
bd166ef1 JA |
1737 | if (rq->tag != -1) |
1738 | return blk_tag_to_qc_t(rq->tag, hctx->queue_num, false); | |
1739 | ||
1740 | return blk_tag_to_qc_t(rq->internal_tag, hctx->queue_num, true); | |
fd2d3326 JA |
1741 | } |
1742 | ||
0f95549c MS |
1743 | static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx, |
1744 | struct request *rq, | |
1745 | blk_qc_t *cookie) | |
f984df1f | 1746 | { |
f984df1f | 1747 | struct request_queue *q = rq->q; |
f984df1f SL |
1748 | struct blk_mq_queue_data bd = { |
1749 | .rq = rq, | |
d945a365 | 1750 | .last = true, |
f984df1f | 1751 | }; |
bd166ef1 | 1752 | blk_qc_t new_cookie; |
f06345ad | 1753 | blk_status_t ret; |
0f95549c MS |
1754 | |
1755 | new_cookie = request_to_qc_t(hctx, rq); | |
1756 | ||
1757 | /* | |
1758 | * For OK queue, we are done. For error, caller may kill it. | |
1759 | * Any other error (busy), just add it to our list as we | |
1760 | * previously would have done. | |
1761 | */ | |
1762 | ret = q->mq_ops->queue_rq(hctx, &bd); | |
1763 | switch (ret) { | |
1764 | case BLK_STS_OK: | |
1765 | *cookie = new_cookie; | |
1766 | break; | |
1767 | case BLK_STS_RESOURCE: | |
86ff7c2a | 1768 | case BLK_STS_DEV_RESOURCE: |
0f95549c MS |
1769 | __blk_mq_requeue_request(rq); |
1770 | break; | |
1771 | default: | |
1772 | *cookie = BLK_QC_T_NONE; | |
1773 | break; | |
1774 | } | |
1775 | ||
1776 | return ret; | |
1777 | } | |
1778 | ||
0f95549c MS |
1779 | static blk_status_t __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx, |
1780 | struct request *rq, | |
396eaf21 ML |
1781 | blk_qc_t *cookie, |
1782 | bool bypass_insert) | |
0f95549c MS |
1783 | { |
1784 | struct request_queue *q = rq->q; | |
d964f04a ML |
1785 | bool run_queue = true; |
1786 | ||
23d4ee19 ML |
1787 | /* |
1788 | * RCU or SRCU read lock is needed before checking quiesced flag. | |
1789 | * | |
1790 | * When queue is stopped or quiesced, ignore 'bypass_insert' from | |
c77ff7fd | 1791 | * blk_mq_request_issue_directly(), and return BLK_STS_OK to caller, |
23d4ee19 ML |
1792 | * and avoid driver to try to dispatch again. |
1793 | */ | |
f4560ffe | 1794 | if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)) { |
d964f04a | 1795 | run_queue = false; |
23d4ee19 | 1796 | bypass_insert = false; |
d964f04a ML |
1797 | goto insert; |
1798 | } | |
f984df1f | 1799 | |
396eaf21 | 1800 | if (q->elevator && !bypass_insert) |
2253efc8 BVA |
1801 | goto insert; |
1802 | ||
0bca799b | 1803 | if (!blk_mq_get_dispatch_budget(hctx)) |
bd166ef1 JA |
1804 | goto insert; |
1805 | ||
0bca799b ML |
1806 | if (!blk_mq_get_driver_tag(rq, NULL, false)) { |
1807 | blk_mq_put_dispatch_budget(hctx); | |
de148297 | 1808 | goto insert; |
88022d72 | 1809 | } |
de148297 | 1810 | |
0f95549c | 1811 | return __blk_mq_issue_directly(hctx, rq, cookie); |
2253efc8 | 1812 | insert: |
396eaf21 ML |
1813 | if (bypass_insert) |
1814 | return BLK_STS_RESOURCE; | |
0f95549c | 1815 | |
23d4ee19 | 1816 | blk_mq_sched_insert_request(rq, false, run_queue, false); |
0f95549c | 1817 | return BLK_STS_OK; |
f984df1f SL |
1818 | } |
1819 | ||
5eb6126e CH |
1820 | static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx, |
1821 | struct request *rq, blk_qc_t *cookie) | |
1822 | { | |
0f95549c | 1823 | blk_status_t ret; |
04ced159 | 1824 | int srcu_idx; |
bf4907c0 | 1825 | |
04ced159 | 1826 | might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING); |
bf4907c0 | 1827 | |
04ced159 | 1828 | hctx_lock(hctx, &srcu_idx); |
0f95549c | 1829 | |
396eaf21 | 1830 | ret = __blk_mq_try_issue_directly(hctx, rq, cookie, false); |
86ff7c2a | 1831 | if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE) |
23d4ee19 | 1832 | blk_mq_sched_insert_request(rq, false, true, false); |
0f95549c MS |
1833 | else if (ret != BLK_STS_OK) |
1834 | blk_mq_end_request(rq, ret); | |
1835 | ||
04ced159 | 1836 | hctx_unlock(hctx, srcu_idx); |
5eb6126e CH |
1837 | } |
1838 | ||
c77ff7fd | 1839 | blk_status_t blk_mq_request_issue_directly(struct request *rq) |
396eaf21 ML |
1840 | { |
1841 | blk_status_t ret; | |
1842 | int srcu_idx; | |
1843 | blk_qc_t unused_cookie; | |
1844 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
1845 | struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(rq->q, ctx->cpu); | |
1846 | ||
1847 | hctx_lock(hctx, &srcu_idx); | |
1848 | ret = __blk_mq_try_issue_directly(hctx, rq, &unused_cookie, true); | |
1849 | hctx_unlock(hctx, srcu_idx); | |
1850 | ||
1851 | return ret; | |
5eb6126e CH |
1852 | } |
1853 | ||
dece1635 | 1854 | static blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio) |
07068d5b | 1855 | { |
ef295ecf | 1856 | const int is_sync = op_is_sync(bio->bi_opf); |
f73f44eb | 1857 | const int is_flush_fua = op_is_flush(bio->bi_opf); |
5a797e00 | 1858 | struct blk_mq_alloc_data data = { .flags = 0 }; |
07068d5b | 1859 | struct request *rq; |
5eb6126e | 1860 | unsigned int request_count = 0; |
f984df1f | 1861 | struct blk_plug *plug; |
5b3f341f | 1862 | struct request *same_queue_rq = NULL; |
7b371636 | 1863 | blk_qc_t cookie; |
87760e5e | 1864 | unsigned int wb_acct; |
07068d5b JA |
1865 | |
1866 | blk_queue_bounce(q, &bio); | |
1867 | ||
af67c31f | 1868 | blk_queue_split(q, &bio); |
f36ea50c | 1869 | |
e23947bd | 1870 | if (!bio_integrity_prep(bio)) |
dece1635 | 1871 | return BLK_QC_T_NONE; |
07068d5b | 1872 | |
87c279e6 OS |
1873 | if (!is_flush_fua && !blk_queue_nomerges(q) && |
1874 | blk_attempt_plug_merge(q, bio, &request_count, &same_queue_rq)) | |
1875 | return BLK_QC_T_NONE; | |
f984df1f | 1876 | |
bd166ef1 JA |
1877 | if (blk_mq_sched_bio_merge(q, bio)) |
1878 | return BLK_QC_T_NONE; | |
1879 | ||
87760e5e JA |
1880 | wb_acct = wbt_wait(q->rq_wb, bio, NULL); |
1881 | ||
bd166ef1 JA |
1882 | trace_block_getrq(q, bio, bio->bi_opf); |
1883 | ||
d2c0d383 | 1884 | rq = blk_mq_get_request(q, bio, bio->bi_opf, &data); |
87760e5e JA |
1885 | if (unlikely(!rq)) { |
1886 | __wbt_done(q->rq_wb, wb_acct); | |
03a07c92 GR |
1887 | if (bio->bi_opf & REQ_NOWAIT) |
1888 | bio_wouldblock_error(bio); | |
dece1635 | 1889 | return BLK_QC_T_NONE; |
87760e5e JA |
1890 | } |
1891 | ||
a8a45941 | 1892 | wbt_track(rq, wb_acct); |
07068d5b | 1893 | |
fd2d3326 | 1894 | cookie = request_to_qc_t(data.hctx, rq); |
07068d5b | 1895 | |
f984df1f | 1896 | plug = current->plug; |
07068d5b | 1897 | if (unlikely(is_flush_fua)) { |
f984df1f | 1898 | blk_mq_put_ctx(data.ctx); |
07068d5b | 1899 | blk_mq_bio_to_request(rq, bio); |
923218f6 ML |
1900 | |
1901 | /* bypass scheduler for flush rq */ | |
1902 | blk_insert_flush(rq); | |
1903 | blk_mq_run_hw_queue(data.hctx, true); | |
a4d907b6 | 1904 | } else if (plug && q->nr_hw_queues == 1) { |
600271d9 SL |
1905 | struct request *last = NULL; |
1906 | ||
b00c53e8 | 1907 | blk_mq_put_ctx(data.ctx); |
e6c4438b | 1908 | blk_mq_bio_to_request(rq, bio); |
0a6219a9 ML |
1909 | |
1910 | /* | |
1911 | * @request_count may become stale because of schedule | |
