]>
Commit | Line | Data |
---|---|---|
75bb4625 JA |
1 | /* |
2 | * Block multiqueue core code | |
3 | * | |
4 | * Copyright (C) 2013-2014 Jens Axboe | |
5 | * Copyright (C) 2013-2014 Christoph Hellwig | |
6 | */ | |
320ae51f JA |
7 | #include <linux/kernel.h> |
8 | #include <linux/module.h> | |
9 | #include <linux/backing-dev.h> | |
10 | #include <linux/bio.h> | |
11 | #include <linux/blkdev.h> | |
f75782e4 | 12 | #include <linux/kmemleak.h> |
320ae51f JA |
13 | #include <linux/mm.h> |
14 | #include <linux/init.h> | |
15 | #include <linux/slab.h> | |
16 | #include <linux/workqueue.h> | |
17 | #include <linux/smp.h> | |
18 | #include <linux/llist.h> | |
19 | #include <linux/list_sort.h> | |
20 | #include <linux/cpu.h> | |
21 | #include <linux/cache.h> | |
22 | #include <linux/sched/sysctl.h> | |
105ab3d8 | 23 | #include <linux/sched/topology.h> |
174cd4b1 | 24 | #include <linux/sched/signal.h> |
320ae51f | 25 | #include <linux/delay.h> |
aedcd72f | 26 | #include <linux/crash_dump.h> |
88c7b2b7 | 27 | #include <linux/prefetch.h> |
320ae51f JA |
28 | |
29 | #include <trace/events/block.h> | |
30 | ||
31 | #include <linux/blk-mq.h> | |
32 | #include "blk.h" | |
33 | #include "blk-mq.h" | |
9c1051aa | 34 | #include "blk-mq-debugfs.h" |
320ae51f | 35 | #include "blk-mq-tag.h" |
cf43e6be | 36 | #include "blk-stat.h" |
87760e5e | 37 | #include "blk-wbt.h" |
bd166ef1 | 38 | #include "blk-mq-sched.h" |
320ae51f | 39 | |
ea435e1b | 40 | static bool blk_mq_poll(struct request_queue *q, blk_qc_t cookie); |
34dbad5d OS |
41 | static void blk_mq_poll_stats_start(struct request_queue *q); |
42 | static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb); | |
43 | ||
720b8ccc SB |
44 | static int blk_mq_poll_stats_bkt(const struct request *rq) |
45 | { | |
46 | int ddir, bytes, bucket; | |
47 | ||
99c749a4 | 48 | ddir = rq_data_dir(rq); |
720b8ccc SB |
49 | bytes = blk_rq_bytes(rq); |
50 | ||
51 | bucket = ddir + 2*(ilog2(bytes) - 9); | |
52 | ||
53 | if (bucket < 0) | |
54 | return -1; | |
55 | else if (bucket >= BLK_MQ_POLL_STATS_BKTS) | |
56 | return ddir + BLK_MQ_POLL_STATS_BKTS - 2; | |
57 | ||
58 | return bucket; | |
59 | } | |
60 | ||
320ae51f JA |
61 | /* |
62 | * Check if any of the ctx's have pending work in this hardware queue | |
63 | */ | |
79f720a7 | 64 | static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx) |
320ae51f | 65 | { |
79f720a7 JA |
66 | return !list_empty_careful(&hctx->dispatch) || |
67 | sbitmap_any_bit_set(&hctx->ctx_map) || | |
bd166ef1 | 68 | blk_mq_sched_has_work(hctx); |
1429d7c9 JA |
69 | } |
70 | ||
320ae51f JA |
71 | /* |
72 | * Mark this ctx as having pending work in this hardware queue | |
73 | */ | |
74 | static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx, | |
75 | struct blk_mq_ctx *ctx) | |
76 | { | |
88459642 OS |
77 | if (!sbitmap_test_bit(&hctx->ctx_map, ctx->index_hw)) |
78 | sbitmap_set_bit(&hctx->ctx_map, ctx->index_hw); | |
1429d7c9 JA |
79 | } |
80 | ||
81 | static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx, | |
82 | struct blk_mq_ctx *ctx) | |
83 | { | |
88459642 | 84 | sbitmap_clear_bit(&hctx->ctx_map, ctx->index_hw); |
320ae51f JA |
85 | } |
86 | ||
f299b7c7 JA |
87 | struct mq_inflight { |
88 | struct hd_struct *part; | |
89 | unsigned int *inflight; | |
90 | }; | |
91 | ||
92 | static void blk_mq_check_inflight(struct blk_mq_hw_ctx *hctx, | |
93 | struct request *rq, void *priv, | |
94 | bool reserved) | |
95 | { | |
96 | struct mq_inflight *mi = priv; | |
97 | ||
67818d25 | 98 | if (blk_mq_rq_state(rq) == MQ_RQ_IN_FLIGHT) { |
f299b7c7 | 99 | /* |
b8d62b3a JA |
100 | * index[0] counts the specific partition that was asked |
101 | * for. index[1] counts the ones that are active on the | |
102 | * whole device, so increment that if mi->part is indeed | |
103 | * a partition, and not a whole device. | |
f299b7c7 | 104 | */ |
b8d62b3a | 105 | if (rq->part == mi->part) |
f299b7c7 | 106 | mi->inflight[0]++; |
b8d62b3a JA |
107 | if (mi->part->partno) |
108 | mi->inflight[1]++; | |
f299b7c7 JA |
109 | } |
110 | } | |
111 | ||
112 | void blk_mq_in_flight(struct request_queue *q, struct hd_struct *part, | |
113 | unsigned int inflight[2]) | |
114 | { | |
115 | struct mq_inflight mi = { .part = part, .inflight = inflight, }; | |
116 | ||
b8d62b3a | 117 | inflight[0] = inflight[1] = 0; |
f299b7c7 JA |
118 | blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi); |
119 | } | |
120 | ||
1671d522 | 121 | void blk_freeze_queue_start(struct request_queue *q) |
43a5e4e2 | 122 | { |
4ecd4fef | 123 | int freeze_depth; |
cddd5d17 | 124 | |
4ecd4fef CH |
125 | freeze_depth = atomic_inc_return(&q->mq_freeze_depth); |
126 | if (freeze_depth == 1) { | |
3ef28e83 | 127 | percpu_ref_kill(&q->q_usage_counter); |
055f6e18 ML |
128 | if (q->mq_ops) |
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) | |
05707b64 | 222 | synchronize_srcu(hctx->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); | |
259 | } | |
260 | ||
320ae51f JA |
261 | bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx) |
262 | { | |
263 | return blk_mq_has_free_tags(hctx->tags); | |
264 | } | |
265 | EXPORT_SYMBOL(blk_mq_can_queue); | |
266 | ||
e4cdf1a1 CH |
267 | static struct request *blk_mq_rq_ctx_init(struct blk_mq_alloc_data *data, |
268 | unsigned int tag, unsigned int op) | |
320ae51f | 269 | { |
e4cdf1a1 CH |
270 | struct blk_mq_tags *tags = blk_mq_tags_from_data(data); |
271 | struct request *rq = tags->static_rqs[tag]; | |
272 | ||
273 | if (data->flags & BLK_MQ_REQ_INTERNAL) { | |
274 | rq->tag = -1; | |
275 | rq->internal_tag = tag; | |
276 | } else { | |
277 | if (blk_mq_tag_busy(data->hctx)) { | |
278 | rq->rq_flags = RQF_MQ_INFLIGHT; | |
279 | atomic_inc(&data->hctx->nr_active); | |
280 | } | |
281 | rq->tag = tag; | |
282 | rq->internal_tag = -1; | |
283 | data->hctx->tags->rqs[rq->tag] = rq; | |
284 | } | |
285 | ||
af76e555 | 286 | /* csd/requeue_work/fifo_time is initialized before use */ |
e4cdf1a1 CH |
287 | rq->q = data->q; |
288 | rq->mq_ctx = data->ctx; | |
7c3fb70f JA |
289 | rq->rq_flags = 0; |
290 | rq->cpu = -1; | |
ef295ecf | 291 | rq->cmd_flags = op; |
1b6d65a0 BVA |
292 | if (data->flags & BLK_MQ_REQ_PREEMPT) |
293 | rq->rq_flags |= RQF_PREEMPT; | |
e4cdf1a1 | 294 | if (blk_queue_io_stat(data->q)) |
e8064021 | 295 | rq->rq_flags |= RQF_IO_STAT; |
7c3fb70f | 296 | INIT_LIST_HEAD(&rq->queuelist); |
af76e555 CH |
297 | INIT_HLIST_NODE(&rq->hash); |
298 | RB_CLEAR_NODE(&rq->rb_node); | |
af76e555 CH |
299 | rq->rq_disk = NULL; |
300 | rq->part = NULL; | |
3ee32372 | 301 | rq->start_time = jiffies; |
af76e555 CH |
302 | rq->nr_phys_segments = 0; |
303 | #if defined(CONFIG_BLK_DEV_INTEGRITY) | |
304 | rq->nr_integrity_segments = 0; | |
305 | #endif | |
af76e555 CH |
306 | rq->special = NULL; |
307 | /* tag was already set */ | |
af76e555 | 308 | rq->extra_len = 0; |
e14575b3 | 309 | rq->__deadline = 0; |
af76e555 | 310 | |
af76e555 | 311 | INIT_LIST_HEAD(&rq->timeout_list); |
f6be4fb4 JA |
312 | rq->timeout = 0; |
313 | ||
af76e555 CH |
314 | rq->end_io = NULL; |
315 | rq->end_io_data = NULL; | |
316 | rq->next_rq = NULL; | |
317 | ||
7c3fb70f JA |
318 | #ifdef CONFIG_BLK_CGROUP |
319 | rq->rl = NULL; | |
320 | set_start_time_ns(rq); | |
321 | rq->io_start_time_ns = 0; | |
322 | #endif | |
323 | ||
e4cdf1a1 CH |
324 | data->ctx->rq_dispatched[op_is_sync(op)]++; |
325 | return rq; | |
5dee8577 CH |
326 | } |
327 | ||
d2c0d383 CH |
328 | static struct request *blk_mq_get_request(struct request_queue *q, |
329 | struct bio *bio, unsigned int op, | |
330 | struct blk_mq_alloc_data *data) | |
331 | { | |
332 | struct elevator_queue *e = q->elevator; | |
333 | struct request *rq; | |
e4cdf1a1 | 334 | unsigned int tag; |
21e768b4 | 335 | bool put_ctx_on_error = false; |
d2c0d383 CH |
336 | |
337 | blk_queue_enter_live(q); | |
338 | data->q = q; | |
21e768b4 BVA |
339 | if (likely(!data->ctx)) { |
340 | data->ctx = blk_mq_get_ctx(q); | |
341 | put_ctx_on_error = true; | |
342 | } | |
d2c0d383 CH |
343 | if (likely(!data->hctx)) |
344 | data->hctx = blk_mq_map_queue(q, data->ctx->cpu); | |
03a07c92 GR |
345 | if (op & REQ_NOWAIT) |
346 | data->flags |= BLK_MQ_REQ_NOWAIT; | |
d2c0d383 CH |
347 | |
348 | if (e) { | |
349 | data->flags |= BLK_MQ_REQ_INTERNAL; | |
350 | ||
351 | /* | |
352 | * Flush requests are special and go directly to the | |
353 | * dispatch list. | |
354 | */ | |
5bbf4e5a CH |
355 | if (!op_is_flush(op) && e->type->ops.mq.limit_depth) |
356 | e->type->ops.mq.limit_depth(op, data); | |
d2c0d383 CH |
357 | } |
358 | ||
e4cdf1a1 CH |
359 | tag = blk_mq_get_tag(data); |
360 | if (tag == BLK_MQ_TAG_FAIL) { | |
21e768b4 BVA |
361 | if (put_ctx_on_error) { |
362 | blk_mq_put_ctx(data->ctx); | |
1ad43c00 ML |
363 | data->ctx = NULL; |
364 | } | |
037cebb8 CH |
365 | blk_queue_exit(q); |
366 | return NULL; | |
d2c0d383 CH |
367 | } |
368 | ||
e4cdf1a1 | 369 | rq = blk_mq_rq_ctx_init(data, tag, op); |
037cebb8 CH |
370 | if (!op_is_flush(op)) { |
371 | rq->elv.icq = NULL; | |
5bbf4e5a | 372 | if (e && e->type->ops.mq.prepare_request) { |
44e8c2bf CH |
373 | if (e->type->icq_cache && rq_ioc(bio)) |
374 | blk_mq_sched_assign_ioc(rq, bio); | |
375 | ||
5bbf4e5a CH |
376 | e->type->ops.mq.prepare_request(rq, bio); |
377 | rq->rq_flags |= RQF_ELVPRIV; | |
44e8c2bf | 378 | } |
037cebb8 CH |
379 | } |
380 | data->hctx->queued++; | |
381 | return rq; | |
d2c0d383 CH |
382 | } |
383 | ||
cd6ce148 | 384 | struct request *blk_mq_alloc_request(struct request_queue *q, unsigned int op, |
9a95e4ef | 385 | blk_mq_req_flags_t flags) |
320ae51f | 386 | { |
5a797e00 | 387 | struct blk_mq_alloc_data alloc_data = { .flags = flags }; |
bd166ef1 | 388 | struct request *rq; |
a492f075 | 389 | int ret; |
320ae51f | 390 | |
3a0a5299 | 391 | ret = blk_queue_enter(q, flags); |
a492f075 JL |
392 | if (ret) |
393 | return ERR_PTR(ret); | |
320ae51f | 394 | |
cd6ce148 | 395 | rq = blk_mq_get_request(q, NULL, op, &alloc_data); |
3280d66a | 396 | blk_queue_exit(q); |
841bac2c | 397 | |
bd166ef1 | 398 | if (!rq) |
a492f075 | 399 | return ERR_PTR(-EWOULDBLOCK); |
0c4de0f3 | 400 | |
1ad43c00 | 401 | blk_mq_put_ctx(alloc_data.ctx); |
1ad43c00 | 402 | |
0c4de0f3 CH |
403 | rq->__data_len = 0; |
404 | rq->__sector = (sector_t) -1; | |
405 | rq->bio = rq->biotail = NULL; | |
320ae51f JA |
406 | return rq; |
407 | } | |
4bb659b1 | 408 | EXPORT_SYMBOL(blk_mq_alloc_request); |
320ae51f | 409 | |
cd6ce148 | 410 | struct request *blk_mq_alloc_request_hctx(struct request_queue *q, |
9a95e4ef | 411 | unsigned int op, blk_mq_req_flags_t flags, unsigned int hctx_idx) |
1f5bd336 | 412 | { |
6d2809d5 | 413 | struct blk_mq_alloc_data alloc_data = { .flags = flags }; |
1f5bd336 | 414 | struct request *rq; |
6d2809d5 | 415 | unsigned int cpu; |
1f5bd336 ML |
416 | int ret; |
417 | ||
418 | /* | |
419 | * If the tag allocator sleeps we could get an allocation for a | |
420 | * different hardware context. No need to complicate the low level | |
421 | * allocator for this for the rare use case of a command tied to | |
422 | * a specific queue. | |
423 | */ | |
424 | if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT))) | |
425 | return ERR_PTR(-EINVAL); | |
426 | ||
427 | if (hctx_idx >= q->nr_hw_queues) | |
428 | return ERR_PTR(-EIO); | |
429 | ||
3a0a5299 | 430 | ret = blk_queue_enter(q, flags); |
1f5bd336 ML |
431 | if (ret) |
432 | return ERR_PTR(ret); | |
433 | ||
c8712c6a CH |
434 | /* |
435 | * Check if the hardware context is actually mapped to anything. | |
436 | * If not tell the caller that it should skip this queue. | |
437 | */ | |
6d2809d5 OS |
438 | alloc_data.hctx = q->queue_hw_ctx[hctx_idx]; |
439 | if (!blk_mq_hw_queue_mapped(alloc_data.hctx)) { | |
440 | blk_queue_exit(q); | |
441 | return ERR_PTR(-EXDEV); | |
c8712c6a | 442 | } |
6d2809d5 OS |
443 | cpu = cpumask_first(alloc_data.hctx->cpumask); |
444 | alloc_data.ctx = __blk_mq_get_ctx(q, cpu); | |
1f5bd336 | 445 | |
cd6ce148 | 446 | rq = blk_mq_get_request(q, NULL, op, &alloc_data); |
3280d66a | 447 | blk_queue_exit(q); |
c8712c6a | 448 | |
6d2809d5 OS |
449 | if (!rq) |
450 | return ERR_PTR(-EWOULDBLOCK); | |
451 | ||
452 | return rq; | |
1f5bd336 ML |
453 | } |
454 | EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx); | |
455 | ||
6af54051 | 456 | void blk_mq_free_request(struct request *rq) |
320ae51f | 457 | { |
320ae51f | 458 | struct request_queue *q = rq->q; |
6af54051 CH |
459 | struct elevator_queue *e = q->elevator; |
460 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
461 | struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, ctx->cpu); | |
462 | const int sched_tag = rq->internal_tag; | |
463 | ||
5bbf4e5a | 464 | if (rq->rq_flags & RQF_ELVPRIV) { |
6af54051 CH |
465 | if (e && e->type->ops.mq.finish_request) |
466 | e->type->ops.mq.finish_request(rq); | |
467 | if (rq->elv.icq) { | |
468 | put_io_context(rq->elv.icq->ioc); | |
469 | rq->elv.icq = NULL; | |
470 | } | |
471 | } | |
320ae51f | 472 | |
6af54051 | 473 | ctx->rq_completed[rq_is_sync(rq)]++; |
e8064021 | 474 | if (rq->rq_flags & RQF_MQ_INFLIGHT) |
0d2602ca | 475 | atomic_dec(&hctx->nr_active); |
87760e5e | 476 | |
7beb2f84 JA |
477 | if (unlikely(laptop_mode && !blk_rq_is_passthrough(rq))) |
478 | laptop_io_completion(q->backing_dev_info); | |
479 | ||
87760e5e | 480 | wbt_done(q->rq_wb, &rq->issue_stat); |
0d2602ca | 481 | |
85acb3ba SL |
482 | if (blk_rq_rl(rq)) |
483 | blk_put_rl(blk_rq_rl(rq)); | |
484 | ||
1d9bd516 | 485 | blk_mq_rq_update_state(rq, MQ_RQ_IDLE); |
bd166ef1 JA |
486 | if (rq->tag != -1) |
487 | blk_mq_put_tag(hctx, hctx->tags, ctx, rq->tag); | |
488 | if (sched_tag != -1) | |
c05f8525 | 489 | blk_mq_put_tag(hctx, hctx->sched_tags, ctx, sched_tag); |
6d8c6c0f | 490 | blk_mq_sched_restart(hctx); |
3ef28e83 | 491 | blk_queue_exit(q); |
320ae51f | 492 | } |
1a3b595a | 493 | EXPORT_SYMBOL_GPL(blk_mq_free_request); |
320ae51f | 494 | |
2a842aca | 495 | inline void __blk_mq_end_request(struct request *rq, blk_status_t error) |
320ae51f | 496 | { |
0d11e6ac ML |
497 | blk_account_io_done(rq); |
498 | ||
91b63639 | 499 | if (rq->end_io) { |
87760e5e | 500 | wbt_done(rq->q->rq_wb, &rq->issue_stat); |
320ae51f | 501 | rq->end_io(rq, error); |
