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; | |
3ee32372 | 312 | rq->start_time = jiffies; |
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; | |
332 | set_start_time_ns(rq); | |
544ccc8d | 333 | rq->cgroup_io_start_time_ns = 0; |
7c3fb70f JA |
334 | #endif |
335 | ||
e4cdf1a1 CH |
336 | data->ctx->rq_dispatched[op_is_sync(op)]++; |
337 | return rq; | |
5dee8577 CH |
338 | } |
339 | ||
d2c0d383 CH |
340 | static struct request *blk_mq_get_request(struct request_queue *q, |
341 | struct bio *bio, unsigned int op, | |
342 | struct blk_mq_alloc_data *data) | |
343 | { | |
344 | struct elevator_queue *e = q->elevator; | |
345 | struct request *rq; | |
e4cdf1a1 | 346 | unsigned int tag; |
21e768b4 | 347 | bool put_ctx_on_error = false; |
d2c0d383 CH |
348 | |
349 | blk_queue_enter_live(q); | |
350 | data->q = q; | |
21e768b4 BVA |
351 | if (likely(!data->ctx)) { |
352 | data->ctx = blk_mq_get_ctx(q); | |
353 | put_ctx_on_error = true; | |
354 | } | |
d2c0d383 CH |
355 | if (likely(!data->hctx)) |
356 | data->hctx = blk_mq_map_queue(q, data->ctx->cpu); | |
03a07c92 GR |
357 | if (op & REQ_NOWAIT) |
358 | data->flags |= BLK_MQ_REQ_NOWAIT; | |
d2c0d383 CH |
359 | |
360 | if (e) { | |
361 | data->flags |= BLK_MQ_REQ_INTERNAL; | |
362 | ||
363 | /* | |
364 | * Flush requests are special and go directly to the | |
365 | * dispatch list. | |
366 | */ | |
5bbf4e5a CH |
367 | if (!op_is_flush(op) && e->type->ops.mq.limit_depth) |
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 | { |
4bc6339a OS |
509 | if (rq->rq_flags & RQF_STATS) { |
510 | blk_mq_poll_stats_start(rq->q); | |
511 | blk_stat_add(rq); | |
512 | } | |
513 | ||
0d11e6ac ML |
514 | blk_account_io_done(rq); |
515 | ||
91b63639 | 516 | if (rq->end_io) { |
a8a45941 | 517 | wbt_done(rq->q->rq_wb, rq); |
320ae51f | 518 | rq->end_io(rq, error); |
91b63639 CH |
519 | } else { |
520 | if (unlikely(blk_bidi_rq(rq))) | |
521 | blk_mq_free_request(rq->next_rq); | |
320ae51f | 522 | blk_mq_free_request(rq); |
91b63639 | 523 | } |
320ae51f | 524 | } |
c8a446ad | 525 | EXPORT_SYMBOL(__blk_mq_end_request); |
63151a44 | 526 | |
2a842aca | 527 | void blk_mq_end_request(struct request *rq, blk_status_t error) |
63151a44 CH |
528 | { |
529 | if (blk_update_request(rq, error, blk_rq_bytes(rq))) | |
530 | BUG(); | |
c8a446ad | 531 | __blk_mq_end_request(rq, error); |
63151a44 | 532 | } |
c8a446ad | 533 | EXPORT_SYMBOL(blk_mq_end_request); |
320ae51f | 534 | |
30a91cb4 | 535 | static void __blk_mq_complete_request_remote(void *data) |
320ae51f | 536 | { |
3d6efbf6 | 537 | struct request *rq = data; |
320ae51f | 538 | |
30a91cb4 | 539 | rq->q->softirq_done_fn(rq); |
320ae51f | 540 | } |
320ae51f | 541 | |
453f8341 | 542 | static void __blk_mq_complete_request(struct request *rq) |
320ae51f JA |
543 | { |
544 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
38535201 | 545 | bool shared = false; |
320ae51f JA |
546 | int cpu; |
547 | ||
1d9bd516 | 548 | WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT); |
5a61c363 | 549 | blk_mq_rq_update_state(rq, MQ_RQ_COMPLETE); |
1d9bd516 | 550 | |
453f8341 CH |
551 | if (rq->internal_tag != -1) |
552 | blk_mq_sched_completed_request(rq); | |
453f8341 | 553 | |
38535201 | 554 | if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) { |
30a91cb4 CH |
555 | rq->q->softirq_done_fn(rq); |
556 | return; | |
557 | } | |
320ae51f JA |
558 | |
559 | cpu = get_cpu(); | |
38535201 CH |
560 | if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags)) |
561 | shared = cpus_share_cache(cpu, ctx->cpu); | |
562 | ||
563 | if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) { | |
30a91cb4 | 564 | rq->csd.func = __blk_mq_complete_request_remote; |
3d6efbf6 CH |
565 | rq->csd.info = rq; |
566 | rq->csd.flags = 0; | |
c46fff2a | 567 | smp_call_function_single_async(ctx->cpu, &rq->csd); |
3d6efbf6 | 568 | } else { |
30a91cb4 | 569 | rq->q->softirq_done_fn(rq); |
3d6efbf6 | 570 | } |
320ae51f JA |
571 | put_cpu(); |
572 | } | |
30a91cb4 | 573 | |
04ced159 | 574 | static void hctx_unlock(struct blk_mq_hw_ctx *hctx, int srcu_idx) |
b7435db8 | 575 | __releases(hctx->srcu) |
04ced159 JA |
576 | { |
577 | if (!(hctx->flags & BLK_MQ_F_BLOCKING)) | |
578 | rcu_read_unlock(); | |
579 | else | |
05707b64 | 580 | srcu_read_unlock(hctx->srcu, srcu_idx); |
04ced159 JA |
581 | } |
582 | ||
583 | static void hctx_lock(struct blk_mq_hw_ctx *hctx, int *srcu_idx) | |
b7435db8 | 584 | __acquires(hctx->srcu) |
04ced159 | 585 | { |
08b5a6e2 JA |
586 | if (!(hctx->flags & BLK_MQ_F_BLOCKING)) { |
587 | /* shut up gcc false positive */ | |
588 | *srcu_idx = 0; | |
04ced159 | 589 | rcu_read_lock(); |
08b5a6e2 | 590 | } else |
05707b64 | 591 | *srcu_idx = srcu_read_lock(hctx->srcu); |
04ced159 JA |
592 | } |
593 | ||
1d9bd516 TH |
594 | static void blk_mq_rq_update_aborted_gstate(struct request *rq, u64 gstate) |
595 | { | |
596 | unsigned long flags; | |
597 | ||
598 | /* | |
599 | * blk_mq_rq_aborted_gstate() is used from the completion path and | |
600 | * can thus be called from irq context. u64_stats_fetch in the | |
601 | * middle of update on the same CPU leads to lockup. Disable irq | |
602 | * while updating. | |
603 | */ | |
604 | local_irq_save(flags); | |
605 | u64_stats_update_begin(&rq->aborted_gstate_sync); | |
606 | rq->aborted_gstate = gstate; | |
607 | u64_stats_update_end(&rq->aborted_gstate_sync); | |
608 | local_irq_restore(flags); | |
609 | } | |
610 | ||
611 | static u64 blk_mq_rq_aborted_gstate(struct request *rq) | |
612 | { | |
613 | unsigned int start; | |
614 | u64 aborted_gstate; | |
615 | ||
616 | do { | |
617 | start = u64_stats_fetch_begin(&rq->aborted_gstate_sync); | |
618 | aborted_gstate = rq->aborted_gstate; | |
619 | } while (u64_stats_fetch_retry(&rq->aborted_gstate_sync, start)); | |
620 | ||
621 | return aborted_gstate; | |
622 | } | |
623 | ||
30a91cb4 CH |
624 | /** |
625 | * blk_mq_complete_request - end I/O on a request | |
626 | * @rq: the request being processed | |
627 | * | |
628 | * Description: | |
629 | * Ends all I/O on a request. It does not handle partial completions. | |
630 | * The actual completion happens out-of-order, through a IPI handler. | |
631 | **/ | |
08e0029a | 632 | void blk_mq_complete_request(struct request *rq) |
30a91cb4 | 633 | { |
95f09684 | 634 | struct request_queue *q = rq->q; |
5197c05e TH |
635 | struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, rq->mq_ctx->cpu); |
636 | int srcu_idx; | |
95f09684 JA |
637 | |
638 | if (unlikely(blk_should_fake_timeout(q))) | |
30a91cb4 | 639 | return; |
5197c05e | 640 | |
1d9bd516 TH |
641 | /* |
642 | * If @rq->aborted_gstate equals the current instance, timeout is | |
643 | * claiming @rq and we lost. This is synchronized through | |
644 | * hctx_lock(). See blk_mq_timeout_work() for details. | |
645 | * | |
646 | * Completion path never blocks and we can directly use RCU here | |
647 | * instead of hctx_lock() which can be either RCU or SRCU. | |
648 | * However, that would complicate paths which want to synchronize | |
649 | * against us. Let stay in sync with the issue path so that | |
650 | * hctx_lock() covers both issue and completion paths. | |
651 | */ | |
5197c05e | 652 | hctx_lock(hctx, &srcu_idx); |
634f9e46 | 653 | if (blk_mq_rq_aborted_gstate(rq) != rq->gstate) |
ed851860 | 654 | __blk_mq_complete_request(rq); |
5197c05e | 655 | hctx_unlock(hctx, srcu_idx); |
30a91cb4 CH |
656 | } |
657 | EXPORT_SYMBOL(blk_mq_complete_request); | |
320ae51f | 658 | |
973c0191 KB |
659 | int blk_mq_request_started(struct request *rq) |
660 | { | |
5a61c363 | 661 | return blk_mq_rq_state(rq) != MQ_RQ_IDLE; |
973c0191 KB |
662 | } |
663 | EXPORT_SYMBOL_GPL(blk_mq_request_started); | |
664 | ||
e2490073 | 665 | void blk_mq_start_request(struct request *rq) |
320ae51f JA |
666 | { |
667 | struct request_queue *q = rq->q; | |
668 | ||
bd166ef1 JA |
669 | blk_mq_sched_started_request(rq); |
670 | ||
320ae51f JA |
671 | trace_block_rq_issue(q, rq); |
672 | ||
cf43e6be | 673 | if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) { |
544ccc8d OS |
674 | rq->io_start_time_ns = ktime_get_ns(); |
675 | #ifdef CONFIG_BLK_DEV_THROTTLING_LOW | |
676 | rq->throtl_size = blk_rq_sectors(rq); | |
677 | #endif | |
cf43e6be | 678 | rq->rq_flags |= RQF_STATS; |
a8a45941 | 679 | wbt_issue(q->rq_wb, rq); |
cf43e6be JA |
680 | } |
681 | ||
1d9bd516 | 682 | WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE); |
538b7534 | 683 | |
87ee7b11 | 684 | /* |
1d9bd516 TH |
685 | * Mark @rq in-flight which also advances the generation number, |
686 | * and register for timeout. Protect with a seqcount to allow the | |
687 | * timeout path to read both @rq->gstate and @rq->deadline | |
688 | * coherently. | |
a7af0af3 | 689 | * |
1d9bd516 TH |
690 | * This is the only place where a request is marked in-flight. If |
691 | * the timeout path reads an in-flight @rq->gstate, the | |
692 | * @rq->deadline it reads together under @rq->gstate_seq is | |
693 | * guaranteed to be the matching one. | |
87ee7b11 | 694 | */ |
1d9bd516 TH |
695 | preempt_disable(); |
696 | write_seqcount_begin(&rq->gstate_seq); | |
697 | ||
698 | blk_mq_rq_update_state(rq, MQ_RQ_IN_FLIGHT); | |
699 | blk_add_timer(rq); | |
700 | ||
701 | write_seqcount_end(&rq->gstate_seq); | |
702 | preempt_enable(); | |
49f5baa5 CH |
703 | |
704 | if (q->dma_drain_size && blk_rq_bytes(rq)) { | |
705 | /* | |
706 | * Make sure space for the drain appears. We know we can do | |
707 | * this because max_hw_segments has been adjusted to be one | |
708 | * fewer than the device can handle. | |
709 | */ | |
710 | rq->nr_phys_segments++; | |
711 | } | |
320ae51f | 712 | } |
e2490073 | 713 | EXPORT_SYMBOL(blk_mq_start_request); |
320ae51f | 714 | |
d9d149a3 | 715 | /* |
5a61c363 TH |
716 | * When we reach here because queue is busy, it's safe to change the state |
717 | * to IDLE without checking @rq->aborted_gstate because we should still be | |
718 | * holding the RCU read lock and thus protected against timeout. | |
d9d149a3 | 719 | */ |
ed0791b2 | 720 | static void __blk_mq_requeue_request(struct request *rq) |
320ae51f JA |
721 | { |
722 | struct request_queue *q = rq->q; | |
723 | ||
923218f6 ML |
724 | blk_mq_put_driver_tag(rq); |
725 | ||
320ae51f | 726 | trace_block_rq_requeue(q, rq); |
a8a45941 | 727 | wbt_requeue(q->rq_wb, rq); |
49f5baa5 | 728 | |
5a61c363 | 729 | if (blk_mq_rq_state(rq) != MQ_RQ_IDLE) { |
1d9bd516 | 730 | blk_mq_rq_update_state(rq, MQ_RQ_IDLE); |
e2490073 CH |
731 | if (q->dma_drain_size && blk_rq_bytes(rq)) |
732 | rq->nr_phys_segments--; | |
733 | } | |
320ae51f JA |
734 | } |
735 | ||
2b053aca | 736 | void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list) |
ed0791b2 | 737 | { |
ed0791b2 | 738 | __blk_mq_requeue_request(rq); |
ed0791b2 | 739 | |
105976f5 ML |
740 | /* this request will be re-inserted to io scheduler queue */ |
741 | blk_mq_sched_requeue_request(rq); | |
742 | ||
ed0791b2 | 743 | BUG_ON(blk_queued_rq(rq)); |
2b053aca | 744 | blk_mq_add_to_requeue_list(rq, true, kick_requeue_list); |
ed0791b2 CH |
745 | } |
746 | EXPORT_SYMBOL(blk_mq_requeue_request); | |
747 | ||
6fca6a61 CH |
748 | static void blk_mq_requeue_work(struct work_struct *work) |
749 | { | |
750 | struct request_queue *q = | |
2849450a | 751 | container_of(work, struct request_queue, requeue_work.work); |
6fca6a61 CH |
752 | LIST_HEAD(rq_list); |
753 | struct request *rq, *next; | |
6fca6a61 | 754 | |
18e9781d | 755 | spin_lock_irq(&q->requeue_lock); |
6fca6a61 | 756 | list_splice_init(&q->requeue_list, &rq_list); |
18e9781d | 757 | spin_unlock_irq(&q->requeue_lock); |
6fca6a61 CH |
758 | |
759 | list_for_each_entry_safe(rq, next, &rq_list, queuelist) { | |
e8064021 | 760 | if (!(rq->rq_flags & RQF_SOFTBARRIER)) |
6fca6a61 CH |
761 | continue; |
762 | ||
e8064021 | 763 | rq->rq_flags &= ~RQF_SOFTBARRIER; |
6fca6a61 | 764 | list_del_init(&rq->queuelist); |
9e97d295 | 765 | blk_mq_sched_insert_request(rq, true, false, false); |
6fca6a61 CH |
766 | } |
767 | ||
768 | while (!list_empty(&rq_list)) { | |
769 | rq = list_entry(rq_list.next, struct request, queuelist); | |
770 | list_del_init(&rq->queuelist); | |
9e97d295 | 771 | blk_mq_sched_insert_request(rq, false, false, false); |
6fca6a61 CH |
772 | } |
773 | ||
52d7f1b5 | 774 | blk_mq_run_hw_queues(q, false); |
6fca6a61 CH |
775 | } |
776 | ||
2b053aca BVA |
777 | void blk_mq_add_to_requeue_list(struct request *rq, bool at_head, |
778 | bool kick_requeue_list) | |
6fca6a61 CH |
779 | { |
780 | struct request_queue *q = rq->q; | |
781 | unsigned long flags; | |
782 | ||
783 | /* | |
784 | * We abuse this flag that is otherwise used by the I/O scheduler to | |
ff821d27 | 785 | * request head insertion from the workqueue. |
6fca6a61 | 786 | */ |
e8064021 | 787 | BUG_ON(rq->rq_flags & RQF_SOFTBARRIER); |
6fca6a61 CH |
788 | |
789 | spin_lock_irqsave(&q->requeue_lock, flags); | |
790 | if (at_head) { | |
e8064021 | 791 | rq->rq_flags |= RQF_SOFTBARRIER; |
6fca6a61 CH |
792 | list_add(&rq->queuelist, &q->requeue_list); |
793 | } else { | |
