Commit | Line | Data |
---|---|---|
75bb4625 JA |
1 | /* |
2 | * Block multiqueue core code | |
3 | * | |
4 | * Copyright (C) 2013-2014 Jens Axboe | |
5 | * Copyright (C) 2013-2014 Christoph Hellwig | |
6 | */ | |
320ae51f JA |
7 | #include <linux/kernel.h> |
8 | #include <linux/module.h> | |
9 | #include <linux/backing-dev.h> | |
10 | #include <linux/bio.h> | |
11 | #include <linux/blkdev.h> | |
f75782e4 | 12 | #include <linux/kmemleak.h> |
320ae51f JA |
13 | #include <linux/mm.h> |
14 | #include <linux/init.h> | |
15 | #include <linux/slab.h> | |
16 | #include <linux/workqueue.h> | |
17 | #include <linux/smp.h> | |
18 | #include <linux/llist.h> | |
19 | #include <linux/list_sort.h> | |
20 | #include <linux/cpu.h> | |
21 | #include <linux/cache.h> | |
22 | #include <linux/sched/sysctl.h> | |
105ab3d8 | 23 | #include <linux/sched/topology.h> |
174cd4b1 | 24 | #include <linux/sched/signal.h> |
320ae51f | 25 | #include <linux/delay.h> |
aedcd72f | 26 | #include <linux/crash_dump.h> |
88c7b2b7 | 27 | #include <linux/prefetch.h> |
320ae51f JA |
28 | |
29 | #include <trace/events/block.h> | |
30 | ||
31 | #include <linux/blk-mq.h> | |
32 | #include "blk.h" | |
33 | #include "blk-mq.h" | |
9c1051aa | 34 | #include "blk-mq-debugfs.h" |
320ae51f | 35 | #include "blk-mq-tag.h" |
cf43e6be | 36 | #include "blk-stat.h" |
87760e5e | 37 | #include "blk-wbt.h" |
bd166ef1 | 38 | #include "blk-mq-sched.h" |
320ae51f | 39 | |
ea435e1b | 40 | static bool blk_mq_poll(struct request_queue *q, blk_qc_t cookie); |
34dbad5d OS |
41 | static void blk_mq_poll_stats_start(struct request_queue *q); |
42 | static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb); | |
43 | ||
720b8ccc SB |
44 | static int blk_mq_poll_stats_bkt(const struct request *rq) |
45 | { | |
46 | int ddir, bytes, bucket; | |
47 | ||
99c749a4 | 48 | ddir = rq_data_dir(rq); |
720b8ccc SB |
49 | bytes = blk_rq_bytes(rq); |
50 | ||
51 | bucket = ddir + 2*(ilog2(bytes) - 9); | |
52 | ||
53 | if (bucket < 0) | |
54 | return -1; | |
55 | else if (bucket >= BLK_MQ_POLL_STATS_BKTS) | |
56 | return ddir + BLK_MQ_POLL_STATS_BKTS - 2; | |
57 | ||
58 | return bucket; | |
59 | } | |
60 | ||
320ae51f JA |
61 | /* |
62 | * Check if any of the ctx's have pending work in this hardware queue | |
63 | */ | |
79f720a7 | 64 | static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx) |
320ae51f | 65 | { |
79f720a7 JA |
66 | return !list_empty_careful(&hctx->dispatch) || |
67 | sbitmap_any_bit_set(&hctx->ctx_map) || | |
bd166ef1 | 68 | blk_mq_sched_has_work(hctx); |
1429d7c9 JA |
69 | } |
70 | ||
320ae51f JA |
71 | /* |
72 | * Mark this ctx as having pending work in this hardware queue | |
73 | */ | |
74 | static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx, | |
75 | struct blk_mq_ctx *ctx) | |
76 | { | |
88459642 OS |
77 | if (!sbitmap_test_bit(&hctx->ctx_map, ctx->index_hw)) |
78 | sbitmap_set_bit(&hctx->ctx_map, ctx->index_hw); | |
1429d7c9 JA |
79 | } |
80 | ||
81 | static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx, | |
82 | struct blk_mq_ctx *ctx) | |
83 | { | |
88459642 | 84 | sbitmap_clear_bit(&hctx->ctx_map, ctx->index_hw); |
320ae51f JA |
85 | } |
86 | ||
f299b7c7 JA |
87 | struct mq_inflight { |
88 | struct hd_struct *part; | |
89 | unsigned int *inflight; | |
90 | }; | |
91 | ||
92 | static void blk_mq_check_inflight(struct blk_mq_hw_ctx *hctx, | |
93 | struct request *rq, void *priv, | |
94 | bool reserved) | |
95 | { | |
96 | struct mq_inflight *mi = priv; | |
97 | ||
67818d25 | 98 | if (blk_mq_rq_state(rq) == MQ_RQ_IN_FLIGHT) { |
f299b7c7 | 99 | /* |
b8d62b3a JA |
100 | * index[0] counts the specific partition that was asked |
101 | * for. index[1] counts the ones that are active on the | |
102 | * whole device, so increment that if mi->part is indeed | |
103 | * a partition, and not a whole device. | |
f299b7c7 | 104 | */ |
b8d62b3a | 105 | if (rq->part == mi->part) |
f299b7c7 | 106 | mi->inflight[0]++; |
b8d62b3a JA |
107 | if (mi->part->partno) |
108 | mi->inflight[1]++; | |
f299b7c7 JA |
109 | } |
110 | } | |
111 | ||
112 | void blk_mq_in_flight(struct request_queue *q, struct hd_struct *part, | |
113 | unsigned int inflight[2]) | |
114 | { | |
115 | struct mq_inflight mi = { .part = part, .inflight = inflight, }; | |
116 | ||
b8d62b3a | 117 | inflight[0] = inflight[1] = 0; |
f299b7c7 JA |
118 | blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi); |
119 | } | |
120 | ||
1671d522 | 121 | void blk_freeze_queue_start(struct request_queue *q) |
43a5e4e2 | 122 | { |
4ecd4fef | 123 | int freeze_depth; |
cddd5d17 | 124 | |
4ecd4fef CH |
125 | freeze_depth = atomic_inc_return(&q->mq_freeze_depth); |
126 | if (freeze_depth == 1) { | |
3ef28e83 | 127 | percpu_ref_kill(&q->q_usage_counter); |
055f6e18 ML |
128 | if (q->mq_ops) |
129 | blk_mq_run_hw_queues(q, false); | |
cddd5d17 | 130 | } |
f3af020b | 131 | } |
1671d522 | 132 | EXPORT_SYMBOL_GPL(blk_freeze_queue_start); |
f3af020b | 133 | |
6bae363e | 134 | void blk_mq_freeze_queue_wait(struct request_queue *q) |
f3af020b | 135 | { |
3ef28e83 | 136 | wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter)); |
43a5e4e2 | 137 | } |
6bae363e | 138 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait); |
43a5e4e2 | 139 | |
f91328c4 KB |
140 | int blk_mq_freeze_queue_wait_timeout(struct request_queue *q, |
141 | unsigned long timeout) | |
142 | { | |
143 | return wait_event_timeout(q->mq_freeze_wq, | |
144 | percpu_ref_is_zero(&q->q_usage_counter), | |
145 | timeout); | |
146 | } | |
147 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout); | |
43a5e4e2 | 148 | |
f3af020b TH |
149 | /* |
150 | * Guarantee no request is in use, so we can change any data structure of | |
151 | * the queue afterward. | |
152 | */ | |
3ef28e83 | 153 | void blk_freeze_queue(struct request_queue *q) |
f3af020b | 154 | { |
3ef28e83 DW |
155 | /* |
156 | * In the !blk_mq case we are only calling this to kill the | |
157 | * q_usage_counter, otherwise this increases the freeze depth | |
158 | * and waits for it to return to zero. For this reason there is | |
159 | * no blk_unfreeze_queue(), and blk_freeze_queue() is not | |
160 | * exported to drivers as the only user for unfreeze is blk_mq. | |
161 | */ | |
1671d522 | 162 | blk_freeze_queue_start(q); |
f3af020b TH |
163 | blk_mq_freeze_queue_wait(q); |
164 | } | |
3ef28e83 DW |
165 | |
166 | void blk_mq_freeze_queue(struct request_queue *q) | |
167 | { | |
168 | /* | |
169 | * ...just an alias to keep freeze and unfreeze actions balanced | |
170 | * in the blk_mq_* namespace | |
171 | */ | |
172 | blk_freeze_queue(q); | |
173 | } | |
c761d96b | 174 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue); |
f3af020b | 175 | |
b4c6a028 | 176 | void blk_mq_unfreeze_queue(struct request_queue *q) |
320ae51f | 177 | { |
4ecd4fef | 178 | int freeze_depth; |
320ae51f | 179 | |
4ecd4fef CH |
180 | freeze_depth = atomic_dec_return(&q->mq_freeze_depth); |
