]>
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 | */ | |
50e1dab8 | 64 | bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx) |
320ae51f | 65 | { |
bd166ef1 JA |
66 | return sbitmap_any_bit_set(&hctx->ctx_map) || |
67 | !list_empty_careful(&hctx->dispatch) || | |
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 | ||
1671d522 | 122 | void blk_freeze_queue_start(struct request_queue *q) |
43a5e4e2 | 123 | { |
4ecd4fef | 124 | int freeze_depth; |
cddd5d17 | 125 | |
4ecd4fef CH |
126 | freeze_depth = atomic_inc_return(&q->mq_freeze_depth); |
127 | if (freeze_depth == 1) { | |
3ef28e83 | 128 | percpu_ref_kill(&q->q_usage_counter); |
b94ec296 | 129 | blk_mq_run_hw_queues(q, false); |
cddd5d17 | 130 | } |
f3af020b | 131 | } |
1671d522 | 132 | EXPORT_SYMBOL_GPL(blk_freeze_queue_start); |
f3af020b | 133 | |
6bae363e | 134 | void blk_mq_freeze_queue_wait(struct request_queue *q) |
f3af020b | 135 | { |
3ef28e83 | 136 | wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter)); |
43a5e4e2 | 137 | } |
6bae363e | 138 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait); |
43a5e4e2 | 139 | |
f91328c4 KB |
140 | int blk_mq_freeze_queue_wait_timeout(struct request_queue *q, |
141 | unsigned long timeout) | |
142 | { | |
143 | return wait_event_timeout(q->mq_freeze_wq, | |
144 | percpu_ref_is_zero(&q->q_usage_counter), | |
145 | timeout); | |
146 | } | |
147 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout); | |
43a5e4e2 | 148 | |
f3af020b TH |
149 | /* |
150 | * Guarantee no request is in use, so we can change any data structure of | |
151 | * the queue afterward. | |
152 | */ | |
3ef28e83 | 153 | void blk_freeze_queue(struct request_queue *q) |
f3af020b | 154 | { |
3ef28e83 DW |
155 | /* |
156 | * In the !blk_mq case we are only calling this to kill the | |
157 | * q_usage_counter, otherwise this increases the freeze depth | |
158 | * and waits for it to return to zero. For this reason there is | |
159 | * no blk_unfreeze_queue(), and blk_freeze_queue() is not | |
160 | * exported to drivers as the only user for unfreeze is blk_mq. | |
161 | */ | |
1671d522 | 162 | blk_freeze_queue_start(q); |
f3af020b TH |
163 | blk_mq_freeze_queue_wait(q); |
164 | } | |
3ef28e83 DW |
165 | |
166 | void blk_mq_freeze_queue(struct request_queue *q) | |
167 | { | |
168 | /* | |
169 | * ...just an alias to keep freeze and unfreeze actions balanced | |
170 | * in the blk_mq_* namespace | |
171 | */ | |
172 | blk_freeze_queue(q); | |
173 | } | |
c761d96b | 174 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue); |
f3af020b | 175 | |
b4c6a028 | 176 | void blk_mq_unfreeze_queue(struct request_queue *q) |
320ae51f | 177 | { |
4ecd4fef | 178 | int freeze_depth; |
320ae51f | 179 | |
4ecd4fef CH |
180 | freeze_depth = atomic_dec_return(&q->mq_freeze_depth); |
181 | WARN_ON_ONCE(freeze_depth < 0); | |
182 | if (!freeze_depth) { | |
3ef28e83 | 183 | percpu_ref_reinit(&q->q_usage_counter); |
320ae51f | 184 | wake_up_all(&q->mq_freeze_wq); |
add703fd | 185 | } |
320ae51f | 186 | } |
b4c6a028 | 187 | EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue); |
320ae51f | 188 | |
852ec809 BVA |
189 | /* |
190 | * FIXME: replace the scsi_internal_device_*block_nowait() calls in the | |
191 | * mpt3sas driver such that this function can be removed. | |
192 | */ | |
193 | void blk_mq_quiesce_queue_nowait(struct request_queue *q) | |
194 | { | |
195 | unsigned long flags; | |
196 | ||
197 | spin_lock_irqsave(q->queue_lock, flags); | |
198 | queue_flag_set(QUEUE_FLAG_QUIESCED, q); | |
199 | spin_unlock_irqrestore(q->queue_lock, flags); | |
200 | } | |
201 | EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait); | |
202 | ||
6a83e74d | 203 | /** |
69e07c4a | 204 | * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished |
6a83e74d BVA |
205 | * @q: request queue. |
206 | * | |
207 | * Note: this function does not prevent that the struct request end_io() | |
69e07c4a ML |
208 | * callback function is invoked. Once this function is returned, we make |
209 | * sure no dispatch can happen until the queue is unquiesced via | |
210 | * blk_mq_unquiesce_queue(). | |
6a83e74d BVA |
211 | */ |
212 | void blk_mq_quiesce_queue(struct request_queue *q) | |
213 | { | |
214 | struct blk_mq_hw_ctx *hctx; | |
215 | unsigned int i; | |
216 | bool rcu = false; | |
217 | ||
1d9e9bc6 | 218 | blk_mq_quiesce_queue_nowait(q); |
f4560ffe | 219 | |
6a83e74d BVA |
220 | queue_for_each_hw_ctx(q, hctx, i) { |
221 | if (hctx->flags & BLK_MQ_F_BLOCKING) | |
07319678 | 222 | synchronize_srcu(hctx->queue_rq_srcu); |
6a83e74d BVA |
223 | else |
224 | rcu = true; | |
225 | } | |
226 | if (rcu) | |
227 | synchronize_rcu(); | |
228 | } | |
229 | EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue); | |
230 | ||
e4e73913 ML |
231 | /* |
232 | * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue() | |
233 | * @q: request queue. | |
234 | * | |
235 | * This function recovers queue into the state before quiescing | |
236 | * which is done by blk_mq_quiesce_queue. | |
237 | */ | |
238 | void blk_mq_unquiesce_queue(struct request_queue *q) | |
239 | { | |
852ec809 BVA |
240 | unsigned long flags; |
241 | ||
242 | spin_lock_irqsave(q->queue_lock, flags); | |
f4560ffe | 243 | queue_flag_clear(QUEUE_FLAG_QUIESCED, q); |
852ec809 | 244 | spin_unlock_irqrestore(q->queue_lock, flags); |
f4560ffe | 245 | |
1d9e9bc6 ML |
246 | /* dispatch requests which are inserted during quiescing */ |
247 | blk_mq_run_hw_queues(q, true); | |
e4e73913 ML |
248 | } |
249 | EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue); | |
250 | ||
aed3ea94 JA |
251 | void blk_mq_wake_waiters(struct request_queue *q) |
252 | { | |
253 | struct blk_mq_hw_ctx *hctx; | |
254 | unsigned int i; | |
255 | ||
256 | queue_for_each_hw_ctx(q, hctx, i) | |
257 | if (blk_mq_hw_queue_mapped(hctx)) | |
258 | blk_mq_tag_wakeup_all(hctx->tags, true); | |
3fd5940c KB |
259 | |
260 | /* | |
261 | * If we are called because the queue has now been marked as | |
262 | * dying, we need to ensure that processes currently waiting on | |
263 | * the queue are notified as well. | |
264 | */ | |
265 | wake_up_all(&q->mq_freeze_wq); | |
aed3ea94 JA |
266 | } |
267 | ||
320ae51f JA |
268 | bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx) |
269 | { | |
270 | return blk_mq_has_free_tags(hctx->tags); | |
271 | } | |
272 | EXPORT_SYMBOL(blk_mq_can_queue); | |
273 | ||
e4cdf1a1 CH |
274 | static struct request *blk_mq_rq_ctx_init(struct blk_mq_alloc_data *data, |
275 | unsigned int tag, unsigned int op) | |
320ae51f | 276 | { |
e4cdf1a1 CH |
277 | struct blk_mq_tags *tags = blk_mq_tags_from_data(data); |
278 | struct request *rq = tags->static_rqs[tag]; | |
279 | ||
c3a148d2 BVA |
280 | rq->rq_flags = 0; |
281 | ||
e4cdf1a1 CH |
282 | if (data->flags & BLK_MQ_REQ_INTERNAL) { |
283 | rq->tag = -1; | |
284 | rq->internal_tag = tag; | |
285 | } else { | |
286 | if (blk_mq_tag_busy(data->hctx)) { | |
287 | rq->rq_flags = RQF_MQ_INFLIGHT; | |
288 | atomic_inc(&data->hctx->nr_active); | |
289 | } | |
290 | rq->tag = tag; | |
291 | rq->internal_tag = -1; | |
292 | data->hctx->tags->rqs[rq->tag] = rq; | |
293 | } | |
294 | ||
af76e555 CH |
295 | INIT_LIST_HEAD(&rq->queuelist); |
296 | /* csd/requeue_work/fifo_time is initialized before use */ | |
e4cdf1a1 CH |
297 | rq->q = data->q; |
298 | rq->mq_ctx = data->ctx; | |
ef295ecf | 299 | rq->cmd_flags = op; |
e4cdf1a1 | 300 | if (blk_queue_io_stat(data->q)) |
e8064021 | 301 | rq->rq_flags |= RQF_IO_STAT; |
af76e555 CH |
302 | /* do not touch atomic flags, it needs atomic ops against the timer */ |
303 | rq->cpu = -1; | |
af76e555 CH |
304 | INIT_HLIST_NODE(&rq->hash); |
305 | RB_CLEAR_NODE(&rq->rb_node); | |
af76e555 CH |
306 | rq->rq_disk = NULL; |
307 | rq->part = NULL; | |
3ee32372 | 308 | rq->start_time = jiffies; |
af76e555 CH |
309 | #ifdef CONFIG_BLK_CGROUP |
310 | rq->rl = NULL; | |
0fec08b4 | 311 | set_start_time_ns(rq); |
af76e555 CH |
312 | rq->io_start_time_ns = 0; |
313 | #endif | |
314 | rq->nr_phys_segments = 0; | |
315 | #if defined(CONFIG_BLK_DEV_INTEGRITY) | |
316 | rq->nr_integrity_segments = 0; | |
317 | #endif | |
af76e555 CH |
318 | rq->special = NULL; |
319 | /* tag was already set */ | |
af76e555 | 320 | rq->extra_len = 0; |
af76e555 | 321 | |
af76e555 | 322 | INIT_LIST_HEAD(&rq->timeout_list); |
f6be4fb4 JA |
323 | rq->timeout = 0; |
324 | ||
af76e555 CH |
325 | rq->end_io = NULL; |
326 | rq->end_io_data = NULL; | |
327 | rq->next_rq = NULL; | |
328 | ||
e4cdf1a1 CH |
329 | data->ctx->rq_dispatched[op_is_sync(op)]++; |
330 | return rq; | |
5dee8577 CH |
331 | } |
332 | ||
d2c0d383 CH |
333 | static struct request *blk_mq_get_request(struct request_queue *q, |
334 | struct bio *bio, unsigned int op, | |
335 | struct blk_mq_alloc_data *data) | |
336 | { | |
337 | struct elevator_queue *e = q->elevator; | |
338 | struct request *rq; | |
e4cdf1a1 | 339 | unsigned int tag; |
21e768b4 | 340 | bool put_ctx_on_error = false; |
d2c0d383 CH |
341 | |
342 | blk_queue_enter_live(q); | |
343 | data->q = q; | |
21e768b4 BVA |
344 | if (likely(!data->ctx)) { |
345 | data->ctx = blk_mq_get_ctx(q); | |
346 | put_ctx_on_error = true; | |
347 | } | |
d2c0d383 CH |
348 | if (likely(!data->hctx)) |
349 | data->hctx = blk_mq_map_queue(q, data->ctx->cpu); | |
03a07c92 GR |
350 | if (op & REQ_NOWAIT) |
351 | data->flags |= BLK_MQ_REQ_NOWAIT; | |
d2c0d383 CH |
352 | |
353 | if (e) { | |
354 | data->flags |= BLK_MQ_REQ_INTERNAL; | |
355 | ||
356 | /* | |
357 | * Flush requests are special and go directly to the | |
358 | * dispatch list. | |
359 | */ | |
5bbf4e5a CH |
360 | if (!op_is_flush(op) && e->type->ops.mq.limit_depth) |
361 | e->type->ops.mq.limit_depth(op, data); | |
d2c0d383 CH |
362 | } |
363 | ||
e4cdf1a1 CH |
364 | tag = blk_mq_get_tag(data); |
365 | if (tag == BLK_MQ_TAG_FAIL) { | |
21e768b4 BVA |
366 | if (put_ctx_on_error) { |
367 | blk_mq_put_ctx(data->ctx); | |
1ad43c00 ML |
368 | data->ctx = NULL; |
369 | } | |
037cebb8 CH |
370 | blk_queue_exit(q); |
371 | return NULL; | |
d2c0d383 CH |
372 | } |
373 | ||
e4cdf1a1 | 374 | rq = blk_mq_rq_ctx_init(data, tag, op); |
037cebb8 CH |
375 | if (!op_is_flush(op)) { |
376 | rq->elv.icq = NULL; | |
5bbf4e5a | 377 | if (e && e->type->ops.mq.prepare_request) { |
44e8c2bf CH |
378 | if (e->type->icq_cache && rq_ioc(bio)) |
379 | blk_mq_sched_assign_ioc(rq, bio); | |
380 | ||
5bbf4e5a CH |
381 | e->type->ops.mq.prepare_request(rq, bio); |
382 | rq->rq_flags |= RQF_ELVPRIV; | |
44e8c2bf | 383 | } |
037cebb8 CH |
384 | } |
385 | data->hctx->queued++; | |
386 | return rq; | |
d2c0d383 CH |
387 | } |
388 | ||
cd6ce148 | 389 | struct request *blk_mq_alloc_request(struct request_queue *q, unsigned int op, |
6f3b0e8b | 390 | unsigned int flags) |
320ae51f | 391 | { |
5a797e00 | 392 | struct blk_mq_alloc_data alloc_data = { .flags = flags }; |
bd166ef1 | 393 | struct request *rq; |
a492f075 | 394 | int ret; |
320ae51f | 395 | |
6f3b0e8b | 396 | ret = blk_queue_enter(q, flags & BLK_MQ_REQ_NOWAIT); |
a492f075 JL |
397 | if (ret) |
398 | return ERR_PTR(ret); | |
320ae51f | 399 | |
cd6ce148 | 400 | rq = blk_mq_get_request(q, NULL, op, &alloc_data); |
3280d66a | 401 | blk_queue_exit(q); |
841bac2c | 402 | |
bd166ef1 | 403 | if (!rq) |
a492f075 | 404 | return ERR_PTR(-EWOULDBLOCK); |
0c4de0f3 | 405 | |
1ad43c00 | 406 | blk_mq_put_ctx(alloc_data.ctx); |
1ad43c00 | 407 | |
0c4de0f3 CH |
408 | rq->__data_len = 0; |
409 | rq->__sector = (sector_t) -1; | |
410 | rq->bio = rq->biotail = NULL; | |
320ae51f JA |
411 | return rq; |
412 | } | |
4bb659b1 | 413 | EXPORT_SYMBOL(blk_mq_alloc_request); |
320ae51f | 414 | |
cd6ce148 BVA |
415 | struct request *blk_mq_alloc_request_hctx(struct request_queue *q, |
416 | unsigned int op, unsigned int flags, unsigned int hctx_idx) | |
1f5bd336 | 417 | { |
6d2809d5 | 418 | struct blk_mq_alloc_data alloc_data = { .flags = flags }; |
1f5bd336 | 419 | struct request *rq; |
6d2809d5 | 420 | unsigned int cpu; |
1f5bd336 ML |
421 | int ret; |
422 | ||
423 | /* | |
424 | * If the tag allocator sleeps we could get an allocation for a | |
425 | * different hardware context. No need to complicate the low level | |
426 | * allocator for this for the rare use case of a command tied to | |
427 | * a specific queue. | |
428 | */ | |
429 | if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT))) | |
430 | return ERR_PTR(-EINVAL); | |
431 | ||
432 | if (hctx_idx >= q->nr_hw_queues) | |
433 | return ERR_PTR(-EIO); | |
434 | ||
435 | ret = blk_queue_enter(q, true); | |
436 | if (ret) | |
437 | return ERR_PTR(ret); | |
438 | ||
c8712c6a CH |
439 | /* |
440 | * Check if the hardware context is actually mapped to anything. | |
441 | * If not tell the caller that it should skip this queue. | |
442 | */ | |
6d2809d5 OS |
443 | alloc_data.hctx = q->queue_hw_ctx[hctx_idx]; |
444 | if (!blk_mq_hw_queue_mapped(alloc_data.hctx)) { | |
445 | blk_queue_exit(q); | |
446 | return ERR_PTR(-EXDEV); | |
c8712c6a | 447 | } |
6d2809d5 OS |
448 | cpu = cpumask_first(alloc_data.hctx->cpumask); |
449 | alloc_data.ctx = __blk_mq_get_ctx(q, cpu); | |
1f5bd336 | 450 | |
cd6ce148 | 451 | rq = blk_mq_get_request(q, NULL, op, &alloc_data); |
3280d66a | 452 | blk_queue_exit(q); |
c8712c6a | 453 | |
6d2809d5 OS |
454 | if (!rq) |
455 | return ERR_PTR(-EWOULDBLOCK); | |
456 | ||
457 | return rq; | |
1f5bd336 ML |
458 | } |
459 | EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx); | |
460 | ||
6af54051 | 461 | void blk_mq_free_request(struct request *rq) |
320ae51f | 462 | { |
320ae51f | 463 | struct request_queue *q = rq->q; |
6af54051 CH |
464 | struct elevator_queue *e = q->elevator; |