91b63639 CH |
502 | } else { |
503 | if (unlikely(blk_bidi_rq(rq))) | |
504 | blk_mq_free_request(rq->next_rq); | |
320ae51f | 505 | blk_mq_free_request(rq); |
91b63639 | 506 | } |
320ae51f | 507 | } |
c8a446ad | 508 | EXPORT_SYMBOL(__blk_mq_end_request); |
63151a44 | 509 | |
2a842aca | 510 | void blk_mq_end_request(struct request *rq, blk_status_t error) |
63151a44 CH |
511 | { |
512 | if (blk_update_request(rq, error, blk_rq_bytes(rq))) | |
513 | BUG(); | |
c8a446ad | 514 | __blk_mq_end_request(rq, error); |
63151a44 | 515 | } |
c8a446ad | 516 | EXPORT_SYMBOL(blk_mq_end_request); |
320ae51f | 517 | |
30a91cb4 | 518 | static void __blk_mq_complete_request_remote(void *data) |
320ae51f | 519 | { |
3d6efbf6 | 520 | struct request *rq = data; |
320ae51f | 521 | |
30a91cb4 | 522 | rq->q->softirq_done_fn(rq); |
320ae51f | 523 | } |
320ae51f | 524 | |
453f8341 | 525 | static void __blk_mq_complete_request(struct request *rq) |
320ae51f JA |
526 | { |
527 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
38535201 | 528 | bool shared = false; |
320ae51f JA |
529 | int cpu; |
530 | ||
1d9bd516 | 531 | WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT); |
5a61c363 | 532 | blk_mq_rq_update_state(rq, MQ_RQ_COMPLETE); |
1d9bd516 | 533 | |
453f8341 CH |
534 | if (rq->internal_tag != -1) |
535 | blk_mq_sched_completed_request(rq); | |
536 | if (rq->rq_flags & RQF_STATS) { | |
537 | blk_mq_poll_stats_start(rq->q); | |
538 | blk_stat_add(rq); | |
539 | } | |
540 | ||
38535201 | 541 | if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) { |
30a91cb4 CH |
542 | rq->q->softirq_done_fn(rq); |
543 | return; | |
544 | } | |
320ae51f JA |
545 | |
546 | cpu = get_cpu(); | |
38535201 CH |
547 | if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags)) |
548 | shared = cpus_share_cache(cpu, ctx->cpu); | |
549 | ||
550 | if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) { | |
30a91cb4 | 551 | rq->csd.func = __blk_mq_complete_request_remote; |
3d6efbf6 CH |
552 | rq->csd.info = rq; |
553 | rq->csd.flags = 0; | |
c46fff2a | 554 | smp_call_function_single_async(ctx->cpu, &rq->csd); |
3d6efbf6 | 555 | } else { |
30a91cb4 | 556 | rq->q->softirq_done_fn(rq); |
3d6efbf6 | 557 | } |
320ae51f JA |
558 | put_cpu(); |
559 | } | |
30a91cb4 | 560 | |
04ced159 JA |
561 | static void hctx_unlock(struct blk_mq_hw_ctx *hctx, int srcu_idx) |
562 | { | |
563 | if (!(hctx->flags & BLK_MQ_F_BLOCKING)) | |
564 | rcu_read_unlock(); | |
565 | else | |
05707b64 | 566 | srcu_read_unlock(hctx->srcu, srcu_idx); |
04ced159 JA |
567 | } |
568 | ||
569 | static void hctx_lock(struct blk_mq_hw_ctx *hctx, int *srcu_idx) | |
570 | { | |
08b5a6e2 JA |
571 | if (!(hctx->flags & BLK_MQ_F_BLOCKING)) { |
572 | /* shut up gcc false positive */ | |
573 | *srcu_idx = 0; | |
04ced159 | 574 | rcu_read_lock(); |
08b5a6e2 | 575 | } else |
05707b64 | 576 | *srcu_idx = srcu_read_lock(hctx->srcu); |
04ced159 JA |
577 | } |
578 | ||
1d9bd516 TH |
579 | static void blk_mq_rq_update_aborted_gstate(struct request *rq, u64 gstate) |
580 | { | |
581 | unsigned long flags; | |
582 | ||
583 | /* | |
584 | * blk_mq_rq_aborted_gstate() is used from the completion path and | |
585 | * can thus be called from irq context. u64_stats_fetch in the | |
586 | * middle of update on the same CPU leads to lockup. Disable irq | |
587 | * while updating. | |
588 | */ | |
589 | local_irq_save(flags); | |
590 | u64_stats_update_begin(&rq->aborted_gstate_sync); | |
591 | rq->aborted_gstate = gstate; | |
592 | u64_stats_update_end(&rq->aborted_gstate_sync); | |
593 | local_irq_restore(flags); | |
594 | } | |
595 | ||
596 | static u64 blk_mq_rq_aborted_gstate(struct request *rq) | |
597 | { | |
598 | unsigned int start; | |
599 | u64 aborted_gstate; | |
600 | ||
601 | do { | |
602 | start = u64_stats_fetch_begin(&rq->aborted_gstate_sync); | |
603 | aborted_gstate = rq->aborted_gstate; | |
604 | } while (u64_stats_fetch_retry(&rq->aborted_gstate_sync, start)); | |
605 | ||
606 | return aborted_gstate; | |
607 | } | |
608 | ||
30a91cb4 CH |
609 | /** |
610 | * blk_mq_complete_request - end I/O on a request | |
611 | * @rq: the request being processed | |
612 | * | |
613 | * Description: | |
614 | * Ends all I/O on a request. It does not handle partial completions. | |
615 | * The actual completion happens out-of-order, through a IPI handler. | |
616 | **/ | |
08e0029a | 617 | void blk_mq_complete_request(struct request *rq) |
30a91cb4 | 618 | { |
95f09684 | 619 | struct request_queue *q = rq->q; |
5197c05e TH |
620 | struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, rq->mq_ctx->cpu); |
621 | int srcu_idx; | |
95f09684 JA |
622 | |
623 | if (unlikely(blk_should_fake_timeout(q))) | |
30a91cb4 | 624 | return; |
5197c05e | 625 | |
1d9bd516 TH |
626 | /* |
627 | * If @rq->aborted_gstate equals the current instance, timeout is | |
628 | * claiming @rq and we lost. This is synchronized through | |
629 | * hctx_lock(). See blk_mq_timeout_work() for details. | |
630 | * | |
631 | * Completion path never blocks and we can directly use RCU here | |
632 | * instead of hctx_lock() which can be either RCU or SRCU. | |
633 | * However, that would complicate paths which want to synchronize | |
634 | * against us. Let stay in sync with the issue path so that | |
635 | * hctx_lock() covers both issue and completion paths. | |
636 | */ | |
5197c05e | 637 | hctx_lock(hctx, &srcu_idx); |
634f9e46 | 638 | if (blk_mq_rq_aborted_gstate(rq) != rq->gstate) |
ed851860 | 639 | __blk_mq_complete_request(rq); |
5197c05e | 640 | hctx_unlock(hctx, srcu_idx); |
30a91cb4 CH |
641 | } |
642 | EXPORT_SYMBOL(blk_mq_complete_request); | |
320ae51f | 643 | |
973c0191 KB |
644 | int blk_mq_request_started(struct request *rq) |
645 | { | |
5a61c363 | 646 | return blk_mq_rq_state(rq) != MQ_RQ_IDLE; |
973c0191 KB |
647 | } |
648 | EXPORT_SYMBOL_GPL(blk_mq_request_started); | |
649 | ||
e2490073 | 650 | void blk_mq_start_request(struct request *rq) |
320ae51f JA |
651 | { |
652 | struct request_queue *q = rq->q; | |
653 | ||
bd166ef1 JA |
654 | blk_mq_sched_started_request(rq); |
655 | ||
320ae51f JA |
656 | trace_block_rq_issue(q, rq); |
657 | ||
cf43e6be | 658 | if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) { |
88eeca49 | 659 | blk_stat_set_issue(&rq->issue_stat, blk_rq_sectors(rq)); |
cf43e6be | 660 | rq->rq_flags |= RQF_STATS; |
87760e5e | 661 | wbt_issue(q->rq_wb, &rq->issue_stat); |
cf43e6be JA |
662 | } |
663 | ||
1d9bd516 | 664 | WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE); |
538b7534 | 665 | |
87ee7b11 | 666 | /* |
1d9bd516 TH |
667 | * Mark @rq in-flight which also advances the generation number, |
668 | * and register for timeout. Protect with a seqcount to allow the | |
669 | * timeout path to read both @rq->gstate and @rq->deadline | |
670 | * coherently. | |
a7af0af3 | 671 | * |
1d9bd516 TH |
672 | * This is the only place where a request is marked in-flight. If |
673 | * the timeout path reads an in-flight @rq->gstate, the | |
674 | * @rq->deadline it reads together under @rq->gstate_seq is | |
675 | * guaranteed to be the matching one. | |
87ee7b11 | 676 | */ |
1d9bd516 TH |
677 | preempt_disable(); |
678 | write_seqcount_begin(&rq->gstate_seq); | |
679 | ||
680 | blk_mq_rq_update_state(rq, MQ_RQ_IN_FLIGHT); | |
681 | blk_add_timer(rq); | |
682 | ||
683 | write_seqcount_end(&rq->gstate_seq); | |
684 | preempt_enable(); | |
685 | ||
49f5baa5 CH |
686 | if (q->dma_drain_size && blk_rq_bytes(rq)) { |
687 | /* | |
688 | * Make sure space for the drain appears. We know we can do | |
689 | * this because max_hw_segments has been adjusted to be one | |
690 | * fewer than the device can handle. | |
691 | */ | |
692 | rq->nr_phys_segments++; | |
693 | } | |
320ae51f | 694 | } |
e2490073 | 695 | EXPORT_SYMBOL(blk_mq_start_request); |
320ae51f | 696 | |
d9d149a3 | 697 | /* |
5a61c363 TH |
698 | * When we reach here because queue is busy, it's safe to change the state |
699 | * to IDLE without checking @rq->aborted_gstate because we should still be | |
700 | * holding the RCU read lock and thus protected against timeout. | |
d9d149a3 | 701 | */ |
ed0791b2 | 702 | static void __blk_mq_requeue_request(struct request *rq) |
320ae51f JA |
703 | { |
704 | struct request_queue *q = rq->q; | |
705 | ||
923218f6 ML |
706 | blk_mq_put_driver_tag(rq); |
707 | ||
320ae51f | 708 | trace_block_rq_requeue(q, rq); |
87760e5e | 709 | wbt_requeue(q->rq_wb, &rq->issue_stat); |
bd166ef1 | 710 | blk_mq_sched_requeue_request(rq); |
49f5baa5 | 711 | |
5a61c363 | 712 | if (blk_mq_rq_state(rq) != MQ_RQ_IDLE) { |
1d9bd516 | 713 | blk_mq_rq_update_state(rq, MQ_RQ_IDLE); |
e2490073 CH |
714 | if (q->dma_drain_size && blk_rq_bytes(rq)) |
715 | rq->nr_phys_segments--; | |
716 | } | |
320ae51f JA |
717 | } |
718 | ||
2b053aca | 719 | void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list) |
ed0791b2 | 720 | { |
ed0791b2 | 721 | __blk_mq_requeue_request(rq); |
ed0791b2 | 722 | |
ed0791b2 | 723 | BUG_ON(blk_queued_rq(rq)); |
2b053aca | 724 | blk_mq_add_to_requeue_list(rq, true, kick_requeue_list); |
ed0791b2 CH |
725 | } |
726 | EXPORT_SYMBOL(blk_mq_requeue_request); | |
727 | ||
6fca6a61 CH |
728 | static void blk_mq_requeue_work(struct work_struct *work) |
729 | { | |
730 | struct request_queue *q = | |
2849450a | 731 | container_of(work, struct request_queue, requeue_work.work); |
6fca6a61 CH |
732 | LIST_HEAD(rq_list); |
733 | struct request *rq, *next; | |
6fca6a61 | 734 | |
18e9781d | 735 | spin_lock_irq(&q->requeue_lock); |
6fca6a61 | 736 | list_splice_init(&q->requeue_list, &rq_list); |
18e9781d | 737 | spin_unlock_irq(&q->requeue_lock); |
6fca6a61 CH |
738 | |
739 | list_for_each_entry_safe(rq, next, &rq_list, queuelist) { | |
e8064021 | 740 | if (!(rq->rq_flags & RQF_SOFTBARRIER)) |
6fca6a61 CH |
741 | continue; |
742 | ||
e8064021 | 743 | rq->rq_flags &= ~RQF_SOFTBARRIER; |
6fca6a61 | 744 | list_del_init(&rq->queuelist); |
bd6737f1 | 745 | blk_mq_sched_insert_request(rq, true, false, false, true); |
6fca6a61 CH |
746 | } |
747 | ||
748 | while (!list_empty(&rq_list)) { | |
749 | rq = list_entry(rq_list.next, struct request, queuelist); | |
750 | list_del_init(&rq->queuelist); | |
bd6737f1 | 751 | blk_mq_sched_insert_request(rq, false, false, false, true); |
6fca6a61 CH |
752 | } |
753 | ||
52d7f1b5 | 754 | blk_mq_run_hw_queues(q, false); |
6fca6a61 CH |
755 | } |
756 | ||
2b053aca BVA |
757 | void blk_mq_add_to_requeue_list(struct request *rq, bool at_head, |
758 | bool kick_requeue_list) | |
6fca6a61 CH |
759 | { |
760 | struct request_queue *q = rq->q; | |
761 | unsigned long flags; | |
762 | ||
763 | /* | |
764 | * We abuse this flag that is otherwise used by the I/O scheduler to | |
ff821d27 | 765 | * request head insertion from the workqueue. |
6fca6a61 | 766 | */ |
e8064021 | 767 | BUG_ON(rq->rq_flags & RQF_SOFTBARRIER); |
6fca6a61 CH |
768 | |
769 | spin_lock_irqsave(&q->requeue_lock, flags); | |
770 | if (at_head) { | |
e8064021 | 771 | rq->rq_flags |= RQF_SOFTBARRIER; |
6fca6a61 CH |
772 | list_add(&rq->queuelist, &q->requeue_list); |
773 | } else { | |
774 | list_add_tail(&rq->queuelist, &q->requeue_list); | |
775 | } | |
776 | spin_unlock_irqrestore(&q->requeue_lock, flags); | |
2b053aca BVA |
777 | |
778 | if (kick_requeue_list) | |
779 | blk_mq_kick_requeue_list(q); | |
6fca6a61 CH |
780 | } |
781 | EXPORT_SYMBOL(blk_mq_add_to_requeue_list); | |
782 | ||
783 | void blk_mq_kick_requeue_list(struct request_queue *q) | |
784 | { | |
2849450a | 785 | kblockd_schedule_delayed_work(&q->requeue_work, 0); |
6fca6a61 CH |
786 | } |
787 | EXPORT_SYMBOL(blk_mq_kick_requeue_list); | |
788 | ||
2849450a MS |
789 | void blk_mq_delay_kick_requeue_list(struct request_queue *q, |
790 | unsigned long msecs) | |
791 | { | |
d4acf365 BVA |
792 | kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, |
793 | msecs_to_jiffies(msecs)); | |
2849450a MS |
794 | } |
795 | EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list); | |
796 | ||
0e62f51f JA |
797 | struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag) |
798 | { | |
88c7b2b7 JA |
799 | if (tag < tags->nr_tags) { |
800 | prefetch(tags->rqs[tag]); | |
4ee86bab | 801 | return tags->rqs[tag]; |
88c7b2b7 | 802 | } |
4ee86bab HR |
803 | |
804 | return NULL; | |
24d2f903 CH |
805 | } |
806 | EXPORT_SYMBOL(blk_mq_tag_to_rq); | |
807 | ||
320ae51f | 808 | struct blk_mq_timeout_data { |
46f92d42 CH |
809 | unsigned long next; |
810 | unsigned int next_set; | |
1d9bd516 | 811 | unsigned int nr_expired; |
320ae51f JA |
812 | }; |
813 | ||
358f70da | 814 | static void blk_mq_rq_timed_out(struct request *req, bool reserved) |
320ae51f | 815 | { |
f8a5b122 | 816 | const struct blk_mq_ops *ops = req->q->mq_ops; |
46f92d42 | 817 | enum blk_eh_timer_return ret = BLK_EH_RESET_TIMER; |
87ee7b11 | 818 | |
634f9e46 TH |
819 | req->rq_flags |= RQF_MQ_TIMEOUT_EXPIRED; |
820 | ||
46f92d42 | 821 | if (ops->timeout) |
0152fb6b | 822 | ret = ops->timeout(req, reserved); |
46f92d42 CH |
823 | |
824 | switch (ret) { | |
825 | case BLK_EH_HANDLED: | |
826 | __blk_mq_complete_request(req); | |
827 | break; | |
828 | case BLK_EH_RESET_TIMER: | |
1d9bd516 TH |
829 | /* |
830 | * As nothing prevents from completion happening while | |
831 | * ->aborted_gstate is set, this may lead to ignored | |
832 | * completions and further spurious timeouts. | |
833 | */ | |
834 | blk_mq_rq_update_aborted_gstate(req, 0); | |
46f92d42 | 835 | blk_add_timer(req); |
46f92d42 CH |
836 | break; |
837 | case BLK_EH_NOT_HANDLED: | |
838 | break; | |
839 | default: | |
840 | printk(KERN_ERR "block: bad eh return: %d\n", ret); | |
841 | break; | |
842 | } | |
87ee7b11 | 843 | } |
5b3f25fc | 844 | |
81481eb4 CH |
845 | static void blk_mq_check_expired(struct blk_mq_hw_ctx *hctx, |
846 | struct request *rq, void *priv, bool reserved) | |
847 | { | |
848 | struct blk_mq_timeout_data *data = priv; | |
1d9bd516 TH |
849 | unsigned long gstate, deadline; |
850 | int start; | |
851 | ||
852 | might_sleep(); | |
87ee7b11 | 853 | |
5a61c363 | 854 | if (rq->rq_flags & RQF_MQ_TIMEOUT_EXPIRED) |
46f92d42 | 855 | return; |
87ee7b11 | 856 | |
1d9bd516 TH |
857 | /* read coherent snapshots of @rq->state_gen and @rq->deadline */ |