181 | WARN_ON_ONCE(freeze_depth < 0); | |
182 | if (!freeze_depth) { | |
3ef28e83 | 183 | percpu_ref_reinit(&q->q_usage_counter); |
320ae51f | 184 | wake_up_all(&q->mq_freeze_wq); |
add703fd | 185 | } |
320ae51f | 186 | } |
b4c6a028 | 187 | EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue); |
320ae51f | 188 | |
852ec809 BVA |
189 | /* |
190 | * FIXME: replace the scsi_internal_device_*block_nowait() calls in the | |
191 | * mpt3sas driver such that this function can be removed. | |
192 | */ | |
193 | void blk_mq_quiesce_queue_nowait(struct request_queue *q) | |
194 | { | |
195 | unsigned long flags; | |
196 | ||
197 | spin_lock_irqsave(q->queue_lock, flags); | |
198 | queue_flag_set(QUEUE_FLAG_QUIESCED, q); | |
199 | spin_unlock_irqrestore(q->queue_lock, flags); | |
200 | } | |
201 | EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait); | |
202 | ||
6a83e74d | 203 | /** |
69e07c4a | 204 | * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished |
6a83e74d BVA |
205 | * @q: request queue. |
206 | * | |
207 | * Note: this function does not prevent that the struct request end_io() | |
69e07c4a ML |
208 | * callback function is invoked. Once this function is returned, we make |
209 | * sure no dispatch can happen until the queue is unquiesced via | |
210 | * blk_mq_unquiesce_queue(). | |
6a83e74d BVA |
211 | */ |
212 | void blk_mq_quiesce_queue(struct request_queue *q) | |
213 | { | |
214 | struct blk_mq_hw_ctx *hctx; | |
215 | unsigned int i; | |
216 | bool rcu = false; | |
217 | ||
1d9e9bc6 | 218 | blk_mq_quiesce_queue_nowait(q); |
f4560ffe | 219 | |
6a83e74d BVA |
220 | queue_for_each_hw_ctx(q, hctx, i) { |
221 | if (hctx->flags & BLK_MQ_F_BLOCKING) | |
05707b64 | 222 | synchronize_srcu(hctx->srcu); |
6a83e74d BVA |
223 | else |
224 | rcu = true; | |
225 | } | |
226 | if (rcu) | |
227 | synchronize_rcu(); | |
228 | } | |
229 | EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue); | |
230 | ||
e4e73913 ML |
231 | /* |
232 | * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue() | |
233 | * @q: request queue. | |
234 | * | |
235 | * This function recovers queue into the state before quiescing | |
236 | * which is done by blk_mq_quiesce_queue. | |
237 | */ | |
238 | void blk_mq_unquiesce_queue(struct request_queue *q) | |
239 | { | |
852ec809 BVA |
240 | unsigned long flags; |
241 | ||
242 | spin_lock_irqsave(q->queue_lock, flags); | |
f4560ffe | 243 | queue_flag_clear(QUEUE_FLAG_QUIESCED, q); |
852ec809 | 244 | spin_unlock_irqrestore(q->queue_lock, flags); |
f4560ffe | 245 | |
1d9e9bc6 ML |
246 | /* dispatch requests which are inserted during quiescing */ |
247 | blk_mq_run_hw_queues(q, true); | |
e4e73913 ML |
248 | } |
249 | EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue); | |
250 | ||
aed3ea94 JA |
251 | void blk_mq_wake_waiters(struct request_queue *q) |
252 | { | |
253 | struct blk_mq_hw_ctx *hctx; | |
254 | unsigned int i; | |
255 | ||
256 | queue_for_each_hw_ctx(q, hctx, i) | |
257 | if (blk_mq_hw_queue_mapped(hctx)) | |
258 | blk_mq_tag_wakeup_all(hctx->tags, true); | |
259 | } | |
260 | ||
320ae51f JA |
261 | bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx) |
262 | { | |
263 | return blk_mq_has_free_tags(hctx->tags); | |
264 | } | |
265 | EXPORT_SYMBOL(blk_mq_can_queue); | |
266 | ||
e4cdf1a1 CH |
267 | static struct request *blk_mq_rq_ctx_init(struct blk_mq_alloc_data *data, |
268 | unsigned int tag, unsigned int op) | |
320ae51f | 269 | { |
e4cdf1a1 CH |
270 | struct blk_mq_tags *tags = blk_mq_tags_from_data(data); |
271 | struct request *rq = tags->static_rqs[tag]; | |
bf9ae8c5 | 272 | req_flags_t rq_flags = 0; |
e4cdf1a1 CH |
273 | |
274 | if (data->flags & BLK_MQ_REQ_INTERNAL) { | |
275 | rq->tag = -1; | |
276 | rq->internal_tag = tag; | |
277 | } else { | |
278 | if (blk_mq_tag_busy(data->hctx)) { | |
bf9ae8c5 | 279 | rq_flags = RQF_MQ_INFLIGHT; |
e4cdf1a1 CH |
280 | atomic_inc(&data->hctx->nr_active); |
281 | } | |
282 | rq->tag = tag; | |
283 | rq->internal_tag = -1; | |
284 | data->hctx->tags->rqs[rq->tag] = rq; | |
285 | } | |
286 | ||
af76e555 | 287 | /* csd/requeue_work/fifo_time is initialized before use */ |
e4cdf1a1 CH |
288 | rq->q = data->q; |
289 | rq->mq_ctx = data->ctx; | |
bf9ae8c5 | 290 | rq->rq_flags = rq_flags; |
7c3fb70f | 291 | rq->cpu = -1; |
ef295ecf | 292 | rq->cmd_flags = op; |
1b6d65a0 BVA |
293 | if (data->flags & BLK_MQ_REQ_PREEMPT) |
294 | rq->rq_flags |= RQF_PREEMPT; | |
e4cdf1a1 | 295 | if (blk_queue_io_stat(data->q)) |
e8064021 | 296 | rq->rq_flags |= RQF_IO_STAT; |
7c3fb70f | 297 | INIT_LIST_HEAD(&rq->queuelist); |
af76e555 CH |
298 | INIT_HLIST_NODE(&rq->hash); |
299 | RB_CLEAR_NODE(&rq->rb_node); | |
af76e555 CH |
300 | rq->rq_disk = NULL; |
301 | rq->part = NULL; | |
3ee32372 | 302 | rq->start_time = jiffies; |
af76e555 CH |
303 | rq->nr_phys_segments = 0; |
304 | #if defined(CONFIG_BLK_DEV_INTEGRITY) | |
305 | rq->nr_integrity_segments = 0; | |
306 | #endif | |
af76e555 CH |
307 | rq->special = NULL; |
308 | /* tag was already set */ | |
af76e555 | 309 | rq->extra_len = 0; |
e14575b3 | 310 | rq->__deadline = 0; |
af76e555 | 311 | |
af76e555 | 312 | INIT_LIST_HEAD(&rq->timeout_list); |
f6be4fb4 JA |
313 | rq->timeout = 0; |
314 | ||
af76e555 CH |
315 | rq->end_io = NULL; |
316 | rq->end_io_data = NULL; | |
317 | rq->next_rq = NULL; | |
318 | ||
7c3fb70f JA |
319 | #ifdef CONFIG_BLK_CGROUP |
320 | rq->rl = NULL; | |
321 | set_start_time_ns(rq); | |
322 | rq->io_start_time_ns = 0; | |
323 | #endif | |
324 | ||
e4cdf1a1 CH |
325 | data->ctx->rq_dispatched[op_is_sync(op)]++; |
326 | return rq; | |
5dee8577 CH |
327 | } |
328 | ||
d2c0d383 CH |
329 | static struct request *blk_mq_get_request(struct request_queue *q, |
330 | struct bio *bio, unsigned int op, | |
331 | struct blk_mq_alloc_data *data) | |
332 | { | |
333 | struct elevator_queue *e = q->elevator; | |
334 | struct request *rq; | |
e4cdf1a1 | 335 | unsigned int tag; |
21e768b4 | 336 | bool put_ctx_on_error = false; |
d2c0d383 CH |
337 | |
338 | blk_queue_enter_live(q); | |
339 | data->q = q; | |
21e768b4 BVA |
340 | if (likely(!data->ctx)) { |
341 | data->ctx = blk_mq_get_ctx(q); | |
342 | put_ctx_on_error = true; | |
343 | } | |
d2c0d383 CH |
344 | if (likely(!data->hctx)) |
345 | data->hctx = blk_mq_map_queue(q, data->ctx->cpu); | |
03a07c92 GR |
346 | if (op & REQ_NOWAIT) |
347 | data->flags |= BLK_MQ_REQ_NOWAIT; | |
d2c0d383 CH |
348 | |
349 | if (e) { | |
350 | data->flags |= BLK_MQ_REQ_INTERNAL; | |
351 | ||
352 | /* | |
353 | * Flush requests are special and go directly to the | |
354 | * dispatch list. | |
355 | */ | |
5bbf4e5a CH |
356 | if (!op_is_flush(op) && e->type->ops.mq.limit_depth) |
357 | e->type->ops.mq.limit_depth(op, data); | |
d2c0d383 CH |
358 | } |
359 | ||
e4cdf1a1 CH |
360 | tag = blk_mq_get_tag(data); |
361 | if (tag == BLK_MQ_TAG_FAIL) { | |
21e768b4 BVA |
362 | if (put_ctx_on_error) { |
363 | blk_mq_put_ctx(data->ctx); | |