465 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
466 | struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, ctx->cpu); | |
467 | const int sched_tag = rq->internal_tag; | |
468 | ||
5bbf4e5a | 469 | if (rq->rq_flags & RQF_ELVPRIV) { |
6af54051 CH |
470 | if (e && e->type->ops.mq.finish_request) |
471 | e->type->ops.mq.finish_request(rq); | |
472 | if (rq->elv.icq) { | |
473 | put_io_context(rq->elv.icq->ioc); | |
474 | rq->elv.icq = NULL; | |
475 | } | |
476 | } | |
320ae51f | 477 | |
6af54051 | 478 | ctx->rq_completed[rq_is_sync(rq)]++; |
e8064021 | 479 | if (rq->rq_flags & RQF_MQ_INFLIGHT) |
0d2602ca | 480 | atomic_dec(&hctx->nr_active); |
87760e5e | 481 | |
7beb2f84 JA |
482 | if (unlikely(laptop_mode && !blk_rq_is_passthrough(rq))) |
483 | laptop_io_completion(q->backing_dev_info); | |
484 | ||
87760e5e | 485 | wbt_done(q->rq_wb, &rq->issue_stat); |
0d2602ca | 486 | |
85acb3ba SL |
487 | if (blk_rq_rl(rq)) |
488 | blk_put_rl(blk_rq_rl(rq)); | |
489 | ||
af76e555 | 490 | clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags); |
06426adf | 491 | clear_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags); |
bd166ef1 JA |
492 | if (rq->tag != -1) |
493 | blk_mq_put_tag(hctx, hctx->tags, ctx, rq->tag); | |
494 | if (sched_tag != -1) | |
c05f8525 | 495 | blk_mq_put_tag(hctx, hctx->sched_tags, ctx, sched_tag); |
6d8c6c0f | 496 | blk_mq_sched_restart(hctx); |
3ef28e83 | 497 | blk_queue_exit(q); |
320ae51f | 498 | } |
1a3b595a | 499 | EXPORT_SYMBOL_GPL(blk_mq_free_request); |
320ae51f | 500 | |
2a842aca | 501 | inline void __blk_mq_end_request(struct request *rq, blk_status_t error) |
320ae51f | 502 | { |
0d11e6ac ML |
503 | blk_account_io_done(rq); |
504 | ||
91b63639 | 505 | if (rq->end_io) { |
87760e5e | 506 | wbt_done(rq->q->rq_wb, &rq->issue_stat); |
320ae51f | 507 | rq->end_io(rq, error); |
91b63639 CH |
508 | } else { |
509 | if (unlikely(blk_bidi_rq(rq))) | |
510 | blk_mq_free_request(rq->next_rq); | |
320ae51f | 511 | blk_mq_free_request(rq); |
91b63639 | 512 | } |
320ae51f | 513 | } |
c8a446ad | 514 | EXPORT_SYMBOL(__blk_mq_end_request); |
63151a44 | 515 | |
2a842aca | 516 | void blk_mq_end_request(struct request *rq, blk_status_t error) |
63151a44 CH |
517 | { |
518 | if (blk_update_request(rq, error, blk_rq_bytes(rq))) | |
519 | BUG(); | |
c8a446ad | 520 | __blk_mq_end_request(rq, error); |
63151a44 | 521 | } |
c8a446ad | 522 | EXPORT_SYMBOL(blk_mq_end_request); |
320ae51f | 523 | |
30a91cb4 | 524 | static void __blk_mq_complete_request_remote(void *data) |
320ae51f | 525 | { |
3d6efbf6 | 526 | struct request *rq = data; |
320ae51f | 527 | |
30a91cb4 | 528 | rq->q->softirq_done_fn(rq); |
320ae51f | 529 | } |
320ae51f | 530 | |
453f8341 | 531 | static void __blk_mq_complete_request(struct request *rq) |
320ae51f JA |
532 | { |
533 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
38535201 | 534 | bool shared = false; |
320ae51f JA |
535 | int cpu; |
536 | ||
453f8341 CH |
537 | if (rq->internal_tag != -1) |
538 | blk_mq_sched_completed_request(rq); | |
539 | if (rq->rq_flags & RQF_STATS) { | |
540 | blk_mq_poll_stats_start(rq->q); | |
541 | blk_stat_add(rq); | |
542 | } | |
543 | ||
38535201 | 544 | if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) { |
30a91cb4 CH |
545 | rq->q->softirq_done_fn(rq); |
546 | return; | |
547 | } | |
320ae51f JA |
548 | |
549 | cpu = get_cpu(); | |
38535201 CH |
550 | if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags)) |
551 | shared = cpus_share_cache(cpu, ctx->cpu); | |
552 | ||
553 | if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) { | |
30a91cb4 | 554 | rq->csd.func = __blk_mq_complete_request_remote; |
3d6efbf6 CH |
555 | rq->csd.info = rq; |
556 | rq->csd.flags = 0; | |
c46fff2a | 557 | smp_call_function_single_async(ctx->cpu, &rq->csd); |
3d6efbf6 | 558 | } else { |
30a91cb4 | 559 | rq->q->softirq_done_fn(rq); |
3d6efbf6 | 560 | } |
320ae51f JA |
561 | put_cpu(); |
562 | } | |
30a91cb4 CH |
563 | |
564 | /** | |
565 | * blk_mq_complete_request - end I/O on a request | |
566 | * @rq: the request being processed | |
567 | * | |
568 | * Description: | |
569 | * Ends all I/O on a request. It does not handle partial completions. | |
570 | * The actual completion happens out-of-order, through a IPI handler. | |
571 | **/ | |
08e0029a | 572 | void blk_mq_complete_request(struct request *rq) |
30a91cb4 | 573 | { |
95f09684 JA |
574 | struct request_queue *q = rq->q; |
575 | ||
576 | if (unlikely(blk_should_fake_timeout(q))) | |
30a91cb4 | 577 | return; |
08e0029a | 578 | if (!blk_mark_rq_complete(rq)) |
ed851860 | 579 | __blk_mq_complete_request(rq); |
30a91cb4 CH |
580 | } |
581 | EXPORT_SYMBOL(blk_mq_complete_request); | |
320ae51f | 582 | |
973c0191 KB |
583 | int blk_mq_request_started(struct request *rq) |
584 | { | |
585 | return test_bit(REQ_ATOM_STARTED, &rq->atomic_flags); | |
586 | } | |
587 | EXPORT_SYMBOL_GPL(blk_mq_request_started); | |
588 | ||
e2490073 | 589 | void blk_mq_start_request(struct request *rq) |
320ae51f JA |
590 | { |
591 | struct request_queue *q = rq->q; | |
592 | ||
bd166ef1 JA |
593 | blk_mq_sched_started_request(rq); |
594 | ||
320ae51f JA |
595 | trace_block_rq_issue(q, rq); |
596 | ||
cf43e6be | 597 | if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) { |
88eeca49 | 598 | blk_stat_set_issue(&rq->issue_stat, blk_rq_sectors(rq)); |
cf43e6be | 599 | rq->rq_flags |= RQF_STATS; |
87760e5e | 600 | wbt_issue(q->rq_wb, &rq->issue_stat); |
cf43e6be JA |
601 | } |
602 | ||
2b8393b4 | 603 | blk_add_timer(rq); |
87ee7b11 | 604 | |
a7af0af3 | 605 | WARN_ON_ONCE(test_bit(REQ_ATOM_STARTED, &rq->atomic_flags)); |
538b7534 | 606 | |
87ee7b11 JA |
607 | /* |
608 | * Mark us as started and clear complete. Complete might have been | |
609 | * set if requeue raced with timeout, which then marked it as | |
610 | * complete. So be sure to clear complete again when we start | |
611 | * the request, otherwise we'll ignore the completion event. | |
a7af0af3 PZ |
612 | * |
613 | * Ensure that ->deadline is visible before we set STARTED, such that | |
614 | * blk_mq_check_expired() is guaranteed to observe our ->deadline when | |
615 | * it observes STARTED. | |
87ee7b11 | 616 | */ |
a7af0af3 PZ |
617 | smp_wmb(); |
618 | set_bit(REQ_ATOM_STARTED, &rq->atomic_flags); | |
619 | if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags)) { | |
620 | /* | |
621 | * Coherence order guarantees these consecutive stores to a | |
622 | * single variable propagate in the specified order. Thus the | |
623 | * clear_bit() is ordered _after_ the set bit. See | |
624 | * blk_mq_check_expired(). | |
625 | * | |
626 | * (the bits must be part of the same byte for this to be | |
627 | * true). | |
628 | */ | |
4b570521 | 629 | clear_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags); |
a7af0af3 | 630 | } |
49f5baa5 CH |
631 | |
632 | if (q->dma_drain_size && blk_rq_bytes(rq)) { | |
633 | /* | |
634 | * Make sure space for the drain appears. We know we can do | |
635 | * this because max_hw_segments has been adjusted to be one | |
636 | * fewer than the device can handle. | |
637 | */ | |
638 | rq->nr_phys_segments++; | |
639 | } | |
320ae51f | 640 | } |
e2490073 | 641 | EXPORT_SYMBOL(blk_mq_start_request); |
320ae51f | 642 | |
d9d149a3 ML |
643 | /* |
644 | * When we reach here because queue is busy, REQ_ATOM_COMPLETE | |
48b99c9d | 645 | * flag isn't set yet, so there may be race with timeout handler, |
d9d149a3 ML |
646 | * but given rq->deadline is just set in .queue_rq() under |
647 | * this situation, the race won't be possible in reality because | |
648 | * rq->timeout should be set as big enough to cover the window | |
649 | * between blk_mq_start_request() called from .queue_rq() and | |
650 | * clearing REQ_ATOM_STARTED here. | |
651 | */ | |
ed0791b2 | 652 | static void __blk_mq_requeue_request(struct request *rq) |
320ae51f JA |
653 | { |
654 | struct request_queue *q = rq->q; | |
655 | ||
923218f6 ML |
656 | blk_mq_put_driver_tag(rq); |
657 | ||
320ae51f | 658 | trace_block_rq_requeue(q, rq); |
87760e5e | 659 | wbt_requeue(q->rq_wb, &rq->issue_stat); |
bd166ef1 | 660 | blk_mq_sched_requeue_request(rq); |
49f5baa5 | 661 | |
e2490073 CH |
662 | if (test_and_clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) { |
663 | if (q->dma_drain_size && blk_rq_bytes(rq)) | |
664 | rq->nr_phys_segments--; | |
665 | } | |
320ae51f JA |
666 | } |
667 | ||
2b053aca | 668 | void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list) |
ed0791b2 | 669 | { |
ed0791b2 | 670 | __blk_mq_requeue_request(rq); |
ed0791b2 | 671 | |
ed0791b2 | 672 | BUG_ON(blk_queued_rq(rq)); |
2b053aca | 673 | blk_mq_add_to_requeue_list(rq, true, kick_requeue_list); |
ed0791b2 CH |
674 | } |
675 | EXPORT_SYMBOL(blk_mq_requeue_request); | |
676 | ||
6fca6a61 CH |
677 | static void blk_mq_requeue_work(struct work_struct *work) |
678 | { | |
679 | struct request_queue *q = | |
2849450a | 680 | container_of(work, struct request_queue, requeue_work.work); |
6fca6a61 CH |
681 | LIST_HEAD(rq_list); |
682 | struct request *rq, *next; | |
6fca6a61 | 683 | |
18e9781d | 684 | spin_lock_irq(&q->requeue_lock); |
6fca6a61 | 685 | list_splice_init(&q->requeue_list, &rq_list); |
18e9781d | 686 | spin_unlock_irq(&q->requeue_lock); |
6fca6a61 CH |
687 | |
688 | list_for_each_entry_safe(rq, next, &rq_list, queuelist) { | |
e8064021 | 689 | if (!(rq->rq_flags & RQF_SOFTBARRIER)) |
6fca6a61 CH |
690 | continue; |
691 | ||
e8064021 | 692 | rq->rq_flags &= ~RQF_SOFTBARRIER; |
6fca6a61 | 693 | list_del_init(&rq->queuelist); |
bd6737f1 | 694 | blk_mq_sched_insert_request(rq, true, false, false, true); |
6fca6a61 CH |
695 | } |
696 | ||
697 | while (!list_empty(&rq_list)) { | |
698 | rq = list_entry(rq_list.next, struct request, queuelist); | |
699 | list_del_init(&rq->queuelist); | |
bd6737f1 | 700 | blk_mq_sched_insert_request(rq, false, false, false, true); |
6fca6a61 CH |
701 | } |
702 | ||
52d7f1b5 | 703 | blk_mq_run_hw_queues(q, false); |
6fca6a61 CH |
704 | } |
705 | ||
2b053aca BVA |
706 | void blk_mq_add_to_requeue_list(struct request *rq, bool at_head, |
707 | bool kick_requeue_list) | |
6fca6a61 CH |
708 | { |
709 | struct request_queue *q = rq->q; | |
710 | unsigned long flags; | |
711 | ||
712 | /* | |
713 | * We abuse this flag that is otherwise used by the I/O scheduler to | |
714 | * request head insertation from the workqueue. | |
715 | */ | |
e8064021 | 716 | BUG_ON(rq->rq_flags & RQF_SOFTBARRIER); |
6fca6a61 CH |
717 | |
718 | spin_lock_irqsave(&q->requeue_lock, flags); | |
719 | if (at_head) { | |
e8064021 | 720 | rq->rq_flags |= RQF_SOFTBARRIER; |
6fca6a61 CH |
721 | list_add(&rq->queuelist, &q->requeue_list); |
722 | } else { | |
723 | list_add_tail(&rq->queuelist, &q->requeue_list); | |
724 | } | |
725 | spin_unlock_irqrestore(&q->requeue_lock, flags); | |
2b053aca BVA |
726 | |
727 | if (kick_requeue_list) | |
728 | blk_mq_kick_requeue_list(q); | |
6fca6a61 CH |
729 | } |
730 | EXPORT_SYMBOL(blk_mq_add_to_requeue_list); | |
731 | ||
732 | void blk_mq_kick_requeue_list(struct request_queue *q) | |
733 | { | |
2849450a | 734 | kblockd_schedule_delayed_work(&q->requeue_work, 0); |
6fca6a61 CH |
735 | } |
736 | EXPORT_SYMBOL(blk_mq_kick_requeue_list); | |
737 | ||
2849450a MS |
738 | void blk_mq_delay_kick_requeue_list(struct request_queue *q, |
739 | unsigned long msecs) | |
740 | { | |
d4acf365 BVA |
741 | kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, |
742 | msecs_to_jiffies(msecs)); | |
2849450a MS |
743 | } |
744 | EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list); | |
745 | ||
0e62f51f JA |
746 | struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag) |
747 | { | |
88c7b2b7 JA |
748 | if (tag < tags->nr_tags) { |
749 | prefetch(tags->rqs[tag]); | |
4ee86bab | 750 | return tags->rqs[tag]; |
88c7b2b7 | 751 | } |
4ee86bab HR |
752 | |
753 | return NULL; | |
24d2f903 CH |
754 | } |
755 | EXPORT_SYMBOL(blk_mq_tag_to_rq); | |
756 | ||
320ae51f | 757 | struct blk_mq_timeout_data { |
46f92d42 CH |
758 | unsigned long next; |
759 | unsigned int next_set; | |
320ae51f JA |
760 | }; |
761 | ||
90415837 | 762 | void blk_mq_rq_timed_out(struct request *req, bool reserved) |
320ae51f | 763 | { |
f8a5b122 | 764 | const struct blk_mq_ops *ops = req->q->mq_ops; |
46f92d42 | 765 | enum blk_eh_timer_return ret = BLK_EH_RESET_TIMER; |
87ee7b11 JA |
766 | |
767 | /* | |
768 | * We know that complete is set at this point. If STARTED isn't set | |
769 | * anymore, then the request isn't active and the "timeout" should | |
770 | * just be ignored. This can happen due to the bitflag ordering. | |
771 | * Timeout first checks if STARTED is set, and if it is, assumes | |
772 | * the request is active. But if we race with completion, then | |
48b99c9d | 773 | * both flags will get cleared. So check here again, and ignore |
87ee7b11 JA |
774 | * a timeout event with a request that isn't active. |
775 | */ | |
46f92d42 CH |
776 | if (!test_bit(REQ_ATOM_STARTED, &req->atomic_flags)) |
777 | return; | |
87ee7b11 | 778 | |
46f92d42 | 779 | if (ops->timeout) |
0152fb6b | 780 | ret = ops->timeout(req, reserved); |
46f92d42 CH |
781 | |
782 | switch (ret) { | |
783 | case BLK_EH_HANDLED: | |
784 | __blk_mq_complete_request(req); | |
785 | break; | |
786 | case BLK_EH_RESET_TIMER: | |
787 | blk_add_timer(req); | |
788 | blk_clear_rq_complete(req); | |
789 | break; | |
790 | case BLK_EH_NOT_HANDLED: | |
791 | break; | |
792 | default: | |
793 | printk(KERN_ERR "block: bad eh return: %d\n", ret); | |
794 | break; | |
795 | } | |
87ee7b11 | 796 | } |
5b3f25fc | 797 | |
