]>
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 | ||
98 | if (test_bit(REQ_ATOM_STARTED, &rq->atomic_flags) && | |
99 | !test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags)) { | |
100 | /* | |
b8d62b3a JA |
101 | * index[0] counts the specific partition that was asked |
102 | * for. index[1] counts the ones that are active on the | |
103 | * whole device, so increment that if mi->part is indeed | |
104 | * a partition, and not a whole device. | |
f299b7c7 | 105 | */ |
b8d62b3a | 106 | if (rq->part == mi->part) |
f299b7c7 | 107 | mi->inflight[0]++; |
b8d62b3a JA |
108 | if (mi->part->partno) |
109 | mi->inflight[1]++; | |
f299b7c7 JA |
110 | } |
111 | } | |
112 | ||
113 | void blk_mq_in_flight(struct request_queue *q, struct hd_struct *part, | |
114 | unsigned int inflight[2]) | |
115 | { | |
116 | struct mq_inflight mi = { .part = part, .inflight = inflight, }; | |
117 | ||
b8d62b3a | 118 | inflight[0] = inflight[1] = 0; |
f299b7c7 JA |
119 | blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi); |
120 | } | |
121 | ||
416d3aa7 OS |
122 | static void blk_mq_check_inflight_rw(struct blk_mq_hw_ctx *hctx, |
123 | struct request *rq, void *priv, | |
124 | bool reserved) | |
125 | { | |
126 | struct mq_inflight *mi = priv; | |
127 | ||
128 | if (rq->part == mi->part) | |
129 | mi->inflight[rq_data_dir(rq)]++; | |
130 | } | |
131 | ||
132 | void blk_mq_in_flight_rw(struct request_queue *q, struct hd_struct *part, | |
133 | unsigned int inflight[2]) | |
134 | { | |
135 | struct mq_inflight mi = { .part = part, .inflight = inflight, }; | |
136 | ||
137 | inflight[0] = inflight[1] = 0; | |
138 | blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight_rw, &mi); | |
139 | } | |
140 | ||
1671d522 | 141 | void blk_freeze_queue_start(struct request_queue *q) |
43a5e4e2 | 142 | { |
4ecd4fef | 143 | int freeze_depth; |
cddd5d17 | 144 | |
4ecd4fef CH |
145 | freeze_depth = atomic_inc_return(&q->mq_freeze_depth); |
146 | if (freeze_depth == 1) { | |
3ef28e83 | 147 | percpu_ref_kill(&q->q_usage_counter); |
055f6e18 ML |
148 | if (q->mq_ops) |
149 | blk_mq_run_hw_queues(q, false); | |
cddd5d17 | 150 | } |
f3af020b | 151 | } |
1671d522 | 152 | EXPORT_SYMBOL_GPL(blk_freeze_queue_start); |
f3af020b | 153 | |
6bae363e | 154 | void blk_mq_freeze_queue_wait(struct request_queue *q) |
f3af020b | 155 | { |
3ef28e83 | 156 | wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter)); |
43a5e4e2 | 157 | } |
6bae363e | 158 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait); |
43a5e4e2 | 159 | |
f91328c4 KB |
160 | int blk_mq_freeze_queue_wait_timeout(struct request_queue *q, |
161 | unsigned long timeout) | |
162 | { | |
163 | return wait_event_timeout(q->mq_freeze_wq, | |
164 | percpu_ref_is_zero(&q->q_usage_counter), | |
165 | timeout); | |
166 | } | |
167 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout); | |
43a5e4e2 | 168 | |
f3af020b TH |
169 | /* |
170 | * Guarantee no request is in use, so we can change any data structure of | |
171 | * the queue afterward. | |
172 | */ | |
3ef28e83 | 173 | void blk_freeze_queue(struct request_queue *q) |
f3af020b | 174 | { |
3ef28e83 DW |
175 | /* |
176 | * In the !blk_mq case we are only calling this to kill the | |
177 | * q_usage_counter, otherwise this increases the freeze depth | |
178 | * and waits for it to return to zero. For this reason there is | |
179 | * no blk_unfreeze_queue(), and blk_freeze_queue() is not | |
180 | * exported to drivers as the only user for unfreeze is blk_mq. | |
181 | */ | |
1671d522 | 182 | blk_freeze_queue_start(q); |
454be724 ML |
183 | if (!q->mq_ops) |
184 | blk_drain_queue(q); | |
f3af020b TH |
185 | blk_mq_freeze_queue_wait(q); |
186 | } | |
3ef28e83 DW |
187 | |
188 | void blk_mq_freeze_queue(struct request_queue *q) | |
189 | { | |
190 | /* | |
191 | * ...just an alias to keep freeze and unfreeze actions balanced | |
192 | * in the blk_mq_* namespace | |
193 | */ | |
194 | blk_freeze_queue(q); | |
195 | } | |
c761d96b | 196 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue); |
f3af020b | 197 | |
b4c6a028 | 198 | void blk_mq_unfreeze_queue(struct request_queue *q) |
320ae51f | 199 | { |
4ecd4fef | 200 | int freeze_depth; |
320ae51f | 201 | |
4ecd4fef CH |
202 | freeze_depth = atomic_dec_return(&q->mq_freeze_depth); |
203 | WARN_ON_ONCE(freeze_depth < 0); | |
204 | if (!freeze_depth) { | |
3ef28e83 | 205 | percpu_ref_reinit(&q->q_usage_counter); |
320ae51f | 206 | wake_up_all(&q->mq_freeze_wq); |
add703fd | 207 | } |
320ae51f | 208 | } |
b4c6a028 | 209 | EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue); |
320ae51f | 210 | |
852ec809 BVA |
211 | /* |
212 | * FIXME: replace the scsi_internal_device_*block_nowait() calls in the | |
213 | * mpt3sas driver such that this function can be removed. | |
214 | */ | |
215 | void blk_mq_quiesce_queue_nowait(struct request_queue *q) | |
216 | { | |
217 | unsigned long flags; | |
218 | ||
219 | spin_lock_irqsave(q->queue_lock, flags); | |
220 | queue_flag_set(QUEUE_FLAG_QUIESCED, q); | |
221 | spin_unlock_irqrestore(q->queue_lock, flags); | |
222 | } | |
223 | EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait); | |
224 | ||
6a83e74d | 225 | /** |
69e07c4a | 226 | * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished |
6a83e74d BVA |
227 | * @q: request queue. |
228 | * | |
229 | * Note: this function does not prevent that the struct request end_io() | |
69e07c4a ML |
230 | * callback function is invoked. Once this function is returned, we make |
231 | * sure no dispatch can happen until the queue is unquiesced via | |
232 | * blk_mq_unquiesce_queue(). | |
6a83e74d BVA |
233 | */ |
234 | void blk_mq_quiesce_queue(struct request_queue *q) | |
235 | { | |
236 | struct blk_mq_hw_ctx *hctx; | |
237 | unsigned int i; | |
238 | bool rcu = false; | |
239 | ||
1d9e9bc6 | 240 | blk_mq_quiesce_queue_nowait(q); |
f4560ffe | 241 | |
6a83e74d BVA |
242 | queue_for_each_hw_ctx(q, hctx, i) { |
243 | if (hctx->flags & BLK_MQ_F_BLOCKING) | |
07319678 | 244 | synchronize_srcu(hctx->queue_rq_srcu); |
6a83e74d BVA |
245 | else |
246 | rcu = true; | |
247 | } | |
248 | if (rcu) | |
249 | synchronize_rcu(); | |
250 | } | |
251 | EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue); | |
252 | ||
e4e73913 ML |
253 | /* |
254 | * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue() | |
255 | * @q: request queue. | |
256 | * | |
257 | * This function recovers queue into the state before quiescing | |
258 | * which is done by blk_mq_quiesce_queue. | |
259 | */ | |
260 | void blk_mq_unquiesce_queue(struct request_queue *q) | |
261 | { | |
852ec809 BVA |
262 | unsigned long flags; |
263 | ||
264 | spin_lock_irqsave(q->queue_lock, flags); | |
f4560ffe | 265 | queue_flag_clear(QUEUE_FLAG_QUIESCED, q); |
852ec809 | 266 | spin_unlock_irqrestore(q->queue_lock, flags); |
f4560ffe | 267 | |
1d9e9bc6 ML |
268 | /* dispatch requests which are inserted during quiescing */ |
269 | blk_mq_run_hw_queues(q, true); | |
e4e73913 ML |
270 | } |
271 | EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue); | |
272 | ||
aed3ea94 JA |
273 | void blk_mq_wake_waiters(struct request_queue *q) |
274 | { | |
275 | struct blk_mq_hw_ctx *hctx; | |
276 | unsigned int i; | |
277 | ||
278 | queue_for_each_hw_ctx(q, hctx, i) | |
279 | if (blk_mq_hw_queue_mapped(hctx)) | |
280 | blk_mq_tag_wakeup_all(hctx->tags, true); | |
281 | } | |
282 | ||
320ae51f JA |
283 | bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx) |
284 | { | |
285 | return blk_mq_has_free_tags(hctx->tags); | |
286 | } | |
287 | EXPORT_SYMBOL(blk_mq_can_queue); | |
288 | ||
e4cdf1a1 CH |
289 | static struct request *blk_mq_rq_ctx_init(struct blk_mq_alloc_data *data, |
290 | unsigned int tag, unsigned int op) | |
320ae51f | 291 | { |
e4cdf1a1 CH |
292 | struct blk_mq_tags *tags = blk_mq_tags_from_data(data); |
293 | struct request *rq = tags->static_rqs[tag]; | |
294 | ||
c3a148d2 BVA |
295 | rq->rq_flags = 0; |
296 | ||
e4cdf1a1 CH |
297 | if (data->flags & BLK_MQ_REQ_INTERNAL) { |
298 | rq->tag = -1; | |
299 | rq->internal_tag = tag; | |
300 | } else { | |
662e29d9 | 301 | if (data->hctx->flags & BLK_MQ_F_TAG_SHARED) { |
e4cdf1a1 CH |
302 | rq->rq_flags = RQF_MQ_INFLIGHT; |
303 | atomic_inc(&data->hctx->nr_active); | |
304 | } | |
305 | rq->tag = tag; | |
306 | rq->internal_tag = -1; | |
307 | data->hctx->tags->rqs[rq->tag] = rq; | |
308 | } | |
309 | ||
af76e555 CH |
310 | INIT_LIST_HEAD(&rq->queuelist); |
311 | /* csd/requeue_work/fifo_time is initialized before use */ | |
e4cdf1a1 CH |
312 | rq->q = data->q; |
313 | rq->mq_ctx = data->ctx; | |
ef295ecf | 314 | rq->cmd_flags = op; |
1b6d65a0 BVA |
315 | if (data->flags & BLK_MQ_REQ_PREEMPT) |
316 | rq->rq_flags |= RQF_PREEMPT; | |
e4cdf1a1 | 317 | if (blk_queue_io_stat(data->q)) |
e8064021 | 318 | rq->rq_flags |= RQF_IO_STAT; |
af76e555 CH |
319 | /* do not touch atomic flags, it needs atomic ops against the timer */ |
320 | rq->cpu = -1; | |
af76e555 CH |
321 | INIT_HLIST_NODE(&rq->hash); |
322 | RB_CLEAR_NODE(&rq->rb_node); | |
af76e555 CH |
323 | rq->rq_disk = NULL; |
324 | rq->part = NULL; | |
3ee32372 | 325 | rq->start_time = jiffies; |
af76e555 CH |
326 | #ifdef CONFIG_BLK_CGROUP |
327 | rq->rl = NULL; | |
0fec08b4 | 328 | set_start_time_ns(rq); |
af76e555 CH |
329 | rq->io_start_time_ns = 0; |
330 | #endif | |
331 | rq->nr_phys_segments = 0; | |
332 | #if defined(CONFIG_BLK_DEV_INTEGRITY) | |
333 | rq->nr_integrity_segments = 0; | |
334 | #endif | |
af76e555 CH |
335 | rq->special = NULL; |
336 | /* tag was already set */ | |
af76e555 | 337 | rq->extra_len = 0; |
af76e555 | 338 | |
af76e555 | 339 | INIT_LIST_HEAD(&rq->timeout_list); |
f6be4fb4 JA |
340 | rq->timeout = 0; |
341 | ||
af76e555 CH |
342 | rq->end_io = NULL; |
343 | rq->end_io_data = NULL; | |
344 | rq->next_rq = NULL; | |
345 | ||
e4cdf1a1 CH |
346 | data->ctx->rq_dispatched[op_is_sync(op)]++; |
347 | return rq; | |
5dee8577 CH |
348 | } |
349 | ||
d2c0d383 CH |
350 | static struct request *blk_mq_get_request(struct request_queue *q, |
351 | struct bio *bio, unsigned int op, | |
352 | struct blk_mq_alloc_data *data) | |
353 | { | |
354 | struct elevator_queue *e = q->elevator; | |
355 | struct request *rq; | |
e4cdf1a1 | 356 | unsigned int tag; |
21e768b4 | 357 | bool put_ctx_on_error = false; |
d2c0d383 CH |
358 | |
359 | blk_queue_enter_live(q); | |
360 | data->q = q; | |
21e768b4 BVA |
361 | if (likely(!data->ctx)) { |
362 | data->ctx = blk_mq_get_ctx(q); | |
363 | put_ctx_on_error = true; | |
364 | } | |
d2c0d383 CH |
365 | if (likely(!data->hctx)) |
366 | data->hctx = blk_mq_map_queue(q, data->ctx->cpu); | |
03a07c92 GR |
367 | if (op & REQ_NOWAIT) |
368 | data->flags |= BLK_MQ_REQ_NOWAIT; | |
d2c0d383 CH |
369 | |
370 | if (e) { | |
371 | data->flags |= BLK_MQ_REQ_INTERNAL; | |
372 | ||
373 | /* | |
374 | * Flush requests are special and go directly to the | |
375 | * dispatch list. | |
376 | */ | |
5bbf4e5a CH |
377 | if (!op_is_flush(op) && e->type->ops.mq.limit_depth) |
378 | e->type->ops.mq.limit_depth(op, data); | |
662e29d9 JW |
379 | } else { |
380 | blk_mq_tag_busy(data->hctx); | |
d2c0d383 CH |
381 | } |
382 | ||
e4cdf1a1 CH |
383 | tag = blk_mq_get_tag(data); |
384 | if (tag == BLK_MQ_TAG_FAIL) { | |
21e768b4 BVA |
385 | if (put_ctx_on_error) { |
386 | blk_mq_put_ctx(data->ctx); | |
1ad43c00 ML |
387 | data->ctx = NULL; |
388 | } | |
037cebb8 CH |
389 | blk_queue_exit(q); |
390 | return NULL; | |
d2c0d383 CH |
391 | } |
392 | ||
e4cdf1a1 | 393 | rq = blk_mq_rq_ctx_init(data, tag, op); |
037cebb8 CH |
394 | if (!op_is_flush(op)) { |
395 | rq->elv.icq = NULL; | |
5bbf4e5a | 396 | if (e && e->type->ops.mq.prepare_request) { |
44e8c2bf CH |
397 | if (e->type->icq_cache && rq_ioc(bio)) |
398 | blk_mq_sched_assign_ioc(rq, bio); | |
399 | ||
5bbf4e5a CH |
400 | e->type->ops.mq.prepare_request(rq, bio); |
401 | rq->rq_flags |= RQF_ELVPRIV; | |
44e8c2bf | 402 | } |
037cebb8 CH |
403 | } |
404 | data->hctx->queued++; | |
405 | return rq; | |
d2c0d383 CH |
406 | } |
407 | ||
cd6ce148 | 408 | struct request *blk_mq_alloc_request(struct request_queue *q, unsigned int op, |
9a95e4ef | 409 | blk_mq_req_flags_t flags) |
320ae51f | 410 | { |
5a797e00 | 411 | struct blk_mq_alloc_data alloc_data = { .flags = flags }; |
bd166ef1 | 412 | struct request *rq; |
a492f075 | 413 | int ret; |
320ae51f | 414 | |
3a0a5299 | 415 | ret = blk_queue_enter(q, flags); |
a492f075 JL |
416 | if (ret) |
417 | return ERR_PTR(ret); | |
320ae51f | 418 | |
cd6ce148 | 419 | rq = blk_mq_get_request(q, NULL, op, &alloc_data); |
3280d66a | 420 | blk_queue_exit(q); |
841bac2c | 421 | |
bd166ef1 | 422 | if (!rq) |
a492f075 | 423 | return ERR_PTR(-EWOULDBLOCK); |
0c4de0f3 | 424 | |
1ad43c00 | 425 | blk_mq_put_ctx(alloc_data.ctx); |
1ad43c00 | 426 | |
0c4de0f3 CH |
427 | rq->__data_len = 0; |
428 | rq->__sector = (sector_t) -1; | |
429 | rq->bio = rq->biotail = NULL; | |
320ae51f JA |
430 | return rq; |
431 | } | |
4bb659b1 | 432 | EXPORT_SYMBOL(blk_mq_alloc_request); |
320ae51f | 433 | |
cd6ce148 | 434 | struct request *blk_mq_alloc_request_hctx(struct request_queue *q, |
9a95e4ef | 435 | unsigned int op, blk_mq_req_flags_t flags, unsigned int hctx_idx) |
1f5bd336 | 436 | { |
6d2809d5 | 437 | struct blk_mq_alloc_data alloc_data = { .flags = flags }; |
1f5bd336 | 438 | struct request *rq; |
6d2809d5 | 439 | unsigned int cpu; |
1f5bd336 ML |
440 | int ret; |
441 | ||
442 | /* | |
443 | * If the tag allocator sleeps we could get an allocation for a | |
444 | * different hardware context. No need to complicate the low level | |
445 | * allocator for this for the rare use case of a command tied to | |
446 | * a specific queue. | |
447 | */ | |
448 | if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT))) | |
449 | return ERR_PTR(-EINVAL); | |
450 | ||
451 | if (hctx_idx >= q->nr_hw_queues) | |
452 | return ERR_PTR(-EIO); | |
453 | ||
3a0a5299 | 454 | ret = blk_queue_enter(q, flags); |
1f5bd336 ML |
455 | if (ret) |
456 | return ERR_PTR(ret); | |
457 | ||
c8712c6a CH |
458 | /* |
459 | * Check if the hardware context is actually mapped to anything. | |
460 | * If not tell the caller that it should skip this queue. | |
461 | */ | |
6d2809d5 OS |
462 | alloc_data.hctx = q->queue_hw_ctx[hctx_idx]; |
463 | if (!blk_mq_hw_queue_mapped(alloc_data.hctx)) { | |
464 | blk_queue_exit(q); | |
465 | return ERR_PTR(-EXDEV); | |
c8712c6a | 466 | } |
8a591698 | 467 | cpu = cpumask_first_and(alloc_data.hctx->cpumask, cpu_online_mask); |