858 | while (true) { | |
859 | start = read_seqcount_begin(&rq->gstate_seq); | |
860 | gstate = READ_ONCE(rq->gstate); | |
0a72e7f4 | 861 | deadline = blk_rq_deadline(rq); |
1d9bd516 TH |
862 | if (!read_seqcount_retry(&rq->gstate_seq, start)) |
863 | break; | |
864 | cond_resched(); | |
865 | } | |
a7af0af3 | 866 | |
1d9bd516 TH |
867 | /* if in-flight && overdue, mark for abortion */ |
868 | if ((gstate & MQ_RQ_STATE_MASK) == MQ_RQ_IN_FLIGHT && | |
869 | time_after_eq(jiffies, deadline)) { | |
870 | blk_mq_rq_update_aborted_gstate(rq, gstate); | |
871 | data->nr_expired++; | |
872 | hctx->nr_expired++; | |
a7af0af3 PZ |
873 | } else if (!data->next_set || time_after(data->next, deadline)) { |
874 | data->next = deadline; | |
46f92d42 CH |
875 | data->next_set = 1; |
876 | } | |
87ee7b11 JA |
877 | } |
878 | ||
1d9bd516 TH |
879 | static void blk_mq_terminate_expired(struct blk_mq_hw_ctx *hctx, |
880 | struct request *rq, void *priv, bool reserved) | |
881 | { | |
882 | /* | |
883 | * We marked @rq->aborted_gstate and waited for RCU. If there were | |
884 | * completions that we lost to, they would have finished and | |
885 | * updated @rq->gstate by now; otherwise, the completion path is | |
886 | * now guaranteed to see @rq->aborted_gstate and yield. If | |
887 | * @rq->aborted_gstate still matches @rq->gstate, @rq is ours. | |
888 | */ | |
634f9e46 TH |
889 | if (!(rq->rq_flags & RQF_MQ_TIMEOUT_EXPIRED) && |
890 | READ_ONCE(rq->gstate) == rq->aborted_gstate) | |
1d9bd516 TH |
891 | blk_mq_rq_timed_out(rq, reserved); |
892 | } | |
893 | ||
287922eb | 894 | static void blk_mq_timeout_work(struct work_struct *work) |
320ae51f | 895 | { |
287922eb CH |
896 | struct request_queue *q = |
897 | container_of(work, struct request_queue, timeout_work); | |
81481eb4 CH |
898 | struct blk_mq_timeout_data data = { |
899 | .next = 0, | |
900 | .next_set = 0, | |
1d9bd516 | 901 | .nr_expired = 0, |
81481eb4 | 902 | }; |
1d9bd516 | 903 | struct blk_mq_hw_ctx *hctx; |
81481eb4 | 904 | int i; |
320ae51f | 905 | |
71f79fb3 GKB |
906 | /* A deadlock might occur if a request is stuck requiring a |
907 | * timeout at the same time a queue freeze is waiting | |
908 | * completion, since the timeout code would not be able to | |
909 | * acquire the queue reference here. | |
910 | * | |
911 | * That's why we don't use blk_queue_enter here; instead, we use | |
912 | * percpu_ref_tryget directly, because we need to be able to | |
913 | * obtain a reference even in the short window between the queue | |
914 | * starting to freeze, by dropping the first reference in | |
1671d522 | 915 | * blk_freeze_queue_start, and the moment the last request is |
71f79fb3 GKB |
916 | * consumed, marked by the instant q_usage_counter reaches |
917 | * zero. | |
918 | */ | |
919 | if (!percpu_ref_tryget(&q->q_usage_counter)) | |
287922eb CH |
920 | return; |
921 | ||
1d9bd516 | 922 | /* scan for the expired ones and set their ->aborted_gstate */ |
0bf6cd5b | 923 | blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &data); |
320ae51f | 924 | |
1d9bd516 TH |
925 | if (data.nr_expired) { |
926 | bool has_rcu = false; | |
927 | ||
928 | /* | |
929 | * Wait till everyone sees ->aborted_gstate. The | |
930 | * sequential waits for SRCUs aren't ideal. If this ever | |
931 | * becomes a problem, we can add per-hw_ctx rcu_head and | |
932 | * wait in parallel. | |
933 | */ | |
934 | queue_for_each_hw_ctx(q, hctx, i) { | |
935 | if (!hctx->nr_expired) | |
936 | continue; | |
937 | ||
938 | if (!(hctx->flags & BLK_MQ_F_BLOCKING)) | |
939 | has_rcu = true; | |
940 | else | |
05707b64 | 941 | synchronize_srcu(hctx->srcu); |
1d9bd516 TH |
942 | |
943 | hctx->nr_expired = 0; | |
944 | } | |
945 | if (has_rcu) | |
946 | synchronize_rcu(); | |
947 | ||
948 | /* terminate the ones we won */ | |
949 | blk_mq_queue_tag_busy_iter(q, blk_mq_terminate_expired, NULL); | |
950 | } | |
951 | ||
81481eb4 CH |
952 | if (data.next_set) { |
953 | data.next = blk_rq_timeout(round_jiffies_up(data.next)); | |
954 | mod_timer(&q->timeout, data.next); | |
0d2602ca | 955 | } else { |
fcd36c36 BVA |
956 | /* |
957 | * Request timeouts are handled as a forward rolling timer. If | |
958 | * we end up here it means that no requests are pending and | |
959 | * also that no request has been pending for a while. Mark | |
960 | * each hctx as idle. | |
961 | */ | |
f054b56c ML |
962 | queue_for_each_hw_ctx(q, hctx, i) { |
963 | /* the hctx may be unmapped, so check it here */ | |
964 | if (blk_mq_hw_queue_mapped(hctx)) | |
965 | blk_mq_tag_idle(hctx); | |
966 | } | |
0d2602ca | 967 | } |
287922eb | 968 | blk_queue_exit(q); |
320ae51f JA |
969 | } |
970 | ||
88459642 OS |
971 | struct flush_busy_ctx_data { |
972 | struct blk_mq_hw_ctx *hctx; | |
973 | struct list_head *list; | |
974 | }; | |
975 | ||
976 | static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data) | |
977 | { | |
978 | struct flush_busy_ctx_data *flush_data = data; | |
979 | struct blk_mq_hw_ctx *hctx = flush_data->hctx; | |
980 | struct blk_mq_ctx *ctx = hctx->ctxs[bitnr]; | |
981 | ||
982 | sbitmap_clear_bit(sb, bitnr); | |
983 | spin_lock(&ctx->lock); | |
984 | list_splice_tail_init(&ctx->rq_list, flush_data->list); | |
985 | spin_unlock(&ctx->lock); | |
986 | return true; | |
987 | } | |
988 | ||
1429d7c9 JA |
989 | /* |
990 | * Process software queues that have been marked busy, splicing them | |
991 | * to the for-dispatch | |
992 | */ | |
2c3ad667 | 993 | void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list) |
1429d7c9 | 994 | { |
88459642 OS |
995 | struct flush_busy_ctx_data data = { |
996 | .hctx = hctx, | |
997 | .list = list, | |
998 | }; | |
1429d7c9 | 999 | |
88459642 | 1000 | sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data); |
1429d7c9 | 1001 | } |
2c3ad667 | 1002 | EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs); |
1429d7c9 | 1003 | |
b347689f ML |
1004 | struct dispatch_rq_data { |
1005 | struct blk_mq_hw_ctx *hctx; | |
1006 | struct request *rq; | |
1007 | }; | |
1008 | ||
1009 | static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr, | |
1010 | void *data) | |
1011 | { | |
1012 | struct dispatch_rq_data *dispatch_data = data; | |
1013 | struct blk_mq_hw_ctx *hctx = dispatch_data->hctx; | |
1014 | struct blk_mq_ctx *ctx = hctx->ctxs[bitnr]; | |
1015 | ||
1016 | spin_lock(&ctx->lock); | |
1017 | if (unlikely(!list_empty(&ctx->rq_list))) { | |
1018 | dispatch_data->rq = list_entry_rq(ctx->rq_list.next); | |
1019 | list_del_init(&dispatch_data->rq->queuelist); | |
1020 | if (list_empty(&ctx->rq_list)) | |
1021 | sbitmap_clear_bit(sb, bitnr); | |
1022 | } | |
1023 | spin_unlock(&ctx->lock); | |
1024 | ||
1025 | return !dispatch_data->rq; | |
1026 | } | |
1027 | ||
1028 | struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx, | |
1029 | struct blk_mq_ctx *start) | |
1030 | { | |
1031 | unsigned off = start ? start->index_hw : 0; | |
1032 | struct dispatch_rq_data data = { | |
1033 | .hctx = hctx, | |
1034 | .rq = NULL, | |
1035 | }; | |
1036 | ||
1037 | __sbitmap_for_each_set(&hctx->ctx_map, off, | |
1038 | dispatch_rq_from_ctx, &data); | |
1039 | ||
1040 | return data.rq; | |
1041 | } | |
1042 | ||
703fd1c0 JA |
1043 | static inline unsigned int queued_to_index(unsigned int queued) |
1044 | { | |
1045 | if (!queued) | |
1046 | return 0; | |
1429d7c9 | 1047 | |
703fd1c0 | 1048 | return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1); |
1429d7c9 JA |
1049 | } |
1050 | ||
bd6737f1 JA |
1051 | bool blk_mq_get_driver_tag(struct request *rq, struct blk_mq_hw_ctx **hctx, |
1052 | bool wait) | |
bd166ef1 JA |
1053 | { |
1054 | struct blk_mq_alloc_data data = { | |
1055 | .q = rq->q, | |
bd166ef1 JA |
1056 | .hctx = blk_mq_map_queue(rq->q, rq->mq_ctx->cpu), |
1057 | .flags = wait ? 0 : BLK_MQ_REQ_NOWAIT, | |
1058 | }; | |
1059 | ||
5feeacdd JA |
1060 | might_sleep_if(wait); |
1061 | ||
81380ca1 OS |
1062 | if (rq->tag != -1) |
1063 | goto done; | |
bd166ef1 | 1064 | |
415b806d SG |
1065 | if (blk_mq_tag_is_reserved(data.hctx->sched_tags, rq->internal_tag)) |
1066 | data.flags |= BLK_MQ_REQ_RESERVED; | |
1067 | ||
bd166ef1 JA |
1068 | rq->tag = blk_mq_get_tag(&data); |
1069 | if (rq->tag >= 0) { | |
200e86b3 JA |
1070 | if (blk_mq_tag_busy(data.hctx)) { |
1071 | rq->rq_flags |= RQF_MQ_INFLIGHT; | |
1072 | atomic_inc(&data.hctx->nr_active); | |
1073 | } | |
bd166ef1 | 1074 | data.hctx->tags->rqs[rq->tag] = rq; |
bd166ef1 JA |
1075 | } |
1076 | ||
81380ca1 OS |
1077 | done: |
1078 | if (hctx) | |
1079 | *hctx = data.hctx; | |
1080 | return rq->tag != -1; | |
bd166ef1 JA |
1081 | } |
1082 | ||
eb619fdb JA |
1083 | static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode, |
1084 | int flags, void *key) | |
da55f2cc OS |
1085 | { |
1086 | struct blk_mq_hw_ctx *hctx; | |
1087 | ||
1088 | hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait); | |
1089 | ||
eb619fdb | 1090 | list_del_init(&wait->entry); |
da55f2cc OS |
1091 | blk_mq_run_hw_queue(hctx, true); |
1092 | return 1; | |
1093 | } | |
1094 | ||
f906a6a0 JA |
1095 | /* |
1096 | * Mark us waiting for a tag. For shared tags, this involves hooking us into | |
ee3e4de5 BVA |
1097 | * the tag wakeups. For non-shared tags, we can simply mark us needing a |
1098 | * restart. For both cases, take care to check the condition again after | |
f906a6a0 JA |
1099 | * marking us as waiting. |
1100 | */ | |
1101 | static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx **hctx, | |
1102 | struct request *rq) | |
da55f2cc | 1103 | { |
eb619fdb | 1104 | struct blk_mq_hw_ctx *this_hctx = *hctx; |
f906a6a0 | 1105 | bool shared_tags = (this_hctx->flags & BLK_MQ_F_TAG_SHARED) != 0; |
da55f2cc | 1106 | struct sbq_wait_state *ws; |
f906a6a0 JA |
1107 | wait_queue_entry_t *wait; |
1108 | bool ret; | |
da55f2cc | 1109 | |
f906a6a0 JA |
1110 | if (!shared_tags) { |
1111 | if (!test_bit(BLK_MQ_S_SCHED_RESTART, &this_hctx->state)) | |
1112 | set_bit(BLK_MQ_S_SCHED_RESTART, &this_hctx->state); | |
1113 | } else { | |
1114 | wait = &this_hctx->dispatch_wait; | |
1115 | if (!list_empty_careful(&wait->entry)) | |
1116 | return false; | |
1117 | ||
1118 | spin_lock(&this_hctx->lock); | |
1119 | if (!list_empty(&wait->entry)) { | |
1120 | spin_unlock(&this_hctx->lock); | |
1121 | return false; | |
1122 | } | |
eb619fdb | 1123 | |
f906a6a0 JA |
1124 | ws = bt_wait_ptr(&this_hctx->tags->bitmap_tags, this_hctx); |
1125 | add_wait_queue(&ws->wait, wait); | |
eb619fdb JA |
1126 | } |
1127 | ||
da55f2cc | 1128 | /* |
eb619fdb JA |
1129 | * It's possible that a tag was freed in the window between the |
1130 | * allocation failure and adding the hardware queue to the wait | |
1131 | * queue. | |
da55f2cc | 1132 | */ |
f906a6a0 JA |
1133 | ret = blk_mq_get_driver_tag(rq, hctx, false); |
1134 | ||
1135 | if (!shared_tags) { | |
1136 | /* | |
1137 | * Don't clear RESTART here, someone else could have set it. | |
1138 | * At most this will cost an extra queue run. | |
1139 | */ | |
1140 | return ret; | |
1141 | } else { | |
1142 | if (!ret) { | |
1143 | spin_unlock(&this_hctx->lock); | |
1144 | return false; | |
1145 | } | |
1146 | ||
1147 | /* | |
1148 | * We got a tag, remove ourselves from the wait queue to ensure | |
1149 | * someone else gets the wakeup. | |
1150 | */ | |
1151 | spin_lock_irq(&ws->wait.lock); | |
1152 | list_del_init(&wait->entry); | |
1153 | spin_unlock_irq(&ws->wait.lock); | |
eb619fdb | 1154 | spin_unlock(&this_hctx->lock); |
f906a6a0 | 1155 | return true; |
eb619fdb | 1156 | } |
da55f2cc OS |
1157 | } |
1158 | ||
de148297 | 1159 | bool blk_mq_dispatch_rq_list(struct request_queue *q, struct list_head *list, |
eb619fdb | 1160 | bool got_budget) |
320ae51f | 1161 | { |
81380ca1 | 1162 | struct blk_mq_hw_ctx *hctx; |
6d6f167c | 1163 | struct request *rq, *nxt; |
eb619fdb | 1164 | bool no_tag = false; |
fc17b653 | 1165 | int errors, queued; |
320ae51f | 1166 | |
81380ca1 OS |
1167 | if (list_empty(list)) |
1168 | return false; | |
1169 | ||
de148297 ML |
1170 | WARN_ON(!list_is_singular(list) && got_budget); |
1171 | ||
320ae51f JA |
1172 | /* |
1173 | * Now process all the entries, sending them to the driver. | |
1174 | */ | |
93efe981 | 1175 | errors = queued = 0; |
81380ca1 | 1176 | do { |
74c45052 | 1177 | struct blk_mq_queue_data bd; |
fc17b653 | 1178 | blk_status_t ret; |
320ae51f | 1179 | |
f04c3df3 | 1180 | rq = list_first_entry(list, struct request, queuelist); |
bd166ef1 | 1181 | if (!blk_mq_get_driver_tag(rq, &hctx, false)) { |
3c782d67 | 1182 | /* |
da55f2cc | 1183 | * The initial allocation attempt failed, so we need to |
eb619fdb JA |
1184 | * rerun the hardware queue when a tag is freed. The |
1185 | * waitqueue takes care of that. If the queue is run | |
1186 | * before we add this entry back on the dispatch list, | |
1187 | * we'll re-run it below. | |
3c782d67 | 1188 | */ |
f906a6a0 | 1189 | if (!blk_mq_mark_tag_wait(&hctx, rq)) { |
de148297 ML |
1190 | if (got_budget) |
1191 | blk_mq_put_dispatch_budget(hctx); | |
f906a6a0 JA |
1192 | /* |
1193 | * For non-shared tags, the RESTART check | |
1194 | * will suffice. | |
1195 | */ | |
1196 | if (hctx->flags & BLK_MQ_F_TAG_SHARED) | |
1197 | no_tag = true; | |
de148297 ML |
1198 | break; |
1199 | } | |
1200 | } | |
1201 | ||
0c6af1cc ML |
1202 | if (!got_budget && !blk_mq_get_dispatch_budget(hctx)) { |
1203 | blk_mq_put_driver_tag(rq); | |
88022d72 | 1204 | break; |
0c6af1cc | 1205 | } |
da55f2cc | 1206 | |
320ae51f | 1207 | list_del_init(&rq->queuelist); |
320ae51f | 1208 | |
74c45052 | 1209 | bd.rq = rq; |
113285b4 JA |
1210 | |
1211 | /* | |
1212 | * Flag last if we have no more requests, or if we have more | |
1213 | * but can't assign a driver tag to it. | |
1214 | */ | |
1215 | if (list_empty(list)) | |
1216 | bd.last = true; | |
1217 | else { | |
113285b4 JA |
1218 | nxt = list_first_entry(list, struct request, queuelist); |
1219 | bd.last = !blk_mq_get_driver_tag(nxt, NULL, false); | |
1220 | } | |
74c45052 JA |
1221 | |
1222 | ret = q->mq_ops->queue_rq(hctx, &bd); | |
fc17b653 | 1223 | if (ret == BLK_STS_RESOURCE) { |
6d6f167c JW |
1224 | /* |
1225 | * If an I/O scheduler has been configured and we got a | |
ff821d27 JA |