81481eb4 CH |
798 | static void blk_mq_check_expired(struct blk_mq_hw_ctx *hctx, |
799 | struct request *rq, void *priv, bool reserved) | |
800 | { | |
801 | struct blk_mq_timeout_data *data = priv; | |
a7af0af3 | 802 | unsigned long deadline; |
87ee7b11 | 803 | |
95a49603 | 804 | if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) |
46f92d42 | 805 | return; |
87ee7b11 | 806 | |
a7af0af3 PZ |
807 | /* |
808 | * Ensures that if we see STARTED we must also see our | |
809 | * up-to-date deadline, see blk_mq_start_request(). | |
810 | */ | |
811 | smp_rmb(); | |
812 | ||
813 | deadline = READ_ONCE(rq->deadline); | |
814 | ||
d9d149a3 ML |
815 | /* |
816 | * The rq being checked may have been freed and reallocated | |
817 | * out already here, we avoid this race by checking rq->deadline | |
818 | * and REQ_ATOM_COMPLETE flag together: | |
819 | * | |
820 | * - if rq->deadline is observed as new value because of | |
821 | * reusing, the rq won't be timed out because of timing. | |
822 | * - if rq->deadline is observed as previous value, | |
823 | * REQ_ATOM_COMPLETE flag won't be cleared in reuse path | |
824 | * because we put a barrier between setting rq->deadline | |
825 | * and clearing the flag in blk_mq_start_request(), so | |
826 | * this rq won't be timed out too. | |
827 | */ | |
a7af0af3 PZ |
828 | if (time_after_eq(jiffies, deadline)) { |
829 | if (!blk_mark_rq_complete(rq)) { | |
830 | /* | |
831 | * Again coherence order ensures that consecutive reads | |
832 | * from the same variable must be in that order. This | |
833 | * ensures that if we see COMPLETE clear, we must then | |
834 | * see STARTED set and we'll ignore this timeout. | |
835 | * | |
836 | * (There's also the MB implied by the test_and_clear()) | |
837 | */ | |
0152fb6b | 838 | blk_mq_rq_timed_out(rq, reserved); |
a7af0af3 PZ |
839 | } |
840 | } else if (!data->next_set || time_after(data->next, deadline)) { | |
841 | data->next = deadline; | |
46f92d42 CH |
842 | data->next_set = 1; |
843 | } | |
87ee7b11 JA |
844 | } |
845 | ||
287922eb | 846 | static void blk_mq_timeout_work(struct work_struct *work) |
320ae51f | 847 | { |
287922eb CH |
848 | struct request_queue *q = |
849 | container_of(work, struct request_queue, timeout_work); | |
81481eb4 CH |
850 | struct blk_mq_timeout_data data = { |
851 | .next = 0, | |
852 | .next_set = 0, | |
853 | }; | |
81481eb4 | 854 | int i; |
320ae51f | 855 | |
71f79fb3 GKB |
856 | /* A deadlock might occur if a request is stuck requiring a |
857 | * timeout at the same time a queue freeze is waiting | |
858 | * completion, since the timeout code would not be able to | |
859 | * acquire the queue reference here. | |
860 | * | |
861 | * That's why we don't use blk_queue_enter here; instead, we use | |
862 | * percpu_ref_tryget directly, because we need to be able to | |
863 | * obtain a reference even in the short window between the queue | |
864 | * starting to freeze, by dropping the first reference in | |
1671d522 | 865 | * blk_freeze_queue_start, and the moment the last request is |
71f79fb3 GKB |
866 | * consumed, marked by the instant q_usage_counter reaches |
867 | * zero. | |
868 | */ | |
869 | if (!percpu_ref_tryget(&q->q_usage_counter)) | |
287922eb CH |
870 | return; |
871 | ||
0bf6cd5b | 872 | blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &data); |
320ae51f | 873 | |
81481eb4 CH |
874 | if (data.next_set) { |
875 | data.next = blk_rq_timeout(round_jiffies_up(data.next)); | |
876 | mod_timer(&q->timeout, data.next); | |
0d2602ca | 877 | } else { |
0bf6cd5b CH |
878 | struct blk_mq_hw_ctx *hctx; |
879 | ||
f054b56c ML |
880 | queue_for_each_hw_ctx(q, hctx, i) { |
881 | /* the hctx may be unmapped, so check it here */ | |
882 | if (blk_mq_hw_queue_mapped(hctx)) | |
883 | blk_mq_tag_idle(hctx); | |
884 | } | |
0d2602ca | 885 | } |
287922eb | 886 | blk_queue_exit(q); |
320ae51f JA |
887 | } |
888 | ||
88459642 OS |
889 | struct flush_busy_ctx_data { |
890 | struct blk_mq_hw_ctx *hctx; | |
891 | struct list_head *list; | |
892 | }; | |
893 | ||
894 | static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data) | |
895 | { | |
896 | struct flush_busy_ctx_data *flush_data = data; | |
897 | struct blk_mq_hw_ctx *hctx = flush_data->hctx; | |
898 | struct blk_mq_ctx *ctx = hctx->ctxs[bitnr]; | |
899 | ||
900 | sbitmap_clear_bit(sb, bitnr); | |
901 | spin_lock(&ctx->lock); | |
902 | list_splice_tail_init(&ctx->rq_list, flush_data->list); | |
903 | spin_unlock(&ctx->lock); | |
904 | return true; | |
905 | } | |
906 | ||
1429d7c9 JA |
907 | /* |
908 | * Process software queues that have been marked busy, splicing them | |
909 | * to the for-dispatch | |
910 | */ | |
2c3ad667 | 911 | void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list) |
1429d7c9 | 912 | { |
88459642 OS |
913 | struct flush_busy_ctx_data data = { |
914 | .hctx = hctx, | |
915 | .list = list, | |
916 | }; | |
1429d7c9 | 917 | |
88459642 | 918 | sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data); |
1429d7c9 | 919 | } |
2c3ad667 | 920 | EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs); |
1429d7c9 | 921 | |
b347689f ML |
922 | struct dispatch_rq_data { |
923 | struct blk_mq_hw_ctx *hctx; | |
924 | struct request *rq; | |
925 | }; | |
926 | ||
927 | static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr, | |
928 | void *data) | |
929 | { | |
930 | struct dispatch_rq_data *dispatch_data = data; | |
931 | struct blk_mq_hw_ctx *hctx = dispatch_data->hctx; | |
932 | struct blk_mq_ctx *ctx = hctx->ctxs[bitnr]; | |
933 | ||
934 | spin_lock(&ctx->lock); | |
935 | if (unlikely(!list_empty(&ctx->rq_list))) { | |
936 | dispatch_data->rq = list_entry_rq(ctx->rq_list.next); | |
937 | list_del_init(&dispatch_data->rq->queuelist); | |
938 | if (list_empty(&ctx->rq_list)) | |
939 | sbitmap_clear_bit(sb, bitnr); | |
940 | } | |
941 | spin_unlock(&ctx->lock); | |
942 | ||
943 | return !dispatch_data->rq; | |
944 | } | |
945 | ||
946 | struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx, | |
947 | struct blk_mq_ctx *start) | |
948 | { | |
949 | unsigned off = start ? start->index_hw : 0; | |
950 | struct dispatch_rq_data data = { | |
951 | .hctx = hctx, | |
952 | .rq = NULL, | |
953 | }; | |
954 | ||
955 | __sbitmap_for_each_set(&hctx->ctx_map, off, | |
956 | dispatch_rq_from_ctx, &data); | |
957 | ||
958 | return data.rq; | |
959 | } | |
960 | ||
703fd1c0 JA |
961 | static inline unsigned int queued_to_index(unsigned int queued) |
962 | { | |
963 | if (!queued) | |
964 | return 0; | |
1429d7c9 | 965 | |
703fd1c0 | 966 | return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1); |
1429d7c9 JA |
967 | } |
968 | ||
bd6737f1 JA |
969 | bool blk_mq_get_driver_tag(struct request *rq, struct blk_mq_hw_ctx **hctx, |
970 | bool wait) | |
bd166ef1 JA |
971 | { |
972 | struct blk_mq_alloc_data data = { | |
973 | .q = rq->q, | |
bd166ef1 JA |
974 | .hctx = blk_mq_map_queue(rq->q, rq->mq_ctx->cpu), |
975 | .flags = wait ? 0 : BLK_MQ_REQ_NOWAIT, | |
976 | }; | |
977 | ||
5feeacdd JA |
978 | might_sleep_if(wait); |
979 | ||
81380ca1 OS |
980 | if (rq->tag != -1) |
981 | goto done; | |
bd166ef1 | 982 | |
415b806d SG |
983 | if (blk_mq_tag_is_reserved(data.hctx->sched_tags, rq->internal_tag)) |
984 | data.flags |= BLK_MQ_REQ_RESERVED; | |
985 | ||
bd166ef1 JA |
986 | rq->tag = blk_mq_get_tag(&data); |
987 | if (rq->tag >= 0) { | |
200e86b3 JA |
988 | if (blk_mq_tag_busy(data.hctx)) { |
989 | rq->rq_flags |= RQF_MQ_INFLIGHT; | |
990 | atomic_inc(&data.hctx->nr_active); | |
991 | } | |
bd166ef1 | 992 | data.hctx->tags->rqs[rq->tag] = rq; |
bd166ef1 JA |
993 | } |
994 | ||
81380ca1 OS |
995 | done: |
996 | if (hctx) | |
997 | *hctx = data.hctx; | |
998 | return rq->tag != -1; | |
bd166ef1 JA |
999 | } |
1000 | ||
eb619fdb JA |
1001 | static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode, |
1002 | int flags, void *key) | |
da55f2cc OS |
1003 | { |
1004 | struct blk_mq_hw_ctx *hctx; | |
1005 | ||
1006 | hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait); | |
1007 | ||
eb619fdb | 1008 | list_del_init(&wait->entry); |
da55f2cc OS |
1009 | blk_mq_run_hw_queue(hctx, true); |
1010 | return 1; | |
1011 | } | |
1012 | ||
eb619fdb JA |
1013 | static bool blk_mq_dispatch_wait_add(struct blk_mq_hw_ctx **hctx, |
1014 | struct request *rq) | |
da55f2cc | 1015 | { |
eb619fdb JA |
1016 | struct blk_mq_hw_ctx *this_hctx = *hctx; |
1017 | wait_queue_entry_t *wait = &this_hctx->dispatch_wait; | |
da55f2cc OS |
1018 | struct sbq_wait_state *ws; |
1019 | ||
eb619fdb JA |
1020 | if (!list_empty_careful(&wait->entry)) |
1021 | return false; | |
1022 | ||
1023 | spin_lock(&this_hctx->lock); | |
1024 | if (!list_empty(&wait->entry)) { | |
1025 | spin_unlock(&this_hctx->lock); | |
1026 | return false; | |
1027 | } | |
1028 | ||
1029 | ws = bt_wait_ptr(&this_hctx->tags->bitmap_tags, this_hctx); | |
1030 | add_wait_queue(&ws->wait, wait); | |
1031 | ||
da55f2cc | 1032 | /* |
eb619fdb JA |
1033 | * It's possible that a tag was freed in the window between the |
1034 | * allocation failure and adding the hardware queue to the wait | |
1035 | * queue. | |
da55f2cc | 1036 | */ |
eb619fdb JA |
1037 | if (!blk_mq_get_driver_tag(rq, hctx, false)) { |
1038 | spin_unlock(&this_hctx->lock); | |
da55f2cc | 1039 | return false; |
eb619fdb | 1040 | } |
da55f2cc OS |
1041 | |
1042 | /* | |
eb619fdb JA |
1043 | * We got a tag, remove ourselves from the wait queue to ensure |
1044 | * someone else gets the wakeup. | |
da55f2cc | 1045 | */ |
eb619fdb JA |
1046 | spin_lock_irq(&ws->wait.lock); |
1047 | list_del_init(&wait->entry); | |
1048 | spin_unlock_irq(&ws->wait.lock); | |
1049 | spin_unlock(&this_hctx->lock); | |
da55f2cc OS |
1050 | return true; |
1051 | } | |
1052 | ||
de148297 | 1053 | bool blk_mq_dispatch_rq_list(struct request_queue *q, struct list_head *list, |
eb619fdb | 1054 | bool got_budget) |
320ae51f | 1055 | { |
81380ca1 | 1056 | struct blk_mq_hw_ctx *hctx; |
6d6f167c | 1057 | struct request *rq, *nxt; |
eb619fdb | 1058 | bool no_tag = false; |
fc17b653 | 1059 | int errors, queued; |
320ae51f | 1060 | |
81380ca1 OS |
1061 | if (list_empty(list)) |
1062 | return false; | |
1063 | ||
de148297 ML |
1064 | WARN_ON(!list_is_singular(list) && got_budget); |
1065 | ||
320ae51f JA |
1066 | /* |
1067 | * Now process all the entries, sending them to the driver. | |
1068 | */ | |
93efe981 | 1069 | errors = queued = 0; |
81380ca1 | 1070 | do { |
74c45052 | 1071 | struct blk_mq_queue_data bd; |
fc17b653 | 1072 | blk_status_t ret; |
320ae51f | 1073 | |
f04c3df3 | 1074 | rq = list_first_entry(list, struct request, queuelist); |
bd166ef1 | 1075 | if (!blk_mq_get_driver_tag(rq, &hctx, false)) { |
3c782d67 | 1076 | /* |
da55f2cc | 1077 | * The initial allocation attempt failed, so we need to |
eb619fdb JA |
1078 | * rerun the hardware queue when a tag is freed. The |
1079 | * waitqueue takes care of that. If the queue is run | |
1080 | * before we add this entry back on the dispatch list, | |
1081 | * we'll re-run it below. | |
3c782d67 | 1082 | */ |
eb619fdb | 1083 | if (!blk_mq_dispatch_wait_add(&hctx, rq)) { |
de148297 ML |
1084 | if (got_budget) |
1085 | blk_mq_put_dispatch_budget(hctx); | |
eb619fdb | 1086 | no_tag = true; |
de148297 ML |
1087 | break; |
1088 | } | |
1089 | } | |
1090 | ||
0c6af1cc ML |
1091 | if (!got_budget && !blk_mq_get_dispatch_budget(hctx)) { |
1092 | blk_mq_put_driver_tag(rq); | |
88022d72 | 1093 | break; |
0c6af1cc | 1094 | } |
da55f2cc | 1095 | |
320ae51f | 1096 | list_del_init(&rq->queuelist); |
320ae51f | 1097 | |
74c45052 | 1098 | bd.rq = rq; |
113285b4 JA |
1099 | |
1100 | /* | |
1101 | * Flag last if we have no more requests, or if we have more | |
1102 | * but can't assign a driver tag to it. | |
1103 | */ | |
1104 | if (list_empty(list)) | |
1105 | bd.last = true; | |
1106 | else { | |
113285b4 JA |
1107 | nxt = list_first_entry(list, struct request, queuelist); |
1108 | bd.last = !blk_mq_get_driver_tag(nxt, NULL, false); | |
1109 | } | |
74c45052 JA |
1110 | |
1111 | ret = q->mq_ops->queue_rq(hctx, &bd); | |
fc17b653 | 1112 | if (ret == BLK_STS_RESOURCE) { |
6d6f167c JW |
1113 | /* |
1114 | * If an I/O scheduler has been configured and we got a | |
1115 | * driver tag for the next request already, free it again. | |
1116 | */ | |
1117 | if (!list_empty(list)) { | |
1118 | nxt = list_first_entry(list, struct request, queuelist); | |
1119 | blk_mq_put_driver_tag(nxt); | |
1120 | } | |
f04c3df3 | 1121 | list_add(&rq->queuelist, list); |
ed0791b2 | 1122 | __blk_mq_requeue_request(rq); |
320ae51f | 1123 | break; |
fc17b653 CH |
1124 | } |
1125 | ||
1126 | if (unlikely(ret != BLK_STS_OK)) { | |
93efe981 | 1127 | errors++; |
2a842aca | 1128 | blk_mq_end_request(rq, BLK_STS_IOERR); |
fc17b653 | 1129 | continue; |
320ae51f JA |
1130 | } |
1131 | ||
fc17b653 | 1132 | queued++; |
81380ca1 | 1133 | } while (!list_empty(list)); |
320ae51f | 1134 | |
703fd1c0 | 1135 | hctx->dispatched[queued_to_index(queued)]++; |
320ae51f JA |
1136 | |
1137 | /* | |
1138 | * Any items that need requeuing? Stuff them into hctx->dispatch, | |
1139 | * that is where we will continue on next queue run. | |
1140 | */ | |
f04c3df3 | 1141 | if (!list_empty(list)) { |
320ae51f | 1142 | spin_lock(&hctx->lock); |
c13660a0 | 1143 | list_splice_init(list, &hctx->dispatch); |
320ae51f | 1144 | spin_unlock(&hctx->lock); |
f04c3df3 | 1145 | |
9ba52e58 | 1146 | /* |
710c785f BVA |
1147 | * If SCHED_RESTART was set by the caller of this function and |
1148 | * it is no longer set that means that it was cleared by another | |
1149 | * thread and hence that a queue rerun is needed. | |