794 | list_add_tail(&rq->queuelist, &q->requeue_list); | |
795 | } | |
796 | spin_unlock_irqrestore(&q->requeue_lock, flags); | |
2b053aca BVA |
797 | |
798 | if (kick_requeue_list) | |
799 | blk_mq_kick_requeue_list(q); | |
6fca6a61 CH |
800 | } |
801 | EXPORT_SYMBOL(blk_mq_add_to_requeue_list); | |
802 | ||
803 | void blk_mq_kick_requeue_list(struct request_queue *q) | |
804 | { | |
ae943d20 | 805 | kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0); |
6fca6a61 CH |
806 | } |
807 | EXPORT_SYMBOL(blk_mq_kick_requeue_list); | |
808 | ||
2849450a MS |
809 | void blk_mq_delay_kick_requeue_list(struct request_queue *q, |
810 | unsigned long msecs) | |
811 | { | |
d4acf365 BVA |
812 | kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, |
813 | msecs_to_jiffies(msecs)); | |
2849450a MS |
814 | } |
815 | EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list); | |
816 | ||
0e62f51f JA |
817 | struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag) |
818 | { | |
88c7b2b7 JA |
819 | if (tag < tags->nr_tags) { |
820 | prefetch(tags->rqs[tag]); | |
4ee86bab | 821 | return tags->rqs[tag]; |
88c7b2b7 | 822 | } |
4ee86bab HR |
823 | |
824 | return NULL; | |
24d2f903 CH |
825 | } |
826 | EXPORT_SYMBOL(blk_mq_tag_to_rq); | |
827 | ||
320ae51f | 828 | struct blk_mq_timeout_data { |
46f92d42 CH |
829 | unsigned long next; |
830 | unsigned int next_set; | |
1d9bd516 | 831 | unsigned int nr_expired; |
320ae51f JA |
832 | }; |
833 | ||
358f70da | 834 | static void blk_mq_rq_timed_out(struct request *req, bool reserved) |
320ae51f | 835 | { |
f8a5b122 | 836 | const struct blk_mq_ops *ops = req->q->mq_ops; |
46f92d42 | 837 | enum blk_eh_timer_return ret = BLK_EH_RESET_TIMER; |
87ee7b11 | 838 | |
634f9e46 | 839 | req->rq_flags |= RQF_MQ_TIMEOUT_EXPIRED; |
87ee7b11 | 840 | |
46f92d42 | 841 | if (ops->timeout) |
0152fb6b | 842 | ret = ops->timeout(req, reserved); |
46f92d42 CH |
843 | |
844 | switch (ret) { | |
845 | case BLK_EH_HANDLED: | |
846 | __blk_mq_complete_request(req); | |
847 | break; | |
848 | case BLK_EH_RESET_TIMER: | |
1d9bd516 TH |
849 | /* |
850 | * As nothing prevents from completion happening while | |
851 | * ->aborted_gstate is set, this may lead to ignored | |
852 | * completions and further spurious timeouts. | |
853 | */ | |
854 | blk_mq_rq_update_aborted_gstate(req, 0); | |
46f92d42 | 855 | blk_add_timer(req); |
46f92d42 CH |
856 | break; |
857 | case BLK_EH_NOT_HANDLED: | |
858 | break; | |
859 | default: | |
860 | printk(KERN_ERR "block: bad eh return: %d\n", ret); | |
861 | break; | |
862 | } | |
87ee7b11 | 863 | } |
5b3f25fc | 864 | |
81481eb4 CH |
865 | static void blk_mq_check_expired(struct blk_mq_hw_ctx *hctx, |
866 | struct request *rq, void *priv, bool reserved) | |
867 | { | |
868 | struct blk_mq_timeout_data *data = priv; | |
1d9bd516 TH |
869 | unsigned long gstate, deadline; |
870 | int start; | |
87ee7b11 | 871 | |
1d9bd516 | 872 | might_sleep(); |
87ee7b11 | 873 | |
5a61c363 | 874 | if (rq->rq_flags & RQF_MQ_TIMEOUT_EXPIRED) |
46f92d42 | 875 | return; |
a7af0af3 | 876 | |
1d9bd516 TH |
877 | /* read coherent snapshots of @rq->state_gen and @rq->deadline */ |
878 | while (true) { | |
879 | start = read_seqcount_begin(&rq->gstate_seq); | |
880 | gstate = READ_ONCE(rq->gstate); | |
0a72e7f4 | 881 | deadline = blk_rq_deadline(rq); |
1d9bd516 TH |
882 | if (!read_seqcount_retry(&rq->gstate_seq, start)) |
883 | break; | |
884 | cond_resched(); | |
885 | } | |
a7af0af3 | 886 | |
1d9bd516 TH |
887 | /* if in-flight && overdue, mark for abortion */ |
888 | if ((gstate & MQ_RQ_STATE_MASK) == MQ_RQ_IN_FLIGHT && | |
889 | time_after_eq(jiffies, deadline)) { | |
890 | blk_mq_rq_update_aborted_gstate(rq, gstate); | |
891 | data->nr_expired++; | |
892 | hctx->nr_expired++; | |
a7af0af3 PZ |
893 | } else if (!data->next_set || time_after(data->next, deadline)) { |
894 | data->next = deadline; | |
46f92d42 CH |
895 | data->next_set = 1; |
896 | } | |
87ee7b11 JA |
897 | } |
898 | ||
1d9bd516 TH |
899 | static void blk_mq_terminate_expired(struct blk_mq_hw_ctx *hctx, |
900 | struct request *rq, void *priv, bool reserved) | |
901 | { | |
902 | /* | |
903 | * We marked @rq->aborted_gstate and waited for RCU. If there were | |
904 | * completions that we lost to, they would have finished and | |
905 | * updated @rq->gstate by now; otherwise, the completion path is | |
906 | * now guaranteed to see @rq->aborted_gstate and yield. If | |
907 | * @rq->aborted_gstate still matches @rq->gstate, @rq is ours. | |
908 | */ | |
634f9e46 TH |
909 | if (!(rq->rq_flags & RQF_MQ_TIMEOUT_EXPIRED) && |
910 | READ_ONCE(rq->gstate) == rq->aborted_gstate) | |
1d9bd516 TH |
911 | blk_mq_rq_timed_out(rq, reserved); |
912 | } | |
913 | ||
287922eb | 914 | static void blk_mq_timeout_work(struct work_struct *work) |
320ae51f | 915 | { |
287922eb CH |
916 | struct request_queue *q = |
917 | container_of(work, struct request_queue, timeout_work); | |
81481eb4 CH |
918 | struct blk_mq_timeout_data data = { |
919 | .next = 0, | |
920 | .next_set = 0, | |
1d9bd516 | 921 | .nr_expired = 0, |
81481eb4 | 922 | }; |
1d9bd516 | 923 | struct blk_mq_hw_ctx *hctx; |
81481eb4 | 924 | int i; |
320ae51f | 925 | |
71f79fb3 GKB |
926 | /* A deadlock might occur if a request is stuck requiring a |
927 | * timeout at the same time a queue freeze is waiting | |
928 | * completion, since the timeout code would not be able to | |
929 | * acquire the queue reference here. | |
930 | * | |
931 | * That's why we don't use blk_queue_enter here; instead, we use | |
932 | * percpu_ref_tryget directly, because we need to be able to | |
933 | * obtain a reference even in the short window between the queue | |
934 | * starting to freeze, by dropping the first reference in | |
1671d522 | 935 | * blk_freeze_queue_start, and the moment the last request is |
71f79fb3 GKB |
936 | * consumed, marked by the instant q_usage_counter reaches |
937 | * zero. | |
938 | */ | |
939 | if (!percpu_ref_tryget(&q->q_usage_counter)) | |
287922eb CH |
940 | return; |
941 | ||
1d9bd516 | 942 | /* scan for the expired ones and set their ->aborted_gstate */ |
0bf6cd5b | 943 | blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &data); |
320ae51f | 944 | |
1d9bd516 TH |
945 | if (data.nr_expired) { |
946 | bool has_rcu = false; | |
947 | ||
948 | /* | |
949 | * Wait till everyone sees ->aborted_gstate. The | |
950 | * sequential waits for SRCUs aren't ideal. If this ever | |
951 | * becomes a problem, we can add per-hw_ctx rcu_head and | |
952 | * wait in parallel. | |
953 | */ | |
954 | queue_for_each_hw_ctx(q, hctx, i) { | |
955 | if (!hctx->nr_expired) | |
956 | continue; | |
957 | ||
958 | if (!(hctx->flags & BLK_MQ_F_BLOCKING)) | |
959 | has_rcu = true; | |
960 | else | |
05707b64 | 961 | synchronize_srcu(hctx->srcu); |
1d9bd516 TH |
962 | |
963 | hctx->nr_expired = 0; | |
964 | } | |
965 | if (has_rcu) | |
966 | synchronize_rcu(); | |
967 | ||
968 | /* terminate the ones we won */ | |
969 | blk_mq_queue_tag_busy_iter(q, blk_mq_terminate_expired, NULL); | |
970 | } | |
971 | ||
81481eb4 CH |
972 | if (data.next_set) { |
973 | data.next = blk_rq_timeout(round_jiffies_up(data.next)); | |
974 | mod_timer(&q->timeout, data.next); | |
0d2602ca | 975 | } else { |
fcd36c36 BVA |
976 | /* |
977 | * Request timeouts are handled as a forward rolling timer. If | |
978 | * we end up here it means that no requests are pending and | |
979 | * also that no request has been pending for a while. Mark | |
980 | * each hctx as idle. | |
981 | */ | |
f054b56c ML |
982 | queue_for_each_hw_ctx(q, hctx, i) { |
983 | /* the hctx may be unmapped, so check it here */ | |
984 | if (blk_mq_hw_queue_mapped(hctx)) | |
985 | blk_mq_tag_idle(hctx); | |
986 | } | |
0d2602ca | 987 | } |
287922eb | 988 | blk_queue_exit(q); |
320ae51f JA |
989 | } |
990 | ||
88459642 OS |
991 | struct flush_busy_ctx_data { |
992 | struct blk_mq_hw_ctx *hctx; | |
993 | struct list_head *list; | |
994 | }; | |
995 | ||
996 | static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data) | |
997 | { | |
998 | struct flush_busy_ctx_data *flush_data = data; | |
999 | struct blk_mq_hw_ctx *hctx = flush_data->hctx; | |
1000 | struct blk_mq_ctx *ctx = hctx->ctxs[bitnr]; | |
1001 | ||
88459642 OS |
1002 | spin_lock(&ctx->lock); |
1003 | list_splice_tail_init(&ctx->rq_list, flush_data->list); | |
e9a99a63 | 1004 | sbitmap_clear_bit(sb, bitnr); |
88459642 OS |
1005 | spin_unlock(&ctx->lock); |
1006 | return true; | |
1007 | } | |
1008 | ||
1429d7c9 JA |
1009 | /* |
1010 | * Process software queues that have been marked busy, splicing them | |
1011 | * to the for-dispatch | |
1012 | */ | |
2c3ad667 | 1013 | void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list) |
1429d7c9 | 1014 | { |
88459642 OS |
1015 | struct flush_busy_ctx_data data = { |
1016 | .hctx = hctx, | |
1017 | .list = list, | |
1018 | }; | |
1429d7c9 | 1019 | |
88459642 | 1020 | sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data); |
1429d7c9 | 1021 | } |
2c3ad667 | 1022 | EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs); |
1429d7c9 | 1023 | |
b347689f ML |
1024 | struct dispatch_rq_data { |
1025 | struct blk_mq_hw_ctx *hctx; | |
1026 | struct request *rq; | |
1027 | }; | |
1028 | ||
1029 | static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr, | |
1030 | void *data) | |
1031 | { | |
1032 | struct dispatch_rq_data *dispatch_data = data; | |
1033 | struct blk_mq_hw_ctx *hctx = dispatch_data->hctx; | |
1034 | struct blk_mq_ctx *ctx = hctx->ctxs[bitnr]; | |
1035 | ||
1036 | spin_lock(&ctx->lock); | |
1037 | if (unlikely(!list_empty(&ctx->rq_list))) { | |
1038 | dispatch_data->rq = list_entry_rq(ctx->rq_list.next); | |
1039 | list_del_init(&dispatch_data->rq->queuelist); | |
1040 | if (list_empty(&ctx->rq_list)) | |
1041 | sbitmap_clear_bit(sb, bitnr); | |
1042 | } | |
1043 | spin_unlock(&ctx->lock); | |
1044 | ||
1045 | return !dispatch_data->rq; | |
1046 | } | |
1047 | ||
1048 | struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx, | |
1049 | struct blk_mq_ctx *start) | |
1050 | { | |
1051 | unsigned off = start ? start->index_hw : 0; | |
1052 | struct dispatch_rq_data data = { | |
1053 | .hctx = hctx, | |
1054 | .rq = NULL, | |
1055 | }; | |
1056 | ||
1057 | __sbitmap_for_each_set(&hctx->ctx_map, off, | |
1058 | dispatch_rq_from_ctx, &data); | |
1059 | ||
1060 | return data.rq; | |
1061 | } | |
1062 | ||
703fd1c0 JA |
1063 | static inline unsigned int queued_to_index(unsigned int queued) |
1064 | { | |
1065 | if (!queued) | |
1066 | return 0; | |
1429d7c9 | 1067 | |
703fd1c0 | 1068 | return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1); |
1429d7c9 JA |
1069 | } |
1070 | ||
bd6737f1 JA |
1071 | bool blk_mq_get_driver_tag(struct request *rq, struct blk_mq_hw_ctx **hctx, |
1072 | bool wait) | |
bd166ef1 JA |
1073 | { |
1074 | struct blk_mq_alloc_data data = { | |
1075 | .q = rq->q, | |
bd166ef1 JA |
1076 | .hctx = blk_mq_map_queue(rq->q, rq->mq_ctx->cpu), |
1077 | .flags = wait ? 0 : BLK_MQ_REQ_NOWAIT, | |
1078 | }; | |
1079 | ||
5feeacdd JA |
1080 | might_sleep_if(wait); |
1081 | ||
81380ca1 OS |
1082 | if (rq->tag != -1) |
1083 | goto done; | |
bd166ef1 | 1084 | |
415b806d SG |
1085 | if (blk_mq_tag_is_reserved(data.hctx->sched_tags, rq->internal_tag)) |
1086 | data.flags |= BLK_MQ_REQ_RESERVED; | |
1087 | ||
bd166ef1 JA |
1088 | rq->tag = blk_mq_get_tag(&data); |
1089 | if (rq->tag >= 0) { | |
200e86b3 JA |
1090 | if (blk_mq_tag_busy(data.hctx)) { |
1091 | rq->rq_flags |= RQF_MQ_INFLIGHT; | |
1092 | atomic_inc(&data.hctx->nr_active); | |
1093 | } | |
bd166ef1 | 1094 | data.hctx->tags->rqs[rq->tag] = rq; |
bd166ef1 JA |
1095 | } |
1096 | ||
81380ca1 OS |
1097 | done: |
1098 | if (hctx) | |
1099 | *hctx = data.hctx; | |
1100 | return rq->tag != -1; | |
bd166ef1 JA |
1101 | } |
1102 | ||
eb619fdb JA |
1103 | static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode, |
1104 | int flags, void *key) | |
da55f2cc OS |
1105 | { |
1106 | struct blk_mq_hw_ctx *hctx; | |
1107 | ||
1108 | hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait); | |
1109 | ||
eb619fdb | 1110 | list_del_init(&wait->entry); |
da55f2cc OS |
1111 | blk_mq_run_hw_queue(hctx, true); |
1112 | return 1; | |
1113 | } | |
1114 | ||
f906a6a0 JA |
1115 | /* |
1116 | * Mark us waiting for a tag. For shared tags, this involves hooking us into | |
ee3e4de5 BVA |
1117 | * the tag wakeups. For non-shared tags, we can simply mark us needing a |
1118 | * restart. For both cases, take care to check the condition again after | |
f906a6a0 JA |
1119 | * marking us as waiting. |
1120 | */ | |
1121 | static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx **hctx, | |
1122 | struct request *rq) | |
da55f2cc | 1123 | { |
eb619fdb | 1124 | struct blk_mq_hw_ctx *this_hctx = *hctx; |
da55f2cc | 1125 | struct sbq_wait_state *ws; |
f906a6a0 JA |
1126 | wait_queue_entry_t *wait; |
1127 | bool ret; | |
da55f2cc | 1128 | |
c27d53fb | 1129 | if (!(this_hctx->flags & BLK_MQ_F_TAG_SHARED)) { |
f906a6a0 JA |
1130 | if (!test_bit(BLK_MQ_S_SCHED_RESTART, &this_hctx->state)) |
1131 | set_bit(BLK_MQ_S_SCHED_RESTART, &this_hctx->state); | |