1ad43c00 ML |
364 | data->ctx = NULL; |
365 | } | |
037cebb8 CH |
366 | blk_queue_exit(q); |
367 | return NULL; | |
d2c0d383 CH |
368 | } |
369 | ||
e4cdf1a1 | 370 | rq = blk_mq_rq_ctx_init(data, tag, op); |
037cebb8 CH |
371 | if (!op_is_flush(op)) { |
372 | rq->elv.icq = NULL; | |
5bbf4e5a | 373 | if (e && e->type->ops.mq.prepare_request) { |
44e8c2bf CH |
374 | if (e->type->icq_cache && rq_ioc(bio)) |
375 | blk_mq_sched_assign_ioc(rq, bio); | |
376 | ||
5bbf4e5a CH |
377 | e->type->ops.mq.prepare_request(rq, bio); |
378 | rq->rq_flags |= RQF_ELVPRIV; | |
44e8c2bf | 379 | } |
037cebb8 CH |
380 | } |
381 | data->hctx->queued++; | |
382 | return rq; | |
d2c0d383 CH |
383 | } |
384 | ||
cd6ce148 | 385 | struct request *blk_mq_alloc_request(struct request_queue *q, unsigned int op, |
9a95e4ef | 386 | blk_mq_req_flags_t flags) |
320ae51f | 387 | { |
5a797e00 | 388 | struct blk_mq_alloc_data alloc_data = { .flags = flags }; |
bd166ef1 | 389 | struct request *rq; |
a492f075 | 390 | int ret; |
320ae51f | 391 | |
3a0a5299 | 392 | ret = blk_queue_enter(q, flags); |
a492f075 JL |
393 | if (ret) |
394 | return ERR_PTR(ret); | |
320ae51f | 395 | |
cd6ce148 | 396 | rq = blk_mq_get_request(q, NULL, op, &alloc_data); |
3280d66a | 397 | blk_queue_exit(q); |
841bac2c | 398 | |
bd166ef1 | 399 | if (!rq) |
a492f075 | 400 | return ERR_PTR(-EWOULDBLOCK); |
0c4de0f3 | 401 | |
1ad43c00 | 402 | blk_mq_put_ctx(alloc_data.ctx); |
1ad43c00 | 403 | |
0c4de0f3 CH |
404 | rq->__data_len = 0; |
405 | rq->__sector = (sector_t) -1; | |
406 | rq->bio = rq->biotail = NULL; | |
320ae51f JA |
407 | return rq; |
408 | } | |
4bb659b1 | 409 | EXPORT_SYMBOL(blk_mq_alloc_request); |
320ae51f | 410 | |
cd6ce148 | 411 | struct request *blk_mq_alloc_request_hctx(struct request_queue *q, |
9a95e4ef | 412 | unsigned int op, blk_mq_req_flags_t flags, unsigned int hctx_idx) |
1f5bd336 | 413 | { |
6d2809d5 | 414 | struct blk_mq_alloc_data alloc_data = { .flags = flags }; |
1f5bd336 | 415 | struct request *rq; |
6d2809d5 | 416 | unsigned int cpu; |
1f5bd336 ML |
417 | int ret; |
418 | ||
419 | /* | |
420 | * If the tag allocator sleeps we could get an allocation for a | |
421 | * different hardware context. No need to complicate the low level | |
422 | * allocator for this for the rare use case of a command tied to | |
423 | * a specific queue. | |
424 | */ | |
425 | if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT))) | |
426 | return ERR_PTR(-EINVAL); | |
427 | ||
428 | if (hctx_idx >= q->nr_hw_queues) | |
429 | return ERR_PTR(-EIO); | |
430 | ||
3a0a5299 | 431 | ret = blk_queue_enter(q, flags); |
1f5bd336 ML |
432 | if (ret) |
433 | return ERR_PTR(ret); | |
434 | ||
c8712c6a CH |
435 | /* |
436 | * Check if the hardware context is actually mapped to anything. | |
437 | * If not tell the caller that it should skip this queue. | |
438 | */ | |
6d2809d5 OS |
439 | alloc_data.hctx = q->queue_hw_ctx[hctx_idx]; |
440 | if (!blk_mq_hw_queue_mapped(alloc_data.hctx)) { | |
441 | blk_queue_exit(q); | |
442 | return ERR_PTR(-EXDEV); | |
c8712c6a | 443 | } |
20e4d813 | 444 | cpu = cpumask_first_and(alloc_data.hctx->cpumask, cpu_online_mask); |
6d2809d5 | 445 | alloc_data.ctx = __blk_mq_get_ctx(q, cpu); |
1f5bd336 | 446 | |
cd6ce148 | 447 | rq = blk_mq_get_request(q, NULL, op, &alloc_data); |
3280d66a | 448 | blk_queue_exit(q); |
c8712c6a | 449 | |
6d2809d5 OS |
450 | if (!rq) |
451 | return ERR_PTR(-EWOULDBLOCK); | |
452 | ||
453 | return rq; | |
1f5bd336 ML |
454 | } |
455 | EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx); | |
456 | ||
6af54051 | 457 | void blk_mq_free_request(struct request *rq) |
320ae51f | 458 | { |
320ae51f | 459 | struct request_queue *q = rq->q; |
6af54051 CH |
460 | struct elevator_queue *e = q->elevator; |
461 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
462 | struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, ctx->cpu); | |
463 | const int sched_tag = rq->internal_tag; | |
464 | ||
5bbf4e5a | 465 | if (rq->rq_flags & RQF_ELVPRIV) { |
6af54051 CH |
466 | if (e && e->type->ops.mq.finish_request) |
467 | e->type->ops.mq.finish_request(rq); | |
468 | if (rq->elv.icq) { | |
469 | put_io_context(rq->elv.icq->ioc); | |
470 | rq->elv.icq = NULL; | |
471 | } | |
472 | } | |
320ae51f | 473 | |
6af54051 | 474 | ctx->rq_completed[rq_is_sync(rq)]++; |
e8064021 | 475 | if (rq->rq_flags & RQF_MQ_INFLIGHT) |
0d2602ca | 476 | atomic_dec(&hctx->nr_active); |
87760e5e | 477 | |
7beb2f84 JA |
478 | if (unlikely(laptop_mode && !blk_rq_is_passthrough(rq))) |
479 | laptop_io_completion(q->backing_dev_info); | |
480 | ||
87760e5e | 481 | wbt_done(q->rq_wb, &rq->issue_stat); |
0d2602ca | 482 | |
85acb3ba SL |
483 | if (blk_rq_rl(rq)) |
484 | blk_put_rl(blk_rq_rl(rq)); | |
485 | ||
1d9bd516 | 486 | blk_mq_rq_update_state(rq, MQ_RQ_IDLE); |
bd166ef1 JA |
487 | if (rq->tag != -1) |
488 | blk_mq_put_tag(hctx, hctx->tags, ctx, rq->tag); | |
489 | if (sched_tag != -1) | |
c05f8525 | 490 | blk_mq_put_tag(hctx, hctx->sched_tags, ctx, sched_tag); |
6d8c6c0f | 491 | blk_mq_sched_restart(hctx); |
3ef28e83 | 492 | blk_queue_exit(q); |
320ae51f | 493 | } |
1a3b595a | 494 | EXPORT_SYMBOL_GPL(blk_mq_free_request); |
320ae51f | 495 | |
2a842aca | 496 | inline void __blk_mq_end_request(struct request *rq, blk_status_t error) |
320ae51f | 497 | { |
0d11e6ac ML |
498 | blk_account_io_done(rq); |
499 | ||
91b63639 | 500 | if (rq->end_io) { |
87760e5e | 501 | wbt_done(rq->q->rq_wb, &rq->issue_stat); |
320ae51f | 502 | rq->end_io(rq, error); |
91b63639 CH |
503 | } else { |
504 | if (unlikely(blk_bidi_rq(rq))) | |
505 | blk_mq_free_request(rq->next_rq); | |
320ae51f | 506 | blk_mq_free_request(rq); |
91b63639 | 507 | } |
320ae51f | 508 | } |
c8a446ad | 509 | EXPORT_SYMBOL(__blk_mq_end_request); |
63151a44 | 510 | |
2a842aca | 511 | void blk_mq_end_request(struct request *rq, blk_status_t error) |
63151a44 CH |
512 | { |
513 | if (blk_update_request(rq, error, blk_rq_bytes(rq))) | |
514 | BUG(); | |
c8a446ad | 515 | __blk_mq_end_request(rq, error); |
63151a44 | 516 | } |
c8a446ad | 517 | EXPORT_SYMBOL(blk_mq_end_request); |
320ae51f | 518 | |
30a91cb4 | 519 | static void __blk_mq_complete_request_remote(void *data) |
320ae51f | 520 | { |
3d6efbf6 | 521 | struct request *rq = data; |
320ae51f | 522 | |
30a91cb4 | 523 | rq->q->softirq_done_fn(rq); |
320ae51f | 524 | } |
320ae51f | 525 | |
453f8341 | 526 | static void __blk_mq_complete_request(struct request *rq) |
320ae51f JA |
527 | { |
528 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
38535201 | 529 | bool shared = false; |
320ae51f JA |
530 | int cpu; |
531 | ||
1d9bd516 | 532 | WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT); |
5a61c363 | 533 | blk_mq_rq_update_state(rq, MQ_RQ_COMPLETE); |
1d9bd516 | 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 | 561 | |
04ced159 | 562 | static void hctx_unlock(struct blk_mq_hw_ctx *hctx, int srcu_idx) |