9ba52e58 | 1150 | * |
eb619fdb JA |
1151 | * If 'no_tag' is set, that means that we failed getting |
1152 | * a driver tag with an I/O scheduler attached. If our dispatch | |
1153 | * waitqueue is no longer active, ensure that we run the queue | |
1154 | * AFTER adding our entries back to the list. | |
bd166ef1 | 1155 | * |
710c785f BVA |
1156 | * If no I/O scheduler has been configured it is possible that |
1157 | * the hardware queue got stopped and restarted before requests | |
1158 | * were pushed back onto the dispatch list. Rerun the queue to | |
1159 | * avoid starvation. Notes: | |
1160 | * - blk_mq_run_hw_queue() checks whether or not a queue has | |
1161 | * been stopped before rerunning a queue. | |
1162 | * - Some but not all block drivers stop a queue before | |
fc17b653 | 1163 | * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq |
710c785f | 1164 | * and dm-rq. |
bd166ef1 | 1165 | */ |
eb619fdb JA |
1166 | if (!blk_mq_sched_needs_restart(hctx) || |
1167 | (no_tag && list_empty_careful(&hctx->dispatch_wait.entry))) | |
bd166ef1 | 1168 | blk_mq_run_hw_queue(hctx, true); |
320ae51f | 1169 | } |
f04c3df3 | 1170 | |
93efe981 | 1171 | return (queued + errors) != 0; |
f04c3df3 JA |
1172 | } |
1173 | ||
6a83e74d BVA |
1174 | static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx) |
1175 | { | |
1176 | int srcu_idx; | |
1177 | ||
b7a71e66 JA |
1178 | /* |
1179 | * We should be running this queue from one of the CPUs that | |
1180 | * are mapped to it. | |
1181 | */ | |
6a83e74d BVA |
1182 | WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) && |
1183 | cpu_online(hctx->next_cpu)); | |
1184 | ||
b7a71e66 JA |
1185 | /* |
1186 | * We can't run the queue inline with ints disabled. Ensure that | |
1187 | * we catch bad users of this early. | |
1188 | */ | |
1189 | WARN_ON_ONCE(in_interrupt()); | |
1190 | ||
6a83e74d BVA |
1191 | if (!(hctx->flags & BLK_MQ_F_BLOCKING)) { |
1192 | rcu_read_lock(); | |
1f460b63 | 1193 | blk_mq_sched_dispatch_requests(hctx); |
6a83e74d BVA |
1194 | rcu_read_unlock(); |
1195 | } else { | |
bf4907c0 JA |
1196 | might_sleep(); |
1197 | ||
07319678 | 1198 | srcu_idx = srcu_read_lock(hctx->queue_rq_srcu); |
1f460b63 | 1199 | blk_mq_sched_dispatch_requests(hctx); |
07319678 | 1200 | srcu_read_unlock(hctx->queue_rq_srcu, srcu_idx); |
6a83e74d BVA |
1201 | } |
1202 | } | |
1203 | ||
506e931f JA |
1204 | /* |
1205 | * It'd be great if the workqueue API had a way to pass | |
1206 | * in a mask and had some smarts for more clever placement. | |
1207 | * For now we just round-robin here, switching for every | |
1208 | * BLK_MQ_CPU_WORK_BATCH queued items. | |
1209 | */ | |
1210 | static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx) | |
1211 | { | |
b657d7e6 CH |
1212 | if (hctx->queue->nr_hw_queues == 1) |
1213 | return WORK_CPU_UNBOUND; | |
506e931f JA |
1214 | |
1215 | if (--hctx->next_cpu_batch <= 0) { | |
c02ebfdd | 1216 | int next_cpu; |
506e931f JA |
1217 | |
1218 | next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask); | |
1219 | if (next_cpu >= nr_cpu_ids) | |
1220 | next_cpu = cpumask_first(hctx->cpumask); | |
1221 | ||
1222 | hctx->next_cpu = next_cpu; | |
1223 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; | |
1224 | } | |
1225 | ||
b657d7e6 | 1226 | return hctx->next_cpu; |
506e931f JA |
1227 | } |
1228 | ||
7587a5ae BVA |
1229 | static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async, |
1230 | unsigned long msecs) | |
320ae51f | 1231 | { |
5435c023 BVA |
1232 | if (WARN_ON_ONCE(!blk_mq_hw_queue_mapped(hctx))) |
1233 | return; | |
1234 | ||
1235 | if (unlikely(blk_mq_hctx_stopped(hctx))) | |
320ae51f JA |
1236 | return; |
1237 | ||
1b792f2f | 1238 | if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) { |
2a90d4aa PB |
1239 | int cpu = get_cpu(); |
1240 | if (cpumask_test_cpu(cpu, hctx->cpumask)) { | |
398205b8 | 1241 | __blk_mq_run_hw_queue(hctx); |
2a90d4aa | 1242 | put_cpu(); |
398205b8 PB |
1243 | return; |
1244 | } | |
e4043dcf | 1245 | |
2a90d4aa | 1246 | put_cpu(); |
e4043dcf | 1247 | } |
398205b8 | 1248 | |
9f993737 JA |
1249 | kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx), |
1250 | &hctx->run_work, | |
1251 | msecs_to_jiffies(msecs)); | |
7587a5ae BVA |
1252 | } |
1253 | ||
1254 | void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs) | |
1255 | { | |
1256 | __blk_mq_delay_run_hw_queue(hctx, true, msecs); | |
1257 | } | |
1258 | EXPORT_SYMBOL(blk_mq_delay_run_hw_queue); | |
1259 | ||
1260 | void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) | |
1261 | { | |
1262 | __blk_mq_delay_run_hw_queue(hctx, async, 0); | |
320ae51f | 1263 | } |
5b727272 | 1264 | EXPORT_SYMBOL(blk_mq_run_hw_queue); |
320ae51f | 1265 | |
b94ec296 | 1266 | void blk_mq_run_hw_queues(struct request_queue *q, bool async) |
320ae51f JA |
1267 | { |
1268 | struct blk_mq_hw_ctx *hctx; | |
1269 | int i; | |
1270 | ||
1271 | queue_for_each_hw_ctx(q, hctx, i) { | |
bd166ef1 | 1272 | if (!blk_mq_hctx_has_pending(hctx) || |
5d1b25c1 | 1273 | blk_mq_hctx_stopped(hctx)) |
320ae51f JA |
1274 | continue; |
1275 | ||
b94ec296 | 1276 | blk_mq_run_hw_queue(hctx, async); |
320ae51f JA |
1277 | } |
1278 | } | |
b94ec296 | 1279 | EXPORT_SYMBOL(blk_mq_run_hw_queues); |
320ae51f | 1280 | |
fd001443 BVA |
1281 | /** |
1282 | * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped | |
1283 | * @q: request queue. | |
1284 | * | |
1285 | * The caller is responsible for serializing this function against | |
1286 | * blk_mq_{start,stop}_hw_queue(). | |
1287 | */ | |
1288 | bool blk_mq_queue_stopped(struct request_queue *q) | |
1289 | { | |
1290 | struct blk_mq_hw_ctx *hctx; | |
1291 | int i; | |
1292 | ||
1293 | queue_for_each_hw_ctx(q, hctx, i) | |
1294 | if (blk_mq_hctx_stopped(hctx)) | |
1295 | return true; | |
1296 | ||
1297 | return false; | |
1298 | } | |
1299 | EXPORT_SYMBOL(blk_mq_queue_stopped); | |
1300 | ||
39a70c76 ML |
1301 | /* |
1302 | * This function is often used for pausing .queue_rq() by driver when | |
1303 | * there isn't enough resource or some conditions aren't satisfied, and | |
4d606219 | 1304 | * BLK_STS_RESOURCE is usually returned. |
39a70c76 ML |
1305 | * |
1306 | * We do not guarantee that dispatch can be drained or blocked | |
1307 | * after blk_mq_stop_hw_queue() returns. Please use | |
1308 | * blk_mq_quiesce_queue() for that requirement. | |
1309 | */ | |
2719aa21 JA |
1310 | void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx) |
1311 | { | |
641a9ed6 | 1312 | cancel_delayed_work(&hctx->run_work); |
280d45f6 | 1313 | |
641a9ed6 | 1314 | set_bit(BLK_MQ_S_STOPPED, &hctx->state); |
2719aa21 | 1315 | } |
641a9ed6 | 1316 | EXPORT_SYMBOL(blk_mq_stop_hw_queue); |
2719aa21 | 1317 | |
39a70c76 ML |
1318 | /* |
1319 | * This function is often used for pausing .queue_rq() by driver when | |
1320 | * there isn't enough resource or some conditions aren't satisfied, and | |
4d606219 | 1321 | * BLK_STS_RESOURCE is usually returned. |
39a70c76 ML |
1322 | * |
1323 | * We do not guarantee that dispatch can be drained or blocked | |
1324 | * after blk_mq_stop_hw_queues() returns. Please use | |
1325 | * blk_mq_quiesce_queue() for that requirement. | |
1326 | */ | |
2719aa21 JA |
1327 | void blk_mq_stop_hw_queues(struct request_queue *q) |
1328 | { | |
641a9ed6 ML |
1329 | struct blk_mq_hw_ctx *hctx; |
1330 | int i; | |
1331 | ||
1332 | queue_for_each_hw_ctx(q, hctx, i) | |
1333 | blk_mq_stop_hw_queue(hctx); | |
280d45f6 CH |
1334 | } |
1335 | EXPORT_SYMBOL(blk_mq_stop_hw_queues); | |
1336 | ||
320ae51f JA |
1337 | void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx) |
1338 | { | |
1339 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
e4043dcf | 1340 | |
0ffbce80 | 1341 | blk_mq_run_hw_queue(hctx, false); |
320ae51f JA |
1342 | } |
1343 | EXPORT_SYMBOL(blk_mq_start_hw_queue); | |
1344 | ||
2f268556 CH |
1345 | void blk_mq_start_hw_queues(struct request_queue *q) |
1346 | { | |
1347 | struct blk_mq_hw_ctx *hctx; | |
1348 | int i; | |
1349 | ||
1350 | queue_for_each_hw_ctx(q, hctx, i) | |
1351 | blk_mq_start_hw_queue(hctx); | |
1352 | } | |
1353 | EXPORT_SYMBOL(blk_mq_start_hw_queues); | |
1354 | ||
ae911c5e JA |
1355 | void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) |
1356 | { | |
1357 | if (!blk_mq_hctx_stopped(hctx)) | |
1358 | return; | |
1359 | ||
1360 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
1361 | blk_mq_run_hw_queue(hctx, async); | |
1362 | } | |
1363 | EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue); | |
1364 | ||
1b4a3258 | 1365 | void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async) |
320ae51f JA |
1366 | { |
1367 | struct blk_mq_hw_ctx *hctx; | |
1368 | int i; | |
1369 | ||
ae911c5e JA |
1370 | queue_for_each_hw_ctx(q, hctx, i) |
1371 | blk_mq_start_stopped_hw_queue(hctx, async); | |
320ae51f JA |
1372 | } |
1373 | EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues); | |
1374 | ||
70f4db63 | 1375 | static void blk_mq_run_work_fn(struct work_struct *work) |
320ae51f JA |
1376 | { |
1377 | struct blk_mq_hw_ctx *hctx; | |
1378 | ||
9f993737 | 1379 | hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work); |
320ae51f | 1380 | |
21c6e939 JA |
1381 | /* |
1382 | * If we are stopped, don't run the queue. The exception is if | |
1383 | * BLK_MQ_S_START_ON_RUN is set. For that case, we auto-clear | |
1384 | * the STOPPED bit and run it. | |
1385 | */ | |
1386 | if (test_bit(BLK_MQ_S_STOPPED, &hctx->state)) { | |
1387 | if (!test_bit(BLK_MQ_S_START_ON_RUN, &hctx->state)) | |
1388 | return; | |
7587a5ae | 1389 | |
21c6e939 JA |
1390 | clear_bit(BLK_MQ_S_START_ON_RUN, &hctx->state); |
1391 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
1392 | } | |
7587a5ae BVA |
1393 | |
1394 | __blk_mq_run_hw_queue(hctx); | |
1395 | } | |
1396 | ||
70f4db63 CH |
1397 | |
1398 | void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs) | |
1399 | { | |
5435c023 | 1400 | if (WARN_ON_ONCE(!blk_mq_hw_queue_mapped(hctx))) |
19c66e59 | 1401 | return; |
70f4db63 | 1402 | |
21c6e939 JA |
1403 | /* |
1404 | * Stop the hw queue, then modify currently delayed work. | |
1405 | * This should prevent us from running the queue prematurely. | |
1406 | * Mark the queue as auto-clearing STOPPED when it runs. | |
1407 | */ | |
7e79dadc | 1408 | blk_mq_stop_hw_queue(hctx); |
21c6e939 JA |
1409 | set_bit(BLK_MQ_S_START_ON_RUN, &hctx->state); |
1410 | kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), | |
1411 | &hctx->run_work, | |
1412 | msecs_to_jiffies(msecs)); | |
70f4db63 CH |
1413 | } |
1414 | EXPORT_SYMBOL(blk_mq_delay_queue); | |
1415 | ||
cfd0c552 | 1416 | static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx, |
cfd0c552 ML |
1417 | struct request *rq, |
1418 | bool at_head) | |
320ae51f | 1419 | { |
e57690fe JA |
1420 | struct blk_mq_ctx *ctx = rq->mq_ctx; |
1421 | ||
7b607814 BVA |
1422 | lockdep_assert_held(&ctx->lock); |
1423 | ||
01b983c9 JA |
1424 | trace_block_rq_insert(hctx->queue, rq); |
1425 | ||
72a0a36e CH |
1426 | if (at_head) |
1427 | list_add(&rq->queuelist, &ctx->rq_list); | |
1428 | else | |
1429 | list_add_tail(&rq->queuelist, &ctx->rq_list); | |
cfd0c552 | 1430 | } |
4bb659b1 | 1431 | |
2c3ad667 JA |
1432 | void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq, |
1433 | bool at_head) | |
cfd0c552 ML |
1434 | { |
1435 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
1436 | ||
7b607814 BVA |
1437 | lockdep_assert_held(&ctx->lock); |
1438 | ||
e57690fe | 1439 | __blk_mq_insert_req_list(hctx, rq, at_head); |
320ae51f | 1440 | blk_mq_hctx_mark_pending(hctx, ctx); |
320ae51f JA |
1441 | } |
1442 | ||
157f377b JA |
1443 | /* |
1444 | * Should only be used carefully, when the caller knows we want to | |
1445 | * bypass a potential IO scheduler on the target device. | |
1446 | */ | |
b0850297 | 1447 | void blk_mq_request_bypass_insert(struct request *rq, bool run_queue) |
157f377b JA |
1448 | { |
1449 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
1450 | struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(rq->q, ctx->cpu); | |
1451 | ||
1452 | spin_lock(&hctx->lock); | |
1453 | list_add_tail(&rq->queuelist, &hctx->dispatch); | |
1454 | spin_unlock(&hctx->lock); | |
1455 | ||
b0850297 ML |
1456 | if (run_queue) |
1457 | blk_mq_run_hw_queue(hctx, false); | |
157f377b JA |
1458 | } |
1459 | ||
bd166ef1 JA |
1460 | void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx, |
1461 | struct list_head *list) | |
320ae51f JA |
1462 | |
1463 | { | |
320ae51f JA |
1464 | /* |
1465 | * preemption doesn't flush plug list, so it's possible ctx->cpu is | |
1466 | * offline now | |
1467 | */ | |
1468 | spin_lock(&ctx->lock); | |
1469 | while (!list_empty(list)) { | |
1470 | struct request *rq; | |
1471 | ||
1472 | rq = list_first_entry(list, struct request, queuelist); | |
e57690fe | 1473 | BUG_ON(rq->mq_ctx != ctx); |
320ae51f | 1474 | list_del_init(&rq->queuelist); |
e57690fe | 1475 | __blk_mq_insert_req_list(hctx, rq, false); |
320ae51f | 1476 | } |
cfd0c552 | 1477 | blk_mq_hctx_mark_pending(hctx, ctx); |
320ae51f | 1478 | spin_unlock(&ctx->lock); |
320ae51f JA |
1479 | } |
1480 | ||
1481 | static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b) | |
1482 | { | |
1483 | struct request *rqa = container_of(a, struct request, queuelist); | |
1484 | struct request *rqb = container_of(b, struct request, queuelist); | |
1485 | ||
1486 | return !(rqa->mq_ctx < rqb->mq_ctx || | |
1487 | (rqa->mq_ctx == rqb->mq_ctx && | |
1488 | blk_rq_pos(rqa) < blk_rq_pos(rqb))); | |
1489 | } | |
1490 | ||
1491 | void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule) | |
1492 | { | |