1226 | * driver tag for the next request already, free it |
1227 | * again. | |
6d6f167c JW |
1228 | */ |
1229 | if (!list_empty(list)) { | |
1230 | nxt = list_first_entry(list, struct request, queuelist); | |
1231 | blk_mq_put_driver_tag(nxt); | |
1232 | } | |
f04c3df3 | 1233 | list_add(&rq->queuelist, list); |
ed0791b2 | 1234 | __blk_mq_requeue_request(rq); |
320ae51f | 1235 | break; |
fc17b653 CH |
1236 | } |
1237 | ||
1238 | if (unlikely(ret != BLK_STS_OK)) { | |
93efe981 | 1239 | errors++; |
2a842aca | 1240 | blk_mq_end_request(rq, BLK_STS_IOERR); |
fc17b653 | 1241 | continue; |
320ae51f JA |
1242 | } |
1243 | ||
fc17b653 | 1244 | queued++; |
81380ca1 | 1245 | } while (!list_empty(list)); |
320ae51f | 1246 | |
703fd1c0 | 1247 | hctx->dispatched[queued_to_index(queued)]++; |
320ae51f JA |
1248 | |
1249 | /* | |
1250 | * Any items that need requeuing? Stuff them into hctx->dispatch, | |
1251 | * that is where we will continue on next queue run. | |
1252 | */ | |
f04c3df3 | 1253 | if (!list_empty(list)) { |
320ae51f | 1254 | spin_lock(&hctx->lock); |
c13660a0 | 1255 | list_splice_init(list, &hctx->dispatch); |
320ae51f | 1256 | spin_unlock(&hctx->lock); |
f04c3df3 | 1257 | |
9ba52e58 | 1258 | /* |
710c785f BVA |
1259 | * If SCHED_RESTART was set by the caller of this function and |
1260 | * it is no longer set that means that it was cleared by another | |
1261 | * thread and hence that a queue rerun is needed. | |
9ba52e58 | 1262 | * |
eb619fdb JA |
1263 | * If 'no_tag' is set, that means that we failed getting |
1264 | * a driver tag with an I/O scheduler attached. If our dispatch | |
1265 | * waitqueue is no longer active, ensure that we run the queue | |
1266 | * AFTER adding our entries back to the list. | |
bd166ef1 | 1267 | * |
710c785f BVA |
1268 | * If no I/O scheduler has been configured it is possible that |
1269 | * the hardware queue got stopped and restarted before requests | |
1270 | * were pushed back onto the dispatch list. Rerun the queue to | |
1271 | * avoid starvation. Notes: | |
1272 | * - blk_mq_run_hw_queue() checks whether or not a queue has | |
1273 | * been stopped before rerunning a queue. | |
1274 | * - Some but not all block drivers stop a queue before | |
fc17b653 | 1275 | * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq |
710c785f | 1276 | * and dm-rq. |
bd166ef1 | 1277 | */ |
eb619fdb JA |
1278 | if (!blk_mq_sched_needs_restart(hctx) || |
1279 | (no_tag && list_empty_careful(&hctx->dispatch_wait.entry))) | |
bd166ef1 | 1280 | blk_mq_run_hw_queue(hctx, true); |
320ae51f | 1281 | } |
f04c3df3 | 1282 | |
93efe981 | 1283 | return (queued + errors) != 0; |
f04c3df3 JA |
1284 | } |
1285 | ||
6a83e74d BVA |
1286 | static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx) |
1287 | { | |
1288 | int srcu_idx; | |
1289 | ||
b7a71e66 JA |
1290 | /* |
1291 | * We should be running this queue from one of the CPUs that | |
1292 | * are mapped to it. | |
1293 | */ | |
6a83e74d BVA |
1294 | WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) && |
1295 | cpu_online(hctx->next_cpu)); | |
1296 | ||
b7a71e66 JA |
1297 | /* |
1298 | * We can't run the queue inline with ints disabled. Ensure that | |
1299 | * we catch bad users of this early. | |
1300 | */ | |
1301 | WARN_ON_ONCE(in_interrupt()); | |
1302 | ||
04ced159 | 1303 | might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING); |
bf4907c0 | 1304 | |
04ced159 JA |
1305 | hctx_lock(hctx, &srcu_idx); |
1306 | blk_mq_sched_dispatch_requests(hctx); | |
1307 | hctx_unlock(hctx, srcu_idx); | |
6a83e74d BVA |
1308 | } |
1309 | ||
506e931f JA |
1310 | /* |
1311 | * It'd be great if the workqueue API had a way to pass | |
1312 | * in a mask and had some smarts for more clever placement. | |
1313 | * For now we just round-robin here, switching for every | |
1314 | * BLK_MQ_CPU_WORK_BATCH queued items. | |
1315 | */ | |
1316 | static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx) | |
1317 | { | |
b657d7e6 CH |
1318 | if (hctx->queue->nr_hw_queues == 1) |
1319 | return WORK_CPU_UNBOUND; | |
506e931f JA |
1320 | |
1321 | if (--hctx->next_cpu_batch <= 0) { | |
c02ebfdd | 1322 | int next_cpu; |
506e931f JA |
1323 | |
1324 | next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask); | |
1325 | if (next_cpu >= nr_cpu_ids) | |
1326 | next_cpu = cpumask_first(hctx->cpumask); | |
1327 | ||
1328 | hctx->next_cpu = next_cpu; | |
1329 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; | |
1330 | } | |
1331 | ||
b657d7e6 | 1332 | return hctx->next_cpu; |
506e931f JA |
1333 | } |
1334 | ||
7587a5ae BVA |
1335 | static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async, |
1336 | unsigned long msecs) | |
320ae51f | 1337 | { |
5435c023 BVA |
1338 | if (WARN_ON_ONCE(!blk_mq_hw_queue_mapped(hctx))) |
1339 | return; | |
1340 | ||
1341 | if (unlikely(blk_mq_hctx_stopped(hctx))) | |
320ae51f JA |
1342 | return; |
1343 | ||
1b792f2f | 1344 | if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) { |
2a90d4aa PB |
1345 | int cpu = get_cpu(); |
1346 | if (cpumask_test_cpu(cpu, hctx->cpumask)) { | |
398205b8 | 1347 | __blk_mq_run_hw_queue(hctx); |
2a90d4aa | 1348 | put_cpu(); |
398205b8 PB |
1349 | return; |
1350 | } | |
e4043dcf | 1351 | |
2a90d4aa | 1352 | put_cpu(); |
e4043dcf | 1353 | } |
398205b8 | 1354 | |
9f993737 JA |
1355 | kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx), |
1356 | &hctx->run_work, | |
1357 | msecs_to_jiffies(msecs)); | |
7587a5ae BVA |
1358 | } |
1359 | ||
1360 | void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs) | |
1361 | { | |
1362 | __blk_mq_delay_run_hw_queue(hctx, true, msecs); | |
1363 | } | |
1364 | EXPORT_SYMBOL(blk_mq_delay_run_hw_queue); | |
1365 | ||
79f720a7 | 1366 | bool blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) |
7587a5ae | 1367 | { |
24f5a90f ML |
1368 | int srcu_idx; |
1369 | bool need_run; | |
1370 | ||
1371 | /* | |
1372 | * When queue is quiesced, we may be switching io scheduler, or | |
1373 | * updating nr_hw_queues, or other things, and we can't run queue | |
1374 | * any more, even __blk_mq_hctx_has_pending() can't be called safely. | |
1375 | * | |
1376 | * And queue will be rerun in blk_mq_unquiesce_queue() if it is | |
1377 | * quiesced. | |
1378 | */ | |
04ced159 JA |
1379 | hctx_lock(hctx, &srcu_idx); |
1380 | need_run = !blk_queue_quiesced(hctx->queue) && | |
1381 | blk_mq_hctx_has_pending(hctx); | |
1382 | hctx_unlock(hctx, srcu_idx); | |
24f5a90f ML |
1383 | |
1384 | if (need_run) { | |
79f720a7 JA |
1385 | __blk_mq_delay_run_hw_queue(hctx, async, 0); |
1386 | return true; | |
1387 | } | |
1388 | ||
1389 | return false; | |
320ae51f | 1390 | } |
5b727272 | 1391 | EXPORT_SYMBOL(blk_mq_run_hw_queue); |
320ae51f | 1392 | |
b94ec296 | 1393 | void blk_mq_run_hw_queues(struct request_queue *q, bool async) |
320ae51f JA |
1394 | { |
1395 | struct blk_mq_hw_ctx *hctx; | |
1396 | int i; | |
1397 | ||
1398 | queue_for_each_hw_ctx(q, hctx, i) { | |
79f720a7 | 1399 | if (blk_mq_hctx_stopped(hctx)) |
320ae51f JA |
1400 | continue; |
1401 | ||
b94ec296 | 1402 | blk_mq_run_hw_queue(hctx, async); |
320ae51f JA |
1403 | } |
1404 | } | |
b94ec296 | 1405 | EXPORT_SYMBOL(blk_mq_run_hw_queues); |
320ae51f | 1406 | |
fd001443 BVA |
1407 | /** |
1408 | * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped | |
1409 | * @q: request queue. | |
1410 | * | |
1411 | * The caller is responsible for serializing this function against | |
1412 | * blk_mq_{start,stop}_hw_queue(). | |
1413 | */ | |
1414 | bool blk_mq_queue_stopped(struct request_queue *q) | |
1415 | { | |
1416 | struct blk_mq_hw_ctx *hctx; | |
1417 | int i; | |
1418 | ||
1419 | queue_for_each_hw_ctx(q, hctx, i) | |
1420 | if (blk_mq_hctx_stopped(hctx)) | |
1421 | return true; | |
1422 | ||
1423 | return false; | |
1424 | } | |
1425 | EXPORT_SYMBOL(blk_mq_queue_stopped); | |
1426 | ||
39a70c76 ML |
1427 | /* |
1428 | * This function is often used for pausing .queue_rq() by driver when | |
1429 | * there isn't enough resource or some conditions aren't satisfied, and | |
4d606219 | 1430 | * BLK_STS_RESOURCE is usually returned. |
39a70c76 ML |
1431 | * |
1432 | * We do not guarantee that dispatch can be drained or blocked | |
1433 | * after blk_mq_stop_hw_queue() returns. Please use | |
1434 | * blk_mq_quiesce_queue() for that requirement. | |
1435 | */ | |
2719aa21 JA |
1436 | void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx) |
1437 | { | |
641a9ed6 | 1438 | cancel_delayed_work(&hctx->run_work); |
280d45f6 | 1439 | |
641a9ed6 | 1440 | set_bit(BLK_MQ_S_STOPPED, &hctx->state); |
2719aa21 | 1441 | } |
641a9ed6 | 1442 | EXPORT_SYMBOL(blk_mq_stop_hw_queue); |
2719aa21 | 1443 | |
39a70c76 ML |
1444 | /* |
1445 | * This function is often used for pausing .queue_rq() by driver when | |
1446 | * there isn't enough resource or some conditions aren't satisfied, and | |
4d606219 | 1447 | * BLK_STS_RESOURCE is usually returned. |
39a70c76 ML |
1448 | * |
1449 | * We do not guarantee that dispatch can be drained or blocked | |
1450 | * after blk_mq_stop_hw_queues() returns. Please use | |
1451 | * blk_mq_quiesce_queue() for that requirement. | |
1452 | */ | |
2719aa21 JA |
1453 | void blk_mq_stop_hw_queues(struct request_queue *q) |
1454 | { | |
641a9ed6 ML |
1455 | struct blk_mq_hw_ctx *hctx; |
1456 | int i; | |
1457 | ||
1458 | queue_for_each_hw_ctx(q, hctx, i) | |
1459 | blk_mq_stop_hw_queue(hctx); | |
280d45f6 CH |
1460 | } |
1461 | EXPORT_SYMBOL(blk_mq_stop_hw_queues); | |
1462 | ||
320ae51f JA |
1463 | void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx) |
1464 | { | |
1465 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
e4043dcf | 1466 | |
0ffbce80 | 1467 | blk_mq_run_hw_queue(hctx, false); |
320ae51f JA |
1468 | } |
1469 | EXPORT_SYMBOL(blk_mq_start_hw_queue); | |
1470 | ||
2f268556 CH |
1471 | void blk_mq_start_hw_queues(struct request_queue *q) |
1472 | { | |
1473 | struct blk_mq_hw_ctx *hctx; | |
1474 | int i; | |
1475 | ||
1476 | queue_for_each_hw_ctx(q, hctx, i) | |
1477 | blk_mq_start_hw_queue(hctx); | |
1478 | } | |
1479 | EXPORT_SYMBOL(blk_mq_start_hw_queues); | |
1480 | ||
ae911c5e JA |
1481 | void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) |
1482 | { | |
1483 | if (!blk_mq_hctx_stopped(hctx)) | |
1484 | return; | |
1485 | ||
1486 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
1487 | blk_mq_run_hw_queue(hctx, async); | |
1488 | } | |
1489 | EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue); | |
1490 | ||
1b4a3258 | 1491 | void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async) |
320ae51f JA |
1492 | { |
1493 | struct blk_mq_hw_ctx *hctx; | |
1494 | int i; | |
1495 | ||
ae911c5e JA |
1496 | queue_for_each_hw_ctx(q, hctx, i) |
1497 | blk_mq_start_stopped_hw_queue(hctx, async); | |
320ae51f JA |
1498 | } |
1499 | EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues); | |
1500 | ||
70f4db63 | 1501 | static void blk_mq_run_work_fn(struct work_struct *work) |
320ae51f JA |
1502 | { |
1503 | struct blk_mq_hw_ctx *hctx; | |
1504 | ||
9f993737 | 1505 | hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work); |
320ae51f | 1506 | |
21c6e939 JA |
1507 | /* |
1508 | * If we are stopped, don't run the queue. The exception is if | |
1509 | * BLK_MQ_S_START_ON_RUN is set. For that case, we auto-clear | |
1510 | * the STOPPED bit and run it. | |
1511 | */ | |
1512 | if (test_bit(BLK_MQ_S_STOPPED, &hctx->state)) { | |
1513 | if (!test_bit(BLK_MQ_S_START_ON_RUN, &hctx->state)) | |
1514 | return; | |
7587a5ae | 1515 | |
21c6e939 JA |
1516 | clear_bit(BLK_MQ_S_START_ON_RUN, &hctx->state); |
1517 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
1518 | } | |
7587a5ae BVA |
1519 | |
1520 | __blk_mq_run_hw_queue(hctx); | |
1521 | } | |
1522 | ||
70f4db63 CH |
1523 | |
1524 | void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs) | |
1525 | { | |
5435c023 | 1526 | if (WARN_ON_ONCE(!blk_mq_hw_queue_mapped(hctx))) |
19c66e59 | 1527 | return; |
70f4db63 | 1528 | |
21c6e939 JA |
1529 | /* |
1530 | * Stop the hw queue, then modify currently delayed work. | |
1531 | * This should prevent us from running the queue prematurely. | |
1532 | * Mark the queue as auto-clearing STOPPED when it runs. | |
1533 | */ | |
7e79dadc | 1534 | blk_mq_stop_hw_queue(hctx); |
21c6e939 JA |
1535 | set_bit(BLK_MQ_S_START_ON_RUN, &hctx->state); |
1536 | kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), | |
1537 | &hctx->run_work, | |
1538 | msecs_to_jiffies(msecs)); | |
70f4db63 CH |
1539 | } |
1540 | EXPORT_SYMBOL(blk_mq_delay_queue); | |
1541 | ||
cfd0c552 | 1542 | static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx, |
cfd0c552 ML |
1543 | struct request *rq, |
1544 | bool at_head) | |
320ae51f | 1545 | { |
e57690fe JA |
1546 | struct blk_mq_ctx *ctx = rq->mq_ctx; |
1547 | ||
7b607814 BVA |
1548 | lockdep_assert_held(&ctx->lock); |
1549 | ||
01b983c9 JA |
1550 | trace_block_rq_insert(hctx->queue, rq); |
1551 | ||
72a0a36e CH |
1552 | if (at_head) |
1553 | list_add(&rq->queuelist, &ctx->rq_list); | |
1554 | else | |
1555 | list_add_tail(&rq->queuelist, &ctx->rq_list); | |
cfd0c552 | 1556 | } |
4bb659b1 | 1557 | |
2c3ad667 JA |
1558 | void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq, |
1559 | bool at_head) | |
cfd0c552 ML |
1560 | { |
1561 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
1562 | ||
7b607814 BVA |
1563 | lockdep_assert_held(&ctx->lock); |
1564 | ||
e57690fe | 1565 | __blk_mq_insert_req_list(hctx, rq, at_head); |
320ae51f | 1566 | blk_mq_hctx_mark_pending(hctx, ctx); |
320ae51f JA |
1567 | } |
1568 | ||
157f377b JA |
1569 | /* |
1570 | * Should only be used carefully, when the caller knows we want to | |
1571 | * bypass a potential IO scheduler on the target device. | |
1572 | */ | |
b0850297 | 1573 | void blk_mq_request_bypass_insert(struct request *rq, bool run_queue) |
157f377b JA |
1574 | { |
1575 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
1576 | struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(rq->q, ctx->cpu); | |
1577 | ||
1578 | spin_lock(&hctx->lock); | |
1579 | list_add_tail(&rq->queuelist, &hctx->dispatch); | |
1580 | spin_unlock(&hctx->lock); | |
1581 | ||
b0850297 ML |
1582 | if (run_queue) |