f906a6a0 | 1132 | |
c27d53fb BVA |
1133 | /* |
1134 | * It's possible that a tag was freed in the window between the | |
1135 | * allocation failure and adding the hardware queue to the wait | |
1136 | * queue. | |
1137 | * | |
1138 | * Don't clear RESTART here, someone else could have set it. | |
1139 | * At most this will cost an extra queue run. | |
1140 | */ | |
1141 | return blk_mq_get_driver_tag(rq, hctx, false); | |
eb619fdb | 1142 | } |
eb619fdb | 1143 | |
c27d53fb BVA |
1144 | wait = &this_hctx->dispatch_wait; |
1145 | if (!list_empty_careful(&wait->entry)) | |
1146 | return false; | |
1147 | ||
1148 | spin_lock(&this_hctx->lock); | |
1149 | if (!list_empty(&wait->entry)) { | |
1150 | spin_unlock(&this_hctx->lock); | |
1151 | return false; | |
eb619fdb JA |
1152 | } |
1153 | ||
c27d53fb BVA |
1154 | ws = bt_wait_ptr(&this_hctx->tags->bitmap_tags, this_hctx); |
1155 | add_wait_queue(&ws->wait, wait); | |
1156 | ||
da55f2cc | 1157 | /* |
eb619fdb JA |
1158 | * It's possible that a tag was freed in the window between the |
1159 | * allocation failure and adding the hardware queue to the wait | |
1160 | * queue. | |
da55f2cc | 1161 | */ |
f906a6a0 | 1162 | ret = blk_mq_get_driver_tag(rq, hctx, false); |
c27d53fb | 1163 | if (!ret) { |
eb619fdb | 1164 | spin_unlock(&this_hctx->lock); |
c27d53fb | 1165 | return false; |
eb619fdb | 1166 | } |
c27d53fb BVA |
1167 | |
1168 | /* | |
1169 | * We got a tag, remove ourselves from the wait queue to ensure | |
1170 | * someone else gets the wakeup. | |
1171 | */ | |
1172 | spin_lock_irq(&ws->wait.lock); | |
1173 | list_del_init(&wait->entry); | |
1174 | spin_unlock_irq(&ws->wait.lock); | |
1175 | spin_unlock(&this_hctx->lock); | |
1176 | ||
1177 | return true; | |
da55f2cc OS |
1178 | } |
1179 | ||
86ff7c2a ML |
1180 | #define BLK_MQ_RESOURCE_DELAY 3 /* ms units */ |
1181 | ||
de148297 | 1182 | bool blk_mq_dispatch_rq_list(struct request_queue *q, struct list_head *list, |
eb619fdb | 1183 | bool got_budget) |
320ae51f | 1184 | { |
81380ca1 | 1185 | struct blk_mq_hw_ctx *hctx; |
6d6f167c | 1186 | struct request *rq, *nxt; |
eb619fdb | 1187 | bool no_tag = false; |
fc17b653 | 1188 | int errors, queued; |
86ff7c2a | 1189 | blk_status_t ret = BLK_STS_OK; |
320ae51f | 1190 | |
81380ca1 OS |
1191 | if (list_empty(list)) |
1192 | return false; | |
1193 | ||
de148297 ML |
1194 | WARN_ON(!list_is_singular(list) && got_budget); |
1195 | ||
320ae51f JA |
1196 | /* |
1197 | * Now process all the entries, sending them to the driver. | |
1198 | */ | |
93efe981 | 1199 | errors = queued = 0; |
81380ca1 | 1200 | do { |
74c45052 | 1201 | struct blk_mq_queue_data bd; |
320ae51f | 1202 | |
f04c3df3 | 1203 | rq = list_first_entry(list, struct request, queuelist); |
0bca799b ML |
1204 | |
1205 | hctx = blk_mq_map_queue(rq->q, rq->mq_ctx->cpu); | |
1206 | if (!got_budget && !blk_mq_get_dispatch_budget(hctx)) | |
1207 | break; | |
1208 | ||
1209 | if (!blk_mq_get_driver_tag(rq, NULL, false)) { | |
3c782d67 | 1210 | /* |
da55f2cc | 1211 | * The initial allocation attempt failed, so we need to |
eb619fdb JA |
1212 | * rerun the hardware queue when a tag is freed. The |
1213 | * waitqueue takes care of that. If the queue is run | |
1214 | * before we add this entry back on the dispatch list, | |
1215 | * we'll re-run it below. | |
3c782d67 | 1216 | */ |
f906a6a0 | 1217 | if (!blk_mq_mark_tag_wait(&hctx, rq)) { |
0bca799b | 1218 | blk_mq_put_dispatch_budget(hctx); |
f906a6a0 JA |
1219 | /* |
1220 | * For non-shared tags, the RESTART check | |
1221 | * will suffice. | |
1222 | */ | |
1223 | if (hctx->flags & BLK_MQ_F_TAG_SHARED) | |
1224 | no_tag = true; | |
de148297 ML |
1225 | break; |
1226 | } | |
1227 | } | |
1228 | ||
320ae51f | 1229 | list_del_init(&rq->queuelist); |
320ae51f | 1230 | |
74c45052 | 1231 | bd.rq = rq; |
113285b4 JA |
1232 | |
1233 | /* | |
1234 | * Flag last if we have no more requests, or if we have more | |
1235 | * but can't assign a driver tag to it. | |
1236 | */ | |
1237 | if (list_empty(list)) | |
1238 | bd.last = true; | |
1239 | else { | |
113285b4 JA |
1240 | nxt = list_first_entry(list, struct request, queuelist); |
1241 | bd.last = !blk_mq_get_driver_tag(nxt, NULL, false); | |
1242 | } | |
74c45052 JA |
1243 | |
1244 | ret = q->mq_ops->queue_rq(hctx, &bd); | |
86ff7c2a | 1245 | if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE) { |
6d6f167c JW |
1246 | /* |
1247 | * If an I/O scheduler has been configured and we got a | |
ff821d27 JA |
1248 | * driver tag for the next request already, free it |
1249 | * again. | |
6d6f167c JW |
1250 | */ |
1251 | if (!list_empty(list)) { | |
1252 | nxt = list_first_entry(list, struct request, queuelist); | |
1253 | blk_mq_put_driver_tag(nxt); | |
1254 | } | |
f04c3df3 | 1255 | list_add(&rq->queuelist, list); |
ed0791b2 | 1256 | __blk_mq_requeue_request(rq); |
320ae51f | 1257 | break; |
fc17b653 CH |
1258 | } |
1259 | ||
1260 | if (unlikely(ret != BLK_STS_OK)) { | |
93efe981 | 1261 | errors++; |
2a842aca | 1262 | blk_mq_end_request(rq, BLK_STS_IOERR); |
fc17b653 | 1263 | continue; |
320ae51f JA |
1264 | } |
1265 | ||
fc17b653 | 1266 | queued++; |
81380ca1 | 1267 | } while (!list_empty(list)); |
320ae51f | 1268 | |
703fd1c0 | 1269 | hctx->dispatched[queued_to_index(queued)]++; |
320ae51f JA |
1270 | |
1271 | /* | |
1272 | * Any items that need requeuing? Stuff them into hctx->dispatch, | |
1273 | * that is where we will continue on next queue run. | |
1274 | */ | |
f04c3df3 | 1275 | if (!list_empty(list)) { |
86ff7c2a ML |
1276 | bool needs_restart; |
1277 | ||
320ae51f | 1278 | spin_lock(&hctx->lock); |
c13660a0 | 1279 | list_splice_init(list, &hctx->dispatch); |
320ae51f | 1280 | spin_unlock(&hctx->lock); |
f04c3df3 | 1281 | |
9ba52e58 | 1282 | /* |
710c785f BVA |
1283 | * If SCHED_RESTART was set by the caller of this function and |
1284 | * it is no longer set that means that it was cleared by another | |
1285 | * thread and hence that a queue rerun is needed. | |
9ba52e58 | 1286 | * |
eb619fdb JA |
1287 | * If 'no_tag' is set, that means that we failed getting |
1288 | * a driver tag with an I/O scheduler attached. If our dispatch | |
1289 | * waitqueue is no longer active, ensure that we run the queue | |
1290 | * AFTER adding our entries back to the list. | |
bd166ef1 | 1291 | * |
710c785f BVA |
1292 | * If no I/O scheduler has been configured it is possible that |
1293 | * the hardware queue got stopped and restarted before requests | |
1294 | * were pushed back onto the dispatch list. Rerun the queue to | |
1295 | * avoid starvation. Notes: | |
1296 | * - blk_mq_run_hw_queue() checks whether or not a queue has | |
1297 | * been stopped before rerunning a queue. | |
1298 | * - Some but not all block drivers stop a queue before | |
fc17b653 | 1299 | * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq |
710c785f | 1300 | * and dm-rq. |
86ff7c2a ML |
1301 | * |
1302 | * If driver returns BLK_STS_RESOURCE and SCHED_RESTART | |
1303 | * bit is set, run queue after a delay to avoid IO stalls | |
1304 | * that could otherwise occur if the queue is idle. | |
bd166ef1 | 1305 | */ |
86ff7c2a ML |
1306 | needs_restart = blk_mq_sched_needs_restart(hctx); |
1307 | if (!needs_restart || | |
eb619fdb | 1308 | (no_tag && list_empty_careful(&hctx->dispatch_wait.entry))) |
bd166ef1 | 1309 | blk_mq_run_hw_queue(hctx, true); |
86ff7c2a ML |
1310 | else if (needs_restart && (ret == BLK_STS_RESOURCE)) |
1311 | blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY); | |
320ae51f | 1312 | } |
f04c3df3 | 1313 | |
93efe981 | 1314 | return (queued + errors) != 0; |
f04c3df3 JA |
1315 | } |
1316 | ||
6a83e74d BVA |
1317 | static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx) |
1318 | { | |
1319 | int srcu_idx; | |
1320 | ||
b7a71e66 JA |
1321 | /* |
1322 | * We should be running this queue from one of the CPUs that | |
1323 | * are mapped to it. | |
7df938fb ML |
1324 | * |
1325 | * There are at least two related races now between setting | |
1326 | * hctx->next_cpu from blk_mq_hctx_next_cpu() and running | |
1327 | * __blk_mq_run_hw_queue(): | |
1328 | * | |
1329 | * - hctx->next_cpu is found offline in blk_mq_hctx_next_cpu(), | |
1330 | * but later it becomes online, then this warning is harmless | |
1331 | * at all | |
1332 | * | |
1333 | * - hctx->next_cpu is found online in blk_mq_hctx_next_cpu(), | |
1334 | * but later it becomes offline, then the warning can't be | |
1335 | * triggered, and we depend on blk-mq timeout handler to | |
1336 | * handle dispatched requests to this hctx | |
b7a71e66 | 1337 | */ |
7df938fb ML |
1338 | if (!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) && |
1339 | cpu_online(hctx->next_cpu)) { | |
1340 | printk(KERN_WARNING "run queue from wrong CPU %d, hctx %s\n", | |
1341 | raw_smp_processor_id(), | |
1342 | cpumask_empty(hctx->cpumask) ? "inactive": "active"); | |
1343 | dump_stack(); | |
1344 | } | |
6a83e74d | 1345 | |
b7a71e66 JA |
1346 | /* |
1347 | * We can't run the queue inline with ints disabled. Ensure that | |
1348 | * we catch bad users of this early. | |
1349 | */ | |
1350 | WARN_ON_ONCE(in_interrupt()); | |
1351 | ||
04ced159 | 1352 | might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING); |
bf4907c0 | 1353 | |
04ced159 JA |
1354 | hctx_lock(hctx, &srcu_idx); |
1355 | blk_mq_sched_dispatch_requests(hctx); | |
1356 | hctx_unlock(hctx, srcu_idx); | |
6a83e74d BVA |
1357 | } |
1358 | ||
f82ddf19 ML |
1359 | static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx) |
1360 | { | |
1361 | int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask); | |
1362 | ||
1363 | if (cpu >= nr_cpu_ids) | |
1364 | cpu = cpumask_first(hctx->cpumask); | |
1365 | return cpu; | |
1366 | } | |
1367 | ||
506e931f JA |
1368 | /* |
1369 | * It'd be great if the workqueue API had a way to pass | |
1370 | * in a mask and had some smarts for more clever placement. | |
1371 | * For now we just round-robin here, switching for every | |
1372 | * BLK_MQ_CPU_WORK_BATCH queued items. | |
1373 | */ | |
1374 | static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx) | |
1375 | { | |
7bed4595 | 1376 | bool tried = false; |
476f8c98 | 1377 | int next_cpu = hctx->next_cpu; |
7bed4595 | 1378 | |
b657d7e6 CH |
1379 | if (hctx->queue->nr_hw_queues == 1) |
1380 | return WORK_CPU_UNBOUND; | |
506e931f JA |
1381 | |
1382 | if (--hctx->next_cpu_batch <= 0) { | |
7bed4595 | 1383 | select_cpu: |
476f8c98 | 1384 | next_cpu = cpumask_next_and(next_cpu, hctx->cpumask, |
20e4d813 | 1385 | cpu_online_mask); |
506e931f | 1386 | if (next_cpu >= nr_cpu_ids) |
f82ddf19 | 1387 | next_cpu = blk_mq_first_mapped_cpu(hctx); |
506e931f JA |
1388 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; |
1389 | } | |
1390 | ||
7bed4595 ML |
1391 | /* |
1392 | * Do unbound schedule if we can't find a online CPU for this hctx, | |
1393 | * and it should only happen in the path of handling CPU DEAD. | |
1394 | */ | |
476f8c98 | 1395 | if (!cpu_online(next_cpu)) { |
7bed4595 ML |
1396 | if (!tried) { |
1397 | tried = true; | |
1398 | goto select_cpu; | |
1399 | } | |
1400 | ||
1401 | /* | |
1402 | * Make sure to re-select CPU next time once after CPUs | |
1403 | * in hctx->cpumask become online again. | |
1404 | */ | |
476f8c98 | 1405 | hctx->next_cpu = next_cpu; |
7bed4595 ML |
1406 | hctx->next_cpu_batch = 1; |
1407 | return WORK_CPU_UNBOUND; | |
1408 | } | |
476f8c98 ML |
1409 | |
1410 | hctx->next_cpu = next_cpu; | |
1411 | return next_cpu; | |
506e931f JA |
1412 | } |
1413 | ||
7587a5ae BVA |
1414 | static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async, |
1415 | unsigned long msecs) | |
320ae51f | 1416 | { |
5435c023 | 1417 | if (unlikely(blk_mq_hctx_stopped(hctx))) |
320ae51f JA |
1418 | return; |
1419 | ||
1b792f2f | 1420 | if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) { |
2a90d4aa PB |
1421 | int cpu = get_cpu(); |
1422 | if (cpumask_test_cpu(cpu, hctx->cpumask)) { | |
398205b8 | 1423 | __blk_mq_run_hw_queue(hctx); |
2a90d4aa | 1424 | put_cpu(); |
398205b8 PB |
1425 | return; |
1426 | } | |
e4043dcf | 1427 | |
2a90d4aa | 1428 | put_cpu(); |
e4043dcf | 1429 | } |
398205b8 | 1430 | |
ae943d20 BVA |
1431 | kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work, |
1432 | msecs_to_jiffies(msecs)); | |
7587a5ae BVA |
1433 | } |
1434 | ||
1435 | void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs) | |
1436 | { | |
1437 | __blk_mq_delay_run_hw_queue(hctx, true, msecs); | |
1438 | } | |
1439 | EXPORT_SYMBOL(blk_mq_delay_run_hw_queue); | |
1440 | ||
79f720a7 | 1441 | bool blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) |
7587a5ae | 1442 | { |
24f5a90f ML |
1443 | int srcu_idx; |
1444 | bool need_run; | |
1445 | ||
1446 | /* | |
1447 | * When queue is quiesced, we may be switching io scheduler, or | |
1448 | * updating nr_hw_queues, or other things, and we can't run queue | |
1449 | * any more, even __blk_mq_hctx_has_pending() can't be called safely. | |
1450 | * | |
1451 | * And queue will be rerun in blk_mq_unquiesce_queue() if it is | |
1452 | * quiesced. | |
1453 | */ | |