1493 | struct blk_mq_ctx *this_ctx; | |
1494 | struct request_queue *this_q; | |
1495 | struct request *rq; | |
1496 | LIST_HEAD(list); | |
1497 | LIST_HEAD(ctx_list); | |
1498 | unsigned int depth; | |
1499 | ||
1500 | list_splice_init(&plug->mq_list, &list); | |
1501 | ||
1502 | list_sort(NULL, &list, plug_ctx_cmp); | |
1503 | ||
1504 | this_q = NULL; | |
1505 | this_ctx = NULL; | |
1506 | depth = 0; | |
1507 | ||
1508 | while (!list_empty(&list)) { | |
1509 | rq = list_entry_rq(list.next); | |
1510 | list_del_init(&rq->queuelist); | |
1511 | BUG_ON(!rq->q); | |
1512 | if (rq->mq_ctx != this_ctx) { | |
1513 | if (this_ctx) { | |
bd166ef1 JA |
1514 | trace_block_unplug(this_q, depth, from_schedule); |
1515 | blk_mq_sched_insert_requests(this_q, this_ctx, | |
1516 | &ctx_list, | |
1517 | from_schedule); | |
320ae51f JA |
1518 | } |
1519 | ||
1520 | this_ctx = rq->mq_ctx; | |
1521 | this_q = rq->q; | |
1522 | depth = 0; | |
1523 | } | |
1524 | ||
1525 | depth++; | |
1526 | list_add_tail(&rq->queuelist, &ctx_list); | |
1527 | } | |
1528 | ||
1529 | /* | |
1530 | * If 'this_ctx' is set, we know we have entries to complete | |
1531 | * on 'ctx_list'. Do those. | |
1532 | */ | |
1533 | if (this_ctx) { | |
bd166ef1 JA |
1534 | trace_block_unplug(this_q, depth, from_schedule); |
1535 | blk_mq_sched_insert_requests(this_q, this_ctx, &ctx_list, | |
1536 | from_schedule); | |
320ae51f JA |
1537 | } |
1538 | } | |
1539 | ||
1540 | static void blk_mq_bio_to_request(struct request *rq, struct bio *bio) | |
1541 | { | |
da8d7f07 | 1542 | blk_init_request_from_bio(rq, bio); |
4b570521 | 1543 | |
85acb3ba SL |
1544 | blk_rq_set_rl(rq, blk_get_rl(rq->q, bio)); |
1545 | ||
6e85eaf3 | 1546 | blk_account_io_start(rq, true); |
320ae51f JA |
1547 | } |
1548 | ||
ab42f35d ML |
1549 | static inline void blk_mq_queue_io(struct blk_mq_hw_ctx *hctx, |
1550 | struct blk_mq_ctx *ctx, | |
1551 | struct request *rq) | |
1552 | { | |
1553 | spin_lock(&ctx->lock); | |
1554 | __blk_mq_insert_request(hctx, rq, false); | |
1555 | spin_unlock(&ctx->lock); | |
07068d5b | 1556 | } |
14ec77f3 | 1557 | |
fd2d3326 JA |
1558 | static blk_qc_t request_to_qc_t(struct blk_mq_hw_ctx *hctx, struct request *rq) |
1559 | { | |
bd166ef1 JA |
1560 | if (rq->tag != -1) |
1561 | return blk_tag_to_qc_t(rq->tag, hctx->queue_num, false); | |
1562 | ||
1563 | return blk_tag_to_qc_t(rq->internal_tag, hctx->queue_num, true); | |
fd2d3326 JA |
1564 | } |
1565 | ||
d964f04a ML |
1566 | static void __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx, |
1567 | struct request *rq, | |
1568 | blk_qc_t *cookie, bool may_sleep) | |
f984df1f | 1569 | { |
f984df1f | 1570 | struct request_queue *q = rq->q; |
f984df1f SL |
1571 | struct blk_mq_queue_data bd = { |
1572 | .rq = rq, | |
d945a365 | 1573 | .last = true, |
f984df1f | 1574 | }; |
bd166ef1 | 1575 | blk_qc_t new_cookie; |
f06345ad | 1576 | blk_status_t ret; |
d964f04a ML |
1577 | bool run_queue = true; |
1578 | ||
f4560ffe ML |
1579 | /* RCU or SRCU read lock is needed before checking quiesced flag */ |
1580 | if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)) { | |
d964f04a ML |
1581 | run_queue = false; |
1582 | goto insert; | |
1583 | } | |
f984df1f | 1584 | |
bd166ef1 | 1585 | if (q->elevator) |
2253efc8 BVA |
1586 | goto insert; |
1587 | ||
d964f04a | 1588 | if (!blk_mq_get_driver_tag(rq, NULL, false)) |
bd166ef1 JA |
1589 | goto insert; |
1590 | ||
88022d72 | 1591 | if (!blk_mq_get_dispatch_budget(hctx)) { |
de148297 ML |
1592 | blk_mq_put_driver_tag(rq); |
1593 | goto insert; | |
88022d72 | 1594 | } |
de148297 | 1595 | |
bd166ef1 JA |
1596 | new_cookie = request_to_qc_t(hctx, rq); |
1597 | ||
f984df1f SL |
1598 | /* |
1599 | * For OK queue, we are done. For error, kill it. Any other | |
1600 | * error (busy), just add it to our list as we previously | |
1601 | * would have done | |
1602 | */ | |
1603 | ret = q->mq_ops->queue_rq(hctx, &bd); | |
fc17b653 CH |
1604 | switch (ret) { |
1605 | case BLK_STS_OK: | |
7b371636 | 1606 | *cookie = new_cookie; |
2253efc8 | 1607 | return; |
fc17b653 CH |
1608 | case BLK_STS_RESOURCE: |
1609 | __blk_mq_requeue_request(rq); | |
1610 | goto insert; | |
1611 | default: | |
7b371636 | 1612 | *cookie = BLK_QC_T_NONE; |
fc17b653 | 1613 | blk_mq_end_request(rq, ret); |
2253efc8 | 1614 | return; |
f984df1f | 1615 | } |
7b371636 | 1616 | |
2253efc8 | 1617 | insert: |
d964f04a | 1618 | blk_mq_sched_insert_request(rq, false, run_queue, false, may_sleep); |
f984df1f SL |
1619 | } |
1620 | ||
5eb6126e CH |
1621 | static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx, |
1622 | struct request *rq, blk_qc_t *cookie) | |
1623 | { | |
1624 | if (!(hctx->flags & BLK_MQ_F_BLOCKING)) { | |
1625 | rcu_read_lock(); | |
d964f04a | 1626 | __blk_mq_try_issue_directly(hctx, rq, cookie, false); |
5eb6126e CH |
1627 | rcu_read_unlock(); |
1628 | } else { | |
bf4907c0 JA |
1629 | unsigned int srcu_idx; |
1630 | ||
1631 | might_sleep(); | |
1632 | ||
07319678 | 1633 | srcu_idx = srcu_read_lock(hctx->queue_rq_srcu); |
d964f04a | 1634 | __blk_mq_try_issue_directly(hctx, rq, cookie, true); |
07319678 | 1635 | srcu_read_unlock(hctx->queue_rq_srcu, srcu_idx); |
5eb6126e CH |
1636 | } |
1637 | } | |
1638 | ||
dece1635 | 1639 | static blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio) |
07068d5b | 1640 | { |
ef295ecf | 1641 | const int is_sync = op_is_sync(bio->bi_opf); |
f73f44eb | 1642 | const int is_flush_fua = op_is_flush(bio->bi_opf); |
5a797e00 | 1643 | struct blk_mq_alloc_data data = { .flags = 0 }; |
07068d5b | 1644 | struct request *rq; |
5eb6126e | 1645 | unsigned int request_count = 0; |
f984df1f | 1646 | struct blk_plug *plug; |
5b3f341f | 1647 | struct request *same_queue_rq = NULL; |
7b371636 | 1648 | blk_qc_t cookie; |
87760e5e | 1649 | unsigned int wb_acct; |
07068d5b JA |
1650 | |
1651 | blk_queue_bounce(q, &bio); | |
1652 | ||
af67c31f | 1653 | blk_queue_split(q, &bio); |
f36ea50c | 1654 | |
e23947bd | 1655 | if (!bio_integrity_prep(bio)) |
dece1635 | 1656 | return BLK_QC_T_NONE; |
07068d5b | 1657 | |
87c279e6 OS |
1658 | if (!is_flush_fua && !blk_queue_nomerges(q) && |
1659 | blk_attempt_plug_merge(q, bio, &request_count, &same_queue_rq)) | |
1660 | return BLK_QC_T_NONE; | |
f984df1f | 1661 | |
bd166ef1 JA |
1662 | if (blk_mq_sched_bio_merge(q, bio)) |
1663 | return BLK_QC_T_NONE; | |
1664 | ||
87760e5e JA |
1665 | wb_acct = wbt_wait(q->rq_wb, bio, NULL); |
1666 | ||
bd166ef1 JA |
1667 | trace_block_getrq(q, bio, bio->bi_opf); |
1668 | ||
d2c0d383 | 1669 | rq = blk_mq_get_request(q, bio, bio->bi_opf, &data); |
87760e5e JA |
1670 | if (unlikely(!rq)) { |
1671 | __wbt_done(q->rq_wb, wb_acct); | |
03a07c92 GR |
1672 | if (bio->bi_opf & REQ_NOWAIT) |
1673 | bio_wouldblock_error(bio); | |
dece1635 | 1674 | return BLK_QC_T_NONE; |
87760e5e JA |
1675 | } |
1676 | ||
1677 | wbt_track(&rq->issue_stat, wb_acct); | |
07068d5b | 1678 | |
fd2d3326 | 1679 | cookie = request_to_qc_t(data.hctx, rq); |
07068d5b | 1680 | |
f984df1f | 1681 | plug = current->plug; |
07068d5b | 1682 | if (unlikely(is_flush_fua)) { |
f984df1f | 1683 | blk_mq_put_ctx(data.ctx); |
07068d5b | 1684 | blk_mq_bio_to_request(rq, bio); |
923218f6 ML |
1685 | |
1686 | /* bypass scheduler for flush rq */ | |
1687 | blk_insert_flush(rq); | |
1688 | blk_mq_run_hw_queue(data.hctx, true); | |
a4d907b6 | 1689 | } else if (plug && q->nr_hw_queues == 1) { |
600271d9 SL |
1690 | struct request *last = NULL; |
1691 | ||
b00c53e8 | 1692 | blk_mq_put_ctx(data.ctx); |
e6c4438b | 1693 | blk_mq_bio_to_request(rq, bio); |
0a6219a9 ML |
1694 | |
1695 | /* | |
1696 | * @request_count may become stale because of schedule | |
1697 | * out, so check the list again. | |
1698 | */ | |
1699 | if (list_empty(&plug->mq_list)) | |
1700 | request_count = 0; | |
254d259d CH |
1701 | else if (blk_queue_nomerges(q)) |
1702 | request_count = blk_plug_queued_count(q); | |
1703 | ||
676d0607 | 1704 | if (!request_count) |
e6c4438b | 1705 | trace_block_plug(q); |
600271d9 SL |
1706 | else |
1707 | last = list_entry_rq(plug->mq_list.prev); | |
b094f89c | 1708 | |
600271d9 SL |
1709 | if (request_count >= BLK_MAX_REQUEST_COUNT || (last && |
1710 | blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) { | |
e6c4438b JM |
1711 | blk_flush_plug_list(plug, false); |
1712 | trace_block_plug(q); | |
320ae51f | 1713 | } |
b094f89c | 1714 | |
e6c4438b | 1715 | list_add_tail(&rq->queuelist, &plug->mq_list); |
2299722c | 1716 | } else if (plug && !blk_queue_nomerges(q)) { |
bd166ef1 | 1717 | blk_mq_bio_to_request(rq, bio); |
07068d5b | 1718 | |
07068d5b | 1719 | /* |
6a83e74d | 1720 | * We do limited plugging. If the bio can be merged, do that. |
f984df1f SL |
1721 | * Otherwise the existing request in the plug list will be |
1722 | * issued. So the plug list will have one request at most | |
2299722c CH |
1723 | * The plug list might get flushed before this. If that happens, |
1724 | * the plug list is empty, and same_queue_rq is invalid. | |
07068d5b | 1725 | */ |
2299722c CH |
1726 | if (list_empty(&plug->mq_list)) |
1727 | same_queue_rq = NULL; | |
1728 | if (same_queue_rq) | |
1729 | list_del_init(&same_queue_rq->queuelist); | |
1730 | list_add_tail(&rq->queuelist, &plug->mq_list); | |
1731 | ||
bf4907c0 JA |
1732 | blk_mq_put_ctx(data.ctx); |
1733 | ||
dad7a3be ML |
1734 | if (same_queue_rq) { |
1735 | data.hctx = blk_mq_map_queue(q, | |
1736 | same_queue_rq->mq_ctx->cpu); | |
2299722c CH |
1737 | blk_mq_try_issue_directly(data.hctx, same_queue_rq, |
1738 | &cookie); | |
dad7a3be | 1739 | } |
a4d907b6 | 1740 | } else if (q->nr_hw_queues > 1 && is_sync) { |
bf4907c0 | 1741 | blk_mq_put_ctx(data.ctx); |
2299722c | 1742 | blk_mq_bio_to_request(rq, bio); |
2299722c | 1743 | blk_mq_try_issue_directly(data.hctx, rq, &cookie); |
a4d907b6 | 1744 | } else if (q->elevator) { |
b00c53e8 | 1745 | blk_mq_put_ctx(data.ctx); |
bd166ef1 | 1746 | blk_mq_bio_to_request(rq, bio); |
a4d907b6 | 1747 | blk_mq_sched_insert_request(rq, false, true, true, true); |
ab42f35d | 1748 | } else { |
b00c53e8 | 1749 | blk_mq_put_ctx(data.ctx); |
ab42f35d ML |
1750 | blk_mq_bio_to_request(rq, bio); |
1751 | blk_mq_queue_io(data.hctx, data.ctx, rq); | |
a4d907b6 | 1752 | blk_mq_run_hw_queue(data.hctx, true); |
ab42f35d | 1753 | } |
320ae51f | 1754 | |
7b371636 | 1755 | return cookie; |
320ae51f JA |
1756 | } |
1757 | ||
cc71a6f4 JA |
1758 | void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags, |
1759 | unsigned int hctx_idx) | |
95363efd | 1760 | { |
e9b267d9 | 1761 | struct page *page; |
320ae51f | 1762 | |
24d2f903 | 1763 | if (tags->rqs && set->ops->exit_request) { |
e9b267d9 | 1764 | int i; |
320ae51f | 1765 | |
24d2f903 | 1766 | for (i = 0; i < tags->nr_tags; i++) { |
2af8cbe3 JA |
1767 | struct request *rq = tags->static_rqs[i]; |
1768 | ||
1769 | if (!rq) | |
e9b267d9 | 1770 | continue; |
d6296d39 | 1771 | set->ops->exit_request(set, rq, hctx_idx); |
2af8cbe3 | 1772 | tags->static_rqs[i] = NULL; |
e9b267d9 | 1773 | } |
320ae51f | 1774 | } |
320ae51f | 1775 | |
24d2f903 CH |
1776 | while (!list_empty(&tags->page_list)) { |
1777 | page = list_first_entry(&tags->page_list, struct page, lru); | |
6753471c | 1778 | list_del_init(&page->lru); |
f75782e4 CM |
1779 | /* |
1780 | * Remove kmemleak object previously allocated in | |
1781 | * blk_mq_init_rq_map(). | |
1782 | */ | |
1783 | kmemleak_free(page_address(page)); | |
320ae51f JA |
1784 | __free_pages(page, page->private); |
1785 | } | |
cc71a6f4 | 1786 | } |
320ae51f | 1787 | |
cc71a6f4 JA |
1788 | void blk_mq_free_rq_map(struct blk_mq_tags *tags) |
1789 | { | |
24d2f903 | 1790 | kfree(tags->rqs); |
cc71a6f4 | 1791 | tags->rqs = NULL; |
2af8cbe3 JA |
1792 | kfree(tags->static_rqs); |
1793 | tags->static_rqs = NULL; | |
320ae51f | 1794 | |
24d2f903 | 1795 | blk_mq_free_tags(tags); |
320ae51f JA |
1796 | } |
1797 | ||
cc71a6f4 JA |
1798 | struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, |
1799 | unsigned int hctx_idx, | |
1800 | unsigned int nr_tags, | |
1801 | unsigned int reserved_tags) | |
320ae51f | 1802 | { |
24d2f903 | 1803 | struct blk_mq_tags *tags; |
59f082e4 | 1804 | int node; |
320ae51f | 1805 | |
59f082e4 SL |
1806 | node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx); |
1807 | if (node == NUMA_NO_NODE) | |
1808 | node = set->numa_node; | |
1809 | ||
1810 | tags = blk_mq_init_tags(nr_tags, reserved_tags, node, | |
24391c0d | 1811 | BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags)); |
24d2f903 CH |
1812 | if (!tags) |
1813 | return NULL; | |
320ae51f | 1814 | |
cc71a6f4 | 1815 | tags->rqs = kzalloc_node(nr_tags * sizeof(struct request *), |
36e1f3d1 | 1816 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, |
59f082e4 | 1817 | node); |
24d2f903 CH |
1818 | if (!tags->rqs) { |
1819 | blk_mq_free_tags(tags); | |
1820 | return NULL; | |
1821 | } | |
320ae51f | 1822 | |
2af8cbe3 JA |
1823 | tags->static_rqs = kzalloc_node(nr_tags * sizeof(struct request *), |
1824 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, | |
59f082e4 | 1825 | node); |
2af8cbe3 JA |
1826 | if (!tags->static_rqs) { |
1827 | kfree(tags->rqs); | |
1828 | blk_mq_free_tags(tags); | |
1829 | return NULL; | |
1830 | } | |
1831 | ||
cc71a6f4 JA |
1832 | return tags; |
1833 | } | |
1834 | ||
1835 | static size_t order_to_size(unsigned int order) | |
1836 | { | |
1837 | return (size_t)PAGE_SIZE << order; | |
1838 | } | |
1839 | ||
1840 | int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags, | |
1841 | unsigned int hctx_idx, unsigned int depth) | |
1842 | { | |
1843 | unsigned int i, j, entries_per_page, max_order = 4; | |
1844 | size_t rq_size, left; | |