1583 | blk_mq_run_hw_queue(hctx, false); | |
157f377b JA |
1584 | } |
1585 | ||
bd166ef1 JA |
1586 | void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx, |
1587 | struct list_head *list) | |
320ae51f JA |
1588 | |
1589 | { | |
320ae51f JA |
1590 | /* |
1591 | * preemption doesn't flush plug list, so it's possible ctx->cpu is | |
1592 | * offline now | |
1593 | */ | |
1594 | spin_lock(&ctx->lock); | |
1595 | while (!list_empty(list)) { | |
1596 | struct request *rq; | |
1597 | ||
1598 | rq = list_first_entry(list, struct request, queuelist); | |
e57690fe | 1599 | BUG_ON(rq->mq_ctx != ctx); |
320ae51f | 1600 | list_del_init(&rq->queuelist); |
e57690fe | 1601 | __blk_mq_insert_req_list(hctx, rq, false); |
320ae51f | 1602 | } |
cfd0c552 | 1603 | blk_mq_hctx_mark_pending(hctx, ctx); |
320ae51f | 1604 | spin_unlock(&ctx->lock); |
320ae51f JA |
1605 | } |
1606 | ||
1607 | static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b) | |
1608 | { | |
1609 | struct request *rqa = container_of(a, struct request, queuelist); | |
1610 | struct request *rqb = container_of(b, struct request, queuelist); | |
1611 | ||
1612 | return !(rqa->mq_ctx < rqb->mq_ctx || | |
1613 | (rqa->mq_ctx == rqb->mq_ctx && | |
1614 | blk_rq_pos(rqa) < blk_rq_pos(rqb))); | |
1615 | } | |
1616 | ||
1617 | void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule) | |
1618 | { | |
1619 | struct blk_mq_ctx *this_ctx; | |
1620 | struct request_queue *this_q; | |
1621 | struct request *rq; | |
1622 | LIST_HEAD(list); | |
1623 | LIST_HEAD(ctx_list); | |
1624 | unsigned int depth; | |
1625 | ||
1626 | list_splice_init(&plug->mq_list, &list); | |
1627 | ||
1628 | list_sort(NULL, &list, plug_ctx_cmp); | |
1629 | ||
1630 | this_q = NULL; | |
1631 | this_ctx = NULL; | |
1632 | depth = 0; | |
1633 | ||
1634 | while (!list_empty(&list)) { | |
1635 | rq = list_entry_rq(list.next); | |
1636 | list_del_init(&rq->queuelist); | |
1637 | BUG_ON(!rq->q); | |
1638 | if (rq->mq_ctx != this_ctx) { | |
1639 | if (this_ctx) { | |
bd166ef1 JA |
1640 | trace_block_unplug(this_q, depth, from_schedule); |
1641 | blk_mq_sched_insert_requests(this_q, this_ctx, | |
1642 | &ctx_list, | |
1643 | from_schedule); | |
320ae51f JA |
1644 | } |
1645 | ||
1646 | this_ctx = rq->mq_ctx; | |
1647 | this_q = rq->q; | |
1648 | depth = 0; | |
1649 | } | |
1650 | ||
1651 | depth++; | |
1652 | list_add_tail(&rq->queuelist, &ctx_list); | |
1653 | } | |
1654 | ||
1655 | /* | |
1656 | * If 'this_ctx' is set, we know we have entries to complete | |
1657 | * on 'ctx_list'. Do those. | |
1658 | */ | |
1659 | if (this_ctx) { | |
bd166ef1 JA |
1660 | trace_block_unplug(this_q, depth, from_schedule); |
1661 | blk_mq_sched_insert_requests(this_q, this_ctx, &ctx_list, | |
1662 | from_schedule); | |
320ae51f JA |
1663 | } |
1664 | } | |
1665 | ||
1666 | static void blk_mq_bio_to_request(struct request *rq, struct bio *bio) | |
1667 | { | |
da8d7f07 | 1668 | blk_init_request_from_bio(rq, bio); |
4b570521 | 1669 | |
85acb3ba SL |
1670 | blk_rq_set_rl(rq, blk_get_rl(rq->q, bio)); |
1671 | ||
6e85eaf3 | 1672 | blk_account_io_start(rq, true); |
320ae51f JA |
1673 | } |
1674 | ||
ab42f35d ML |
1675 | static inline void blk_mq_queue_io(struct blk_mq_hw_ctx *hctx, |
1676 | struct blk_mq_ctx *ctx, | |
1677 | struct request *rq) | |
1678 | { | |
1679 | spin_lock(&ctx->lock); | |
1680 | __blk_mq_insert_request(hctx, rq, false); | |
1681 | spin_unlock(&ctx->lock); | |
07068d5b | 1682 | } |
14ec77f3 | 1683 | |
fd2d3326 JA |
1684 | static blk_qc_t request_to_qc_t(struct blk_mq_hw_ctx *hctx, struct request *rq) |
1685 | { | |
bd166ef1 JA |
1686 | if (rq->tag != -1) |
1687 | return blk_tag_to_qc_t(rq->tag, hctx->queue_num, false); | |
1688 | ||
1689 | return blk_tag_to_qc_t(rq->internal_tag, hctx->queue_num, true); | |
fd2d3326 JA |
1690 | } |
1691 | ||
d964f04a ML |
1692 | static void __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx, |
1693 | struct request *rq, | |
04ced159 | 1694 | blk_qc_t *cookie) |
f984df1f | 1695 | { |
f984df1f | 1696 | struct request_queue *q = rq->q; |
f984df1f SL |
1697 | struct blk_mq_queue_data bd = { |
1698 | .rq = rq, | |
d945a365 | 1699 | .last = true, |
f984df1f | 1700 | }; |
bd166ef1 | 1701 | blk_qc_t new_cookie; |
f06345ad | 1702 | blk_status_t ret; |
d964f04a ML |
1703 | bool run_queue = true; |
1704 | ||
f4560ffe ML |
1705 | /* RCU or SRCU read lock is needed before checking quiesced flag */ |
1706 | if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)) { | |
d964f04a ML |
1707 | run_queue = false; |
1708 | goto insert; | |
1709 | } | |
f984df1f | 1710 | |
bd166ef1 | 1711 | if (q->elevator) |
2253efc8 BVA |
1712 | goto insert; |
1713 | ||
d964f04a | 1714 | if (!blk_mq_get_driver_tag(rq, NULL, false)) |
bd166ef1 JA |
1715 | goto insert; |
1716 | ||
88022d72 | 1717 | if (!blk_mq_get_dispatch_budget(hctx)) { |
de148297 ML |
1718 | blk_mq_put_driver_tag(rq); |
1719 | goto insert; | |
88022d72 | 1720 | } |
de148297 | 1721 | |
bd166ef1 JA |
1722 | new_cookie = request_to_qc_t(hctx, rq); |
1723 | ||
f984df1f SL |
1724 | /* |
1725 | * For OK queue, we are done. For error, kill it. Any other | |
1726 | * error (busy), just add it to our list as we previously | |
1727 | * would have done | |
1728 | */ | |
1729 | ret = q->mq_ops->queue_rq(hctx, &bd); | |
fc17b653 CH |
1730 | switch (ret) { |
1731 | case BLK_STS_OK: | |
7b371636 | 1732 | *cookie = new_cookie; |
2253efc8 | 1733 | return; |
fc17b653 CH |
1734 | case BLK_STS_RESOURCE: |
1735 | __blk_mq_requeue_request(rq); | |
1736 | goto insert; | |
1737 | default: | |
7b371636 | 1738 | *cookie = BLK_QC_T_NONE; |
fc17b653 | 1739 | blk_mq_end_request(rq, ret); |
2253efc8 | 1740 | return; |
f984df1f | 1741 | } |
7b371636 | 1742 | |
2253efc8 | 1743 | insert: |
04ced159 JA |
1744 | blk_mq_sched_insert_request(rq, false, run_queue, false, |
1745 | hctx->flags & BLK_MQ_F_BLOCKING); | |
f984df1f SL |
1746 | } |
1747 | ||
5eb6126e CH |
1748 | static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx, |
1749 | struct request *rq, blk_qc_t *cookie) | |
1750 | { | |
04ced159 | 1751 | int srcu_idx; |
bf4907c0 | 1752 | |
04ced159 | 1753 | might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING); |
bf4907c0 | 1754 | |
04ced159 JA |
1755 | hctx_lock(hctx, &srcu_idx); |
1756 | __blk_mq_try_issue_directly(hctx, rq, cookie); | |
1757 | hctx_unlock(hctx, srcu_idx); | |
5eb6126e CH |
1758 | } |
1759 | ||
dece1635 | 1760 | static blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio) |
07068d5b | 1761 | { |
ef295ecf | 1762 | const int is_sync = op_is_sync(bio->bi_opf); |
f73f44eb | 1763 | const int is_flush_fua = op_is_flush(bio->bi_opf); |
5a797e00 | 1764 | struct blk_mq_alloc_data data = { .flags = 0 }; |
07068d5b | 1765 | struct request *rq; |
5eb6126e | 1766 | unsigned int request_count = 0; |
f984df1f | 1767 | struct blk_plug *plug; |
5b3f341f | 1768 | struct request *same_queue_rq = NULL; |
7b371636 | 1769 | blk_qc_t cookie; |
87760e5e | 1770 | unsigned int wb_acct; |
07068d5b JA |
1771 | |
1772 | blk_queue_bounce(q, &bio); | |
1773 | ||
af67c31f | 1774 | blk_queue_split(q, &bio); |
f36ea50c | 1775 | |
e23947bd | 1776 | if (!bio_integrity_prep(bio)) |
dece1635 | 1777 | return BLK_QC_T_NONE; |
07068d5b | 1778 | |
87c279e6 OS |
1779 | if (!is_flush_fua && !blk_queue_nomerges(q) && |
1780 | blk_attempt_plug_merge(q, bio, &request_count, &same_queue_rq)) | |
1781 | return BLK_QC_T_NONE; | |
f984df1f | 1782 | |
bd166ef1 JA |
1783 | if (blk_mq_sched_bio_merge(q, bio)) |
1784 | return BLK_QC_T_NONE; | |
1785 | ||
87760e5e JA |
1786 | wb_acct = wbt_wait(q->rq_wb, bio, NULL); |
1787 | ||
bd166ef1 JA |
1788 | trace_block_getrq(q, bio, bio->bi_opf); |
1789 | ||
d2c0d383 | 1790 | rq = blk_mq_get_request(q, bio, bio->bi_opf, &data); |
87760e5e JA |
1791 | if (unlikely(!rq)) { |
1792 | __wbt_done(q->rq_wb, wb_acct); | |
03a07c92 GR |
1793 | if (bio->bi_opf & REQ_NOWAIT) |
1794 | bio_wouldblock_error(bio); | |
dece1635 | 1795 | return BLK_QC_T_NONE; |
87760e5e JA |
1796 | } |
1797 | ||
1798 | wbt_track(&rq->issue_stat, wb_acct); | |
07068d5b | 1799 | |
fd2d3326 | 1800 | cookie = request_to_qc_t(data.hctx, rq); |
07068d5b | 1801 | |
f984df1f | 1802 | plug = current->plug; |
07068d5b | 1803 | if (unlikely(is_flush_fua)) { |
f984df1f | 1804 | blk_mq_put_ctx(data.ctx); |
07068d5b | 1805 | blk_mq_bio_to_request(rq, bio); |
923218f6 ML |
1806 | |
1807 | /* bypass scheduler for flush rq */ | |
1808 | blk_insert_flush(rq); | |
1809 | blk_mq_run_hw_queue(data.hctx, true); | |
a4d907b6 | 1810 | } else if (plug && q->nr_hw_queues == 1) { |
600271d9 SL |
1811 | struct request *last = NULL; |
1812 | ||
b00c53e8 | 1813 | blk_mq_put_ctx(data.ctx); |
e6c4438b | 1814 | blk_mq_bio_to_request(rq, bio); |
0a6219a9 ML |
1815 | |
1816 | /* | |
1817 | * @request_count may become stale because of schedule | |
1818 | * out, so check the list again. | |
1819 | */ | |
1820 | if (list_empty(&plug->mq_list)) | |
1821 | request_count = 0; | |
254d259d CH |
1822 | else if (blk_queue_nomerges(q)) |
1823 | request_count = blk_plug_queued_count(q); | |
1824 | ||
676d0607 | 1825 | if (!request_count) |
e6c4438b | 1826 | trace_block_plug(q); |
600271d9 SL |
1827 | else |
1828 | last = list_entry_rq(plug->mq_list.prev); | |
b094f89c | 1829 | |
600271d9 SL |
1830 | if (request_count >= BLK_MAX_REQUEST_COUNT || (last && |
1831 | blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) { | |
e6c4438b JM |
1832 | blk_flush_plug_list(plug, false); |
1833 | trace_block_plug(q); | |
320ae51f | 1834 | } |
b094f89c | 1835 | |
e6c4438b | 1836 | list_add_tail(&rq->queuelist, &plug->mq_list); |
2299722c | 1837 | } else if (plug && !blk_queue_nomerges(q)) { |
bd166ef1 | 1838 | blk_mq_bio_to_request(rq, bio); |
07068d5b | 1839 | |
07068d5b | 1840 | /* |
6a83e74d | 1841 | * We do limited plugging. If the bio can be merged, do that. |
f984df1f SL |
1842 | * Otherwise the existing request in the plug list will be |
1843 | * issued. So the plug list will have one request at most | |
2299722c CH |
1844 | * The plug list might get flushed before this. If that happens, |
1845 | * the plug list is empty, and same_queue_rq is invalid. | |
07068d5b | 1846 | */ |
2299722c CH |
1847 | if (list_empty(&plug->mq_list)) |
1848 | same_queue_rq = NULL; | |
1849 | if (same_queue_rq) | |
1850 | list_del_init(&same_queue_rq->queuelist); | |
1851 | list_add_tail(&rq->queuelist, &plug->mq_list); | |
1852 | ||
bf4907c0 JA |
1853 | blk_mq_put_ctx(data.ctx); |
1854 | ||
dad7a3be ML |
1855 | if (same_queue_rq) { |
1856 | data.hctx = blk_mq_map_queue(q, | |
1857 | same_queue_rq->mq_ctx->cpu); | |
2299722c CH |
1858 | blk_mq_try_issue_directly(data.hctx, same_queue_rq, |
1859 | &cookie); | |
dad7a3be | 1860 | } |
a4d907b6 | 1861 | } else if (q->nr_hw_queues > 1 && is_sync) { |
bf4907c0 | 1862 | blk_mq_put_ctx(data.ctx); |
2299722c | 1863 | blk_mq_bio_to_request(rq, bio); |
2299722c | 1864 | blk_mq_try_issue_directly(data.hctx, rq, &cookie); |
a4d907b6 | 1865 | } else if (q->elevator) { |
b00c53e8 | 1866 | blk_mq_put_ctx(data.ctx); |
bd166ef1 | 1867 | blk_mq_bio_to_request(rq, bio); |
a4d907b6 | 1868 | blk_mq_sched_insert_request(rq, false, true, true, true); |
ab42f35d | 1869 | } else { |
b00c53e8 | 1870 | blk_mq_put_ctx(data.ctx); |
ab42f35d ML |
1871 | blk_mq_bio_to_request(rq, bio); |
1872 | blk_mq_queue_io(data.hctx, data.ctx, rq); | |
a4d907b6 | 1873 | blk_mq_run_hw_queue(data.hctx, true); |
ab42f35d | 1874 | } |
320ae51f | 1875 | |
7b371636 | 1876 | return cookie; |
320ae51f JA |
1877 | } |
1878 | ||
cc71a6f4 JA |
1879 | void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags, |
1880 | unsigned int hctx_idx) | |
95363efd | 1881 | { |
e9b267d9 | 1882 | struct page *page; |
320ae51f | 1883 | |
24d2f903 | 1884 | if (tags->rqs && set->ops->exit_request) { |
e9b267d9 | 1885 | int i; |
320ae51f | 1886 | |
24d2f903 | 1887 | for (i = 0; i < tags->nr_tags; i++) { |
2af8cbe3 JA |
1888 | struct request *rq = tags->static_rqs[i]; |
1889 | ||
1890 | if (!rq) | |
e9b267d9 | 1891 | continue; |
d6296d39 | 1892 | set->ops->exit_request(set, rq, hctx_idx); |
2af8cbe3 | 1893 | tags->static_rqs[i] = NULL; |
e9b267d9 | 1894 | } |
320ae51f | 1895 | } |
320ae51f | 1896 | |
24d2f903 CH |
1897 | while (!list_empty(&tags->page_list)) { |
1898 | page = list_first_entry(&tags->page_list, struct page, lru); | |
6753471c | 1899 | list_del_init(&page->lru); |
f75782e4 CM |
1900 | /* |
1901 | * Remove kmemleak object previously allocated in | |
1902 | * blk_mq_init_rq_map(). | |
1903 | */ | |
1904 | kmemleak_free(page_address(page)); | |
320ae51f JA |
1905 | __free_pages(page, page->private); |
1906 | } | |
cc71a6f4 | 1907 | } |
320ae51f | 1908 | |
cc71a6f4 JA |
1909 | void blk_mq_free_rq_map(struct blk_mq_tags *tags) |
1910 | { | |
24d2f903 | 1911 | kfree(tags->rqs); |
cc71a6f4 | 1912 | tags->rqs = NULL; |
2af8cbe3 JA |
1913 | kfree(tags->static_rqs); |
1914 | tags->static_rqs = NULL; | |
320ae51f | 1915 | |
24d2f903 | 1916 | blk_mq_free_tags(tags); |
320ae51f JA |
1917 | } |
1918 | ||
cc71a6f4 JA |
1919 | struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, |
1920 | unsigned int hctx_idx, | |
1921 | unsigned int nr_tags, | |
1922 | unsigned int reserved_tags) | |
320ae51f | 1923 | { |
24d2f903 | 1924 | struct blk_mq_tags *tags; |
59f082e4 | 1925 | int node; |
320ae51f | 1926 | |
59f082e4 SL |
1927 | node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx); |
1928 | if (node == NUMA_NO_NODE) | |
1929 | node = set->numa_node; | |
1930 | ||
1931 | tags = blk_mq_init_tags(nr_tags, reserved_tags, node, | |
24391c0d | 1932 | BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags)); |
24d2f903 CH |
1933 | if (!tags) |
1934 | return NULL; | |
320ae51f | 1935 | |
cc71a6f4 | 1936 | tags->rqs = kzalloc_node(nr_tags * sizeof(struct request *), |
36e1f3d1 | 1937 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, |
59f082e4 | 1938 | node); |
24d2f903 CH |
1939 | if (!tags->rqs) { |
1940 | blk_mq_free_tags(tags); | |
1941 | return NULL; | |
1942 | } | |
320ae51f | 1943 | |
2af8cbe3 JA |
1944 | tags->static_rqs = kzalloc_node(nr_tags * sizeof(struct request *), |
1945 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, | |
59f082e4 | 1946 | node); |
