]>
Commit | Line | Data |
---|---|---|
75bb4625 JA |
1 | /* |
2 | * Block multiqueue core code | |
3 | * | |
4 | * Copyright (C) 2013-2014 Jens Axboe | |
5 | * Copyright (C) 2013-2014 Christoph Hellwig | |
6 | */ | |
320ae51f JA |
7 | #include <linux/kernel.h> |
8 | #include <linux/module.h> | |
9 | #include <linux/backing-dev.h> | |
10 | #include <linux/bio.h> | |
11 | #include <linux/blkdev.h> | |
f75782e4 | 12 | #include <linux/kmemleak.h> |
320ae51f JA |
13 | #include <linux/mm.h> |
14 | #include <linux/init.h> | |
15 | #include <linux/slab.h> | |
16 | #include <linux/workqueue.h> | |
17 | #include <linux/smp.h> | |
18 | #include <linux/llist.h> | |
19 | #include <linux/list_sort.h> | |
20 | #include <linux/cpu.h> | |
21 | #include <linux/cache.h> | |
22 | #include <linux/sched/sysctl.h> | |
105ab3d8 | 23 | #include <linux/sched/topology.h> |
174cd4b1 | 24 | #include <linux/sched/signal.h> |
320ae51f | 25 | #include <linux/delay.h> |
aedcd72f | 26 | #include <linux/crash_dump.h> |
88c7b2b7 | 27 | #include <linux/prefetch.h> |
320ae51f JA |
28 | |
29 | #include <trace/events/block.h> | |
30 | ||
31 | #include <linux/blk-mq.h> | |
32 | #include "blk.h" | |
33 | #include "blk-mq.h" | |
9c1051aa | 34 | #include "blk-mq-debugfs.h" |
320ae51f | 35 | #include "blk-mq-tag.h" |
cf43e6be | 36 | #include "blk-stat.h" |
87760e5e | 37 | #include "blk-wbt.h" |
bd166ef1 | 38 | #include "blk-mq-sched.h" |
320ae51f | 39 | |
ea435e1b | 40 | static bool blk_mq_poll(struct request_queue *q, blk_qc_t cookie); |
34dbad5d OS |
41 | static void blk_mq_poll_stats_start(struct request_queue *q); |
42 | static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb); | |
43 | ||
720b8ccc SB |
44 | static int blk_mq_poll_stats_bkt(const struct request *rq) |
45 | { | |
46 | int ddir, bytes, bucket; | |
47 | ||
99c749a4 | 48 | ddir = rq_data_dir(rq); |
720b8ccc SB |
49 | bytes = blk_rq_bytes(rq); |
50 | ||
51 | bucket = ddir + 2*(ilog2(bytes) - 9); | |
52 | ||
53 | if (bucket < 0) | |
54 | return -1; | |
55 | else if (bucket >= BLK_MQ_POLL_STATS_BKTS) | |
56 | return ddir + BLK_MQ_POLL_STATS_BKTS - 2; | |
57 | ||
58 | return bucket; | |
59 | } | |
60 | ||
320ae51f JA |
61 | /* |
62 | * Check if any of the ctx's have pending work in this hardware queue | |
63 | */ | |
79f720a7 | 64 | static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx) |
320ae51f | 65 | { |
79f720a7 JA |
66 | return !list_empty_careful(&hctx->dispatch) || |
67 | sbitmap_any_bit_set(&hctx->ctx_map) || | |
bd166ef1 | 68 | blk_mq_sched_has_work(hctx); |
1429d7c9 JA |
69 | } |
70 | ||
320ae51f JA |
71 | /* |
72 | * Mark this ctx as having pending work in this hardware queue | |
73 | */ | |
74 | static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx, | |
75 | struct blk_mq_ctx *ctx) | |
76 | { | |
88459642 OS |
77 | if (!sbitmap_test_bit(&hctx->ctx_map, ctx->index_hw)) |
78 | sbitmap_set_bit(&hctx->ctx_map, ctx->index_hw); | |
1429d7c9 JA |
79 | } |
80 | ||
81 | static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx, | |
82 | struct blk_mq_ctx *ctx) | |
83 | { | |
88459642 | 84 | sbitmap_clear_bit(&hctx->ctx_map, ctx->index_hw); |
320ae51f JA |
85 | } |
86 | ||
f299b7c7 JA |
87 | struct mq_inflight { |
88 | struct hd_struct *part; | |
89 | unsigned int *inflight; | |
90 | }; | |
91 | ||
92 | static void blk_mq_check_inflight(struct blk_mq_hw_ctx *hctx, | |
93 | struct request *rq, void *priv, | |
94 | bool reserved) | |
95 | { | |
96 | struct mq_inflight *mi = priv; | |
97 | ||
98 | if (test_bit(REQ_ATOM_STARTED, &rq->atomic_flags) && | |
99 | !test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags)) { | |
100 | /* | |
b8d62b3a JA |
101 | * index[0] counts the specific partition that was asked |
102 | * for. index[1] counts the ones that are active on the | |
103 | * whole device, so increment that if mi->part is indeed | |
104 | * a partition, and not a whole device. | |
f299b7c7 | 105 | */ |
b8d62b3a | 106 | if (rq->part == mi->part) |
f299b7c7 | 107 | mi->inflight[0]++; |
b8d62b3a JA |
108 | if (mi->part->partno) |
109 | mi->inflight[1]++; | |
f299b7c7 JA |
110 | } |
111 | } | |
112 | ||
113 | void blk_mq_in_flight(struct request_queue *q, struct hd_struct *part, | |
114 | unsigned int inflight[2]) | |
115 | { | |
116 | struct mq_inflight mi = { .part = part, .inflight = inflight, }; | |
117 | ||
b8d62b3a | 118 | inflight[0] = inflight[1] = 0; |
f299b7c7 JA |
119 | blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi); |
120 | } | |
121 | ||
1671d522 | 122 | void blk_freeze_queue_start(struct request_queue *q) |
43a5e4e2 | 123 | { |
4ecd4fef | 124 | int freeze_depth; |
cddd5d17 | 125 | |
4ecd4fef CH |
126 | freeze_depth = atomic_inc_return(&q->mq_freeze_depth); |
127 | if (freeze_depth == 1) { | |
3ef28e83 | 128 | percpu_ref_kill(&q->q_usage_counter); |
055f6e18 ML |
129 | if (q->mq_ops) |
130 | blk_mq_run_hw_queues(q, false); | |
cddd5d17 | 131 | } |
f3af020b | 132 | } |
1671d522 | 133 | EXPORT_SYMBOL_GPL(blk_freeze_queue_start); |
f3af020b | 134 | |
6bae363e | 135 | void blk_mq_freeze_queue_wait(struct request_queue *q) |
f3af020b | 136 | { |
3ef28e83 | 137 | wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter)); |
43a5e4e2 | 138 | } |
6bae363e | 139 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait); |
43a5e4e2 | 140 | |
f91328c4 KB |
141 | int blk_mq_freeze_queue_wait_timeout(struct request_queue *q, |
142 | unsigned long timeout) | |
143 | { | |
144 | return wait_event_timeout(q->mq_freeze_wq, | |
145 | percpu_ref_is_zero(&q->q_usage_counter), | |
146 | timeout); | |
147 | } | |
148 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout); | |
43a5e4e2 | 149 | |
f3af020b TH |
150 | /* |
151 | * Guarantee no request is in use, so we can change any data structure of | |
152 | * the queue afterward. | |
153 | */ | |
3ef28e83 | 154 | void blk_freeze_queue(struct request_queue *q) |
f3af020b | 155 | { |
3ef28e83 DW |
156 | /* |
157 | * In the !blk_mq case we are only calling this to kill the | |
158 | * q_usage_counter, otherwise this increases the freeze depth | |
159 | * and waits for it to return to zero. For this reason there is | |
160 | * no blk_unfreeze_queue(), and blk_freeze_queue() is not | |
161 | * exported to drivers as the only user for unfreeze is blk_mq. | |
162 | */ | |
1671d522 | 163 | blk_freeze_queue_start(q); |
454be724 ML |
164 | if (!q->mq_ops) |
165 | blk_drain_queue(q); | |
f3af020b TH |
166 | blk_mq_freeze_queue_wait(q); |
167 | } | |
3ef28e83 DW |
168 | |
169 | void blk_mq_freeze_queue(struct request_queue *q) | |
170 | { | |
171 | /* | |
172 | * ...just an alias to keep freeze and unfreeze actions balanced | |
173 | * in the blk_mq_* namespace | |
174 | */ | |
175 | blk_freeze_queue(q); | |
176 | } | |
c761d96b | 177 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue); |
f3af020b | 178 | |
b4c6a028 | 179 | void blk_mq_unfreeze_queue(struct request_queue *q) |
320ae51f | 180 | { |
4ecd4fef | 181 | int freeze_depth; |
320ae51f | 182 | |
4ecd4fef CH |
183 | freeze_depth = atomic_dec_return(&q->mq_freeze_depth); |
184 | WARN_ON_ONCE(freeze_depth < 0); | |
185 | if (!freeze_depth) { | |
3ef28e83 | 186 | percpu_ref_reinit(&q->q_usage_counter); |
320ae51f | 187 | wake_up_all(&q->mq_freeze_wq); |
add703fd | 188 | } |
320ae51f | 189 | } |
b4c6a028 | 190 | EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue); |
320ae51f | 191 | |
852ec809 BVA |
192 | /* |
193 | * FIXME: replace the scsi_internal_device_*block_nowait() calls in the | |
194 | * mpt3sas driver such that this function can be removed. | |
195 | */ | |
196 | void blk_mq_quiesce_queue_nowait(struct request_queue *q) | |
197 | { | |
198 | unsigned long flags; | |
199 | ||
200 | spin_lock_irqsave(q->queue_lock, flags); | |
201 | queue_flag_set(QUEUE_FLAG_QUIESCED, q); | |
202 | spin_unlock_irqrestore(q->queue_lock, flags); | |
203 | } | |
204 | EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait); | |
205 | ||
6a83e74d | 206 | /** |
69e07c4a | 207 | * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished |
6a83e74d BVA |
208 | * @q: request queue. |
209 | * | |
210 | * Note: this function does not prevent that the struct request end_io() | |
69e07c4a ML |
211 | * callback function is invoked. Once this function is returned, we make |
212 | * sure no dispatch can happen until the queue is unquiesced via | |
213 | * blk_mq_unquiesce_queue(). | |
6a83e74d BVA |
214 | */ |
215 | void blk_mq_quiesce_queue(struct request_queue *q) | |
216 | { | |
217 | struct blk_mq_hw_ctx *hctx; | |
218 | unsigned int i; | |
219 | bool rcu = false; | |
220 | ||
1d9e9bc6 | 221 | blk_mq_quiesce_queue_nowait(q); |
f4560ffe | 222 | |
6a83e74d BVA |
223 | queue_for_each_hw_ctx(q, hctx, i) { |
224 | if (hctx->flags & BLK_MQ_F_BLOCKING) | |
07319678 | 225 | synchronize_srcu(hctx->queue_rq_srcu); |
6a83e74d BVA |
226 | else |
227 | rcu = true; | |
228 | } | |
229 | if (rcu) | |
230 | synchronize_rcu(); | |
231 | } | |
232 | EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue); | |
233 | ||
e4e73913 ML |
234 | /* |
235 | * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue() | |
236 | * @q: request queue. | |
237 | * | |
238 | * This function recovers queue into the state before quiescing | |
239 | * which is done by blk_mq_quiesce_queue. | |
240 | */ | |
241 | void blk_mq_unquiesce_queue(struct request_queue *q) | |
242 | { | |
852ec809 BVA |
243 | unsigned long flags; |
244 | ||
245 | spin_lock_irqsave(q->queue_lock, flags); | |
f4560ffe | 246 | queue_flag_clear(QUEUE_FLAG_QUIESCED, q); |
852ec809 | 247 | spin_unlock_irqrestore(q->queue_lock, flags); |
f4560ffe | 248 | |
1d9e9bc6 ML |
249 | /* dispatch requests which are inserted during quiescing */ |
250 | blk_mq_run_hw_queues(q, true); | |
e4e73913 ML |
251 | } |
252 | EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue); | |
253 | ||
aed3ea94 JA |
254 | void blk_mq_wake_waiters(struct request_queue *q) |
255 | { | |
256 | struct blk_mq_hw_ctx *hctx; | |
257 | unsigned int i; | |
258 | ||
259 | queue_for_each_hw_ctx(q, hctx, i) | |
260 | if (blk_mq_hw_queue_mapped(hctx)) | |
261 | blk_mq_tag_wakeup_all(hctx->tags, true); | |
262 | } | |
263 | ||
320ae51f JA |
264 | bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx) |
265 | { | |
266 | return blk_mq_has_free_tags(hctx->tags); | |
267 | } | |
268 | EXPORT_SYMBOL(blk_mq_can_queue); | |
269 | ||
e4cdf1a1 CH |
270 | static struct request *blk_mq_rq_ctx_init(struct blk_mq_alloc_data *data, |
271 | unsigned int tag, unsigned int op) | |
320ae51f | 272 | { |
e4cdf1a1 CH |
273 | struct blk_mq_tags *tags = blk_mq_tags_from_data(data); |
274 | struct request *rq = tags->static_rqs[tag]; | |
275 | ||
c3a148d2 BVA |
276 | rq->rq_flags = 0; |
277 | ||
e4cdf1a1 CH |
278 | if (data->flags & BLK_MQ_REQ_INTERNAL) { |
279 | rq->tag = -1; | |
280 | rq->internal_tag = tag; | |
281 | } else { | |
282 | if (blk_mq_tag_busy(data->hctx)) { | |
283 | rq->rq_flags = RQF_MQ_INFLIGHT; | |
284 | atomic_inc(&data->hctx->nr_active); | |
285 | } | |
286 | rq->tag = tag; | |
287 | rq->internal_tag = -1; | |
288 | data->hctx->tags->rqs[rq->tag] = rq; | |
289 | } | |
290 | ||
af76e555 CH |
291 | INIT_LIST_HEAD(&rq->queuelist); |
292 | /* csd/requeue_work/fifo_time is initialized before use */ | |
e4cdf1a1 CH |
293 | rq->q = data->q; |
294 | rq->mq_ctx = data->ctx; | |
ef295ecf | 295 | rq->cmd_flags = op; |
1b6d65a0 BVA |
296 | if (data->flags & BLK_MQ_REQ_PREEMPT) |
297 | rq->rq_flags |= RQF_PREEMPT; | |
e4cdf1a1 | 298 | if (blk_queue_io_stat(data->q)) |
e8064021 | 299 | rq->rq_flags |= RQF_IO_STAT; |
af76e555 CH |
300 | /* do not touch atomic flags, it needs atomic ops against the timer */ |
301 | rq->cpu = -1; | |
af76e555 CH |
302 | INIT_HLIST_NODE(&rq->hash); |
303 | RB_CLEAR_NODE(&rq->rb_node); | |
af76e555 CH |
304 | rq->rq_disk = NULL; |
305 | rq->part = NULL; | |
3ee32372 | 306 | rq->start_time = jiffies; |
af76e555 CH |
307 | #ifdef CONFIG_BLK_CGROUP |
308 | rq->rl = NULL; | |
0fec08b4 | 309 | set_start_time_ns(rq); |
af76e555 CH |
310 | rq->io_start_time_ns = 0; |
311 | #endif | |
312 | rq->nr_phys_segments = 0; | |
313 | #if defined(CONFIG_BLK_DEV_INTEGRITY) | |
314 | rq->nr_integrity_segments = 0; | |
315 | #endif | |
af76e555 CH |
316 | rq->special = NULL; |
317 | /* tag was already set */ | |
af76e555 | 318 | rq->extra_len = 0; |
af76e555 | 319 | |
af76e555 | 320 | INIT_LIST_HEAD(&rq->timeout_list); |
f6be4fb4 JA |
321 | rq->timeout = 0; |
322 | ||
af76e555 CH |
323 | rq->end_io = NULL; |
324 | rq->end_io_data = NULL; | |
325 | rq->next_rq = NULL; | |
326 | ||
e4cdf1a1 CH |
327 | data->ctx->rq_dispatched[op_is_sync(op)]++; |
328 | return rq; | |
5dee8577 CH |
329 | } |
330 | ||
d2c0d383 CH |
331 | static struct request *blk_mq_get_request(struct request_queue *q, |
332 | struct bio *bio, unsigned int op, | |
333 | struct blk_mq_alloc_data *data) | |
334 | { | |
335 | struct elevator_queue *e = q->elevator; | |
336 | struct request *rq; | |
e4cdf1a1 | 337 | unsigned int tag; |
21e768b4 | 338 | bool put_ctx_on_error = false; |
d2c0d383 CH |
339 | |
340 | blk_queue_enter_live(q); | |
341 | data->q = q; | |
21e768b4 BVA |
342 | if (likely(!data->ctx)) { |
343 | data->ctx = blk_mq_get_ctx(q); | |
344 | put_ctx_on_error = true; | |
345 | } | |
d2c0d383 CH |
346 | if (likely(!data->hctx)) |
347 | data->hctx = blk_mq_map_queue(q, data->ctx->cpu); | |
03a07c92 GR |
348 | if (op & REQ_NOWAIT) |
349 | data->flags |= BLK_MQ_REQ_NOWAIT; | |
d2c0d383 CH |
350 | |
351 | if (e) { | |
352 | data->flags |= BLK_MQ_REQ_INTERNAL; | |
353 | ||
354 | /* | |
355 | * Flush requests are special and go directly to the | |
356 | * dispatch list. | |
357 | */ | |
5bbf4e5a CH |
358 | if (!op_is_flush(op) && e->type->ops.mq.limit_depth) |
359 | e->type->ops.mq.limit_depth(op, data); | |
d2c0d383 CH |
360 | } |
361 | ||
e4cdf1a1 CH |
362 | tag = blk_mq_get_tag(data); |
363 | if (tag == BLK_MQ_TAG_FAIL) { | |
21e768b4 BVA |
364 | if (put_ctx_on_error) { |
365 | blk_mq_put_ctx(data->ctx); | |
1ad43c00 ML |
366 | data->ctx = NULL; |
367 | } | |
037cebb8 CH |
368 | blk_queue_exit(q); |
369 | return NULL; | |
d2c0d383 CH |
370 | } |
371 | ||
e4cdf1a1 | 372 | rq = blk_mq_rq_ctx_init(data, tag, op); |
037cebb8 CH |
373 | if (!op_is_flush(op)) { |
374 | rq->elv.icq = NULL; | |
5bbf4e5a | 375 | if (e && e->type->ops.mq.prepare_request) { |
44e8c2bf CH |
376 | if (e->type->icq_cache && rq_ioc(bio)) |
377 | blk_mq_sched_assign_ioc(rq, bio); | |
378 | ||
5bbf4e5a CH |
379 | e->type->ops.mq.prepare_request(rq, bio); |