59f082e4 SL |
1845 | int node; |
1846 | ||
1847 | node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx); | |
1848 | if (node == NUMA_NO_NODE) | |
1849 | node = set->numa_node; | |
cc71a6f4 JA |
1850 | |
1851 | INIT_LIST_HEAD(&tags->page_list); | |
1852 | ||
320ae51f JA |
1853 | /* |
1854 | * rq_size is the size of the request plus driver payload, rounded | |
1855 | * to the cacheline size | |
1856 | */ | |
24d2f903 | 1857 | rq_size = round_up(sizeof(struct request) + set->cmd_size, |
320ae51f | 1858 | cache_line_size()); |
cc71a6f4 | 1859 | left = rq_size * depth; |
320ae51f | 1860 | |
cc71a6f4 | 1861 | for (i = 0; i < depth; ) { |
320ae51f JA |
1862 | int this_order = max_order; |
1863 | struct page *page; | |
1864 | int to_do; | |
1865 | void *p; | |
1866 | ||
b3a834b1 | 1867 | while (this_order && left < order_to_size(this_order - 1)) |
320ae51f JA |
1868 | this_order--; |
1869 | ||
1870 | do { | |
59f082e4 | 1871 | page = alloc_pages_node(node, |
36e1f3d1 | 1872 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO, |
a5164405 | 1873 | this_order); |
320ae51f JA |
1874 | if (page) |
1875 | break; | |
1876 | if (!this_order--) | |
1877 | break; | |
1878 | if (order_to_size(this_order) < rq_size) | |
1879 | break; | |
1880 | } while (1); | |
1881 | ||
1882 | if (!page) | |
24d2f903 | 1883 | goto fail; |
320ae51f JA |
1884 | |
1885 | page->private = this_order; | |
24d2f903 | 1886 | list_add_tail(&page->lru, &tags->page_list); |
320ae51f JA |
1887 | |
1888 | p = page_address(page); | |
f75782e4 CM |
1889 | /* |
1890 | * Allow kmemleak to scan these pages as they contain pointers | |
1891 | * to additional allocations like via ops->init_request(). | |
1892 | */ | |
36e1f3d1 | 1893 | kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO); |
320ae51f | 1894 | entries_per_page = order_to_size(this_order) / rq_size; |
cc71a6f4 | 1895 | to_do = min(entries_per_page, depth - i); |
320ae51f JA |
1896 | left -= to_do * rq_size; |
1897 | for (j = 0; j < to_do; j++) { | |
2af8cbe3 JA |
1898 | struct request *rq = p; |
1899 | ||
1900 | tags->static_rqs[i] = rq; | |
24d2f903 | 1901 | if (set->ops->init_request) { |
d6296d39 | 1902 | if (set->ops->init_request(set, rq, hctx_idx, |
59f082e4 | 1903 | node)) { |
2af8cbe3 | 1904 | tags->static_rqs[i] = NULL; |
24d2f903 | 1905 | goto fail; |
a5164405 | 1906 | } |
e9b267d9 CH |
1907 | } |
1908 | ||
320ae51f JA |
1909 | p += rq_size; |
1910 | i++; | |
1911 | } | |
1912 | } | |
cc71a6f4 | 1913 | return 0; |
320ae51f | 1914 | |
24d2f903 | 1915 | fail: |
cc71a6f4 JA |
1916 | blk_mq_free_rqs(set, tags, hctx_idx); |
1917 | return -ENOMEM; | |
320ae51f JA |
1918 | } |
1919 | ||
e57690fe JA |
1920 | /* |
1921 | * 'cpu' is going away. splice any existing rq_list entries from this | |
1922 | * software queue to the hw queue dispatch list, and ensure that it | |
1923 | * gets run. | |
1924 | */ | |
9467f859 | 1925 | static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node) |
484b4061 | 1926 | { |
9467f859 | 1927 | struct blk_mq_hw_ctx *hctx; |
484b4061 JA |
1928 | struct blk_mq_ctx *ctx; |
1929 | LIST_HEAD(tmp); | |
1930 | ||
9467f859 | 1931 | hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead); |
e57690fe | 1932 | ctx = __blk_mq_get_ctx(hctx->queue, cpu); |
484b4061 JA |
1933 | |
1934 | spin_lock(&ctx->lock); | |
1935 | if (!list_empty(&ctx->rq_list)) { | |
1936 | list_splice_init(&ctx->rq_list, &tmp); | |
1937 | blk_mq_hctx_clear_pending(hctx, ctx); | |
1938 | } | |
1939 | spin_unlock(&ctx->lock); | |
1940 | ||
1941 | if (list_empty(&tmp)) | |
9467f859 | 1942 | return 0; |
484b4061 | 1943 | |
e57690fe JA |
1944 | spin_lock(&hctx->lock); |
1945 | list_splice_tail_init(&tmp, &hctx->dispatch); | |
1946 | spin_unlock(&hctx->lock); | |
484b4061 JA |
1947 | |
1948 | blk_mq_run_hw_queue(hctx, true); | |
9467f859 | 1949 | return 0; |
484b4061 JA |
1950 | } |
1951 | ||
9467f859 | 1952 | static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx) |
484b4061 | 1953 | { |
9467f859 TG |
1954 | cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD, |
1955 | &hctx->cpuhp_dead); | |
484b4061 JA |
1956 | } |
1957 | ||
c3b4afca | 1958 | /* hctx->ctxs will be freed in queue's release handler */ |
08e98fc6 ML |
1959 | static void blk_mq_exit_hctx(struct request_queue *q, |
1960 | struct blk_mq_tag_set *set, | |
1961 | struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx) | |
1962 | { | |
9c1051aa OS |
1963 | blk_mq_debugfs_unregister_hctx(hctx); |
1964 | ||
08e98fc6 ML |
1965 | blk_mq_tag_idle(hctx); |
1966 | ||
f70ced09 | 1967 | if (set->ops->exit_request) |
d6296d39 | 1968 | set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx); |
f70ced09 | 1969 | |
93252632 OS |
1970 | blk_mq_sched_exit_hctx(q, hctx, hctx_idx); |
1971 | ||
08e98fc6 ML |
1972 | if (set->ops->exit_hctx) |
1973 | set->ops->exit_hctx(hctx, hctx_idx); | |
1974 | ||
6a83e74d | 1975 | if (hctx->flags & BLK_MQ_F_BLOCKING) |
07319678 | 1976 | cleanup_srcu_struct(hctx->queue_rq_srcu); |
6a83e74d | 1977 | |
9467f859 | 1978 | blk_mq_remove_cpuhp(hctx); |
f70ced09 | 1979 | blk_free_flush_queue(hctx->fq); |
88459642 | 1980 | sbitmap_free(&hctx->ctx_map); |
08e98fc6 ML |
1981 | } |
1982 | ||
624dbe47 ML |
1983 | static void blk_mq_exit_hw_queues(struct request_queue *q, |
1984 | struct blk_mq_tag_set *set, int nr_queue) | |
1985 | { | |
1986 | struct blk_mq_hw_ctx *hctx; | |
1987 | unsigned int i; | |
1988 | ||
1989 | queue_for_each_hw_ctx(q, hctx, i) { | |
1990 | if (i == nr_queue) | |
1991 | break; | |
08e98fc6 | 1992 | blk_mq_exit_hctx(q, set, hctx, i); |
624dbe47 | 1993 | } |
624dbe47 ML |
1994 | } |
1995 | ||
08e98fc6 ML |
1996 | static int blk_mq_init_hctx(struct request_queue *q, |
1997 | struct blk_mq_tag_set *set, | |
1998 | struct blk_mq_hw_ctx *hctx, unsigned hctx_idx) | |
320ae51f | 1999 | { |
08e98fc6 ML |
2000 | int node; |
2001 | ||
2002 | node = hctx->numa_node; | |
2003 | if (node == NUMA_NO_NODE) | |
2004 | node = hctx->numa_node = set->numa_node; | |
2005 | ||
9f993737 | 2006 | INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn); |
08e98fc6 ML |
2007 | spin_lock_init(&hctx->lock); |
2008 | INIT_LIST_HEAD(&hctx->dispatch); | |
2009 | hctx->queue = q; | |
2404e607 | 2010 | hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED; |
08e98fc6 | 2011 | |
9467f859 | 2012 | cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead); |
08e98fc6 ML |
2013 | |
2014 | hctx->tags = set->tags[hctx_idx]; | |
320ae51f JA |
2015 | |
2016 | /* | |
08e98fc6 ML |
2017 | * Allocate space for all possible cpus to avoid allocation at |
2018 | * runtime | |
320ae51f | 2019 | */ |
08e98fc6 ML |
2020 | hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *), |
2021 | GFP_KERNEL, node); | |
2022 | if (!hctx->ctxs) | |
2023 | goto unregister_cpu_notifier; | |
320ae51f | 2024 | |
88459642 OS |
2025 | if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8), GFP_KERNEL, |
2026 | node)) | |
08e98fc6 | 2027 | goto free_ctxs; |
320ae51f | 2028 | |
08e98fc6 | 2029 | hctx->nr_ctx = 0; |
320ae51f | 2030 | |
eb619fdb JA |
2031 | init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake); |
2032 | INIT_LIST_HEAD(&hctx->dispatch_wait.entry); | |
2033 | ||
08e98fc6 ML |
2034 | if (set->ops->init_hctx && |
2035 | set->ops->init_hctx(hctx, set->driver_data, hctx_idx)) | |
2036 | goto free_bitmap; | |
320ae51f | 2037 | |
93252632 OS |
2038 | if (blk_mq_sched_init_hctx(q, hctx, hctx_idx)) |
2039 | goto exit_hctx; | |
2040 | ||
f70ced09 ML |
2041 | hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size); |
2042 | if (!hctx->fq) | |
93252632 | 2043 | goto sched_exit_hctx; |
320ae51f | 2044 | |
f70ced09 | 2045 | if (set->ops->init_request && |
d6296d39 CH |
2046 | set->ops->init_request(set, hctx->fq->flush_rq, hctx_idx, |
2047 | node)) | |
f70ced09 | 2048 | goto free_fq; |
320ae51f | 2049 | |
6a83e74d | 2050 | if (hctx->flags & BLK_MQ_F_BLOCKING) |
07319678 | 2051 | init_srcu_struct(hctx->queue_rq_srcu); |
6a83e74d | 2052 | |
9c1051aa OS |
2053 | blk_mq_debugfs_register_hctx(q, hctx); |
2054 | ||
08e98fc6 | 2055 | return 0; |
320ae51f | 2056 | |
f70ced09 ML |
2057 | free_fq: |
2058 | kfree(hctx->fq); | |
93252632 OS |
2059 | sched_exit_hctx: |
2060 | blk_mq_sched_exit_hctx(q, hctx, hctx_idx); | |
f70ced09 ML |
2061 | exit_hctx: |
2062 | if (set->ops->exit_hctx) | |
2063 | set->ops->exit_hctx(hctx, hctx_idx); | |
08e98fc6 | 2064 | free_bitmap: |
88459642 | 2065 | sbitmap_free(&hctx->ctx_map); |
08e98fc6 ML |
2066 | free_ctxs: |
2067 | kfree(hctx->ctxs); | |
2068 | unregister_cpu_notifier: | |
9467f859 | 2069 | blk_mq_remove_cpuhp(hctx); |
08e98fc6 ML |
2070 | return -1; |
2071 | } | |
320ae51f | 2072 | |
320ae51f JA |
2073 | static void blk_mq_init_cpu_queues(struct request_queue *q, |
2074 | unsigned int nr_hw_queues) | |
2075 | { | |
2076 | unsigned int i; | |
2077 | ||
2078 | for_each_possible_cpu(i) { | |
2079 | struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i); | |
2080 | struct blk_mq_hw_ctx *hctx; | |
2081 | ||
320ae51f JA |
2082 | __ctx->cpu = i; |
2083 | spin_lock_init(&__ctx->lock); | |
2084 | INIT_LIST_HEAD(&__ctx->rq_list); | |
2085 | __ctx->queue = q; | |
2086 | ||
4b855ad3 CH |
2087 | /* If the cpu isn't present, the cpu is mapped to first hctx */ |
2088 | if (!cpu_present(i)) | |
320ae51f JA |
2089 | continue; |
2090 | ||
7d7e0f90 | 2091 | hctx = blk_mq_map_queue(q, i); |
e4043dcf | 2092 | |
320ae51f JA |
2093 | /* |
2094 | * Set local node, IFF we have more than one hw queue. If | |
2095 | * not, we remain on the home node of the device | |
2096 | */ | |
2097 | if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE) | |
bffed457 | 2098 | hctx->numa_node = local_memory_node(cpu_to_node(i)); |
320ae51f JA |
2099 | } |
2100 | } | |
2101 | ||
cc71a6f4 JA |
2102 | static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, int hctx_idx) |
2103 | { | |
2104 | int ret = 0; | |
2105 | ||
2106 | set->tags[hctx_idx] = blk_mq_alloc_rq_map(set, hctx_idx, | |
2107 | set->queue_depth, set->reserved_tags); | |
2108 | if (!set->tags[hctx_idx]) | |
2109 | return false; | |
2110 | ||
2111 | ret = blk_mq_alloc_rqs(set, set->tags[hctx_idx], hctx_idx, | |
2112 | set->queue_depth); | |
2113 | if (!ret) | |
2114 | return true; | |
2115 | ||
2116 | blk_mq_free_rq_map(set->tags[hctx_idx]); | |
2117 | set->tags[hctx_idx] = NULL; | |
2118 | return false; | |
2119 | } | |
2120 | ||
2121 | static void blk_mq_free_map_and_requests(struct blk_mq_tag_set *set, | |
2122 | unsigned int hctx_idx) | |
2123 | { | |
bd166ef1 JA |
2124 | if (set->tags[hctx_idx]) { |
2125 | blk_mq_free_rqs(set, set->tags[hctx_idx], hctx_idx); | |
2126 | blk_mq_free_rq_map(set->tags[hctx_idx]); | |
2127 | set->tags[hctx_idx] = NULL; | |
2128 | } | |
cc71a6f4 JA |
2129 | } |
2130 | ||
4b855ad3 | 2131 | static void blk_mq_map_swqueue(struct request_queue *q) |
320ae51f | 2132 | { |
d1b1cea1 | 2133 | unsigned int i, hctx_idx; |
320ae51f JA |
2134 | struct blk_mq_hw_ctx *hctx; |
2135 | struct blk_mq_ctx *ctx; | |
2a34c087 | 2136 | struct blk_mq_tag_set *set = q->tag_set; |
320ae51f | 2137 | |
60de074b AM |
2138 | /* |
2139 | * Avoid others reading imcomplete hctx->cpumask through sysfs | |
2140 | */ | |
2141 | mutex_lock(&q->sysfs_lock); | |
2142 | ||
320ae51f | 2143 | queue_for_each_hw_ctx(q, hctx, i) { |
e4043dcf | 2144 | cpumask_clear(hctx->cpumask); |
320ae51f JA |
2145 | hctx->nr_ctx = 0; |
2146 | } | |
2147 | ||
2148 | /* | |
4b855ad3 CH |
2149 | * Map software to hardware queues. |
2150 | * | |
2151 | * If the cpu isn't present, the cpu is mapped to first hctx. | |
320ae51f | 2152 | */ |
4b855ad3 | 2153 | for_each_present_cpu(i) { |
d1b1cea1 GKB |
2154 | hctx_idx = q->mq_map[i]; |
2155 | /* unmapped hw queue can be remapped after CPU topo changed */ | |
cc71a6f4 JA |
2156 | if (!set->tags[hctx_idx] && |
2157 | !__blk_mq_alloc_rq_map(set, hctx_idx)) { | |
d1b1cea1 GKB |
2158 | /* |
2159 | * If tags initialization fail for some hctx, | |
2160 | * that hctx won't be brought online. In this | |
2161 | * case, remap the current ctx to hctx[0] which | |
2162 | * is guaranteed to always have tags allocated | |
2163 | */ | |
cc71a6f4 | 2164 | q->mq_map[i] = 0; |
d1b1cea1 GKB |
2165 | } |
2166 | ||
897bb0c7 | 2167 | ctx = per_cpu_ptr(q->queue_ctx, i); |
7d7e0f90 | 2168 | hctx = blk_mq_map_queue(q, i); |
868f2f0b | 2169 | |
e4043dcf | 2170 | cpumask_set_cpu(i, hctx->cpumask); |
320ae51f JA |
2171 | ctx->index_hw = hctx->nr_ctx; |
2172 | hctx->ctxs[hctx->nr_ctx++] = ctx; | |
2173 | } | |
506e931f | 2174 | |
60de074b AM |
2175 | mutex_unlock(&q->sysfs_lock); |
2176 | ||
506e931f | 2177 | queue_for_each_hw_ctx(q, hctx, i) { |
484b4061 | 2178 | /* |
a68aafa5 JA |
2179 | * If no software queues are mapped to this hardware queue, |
2180 | * disable it and free the request entries. | |
484b4061 JA |
2181 | */ |
2182 | if (!hctx->nr_ctx) { | |
d1b1cea1 GKB |
2183 | /* Never unmap queue 0. We need it as a |
2184 | * fallback in case of a new remap fails | |
2185 | * allocation | |
2186 | */ | |
cc71a6f4 JA |
2187 | if (i && set->tags[i]) |
2188 | blk_mq_free_map_and_requests(set, i); | |
2189 | ||
2a34c087 | 2190 | hctx->tags = NULL; |
484b4061 JA |
2191 | continue; |
2192 | } | |
2193 | ||
2a34c087 ML |
2194 | hctx->tags = set->tags[i]; |
2195 | WARN_ON(!hctx->tags); | |
2196 | ||
889fa31f CY |
2197 | /* |
2198 | * Set the map size to the number of mapped software queues. | |
2199 | * This is more accurate and more efficient than looping | |
2200 | * over all possibly mapped software queues. | |
2201 | */ | |
88459642 | 2202 | sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx); |
889fa31f | 2203 | |