2af8cbe3 JA |
1947 | if (!tags->static_rqs) { |
1948 | kfree(tags->rqs); | |
1949 | blk_mq_free_tags(tags); | |
1950 | return NULL; | |
1951 | } | |
1952 | ||
cc71a6f4 JA |
1953 | return tags; |
1954 | } | |
1955 | ||
1956 | static size_t order_to_size(unsigned int order) | |
1957 | { | |
1958 | return (size_t)PAGE_SIZE << order; | |
1959 | } | |
1960 | ||
1d9bd516 TH |
1961 | static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq, |
1962 | unsigned int hctx_idx, int node) | |
1963 | { | |
1964 | int ret; | |
1965 | ||
1966 | if (set->ops->init_request) { | |
1967 | ret = set->ops->init_request(set, rq, hctx_idx, node); | |
1968 | if (ret) | |
1969 | return ret; | |
1970 | } | |
1971 | ||
1972 | seqcount_init(&rq->gstate_seq); | |
1973 | u64_stats_init(&rq->aborted_gstate_sync); | |
1974 | return 0; | |
1975 | } | |
1976 | ||
cc71a6f4 JA |
1977 | int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags, |
1978 | unsigned int hctx_idx, unsigned int depth) | |
1979 | { | |
1980 | unsigned int i, j, entries_per_page, max_order = 4; | |
1981 | size_t rq_size, left; | |
59f082e4 SL |
1982 | int node; |
1983 | ||
1984 | node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx); | |
1985 | if (node == NUMA_NO_NODE) | |
1986 | node = set->numa_node; | |
cc71a6f4 JA |
1987 | |
1988 | INIT_LIST_HEAD(&tags->page_list); | |
1989 | ||
320ae51f JA |
1990 | /* |
1991 | * rq_size is the size of the request plus driver payload, rounded | |
1992 | * to the cacheline size | |
1993 | */ | |
24d2f903 | 1994 | rq_size = round_up(sizeof(struct request) + set->cmd_size, |
320ae51f | 1995 | cache_line_size()); |
cc71a6f4 | 1996 | left = rq_size * depth; |
320ae51f | 1997 | |
cc71a6f4 | 1998 | for (i = 0; i < depth; ) { |
320ae51f JA |
1999 | int this_order = max_order; |
2000 | struct page *page; | |
2001 | int to_do; | |
2002 | void *p; | |
2003 | ||
b3a834b1 | 2004 | while (this_order && left < order_to_size(this_order - 1)) |
320ae51f JA |
2005 | this_order--; |
2006 | ||
2007 | do { | |
59f082e4 | 2008 | page = alloc_pages_node(node, |
36e1f3d1 | 2009 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO, |
a5164405 | 2010 | this_order); |
320ae51f JA |
2011 | if (page) |
2012 | break; | |
2013 | if (!this_order--) | |
2014 | break; | |
2015 | if (order_to_size(this_order) < rq_size) | |
2016 | break; | |
2017 | } while (1); | |
2018 | ||
2019 | if (!page) | |
24d2f903 | 2020 | goto fail; |
320ae51f JA |
2021 | |
2022 | page->private = this_order; | |
24d2f903 | 2023 | list_add_tail(&page->lru, &tags->page_list); |
320ae51f JA |
2024 | |
2025 | p = page_address(page); | |
f75782e4 CM |
2026 | /* |
2027 | * Allow kmemleak to scan these pages as they contain pointers | |
2028 | * to additional allocations like via ops->init_request(). | |
2029 | */ | |
36e1f3d1 | 2030 | kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO); |
320ae51f | 2031 | entries_per_page = order_to_size(this_order) / rq_size; |
cc71a6f4 | 2032 | to_do = min(entries_per_page, depth - i); |
320ae51f JA |
2033 | left -= to_do * rq_size; |
2034 | for (j = 0; j < to_do; j++) { | |
2af8cbe3 JA |
2035 | struct request *rq = p; |
2036 | ||
2037 | tags->static_rqs[i] = rq; | |
1d9bd516 TH |
2038 | if (blk_mq_init_request(set, rq, hctx_idx, node)) { |
2039 | tags->static_rqs[i] = NULL; | |
2040 | goto fail; | |
e9b267d9 CH |
2041 | } |
2042 | ||
320ae51f JA |
2043 | p += rq_size; |
2044 | i++; | |
2045 | } | |
2046 | } | |
cc71a6f4 | 2047 | return 0; |
320ae51f | 2048 | |
24d2f903 | 2049 | fail: |
cc71a6f4 JA |
2050 | blk_mq_free_rqs(set, tags, hctx_idx); |
2051 | return -ENOMEM; | |
320ae51f JA |
2052 | } |
2053 | ||
e57690fe JA |
2054 | /* |
2055 | * 'cpu' is going away. splice any existing rq_list entries from this | |
2056 | * software queue to the hw queue dispatch list, and ensure that it | |
2057 | * gets run. | |
2058 | */ | |
9467f859 | 2059 | static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node) |
484b4061 | 2060 | { |
9467f859 | 2061 | struct blk_mq_hw_ctx *hctx; |
484b4061 JA |
2062 | struct blk_mq_ctx *ctx; |
2063 | LIST_HEAD(tmp); | |
2064 | ||
9467f859 | 2065 | hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead); |
e57690fe | 2066 | ctx = __blk_mq_get_ctx(hctx->queue, cpu); |
484b4061 JA |
2067 | |
2068 | spin_lock(&ctx->lock); | |
2069 | if (!list_empty(&ctx->rq_list)) { | |
2070 | list_splice_init(&ctx->rq_list, &tmp); | |
2071 | blk_mq_hctx_clear_pending(hctx, ctx); | |
2072 | } | |
2073 | spin_unlock(&ctx->lock); | |
2074 | ||
2075 | if (list_empty(&tmp)) | |
9467f859 | 2076 | return 0; |
484b4061 | 2077 | |
e57690fe JA |
2078 | spin_lock(&hctx->lock); |
2079 | list_splice_tail_init(&tmp, &hctx->dispatch); | |
2080 | spin_unlock(&hctx->lock); | |
484b4061 JA |
2081 | |
2082 | blk_mq_run_hw_queue(hctx, true); | |
9467f859 | 2083 | return 0; |
484b4061 JA |
2084 | } |
2085 | ||
9467f859 | 2086 | static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx) |
484b4061 | 2087 | { |
9467f859 TG |
2088 | cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD, |
2089 | &hctx->cpuhp_dead); | |
484b4061 JA |
2090 | } |
2091 | ||
c3b4afca | 2092 | /* hctx->ctxs will be freed in queue's release handler */ |
08e98fc6 ML |
2093 | static void blk_mq_exit_hctx(struct request_queue *q, |
2094 | struct blk_mq_tag_set *set, | |
2095 | struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx) | |
2096 | { | |
9c1051aa OS |
2097 | blk_mq_debugfs_unregister_hctx(hctx); |
2098 | ||
8ab0b7dc ML |
2099 | if (blk_mq_hw_queue_mapped(hctx)) |
2100 | blk_mq_tag_idle(hctx); | |
08e98fc6 | 2101 | |
f70ced09 | 2102 | if (set->ops->exit_request) |
d6296d39 | 2103 | set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx); |
f70ced09 | 2104 | |
93252632 OS |
2105 | blk_mq_sched_exit_hctx(q, hctx, hctx_idx); |
2106 | ||
08e98fc6 ML |
2107 | if (set->ops->exit_hctx) |
2108 | set->ops->exit_hctx(hctx, hctx_idx); | |
2109 | ||
6a83e74d | 2110 | if (hctx->flags & BLK_MQ_F_BLOCKING) |
05707b64 | 2111 | cleanup_srcu_struct(hctx->srcu); |
6a83e74d | 2112 | |
9467f859 | 2113 | blk_mq_remove_cpuhp(hctx); |
f70ced09 | 2114 | blk_free_flush_queue(hctx->fq); |
88459642 | 2115 | sbitmap_free(&hctx->ctx_map); |
08e98fc6 ML |
2116 | } |
2117 | ||
624dbe47 ML |
2118 | static void blk_mq_exit_hw_queues(struct request_queue *q, |
2119 | struct blk_mq_tag_set *set, int nr_queue) | |
2120 | { | |
2121 | struct blk_mq_hw_ctx *hctx; | |
2122 | unsigned int i; | |
2123 | ||
2124 | queue_for_each_hw_ctx(q, hctx, i) { | |
2125 | if (i == nr_queue) | |
2126 | break; | |
08e98fc6 | 2127 | blk_mq_exit_hctx(q, set, hctx, i); |
624dbe47 | 2128 | } |
624dbe47 ML |
2129 | } |
2130 | ||
08e98fc6 ML |
2131 | static int blk_mq_init_hctx(struct request_queue *q, |
2132 | struct blk_mq_tag_set *set, | |
2133 | struct blk_mq_hw_ctx *hctx, unsigned hctx_idx) | |
320ae51f | 2134 | { |
08e98fc6 ML |
2135 | int node; |
2136 | ||
2137 | node = hctx->numa_node; | |
2138 | if (node == NUMA_NO_NODE) | |
2139 | node = hctx->numa_node = set->numa_node; | |
2140 | ||
9f993737 | 2141 | INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn); |
08e98fc6 ML |
2142 | spin_lock_init(&hctx->lock); |
2143 | INIT_LIST_HEAD(&hctx->dispatch); | |
2144 | hctx->queue = q; | |
2404e607 | 2145 | hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED; |
08e98fc6 | 2146 | |
9467f859 | 2147 | cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead); |
08e98fc6 ML |
2148 | |
2149 | hctx->tags = set->tags[hctx_idx]; | |
320ae51f JA |
2150 | |
2151 | /* | |
08e98fc6 ML |
2152 | * Allocate space for all possible cpus to avoid allocation at |
2153 | * runtime | |
320ae51f | 2154 | */ |
d904bfa7 | 2155 | hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *), |
08e98fc6 ML |
2156 | GFP_KERNEL, node); |
2157 | if (!hctx->ctxs) | |
2158 | goto unregister_cpu_notifier; | |
320ae51f | 2159 | |
88459642 OS |
2160 | if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8), GFP_KERNEL, |
2161 | node)) | |
08e98fc6 | 2162 | goto free_ctxs; |
320ae51f | 2163 | |
08e98fc6 | 2164 | hctx->nr_ctx = 0; |
320ae51f | 2165 | |
eb619fdb JA |
2166 | init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake); |
2167 | INIT_LIST_HEAD(&hctx->dispatch_wait.entry); | |
2168 | ||
08e98fc6 ML |
2169 | if (set->ops->init_hctx && |
2170 | set->ops->init_hctx(hctx, set->driver_data, hctx_idx)) | |
2171 | goto free_bitmap; | |
320ae51f | 2172 | |
93252632 OS |
2173 | if (blk_mq_sched_init_hctx(q, hctx, hctx_idx)) |
2174 | goto exit_hctx; | |
2175 | ||
f70ced09 ML |
2176 | hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size); |
2177 | if (!hctx->fq) | |
93252632 | 2178 | goto sched_exit_hctx; |
320ae51f | 2179 | |
1d9bd516 | 2180 | if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx, node)) |
f70ced09 | 2181 | goto free_fq; |
320ae51f | 2182 | |
6a83e74d | 2183 | if (hctx->flags & BLK_MQ_F_BLOCKING) |
05707b64 | 2184 | init_srcu_struct(hctx->srcu); |
6a83e74d | 2185 | |
9c1051aa OS |
2186 | blk_mq_debugfs_register_hctx(q, hctx); |
2187 | ||
08e98fc6 | 2188 | return 0; |
320ae51f | 2189 | |
f70ced09 ML |
2190 | free_fq: |
2191 | kfree(hctx->fq); | |
93252632 OS |
2192 | sched_exit_hctx: |
2193 | blk_mq_sched_exit_hctx(q, hctx, hctx_idx); | |
f70ced09 ML |
2194 | exit_hctx: |
2195 | if (set->ops->exit_hctx) | |
2196 | set->ops->exit_hctx(hctx, hctx_idx); | |
08e98fc6 | 2197 | free_bitmap: |
88459642 | 2198 | sbitmap_free(&hctx->ctx_map); |
08e98fc6 ML |
2199 | free_ctxs: |
2200 | kfree(hctx->ctxs); | |
2201 | unregister_cpu_notifier: | |
9467f859 | 2202 | blk_mq_remove_cpuhp(hctx); |
08e98fc6 ML |
2203 | return -1; |
2204 | } | |
320ae51f | 2205 | |
320ae51f JA |
2206 | static void blk_mq_init_cpu_queues(struct request_queue *q, |
2207 | unsigned int nr_hw_queues) | |
2208 | { | |
2209 | unsigned int i; | |
2210 | ||
2211 | for_each_possible_cpu(i) { | |
2212 | struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i); | |
2213 | struct blk_mq_hw_ctx *hctx; | |
2214 | ||
320ae51f JA |
2215 | __ctx->cpu = i; |
2216 | spin_lock_init(&__ctx->lock); | |
2217 | INIT_LIST_HEAD(&__ctx->rq_list); | |
2218 | __ctx->queue = q; | |
2219 | ||
4b855ad3 CH |
2220 | /* If the cpu isn't present, the cpu is mapped to first hctx */ |
2221 | if (!cpu_present(i)) | |
320ae51f JA |
2222 | continue; |
2223 | ||
7d7e0f90 | 2224 | hctx = blk_mq_map_queue(q, i); |
e4043dcf | 2225 | |
320ae51f JA |
2226 | /* |
2227 | * Set local node, IFF we have more than one hw queue. If | |
2228 | * not, we remain on the home node of the device | |
2229 | */ | |
2230 | if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE) | |
bffed457 | 2231 | hctx->numa_node = local_memory_node(cpu_to_node(i)); |
320ae51f JA |
2232 | } |
2233 | } | |
2234 | ||
cc71a6f4 JA |
2235 | static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, int hctx_idx) |
2236 | { | |
2237 | int ret = 0; | |
2238 | ||
2239 | set->tags[hctx_idx] = blk_mq_alloc_rq_map(set, hctx_idx, | |
2240 | set->queue_depth, set->reserved_tags); | |
2241 | if (!set->tags[hctx_idx]) | |
2242 | return false; | |
2243 | ||
2244 | ret = blk_mq_alloc_rqs(set, set->tags[hctx_idx], hctx_idx, | |
2245 | set->queue_depth); | |
2246 | if (!ret) | |
2247 | return true; | |
2248 | ||
2249 | blk_mq_free_rq_map(set->tags[hctx_idx]); | |
2250 | set->tags[hctx_idx] = NULL; | |
2251 | return false; | |
2252 | } | |
2253 | ||
2254 | static void blk_mq_free_map_and_requests(struct blk_mq_tag_set *set, | |
2255 | unsigned int hctx_idx) | |
2256 | { | |
bd166ef1 JA |
2257 | if (set->tags[hctx_idx]) { |
2258 | blk_mq_free_rqs(set, set->tags[hctx_idx], hctx_idx); | |
2259 | blk_mq_free_rq_map(set->tags[hctx_idx]); | |
2260 | set->tags[hctx_idx] = NULL; | |
2261 | } | |
cc71a6f4 JA |
2262 | } |
2263 | ||
4b855ad3 | 2264 | static void blk_mq_map_swqueue(struct request_queue *q) |
320ae51f | 2265 | { |
d1b1cea1 | 2266 | unsigned int i, hctx_idx; |
320ae51f JA |
2267 | struct blk_mq_hw_ctx *hctx; |
2268 | struct blk_mq_ctx *ctx; | |
2a34c087 | 2269 | struct blk_mq_tag_set *set = q->tag_set; |
320ae51f | 2270 | |
60de074b AM |
2271 | /* |
2272 | * Avoid others reading imcomplete hctx->cpumask through sysfs | |
2273 | */ | |
2274 | mutex_lock(&q->sysfs_lock); | |
2275 | ||
320ae51f | 2276 | queue_for_each_hw_ctx(q, hctx, i) { |
e4043dcf | 2277 | cpumask_clear(hctx->cpumask); |
320ae51f JA |
2278 | hctx->nr_ctx = 0; |
2279 | } | |
2280 | ||
2281 | /* | |
4b855ad3 CH |
2282 | * Map software to hardware queues. |
2283 | * | |
2284 | * If the cpu isn't present, the cpu is mapped to first hctx. | |
320ae51f | 2285 | */ |
4b855ad3 | 2286 | for_each_present_cpu(i) { |
d1b1cea1 GKB |
2287 | hctx_idx = q->mq_map[i]; |
2288 | /* unmapped hw queue can be remapped after CPU topo changed */ | |
cc71a6f4 JA |
2289 | if (!set->tags[hctx_idx] && |
2290 | !__blk_mq_alloc_rq_map(set, hctx_idx)) { | |
d1b1cea1 GKB |
2291 | /* |
2292 | * If tags initialization fail for some hctx, | |
2293 | * that hctx won't be brought online. In this | |
2294 | * case, remap the current ctx to hctx[0] which | |
2295 | * is guaranteed to always have tags allocated | |
2296 | */ | |
cc71a6f4 | 2297 | q->mq_map[i] = 0; |
d1b1cea1 GKB |
2298 | } |
2299 | ||
897bb0c7 | 2300 | ctx = per_cpu_ptr(q->queue_ctx, i); |
7d7e0f90 | 2301 | hctx = blk_mq_map_queue(q, i); |
868f2f0b | 2302 | |
e4043dcf | 2303 | cpumask_set_cpu(i, hctx->cpumask); |
320ae51f JA |
2304 | ctx->index_hw = hctx->nr_ctx; |
2305 | hctx->ctxs[hctx->nr_ctx++] = ctx; | |
2306 | } | |
506e931f | 2307 | |
60de074b AM |
2308 | mutex_unlock(&q->sysfs_lock); |
2309 | ||
506e931f | 2310 | queue_for_each_hw_ctx(q, hctx, i) { |
484b4061 | 2311 | /* |
a68aafa5 JA |
2312 | * If no software queues are mapped to this hardware queue, |
2313 | * disable it and free the request entries. | |
484b4061 JA |
2314 | */ |
2315 | if (!hctx->nr_ctx) { | |
d1b1cea1 GKB |
2316 | /* Never unmap queue 0. We need it as a |
2317 | * fallback in case of a new remap fails | |
2318 | * allocation | |
2319 | */ | |
cc71a6f4 JA |
2320 | if (i && set->tags[i]) |
2321 | blk_mq_free_map_and_requests(set, i); | |
2322 | ||
2a34c087 | 2323 | hctx->tags = NULL; |
484b4061 JA |
2324 | continue; |
2325 | } | |
2326 | ||
2a34c087 ML |