b7435db8 | 563 | __releases(hctx->srcu) |
04ced159 JA |
564 | { |
565 | if (!(hctx->flags & BLK_MQ_F_BLOCKING)) | |
566 | rcu_read_unlock(); | |
567 | else | |
05707b64 | 568 | srcu_read_unlock(hctx->srcu, srcu_idx); |
04ced159 JA |
569 | } |
570 | ||
571 | static void hctx_lock(struct blk_mq_hw_ctx *hctx, int *srcu_idx) | |
b7435db8 | 572 | __acquires(hctx->srcu) |
04ced159 | 573 | { |
08b5a6e2 JA |
574 | if (!(hctx->flags & BLK_MQ_F_BLOCKING)) { |
575 | /* shut up gcc false positive */ | |
576 | *srcu_idx = 0; | |
04ced159 | 577 | rcu_read_lock(); |
08b5a6e2 | 578 | } else |
05707b64 | 579 | *srcu_idx = srcu_read_lock(hctx->srcu); |
04ced159 JA |
580 | } |
581 | ||
1d9bd516 TH |
582 | static void blk_mq_rq_update_aborted_gstate(struct request *rq, u64 gstate) |
583 | { | |
584 | unsigned long flags; | |
585 | ||
586 | /* | |
587 | * blk_mq_rq_aborted_gstate() is used from the completion path and | |
588 | * can thus be called from irq context. u64_stats_fetch in the | |
589 | * middle of update on the same CPU leads to lockup. Disable irq | |
590 | * while updating. | |
591 | */ | |
592 | local_irq_save(flags); | |
593 | u64_stats_update_begin(&rq->aborted_gstate_sync); | |
594 | rq->aborted_gstate = gstate; | |
595 | u64_stats_update_end(&rq->aborted_gstate_sync); | |
596 | local_irq_restore(flags); | |
597 | } | |
598 | ||
599 | static u64 blk_mq_rq_aborted_gstate(struct request *rq) | |
600 | { | |
601 | unsigned int start; | |
602 | u64 aborted_gstate; | |
603 | ||
604 | do { | |
605 | start = u64_stats_fetch_begin(&rq->aborted_gstate_sync); | |
606 | aborted_gstate = rq->aborted_gstate; | |
607 | } while (u64_stats_fetch_retry(&rq->aborted_gstate_sync, start)); | |
608 | ||
609 | return aborted_gstate; | |
610 | } | |
611 | ||
30a91cb4 CH |
612 | /** |
613 | * blk_mq_complete_request - end I/O on a request | |
614 | * @rq: the request being processed | |
615 | * | |
616 | * Description: | |
617 | * Ends all I/O on a request. It does not handle partial completions. | |
618 | * The actual completion happens out-of-order, through a IPI handler. | |
619 | **/ | |
08e0029a | 620 | void blk_mq_complete_request(struct request *rq) |
30a91cb4 | 621 | { |
95f09684 | 622 | struct request_queue *q = rq->q; |
5197c05e TH |
623 | struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, rq->mq_ctx->cpu); |
624 | int srcu_idx; | |
95f09684 JA |
625 | |
626 | if (unlikely(blk_should_fake_timeout(q))) | |
30a91cb4 | 627 | return; |
5197c05e | 628 | |
1d9bd516 TH |
629 | /* |
630 | * If @rq->aborted_gstate equals the current instance, timeout is | |
631 | * claiming @rq and we lost. This is synchronized through | |
632 | * hctx_lock(). See blk_mq_timeout_work() for details. | |
633 | * | |
634 | * Completion path never blocks and we can directly use RCU here | |
635 | * instead of hctx_lock() which can be either RCU or SRCU. | |
636 | * However, that would complicate paths which want to synchronize | |
637 | * against us. Let stay in sync with the issue path so that | |
638 | * hctx_lock() covers both issue and completion paths. | |
639 | */ | |
5197c05e | 640 | hctx_lock(hctx, &srcu_idx); |
634f9e46 | 641 | if (blk_mq_rq_aborted_gstate(rq) != rq->gstate) |
ed851860 | 642 | __blk_mq_complete_request(rq); |
5197c05e | 643 | hctx_unlock(hctx, srcu_idx); |
30a91cb4 CH |
644 | } |
645 | EXPORT_SYMBOL(blk_mq_complete_request); | |
320ae51f | 646 | |
973c0191 KB |
647 | int blk_mq_request_started(struct request *rq) |
648 | { | |
5a61c363 | 649 | return blk_mq_rq_state(rq) != MQ_RQ_IDLE; |
973c0191 KB |
650 | } |
651 | EXPORT_SYMBOL_GPL(blk_mq_request_started); | |
652 | ||
e2490073 | 653 | void blk_mq_start_request(struct request *rq) |
320ae51f JA |
654 | { |
655 | struct request_queue *q = rq->q; | |
656 | ||
bd166ef1 JA |
657 | blk_mq_sched_started_request(rq); |
658 | ||
320ae51f JA |
659 | trace_block_rq_issue(q, rq); |
660 | ||
cf43e6be | 661 | if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) { |
88eeca49 | 662 | blk_stat_set_issue(&rq->issue_stat, blk_rq_sectors(rq)); |
cf43e6be | 663 | rq->rq_flags |= RQF_STATS; |
87760e5e | 664 | wbt_issue(q->rq_wb, &rq->issue_stat); |
cf43e6be JA |
665 | } |
666 | ||
1d9bd516 | 667 | WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE); |
538b7534 | 668 | |
87ee7b11 | 669 | /* |
1d9bd516 TH |
670 | * Mark @rq in-flight which also advances the generation number, |
671 | * and register for timeout. Protect with a seqcount to allow the | |
672 | * timeout path to read both @rq->gstate and @rq->deadline | |
673 | * coherently. | |
a7af0af3 | 674 | * |
1d9bd516 TH |
675 | * This is the only place where a request is marked in-flight. If |
676 | * the timeout path reads an in-flight @rq->gstate, the | |
677 | * @rq->deadline it reads together under @rq->gstate_seq is | |
678 | * guaranteed to be the matching one. | |
87ee7b11 | 679 | */ |
1d9bd516 TH |
680 | preempt_disable(); |
681 | write_seqcount_begin(&rq->gstate_seq); | |
682 | ||
683 | blk_mq_rq_update_state(rq, MQ_RQ_IN_FLIGHT); | |
684 | blk_add_timer(rq); | |
685 | ||
686 | write_seqcount_end(&rq->gstate_seq); | |
687 | preempt_enable(); | |
688 | ||
49f5baa5 CH |
689 | if (q->dma_drain_size && blk_rq_bytes(rq)) { |
690 | /* | |
691 | * Make sure space for the drain appears. We know we can do | |
692 | * this because max_hw_segments has been adjusted to be one | |
693 | * fewer than the device can handle. | |
694 | */ | |
695 | rq->nr_phys_segments++; | |
696 | } | |
320ae51f | 697 | } |
e2490073 | 698 | EXPORT_SYMBOL(blk_mq_start_request); |
320ae51f | 699 | |
d9d149a3 | 700 | /* |
5a61c363 TH |
701 | * When we reach here because queue is busy, it's safe to change the state |
702 | * to IDLE without checking @rq->aborted_gstate because we should still be | |
703 | * holding the RCU read lock and thus protected against timeout. | |
d9d149a3 | 704 | */ |
ed0791b2 | 705 | static void __blk_mq_requeue_request(struct request *rq) |
320ae51f JA |
706 | { |
707 | struct request_queue *q = rq->q; | |
708 | ||
923218f6 ML |
709 | blk_mq_put_driver_tag(rq); |
710 | ||
320ae51f | 711 | trace_block_rq_requeue(q, rq); |
87760e5e | 712 | wbt_requeue(q->rq_wb, &rq->issue_stat); |
bd166ef1 | 713 | blk_mq_sched_requeue_request(rq); |
49f5baa5 | 714 | |
5a61c363 | 715 | if (blk_mq_rq_state(rq) != MQ_RQ_IDLE) { |
1d9bd516 | 716 | blk_mq_rq_update_state(rq, MQ_RQ_IDLE); |
e2490073 CH |
717 | if (q->dma_drain_size && blk_rq_bytes(rq)) |
718 | rq->nr_phys_segments--; | |
719 | } | |
320ae51f JA |
720 | } |
721 | ||
2b053aca | 722 | void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list) |
ed0791b2 | 723 | { |
ed0791b2 | 724 | __blk_mq_requeue_request(rq); |
ed0791b2 | 725 | |
ed0791b2 | 726 | BUG_ON(blk_queued_rq(rq)); |
2b053aca | 727 | blk_mq_add_to_requeue_list(rq, true, kick_requeue_list); |
ed0791b2 CH |
728 | } |
729 | EXPORT_SYMBOL(blk_mq_requeue_request); | |
730 | ||
6fca6a61 CH |
731 | static void blk_mq_requeue_work(struct work_struct *work) |