1912 | * out, so check the list again. | |
1913 | */ | |
1914 | if (list_empty(&plug->mq_list)) | |
1915 | request_count = 0; | |
254d259d CH |
1916 | else if (blk_queue_nomerges(q)) |
1917 | request_count = blk_plug_queued_count(q); | |
1918 | ||
676d0607 | 1919 | if (!request_count) |
e6c4438b | 1920 | trace_block_plug(q); |
600271d9 SL |
1921 | else |
1922 | last = list_entry_rq(plug->mq_list.prev); | |
b094f89c | 1923 | |
600271d9 SL |
1924 | if (request_count >= BLK_MAX_REQUEST_COUNT || (last && |
1925 | blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) { | |
e6c4438b JM |
1926 | blk_flush_plug_list(plug, false); |
1927 | trace_block_plug(q); | |
320ae51f | 1928 | } |
b094f89c | 1929 | |
e6c4438b | 1930 | list_add_tail(&rq->queuelist, &plug->mq_list); |
2299722c | 1931 | } else if (plug && !blk_queue_nomerges(q)) { |
bd166ef1 | 1932 | blk_mq_bio_to_request(rq, bio); |
07068d5b | 1933 | |
07068d5b | 1934 | /* |
6a83e74d | 1935 | * We do limited plugging. If the bio can be merged, do that. |
f984df1f SL |
1936 | * Otherwise the existing request in the plug list will be |
1937 | * issued. So the plug list will have one request at most | |
2299722c CH |
1938 | * The plug list might get flushed before this. If that happens, |
1939 | * the plug list is empty, and same_queue_rq is invalid. | |
07068d5b | 1940 | */ |
2299722c CH |
1941 | if (list_empty(&plug->mq_list)) |
1942 | same_queue_rq = NULL; | |
1943 | if (same_queue_rq) | |
1944 | list_del_init(&same_queue_rq->queuelist); | |
1945 | list_add_tail(&rq->queuelist, &plug->mq_list); | |
1946 | ||
bf4907c0 JA |
1947 | blk_mq_put_ctx(data.ctx); |
1948 | ||
dad7a3be ML |
1949 | if (same_queue_rq) { |
1950 | data.hctx = blk_mq_map_queue(q, | |
1951 | same_queue_rq->mq_ctx->cpu); | |
2299722c CH |
1952 | blk_mq_try_issue_directly(data.hctx, same_queue_rq, |
1953 | &cookie); | |
dad7a3be | 1954 | } |
a4d907b6 | 1955 | } else if (q->nr_hw_queues > 1 && is_sync) { |
bf4907c0 | 1956 | blk_mq_put_ctx(data.ctx); |
2299722c | 1957 | blk_mq_bio_to_request(rq, bio); |
2299722c | 1958 | blk_mq_try_issue_directly(data.hctx, rq, &cookie); |
a4d907b6 | 1959 | } else if (q->elevator) { |
b00c53e8 | 1960 | blk_mq_put_ctx(data.ctx); |
bd166ef1 | 1961 | blk_mq_bio_to_request(rq, bio); |
9e97d295 | 1962 | blk_mq_sched_insert_request(rq, false, true, true); |
ab42f35d | 1963 | } else { |
b00c53e8 | 1964 | blk_mq_put_ctx(data.ctx); |
ab42f35d ML |
1965 | blk_mq_bio_to_request(rq, bio); |
1966 | blk_mq_queue_io(data.hctx, data.ctx, rq); | |
a4d907b6 | 1967 | blk_mq_run_hw_queue(data.hctx, true); |
ab42f35d | 1968 | } |
320ae51f | 1969 | |
7b371636 | 1970 | return cookie; |
320ae51f JA |
1971 | } |
1972 | ||
cc71a6f4 JA |
1973 | void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags, |
1974 | unsigned int hctx_idx) | |
95363efd | 1975 | { |
e9b267d9 | 1976 | struct page *page; |
320ae51f | 1977 | |
24d2f903 | 1978 | if (tags->rqs && set->ops->exit_request) { |
e9b267d9 | 1979 | int i; |
320ae51f | 1980 | |
24d2f903 | 1981 | for (i = 0; i < tags->nr_tags; i++) { |
2af8cbe3 JA |
1982 | struct request *rq = tags->static_rqs[i]; |
1983 | ||
1984 | if (!rq) | |
e9b267d9 | 1985 | continue; |
d6296d39 | 1986 | set->ops->exit_request(set, rq, hctx_idx); |
2af8cbe3 | 1987 | tags->static_rqs[i] = NULL; |
e9b267d9 | 1988 | } |
320ae51f | 1989 | } |
320ae51f | 1990 | |
24d2f903 CH |
1991 | while (!list_empty(&tags->page_list)) { |
1992 | page = list_first_entry(&tags->page_list, struct page, lru); | |
6753471c | 1993 | list_del_init(&page->lru); |
f75782e4 CM |
1994 | /* |
1995 | * Remove kmemleak object previously allocated in | |
1996 | * blk_mq_init_rq_map(). | |
1997 | */ | |
1998 | kmemleak_free(page_address(page)); | |
320ae51f JA |
1999 | __free_pages(page, page->private); |
2000 | } | |
cc71a6f4 | 2001 | } |
320ae51f | 2002 | |
cc71a6f4 JA |
2003 | void blk_mq_free_rq_map(struct blk_mq_tags *tags) |
2004 | { | |
24d2f903 | 2005 | kfree(tags->rqs); |
cc71a6f4 | 2006 | tags->rqs = NULL; |
2af8cbe3 JA |
2007 | kfree(tags->static_rqs); |
2008 | tags->static_rqs = NULL; | |
320ae51f | 2009 | |
24d2f903 | 2010 | blk_mq_free_tags(tags); |
320ae51f JA |
2011 | } |
2012 | ||
cc71a6f4 JA |
2013 | struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, |
2014 | unsigned int hctx_idx, | |
2015 | unsigned int nr_tags, | |
2016 | unsigned int reserved_tags) | |
320ae51f | 2017 | { |
24d2f903 | 2018 | struct blk_mq_tags *tags; |
59f082e4 | 2019 | int node; |
320ae51f | 2020 | |
59f082e4 SL |
2021 | node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx); |
2022 | if (node == NUMA_NO_NODE) | |
2023 | node = set->numa_node; | |
2024 | ||
2025 | tags = blk_mq_init_tags(nr_tags, reserved_tags, node, | |
24391c0d | 2026 | BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags)); |
24d2f903 CH |
2027 | if (!tags) |
2028 | return NULL; | |
320ae51f | 2029 | |
cc71a6f4 | 2030 | tags->rqs = kzalloc_node(nr_tags * sizeof(struct request *), |
36e1f3d1 | 2031 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, |
59f082e4 | 2032 | node); |
24d2f903 CH |
2033 | if (!tags->rqs) { |
2034 | blk_mq_free_tags(tags); | |
2035 | return NULL; | |
2036 | } | |
320ae51f | 2037 | |
2af8cbe3 JA |
2038 | tags->static_rqs = kzalloc_node(nr_tags * sizeof(struct request *), |
2039 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, | |
59f082e4 | 2040 | node); |
2af8cbe3 JA |
2041 | if (!tags->static_rqs) { |
2042 | kfree(tags->rqs); | |
2043 | blk_mq_free_tags(tags); | |
2044 | return NULL; | |
2045 | } | |
2046 | ||
cc71a6f4 JA |
2047 | return tags; |
2048 | } | |
2049 | ||
2050 | static size_t order_to_size(unsigned int order) | |
2051 | { | |
2052 | return (size_t)PAGE_SIZE << order; | |
2053 | } | |
2054 | ||
1d9bd516 TH |
2055 | static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq, |
2056 | unsigned int hctx_idx, int node) | |
2057 | { | |
2058 | int ret; | |
2059 | ||
2060 | if (set->ops->init_request) { | |
2061 | ret = set->ops->init_request(set, rq, hctx_idx, node); | |
2062 | if (ret) | |
2063 | return ret; | |
2064 | } | |
2065 | ||
2066 | seqcount_init(&rq->gstate_seq); | |
2067 | u64_stats_init(&rq->aborted_gstate_sync); | |
f4560231 JW |
2068 | /* |
2069 | * start gstate with gen 1 instead of 0, otherwise it will be equal | |
2070 | * to aborted_gstate, and be identified timed out by | |
2071 | * blk_mq_terminate_expired. | |
2072 | */ | |
2073 | WRITE_ONCE(rq->gstate, MQ_RQ_GEN_INC); | |
2074 | ||
1d9bd516 TH |
2075 | return 0; |
2076 | } | |
2077 | ||
cc71a6f4 JA |
2078 | int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags, |
2079 | unsigned int hctx_idx, unsigned int depth) | |
2080 | { | |
2081 | unsigned int i, j, entries_per_page, max_order = 4; | |
2082 | size_t rq_size, left; | |
59f082e4 SL |
2083 | int node; |
2084 | ||
2085 | node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx); | |