04ced159 JA |
1454 | hctx_lock(hctx, &srcu_idx); |
1455 | need_run = !blk_queue_quiesced(hctx->queue) && | |
1456 | blk_mq_hctx_has_pending(hctx); | |
1457 | hctx_unlock(hctx, srcu_idx); | |
24f5a90f ML |
1458 | |
1459 | if (need_run) { | |
79f720a7 JA |
1460 | __blk_mq_delay_run_hw_queue(hctx, async, 0); |
1461 | return true; | |
1462 | } | |
1463 | ||
1464 | return false; | |
320ae51f | 1465 | } |
5b727272 | 1466 | EXPORT_SYMBOL(blk_mq_run_hw_queue); |
320ae51f | 1467 | |
b94ec296 | 1468 | void blk_mq_run_hw_queues(struct request_queue *q, bool async) |
320ae51f JA |
1469 | { |
1470 | struct blk_mq_hw_ctx *hctx; | |
1471 | int i; | |
1472 | ||
1473 | queue_for_each_hw_ctx(q, hctx, i) { | |
79f720a7 | 1474 | if (blk_mq_hctx_stopped(hctx)) |
320ae51f JA |
1475 | continue; |
1476 | ||
b94ec296 | 1477 | blk_mq_run_hw_queue(hctx, async); |
320ae51f JA |
1478 | } |
1479 | } | |
b94ec296 | 1480 | EXPORT_SYMBOL(blk_mq_run_hw_queues); |
320ae51f | 1481 | |
fd001443 BVA |
1482 | /** |
1483 | * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped | |
1484 | * @q: request queue. | |
1485 | * | |
1486 | * The caller is responsible for serializing this function against | |
1487 | * blk_mq_{start,stop}_hw_queue(). | |
1488 | */ | |
1489 | bool blk_mq_queue_stopped(struct request_queue *q) | |
1490 | { | |
1491 | struct blk_mq_hw_ctx *hctx; | |
1492 | int i; | |
1493 | ||
1494 | queue_for_each_hw_ctx(q, hctx, i) | |
1495 | if (blk_mq_hctx_stopped(hctx)) | |
1496 | return true; | |
1497 | ||
1498 | return false; | |
1499 | } | |
1500 | EXPORT_SYMBOL(blk_mq_queue_stopped); | |
1501 | ||
39a70c76 ML |
1502 | /* |
1503 | * This function is often used for pausing .queue_rq() by driver when | |
1504 | * there isn't enough resource or some conditions aren't satisfied, and | |
4d606219 | 1505 | * BLK_STS_RESOURCE is usually returned. |
39a70c76 ML |
1506 | * |
1507 | * We do not guarantee that dispatch can be drained or blocked | |
1508 | * after blk_mq_stop_hw_queue() returns. Please use | |
1509 | * blk_mq_quiesce_queue() for that requirement. | |
1510 | */ | |
2719aa21 JA |
1511 | void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx) |
1512 | { | |
641a9ed6 | 1513 | cancel_delayed_work(&hctx->run_work); |
280d45f6 | 1514 | |
641a9ed6 | 1515 | set_bit(BLK_MQ_S_STOPPED, &hctx->state); |
2719aa21 | 1516 | } |
641a9ed6 | 1517 | EXPORT_SYMBOL(blk_mq_stop_hw_queue); |
2719aa21 | 1518 | |
39a70c76 ML |
1519 | /* |
1520 | * This function is often used for pausing .queue_rq() by driver when | |
1521 | * there isn't enough resource or some conditions aren't satisfied, and | |
4d606219 | 1522 | * BLK_STS_RESOURCE is usually returned. |
39a70c76 ML |
1523 | * |
1524 | * We do not guarantee that dispatch can be drained or blocked | |
1525 | * after blk_mq_stop_hw_queues() returns. Please use | |
1526 | * blk_mq_quiesce_queue() for that requirement. | |
1527 | */ | |
2719aa21 JA |
1528 | void blk_mq_stop_hw_queues(struct request_queue *q) |
1529 | { | |
641a9ed6 ML |
1530 | struct blk_mq_hw_ctx *hctx; |
1531 | int i; | |
1532 | ||
1533 | queue_for_each_hw_ctx(q, hctx, i) | |
1534 | blk_mq_stop_hw_queue(hctx); | |
280d45f6 CH |
1535 | } |
1536 | EXPORT_SYMBOL(blk_mq_stop_hw_queues); | |
1537 | ||
320ae51f JA |
1538 | void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx) |
1539 | { | |
1540 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
e4043dcf | 1541 | |
0ffbce80 | 1542 | blk_mq_run_hw_queue(hctx, false); |
320ae51f JA |
1543 | } |
1544 | EXPORT_SYMBOL(blk_mq_start_hw_queue); | |
1545 | ||
2f268556 CH |
1546 | void blk_mq_start_hw_queues(struct request_queue *q) |
1547 | { | |
1548 | struct blk_mq_hw_ctx *hctx; | |
1549 | int i; | |
1550 | ||
1551 | queue_for_each_hw_ctx(q, hctx, i) | |
1552 | blk_mq_start_hw_queue(hctx); | |
1553 | } | |
1554 | EXPORT_SYMBOL(blk_mq_start_hw_queues); | |
1555 | ||
ae911c5e JA |
1556 | void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) |
1557 | { | |
1558 | if (!blk_mq_hctx_stopped(hctx)) | |
1559 | return; | |
1560 | ||
1561 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
1562 | blk_mq_run_hw_queue(hctx, async); | |
1563 | } | |
1564 | EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue); | |
1565 | ||
1b4a3258 | 1566 | void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async) |
320ae51f JA |
1567 | { |
1568 | struct blk_mq_hw_ctx *hctx; | |
1569 | int i; | |
1570 | ||
ae911c5e JA |
1571 | queue_for_each_hw_ctx(q, hctx, i) |
1572 | blk_mq_start_stopped_hw_queue(hctx, async); | |
320ae51f JA |
1573 | } |
1574 | EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues); | |
1575 | ||
70f4db63 | 1576 | static void blk_mq_run_work_fn(struct work_struct *work) |
320ae51f JA |
1577 | { |
1578 | struct blk_mq_hw_ctx *hctx; | |
1579 | ||
9f993737 | 1580 | hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work); |
320ae51f | 1581 | |
21c6e939 | 1582 | /* |
15fe8a90 | 1583 | * If we are stopped, don't run the queue. |
21c6e939 | 1584 | */ |
15fe8a90 | 1585 | if (test_bit(BLK_MQ_S_STOPPED, &hctx->state)) |
21c6e939 | 1586 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); |
7587a5ae BVA |
1587 | |
1588 | __blk_mq_run_hw_queue(hctx); | |
1589 | } | |
1590 | ||
cfd0c552 | 1591 | static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx, |
cfd0c552 ML |
1592 | struct request *rq, |
1593 | bool at_head) | |
320ae51f | 1594 | { |
e57690fe JA |
1595 | struct blk_mq_ctx *ctx = rq->mq_ctx; |
1596 | ||
7b607814 BVA |
1597 | lockdep_assert_held(&ctx->lock); |
1598 | ||
01b983c9 JA |
1599 | trace_block_rq_insert(hctx->queue, rq); |
1600 | ||
72a0a36e CH |
1601 | if (at_head) |
1602 | list_add(&rq->queuelist, &ctx->rq_list); | |
1603 | else | |
1604 | list_add_tail(&rq->queuelist, &ctx->rq_list); | |
cfd0c552 | 1605 | } |
4bb659b1 | 1606 | |
2c3ad667 JA |
1607 | void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq, |
1608 | bool at_head) | |
cfd0c552 ML |
1609 | { |
1610 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
1611 | ||
7b607814 BVA |
1612 | lockdep_assert_held(&ctx->lock); |
1613 | ||
e57690fe | 1614 | __blk_mq_insert_req_list(hctx, rq, at_head); |
320ae51f | 1615 | blk_mq_hctx_mark_pending(hctx, ctx); |
320ae51f JA |
1616 | } |
1617 | ||
157f377b JA |
1618 | /* |
1619 | * Should only be used carefully, when the caller knows we want to | |
1620 | * bypass a potential IO scheduler on the target device. | |
1621 | */ | |
b0850297 | 1622 | void blk_mq_request_bypass_insert(struct request *rq, bool run_queue) |
157f377b JA |
1623 | { |
1624 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
1625 | struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(rq->q, ctx->cpu); | |
1626 | ||
1627 | spin_lock(&hctx->lock); | |
1628 | list_add_tail(&rq->queuelist, &hctx->dispatch); | |
1629 | spin_unlock(&hctx->lock); | |
1630 | ||
b0850297 ML |
1631 | if (run_queue) |
1632 | blk_mq_run_hw_queue(hctx, false); | |
157f377b JA |
1633 | } |
1634 | ||
bd166ef1 JA |
1635 | void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx, |
1636 | struct list_head *list) | |
320ae51f JA |
1637 | |
1638 | { | |
320ae51f JA |
1639 | /* |
1640 | * preemption doesn't flush plug list, so it's possible ctx->cpu is | |
1641 | * offline now | |
1642 | */ | |
1643 | spin_lock(&ctx->lock); | |
1644 | while (!list_empty(list)) { | |
1645 | struct request *rq; | |
1646 | ||
1647 | rq = list_first_entry(list, struct request, queuelist); | |
e57690fe | 1648 | BUG_ON(rq->mq_ctx != ctx); |
320ae51f | 1649 | list_del_init(&rq->queuelist); |
e57690fe | 1650 | __blk_mq_insert_req_list(hctx, rq, false); |
320ae51f | 1651 | } |
cfd0c552 | 1652 | blk_mq_hctx_mark_pending(hctx, ctx); |
320ae51f | 1653 | spin_unlock(&ctx->lock); |
320ae51f JA |
1654 | } |
1655 | ||
1656 | static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b) | |
1657 | { | |
1658 | struct request *rqa = container_of(a, struct request, queuelist); | |
1659 | struct request *rqb = container_of(b, struct request, queuelist); | |
1660 | ||
1661 | return !(rqa->mq_ctx < rqb->mq_ctx || | |
1662 | (rqa->mq_ctx == rqb->mq_ctx && | |
1663 | blk_rq_pos(rqa) < blk_rq_pos(rqb))); | |
1664 | } | |
1665 | ||
1666 | void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule) | |
1667 | { | |
1668 | struct blk_mq_ctx *this_ctx; | |
1669 | struct request_queue *this_q; | |
1670 | struct request *rq; | |
1671 | LIST_HEAD(list); | |
1672 | LIST_HEAD(ctx_list); | |
1673 | unsigned int depth; | |
1674 | ||
1675 | list_splice_init(&plug->mq_list, &list); | |
1676 | ||
1677 | list_sort(NULL, &list, plug_ctx_cmp); | |
1678 | ||
1679 | this_q = NULL; | |
1680 | this_ctx = NULL; | |
1681 | depth = 0; | |
1682 | ||
1683 | while (!list_empty(&list)) { | |
1684 | rq = list_entry_rq(list.next); | |
1685 | list_del_init(&rq->queuelist); | |
1686 | BUG_ON(!rq->q); | |
1687 | if (rq->mq_ctx != this_ctx) { | |
1688 | if (this_ctx) { | |
bd166ef1 JA |
1689 | trace_block_unplug(this_q, depth, from_schedule); |
1690 | blk_mq_sched_insert_requests(this_q, this_ctx, | |
1691 | &ctx_list, | |
1692 | from_schedule); | |
320ae51f JA |
1693 | } |
1694 | ||
1695 | this_ctx = rq->mq_ctx; | |
1696 | this_q = rq->q; | |
1697 | depth = 0; | |
1698 | } | |
1699 | ||
1700 | depth++; | |
1701 | list_add_tail(&rq->queuelist, &ctx_list); | |
1702 | } | |
1703 | ||
1704 | /* | |
1705 | * If 'this_ctx' is set, we know we have entries to complete | |
1706 | * on 'ctx_list'. Do those. | |
1707 | */ | |
1708 | if (this_ctx) { | |
bd166ef1 JA |
1709 | trace_block_unplug(this_q, depth, from_schedule); |
1710 | blk_mq_sched_insert_requests(this_q, this_ctx, &ctx_list, | |
1711 | from_schedule); | |
320ae51f JA |
1712 | } |
1713 | } | |
1714 | ||
1715 | static void blk_mq_bio_to_request(struct request *rq, struct bio *bio) | |
1716 | { | |
da8d7f07 | 1717 | blk_init_request_from_bio(rq, bio); |
4b570521 | 1718 | |
85acb3ba SL |
1719 | blk_rq_set_rl(rq, blk_get_rl(rq->q, bio)); |
1720 | ||
6e85eaf3 | 1721 | blk_account_io_start(rq, true); |
320ae51f JA |
1722 | } |
1723 | ||
ab42f35d ML |
1724 | static inline void blk_mq_queue_io(struct blk_mq_hw_ctx *hctx, |
1725 | struct blk_mq_ctx *ctx, | |
1726 | struct request *rq) | |
1727 | { | |
1728 | spin_lock(&ctx->lock); | |
1729 | __blk_mq_insert_request(hctx, rq, false); | |
1730 | spin_unlock(&ctx->lock); | |
07068d5b | 1731 | } |
14ec77f3 | 1732 | |
fd2d3326 JA |
1733 | static blk_qc_t request_to_qc_t(struct blk_mq_hw_ctx *hctx, struct request *rq) |
1734 | { | |
bd166ef1 JA |
1735 | if (rq->tag != -1) |
1736 | return blk_tag_to_qc_t(rq->tag, hctx->queue_num, false); | |
1737 | ||
1738 | return blk_tag_to_qc_t(rq->internal_tag, hctx->queue_num, true); | |
fd2d3326 JA |
1739 | } |
1740 | ||
0f95549c MS |
1741 | static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx, |
1742 | struct request *rq, | |
1743 | blk_qc_t *cookie) | |
f984df1f | 1744 | { |
f984df1f | 1745 | struct request_queue *q = rq->q; |
f984df1f SL |
1746 | struct blk_mq_queue_data bd = { |
1747 | .rq = rq, | |
d945a365 | 1748 | .last = true, |
f984df1f | 1749 | }; |
bd166ef1 | 1750 | blk_qc_t new_cookie; |
f06345ad | 1751 | blk_status_t ret; |
0f95549c MS |
1752 | |
1753 | new_cookie = request_to_qc_t(hctx, rq); | |
1754 | ||
1755 | /* | |
1756 | * For OK queue, we are done. For error, caller may kill it. | |
1757 | * Any other error (busy), just add it to our list as we | |
1758 | * previously would have done. | |
1759 | */ | |
1760 | ret = q->mq_ops->queue_rq(hctx, &bd); | |
1761 | switch (ret) { | |
1762 | case BLK_STS_OK: | |
1763 | *cookie = new_cookie; | |
1764 | break; | |
1765 | case BLK_STS_RESOURCE: | |
86ff7c2a | 1766 | case BLK_STS_DEV_RESOURCE: |
0f95549c MS |
1767 | __blk_mq_requeue_request(rq); |
1768 | break; | |
1769 | default: | |
1770 | *cookie = BLK_QC_T_NONE; | |
1771 | break; | |
1772 | } | |
1773 | ||
1774 | return ret; | |
1775 | } | |
1776 | ||
0f95549c MS |
1777 | static blk_status_t __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx, |
1778 | struct request *rq, | |
396eaf21 ML |
1779 | blk_qc_t *cookie, |
1780 | bool bypass_insert) | |
0f95549c MS |
1781 | { |
1782 | struct request_queue *q = rq->q; | |
d964f04a ML |
1783 | bool run_queue = true; |
1784 | ||
23d4ee19 ML |
1785 | /* |
1786 | * RCU or SRCU read lock is needed before checking quiesced flag. | |
1787 | * | |
1788 | * When queue is stopped or quiesced, ignore 'bypass_insert' from | |
c77ff7fd | 1789 | * blk_mq_request_issue_directly(), and return BLK_STS_OK to caller, |
23d4ee19 ML |
1790 | * and avoid driver to try to dispatch again. |
1791 | */ | |
f4560ffe | 1792 | if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)) { |
d964f04a | 1793 | run_queue = false; |
23d4ee19 | 1794 | bypass_insert = false; |
d964f04a ML |
1795 | goto insert; |
1796 | } | |
f984df1f | 1797 | |
396eaf21 | 1798 | if (q->elevator && !bypass_insert) |
2253efc8 BVA |
1799 | goto insert; |
1800 | ||
0bca799b | 1801 | if (!blk_mq_get_dispatch_budget(hctx)) |
bd166ef1 JA |
1802 | goto insert; |
1803 | ||
0bca799b ML |
1804 | if (!blk_mq_get_driver_tag(rq, NULL, false)) { |
1805 | blk_mq_put_dispatch_budget(hctx); | |