732 | { | |
733 | struct request_queue *q = | |
2849450a | 734 | container_of(work, struct request_queue, requeue_work.work); |
6fca6a61 CH |
735 | LIST_HEAD(rq_list); |
736 | struct request *rq, *next; | |
6fca6a61 | 737 | |
18e9781d | 738 | spin_lock_irq(&q->requeue_lock); |
6fca6a61 | 739 | list_splice_init(&q->requeue_list, &rq_list); |
18e9781d | 740 | spin_unlock_irq(&q->requeue_lock); |
6fca6a61 CH |
741 | |
742 | list_for_each_entry_safe(rq, next, &rq_list, queuelist) { | |
e8064021 | 743 | if (!(rq->rq_flags & RQF_SOFTBARRIER)) |
6fca6a61 CH |
744 | continue; |
745 | ||
e8064021 | 746 | rq->rq_flags &= ~RQF_SOFTBARRIER; |
6fca6a61 | 747 | list_del_init(&rq->queuelist); |
9e97d295 | 748 | blk_mq_sched_insert_request(rq, true, false, false); |
6fca6a61 CH |
749 | } |
750 | ||
751 | while (!list_empty(&rq_list)) { | |
752 | rq = list_entry(rq_list.next, struct request, queuelist); | |
753 | list_del_init(&rq->queuelist); | |
9e97d295 | 754 | blk_mq_sched_insert_request(rq, false, false, false); |
6fca6a61 CH |
755 | } |
756 | ||
52d7f1b5 | 757 | blk_mq_run_hw_queues(q, false); |
6fca6a61 CH |
758 | } |
759 | ||
2b053aca BVA |
760 | void blk_mq_add_to_requeue_list(struct request *rq, bool at_head, |
761 | bool kick_requeue_list) | |
6fca6a61 CH |
762 | { |
763 | struct request_queue *q = rq->q; | |
764 | unsigned long flags; | |
765 | ||
766 | /* | |
767 | * We abuse this flag that is otherwise used by the I/O scheduler to | |
ff821d27 | 768 | * request head insertion from the workqueue. |
6fca6a61 | 769 | */ |
e8064021 | 770 | BUG_ON(rq->rq_flags & RQF_SOFTBARRIER); |
6fca6a61 CH |
771 | |
772 | spin_lock_irqsave(&q->requeue_lock, flags); | |
773 | if (at_head) { | |
e8064021 | 774 | rq->rq_flags |= RQF_SOFTBARRIER; |
6fca6a61 CH |
775 | list_add(&rq->queuelist, &q->requeue_list); |
776 | } else { | |
777 | list_add_tail(&rq->queuelist, &q->requeue_list); | |
778 | } | |
779 | spin_unlock_irqrestore(&q->requeue_lock, flags); | |
2b053aca BVA |
780 | |
781 | if (kick_requeue_list) | |
782 | blk_mq_kick_requeue_list(q); | |
6fca6a61 CH |
783 | } |
784 | EXPORT_SYMBOL(blk_mq_add_to_requeue_list); | |
785 | ||
786 | void blk_mq_kick_requeue_list(struct request_queue *q) | |
787 | { | |
2849450a | 788 | kblockd_schedule_delayed_work(&q->requeue_work, 0); |
6fca6a61 CH |
789 | } |
790 | EXPORT_SYMBOL(blk_mq_kick_requeue_list); | |
791 | ||
2849450a MS |
792 | void blk_mq_delay_kick_requeue_list(struct request_queue *q, |
793 | unsigned long msecs) | |
794 | { | |
d4acf365 BVA |
795 | kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, |
796 | msecs_to_jiffies(msecs)); | |
2849450a MS |
797 | } |
798 | EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list); | |
799 | ||
0e62f51f JA |
800 | struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag) |
801 | { | |
88c7b2b7 JA |
802 | if (tag < tags->nr_tags) { |
803 | prefetch(tags->rqs[tag]); | |
4ee86bab | 804 | return tags->rqs[tag]; |
88c7b2b7 | 805 | } |
4ee86bab HR |
806 | |
807 | return NULL; | |
24d2f903 CH |
808 | } |
809 | EXPORT_SYMBOL(blk_mq_tag_to_rq); | |
810 | ||
320ae51f | 811 | struct blk_mq_timeout_data { |
46f92d42 CH |
812 | unsigned long next; |
813 | unsigned int next_set; | |
1d9bd516 | 814 | unsigned int nr_expired; |
320ae51f JA |
815 | }; |
816 | ||
358f70da | 817 | static void blk_mq_rq_timed_out(struct request *req, bool reserved) |
320ae51f | 818 | { |
f8a5b122 | 819 | const struct blk_mq_ops *ops = req->q->mq_ops; |
46f92d42 | 820 | enum blk_eh_timer_return ret = BLK_EH_RESET_TIMER; |
87ee7b11 | 821 | |
634f9e46 TH |
822 | req->rq_flags |= RQF_MQ_TIMEOUT_EXPIRED; |
823 | ||
46f92d42 | 824 | if (ops->timeout) |
0152fb6b | 825 | ret = ops->timeout(req, reserved); |
46f92d42 CH |
826 | |
827 | switch (ret) { | |
828 | case BLK_EH_HANDLED: | |
829 | __blk_mq_complete_request(req); | |
830 | break; | |
831 | case BLK_EH_RESET_TIMER: | |
1d9bd516 TH |
832 | /* |
833 | * As nothing prevents from completion happening while | |
834 | * ->aborted_gstate is set, this may lead to ignored | |
835 | * completions and further spurious timeouts. | |
836 | */ | |
837 | blk_mq_rq_update_aborted_gstate(req, 0); | |
46f92d42 | 838 | blk_add_timer(req); |
46f92d42 CH |
839 | break; |
840 | case BLK_EH_NOT_HANDLED: | |
841 | break; | |
842 | default: | |
843 | printk(KERN_ERR "block: bad eh return: %d\n", ret); | |
844 | break; | |
845 | } | |
87ee7b11 | 846 | } |
5b3f25fc | 847 | |
81481eb4 CH |
848 | static void blk_mq_check_expired(struct blk_mq_hw_ctx *hctx, |
849 | struct request *rq, void *priv, bool reserved) | |
850 | { | |
851 | struct blk_mq_timeout_data *data = priv; | |
1d9bd516 TH |
852 | unsigned long gstate, deadline; |
853 | int start; | |
854 | ||
855 | might_sleep(); | |
87ee7b11 | 856 | |
5a61c363 | 857 | if (rq->rq_flags & RQF_MQ_TIMEOUT_EXPIRED) |
46f92d42 | 858 | return; |
87ee7b11 | 859 | |
1d9bd516 TH |
860 | /* read coherent snapshots of @rq->state_gen and @rq->deadline */ |
861 | while (true) { | |
862 | start = read_seqcount_begin(&rq->gstate_seq); | |
863 | gstate = READ_ONCE(rq->gstate); | |
0a72e7f4 | 864 | deadline = blk_rq_deadline(rq); |
1d9bd516 TH |
865 | if (!read_seqcount_retry(&rq->gstate_seq, start)) |
866 | break; | |
867 | cond_resched(); | |
868 | } | |
a7af0af3 | 869 | |
1d9bd516 TH |
870 | /* if in-flight && overdue, mark for abortion */ |
871 | if ((gstate & MQ_RQ_STATE_MASK) == MQ_RQ_IN_FLIGHT && | |
872 | time_after_eq(jiffies, deadline)) { | |
873 | blk_mq_rq_update_aborted_gstate(rq, gstate); | |
874 | data->nr_expired++; | |
875 | hctx->nr_expired++; | |
a7af0af3 PZ |
876 | } else if (!data->next_set || time_after(data->next, deadline)) { |
877 | data->next = deadline; | |
46f92d42 CH |
878 | data->next_set = 1; |
879 | } | |
87ee7b11 JA |
880 | } |
881 | ||
1d9bd516 TH |
882 | static void blk_mq_terminate_expired(struct blk_mq_hw_ctx *hctx, |
883 | struct request *rq, void *priv, bool reserved) | |
884 | { | |
885 | /* | |
886 | * We marked @rq->aborted_gstate and waited for RCU. If there were | |
887 | * completions that we lost to, they would have finished and | |
888 | * updated @rq->gstate by now; otherwise, the completion path is | |
889 | * now guaranteed to see @rq->aborted_gstate and yield. If | |
890 | * @rq->aborted_gstate still matches @rq->gstate, @rq is ours. | |
891 | */ | |
634f9e46 TH |
892 | if (!(rq->rq_flags & RQF_MQ_TIMEOUT_EXPIRED) && |
893 | READ_ONCE(rq->gstate) == rq->aborted_gstate) | |
1d9bd516 TH |
894 | blk_mq_rq_timed_out(rq, reserved); |
895 | } | |
896 | ||
287922eb | 897 | static void blk_mq_timeout_work(struct work_struct *work) |
320ae51f | 898 | { |
287922eb CH |
899 | struct request_queue *q = |
900 | container_of(work, struct request_queue, timeout_work); | |
81481eb4 CH |
901 | struct blk_mq_timeout_data data = { |
902 | .next = 0, | |
903 | .next_set = 0, | |
1d9bd516 | 904 | .nr_expired = 0, |
81481eb4 | 905 | }; |