6d2809d5 | 468 | alloc_data.ctx = __blk_mq_get_ctx(q, cpu); |
1f5bd336 | 469 | |
cd6ce148 | 470 | rq = blk_mq_get_request(q, NULL, op, &alloc_data); |
3280d66a | 471 | blk_queue_exit(q); |
c8712c6a | 472 | |
6d2809d5 OS |
473 | if (!rq) |
474 | return ERR_PTR(-EWOULDBLOCK); | |
475 | ||
476 | return rq; | |
1f5bd336 ML |
477 | } |
478 | EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx); | |
479 | ||
6af54051 | 480 | void blk_mq_free_request(struct request *rq) |
320ae51f | 481 | { |
320ae51f | 482 | struct request_queue *q = rq->q; |
6af54051 CH |
483 | struct elevator_queue *e = q->elevator; |
484 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
485 | struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, ctx->cpu); | |
486 | const int sched_tag = rq->internal_tag; | |
487 | ||
5bbf4e5a | 488 | if (rq->rq_flags & RQF_ELVPRIV) { |
6af54051 CH |
489 | if (e && e->type->ops.mq.finish_request) |
490 | e->type->ops.mq.finish_request(rq); | |
491 | if (rq->elv.icq) { | |
492 | put_io_context(rq->elv.icq->ioc); | |
493 | rq->elv.icq = NULL; | |
494 | } | |
495 | } | |
320ae51f | 496 | |
6af54051 | 497 | ctx->rq_completed[rq_is_sync(rq)]++; |
e8064021 | 498 | if (rq->rq_flags & RQF_MQ_INFLIGHT) |
0d2602ca | 499 | atomic_dec(&hctx->nr_active); |
87760e5e | 500 | |
7beb2f84 JA |
501 | if (unlikely(laptop_mode && !blk_rq_is_passthrough(rq))) |
502 | laptop_io_completion(q->backing_dev_info); | |
503 | ||
87760e5e | 504 | wbt_done(q->rq_wb, &rq->issue_stat); |
0d2602ca | 505 | |
85acb3ba SL |
506 | if (blk_rq_rl(rq)) |
507 | blk_put_rl(blk_rq_rl(rq)); | |
508 | ||
af76e555 | 509 | clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags); |
06426adf | 510 | clear_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags); |
bd166ef1 JA |
511 | if (rq->tag != -1) |
512 | blk_mq_put_tag(hctx, hctx->tags, ctx, rq->tag); | |
513 | if (sched_tag != -1) | |
c05f8525 | 514 | blk_mq_put_tag(hctx, hctx->sched_tags, ctx, sched_tag); |
6d8c6c0f | 515 | blk_mq_sched_restart(hctx); |
3ef28e83 | 516 | blk_queue_exit(q); |
320ae51f | 517 | } |
1a3b595a | 518 | EXPORT_SYMBOL_GPL(blk_mq_free_request); |
320ae51f | 519 | |
2a842aca | 520 | inline void __blk_mq_end_request(struct request *rq, blk_status_t error) |
320ae51f | 521 | { |
0d11e6ac ML |
522 | blk_account_io_done(rq); |
523 | ||
91b63639 | 524 | if (rq->end_io) { |
87760e5e | 525 | wbt_done(rq->q->rq_wb, &rq->issue_stat); |
320ae51f | 526 | rq->end_io(rq, error); |
91b63639 CH |
527 | } else { |
528 | if (unlikely(blk_bidi_rq(rq))) | |
529 | blk_mq_free_request(rq->next_rq); | |
320ae51f | 530 | blk_mq_free_request(rq); |
91b63639 | 531 | } |
320ae51f | 532 | } |
c8a446ad | 533 | EXPORT_SYMBOL(__blk_mq_end_request); |
63151a44 | 534 | |
2a842aca | 535 | void blk_mq_end_request(struct request *rq, blk_status_t error) |
63151a44 CH |
536 | { |
537 | if (blk_update_request(rq, error, blk_rq_bytes(rq))) | |
538 | BUG(); | |
c8a446ad | 539 | __blk_mq_end_request(rq, error); |
63151a44 | 540 | } |
c8a446ad | 541 | EXPORT_SYMBOL(blk_mq_end_request); |
320ae51f | 542 | |
30a91cb4 | 543 | static void __blk_mq_complete_request_remote(void *data) |
320ae51f | 544 | { |
3d6efbf6 | 545 | struct request *rq = data; |
320ae51f | 546 | |
30a91cb4 | 547 | rq->q->softirq_done_fn(rq); |
320ae51f | 548 | } |
320ae51f | 549 | |
453f8341 | 550 | static void __blk_mq_complete_request(struct request *rq) |
320ae51f JA |
551 | { |
552 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
38535201 | 553 | bool shared = false; |
320ae51f JA |
554 | int cpu; |
555 | ||
453f8341 CH |
556 | if (rq->internal_tag != -1) |
557 | blk_mq_sched_completed_request(rq); | |
558 | if (rq->rq_flags & RQF_STATS) { | |
559 | blk_mq_poll_stats_start(rq->q); | |
560 | blk_stat_add(rq); | |
561 | } | |
562 | ||
38535201 | 563 | if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) { |
30a91cb4 CH |
564 | rq->q->softirq_done_fn(rq); |
565 | return; | |
566 | } | |
320ae51f JA |
567 | |
568 | cpu = get_cpu(); | |
38535201 CH |
569 | if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags)) |
570 | shared = cpus_share_cache(cpu, ctx->cpu); | |
571 | ||
572 | if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) { | |
30a91cb4 | 573 | rq->csd.func = __blk_mq_complete_request_remote; |
3d6efbf6 CH |
574 | rq->csd.info = rq; |
575 | rq->csd.flags = 0; | |
c46fff2a | 576 | smp_call_function_single_async(ctx->cpu, &rq->csd); |
3d6efbf6 | 577 | } else { |
30a91cb4 | 578 | rq->q->softirq_done_fn(rq); |
3d6efbf6 | 579 | } |
320ae51f JA |
580 | put_cpu(); |
581 | } | |
30a91cb4 | 582 | |
74e11d8d JA |
583 | static void hctx_unlock(struct blk_mq_hw_ctx *hctx, int srcu_idx) |
584 | { | |
585 | if (!(hctx->flags & BLK_MQ_F_BLOCKING)) | |
586 | rcu_read_unlock(); | |
587 | else | |
588 | srcu_read_unlock(hctx->queue_rq_srcu, srcu_idx); | |
589 | } | |
590 | ||
591 | static void hctx_lock(struct blk_mq_hw_ctx *hctx, int *srcu_idx) | |
592 | { | |
593 | if (!(hctx->flags & BLK_MQ_F_BLOCKING)) | |
594 | rcu_read_lock(); | |
595 | else | |
596 | *srcu_idx = srcu_read_lock(hctx->queue_rq_srcu); | |
597 | } | |
598 | ||
30a91cb4 CH |
599 | /** |
600 | * blk_mq_complete_request - end I/O on a request | |
601 | * @rq: the request being processed | |
602 | * | |
603 | * Description: | |
604 | * Ends all I/O on a request. It does not handle partial completions. | |
605 | * The actual completion happens out-of-order, through a IPI handler. | |
606 | **/ | |
08e0029a | 607 | void blk_mq_complete_request(struct request *rq) |
30a91cb4 | 608 | { |
95f09684 JA |
609 | struct request_queue *q = rq->q; |
610 | ||
611 | if (unlikely(blk_should_fake_timeout(q))) | |
30a91cb4 | 612 | return; |
08e0029a | 613 | if (!blk_mark_rq_complete(rq)) |
ed851860 | 614 | __blk_mq_complete_request(rq); |
30a91cb4 CH |
615 | } |
616 | EXPORT_SYMBOL(blk_mq_complete_request); | |
320ae51f | 617 | |
973c0191 KB |
618 | int blk_mq_request_started(struct request *rq) |
619 | { | |
620 | return test_bit(REQ_ATOM_STARTED, &rq->atomic_flags); | |
621 | } | |
622 | EXPORT_SYMBOL_GPL(blk_mq_request_started); | |
623 | ||
e2490073 | 624 | void blk_mq_start_request(struct request *rq) |
320ae51f JA |
625 | { |
626 | struct request_queue *q = rq->q; | |
627 | ||
bd166ef1 JA |
628 | blk_mq_sched_started_request(rq); |
629 | ||
320ae51f JA |
630 | trace_block_rq_issue(q, rq); |
631 | ||
cf43e6be | 632 | if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) { |
88eeca49 | 633 | blk_stat_set_issue(&rq->issue_stat, blk_rq_sectors(rq)); |
cf43e6be | 634 | rq->rq_flags |= RQF_STATS; |
87760e5e | 635 | wbt_issue(q->rq_wb, &rq->issue_stat); |
cf43e6be JA |
636 | } |
637 | ||
2b8393b4 | 638 | blk_add_timer(rq); |
87ee7b11 | 639 | |
a7af0af3 | 640 | WARN_ON_ONCE(test_bit(REQ_ATOM_STARTED, &rq->atomic_flags)); |
538b7534 | 641 | |
87ee7b11 JA |
642 | /* |
643 | * Mark us as started and clear complete. Complete might have been | |
644 | * set if requeue raced with timeout, which then marked it as | |
645 | * complete. So be sure to clear complete again when we start | |
646 | * the request, otherwise we'll ignore the completion event. | |
a7af0af3 PZ |
647 | * |
648 | * Ensure that ->deadline is visible before we set STARTED, such that | |
649 | * blk_mq_check_expired() is guaranteed to observe our ->deadline when | |
650 | * it observes STARTED. | |
87ee7b11 | 651 | */ |
a7af0af3 PZ |
652 | smp_wmb(); |
653 | set_bit(REQ_ATOM_STARTED, &rq->atomic_flags); | |
654 | if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags)) { | |
655 | /* | |
656 | * Coherence order guarantees these consecutive stores to a | |
657 | * single variable propagate in the specified order. Thus the | |
658 | * clear_bit() is ordered _after_ the set bit. See | |
659 | * blk_mq_check_expired(). | |
660 | * | |
661 | * (the bits must be part of the same byte for this to be | |
662 | * true). | |
663 | */ | |
4b570521 | 664 | clear_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags); |
a7af0af3 | 665 | } |
49f5baa5 CH |
666 | |
667 | if (q->dma_drain_size && blk_rq_bytes(rq)) { | |
668 | /* | |
669 | * Make sure space for the drain appears. We know we can do | |
670 | * this because max_hw_segments has been adjusted to be one | |
671 | * fewer than the device can handle. | |
672 | */ | |
673 | rq->nr_phys_segments++; | |
674 | } | |
320ae51f | 675 | } |
e2490073 | 676 | EXPORT_SYMBOL(blk_mq_start_request); |
320ae51f | 677 | |
d9d149a3 ML |
678 | /* |
679 | * When we reach here because queue is busy, REQ_ATOM_COMPLETE | |
48b99c9d | 680 | * flag isn't set yet, so there may be race with timeout handler, |
d9d149a3 ML |
681 | * but given rq->deadline is just set in .queue_rq() under |
682 | * this situation, the race won't be possible in reality because | |
683 | * rq->timeout should be set as big enough to cover the window | |
684 | * between blk_mq_start_request() called from .queue_rq() and | |
685 | * clearing REQ_ATOM_STARTED here. | |
686 | */ | |
ed0791b2 | 687 | static void __blk_mq_requeue_request(struct request *rq) |
320ae51f JA |
688 | { |
689 | struct request_queue *q = rq->q; | |
690 | ||
923218f6 ML |
691 | blk_mq_put_driver_tag(rq); |
692 | ||
320ae51f | 693 | trace_block_rq_requeue(q, rq); |
87760e5e | 694 | wbt_requeue(q->rq_wb, &rq->issue_stat); |
49f5baa5 | 695 | |
e2490073 CH |
696 | if (test_and_clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) { |
697 | if (q->dma_drain_size && blk_rq_bytes(rq)) | |
698 | rq->nr_phys_segments--; | |
699 | } | |
320ae51f JA |
700 | } |
701 | ||
2b053aca | 702 | void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list) |
ed0791b2 | 703 | { |
ed0791b2 | 704 | __blk_mq_requeue_request(rq); |
ed0791b2 | 705 | |
4c1f67e5 ML |
706 | /* this request will be re-inserted to io scheduler queue */ |
707 | blk_mq_sched_requeue_request(rq); | |
708 | ||
ed0791b2 | 709 | BUG_ON(blk_queued_rq(rq)); |
2b053aca | 710 | blk_mq_add_to_requeue_list(rq, true, kick_requeue_list); |
ed0791b2 CH |
711 | } |
712 | EXPORT_SYMBOL(blk_mq_requeue_request); | |
713 | ||
6fca6a61 CH |
714 | static void blk_mq_requeue_work(struct work_struct *work) |
715 | { | |
716 | struct request_queue *q = | |
2849450a | 717 | container_of(work, struct request_queue, requeue_work.work); |
6fca6a61 CH |
718 | LIST_HEAD(rq_list); |
719 | struct request *rq, *next; | |
6fca6a61 | 720 | |
18e9781d | 721 | spin_lock_irq(&q->requeue_lock); |
6fca6a61 | 722 | list_splice_init(&q->requeue_list, &rq_list); |
18e9781d | 723 | spin_unlock_irq(&q->requeue_lock); |
6fca6a61 CH |
724 | |
725 | list_for_each_entry_safe(rq, next, &rq_list, queuelist) { | |
e4e131dd | 726 | if (!(rq->rq_flags & (RQF_SOFTBARRIER | RQF_DONTPREP))) |
6fca6a61 CH |
727 | continue; |
728 | ||
e8064021 | 729 | rq->rq_flags &= ~RQF_SOFTBARRIER; |
6fca6a61 | 730 | list_del_init(&rq->queuelist); |
e4e131dd JW |
731 | /* |
732 | * If RQF_DONTPREP, rq has contained some driver specific | |
733 | * data, so insert it to hctx dispatch list to avoid any | |
734 | * merge. | |
735 | */ | |
736 | if (rq->rq_flags & RQF_DONTPREP) | |
737 | blk_mq_request_bypass_insert(rq, false); | |
738 | else | |
739 | blk_mq_sched_insert_request(rq, true, false, false, true); | |
6fca6a61 CH |
740 | } |
741 | ||
742 | while (!list_empty(&rq_list)) { | |
743 | rq = list_entry(rq_list.next, struct request, queuelist); | |
744 | list_del_init(&rq->queuelist); | |
bd6737f1 | 745 | blk_mq_sched_insert_request(rq, false, false, false, true); |
6fca6a61 CH |
746 | } |
747 | ||
52d7f1b5 | 748 | blk_mq_run_hw_queues(q, false); |
6fca6a61 CH |
749 | } |
750 | ||
2b053aca BVA |
751 | void blk_mq_add_to_requeue_list(struct request *rq, bool at_head, |
752 | bool kick_requeue_list) | |
6fca6a61 CH |
753 | { |
754 | struct request_queue *q = rq->q; | |
755 | unsigned long flags; | |
756 | ||
757 | /* | |
758 | * We abuse this flag that is otherwise used by the I/O scheduler to | |
ff821d27 | 759 | * request head insertion from the workqueue. |
6fca6a61 | 760 | */ |
e8064021 | 761 | BUG_ON(rq->rq_flags & RQF_SOFTBARRIER); |
6fca6a61 CH |
762 | |
763 | spin_lock_irqsave(&q->requeue_lock, flags); | |
764 | if (at_head) { | |
e8064021 | 765 | rq->rq_flags |= RQF_SOFTBARRIER; |
6fca6a61 CH |
766 | list_add(&rq->queuelist, &q->requeue_list); |
767 | } else { | |
768 | list_add_tail(&rq->queuelist, &q->requeue_list); | |
769 | } | |
770 | spin_unlock_irqrestore(&q->requeue_lock, flags); | |
2b053aca BVA |
771 | |
772 | if (kick_requeue_list) | |
773 | blk_mq_kick_requeue_list(q); | |
6fca6a61 CH |
774 | } |
775 | EXPORT_SYMBOL(blk_mq_add_to_requeue_list); | |
776 | ||
777 | void blk_mq_kick_requeue_list(struct request_queue *q) | |
778 | { | |
81517da9 | 779 | kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0); |
6fca6a61 CH |
780 | } |
781 | EXPORT_SYMBOL(blk_mq_kick_requeue_list); | |
782 | ||
2849450a MS |
783 | void blk_mq_delay_kick_requeue_list(struct request_queue *q, |
784 | unsigned long msecs) | |
785 | { | |
d4acf365 BVA |
786 | kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, |
787 | msecs_to_jiffies(msecs)); | |
2849450a MS |
788 | } |
789 | EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list); | |
790 | ||
0e62f51f JA |
791 | struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag) |
792 | { | |
88c7b2b7 JA |
793 | if (tag < tags->nr_tags) { |
794 | prefetch(tags->rqs[tag]); | |
4ee86bab | 795 | return tags->rqs[tag]; |
88c7b2b7 | 796 | } |
4ee86bab HR |
797 | |
798 | return NULL; | |
24d2f903 CH |
799 | } |
800 | EXPORT_SYMBOL(blk_mq_tag_to_rq); | |
801 | ||
320ae51f | 802 | struct blk_mq_timeout_data { |
46f92d42 CH |
803 | unsigned long next; |
804 | unsigned int next_set; | |
320ae51f JA |
805 | }; |
806 | ||
90415837 | 807 | void blk_mq_rq_timed_out(struct request *req, bool reserved) |
320ae51f | 808 | { |
f8a5b122 | 809 | const struct blk_mq_ops *ops = req->q->mq_ops; |
46f92d42 | 810 | enum blk_eh_timer_return ret = BLK_EH_RESET_TIMER; |
87ee7b11 JA |
811 | |
812 | /* | |
813 | * We know that complete is set at this point. If STARTED isn't set | |
814 | * anymore, then the request isn't active and the "timeout" should | |
815 | * just be ignored. This can happen due to the bitflag ordering. | |
816 | * Timeout first checks if STARTED is set, and if it is, assumes | |
817 | * the request is active. But if we race with completion, then | |
48b99c9d | 818 | * both flags will get cleared. So check here again, and ignore |
87ee7b11 JA |
819 | * a timeout event with a request that isn't active. |