2327 | hctx->tags = set->tags[i]; |
2328 | WARN_ON(!hctx->tags); | |
2329 | ||
889fa31f CY |
2330 | /* |
2331 | * Set the map size to the number of mapped software queues. | |
2332 | * This is more accurate and more efficient than looping | |
2333 | * over all possibly mapped software queues. | |
2334 | */ | |
88459642 | 2335 | sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx); |
889fa31f | 2336 | |
484b4061 JA |
2337 | /* |
2338 | * Initialize batch roundrobin counts | |
2339 | */ | |
506e931f JA |
2340 | hctx->next_cpu = cpumask_first(hctx->cpumask); |
2341 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; | |
2342 | } | |
320ae51f JA |
2343 | } |
2344 | ||
8e8320c9 JA |
2345 | /* |
2346 | * Caller needs to ensure that we're either frozen/quiesced, or that | |
2347 | * the queue isn't live yet. | |
2348 | */ | |
2404e607 | 2349 | static void queue_set_hctx_shared(struct request_queue *q, bool shared) |
0d2602ca JA |
2350 | { |
2351 | struct blk_mq_hw_ctx *hctx; | |
0d2602ca JA |
2352 | int i; |
2353 | ||
2404e607 | 2354 | queue_for_each_hw_ctx(q, hctx, i) { |
8e8320c9 JA |
2355 | if (shared) { |
2356 | if (test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state)) | |
2357 | atomic_inc(&q->shared_hctx_restart); | |
2404e607 | 2358 | hctx->flags |= BLK_MQ_F_TAG_SHARED; |
8e8320c9 JA |
2359 | } else { |
2360 | if (test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state)) | |
2361 | atomic_dec(&q->shared_hctx_restart); | |
2404e607 | 2362 | hctx->flags &= ~BLK_MQ_F_TAG_SHARED; |
8e8320c9 | 2363 | } |
2404e607 JM |
2364 | } |
2365 | } | |
2366 | ||
8e8320c9 JA |
2367 | static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set, |
2368 | bool shared) | |
2404e607 JM |
2369 | { |
2370 | struct request_queue *q; | |
0d2602ca | 2371 | |
705cda97 BVA |
2372 | lockdep_assert_held(&set->tag_list_lock); |
2373 | ||
0d2602ca JA |
2374 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
2375 | blk_mq_freeze_queue(q); | |
2404e607 | 2376 | queue_set_hctx_shared(q, shared); |
0d2602ca JA |
2377 | blk_mq_unfreeze_queue(q); |
2378 | } | |
2379 | } | |
2380 | ||
2381 | static void blk_mq_del_queue_tag_set(struct request_queue *q) | |
2382 | { | |
2383 | struct blk_mq_tag_set *set = q->tag_set; | |
2384 | ||
0d2602ca | 2385 | mutex_lock(&set->tag_list_lock); |
705cda97 BVA |
2386 | list_del_rcu(&q->tag_set_list); |
2387 | INIT_LIST_HEAD(&q->tag_set_list); | |
2404e607 JM |
2388 | if (list_is_singular(&set->tag_list)) { |
2389 | /* just transitioned to unshared */ | |
2390 | set->flags &= ~BLK_MQ_F_TAG_SHARED; | |
2391 | /* update existing queue */ | |
2392 | blk_mq_update_tag_set_depth(set, false); | |
2393 | } | |
0d2602ca | 2394 | mutex_unlock(&set->tag_list_lock); |
705cda97 BVA |
2395 | |
2396 | synchronize_rcu(); | |
0d2602ca JA |
2397 | } |
2398 | ||
2399 | static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set, | |
2400 | struct request_queue *q) | |
2401 | { | |
2402 | q->tag_set = set; | |
2403 | ||
2404 | mutex_lock(&set->tag_list_lock); | |
2404e607 | 2405 | |
ff821d27 JA |
2406 | /* |
2407 | * Check to see if we're transitioning to shared (from 1 to 2 queues). | |
2408 | */ | |
2409 | if (!list_empty(&set->tag_list) && | |
2410 | !(set->flags & BLK_MQ_F_TAG_SHARED)) { | |
2404e607 JM |
2411 | set->flags |= BLK_MQ_F_TAG_SHARED; |
2412 | /* update existing queue */ | |
2413 | blk_mq_update_tag_set_depth(set, true); | |
2414 | } | |
2415 | if (set->flags & BLK_MQ_F_TAG_SHARED) | |
2416 | queue_set_hctx_shared(q, true); | |
705cda97 | 2417 | list_add_tail_rcu(&q->tag_set_list, &set->tag_list); |
2404e607 | 2418 | |
0d2602ca JA |
2419 | mutex_unlock(&set->tag_list_lock); |
2420 | } | |
2421 | ||
e09aae7e ML |
2422 | /* |
2423 | * It is the actual release handler for mq, but we do it from | |
2424 | * request queue's release handler for avoiding use-after-free | |
2425 | * and headache because q->mq_kobj shouldn't have been introduced, | |
2426 | * but we can't group ctx/kctx kobj without it. | |
2427 | */ | |
2428 | void blk_mq_release(struct request_queue *q) | |
2429 | { | |
2430 | struct blk_mq_hw_ctx *hctx; | |
2431 | unsigned int i; | |
2432 | ||
2433 | /* hctx kobj stays in hctx */ | |
c3b4afca ML |
2434 | queue_for_each_hw_ctx(q, hctx, i) { |
2435 | if (!hctx) | |
2436 | continue; | |
6c8b232e | 2437 | kobject_put(&hctx->kobj); |
c3b4afca | 2438 | } |
e09aae7e | 2439 | |
a723bab3 AM |
2440 | q->mq_map = NULL; |
2441 | ||
e09aae7e ML |
2442 | kfree(q->queue_hw_ctx); |
2443 | ||
7ea5fe31 ML |
2444 | /* |
2445 | * release .mq_kobj and sw queue's kobject now because | |
2446 | * both share lifetime with request queue. | |
2447 | */ | |
2448 | blk_mq_sysfs_deinit(q); | |
2449 | ||
e09aae7e ML |
2450 | free_percpu(q->queue_ctx); |
2451 | } | |
2452 | ||
24d2f903 | 2453 | struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set) |
b62c21b7 MS |
2454 | { |
2455 | struct request_queue *uninit_q, *q; | |
2456 | ||
2457 | uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node); | |
2458 | if (!uninit_q) | |
2459 | return ERR_PTR(-ENOMEM); | |
2460 | ||
2461 | q = blk_mq_init_allocated_queue(set, uninit_q); | |
2462 | if (IS_ERR(q)) | |
2463 | blk_cleanup_queue(uninit_q); | |
2464 | ||
2465 | return q; | |
2466 | } | |
2467 | EXPORT_SYMBOL(blk_mq_init_queue); | |
2468 | ||
07319678 BVA |
2469 | static int blk_mq_hw_ctx_size(struct blk_mq_tag_set *tag_set) |
2470 | { | |
2471 | int hw_ctx_size = sizeof(struct blk_mq_hw_ctx); | |
2472 | ||
05707b64 | 2473 | BUILD_BUG_ON(ALIGN(offsetof(struct blk_mq_hw_ctx, srcu), |
07319678 BVA |
2474 | __alignof__(struct blk_mq_hw_ctx)) != |
2475 | sizeof(struct blk_mq_hw_ctx)); | |
2476 | ||
2477 | if (tag_set->flags & BLK_MQ_F_BLOCKING) | |
2478 | hw_ctx_size += sizeof(struct srcu_struct); | |
2479 | ||
2480 | return hw_ctx_size; | |
2481 | } | |
2482 | ||
868f2f0b KB |
2483 | static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set, |
2484 | struct request_queue *q) | |
320ae51f | 2485 | { |
868f2f0b KB |
2486 | int i, j; |
2487 | struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx; | |
f14bbe77 | 2488 | |
868f2f0b | 2489 | blk_mq_sysfs_unregister(q); |
fb350e0a ML |
2490 | |
2491 | /* protect against switching io scheduler */ | |
2492 | mutex_lock(&q->sysfs_lock); | |
24d2f903 | 2493 | for (i = 0; i < set->nr_hw_queues; i++) { |
868f2f0b | 2494 | int node; |
f14bbe77 | 2495 | |
868f2f0b KB |
2496 | if (hctxs[i]) |
2497 | continue; | |
2498 | ||
2499 | node = blk_mq_hw_queue_to_node(q->mq_map, i); | |
07319678 | 2500 | hctxs[i] = kzalloc_node(blk_mq_hw_ctx_size(set), |
cdef54dd | 2501 | GFP_KERNEL, node); |
320ae51f | 2502 | if (!hctxs[i]) |
868f2f0b | 2503 | break; |
320ae51f | 2504 | |
a86073e4 | 2505 | if (!zalloc_cpumask_var_node(&hctxs[i]->cpumask, GFP_KERNEL, |
868f2f0b KB |
2506 | node)) { |
2507 | kfree(hctxs[i]); | |
2508 | hctxs[i] = NULL; | |
2509 | break; | |
2510 | } | |
e4043dcf | 2511 | |
0d2602ca | 2512 | atomic_set(&hctxs[i]->nr_active, 0); |
f14bbe77 | 2513 | hctxs[i]->numa_node = node; |
320ae51f | 2514 | hctxs[i]->queue_num = i; |
868f2f0b KB |
2515 | |
2516 | if (blk_mq_init_hctx(q, set, hctxs[i], i)) { | |
2517 | free_cpumask_var(hctxs[i]->cpumask); | |
2518 | kfree(hctxs[i]); | |
2519 | hctxs[i] = NULL; | |
2520 | break; | |
2521 | } | |
2522 | blk_mq_hctx_kobj_init(hctxs[i]); | |
320ae51f | 2523 | } |
868f2f0b KB |
2524 | for (j = i; j < q->nr_hw_queues; j++) { |
2525 | struct blk_mq_hw_ctx *hctx = hctxs[j]; | |
2526 | ||
2527 | if (hctx) { | |
cc71a6f4 JA |
2528 | if (hctx->tags) |
2529 | blk_mq_free_map_and_requests(set, j); | |
868f2f0b | 2530 | blk_mq_exit_hctx(q, set, hctx, j); |
868f2f0b | 2531 | kobject_put(&hctx->kobj); |
868f2f0b KB |
2532 | hctxs[j] = NULL; |
2533 | ||
2534 | } | |
2535 | } | |
2536 | q->nr_hw_queues = i; | |
fb350e0a | 2537 | mutex_unlock(&q->sysfs_lock); |
868f2f0b KB |
2538 | blk_mq_sysfs_register(q); |
2539 | } | |
2540 | ||
2541 | struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set, | |
2542 | struct request_queue *q) | |
2543 | { | |
66841672 ML |
2544 | /* mark the queue as mq asap */ |
2545 | q->mq_ops = set->ops; | |
2546 | ||
34dbad5d | 2547 | q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn, |
720b8ccc SB |
2548 | blk_mq_poll_stats_bkt, |
2549 | BLK_MQ_POLL_STATS_BKTS, q); | |
34dbad5d OS |
2550 | if (!q->poll_cb) |
2551 | goto err_exit; | |
2552 | ||
868f2f0b KB |
2553 | q->queue_ctx = alloc_percpu(struct blk_mq_ctx); |
2554 | if (!q->queue_ctx) | |
c7de5726 | 2555 | goto err_exit; |
868f2f0b | 2556 | |
737f98cf ML |
2557 | /* init q->mq_kobj and sw queues' kobjects */ |
2558 | blk_mq_sysfs_init(q); | |
2559 | ||
868f2f0b KB |
2560 | q->queue_hw_ctx = kzalloc_node(nr_cpu_ids * sizeof(*(q->queue_hw_ctx)), |
2561 | GFP_KERNEL, set->numa_node); | |
2562 | if (!q->queue_hw_ctx) | |
2563 | goto err_percpu; | |
2564 | ||
bdd17e75 | 2565 | q->mq_map = set->mq_map; |
868f2f0b KB |
2566 | |
2567 | blk_mq_realloc_hw_ctxs(set, q); | |
2568 | if (!q->nr_hw_queues) | |
2569 | goto err_hctxs; | |
320ae51f | 2570 | |
287922eb | 2571 | INIT_WORK(&q->timeout_work, blk_mq_timeout_work); |
e56f698b | 2572 | blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ); |
320ae51f JA |
2573 | |
2574 | q->nr_queues = nr_cpu_ids; | |
320ae51f | 2575 | |
94eddfbe | 2576 | q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT; |
320ae51f | 2577 | |
05f1dd53 JA |
2578 | if (!(set->flags & BLK_MQ_F_SG_MERGE)) |
2579 | q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE; | |
2580 | ||
1be036e9 CH |
2581 | q->sg_reserved_size = INT_MAX; |
2582 | ||
2849450a | 2583 | INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work); |
6fca6a61 CH |
2584 | INIT_LIST_HEAD(&q->requeue_list); |
2585 | spin_lock_init(&q->requeue_lock); | |
2586 | ||
254d259d | 2587 | blk_queue_make_request(q, blk_mq_make_request); |
ea435e1b CH |
2588 | if (q->mq_ops->poll) |
2589 | q->poll_fn = blk_mq_poll; | |
07068d5b | 2590 | |
eba71768 JA |
2591 | /* |
2592 | * Do this after blk_queue_make_request() overrides it... | |
2593 | */ | |
2594 | q->nr_requests = set->queue_depth; | |
2595 | ||
64f1c21e JA |
2596 | /* |
2597 | * Default to classic polling | |
2598 | */ | |
2599 | q->poll_nsec = -1; | |
2600 | ||
24d2f903 CH |
2601 | if (set->ops->complete) |
2602 | blk_queue_softirq_done(q, set->ops->complete); | |
30a91cb4 | 2603 | |
24d2f903 | 2604 | blk_mq_init_cpu_queues(q, set->nr_hw_queues); |
0d2602ca | 2605 | blk_mq_add_queue_tag_set(set, q); |
4b855ad3 | 2606 | blk_mq_map_swqueue(q); |
4593fdbe | 2607 | |
d3484991 JA |
2608 | if (!(set->flags & BLK_MQ_F_NO_SCHED)) { |
2609 | int ret; | |
2610 | ||
2611 | ret = blk_mq_sched_init(q); | |
2612 | if (ret) | |
2613 | return ERR_PTR(ret); | |
2614 | } | |
2615 | ||
320ae51f | 2616 | return q; |
18741986 | 2617 | |
320ae51f | 2618 | err_hctxs: |
868f2f0b | 2619 | kfree(q->queue_hw_ctx); |
320ae51f | 2620 | err_percpu: |
868f2f0b | 2621 | free_percpu(q->queue_ctx); |
c7de5726 ML |
2622 | err_exit: |
2623 | q->mq_ops = NULL; | |
320ae51f JA |
2624 | return ERR_PTR(-ENOMEM); |
2625 | } | |
b62c21b7 | 2626 | EXPORT_SYMBOL(blk_mq_init_allocated_queue); |
320ae51f JA |
2627 | |
2628 | void blk_mq_free_queue(struct request_queue *q) | |
2629 | { | |
624dbe47 | 2630 | struct blk_mq_tag_set *set = q->tag_set; |
320ae51f | 2631 | |
0d2602ca | 2632 | blk_mq_del_queue_tag_set(q); |
624dbe47 | 2633 | blk_mq_exit_hw_queues(q, set, set->nr_hw_queues); |
320ae51f | 2634 | } |
320ae51f JA |
2635 | |
2636 | /* Basically redo blk_mq_init_queue with queue frozen */ | |
4b855ad3 | 2637 | static void blk_mq_queue_reinit(struct request_queue *q) |
320ae51f | 2638 | { |
4ecd4fef | 2639 | WARN_ON_ONCE(!atomic_read(&q->mq_freeze_depth)); |
320ae51f | 2640 | |
9c1051aa | 2641 | blk_mq_debugfs_unregister_hctxs(q); |
67aec14c JA |
2642 | blk_mq_sysfs_unregister(q); |
2643 | ||
320ae51f JA |
2644 | /* |
2645 | * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe | |
ff821d27 JA |
2646 | * we should change hctx numa_node according to the new topology (this |
2647 | * involves freeing and re-allocating memory, worth doing?) | |
320ae51f | 2648 | */ |
4b855ad3 | 2649 | blk_mq_map_swqueue(q); |
320ae51f | 2650 | |
67aec14c | 2651 | blk_mq_sysfs_register(q); |
9c1051aa | 2652 | blk_mq_debugfs_register_hctxs(q); |
320ae51f JA |
2653 | } |
2654 | ||
a5164405 JA |
2655 | static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set) |
2656 | { | |
2657 | int i; | |
2658 | ||
cc71a6f4 JA |
2659 | for (i = 0; i < set->nr_hw_queues; i++) |
2660 | if (!__blk_mq_alloc_rq_map(set, i)) | |
a5164405 | 2661 | goto out_unwind; |
a5164405 JA |
2662 | |
2663 | return 0; | |
2664 | ||
2665 | out_unwind: | |
2666 | while (--i >= 0) | |
cc71a6f4 | 2667 | blk_mq_free_rq_map(set->tags[i]); |
a5164405 | 2668 | |
a5164405 JA |
2669 | return -ENOMEM; |
2670 | } | |
2671 | ||
2672 | /* | |
2673 | * Allocate the request maps associated with this tag_set. Note that this | |
2674 | * may reduce the depth asked for, if memory is tight. set->queue_depth | |
2675 | * will be updated to reflect the allocated depth. | |
2676 | */ | |
2677 | static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set) | |
2678 | { | |
2679 | unsigned int depth; | |
2680 | int err; | |
2681 | ||
2682 | depth = set->queue_depth; | |
2683 | do { | |
2684 | err = __blk_mq_alloc_rq_maps(set); | |
2685 | if (!err) | |
2686 | break; | |
2687 | ||
2688 | set->queue_depth >>= 1; | |
2689 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) { | |
2690 | err = -ENOMEM; | |
2691 | break; | |
2692 | } | |
2693 | } while (set->queue_depth); | |
2694 | ||
2695 | if (!set->queue_depth || err) { | |
2696 | pr_err("blk-mq: failed to allocate request map\n"); | |
2697 | return -ENOMEM; | |
2698 | } | |
2699 | ||
2700 | if (depth != set->queue_depth) | |
2701 | pr_info("blk-mq: reduced tag depth (%u -> %u)\n", | |
2702 | depth, set->queue_depth); | |
2703 | ||
2704 | return 0; | |
2705 | } | |
2706 | ||
ebe8bddb OS |
2707 | static int blk_mq_update_queue_map(struct blk_mq_tag_set *set) |
2708 | { | |
7d4901a9 ML |
2709 | if (set->ops->map_queues) { |
2710 | int cpu; | |
2711 | /* | |
2712 | * transport .map_queues is usually done in the following | |
2713 | * way: | |