de148297 | 1806 | goto insert; |
88022d72 | 1807 | } |
de148297 | 1808 | |
0f95549c | 1809 | return __blk_mq_issue_directly(hctx, rq, cookie); |
2253efc8 | 1810 | insert: |
396eaf21 ML |
1811 | if (bypass_insert) |
1812 | return BLK_STS_RESOURCE; | |
0f95549c | 1813 | |
23d4ee19 | 1814 | blk_mq_sched_insert_request(rq, false, run_queue, false); |
0f95549c | 1815 | return BLK_STS_OK; |
f984df1f SL |
1816 | } |
1817 | ||
5eb6126e CH |
1818 | static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx, |
1819 | struct request *rq, blk_qc_t *cookie) | |
1820 | { | |
0f95549c | 1821 | blk_status_t ret; |
04ced159 | 1822 | int srcu_idx; |
bf4907c0 | 1823 | |
04ced159 | 1824 | might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING); |
bf4907c0 | 1825 | |
04ced159 | 1826 | hctx_lock(hctx, &srcu_idx); |
0f95549c | 1827 | |
396eaf21 | 1828 | ret = __blk_mq_try_issue_directly(hctx, rq, cookie, false); |
86ff7c2a | 1829 | if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE) |
23d4ee19 | 1830 | blk_mq_sched_insert_request(rq, false, true, false); |
0f95549c MS |
1831 | else if (ret != BLK_STS_OK) |
1832 | blk_mq_end_request(rq, ret); | |
1833 | ||
04ced159 | 1834 | hctx_unlock(hctx, srcu_idx); |
5eb6126e CH |
1835 | } |
1836 | ||
c77ff7fd | 1837 | blk_status_t blk_mq_request_issue_directly(struct request *rq) |
396eaf21 ML |
1838 | { |
1839 | blk_status_t ret; | |
1840 | int srcu_idx; | |
1841 | blk_qc_t unused_cookie; | |
1842 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
1843 | struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(rq->q, ctx->cpu); | |
1844 | ||
1845 | hctx_lock(hctx, &srcu_idx); | |
1846 | ret = __blk_mq_try_issue_directly(hctx, rq, &unused_cookie, true); | |
1847 | hctx_unlock(hctx, srcu_idx); | |
1848 | ||
1849 | return ret; | |
5eb6126e CH |
1850 | } |
1851 | ||
dece1635 | 1852 | static blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio) |
07068d5b | 1853 | { |
ef295ecf | 1854 | const int is_sync = op_is_sync(bio->bi_opf); |
f73f44eb | 1855 | const int is_flush_fua = op_is_flush(bio->bi_opf); |
5a797e00 | 1856 | struct blk_mq_alloc_data data = { .flags = 0 }; |
07068d5b | 1857 | struct request *rq; |
5eb6126e | 1858 | unsigned int request_count = 0; |
f984df1f | 1859 | struct blk_plug *plug; |
5b3f341f | 1860 | struct request *same_queue_rq = NULL; |
7b371636 | 1861 | blk_qc_t cookie; |
87760e5e | 1862 | unsigned int wb_acct; |
07068d5b JA |
1863 | |
1864 | blk_queue_bounce(q, &bio); | |
1865 | ||
af67c31f | 1866 | blk_queue_split(q, &bio); |
f36ea50c | 1867 | |
e23947bd | 1868 | if (!bio_integrity_prep(bio)) |
dece1635 | 1869 | return BLK_QC_T_NONE; |
07068d5b | 1870 | |
87c279e6 OS |
1871 | if (!is_flush_fua && !blk_queue_nomerges(q) && |
1872 | blk_attempt_plug_merge(q, bio, &request_count, &same_queue_rq)) | |
1873 | return BLK_QC_T_NONE; | |
f984df1f | 1874 | |
bd166ef1 JA |
1875 | if (blk_mq_sched_bio_merge(q, bio)) |
1876 | return BLK_QC_T_NONE; | |
1877 | ||
87760e5e JA |
1878 | wb_acct = wbt_wait(q->rq_wb, bio, NULL); |
1879 | ||
bd166ef1 JA |
1880 | trace_block_getrq(q, bio, bio->bi_opf); |
1881 | ||
d2c0d383 | 1882 | rq = blk_mq_get_request(q, bio, bio->bi_opf, &data); |
87760e5e JA |
1883 | if (unlikely(!rq)) { |
1884 | __wbt_done(q->rq_wb, wb_acct); | |
03a07c92 GR |
1885 | if (bio->bi_opf & REQ_NOWAIT) |
1886 | bio_wouldblock_error(bio); | |
dece1635 | 1887 | return BLK_QC_T_NONE; |
87760e5e JA |
1888 | } |
1889 | ||
a8a45941 | 1890 | wbt_track(rq, wb_acct); |
07068d5b | 1891 | |
fd2d3326 | 1892 | cookie = request_to_qc_t(data.hctx, rq); |
07068d5b | 1893 | |
f984df1f | 1894 | plug = current->plug; |
07068d5b | 1895 | if (unlikely(is_flush_fua)) { |
f984df1f | 1896 | blk_mq_put_ctx(data.ctx); |
07068d5b | 1897 | blk_mq_bio_to_request(rq, bio); |
923218f6 ML |
1898 | |
1899 | /* bypass scheduler for flush rq */ | |
1900 | blk_insert_flush(rq); | |
1901 | blk_mq_run_hw_queue(data.hctx, true); | |
a4d907b6 | 1902 | } else if (plug && q->nr_hw_queues == 1) { |
600271d9 SL |
1903 | struct request *last = NULL; |
1904 | ||
b00c53e8 | 1905 | blk_mq_put_ctx(data.ctx); |
e6c4438b | 1906 | blk_mq_bio_to_request(rq, bio); |
0a6219a9 ML |
1907 | |
1908 | /* | |
1909 | * @request_count may become stale because of schedule | |
1910 | * out, so check the list again. | |
1911 | */ | |
1912 | if (list_empty(&plug->mq_list)) | |
1913 | request_count = 0; | |
254d259d CH |
1914 | else if (blk_queue_nomerges(q)) |
1915 | request_count = blk_plug_queued_count(q); | |
1916 | ||
676d0607 | 1917 | if (!request_count) |
e6c4438b | 1918 | trace_block_plug(q); |
600271d9 SL |
1919 | else |
1920 | last = list_entry_rq(plug->mq_list.prev); | |
b094f89c | 1921 | |
600271d9 SL |
1922 | if (request_count >= BLK_MAX_REQUEST_COUNT || (last && |
1923 | blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) { | |
e6c4438b JM |
1924 | blk_flush_plug_list(plug, false); |
1925 | trace_block_plug(q); | |
320ae51f | 1926 | } |
b094f89c | 1927 | |
e6c4438b | 1928 | list_add_tail(&rq->queuelist, &plug->mq_list); |
2299722c | 1929 | } else if (plug && !blk_queue_nomerges(q)) { |
bd166ef1 | 1930 | blk_mq_bio_to_request(rq, bio); |
07068d5b | 1931 | |
07068d5b | 1932 | /* |
6a83e74d | 1933 | * We do limited plugging. If the bio can be merged, do that. |
f984df1f SL |
1934 | * Otherwise the existing request in the plug list will be |
1935 | * issued. So the plug list will have one request at most | |
2299722c CH |
1936 | * The plug list might get flushed before this. If that happens, |
1937 | * the plug list is empty, and same_queue_rq is invalid. | |
07068d5b | 1938 | */ |
2299722c CH |
1939 | if (list_empty(&plug->mq_list)) |
1940 | same_queue_rq = NULL; | |
1941 | if (same_queue_rq) | |
1942 | list_del_init(&same_queue_rq->queuelist); | |
1943 | list_add_tail(&rq->queuelist, &plug->mq_list); | |
1944 | ||
bf4907c0 JA |
1945 | blk_mq_put_ctx(data.ctx); |
1946 | ||
dad7a3be ML |
1947 | if (same_queue_rq) { |
1948 | data.hctx = blk_mq_map_queue(q, | |
1949 | same_queue_rq->mq_ctx->cpu); | |
2299722c CH |
1950 | blk_mq_try_issue_directly(data.hctx, same_queue_rq, |
1951 | &cookie); | |
dad7a3be | 1952 | } |
a4d907b6 | 1953 | } else if (q->nr_hw_queues > 1 && is_sync) { |
bf4907c0 | 1954 | blk_mq_put_ctx(data.ctx); |
2299722c | 1955 | blk_mq_bio_to_request(rq, bio); |
2299722c | 1956 | blk_mq_try_issue_directly(data.hctx, rq, &cookie); |
a4d907b6 | 1957 | } else if (q->elevator) { |
b00c53e8 | 1958 | blk_mq_put_ctx(data.ctx); |
bd166ef1 | 1959 | blk_mq_bio_to_request(rq, bio); |
9e97d295 | 1960 | blk_mq_sched_insert_request(rq, false, true, true); |
ab42f35d | 1961 | } else { |
b00c53e8 | 1962 | blk_mq_put_ctx(data.ctx); |
ab42f35d ML |
1963 | blk_mq_bio_to_request(rq, bio); |
1964 | blk_mq_queue_io(data.hctx, data.ctx, rq); | |
a4d907b6 | 1965 | blk_mq_run_hw_queue(data.hctx, true); |
ab42f35d | 1966 | } |
320ae51f | 1967 | |
7b371636 | 1968 | return cookie; |
320ae51f JA |
1969 | } |
1970 | ||
cc71a6f4 JA |
1971 | void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags, |
1972 | unsigned int hctx_idx) | |
95363efd | 1973 | { |
e9b267d9 | 1974 | struct page *page; |
320ae51f | 1975 | |
24d2f903 | 1976 | if (tags->rqs && set->ops->exit_request) { |
e9b267d9 | 1977 | int i; |
320ae51f | 1978 | |
24d2f903 | 1979 | for (i = 0; i < tags->nr_tags; i++) { |
2af8cbe3 JA |
1980 | struct request *rq = tags->static_rqs[i]; |
1981 | ||
1982 | if (!rq) | |
e9b267d9 | 1983 | continue; |
d6296d39 | 1984 | set->ops->exit_request(set, rq, hctx_idx); |
2af8cbe3 | 1985 | tags->static_rqs[i] = NULL; |
e9b267d9 | 1986 | } |
320ae51f | 1987 | } |
320ae51f | 1988 | |
24d2f903 CH |
1989 | while (!list_empty(&tags->page_list)) { |
1990 | page = list_first_entry(&tags->page_list, struct page, lru); | |
6753471c | 1991 | list_del_init(&page->lru); |
f75782e4 CM |
1992 | /* |
1993 | * Remove kmemleak object previously allocated in | |
1994 | * blk_mq_init_rq_map(). | |
1995 | */ | |
1996 | kmemleak_free(page_address(page)); | |
320ae51f JA |
1997 | __free_pages(page, page->private); |
1998 | } | |
cc71a6f4 | 1999 | } |
320ae51f | 2000 | |
cc71a6f4 JA |
2001 | void blk_mq_free_rq_map(struct blk_mq_tags *tags) |
2002 | { | |
24d2f903 | 2003 | kfree(tags->rqs); |
cc71a6f4 | 2004 | tags->rqs = NULL; |
2af8cbe3 JA |
2005 | kfree(tags->static_rqs); |
2006 | tags->static_rqs = NULL; | |
320ae51f | 2007 | |
24d2f903 | 2008 | blk_mq_free_tags(tags); |
320ae51f JA |
2009 | } |
2010 | ||
cc71a6f4 JA |
2011 | struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, |
2012 | unsigned int hctx_idx, | |
2013 | unsigned int nr_tags, | |
2014 | unsigned int reserved_tags) | |
320ae51f | 2015 | { |
24d2f903 | 2016 | struct blk_mq_tags *tags; |
59f082e4 | 2017 | int node; |
320ae51f | 2018 | |
59f082e4 SL |
2019 | node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx); |
2020 | if (node == NUMA_NO_NODE) | |
2021 | node = set->numa_node; | |
2022 | ||
2023 | tags = blk_mq_init_tags(nr_tags, reserved_tags, node, | |
24391c0d | 2024 | BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags)); |
24d2f903 CH |
2025 | if (!tags) |
2026 | return NULL; | |
320ae51f | 2027 | |
cc71a6f4 | 2028 | tags->rqs = kzalloc_node(nr_tags * sizeof(struct request *), |
36e1f3d1 | 2029 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, |
59f082e4 | 2030 | node); |
24d2f903 CH |
2031 | if (!tags->rqs) { |
2032 | blk_mq_free_tags(tags); | |
2033 | return NULL; | |
2034 | } | |
320ae51f | 2035 | |
2af8cbe3 JA |
2036 | tags->static_rqs = kzalloc_node(nr_tags * sizeof(struct request *), |
2037 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, | |
59f082e4 | 2038 | node); |
2af8cbe3 JA |
2039 | if (!tags->static_rqs) { |
2040 | kfree(tags->rqs); | |
2041 | blk_mq_free_tags(tags); | |
2042 | return NULL; | |
2043 | } | |
2044 | ||
cc71a6f4 JA |
2045 | return tags; |
2046 | } | |
2047 | ||
2048 | static size_t order_to_size(unsigned int order) | |
2049 | { | |
2050 | return (size_t)PAGE_SIZE << order; | |
2051 | } | |
2052 | ||
1d9bd516 TH |
2053 | static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq, |
2054 | unsigned int hctx_idx, int node) | |
2055 | { | |
2056 | int ret; | |
2057 | ||
2058 | if (set->ops->init_request) { | |
2059 | ret = set->ops->init_request(set, rq, hctx_idx, node); | |
2060 | if (ret) | |
2061 | return ret; | |
2062 | } | |
2063 | ||
2064 | seqcount_init(&rq->gstate_seq); | |
2065 | u64_stats_init(&rq->aborted_gstate_sync); | |
f4560231 JW |
2066 | /* |
2067 | * start gstate with gen 1 instead of 0, otherwise it will be equal | |
2068 | * to aborted_gstate, and be identified timed out by | |
2069 | * blk_mq_terminate_expired. | |
2070 | */ | |
2071 | WRITE_ONCE(rq->gstate, MQ_RQ_GEN_INC); | |
2072 | ||
1d9bd516 TH |
2073 | return 0; |
2074 | } | |
2075 | ||
cc71a6f4 JA |
2076 | int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags, |
2077 | unsigned int hctx_idx, unsigned int depth) | |
2078 | { | |
2079 | unsigned int i, j, entries_per_page, max_order = 4; | |
2080 | size_t rq_size, left; | |
59f082e4 SL |
2081 | int node; |
2082 | ||
2083 | node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx); | |
2084 | if (node == NUMA_NO_NODE) | |
2085 | node = set->numa_node; | |
cc71a6f4 JA |
2086 | |
2087 | INIT_LIST_HEAD(&tags->page_list); | |
2088 | ||
320ae51f JA |
2089 | /* |
2090 | * rq_size is the size of the request plus driver payload, rounded | |
2091 | * to the cacheline size | |
2092 | */ | |
24d2f903 | 2093 | rq_size = round_up(sizeof(struct request) + set->cmd_size, |
320ae51f | 2094 | cache_line_size()); |
cc71a6f4 | 2095 | left = rq_size * depth; |
320ae51f | 2096 | |
cc71a6f4 | 2097 | for (i = 0; i < depth; ) { |
320ae51f JA |
2098 | int this_order = max_order; |
2099 | struct page *page; | |
2100 | int to_do; | |
2101 | void *p; | |
2102 | ||
b3a834b1 | 2103 | while (this_order && left < order_to_size(this_order - 1)) |
320ae51f JA |
2104 | this_order--; |
2105 | ||
2106 | do { | |
59f082e4 | 2107 | page = alloc_pages_node(node, |
36e1f3d1 | 2108 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO, |
a5164405 | 2109 | this_order); |
320ae51f JA |
2110 | if (page) |
2111 | break; | |
2112 | if (!this_order--) | |
2113 | break; | |
2114 | if (order_to_size(this_order) < rq_size) | |
2115 | break; | |
2116 | } while (1); | |
2117 | ||
2118 | if (!page) | |
24d2f903 | 2119 | goto fail; |
320ae51f JA |
2120 | |
2121 | page->private = this_order; | |
24d2f903 | 2122 | list_add_tail(&page->lru, &tags->page_list); |
320ae51f JA |
2123 | |
2124 | p = page_address(page); | |
f75782e4 CM |
2125 | /* |
2126 | * Allow kmemleak to scan these pages as they contain pointers | |
2127 | * to additional allocations like via ops->init_request(). | |
2128 | */ | |
36e1f3d1 | 2129 | kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO); |
320ae51f | 2130 | entries_per_page = order_to_size(this_order) / rq_size; |
cc71a6f4 | 2131 | to_do = min(entries_per_page, depth - i); |
320ae51f JA |
2132 | left -= to_do * rq_size; |
2133 | for (j = 0; j < to_do; j++) { | |
2af8cbe3 JA |
2134 | struct request *rq = p; |
2135 | ||