2086 | if (node == NUMA_NO_NODE) | |
2087 | node = set->numa_node; | |
cc71a6f4 JA |
2088 | |
2089 | INIT_LIST_HEAD(&tags->page_list); | |
2090 | ||
320ae51f JA |
2091 | /* |
2092 | * rq_size is the size of the request plus driver payload, rounded | |
2093 | * to the cacheline size | |
2094 | */ | |
24d2f903 | 2095 | rq_size = round_up(sizeof(struct request) + set->cmd_size, |
320ae51f | 2096 | cache_line_size()); |
cc71a6f4 | 2097 | left = rq_size * depth; |
320ae51f | 2098 | |
cc71a6f4 | 2099 | for (i = 0; i < depth; ) { |
320ae51f JA |
2100 | int this_order = max_order; |
2101 | struct page *page; | |
2102 | int to_do; | |
2103 | void *p; | |
2104 | ||
b3a834b1 | 2105 | while (this_order && left < order_to_size(this_order - 1)) |
320ae51f JA |
2106 | this_order--; |
2107 | ||
2108 | do { | |
59f082e4 | 2109 | page = alloc_pages_node(node, |
36e1f3d1 | 2110 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO, |
a5164405 | 2111 | this_order); |
320ae51f JA |
2112 | if (page) |
2113 | break; | |
2114 | if (!this_order--) | |
2115 | break; | |
2116 | if (order_to_size(this_order) < rq_size) | |
2117 | break; | |
2118 | } while (1); | |
2119 | ||
2120 | if (!page) | |
24d2f903 | 2121 | goto fail; |
320ae51f JA |
2122 | |
2123 | page->private = this_order; | |
24d2f903 | 2124 | list_add_tail(&page->lru, &tags->page_list); |
320ae51f JA |
2125 | |
2126 | p = page_address(page); | |
f75782e4 CM |
2127 | /* |
2128 | * Allow kmemleak to scan these pages as they contain pointers | |
2129 | * to additional allocations like via ops->init_request(). | |
2130 | */ | |
36e1f3d1 | 2131 | kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO); |
320ae51f | 2132 | entries_per_page = order_to_size(this_order) / rq_size; |
cc71a6f4 | 2133 | to_do = min(entries_per_page, depth - i); |
320ae51f JA |
2134 | left -= to_do * rq_size; |
2135 | for (j = 0; j < to_do; j++) { | |
2af8cbe3 JA |
2136 | struct request *rq = p; |
2137 | ||
2138 | tags->static_rqs[i] = rq; | |
1d9bd516 TH |
2139 | if (blk_mq_init_request(set, rq, hctx_idx, node)) { |
2140 | tags->static_rqs[i] = NULL; | |
2141 | goto fail; | |
e9b267d9 CH |
2142 | } |
2143 | ||
320ae51f JA |
2144 | p += rq_size; |
2145 | i++; | |
2146 | } | |
2147 | } | |
cc71a6f4 | 2148 | return 0; |
320ae51f | 2149 | |
24d2f903 | 2150 | fail: |
cc71a6f4 JA |
2151 | blk_mq_free_rqs(set, tags, hctx_idx); |
2152 | return -ENOMEM; | |
320ae51f JA |
2153 | } |
2154 | ||
e57690fe JA |
2155 | /* |
2156 | * 'cpu' is going away. splice any existing rq_list entries from this | |
2157 | * software queue to the hw queue dispatch list, and ensure that it | |
2158 | * gets run. | |
2159 | */ | |
9467f859 | 2160 | static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node) |
484b4061 | 2161 | { |
9467f859 | 2162 | struct blk_mq_hw_ctx *hctx; |
484b4061 JA |
2163 | struct blk_mq_ctx *ctx; |
2164 | LIST_HEAD(tmp); | |
2165 | ||
9467f859 | 2166 | hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead); |
e57690fe | 2167 | ctx = __blk_mq_get_ctx(hctx->queue, cpu); |
484b4061 JA |
2168 | |
2169 | spin_lock(&ctx->lock); | |
2170 | if (!list_empty(&ctx->rq_list)) { | |
2171 | list_splice_init(&ctx->rq_list, &tmp); | |
2172 | blk_mq_hctx_clear_pending(hctx, ctx); | |
2173 | } | |
2174 | spin_unlock(&ctx->lock); | |
2175 | ||
2176 | if (list_empty(&tmp)) | |
9467f859 | 2177 | return 0; |
484b4061 | 2178 | |
e57690fe JA |
2179 | spin_lock(&hctx->lock); |
2180 | list_splice_tail_init(&tmp, &hctx->dispatch); | |
2181 | spin_unlock(&hctx->lock); | |
484b4061 JA |
2182 | |
2183 | blk_mq_run_hw_queue(hctx, true); | |
9467f859 | 2184 | return 0; |
484b4061 JA |
2185 | } |
2186 | ||
9467f859 | 2187 | static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx) |
484b4061 | 2188 | { |
9467f859 TG |
2189 | cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD, |
2190 | &hctx->cpuhp_dead); | |
484b4061 JA |
2191 | } |
2192 | ||
c3b4afca | 2193 | /* hctx->ctxs will be freed in queue's release handler */ |
08e98fc6 ML |
2194 | static void blk_mq_exit_hctx(struct request_queue *q, |
2195 | struct blk_mq_tag_set *set, | |
2196 | struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx) | |
2197 | { | |
9c1051aa OS |
2198 | blk_mq_debugfs_unregister_hctx(hctx); |
2199 | ||
8ab0b7dc ML |
2200 | if (blk_mq_hw_queue_mapped(hctx)) |
2201 | blk_mq_tag_idle(hctx); | |
08e98fc6 | 2202 | |
f70ced09 | 2203 | if (set->ops->exit_request) |
d6296d39 | 2204 | set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx); |
f70ced09 | 2205 | |
93252632 OS |
2206 | blk_mq_sched_exit_hctx(q, hctx, hctx_idx); |
2207 | ||
08e98fc6 ML |
2208 | if (set->ops->exit_hctx) |
2209 | set->ops->exit_hctx(hctx, hctx_idx); | |
2210 | ||
6a83e74d | 2211 | if (hctx->flags & BLK_MQ_F_BLOCKING) |
05707b64 | 2212 | cleanup_srcu_struct(hctx->srcu); |
6a83e74d | 2213 | |
9467f859 | 2214 | blk_mq_remove_cpuhp(hctx); |
f70ced09 | 2215 | blk_free_flush_queue(hctx->fq); |
88459642 | 2216 | sbitmap_free(&hctx->ctx_map); |
08e98fc6 ML |
2217 | } |
2218 | ||
624dbe47 ML |
2219 | static void blk_mq_exit_hw_queues(struct request_queue *q, |
2220 | struct blk_mq_tag_set *set, int nr_queue) | |
2221 | { | |
2222 | struct blk_mq_hw_ctx *hctx; | |
2223 | unsigned int i; | |
2224 | ||
2225 | queue_for_each_hw_ctx(q, hctx, i) { | |
2226 | if (i == nr_queue) | |
2227 | break; | |
08e98fc6 | 2228 | blk_mq_exit_hctx(q, set, hctx, i); |
624dbe47 | 2229 | } |
624dbe47 ML |
2230 | } |
2231 | ||
08e98fc6 ML |
2232 | static int blk_mq_init_hctx(struct request_queue *q, |
2233 | struct blk_mq_tag_set *set, | |
2234 | struct blk_mq_hw_ctx *hctx, unsigned hctx_idx) | |
320ae51f | 2235 | { |
08e98fc6 ML |
2236 | int node; |
2237 | ||
2238 | node = hctx->numa_node; | |
2239 | if (node == NUMA_NO_NODE) | |
2240 | node = hctx->numa_node = set->numa_node; | |
2241 | ||
9f993737 | 2242 | INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn); |
08e98fc6 ML |
2243 | spin_lock_init(&hctx->lock); |
2244 | INIT_LIST_HEAD(&hctx->dispatch); | |
2245 | hctx->queue = q; | |
2404e607 | 2246 | hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED; |
08e98fc6 | 2247 | |
9467f859 | 2248 | cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead); |
08e98fc6 ML |
2249 | |
2250 | hctx->tags = set->tags[hctx_idx]; | |
320ae51f JA |
2251 | |
2252 | /* | |
08e98fc6 ML |
2253 | * Allocate space for all possible cpus to avoid allocation at |
2254 | * runtime | |
320ae51f | 2255 | */ |
d904bfa7 | 2256 | hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *), |
08e98fc6 ML |
2257 | GFP_KERNEL, node); |
2258 | if (!hctx->ctxs) | |
2259 | goto unregister_cpu_notifier; | |
320ae51f | 2260 | |
88459642 OS |
2261 | if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8), GFP_KERNEL, |
2262 | node)) | |
08e98fc6 | 2263 | goto free_ctxs; |
320ae51f | 2264 | |
08e98fc6 | 2265 | hctx->nr_ctx = 0; |