484b4061 JA |
2204 | /* |
2205 | * Initialize batch roundrobin counts | |
2206 | */ | |
506e931f JA |
2207 | hctx->next_cpu = cpumask_first(hctx->cpumask); |
2208 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; | |
2209 | } | |
320ae51f JA |
2210 | } |
2211 | ||
8e8320c9 JA |
2212 | /* |
2213 | * Caller needs to ensure that we're either frozen/quiesced, or that | |
2214 | * the queue isn't live yet. | |
2215 | */ | |
2404e607 | 2216 | static void queue_set_hctx_shared(struct request_queue *q, bool shared) |
0d2602ca JA |
2217 | { |
2218 | struct blk_mq_hw_ctx *hctx; | |
0d2602ca JA |
2219 | int i; |
2220 | ||
2404e607 | 2221 | queue_for_each_hw_ctx(q, hctx, i) { |
8e8320c9 JA |
2222 | if (shared) { |
2223 | if (test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state)) | |
2224 | atomic_inc(&q->shared_hctx_restart); | |
2404e607 | 2225 | hctx->flags |= BLK_MQ_F_TAG_SHARED; |
8e8320c9 JA |
2226 | } else { |
2227 | if (test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state)) | |
2228 | atomic_dec(&q->shared_hctx_restart); | |
2404e607 | 2229 | hctx->flags &= ~BLK_MQ_F_TAG_SHARED; |
8e8320c9 | 2230 | } |
2404e607 JM |
2231 | } |
2232 | } | |
2233 | ||
8e8320c9 JA |
2234 | static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set, |
2235 | bool shared) | |
2404e607 JM |
2236 | { |
2237 | struct request_queue *q; | |
0d2602ca | 2238 | |
705cda97 BVA |
2239 | lockdep_assert_held(&set->tag_list_lock); |
2240 | ||
0d2602ca JA |
2241 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
2242 | blk_mq_freeze_queue(q); | |
2404e607 | 2243 | queue_set_hctx_shared(q, shared); |
0d2602ca JA |
2244 | blk_mq_unfreeze_queue(q); |
2245 | } | |
2246 | } | |
2247 | ||
2248 | static void blk_mq_del_queue_tag_set(struct request_queue *q) | |
2249 | { | |
2250 | struct blk_mq_tag_set *set = q->tag_set; | |
2251 | ||
0d2602ca | 2252 | mutex_lock(&set->tag_list_lock); |
705cda97 BVA |
2253 | list_del_rcu(&q->tag_set_list); |
2254 | INIT_LIST_HEAD(&q->tag_set_list); | |
2404e607 JM |
2255 | if (list_is_singular(&set->tag_list)) { |
2256 | /* just transitioned to unshared */ | |
2257 | set->flags &= ~BLK_MQ_F_TAG_SHARED; | |
2258 | /* update existing queue */ | |
2259 | blk_mq_update_tag_set_depth(set, false); | |
2260 | } | |
0d2602ca | 2261 | mutex_unlock(&set->tag_list_lock); |
705cda97 BVA |
2262 | |
2263 | synchronize_rcu(); | |
0d2602ca JA |
2264 | } |
2265 | ||
2266 | static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set, | |
2267 | struct request_queue *q) | |
2268 | { | |
2269 | q->tag_set = set; | |
2270 | ||
2271 | mutex_lock(&set->tag_list_lock); | |
2404e607 JM |
2272 | |
2273 | /* Check to see if we're transitioning to shared (from 1 to 2 queues). */ | |
2274 | if (!list_empty(&set->tag_list) && !(set->flags & BLK_MQ_F_TAG_SHARED)) { | |
2275 | set->flags |= BLK_MQ_F_TAG_SHARED; | |
2276 | /* update existing queue */ | |
2277 | blk_mq_update_tag_set_depth(set, true); | |
2278 | } | |
2279 | if (set->flags & BLK_MQ_F_TAG_SHARED) | |
2280 | queue_set_hctx_shared(q, true); | |
705cda97 | 2281 | list_add_tail_rcu(&q->tag_set_list, &set->tag_list); |
2404e607 | 2282 | |
0d2602ca JA |
2283 | mutex_unlock(&set->tag_list_lock); |
2284 | } | |
2285 | ||
e09aae7e ML |
2286 | /* |
2287 | * It is the actual release handler for mq, but we do it from | |
2288 | * request queue's release handler for avoiding use-after-free | |
2289 | * and headache because q->mq_kobj shouldn't have been introduced, | |
2290 | * but we can't group ctx/kctx kobj without it. | |
2291 | */ | |
2292 | void blk_mq_release(struct request_queue *q) | |
2293 | { | |
2294 | struct blk_mq_hw_ctx *hctx; | |
2295 | unsigned int i; | |
2296 | ||
2297 | /* hctx kobj stays in hctx */ | |
c3b4afca ML |
2298 | queue_for_each_hw_ctx(q, hctx, i) { |
2299 | if (!hctx) | |
2300 | continue; | |
6c8b232e | 2301 | kobject_put(&hctx->kobj); |
c3b4afca | 2302 | } |
e09aae7e | 2303 | |
a723bab3 AM |
2304 | q->mq_map = NULL; |
2305 | ||
e09aae7e ML |
2306 | kfree(q->queue_hw_ctx); |
2307 | ||
7ea5fe31 ML |
2308 | /* |
2309 | * release .mq_kobj and sw queue's kobject now because | |
2310 | * both share lifetime with request queue. | |
2311 | */ | |
2312 | blk_mq_sysfs_deinit(q); | |
2313 | ||
e09aae7e ML |
2314 | free_percpu(q->queue_ctx); |
2315 | } | |
2316 | ||
24d2f903 | 2317 | struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set) |
b62c21b7 MS |
2318 | { |
2319 | struct request_queue *uninit_q, *q; | |
2320 | ||
2321 | uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node); | |
2322 | if (!uninit_q) | |
2323 | return ERR_PTR(-ENOMEM); | |
2324 | ||
2325 | q = blk_mq_init_allocated_queue(set, uninit_q); | |
2326 | if (IS_ERR(q)) | |
2327 | blk_cleanup_queue(uninit_q); | |
2328 | ||
2329 | return q; | |
2330 | } | |
2331 | EXPORT_SYMBOL(blk_mq_init_queue); | |
2332 | ||
07319678 BVA |
2333 | static int blk_mq_hw_ctx_size(struct blk_mq_tag_set *tag_set) |
2334 | { | |
2335 | int hw_ctx_size = sizeof(struct blk_mq_hw_ctx); | |
2336 | ||
2337 | BUILD_BUG_ON(ALIGN(offsetof(struct blk_mq_hw_ctx, queue_rq_srcu), | |
2338 | __alignof__(struct blk_mq_hw_ctx)) != | |
2339 | sizeof(struct blk_mq_hw_ctx)); | |
2340 | ||
2341 | if (tag_set->flags & BLK_MQ_F_BLOCKING) | |
2342 | hw_ctx_size += sizeof(struct srcu_struct); | |
2343 | ||
2344 | return hw_ctx_size; | |
2345 | } | |
2346 | ||
868f2f0b KB |
2347 | static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set, |
2348 | struct request_queue *q) | |
320ae51f | 2349 | { |
868f2f0b KB |
2350 | int i, j; |
2351 | struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx; | |
f14bbe77 | 2352 | |
868f2f0b | 2353 | blk_mq_sysfs_unregister(q); |
24d2f903 | 2354 | for (i = 0; i < set->nr_hw_queues; i++) { |
868f2f0b | 2355 | int node; |
f14bbe77 | 2356 | |
868f2f0b KB |
2357 | if (hctxs[i]) |
2358 | continue; | |
2359 | ||
2360 | node = blk_mq_hw_queue_to_node(q->mq_map, i); | |
07319678 | 2361 | hctxs[i] = kzalloc_node(blk_mq_hw_ctx_size(set), |
cdef54dd | 2362 | GFP_KERNEL, node); |
320ae51f | 2363 | if (!hctxs[i]) |
868f2f0b | 2364 | break; |
320ae51f | 2365 | |
a86073e4 | 2366 | if (!zalloc_cpumask_var_node(&hctxs[i]->cpumask, GFP_KERNEL, |
868f2f0b KB |
2367 | node)) { |
2368 | kfree(hctxs[i]); | |
2369 | hctxs[i] = NULL; | |
2370 | break; | |
2371 | } | |
e4043dcf | 2372 | |
0d2602ca | 2373 | atomic_set(&hctxs[i]->nr_active, 0); |
f14bbe77 | 2374 | hctxs[i]->numa_node = node; |
320ae51f | 2375 | hctxs[i]->queue_num = i; |
868f2f0b KB |
2376 | |
2377 | if (blk_mq_init_hctx(q, set, hctxs[i], i)) { | |
2378 | free_cpumask_var(hctxs[i]->cpumask); | |
2379 | kfree(hctxs[i]); | |
2380 | hctxs[i] = NULL; | |
2381 | break; | |
2382 | } | |
2383 | blk_mq_hctx_kobj_init(hctxs[i]); | |
320ae51f | 2384 | } |
868f2f0b KB |
2385 | for (j = i; j < q->nr_hw_queues; j++) { |
2386 | struct blk_mq_hw_ctx *hctx = hctxs[j]; | |
2387 | ||
2388 | if (hctx) { | |
cc71a6f4 JA |
2389 | if (hctx->tags) |
2390 | blk_mq_free_map_and_requests(set, j); | |
868f2f0b | 2391 | blk_mq_exit_hctx(q, set, hctx, j); |
868f2f0b | 2392 | kobject_put(&hctx->kobj); |
868f2f0b KB |
2393 | hctxs[j] = NULL; |
2394 | ||
2395 | } | |
2396 | } | |
2397 | q->nr_hw_queues = i; | |
2398 | blk_mq_sysfs_register(q); | |
2399 | } | |
2400 | ||
2401 | struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set, | |
2402 | struct request_queue *q) | |
2403 | { | |
66841672 ML |
2404 | /* mark the queue as mq asap */ |
2405 | q->mq_ops = set->ops; | |
2406 | ||
34dbad5d | 2407 | q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn, |
720b8ccc SB |
2408 | blk_mq_poll_stats_bkt, |
2409 | BLK_MQ_POLL_STATS_BKTS, q); | |
34dbad5d OS |
2410 | if (!q->poll_cb) |
2411 | goto err_exit; | |
2412 | ||
868f2f0b KB |
2413 | q->queue_ctx = alloc_percpu(struct blk_mq_ctx); |
2414 | if (!q->queue_ctx) | |
c7de5726 | 2415 | goto err_exit; |
868f2f0b | 2416 | |
737f98cf ML |
2417 | /* init q->mq_kobj and sw queues' kobjects */ |
2418 | blk_mq_sysfs_init(q); | |
2419 | ||
868f2f0b KB |
2420 | q->queue_hw_ctx = kzalloc_node(nr_cpu_ids * sizeof(*(q->queue_hw_ctx)), |
2421 | GFP_KERNEL, set->numa_node); | |
2422 | if (!q->queue_hw_ctx) | |
2423 | goto err_percpu; | |
2424 | ||
bdd17e75 | 2425 | q->mq_map = set->mq_map; |
868f2f0b KB |
2426 | |
2427 | blk_mq_realloc_hw_ctxs(set, q); | |
2428 | if (!q->nr_hw_queues) | |
2429 | goto err_hctxs; | |
320ae51f | 2430 | |
287922eb | 2431 | INIT_WORK(&q->timeout_work, blk_mq_timeout_work); |
e56f698b | 2432 | blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ); |
320ae51f JA |
2433 | |
2434 | q->nr_queues = nr_cpu_ids; | |
320ae51f | 2435 | |
94eddfbe | 2436 | q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT; |
320ae51f | 2437 | |
05f1dd53 JA |
2438 | if (!(set->flags & BLK_MQ_F_SG_MERGE)) |
2439 | q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE; | |
2440 | ||
1be036e9 CH |
2441 | q->sg_reserved_size = INT_MAX; |
2442 | ||
2849450a | 2443 | INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work); |
6fca6a61 CH |
2444 | INIT_LIST_HEAD(&q->requeue_list); |
2445 | spin_lock_init(&q->requeue_lock); | |
2446 | ||
254d259d | 2447 | blk_queue_make_request(q, blk_mq_make_request); |
ea435e1b CH |
2448 | if (q->mq_ops->poll) |
2449 | q->poll_fn = blk_mq_poll; | |
07068d5b | 2450 | |
eba71768 JA |
2451 | /* |
2452 | * Do this after blk_queue_make_request() overrides it... | |
2453 | */ | |
2454 | q->nr_requests = set->queue_depth; | |
2455 | ||
64f1c21e JA |
2456 | /* |
2457 | * Default to classic polling | |
2458 | */ | |
2459 | q->poll_nsec = -1; | |
2460 | ||
24d2f903 CH |
2461 | if (set->ops->complete) |
2462 | blk_queue_softirq_done(q, set->ops->complete); | |
30a91cb4 | 2463 | |
24d2f903 | 2464 | blk_mq_init_cpu_queues(q, set->nr_hw_queues); |
0d2602ca | 2465 | blk_mq_add_queue_tag_set(set, q); |
4b855ad3 | 2466 | blk_mq_map_swqueue(q); |
4593fdbe | 2467 | |
d3484991 JA |
2468 | if (!(set->flags & BLK_MQ_F_NO_SCHED)) { |
2469 | int ret; | |
2470 | ||
2471 | ret = blk_mq_sched_init(q); | |
2472 | if (ret) | |
2473 | return ERR_PTR(ret); | |
2474 | } | |
2475 | ||
320ae51f | 2476 | return q; |
18741986 | 2477 | |
320ae51f | 2478 | err_hctxs: |
868f2f0b | 2479 | kfree(q->queue_hw_ctx); |
320ae51f | 2480 | err_percpu: |
868f2f0b | 2481 | free_percpu(q->queue_ctx); |
c7de5726 ML |
2482 | err_exit: |
2483 | q->mq_ops = NULL; | |
320ae51f JA |
2484 | return ERR_PTR(-ENOMEM); |
2485 | } | |
b62c21b7 | 2486 | EXPORT_SYMBOL(blk_mq_init_allocated_queue); |
320ae51f JA |
2487 | |
2488 | void blk_mq_free_queue(struct request_queue *q) | |
2489 | { | |
624dbe47 | 2490 | struct blk_mq_tag_set *set = q->tag_set; |
320ae51f | 2491 | |
0d2602ca | 2492 | blk_mq_del_queue_tag_set(q); |
624dbe47 | 2493 | blk_mq_exit_hw_queues(q, set, set->nr_hw_queues); |
320ae51f | 2494 | } |
320ae51f JA |
2495 | |
2496 | /* Basically redo blk_mq_init_queue with queue frozen */ | |
4b855ad3 | 2497 | static void blk_mq_queue_reinit(struct request_queue *q) |
320ae51f | 2498 | { |
4ecd4fef | 2499 | WARN_ON_ONCE(!atomic_read(&q->mq_freeze_depth)); |
320ae51f | 2500 | |
9c1051aa | 2501 | blk_mq_debugfs_unregister_hctxs(q); |
67aec14c JA |
2502 | blk_mq_sysfs_unregister(q); |
2503 | ||
320ae51f JA |
2504 | /* |
2505 | * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe | |
2506 | * we should change hctx numa_node according to new topology (this | |
2507 | * involves free and re-allocate memory, worthy doing?) | |
2508 | */ | |
2509 | ||
4b855ad3 | 2510 | blk_mq_map_swqueue(q); |
320ae51f | 2511 | |
67aec14c | 2512 | blk_mq_sysfs_register(q); |
9c1051aa | 2513 | blk_mq_debugfs_register_hctxs(q); |
320ae51f JA |
2514 | } |
2515 | ||
a5164405 JA |
2516 | static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set) |
2517 | { | |
2518 | int i; | |
2519 | ||
cc71a6f4 JA |
2520 | for (i = 0; i < set->nr_hw_queues; i++) |
2521 | if (!__blk_mq_alloc_rq_map(set, i)) | |
a5164405 | 2522 | goto out_unwind; |
a5164405 JA |
2523 | |
2524 | return 0; | |
2525 | ||
2526 | out_unwind: | |
2527 | while (--i >= 0) | |
cc71a6f4 | 2528 | blk_mq_free_rq_map(set->tags[i]); |
a5164405 | 2529 | |
a5164405 JA |
2530 | return -ENOMEM; |
2531 | } | |
2532 | ||
2533 | /* | |
2534 | * Allocate the request maps associated with this tag_set. Note that this | |
2535 | * may reduce the depth asked for, if memory is tight. set->queue_depth | |
2536 | * will be updated to reflect the allocated depth. | |
2537 | */ | |
2538 | static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set) | |
2539 | { | |
2540 | unsigned int depth; | |
2541 | int err; | |
2542 | ||
2543 | depth = set->queue_depth; | |
2544 | do { | |
2545 | err = __blk_mq_alloc_rq_maps(set); | |
2546 | if (!err) | |
2547 | break; | |
2548 | ||
2549 | set->queue_depth >>= 1; | |
2550 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) { | |
2551 | err = -ENOMEM; | |
2552 | break; | |
2553 | } | |
2554 | } while (set->queue_depth); | |
2555 | ||
2556 | if (!set->queue_depth || err) { | |
2557 | pr_err("blk-mq: failed to allocate request map\n"); | |
2558 | return -ENOMEM; | |
2559 | } | |
2560 | ||
2561 | if (depth != set->queue_depth) | |
2562 | pr_info("blk-mq: reduced tag depth (%u -> %u)\n", | |
2563 | depth, set->queue_depth); | |
2564 | ||
2565 | return 0; | |
2566 | } | |
2567 | ||
ebe8bddb OS |
2568 | static int blk_mq_update_queue_map(struct blk_mq_tag_set *set) |
2569 | { | |
2570 | if (set->ops->map_queues) | |
2571 | return set->ops->map_queues(set); | |
2572 | else | |
2573 | return blk_mq_map_queues(set); | |
2574 | } | |
2575 | ||
a4391c64 JA |
2576 | /* |
2577 | * Alloc a tag set to be associated with one or more request queues. | |