2136 | tags->static_rqs[i] = rq; | |
1d9bd516 TH |
2137 | if (blk_mq_init_request(set, rq, hctx_idx, node)) { |
2138 | tags->static_rqs[i] = NULL; | |
2139 | goto fail; | |
e9b267d9 CH |
2140 | } |
2141 | ||
320ae51f JA |
2142 | p += rq_size; |
2143 | i++; | |
2144 | } | |
2145 | } | |
cc71a6f4 | 2146 | return 0; |
320ae51f | 2147 | |
24d2f903 | 2148 | fail: |
cc71a6f4 JA |
2149 | blk_mq_free_rqs(set, tags, hctx_idx); |
2150 | return -ENOMEM; | |
320ae51f JA |
2151 | } |
2152 | ||
e57690fe JA |
2153 | /* |
2154 | * 'cpu' is going away. splice any existing rq_list entries from this | |
2155 | * software queue to the hw queue dispatch list, and ensure that it | |
2156 | * gets run. | |
2157 | */ | |
9467f859 | 2158 | static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node) |
484b4061 | 2159 | { |
9467f859 | 2160 | struct blk_mq_hw_ctx *hctx; |
484b4061 JA |
2161 | struct blk_mq_ctx *ctx; |
2162 | LIST_HEAD(tmp); | |
2163 | ||
9467f859 | 2164 | hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead); |
e57690fe | 2165 | ctx = __blk_mq_get_ctx(hctx->queue, cpu); |
484b4061 JA |
2166 | |
2167 | spin_lock(&ctx->lock); | |
2168 | if (!list_empty(&ctx->rq_list)) { | |
2169 | list_splice_init(&ctx->rq_list, &tmp); | |
2170 | blk_mq_hctx_clear_pending(hctx, ctx); | |
2171 | } | |
2172 | spin_unlock(&ctx->lock); | |
2173 | ||
2174 | if (list_empty(&tmp)) | |
9467f859 | 2175 | return 0; |
484b4061 | 2176 | |
e57690fe JA |
2177 | spin_lock(&hctx->lock); |
2178 | list_splice_tail_init(&tmp, &hctx->dispatch); | |
2179 | spin_unlock(&hctx->lock); | |
484b4061 JA |
2180 | |
2181 | blk_mq_run_hw_queue(hctx, true); | |
9467f859 | 2182 | return 0; |
484b4061 JA |
2183 | } |
2184 | ||
9467f859 | 2185 | static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx) |
484b4061 | 2186 | { |
9467f859 TG |
2187 | cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD, |
2188 | &hctx->cpuhp_dead); | |
484b4061 JA |
2189 | } |
2190 | ||
c3b4afca | 2191 | /* hctx->ctxs will be freed in queue's release handler */ |
08e98fc6 ML |
2192 | static void blk_mq_exit_hctx(struct request_queue *q, |
2193 | struct blk_mq_tag_set *set, | |
2194 | struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx) | |
2195 | { | |
9c1051aa OS |
2196 | blk_mq_debugfs_unregister_hctx(hctx); |
2197 | ||
8ab0b7dc ML |
2198 | if (blk_mq_hw_queue_mapped(hctx)) |
2199 | blk_mq_tag_idle(hctx); | |
08e98fc6 | 2200 | |
f70ced09 | 2201 | if (set->ops->exit_request) |
d6296d39 | 2202 | set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx); |
f70ced09 | 2203 | |
93252632 OS |
2204 | blk_mq_sched_exit_hctx(q, hctx, hctx_idx); |
2205 | ||
08e98fc6 ML |
2206 | if (set->ops->exit_hctx) |
2207 | set->ops->exit_hctx(hctx, hctx_idx); | |
2208 | ||
6a83e74d | 2209 | if (hctx->flags & BLK_MQ_F_BLOCKING) |
05707b64 | 2210 | cleanup_srcu_struct(hctx->srcu); |
6a83e74d | 2211 | |
9467f859 | 2212 | blk_mq_remove_cpuhp(hctx); |
f70ced09 | 2213 | blk_free_flush_queue(hctx->fq); |
88459642 | 2214 | sbitmap_free(&hctx->ctx_map); |
08e98fc6 ML |
2215 | } |
2216 | ||
624dbe47 ML |
2217 | static void blk_mq_exit_hw_queues(struct request_queue *q, |
2218 | struct blk_mq_tag_set *set, int nr_queue) | |
2219 | { | |
2220 | struct blk_mq_hw_ctx *hctx; | |
2221 | unsigned int i; | |
2222 | ||
2223 | queue_for_each_hw_ctx(q, hctx, i) { | |
2224 | if (i == nr_queue) | |
2225 | break; | |
08e98fc6 | 2226 | blk_mq_exit_hctx(q, set, hctx, i); |
624dbe47 | 2227 | } |
624dbe47 ML |
2228 | } |
2229 | ||
08e98fc6 ML |
2230 | static int blk_mq_init_hctx(struct request_queue *q, |
2231 | struct blk_mq_tag_set *set, | |
2232 | struct blk_mq_hw_ctx *hctx, unsigned hctx_idx) | |
320ae51f | 2233 | { |
08e98fc6 ML |
2234 | int node; |
2235 | ||
2236 | node = hctx->numa_node; | |
2237 | if (node == NUMA_NO_NODE) | |
2238 | node = hctx->numa_node = set->numa_node; | |
2239 | ||
9f993737 | 2240 | INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn); |
08e98fc6 ML |
2241 | spin_lock_init(&hctx->lock); |
2242 | INIT_LIST_HEAD(&hctx->dispatch); | |
2243 | hctx->queue = q; | |
2404e607 | 2244 | hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED; |
08e98fc6 | 2245 | |
9467f859 | 2246 | cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead); |
08e98fc6 ML |
2247 | |
2248 | hctx->tags = set->tags[hctx_idx]; | |
320ae51f JA |
2249 | |
2250 | /* | |
08e98fc6 ML |
2251 | * Allocate space for all possible cpus to avoid allocation at |
2252 | * runtime | |
320ae51f | 2253 | */ |
d904bfa7 | 2254 | hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *), |
08e98fc6 ML |
2255 | GFP_KERNEL, node); |
2256 | if (!hctx->ctxs) | |
2257 | goto unregister_cpu_notifier; | |
320ae51f | 2258 | |
88459642 OS |
2259 | if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8), GFP_KERNEL, |
2260 | node)) | |
08e98fc6 | 2261 | goto free_ctxs; |
320ae51f | 2262 | |
08e98fc6 | 2263 | hctx->nr_ctx = 0; |
320ae51f | 2264 | |
eb619fdb JA |
2265 | init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake); |
2266 | INIT_LIST_HEAD(&hctx->dispatch_wait.entry); | |
2267 | ||
08e98fc6 ML |
2268 | if (set->ops->init_hctx && |
2269 | set->ops->init_hctx(hctx, set->driver_data, hctx_idx)) | |
2270 | goto free_bitmap; | |
320ae51f | 2271 | |
93252632 OS |
2272 | if (blk_mq_sched_init_hctx(q, hctx, hctx_idx)) |
2273 | goto exit_hctx; | |
2274 | ||
f70ced09 ML |
2275 | hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size); |
2276 | if (!hctx->fq) | |
93252632 | 2277 | goto sched_exit_hctx; |
320ae51f | 2278 | |
1d9bd516 | 2279 | if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx, node)) |
f70ced09 | 2280 | goto free_fq; |
320ae51f | 2281 | |
6a83e74d | 2282 | if (hctx->flags & BLK_MQ_F_BLOCKING) |
05707b64 | 2283 | init_srcu_struct(hctx->srcu); |
6a83e74d | 2284 | |
9c1051aa OS |
2285 | blk_mq_debugfs_register_hctx(q, hctx); |
2286 | ||
08e98fc6 | 2287 | return 0; |
320ae51f | 2288 | |
f70ced09 ML |
2289 | free_fq: |
2290 | kfree(hctx->fq); | |
93252632 OS |
2291 | sched_exit_hctx: |
2292 | blk_mq_sched_exit_hctx(q, hctx, hctx_idx); | |
f70ced09 ML |
2293 | exit_hctx: |
2294 | if (set->ops->exit_hctx) | |
2295 | set->ops->exit_hctx(hctx, hctx_idx); | |
08e98fc6 | 2296 | free_bitmap: |
88459642 | 2297 | sbitmap_free(&hctx->ctx_map); |
08e98fc6 ML |
2298 | free_ctxs: |
2299 | kfree(hctx->ctxs); | |
2300 | unregister_cpu_notifier: | |
9467f859 | 2301 | blk_mq_remove_cpuhp(hctx); |
08e98fc6 ML |
2302 | return -1; |
2303 | } | |
320ae51f | 2304 | |
320ae51f JA |
2305 | static void blk_mq_init_cpu_queues(struct request_queue *q, |
2306 | unsigned int nr_hw_queues) | |
2307 | { | |
2308 | unsigned int i; | |
2309 | ||
2310 | for_each_possible_cpu(i) { | |
2311 | struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i); | |
2312 | struct blk_mq_hw_ctx *hctx; | |
2313 | ||
320ae51f JA |
2314 | __ctx->cpu = i; |
2315 | spin_lock_init(&__ctx->lock); | |
2316 | INIT_LIST_HEAD(&__ctx->rq_list); | |
2317 | __ctx->queue = q; | |
2318 | ||
320ae51f JA |
2319 | /* |
2320 | * Set local node, IFF we have more than one hw queue. If | |
2321 | * not, we remain on the home node of the device | |
2322 | */ | |
20e4d813 | 2323 | hctx = blk_mq_map_queue(q, i); |
320ae51f | 2324 | if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE) |
bffed457 | 2325 | hctx->numa_node = local_memory_node(cpu_to_node(i)); |
320ae51f JA |
2326 | } |
2327 | } | |
2328 | ||
cc71a6f4 JA |
2329 | static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, int hctx_idx) |
2330 | { | |
2331 | int ret = 0; | |
2332 | ||
2333 | set->tags[hctx_idx] = blk_mq_alloc_rq_map(set, hctx_idx, | |
2334 | set->queue_depth, set->reserved_tags); | |
2335 | if (!set->tags[hctx_idx]) | |
2336 | return false; | |
2337 | ||
2338 | ret = blk_mq_alloc_rqs(set, set->tags[hctx_idx], hctx_idx, | |
2339 | set->queue_depth); | |
2340 | if (!ret) | |
2341 | return true; | |
2342 | ||
2343 | blk_mq_free_rq_map(set->tags[hctx_idx]); | |
2344 | set->tags[hctx_idx] = NULL; | |
2345 | return false; | |
2346 | } | |
2347 | ||
2348 | static void blk_mq_free_map_and_requests(struct blk_mq_tag_set *set, | |
2349 | unsigned int hctx_idx) | |
2350 | { | |
bd166ef1 JA |
2351 | if (set->tags[hctx_idx]) { |
2352 | blk_mq_free_rqs(set, set->tags[hctx_idx], hctx_idx); | |
2353 | blk_mq_free_rq_map(set->tags[hctx_idx]); | |
2354 | set->tags[hctx_idx] = NULL; | |
2355 | } | |
cc71a6f4 JA |
2356 | } |
2357 | ||
4b855ad3 | 2358 | static void blk_mq_map_swqueue(struct request_queue *q) |
320ae51f | 2359 | { |
4412efec | 2360 | unsigned int i, hctx_idx; |
320ae51f JA |
2361 | struct blk_mq_hw_ctx *hctx; |
2362 | struct blk_mq_ctx *ctx; | |
2a34c087 | 2363 | struct blk_mq_tag_set *set = q->tag_set; |
320ae51f | 2364 | |
60de074b AM |
2365 | /* |
2366 | * Avoid others reading imcomplete hctx->cpumask through sysfs | |
2367 | */ | |
2368 | mutex_lock(&q->sysfs_lock); | |
2369 | ||
320ae51f | 2370 | queue_for_each_hw_ctx(q, hctx, i) { |
e4043dcf | 2371 | cpumask_clear(hctx->cpumask); |
320ae51f JA |
2372 | hctx->nr_ctx = 0; |
2373 | } | |
2374 | ||
2375 | /* | |
4b855ad3 | 2376 | * Map software to hardware queues. |
4412efec ML |
2377 | * |
2378 | * If the cpu isn't present, the cpu is mapped to first hctx. | |
320ae51f | 2379 | */ |
20e4d813 | 2380 | for_each_possible_cpu(i) { |
4412efec ML |
2381 | hctx_idx = q->mq_map[i]; |
2382 | /* unmapped hw queue can be remapped after CPU topo changed */ | |
2383 | if (!set->tags[hctx_idx] && | |
2384 | !__blk_mq_alloc_rq_map(set, hctx_idx)) { | |
2385 | /* | |
2386 | * If tags initialization fail for some hctx, | |
2387 | * that hctx won't be brought online. In this | |
2388 | * case, remap the current ctx to hctx[0] which | |
2389 | * is guaranteed to always have tags allocated | |
2390 | */ | |
2391 | q->mq_map[i] = 0; | |
2392 | } | |
2393 | ||
897bb0c7 | 2394 | ctx = per_cpu_ptr(q->queue_ctx, i); |
7d7e0f90 | 2395 | hctx = blk_mq_map_queue(q, i); |
868f2f0b | 2396 | |
e4043dcf | 2397 | cpumask_set_cpu(i, hctx->cpumask); |
320ae51f JA |
2398 | ctx->index_hw = hctx->nr_ctx; |
2399 | hctx->ctxs[hctx->nr_ctx++] = ctx; | |
2400 | } | |
506e931f | 2401 | |
60de074b AM |
2402 | mutex_unlock(&q->sysfs_lock); |
2403 | ||
506e931f | 2404 | queue_for_each_hw_ctx(q, hctx, i) { |
4412efec ML |
2405 | /* |
2406 | * If no software queues are mapped to this hardware queue, | |
2407 | * disable it and free the request entries. | |
2408 | */ | |
2409 | if (!hctx->nr_ctx) { | |
2410 | /* Never unmap queue 0. We need it as a | |
2411 | * fallback in case of a new remap fails | |
2412 | * allocation | |
2413 | */ | |
2414 | if (i && set->tags[i]) | |
2415 | blk_mq_free_map_and_requests(set, i); | |
2416 | ||
2417 | hctx->tags = NULL; | |
2418 | continue; | |
2419 | } | |
484b4061 | 2420 | |
2a34c087 ML |
2421 | hctx->tags = set->tags[i]; |
2422 | WARN_ON(!hctx->tags); | |
2423 | ||
889fa31f CY |
2424 | /* |
2425 | * Set the map size to the number of mapped software queues. | |
2426 | * This is more accurate and more efficient than looping | |
2427 | * over all possibly mapped software queues. | |
2428 | */ | |
88459642 | 2429 | sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx); |
889fa31f | 2430 | |
484b4061 JA |
2431 | /* |
2432 | * Initialize batch roundrobin counts | |
2433 | */ | |
f82ddf19 | 2434 | hctx->next_cpu = blk_mq_first_mapped_cpu(hctx); |
506e931f JA |
2435 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; |
2436 | } | |
320ae51f JA |
2437 | } |
2438 | ||
8e8320c9 JA |
2439 | /* |
2440 | * Caller needs to ensure that we're either frozen/quiesced, or that | |
2441 | * the queue isn't live yet. | |
2442 | */ | |
2404e607 | 2443 | static void queue_set_hctx_shared(struct request_queue *q, bool shared) |
0d2602ca JA |
2444 | { |
2445 | struct blk_mq_hw_ctx *hctx; | |
0d2602ca JA |
2446 | int i; |
2447 | ||
2404e607 | 2448 | queue_for_each_hw_ctx(q, hctx, i) { |
8e8320c9 JA |
2449 | if (shared) { |
2450 | if (test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state)) | |
2451 | atomic_inc(&q->shared_hctx_restart); | |
2404e607 | 2452 | hctx->flags |= BLK_MQ_F_TAG_SHARED; |
8e8320c9 JA |
2453 | } else { |
2454 | if (test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state)) | |
2455 | atomic_dec(&q->shared_hctx_restart); | |
2404e607 | 2456 | hctx->flags &= ~BLK_MQ_F_TAG_SHARED; |
8e8320c9 | 2457 | } |
2404e607 JM |
2458 | } |
2459 | } | |
2460 | ||
8e8320c9 JA |
2461 | static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set, |
2462 | bool shared) | |
2404e607 JM |
2463 | { |
2464 | struct request_queue *q; | |
0d2602ca | 2465 | |
705cda97 BVA |
2466 | lockdep_assert_held(&set->tag_list_lock); |
2467 | ||
0d2602ca JA |
2468 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
2469 | blk_mq_freeze_queue(q); | |
2404e607 | 2470 | queue_set_hctx_shared(q, shared); |
0d2602ca JA |
2471 | blk_mq_unfreeze_queue(q); |
2472 | } | |
2473 | } | |
2474 | ||
2475 | static void blk_mq_del_queue_tag_set(struct request_queue *q) | |
2476 | { | |
2477 | struct blk_mq_tag_set *set = q->tag_set; | |
2478 | ||
0d2602ca | 2479 | mutex_lock(&set->tag_list_lock); |
705cda97 BVA |