320ae51f | 2266 | |
eb619fdb JA |
2267 | init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake); |
2268 | INIT_LIST_HEAD(&hctx->dispatch_wait.entry); | |
2269 | ||
08e98fc6 ML |
2270 | if (set->ops->init_hctx && |
2271 | set->ops->init_hctx(hctx, set->driver_data, hctx_idx)) | |
2272 | goto free_bitmap; | |
320ae51f | 2273 | |
93252632 OS |
2274 | if (blk_mq_sched_init_hctx(q, hctx, hctx_idx)) |
2275 | goto exit_hctx; | |
2276 | ||
f70ced09 ML |
2277 | hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size); |
2278 | if (!hctx->fq) | |
93252632 | 2279 | goto sched_exit_hctx; |
320ae51f | 2280 | |
1d9bd516 | 2281 | if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx, node)) |
f70ced09 | 2282 | goto free_fq; |
320ae51f | 2283 | |
6a83e74d | 2284 | if (hctx->flags & BLK_MQ_F_BLOCKING) |
05707b64 | 2285 | init_srcu_struct(hctx->srcu); |
6a83e74d | 2286 | |
9c1051aa OS |
2287 | blk_mq_debugfs_register_hctx(q, hctx); |
2288 | ||
08e98fc6 | 2289 | return 0; |
320ae51f | 2290 | |
f70ced09 ML |
2291 | free_fq: |
2292 | kfree(hctx->fq); | |
93252632 OS |
2293 | sched_exit_hctx: |
2294 | blk_mq_sched_exit_hctx(q, hctx, hctx_idx); | |
f70ced09 ML |
2295 | exit_hctx: |
2296 | if (set->ops->exit_hctx) | |
2297 | set->ops->exit_hctx(hctx, hctx_idx); | |
08e98fc6 | 2298 | free_bitmap: |
88459642 | 2299 | sbitmap_free(&hctx->ctx_map); |
08e98fc6 ML |
2300 | free_ctxs: |
2301 | kfree(hctx->ctxs); | |
2302 | unregister_cpu_notifier: | |
9467f859 | 2303 | blk_mq_remove_cpuhp(hctx); |
08e98fc6 ML |
2304 | return -1; |
2305 | } | |
320ae51f | 2306 | |
320ae51f JA |
2307 | static void blk_mq_init_cpu_queues(struct request_queue *q, |
2308 | unsigned int nr_hw_queues) | |
2309 | { | |
2310 | unsigned int i; | |
2311 | ||
2312 | for_each_possible_cpu(i) { | |
2313 | struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i); | |
2314 | struct blk_mq_hw_ctx *hctx; | |
2315 | ||
320ae51f JA |
2316 | __ctx->cpu = i; |
2317 | spin_lock_init(&__ctx->lock); | |
2318 | INIT_LIST_HEAD(&__ctx->rq_list); | |
2319 | __ctx->queue = q; | |
2320 | ||
320ae51f JA |
2321 | /* |
2322 | * Set local node, IFF we have more than one hw queue. If | |
2323 | * not, we remain on the home node of the device | |
2324 | */ | |
20e4d813 | 2325 | hctx = blk_mq_map_queue(q, i); |
320ae51f | 2326 | if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE) |
bffed457 | 2327 | hctx->numa_node = local_memory_node(cpu_to_node(i)); |
320ae51f JA |
2328 | } |
2329 | } | |
2330 | ||
cc71a6f4 JA |
2331 | static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, int hctx_idx) |
2332 | { | |
2333 | int ret = 0; | |
2334 | ||
2335 | set->tags[hctx_idx] = blk_mq_alloc_rq_map(set, hctx_idx, | |
2336 | set->queue_depth, set->reserved_tags); | |
2337 | if (!set->tags[hctx_idx]) | |
2338 | return false; | |
2339 | ||
2340 | ret = blk_mq_alloc_rqs(set, set->tags[hctx_idx], hctx_idx, | |
2341 | set->queue_depth); | |
2342 | if (!ret) | |
2343 | return true; | |
2344 | ||
2345 | blk_mq_free_rq_map(set->tags[hctx_idx]); | |
2346 | set->tags[hctx_idx] = NULL; | |
2347 | return false; | |
2348 | } | |
2349 | ||
2350 | static void blk_mq_free_map_and_requests(struct blk_mq_tag_set *set, | |
2351 | unsigned int hctx_idx) | |
2352 | { | |
bd166ef1 JA |
2353 | if (set->tags[hctx_idx]) { |
2354 | blk_mq_free_rqs(set, set->tags[hctx_idx], hctx_idx); | |
2355 | blk_mq_free_rq_map(set->tags[hctx_idx]); | |
2356 | set->tags[hctx_idx] = NULL; | |
2357 | } | |
cc71a6f4 JA |
2358 | } |
2359 | ||
4b855ad3 | 2360 | static void blk_mq_map_swqueue(struct request_queue *q) |
320ae51f | 2361 | { |
4412efec | 2362 | unsigned int i, hctx_idx; |
320ae51f JA |
2363 | struct blk_mq_hw_ctx *hctx; |
2364 | struct blk_mq_ctx *ctx; | |
2a34c087 | 2365 | struct blk_mq_tag_set *set = q->tag_set; |
320ae51f | 2366 | |
60de074b AM |
2367 | /* |
2368 | * Avoid others reading imcomplete hctx->cpumask through sysfs | |
2369 | */ | |
2370 | mutex_lock(&q->sysfs_lock); | |
2371 | ||
320ae51f | 2372 | queue_for_each_hw_ctx(q, hctx, i) { |
e4043dcf | 2373 | cpumask_clear(hctx->cpumask); |
320ae51f JA |
2374 | hctx->nr_ctx = 0; |
2375 | } | |
2376 | ||
2377 | /* | |
4b855ad3 | 2378 | * Map software to hardware queues. |
4412efec ML |
2379 | * |
2380 | * If the cpu isn't present, the cpu is mapped to first hctx. | |
320ae51f | 2381 | */ |
20e4d813 | 2382 | for_each_possible_cpu(i) { |
4412efec ML |
2383 | hctx_idx = q->mq_map[i]; |
2384 | /* unmapped hw queue can be remapped after CPU topo changed */ | |
2385 | if (!set->tags[hctx_idx] && | |
2386 | !__blk_mq_alloc_rq_map(set, hctx_idx)) { | |
2387 | /* | |
2388 | * If tags initialization fail for some hctx, | |
2389 | * that hctx won't be brought online. In this | |
2390 | * case, remap the current ctx to hctx[0] which | |
2391 | * is guaranteed to always have tags allocated | |
2392 | */ | |
2393 | q->mq_map[i] = 0; | |
2394 | } | |
2395 | ||
897bb0c7 | 2396 | ctx = per_cpu_ptr(q->queue_ctx, i); |
7d7e0f90 | 2397 | hctx = blk_mq_map_queue(q, i); |
868f2f0b | 2398 | |
e4043dcf | 2399 | cpumask_set_cpu(i, hctx->cpumask); |
320ae51f JA |
2400 | ctx->index_hw = hctx->nr_ctx; |
2401 | hctx->ctxs[hctx->nr_ctx++] = ctx; | |
2402 | } | |
506e931f | 2403 | |
60de074b AM |
2404 | mutex_unlock(&q->sysfs_lock); |
2405 | ||
506e931f | 2406 | queue_for_each_hw_ctx(q, hctx, i) { |
4412efec ML |
2407 | /* |
2408 | * If no software queues are mapped to this hardware queue, | |
2409 | * disable it and free the request entries. | |
2410 | */ | |
2411 | if (!hctx->nr_ctx) { | |
2412 | /* Never unmap queue 0. We need it as a | |
2413 | * fallback in case of a new remap fails | |
2414 | * allocation | |
2415 | */ | |
2416 | if (i && set->tags[i]) | |
2417 | blk_mq_free_map_and_requests(set, i); | |
2418 | ||
2419 | hctx->tags = NULL; | |
2420 | continue; | |
2421 | } | |
484b4061 | 2422 | |
2a34c087 ML |
2423 | hctx->tags = set->tags[i]; |
2424 | WARN_ON(!hctx->tags); | |
2425 | ||
889fa31f CY |
2426 | /* |
2427 | * Set the map size to the number of mapped software queues. | |
2428 | * This is more accurate and more efficient than looping | |
2429 | * over all possibly mapped software queues. | |
2430 | */ | |
88459642 | 2431 | sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx); |
889fa31f | 2432 | |
484b4061 JA |
2433 | /* |
2434 | * Initialize batch roundrobin counts | |
2435 | */ | |
f82ddf19 | 2436 | hctx->next_cpu = blk_mq_first_mapped_cpu(hctx); |
506e931f JA |
2437 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; |
2438 | } | |
320ae51f JA |
2439 | } |
2440 | ||
8e8320c9 JA |
2441 | /* |
2442 | * Caller needs to ensure that we're either frozen/quiesced, or that | |
2443 | * the queue isn't live yet. | |
2444 | */ | |
2404e607 | 2445 | static void queue_set_hctx_shared(struct request_queue *q, bool shared) |
0d2602ca JA |
2446 | { |
2447 | struct blk_mq_hw_ctx *hctx; | |
0d2602ca JA |