380 | rq->rq_flags |= RQF_ELVPRIV; | |
44e8c2bf | 381 | } |
037cebb8 CH |
382 | } |
383 | data->hctx->queued++; | |
384 | return rq; | |
d2c0d383 CH |
385 | } |
386 | ||
cd6ce148 | 387 | struct request *blk_mq_alloc_request(struct request_queue *q, unsigned int op, |
9a95e4ef | 388 | blk_mq_req_flags_t flags) |
320ae51f | 389 | { |
5a797e00 | 390 | struct blk_mq_alloc_data alloc_data = { .flags = flags }; |
bd166ef1 | 391 | struct request *rq; |
a492f075 | 392 | int ret; |
320ae51f | 393 | |
3a0a5299 | 394 | ret = blk_queue_enter(q, flags); |
a492f075 JL |
395 | if (ret) |
396 | return ERR_PTR(ret); | |
320ae51f | 397 | |
cd6ce148 | 398 | rq = blk_mq_get_request(q, NULL, op, &alloc_data); |
3280d66a | 399 | blk_queue_exit(q); |
841bac2c | 400 | |
bd166ef1 | 401 | if (!rq) |
a492f075 | 402 | return ERR_PTR(-EWOULDBLOCK); |
0c4de0f3 | 403 | |
1ad43c00 | 404 | blk_mq_put_ctx(alloc_data.ctx); |
1ad43c00 | 405 | |
0c4de0f3 CH |
406 | rq->__data_len = 0; |
407 | rq->__sector = (sector_t) -1; | |
408 | rq->bio = rq->biotail = NULL; | |
320ae51f JA |
409 | return rq; |
410 | } | |
4bb659b1 | 411 | EXPORT_SYMBOL(blk_mq_alloc_request); |
320ae51f | 412 | |
cd6ce148 | 413 | struct request *blk_mq_alloc_request_hctx(struct request_queue *q, |
9a95e4ef | 414 | unsigned int op, blk_mq_req_flags_t flags, unsigned int hctx_idx) |
1f5bd336 | 415 | { |
6d2809d5 | 416 | struct blk_mq_alloc_data alloc_data = { .flags = flags }; |
1f5bd336 | 417 | struct request *rq; |
6d2809d5 | 418 | unsigned int cpu; |
1f5bd336 ML |
419 | int ret; |
420 | ||
421 | /* | |
422 | * If the tag allocator sleeps we could get an allocation for a | |
423 | * different hardware context. No need to complicate the low level | |
424 | * allocator for this for the rare use case of a command tied to | |
425 | * a specific queue. | |
426 | */ | |
427 | if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT))) | |
428 | return ERR_PTR(-EINVAL); | |
429 | ||
430 | if (hctx_idx >= q->nr_hw_queues) | |
431 | return ERR_PTR(-EIO); | |
432 | ||
3a0a5299 | 433 | ret = blk_queue_enter(q, flags); |
1f5bd336 ML |
434 | if (ret) |
435 | return ERR_PTR(ret); | |
436 | ||
c8712c6a CH |
437 | /* |
438 | * Check if the hardware context is actually mapped to anything. | |
439 | * If not tell the caller that it should skip this queue. | |
440 | */ | |
6d2809d5 OS |
441 | alloc_data.hctx = q->queue_hw_ctx[hctx_idx]; |
442 | if (!blk_mq_hw_queue_mapped(alloc_data.hctx)) { | |
443 | blk_queue_exit(q); | |
444 | return ERR_PTR(-EXDEV); | |
c8712c6a | 445 | } |
18a15d75 | 446 | cpu = cpumask_first(alloc_data.hctx->cpumask); |
6d2809d5 | 447 | alloc_data.ctx = __blk_mq_get_ctx(q, cpu); |
1f5bd336 | 448 | |
cd6ce148 | 449 | rq = blk_mq_get_request(q, NULL, op, &alloc_data); |
3280d66a | 450 | blk_queue_exit(q); |
c8712c6a | 451 | |
6d2809d5 OS |
452 | if (!rq) |
453 | return ERR_PTR(-EWOULDBLOCK); | |
454 | ||
455 | return rq; | |
1f5bd336 ML |
456 | } |
457 | EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx); | |
458 | ||
6af54051 | 459 | void blk_mq_free_request(struct request *rq) |
320ae51f | 460 | { |
320ae51f | 461 | struct request_queue *q = rq->q; |
6af54051 CH |
462 | struct elevator_queue *e = q->elevator; |
463 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
464 | struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, ctx->cpu); | |
465 | const int sched_tag = rq->internal_tag; | |
466 | ||
5bbf4e5a | 467 | if (rq->rq_flags & RQF_ELVPRIV) { |
6af54051 CH |
468 | if (e && e->type->ops.mq.finish_request) |
469 | e->type->ops.mq.finish_request(rq); | |
470 | if (rq->elv.icq) { | |
471 | put_io_context(rq->elv.icq->ioc); | |
472 | rq->elv.icq = NULL; | |
473 | } | |
474 | } | |
320ae51f | 475 | |
6af54051 | 476 | ctx->rq_completed[rq_is_sync(rq)]++; |
e8064021 | 477 | if (rq->rq_flags & RQF_MQ_INFLIGHT) |
0d2602ca | 478 | atomic_dec(&hctx->nr_active); |
87760e5e | 479 | |
7beb2f84 JA |
480 | if (unlikely(laptop_mode && !blk_rq_is_passthrough(rq))) |
481 | laptop_io_completion(q->backing_dev_info); | |
482 | ||
87760e5e | 483 | wbt_done(q->rq_wb, &rq->issue_stat); |
0d2602ca | 484 | |
85acb3ba SL |
485 | if (blk_rq_rl(rq)) |
486 | blk_put_rl(blk_rq_rl(rq)); | |
487 | ||
af76e555 | 488 | clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags); |
06426adf | 489 | clear_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags); |
bd166ef1 JA |
490 | if (rq->tag != -1) |
491 | blk_mq_put_tag(hctx, hctx->tags, ctx, rq->tag); | |
492 | if (sched_tag != -1) | |
c05f8525 | 493 | blk_mq_put_tag(hctx, hctx->sched_tags, ctx, sched_tag); |
6d8c6c0f | 494 | blk_mq_sched_restart(hctx); |
3ef28e83 | 495 | blk_queue_exit(q); |
320ae51f | 496 | } |
1a3b595a | 497 | EXPORT_SYMBOL_GPL(blk_mq_free_request); |
320ae51f | 498 | |
2a842aca | 499 | inline void __blk_mq_end_request(struct request *rq, blk_status_t error) |
320ae51f | 500 | { |
0d11e6ac ML |
501 | blk_account_io_done(rq); |
502 | ||
91b63639 | 503 | if (rq->end_io) { |
87760e5e | 504 | wbt_done(rq->q->rq_wb, &rq->issue_stat); |
320ae51f | 505 | rq->end_io(rq, error); |
91b63639 CH |
506 | } else { |
507 | if (unlikely(blk_bidi_rq(rq))) | |
508 | blk_mq_free_request(rq->next_rq); | |
320ae51f | 509 | blk_mq_free_request(rq); |
91b63639 | 510 | } |
320ae51f | 511 | } |
c8a446ad | 512 | EXPORT_SYMBOL(__blk_mq_end_request); |
63151a44 | 513 | |
2a842aca | 514 | void blk_mq_end_request(struct request *rq, blk_status_t error) |
63151a44 CH |
515 | { |
516 | if (blk_update_request(rq, error, blk_rq_bytes(rq))) | |
517 | BUG(); | |
c8a446ad | 518 | __blk_mq_end_request(rq, error); |
63151a44 | 519 | } |
c8a446ad | 520 | EXPORT_SYMBOL(blk_mq_end_request); |
320ae51f | 521 | |
30a91cb4 | 522 | static void __blk_mq_complete_request_remote(void *data) |
320ae51f | 523 | { |
3d6efbf6 | 524 | struct request *rq = data; |
320ae51f | 525 | |
30a91cb4 | 526 | rq->q->softirq_done_fn(rq); |
320ae51f | 527 | } |
320ae51f | 528 | |
453f8341 | 529 | static void __blk_mq_complete_request(struct request *rq) |
320ae51f JA |
530 | { |
531 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
38535201 | 532 | bool shared = false; |
320ae51f JA |
533 | int cpu; |
534 | ||
453f8341 CH |
535 | if (rq->internal_tag != -1) |
536 | blk_mq_sched_completed_request(rq); | |
537 | if (rq->rq_flags & RQF_STATS) { | |
538 | blk_mq_poll_stats_start(rq->q); | |
539 | blk_stat_add(rq); | |
540 | } | |
541 | ||
38535201 | 542 | if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) { |
30a91cb4 CH |
543 | rq->q->softirq_done_fn(rq); |
544 | return; | |
545 | } | |
320ae51f JA |
546 | |
547 | cpu = get_cpu(); | |
38535201 CH |
548 | if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags)) |
549 | shared = cpus_share_cache(cpu, ctx->cpu); | |
550 | ||
551 | if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) { | |
30a91cb4 | 552 | rq->csd.func = __blk_mq_complete_request_remote; |
3d6efbf6 CH |
553 | rq->csd.info = rq; |
554 | rq->csd.flags = 0; | |
c46fff2a | 555 | smp_call_function_single_async(ctx->cpu, &rq->csd); |
3d6efbf6 | 556 | } else { |
30a91cb4 | 557 | rq->q->softirq_done_fn(rq); |
3d6efbf6 | 558 | } |
320ae51f JA |
559 | put_cpu(); |
560 | } | |
30a91cb4 CH |
561 | |
562 | /** | |
563 | * blk_mq_complete_request - end I/O on a request | |
564 | * @rq: the request being processed | |
565 | * | |
566 | * Description: | |
567 | * Ends all I/O on a request. It does not handle partial completions. | |
568 | * The actual completion happens out-of-order, through a IPI handler. | |
569 | **/ | |
08e0029a | 570 | void blk_mq_complete_request(struct request *rq) |
30a91cb4 | 571 | { |
95f09684 JA |
572 | struct request_queue *q = rq->q; |
573 | ||
574 | if (unlikely(blk_should_fake_timeout(q))) | |
30a91cb4 | 575 | return; |
08e0029a | 576 | if (!blk_mark_rq_complete(rq)) |
ed851860 | 577 | __blk_mq_complete_request(rq); |
30a91cb4 CH |
578 | } |
579 | EXPORT_SYMBOL(blk_mq_complete_request); | |
320ae51f | 580 | |
973c0191 KB |
581 | int blk_mq_request_started(struct request *rq) |
582 | { | |
583 | return test_bit(REQ_ATOM_STARTED, &rq->atomic_flags); | |
584 | } | |
585 | EXPORT_SYMBOL_GPL(blk_mq_request_started); | |
586 | ||
e2490073 | 587 | void blk_mq_start_request(struct request *rq) |
320ae51f JA |
588 | { |
589 | struct request_queue *q = rq->q; | |
590 | ||
bd166ef1 JA |
591 | blk_mq_sched_started_request(rq); |
592 | ||
320ae51f JA |
593 | trace_block_rq_issue(q, rq); |
594 | ||
cf43e6be | 595 | if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) { |
88eeca49 | 596 | blk_stat_set_issue(&rq->issue_stat, blk_rq_sectors(rq)); |
cf43e6be | 597 | rq->rq_flags |= RQF_STATS; |
87760e5e | 598 | wbt_issue(q->rq_wb, &rq->issue_stat); |
cf43e6be JA |
599 | } |
600 | ||
2b8393b4 | 601 | blk_add_timer(rq); |
87ee7b11 | 602 | |
a7af0af3 | 603 | WARN_ON_ONCE(test_bit(REQ_ATOM_STARTED, &rq->atomic_flags)); |
538b7534 | 604 | |
87ee7b11 JA |
605 | /* |
606 | * Mark us as started and clear complete. Complete might have been | |
607 | * set if requeue raced with timeout, which then marked it as | |
608 | * complete. So be sure to clear complete again when we start | |
609 | * the request, otherwise we'll ignore the completion event. | |
a7af0af3 PZ |
610 | * |
611 | * Ensure that ->deadline is visible before we set STARTED, such that | |
612 | * blk_mq_check_expired() is guaranteed to observe our ->deadline when | |
613 | * it observes STARTED. | |
87ee7b11 | 614 | */ |
a7af0af3 PZ |
615 | smp_wmb(); |
616 | set_bit(REQ_ATOM_STARTED, &rq->atomic_flags); | |
617 | if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags)) { | |
618 | /* | |
619 | * Coherence order guarantees these consecutive stores to a | |
620 | * single variable propagate in the specified order. Thus the | |
621 | * clear_bit() is ordered _after_ the set bit. See | |
622 | * blk_mq_check_expired(). | |
623 | * | |
624 | * (the bits must be part of the same byte for this to be | |
625 | * true). | |
626 | */ | |
4b570521 | 627 | clear_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags); |
a7af0af3 | 628 | } |
49f5baa5 CH |
629 | |
630 | if (q->dma_drain_size && blk_rq_bytes(rq)) { | |
631 | /* | |
632 | * Make sure space for the drain appears. We know we can do | |
633 | * this because max_hw_segments has been adjusted to be one | |
634 | * fewer than the device can handle. | |
635 | */ | |
636 | rq->nr_phys_segments++; | |
637 | } | |
320ae51f | 638 | } |
e2490073 | 639 | EXPORT_SYMBOL(blk_mq_start_request); |
320ae51f | 640 | |
d9d149a3 ML |
641 | /* |
642 | * When we reach here because queue is busy, REQ_ATOM_COMPLETE | |
48b99c9d | 643 | * flag isn't set yet, so there may be race with timeout handler, |
d9d149a3 ML |
644 | * but given rq->deadline is just set in .queue_rq() under |
645 | * this situation, the race won't be possible in reality because | |
646 | * rq->timeout should be set as big enough to cover the window | |
647 | * between blk_mq_start_request() called from .queue_rq() and | |
648 | * clearing REQ_ATOM_STARTED here. | |
649 | */ | |
ed0791b2 | 650 | static void __blk_mq_requeue_request(struct request *rq) |
320ae51f JA |
651 | { |
652 | struct request_queue *q = rq->q; | |
653 | ||
923218f6 ML |
654 | blk_mq_put_driver_tag(rq); |
655 | ||
320ae51f | 656 | trace_block_rq_requeue(q, rq); |
87760e5e | 657 | wbt_requeue(q->rq_wb, &rq->issue_stat); |
49f5baa5 | 658 | |
e2490073 CH |
659 | if (test_and_clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) { |
660 | if (q->dma_drain_size && blk_rq_bytes(rq)) | |
661 | rq->nr_phys_segments--; | |
662 | } | |
320ae51f JA |
663 | } |
664 | ||
2b053aca | 665 | void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list) |
ed0791b2 | 666 | { |
ed0791b2 | 667 | __blk_mq_requeue_request(rq); |
ed0791b2 | 668 | |
4c1f67e5 ML |
669 | /* this request will be re-inserted to io scheduler queue */ |
670 | blk_mq_sched_requeue_request(rq); | |
671 | ||
ed0791b2 | 672 | BUG_ON(blk_queued_rq(rq)); |
2b053aca | 673 | blk_mq_add_to_requeue_list(rq, true, kick_requeue_list); |
ed0791b2 CH |
674 | } |
675 | EXPORT_SYMBOL(blk_mq_requeue_request); | |
676 | ||
6fca6a61 CH |
677 | static void blk_mq_requeue_work(struct work_struct *work) |
678 | { | |
679 | struct request_queue *q = | |
2849450a | 680 | container_of(work, struct request_queue, requeue_work.work); |
6fca6a61 CH |
681 | LIST_HEAD(rq_list); |
682 | struct request *rq, *next; | |
6fca6a61 | 683 | |
18e9781d | 684 | spin_lock_irq(&q->requeue_lock); |
6fca6a61 | 685 | list_splice_init(&q->requeue_list, &rq_list); |
18e9781d | 686 | spin_unlock_irq(&q->requeue_lock); |
6fca6a61 CH |
687 | |
688 | list_for_each_entry_safe(rq, next, &rq_list, queuelist) { | |
e8064021 | 689 | if (!(rq->rq_flags & RQF_SOFTBARRIER)) |
6fca6a61 CH |
690 | continue; |
691 | ||
e8064021 | 692 | rq->rq_flags &= ~RQF_SOFTBARRIER; |
6fca6a61 | 693 | list_del_init(&rq->queuelist); |
bd6737f1 | 694 | blk_mq_sched_insert_request(rq, true, false, false, true); |
6fca6a61 CH |
695 | } |
696 | ||
697 | while (!list_empty(&rq_list)) { | |
698 | rq = list_entry(rq_list.next, struct request, queuelist); | |
699 | list_del_init(&rq->queuelist); | |
bd6737f1 | 700 | blk_mq_sched_insert_request(rq, false, false, false, true); |
6fca6a61 CH |
701 | } |
702 | ||
52d7f1b5 | 703 | blk_mq_run_hw_queues(q, false); |
6fca6a61 CH |
704 | } |
705 | ||
2b053aca BVA |
706 | void blk_mq_add_to_requeue_list(struct request *rq, bool at_head, |
707 | bool kick_requeue_list) | |
6fca6a61 CH |
708 | { |
709 | struct request_queue *q = rq->q; | |
710 | unsigned long flags; | |
711 | ||
712 | /* | |
713 | * We abuse this flag that is otherwise used by the I/O scheduler to | |
ff821d27 | 714 | * request head insertion from the workqueue. |
6fca6a61 | 715 | */ |
e8064021 | 716 | BUG_ON(rq->rq_flags & RQF_SOFTBARRIER); |
6fca6a61 CH |
717 | |
718 | spin_lock_irqsave(&q->requeue_lock, flags); | |
719 | if (at_head) { | |
e8064021 | 720 | rq->rq_flags |= RQF_SOFTBARRIER; |
6fca6a61 CH |
721 | list_add(&rq->queuelist, &q->requeue_list); |
722 | } else { | |
723 | list_add_tail(&rq->queuelist, &q->requeue_list); | |
724 | } | |
725 | spin_unlock_irqrestore(&q->requeue_lock, flags); | |
2b053aca BVA |
726 | |
727 | if (kick_requeue_list) | |
728 | blk_mq_kick_requeue_list(q); | |
6fca6a61 CH |
729 | } |
730 | EXPORT_SYMBOL(blk_mq_add_to_requeue_list); | |
731 | ||
732 | void blk_mq_kick_requeue_list(struct request_queue *q) | |
733 | { | |
2849450a | 734 | kblockd_schedule_delayed_work(&q->requeue_work, 0); |
6fca6a61 CH |
735 | } |
736 | EXPORT_SYMBOL(blk_mq_kick_requeue_list); | |