1d9bd516 | 906 | struct blk_mq_hw_ctx *hctx; |
81481eb4 | 907 | int i; |
320ae51f | 908 | |
71f79fb3 GKB |
909 | /* A deadlock might occur if a request is stuck requiring a |
910 | * timeout at the same time a queue freeze is waiting | |
911 | * completion, since the timeout code would not be able to | |
912 | * acquire the queue reference here. | |
913 | * | |
914 | * That's why we don't use blk_queue_enter here; instead, we use | |
915 | * percpu_ref_tryget directly, because we need to be able to | |
916 | * obtain a reference even in the short window between the queue | |
917 | * starting to freeze, by dropping the first reference in | |
1671d522 | 918 | * blk_freeze_queue_start, and the moment the last request is |
71f79fb3 GKB |
919 | * consumed, marked by the instant q_usage_counter reaches |
920 | * zero. | |
921 | */ | |
922 | if (!percpu_ref_tryget(&q->q_usage_counter)) | |
287922eb CH |
923 | return; |
924 | ||
1d9bd516 | 925 | /* scan for the expired ones and set their ->aborted_gstate */ |
0bf6cd5b | 926 | blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &data); |
320ae51f | 927 | |
1d9bd516 TH |
928 | if (data.nr_expired) { |
929 | bool has_rcu = false; | |
930 | ||
931 | /* | |
932 | * Wait till everyone sees ->aborted_gstate. The | |
933 | * sequential waits for SRCUs aren't ideal. If this ever | |
934 | * becomes a problem, we can add per-hw_ctx rcu_head and | |
935 | * wait in parallel. | |
936 | */ | |
937 | queue_for_each_hw_ctx(q, hctx, i) { | |
938 | if (!hctx->nr_expired) | |
939 | continue; | |
940 | ||
941 | if (!(hctx->flags & BLK_MQ_F_BLOCKING)) | |
942 | has_rcu = true; | |
943 | else | |
05707b64 | 944 | synchronize_srcu(hctx->srcu); |
1d9bd516 TH |
945 | |
946 | hctx->nr_expired = 0; | |
947 | } | |
948 | if (has_rcu) | |
949 | synchronize_rcu(); | |
950 | ||
951 | /* terminate the ones we won */ | |
952 | blk_mq_queue_tag_busy_iter(q, blk_mq_terminate_expired, NULL); | |
953 | } | |
954 | ||
81481eb4 CH |
955 | if (data.next_set) { |
956 | data.next = blk_rq_timeout(round_jiffies_up(data.next)); | |
957 | mod_timer(&q->timeout, data.next); | |
0d2602ca | 958 | } else { |
fcd36c36 BVA |
959 | /* |
960 | * Request timeouts are handled as a forward rolling timer. If | |
961 | * we end up here it means that no requests are pending and | |
962 | * also that no request has been pending for a while. Mark | |
963 | * each hctx as idle. | |
964 | */ | |
f054b56c ML |
965 | queue_for_each_hw_ctx(q, hctx, i) { |
966 | /* the hctx may be unmapped, so check it here */ | |
967 | if (blk_mq_hw_queue_mapped(hctx)) | |
968 | blk_mq_tag_idle(hctx); | |
969 | } | |
0d2602ca | 970 | } |
287922eb | 971 | blk_queue_exit(q); |
320ae51f JA |
972 | } |
973 | ||
88459642 OS |
974 | struct flush_busy_ctx_data { |
975 | struct blk_mq_hw_ctx *hctx; | |
976 | struct list_head *list; | |
977 | }; | |
978 | ||
979 | static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data) | |
980 | { | |
981 | struct flush_busy_ctx_data *flush_data = data; | |
982 | struct blk_mq_hw_ctx *hctx = flush_data->hctx; | |
983 | struct blk_mq_ctx *ctx = hctx->ctxs[bitnr]; | |
984 | ||
985 | sbitmap_clear_bit(sb, bitnr); | |
986 | spin_lock(&ctx->lock); | |
987 | list_splice_tail_init(&ctx->rq_list, flush_data->list); | |
988 | spin_unlock(&ctx->lock); | |
989 | return true; | |
990 | } | |
991 | ||
1429d7c9 JA |
992 | /* |
993 | * Process software queues that have been marked busy, splicing them | |
994 | * to the for-dispatch | |
995 | */ | |
2c3ad667 | 996 | void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list) |
1429d7c9 | 997 | { |
88459642 OS |
998 | struct flush_busy_ctx_data data = { |
999 | .hctx = hctx, | |
1000 | .list = list, | |
1001 | }; | |
1429d7c9 | 1002 | |
88459642 | 1003 | sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data); |
1429d7c9 | 1004 | } |
2c3ad667 | 1005 | EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs); |
1429d7c9 | 1006 | |
b347689f ML |
1007 | struct dispatch_rq_data { |
1008 | struct blk_mq_hw_ctx *hctx; | |
1009 | struct request *rq; | |
1010 | }; | |
1011 | ||
1012 | static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr, | |
1013 | void *data) | |
1014 | { | |
1015 | struct dispatch_rq_data *dispatch_data = data; | |
1016 | struct blk_mq_hw_ctx *hctx = dispatch_data->hctx; | |
1017 | struct blk_mq_ctx *ctx = hctx->ctxs[bitnr]; | |
1018 | ||
1019 | spin_lock(&ctx->lock); | |
1020 | if (unlikely(!list_empty(&ctx->rq_list))) { | |
1021 | dispatch_data->rq = list_entry_rq(ctx->rq_list.next); | |
1022 | list_del_init(&dispatch_data->rq->queuelist); | |
1023 | if (list_empty(&ctx->rq_list)) | |
1024 | sbitmap_clear_bit(sb, bitnr); | |
1025 | } | |
1026 | spin_unlock(&ctx->lock); | |
1027 | ||
1028 | return !dispatch_data->rq; | |
1029 | } | |
1030 | ||
1031 | struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx, | |
1032 | struct blk_mq_ctx *start) | |
1033 | { | |
1034 | unsigned off = start ? start->index_hw : 0; | |
1035 | struct dispatch_rq_data data = { | |
1036 | .hctx = hctx, | |
1037 | .rq = NULL, | |
1038 | }; | |
1039 | ||
1040 | __sbitmap_for_each_set(&hctx->ctx_map, off, | |
1041 | dispatch_rq_from_ctx, &data); | |
1042 | ||
1043 | return data.rq; | |
1044 | } | |
1045 | ||
703fd1c0 JA |
1046 | static inline unsigned int queued_to_index(unsigned int queued) |
1047 | { | |
1048 | if (!queued) | |
1049 | return 0; | |
1429d7c9 | 1050 | |
703fd1c0 | 1051 | return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1); |
1429d7c9 JA |
1052 | } |
1053 | ||
bd6737f1 JA |
1054 | bool blk_mq_get_driver_tag(struct request *rq, struct blk_mq_hw_ctx **hctx, |
1055 | bool wait) | |
bd166ef1 JA |
1056 | { |
1057 | struct blk_mq_alloc_data data = { | |
1058 | .q = rq->q, | |
bd166ef1 JA |
1059 | .hctx = blk_mq_map_queue(rq->q, rq->mq_ctx->cpu), |
1060 | .flags = wait ? 0 : BLK_MQ_REQ_NOWAIT, | |
1061 | }; | |
1062 | ||
5feeacdd JA |
1063 | might_sleep_if(wait); |
1064 | ||
81380ca1 OS |
1065 | if (rq->tag != -1) |
1066 | goto done; | |
bd166ef1 | 1067 | |
415b806d SG |
1068 | if (blk_mq_tag_is_reserved(data.hctx->sched_tags, rq->internal_tag)) |
1069 | data.flags |= BLK_MQ_REQ_RESERVED; | |
1070 | ||
bd166ef1 JA |
1071 | rq->tag = blk_mq_get_tag(&data); |
1072 | if (rq->tag >= 0) { | |
200e86b3 JA |
1073 | if (blk_mq_tag_busy(data.hctx)) { |
1074 | rq->rq_flags |= RQF_MQ_INFLIGHT; | |
1075 | atomic_inc(&data.hctx->nr_active); | |
1076 | } | |
bd166ef1 | 1077 | data.hctx->tags->rqs[rq->tag] = rq; |
bd166ef1 JA |
1078 | } |
1079 | ||
81380ca1 OS |
1080 | done: |
1081 | if (hctx) | |
1082 | *hctx = data.hctx; | |
1083 | return rq->tag != -1; | |
bd166ef1 JA |
1084 | } |
1085 | ||
eb619fdb JA |
1086 | static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode, |
1087 | int flags, void *key) | |
da55f2cc OS |
1088 | { |
1089 | struct blk_mq_hw_ctx *hctx; | |
1090 | ||
1091 | hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait); | |