2480 | list_del_rcu(&q->tag_set_list); |
2481 | INIT_LIST_HEAD(&q->tag_set_list); | |
2404e607 JM |
2482 | if (list_is_singular(&set->tag_list)) { |
2483 | /* just transitioned to unshared */ | |
2484 | set->flags &= ~BLK_MQ_F_TAG_SHARED; | |
2485 | /* update existing queue */ | |
2486 | blk_mq_update_tag_set_depth(set, false); | |
2487 | } | |
0d2602ca | 2488 | mutex_unlock(&set->tag_list_lock); |
705cda97 BVA |
2489 | |
2490 | synchronize_rcu(); | |
0d2602ca JA |
2491 | } |
2492 | ||
2493 | static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set, | |
2494 | struct request_queue *q) | |
2495 | { | |
2496 | q->tag_set = set; | |
2497 | ||
2498 | mutex_lock(&set->tag_list_lock); | |
2404e607 | 2499 | |
ff821d27 JA |
2500 | /* |
2501 | * Check to see if we're transitioning to shared (from 1 to 2 queues). | |
2502 | */ | |
2503 | if (!list_empty(&set->tag_list) && | |
2504 | !(set->flags & BLK_MQ_F_TAG_SHARED)) { | |
2404e607 JM |
2505 | set->flags |= BLK_MQ_F_TAG_SHARED; |
2506 | /* update existing queue */ | |
2507 | blk_mq_update_tag_set_depth(set, true); | |
2508 | } | |
2509 | if (set->flags & BLK_MQ_F_TAG_SHARED) | |
2510 | queue_set_hctx_shared(q, true); | |
705cda97 | 2511 | list_add_tail_rcu(&q->tag_set_list, &set->tag_list); |
2404e607 | 2512 | |
0d2602ca JA |
2513 | mutex_unlock(&set->tag_list_lock); |
2514 | } | |
2515 | ||
e09aae7e ML |
2516 | /* |
2517 | * It is the actual release handler for mq, but we do it from | |
2518 | * request queue's release handler for avoiding use-after-free | |
2519 | * and headache because q->mq_kobj shouldn't have been introduced, | |
2520 | * but we can't group ctx/kctx kobj without it. | |
2521 | */ | |
2522 | void blk_mq_release(struct request_queue *q) | |
2523 | { | |
2524 | struct blk_mq_hw_ctx *hctx; | |
2525 | unsigned int i; | |
2526 | ||
2527 | /* hctx kobj stays in hctx */ | |
c3b4afca ML |
2528 | queue_for_each_hw_ctx(q, hctx, i) { |
2529 | if (!hctx) | |
2530 | continue; | |
6c8b232e | 2531 | kobject_put(&hctx->kobj); |
c3b4afca | 2532 | } |
e09aae7e | 2533 | |
a723bab3 AM |
2534 | q->mq_map = NULL; |
2535 | ||
e09aae7e ML |
2536 | kfree(q->queue_hw_ctx); |
2537 | ||
7ea5fe31 ML |
2538 | /* |
2539 | * release .mq_kobj and sw queue's kobject now because | |
2540 | * both share lifetime with request queue. | |
2541 | */ | |
2542 | blk_mq_sysfs_deinit(q); | |
2543 | ||
e09aae7e ML |
2544 | free_percpu(q->queue_ctx); |
2545 | } | |
2546 | ||
24d2f903 | 2547 | struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set) |
b62c21b7 MS |
2548 | { |
2549 | struct request_queue *uninit_q, *q; | |
2550 | ||
5ee0524b | 2551 | uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node, NULL); |
b62c21b7 MS |
2552 | if (!uninit_q) |
2553 | return ERR_PTR(-ENOMEM); | |
2554 | ||
2555 | q = blk_mq_init_allocated_queue(set, uninit_q); | |
2556 | if (IS_ERR(q)) | |
2557 | blk_cleanup_queue(uninit_q); | |
2558 | ||
2559 | return q; | |
2560 | } | |
2561 | EXPORT_SYMBOL(blk_mq_init_queue); | |
2562 | ||
07319678 BVA |
2563 | static int blk_mq_hw_ctx_size(struct blk_mq_tag_set *tag_set) |
2564 | { | |
2565 | int hw_ctx_size = sizeof(struct blk_mq_hw_ctx); | |
2566 | ||
05707b64 | 2567 | BUILD_BUG_ON(ALIGN(offsetof(struct blk_mq_hw_ctx, srcu), |
07319678 BVA |
2568 | __alignof__(struct blk_mq_hw_ctx)) != |
2569 | sizeof(struct blk_mq_hw_ctx)); | |
2570 | ||
2571 | if (tag_set->flags & BLK_MQ_F_BLOCKING) | |
2572 | hw_ctx_size += sizeof(struct srcu_struct); | |
2573 | ||
2574 | return hw_ctx_size; | |
2575 | } | |
2576 | ||
868f2f0b KB |
2577 | static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set, |
2578 | struct request_queue *q) | |
320ae51f | 2579 | { |
868f2f0b KB |
2580 | int i, j; |
2581 | struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx; | |
f14bbe77 | 2582 | |
868f2f0b | 2583 | blk_mq_sysfs_unregister(q); |
fb350e0a ML |
2584 | |
2585 | /* protect against switching io scheduler */ | |
2586 | mutex_lock(&q->sysfs_lock); | |
24d2f903 | 2587 | for (i = 0; i < set->nr_hw_queues; i++) { |
868f2f0b | 2588 | int node; |
f14bbe77 | 2589 | |
868f2f0b KB |
2590 | if (hctxs[i]) |
2591 | continue; | |
2592 | ||
2593 | node = blk_mq_hw_queue_to_node(q->mq_map, i); | |
07319678 | 2594 | hctxs[i] = kzalloc_node(blk_mq_hw_ctx_size(set), |
cdef54dd | 2595 | GFP_KERNEL, node); |
320ae51f | 2596 | if (!hctxs[i]) |
868f2f0b | 2597 | break; |
320ae51f | 2598 | |
a86073e4 | 2599 | if (!zalloc_cpumask_var_node(&hctxs[i]->cpumask, GFP_KERNEL, |
868f2f0b KB |
2600 | node)) { |
2601 | kfree(hctxs[i]); | |
2602 | hctxs[i] = NULL; | |
2603 | break; | |
2604 | } | |
e4043dcf | 2605 | |
0d2602ca | 2606 | atomic_set(&hctxs[i]->nr_active, 0); |
f14bbe77 | 2607 | hctxs[i]->numa_node = node; |
320ae51f | 2608 | hctxs[i]->queue_num = i; |
868f2f0b KB |
2609 | |
2610 | if (blk_mq_init_hctx(q, set, hctxs[i], i)) { | |
2611 | free_cpumask_var(hctxs[i]->cpumask); | |
2612 | kfree(hctxs[i]); | |
2613 | hctxs[i] = NULL; | |
2614 | break; | |
2615 | } | |
2616 | blk_mq_hctx_kobj_init(hctxs[i]); | |
320ae51f | 2617 | } |
868f2f0b KB |
2618 | for (j = i; j < q->nr_hw_queues; j++) { |
2619 | struct blk_mq_hw_ctx *hctx = hctxs[j]; | |
2620 | ||
2621 | if (hctx) { | |
cc71a6f4 JA |
2622 | if (hctx->tags) |
2623 | blk_mq_free_map_and_requests(set, j); | |
868f2f0b | 2624 | blk_mq_exit_hctx(q, set, hctx, j); |
868f2f0b | 2625 | kobject_put(&hctx->kobj); |
868f2f0b KB |
2626 | hctxs[j] = NULL; |
2627 | ||
2628 | } | |
2629 | } | |
2630 | q->nr_hw_queues = i; | |
fb350e0a | 2631 | mutex_unlock(&q->sysfs_lock); |
868f2f0b KB |
2632 | blk_mq_sysfs_register(q); |
2633 | } | |
2634 | ||
2635 | struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set, | |
2636 | struct request_queue *q) | |
2637 | { | |
66841672 ML |
2638 | /* mark the queue as mq asap */ |
2639 | q->mq_ops = set->ops; | |
2640 | ||
34dbad5d | 2641 | q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn, |
720b8ccc SB |
2642 | blk_mq_poll_stats_bkt, |
2643 | BLK_MQ_POLL_STATS_BKTS, q); | |
34dbad5d OS |
2644 | if (!q->poll_cb) |
2645 | goto err_exit; | |
2646 | ||
868f2f0b KB |
2647 | q->queue_ctx = alloc_percpu(struct blk_mq_ctx); |
2648 | if (!q->queue_ctx) | |
c7de5726 | 2649 | goto err_exit; |
868f2f0b | 2650 | |
737f98cf ML |
2651 | /* init q->mq_kobj and sw queues' kobjects */ |
2652 | blk_mq_sysfs_init(q); | |
2653 | ||
868f2f0b KB |
2654 | q->queue_hw_ctx = kzalloc_node(nr_cpu_ids * sizeof(*(q->queue_hw_ctx)), |
2655 | GFP_KERNEL, set->numa_node); | |
2656 | if (!q->queue_hw_ctx) | |
2657 | goto err_percpu; | |
2658 | ||
bdd17e75 | 2659 | q->mq_map = set->mq_map; |
868f2f0b KB |
2660 | |
2661 | blk_mq_realloc_hw_ctxs(set, q); | |
2662 | if (!q->nr_hw_queues) | |
2663 | goto err_hctxs; | |
320ae51f | 2664 | |
287922eb | 2665 | INIT_WORK(&q->timeout_work, blk_mq_timeout_work); |
e56f698b | 2666 | blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ); |
320ae51f JA |
2667 | |
2668 | q->nr_queues = nr_cpu_ids; | |
320ae51f | 2669 | |
94eddfbe | 2670 | q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT; |
320ae51f | 2671 | |
05f1dd53 | 2672 | if (!(set->flags & BLK_MQ_F_SG_MERGE)) |
f78bac2c | 2673 | queue_flag_set_unlocked(QUEUE_FLAG_NO_SG_MERGE, q); |
05f1dd53 | 2674 | |
1be036e9 CH |
2675 | q->sg_reserved_size = INT_MAX; |
2676 | ||
2849450a | 2677 | INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work); |
6fca6a61 CH |
2678 | INIT_LIST_HEAD(&q->requeue_list); |
2679 | spin_lock_init(&q->requeue_lock); | |
2680 | ||
254d259d | 2681 | blk_queue_make_request(q, blk_mq_make_request); |
ea435e1b CH |
2682 | if (q->mq_ops->poll) |
2683 | q->poll_fn = blk_mq_poll; | |
07068d5b | 2684 | |
eba71768 JA |
2685 | /* |
2686 | * Do this after blk_queue_make_request() overrides it... | |
2687 | */ | |
2688 | q->nr_requests = set->queue_depth; | |
2689 | ||
64f1c21e JA |
2690 | /* |
2691 | * Default to classic polling | |
2692 | */ | |
2693 | q->poll_nsec = -1; | |
2694 | ||
24d2f903 CH |
2695 | if (set->ops->complete) |
2696 | blk_queue_softirq_done(q, set->ops->complete); | |
30a91cb4 | 2697 | |
24d2f903 | 2698 | blk_mq_init_cpu_queues(q, set->nr_hw_queues); |
0d2602ca | 2699 | blk_mq_add_queue_tag_set(set, q); |
4b855ad3 | 2700 | blk_mq_map_swqueue(q); |
4593fdbe | 2701 | |
d3484991 JA |
2702 | if (!(set->flags & BLK_MQ_F_NO_SCHED)) { |
2703 | int ret; | |
2704 | ||
2705 | ret = blk_mq_sched_init(q); | |
2706 | if (ret) | |
2707 | return ERR_PTR(ret); | |
2708 | } | |
2709 | ||
320ae51f | 2710 | return q; |
18741986 | 2711 | |
320ae51f | 2712 | err_hctxs: |
868f2f0b | 2713 | kfree(q->queue_hw_ctx); |
320ae51f | 2714 | err_percpu: |
868f2f0b | 2715 | free_percpu(q->queue_ctx); |
c7de5726 ML |
2716 | err_exit: |
2717 | q->mq_ops = NULL; | |
320ae51f JA |
2718 | return ERR_PTR(-ENOMEM); |
2719 | } | |
b62c21b7 | 2720 | EXPORT_SYMBOL(blk_mq_init_allocated_queue); |
320ae51f JA |
2721 | |
2722 | void blk_mq_free_queue(struct request_queue *q) | |
2723 | { | |
624dbe47 | 2724 | struct blk_mq_tag_set *set = q->tag_set; |
320ae51f | 2725 | |
0d2602ca | 2726 | blk_mq_del_queue_tag_set(q); |
624dbe47 | 2727 | blk_mq_exit_hw_queues(q, set, set->nr_hw_queues); |
320ae51f | 2728 | } |
320ae51f JA |
2729 | |
2730 | /* Basically redo blk_mq_init_queue with queue frozen */ | |
4b855ad3 | 2731 | static void blk_mq_queue_reinit(struct request_queue *q) |
320ae51f | 2732 | { |
4ecd4fef | 2733 | WARN_ON_ONCE(!atomic_read(&q->mq_freeze_depth)); |
320ae51f | 2734 | |
9c1051aa | 2735 | blk_mq_debugfs_unregister_hctxs(q); |
67aec14c JA |
2736 | blk_mq_sysfs_unregister(q); |
2737 | ||
320ae51f JA |
2738 | /* |
2739 | * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe | |
ff821d27 JA |
2740 | * we should change hctx numa_node according to the new topology (this |
2741 | * involves freeing and re-allocating memory, worth doing?) | |
320ae51f | 2742 | */ |
4b855ad3 | 2743 | blk_mq_map_swqueue(q); |
320ae51f | 2744 | |
67aec14c | 2745 | blk_mq_sysfs_register(q); |
9c1051aa | 2746 | blk_mq_debugfs_register_hctxs(q); |
320ae51f JA |
2747 | } |
2748 | ||
a5164405 JA |
2749 | static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set) |
2750 | { | |
2751 | int i; | |
2752 | ||
cc71a6f4 JA |
2753 | for (i = 0; i < set->nr_hw_queues; i++) |
2754 | if (!__blk_mq_alloc_rq_map(set, i)) | |
a5164405 | 2755 | goto out_unwind; |
a5164405 JA |
2756 | |
2757 | return 0; | |
2758 | ||
2759 | out_unwind: | |
2760 | while (--i >= 0) | |
cc71a6f4 | 2761 | blk_mq_free_rq_map(set->tags[i]); |
a5164405 | 2762 | |
a5164405 JA |
2763 | return -ENOMEM; |
2764 | } | |
2765 | ||
2766 | /* | |
2767 | * Allocate the request maps associated with this tag_set. Note that this | |
2768 | * may reduce the depth asked for, if memory is tight. set->queue_depth | |
2769 | * will be updated to reflect the allocated depth. | |
2770 | */ | |
2771 | static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set) | |
2772 | { | |
2773 | unsigned int depth; | |
2774 | int err; | |
2775 | ||
2776 | depth = set->queue_depth; | |
2777 | do { | |
2778 | err = __blk_mq_alloc_rq_maps(set); | |
2779 | if (!err) | |
2780 | break; | |
2781 | ||
2782 | set->queue_depth >>= 1; | |
2783 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) { | |
2784 | err = -ENOMEM; | |
2785 | break; | |
2786 | } | |
2787 | } while (set->queue_depth); | |
2788 | ||
2789 | if (!set->queue_depth || err) { | |
2790 | pr_err("blk-mq: failed to allocate request map\n"); | |
2791 | return -ENOMEM; | |
2792 | } | |
2793 | ||
2794 | if (depth != set->queue_depth) | |
2795 | pr_info("blk-mq: reduced tag depth (%u -> %u)\n", | |
2796 | depth, set->queue_depth); | |
2797 | ||
2798 | return 0; | |
2799 | } | |
2800 | ||
ebe8bddb OS |
2801 | static int blk_mq_update_queue_map(struct blk_mq_tag_set *set) |
2802 | { | |
7d4901a9 ML |
2803 | if (set->ops->map_queues) { |
2804 | int cpu; | |
2805 | /* | |
2806 | * transport .map_queues is usually done in the following | |
2807 | * way: | |
2808 | * | |
2809 | * for (queue = 0; queue < set->nr_hw_queues; queue++) { | |
2810 | * mask = get_cpu_mask(queue) | |
2811 | * for_each_cpu(cpu, mask) | |
2812 | * set->mq_map[cpu] = queue; | |
2813 | * } | |
2814 | * | |
2815 | * When we need to remap, the table has to be cleared for | |
2816 | * killing stale mapping since one CPU may not be mapped | |
2817 | * to any hw queue. | |
2818 | */ | |
2819 | for_each_possible_cpu(cpu) | |
2820 | set->mq_map[cpu] = 0; | |
2821 | ||
ebe8bddb | 2822 | return set->ops->map_queues(set); |
7d4901a9 | 2823 | } else |
ebe8bddb OS |
2824 | return blk_mq_map_queues(set); |
2825 | } | |
2826 | ||
a4391c64 JA |
2827 | /* |
2828 | * Alloc a tag set to be associated with one or more request queues. | |
2829 | * May fail with EINVAL for various error conditions. May adjust the | |
2830 | * requested depth down, if if it too large. In that case, the set | |
2831 | * value will be stored in set->queue_depth. | |
2832 | */ | |
24d2f903 CH |
2833 | int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set) |
2834 | { | |
da695ba2 CH |
2835 | int ret; |
2836 | ||
205fb5f5 BVA |
2837 | BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS); |