737 | ||
2849450a MS |
738 | void blk_mq_delay_kick_requeue_list(struct request_queue *q, |
739 | unsigned long msecs) | |
740 | { | |
d4acf365 BVA |
741 | kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, |
742 | msecs_to_jiffies(msecs)); | |
2849450a MS |
743 | } |
744 | EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list); | |
745 | ||
0e62f51f JA |
746 | struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag) |
747 | { | |
88c7b2b7 JA |
748 | if (tag < tags->nr_tags) { |
749 | prefetch(tags->rqs[tag]); | |
4ee86bab | 750 | return tags->rqs[tag]; |
88c7b2b7 | 751 | } |
4ee86bab HR |
752 | |
753 | return NULL; | |
24d2f903 CH |
754 | } |
755 | EXPORT_SYMBOL(blk_mq_tag_to_rq); | |
756 | ||
320ae51f | 757 | struct blk_mq_timeout_data { |
46f92d42 CH |
758 | unsigned long next; |
759 | unsigned int next_set; | |
320ae51f JA |
760 | }; |
761 | ||
90415837 | 762 | void blk_mq_rq_timed_out(struct request *req, bool reserved) |
320ae51f | 763 | { |
f8a5b122 | 764 | const struct blk_mq_ops *ops = req->q->mq_ops; |
46f92d42 | 765 | enum blk_eh_timer_return ret = BLK_EH_RESET_TIMER; |
87ee7b11 JA |
766 | |
767 | /* | |
768 | * We know that complete is set at this point. If STARTED isn't set | |
769 | * anymore, then the request isn't active and the "timeout" should | |
770 | * just be ignored. This can happen due to the bitflag ordering. | |
771 | * Timeout first checks if STARTED is set, and if it is, assumes | |
772 | * the request is active. But if we race with completion, then | |
48b99c9d | 773 | * both flags will get cleared. So check here again, and ignore |
87ee7b11 JA |
774 | * a timeout event with a request that isn't active. |
775 | */ | |
46f92d42 CH |
776 | if (!test_bit(REQ_ATOM_STARTED, &req->atomic_flags)) |
777 | return; | |
87ee7b11 | 778 | |
46f92d42 | 779 | if (ops->timeout) |
0152fb6b | 780 | ret = ops->timeout(req, reserved); |
46f92d42 CH |
781 | |
782 | switch (ret) { | |
783 | case BLK_EH_HANDLED: | |
784 | __blk_mq_complete_request(req); | |
785 | break; | |
786 | case BLK_EH_RESET_TIMER: | |
787 | blk_add_timer(req); | |
788 | blk_clear_rq_complete(req); | |
789 | break; | |
790 | case BLK_EH_NOT_HANDLED: | |
791 | break; | |
792 | default: | |
793 | printk(KERN_ERR "block: bad eh return: %d\n", ret); | |
794 | break; | |
795 | } | |
87ee7b11 | 796 | } |
5b3f25fc | 797 | |
81481eb4 CH |
798 | static void blk_mq_check_expired(struct blk_mq_hw_ctx *hctx, |
799 | struct request *rq, void *priv, bool reserved) | |
800 | { | |
801 | struct blk_mq_timeout_data *data = priv; | |
a7af0af3 | 802 | unsigned long deadline; |
87ee7b11 | 803 | |
95a49603 | 804 | if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) |
46f92d42 | 805 | return; |
87ee7b11 | 806 | |
a7af0af3 PZ |
807 | /* |
808 | * Ensures that if we see STARTED we must also see our | |
809 | * up-to-date deadline, see blk_mq_start_request(). | |
810 | */ | |
811 | smp_rmb(); | |
812 | ||
813 | deadline = READ_ONCE(rq->deadline); | |
814 | ||
d9d149a3 ML |
815 | /* |
816 | * The rq being checked may have been freed and reallocated | |
817 | * out already here, we avoid this race by checking rq->deadline | |
818 | * and REQ_ATOM_COMPLETE flag together: | |
819 | * | |
820 | * - if rq->deadline is observed as new value because of | |
821 | * reusing, the rq won't be timed out because of timing. | |
822 | * - if rq->deadline is observed as previous value, | |
823 | * REQ_ATOM_COMPLETE flag won't be cleared in reuse path | |
824 | * because we put a barrier between setting rq->deadline | |
825 | * and clearing the flag in blk_mq_start_request(), so | |
826 | * this rq won't be timed out too. | |
827 | */ | |
a7af0af3 PZ |
828 | if (time_after_eq(jiffies, deadline)) { |
829 | if (!blk_mark_rq_complete(rq)) { | |
830 | /* | |
831 | * Again coherence order ensures that consecutive reads | |
832 | * from the same variable must be in that order. This | |
833 | * ensures that if we see COMPLETE clear, we must then | |
834 | * see STARTED set and we'll ignore this timeout. | |
835 | * | |
836 | * (There's also the MB implied by the test_and_clear()) | |
837 | */ | |
0152fb6b | 838 | blk_mq_rq_timed_out(rq, reserved); |
a7af0af3 PZ |
839 | } |
840 | } else if (!data->next_set || time_after(data->next, deadline)) { | |
841 | data->next = deadline; | |
46f92d42 CH |
842 | data->next_set = 1; |
843 | } | |
87ee7b11 JA |
844 | } |
845 | ||
287922eb | 846 | static void blk_mq_timeout_work(struct work_struct *work) |
320ae51f | 847 | { |
287922eb CH |
848 | struct request_queue *q = |
849 | container_of(work, struct request_queue, timeout_work); | |
81481eb4 CH |
850 | struct blk_mq_timeout_data data = { |
851 | .next = 0, | |
852 | .next_set = 0, | |
853 | }; | |
81481eb4 | 854 | int i; |
320ae51f | 855 | |
71f79fb3 GKB |
856 | /* A deadlock might occur if a request is stuck requiring a |
857 | * timeout at the same time a queue freeze is waiting | |
858 | * completion, since the timeout code would not be able to | |
859 | * acquire the queue reference here. | |
860 | * | |
861 | * That's why we don't use blk_queue_enter here; instead, we use | |
862 | * percpu_ref_tryget directly, because we need to be able to | |
863 | * obtain a reference even in the short window between the queue | |
864 | * starting to freeze, by dropping the first reference in | |
1671d522 | 865 | * blk_freeze_queue_start, and the moment the last request is |
71f79fb3 GKB |
866 | * consumed, marked by the instant q_usage_counter reaches |
867 | * zero. | |
868 | */ | |
869 | if (!percpu_ref_tryget(&q->q_usage_counter)) | |
287922eb CH |
870 | return; |
871 | ||
0bf6cd5b | 872 | blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &data); |
320ae51f | 873 | |
81481eb4 CH |
874 | if (data.next_set) { |
875 | data.next = blk_rq_timeout(round_jiffies_up(data.next)); | |
876 | mod_timer(&q->timeout, data.next); | |
0d2602ca | 877 | } else { |
0bf6cd5b CH |
878 | struct blk_mq_hw_ctx *hctx; |
879 | ||
f054b56c ML |
880 | queue_for_each_hw_ctx(q, hctx, i) { |
881 | /* the hctx may be unmapped, so check it here */ | |
882 | if (blk_mq_hw_queue_mapped(hctx)) | |
883 | blk_mq_tag_idle(hctx); | |
884 | } | |
0d2602ca | 885 | } |
287922eb | 886 | blk_queue_exit(q); |
320ae51f JA |
887 | } |
888 | ||
88459642 OS |
889 | struct flush_busy_ctx_data { |
890 | struct blk_mq_hw_ctx *hctx; | |
891 | struct list_head *list; | |
892 | }; | |
893 | ||
894 | static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data) | |
895 | { | |
896 | struct flush_busy_ctx_data *flush_data = data; | |
897 | struct blk_mq_hw_ctx *hctx = flush_data->hctx; | |
898 | struct blk_mq_ctx *ctx = hctx->ctxs[bitnr]; | |
899 | ||
900 | sbitmap_clear_bit(sb, bitnr); | |
901 | spin_lock(&ctx->lock); | |
902 | list_splice_tail_init(&ctx->rq_list, flush_data->list); | |
903 | spin_unlock(&ctx->lock); | |
904 | return true; | |
905 | } | |
906 | ||
1429d7c9 JA |
907 | /* |
908 | * Process software queues that have been marked busy, splicing them | |
909 | * to the for-dispatch | |
910 | */ | |
2c3ad667 | 911 | void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list) |
1429d7c9 | 912 | { |
88459642 OS |
913 | struct flush_busy_ctx_data data = { |
914 | .hctx = hctx, | |
915 | .list = list, | |
916 | }; | |
1429d7c9 | 917 | |
88459642 | 918 | sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data); |
1429d7c9 | 919 | } |
2c3ad667 | 920 | EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs); |
1429d7c9 | 921 | |
b347689f ML |
922 | struct dispatch_rq_data { |
923 | struct blk_mq_hw_ctx *hctx; | |
924 | struct request *rq; | |
925 | }; | |
926 | ||
927 | static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr, | |
928 | void *data) | |
929 | { | |
930 | struct dispatch_rq_data *dispatch_data = data; | |
931 | struct blk_mq_hw_ctx *hctx = dispatch_data->hctx; | |
932 | struct blk_mq_ctx *ctx = hctx->ctxs[bitnr]; | |
933 | ||
934 | spin_lock(&ctx->lock); | |
935 | if (unlikely(!list_empty(&ctx->rq_list))) { | |
936 | dispatch_data->rq = list_entry_rq(ctx->rq_list.next); | |
937 | list_del_init(&dispatch_data->rq->queuelist); | |
938 | if (list_empty(&ctx->rq_list)) | |
939 | sbitmap_clear_bit(sb, bitnr); | |
940 | } | |
941 | spin_unlock(&ctx->lock); | |
942 | ||
943 | return !dispatch_data->rq; | |
944 | } | |
945 | ||
946 | struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx, | |
947 | struct blk_mq_ctx *start) | |
948 | { | |
949 | unsigned off = start ? start->index_hw : 0; | |
950 | struct dispatch_rq_data data = { | |
951 | .hctx = hctx, | |
952 | .rq = NULL, | |
953 | }; | |
954 | ||
955 | __sbitmap_for_each_set(&hctx->ctx_map, off, | |
956 | dispatch_rq_from_ctx, &data); | |
957 | ||
958 | return data.rq; | |
959 | } | |
960 | ||
703fd1c0 JA |
961 | static inline unsigned int queued_to_index(unsigned int queued) |
962 | { | |
963 | if (!queued) | |
964 | return 0; | |
1429d7c9 | 965 | |
703fd1c0 | 966 | return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1); |
1429d7c9 JA |
967 | } |
968 | ||
bd6737f1 JA |
969 | bool blk_mq_get_driver_tag(struct request *rq, struct blk_mq_hw_ctx **hctx, |
970 | bool wait) | |
bd166ef1 JA |
971 | { |
972 | struct blk_mq_alloc_data data = { | |
973 | .q = rq->q, | |
bd166ef1 JA |
974 | .hctx = blk_mq_map_queue(rq->q, rq->mq_ctx->cpu), |
975 | .flags = wait ? 0 : BLK_MQ_REQ_NOWAIT, | |
976 | }; | |
977 | ||
5feeacdd JA |
978 | might_sleep_if(wait); |
979 | ||
81380ca1 OS |
980 | if (rq->tag != -1) |
981 | goto done; | |
bd166ef1 | 982 | |
415b806d SG |
983 | if (blk_mq_tag_is_reserved(data.hctx->sched_tags, rq->internal_tag)) |
984 | data.flags |= BLK_MQ_REQ_RESERVED; | |
985 | ||
bd166ef1 JA |
986 | rq->tag = blk_mq_get_tag(&data); |
987 | if (rq->tag >= 0) { | |
200e86b3 JA |
988 | if (blk_mq_tag_busy(data.hctx)) { |
989 | rq->rq_flags |= RQF_MQ_INFLIGHT; | |
990 | atomic_inc(&data.hctx->nr_active); | |
991 | } | |
bd166ef1 | 992 | data.hctx->tags->rqs[rq->tag] = rq; |
bd166ef1 JA |
993 | } |
994 | ||
81380ca1 OS |
995 | done: |
996 | if (hctx) | |
997 | *hctx = data.hctx; | |
998 | return rq->tag != -1; | |
bd166ef1 JA |
999 | } |
1000 | ||
eb619fdb JA |
1001 | static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode, |
1002 | int flags, void *key) | |
da55f2cc OS |
1003 | { |
1004 | struct blk_mq_hw_ctx *hctx; | |
1005 | ||
1006 | hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait); | |
1007 | ||
eb619fdb | 1008 | list_del_init(&wait->entry); |
da55f2cc OS |
1009 | blk_mq_run_hw_queue(hctx, true); |
1010 | return 1; | |
1011 | } | |
1012 | ||
f906a6a0 JA |
1013 | /* |
1014 | * Mark us waiting for a tag. For shared tags, this involves hooking us into | |
1015 | * the tag wakeups. For non-shared tags, we can simply mark us nedeing a | |
1016 | * restart. For both caes, take care to check the condition again after | |
1017 | * marking us as waiting. | |
1018 | */ | |
1019 | static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx **hctx, | |
1020 | struct request *rq) | |
da55f2cc | 1021 | { |
eb619fdb | 1022 | struct blk_mq_hw_ctx *this_hctx = *hctx; |
f906a6a0 | 1023 | bool shared_tags = (this_hctx->flags & BLK_MQ_F_TAG_SHARED) != 0; |
da55f2cc | 1024 | struct sbq_wait_state *ws; |
f906a6a0 JA |
1025 | wait_queue_entry_t *wait; |
1026 | bool ret; | |
da55f2cc | 1027 | |
f906a6a0 JA |
1028 | if (!shared_tags) { |
1029 | if (!test_bit(BLK_MQ_S_SCHED_RESTART, &this_hctx->state)) | |
1030 | set_bit(BLK_MQ_S_SCHED_RESTART, &this_hctx->state); | |
1031 | } else { | |
1032 | wait = &this_hctx->dispatch_wait; | |
1033 | if (!list_empty_careful(&wait->entry)) | |
1034 | return false; | |
1035 | ||
1036 | spin_lock(&this_hctx->lock); | |
1037 | if (!list_empty(&wait->entry)) { | |
1038 | spin_unlock(&this_hctx->lock); | |
1039 | return false; | |
1040 | } | |
eb619fdb | 1041 | |
f906a6a0 JA |
1042 | ws = bt_wait_ptr(&this_hctx->tags->bitmap_tags, this_hctx); |
1043 | add_wait_queue(&ws->wait, wait); | |
eb619fdb JA |
1044 | } |
1045 | ||
da55f2cc | 1046 | /* |
eb619fdb JA |
1047 | * It's possible that a tag was freed in the window between the |
1048 | * allocation failure and adding the hardware queue to the wait | |
1049 | * queue. | |
da55f2cc | 1050 | */ |
f906a6a0 JA |
1051 | ret = blk_mq_get_driver_tag(rq, hctx, false); |
1052 | ||
1053 | if (!shared_tags) { | |
1054 | /* | |
1055 | * Don't clear RESTART here, someone else could have set it. | |
1056 | * At most this will cost an extra queue run. | |
1057 | */ | |
1058 | return ret; | |
1059 | } else { | |
1060 | if (!ret) { | |
1061 | spin_unlock(&this_hctx->lock); | |
1062 | return false; | |
1063 | } | |
1064 | ||
1065 | /* | |
1066 | * We got a tag, remove ourselves from the wait queue to ensure | |
1067 | * someone else gets the wakeup. | |
1068 | */ | |
1069 | spin_lock_irq(&ws->wait.lock); | |
1070 | list_del_init(&wait->entry); | |
1071 | spin_unlock_irq(&ws->wait.lock); | |
eb619fdb | 1072 | spin_unlock(&this_hctx->lock); |
f906a6a0 | 1073 | return true; |
eb619fdb | 1074 | } |
da55f2cc OS |
1075 | } |
1076 | ||
de148297 | 1077 | bool blk_mq_dispatch_rq_list(struct request_queue *q, struct list_head *list, |
eb619fdb | 1078 | bool got_budget) |
320ae51f | 1079 | { |
81380ca1 | 1080 | struct blk_mq_hw_ctx *hctx; |
6d6f167c | 1081 | struct request *rq, *nxt; |
eb619fdb | 1082 | bool no_tag = false; |
fc17b653 | 1083 | int errors, queued; |
320ae51f | 1084 | |
81380ca1 OS |
1085 | if (list_empty(list)) |
1086 | return false; | |
1087 | ||
de148297 ML |
1088 | WARN_ON(!list_is_singular(list) && got_budget); |
1089 | ||
320ae51f JA |
1090 | /* |
1091 | * Now process all the entries, sending them to the driver. | |
1092 | */ | |
93efe981 | 1093 | errors = queued = 0; |
81380ca1 | 1094 | do { |
74c45052 | 1095 | struct blk_mq_queue_data bd; |
fc17b653 | 1096 | blk_status_t ret; |
320ae51f | 1097 | |
f04c3df3 | 1098 | rq = list_first_entry(list, struct request, queuelist); |
bd166ef1 | 1099 | if (!blk_mq_get_driver_tag(rq, &hctx, false)) { |
3c782d67 | 1100 | /* |
da55f2cc | 1101 | * The initial allocation attempt failed, so we need to |
eb619fdb JA |
1102 | * rerun the hardware queue when a tag is freed. The |
1103 | * waitqueue takes care of that. If the queue is run | |
1104 | * before we add this entry back on the dispatch list, | |
1105 | * we'll re-run it below. | |
3c782d67 | 1106 | */ |
f906a6a0 | 1107 | if (!blk_mq_mark_tag_wait(&hctx, rq)) { |
de148297 ML |
1108 | if (got_budget) |
1109 | blk_mq_put_dispatch_budget(hctx); | |
f906a6a0 JA |
1110 | /* |