2448 | int i; |
2449 | ||
2404e607 | 2450 | queue_for_each_hw_ctx(q, hctx, i) { |
8e8320c9 JA |
2451 | if (shared) { |
2452 | if (test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state)) | |
2453 | atomic_inc(&q->shared_hctx_restart); | |
2404e607 | 2454 | hctx->flags |= BLK_MQ_F_TAG_SHARED; |
8e8320c9 JA |
2455 | } else { |
2456 | if (test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state)) | |
2457 | atomic_dec(&q->shared_hctx_restart); | |
2404e607 | 2458 | hctx->flags &= ~BLK_MQ_F_TAG_SHARED; |
8e8320c9 | 2459 | } |
2404e607 JM |
2460 | } |
2461 | } | |
2462 | ||
8e8320c9 JA |
2463 | static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set, |
2464 | bool shared) | |
2404e607 JM |
2465 | { |
2466 | struct request_queue *q; | |
0d2602ca | 2467 | |
705cda97 BVA |
2468 | lockdep_assert_held(&set->tag_list_lock); |
2469 | ||
0d2602ca JA |
2470 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
2471 | blk_mq_freeze_queue(q); | |
2404e607 | 2472 | queue_set_hctx_shared(q, shared); |
0d2602ca JA |
2473 | blk_mq_unfreeze_queue(q); |
2474 | } | |
2475 | } | |
2476 | ||
2477 | static void blk_mq_del_queue_tag_set(struct request_queue *q) | |
2478 | { | |
2479 | struct blk_mq_tag_set *set = q->tag_set; | |
2480 | ||
0d2602ca | 2481 | mutex_lock(&set->tag_list_lock); |
705cda97 BVA |
2482 | list_del_rcu(&q->tag_set_list); |
2483 | INIT_LIST_HEAD(&q->tag_set_list); | |
2404e607 JM |
2484 | if (list_is_singular(&set->tag_list)) { |
2485 | /* just transitioned to unshared */ | |
2486 | set->flags &= ~BLK_MQ_F_TAG_SHARED; | |
2487 | /* update existing queue */ | |
2488 | blk_mq_update_tag_set_depth(set, false); | |
2489 | } | |
0d2602ca | 2490 | mutex_unlock(&set->tag_list_lock); |
705cda97 BVA |
2491 | |
2492 | synchronize_rcu(); | |
0d2602ca JA |
2493 | } |
2494 | ||
2495 | static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set, | |
2496 | struct request_queue *q) | |
2497 | { | |
2498 | q->tag_set = set; | |
2499 | ||
2500 | mutex_lock(&set->tag_list_lock); | |
2404e607 | 2501 | |
ff821d27 JA |
2502 | /* |
2503 | * Check to see if we're transitioning to shared (from 1 to 2 queues). | |
2504 | */ | |
2505 | if (!list_empty(&set->tag_list) && | |
2506 | !(set->flags & BLK_MQ_F_TAG_SHARED)) { | |
2404e607 JM |
2507 | set->flags |= BLK_MQ_F_TAG_SHARED; |
2508 | /* update existing queue */ | |
2509 | blk_mq_update_tag_set_depth(set, true); | |
2510 | } | |
2511 | if (set->flags & BLK_MQ_F_TAG_SHARED) | |
2512 | queue_set_hctx_shared(q, true); | |
705cda97 | 2513 | list_add_tail_rcu(&q->tag_set_list, &set->tag_list); |
2404e607 | 2514 | |
0d2602ca JA |
2515 | mutex_unlock(&set->tag_list_lock); |
2516 | } | |
2517 | ||
e09aae7e ML |
2518 | /* |
2519 | * It is the actual release handler for mq, but we do it from | |
2520 | * request queue's release handler for avoiding use-after-free | |
2521 | * and headache because q->mq_kobj shouldn't have been introduced, | |
2522 | * but we can't group ctx/kctx kobj without it. | |
2523 | */ | |
2524 | void blk_mq_release(struct request_queue *q) | |
2525 | { | |
2526 | struct blk_mq_hw_ctx *hctx; | |
2527 | unsigned int i; | |
2528 | ||
2529 | /* hctx kobj stays in hctx */ | |
c3b4afca ML |
2530 | queue_for_each_hw_ctx(q, hctx, i) { |
2531 | if (!hctx) | |
2532 | continue; | |
6c8b232e | 2533 | kobject_put(&hctx->kobj); |
c3b4afca | 2534 | } |
e09aae7e | 2535 | |
a723bab3 AM |
2536 | q->mq_map = NULL; |
2537 | ||
e09aae7e ML |
2538 | kfree(q->queue_hw_ctx); |
2539 | ||
7ea5fe31 ML |
2540 | /* |
2541 | * release .mq_kobj and sw queue's kobject now because | |
2542 | * both share lifetime with request queue. | |
2543 | */ | |
2544 | blk_mq_sysfs_deinit(q); | |
2545 | ||
e09aae7e ML |
2546 | free_percpu(q->queue_ctx); |
2547 | } | |
2548 | ||
24d2f903 | 2549 | struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set) |
b62c21b7 MS |
2550 | { |
2551 | struct request_queue *uninit_q, *q; | |
2552 | ||
5ee0524b | 2553 | uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node, NULL); |
b62c21b7 MS |
2554 | if (!uninit_q) |
2555 | return ERR_PTR(-ENOMEM); | |
2556 | ||
2557 | q = blk_mq_init_allocated_queue(set, uninit_q); | |
2558 | if (IS_ERR(q)) | |
2559 | blk_cleanup_queue(uninit_q); | |
2560 | ||
2561 | return q; | |
2562 | } | |
2563 | EXPORT_SYMBOL(blk_mq_init_queue); | |
2564 | ||
07319678 BVA |
2565 | static int blk_mq_hw_ctx_size(struct blk_mq_tag_set *tag_set) |
2566 | { | |
2567 | int hw_ctx_size = sizeof(struct blk_mq_hw_ctx); | |
2568 | ||
05707b64 | 2569 | BUILD_BUG_ON(ALIGN(offsetof(struct blk_mq_hw_ctx, srcu), |
07319678 BVA |
2570 | __alignof__(struct blk_mq_hw_ctx)) != |
2571 | sizeof(struct blk_mq_hw_ctx)); | |
2572 | ||
2573 | if (tag_set->flags & BLK_MQ_F_BLOCKING) | |
2574 | hw_ctx_size += sizeof(struct srcu_struct); | |
2575 | ||
2576 | return hw_ctx_size; | |
2577 | } | |
2578 | ||
868f2f0b KB |
2579 | static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set, |
2580 | struct request_queue *q) | |
320ae51f | 2581 | { |
868f2f0b KB |
2582 | int i, j; |
2583 | struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx; | |
f14bbe77 | 2584 | |
868f2f0b | 2585 | blk_mq_sysfs_unregister(q); |
fb350e0a ML |
2586 | |
2587 | /* protect against switching io scheduler */ | |
2588 | mutex_lock(&q->sysfs_lock); | |
24d2f903 | 2589 | for (i = 0; i < set->nr_hw_queues; i++) { |
868f2f0b | 2590 | int node; |
f14bbe77 | 2591 | |
868f2f0b KB |
2592 | if (hctxs[i]) |
2593 | continue; | |
2594 | ||
2595 | node = blk_mq_hw_queue_to_node(q->mq_map, i); | |
07319678 | 2596 | hctxs[i] = kzalloc_node(blk_mq_hw_ctx_size(set), |
cdef54dd | 2597 | GFP_KERNEL, node); |
320ae51f | 2598 | if (!hctxs[i]) |
868f2f0b | 2599 | break; |
320ae51f | 2600 | |
a86073e4 | 2601 | if (!zalloc_cpumask_var_node(&hctxs[i]->cpumask, GFP_KERNEL, |
868f2f0b KB |
2602 | node)) { |
2603 | kfree(hctxs[i]); | |
2604 | hctxs[i] = NULL; | |
2605 | break; | |
2606 | } | |
e4043dcf | 2607 | |
0d2602ca | 2608 | atomic_set(&hctxs[i]->nr_active, 0); |
f14bbe77 | 2609 | hctxs[i]->numa_node = node; |
320ae51f | 2610 | hctxs[i]->queue_num = i; |
868f2f0b KB |
2611 | |
2612 | if (blk_mq_init_hctx(q, set, hctxs[i], i)) { | |
2613 | free_cpumask_var(hctxs[i]->cpumask); | |
2614 | kfree(hctxs[i]); | |
2615 | hctxs[i] = NULL; | |
2616 | break; | |
2617 | } | |
2618 | blk_mq_hctx_kobj_init(hctxs[i]); | |
320ae51f | 2619 | } |
868f2f0b KB |
2620 | for (j = i; j < q->nr_hw_queues; j++) { |
2621 | struct blk_mq_hw_ctx *hctx = hctxs[j]; | |
2622 | ||
2623 | if (hctx) { | |
cc71a6f4 JA |
2624 | if (hctx->tags) |
2625 | blk_mq_free_map_and_requests(set, j); | |
868f2f0b | 2626 | blk_mq_exit_hctx(q, set, hctx, j); |
868f2f0b | 2627 | kobject_put(&hctx->kobj); |
868f2f0b KB |
2628 | hctxs[j] = NULL; |
2629 | ||
2630 | } | |
2631 | } | |
2632 | q->nr_hw_queues = i; | |
fb350e0a | 2633 | mutex_unlock(&q->sysfs_lock); |
868f2f0b KB |