820 | */ | |
46f92d42 CH |
821 | if (!test_bit(REQ_ATOM_STARTED, &req->atomic_flags)) |
822 | return; | |
87ee7b11 | 823 | |
46f92d42 | 824 | if (ops->timeout) |
0152fb6b | 825 | ret = ops->timeout(req, reserved); |
46f92d42 CH |
826 | |
827 | switch (ret) { | |
828 | case BLK_EH_HANDLED: | |
829 | __blk_mq_complete_request(req); | |
830 | break; | |
831 | case BLK_EH_RESET_TIMER: | |
832 | blk_add_timer(req); | |
833 | blk_clear_rq_complete(req); | |
834 | break; | |
835 | case BLK_EH_NOT_HANDLED: | |
836 | break; | |
837 | default: | |
838 | printk(KERN_ERR "block: bad eh return: %d\n", ret); | |
839 | break; | |
840 | } | |
87ee7b11 | 841 | } |
5b3f25fc | 842 | |
81481eb4 CH |
843 | static void blk_mq_check_expired(struct blk_mq_hw_ctx *hctx, |
844 | struct request *rq, void *priv, bool reserved) | |
845 | { | |
846 | struct blk_mq_timeout_data *data = priv; | |
a7af0af3 | 847 | unsigned long deadline; |
87ee7b11 | 848 | |
95a49603 | 849 | if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) |
46f92d42 | 850 | return; |
87ee7b11 | 851 | |
a7af0af3 PZ |
852 | /* |
853 | * Ensures that if we see STARTED we must also see our | |
854 | * up-to-date deadline, see blk_mq_start_request(). | |
855 | */ | |
856 | smp_rmb(); | |
857 | ||
858 | deadline = READ_ONCE(rq->deadline); | |
859 | ||
d9d149a3 ML |
860 | /* |
861 | * The rq being checked may have been freed and reallocated | |
862 | * out already here, we avoid this race by checking rq->deadline | |
863 | * and REQ_ATOM_COMPLETE flag together: | |
864 | * | |
865 | * - if rq->deadline is observed as new value because of | |
866 | * reusing, the rq won't be timed out because of timing. | |
867 | * - if rq->deadline is observed as previous value, | |
868 | * REQ_ATOM_COMPLETE flag won't be cleared in reuse path | |
869 | * because we put a barrier between setting rq->deadline | |
870 | * and clearing the flag in blk_mq_start_request(), so | |
871 | * this rq won't be timed out too. | |
872 | */ | |
a7af0af3 PZ |
873 | if (time_after_eq(jiffies, deadline)) { |
874 | if (!blk_mark_rq_complete(rq)) { | |
875 | /* | |
876 | * Again coherence order ensures that consecutive reads | |
877 | * from the same variable must be in that order. This | |
878 | * ensures that if we see COMPLETE clear, we must then | |
879 | * see STARTED set and we'll ignore this timeout. | |
880 | * | |
881 | * (There's also the MB implied by the test_and_clear()) | |
882 | */ | |
0152fb6b | 883 | blk_mq_rq_timed_out(rq, reserved); |
a7af0af3 PZ |
884 | } |
885 | } else if (!data->next_set || time_after(data->next, deadline)) { | |
886 | data->next = deadline; | |
46f92d42 CH |
887 | data->next_set = 1; |
888 | } | |
87ee7b11 JA |
889 | } |
890 | ||
287922eb | 891 | static void blk_mq_timeout_work(struct work_struct *work) |
320ae51f | 892 | { |
287922eb CH |
893 | struct request_queue *q = |
894 | container_of(work, struct request_queue, timeout_work); | |
81481eb4 CH |
895 | struct blk_mq_timeout_data data = { |
896 | .next = 0, | |
897 | .next_set = 0, | |
898 | }; | |
81481eb4 | 899 | int i; |
320ae51f | 900 | |
71f79fb3 GKB |
901 | /* A deadlock might occur if a request is stuck requiring a |
902 | * timeout at the same time a queue freeze is waiting | |
903 | * completion, since the timeout code would not be able to | |
904 | * acquire the queue reference here. | |
905 | * | |
906 | * That's why we don't use blk_queue_enter here; instead, we use | |
907 | * percpu_ref_tryget directly, because we need to be able to | |
908 | * obtain a reference even in the short window between the queue | |
909 | * starting to freeze, by dropping the first reference in | |
1671d522 | 910 | * blk_freeze_queue_start, and the moment the last request is |
71f79fb3 GKB |
911 | * consumed, marked by the instant q_usage_counter reaches |
912 | * zero. | |
913 | */ | |
914 | if (!percpu_ref_tryget(&q->q_usage_counter)) | |
287922eb CH |
915 | return; |
916 | ||
0bf6cd5b | 917 | blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &data); |
320ae51f | 918 | |
81481eb4 CH |
919 | if (data.next_set) { |
920 | data.next = blk_rq_timeout(round_jiffies_up(data.next)); | |
921 | mod_timer(&q->timeout, data.next); | |
0d2602ca | 922 | } else { |
0bf6cd5b CH |
923 | struct blk_mq_hw_ctx *hctx; |
924 | ||
f054b56c ML |
925 | queue_for_each_hw_ctx(q, hctx, i) { |
926 | /* the hctx may be unmapped, so check it here */ | |
927 | if (blk_mq_hw_queue_mapped(hctx)) | |
928 | blk_mq_tag_idle(hctx); | |
929 | } | |
0d2602ca | 930 | } |
287922eb | 931 | blk_queue_exit(q); |
320ae51f JA |
932 | } |
933 | ||
88459642 OS |
934 | struct flush_busy_ctx_data { |
935 | struct blk_mq_hw_ctx *hctx; | |
936 | struct list_head *list; | |
937 | }; | |
938 | ||
939 | static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data) | |
940 | { | |
941 | struct flush_busy_ctx_data *flush_data = data; | |
942 | struct blk_mq_hw_ctx *hctx = flush_data->hctx; | |
943 | struct blk_mq_ctx *ctx = hctx->ctxs[bitnr]; | |
944 | ||
945 | sbitmap_clear_bit(sb, bitnr); | |
946 | spin_lock(&ctx->lock); | |
947 | list_splice_tail_init(&ctx->rq_list, flush_data->list); | |
948 | spin_unlock(&ctx->lock); | |
949 | return true; | |
950 | } | |
951 | ||
1429d7c9 JA |
952 | /* |
953 | * Process software queues that have been marked busy, splicing them | |
954 | * to the for-dispatch | |
955 | */ | |
2c3ad667 | 956 | void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list) |
1429d7c9 | 957 | { |
88459642 OS |
958 | struct flush_busy_ctx_data data = { |
959 | .hctx = hctx, | |
960 | .list = list, | |
961 | }; | |
1429d7c9 | 962 | |
88459642 | 963 | sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data); |
1429d7c9 | 964 | } |
2c3ad667 | 965 | EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs); |
1429d7c9 | 966 | |
b347689f ML |
967 | struct dispatch_rq_data { |
968 | struct blk_mq_hw_ctx *hctx; | |
969 | struct request *rq; | |
970 | }; | |
971 | ||
972 | static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr, | |
973 | void *data) | |
974 | { | |
975 | struct dispatch_rq_data *dispatch_data = data; | |
976 | struct blk_mq_hw_ctx *hctx = dispatch_data->hctx; | |
977 | struct blk_mq_ctx *ctx = hctx->ctxs[bitnr]; | |
978 | ||
979 | spin_lock(&ctx->lock); | |
980 | if (unlikely(!list_empty(&ctx->rq_list))) { | |
981 | dispatch_data->rq = list_entry_rq(ctx->rq_list.next); | |
982 | list_del_init(&dispatch_data->rq->queuelist); | |
983 | if (list_empty(&ctx->rq_list)) | |
984 | sbitmap_clear_bit(sb, bitnr); | |
985 | } | |
986 | spin_unlock(&ctx->lock); | |
987 | ||
988 | return !dispatch_data->rq; | |
989 | } | |
990 | ||
991 | struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx, | |
992 | struct blk_mq_ctx *start) | |
993 | { | |
994 | unsigned off = start ? start->index_hw : 0; | |
995 | struct dispatch_rq_data data = { | |
996 | .hctx = hctx, | |
997 | .rq = NULL, | |
998 | }; | |
999 | ||
1000 | __sbitmap_for_each_set(&hctx->ctx_map, off, | |
1001 | dispatch_rq_from_ctx, &data); | |
1002 | ||
1003 | return data.rq; | |
1004 | } | |
1005 | ||
703fd1c0 JA |
1006 | static inline unsigned int queued_to_index(unsigned int queued) |
1007 | { | |
1008 | if (!queued) | |
1009 | return 0; | |
1429d7c9 | 1010 | |
703fd1c0 | 1011 | return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1); |
1429d7c9 JA |
1012 | } |
1013 | ||
bd6737f1 JA |
1014 | bool blk_mq_get_driver_tag(struct request *rq, struct blk_mq_hw_ctx **hctx, |
1015 | bool wait) | |
bd166ef1 JA |
1016 | { |
1017 | struct blk_mq_alloc_data data = { | |
1018 | .q = rq->q, | |
bd166ef1 JA |
1019 | .hctx = blk_mq_map_queue(rq->q, rq->mq_ctx->cpu), |
1020 | .flags = wait ? 0 : BLK_MQ_REQ_NOWAIT, | |
1021 | }; | |
662e29d9 | 1022 | bool shared; |
bd166ef1 | 1023 | |
5feeacdd JA |
1024 | might_sleep_if(wait); |
1025 | ||
81380ca1 OS |
1026 | if (rq->tag != -1) |
1027 | goto done; | |
bd166ef1 | 1028 | |
415b806d SG |
1029 | if (blk_mq_tag_is_reserved(data.hctx->sched_tags, rq->internal_tag)) |
1030 | data.flags |= BLK_MQ_REQ_RESERVED; | |
1031 | ||
662e29d9 | 1032 | shared = blk_mq_tag_busy(data.hctx); |
bd166ef1 JA |
1033 | rq->tag = blk_mq_get_tag(&data); |
1034 | if (rq->tag >= 0) { | |
662e29d9 | 1035 | if (shared) { |
200e86b3 JA |
1036 | rq->rq_flags |= RQF_MQ_INFLIGHT; |
1037 | atomic_inc(&data.hctx->nr_active); | |
1038 | } | |
bd166ef1 | 1039 | data.hctx->tags->rqs[rq->tag] = rq; |
bd166ef1 JA |
1040 | } |
1041 | ||
81380ca1 OS |
1042 | done: |
1043 | if (hctx) | |
1044 | *hctx = data.hctx; | |
1045 | return rq->tag != -1; | |
bd166ef1 JA |
1046 | } |
1047 | ||
eb619fdb JA |
1048 | static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode, |
1049 | int flags, void *key) | |
da55f2cc OS |
1050 | { |
1051 | struct blk_mq_hw_ctx *hctx; | |
1052 | ||
1053 | hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait); | |
1054 | ||
eb619fdb | 1055 | list_del_init(&wait->entry); |
da55f2cc OS |
1056 | blk_mq_run_hw_queue(hctx, true); |
1057 | return 1; | |
1058 | } | |
1059 | ||
f906a6a0 JA |
1060 | /* |
1061 | * Mark us waiting for a tag. For shared tags, this involves hooking us into | |
1062 | * the tag wakeups. For non-shared tags, we can simply mark us nedeing a | |
1063 | * restart. For both caes, take care to check the condition again after | |
1064 | * marking us as waiting. | |
1065 | */ | |
1066 | static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx **hctx, | |
1067 | struct request *rq) | |
da55f2cc | 1068 | { |
eb619fdb | 1069 | struct blk_mq_hw_ctx *this_hctx = *hctx; |
f906a6a0 | 1070 | bool shared_tags = (this_hctx->flags & BLK_MQ_F_TAG_SHARED) != 0; |
da55f2cc | 1071 | struct sbq_wait_state *ws; |
f906a6a0 JA |
1072 | wait_queue_entry_t *wait; |
1073 | bool ret; | |
da55f2cc | 1074 | |
f906a6a0 JA |
1075 | if (!shared_tags) { |
1076 | if (!test_bit(BLK_MQ_S_SCHED_RESTART, &this_hctx->state)) | |
1077 | set_bit(BLK_MQ_S_SCHED_RESTART, &this_hctx->state); | |
1078 | } else { | |
1079 | wait = &this_hctx->dispatch_wait; | |
1080 | if (!list_empty_careful(&wait->entry)) | |
1081 | return false; | |
1082 | ||
1083 | spin_lock(&this_hctx->lock); | |
1084 | if (!list_empty(&wait->entry)) { | |
1085 | spin_unlock(&this_hctx->lock); | |
1086 | return false; | |
1087 | } | |
eb619fdb | 1088 | |
f906a6a0 JA |
1089 | ws = bt_wait_ptr(&this_hctx->tags->bitmap_tags, this_hctx); |
1090 | add_wait_queue(&ws->wait, wait); | |
eb619fdb JA |
1091 | } |
1092 | ||
da55f2cc | 1093 | /* |
eb619fdb JA |
1094 | * It's possible that a tag was freed in the window between the |
1095 | * allocation failure and adding the hardware queue to the wait | |
1096 | * queue. | |
da55f2cc | 1097 | */ |
f906a6a0 JA |
1098 | ret = blk_mq_get_driver_tag(rq, hctx, false); |
1099 | ||
1100 | if (!shared_tags) { | |
1101 | /* | |
1102 | * Don't clear RESTART here, someone else could have set it. | |
1103 | * At most this will cost an extra queue run. | |
1104 | */ | |
1105 | return ret; | |
1106 | } else { | |
1107 | if (!ret) { | |
1108 | spin_unlock(&this_hctx->lock); | |
1109 | return false; | |
1110 | } | |
1111 | ||
1112 | /* | |
1113 | * We got a tag, remove ourselves from the wait queue to ensure | |
1114 | * someone else gets the wakeup. | |
1115 | */ | |
1116 | spin_lock_irq(&ws->wait.lock); | |
1117 | list_del_init(&wait->entry); | |
1118 | spin_unlock_irq(&ws->wait.lock); | |
eb619fdb | 1119 | spin_unlock(&this_hctx->lock); |
f906a6a0 | 1120 | return true; |
eb619fdb | 1121 | } |
da55f2cc OS |
1122 | } |
1123 | ||
de148297 | 1124 | bool blk_mq_dispatch_rq_list(struct request_queue *q, struct list_head *list, |
eb619fdb | 1125 | bool got_budget) |
320ae51f | 1126 | { |
81380ca1 | 1127 | struct blk_mq_hw_ctx *hctx; |
6d6f167c | 1128 | struct request *rq, *nxt; |
eb619fdb | 1129 | bool no_tag = false; |
fc17b653 | 1130 | int errors, queued; |
320ae51f | 1131 | |
81380ca1 OS |
1132 | if (list_empty(list)) |
1133 | return false; | |
1134 | ||
de148297 ML |
1135 | WARN_ON(!list_is_singular(list) && got_budget); |
1136 | ||
320ae51f JA |
1137 | /* |
1138 | * Now process all the entries, sending them to the driver. | |
1139 | */ | |
93efe981 | 1140 | errors = queued = 0; |
81380ca1 | 1141 | do { |
74c45052 | 1142 | struct blk_mq_queue_data bd; |
fc17b653 | 1143 | blk_status_t ret; |
320ae51f | 1144 | |
f04c3df3 | 1145 | rq = list_first_entry(list, struct request, queuelist); |
462131c7 ML |
1146 | |
1147 | hctx = blk_mq_map_queue(rq->q, rq->mq_ctx->cpu); | |
1148 | if (!got_budget && !blk_mq_get_dispatch_budget(hctx)) | |
1149 | break; | |
1150 | ||
1151 | if (!blk_mq_get_driver_tag(rq, NULL, false)) { | |
3c782d67 | 1152 | /* |
da55f2cc | 1153 | * The initial allocation attempt failed, so we need to |
eb619fdb JA |
1154 | * rerun the hardware queue when a tag is freed. The |
1155 | * waitqueue takes care of that. If the queue is run | |
1156 | * before we add this entry back on the dispatch list, | |
1157 | * we'll re-run it below. | |
3c782d67 | 1158 | */ |
f906a6a0 | 1159 | if (!blk_mq_mark_tag_wait(&hctx, rq)) { |
462131c7 | 1160 | blk_mq_put_dispatch_budget(hctx); |
f906a6a0 JA |
1161 | /* |
1162 | * For non-shared tags, the RESTART check | |
1163 | * will suffice. | |
1164 | */ | |
1165 | if (hctx->flags & BLK_MQ_F_TAG_SHARED) | |
1166 | no_tag = true; | |
de148297 ML |
1167 | break; |
1168 | } | |
1169 | } | |
1170 | ||
320ae51f | 1171 | list_del_init(&rq->queuelist); |
320ae51f | 1172 | |
74c45052 | 1173 | bd.rq = rq; |
113285b4 JA |
1174 | |
1175 | /* | |
1176 | * Flag last if we have no more requests, or if we have more | |
1177 | * but can't assign a driver tag to it. | |
1178 | */ | |
1179 | if (list_empty(list)) | |
1180 | bd.last = true; | |
1181 | else { | |
113285b4 JA |
1182 | nxt = list_first_entry(list, struct request, queuelist); |
1183 | bd.last = !blk_mq_get_driver_tag(nxt, NULL, false); | |
1184 | } | |
74c45052 JA |
1185 | |
1186 | ret = q->mq_ops->queue_rq(hctx, &bd); | |
fc17b653 | 1187 | if (ret == BLK_STS_RESOURCE) { |
6d6f167c JW |
1188 | /* |
1189 | * If an I/O scheduler has been configured and we got a | |
ff821d27 JA |
1190 | * driver tag for the next request already, free it |
1191 | * again. | |
6d6f167c JW |
1192 | */ |
1193 | if (!list_empty(list)) { | |
1194 | nxt = list_first_entry(list, struct request, queuelist); | |
1195 | blk_mq_put_driver_tag(nxt); | |
1196 | } | |
f04c3df3 | 1197 | list_add(&rq->queuelist, list); |
ed0791b2 | 1198 | __blk_mq_requeue_request(rq); |
320ae51f | 1199 | break; |
fc17b653 CH |
1200 | } |
1201 | ||