1111 | * For non-shared tags, the RESTART check | |
1112 | * will suffice. | |
1113 | */ | |
1114 | if (hctx->flags & BLK_MQ_F_TAG_SHARED) | |
1115 | no_tag = true; | |
de148297 ML |
1116 | break; |
1117 | } | |
1118 | } | |
1119 | ||
0c6af1cc ML |
1120 | if (!got_budget && !blk_mq_get_dispatch_budget(hctx)) { |
1121 | blk_mq_put_driver_tag(rq); | |
88022d72 | 1122 | break; |
0c6af1cc | 1123 | } |
da55f2cc | 1124 | |
320ae51f | 1125 | list_del_init(&rq->queuelist); |
320ae51f | 1126 | |
74c45052 | 1127 | bd.rq = rq; |
113285b4 JA |
1128 | |
1129 | /* | |
1130 | * Flag last if we have no more requests, or if we have more | |
1131 | * but can't assign a driver tag to it. | |
1132 | */ | |
1133 | if (list_empty(list)) | |
1134 | bd.last = true; | |
1135 | else { | |
113285b4 JA |
1136 | nxt = list_first_entry(list, struct request, queuelist); |
1137 | bd.last = !blk_mq_get_driver_tag(nxt, NULL, false); | |
1138 | } | |
74c45052 JA |
1139 | |
1140 | ret = q->mq_ops->queue_rq(hctx, &bd); | |
fc17b653 | 1141 | if (ret == BLK_STS_RESOURCE) { |
6d6f167c JW |
1142 | /* |
1143 | * If an I/O scheduler has been configured and we got a | |
ff821d27 JA |
1144 | * driver tag for the next request already, free it |
1145 | * again. | |
6d6f167c JW |
1146 | */ |
1147 | if (!list_empty(list)) { | |
1148 | nxt = list_first_entry(list, struct request, queuelist); | |
1149 | blk_mq_put_driver_tag(nxt); | |
1150 | } | |
f04c3df3 | 1151 | list_add(&rq->queuelist, list); |
ed0791b2 | 1152 | __blk_mq_requeue_request(rq); |
320ae51f | 1153 | break; |
fc17b653 CH |
1154 | } |
1155 | ||
1156 | if (unlikely(ret != BLK_STS_OK)) { | |
93efe981 | 1157 | errors++; |
2a842aca | 1158 | blk_mq_end_request(rq, BLK_STS_IOERR); |
fc17b653 | 1159 | continue; |
320ae51f JA |
1160 | } |
1161 | ||
fc17b653 | 1162 | queued++; |
81380ca1 | 1163 | } while (!list_empty(list)); |
320ae51f | 1164 | |
703fd1c0 | 1165 | hctx->dispatched[queued_to_index(queued)]++; |
320ae51f JA |
1166 | |
1167 | /* | |
1168 | * Any items that need requeuing? Stuff them into hctx->dispatch, | |
1169 | * that is where we will continue on next queue run. | |
1170 | */ | |
f04c3df3 | 1171 | if (!list_empty(list)) { |
320ae51f | 1172 | spin_lock(&hctx->lock); |
c13660a0 | 1173 | list_splice_init(list, &hctx->dispatch); |
320ae51f | 1174 | spin_unlock(&hctx->lock); |
f04c3df3 | 1175 | |
9ba52e58 | 1176 | /* |
710c785f BVA |
1177 | * If SCHED_RESTART was set by the caller of this function and |
1178 | * it is no longer set that means that it was cleared by another | |
1179 | * thread and hence that a queue rerun is needed. | |
9ba52e58 | 1180 | * |
eb619fdb JA |
1181 | * If 'no_tag' is set, that means that we failed getting |
1182 | * a driver tag with an I/O scheduler attached. If our dispatch | |
1183 | * waitqueue is no longer active, ensure that we run the queue | |
1184 | * AFTER adding our entries back to the list. | |
bd166ef1 | 1185 | * |
710c785f BVA |
1186 | * If no I/O scheduler has been configured it is possible that |
1187 | * the hardware queue got stopped and restarted before requests | |
1188 | * were pushed back onto the dispatch list. Rerun the queue to | |
1189 | * avoid starvation. Notes: | |
1190 | * - blk_mq_run_hw_queue() checks whether or not a queue has | |
1191 | * been stopped before rerunning a queue. | |
1192 | * - Some but not all block drivers stop a queue before | |
fc17b653 | 1193 | * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq |
710c785f | 1194 | * and dm-rq. |
bd166ef1 | 1195 | */ |
eb619fdb JA |
1196 | if (!blk_mq_sched_needs_restart(hctx) || |
1197 | (no_tag && list_empty_careful(&hctx->dispatch_wait.entry))) | |
bd166ef1 | 1198 | blk_mq_run_hw_queue(hctx, true); |
320ae51f | 1199 | } |
f04c3df3 | 1200 | |
93efe981 | 1201 | return (queued + errors) != 0; |
f04c3df3 JA |
1202 | } |
1203 | ||
6a83e74d BVA |
1204 | static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx) |
1205 | { | |
1206 | int srcu_idx; | |
1207 | ||
b7a71e66 JA |
1208 | /* |
1209 | * We should be running this queue from one of the CPUs that | |
1210 | * are mapped to it. | |
54011168 ML |
1211 | * |
1212 | * There are at least two related races now between setting | |
1213 | * hctx->next_cpu from blk_mq_hctx_next_cpu() and running | |
1214 | * __blk_mq_run_hw_queue(): | |
1215 | * | |
1216 | * - hctx->next_cpu is found offline in blk_mq_hctx_next_cpu(), | |
1217 | * but later it becomes online, then this warning is harmless | |
1218 | * at all | |
1219 | * | |
1220 | * - hctx->next_cpu is found online in blk_mq_hctx_next_cpu(), | |
1221 | * but later it becomes offline, then the warning can't be | |
1222 | * triggered, and we depend on blk-mq timeout handler to | |
1223 | * handle dispatched requests to this hctx | |
b7a71e66 | 1224 | */ |
54011168 ML |
1225 | if (!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) && |
1226 | cpu_online(hctx->next_cpu)) { | |
1227 | printk(KERN_WARNING "run queue from wrong CPU %d, hctx %s\n", | |
1228 | raw_smp_processor_id(), | |
1229 | cpumask_empty(hctx->cpumask) ? "inactive": "active"); | |
1230 | dump_stack(); | |
1231 | } | |
6a83e74d | 1232 | |
b7a71e66 JA |
1233 | /* |
1234 | * We can't run the queue inline with ints disabled. Ensure that | |
1235 | * we catch bad users of this early. | |
1236 | */ | |
1237 | WARN_ON_ONCE(in_interrupt()); | |
1238 | ||
6a83e74d BVA |
1239 | if (!(hctx->flags & BLK_MQ_F_BLOCKING)) { |
1240 | rcu_read_lock(); | |
1f460b63 | 1241 | blk_mq_sched_dispatch_requests(hctx); |
6a83e74d BVA |
1242 | rcu_read_unlock(); |
1243 | } else { | |
bf4907c0 JA |
1244 | might_sleep(); |
1245 | ||
07319678 | 1246 | srcu_idx = srcu_read_lock(hctx->queue_rq_srcu); |
1f460b63 | 1247 | blk_mq_sched_dispatch_requests(hctx); |
07319678 | 1248 | srcu_read_unlock(hctx->queue_rq_srcu, srcu_idx); |
6a83e74d BVA |
1249 | } |
1250 | } | |
1251 | ||
506e931f JA |
1252 | /* |
1253 | * It'd be great if the workqueue API had a way to pass | |
1254 | * in a mask and had some smarts for more clever placement. | |
1255 | * For now we just round-robin here, switching for every | |
1256 | * BLK_MQ_CPU_WORK_BATCH queued items. | |
1257 | */ | |
1258 | static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx) | |
1259 | { | |
b657d7e6 CH |
1260 | if (hctx->queue->nr_hw_queues == 1) |
1261 | return WORK_CPU_UNBOUND; | |
506e931f JA |
1262 | |
1263 | if (--hctx->next_cpu_batch <= 0) { | |
c02ebfdd | 1264 | int next_cpu; |
506e931f | 1265 | |
18a15d75 | 1266 | next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask); |
506e931f | 1267 | if (next_cpu >= nr_cpu_ids) |
18a15d75 | 1268 | next_cpu = cpumask_first(hctx->cpumask); |
506e931f JA |
1269 | |
1270 | hctx->next_cpu = next_cpu; | |
1271 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; | |
1272 | } | |
1273 | ||
b657d7e6 | 1274 | return hctx->next_cpu; |
506e931f JA |
1275 | } |
1276 | ||
7587a5ae BVA |
1277 | static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async, |
1278 | unsigned long msecs) | |
320ae51f | 1279 | { |
5435c023 BVA |
1280 | if (WARN_ON_ONCE(!blk_mq_hw_queue_mapped(hctx))) |
1281 | return; | |
1282 | ||
1283 | if (unlikely(blk_mq_hctx_stopped(hctx))) | |
320ae51f JA |
1284 | return; |
1285 | ||
1b792f2f | 1286 | if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) { |
2a90d4aa PB |
1287 | int cpu = get_cpu(); |
1288 | if (cpumask_test_cpu(cpu, hctx->cpumask)) { | |
398205b8 | 1289 | __blk_mq_run_hw_queue(hctx); |
2a90d4aa | 1290 | put_cpu(); |
398205b8 PB |
1291 | return; |
1292 | } | |
e4043dcf | 1293 | |
2a90d4aa | 1294 | put_cpu(); |
e4043dcf | 1295 | } |
398205b8 | 1296 | |
9f993737 JA |
1297 | kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx), |
1298 | &hctx->run_work, | |
1299 | msecs_to_jiffies(msecs)); | |
7587a5ae BVA |
1300 | } |
1301 | ||
1302 | void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs) | |
1303 | { | |
1304 | __blk_mq_delay_run_hw_queue(hctx, true, msecs); | |
1305 | } | |
1306 | EXPORT_SYMBOL(blk_mq_delay_run_hw_queue); | |
1307 | ||
79f720a7 | 1308 | bool blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) |
7587a5ae | 1309 | { |
79f720a7 JA |
1310 | if (blk_mq_hctx_has_pending(hctx)) { |
1311 | __blk_mq_delay_run_hw_queue(hctx, async, 0); | |
1312 | return true; | |
1313 | } | |
1314 | ||
1315 | return false; | |
320ae51f | 1316 | } |
5b727272 | 1317 | EXPORT_SYMBOL(blk_mq_run_hw_queue); |
320ae51f | 1318 | |
b94ec296 | 1319 | void blk_mq_run_hw_queues(struct request_queue *q, bool async) |
320ae51f JA |
1320 | { |
1321 | struct blk_mq_hw_ctx *hctx; | |
1322 | int i; | |
1323 | ||
1324 | queue_for_each_hw_ctx(q, hctx, i) { | |
79f720a7 | 1325 | if (blk_mq_hctx_stopped(hctx)) |
320ae51f JA |
1326 | continue; |
1327 | ||
b94ec296 | 1328 | blk_mq_run_hw_queue(hctx, async); |
320ae51f JA |
1329 | } |
1330 | } | |
b94ec296 | 1331 | EXPORT_SYMBOL(blk_mq_run_hw_queues); |
320ae51f | 1332 | |
fd001443 BVA |
1333 | /** |
1334 | * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped | |
1335 | * @q: request queue. | |
1336 | * | |
1337 | * The caller is responsible for serializing this function against | |
1338 | * blk_mq_{start,stop}_hw_queue(). | |
1339 | */ | |
1340 | bool blk_mq_queue_stopped(struct request_queue *q) | |
1341 | { | |
1342 | struct blk_mq_hw_ctx *hctx; | |
1343 | int i; | |
1344 | ||
1345 | queue_for_each_hw_ctx(q, hctx, i) | |
1346 | if (blk_mq_hctx_stopped(hctx)) | |
1347 | return true; | |
1348 | ||
1349 | return false; | |
1350 | } | |
1351 | EXPORT_SYMBOL(blk_mq_queue_stopped); | |
1352 | ||
39a70c76 ML |
1353 | /* |
1354 | * This function is often used for pausing .queue_rq() by driver when | |
1355 | * there isn't enough resource or some conditions aren't satisfied, and | |
4d606219 | 1356 | * BLK_STS_RESOURCE is usually returned. |
39a70c76 ML |
1357 | * |
1358 | * We do not guarantee that dispatch can be drained or blocked | |
1359 | * after blk_mq_stop_hw_queue() returns. Please use | |
1360 | * blk_mq_quiesce_queue() for that requirement. | |
1361 | */ | |
2719aa21 JA |
1362 | void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx) |
1363 | { | |
641a9ed6 | 1364 | cancel_delayed_work(&hctx->run_work); |
280d45f6 | 1365 | |
641a9ed6 | 1366 | set_bit(BLK_MQ_S_STOPPED, &hctx->state); |
2719aa21 | 1367 | } |
641a9ed6 | 1368 | EXPORT_SYMBOL(blk_mq_stop_hw_queue); |
2719aa21 | 1369 | |
39a70c76 ML |
1370 | /* |
1371 | * This function is often used for pausing .queue_rq() by driver when | |
1372 | * there isn't enough resource or some conditions aren't satisfied, and | |
4d606219 | 1373 | * BLK_STS_RESOURCE is usually returned. |
39a70c76 ML |
1374 | * |
1375 | * We do not guarantee that dispatch can be drained or blocked | |
1376 | * after blk_mq_stop_hw_queues() returns. Please use | |
1377 | * blk_mq_quiesce_queue() for that requirement. | |
1378 | */ | |
2719aa21 JA |
1379 | void blk_mq_stop_hw_queues(struct request_queue *q) |
1380 | { | |
641a9ed6 ML |
1381 | struct blk_mq_hw_ctx *hctx; |
1382 | int i; | |
1383 | ||
1384 | queue_for_each_hw_ctx(q, hctx, i) | |
1385 | blk_mq_stop_hw_queue(hctx); | |
280d45f6 CH |
1386 | } |
1387 | EXPORT_SYMBOL(blk_mq_stop_hw_queues); | |
1388 | ||
320ae51f JA |
1389 | void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx) |
1390 | { | |
1391 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
e4043dcf | 1392 | |
0ffbce80 | 1393 | blk_mq_run_hw_queue(hctx, false); |
320ae51f JA |
1394 | } |
1395 | EXPORT_SYMBOL(blk_mq_start_hw_queue); | |
1396 | ||
2f268556 CH |
1397 | void blk_mq_start_hw_queues(struct request_queue *q) |
1398 | { | |
1399 | struct blk_mq_hw_ctx *hctx; | |
1400 | int i; | |
1401 | ||
1402 | queue_for_each_hw_ctx(q, hctx, i) | |
1403 | blk_mq_start_hw_queue(hctx); | |
1404 | } | |
1405 | EXPORT_SYMBOL(blk_mq_start_hw_queues); | |
1406 | ||
ae911c5e JA |
1407 | void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) |
1408 | { | |
1409 | if (!blk_mq_hctx_stopped(hctx)) | |
1410 | return; | |
1411 | ||
1412 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
1413 | blk_mq_run_hw_queue(hctx, async); | |
1414 | } | |
1415 | EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue); | |
1416 | ||
1b4a3258 | 1417 | void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async) |
320ae51f JA |
1418 | { |
1419 | struct blk_mq_hw_ctx *hctx; | |
1420 | int i; | |
1421 | ||
ae911c5e JA |
1422 | queue_for_each_hw_ctx(q, hctx, i) |
1423 | blk_mq_start_stopped_hw_queue(hctx, async); | |
320ae51f JA |
1424 | } |
1425 | EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues); | |
1426 | ||
70f4db63 | 1427 | static void blk_mq_run_work_fn(struct work_struct *work) |
320ae51f JA |
1428 | { |
1429 | struct blk_mq_hw_ctx *hctx; | |
1430 | ||
9f993737 | 1431 | hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work); |
320ae51f | 1432 | |
21c6e939 JA |
1433 | /* |
1434 | * If we are stopped, don't run the queue. The exception is if | |
1435 | * BLK_MQ_S_START_ON_RUN is set. For that case, we auto-clear | |
1436 | * the STOPPED bit and run it. | |
1437 | */ | |
1438 | if (test_bit(BLK_MQ_S_STOPPED, &hctx->state)) { | |
1439 | if (!test_bit(BLK_MQ_S_START_ON_RUN, &hctx->state)) | |
1440 | return; | |
7587a5ae | 1441 | |
21c6e939 JA |
1442 | clear_bit(BLK_MQ_S_START_ON_RUN, &hctx->state); |
1443 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
1444 | } | |
7587a5ae BVA |
1445 | |
1446 | __blk_mq_run_hw_queue(hctx); | |
1447 | } | |
1448 | ||
70f4db63 CH |
1449 | |
1450 | void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs) | |
1451 | { | |
5435c023 | 1452 | if (WARN_ON_ONCE(!blk_mq_hw_queue_mapped(hctx))) |
19c66e59 | 1453 | return; |
70f4db63 | 1454 | |
21c6e939 JA |
1455 | /* |
1456 | * Stop the hw queue, then modify currently delayed work. | |
1457 | * This should prevent us from running the queue prematurely. | |
1458 | * Mark the queue as auto-clearing STOPPED when it runs. | |
1459 | */ | |
7e79dadc | 1460 | blk_mq_stop_hw_queue(hctx); |
21c6e939 JA |
1461 | set_bit(BLK_MQ_S_START_ON_RUN, &hctx->state); |
1462 | kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), | |
1463 | &hctx->run_work, | |
1464 | msecs_to_jiffies(msecs)); | |
70f4db63 CH |
1465 | } |
1466 | EXPORT_SYMBOL(blk_mq_delay_queue); | |
1467 | ||
cfd0c552 | 1468 | static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx, |
cfd0c552 ML |
1469 | struct request *rq, |
1470 | bool at_head) | |
320ae51f | 1471 | { |
e57690fe JA |
1472 | struct blk_mq_ctx *ctx = rq->mq_ctx; |
1473 | ||
7b607814 BVA |
1474 | lockdep_assert_held(&ctx->lock); |
1475 | ||
01b983c9 JA |