2714 | * | |
2715 | * for (queue = 0; queue < set->nr_hw_queues; queue++) { | |
2716 | * mask = get_cpu_mask(queue) | |
2717 | * for_each_cpu(cpu, mask) | |
2718 | * set->mq_map[cpu] = queue; | |
2719 | * } | |
2720 | * | |
2721 | * When we need to remap, the table has to be cleared for | |
2722 | * killing stale mapping since one CPU may not be mapped | |
2723 | * to any hw queue. | |
2724 | */ | |
2725 | for_each_possible_cpu(cpu) | |
2726 | set->mq_map[cpu] = 0; | |
2727 | ||
ebe8bddb | 2728 | return set->ops->map_queues(set); |
7d4901a9 | 2729 | } else |
ebe8bddb OS |
2730 | return blk_mq_map_queues(set); |
2731 | } | |
2732 | ||
a4391c64 JA |
2733 | /* |
2734 | * Alloc a tag set to be associated with one or more request queues. | |
2735 | * May fail with EINVAL for various error conditions. May adjust the | |
2736 | * requested depth down, if if it too large. In that case, the set | |
2737 | * value will be stored in set->queue_depth. | |
2738 | */ | |
24d2f903 CH |
2739 | int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set) |
2740 | { | |
da695ba2 CH |
2741 | int ret; |
2742 | ||
205fb5f5 BVA |
2743 | BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS); |
2744 | ||
24d2f903 CH |
2745 | if (!set->nr_hw_queues) |
2746 | return -EINVAL; | |
a4391c64 | 2747 | if (!set->queue_depth) |
24d2f903 CH |
2748 | return -EINVAL; |
2749 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) | |
2750 | return -EINVAL; | |
2751 | ||
7d7e0f90 | 2752 | if (!set->ops->queue_rq) |
24d2f903 CH |
2753 | return -EINVAL; |
2754 | ||
de148297 ML |
2755 | if (!set->ops->get_budget ^ !set->ops->put_budget) |
2756 | return -EINVAL; | |
2757 | ||
a4391c64 JA |
2758 | if (set->queue_depth > BLK_MQ_MAX_DEPTH) { |
2759 | pr_info("blk-mq: reduced tag depth to %u\n", | |
2760 | BLK_MQ_MAX_DEPTH); | |
2761 | set->queue_depth = BLK_MQ_MAX_DEPTH; | |
2762 | } | |
24d2f903 | 2763 | |
6637fadf SL |
2764 | /* |
2765 | * If a crashdump is active, then we are potentially in a very | |
2766 | * memory constrained environment. Limit us to 1 queue and | |
2767 | * 64 tags to prevent using too much memory. | |
2768 | */ | |
2769 | if (is_kdump_kernel()) { | |
2770 | set->nr_hw_queues = 1; | |
2771 | set->queue_depth = min(64U, set->queue_depth); | |
2772 | } | |
868f2f0b KB |
2773 | /* |
2774 | * There is no use for more h/w queues than cpus. | |
2775 | */ | |
2776 | if (set->nr_hw_queues > nr_cpu_ids) | |
2777 | set->nr_hw_queues = nr_cpu_ids; | |
6637fadf | 2778 | |
868f2f0b | 2779 | set->tags = kzalloc_node(nr_cpu_ids * sizeof(struct blk_mq_tags *), |
24d2f903 CH |
2780 | GFP_KERNEL, set->numa_node); |
2781 | if (!set->tags) | |
a5164405 | 2782 | return -ENOMEM; |
24d2f903 | 2783 | |
da695ba2 CH |
2784 | ret = -ENOMEM; |
2785 | set->mq_map = kzalloc_node(sizeof(*set->mq_map) * nr_cpu_ids, | |
2786 | GFP_KERNEL, set->numa_node); | |
bdd17e75 CH |
2787 | if (!set->mq_map) |
2788 | goto out_free_tags; | |
2789 | ||
ebe8bddb | 2790 | ret = blk_mq_update_queue_map(set); |
da695ba2 CH |
2791 | if (ret) |
2792 | goto out_free_mq_map; | |
2793 | ||
2794 | ret = blk_mq_alloc_rq_maps(set); | |
2795 | if (ret) | |
bdd17e75 | 2796 | goto out_free_mq_map; |
24d2f903 | 2797 | |
0d2602ca JA |
2798 | mutex_init(&set->tag_list_lock); |
2799 | INIT_LIST_HEAD(&set->tag_list); | |
2800 | ||
24d2f903 | 2801 | return 0; |
bdd17e75 CH |
2802 | |
2803 | out_free_mq_map: | |
2804 | kfree(set->mq_map); | |
2805 | set->mq_map = NULL; | |
2806 | out_free_tags: | |
5676e7b6 RE |
2807 | kfree(set->tags); |
2808 | set->tags = NULL; | |
da695ba2 | 2809 | return ret; |
24d2f903 CH |
2810 | } |
2811 | EXPORT_SYMBOL(blk_mq_alloc_tag_set); | |
2812 | ||
2813 | void blk_mq_free_tag_set(struct blk_mq_tag_set *set) | |
2814 | { | |
2815 | int i; | |
2816 | ||
cc71a6f4 JA |
2817 | for (i = 0; i < nr_cpu_ids; i++) |
2818 | blk_mq_free_map_and_requests(set, i); | |
484b4061 | 2819 | |
bdd17e75 CH |
2820 | kfree(set->mq_map); |
2821 | set->mq_map = NULL; | |
2822 | ||
981bd189 | 2823 | kfree(set->tags); |
5676e7b6 | 2824 | set->tags = NULL; |
24d2f903 CH |
2825 | } |
2826 | EXPORT_SYMBOL(blk_mq_free_tag_set); | |
2827 | ||
e3a2b3f9 JA |
2828 | int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr) |
2829 | { | |
2830 | struct blk_mq_tag_set *set = q->tag_set; | |
2831 | struct blk_mq_hw_ctx *hctx; | |
2832 | int i, ret; | |
2833 | ||
bd166ef1 | 2834 | if (!set) |
e3a2b3f9 JA |
2835 | return -EINVAL; |
2836 | ||
70f36b60 | 2837 | blk_mq_freeze_queue(q); |
24f5a90f | 2838 | blk_mq_quiesce_queue(q); |
70f36b60 | 2839 | |
e3a2b3f9 JA |
2840 | ret = 0; |
2841 | queue_for_each_hw_ctx(q, hctx, i) { | |
e9137d4b KB |
2842 | if (!hctx->tags) |
2843 | continue; | |
bd166ef1 JA |
2844 | /* |
2845 | * If we're using an MQ scheduler, just update the scheduler | |
2846 | * queue depth. This is similar to what the old code would do. | |
2847 | */ | |
70f36b60 | 2848 | if (!hctx->sched_tags) { |
c2e82a23 | 2849 | ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr, |
70f36b60 JA |
2850 | false); |
2851 | } else { | |
2852 | ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags, | |
2853 | nr, true); | |
2854 | } | |
e3a2b3f9 JA |
2855 | if (ret) |
2856 | break; | |
2857 | } | |
2858 | ||
2859 | if (!ret) | |
2860 | q->nr_requests = nr; | |
2861 | ||
24f5a90f | 2862 | blk_mq_unquiesce_queue(q); |
70f36b60 | 2863 | blk_mq_unfreeze_queue(q); |
70f36b60 | 2864 | |
e3a2b3f9 JA |
2865 | return ret; |
2866 | } | |
2867 | ||
e4dc2b32 KB |
2868 | static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, |
2869 | int nr_hw_queues) | |
868f2f0b KB |
2870 | { |
2871 | struct request_queue *q; | |
2872 | ||
705cda97 BVA |
2873 | lockdep_assert_held(&set->tag_list_lock); |
2874 | ||
868f2f0b KB |
2875 | if (nr_hw_queues > nr_cpu_ids) |
2876 | nr_hw_queues = nr_cpu_ids; | |
2877 | if (nr_hw_queues < 1 || nr_hw_queues == set->nr_hw_queues) | |
2878 | return; | |
2879 | ||
2880 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
2881 | blk_mq_freeze_queue(q); | |
2882 | ||
2883 | set->nr_hw_queues = nr_hw_queues; | |
ebe8bddb | 2884 | blk_mq_update_queue_map(set); |
868f2f0b KB |
2885 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
2886 | blk_mq_realloc_hw_ctxs(set, q); | |
4b855ad3 | 2887 | blk_mq_queue_reinit(q); |
868f2f0b KB |
2888 | } |
2889 | ||
2890 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
2891 | blk_mq_unfreeze_queue(q); | |
2892 | } | |
e4dc2b32 KB |
2893 | |
2894 | void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues) | |
2895 | { | |
2896 | mutex_lock(&set->tag_list_lock); | |
2897 | __blk_mq_update_nr_hw_queues(set, nr_hw_queues); | |
2898 | mutex_unlock(&set->tag_list_lock); | |
2899 | } | |
868f2f0b KB |
2900 | EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues); |
2901 | ||
34dbad5d OS |
2902 | /* Enable polling stats and return whether they were already enabled. */ |
2903 | static bool blk_poll_stats_enable(struct request_queue *q) | |
2904 | { | |
2905 | if (test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) || | |
2906 | test_and_set_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags)) | |
2907 | return true; | |
2908 | blk_stat_add_callback(q, q->poll_cb); | |
2909 | return false; | |
2910 | } | |
2911 | ||
2912 | static void blk_mq_poll_stats_start(struct request_queue *q) | |
2913 | { | |
2914 | /* | |
2915 | * We don't arm the callback if polling stats are not enabled or the | |
2916 | * callback is already active. | |
2917 | */ | |
2918 | if (!test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) || | |
2919 | blk_stat_is_active(q->poll_cb)) | |
2920 | return; | |
2921 | ||
2922 | blk_stat_activate_msecs(q->poll_cb, 100); | |
2923 | } | |
2924 | ||
2925 | static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb) | |
2926 | { | |
2927 | struct request_queue *q = cb->data; | |
720b8ccc | 2928 | int bucket; |
34dbad5d | 2929 | |
720b8ccc SB |
2930 | for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) { |
2931 | if (cb->stat[bucket].nr_samples) | |
2932 | q->poll_stat[bucket] = cb->stat[bucket]; | |
2933 | } | |
34dbad5d OS |
2934 | } |
2935 | ||
64f1c21e JA |
2936 | static unsigned long blk_mq_poll_nsecs(struct request_queue *q, |
2937 | struct blk_mq_hw_ctx *hctx, | |
2938 | struct request *rq) | |
2939 | { | |
64f1c21e | 2940 | unsigned long ret = 0; |
720b8ccc | 2941 | int bucket; |
64f1c21e JA |
2942 | |
2943 | /* | |
2944 | * If stats collection isn't on, don't sleep but turn it on for | |
2945 | * future users | |
2946 | */ | |
34dbad5d | 2947 | if (!blk_poll_stats_enable(q)) |
64f1c21e JA |
2948 | return 0; |
2949 | ||
64f1c21e JA |
2950 | /* |
2951 | * As an optimistic guess, use half of the mean service time | |
2952 | * for this type of request. We can (and should) make this smarter. | |
2953 | * For instance, if the completion latencies are tight, we can | |
2954 | * get closer than just half the mean. This is especially | |
2955 | * important on devices where the completion latencies are longer | |
720b8ccc SB |
2956 | * than ~10 usec. We do use the stats for the relevant IO size |
2957 | * if available which does lead to better estimates. | |
64f1c21e | 2958 | */ |
720b8ccc SB |
2959 | bucket = blk_mq_poll_stats_bkt(rq); |
2960 | if (bucket < 0) | |
2961 | return ret; | |
2962 | ||
2963 | if (q->poll_stat[bucket].nr_samples) | |
2964 | ret = (q->poll_stat[bucket].mean + 1) / 2; | |
64f1c21e JA |
2965 | |
2966 | return ret; | |
2967 | } | |
2968 | ||
06426adf | 2969 | static bool blk_mq_poll_hybrid_sleep(struct request_queue *q, |
64f1c21e | 2970 | struct blk_mq_hw_ctx *hctx, |
06426adf JA |
2971 | struct request *rq) |
2972 | { | |
2973 | struct hrtimer_sleeper hs; | |
2974 | enum hrtimer_mode mode; | |
64f1c21e | 2975 | unsigned int nsecs; |
06426adf JA |
2976 | ktime_t kt; |
2977 | ||
76a86f9d | 2978 | if (rq->rq_flags & RQF_MQ_POLL_SLEPT) |
64f1c21e JA |
2979 | return false; |
2980 | ||
2981 | /* | |
2982 | * poll_nsec can be: | |
2983 | * | |
2984 | * -1: don't ever hybrid sleep | |
2985 | * 0: use half of prev avg | |
2986 | * >0: use this specific value | |
2987 | */ | |
2988 | if (q->poll_nsec == -1) | |
2989 | return false; | |
2990 | else if (q->poll_nsec > 0) | |
2991 | nsecs = q->poll_nsec; | |
2992 | else | |
2993 | nsecs = blk_mq_poll_nsecs(q, hctx, rq); | |
2994 | ||
2995 | if (!nsecs) | |
06426adf JA |
2996 | return false; |
2997 | ||
76a86f9d | 2998 | rq->rq_flags |= RQF_MQ_POLL_SLEPT; |
06426adf JA |
2999 | |
3000 | /* | |
3001 | * This will be replaced with the stats tracking code, using | |
3002 | * 'avg_completion_time / 2' as the pre-sleep target. | |
3003 | */ | |
8b0e1953 | 3004 | kt = nsecs; |
06426adf JA |
3005 | |
3006 | mode = HRTIMER_MODE_REL; | |
3007 | hrtimer_init_on_stack(&hs.timer, CLOCK_MONOTONIC, mode); | |
3008 | hrtimer_set_expires(&hs.timer, kt); | |
3009 | ||
3010 | hrtimer_init_sleeper(&hs, current); | |
3011 | do { | |
5a61c363 | 3012 | if (blk_mq_rq_state(rq) == MQ_RQ_COMPLETE) |
06426adf JA |
3013 | break; |
3014 | set_current_state(TASK_UNINTERRUPTIBLE); | |
3015 | hrtimer_start_expires(&hs.timer, mode); | |
3016 | if (hs.task) | |
3017 | io_schedule(); | |
3018 | hrtimer_cancel(&hs.timer); | |
3019 | mode = HRTIMER_MODE_ABS; | |
3020 | } while (hs.task && !signal_pending(current)); | |
3021 | ||
3022 | __set_current_state(TASK_RUNNING); | |
3023 | destroy_hrtimer_on_stack(&hs.timer); | |
3024 | return true; | |
3025 | } | |
3026 | ||
bbd7bb70 JA |
3027 | static bool __blk_mq_poll(struct blk_mq_hw_ctx *hctx, struct request *rq) |
3028 | { | |
3029 | struct request_queue *q = hctx->queue; | |
3030 | long state; | |
3031 | ||
06426adf JA |
3032 | /* |
3033 | * If we sleep, have the caller restart the poll loop to reset | |
3034 | * the state. Like for the other success return cases, the | |
3035 | * caller is responsible for checking if the IO completed. If | |
3036 | * the IO isn't complete, we'll get called again and will go | |
3037 | * straight to the busy poll loop. | |
3038 | */ | |
64f1c21e | 3039 | if (blk_mq_poll_hybrid_sleep(q, hctx, rq)) |
06426adf JA |
3040 | return true; |
3041 | ||
bbd7bb70 JA |
3042 | hctx->poll_considered++; |
3043 | ||
3044 | state = current->state; | |
3045 | while (!need_resched()) { | |
3046 | int ret; | |
3047 | ||
3048 | hctx->poll_invoked++; | |
3049 | ||
3050 | ret = q->mq_ops->poll(hctx, rq->tag); | |
3051 | if (ret > 0) { | |
3052 | hctx->poll_success++; | |
3053 | set_current_state(TASK_RUNNING); | |
3054 | return true; | |
3055 | } | |
3056 | ||
3057 | if (signal_pending_state(state, current)) | |
3058 | set_current_state(TASK_RUNNING); | |
3059 | ||
3060 | if (current->state == TASK_RUNNING) | |
3061 | return true; | |
3062 | if (ret < 0) | |
3063 | break; | |
3064 | cpu_relax(); | |
3065 | } | |
3066 | ||
3067 | return false; | |
3068 | } | |
3069 | ||
ea435e1b | 3070 | static bool blk_mq_poll(struct request_queue *q, blk_qc_t cookie) |
bbd7bb70 JA |
3071 | { |
3072 | struct blk_mq_hw_ctx *hctx; | |
bbd7bb70 JA |
3073 | struct request *rq; |
3074 | ||
ea435e1b | 3075 | if (!test_bit(QUEUE_FLAG_POLL, &q->queue_flags)) |
bbd7bb70 JA |
3076 | return false; |
3077 | ||
bbd7bb70 | 3078 | hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)]; |
bd166ef1 JA |
3079 | if (!blk_qc_t_is_internal(cookie)) |
3080 | rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie)); | |
3a07bb1d | 3081 | else { |
bd166ef1 | 3082 | rq = blk_mq_tag_to_rq(hctx->sched_tags, blk_qc_t_to_tag(cookie)); |
3a07bb1d JA |
3083 | /* |
3084 | * With scheduling, if the request has completed, we'll | |
3085 | * get a NULL return here, as we clear the sched tag when | |
3086 | * that happens. The request still remains valid, like always, | |
3087 | * so we should be safe with just the NULL check. | |
3088 | */ | |
3089 | if (!rq) | |
3090 | return false; | |
3091 | } | |
bbd7bb70 JA |
3092 | |
3093 | return __blk_mq_poll(hctx, rq); | |
3094 | } | |
bbd7bb70 | 3095 | |
320ae51f JA |
3096 | static int __init blk_mq_init(void) |
3097 | { | |
9467f859 TG |
3098 | cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL, |
3099 | blk_mq_hctx_notify_dead); | |
320ae51f JA |
3100 | return 0; |
3101 | } | |
3102 | subsys_initcall(blk_mq_init); |