2838 | ||
24d2f903 CH |
2839 | if (!set->nr_hw_queues) |
2840 | return -EINVAL; | |
a4391c64 | 2841 | if (!set->queue_depth) |
24d2f903 CH |
2842 | return -EINVAL; |
2843 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) | |
2844 | return -EINVAL; | |
2845 | ||
7d7e0f90 | 2846 | if (!set->ops->queue_rq) |
24d2f903 CH |
2847 | return -EINVAL; |
2848 | ||
de148297 ML |
2849 | if (!set->ops->get_budget ^ !set->ops->put_budget) |
2850 | return -EINVAL; | |
2851 | ||
a4391c64 JA |
2852 | if (set->queue_depth > BLK_MQ_MAX_DEPTH) { |
2853 | pr_info("blk-mq: reduced tag depth to %u\n", | |
2854 | BLK_MQ_MAX_DEPTH); | |
2855 | set->queue_depth = BLK_MQ_MAX_DEPTH; | |
2856 | } | |
24d2f903 | 2857 | |
6637fadf SL |
2858 | /* |
2859 | * If a crashdump is active, then we are potentially in a very | |
2860 | * memory constrained environment. Limit us to 1 queue and | |
2861 | * 64 tags to prevent using too much memory. | |
2862 | */ | |
2863 | if (is_kdump_kernel()) { | |
2864 | set->nr_hw_queues = 1; | |
2865 | set->queue_depth = min(64U, set->queue_depth); | |
2866 | } | |
868f2f0b KB |
2867 | /* |
2868 | * There is no use for more h/w queues than cpus. | |
2869 | */ | |
2870 | if (set->nr_hw_queues > nr_cpu_ids) | |
2871 | set->nr_hw_queues = nr_cpu_ids; | |
6637fadf | 2872 | |
868f2f0b | 2873 | set->tags = kzalloc_node(nr_cpu_ids * sizeof(struct blk_mq_tags *), |
24d2f903 CH |
2874 | GFP_KERNEL, set->numa_node); |
2875 | if (!set->tags) | |
a5164405 | 2876 | return -ENOMEM; |
24d2f903 | 2877 | |
da695ba2 CH |
2878 | ret = -ENOMEM; |
2879 | set->mq_map = kzalloc_node(sizeof(*set->mq_map) * nr_cpu_ids, | |
2880 | GFP_KERNEL, set->numa_node); | |
bdd17e75 CH |
2881 | if (!set->mq_map) |
2882 | goto out_free_tags; | |
2883 | ||
ebe8bddb | 2884 | ret = blk_mq_update_queue_map(set); |
da695ba2 CH |
2885 | if (ret) |
2886 | goto out_free_mq_map; | |
2887 | ||
2888 | ret = blk_mq_alloc_rq_maps(set); | |
2889 | if (ret) | |
bdd17e75 | 2890 | goto out_free_mq_map; |
24d2f903 | 2891 | |
0d2602ca JA |
2892 | mutex_init(&set->tag_list_lock); |
2893 | INIT_LIST_HEAD(&set->tag_list); | |
2894 | ||
24d2f903 | 2895 | return 0; |
bdd17e75 CH |
2896 | |
2897 | out_free_mq_map: | |
2898 | kfree(set->mq_map); | |
2899 | set->mq_map = NULL; | |
2900 | out_free_tags: | |
5676e7b6 RE |
2901 | kfree(set->tags); |
2902 | set->tags = NULL; | |
da695ba2 | 2903 | return ret; |
24d2f903 CH |
2904 | } |
2905 | EXPORT_SYMBOL(blk_mq_alloc_tag_set); | |
2906 | ||
2907 | void blk_mq_free_tag_set(struct blk_mq_tag_set *set) | |
2908 | { | |
2909 | int i; | |
2910 | ||
cc71a6f4 JA |
2911 | for (i = 0; i < nr_cpu_ids; i++) |
2912 | blk_mq_free_map_and_requests(set, i); | |
484b4061 | 2913 | |
bdd17e75 CH |
2914 | kfree(set->mq_map); |
2915 | set->mq_map = NULL; | |
2916 | ||
981bd189 | 2917 | kfree(set->tags); |
5676e7b6 | 2918 | set->tags = NULL; |
24d2f903 CH |
2919 | } |
2920 | EXPORT_SYMBOL(blk_mq_free_tag_set); | |
2921 | ||
e3a2b3f9 JA |
2922 | int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr) |
2923 | { | |
2924 | struct blk_mq_tag_set *set = q->tag_set; | |
2925 | struct blk_mq_hw_ctx *hctx; | |
2926 | int i, ret; | |
2927 | ||
bd166ef1 | 2928 | if (!set) |
e3a2b3f9 JA |
2929 | return -EINVAL; |
2930 | ||
70f36b60 | 2931 | blk_mq_freeze_queue(q); |
24f5a90f | 2932 | blk_mq_quiesce_queue(q); |
70f36b60 | 2933 | |
e3a2b3f9 JA |
2934 | ret = 0; |
2935 | queue_for_each_hw_ctx(q, hctx, i) { | |
e9137d4b KB |
2936 | if (!hctx->tags) |
2937 | continue; | |
bd166ef1 JA |
2938 | /* |
2939 | * If we're using an MQ scheduler, just update the scheduler | |
2940 | * queue depth. This is similar to what the old code would do. | |
2941 | */ | |
70f36b60 | 2942 | if (!hctx->sched_tags) { |
c2e82a23 | 2943 | ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr, |
70f36b60 JA |
2944 | false); |
2945 | } else { | |
2946 | ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags, | |
2947 | nr, true); | |
2948 | } | |
e3a2b3f9 JA |
2949 | if (ret) |
2950 | break; | |
2951 | } | |
2952 | ||
2953 | if (!ret) | |
2954 | q->nr_requests = nr; | |
2955 | ||
24f5a90f | 2956 | blk_mq_unquiesce_queue(q); |
70f36b60 | 2957 | blk_mq_unfreeze_queue(q); |
70f36b60 | 2958 | |
e3a2b3f9 JA |
2959 | return ret; |
2960 | } | |
2961 | ||
e4dc2b32 KB |
2962 | static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, |
2963 | int nr_hw_queues) | |
868f2f0b KB |
2964 | { |
2965 | struct request_queue *q; | |
2966 | ||
705cda97 BVA |
2967 | lockdep_assert_held(&set->tag_list_lock); |
2968 | ||
868f2f0b KB |
2969 | if (nr_hw_queues > nr_cpu_ids) |
2970 | nr_hw_queues = nr_cpu_ids; | |
2971 | if (nr_hw_queues < 1 || nr_hw_queues == set->nr_hw_queues) | |
2972 | return; | |
2973 | ||
2974 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
2975 | blk_mq_freeze_queue(q); | |
2976 | ||
2977 | set->nr_hw_queues = nr_hw_queues; | |
ebe8bddb | 2978 | blk_mq_update_queue_map(set); |
868f2f0b KB |
2979 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
2980 | blk_mq_realloc_hw_ctxs(set, q); | |
4b855ad3 | 2981 | blk_mq_queue_reinit(q); |
868f2f0b KB |
2982 | } |
2983 | ||
2984 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
2985 | blk_mq_unfreeze_queue(q); | |
2986 | } | |
e4dc2b32 KB |
2987 | |
2988 | void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues) | |
2989 | { | |
2990 | mutex_lock(&set->tag_list_lock); | |
2991 | __blk_mq_update_nr_hw_queues(set, nr_hw_queues); | |
2992 | mutex_unlock(&set->tag_list_lock); | |
2993 | } | |
868f2f0b KB |
2994 | EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues); |
2995 | ||
34dbad5d OS |
2996 | /* Enable polling stats and return whether they were already enabled. */ |
2997 | static bool blk_poll_stats_enable(struct request_queue *q) | |
2998 | { | |
2999 | if (test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) || | |
7dfdbc73 | 3000 | blk_queue_flag_test_and_set(QUEUE_FLAG_POLL_STATS, q)) |
34dbad5d OS |
3001 | return true; |
3002 | blk_stat_add_callback(q, q->poll_cb); | |
3003 | return false; | |
3004 | } | |
3005 | ||
3006 | static void blk_mq_poll_stats_start(struct request_queue *q) | |
3007 | { | |
3008 | /* | |
3009 | * We don't arm the callback if polling stats are not enabled or the | |
3010 | * callback is already active. | |
3011 | */ | |
3012 | if (!test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) || | |
3013 | blk_stat_is_active(q->poll_cb)) | |
3014 | return; | |
3015 | ||
3016 | blk_stat_activate_msecs(q->poll_cb, 100); | |
3017 | } | |
3018 | ||
3019 | static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb) | |
3020 | { | |
3021 | struct request_queue *q = cb->data; | |
720b8ccc | 3022 | int bucket; |
34dbad5d | 3023 | |
720b8ccc SB |
3024 | for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) { |
3025 | if (cb->stat[bucket].nr_samples) | |
3026 | q->poll_stat[bucket] = cb->stat[bucket]; | |
3027 | } | |
34dbad5d OS |
3028 | } |
3029 | ||
64f1c21e JA |
3030 | static unsigned long blk_mq_poll_nsecs(struct request_queue *q, |
3031 | struct blk_mq_hw_ctx *hctx, | |
3032 | struct request *rq) | |
3033 | { | |
64f1c21e | 3034 | unsigned long ret = 0; |
720b8ccc | 3035 | int bucket; |
64f1c21e JA |
3036 | |
3037 | /* | |
3038 | * If stats collection isn't on, don't sleep but turn it on for | |
3039 | * future users | |
3040 | */ | |
34dbad5d | 3041 | if (!blk_poll_stats_enable(q)) |
64f1c21e JA |
3042 | return 0; |
3043 | ||
64f1c21e JA |
3044 | /* |
3045 | * As an optimistic guess, use half of the mean service time | |
3046 | * for this type of request. We can (and should) make this smarter. | |
3047 | * For instance, if the completion latencies are tight, we can | |
3048 | * get closer than just half the mean. This is especially | |
3049 | * important on devices where the completion latencies are longer | |
720b8ccc SB |
3050 | * than ~10 usec. We do use the stats for the relevant IO size |
3051 | * if available which does lead to better estimates. | |
64f1c21e | 3052 | */ |
720b8ccc SB |
3053 | bucket = blk_mq_poll_stats_bkt(rq); |
3054 | if (bucket < 0) | |
3055 | return ret; | |
3056 | ||
3057 | if (q->poll_stat[bucket].nr_samples) | |
3058 | ret = (q->poll_stat[bucket].mean + 1) / 2; | |
64f1c21e JA |
3059 | |
3060 | return ret; | |
3061 | } | |
3062 | ||
06426adf | 3063 | static bool blk_mq_poll_hybrid_sleep(struct request_queue *q, |
64f1c21e | 3064 | struct blk_mq_hw_ctx *hctx, |
06426adf JA |
3065 | struct request *rq) |
3066 | { | |
3067 | struct hrtimer_sleeper hs; | |
3068 | enum hrtimer_mode mode; | |
64f1c21e | 3069 | unsigned int nsecs; |
06426adf JA |
3070 | ktime_t kt; |
3071 | ||
76a86f9d | 3072 | if (rq->rq_flags & RQF_MQ_POLL_SLEPT) |
64f1c21e JA |
3073 | return false; |
3074 | ||
3075 | /* | |
3076 | * poll_nsec can be: | |
3077 | * | |
3078 | * -1: don't ever hybrid sleep | |
3079 | * 0: use half of prev avg | |
3080 | * >0: use this specific value | |
3081 | */ | |
3082 | if (q->poll_nsec == -1) | |
3083 | return false; | |
3084 | else if (q->poll_nsec > 0) | |
3085 | nsecs = q->poll_nsec; | |
3086 | else | |
3087 | nsecs = blk_mq_poll_nsecs(q, hctx, rq); | |
3088 | ||
3089 | if (!nsecs) | |
06426adf JA |
3090 | return false; |
3091 | ||
76a86f9d | 3092 | rq->rq_flags |= RQF_MQ_POLL_SLEPT; |
06426adf JA |
3093 | |
3094 | /* | |
3095 | * This will be replaced with the stats tracking code, using | |
3096 | * 'avg_completion_time / 2' as the pre-sleep target. | |
3097 | */ | |
8b0e1953 | 3098 | kt = nsecs; |
06426adf JA |
3099 | |
3100 | mode = HRTIMER_MODE_REL; | |
3101 | hrtimer_init_on_stack(&hs.timer, CLOCK_MONOTONIC, mode); | |
3102 | hrtimer_set_expires(&hs.timer, kt); | |
3103 | ||
3104 | hrtimer_init_sleeper(&hs, current); | |
3105 | do { | |
5a61c363 | 3106 | if (blk_mq_rq_state(rq) == MQ_RQ_COMPLETE) |
06426adf JA |
3107 | break; |
3108 | set_current_state(TASK_UNINTERRUPTIBLE); | |
3109 | hrtimer_start_expires(&hs.timer, mode); | |
3110 | if (hs.task) | |
3111 | io_schedule(); | |
3112 | hrtimer_cancel(&hs.timer); | |
3113 | mode = HRTIMER_MODE_ABS; | |
3114 | } while (hs.task && !signal_pending(current)); | |
3115 | ||
3116 | __set_current_state(TASK_RUNNING); | |
3117 | destroy_hrtimer_on_stack(&hs.timer); | |
3118 | return true; | |
3119 | } | |
3120 | ||
bbd7bb70 JA |
3121 | static bool __blk_mq_poll(struct blk_mq_hw_ctx *hctx, struct request *rq) |
3122 | { | |
3123 | struct request_queue *q = hctx->queue; | |
3124 | long state; | |
3125 | ||
06426adf JA |
3126 | /* |
3127 | * If we sleep, have the caller restart the poll loop to reset | |
3128 | * the state. Like for the other success return cases, the | |
3129 | * caller is responsible for checking if the IO completed. If | |
3130 | * the IO isn't complete, we'll get called again and will go | |
3131 | * straight to the busy poll loop. | |
3132 | */ | |
64f1c21e | 3133 | if (blk_mq_poll_hybrid_sleep(q, hctx, rq)) |
06426adf JA |
3134 | return true; |
3135 | ||
bbd7bb70 JA |
3136 | hctx->poll_considered++; |
3137 | ||
3138 | state = current->state; | |
3139 | while (!need_resched()) { | |
3140 | int ret; | |
3141 | ||
3142 | hctx->poll_invoked++; | |
3143 | ||
3144 | ret = q->mq_ops->poll(hctx, rq->tag); | |
3145 | if (ret > 0) { | |
3146 | hctx->poll_success++; | |
3147 | set_current_state(TASK_RUNNING); | |
3148 | return true; | |
3149 | } | |
3150 | ||
3151 | if (signal_pending_state(state, current)) | |
3152 | set_current_state(TASK_RUNNING); | |
3153 | ||
3154 | if (current->state == TASK_RUNNING) | |
3155 | return true; | |
3156 | if (ret < 0) | |
3157 | break; | |
3158 | cpu_relax(); | |
3159 | } | |
3160 | ||
67b4110f | 3161 | __set_current_state(TASK_RUNNING); |
bbd7bb70 JA |
3162 | return false; |
3163 | } | |
3164 | ||
ea435e1b | 3165 | static bool blk_mq_poll(struct request_queue *q, blk_qc_t cookie) |
bbd7bb70 JA |
3166 | { |
3167 | struct blk_mq_hw_ctx *hctx; | |
bbd7bb70 JA |
3168 | struct request *rq; |
3169 | ||
ea435e1b | 3170 | if (!test_bit(QUEUE_FLAG_POLL, &q->queue_flags)) |
bbd7bb70 JA |
3171 | return false; |
3172 | ||
bbd7bb70 | 3173 | hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)]; |
bd166ef1 JA |
3174 | if (!blk_qc_t_is_internal(cookie)) |
3175 | rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie)); | |
3a07bb1d | 3176 | else { |
bd166ef1 | 3177 | rq = blk_mq_tag_to_rq(hctx->sched_tags, blk_qc_t_to_tag(cookie)); |
3a07bb1d JA |
3178 | /* |
3179 | * With scheduling, if the request has completed, we'll | |
3180 | * get a NULL return here, as we clear the sched tag when | |
3181 | * that happens. The request still remains valid, like always, | |
3182 | * so we should be safe with just the NULL check. | |
3183 | */ | |
3184 | if (!rq) | |
3185 | return false; | |
3186 | } | |
bbd7bb70 JA |
3187 | |
3188 | return __blk_mq_poll(hctx, rq); | |
3189 | } | |
bbd7bb70 | 3190 | |
320ae51f JA |
3191 | static int __init blk_mq_init(void) |
3192 | { | |
9467f859 TG |
3193 | cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL, |
3194 | blk_mq_hctx_notify_dead); | |
320ae51f JA |
3195 | return 0; |
3196 | } | |
3197 | subsys_initcall(blk_mq_init); |