1476 | trace_block_rq_insert(hctx->queue, rq); |
1477 | ||
72a0a36e CH |
1478 | if (at_head) |
1479 | list_add(&rq->queuelist, &ctx->rq_list); | |
1480 | else | |
1481 | list_add_tail(&rq->queuelist, &ctx->rq_list); | |
cfd0c552 | 1482 | } |
4bb659b1 | 1483 | |
2c3ad667 JA |
1484 | void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq, |
1485 | bool at_head) | |
cfd0c552 ML |
1486 | { |
1487 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
1488 | ||
7b607814 BVA |
1489 | lockdep_assert_held(&ctx->lock); |
1490 | ||
e57690fe | 1491 | __blk_mq_insert_req_list(hctx, rq, at_head); |
320ae51f | 1492 | blk_mq_hctx_mark_pending(hctx, ctx); |
320ae51f JA |
1493 | } |
1494 | ||
157f377b JA |
1495 | /* |
1496 | * Should only be used carefully, when the caller knows we want to | |
1497 | * bypass a potential IO scheduler on the target device. | |
1498 | */ | |
b0850297 | 1499 | void blk_mq_request_bypass_insert(struct request *rq, bool run_queue) |
157f377b JA |
1500 | { |
1501 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
1502 | struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(rq->q, ctx->cpu); | |
1503 | ||
1504 | spin_lock(&hctx->lock); | |
1505 | list_add_tail(&rq->queuelist, &hctx->dispatch); | |
1506 | spin_unlock(&hctx->lock); | |
1507 | ||
b0850297 ML |
1508 | if (run_queue) |
1509 | blk_mq_run_hw_queue(hctx, false); | |
157f377b JA |
1510 | } |
1511 | ||
bd166ef1 JA |
1512 | void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx, |
1513 | struct list_head *list) | |
320ae51f JA |
1514 | |
1515 | { | |
320ae51f JA |
1516 | /* |
1517 | * preemption doesn't flush plug list, so it's possible ctx->cpu is | |
1518 | * offline now | |
1519 | */ | |
1520 | spin_lock(&ctx->lock); | |
1521 | while (!list_empty(list)) { | |
1522 | struct request *rq; | |
1523 | ||
1524 | rq = list_first_entry(list, struct request, queuelist); | |
e57690fe | 1525 | BUG_ON(rq->mq_ctx != ctx); |
320ae51f | 1526 | list_del_init(&rq->queuelist); |
e57690fe | 1527 | __blk_mq_insert_req_list(hctx, rq, false); |
320ae51f | 1528 | } |
cfd0c552 | 1529 | blk_mq_hctx_mark_pending(hctx, ctx); |
320ae51f | 1530 | spin_unlock(&ctx->lock); |
320ae51f JA |
1531 | } |
1532 | ||
1533 | static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b) | |
1534 | { | |
1535 | struct request *rqa = container_of(a, struct request, queuelist); | |
1536 | struct request *rqb = container_of(b, struct request, queuelist); | |
1537 | ||
1538 | return !(rqa->mq_ctx < rqb->mq_ctx || | |
1539 | (rqa->mq_ctx == rqb->mq_ctx && | |
1540 | blk_rq_pos(rqa) < blk_rq_pos(rqb))); | |
1541 | } | |
1542 | ||
1543 | void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule) | |
1544 | { | |
1545 | struct blk_mq_ctx *this_ctx; | |
1546 | struct request_queue *this_q; | |
1547 | struct request *rq; | |
1548 | LIST_HEAD(list); | |
1549 | LIST_HEAD(ctx_list); | |
1550 | unsigned int depth; | |
1551 | ||
1552 | list_splice_init(&plug->mq_list, &list); | |
1553 | ||
1554 | list_sort(NULL, &list, plug_ctx_cmp); | |
1555 | ||
1556 | this_q = NULL; | |
1557 | this_ctx = NULL; | |
1558 | depth = 0; | |
1559 | ||
1560 | while (!list_empty(&list)) { | |
1561 | rq = list_entry_rq(list.next); | |
1562 | list_del_init(&rq->queuelist); | |
1563 | BUG_ON(!rq->q); | |
1564 | if (rq->mq_ctx != this_ctx) { | |
1565 | if (this_ctx) { | |
bd166ef1 JA |
1566 | trace_block_unplug(this_q, depth, from_schedule); |
1567 | blk_mq_sched_insert_requests(this_q, this_ctx, | |
1568 | &ctx_list, | |
1569 | from_schedule); | |
320ae51f JA |
1570 | } |
1571 | ||
1572 | this_ctx = rq->mq_ctx; | |
1573 | this_q = rq->q; | |
1574 | depth = 0; | |
1575 | } | |
1576 | ||
1577 | depth++; | |
1578 | list_add_tail(&rq->queuelist, &ctx_list); | |
1579 | } | |
1580 | ||
1581 | /* | |
1582 | * If 'this_ctx' is set, we know we have entries to complete | |
1583 | * on 'ctx_list'. Do those. | |
1584 | */ | |
1585 | if (this_ctx) { | |
bd166ef1 JA |
1586 | trace_block_unplug(this_q, depth, from_schedule); |
1587 | blk_mq_sched_insert_requests(this_q, this_ctx, &ctx_list, | |
1588 | from_schedule); | |
320ae51f JA |
1589 | } |
1590 | } | |
1591 | ||
1592 | static void blk_mq_bio_to_request(struct request *rq, struct bio *bio) | |
1593 | { | |
da8d7f07 | 1594 | blk_init_request_from_bio(rq, bio); |
4b570521 | 1595 | |
85acb3ba SL |
1596 | blk_rq_set_rl(rq, blk_get_rl(rq->q, bio)); |
1597 | ||
6e85eaf3 | 1598 | blk_account_io_start(rq, true); |
320ae51f JA |
1599 | } |
1600 | ||
ab42f35d ML |
1601 | static inline void blk_mq_queue_io(struct blk_mq_hw_ctx *hctx, |
1602 | struct blk_mq_ctx *ctx, | |
1603 | struct request *rq) | |
1604 | { | |
1605 | spin_lock(&ctx->lock); | |
1606 | __blk_mq_insert_request(hctx, rq, false); | |
1607 | spin_unlock(&ctx->lock); | |
07068d5b | 1608 | } |
14ec77f3 | 1609 | |
fd2d3326 JA |
1610 | static blk_qc_t request_to_qc_t(struct blk_mq_hw_ctx *hctx, struct request *rq) |
1611 | { | |
bd166ef1 JA |
1612 | if (rq->tag != -1) |
1613 | return blk_tag_to_qc_t(rq->tag, hctx->queue_num, false); | |
1614 | ||
1615 | return blk_tag_to_qc_t(rq->internal_tag, hctx->queue_num, true); | |
fd2d3326 JA |
1616 | } |
1617 | ||
d964f04a ML |
1618 | static void __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx, |
1619 | struct request *rq, | |
1620 | blk_qc_t *cookie, bool may_sleep) | |
f984df1f | 1621 | { |
f984df1f | 1622 | struct request_queue *q = rq->q; |
f984df1f SL |
1623 | struct blk_mq_queue_data bd = { |
1624 | .rq = rq, | |
d945a365 | 1625 | .last = true, |
f984df1f | 1626 | }; |
bd166ef1 | 1627 | blk_qc_t new_cookie; |
f06345ad | 1628 | blk_status_t ret; |
d964f04a ML |
1629 | bool run_queue = true; |
1630 | ||
f4560ffe ML |
1631 | /* RCU or SRCU read lock is needed before checking quiesced flag */ |
1632 | if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)) { | |
d964f04a ML |
1633 | run_queue = false; |
1634 | goto insert; | |
1635 | } | |
f984df1f | 1636 | |
bd166ef1 | 1637 | if (q->elevator) |
2253efc8 BVA |
1638 | goto insert; |
1639 | ||
d964f04a | 1640 | if (!blk_mq_get_driver_tag(rq, NULL, false)) |
bd166ef1 JA |
1641 | goto insert; |
1642 | ||
88022d72 | 1643 | if (!blk_mq_get_dispatch_budget(hctx)) { |
de148297 ML |
1644 | blk_mq_put_driver_tag(rq); |
1645 | goto insert; | |
88022d72 | 1646 | } |
de148297 | 1647 | |
bd166ef1 JA |
1648 | new_cookie = request_to_qc_t(hctx, rq); |
1649 | ||
f984df1f SL |
1650 | /* |
1651 | * For OK queue, we are done. For error, kill it. Any other | |
1652 | * error (busy), just add it to our list as we previously | |
1653 | * would have done | |
1654 | */ | |
1655 | ret = q->mq_ops->queue_rq(hctx, &bd); | |
fc17b653 CH |
1656 | switch (ret) { |
1657 | case BLK_STS_OK: | |
7b371636 | 1658 | *cookie = new_cookie; |
2253efc8 | 1659 | return; |
fc17b653 CH |
1660 | case BLK_STS_RESOURCE: |
1661 | __blk_mq_requeue_request(rq); | |
1662 | goto insert; | |
1663 | default: | |
7b371636 | 1664 | *cookie = BLK_QC_T_NONE; |
fc17b653 | 1665 | blk_mq_end_request(rq, ret); |
2253efc8 | 1666 | return; |
f984df1f | 1667 | } |
7b371636 | 1668 | |
2253efc8 | 1669 | insert: |
d964f04a | 1670 | blk_mq_sched_insert_request(rq, false, run_queue, false, may_sleep); |
f984df1f SL |
1671 | } |
1672 | ||
5eb6126e CH |
1673 | static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx, |
1674 | struct request *rq, blk_qc_t *cookie) | |
1675 | { | |
1676 | if (!(hctx->flags & BLK_MQ_F_BLOCKING)) { | |
1677 | rcu_read_lock(); | |
d964f04a | 1678 | __blk_mq_try_issue_directly(hctx, rq, cookie, false); |
5eb6126e CH |
1679 | rcu_read_unlock(); |
1680 | } else { | |
bf4907c0 JA |
1681 | unsigned int srcu_idx; |
1682 | ||
1683 | might_sleep(); | |
1684 | ||
07319678 | 1685 | srcu_idx = srcu_read_lock(hctx->queue_rq_srcu); |
d964f04a | 1686 | __blk_mq_try_issue_directly(hctx, rq, cookie, true); |
07319678 | 1687 | srcu_read_unlock(hctx->queue_rq_srcu, srcu_idx); |
5eb6126e CH |
1688 | } |
1689 | } | |
1690 | ||
dece1635 | 1691 | static blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio) |
07068d5b | 1692 | { |
ef295ecf | 1693 | const int is_sync = op_is_sync(bio->bi_opf); |
f73f44eb | 1694 | const int is_flush_fua = op_is_flush(bio->bi_opf); |
5a797e00 | 1695 | struct blk_mq_alloc_data data = { .flags = 0 }; |
07068d5b | 1696 | struct request *rq; |
5eb6126e | 1697 | unsigned int request_count = 0; |
f984df1f | 1698 | struct blk_plug *plug; |
5b3f341f | 1699 | struct request *same_queue_rq = NULL; |
7b371636 | 1700 | blk_qc_t cookie; |
87760e5e | 1701 | unsigned int wb_acct; |
07068d5b JA |
1702 | |
1703 | blk_queue_bounce(q, &bio); | |
1704 | ||
af67c31f | 1705 | blk_queue_split(q, &bio); |
f36ea50c | 1706 | |
e23947bd | 1707 | if (!bio_integrity_prep(bio)) |
dece1635 | 1708 | return BLK_QC_T_NONE; |
07068d5b | 1709 | |
87c279e6 OS |
1710 | if (!is_flush_fua && !blk_queue_nomerges(q) && |
1711 | blk_attempt_plug_merge(q, bio, &request_count, &same_queue_rq)) | |
1712 | return BLK_QC_T_NONE; | |
f984df1f | 1713 | |
bd166ef1 JA |
1714 | if (blk_mq_sched_bio_merge(q, bio)) |
1715 | return BLK_QC_T_NONE; | |
1716 | ||
87760e5e JA |
1717 | wb_acct = wbt_wait(q->rq_wb, bio, NULL); |
1718 | ||
bd166ef1 JA |
1719 | trace_block_getrq(q, bio, bio->bi_opf); |
1720 | ||
d2c0d383 | 1721 | rq = blk_mq_get_request(q, bio, bio->bi_opf, &data); |
87760e5e JA |
1722 | if (unlikely(!rq)) { |
1723 | __wbt_done(q->rq_wb, wb_acct); | |
03a07c92 GR |
1724 | if (bio->bi_opf & REQ_NOWAIT) |
1725 | bio_wouldblock_error(bio); | |
dece1635 | 1726 | return BLK_QC_T_NONE; |
87760e5e JA |
1727 | } |
1728 | ||
1729 | wbt_track(&rq->issue_stat, wb_acct); | |
07068d5b | 1730 | |
fd2d3326 | 1731 | cookie = request_to_qc_t(data.hctx, rq); |
07068d5b | 1732 | |
f984df1f | 1733 | plug = current->plug; |
07068d5b | 1734 | if (unlikely(is_flush_fua)) { |
f984df1f | 1735 | blk_mq_put_ctx(data.ctx); |
07068d5b | 1736 | blk_mq_bio_to_request(rq, bio); |
923218f6 ML |
1737 | |
1738 | /* bypass scheduler for flush rq */ | |
1739 | blk_insert_flush(rq); | |
1740 | blk_mq_run_hw_queue(data.hctx, true); | |
a4d907b6 | 1741 | } else if (plug && q->nr_hw_queues == 1) { |
600271d9 SL |
1742 | struct request *last = NULL; |
1743 | ||
b00c53e8 | 1744 | blk_mq_put_ctx(data.ctx); |
e6c4438b | 1745 | blk_mq_bio_to_request(rq, bio); |
0a6219a9 ML |
1746 | |
1747 | /* | |
1748 | * @request_count may become stale because of schedule | |
1749 | * out, so check the list again. | |
1750 | */ | |
1751 | if (list_empty(&plug->mq_list)) | |
1752 | request_count = 0; | |
254d259d CH |
1753 | else if (blk_queue_nomerges(q)) |
1754 | request_count = blk_plug_queued_count(q); | |
1755 | ||
676d0607 | 1756 | if (!request_count) |
e6c4438b | 1757 | trace_block_plug(q); |
600271d9 SL |
1758 | else |
1759 | last = list_entry_rq(plug->mq_list.prev); | |
b094f89c | 1760 | |
600271d9 SL |
1761 | if (request_count >= BLK_MAX_REQUEST_COUNT || (last && |
1762 | blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) { | |
e6c4438b JM |
1763 | blk_flush_plug_list(plug, false); |
1764 | trace_block_plug(q); | |
320ae51f | 1765 | } |
b094f89c | 1766 | |
e6c4438b | 1767 | list_add_tail(&rq->queuelist, &plug->mq_list); |
2299722c | 1768 | } else if (plug && !blk_queue_nomerges(q)) { |
bd166ef1 | 1769 | blk_mq_bio_to_request(rq, bio); |
07068d5b | 1770 | |
07068d5b | 1771 | /* |
6a83e74d | 1772 | * We do limited plugging. If the bio can be merged, do that. |
f984df1f SL |
1773 | * Otherwise the existing request in the plug list will be |
1774 | * issued. So the plug list will have one request at most | |
2299722c CH |
1775 | * The plug list might get flushed before this. If that happens, |
1776 | * the plug list is empty, and same_queue_rq is invalid. | |
07068d5b | 1777 | */ |
2299722c CH |
1778 | if (list_empty(&plug->mq_list)) |
1779 | same_queue_rq = NULL; | |
1780 | if (same_queue_rq) | |
1781 | list_del_init(&same_queue_rq->queuelist); | |
1782 | list_add_tail(&rq->queuelist, &plug->mq_list); | |
1783 | ||
bf4907c0 JA |
1784 | blk_mq_put_ctx(data.ctx); |
1785 | ||
dad7a3be ML |
1786 | if (same_queue_rq) { |
1787 | data.hctx = blk_mq_map_queue(q, | |
1788 | same_queue_rq->mq_ctx->cpu); | |
2299722c CH |
1789 | blk_mq_try_issue_directly(data.hctx, same_queue_rq, |
1790 | &cookie); | |
dad7a3be | 1791 | } |
a4d907b6 | 1792 | } else if (q->nr_hw_queues > 1 && is_sync) { |
bf4907c0 | 1793 | blk_mq_put_ctx(data.ctx); |
2299722c | 1794 | blk_mq_bio_to_request(rq, bio); |
2299722c | 1795 | blk_mq_try_issue_directly(data.hctx, rq, &cookie); |
a4d907b6 | 1796 | } else if (q->elevator) { |
b00c53e8 | 1797 | blk_mq_put_ctx(data.ctx); |
bd166ef1 | 1798 | blk_mq_bio_to_request(rq, bio); |
a4d907b6 | 1799 | blk_mq_sched_insert_request(rq, false, true, true, true); |
ab42f35d | 1800 | } else { |
b00c53e8 | 1801 | blk_mq_put_ctx(data.ctx); |
ab42f35d ML |
1802 | blk_mq_bio_to_request(rq, bio); |
1803 | blk_mq_queue_io(data.hctx, data.ctx, rq); | |
a4d907b6 | 1804 | blk_mq_run_hw_queue(data.hctx, true); |
ab42f35d | 1805 | } |
320ae51f | 1806 | |
7b371636 | 1807 | return cookie; |
320ae51f JA |
1808 | } |
1809 | ||
cc71a6f4 JA |
1810 | void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags, |
1811 | unsigned int hctx_idx) | |
95363efd | 1812 | { |
e9b267d9 | 1813 | struct page *page; |
320ae51f | 1814 | |
24d2f903 | 1815 | if (tags->rqs && set->ops->exit_request) { |
e9b267d9 | 1816 | int i; |
320ae51f | 1817 | |
24d2f903 | 1818 | for (i = 0; i < tags->nr_tags; i++) { |
2af8cbe3 JA |
1819 | struct request *rq = tags->static_rqs[i]; |
1820 | ||
1821 | if (!rq) | |
e9b267d9 | 1822 | continue; |
d6296d39 | 1823 | set->ops->exit_request(set, rq, hctx_idx); |
2af8cbe3 | 1824 | tags->static_rqs[i] = NULL; |
e9b267d9 | 1825 | } |
320ae51f | 1826 | } |
320ae51f | 1827 | |
24d2f903 CH |
1828 | while (!list_empty(&tags->page_list)) { |
1829 | page = list_first_entry(&tags->page_list, struct page, lru); | |
6753471c | 1830 | list_del_init(&page->lru); |
f75782e4 CM |
1831 | /* |
1832 | * Remove kmemleak object previously allocated in | |
1833 | * blk_mq_init_rq_map(). | |
1834 | */ | |
1835 | kmemleak_free(page_address(page)); | |
320ae51f JA |
1836 | __free_pages(page, page->private); |
1837 | } | |
cc71a6f4 | 1838 | } |
320ae51f | 1839 | |
cc71a6f4 JA |
1840 | void blk_mq_free_rq_map(struct blk_mq_tags *tags) |
1841 | { | |
24d2f903 | 1842 | kfree(tags->rqs); |
cc71a6f4 | 1843 | tags->rqs = NULL; |
2af8cbe3 JA |
1844 | kfree(tags->static_rqs); |
1845 | tags->static_rqs = NULL; | |
320ae51f | 1846 | |
24d2f903 | 1847 | blk_mq_free_tags(tags); |
320ae51f JA |
1848 | } |
1849 | ||