2578 | * May fail with EINVAL for various error conditions. May adjust the | |
2579 | * requested depth down, if if it too large. In that case, the set | |
2580 | * value will be stored in set->queue_depth. | |
2581 | */ | |
24d2f903 CH |
2582 | int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set) |
2583 | { | |
da695ba2 CH |
2584 | int ret; |
2585 | ||
205fb5f5 BVA |
2586 | BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS); |
2587 | ||
24d2f903 CH |
2588 | if (!set->nr_hw_queues) |
2589 | return -EINVAL; | |
a4391c64 | 2590 | if (!set->queue_depth) |
24d2f903 CH |
2591 | return -EINVAL; |
2592 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) | |
2593 | return -EINVAL; | |
2594 | ||
7d7e0f90 | 2595 | if (!set->ops->queue_rq) |
24d2f903 CH |
2596 | return -EINVAL; |
2597 | ||
de148297 ML |
2598 | if (!set->ops->get_budget ^ !set->ops->put_budget) |
2599 | return -EINVAL; | |
2600 | ||
a4391c64 JA |
2601 | if (set->queue_depth > BLK_MQ_MAX_DEPTH) { |
2602 | pr_info("blk-mq: reduced tag depth to %u\n", | |
2603 | BLK_MQ_MAX_DEPTH); | |
2604 | set->queue_depth = BLK_MQ_MAX_DEPTH; | |
2605 | } | |
24d2f903 | 2606 | |
6637fadf SL |
2607 | /* |
2608 | * If a crashdump is active, then we are potentially in a very | |
2609 | * memory constrained environment. Limit us to 1 queue and | |
2610 | * 64 tags to prevent using too much memory. | |
2611 | */ | |
2612 | if (is_kdump_kernel()) { | |
2613 | set->nr_hw_queues = 1; | |
2614 | set->queue_depth = min(64U, set->queue_depth); | |
2615 | } | |
868f2f0b KB |
2616 | /* |
2617 | * There is no use for more h/w queues than cpus. | |
2618 | */ | |
2619 | if (set->nr_hw_queues > nr_cpu_ids) | |
2620 | set->nr_hw_queues = nr_cpu_ids; | |
6637fadf | 2621 | |
868f2f0b | 2622 | set->tags = kzalloc_node(nr_cpu_ids * sizeof(struct blk_mq_tags *), |
24d2f903 CH |
2623 | GFP_KERNEL, set->numa_node); |
2624 | if (!set->tags) | |
a5164405 | 2625 | return -ENOMEM; |
24d2f903 | 2626 | |
da695ba2 CH |
2627 | ret = -ENOMEM; |
2628 | set->mq_map = kzalloc_node(sizeof(*set->mq_map) * nr_cpu_ids, | |
2629 | GFP_KERNEL, set->numa_node); | |
bdd17e75 CH |
2630 | if (!set->mq_map) |
2631 | goto out_free_tags; | |
2632 | ||
ebe8bddb | 2633 | ret = blk_mq_update_queue_map(set); |
da695ba2 CH |
2634 | if (ret) |
2635 | goto out_free_mq_map; | |
2636 | ||
2637 | ret = blk_mq_alloc_rq_maps(set); | |
2638 | if (ret) | |
bdd17e75 | 2639 | goto out_free_mq_map; |
24d2f903 | 2640 | |
0d2602ca JA |
2641 | mutex_init(&set->tag_list_lock); |
2642 | INIT_LIST_HEAD(&set->tag_list); | |
2643 | ||
24d2f903 | 2644 | return 0; |
bdd17e75 CH |
2645 | |
2646 | out_free_mq_map: | |
2647 | kfree(set->mq_map); | |
2648 | set->mq_map = NULL; | |
2649 | out_free_tags: | |
5676e7b6 RE |
2650 | kfree(set->tags); |
2651 | set->tags = NULL; | |
da695ba2 | 2652 | return ret; |
24d2f903 CH |
2653 | } |
2654 | EXPORT_SYMBOL(blk_mq_alloc_tag_set); | |
2655 | ||
2656 | void blk_mq_free_tag_set(struct blk_mq_tag_set *set) | |
2657 | { | |
2658 | int i; | |
2659 | ||
cc71a6f4 JA |
2660 | for (i = 0; i < nr_cpu_ids; i++) |
2661 | blk_mq_free_map_and_requests(set, i); | |
484b4061 | 2662 | |
bdd17e75 CH |
2663 | kfree(set->mq_map); |
2664 | set->mq_map = NULL; | |
2665 | ||
981bd189 | 2666 | kfree(set->tags); |
5676e7b6 | 2667 | set->tags = NULL; |
24d2f903 CH |
2668 | } |
2669 | EXPORT_SYMBOL(blk_mq_free_tag_set); | |
2670 | ||
e3a2b3f9 JA |
2671 | int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr) |
2672 | { | |
2673 | struct blk_mq_tag_set *set = q->tag_set; | |
2674 | struct blk_mq_hw_ctx *hctx; | |
2675 | int i, ret; | |
2676 | ||
bd166ef1 | 2677 | if (!set) |
e3a2b3f9 JA |
2678 | return -EINVAL; |
2679 | ||
70f36b60 | 2680 | blk_mq_freeze_queue(q); |
70f36b60 | 2681 | |
e3a2b3f9 JA |
2682 | ret = 0; |
2683 | queue_for_each_hw_ctx(q, hctx, i) { | |
e9137d4b KB |
2684 | if (!hctx->tags) |
2685 | continue; | |
bd166ef1 JA |
2686 | /* |
2687 | * If we're using an MQ scheduler, just update the scheduler | |
2688 | * queue depth. This is similar to what the old code would do. | |
2689 | */ | |
70f36b60 | 2690 | if (!hctx->sched_tags) { |
c2e82a23 | 2691 | ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr, |
70f36b60 JA |
2692 | false); |
2693 | } else { | |
2694 | ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags, | |
2695 | nr, true); | |
2696 | } | |
e3a2b3f9 JA |
2697 | if (ret) |
2698 | break; | |
2699 | } | |
2700 | ||
2701 | if (!ret) | |
2702 | q->nr_requests = nr; | |
2703 | ||
70f36b60 | 2704 | blk_mq_unfreeze_queue(q); |
70f36b60 | 2705 | |
e3a2b3f9 JA |
2706 | return ret; |
2707 | } | |
2708 | ||
e4dc2b32 KB |
2709 | static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, |
2710 | int nr_hw_queues) | |
868f2f0b KB |
2711 | { |
2712 | struct request_queue *q; | |
2713 | ||
705cda97 BVA |
2714 | lockdep_assert_held(&set->tag_list_lock); |
2715 | ||
868f2f0b KB |
2716 | if (nr_hw_queues > nr_cpu_ids) |
2717 | nr_hw_queues = nr_cpu_ids; | |
2718 | if (nr_hw_queues < 1 || nr_hw_queues == set->nr_hw_queues) | |
2719 | return; | |
2720 | ||
2721 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
2722 | blk_mq_freeze_queue(q); | |
2723 | ||
2724 | set->nr_hw_queues = nr_hw_queues; | |
ebe8bddb | 2725 | blk_mq_update_queue_map(set); |
868f2f0b KB |
2726 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
2727 | blk_mq_realloc_hw_ctxs(set, q); | |
4b855ad3 | 2728 | blk_mq_queue_reinit(q); |
868f2f0b KB |
2729 | } |
2730 | ||
2731 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
2732 | blk_mq_unfreeze_queue(q); | |
2733 | } | |
e4dc2b32 KB |
2734 | |
2735 | void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues) | |
2736 | { | |
2737 | mutex_lock(&set->tag_list_lock); | |
2738 | __blk_mq_update_nr_hw_queues(set, nr_hw_queues); | |
2739 | mutex_unlock(&set->tag_list_lock); | |
2740 | } | |
868f2f0b KB |
2741 | EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues); |
2742 | ||
34dbad5d OS |
2743 | /* Enable polling stats and return whether they were already enabled. */ |
2744 | static bool blk_poll_stats_enable(struct request_queue *q) | |
2745 | { | |
2746 | if (test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) || | |
2747 | test_and_set_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags)) | |
2748 | return true; | |
2749 | blk_stat_add_callback(q, q->poll_cb); | |
2750 | return false; | |
2751 | } | |
2752 | ||
2753 | static void blk_mq_poll_stats_start(struct request_queue *q) | |
2754 | { | |
2755 | /* | |
2756 | * We don't arm the callback if polling stats are not enabled or the | |
2757 | * callback is already active. | |
2758 | */ | |
2759 | if (!test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) || | |
2760 | blk_stat_is_active(q->poll_cb)) | |
2761 | return; | |
2762 | ||
2763 | blk_stat_activate_msecs(q->poll_cb, 100); | |
2764 | } | |
2765 | ||
2766 | static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb) | |
2767 | { | |
2768 | struct request_queue *q = cb->data; | |
720b8ccc | 2769 | int bucket; |
34dbad5d | 2770 | |
720b8ccc SB |
2771 | for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) { |
2772 | if (cb->stat[bucket].nr_samples) | |
2773 | q->poll_stat[bucket] = cb->stat[bucket]; | |
2774 | } | |
34dbad5d OS |
2775 | } |
2776 | ||
64f1c21e JA |
2777 | static unsigned long blk_mq_poll_nsecs(struct request_queue *q, |
2778 | struct blk_mq_hw_ctx *hctx, | |
2779 | struct request *rq) | |
2780 | { | |
64f1c21e | 2781 | unsigned long ret = 0; |
720b8ccc | 2782 | int bucket; |
64f1c21e JA |
2783 | |
2784 | /* | |
2785 | * If stats collection isn't on, don't sleep but turn it on for | |
2786 | * future users | |
2787 | */ | |
34dbad5d | 2788 | if (!blk_poll_stats_enable(q)) |
64f1c21e JA |
2789 | return 0; |
2790 | ||
64f1c21e JA |
2791 | /* |
2792 | * As an optimistic guess, use half of the mean service time | |
2793 | * for this type of request. We can (and should) make this smarter. | |
2794 | * For instance, if the completion latencies are tight, we can | |
2795 | * get closer than just half the mean. This is especially | |
2796 | * important on devices where the completion latencies are longer | |
720b8ccc SB |
2797 | * than ~10 usec. We do use the stats for the relevant IO size |
2798 | * if available which does lead to better estimates. | |
64f1c21e | 2799 | */ |
720b8ccc SB |
2800 | bucket = blk_mq_poll_stats_bkt(rq); |
2801 | if (bucket < 0) | |
2802 | return ret; | |
2803 | ||
2804 | if (q->poll_stat[bucket].nr_samples) | |
2805 | ret = (q->poll_stat[bucket].mean + 1) / 2; | |
64f1c21e JA |
2806 | |
2807 | return ret; | |
2808 | } | |
2809 | ||
06426adf | 2810 | static bool blk_mq_poll_hybrid_sleep(struct request_queue *q, |
64f1c21e | 2811 | struct blk_mq_hw_ctx *hctx, |
06426adf JA |
2812 | struct request *rq) |
2813 | { | |
2814 | struct hrtimer_sleeper hs; | |
2815 | enum hrtimer_mode mode; | |
64f1c21e | 2816 | unsigned int nsecs; |
06426adf JA |
2817 | ktime_t kt; |
2818 | ||
64f1c21e JA |
2819 | if (test_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags)) |
2820 | return false; | |
2821 | ||
2822 | /* | |
2823 | * poll_nsec can be: | |
2824 | * | |
2825 | * -1: don't ever hybrid sleep | |
2826 | * 0: use half of prev avg | |
2827 | * >0: use this specific value | |
2828 | */ | |
2829 | if (q->poll_nsec == -1) | |
2830 | return false; | |
2831 | else if (q->poll_nsec > 0) | |
2832 | nsecs = q->poll_nsec; | |
2833 | else | |
2834 | nsecs = blk_mq_poll_nsecs(q, hctx, rq); | |
2835 | ||
2836 | if (!nsecs) | |
06426adf JA |
2837 | return false; |
2838 | ||
2839 | set_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags); | |
2840 | ||
2841 | /* | |
2842 | * This will be replaced with the stats tracking code, using | |
2843 | * 'avg_completion_time / 2' as the pre-sleep target. | |
2844 | */ | |
8b0e1953 | 2845 | kt = nsecs; |
06426adf JA |
2846 | |
2847 | mode = HRTIMER_MODE_REL; | |
2848 | hrtimer_init_on_stack(&hs.timer, CLOCK_MONOTONIC, mode); | |
2849 | hrtimer_set_expires(&hs.timer, kt); | |
2850 | ||
2851 | hrtimer_init_sleeper(&hs, current); | |
2852 | do { | |
2853 | if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags)) | |
2854 | break; | |
2855 | set_current_state(TASK_UNINTERRUPTIBLE); | |
2856 | hrtimer_start_expires(&hs.timer, mode); | |
2857 | if (hs.task) | |
2858 | io_schedule(); | |
2859 | hrtimer_cancel(&hs.timer); | |
2860 | mode = HRTIMER_MODE_ABS; | |
2861 | } while (hs.task && !signal_pending(current)); | |
2862 | ||
2863 | __set_current_state(TASK_RUNNING); | |
2864 | destroy_hrtimer_on_stack(&hs.timer); | |
2865 | return true; | |
2866 | } | |
2867 | ||
bbd7bb70 JA |
2868 | static bool __blk_mq_poll(struct blk_mq_hw_ctx *hctx, struct request *rq) |
2869 | { | |
2870 | struct request_queue *q = hctx->queue; | |
2871 | long state; | |
2872 | ||
06426adf JA |
2873 | /* |
2874 | * If we sleep, have the caller restart the poll loop to reset | |
2875 | * the state. Like for the other success return cases, the | |
2876 | * caller is responsible for checking if the IO completed. If | |
2877 | * the IO isn't complete, we'll get called again and will go | |
2878 | * straight to the busy poll loop. | |
2879 | */ | |
64f1c21e | 2880 | if (blk_mq_poll_hybrid_sleep(q, hctx, rq)) |
06426adf JA |
2881 | return true; |
2882 | ||
bbd7bb70 JA |
2883 | hctx->poll_considered++; |
2884 | ||
2885 | state = current->state; | |
2886 | while (!need_resched()) { | |
2887 | int ret; | |
2888 | ||
2889 | hctx->poll_invoked++; | |
2890 | ||
2891 | ret = q->mq_ops->poll(hctx, rq->tag); | |
2892 | if (ret > 0) { | |
2893 | hctx->poll_success++; | |
2894 | set_current_state(TASK_RUNNING); | |
2895 | return true; | |
2896 | } | |
2897 | ||
2898 | if (signal_pending_state(state, current)) | |
2899 | set_current_state(TASK_RUNNING); | |
2900 | ||
2901 | if (current->state == TASK_RUNNING) | |
2902 | return true; | |
2903 | if (ret < 0) | |
2904 | break; | |
2905 | cpu_relax(); | |
2906 | } | |
2907 | ||
2908 | return false; | |
2909 | } | |
2910 | ||
ea435e1b | 2911 | static bool blk_mq_poll(struct request_queue *q, blk_qc_t cookie) |
bbd7bb70 JA |
2912 | { |
2913 | struct blk_mq_hw_ctx *hctx; | |
bbd7bb70 JA |
2914 | struct request *rq; |
2915 | ||
ea435e1b | 2916 | if (!test_bit(QUEUE_FLAG_POLL, &q->queue_flags)) |
bbd7bb70 JA |
2917 | return false; |
2918 | ||
bbd7bb70 | 2919 | hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)]; |
bd166ef1 JA |
2920 | if (!blk_qc_t_is_internal(cookie)) |
2921 | rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie)); | |
3a07bb1d | 2922 | else { |
bd166ef1 | 2923 | rq = blk_mq_tag_to_rq(hctx->sched_tags, blk_qc_t_to_tag(cookie)); |
3a07bb1d JA |
2924 | /* |
2925 | * With scheduling, if the request has completed, we'll | |
2926 | * get a NULL return here, as we clear the sched tag when | |
2927 | * that happens. The request still remains valid, like always, | |
2928 | * so we should be safe with just the NULL check. | |
2929 | */ | |
2930 | if (!rq) | |
2931 | return false; | |
2932 | } | |
bbd7bb70 JA |
2933 | |
2934 | return __blk_mq_poll(hctx, rq); | |
2935 | } | |
bbd7bb70 | 2936 | |
320ae51f JA |
2937 | static int __init blk_mq_init(void) |
2938 | { | |
fc13457f JA |
2939 | /* |
2940 | * See comment in block/blk.h rq_atomic_flags enum | |
2941 | */ | |
2942 | BUILD_BUG_ON((REQ_ATOM_STARTED / BITS_PER_BYTE) != | |
2943 | (REQ_ATOM_COMPLETE / BITS_PER_BYTE)); | |
2944 | ||
9467f859 TG |
2945 | cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL, |
2946 | blk_mq_hctx_notify_dead); | |
320ae51f JA |
2947 | return 0; |
2948 | } | |
2949 | subsys_initcall(blk_mq_init); |