1202 | if (unlikely(ret != BLK_STS_OK)) { | |
93efe981 | 1203 | errors++; |
2a842aca | 1204 | blk_mq_end_request(rq, BLK_STS_IOERR); |
fc17b653 | 1205 | continue; |
320ae51f JA |
1206 | } |
1207 | ||
fc17b653 | 1208 | queued++; |
81380ca1 | 1209 | } while (!list_empty(list)); |
320ae51f | 1210 | |
703fd1c0 | 1211 | hctx->dispatched[queued_to_index(queued)]++; |
320ae51f JA |
1212 | |
1213 | /* | |
1214 | * Any items that need requeuing? Stuff them into hctx->dispatch, | |
1215 | * that is where we will continue on next queue run. | |
1216 | */ | |
f04c3df3 | 1217 | if (!list_empty(list)) { |
320ae51f | 1218 | spin_lock(&hctx->lock); |
c13660a0 | 1219 | list_splice_init(list, &hctx->dispatch); |
320ae51f | 1220 | spin_unlock(&hctx->lock); |
f04c3df3 | 1221 | |
9ba52e58 | 1222 | /* |
710c785f BVA |
1223 | * If SCHED_RESTART was set by the caller of this function and |
1224 | * it is no longer set that means that it was cleared by another | |
1225 | * thread and hence that a queue rerun is needed. | |
9ba52e58 | 1226 | * |
eb619fdb JA |
1227 | * If 'no_tag' is set, that means that we failed getting |
1228 | * a driver tag with an I/O scheduler attached. If our dispatch | |
1229 | * waitqueue is no longer active, ensure that we run the queue | |
1230 | * AFTER adding our entries back to the list. | |
bd166ef1 | 1231 | * |
710c785f BVA |
1232 | * If no I/O scheduler has been configured it is possible that |
1233 | * the hardware queue got stopped and restarted before requests | |
1234 | * were pushed back onto the dispatch list. Rerun the queue to | |
1235 | * avoid starvation. Notes: | |
1236 | * - blk_mq_run_hw_queue() checks whether or not a queue has | |
1237 | * been stopped before rerunning a queue. | |
1238 | * - Some but not all block drivers stop a queue before | |
fc17b653 | 1239 | * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq |
710c785f | 1240 | * and dm-rq. |
bd166ef1 | 1241 | */ |
eb619fdb JA |
1242 | if (!blk_mq_sched_needs_restart(hctx) || |
1243 | (no_tag && list_empty_careful(&hctx->dispatch_wait.entry))) | |
bd166ef1 | 1244 | blk_mq_run_hw_queue(hctx, true); |
320ae51f | 1245 | } |
f04c3df3 | 1246 | |
93efe981 | 1247 | return (queued + errors) != 0; |
f04c3df3 JA |
1248 | } |
1249 | ||
6a83e74d BVA |
1250 | static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx) |
1251 | { | |
1252 | int srcu_idx; | |
1253 | ||
b7a71e66 JA |
1254 | /* |
1255 | * We should be running this queue from one of the CPUs that | |
1256 | * are mapped to it. | |
54011168 ML |
1257 | * |
1258 | * There are at least two related races now between setting | |
1259 | * hctx->next_cpu from blk_mq_hctx_next_cpu() and running | |
1260 | * __blk_mq_run_hw_queue(): | |
1261 | * | |
1262 | * - hctx->next_cpu is found offline in blk_mq_hctx_next_cpu(), | |
1263 | * but later it becomes online, then this warning is harmless | |
1264 | * at all | |
1265 | * | |
1266 | * - hctx->next_cpu is found online in blk_mq_hctx_next_cpu(), | |
1267 | * but later it becomes offline, then the warning can't be | |
1268 | * triggered, and we depend on blk-mq timeout handler to | |
1269 | * handle dispatched requests to this hctx | |
b7a71e66 | 1270 | */ |
54011168 ML |
1271 | if (!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) && |
1272 | cpu_online(hctx->next_cpu)) { | |
1273 | printk(KERN_WARNING "run queue from wrong CPU %d, hctx %s\n", | |
1274 | raw_smp_processor_id(), | |
1275 | cpumask_empty(hctx->cpumask) ? "inactive": "active"); | |
1276 | dump_stack(); | |
1277 | } | |
6a83e74d | 1278 | |
b7a71e66 JA |
1279 | /* |
1280 | * We can't run the queue inline with ints disabled. Ensure that | |
1281 | * we catch bad users of this early. | |
1282 | */ | |
1283 | WARN_ON_ONCE(in_interrupt()); | |
1284 | ||
74e11d8d | 1285 | might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING); |
bf4907c0 | 1286 | |
74e11d8d JA |
1287 | hctx_lock(hctx, &srcu_idx); |
1288 | blk_mq_sched_dispatch_requests(hctx); | |
1289 | hctx_unlock(hctx, srcu_idx); | |
6a83e74d BVA |
1290 | } |
1291 | ||
6204c099 ML |
1292 | static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx) |
1293 | { | |
1294 | int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask); | |
1295 | ||
1296 | if (cpu >= nr_cpu_ids) | |
1297 | cpu = cpumask_first(hctx->cpumask); | |
1298 | return cpu; | |
1299 | } | |
1300 | ||
506e931f JA |
1301 | /* |
1302 | * It'd be great if the workqueue API had a way to pass | |
1303 | * in a mask and had some smarts for more clever placement. | |
1304 | * For now we just round-robin here, switching for every | |
1305 | * BLK_MQ_CPU_WORK_BATCH queued items. | |
1306 | */ | |
1307 | static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx) | |
1308 | { | |
9027a417 | 1309 | bool tried = false; |
88f4a962 | 1310 | int next_cpu = hctx->next_cpu; |
9027a417 | 1311 | |
b657d7e6 CH |
1312 | if (hctx->queue->nr_hw_queues == 1) |
1313 | return WORK_CPU_UNBOUND; | |
506e931f JA |
1314 | |
1315 | if (--hctx->next_cpu_batch <= 0) { | |
9027a417 | 1316 | select_cpu: |
88f4a962 | 1317 | next_cpu = cpumask_next_and(next_cpu, hctx->cpumask, |
8a591698 | 1318 | cpu_online_mask); |
506e931f | 1319 | if (next_cpu >= nr_cpu_ids) |
6204c099 | 1320 | next_cpu = blk_mq_first_mapped_cpu(hctx); |
506e931f JA |
1321 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; |
1322 | } | |
1323 | ||
9027a417 ML |
1324 | /* |
1325 | * Do unbound schedule if we can't find a online CPU for this hctx, | |
1326 | * and it should only happen in the path of handling CPU DEAD. | |
1327 | */ | |
88f4a962 | 1328 | if (!cpu_online(next_cpu)) { |
9027a417 ML |
1329 | if (!tried) { |
1330 | tried = true; | |
1331 | goto select_cpu; | |
1332 | } | |
1333 | ||
1334 | /* | |
1335 | * Make sure to re-select CPU next time once after CPUs | |
1336 | * in hctx->cpumask become online again. | |
1337 | */ | |
88f4a962 | 1338 | hctx->next_cpu = next_cpu; |
9027a417 ML |
1339 | hctx->next_cpu_batch = 1; |
1340 | return WORK_CPU_UNBOUND; | |
1341 | } | |
88f4a962 ML |
1342 | |
1343 | hctx->next_cpu = next_cpu; | |
1344 | return next_cpu; | |
506e931f JA |
1345 | } |
1346 | ||
7587a5ae BVA |
1347 | static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async, |
1348 | unsigned long msecs) | |
320ae51f | 1349 | { |
5435c023 | 1350 | if (unlikely(blk_mq_hctx_stopped(hctx))) |
320ae51f JA |
1351 | return; |
1352 | ||
1b792f2f | 1353 | if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) { |
2a90d4aa PB |
1354 | int cpu = get_cpu(); |
1355 | if (cpumask_test_cpu(cpu, hctx->cpumask)) { | |
398205b8 | 1356 | __blk_mq_run_hw_queue(hctx); |
2a90d4aa | 1357 | put_cpu(); |
398205b8 PB |
1358 | return; |
1359 | } | |
e4043dcf | 1360 | |
2a90d4aa | 1361 | put_cpu(); |
e4043dcf | 1362 | } |
398205b8 | 1363 | |
81517da9 BVA |
1364 | kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work, |
1365 | msecs_to_jiffies(msecs)); | |
7587a5ae BVA |
1366 | } |
1367 | ||
1368 | void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs) | |
1369 | { | |
1370 | __blk_mq_delay_run_hw_queue(hctx, true, msecs); | |
1371 | } | |
1372 | EXPORT_SYMBOL(blk_mq_delay_run_hw_queue); | |
1373 | ||
79f720a7 | 1374 | bool blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) |
7587a5ae | 1375 | { |
d2265003 ML |
1376 | int srcu_idx; |
1377 | bool need_run; | |
1378 | ||
1379 | /* | |
1380 | * When queue is quiesced, we may be switching io scheduler, or | |
1381 | * updating nr_hw_queues, or other things, and we can't run queue | |
1382 | * any more, even __blk_mq_hctx_has_pending() can't be called safely. | |
1383 | * | |
1384 | * And queue will be rerun in blk_mq_unquiesce_queue() if it is | |
1385 | * quiesced. | |
1386 | */ | |
74e11d8d JA |
1387 | hctx_lock(hctx, &srcu_idx); |
1388 | need_run = !blk_queue_quiesced(hctx->queue) && | |
1389 | blk_mq_hctx_has_pending(hctx); | |
1390 | hctx_unlock(hctx, srcu_idx); | |
d2265003 ML |
1391 | |
1392 | if (need_run) { | |
79f720a7 JA |
1393 | __blk_mq_delay_run_hw_queue(hctx, async, 0); |
1394 | return true; | |
1395 | } | |
1396 | ||
1397 | return false; | |
320ae51f | 1398 | } |
5b727272 | 1399 | EXPORT_SYMBOL(blk_mq_run_hw_queue); |
320ae51f | 1400 | |
b94ec296 | 1401 | void blk_mq_run_hw_queues(struct request_queue *q, bool async) |
320ae51f JA |
1402 | { |
1403 | struct blk_mq_hw_ctx *hctx; | |
1404 | int i; | |
1405 | ||
1406 | queue_for_each_hw_ctx(q, hctx, i) { | |
79f720a7 | 1407 | if (blk_mq_hctx_stopped(hctx)) |
320ae51f JA |
1408 | continue; |
1409 | ||
b94ec296 | 1410 | blk_mq_run_hw_queue(hctx, async); |
320ae51f JA |
1411 | } |
1412 | } | |
b94ec296 | 1413 | EXPORT_SYMBOL(blk_mq_run_hw_queues); |
320ae51f | 1414 | |
fd001443 BVA |
1415 | /** |
1416 | * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped | |
1417 | * @q: request queue. | |
1418 | * | |
1419 | * The caller is responsible for serializing this function against | |
1420 | * blk_mq_{start,stop}_hw_queue(). | |
1421 | */ | |
1422 | bool blk_mq_queue_stopped(struct request_queue *q) | |
1423 | { | |
1424 | struct blk_mq_hw_ctx *hctx; | |
1425 | int i; | |
1426 | ||
1427 | queue_for_each_hw_ctx(q, hctx, i) | |
1428 | if (blk_mq_hctx_stopped(hctx)) | |
1429 | return true; | |
1430 | ||
1431 | return false; | |
1432 | } | |
1433 | EXPORT_SYMBOL(blk_mq_queue_stopped); | |
1434 | ||
39a70c76 ML |
1435 | /* |
1436 | * This function is often used for pausing .queue_rq() by driver when | |
1437 | * there isn't enough resource or some conditions aren't satisfied, and | |
4d606219 | 1438 | * BLK_STS_RESOURCE is usually returned. |
39a70c76 ML |
1439 | * |
1440 | * We do not guarantee that dispatch can be drained or blocked | |
1441 | * after blk_mq_stop_hw_queue() returns. Please use | |
1442 | * blk_mq_quiesce_queue() for that requirement. | |
1443 | */ | |
2719aa21 JA |
1444 | void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx) |
1445 | { | |
641a9ed6 | 1446 | cancel_delayed_work(&hctx->run_work); |
280d45f6 | 1447 | |
641a9ed6 | 1448 | set_bit(BLK_MQ_S_STOPPED, &hctx->state); |
2719aa21 | 1449 | } |
641a9ed6 | 1450 | EXPORT_SYMBOL(blk_mq_stop_hw_queue); |
2719aa21 | 1451 | |
39a70c76 ML |
1452 | /* |
1453 | * This function is often used for pausing .queue_rq() by driver when | |
1454 | * there isn't enough resource or some conditions aren't satisfied, and | |
4d606219 | 1455 | * BLK_STS_RESOURCE is usually returned. |
39a70c76 ML |
1456 | * |
1457 | * We do not guarantee that dispatch can be drained or blocked | |
1458 | * after blk_mq_stop_hw_queues() returns. Please use | |
1459 | * blk_mq_quiesce_queue() for that requirement. | |
1460 | */ | |
2719aa21 JA |
1461 | void blk_mq_stop_hw_queues(struct request_queue *q) |
1462 | { | |
641a9ed6 ML |
1463 | struct blk_mq_hw_ctx *hctx; |
1464 | int i; | |
1465 | ||
1466 | queue_for_each_hw_ctx(q, hctx, i) | |
1467 | blk_mq_stop_hw_queue(hctx); | |
280d45f6 CH |
1468 | } |
1469 | EXPORT_SYMBOL(blk_mq_stop_hw_queues); | |
1470 | ||
320ae51f JA |
1471 | void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx) |
1472 | { | |
1473 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
e4043dcf | 1474 | |
0ffbce80 | 1475 | blk_mq_run_hw_queue(hctx, false); |
320ae51f JA |
1476 | } |
1477 | EXPORT_SYMBOL(blk_mq_start_hw_queue); | |
1478 | ||
2f268556 CH |
1479 | void blk_mq_start_hw_queues(struct request_queue *q) |
1480 | { | |
1481 | struct blk_mq_hw_ctx *hctx; | |
1482 | int i; | |
1483 | ||
1484 | queue_for_each_hw_ctx(q, hctx, i) | |
1485 | blk_mq_start_hw_queue(hctx); | |
1486 | } | |
1487 | EXPORT_SYMBOL(blk_mq_start_hw_queues); | |
1488 | ||
ae911c5e JA |
1489 | void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) |
1490 | { | |
1491 | if (!blk_mq_hctx_stopped(hctx)) | |
1492 | return; | |
1493 | ||
1494 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
1495 | blk_mq_run_hw_queue(hctx, async); | |
1496 | } | |
1497 | EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue); | |
1498 | ||
1b4a3258 | 1499 | void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async) |
320ae51f JA |
1500 | { |
1501 | struct blk_mq_hw_ctx *hctx; | |
1502 | int i; | |
1503 | ||
ae911c5e JA |
1504 | queue_for_each_hw_ctx(q, hctx, i) |
1505 | blk_mq_start_stopped_hw_queue(hctx, async); | |
320ae51f JA |
1506 | } |
1507 | EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues); | |
1508 | ||
70f4db63 | 1509 | static void blk_mq_run_work_fn(struct work_struct *work) |
320ae51f JA |
1510 | { |
1511 | struct blk_mq_hw_ctx *hctx; | |
1512 | ||
9f993737 | 1513 | hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work); |
320ae51f | 1514 | |
21c6e939 JA |
1515 | /* |
1516 | * If we are stopped, don't run the queue. The exception is if | |
1517 | * BLK_MQ_S_START_ON_RUN is set. For that case, we auto-clear | |
1518 | * the STOPPED bit and run it. | |
1519 | */ | |
1520 | if (test_bit(BLK_MQ_S_STOPPED, &hctx->state)) { | |
1521 | if (!test_bit(BLK_MQ_S_START_ON_RUN, &hctx->state)) | |
1522 | return; | |
7587a5ae | 1523 | |
21c6e939 JA |
1524 | clear_bit(BLK_MQ_S_START_ON_RUN, &hctx->state); |
1525 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
1526 | } | |
7587a5ae BVA |
1527 | |
1528 | __blk_mq_run_hw_queue(hctx); | |
1529 | } | |
1530 | ||
70f4db63 CH |
1531 | |
1532 | void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs) | |
1533 | { | |
5435c023 | 1534 | if (WARN_ON_ONCE(!blk_mq_hw_queue_mapped(hctx))) |
19c66e59 | 1535 | return; |
70f4db63 | 1536 | |
21c6e939 JA |
1537 | /* |
1538 | * Stop the hw queue, then modify currently delayed work. | |
1539 | * This should prevent us from running the queue prematurely. | |
1540 | * Mark the queue as auto-clearing STOPPED when it runs. | |
1541 | */ | |
7e79dadc | 1542 | blk_mq_stop_hw_queue(hctx); |
21c6e939 JA |
1543 | set_bit(BLK_MQ_S_START_ON_RUN, &hctx->state); |
1544 | kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), | |
1545 | &hctx->run_work, | |
1546 | msecs_to_jiffies(msecs)); | |
70f4db63 CH |
1547 | } |
1548 | EXPORT_SYMBOL(blk_mq_delay_queue); | |
1549 | ||
cfd0c552 | 1550 | static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx, |
cfd0c552 ML |
1551 | struct request *rq, |
1552 | bool at_head) | |
320ae51f | 1553 | { |
e57690fe JA |
1554 | struct blk_mq_ctx *ctx = rq->mq_ctx; |
1555 | ||
7b607814 BVA |
1556 | lockdep_assert_held(&ctx->lock); |
1557 | ||
01b983c9 JA |
1558 | trace_block_rq_insert(hctx->queue, rq); |
1559 | ||
72a0a36e CH |
1560 | if (at_head) |
1561 | list_add(&rq->queuelist, &ctx->rq_list); | |
1562 | else | |