2634 | blk_mq_sysfs_register(q); |
2635 | } | |
2636 | ||
2637 | struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set, | |
2638 | struct request_queue *q) | |
2639 | { | |
66841672 ML |
2640 | /* mark the queue as mq asap */ |
2641 | q->mq_ops = set->ops; | |
2642 | ||
34dbad5d | 2643 | q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn, |
720b8ccc SB |
2644 | blk_mq_poll_stats_bkt, |
2645 | BLK_MQ_POLL_STATS_BKTS, q); | |
34dbad5d OS |
2646 | if (!q->poll_cb) |
2647 | goto err_exit; | |
2648 | ||
868f2f0b KB |
2649 | q->queue_ctx = alloc_percpu(struct blk_mq_ctx); |
2650 | if (!q->queue_ctx) | |
c7de5726 | 2651 | goto err_exit; |
868f2f0b | 2652 | |
737f98cf ML |
2653 | /* init q->mq_kobj and sw queues' kobjects */ |
2654 | blk_mq_sysfs_init(q); | |
2655 | ||
868f2f0b KB |
2656 | q->queue_hw_ctx = kzalloc_node(nr_cpu_ids * sizeof(*(q->queue_hw_ctx)), |
2657 | GFP_KERNEL, set->numa_node); | |
2658 | if (!q->queue_hw_ctx) | |
2659 | goto err_percpu; | |
2660 | ||
bdd17e75 | 2661 | q->mq_map = set->mq_map; |
868f2f0b KB |
2662 | |
2663 | blk_mq_realloc_hw_ctxs(set, q); | |
2664 | if (!q->nr_hw_queues) | |
2665 | goto err_hctxs; | |
320ae51f | 2666 | |
287922eb | 2667 | INIT_WORK(&q->timeout_work, blk_mq_timeout_work); |
e56f698b | 2668 | blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ); |
320ae51f JA |
2669 | |
2670 | q->nr_queues = nr_cpu_ids; | |
320ae51f | 2671 | |
94eddfbe | 2672 | q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT; |
320ae51f | 2673 | |
05f1dd53 | 2674 | if (!(set->flags & BLK_MQ_F_SG_MERGE)) |
f78bac2c | 2675 | queue_flag_set_unlocked(QUEUE_FLAG_NO_SG_MERGE, q); |
05f1dd53 | 2676 | |
1be036e9 CH |
2677 | q->sg_reserved_size = INT_MAX; |
2678 | ||
2849450a | 2679 | INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work); |
6fca6a61 CH |
2680 | INIT_LIST_HEAD(&q->requeue_list); |
2681 | spin_lock_init(&q->requeue_lock); | |
2682 | ||
254d259d | 2683 | blk_queue_make_request(q, blk_mq_make_request); |
ea435e1b CH |
2684 | if (q->mq_ops->poll) |
2685 | q->poll_fn = blk_mq_poll; | |
07068d5b | 2686 | |
eba71768 JA |
2687 | /* |
2688 | * Do this after blk_queue_make_request() overrides it... | |
2689 | */ | |
2690 | q->nr_requests = set->queue_depth; | |
2691 | ||
64f1c21e JA |
2692 | /* |
2693 | * Default to classic polling | |
2694 | */ | |
2695 | q->poll_nsec = -1; | |
2696 | ||
24d2f903 CH |
2697 | if (set->ops->complete) |
2698 | blk_queue_softirq_done(q, set->ops->complete); | |
30a91cb4 | 2699 | |
24d2f903 | 2700 | blk_mq_init_cpu_queues(q, set->nr_hw_queues); |
0d2602ca | 2701 | blk_mq_add_queue_tag_set(set, q); |
4b855ad3 | 2702 | blk_mq_map_swqueue(q); |
4593fdbe | 2703 | |
d3484991 JA |
2704 | if (!(set->flags & BLK_MQ_F_NO_SCHED)) { |
2705 | int ret; | |
2706 | ||
2707 | ret = blk_mq_sched_init(q); | |
2708 | if (ret) | |
2709 | return ERR_PTR(ret); | |
2710 | } | |
2711 | ||
320ae51f | 2712 | return q; |
18741986 | 2713 | |
320ae51f | 2714 | err_hctxs: |
868f2f0b | 2715 | kfree(q->queue_hw_ctx); |
320ae51f | 2716 | err_percpu: |
868f2f0b | 2717 | free_percpu(q->queue_ctx); |
c7de5726 ML |
2718 | err_exit: |
2719 | q->mq_ops = NULL; | |
320ae51f JA |
2720 | return ERR_PTR(-ENOMEM); |
2721 | } | |
b62c21b7 | 2722 | EXPORT_SYMBOL(blk_mq_init_allocated_queue); |
320ae51f JA |
2723 | |
2724 | void blk_mq_free_queue(struct request_queue *q) | |
2725 | { | |
624dbe47 | 2726 | struct blk_mq_tag_set *set = q->tag_set; |
320ae51f | 2727 | |
0d2602ca | 2728 | blk_mq_del_queue_tag_set(q); |
624dbe47 | 2729 | blk_mq_exit_hw_queues(q, set, set->nr_hw_queues); |
320ae51f | 2730 | } |
320ae51f JA |
2731 | |
2732 | /* Basically redo blk_mq_init_queue with queue frozen */ | |
4b855ad3 | 2733 | static void blk_mq_queue_reinit(struct request_queue *q) |
320ae51f | 2734 | { |
4ecd4fef | 2735 | WARN_ON_ONCE(!atomic_read(&q->mq_freeze_depth)); |
320ae51f | 2736 | |
9c1051aa | 2737 | blk_mq_debugfs_unregister_hctxs(q); |
67aec14c JA |
2738 | blk_mq_sysfs_unregister(q); |
2739 | ||
320ae51f JA |
2740 | /* |
2741 | * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe | |
ff821d27 JA |
2742 | * we should change hctx numa_node according to the new topology (this |
2743 | * involves freeing and re-allocating memory, worth doing?) | |
320ae51f | 2744 | */ |
4b855ad3 | 2745 | blk_mq_map_swqueue(q); |
320ae51f | 2746 | |
67aec14c | 2747 | blk_mq_sysfs_register(q); |
9c1051aa | 2748 | blk_mq_debugfs_register_hctxs(q); |
320ae51f JA |
2749 | } |
2750 | ||
a5164405 JA |
2751 | static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set) |
2752 | { | |
2753 | int i; | |
2754 | ||
cc71a6f4 JA |
2755 | for (i = 0; i < set->nr_hw_queues; i++) |
2756 | if (!__blk_mq_alloc_rq_map(set, i)) | |
a5164405 | 2757 | goto out_unwind; |
a5164405 JA |
2758 | |
2759 | return 0; | |
2760 | ||
2761 | out_unwind: | |
2762 | while (--i >= 0) | |
cc71a6f4 | 2763 | blk_mq_free_rq_map(set->tags[i]); |
a5164405 | 2764 | |
a5164405 JA |
2765 | return -ENOMEM; |
2766 | } | |
2767 | ||
2768 | /* | |
2769 | * Allocate the request maps associated with this tag_set. Note that this | |
2770 | * may reduce the depth asked for, if memory is tight. set->queue_depth | |
2771 | * will be updated to reflect the allocated depth. | |
2772 | */ | |
2773 | static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set) | |
2774 | { | |
2775 | unsigned int depth; | |
2776 | int err; | |
2777 | ||
2778 | depth = set->queue_depth; | |
2779 | do { | |
2780 | err = __blk_mq_alloc_rq_maps(set); | |
2781 | if (!err) | |
2782 | break; | |
2783 | ||
2784 | set->queue_depth >>= 1; | |
2785 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) { | |
2786 | err = -ENOMEM; | |
2787 | break; | |
2788 | } | |
2789 | } while (set->queue_depth); | |
2790 | ||
2791 | if (!set->queue_depth || err) { | |
2792 | pr_err("blk-mq: failed to allocate request map\n"); | |
2793 | return -ENOMEM; | |
2794 | } | |
2795 | ||
2796 | if (depth != set->queue_depth) | |
2797 | pr_info("blk-mq: reduced tag depth (%u -> %u)\n", | |
2798 | depth, set->queue_depth); | |
2799 | ||
2800 | return 0; | |
2801 | } | |
2802 | ||
ebe8bddb OS |
2803 | static int blk_mq_update_queue_map(struct blk_mq_tag_set *set) |
2804 | { | |
7d4901a9 ML |
2805 | if (set->ops->map_queues) { |
2806 | int cpu; | |
2807 | /* | |
2808 | * transport .map_queues is usually done in the following | |
2809 | * way: | |
2810 | * | |
2811 | * for (queue = 0; queue < set->nr_hw_queues; queue++) { | |
2812 | * mask = get_cpu_mask(queue) | |
2813 | * for_each_cpu(cpu, mask) | |
2814 | * set->mq_map[cpu] = queue; | |
2815 | * } | |
2816 | * | |
2817 | * When we need to remap, the table has to be cleared for | |
2818 | * killing stale mapping since one CPU may not be mapped | |
2819 | * to any hw queue. | |
2820 | */ | |
2821 | for_each_possible_cpu(cpu) | |
2822 | set->mq_map[cpu] = 0; | |
2823 | ||