1092 | ||
eb619fdb | 1093 | list_del_init(&wait->entry); |
da55f2cc OS |
1094 | blk_mq_run_hw_queue(hctx, true); |
1095 | return 1; | |
1096 | } | |
1097 | ||
f906a6a0 JA |
1098 | /* |
1099 | * Mark us waiting for a tag. For shared tags, this involves hooking us into | |
ee3e4de5 BVA |
1100 | * the tag wakeups. For non-shared tags, we can simply mark us needing a |
1101 | * restart. For both cases, take care to check the condition again after | |
f906a6a0 JA |
1102 | * marking us as waiting. |
1103 | */ | |
1104 | static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx **hctx, | |
1105 | struct request *rq) | |
da55f2cc | 1106 | { |
eb619fdb | 1107 | struct blk_mq_hw_ctx *this_hctx = *hctx; |
da55f2cc | 1108 | struct sbq_wait_state *ws; |
f906a6a0 JA |
1109 | wait_queue_entry_t *wait; |
1110 | bool ret; | |
da55f2cc | 1111 | |
c27d53fb | 1112 | if (!(this_hctx->flags & BLK_MQ_F_TAG_SHARED)) { |
f906a6a0 JA |
1113 | if (!test_bit(BLK_MQ_S_SCHED_RESTART, &this_hctx->state)) |
1114 | set_bit(BLK_MQ_S_SCHED_RESTART, &this_hctx->state); | |
eb619fdb | 1115 | |
c27d53fb BVA |
1116 | /* |
1117 | * It's possible that a tag was freed in the window between the | |
1118 | * allocation failure and adding the hardware queue to the wait | |
1119 | * queue. | |
1120 | * | |
1121 | * Don't clear RESTART here, someone else could have set it. | |
1122 | * At most this will cost an extra queue run. | |
1123 | */ | |
1124 | return blk_mq_get_driver_tag(rq, hctx, false); | |
eb619fdb JA |
1125 | } |
1126 | ||
c27d53fb BVA |
1127 | wait = &this_hctx->dispatch_wait; |
1128 | if (!list_empty_careful(&wait->entry)) | |
1129 | return false; | |
1130 | ||
1131 | spin_lock(&this_hctx->lock); | |
1132 | if (!list_empty(&wait->entry)) { | |
1133 | spin_unlock(&this_hctx->lock); | |
1134 | return false; | |
1135 | } | |
1136 | ||
1137 | ws = bt_wait_ptr(&this_hctx->tags->bitmap_tags, this_hctx); | |
1138 | add_wait_queue(&ws->wait, wait); | |
1139 | ||
da55f2cc | 1140 | /* |
eb619fdb JA |
1141 | * It's possible that a tag was freed in the window between the |
1142 | * allocation failure and adding the hardware queue to the wait | |
1143 | * queue. | |
da55f2cc | 1144 | */ |
f906a6a0 | 1145 | ret = blk_mq_get_driver_tag(rq, hctx, false); |
c27d53fb | 1146 | if (!ret) { |
eb619fdb | 1147 | spin_unlock(&this_hctx->lock); |
c27d53fb | 1148 | return false; |
eb619fdb | 1149 | } |
c27d53fb BVA |
1150 | |
1151 | /* | |
1152 | * We got a tag, remove ourselves from the wait queue to ensure | |
1153 | * someone else gets the wakeup. | |
1154 | */ | |
1155 | spin_lock_irq(&ws->wait.lock); | |
1156 | list_del_init(&wait->entry); | |
1157 | spin_unlock_irq(&ws->wait.lock); | |
1158 | spin_unlock(&this_hctx->lock); | |
1159 | ||
1160 | return true; | |
da55f2cc OS |
1161 | } |
1162 | ||
de148297 | 1163 | bool blk_mq_dispatch_rq_list(struct request_queue *q, struct list_head *list, |
eb619fdb | 1164 | bool got_budget) |
320ae51f | 1165 | { |
81380ca1 | 1166 | struct blk_mq_hw_ctx *hctx; |
6d6f167c | 1167 | struct request *rq, *nxt; |
eb619fdb | 1168 | bool no_tag = false; |
fc17b653 | 1169 | int errors, queued; |
320ae51f | 1170 | |
81380ca1 OS |
1171 | if (list_empty(list)) |
1172 | return false; | |
1173 | ||
de148297 ML |
1174 | WARN_ON(!list_is_singular(list) && got_budget); |
1175 | ||
320ae51f JA |
1176 | /* |
1177 | * Now process all the entries, sending them to the driver. | |
1178 | */ | |
93efe981 | 1179 | errors = queued = 0; |
81380ca1 | 1180 | do { |
74c45052 | 1181 | struct blk_mq_queue_data bd; |
fc17b653 | 1182 | blk_status_t ret; |
320ae51f | 1183 | |
f04c3df3 | 1184 | rq = list_first_entry(list, struct request, queuelist); |
bd166ef1 | 1185 | if (!blk_mq_get_driver_tag(rq, &hctx, false)) { |
3c782d67 | 1186 | /* |
da55f2cc | 1187 | * The initial allocation attempt failed, so we need to |
eb619fdb JA |
1188 | * rerun the hardware queue when a tag is freed. The |
1189 | * waitqueue takes care of that. If the queue is run | |
1190 | * before we add this entry back on the dispatch list, | |
1191 | * we'll re-run it below. | |
3c782d67 | 1192 | */ |
f906a6a0 | 1193 | if (!blk_mq_mark_tag_wait(&hctx, rq)) { |
de148297 ML |
1194 | if (got_budget) |
1195 | blk_mq_put_dispatch_budget(hctx); | |
f906a6a0 JA |
1196 | /* |
1197 | * For non-shared tags, the RESTART check | |
1198 | * will suffice. | |
1199 | */ | |
1200 | if (hctx->flags & BLK_MQ_F_TAG_SHARED) | |
1201 | no_tag = true; | |
de148297 ML |
1202 | break; |
1203 | } | |
1204 | } | |
1205 | ||
0c6af1cc ML |
1206 | if (!got_budget && !blk_mq_get_dispatch_budget(hctx)) { |
1207 | blk_mq_put_driver_tag(rq); | |
88022d72 | 1208 | break; |
0c6af1cc | 1209 | } |
da55f2cc | 1210 | |
320ae51f | 1211 | list_del_init(&rq->queuelist); |
320ae51f | 1212 | |
74c45052 | 1213 | bd.rq = rq; |
113285b4 JA |
1214 | |
1215 | /* | |
1216 | * Flag last if we have no more requests, or if we have more | |
1217 | * but can't assign a driver tag to it. | |
1218 | */ | |
1219 | if (list_empty(list)) | |
1220 | bd.last = true; | |
1221 | else { | |
113285b4 JA |
1222 | nxt = list_first_entry(list, struct request, queuelist); |
1223 | bd.last = !blk_mq_get_driver_tag(nxt, NULL, false); | |
1224 | } | |
74c45052 JA |
1225 | |
1226 | ret = q->mq_ops->queue_rq(hctx, &bd); | |
fc17b653 | 1227 | if (ret == BLK_STS_RESOURCE) { |
6d6f167c JW |
1228 | /* |
1229 | * If an I/O scheduler has been configured and we got a | |
ff821d27 JA |
1230 | * driver tag for the next request already, free it |
1231 | * again. | |
6d6f167c JW |
1232 | */ |
1233 | if (!list_empty(list)) { | |
1234 | nxt = list_first_entry(list, struct request, queuelist); | |
1235 | blk_mq_put_driver_tag(nxt); | |
1236 | } | |
f04c3df3 | 1237 | list_add(&rq->queuelist, list); |
ed0791b2 | 1238 | __blk_mq_requeue_request(rq); |
320ae51f | 1239 | break; |
fc17b653 CH |
1240 | } |
1241 | ||
1242 | if (unlikely(ret != BLK_STS_OK)) { | |
93efe981 | 1243 | errors++; |
2a842aca | 1244 | blk_mq_end_request(rq, BLK_STS_IOERR); |
fc17b653 | 1245 | continue; |
320ae51f JA |
1246 | } |
1247 | ||
fc17b653 | 1248 | queued++; |
81380ca1 | 1249 | } while (!list_empty(list)); |
320ae51f | 1250 | |
703fd1c0 | 1251 | hctx->dispatched[queued_to_index(queued)]++; |
320ae51f JA |
1252 | |
1253 | /* | |
1254 | * Any items that need requeuing? Stuff them into hctx->dispatch, | |
1255 | * that is where we will continue on next queue run. | |
1256 | */ | |
f04c3df3 | 1257 | if (!list_empty(list)) { |
320ae51f | 1258 | spin_lock(&hctx->lock); |
c13660a0 | 1259 | list_splice_init(list, &hctx->dispatch); |
320ae51f | 1260 | spin_unlock(&hctx->lock); |
f04c3df3 | 1261 | |
9ba52e58 | 1262 | /* |
710c785f BVA |
1263 | * If SCHED_RESTART was set by the caller of this function and |
1264 | * it is no longer set that means that it was cleared by another | |
1265 | * thread and hence that a queue rerun is needed. | |
9ba52e58 | 1266 | * |
eb619fdb JA |