1563 | list_add_tail(&rq->queuelist, &ctx->rq_list); | |
cfd0c552 | 1564 | } |
4bb659b1 | 1565 | |
2c3ad667 JA |
1566 | void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq, |
1567 | bool at_head) | |
cfd0c552 ML |
1568 | { |
1569 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
1570 | ||
7b607814 BVA |
1571 | lockdep_assert_held(&ctx->lock); |
1572 | ||
e57690fe | 1573 | __blk_mq_insert_req_list(hctx, rq, at_head); |
320ae51f | 1574 | blk_mq_hctx_mark_pending(hctx, ctx); |
320ae51f JA |
1575 | } |
1576 | ||
157f377b JA |
1577 | /* |
1578 | * Should only be used carefully, when the caller knows we want to | |
1579 | * bypass a potential IO scheduler on the target device. | |
1580 | */ | |
b0850297 | 1581 | void blk_mq_request_bypass_insert(struct request *rq, bool run_queue) |
157f377b JA |
1582 | { |
1583 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
1584 | struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(rq->q, ctx->cpu); | |
1585 | ||
1586 | spin_lock(&hctx->lock); | |
1587 | list_add_tail(&rq->queuelist, &hctx->dispatch); | |
1588 | spin_unlock(&hctx->lock); | |
1589 | ||
b0850297 ML |
1590 | if (run_queue) |
1591 | blk_mq_run_hw_queue(hctx, false); | |
157f377b JA |
1592 | } |
1593 | ||
bd166ef1 JA |
1594 | void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx, |
1595 | struct list_head *list) | |
320ae51f JA |
1596 | |
1597 | { | |
320ae51f JA |
1598 | /* |
1599 | * preemption doesn't flush plug list, so it's possible ctx->cpu is | |
1600 | * offline now | |
1601 | */ | |
1602 | spin_lock(&ctx->lock); | |
1603 | while (!list_empty(list)) { | |
1604 | struct request *rq; | |
1605 | ||
1606 | rq = list_first_entry(list, struct request, queuelist); | |
e57690fe | 1607 | BUG_ON(rq->mq_ctx != ctx); |
320ae51f | 1608 | list_del_init(&rq->queuelist); |
e57690fe | 1609 | __blk_mq_insert_req_list(hctx, rq, false); |
320ae51f | 1610 | } |
cfd0c552 | 1611 | blk_mq_hctx_mark_pending(hctx, ctx); |
320ae51f | 1612 | spin_unlock(&ctx->lock); |
320ae51f JA |
1613 | } |
1614 | ||
1615 | static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b) | |
1616 | { | |
1617 | struct request *rqa = container_of(a, struct request, queuelist); | |
1618 | struct request *rqb = container_of(b, struct request, queuelist); | |
1619 | ||
1620 | return !(rqa->mq_ctx < rqb->mq_ctx || | |
1621 | (rqa->mq_ctx == rqb->mq_ctx && | |
1622 | blk_rq_pos(rqa) < blk_rq_pos(rqb))); | |
1623 | } | |
1624 | ||
1625 | void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule) | |
1626 | { | |
1627 | struct blk_mq_ctx *this_ctx; | |
1628 | struct request_queue *this_q; | |
1629 | struct request *rq; | |
1630 | LIST_HEAD(list); | |
1631 | LIST_HEAD(ctx_list); | |
1632 | unsigned int depth; | |
1633 | ||
1634 | list_splice_init(&plug->mq_list, &list); | |
1635 | ||
1636 | list_sort(NULL, &list, plug_ctx_cmp); | |
1637 | ||
1638 | this_q = NULL; | |
1639 | this_ctx = NULL; | |
1640 | depth = 0; | |
1641 | ||
1642 | while (!list_empty(&list)) { | |
1643 | rq = list_entry_rq(list.next); | |
1644 | list_del_init(&rq->queuelist); | |
1645 | BUG_ON(!rq->q); | |
1646 | if (rq->mq_ctx != this_ctx) { | |
1647 | if (this_ctx) { | |
0f7cb0db | 1648 | trace_block_unplug(this_q, depth, !from_schedule); |
bd166ef1 JA |
1649 | blk_mq_sched_insert_requests(this_q, this_ctx, |
1650 | &ctx_list, | |
1651 | from_schedule); | |
320ae51f JA |
1652 | } |
1653 | ||
1654 | this_ctx = rq->mq_ctx; | |
1655 | this_q = rq->q; | |
1656 | depth = 0; | |
1657 | } | |
1658 | ||
1659 | depth++; | |
1660 | list_add_tail(&rq->queuelist, &ctx_list); | |
1661 | } | |
1662 | ||
1663 | /* | |
1664 | * If 'this_ctx' is set, we know we have entries to complete | |
1665 | * on 'ctx_list'. Do those. | |
1666 | */ | |
1667 | if (this_ctx) { | |
0f7cb0db | 1668 | trace_block_unplug(this_q, depth, !from_schedule); |
bd166ef1 JA |
1669 | blk_mq_sched_insert_requests(this_q, this_ctx, &ctx_list, |
1670 | from_schedule); | |
320ae51f JA |
1671 | } |
1672 | } | |
1673 | ||
1674 | static void blk_mq_bio_to_request(struct request *rq, struct bio *bio) | |
1675 | { | |
da8d7f07 | 1676 | blk_init_request_from_bio(rq, bio); |
4b570521 | 1677 | |
85acb3ba SL |
1678 | blk_rq_set_rl(rq, blk_get_rl(rq->q, bio)); |
1679 | ||
6e85eaf3 | 1680 | blk_account_io_start(rq, true); |
320ae51f JA |
1681 | } |
1682 | ||
ab42f35d ML |
1683 | static inline void blk_mq_queue_io(struct blk_mq_hw_ctx *hctx, |
1684 | struct blk_mq_ctx *ctx, | |
1685 | struct request *rq) | |
1686 | { | |
1687 | spin_lock(&ctx->lock); | |
1688 | __blk_mq_insert_request(hctx, rq, false); | |
1689 | spin_unlock(&ctx->lock); | |
07068d5b | 1690 | } |
14ec77f3 | 1691 | |
fd2d3326 JA |
1692 | static blk_qc_t request_to_qc_t(struct blk_mq_hw_ctx *hctx, struct request *rq) |
1693 | { | |
bd166ef1 JA |
1694 | if (rq->tag != -1) |
1695 | return blk_tag_to_qc_t(rq->tag, hctx->queue_num, false); | |
1696 | ||
1697 | return blk_tag_to_qc_t(rq->internal_tag, hctx->queue_num, true); | |
fd2d3326 JA |
1698 | } |
1699 | ||
d964f04a ML |
1700 | static void __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx, |
1701 | struct request *rq, | |
74e11d8d | 1702 | blk_qc_t *cookie) |
f984df1f | 1703 | { |
f984df1f | 1704 | struct request_queue *q = rq->q; |
f984df1f SL |
1705 | struct blk_mq_queue_data bd = { |
1706 | .rq = rq, | |
d945a365 | 1707 | .last = true, |
f984df1f | 1708 | }; |
bd166ef1 | 1709 | blk_qc_t new_cookie; |
f06345ad | 1710 | blk_status_t ret; |
d964f04a ML |
1711 | bool run_queue = true; |
1712 | ||
f4560ffe ML |
1713 | /* RCU or SRCU read lock is needed before checking quiesced flag */ |
1714 | if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)) { | |
d964f04a ML |
1715 | run_queue = false; |
1716 | goto insert; | |
1717 | } | |
f984df1f | 1718 | |
bd166ef1 | 1719 | if (q->elevator) |
2253efc8 BVA |
1720 | goto insert; |
1721 | ||
462131c7 | 1722 | if (!blk_mq_get_dispatch_budget(hctx)) |
bd166ef1 JA |
1723 | goto insert; |
1724 | ||
462131c7 ML |
1725 | if (!blk_mq_get_driver_tag(rq, NULL, false)) { |
1726 | blk_mq_put_dispatch_budget(hctx); | |
de148297 | 1727 | goto insert; |
88022d72 | 1728 | } |
de148297 | 1729 | |
bd166ef1 JA |
1730 | new_cookie = request_to_qc_t(hctx, rq); |
1731 | ||
f984df1f SL |
1732 | /* |
1733 | * For OK queue, we are done. For error, kill it. Any other | |
1734 | * error (busy), just add it to our list as we previously | |
1735 | * would have done | |
1736 | */ | |
1737 | ret = q->mq_ops->queue_rq(hctx, &bd); | |
fc17b653 CH |
1738 | switch (ret) { |
1739 | case BLK_STS_OK: | |
7b371636 | 1740 | *cookie = new_cookie; |
2253efc8 | 1741 | return; |
fc17b653 CH |
1742 | case BLK_STS_RESOURCE: |
1743 | __blk_mq_requeue_request(rq); | |
1744 | goto insert; | |
1745 | default: | |
7b371636 | 1746 | *cookie = BLK_QC_T_NONE; |
fc17b653 | 1747 | blk_mq_end_request(rq, ret); |
2253efc8 | 1748 | return; |
f984df1f | 1749 | } |
7b371636 | 1750 | |
2253efc8 | 1751 | insert: |
74e11d8d JA |
1752 | blk_mq_sched_insert_request(rq, false, run_queue, false, |
1753 | hctx->flags & BLK_MQ_F_BLOCKING); | |
f984df1f SL |
1754 | } |
1755 | ||
5eb6126e CH |
1756 | static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx, |
1757 | struct request *rq, blk_qc_t *cookie) | |
1758 | { | |
74e11d8d | 1759 | int srcu_idx; |
bf4907c0 | 1760 | |
74e11d8d | 1761 | might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING); |
bf4907c0 | 1762 | |
74e11d8d JA |
1763 | hctx_lock(hctx, &srcu_idx); |
1764 | __blk_mq_try_issue_directly(hctx, rq, cookie); | |
1765 | hctx_unlock(hctx, srcu_idx); | |
5eb6126e CH |
1766 | } |
1767 | ||
dece1635 | 1768 | static blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio) |
07068d5b | 1769 | { |
ef295ecf | 1770 | const int is_sync = op_is_sync(bio->bi_opf); |
f73f44eb | 1771 | const int is_flush_fua = op_is_flush(bio->bi_opf); |
5a797e00 | 1772 | struct blk_mq_alloc_data data = { .flags = 0 }; |
07068d5b | 1773 | struct request *rq; |
5eb6126e | 1774 | unsigned int request_count = 0; |
f984df1f | 1775 | struct blk_plug *plug; |
5b3f341f | 1776 | struct request *same_queue_rq = NULL; |
7b371636 | 1777 | blk_qc_t cookie; |
87760e5e | 1778 | unsigned int wb_acct; |
07068d5b JA |
1779 | |
1780 | blk_queue_bounce(q, &bio); | |
1781 | ||
af67c31f | 1782 | blk_queue_split(q, &bio); |
f36ea50c | 1783 | |
e23947bd | 1784 | if (!bio_integrity_prep(bio)) |
dece1635 | 1785 | return BLK_QC_T_NONE; |
07068d5b | 1786 | |
87c279e6 OS |
1787 | if (!is_flush_fua && !blk_queue_nomerges(q) && |
1788 | blk_attempt_plug_merge(q, bio, &request_count, &same_queue_rq)) | |
1789 | return BLK_QC_T_NONE; | |
f984df1f | 1790 | |
bd166ef1 JA |
1791 | if (blk_mq_sched_bio_merge(q, bio)) |
1792 | return BLK_QC_T_NONE; | |
1793 | ||
87760e5e JA |
1794 | wb_acct = wbt_wait(q->rq_wb, bio, NULL); |
1795 | ||
bd166ef1 JA |
1796 | trace_block_getrq(q, bio, bio->bi_opf); |
1797 | ||
d2c0d383 | 1798 | rq = blk_mq_get_request(q, bio, bio->bi_opf, &data); |
87760e5e JA |
1799 | if (unlikely(!rq)) { |
1800 | __wbt_done(q->rq_wb, wb_acct); | |
03a07c92 GR |
1801 | if (bio->bi_opf & REQ_NOWAIT) |
1802 | bio_wouldblock_error(bio); | |
dece1635 | 1803 | return BLK_QC_T_NONE; |
87760e5e JA |
1804 | } |
1805 | ||
1806 | wbt_track(&rq->issue_stat, wb_acct); | |
07068d5b | 1807 | |
fd2d3326 | 1808 | cookie = request_to_qc_t(data.hctx, rq); |
07068d5b | 1809 | |
f984df1f | 1810 | plug = current->plug; |
07068d5b | 1811 | if (unlikely(is_flush_fua)) { |
f984df1f | 1812 | blk_mq_put_ctx(data.ctx); |
07068d5b | 1813 | blk_mq_bio_to_request(rq, bio); |
923218f6 ML |
1814 | |
1815 | /* bypass scheduler for flush rq */ | |
1816 | blk_insert_flush(rq); | |
1817 | blk_mq_run_hw_queue(data.hctx, true); | |
a4d907b6 | 1818 | } else if (plug && q->nr_hw_queues == 1) { |
600271d9 SL |
1819 | struct request *last = NULL; |
1820 | ||
b00c53e8 | 1821 | blk_mq_put_ctx(data.ctx); |
e6c4438b | 1822 | blk_mq_bio_to_request(rq, bio); |
0a6219a9 ML |
1823 | |
1824 | /* | |
1825 | * @request_count may become stale because of schedule | |
1826 | * out, so check the list again. | |
1827 | */ | |
1828 | if (list_empty(&plug->mq_list)) | |
1829 | request_count = 0; | |
254d259d CH |
1830 | else if (blk_queue_nomerges(q)) |
1831 | request_count = blk_plug_queued_count(q); | |
1832 | ||
676d0607 | 1833 | if (!request_count) |
e6c4438b | 1834 | trace_block_plug(q); |
600271d9 SL |
1835 | else |
1836 | last = list_entry_rq(plug->mq_list.prev); | |
b094f89c | 1837 | |
600271d9 SL |
1838 | if (request_count >= BLK_MAX_REQUEST_COUNT || (last && |
1839 | blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) { | |
e6c4438b JM |
1840 | blk_flush_plug_list(plug, false); |
1841 | trace_block_plug(q); | |
320ae51f | 1842 | } |
b094f89c | 1843 | |
e6c4438b | 1844 | list_add_tail(&rq->queuelist, &plug->mq_list); |
2299722c | 1845 | } else if (plug && !blk_queue_nomerges(q)) { |
bd166ef1 | 1846 | blk_mq_bio_to_request(rq, bio); |
07068d5b | 1847 | |
07068d5b | 1848 | /* |
6a83e74d | 1849 | * We do limited plugging. If the bio can be merged, do that. |
f984df1f SL |
1850 | * Otherwise the existing request in the plug list will be |
1851 | * issued. So the plug list will have one request at most | |
2299722c CH |
1852 | * The plug list might get flushed before this. If that happens, |
1853 | * the plug list is empty, and same_queue_rq is invalid. | |
07068d5b | 1854 | */ |
2299722c CH |
1855 | if (list_empty(&plug->mq_list)) |
1856 | same_queue_rq = NULL; | |
1857 | if (same_queue_rq) | |
1858 | list_del_init(&same_queue_rq->queuelist); | |
1859 | list_add_tail(&rq->queuelist, &plug->mq_list); | |
1860 | ||
bf4907c0 JA |
1861 | blk_mq_put_ctx(data.ctx); |
1862 | ||
dad7a3be ML |
1863 | if (same_queue_rq) { |
1864 | data.hctx = blk_mq_map_queue(q, | |
1865 | same_queue_rq->mq_ctx->cpu); | |
2299722c CH |
1866 | blk_mq_try_issue_directly(data.hctx, same_queue_rq, |
1867 | &cookie); | |
dad7a3be | 1868 | } |
a4d907b6 | 1869 | } else if (q->nr_hw_queues > 1 && is_sync) { |
bf4907c0 | 1870 | blk_mq_put_ctx(data.ctx); |
2299722c | 1871 | blk_mq_bio_to_request(rq, bio); |
2299722c | 1872 | blk_mq_try_issue_directly(data.hctx, rq, &cookie); |
a4d907b6 | 1873 | } else if (q->elevator) { |
b00c53e8 | 1874 | blk_mq_put_ctx(data.ctx); |
bd166ef1 | 1875 | blk_mq_bio_to_request(rq, bio); |
a4d907b6 | 1876 | blk_mq_sched_insert_request(rq, false, true, true, true); |
ab42f35d | 1877 | } else { |
b00c53e8 | 1878 | blk_mq_put_ctx(data.ctx); |
ab42f35d ML |
1879 | blk_mq_bio_to_request(rq, bio); |
1880 | blk_mq_queue_io(data.hctx, data.ctx, rq); | |
a4d907b6 | 1881 | blk_mq_run_hw_queue(data.hctx, true); |
ab42f35d | 1882 | } |
320ae51f | 1883 | |
7b371636 | 1884 | return cookie; |
320ae51f JA |
1885 | } |
1886 | ||
cc71a6f4 JA |
1887 | void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags, |
1888 | unsigned int hctx_idx) | |
95363efd | 1889 | { |
e9b267d9 | 1890 | struct page *page; |
320ae51f | 1891 | |
24d2f903 | 1892 | if (tags->rqs && set->ops->exit_request) { |
e9b267d9 | 1893 | int i; |
320ae51f | 1894 | |
24d2f903 | 1895 | for (i = 0; i < tags->nr_tags; i++) { |
2af8cbe3 JA |
1896 | struct request *rq = tags->static_rqs[i]; |
1897 | ||
1898 | if (!rq) | |
e9b267d9 | 1899 | continue; |
d6296d39 | 1900 | set->ops->exit_request(set, rq, hctx_idx); |
2af8cbe3 | 1901 | tags->static_rqs[i] = NULL; |
e9b267d9 | 1902 | } |
320ae51f | 1903 | } |
320ae51f | 1904 | |
24d2f903 CH |
1905 | while (!list_empty(&tags->page_list)) { |
1906 | page = list_first_entry(&tags->page_list, struct page, lru); | |
6753471c | 1907 | list_del_init(&page->lru); |
f75782e4 CM |
1908 | /* |
1909 | * Remove kmemleak object previously allocated in | |
1910 | * blk_mq_init_rq_map(). | |
1911 | */ | |
1912 | kmemleak_free(page_address(page)); | |
320ae51f JA |
1913 | __free_pages(page, page->private); |
1914 | } | |
cc71a6f4 | 1915 | } |
320ae51f | 1916 | |
cc71a6f4 JA |
1917 | void blk_mq_free_rq_map(struct blk_mq_tags *tags) |
1918 | { | |
24d2f903 | 1919 | kfree(tags->rqs); |
cc71a6f4 | 1920 | tags->rqs = NULL; |
2af8cbe3 JA |
1921 | kfree(tags->static_rqs); |
1922 | tags->static_rqs = NULL; | |
320ae51f | 1923 | |
24d2f903 | 1924 | blk_mq_free_tags(tags); |
320ae51f JA |
1925 | } |
1926 | ||
cc71a6f4 JA |
1927 | struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, |
1928 | unsigned int hctx_idx, | |
1929 | unsigned int nr_tags, | |
1930 | unsigned int reserved_tags) | |
320ae51f | 1931 | { |
24d2f903 | 1932 | struct blk_mq_tags *tags; |
59f082e4 | 1933 | int node; |
320ae51f | 1934 | |