cc71a6f4 JA |
1850 | struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, |
1851 | unsigned int hctx_idx, | |
1852 | unsigned int nr_tags, | |
1853 | unsigned int reserved_tags) | |
320ae51f | 1854 | { |
24d2f903 | 1855 | struct blk_mq_tags *tags; |
59f082e4 | 1856 | int node; |
320ae51f | 1857 | |
59f082e4 SL |
1858 | node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx); |
1859 | if (node == NUMA_NO_NODE) | |
1860 | node = set->numa_node; | |
1861 | ||
1862 | tags = blk_mq_init_tags(nr_tags, reserved_tags, node, | |
24391c0d | 1863 | BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags)); |
24d2f903 CH |
1864 | if (!tags) |
1865 | return NULL; | |
320ae51f | 1866 | |
cc71a6f4 | 1867 | tags->rqs = kzalloc_node(nr_tags * sizeof(struct request *), |
36e1f3d1 | 1868 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, |
59f082e4 | 1869 | node); |
24d2f903 CH |
1870 | if (!tags->rqs) { |
1871 | blk_mq_free_tags(tags); | |
1872 | return NULL; | |
1873 | } | |
320ae51f | 1874 | |
2af8cbe3 JA |
1875 | tags->static_rqs = kzalloc_node(nr_tags * sizeof(struct request *), |
1876 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, | |
59f082e4 | 1877 | node); |
2af8cbe3 JA |
1878 | if (!tags->static_rqs) { |
1879 | kfree(tags->rqs); | |
1880 | blk_mq_free_tags(tags); | |
1881 | return NULL; | |
1882 | } | |
1883 | ||
cc71a6f4 JA |
1884 | return tags; |
1885 | } | |
1886 | ||
1887 | static size_t order_to_size(unsigned int order) | |
1888 | { | |
1889 | return (size_t)PAGE_SIZE << order; | |
1890 | } | |
1891 | ||
1892 | int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags, | |
1893 | unsigned int hctx_idx, unsigned int depth) | |
1894 | { | |
1895 | unsigned int i, j, entries_per_page, max_order = 4; | |
1896 | size_t rq_size, left; | |
59f082e4 SL |
1897 | int node; |
1898 | ||
1899 | node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx); | |
1900 | if (node == NUMA_NO_NODE) | |
1901 | node = set->numa_node; | |
cc71a6f4 JA |
1902 | |
1903 | INIT_LIST_HEAD(&tags->page_list); | |
1904 | ||
320ae51f JA |
1905 | /* |
1906 | * rq_size is the size of the request plus driver payload, rounded | |
1907 | * to the cacheline size | |
1908 | */ | |
24d2f903 | 1909 | rq_size = round_up(sizeof(struct request) + set->cmd_size, |
320ae51f | 1910 | cache_line_size()); |
cc71a6f4 | 1911 | left = rq_size * depth; |
320ae51f | 1912 | |
cc71a6f4 | 1913 | for (i = 0; i < depth; ) { |
320ae51f JA |
1914 | int this_order = max_order; |
1915 | struct page *page; | |
1916 | int to_do; | |
1917 | void *p; | |
1918 | ||
b3a834b1 | 1919 | while (this_order && left < order_to_size(this_order - 1)) |
320ae51f JA |
1920 | this_order--; |
1921 | ||
1922 | do { | |
59f082e4 | 1923 | page = alloc_pages_node(node, |
36e1f3d1 | 1924 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO, |
a5164405 | 1925 | this_order); |
320ae51f JA |
1926 | if (page) |
1927 | break; | |
1928 | if (!this_order--) | |
1929 | break; | |
1930 | if (order_to_size(this_order) < rq_size) | |
1931 | break; | |
1932 | } while (1); | |
1933 | ||
1934 | if (!page) | |
24d2f903 | 1935 | goto fail; |
320ae51f JA |
1936 | |
1937 | page->private = this_order; | |
24d2f903 | 1938 | list_add_tail(&page->lru, &tags->page_list); |
320ae51f JA |
1939 | |
1940 | p = page_address(page); | |
f75782e4 CM |
1941 | /* |
1942 | * Allow kmemleak to scan these pages as they contain pointers | |
1943 | * to additional allocations like via ops->init_request(). | |
1944 | */ | |
36e1f3d1 | 1945 | kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO); |
320ae51f | 1946 | entries_per_page = order_to_size(this_order) / rq_size; |
cc71a6f4 | 1947 | to_do = min(entries_per_page, depth - i); |
320ae51f JA |
1948 | left -= to_do * rq_size; |
1949 | for (j = 0; j < to_do; j++) { | |
2af8cbe3 JA |
1950 | struct request *rq = p; |
1951 | ||
1952 | tags->static_rqs[i] = rq; | |
24d2f903 | 1953 | if (set->ops->init_request) { |
d6296d39 | 1954 | if (set->ops->init_request(set, rq, hctx_idx, |
59f082e4 | 1955 | node)) { |
2af8cbe3 | 1956 | tags->static_rqs[i] = NULL; |
24d2f903 | 1957 | goto fail; |
a5164405 | 1958 | } |
e9b267d9 CH |
1959 | } |
1960 | ||
320ae51f JA |
1961 | p += rq_size; |
1962 | i++; | |
1963 | } | |
1964 | } | |
cc71a6f4 | 1965 | return 0; |
320ae51f | 1966 | |
24d2f903 | 1967 | fail: |
cc71a6f4 JA |
1968 | blk_mq_free_rqs(set, tags, hctx_idx); |
1969 | return -ENOMEM; | |
320ae51f JA |
1970 | } |
1971 | ||
e57690fe JA |
1972 | /* |
1973 | * 'cpu' is going away. splice any existing rq_list entries from this | |
1974 | * software queue to the hw queue dispatch list, and ensure that it | |
1975 | * gets run. | |
1976 | */ | |
9467f859 | 1977 | static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node) |
484b4061 | 1978 | { |
9467f859 | 1979 | struct blk_mq_hw_ctx *hctx; |
484b4061 JA |
1980 | struct blk_mq_ctx *ctx; |
1981 | LIST_HEAD(tmp); | |
1982 | ||
9467f859 | 1983 | hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead); |
e57690fe | 1984 | ctx = __blk_mq_get_ctx(hctx->queue, cpu); |
484b4061 JA |
1985 | |
1986 | spin_lock(&ctx->lock); | |
1987 | if (!list_empty(&ctx->rq_list)) { | |
1988 | list_splice_init(&ctx->rq_list, &tmp); | |
1989 | blk_mq_hctx_clear_pending(hctx, ctx); | |
1990 | } | |
1991 | spin_unlock(&ctx->lock); | |
1992 | ||
1993 | if (list_empty(&tmp)) | |
9467f859 | 1994 | return 0; |
484b4061 | 1995 | |
e57690fe JA |
1996 | spin_lock(&hctx->lock); |
1997 | list_splice_tail_init(&tmp, &hctx->dispatch); | |
1998 | spin_unlock(&hctx->lock); | |
484b4061 JA |
1999 | |
2000 | blk_mq_run_hw_queue(hctx, true); | |
9467f859 | 2001 | return 0; |
484b4061 JA |
2002 | } |
2003 | ||
9467f859 | 2004 | static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx) |
484b4061 | 2005 | { |
9467f859 TG |
2006 | cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD, |
2007 | &hctx->cpuhp_dead); | |
484b4061 JA |
2008 | } |
2009 | ||
c3b4afca | 2010 | /* hctx->ctxs will be freed in queue's release handler */ |
08e98fc6 ML |
2011 | static void blk_mq_exit_hctx(struct request_queue *q, |
2012 | struct blk_mq_tag_set *set, | |
2013 | struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx) | |
2014 | { | |
9c1051aa OS |
2015 | blk_mq_debugfs_unregister_hctx(hctx); |
2016 | ||
25a2bfb9 ML |
2017 | if (blk_mq_hw_queue_mapped(hctx)) |
2018 | blk_mq_tag_idle(hctx); | |
08e98fc6 | 2019 | |
f70ced09 | 2020 | if (set->ops->exit_request) |
d6296d39 | 2021 | set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx); |
f70ced09 | 2022 | |
93252632 OS |
2023 | blk_mq_sched_exit_hctx(q, hctx, hctx_idx); |
2024 | ||
08e98fc6 ML |
2025 | if (set->ops->exit_hctx) |
2026 | set->ops->exit_hctx(hctx, hctx_idx); | |
2027 | ||
6a83e74d | 2028 | if (hctx->flags & BLK_MQ_F_BLOCKING) |
07319678 | 2029 | cleanup_srcu_struct(hctx->queue_rq_srcu); |
6a83e74d | 2030 | |
9467f859 | 2031 | blk_mq_remove_cpuhp(hctx); |
f70ced09 | 2032 | blk_free_flush_queue(hctx->fq); |
88459642 | 2033 | sbitmap_free(&hctx->ctx_map); |
08e98fc6 ML |
2034 | } |
2035 | ||
624dbe47 ML |
2036 | static void blk_mq_exit_hw_queues(struct request_queue *q, |
2037 | struct blk_mq_tag_set *set, int nr_queue) | |
2038 | { | |
2039 | struct blk_mq_hw_ctx *hctx; | |
2040 | unsigned int i; | |
2041 | ||
2042 | queue_for_each_hw_ctx(q, hctx, i) { | |
2043 | if (i == nr_queue) | |
2044 | break; | |
08e98fc6 | 2045 | blk_mq_exit_hctx(q, set, hctx, i); |
624dbe47 | 2046 | } |
624dbe47 ML |
2047 | } |
2048 | ||
08e98fc6 ML |
2049 | static int blk_mq_init_hctx(struct request_queue *q, |
2050 | struct blk_mq_tag_set *set, | |
2051 | struct blk_mq_hw_ctx *hctx, unsigned hctx_idx) | |
320ae51f | 2052 | { |
08e98fc6 ML |
2053 | int node; |
2054 | ||
2055 | node = hctx->numa_node; | |
2056 | if (node == NUMA_NO_NODE) | |
2057 | node = hctx->numa_node = set->numa_node; | |
2058 | ||
9f993737 | 2059 | INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn); |
08e98fc6 ML |
2060 | spin_lock_init(&hctx->lock); |
2061 | INIT_LIST_HEAD(&hctx->dispatch); | |
2062 | hctx->queue = q; | |
2404e607 | 2063 | hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED; |
08e98fc6 | 2064 | |
9467f859 | 2065 | cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead); |
08e98fc6 ML |
2066 | |
2067 | hctx->tags = set->tags[hctx_idx]; | |
320ae51f JA |
2068 | |
2069 | /* | |
08e98fc6 ML |
2070 | * Allocate space for all possible cpus to avoid allocation at |
2071 | * runtime | |
320ae51f | 2072 | */ |
d904bfa7 | 2073 | hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *), |
08e98fc6 ML |
2074 | GFP_KERNEL, node); |
2075 | if (!hctx->ctxs) | |
2076 | goto unregister_cpu_notifier; | |
320ae51f | 2077 | |
88459642 OS |
2078 | if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8), GFP_KERNEL, |
2079 | node)) | |
08e98fc6 | 2080 | goto free_ctxs; |
320ae51f | 2081 | |
08e98fc6 | 2082 | hctx->nr_ctx = 0; |
320ae51f | 2083 | |
eb619fdb JA |
2084 | init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake); |
2085 | INIT_LIST_HEAD(&hctx->dispatch_wait.entry); | |
2086 | ||
08e98fc6 ML |
2087 | if (set->ops->init_hctx && |
2088 | set->ops->init_hctx(hctx, set->driver_data, hctx_idx)) | |
2089 | goto free_bitmap; | |
320ae51f | 2090 | |
93252632 OS |
2091 | if (blk_mq_sched_init_hctx(q, hctx, hctx_idx)) |
2092 | goto exit_hctx; | |
2093 | ||
f70ced09 ML |
2094 | hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size); |
2095 | if (!hctx->fq) | |
93252632 | 2096 | goto sched_exit_hctx; |
320ae51f | 2097 | |
f70ced09 | 2098 | if (set->ops->init_request && |
d6296d39 CH |
2099 | set->ops->init_request(set, hctx->fq->flush_rq, hctx_idx, |
2100 | node)) | |
f70ced09 | 2101 | goto free_fq; |
320ae51f | 2102 | |
6a83e74d | 2103 | if (hctx->flags & BLK_MQ_F_BLOCKING) |
07319678 | 2104 | init_srcu_struct(hctx->queue_rq_srcu); |
6a83e74d | 2105 | |
9c1051aa OS |
2106 | blk_mq_debugfs_register_hctx(q, hctx); |
2107 | ||
08e98fc6 | 2108 | return 0; |
320ae51f | 2109 | |
f70ced09 ML |
2110 | free_fq: |
2111 | kfree(hctx->fq); | |
93252632 OS |
2112 | sched_exit_hctx: |
2113 | blk_mq_sched_exit_hctx(q, hctx, hctx_idx); | |
f70ced09 ML |
2114 | exit_hctx: |
2115 | if (set->ops->exit_hctx) | |
2116 | set->ops->exit_hctx(hctx, hctx_idx); | |
08e98fc6 | 2117 | free_bitmap: |
88459642 | 2118 | sbitmap_free(&hctx->ctx_map); |
08e98fc6 ML |
2119 | free_ctxs: |
2120 | kfree(hctx->ctxs); | |
2121 | unregister_cpu_notifier: | |
9467f859 | 2122 | blk_mq_remove_cpuhp(hctx); |
08e98fc6 ML |
2123 | return -1; |
2124 | } | |
320ae51f | 2125 | |
320ae51f JA |
2126 | static void blk_mq_init_cpu_queues(struct request_queue *q, |
2127 | unsigned int nr_hw_queues) | |
2128 | { | |
2129 | unsigned int i; | |
2130 | ||
2131 | for_each_possible_cpu(i) { | |
2132 | struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i); | |
2133 | struct blk_mq_hw_ctx *hctx; | |
2134 | ||
320ae51f JA |
2135 | __ctx->cpu = i; |
2136 | spin_lock_init(&__ctx->lock); | |
2137 | INIT_LIST_HEAD(&__ctx->rq_list); | |
2138 | __ctx->queue = q; | |
2139 | ||
18a15d75 SF |
2140 | /* If the cpu isn't present, the cpu is mapped to first hctx */ |
2141 | if (!cpu_present(i)) | |
2142 | continue; | |
2143 | ||
2144 | hctx = blk_mq_map_queue(q, i); | |
2145 | ||
320ae51f JA |
2146 | /* |
2147 | * Set local node, IFF we have more than one hw queue. If | |
2148 | * not, we remain on the home node of the device | |
2149 | */ | |
2150 | if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE) | |
bffed457 | 2151 | hctx->numa_node = local_memory_node(cpu_to_node(i)); |
320ae51f JA |
2152 | } |
2153 | } | |
2154 | ||
cc71a6f4 JA |
2155 | static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, int hctx_idx) |
2156 | { | |
2157 | int ret = 0; | |
2158 | ||
2159 | set->tags[hctx_idx] = blk_mq_alloc_rq_map(set, hctx_idx, | |
2160 | set->queue_depth, set->reserved_tags); | |
2161 | if (!set->tags[hctx_idx]) | |
2162 | return false; | |
2163 | ||
2164 | ret = blk_mq_alloc_rqs(set, set->tags[hctx_idx], hctx_idx, | |
2165 | set->queue_depth); | |
2166 | if (!ret) | |
2167 | return true; | |
2168 | ||
2169 | blk_mq_free_rq_map(set->tags[hctx_idx]); | |
2170 | set->tags[hctx_idx] = NULL; | |
2171 | return false; | |
2172 | } | |
2173 | ||
2174 | static void blk_mq_free_map_and_requests(struct blk_mq_tag_set *set, | |
2175 | unsigned int hctx_idx) | |
2176 | { | |
bd166ef1 JA |
2177 | if (set->tags[hctx_idx]) { |
2178 | blk_mq_free_rqs(set, set->tags[hctx_idx], hctx_idx); | |
2179 | blk_mq_free_rq_map(set->tags[hctx_idx]); | |
2180 | set->tags[hctx_idx] = NULL; | |
2181 | } | |
cc71a6f4 JA |
2182 | } |
2183 | ||
4b855ad3 | 2184 | static void blk_mq_map_swqueue(struct request_queue *q) |
320ae51f | 2185 | { |
d1b1cea1 | 2186 | unsigned int i, hctx_idx; |
320ae51f JA |
2187 | struct blk_mq_hw_ctx *hctx; |
2188 | struct blk_mq_ctx *ctx; | |
2a34c087 | 2189 | struct blk_mq_tag_set *set = q->tag_set; |
320ae51f | 2190 | |
60de074b AM |
2191 | /* |
2192 | * Avoid others reading imcomplete hctx->cpumask through sysfs | |
2193 | */ | |
2194 | mutex_lock(&q->sysfs_lock); | |
2195 | ||
320ae51f | 2196 | queue_for_each_hw_ctx(q, hctx, i) { |
e4043dcf | 2197 | cpumask_clear(hctx->cpumask); |
320ae51f JA |
2198 | hctx->nr_ctx = 0; |
2199 | } | |
2200 | ||
2201 | /* | |
4b855ad3 CH |
2202 | * Map software to hardware queues. |
2203 | * | |
2204 | * If the cpu isn't present, the cpu is mapped to first hctx. | |
320ae51f | 2205 | */ |
18a15d75 | 2206 | for_each_present_cpu(i) { |
d1b1cea1 GKB |
2207 | hctx_idx = q->mq_map[i]; |
2208 | /* unmapped hw queue can be remapped after CPU topo changed */ | |
cc71a6f4 JA |
2209 | if (!set->tags[hctx_idx] && |
2210 | !__blk_mq_alloc_rq_map(set, hctx_idx)) { | |
d1b1cea1 GKB |
2211 | /* |
2212 | * If tags initialization fail for some hctx, | |
2213 | * that hctx won't be brought online. In this | |
2214 | * case, remap the current ctx to hctx[0] which | |
2215 | * is guaranteed to always have tags allocated | |
2216 | */ | |
cc71a6f4 | 2217 | q->mq_map[i] = 0; |
d1b1cea1 GKB |
2218 | } |
2219 | ||
897bb0c7 | 2220 | ctx = per_cpu_ptr(q->queue_ctx, i); |
7d7e0f90 | 2221 | hctx = blk_mq_map_queue(q, i); |
868f2f0b | 2222 | |
e4043dcf | 2223 | cpumask_set_cpu(i, hctx->cpumask); |
320ae51f JA |
2224 | ctx->index_hw = hctx->nr_ctx; |
2225 | hctx->ctxs[hctx->nr_ctx++] = ctx; | |
2226 | } | |
506e931f | 2227 | |
60de074b AM |
2228 | mutex_unlock(&q->sysfs_lock); |
2229 | ||
506e931f | 2230 | queue_for_each_hw_ctx(q, hctx, i) { |
484b4061 | 2231 | /* |
a68aafa5 JA |
2232 | * If no software queues are mapped to this hardware queue, |