ebe8bddb | 2824 | return set->ops->map_queues(set); |
7d4901a9 | 2825 | } else |
ebe8bddb OS |
2826 | return blk_mq_map_queues(set); |
2827 | } | |
2828 | ||
a4391c64 JA |
2829 | /* |
2830 | * Alloc a tag set to be associated with one or more request queues. | |
2831 | * May fail with EINVAL for various error conditions. May adjust the | |
2832 | * requested depth down, if if it too large. In that case, the set | |
2833 | * value will be stored in set->queue_depth. | |
2834 | */ | |
24d2f903 CH |
2835 | int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set) |
2836 | { | |
da695ba2 CH |
2837 | int ret; |
2838 | ||
205fb5f5 BVA |
2839 | BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS); |
2840 | ||
24d2f903 CH |
2841 | if (!set->nr_hw_queues) |
2842 | return -EINVAL; | |
a4391c64 | 2843 | if (!set->queue_depth) |
24d2f903 CH |
2844 | return -EINVAL; |
2845 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) | |
2846 | return -EINVAL; | |
2847 | ||
7d7e0f90 | 2848 | if (!set->ops->queue_rq) |
24d2f903 CH |
2849 | return -EINVAL; |
2850 | ||
de148297 ML |
2851 | if (!set->ops->get_budget ^ !set->ops->put_budget) |
2852 | return -EINVAL; | |
2853 | ||
a4391c64 JA |
2854 | if (set->queue_depth > BLK_MQ_MAX_DEPTH) { |
2855 | pr_info("blk-mq: reduced tag depth to %u\n", | |
2856 | BLK_MQ_MAX_DEPTH); | |
2857 | set->queue_depth = BLK_MQ_MAX_DEPTH; | |
2858 | } | |
24d2f903 | 2859 | |
6637fadf SL |
2860 | /* |
2861 | * If a crashdump is active, then we are potentially in a very | |
2862 | * memory constrained environment. Limit us to 1 queue and | |
2863 | * 64 tags to prevent using too much memory. | |
2864 | */ | |
2865 | if (is_kdump_kernel()) { | |
2866 | set->nr_hw_queues = 1; | |
2867 | set->queue_depth = min(64U, set->queue_depth); | |
2868 | } | |
868f2f0b KB |
2869 | /* |
2870 | * There is no use for more h/w queues than cpus. | |
2871 | */ | |
2872 | if (set->nr_hw_queues > nr_cpu_ids) | |
2873 | set->nr_hw_queues = nr_cpu_ids; | |
6637fadf | 2874 | |
868f2f0b | 2875 | set->tags = kzalloc_node(nr_cpu_ids * sizeof(struct blk_mq_tags *), |
24d2f903 CH |
2876 | GFP_KERNEL, set->numa_node); |
2877 | if (!set->tags) | |
a5164405 | 2878 | return -ENOMEM; |
24d2f903 | 2879 | |
da695ba2 CH |
2880 | ret = -ENOMEM; |
2881 | set->mq_map = kzalloc_node(sizeof(*set->mq_map) * nr_cpu_ids, | |
2882 | GFP_KERNEL, set->numa_node); | |
bdd17e75 CH |
2883 | if (!set->mq_map) |
2884 | goto out_free_tags; | |
2885 | ||
ebe8bddb | 2886 | ret = blk_mq_update_queue_map(set); |
da695ba2 CH |
2887 | if (ret) |
2888 | goto out_free_mq_map; | |
2889 | ||
2890 | ret = blk_mq_alloc_rq_maps(set); | |
2891 | if (ret) | |
bdd17e75 | 2892 | goto out_free_mq_map; |
24d2f903 | 2893 | |
0d2602ca JA |
2894 | mutex_init(&set->tag_list_lock); |
2895 | INIT_LIST_HEAD(&set->tag_list); | |
2896 | ||
24d2f903 | 2897 | return 0; |
bdd17e75 CH |
2898 | |
2899 | out_free_mq_map: | |
2900 | kfree(set->mq_map); | |
2901 | set->mq_map = NULL; | |
2902 | out_free_tags: | |
5676e7b6 RE |
2903 | kfree(set->tags); |
2904 | set->tags = NULL; | |
da695ba2 | 2905 | return ret; |
24d2f903 CH |
2906 | } |
2907 | EXPORT_SYMBOL(blk_mq_alloc_tag_set); | |
2908 | ||
2909 | void blk_mq_free_tag_set(struct blk_mq_tag_set *set) | |
2910 | { | |
2911 | int i; | |
2912 | ||
cc71a6f4 JA |
2913 | for (i = 0; i < nr_cpu_ids; i++) |
2914 | blk_mq_free_map_and_requests(set, i); | |
484b4061 | 2915 | |
bdd17e75 CH |
2916 | kfree(set->mq_map); |
2917 | set->mq_map = NULL; | |
2918 | ||
981bd189 | 2919 | kfree(set->tags); |
5676e7b6 | 2920 | set->tags = NULL; |
24d2f903 CH |
2921 | } |
2922 | EXPORT_SYMBOL(blk_mq_free_tag_set); | |
2923 | ||
e3a2b3f9 JA |
2924 | int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr) |
2925 | { | |
2926 | struct blk_mq_tag_set *set = q->tag_set; | |
2927 | struct blk_mq_hw_ctx *hctx; | |
2928 | int i, ret; | |
2929 | ||
bd166ef1 | 2930 | if (!set) |
e3a2b3f9 JA |
2931 | return -EINVAL; |
2932 | ||
70f36b60 | 2933 | blk_mq_freeze_queue(q); |
24f5a90f | 2934 | blk_mq_quiesce_queue(q); |
70f36b60 | 2935 | |
e3a2b3f9 JA |
2936 | ret = 0; |
2937 | queue_for_each_hw_ctx(q, hctx, i) { | |
e9137d4b KB |
2938 | if (!hctx->tags) |
2939 | continue; | |
bd166ef1 JA |
2940 | /* |
2941 | * If we're using an MQ scheduler, just update the scheduler | |
2942 | * queue depth. This is similar to what the old code would do. | |
2943 | */ | |
70f36b60 | 2944 | if (!hctx->sched_tags) { |
c2e82a23 | 2945 | ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr, |
70f36b60 JA |
2946 | false); |
2947 | } else { | |
2948 | ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags, | |
2949 | nr, true); | |
2950 | } | |
e3a2b3f9 JA |
2951 | if (ret) |
2952 | break; | |
2953 | } | |
2954 | ||
2955 | if (!ret) | |
2956 | q->nr_requests = nr; | |
2957 | ||
24f5a90f | 2958 | blk_mq_unquiesce_queue(q); |
70f36b60 | 2959 | blk_mq_unfreeze_queue(q); |
70f36b60 | 2960 | |
e3a2b3f9 JA |
2961 | return ret; |
2962 | } | |
2963 | ||
e4dc2b32 KB |
2964 | static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, |
2965 | int nr_hw_queues) | |
868f2f0b KB |
2966 | { |
2967 | struct request_queue *q; | |
2968 | ||
705cda97 BVA |
2969 | lockdep_assert_held(&set->tag_list_lock); |
2970 | ||
868f2f0b KB |
2971 | if (nr_hw_queues > nr_cpu_ids) |
2972 | nr_hw_queues = nr_cpu_ids; | |
2973 | if (nr_hw_queues < 1 || nr_hw_queues == set->nr_hw_queues) | |
2974 | return; | |
2975 | ||
2976 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
2977 | blk_mq_freeze_queue(q); | |
2978 | ||
2979 | set->nr_hw_queues = nr_hw_queues; | |
ebe8bddb | 2980 | blk_mq_update_queue_map(set); |
868f2f0b KB |
2981 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
2982 | blk_mq_realloc_hw_ctxs(set, q); | |
4b855ad3 | 2983 | blk_mq_queue_reinit(q); |
868f2f0b KB |
2984 | } |
2985 | ||
2986 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
2987 | blk_mq_unfreeze_queue(q); | |
2988 | } | |
e4dc2b32 KB |
2989 | |
2990 | void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues) | |
2991 | { | |
2992 | mutex_lock(&set->tag_list_lock); | |
2993 | __blk_mq_update_nr_hw_queues(set, nr_hw_queues); | |
2994 | mutex_unlock(&set->tag_list_lock); | |
2995 | } | |
868f2f0b KB |
2996 | EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues); |
2997 | ||
34dbad5d OS |
2998 | /* Enable polling stats and return whether they were already enabled. */ |
2999 | static bool blk_poll_stats_enable(struct request_queue *q) | |
3000 | { | |
3001 | if (test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) || | |
7dfdbc73 | 3002 | blk_queue_flag_test_and_set(QUEUE_FLAG_POLL_STATS, q)) |
34dbad5d OS |
3003 | return true; |
3004 | blk_stat_add_callback(q, q->poll_cb); | |
3005 | return false; | |
3006 | } | |
3007 | ||
3008 | static void blk_mq_poll_stats_start(struct request_queue *q) | |
3009 | { | |
3010 | /* | |
3011 | * We don't arm the callback if polling stats are not enabled or the | |