59f082e4 SL |
1935 | node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx); |
1936 | if (node == NUMA_NO_NODE) | |
1937 | node = set->numa_node; | |
1938 | ||
1939 | tags = blk_mq_init_tags(nr_tags, reserved_tags, node, | |
24391c0d | 1940 | BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags)); |
24d2f903 CH |
1941 | if (!tags) |
1942 | return NULL; | |
320ae51f | 1943 | |
cc71a6f4 | 1944 | tags->rqs = kzalloc_node(nr_tags * sizeof(struct request *), |
36e1f3d1 | 1945 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, |
59f082e4 | 1946 | node); |
24d2f903 CH |
1947 | if (!tags->rqs) { |
1948 | blk_mq_free_tags(tags); | |
1949 | return NULL; | |
1950 | } | |
320ae51f | 1951 | |
2af8cbe3 JA |
1952 | tags->static_rqs = kzalloc_node(nr_tags * sizeof(struct request *), |
1953 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, | |
59f082e4 | 1954 | node); |
2af8cbe3 JA |
1955 | if (!tags->static_rqs) { |
1956 | kfree(tags->rqs); | |
1957 | blk_mq_free_tags(tags); | |
1958 | return NULL; | |
1959 | } | |
1960 | ||
cc71a6f4 JA |
1961 | return tags; |
1962 | } | |
1963 | ||
1964 | static size_t order_to_size(unsigned int order) | |
1965 | { | |
1966 | return (size_t)PAGE_SIZE << order; | |
1967 | } | |
1968 | ||
1969 | int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags, | |
1970 | unsigned int hctx_idx, unsigned int depth) | |
1971 | { | |
1972 | unsigned int i, j, entries_per_page, max_order = 4; | |
1973 | size_t rq_size, left; | |
59f082e4 SL |
1974 | int node; |
1975 | ||
1976 | node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx); | |
1977 | if (node == NUMA_NO_NODE) | |
1978 | node = set->numa_node; | |
cc71a6f4 JA |
1979 | |
1980 | INIT_LIST_HEAD(&tags->page_list); | |
1981 | ||
320ae51f JA |
1982 | /* |
1983 | * rq_size is the size of the request plus driver payload, rounded | |
1984 | * to the cacheline size | |
1985 | */ | |
24d2f903 | 1986 | rq_size = round_up(sizeof(struct request) + set->cmd_size, |
320ae51f | 1987 | cache_line_size()); |
cc71a6f4 | 1988 | left = rq_size * depth; |
320ae51f | 1989 | |
cc71a6f4 | 1990 | for (i = 0; i < depth; ) { |
320ae51f JA |
1991 | int this_order = max_order; |
1992 | struct page *page; | |
1993 | int to_do; | |
1994 | void *p; | |
1995 | ||
b3a834b1 | 1996 | while (this_order && left < order_to_size(this_order - 1)) |
320ae51f JA |
1997 | this_order--; |
1998 | ||
1999 | do { | |
59f082e4 | 2000 | page = alloc_pages_node(node, |
36e1f3d1 | 2001 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO, |
a5164405 | 2002 | this_order); |
320ae51f JA |
2003 | if (page) |
2004 | break; | |
2005 | if (!this_order--) | |
2006 | break; | |
2007 | if (order_to_size(this_order) < rq_size) | |
2008 | break; | |
2009 | } while (1); | |
2010 | ||
2011 | if (!page) | |
24d2f903 | 2012 | goto fail; |
320ae51f JA |
2013 | |
2014 | page->private = this_order; | |
24d2f903 | 2015 | list_add_tail(&page->lru, &tags->page_list); |
320ae51f JA |
2016 | |
2017 | p = page_address(page); | |
f75782e4 CM |
2018 | /* |
2019 | * Allow kmemleak to scan these pages as they contain pointers | |
2020 | * to additional allocations like via ops->init_request(). | |
2021 | */ | |
36e1f3d1 | 2022 | kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO); |
320ae51f | 2023 | entries_per_page = order_to_size(this_order) / rq_size; |
cc71a6f4 | 2024 | to_do = min(entries_per_page, depth - i); |
320ae51f JA |
2025 | left -= to_do * rq_size; |
2026 | for (j = 0; j < to_do; j++) { | |
2af8cbe3 JA |
2027 | struct request *rq = p; |
2028 | ||
2029 | tags->static_rqs[i] = rq; | |
24d2f903 | 2030 | if (set->ops->init_request) { |
d6296d39 | 2031 | if (set->ops->init_request(set, rq, hctx_idx, |
59f082e4 | 2032 | node)) { |
2af8cbe3 | 2033 | tags->static_rqs[i] = NULL; |
24d2f903 | 2034 | goto fail; |
a5164405 | 2035 | } |
e9b267d9 CH |
2036 | } |
2037 | ||
320ae51f JA |
2038 | p += rq_size; |
2039 | i++; | |
2040 | } | |
2041 | } | |
cc71a6f4 | 2042 | return 0; |
320ae51f | 2043 | |
24d2f903 | 2044 | fail: |
cc71a6f4 JA |
2045 | blk_mq_free_rqs(set, tags, hctx_idx); |
2046 | return -ENOMEM; | |
320ae51f JA |
2047 | } |
2048 | ||
e57690fe JA |
2049 | /* |
2050 | * 'cpu' is going away. splice any existing rq_list entries from this | |
2051 | * software queue to the hw queue dispatch list, and ensure that it | |
2052 | * gets run. | |
2053 | */ | |
9467f859 | 2054 | static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node) |
484b4061 | 2055 | { |
9467f859 | 2056 | struct blk_mq_hw_ctx *hctx; |
484b4061 JA |
2057 | struct blk_mq_ctx *ctx; |
2058 | LIST_HEAD(tmp); | |
2059 | ||
9467f859 | 2060 | hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead); |
e57690fe | 2061 | ctx = __blk_mq_get_ctx(hctx->queue, cpu); |
484b4061 JA |
2062 | |
2063 | spin_lock(&ctx->lock); | |
2064 | if (!list_empty(&ctx->rq_list)) { | |
2065 | list_splice_init(&ctx->rq_list, &tmp); | |
2066 | blk_mq_hctx_clear_pending(hctx, ctx); | |
2067 | } | |
2068 | spin_unlock(&ctx->lock); | |
2069 | ||
2070 | if (list_empty(&tmp)) | |
9467f859 | 2071 | return 0; |
484b4061 | 2072 | |
e57690fe JA |
2073 | spin_lock(&hctx->lock); |
2074 | list_splice_tail_init(&tmp, &hctx->dispatch); | |
2075 | spin_unlock(&hctx->lock); | |
484b4061 JA |
2076 | |
2077 | blk_mq_run_hw_queue(hctx, true); | |
9467f859 | 2078 | return 0; |
484b4061 JA |
2079 | } |
2080 | ||
9467f859 | 2081 | static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx) |
484b4061 | 2082 | { |
9467f859 TG |
2083 | cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD, |
2084 | &hctx->cpuhp_dead); | |
484b4061 JA |
2085 | } |
2086 | ||
c3b4afca | 2087 | /* hctx->ctxs will be freed in queue's release handler */ |
08e98fc6 ML |
2088 | static void blk_mq_exit_hctx(struct request_queue *q, |
2089 | struct blk_mq_tag_set *set, | |
2090 | struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx) | |
2091 | { | |
9c1051aa OS |
2092 | blk_mq_debugfs_unregister_hctx(hctx); |
2093 | ||
25a2bfb9 ML |
2094 | if (blk_mq_hw_queue_mapped(hctx)) |
2095 | blk_mq_tag_idle(hctx); | |
08e98fc6 | 2096 | |
f70ced09 | 2097 | if (set->ops->exit_request) |
d6296d39 | 2098 | set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx); |
f70ced09 | 2099 | |
93252632 OS |
2100 | blk_mq_sched_exit_hctx(q, hctx, hctx_idx); |
2101 | ||
08e98fc6 ML |
2102 | if (set->ops->exit_hctx) |
2103 | set->ops->exit_hctx(hctx, hctx_idx); | |
2104 | ||
9467f859 | 2105 | blk_mq_remove_cpuhp(hctx); |
08e98fc6 ML |
2106 | } |
2107 | ||
624dbe47 ML |
2108 | static void blk_mq_exit_hw_queues(struct request_queue *q, |
2109 | struct blk_mq_tag_set *set, int nr_queue) | |
2110 | { | |
2111 | struct blk_mq_hw_ctx *hctx; | |
2112 | unsigned int i; | |
2113 | ||
2114 | queue_for_each_hw_ctx(q, hctx, i) { | |
2115 | if (i == nr_queue) | |
2116 | break; | |
08e98fc6 | 2117 | blk_mq_exit_hctx(q, set, hctx, i); |
624dbe47 | 2118 | } |
624dbe47 ML |
2119 | } |
2120 | ||
08e98fc6 ML |
2121 | static int blk_mq_init_hctx(struct request_queue *q, |
2122 | struct blk_mq_tag_set *set, | |
2123 | struct blk_mq_hw_ctx *hctx, unsigned hctx_idx) | |
320ae51f | 2124 | { |
08e98fc6 ML |
2125 | int node; |
2126 | ||
2127 | node = hctx->numa_node; | |
2128 | if (node == NUMA_NO_NODE) | |
2129 | node = hctx->numa_node = set->numa_node; | |
2130 | ||
9f993737 | 2131 | INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn); |
08e98fc6 ML |
2132 | spin_lock_init(&hctx->lock); |
2133 | INIT_LIST_HEAD(&hctx->dispatch); | |
2134 | hctx->queue = q; | |
2404e607 | 2135 | hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED; |
08e98fc6 | 2136 | |
9467f859 | 2137 | cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead); |
08e98fc6 ML |
2138 | |
2139 | hctx->tags = set->tags[hctx_idx]; | |
320ae51f JA |
2140 | |
2141 | /* | |
08e98fc6 ML |
2142 | * Allocate space for all possible cpus to avoid allocation at |
2143 | * runtime | |
320ae51f | 2144 | */ |
d904bfa7 | 2145 | hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *), |
3bcd45bf | 2146 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, node); |
08e98fc6 ML |
2147 | if (!hctx->ctxs) |
2148 | goto unregister_cpu_notifier; | |
320ae51f | 2149 | |
3bcd45bf JW |
2150 | if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8), |
2151 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, node)) | |
08e98fc6 | 2152 | goto free_ctxs; |
320ae51f | 2153 | |
08e98fc6 | 2154 | hctx->nr_ctx = 0; |
320ae51f | 2155 | |
eb619fdb JA |
2156 | init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake); |
2157 | INIT_LIST_HEAD(&hctx->dispatch_wait.entry); | |
2158 | ||
08e98fc6 ML |
2159 | if (set->ops->init_hctx && |
2160 | set->ops->init_hctx(hctx, set->driver_data, hctx_idx)) | |
2161 | goto free_bitmap; | |
320ae51f | 2162 | |
93252632 OS |
2163 | if (blk_mq_sched_init_hctx(q, hctx, hctx_idx)) |
2164 | goto exit_hctx; | |
2165 | ||
3bcd45bf JW |
2166 | hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size, |
2167 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY); | |
f70ced09 | 2168 | if (!hctx->fq) |
93252632 | 2169 | goto sched_exit_hctx; |
320ae51f | 2170 | |
f70ced09 | 2171 | if (set->ops->init_request && |
d6296d39 CH |
2172 | set->ops->init_request(set, hctx->fq->flush_rq, hctx_idx, |
2173 | node)) | |
f70ced09 | 2174 | goto free_fq; |
320ae51f | 2175 | |
6a83e74d | 2176 | if (hctx->flags & BLK_MQ_F_BLOCKING) |
07319678 | 2177 | init_srcu_struct(hctx->queue_rq_srcu); |
6a83e74d | 2178 | |
9c1051aa OS |
2179 | blk_mq_debugfs_register_hctx(q, hctx); |
2180 | ||
08e98fc6 | 2181 | return 0; |
320ae51f | 2182 | |
f70ced09 | 2183 | free_fq: |
aa82ef2f | 2184 | blk_free_flush_queue(hctx->fq); |
93252632 OS |
2185 | sched_exit_hctx: |
2186 | blk_mq_sched_exit_hctx(q, hctx, hctx_idx); | |
f70ced09 ML |
2187 | exit_hctx: |
2188 | if (set->ops->exit_hctx) | |
2189 | set->ops->exit_hctx(hctx, hctx_idx); | |
08e98fc6 | 2190 | free_bitmap: |
88459642 | 2191 | sbitmap_free(&hctx->ctx_map); |
08e98fc6 ML |
2192 | free_ctxs: |
2193 | kfree(hctx->ctxs); | |
2194 | unregister_cpu_notifier: | |
9467f859 | 2195 | blk_mq_remove_cpuhp(hctx); |
08e98fc6 ML |
2196 | return -1; |
2197 | } | |
320ae51f | 2198 | |
320ae51f JA |
2199 | static void blk_mq_init_cpu_queues(struct request_queue *q, |
2200 | unsigned int nr_hw_queues) | |
2201 | { | |
2202 | unsigned int i; | |
2203 | ||
2204 | for_each_possible_cpu(i) { | |
2205 | struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i); | |
2206 | struct blk_mq_hw_ctx *hctx; | |
2207 | ||
320ae51f JA |
2208 | __ctx->cpu = i; |
2209 | spin_lock_init(&__ctx->lock); | |
2210 | INIT_LIST_HEAD(&__ctx->rq_list); | |
2211 | __ctx->queue = q; | |
2212 | ||
320ae51f JA |
2213 | /* |
2214 | * Set local node, IFF we have more than one hw queue. If | |
2215 | * not, we remain on the home node of the device | |
2216 | */ | |
8a591698 | 2217 | hctx = blk_mq_map_queue(q, i); |
320ae51f | 2218 | if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE) |
bffed457 | 2219 | hctx->numa_node = local_memory_node(cpu_to_node(i)); |
320ae51f JA |
2220 | } |
2221 | } | |
2222 | ||
cc71a6f4 JA |
2223 | static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, int hctx_idx) |
2224 | { | |
2225 | int ret = 0; | |
2226 | ||
2227 | set->tags[hctx_idx] = blk_mq_alloc_rq_map(set, hctx_idx, | |
2228 | set->queue_depth, set->reserved_tags); | |
2229 | if (!set->tags[hctx_idx]) | |
2230 | return false; | |
2231 | ||
2232 | ret = blk_mq_alloc_rqs(set, set->tags[hctx_idx], hctx_idx, | |
2233 | set->queue_depth); | |
2234 | if (!ret) | |
2235 | return true; | |
2236 | ||
2237 | blk_mq_free_rq_map(set->tags[hctx_idx]); | |
2238 | set->tags[hctx_idx] = NULL; | |
2239 | return false; | |
2240 | } | |
2241 | ||
2242 | static void blk_mq_free_map_and_requests(struct blk_mq_tag_set *set, | |
2243 | unsigned int hctx_idx) | |
2244 | { | |
bd166ef1 JA |
2245 | if (set->tags[hctx_idx]) { |
2246 | blk_mq_free_rqs(set, set->tags[hctx_idx], hctx_idx); | |
2247 | blk_mq_free_rq_map(set->tags[hctx_idx]); | |
2248 | set->tags[hctx_idx] = NULL; | |
2249 | } | |
cc71a6f4 JA |
2250 | } |
2251 | ||
4b855ad3 | 2252 | static void blk_mq_map_swqueue(struct request_queue *q) |
320ae51f | 2253 | { |
d1b1cea1 | 2254 | unsigned int i, hctx_idx; |
320ae51f JA |
2255 | struct blk_mq_hw_ctx *hctx; |
2256 | struct blk_mq_ctx *ctx; | |
2a34c087 | 2257 | struct blk_mq_tag_set *set = q->tag_set; |
320ae51f | 2258 | |
60de074b AM |
2259 | /* |
2260 | * Avoid others reading imcomplete hctx->cpumask through sysfs | |
2261 | */ | |
2262 | mutex_lock(&q->sysfs_lock); | |
2263 | ||
320ae51f | 2264 | queue_for_each_hw_ctx(q, hctx, i) { |
e4043dcf | 2265 | cpumask_clear(hctx->cpumask); |
320ae51f JA |
2266 | hctx->nr_ctx = 0; |
2267 | } | |
2268 | ||
2269 | /* | |
4b855ad3 CH |
2270 | * Map software to hardware queues. |
2271 | * | |
2272 | * If the cpu isn't present, the cpu is mapped to first hctx. | |
320ae51f | 2273 | */ |
8a591698 | 2274 | for_each_possible_cpu(i) { |
d1b1cea1 GKB |
2275 | hctx_idx = q->mq_map[i]; |
2276 | /* unmapped hw queue can be remapped after CPU topo changed */ | |
cc71a6f4 JA |
2277 | if (!set->tags[hctx_idx] && |
2278 | !__blk_mq_alloc_rq_map(set, hctx_idx)) { | |
d1b1cea1 GKB |
2279 | /* |
2280 | * If tags initialization fail for some hctx, | |
2281 | * that hctx won't be brought online. In this | |
2282 | * case, remap the current ctx to hctx[0] which | |
2283 | * is guaranteed to always have tags allocated | |
2284 | */ | |
cc71a6f4 | 2285 | q->mq_map[i] = 0; |
d1b1cea1 GKB |
2286 | } |
2287 | ||
897bb0c7 | 2288 | ctx = per_cpu_ptr(q->queue_ctx, i); |
7d7e0f90 | 2289 | hctx = blk_mq_map_queue(q, i); |
868f2f0b | 2290 | |
e4043dcf | 2291 | cpumask_set_cpu(i, hctx->cpumask); |
320ae51f JA |
2292 | ctx->index_hw = hctx->nr_ctx; |
2293 | hctx->ctxs[hctx->nr_ctx++] = ctx; | |
2294 | } | |
506e931f | 2295 | |
60de074b AM |
2296 | mutex_unlock(&q->sysfs_lock); |
2297 | ||
506e931f | 2298 | queue_for_each_hw_ctx(q, hctx, i) { |
484b4061 | 2299 | /* |
a68aafa5 JA |
2300 | * If no software queues are mapped to this hardware queue, |
2301 | * disable it and free the request entries. | |
484b4061 JA |
2302 | */ |
2303 | if (!hctx->nr_ctx) { | |
d1b1cea1 GKB |
2304 | /* Never unmap queue 0. We need it as a |
2305 | * fallback in case of a new remap fails | |
2306 | * allocation | |
2307 | */ | |
cc71a6f4 JA |
2308 | if (i && set->tags[i]) |
2309 | blk_mq_free_map_and_requests(set, i); | |
2310 | ||
2a34c087 | 2311 | hctx->tags = NULL; |
484b4061 JA |