2233 | * disable it and free the request entries. | |
484b4061 JA |
2234 | */ |
2235 | if (!hctx->nr_ctx) { | |
d1b1cea1 GKB |
2236 | /* Never unmap queue 0. We need it as a |
2237 | * fallback in case of a new remap fails | |
2238 | * allocation | |
2239 | */ | |
cc71a6f4 JA |
2240 | if (i && set->tags[i]) |
2241 | blk_mq_free_map_and_requests(set, i); | |
2242 | ||
2a34c087 | 2243 | hctx->tags = NULL; |
484b4061 JA |
2244 | continue; |
2245 | } | |
2246 | ||
2a34c087 ML |
2247 | hctx->tags = set->tags[i]; |
2248 | WARN_ON(!hctx->tags); | |
2249 | ||
889fa31f CY |
2250 | /* |
2251 | * Set the map size to the number of mapped software queues. | |
2252 | * This is more accurate and more efficient than looping | |
2253 | * over all possibly mapped software queues. | |
2254 | */ | |
88459642 | 2255 | sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx); |
889fa31f | 2256 | |
484b4061 JA |
2257 | /* |
2258 | * Initialize batch roundrobin counts | |
2259 | */ | |
18a15d75 | 2260 | hctx->next_cpu = cpumask_first(hctx->cpumask); |
506e931f JA |
2261 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; |
2262 | } | |
320ae51f JA |
2263 | } |
2264 | ||
8e8320c9 JA |
2265 | /* |
2266 | * Caller needs to ensure that we're either frozen/quiesced, or that | |
2267 | * the queue isn't live yet. | |
2268 | */ | |
2404e607 | 2269 | static void queue_set_hctx_shared(struct request_queue *q, bool shared) |
0d2602ca JA |
2270 | { |
2271 | struct blk_mq_hw_ctx *hctx; | |
0d2602ca JA |
2272 | int i; |
2273 | ||
2404e607 | 2274 | queue_for_each_hw_ctx(q, hctx, i) { |
8e8320c9 JA |
2275 | if (shared) { |
2276 | if (test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state)) | |
2277 | atomic_inc(&q->shared_hctx_restart); | |
2404e607 | 2278 | hctx->flags |= BLK_MQ_F_TAG_SHARED; |
8e8320c9 JA |
2279 | } else { |
2280 | if (test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state)) | |
2281 | atomic_dec(&q->shared_hctx_restart); | |
2404e607 | 2282 | hctx->flags &= ~BLK_MQ_F_TAG_SHARED; |
8e8320c9 | 2283 | } |
2404e607 JM |
2284 | } |
2285 | } | |
2286 | ||
8e8320c9 JA |
2287 | static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set, |
2288 | bool shared) | |
2404e607 JM |
2289 | { |
2290 | struct request_queue *q; | |
0d2602ca | 2291 | |
705cda97 BVA |
2292 | lockdep_assert_held(&set->tag_list_lock); |
2293 | ||
0d2602ca JA |
2294 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
2295 | blk_mq_freeze_queue(q); | |
2404e607 | 2296 | queue_set_hctx_shared(q, shared); |
0d2602ca JA |
2297 | blk_mq_unfreeze_queue(q); |
2298 | } | |
2299 | } | |
2300 | ||
2301 | static void blk_mq_del_queue_tag_set(struct request_queue *q) | |
2302 | { | |
2303 | struct blk_mq_tag_set *set = q->tag_set; | |
2304 | ||
0d2602ca | 2305 | mutex_lock(&set->tag_list_lock); |
705cda97 BVA |
2306 | list_del_rcu(&q->tag_set_list); |
2307 | INIT_LIST_HEAD(&q->tag_set_list); | |
2404e607 JM |
2308 | if (list_is_singular(&set->tag_list)) { |
2309 | /* just transitioned to unshared */ | |
2310 | set->flags &= ~BLK_MQ_F_TAG_SHARED; | |
2311 | /* update existing queue */ | |
2312 | blk_mq_update_tag_set_depth(set, false); | |
2313 | } | |
0d2602ca | 2314 | mutex_unlock(&set->tag_list_lock); |
705cda97 BVA |
2315 | |
2316 | synchronize_rcu(); | |
0d2602ca JA |
2317 | } |
2318 | ||
2319 | static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set, | |
2320 | struct request_queue *q) | |
2321 | { | |
2322 | q->tag_set = set; | |
2323 | ||
2324 | mutex_lock(&set->tag_list_lock); | |
2404e607 | 2325 | |
ff821d27 JA |
2326 | /* |
2327 | * Check to see if we're transitioning to shared (from 1 to 2 queues). | |
2328 | */ | |
2329 | if (!list_empty(&set->tag_list) && | |
2330 | !(set->flags & BLK_MQ_F_TAG_SHARED)) { | |
2404e607 JM |
2331 | set->flags |= BLK_MQ_F_TAG_SHARED; |
2332 | /* update existing queue */ | |
2333 | blk_mq_update_tag_set_depth(set, true); | |
2334 | } | |
2335 | if (set->flags & BLK_MQ_F_TAG_SHARED) | |
2336 | queue_set_hctx_shared(q, true); | |
705cda97 | 2337 | list_add_tail_rcu(&q->tag_set_list, &set->tag_list); |
2404e607 | 2338 | |
0d2602ca JA |
2339 | mutex_unlock(&set->tag_list_lock); |
2340 | } | |
2341 | ||
e09aae7e ML |
2342 | /* |
2343 | * It is the actual release handler for mq, but we do it from | |
2344 | * request queue's release handler for avoiding use-after-free | |
2345 | * and headache because q->mq_kobj shouldn't have been introduced, | |
2346 | * but we can't group ctx/kctx kobj without it. | |
2347 | */ | |
2348 | void blk_mq_release(struct request_queue *q) | |
2349 | { | |
2350 | struct blk_mq_hw_ctx *hctx; | |
2351 | unsigned int i; | |
2352 | ||
2353 | /* hctx kobj stays in hctx */ | |
c3b4afca ML |
2354 | queue_for_each_hw_ctx(q, hctx, i) { |
2355 | if (!hctx) | |
2356 | continue; | |
6c8b232e | 2357 | kobject_put(&hctx->kobj); |
c3b4afca | 2358 | } |
e09aae7e | 2359 | |
a723bab3 AM |
2360 | q->mq_map = NULL; |
2361 | ||
e09aae7e ML |
2362 | kfree(q->queue_hw_ctx); |
2363 | ||
7ea5fe31 ML |
2364 | /* |
2365 | * release .mq_kobj and sw queue's kobject now because | |
2366 | * both share lifetime with request queue. | |
2367 | */ | |
2368 | blk_mq_sysfs_deinit(q); | |
2369 | ||
e09aae7e ML |
2370 | free_percpu(q->queue_ctx); |
2371 | } | |
2372 | ||
24d2f903 | 2373 | struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set) |
b62c21b7 MS |
2374 | { |
2375 | struct request_queue *uninit_q, *q; | |
2376 | ||
2377 | uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node); | |
2378 | if (!uninit_q) | |
2379 | return ERR_PTR(-ENOMEM); | |
2380 | ||
2381 | q = blk_mq_init_allocated_queue(set, uninit_q); | |
2382 | if (IS_ERR(q)) | |
2383 | blk_cleanup_queue(uninit_q); | |
2384 | ||
2385 | return q; | |
2386 | } | |
2387 | EXPORT_SYMBOL(blk_mq_init_queue); | |
2388 | ||
07319678 BVA |
2389 | static int blk_mq_hw_ctx_size(struct blk_mq_tag_set *tag_set) |
2390 | { | |
2391 | int hw_ctx_size = sizeof(struct blk_mq_hw_ctx); | |
2392 | ||
2393 | BUILD_BUG_ON(ALIGN(offsetof(struct blk_mq_hw_ctx, queue_rq_srcu), | |
2394 | __alignof__(struct blk_mq_hw_ctx)) != | |
2395 | sizeof(struct blk_mq_hw_ctx)); | |
2396 | ||
2397 | if (tag_set->flags & BLK_MQ_F_BLOCKING) | |
2398 | hw_ctx_size += sizeof(struct srcu_struct); | |
2399 | ||
2400 | return hw_ctx_size; | |
2401 | } | |
2402 | ||
868f2f0b KB |
2403 | static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set, |
2404 | struct request_queue *q) | |
320ae51f | 2405 | { |
868f2f0b KB |
2406 | int i, j; |
2407 | struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx; | |
f14bbe77 | 2408 | |
868f2f0b | 2409 | blk_mq_sysfs_unregister(q); |
a7e54b6a ML |
2410 | |
2411 | /* protect against switching io scheduler */ | |
2412 | mutex_lock(&q->sysfs_lock); | |
24d2f903 | 2413 | for (i = 0; i < set->nr_hw_queues; i++) { |
868f2f0b | 2414 | int node; |
f14bbe77 | 2415 | |
868f2f0b KB |
2416 | if (hctxs[i]) |
2417 | continue; | |
2418 | ||
2419 | node = blk_mq_hw_queue_to_node(q->mq_map, i); | |
07319678 | 2420 | hctxs[i] = kzalloc_node(blk_mq_hw_ctx_size(set), |
cdef54dd | 2421 | GFP_KERNEL, node); |
320ae51f | 2422 | if (!hctxs[i]) |
868f2f0b | 2423 | break; |
320ae51f | 2424 | |
a86073e4 | 2425 | if (!zalloc_cpumask_var_node(&hctxs[i]->cpumask, GFP_KERNEL, |
868f2f0b KB |
2426 | node)) { |
2427 | kfree(hctxs[i]); | |
2428 | hctxs[i] = NULL; | |
2429 | break; | |
2430 | } | |
e4043dcf | 2431 | |
0d2602ca | 2432 | atomic_set(&hctxs[i]->nr_active, 0); |
f14bbe77 | 2433 | hctxs[i]->numa_node = node; |
320ae51f | 2434 | hctxs[i]->queue_num = i; |
868f2f0b KB |
2435 | |
2436 | if (blk_mq_init_hctx(q, set, hctxs[i], i)) { | |
2437 | free_cpumask_var(hctxs[i]->cpumask); | |
2438 | kfree(hctxs[i]); | |
2439 | hctxs[i] = NULL; | |
2440 | break; | |
2441 | } | |
2442 | blk_mq_hctx_kobj_init(hctxs[i]); | |
320ae51f | 2443 | } |
868f2f0b KB |
2444 | for (j = i; j < q->nr_hw_queues; j++) { |
2445 | struct blk_mq_hw_ctx *hctx = hctxs[j]; | |
2446 | ||
2447 | if (hctx) { | |
cc71a6f4 JA |
2448 | if (hctx->tags) |
2449 | blk_mq_free_map_and_requests(set, j); | |
868f2f0b | 2450 | blk_mq_exit_hctx(q, set, hctx, j); |
868f2f0b | 2451 | kobject_put(&hctx->kobj); |
868f2f0b KB |
2452 | hctxs[j] = NULL; |
2453 | ||
2454 | } | |
2455 | } | |
2456 | q->nr_hw_queues = i; | |
a7e54b6a | 2457 | mutex_unlock(&q->sysfs_lock); |
868f2f0b KB |
2458 | blk_mq_sysfs_register(q); |
2459 | } | |
2460 | ||
2461 | struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set, | |
2462 | struct request_queue *q) | |
2463 | { | |
66841672 ML |
2464 | /* mark the queue as mq asap */ |
2465 | q->mq_ops = set->ops; | |
2466 | ||
34dbad5d | 2467 | q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn, |
720b8ccc SB |
2468 | blk_mq_poll_stats_bkt, |
2469 | BLK_MQ_POLL_STATS_BKTS, q); | |
34dbad5d OS |
2470 | if (!q->poll_cb) |
2471 | goto err_exit; | |
2472 | ||
868f2f0b KB |
2473 | q->queue_ctx = alloc_percpu(struct blk_mq_ctx); |
2474 | if (!q->queue_ctx) | |
c7de5726 | 2475 | goto err_exit; |
868f2f0b | 2476 | |
737f98cf ML |
2477 | /* init q->mq_kobj and sw queues' kobjects */ |
2478 | blk_mq_sysfs_init(q); | |
2479 | ||
868f2f0b KB |
2480 | q->queue_hw_ctx = kzalloc_node(nr_cpu_ids * sizeof(*(q->queue_hw_ctx)), |
2481 | GFP_KERNEL, set->numa_node); | |
2482 | if (!q->queue_hw_ctx) | |
2483 | goto err_percpu; | |
2484 | ||
bdd17e75 | 2485 | q->mq_map = set->mq_map; |
868f2f0b KB |
2486 | |
2487 | blk_mq_realloc_hw_ctxs(set, q); | |
2488 | if (!q->nr_hw_queues) | |
2489 | goto err_hctxs; | |
320ae51f | 2490 | |
287922eb | 2491 | INIT_WORK(&q->timeout_work, blk_mq_timeout_work); |
e56f698b | 2492 | blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ); |
320ae51f JA |
2493 | |
2494 | q->nr_queues = nr_cpu_ids; | |
320ae51f | 2495 | |
94eddfbe | 2496 | q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT; |
320ae51f | 2497 | |
05f1dd53 JA |
2498 | if (!(set->flags & BLK_MQ_F_SG_MERGE)) |
2499 | q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE; | |
2500 | ||
1be036e9 CH |
2501 | q->sg_reserved_size = INT_MAX; |
2502 | ||
2849450a | 2503 | INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work); |
6fca6a61 CH |
2504 | INIT_LIST_HEAD(&q->requeue_list); |
2505 | spin_lock_init(&q->requeue_lock); | |
2506 | ||
254d259d | 2507 | blk_queue_make_request(q, blk_mq_make_request); |
ea435e1b CH |
2508 | if (q->mq_ops->poll) |
2509 | q->poll_fn = blk_mq_poll; | |
07068d5b | 2510 | |
eba71768 JA |
2511 | /* |
2512 | * Do this after blk_queue_make_request() overrides it... | |
2513 | */ | |
2514 | q->nr_requests = set->queue_depth; | |
2515 | ||
64f1c21e JA |
2516 | /* |
2517 | * Default to classic polling | |
2518 | */ | |
2519 | q->poll_nsec = -1; | |
2520 | ||
24d2f903 CH |
2521 | if (set->ops->complete) |
2522 | blk_queue_softirq_done(q, set->ops->complete); | |
30a91cb4 | 2523 | |
24d2f903 | 2524 | blk_mq_init_cpu_queues(q, set->nr_hw_queues); |
0d2602ca | 2525 | blk_mq_add_queue_tag_set(set, q); |
4b855ad3 | 2526 | blk_mq_map_swqueue(q); |
4593fdbe | 2527 | |
d3484991 JA |
2528 | if (!(set->flags & BLK_MQ_F_NO_SCHED)) { |
2529 | int ret; | |
2530 | ||
2531 | ret = blk_mq_sched_init(q); | |
2532 | if (ret) | |
2533 | return ERR_PTR(ret); | |
2534 | } | |
2535 | ||
320ae51f | 2536 | return q; |
18741986 | 2537 | |
320ae51f | 2538 | err_hctxs: |
868f2f0b | 2539 | kfree(q->queue_hw_ctx); |
320ae51f | 2540 | err_percpu: |
868f2f0b | 2541 | free_percpu(q->queue_ctx); |
c7de5726 ML |
2542 | err_exit: |
2543 | q->mq_ops = NULL; | |
320ae51f JA |
2544 | return ERR_PTR(-ENOMEM); |
2545 | } | |
b62c21b7 | 2546 | EXPORT_SYMBOL(blk_mq_init_allocated_queue); |
320ae51f JA |
2547 | |
2548 | void blk_mq_free_queue(struct request_queue *q) | |
2549 | { | |
624dbe47 | 2550 | struct blk_mq_tag_set *set = q->tag_set; |
320ae51f | 2551 | |
0d2602ca | 2552 | blk_mq_del_queue_tag_set(q); |
624dbe47 | 2553 | blk_mq_exit_hw_queues(q, set, set->nr_hw_queues); |
320ae51f | 2554 | } |
320ae51f JA |
2555 | |
2556 | /* Basically redo blk_mq_init_queue with queue frozen */ | |
4b855ad3 | 2557 | static void blk_mq_queue_reinit(struct request_queue *q) |
320ae51f | 2558 | { |
4ecd4fef | 2559 | WARN_ON_ONCE(!atomic_read(&q->mq_freeze_depth)); |
320ae51f | 2560 | |
9c1051aa | 2561 | blk_mq_debugfs_unregister_hctxs(q); |
67aec14c JA |
2562 | blk_mq_sysfs_unregister(q); |
2563 | ||
320ae51f JA |
2564 | /* |
2565 | * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe | |
ff821d27 JA |
2566 | * we should change hctx numa_node according to the new topology (this |
2567 | * involves freeing and re-allocating memory, worth doing?) | |
320ae51f | 2568 | */ |
4b855ad3 | 2569 | blk_mq_map_swqueue(q); |
320ae51f | 2570 | |
67aec14c | 2571 | blk_mq_sysfs_register(q); |
9c1051aa | 2572 | blk_mq_debugfs_register_hctxs(q); |
320ae51f JA |
2573 | } |
2574 | ||
a5164405 JA |
2575 | static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set) |
2576 | { | |
2577 | int i; | |
2578 | ||
cc71a6f4 JA |
2579 | for (i = 0; i < set->nr_hw_queues; i++) |
2580 | if (!__blk_mq_alloc_rq_map(set, i)) | |
a5164405 | 2581 | goto out_unwind; |
a5164405 JA |
2582 | |
2583 | return 0; | |
2584 | ||
2585 | out_unwind: | |
2586 | while (--i >= 0) | |
cc71a6f4 | 2587 | blk_mq_free_rq_map(set->tags[i]); |
a5164405 | 2588 | |
a5164405 JA |
2589 | return -ENOMEM; |
2590 | } | |
2591 | ||
2592 | /* | |
2593 | * Allocate the request maps associated with this tag_set. Note that this | |
2594 | * may reduce the depth asked for, if memory is tight. set->queue_depth | |
2595 | * will be updated to reflect the allocated depth. | |
2596 | */ | |
2597 | static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set) | |
2598 | { | |
2599 | unsigned int depth; | |
2600 | int err; | |
2601 | ||
2602 | depth = set->queue_depth; | |
2603 | do { | |
2604 | err = __blk_mq_alloc_rq_maps(set); | |
2605 | if (!err) | |
2606 | break; | |
2607 | ||
2608 | set->queue_depth >>= 1; | |
2609 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) { | |
2610 | err = -ENOMEM; | |
2611 | break; | |
2612 | } | |
2613 | } while (set->queue_depth); | |
2614 | ||
2615 | if (!set->queue_depth || err) { | |
2616 | pr_err("blk-mq: failed to allocate request map\n"); | |
2617 | return -ENOMEM; | |
2618 | } | |
2619 | ||
2620 | if (depth != set->queue_depth) | |
2621 | pr_info("blk-mq: reduced tag depth (%u -> %u)\n", | |
2622 | depth, set->queue_depth); | |
2623 | ||
2624 | return 0; | |
2625 | } | |
2626 | ||
ebe8bddb OS |
2627 | static int blk_mq_update_queue_map(struct blk_mq_tag_set *set) |