3012 | * callback is already active. | |
3013 | */ | |
3014 | if (!test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) || | |
3015 | blk_stat_is_active(q->poll_cb)) | |
3016 | return; | |
3017 | ||
3018 | blk_stat_activate_msecs(q->poll_cb, 100); | |
3019 | } | |
3020 | ||
3021 | static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb) | |
3022 | { | |
3023 | struct request_queue *q = cb->data; | |
720b8ccc | 3024 | int bucket; |
34dbad5d | 3025 | |
720b8ccc SB |
3026 | for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) { |
3027 | if (cb->stat[bucket].nr_samples) | |
3028 | q->poll_stat[bucket] = cb->stat[bucket]; | |
3029 | } | |
34dbad5d OS |
3030 | } |
3031 | ||
64f1c21e JA |
3032 | static unsigned long blk_mq_poll_nsecs(struct request_queue *q, |
3033 | struct blk_mq_hw_ctx *hctx, | |
3034 | struct request *rq) | |
3035 | { | |
64f1c21e | 3036 | unsigned long ret = 0; |
720b8ccc | 3037 | int bucket; |
64f1c21e JA |
3038 | |
3039 | /* | |
3040 | * If stats collection isn't on, don't sleep but turn it on for | |
3041 | * future users | |
3042 | */ | |
34dbad5d | 3043 | if (!blk_poll_stats_enable(q)) |
64f1c21e JA |
3044 | return 0; |
3045 | ||
64f1c21e JA |
3046 | /* |
3047 | * As an optimistic guess, use half of the mean service time | |
3048 | * for this type of request. We can (and should) make this smarter. | |
3049 | * For instance, if the completion latencies are tight, we can | |
3050 | * get closer than just half the mean. This is especially | |
3051 | * important on devices where the completion latencies are longer | |
720b8ccc SB |
3052 | * than ~10 usec. We do use the stats for the relevant IO size |
3053 | * if available which does lead to better estimates. | |
64f1c21e | 3054 | */ |
720b8ccc SB |
3055 | bucket = blk_mq_poll_stats_bkt(rq); |
3056 | if (bucket < 0) | |
3057 | return ret; | |
3058 | ||
3059 | if (q->poll_stat[bucket].nr_samples) | |
3060 | ret = (q->poll_stat[bucket].mean + 1) / 2; | |
64f1c21e JA |
3061 | |
3062 | return ret; | |
3063 | } | |
3064 | ||
06426adf | 3065 | static bool blk_mq_poll_hybrid_sleep(struct request_queue *q, |
64f1c21e | 3066 | struct blk_mq_hw_ctx *hctx, |
06426adf JA |
3067 | struct request *rq) |
3068 | { | |
3069 | struct hrtimer_sleeper hs; | |
3070 | enum hrtimer_mode mode; | |
64f1c21e | 3071 | unsigned int nsecs; |
06426adf JA |
3072 | ktime_t kt; |
3073 | ||
76a86f9d | 3074 | if (rq->rq_flags & RQF_MQ_POLL_SLEPT) |
64f1c21e JA |
3075 | return false; |
3076 | ||
3077 | /* | |
3078 | * poll_nsec can be: | |
3079 | * | |
3080 | * -1: don't ever hybrid sleep | |
3081 | * 0: use half of prev avg | |
3082 | * >0: use this specific value | |
3083 | */ | |
3084 | if (q->poll_nsec == -1) | |
3085 | return false; | |
3086 | else if (q->poll_nsec > 0) | |
3087 | nsecs = q->poll_nsec; | |
3088 | else | |
3089 | nsecs = blk_mq_poll_nsecs(q, hctx, rq); | |
3090 | ||
3091 | if (!nsecs) | |
06426adf JA |
3092 | return false; |
3093 | ||
76a86f9d | 3094 | rq->rq_flags |= RQF_MQ_POLL_SLEPT; |
06426adf JA |
3095 | |
3096 | /* | |
3097 | * This will be replaced with the stats tracking code, using | |
3098 | * 'avg_completion_time / 2' as the pre-sleep target. | |
3099 | */ | |
8b0e1953 | 3100 | kt = nsecs; |
06426adf JA |
3101 | |
3102 | mode = HRTIMER_MODE_REL; | |
3103 | hrtimer_init_on_stack(&hs.timer, CLOCK_MONOTONIC, mode); | |
3104 | hrtimer_set_expires(&hs.timer, kt); | |
3105 | ||
3106 | hrtimer_init_sleeper(&hs, current); | |
3107 | do { | |
5a61c363 | 3108 | if (blk_mq_rq_state(rq) == MQ_RQ_COMPLETE) |
06426adf JA |
3109 | break; |
3110 | set_current_state(TASK_UNINTERRUPTIBLE); | |
3111 | hrtimer_start_expires(&hs.timer, mode); | |
3112 | if (hs.task) | |
3113 | io_schedule(); | |
3114 | hrtimer_cancel(&hs.timer); | |
3115 | mode = HRTIMER_MODE_ABS; | |
3116 | } while (hs.task && !signal_pending(current)); | |
3117 | ||
3118 | __set_current_state(TASK_RUNNING); | |
3119 | destroy_hrtimer_on_stack(&hs.timer); | |
3120 | return true; | |
3121 | } | |
3122 | ||
bbd7bb70 JA |
3123 | static bool __blk_mq_poll(struct blk_mq_hw_ctx *hctx, struct request *rq) |
3124 | { | |
3125 | struct request_queue *q = hctx->queue; | |
3126 | long state; | |
3127 | ||
06426adf JA |
3128 | /* |
3129 | * If we sleep, have the caller restart the poll loop to reset | |
3130 | * the state. Like for the other success return cases, the | |
3131 | * caller is responsible for checking if the IO completed. If | |
3132 | * the IO isn't complete, we'll get called again and will go | |
3133 | * straight to the busy poll loop. | |
3134 | */ | |
64f1c21e | 3135 | if (blk_mq_poll_hybrid_sleep(q, hctx, rq)) |
06426adf JA |
3136 | return true; |
3137 | ||
bbd7bb70 JA |
3138 | hctx->poll_considered++; |
3139 | ||
3140 | state = current->state; | |
3141 | while (!need_resched()) { | |
3142 | int ret; | |
3143 | ||
3144 | hctx->poll_invoked++; | |
3145 | ||
3146 | ret = q->mq_ops->poll(hctx, rq->tag); | |
3147 | if (ret > 0) { | |
3148 | hctx->poll_success++; | |
3149 | set_current_state(TASK_RUNNING); | |
3150 | return true; | |
3151 | } | |
3152 | ||
3153 | if (signal_pending_state(state, current)) | |
3154 | set_current_state(TASK_RUNNING); | |
3155 | ||
3156 | if (current->state == TASK_RUNNING) | |
3157 | return true; | |
3158 | if (ret < 0) | |
3159 | break; | |
3160 | cpu_relax(); | |
3161 | } | |
3162 | ||
67b4110f | 3163 | __set_current_state(TASK_RUNNING); |
bbd7bb70 JA |
3164 | return false; |
3165 | } | |
3166 | ||
ea435e1b | 3167 | static bool blk_mq_poll(struct request_queue *q, blk_qc_t cookie) |
bbd7bb70 JA |
3168 | { |
3169 | struct blk_mq_hw_ctx *hctx; | |
bbd7bb70 JA |
3170 | struct request *rq; |
3171 | ||
ea435e1b | 3172 | if (!test_bit(QUEUE_FLAG_POLL, &q->queue_flags)) |
bbd7bb70 JA |
3173 | return false; |
3174 | ||
bbd7bb70 | 3175 | hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)]; |
bd166ef1 JA |
3176 | if (!blk_qc_t_is_internal(cookie)) |
3177 | rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie)); | |
3a07bb1d | 3178 | else { |
bd166ef1 | 3179 | rq = blk_mq_tag_to_rq(hctx->sched_tags, blk_qc_t_to_tag(cookie)); |
3a07bb1d JA |
3180 | /* |
3181 | * With scheduling, if the request has completed, we'll | |
3182 | * get a NULL return here, as we clear the sched tag when | |
3183 | * that happens. The request still remains valid, like always, | |
3184 | * so we should be safe with just the NULL check. | |
3185 | */ | |
3186 | if (!rq) | |
3187 | return false; | |
3188 | } | |
bbd7bb70 JA |
3189 | |
3190 | return __blk_mq_poll(hctx, rq); | |
3191 | } | |
bbd7bb70 | 3192 | |
320ae51f JA |
3193 | static int __init blk_mq_init(void) |
3194 | { | |
9467f859 TG |
3195 | cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL, |
3196 | blk_mq_hctx_notify_dead); | |
320ae51f JA |
3197 | return 0; |
3198 | } | |
3199 | subsys_initcall(blk_mq_init); |