2312 | continue; |
2313 | } | |
2314 | ||
2a34c087 ML |
2315 | hctx->tags = set->tags[i]; |
2316 | WARN_ON(!hctx->tags); | |
2317 | ||
889fa31f CY |
2318 | /* |
2319 | * Set the map size to the number of mapped software queues. | |
2320 | * This is more accurate and more efficient than looping | |
2321 | * over all possibly mapped software queues. | |
2322 | */ | |
88459642 | 2323 | sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx); |
889fa31f | 2324 | |
484b4061 JA |
2325 | /* |
2326 | * Initialize batch roundrobin counts | |
2327 | */ | |
6204c099 | 2328 | hctx->next_cpu = blk_mq_first_mapped_cpu(hctx); |
506e931f JA |
2329 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; |
2330 | } | |
320ae51f JA |
2331 | } |
2332 | ||
8e8320c9 JA |
2333 | /* |
2334 | * Caller needs to ensure that we're either frozen/quiesced, or that | |
2335 | * the queue isn't live yet. | |
2336 | */ | |
2404e607 | 2337 | static void queue_set_hctx_shared(struct request_queue *q, bool shared) |
0d2602ca JA |
2338 | { |
2339 | struct blk_mq_hw_ctx *hctx; | |
0d2602ca JA |
2340 | int i; |
2341 | ||
2404e607 | 2342 | queue_for_each_hw_ctx(q, hctx, i) { |
8e8320c9 JA |
2343 | if (shared) { |
2344 | if (test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state)) | |
2345 | atomic_inc(&q->shared_hctx_restart); | |
2404e607 | 2346 | hctx->flags |= BLK_MQ_F_TAG_SHARED; |
8e8320c9 JA |
2347 | } else { |
2348 | if (test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state)) | |
2349 | atomic_dec(&q->shared_hctx_restart); | |
2404e607 | 2350 | hctx->flags &= ~BLK_MQ_F_TAG_SHARED; |
8e8320c9 | 2351 | } |
2404e607 JM |
2352 | } |
2353 | } | |
2354 | ||
8e8320c9 JA |
2355 | static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set, |
2356 | bool shared) | |
2404e607 JM |
2357 | { |
2358 | struct request_queue *q; | |
0d2602ca | 2359 | |
705cda97 BVA |
2360 | lockdep_assert_held(&set->tag_list_lock); |
2361 | ||
0d2602ca JA |
2362 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
2363 | blk_mq_freeze_queue(q); | |
2404e607 | 2364 | queue_set_hctx_shared(q, shared); |
0d2602ca JA |
2365 | blk_mq_unfreeze_queue(q); |
2366 | } | |
2367 | } | |
2368 | ||
2369 | static void blk_mq_del_queue_tag_set(struct request_queue *q) | |
2370 | { | |
2371 | struct blk_mq_tag_set *set = q->tag_set; | |
2372 | ||
0d2602ca | 2373 | mutex_lock(&set->tag_list_lock); |
705cda97 | 2374 | list_del_rcu(&q->tag_set_list); |
2404e607 JM |
2375 | if (list_is_singular(&set->tag_list)) { |
2376 | /* just transitioned to unshared */ | |
2377 | set->flags &= ~BLK_MQ_F_TAG_SHARED; | |
2378 | /* update existing queue */ | |
2379 | blk_mq_update_tag_set_depth(set, false); | |
2380 | } | |
0d2602ca | 2381 | mutex_unlock(&set->tag_list_lock); |
705cda97 | 2382 | synchronize_rcu(); |
7c68628e | 2383 | INIT_LIST_HEAD(&q->tag_set_list); |
0d2602ca JA |
2384 | } |
2385 | ||
2386 | static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set, | |
2387 | struct request_queue *q) | |
2388 | { | |
2389 | q->tag_set = set; | |
2390 | ||
2391 | mutex_lock(&set->tag_list_lock); | |
2404e607 | 2392 | |
ff821d27 JA |
2393 | /* |
2394 | * Check to see if we're transitioning to shared (from 1 to 2 queues). | |
2395 | */ | |
2396 | if (!list_empty(&set->tag_list) && | |
2397 | !(set->flags & BLK_MQ_F_TAG_SHARED)) { | |
2404e607 JM |
2398 | set->flags |= BLK_MQ_F_TAG_SHARED; |
2399 | /* update existing queue */ | |
2400 | blk_mq_update_tag_set_depth(set, true); | |
2401 | } | |
2402 | if (set->flags & BLK_MQ_F_TAG_SHARED) | |
2403 | queue_set_hctx_shared(q, true); | |
705cda97 | 2404 | list_add_tail_rcu(&q->tag_set_list, &set->tag_list); |
2404e607 | 2405 | |
0d2602ca JA |
2406 | mutex_unlock(&set->tag_list_lock); |
2407 | } | |
2408 | ||
e09aae7e ML |
2409 | /* |
2410 | * It is the actual release handler for mq, but we do it from | |
2411 | * request queue's release handler for avoiding use-after-free | |
2412 | * and headache because q->mq_kobj shouldn't have been introduced, | |
2413 | * but we can't group ctx/kctx kobj without it. | |
2414 | */ | |
2415 | void blk_mq_release(struct request_queue *q) | |
2416 | { | |
2417 | struct blk_mq_hw_ctx *hctx; | |
2418 | unsigned int i; | |
2419 | ||
2420 | /* hctx kobj stays in hctx */ | |
c3b4afca ML |
2421 | queue_for_each_hw_ctx(q, hctx, i) { |
2422 | if (!hctx) | |
2423 | continue; | |
6c8b232e | 2424 | kobject_put(&hctx->kobj); |
c3b4afca | 2425 | } |
e09aae7e | 2426 | |
a723bab3 AM |
2427 | q->mq_map = NULL; |
2428 | ||
e09aae7e ML |
2429 | kfree(q->queue_hw_ctx); |
2430 | ||
7ea5fe31 ML |
2431 | /* |
2432 | * release .mq_kobj and sw queue's kobject now because | |
2433 | * both share lifetime with request queue. | |
2434 | */ | |
2435 | blk_mq_sysfs_deinit(q); | |
2436 | ||
e09aae7e ML |
2437 | free_percpu(q->queue_ctx); |
2438 | } | |
2439 | ||
24d2f903 | 2440 | struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set) |
b62c21b7 MS |
2441 | { |
2442 | struct request_queue *uninit_q, *q; | |
2443 | ||
2444 | uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node); | |
2445 | if (!uninit_q) | |
2446 | return ERR_PTR(-ENOMEM); | |
2447 | ||
2448 | q = blk_mq_init_allocated_queue(set, uninit_q); | |
2449 | if (IS_ERR(q)) | |
2450 | blk_cleanup_queue(uninit_q); | |
2451 | ||
2452 | return q; | |
2453 | } | |
2454 | EXPORT_SYMBOL(blk_mq_init_queue); | |
2455 | ||
07319678 BVA |
2456 | static int blk_mq_hw_ctx_size(struct blk_mq_tag_set *tag_set) |
2457 | { | |
2458 | int hw_ctx_size = sizeof(struct blk_mq_hw_ctx); | |
2459 | ||
2460 | BUILD_BUG_ON(ALIGN(offsetof(struct blk_mq_hw_ctx, queue_rq_srcu), | |
2461 | __alignof__(struct blk_mq_hw_ctx)) != | |
2462 | sizeof(struct blk_mq_hw_ctx)); | |
2463 | ||
2464 | if (tag_set->flags & BLK_MQ_F_BLOCKING) | |
2465 | hw_ctx_size += sizeof(struct srcu_struct); | |
2466 | ||
2467 | return hw_ctx_size; | |
2468 | } | |
2469 | ||
868f2f0b KB |
2470 | static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set, |
2471 | struct request_queue *q) | |
320ae51f | 2472 | { |
868f2f0b KB |
2473 | int i, j; |
2474 | struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx; | |
f14bbe77 | 2475 | |
868f2f0b | 2476 | blk_mq_sysfs_unregister(q); |
a7e54b6a ML |
2477 | |
2478 | /* protect against switching io scheduler */ | |
2479 | mutex_lock(&q->sysfs_lock); | |
24d2f903 | 2480 | for (i = 0; i < set->nr_hw_queues; i++) { |
868f2f0b | 2481 | int node; |
f14bbe77 | 2482 | |
868f2f0b KB |
2483 | if (hctxs[i]) |
2484 | continue; | |
2485 | ||
2486 | node = blk_mq_hw_queue_to_node(q->mq_map, i); | |
07319678 | 2487 | hctxs[i] = kzalloc_node(blk_mq_hw_ctx_size(set), |
3bcd45bf JW |
2488 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, |
2489 | node); | |
320ae51f | 2490 | if (!hctxs[i]) |
868f2f0b | 2491 | break; |
320ae51f | 2492 | |
3bcd45bf JW |
2493 | if (!zalloc_cpumask_var_node(&hctxs[i]->cpumask, |
2494 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, | |
2495 | node)) { | |
868f2f0b KB |
2496 | kfree(hctxs[i]); |
2497 | hctxs[i] = NULL; | |
2498 | break; | |
2499 | } | |
e4043dcf | 2500 | |
0d2602ca | 2501 | atomic_set(&hctxs[i]->nr_active, 0); |
f14bbe77 | 2502 | hctxs[i]->numa_node = node; |
320ae51f | 2503 | hctxs[i]->queue_num = i; |
868f2f0b KB |
2504 | |
2505 | if (blk_mq_init_hctx(q, set, hctxs[i], i)) { | |
2506 | free_cpumask_var(hctxs[i]->cpumask); | |
2507 | kfree(hctxs[i]); | |
2508 | hctxs[i] = NULL; | |
2509 | break; | |
2510 | } | |
2511 | blk_mq_hctx_kobj_init(hctxs[i]); | |
320ae51f | 2512 | } |
868f2f0b KB |
2513 | for (j = i; j < q->nr_hw_queues; j++) { |
2514 | struct blk_mq_hw_ctx *hctx = hctxs[j]; | |
2515 | ||
2516 | if (hctx) { | |
cc71a6f4 JA |
2517 | if (hctx->tags) |
2518 | blk_mq_free_map_and_requests(set, j); | |
868f2f0b | 2519 | blk_mq_exit_hctx(q, set, hctx, j); |
868f2f0b | 2520 | kobject_put(&hctx->kobj); |
868f2f0b KB |
2521 | hctxs[j] = NULL; |
2522 | ||
2523 | } | |
2524 | } | |
2525 | q->nr_hw_queues = i; | |
a7e54b6a | 2526 | mutex_unlock(&q->sysfs_lock); |
868f2f0b KB |
2527 | blk_mq_sysfs_register(q); |
2528 | } | |
2529 | ||
2530 | struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set, | |
2531 | struct request_queue *q) | |
2532 | { | |
66841672 ML |
2533 | /* mark the queue as mq asap */ |
2534 | q->mq_ops = set->ops; | |
2535 | ||
34dbad5d | 2536 | q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn, |
720b8ccc SB |
2537 | blk_mq_poll_stats_bkt, |
2538 | BLK_MQ_POLL_STATS_BKTS, q); | |
34dbad5d OS |
2539 | if (!q->poll_cb) |
2540 | goto err_exit; | |
2541 | ||
868f2f0b KB |
2542 | q->queue_ctx = alloc_percpu(struct blk_mq_ctx); |
2543 | if (!q->queue_ctx) | |
c7de5726 | 2544 | goto err_exit; |
868f2f0b | 2545 | |
737f98cf ML |
2546 | /* init q->mq_kobj and sw queues' kobjects */ |
2547 | blk_mq_sysfs_init(q); | |
2548 | ||
868f2f0b KB |
2549 | q->queue_hw_ctx = kzalloc_node(nr_cpu_ids * sizeof(*(q->queue_hw_ctx)), |
2550 | GFP_KERNEL, set->numa_node); | |
2551 | if (!q->queue_hw_ctx) | |
2552 | goto err_percpu; | |
2553 | ||
bdd17e75 | 2554 | q->mq_map = set->mq_map; |
868f2f0b KB |
2555 | |
2556 | blk_mq_realloc_hw_ctxs(set, q); | |
2557 | if (!q->nr_hw_queues) | |
2558 | goto err_hctxs; | |
320ae51f | 2559 | |
287922eb | 2560 | INIT_WORK(&q->timeout_work, blk_mq_timeout_work); |
e56f698b | 2561 | blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ); |
320ae51f JA |
2562 | |
2563 | q->nr_queues = nr_cpu_ids; | |
320ae51f | 2564 | |
94eddfbe | 2565 | q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT; |
320ae51f | 2566 | |
05f1dd53 JA |
2567 | if (!(set->flags & BLK_MQ_F_SG_MERGE)) |
2568 | q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE; | |
2569 | ||
1be036e9 CH |
2570 | q->sg_reserved_size = INT_MAX; |
2571 | ||
2849450a | 2572 | INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work); |
6fca6a61 CH |
2573 | INIT_LIST_HEAD(&q->requeue_list); |
2574 | spin_lock_init(&q->requeue_lock); | |
2575 | ||
254d259d | 2576 | blk_queue_make_request(q, blk_mq_make_request); |
ea435e1b CH |
2577 | if (q->mq_ops->poll) |
2578 | q->poll_fn = blk_mq_poll; | |
07068d5b | 2579 | |
eba71768 JA |
2580 | /* |
2581 | * Do this after blk_queue_make_request() overrides it... | |
2582 | */ | |
2583 | q->nr_requests = set->queue_depth; | |
2584 | ||
64f1c21e JA |
2585 | /* |
2586 | * Default to classic polling | |
2587 | */ | |
2588 | q->poll_nsec = -1; | |
2589 | ||
24d2f903 CH |
2590 | if (set->ops->complete) |
2591 | blk_queue_softirq_done(q, set->ops->complete); | |
30a91cb4 | 2592 | |
24d2f903 | 2593 | blk_mq_init_cpu_queues(q, set->nr_hw_queues); |
0d2602ca | 2594 | blk_mq_add_queue_tag_set(set, q); |
4b855ad3 | 2595 | blk_mq_map_swqueue(q); |
4593fdbe | 2596 | |
d3484991 JA |
2597 | if (!(set->flags & BLK_MQ_F_NO_SCHED)) { |
2598 | int ret; | |
2599 | ||
2600 | ret = blk_mq_sched_init(q); | |
2601 | if (ret) | |
2602 | return ERR_PTR(ret); | |
2603 | } | |
2604 | ||
320ae51f | 2605 | return q; |
18741986 | 2606 | |
320ae51f | 2607 | err_hctxs: |
868f2f0b | 2608 | kfree(q->queue_hw_ctx); |
320ae51f | 2609 | err_percpu: |
868f2f0b | 2610 | free_percpu(q->queue_ctx); |
c7de5726 ML |
2611 | err_exit: |
2612 | q->mq_ops = NULL; | |
320ae51f JA |
2613 | return ERR_PTR(-ENOMEM); |
2614 | } | |
b62c21b7 | 2615 | EXPORT_SYMBOL(blk_mq_init_allocated_queue); |
320ae51f | 2616 | |
84c8ff7f ML |
2617 | /* tags can _not_ be used after returning from blk_mq_exit_queue */ |
2618 | void blk_mq_exit_queue(struct request_queue *q) | |
320ae51f | 2619 | { |
624dbe47 | 2620 | struct blk_mq_tag_set *set = q->tag_set; |
320ae51f | 2621 | |
0d2602ca | 2622 | blk_mq_del_queue_tag_set(q); |
624dbe47 | 2623 | blk_mq_exit_hw_queues(q, set, set->nr_hw_queues); |
320ae51f | 2624 | } |
320ae51f JA |
2625 | |
2626 | /* Basically redo blk_mq_init_queue with queue frozen */ | |
4b855ad3 | 2627 | static void blk_mq_queue_reinit(struct request_queue *q) |
320ae51f | 2628 | { |
4ecd4fef | 2629 | WARN_ON_ONCE(!atomic_read(&q->mq_freeze_depth)); |
320ae51f | 2630 | |
9c1051aa | 2631 | blk_mq_debugfs_unregister_hctxs(q); |
67aec14c JA |
2632 | blk_mq_sysfs_unregister(q); |
2633 | ||
320ae51f JA |
2634 | /* |
2635 | * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe | |
ff821d27 JA |
2636 | * we should change hctx numa_node according to the new topology (this |
2637 | * involves freeing and re-allocating memory, worth doing?) | |
320ae51f | 2638 | */ |
4b855ad3 | 2639 | blk_mq_map_swqueue(q); |
320ae51f | 2640 | |
67aec14c | 2641 | blk_mq_sysfs_register(q); |
9c1051aa | 2642 | blk_mq_debugfs_register_hctxs(q); |
320ae51f JA |
2643 | } |
2644 | ||
a5164405 JA |
2645 | static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set) |
2646 | { | |
2647 | int i; | |
2648 | ||
cc71a6f4 JA |
2649 | for (i = 0; i < set->nr_hw_queues; i++) |
2650 | if (!__blk_mq_alloc_rq_map(set, i)) | |
a5164405 | 2651 | goto out_unwind; |
a5164405 JA |
2652 | |
2653 | return 0; | |
2654 | ||
2655 | out_unwind: | |
2656 | while (--i >= 0) | |
cc71a6f4 | 2657 | blk_mq_free_rq_map(set->tags[i]); |
a5164405 | 2658 | |
a5164405 JA |
2659 | return -ENOMEM; |
2660 | } | |
2661 | ||
2662 | /* | |
2663 | * Allocate the request maps associated with this tag_set. Note that this | |
2664 | * may reduce the depth asked for, if memory is tight. set->queue_depth | |
2665 | * will be updated to reflect the allocated depth. | |
2666 | */ | |
2667 | static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set) | |
2668 | { | |
2669 | unsigned int depth; | |
2670 | int err; | |
2671 | ||
2672 | depth = set->queue_depth; | |
2673 | do { | |
2674 | err = __blk_mq_alloc_rq_maps(set); | |
2675 | if (!err) | |
2676 | break; | |
2677 | ||
2678 | set->queue_depth >>= 1; | |
2679 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) { | |
2680 | err = -ENOMEM; | |
2681 | break; | |
2682 | } | |
2683 | } while (set->queue_depth); | |
2684 | ||
2685 | if (!set->queue_depth || err) { | |
2686 | pr_err("blk-mq: failed to allocate request map\n"); | |
2687 | return -ENOMEM; | |
2688 | } | |
2689 | ||
2690 | if (depth != set->queue_depth) | |
2691 | pr_info("blk-mq: reduced tag depth (%u -> %u)\n", | |
2692 | depth, set->queue_depth); | |
2693 | ||
2694 | return 0; | |
2695 | } | |
2696 | ||
ebe8bddb OS |
2697 | static int blk_mq_update_queue_map(struct blk_mq_tag_set *set) |
2698 | { | |
6a684e3d ML |
2699 | if (set->ops->map_queues) { |
2700 | int cpu; | |
2701 | /* | |
2702 | * transport .map_queues is usually done in the following | |
2703 | * way: | |
2704 | * | |