2628 | { | |
6a684e3d ML |
2629 | if (set->ops->map_queues) { |
2630 | int cpu; | |
2631 | /* | |
2632 | * transport .map_queues is usually done in the following | |
2633 | * way: | |
2634 | * | |
2635 | * for (queue = 0; queue < set->nr_hw_queues; queue++) { | |
2636 | * mask = get_cpu_mask(queue) | |
2637 | * for_each_cpu(cpu, mask) | |
2638 | * set->mq_map[cpu] = queue; | |
2639 | * } | |
2640 | * | |
2641 | * When we need to remap, the table has to be cleared for | |
2642 | * killing stale mapping since one CPU may not be mapped | |
2643 | * to any hw queue. | |
2644 | */ | |
2645 | for_each_possible_cpu(cpu) | |
2646 | set->mq_map[cpu] = 0; | |
2647 | ||
ebe8bddb | 2648 | return set->ops->map_queues(set); |
6a684e3d | 2649 | } else |
ebe8bddb OS |
2650 | return blk_mq_map_queues(set); |
2651 | } | |
2652 | ||
a4391c64 JA |
2653 | /* |
2654 | * Alloc a tag set to be associated with one or more request queues. | |
2655 | * May fail with EINVAL for various error conditions. May adjust the | |
2656 | * requested depth down, if if it too large. In that case, the set | |
2657 | * value will be stored in set->queue_depth. | |
2658 | */ | |
24d2f903 CH |
2659 | int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set) |
2660 | { | |
da695ba2 CH |
2661 | int ret; |
2662 | ||
205fb5f5 BVA |
2663 | BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS); |
2664 | ||
24d2f903 CH |
2665 | if (!set->nr_hw_queues) |
2666 | return -EINVAL; | |
a4391c64 | 2667 | if (!set->queue_depth) |
24d2f903 CH |
2668 | return -EINVAL; |
2669 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) | |
2670 | return -EINVAL; | |
2671 | ||
7d7e0f90 | 2672 | if (!set->ops->queue_rq) |
24d2f903 CH |
2673 | return -EINVAL; |
2674 | ||
de148297 ML |
2675 | if (!set->ops->get_budget ^ !set->ops->put_budget) |
2676 | return -EINVAL; | |
2677 | ||
a4391c64 JA |
2678 | if (set->queue_depth > BLK_MQ_MAX_DEPTH) { |
2679 | pr_info("blk-mq: reduced tag depth to %u\n", | |
2680 | BLK_MQ_MAX_DEPTH); | |
2681 | set->queue_depth = BLK_MQ_MAX_DEPTH; | |
2682 | } | |
24d2f903 | 2683 | |
6637fadf SL |
2684 | /* |
2685 | * If a crashdump is active, then we are potentially in a very | |
2686 | * memory constrained environment. Limit us to 1 queue and | |
2687 | * 64 tags to prevent using too much memory. | |
2688 | */ | |
2689 | if (is_kdump_kernel()) { | |
2690 | set->nr_hw_queues = 1; | |
2691 | set->queue_depth = min(64U, set->queue_depth); | |
2692 | } | |
868f2f0b KB |
2693 | /* |
2694 | * There is no use for more h/w queues than cpus. | |
2695 | */ | |
2696 | if (set->nr_hw_queues > nr_cpu_ids) | |
2697 | set->nr_hw_queues = nr_cpu_ids; | |
6637fadf | 2698 | |
868f2f0b | 2699 | set->tags = kzalloc_node(nr_cpu_ids * sizeof(struct blk_mq_tags *), |
24d2f903 CH |
2700 | GFP_KERNEL, set->numa_node); |
2701 | if (!set->tags) | |
a5164405 | 2702 | return -ENOMEM; |
24d2f903 | 2703 | |
da695ba2 CH |
2704 | ret = -ENOMEM; |
2705 | set->mq_map = kzalloc_node(sizeof(*set->mq_map) * nr_cpu_ids, | |
2706 | GFP_KERNEL, set->numa_node); | |
bdd17e75 CH |
2707 | if (!set->mq_map) |
2708 | goto out_free_tags; | |
2709 | ||
ebe8bddb | 2710 | ret = blk_mq_update_queue_map(set); |
da695ba2 CH |
2711 | if (ret) |
2712 | goto out_free_mq_map; | |
2713 | ||
2714 | ret = blk_mq_alloc_rq_maps(set); | |
2715 | if (ret) | |
bdd17e75 | 2716 | goto out_free_mq_map; |
24d2f903 | 2717 | |
0d2602ca JA |
2718 | mutex_init(&set->tag_list_lock); |
2719 | INIT_LIST_HEAD(&set->tag_list); | |
2720 | ||
24d2f903 | 2721 | return 0; |
bdd17e75 CH |
2722 | |
2723 | out_free_mq_map: | |
2724 | kfree(set->mq_map); | |
2725 | set->mq_map = NULL; | |
2726 | out_free_tags: | |
5676e7b6 RE |
2727 | kfree(set->tags); |
2728 | set->tags = NULL; | |
da695ba2 | 2729 | return ret; |
24d2f903 CH |
2730 | } |
2731 | EXPORT_SYMBOL(blk_mq_alloc_tag_set); | |
2732 | ||
2733 | void blk_mq_free_tag_set(struct blk_mq_tag_set *set) | |
2734 | { | |
2735 | int i; | |
2736 | ||
cc71a6f4 JA |
2737 | for (i = 0; i < nr_cpu_ids; i++) |
2738 | blk_mq_free_map_and_requests(set, i); | |
484b4061 | 2739 | |
bdd17e75 CH |
2740 | kfree(set->mq_map); |
2741 | set->mq_map = NULL; | |
2742 | ||
981bd189 | 2743 | kfree(set->tags); |
5676e7b6 | 2744 | set->tags = NULL; |
24d2f903 CH |
2745 | } |
2746 | EXPORT_SYMBOL(blk_mq_free_tag_set); | |
2747 | ||
e3a2b3f9 JA |
2748 | int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr) |
2749 | { | |
2750 | struct blk_mq_tag_set *set = q->tag_set; | |
2751 | struct blk_mq_hw_ctx *hctx; | |
2752 | int i, ret; | |
2753 | ||
bd166ef1 | 2754 | if (!set) |
e3a2b3f9 JA |
2755 | return -EINVAL; |
2756 | ||
70f36b60 | 2757 | blk_mq_freeze_queue(q); |
70f36b60 | 2758 | |
e3a2b3f9 JA |
2759 | ret = 0; |
2760 | queue_for_each_hw_ctx(q, hctx, i) { | |
e9137d4b KB |
2761 | if (!hctx->tags) |
2762 | continue; | |
bd166ef1 JA |
2763 | /* |
2764 | * If we're using an MQ scheduler, just update the scheduler | |
2765 | * queue depth. This is similar to what the old code would do. | |
2766 | */ | |
70f36b60 | 2767 | if (!hctx->sched_tags) { |
c2e82a23 | 2768 | ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr, |
70f36b60 JA |
2769 | false); |
2770 | } else { | |
2771 | ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags, | |
2772 | nr, true); | |
2773 | } | |
e3a2b3f9 JA |
2774 | if (ret) |
2775 | break; | |
2776 | } | |
2777 | ||
2778 | if (!ret) | |
2779 | q->nr_requests = nr; | |
2780 | ||
70f36b60 | 2781 | blk_mq_unfreeze_queue(q); |
70f36b60 | 2782 | |
e3a2b3f9 JA |
2783 | return ret; |
2784 | } | |
2785 | ||
e4dc2b32 KB |
2786 | static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, |
2787 | int nr_hw_queues) | |
868f2f0b KB |
2788 | { |
2789 | struct request_queue *q; | |
2790 | ||
705cda97 BVA |
2791 | lockdep_assert_held(&set->tag_list_lock); |
2792 | ||
868f2f0b KB |
2793 | if (nr_hw_queues > nr_cpu_ids) |
2794 | nr_hw_queues = nr_cpu_ids; | |
2795 | if (nr_hw_queues < 1 || nr_hw_queues == set->nr_hw_queues) | |
2796 | return; | |
2797 | ||
2798 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
2799 | blk_mq_freeze_queue(q); | |
2800 | ||
2801 | set->nr_hw_queues = nr_hw_queues; | |
ebe8bddb | 2802 | blk_mq_update_queue_map(set); |
868f2f0b KB |
2803 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
2804 | blk_mq_realloc_hw_ctxs(set, q); | |
4b855ad3 | 2805 | blk_mq_queue_reinit(q); |
868f2f0b KB |
2806 | } |
2807 | ||
2808 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
2809 | blk_mq_unfreeze_queue(q); | |
2810 | } | |
e4dc2b32 KB |
2811 | |
2812 | void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues) | |
2813 | { | |
2814 | mutex_lock(&set->tag_list_lock); | |
2815 | __blk_mq_update_nr_hw_queues(set, nr_hw_queues); | |
2816 | mutex_unlock(&set->tag_list_lock); | |
2817 | } | |
868f2f0b KB |
2818 | EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues); |
2819 | ||
34dbad5d OS |
2820 | /* Enable polling stats and return whether they were already enabled. */ |
2821 | static bool blk_poll_stats_enable(struct request_queue *q) | |
2822 | { | |
2823 | if (test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) || | |
2824 | test_and_set_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags)) | |
2825 | return true; | |
2826 | blk_stat_add_callback(q, q->poll_cb); | |
2827 | return false; | |
2828 | } | |
2829 | ||
2830 | static void blk_mq_poll_stats_start(struct request_queue *q) | |
2831 | { | |
2832 | /* | |
2833 | * We don't arm the callback if polling stats are not enabled or the | |
2834 | * callback is already active. | |
2835 | */ | |
2836 | if (!test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) || | |
2837 | blk_stat_is_active(q->poll_cb)) | |
2838 | return; | |
2839 | ||
2840 | blk_stat_activate_msecs(q->poll_cb, 100); | |
2841 | } | |
2842 | ||
2843 | static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb) | |
2844 | { | |
2845 | struct request_queue *q = cb->data; | |
720b8ccc | 2846 | int bucket; |
34dbad5d | 2847 | |
720b8ccc SB |
2848 | for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) { |
2849 | if (cb->stat[bucket].nr_samples) | |
2850 | q->poll_stat[bucket] = cb->stat[bucket]; | |
2851 | } | |
34dbad5d OS |
2852 | } |
2853 | ||
64f1c21e JA |
2854 | static unsigned long blk_mq_poll_nsecs(struct request_queue *q, |
2855 | struct blk_mq_hw_ctx *hctx, | |
2856 | struct request *rq) | |
2857 | { | |
64f1c21e | 2858 | unsigned long ret = 0; |
720b8ccc | 2859 | int bucket; |
64f1c21e JA |
2860 | |
2861 | /* | |
2862 | * If stats collection isn't on, don't sleep but turn it on for | |
2863 | * future users | |
2864 | */ | |
34dbad5d | 2865 | if (!blk_poll_stats_enable(q)) |
64f1c21e JA |
2866 | return 0; |
2867 | ||
64f1c21e JA |
2868 | /* |
2869 | * As an optimistic guess, use half of the mean service time | |
2870 | * for this type of request. We can (and should) make this smarter. | |
2871 | * For instance, if the completion latencies are tight, we can | |
2872 | * get closer than just half the mean. This is especially | |
2873 | * important on devices where the completion latencies are longer | |
720b8ccc SB |
2874 | * than ~10 usec. We do use the stats for the relevant IO size |
2875 | * if available which does lead to better estimates. | |
64f1c21e | 2876 | */ |
720b8ccc SB |
2877 | bucket = blk_mq_poll_stats_bkt(rq); |
2878 | if (bucket < 0) | |
2879 | return ret; | |
2880 | ||
2881 | if (q->poll_stat[bucket].nr_samples) | |
2882 | ret = (q->poll_stat[bucket].mean + 1) / 2; | |
64f1c21e JA |
2883 | |
2884 | return ret; | |
2885 | } | |
2886 | ||
06426adf | 2887 | static bool blk_mq_poll_hybrid_sleep(struct request_queue *q, |
64f1c21e | 2888 | struct blk_mq_hw_ctx *hctx, |
06426adf JA |
2889 | struct request *rq) |
2890 | { | |
2891 | struct hrtimer_sleeper hs; | |
2892 | enum hrtimer_mode mode; | |
64f1c21e | 2893 | unsigned int nsecs; |
06426adf JA |
2894 | ktime_t kt; |
2895 | ||
64f1c21e JA |
2896 | if (test_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags)) |
2897 | return false; | |
2898 | ||
2899 | /* | |
2900 | * poll_nsec can be: | |
2901 | * | |
2902 | * -1: don't ever hybrid sleep | |
2903 | * 0: use half of prev avg | |
2904 | * >0: use this specific value | |
2905 | */ | |
2906 | if (q->poll_nsec == -1) | |
2907 | return false; | |
2908 | else if (q->poll_nsec > 0) | |
2909 | nsecs = q->poll_nsec; | |
2910 | else | |
2911 | nsecs = blk_mq_poll_nsecs(q, hctx, rq); | |
2912 | ||
2913 | if (!nsecs) | |
06426adf JA |
2914 | return false; |
2915 | ||
2916 | set_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags); | |
2917 | ||
2918 | /* | |
2919 | * This will be replaced with the stats tracking code, using | |
2920 | * 'avg_completion_time / 2' as the pre-sleep target. | |
2921 | */ | |
8b0e1953 | 2922 | kt = nsecs; |
06426adf JA |
2923 | |
2924 | mode = HRTIMER_MODE_REL; | |
2925 | hrtimer_init_on_stack(&hs.timer, CLOCK_MONOTONIC, mode); | |
2926 | hrtimer_set_expires(&hs.timer, kt); | |
2927 | ||
2928 | hrtimer_init_sleeper(&hs, current); | |
2929 | do { | |
2930 | if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags)) | |
2931 | break; | |
2932 | set_current_state(TASK_UNINTERRUPTIBLE); | |
2933 | hrtimer_start_expires(&hs.timer, mode); | |
2934 | if (hs.task) | |
2935 | io_schedule(); | |
2936 | hrtimer_cancel(&hs.timer); | |
2937 | mode = HRTIMER_MODE_ABS; | |
2938 | } while (hs.task && !signal_pending(current)); | |
2939 | ||
2940 | __set_current_state(TASK_RUNNING); | |
2941 | destroy_hrtimer_on_stack(&hs.timer); | |
2942 | return true; | |
2943 | } | |
2944 | ||
bbd7bb70 JA |
2945 | static bool __blk_mq_poll(struct blk_mq_hw_ctx *hctx, struct request *rq) |
2946 | { | |
2947 | struct request_queue *q = hctx->queue; | |
2948 | long state; | |
2949 | ||
06426adf JA |
2950 | /* |
2951 | * If we sleep, have the caller restart the poll loop to reset | |
2952 | * the state. Like for the other success return cases, the | |
2953 | * caller is responsible for checking if the IO completed. If | |
2954 | * the IO isn't complete, we'll get called again and will go | |
2955 | * straight to the busy poll loop. | |
2956 | */ | |
64f1c21e | 2957 | if (blk_mq_poll_hybrid_sleep(q, hctx, rq)) |
06426adf JA |
2958 | return true; |
2959 | ||
bbd7bb70 JA |
2960 | hctx->poll_considered++; |
2961 | ||
2962 | state = current->state; | |
2963 | while (!need_resched()) { | |
2964 | int ret; | |
2965 | ||
2966 | hctx->poll_invoked++; | |
2967 | ||
2968 | ret = q->mq_ops->poll(hctx, rq->tag); | |
2969 | if (ret > 0) { | |
2970 | hctx->poll_success++; | |
2971 | set_current_state(TASK_RUNNING); | |
2972 | return true; | |
2973 | } | |
2974 | ||
2975 | if (signal_pending_state(state, current)) | |
2976 | set_current_state(TASK_RUNNING); | |
2977 | ||
2978 | if (current->state == TASK_RUNNING) | |
2979 | return true; | |
2980 | if (ret < 0) | |
2981 | break; | |
2982 | cpu_relax(); | |
2983 | } | |
2984 | ||
2985 | return false; | |
2986 | } | |
2987 | ||
ea435e1b | 2988 | static bool blk_mq_poll(struct request_queue *q, blk_qc_t cookie) |
bbd7bb70 JA |
2989 | { |
2990 | struct blk_mq_hw_ctx *hctx; | |
bbd7bb70 JA |
2991 | struct request *rq; |
2992 | ||
ea435e1b | 2993 | if (!test_bit(QUEUE_FLAG_POLL, &q->queue_flags)) |
bbd7bb70 JA |
2994 | return false; |
2995 | ||
bbd7bb70 | 2996 | hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)]; |
bd166ef1 JA |
2997 | if (!blk_qc_t_is_internal(cookie)) |
2998 | rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie)); | |
3a07bb1d | 2999 | else { |
bd166ef1 | 3000 | rq = blk_mq_tag_to_rq(hctx->sched_tags, blk_qc_t_to_tag(cookie)); |
3a07bb1d JA |
3001 | /* |
3002 | * With scheduling, if the request has completed, we'll | |
3003 | * get a NULL return here, as we clear the sched tag when | |
3004 | * that happens. The request still remains valid, like always, | |
3005 | * so we should be safe with just the NULL check. | |
3006 | */ | |
3007 | if (!rq) | |
3008 | return false; | |
3009 | } | |
bbd7bb70 JA |
3010 | |
3011 | return __blk_mq_poll(hctx, rq); | |
3012 | } | |
bbd7bb70 | 3013 | |
320ae51f JA |
3014 | static int __init blk_mq_init(void) |
3015 | { | |
fc13457f JA |
3016 | /* |
3017 | * See comment in block/blk.h rq_atomic_flags enum | |
3018 | */ | |
3019 | BUILD_BUG_ON((REQ_ATOM_STARTED / BITS_PER_BYTE) != | |
3020 | (REQ_ATOM_COMPLETE / BITS_PER_BYTE)); | |
3021 | ||
9467f859 TG |
3022 | cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL, |
3023 | blk_mq_hctx_notify_dead); | |
320ae51f JA |
3024 | return 0; |
3025 | } | |
3026 | subsys_initcall(blk_mq_init); |