2705 | * for (queue = 0; queue < set->nr_hw_queues; queue++) { | |
2706 | * mask = get_cpu_mask(queue) | |
2707 | * for_each_cpu(cpu, mask) | |
2708 | * set->mq_map[cpu] = queue; | |
2709 | * } | |
2710 | * | |
2711 | * When we need to remap, the table has to be cleared for | |
2712 | * killing stale mapping since one CPU may not be mapped | |
2713 | * to any hw queue. | |
2714 | */ | |
2715 | for_each_possible_cpu(cpu) | |
2716 | set->mq_map[cpu] = 0; | |
2717 | ||
ebe8bddb | 2718 | return set->ops->map_queues(set); |
6a684e3d | 2719 | } else |
ebe8bddb OS |
2720 | return blk_mq_map_queues(set); |
2721 | } | |
2722 | ||
a4391c64 JA |
2723 | /* |
2724 | * Alloc a tag set to be associated with one or more request queues. | |
2725 | * May fail with EINVAL for various error conditions. May adjust the | |
2726 | * requested depth down, if if it too large. In that case, the set | |
2727 | * value will be stored in set->queue_depth. | |
2728 | */ | |
24d2f903 CH |
2729 | int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set) |
2730 | { | |
da695ba2 CH |
2731 | int ret; |
2732 | ||
205fb5f5 BVA |
2733 | BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS); |
2734 | ||
24d2f903 CH |
2735 | if (!set->nr_hw_queues) |
2736 | return -EINVAL; | |
a4391c64 | 2737 | if (!set->queue_depth) |
24d2f903 CH |
2738 | return -EINVAL; |
2739 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) | |
2740 | return -EINVAL; | |
2741 | ||
7d7e0f90 | 2742 | if (!set->ops->queue_rq) |
24d2f903 CH |
2743 | return -EINVAL; |
2744 | ||
de148297 ML |
2745 | if (!set->ops->get_budget ^ !set->ops->put_budget) |
2746 | return -EINVAL; | |
2747 | ||
a4391c64 JA |
2748 | if (set->queue_depth > BLK_MQ_MAX_DEPTH) { |
2749 | pr_info("blk-mq: reduced tag depth to %u\n", | |
2750 | BLK_MQ_MAX_DEPTH); | |
2751 | set->queue_depth = BLK_MQ_MAX_DEPTH; | |
2752 | } | |
24d2f903 | 2753 | |
6637fadf SL |
2754 | /* |
2755 | * If a crashdump is active, then we are potentially in a very | |
2756 | * memory constrained environment. Limit us to 1 queue and | |
2757 | * 64 tags to prevent using too much memory. | |
2758 | */ | |
2759 | if (is_kdump_kernel()) { | |
2760 | set->nr_hw_queues = 1; | |
2761 | set->queue_depth = min(64U, set->queue_depth); | |
2762 | } | |
868f2f0b KB |
2763 | /* |
2764 | * There is no use for more h/w queues than cpus. | |
2765 | */ | |
2766 | if (set->nr_hw_queues > nr_cpu_ids) | |
2767 | set->nr_hw_queues = nr_cpu_ids; | |
6637fadf | 2768 | |
868f2f0b | 2769 | set->tags = kzalloc_node(nr_cpu_ids * sizeof(struct blk_mq_tags *), |
24d2f903 CH |
2770 | GFP_KERNEL, set->numa_node); |
2771 | if (!set->tags) | |
a5164405 | 2772 | return -ENOMEM; |
24d2f903 | 2773 | |
da695ba2 CH |
2774 | ret = -ENOMEM; |
2775 | set->mq_map = kzalloc_node(sizeof(*set->mq_map) * nr_cpu_ids, | |
2776 | GFP_KERNEL, set->numa_node); | |
bdd17e75 CH |
2777 | if (!set->mq_map) |
2778 | goto out_free_tags; | |
2779 | ||
ebe8bddb | 2780 | ret = blk_mq_update_queue_map(set); |
da695ba2 CH |
2781 | if (ret) |
2782 | goto out_free_mq_map; | |
2783 | ||
2784 | ret = blk_mq_alloc_rq_maps(set); | |
2785 | if (ret) | |
bdd17e75 | 2786 | goto out_free_mq_map; |
24d2f903 | 2787 | |
0d2602ca JA |
2788 | mutex_init(&set->tag_list_lock); |
2789 | INIT_LIST_HEAD(&set->tag_list); | |
2790 | ||
24d2f903 | 2791 | return 0; |
bdd17e75 CH |
2792 | |
2793 | out_free_mq_map: | |
2794 | kfree(set->mq_map); | |
2795 | set->mq_map = NULL; | |
2796 | out_free_tags: | |
5676e7b6 RE |
2797 | kfree(set->tags); |
2798 | set->tags = NULL; | |
da695ba2 | 2799 | return ret; |
24d2f903 CH |
2800 | } |
2801 | EXPORT_SYMBOL(blk_mq_alloc_tag_set); | |
2802 | ||
2803 | void blk_mq_free_tag_set(struct blk_mq_tag_set *set) | |
2804 | { | |
2805 | int i; | |
2806 | ||
cc71a6f4 JA |
2807 | for (i = 0; i < nr_cpu_ids; i++) |
2808 | blk_mq_free_map_and_requests(set, i); | |
484b4061 | 2809 | |
bdd17e75 CH |
2810 | kfree(set->mq_map); |
2811 | set->mq_map = NULL; | |
2812 | ||
981bd189 | 2813 | kfree(set->tags); |
5676e7b6 | 2814 | set->tags = NULL; |
24d2f903 CH |
2815 | } |
2816 | EXPORT_SYMBOL(blk_mq_free_tag_set); | |
2817 | ||
e3a2b3f9 JA |
2818 | int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr) |
2819 | { | |
2820 | struct blk_mq_tag_set *set = q->tag_set; | |
2821 | struct blk_mq_hw_ctx *hctx; | |
2822 | int i, ret; | |
2823 | ||
bd166ef1 | 2824 | if (!set) |
e3a2b3f9 JA |
2825 | return -EINVAL; |
2826 | ||
ac99bae9 AZ |
2827 | if (q->nr_requests == nr) |
2828 | return 0; | |
2829 | ||
70f36b60 | 2830 | blk_mq_freeze_queue(q); |
d2265003 | 2831 | blk_mq_quiesce_queue(q); |
70f36b60 | 2832 | |
e3a2b3f9 JA |
2833 | ret = 0; |
2834 | queue_for_each_hw_ctx(q, hctx, i) { | |
e9137d4b KB |
2835 | if (!hctx->tags) |
2836 | continue; | |
bd166ef1 JA |
2837 | /* |
2838 | * If we're using an MQ scheduler, just update the scheduler | |
2839 | * queue depth. This is similar to what the old code would do. | |
2840 | */ | |
70f36b60 | 2841 | if (!hctx->sched_tags) { |
c2e82a23 | 2842 | ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr, |
70f36b60 JA |
2843 | false); |
2844 | } else { | |
2845 | ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags, | |
2846 | nr, true); | |
2847 | } | |
e3a2b3f9 JA |
2848 | if (ret) |
2849 | break; | |
2850 | } | |
2851 | ||
2852 | if (!ret) | |
2853 | q->nr_requests = nr; | |
2854 | ||
d2265003 | 2855 | blk_mq_unquiesce_queue(q); |
70f36b60 | 2856 | blk_mq_unfreeze_queue(q); |
70f36b60 | 2857 | |
e3a2b3f9 JA |
2858 | return ret; |
2859 | } | |
2860 | ||
e4dc2b32 KB |
2861 | static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, |
2862 | int nr_hw_queues) | |
868f2f0b KB |
2863 | { |
2864 | struct request_queue *q; | |
2865 | ||
705cda97 BVA |
2866 | lockdep_assert_held(&set->tag_list_lock); |
2867 | ||
868f2f0b KB |
2868 | if (nr_hw_queues > nr_cpu_ids) |
2869 | nr_hw_queues = nr_cpu_ids; | |
2870 | if (nr_hw_queues < 1 || nr_hw_queues == set->nr_hw_queues) | |
2871 | return; | |
2872 | ||
2873 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
2874 | blk_mq_freeze_queue(q); | |
2875 | ||
2876 | set->nr_hw_queues = nr_hw_queues; | |
ebe8bddb | 2877 | blk_mq_update_queue_map(set); |
868f2f0b KB |
2878 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
2879 | blk_mq_realloc_hw_ctxs(set, q); | |
4b855ad3 | 2880 | blk_mq_queue_reinit(q); |
868f2f0b KB |
2881 | } |
2882 | ||
2883 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
2884 | blk_mq_unfreeze_queue(q); | |
2885 | } | |
e4dc2b32 KB |
2886 | |
2887 | void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues) | |
2888 | { | |
2889 | mutex_lock(&set->tag_list_lock); | |
2890 | __blk_mq_update_nr_hw_queues(set, nr_hw_queues); | |
2891 | mutex_unlock(&set->tag_list_lock); | |
2892 | } | |
868f2f0b KB |
2893 | EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues); |
2894 | ||
34dbad5d OS |
2895 | /* Enable polling stats and return whether they were already enabled. */ |
2896 | static bool blk_poll_stats_enable(struct request_queue *q) | |
2897 | { | |
2898 | if (test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) || | |
2899 | test_and_set_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags)) | |
2900 | return true; | |
2901 | blk_stat_add_callback(q, q->poll_cb); | |
2902 | return false; | |
2903 | } | |
2904 | ||
2905 | static void blk_mq_poll_stats_start(struct request_queue *q) | |
2906 | { | |
2907 | /* | |
2908 | * We don't arm the callback if polling stats are not enabled or the | |
2909 | * callback is already active. | |
2910 | */ | |
2911 | if (!test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) || | |
2912 | blk_stat_is_active(q->poll_cb)) | |
2913 | return; | |
2914 | ||
2915 | blk_stat_activate_msecs(q->poll_cb, 100); | |
2916 | } | |
2917 | ||
2918 | static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb) | |
2919 | { | |
2920 | struct request_queue *q = cb->data; | |
720b8ccc | 2921 | int bucket; |
34dbad5d | 2922 | |
720b8ccc SB |
2923 | for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) { |
2924 | if (cb->stat[bucket].nr_samples) | |
2925 | q->poll_stat[bucket] = cb->stat[bucket]; | |
2926 | } | |
34dbad5d OS |
2927 | } |
2928 | ||
64f1c21e JA |
2929 | static unsigned long blk_mq_poll_nsecs(struct request_queue *q, |
2930 | struct blk_mq_hw_ctx *hctx, | |
2931 | struct request *rq) | |
2932 | { | |
64f1c21e | 2933 | unsigned long ret = 0; |
720b8ccc | 2934 | int bucket; |
64f1c21e JA |
2935 | |
2936 | /* | |
2937 | * If stats collection isn't on, don't sleep but turn it on for | |
2938 | * future users | |
2939 | */ | |
34dbad5d | 2940 | if (!blk_poll_stats_enable(q)) |
64f1c21e JA |
2941 | return 0; |
2942 | ||
64f1c21e JA |
2943 | /* |
2944 | * As an optimistic guess, use half of the mean service time | |
2945 | * for this type of request. We can (and should) make this smarter. | |
2946 | * For instance, if the completion latencies are tight, we can | |
2947 | * get closer than just half the mean. This is especially | |
2948 | * important on devices where the completion latencies are longer | |
720b8ccc SB |
2949 | * than ~10 usec. We do use the stats for the relevant IO size |
2950 | * if available which does lead to better estimates. | |
64f1c21e | 2951 | */ |
720b8ccc SB |
2952 | bucket = blk_mq_poll_stats_bkt(rq); |
2953 | if (bucket < 0) | |
2954 | return ret; | |
2955 | ||
2956 | if (q->poll_stat[bucket].nr_samples) | |
2957 | ret = (q->poll_stat[bucket].mean + 1) / 2; | |
64f1c21e JA |
2958 | |
2959 | return ret; | |
2960 | } | |
2961 | ||
06426adf | 2962 | static bool blk_mq_poll_hybrid_sleep(struct request_queue *q, |
64f1c21e | 2963 | struct blk_mq_hw_ctx *hctx, |
06426adf JA |
2964 | struct request *rq) |
2965 | { | |
2966 | struct hrtimer_sleeper hs; | |
2967 | enum hrtimer_mode mode; | |
64f1c21e | 2968 | unsigned int nsecs; |
06426adf JA |
2969 | ktime_t kt; |
2970 | ||
64f1c21e JA |
2971 | if (test_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags)) |
2972 | return false; | |
2973 | ||
2974 | /* | |
2975 | * poll_nsec can be: | |
2976 | * | |
2977 | * -1: don't ever hybrid sleep | |
2978 | * 0: use half of prev avg | |
2979 | * >0: use this specific value | |
2980 | */ | |
2981 | if (q->poll_nsec == -1) | |
2982 | return false; | |
2983 | else if (q->poll_nsec > 0) | |
2984 | nsecs = q->poll_nsec; | |
2985 | else | |
2986 | nsecs = blk_mq_poll_nsecs(q, hctx, rq); | |
2987 | ||
2988 | if (!nsecs) | |
06426adf JA |
2989 | return false; |
2990 | ||
2991 | set_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags); | |
2992 | ||
2993 | /* | |
2994 | * This will be replaced with the stats tracking code, using | |
2995 | * 'avg_completion_time / 2' as the pre-sleep target. | |
2996 | */ | |
8b0e1953 | 2997 | kt = nsecs; |
06426adf JA |
2998 | |
2999 | mode = HRTIMER_MODE_REL; | |
3000 | hrtimer_init_on_stack(&hs.timer, CLOCK_MONOTONIC, mode); | |
3001 | hrtimer_set_expires(&hs.timer, kt); | |
3002 | ||
3003 | hrtimer_init_sleeper(&hs, current); | |
3004 | do { | |
3005 | if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags)) | |
3006 | break; | |
3007 | set_current_state(TASK_UNINTERRUPTIBLE); | |
3008 | hrtimer_start_expires(&hs.timer, mode); | |
3009 | if (hs.task) | |
3010 | io_schedule(); | |
3011 | hrtimer_cancel(&hs.timer); | |
3012 | mode = HRTIMER_MODE_ABS; | |
3013 | } while (hs.task && !signal_pending(current)); | |
3014 | ||
3015 | __set_current_state(TASK_RUNNING); | |
3016 | destroy_hrtimer_on_stack(&hs.timer); | |
3017 | return true; | |
3018 | } | |
3019 | ||
bbd7bb70 JA |
3020 | static bool __blk_mq_poll(struct blk_mq_hw_ctx *hctx, struct request *rq) |
3021 | { | |
3022 | struct request_queue *q = hctx->queue; | |
3023 | long state; | |
3024 | ||
06426adf JA |
3025 | /* |
3026 | * If we sleep, have the caller restart the poll loop to reset | |
3027 | * the state. Like for the other success return cases, the | |
3028 | * caller is responsible for checking if the IO completed. If | |
3029 | * the IO isn't complete, we'll get called again and will go | |
3030 | * straight to the busy poll loop. | |
3031 | */ | |
64f1c21e | 3032 | if (blk_mq_poll_hybrid_sleep(q, hctx, rq)) |
06426adf JA |
3033 | return true; |
3034 | ||
bbd7bb70 JA |
3035 | hctx->poll_considered++; |
3036 | ||
3037 | state = current->state; | |
3038 | while (!need_resched()) { | |
3039 | int ret; | |
3040 | ||
3041 | hctx->poll_invoked++; | |
3042 | ||
3043 | ret = q->mq_ops->poll(hctx, rq->tag); | |
3044 | if (ret > 0) { | |
3045 | hctx->poll_success++; | |
3046 | set_current_state(TASK_RUNNING); | |
3047 | return true; | |
3048 | } | |
3049 | ||
3050 | if (signal_pending_state(state, current)) | |
3051 | set_current_state(TASK_RUNNING); | |
3052 | ||
3053 | if (current->state == TASK_RUNNING) | |
3054 | return true; | |
3055 | if (ret < 0) | |
3056 | break; | |
3057 | cpu_relax(); | |
3058 | } | |
3059 | ||
3060 | return false; | |
3061 | } | |
3062 | ||
ea435e1b | 3063 | static bool blk_mq_poll(struct request_queue *q, blk_qc_t cookie) |
bbd7bb70 JA |
3064 | { |
3065 | struct blk_mq_hw_ctx *hctx; | |
bbd7bb70 JA |
3066 | struct request *rq; |
3067 | ||
ea435e1b | 3068 | if (!test_bit(QUEUE_FLAG_POLL, &q->queue_flags)) |
bbd7bb70 JA |
3069 | return false; |
3070 | ||
bbd7bb70 | 3071 | hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)]; |
bd166ef1 JA |
3072 | if (!blk_qc_t_is_internal(cookie)) |
3073 | rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie)); | |
3a07bb1d | 3074 | else { |
bd166ef1 | 3075 | rq = blk_mq_tag_to_rq(hctx->sched_tags, blk_qc_t_to_tag(cookie)); |
3a07bb1d JA |
3076 | /* |
3077 | * With scheduling, if the request has completed, we'll | |
3078 | * get a NULL return here, as we clear the sched tag when | |
3079 | * that happens. The request still remains valid, like always, | |
3080 | * so we should be safe with just the NULL check. | |
3081 | */ | |
3082 | if (!rq) | |
3083 | return false; | |
3084 | } | |
bbd7bb70 JA |
3085 | |
3086 | return __blk_mq_poll(hctx, rq); | |
3087 | } | |
bbd7bb70 | 3088 | |
320ae51f JA |
3089 | static int __init blk_mq_init(void) |
3090 | { | |
fc13457f JA |
3091 | /* |
3092 | * See comment in block/blk.h rq_atomic_flags enum | |
3093 | */ | |
3094 | BUILD_BUG_ON((REQ_ATOM_STARTED / BITS_PER_BYTE) != | |
3095 | (REQ_ATOM_COMPLETE / BITS_PER_BYTE)); | |
3096 | ||
9467f859 TG |
3097 | cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL, |
3098 | blk_mq_hctx_notify_dead); | |
320ae51f JA |
3099 | return 0; |
3100 | } | |
3101 | subsys_initcall(blk_mq_init); |