1267 | * If 'no_tag' is set, that means that we failed getting |
1268 | * a driver tag with an I/O scheduler attached. If our dispatch | |
1269 | * waitqueue is no longer active, ensure that we run the queue | |
1270 | * AFTER adding our entries back to the list. | |
bd166ef1 | 1271 | * |
710c785f BVA |
1272 | * If no I/O scheduler has been configured it is possible that |
1273 | * the hardware queue got stopped and restarted before requests | |
1274 | * were pushed back onto the dispatch list. Rerun the queue to | |
1275 | * avoid starvation. Notes: | |
1276 | * - blk_mq_run_hw_queue() checks whether or not a queue has | |
1277 | * been stopped before rerunning a queue. | |
1278 | * - Some but not all block drivers stop a queue before | |
fc17b653 | 1279 | * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq |
710c785f | 1280 | * and dm-rq. |
bd166ef1 | 1281 | */ |
eb619fdb JA |
1282 | if (!blk_mq_sched_needs_restart(hctx) || |
1283 | (no_tag && list_empty_careful(&hctx->dispatch_wait.entry))) | |
bd166ef1 | 1284 | blk_mq_run_hw_queue(hctx, true); |
320ae51f | 1285 | } |
f04c3df3 | 1286 | |
93efe981 | 1287 | return (queued + errors) != 0; |
f04c3df3 JA |
1288 | } |
1289 | ||
6a83e74d BVA |
1290 | static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx) |
1291 | { | |
1292 | int srcu_idx; | |
1293 | ||
b7a71e66 JA |
1294 | /* |
1295 | * We should be running this queue from one of the CPUs that | |
1296 | * are mapped to it. | |
7df938fb ML |
1297 | * |
1298 | * There are at least two related races now between setting | |
1299 | * hctx->next_cpu from blk_mq_hctx_next_cpu() and running | |
1300 | * __blk_mq_run_hw_queue(): | |
1301 | * | |
1302 | * - hctx->next_cpu is found offline in blk_mq_hctx_next_cpu(), | |
1303 | * but later it becomes online, then this warning is harmless | |
1304 | * at all | |
1305 | * | |
1306 | * - hctx->next_cpu is found online in blk_mq_hctx_next_cpu(), | |
1307 | * but later it becomes offline, then the warning can't be | |
1308 | * triggered, and we depend on blk-mq timeout handler to | |
1309 | * handle dispatched requests to this hctx | |
b7a71e66 | 1310 | */ |
7df938fb ML |
1311 | if (!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) && |
1312 | cpu_online(hctx->next_cpu)) { | |
1313 | printk(KERN_WARNING "run queue from wrong CPU %d, hctx %s\n", | |
1314 | raw_smp_processor_id(), | |
1315 | cpumask_empty(hctx->cpumask) ? "inactive": "active"); | |
1316 | dump_stack(); | |
1317 | } | |
6a83e74d | 1318 | |
b7a71e66 JA |
1319 | /* |
1320 | * We can't run the queue inline with ints disabled. Ensure that | |
1321 | * we catch bad users of this early. | |
1322 | */ | |
1323 | WARN_ON_ONCE(in_interrupt()); | |
1324 | ||
04ced159 | 1325 | might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING); |
bf4907c0 | 1326 | |
04ced159 JA |
1327 | hctx_lock(hctx, &srcu_idx); |
1328 | blk_mq_sched_dispatch_requests(hctx); | |
1329 | hctx_unlock(hctx, srcu_idx); | |
6a83e74d BVA |
1330 | } |
1331 | ||
506e931f JA |
1332 | /* |
1333 | * It'd be great if the workqueue API had a way to pass | |
1334 | * in a mask and had some smarts for more clever placement. | |
1335 | * For now we just round-robin here, switching for every | |
1336 | * BLK_MQ_CPU_WORK_BATCH queued items. | |
1337 | */ | |
1338 | static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx) | |
1339 | { | |
7bed4595 ML |
1340 | bool tried = false; |
1341 | ||
b657d7e6 CH |
1342 | if (hctx->queue->nr_hw_queues == 1) |
1343 | return WORK_CPU_UNBOUND; | |
506e931f JA |
1344 | |
1345 | if (--hctx->next_cpu_batch <= 0) { | |
c02ebfdd | 1346 | int next_cpu; |
7bed4595 | 1347 | select_cpu: |
20e4d813 CH |
1348 | next_cpu = cpumask_next_and(hctx->next_cpu, hctx->cpumask, |
1349 | cpu_online_mask); | |
506e931f | 1350 | if (next_cpu >= nr_cpu_ids) |
20e4d813 | 1351 | next_cpu = cpumask_first_and(hctx->cpumask,cpu_online_mask); |
506e931f | 1352 | |
7bed4595 ML |
1353 | /* |
1354 | * No online CPU is found, so have to make sure hctx->next_cpu | |
1355 | * is set correctly for not breaking workqueue. | |
1356 | */ | |
1357 | if (next_cpu >= nr_cpu_ids) | |
1358 | hctx->next_cpu = cpumask_first(hctx->cpumask); | |
1359 | else | |
1360 | hctx->next_cpu = next_cpu; | |
506e931f JA |
1361 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; |
1362 | } | |
1363 | ||
7bed4595 ML |
1364 | /* |
1365 | * Do unbound schedule if we can't find a online CPU for this hctx, | |
1366 | * and it should only happen in the path of handling CPU DEAD. | |
1367 | */ | |
1368 | if (!cpu_online(hctx->next_cpu)) { | |
1369 | if (!tried) { | |
1370 | tried = true; | |
1371 | goto select_cpu; | |
1372 | } | |
1373 | ||
1374 | /* | |
1375 | * Make sure to re-select CPU next time once after CPUs | |
1376 | * in hctx->cpumask become online again. | |
1377 | */ | |
1378 | hctx->next_cpu_batch = 1; | |
1379 | return WORK_CPU_UNBOUND; | |
1380 | } | |
b657d7e6 | 1381 | return hctx->next_cpu; |
506e931f JA |
1382 | } |
1383 | ||
7587a5ae BVA |
1384 | static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async, |
1385 | unsigned long msecs) | |
320ae51f | 1386 | { |
5435c023 BVA |
1387 | if (WARN_ON_ONCE(!blk_mq_hw_queue_mapped(hctx))) |
1388 | return; | |
1389 | ||
1390 | if (unlikely(blk_mq_hctx_stopped(hctx))) | |
320ae51f JA |
1391 | return; |
1392 | ||
1b792f2f | 1393 | if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) { |
2a90d4aa PB |
1394 | int cpu = get_cpu(); |
1395 | if (cpumask_test_cpu(cpu, hctx->cpumask)) { | |
398205b8 | 1396 | __blk_mq_run_hw_queue(hctx); |
2a90d4aa | 1397 | put_cpu(); |
398205b8 PB |
1398 | return; |
1399 | } | |
e4043dcf | 1400 | |
2a90d4aa | 1401 | put_cpu(); |
e4043dcf | 1402 | } |
398205b8 | 1403 | |
9f993737 JA |
1404 | kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx), |
1405 | &hctx->run_work, | |
1406 | msecs_to_jiffies(msecs)); | |
7587a5ae BVA |
1407 | } |
1408 | ||
1409 | void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs) | |
1410 | { | |
1411 | __blk_mq_delay_run_hw_queue(hctx, true, msecs); | |
1412 | } | |
1413 | EXPORT_SYMBOL(blk_mq_delay_run_hw_queue); | |
1414 | ||
79f720a7 | 1415 | bool blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) |
7587a5ae | 1416 | { |
24f5a90f ML |
1417 | int srcu_idx; |
1418 | bool need_run; | |
1419 | ||
1420 | /* | |
1421 | * When queue is quiesced, we may be switching io scheduler, or | |
1422 | * updating nr_hw_queues, or other things, and we can't run queue | |
1423 | * any more, even __blk_mq_hctx_has_pending() can't be called safely. | |
1424 | * | |
1425 | * And queue will be rerun in blk_mq_unquiesce_queue() if it is | |
1426 | * quiesced. | |
1427 | */ | |
04ced159 JA |
1428 | hctx_lock(hctx, &srcu_idx); |
1429 | need_run = !blk_queue_quiesced(hctx->queue) && | |
1430 | blk_mq_hctx_has_pending(hctx); | |
1431 | hctx_unlock(hctx, srcu_idx); | |
24f5a90f ML |
1432 | |
1433 | if (need_run) { | |
79f720a7 JA |
1434 | __blk_mq_delay_run_hw_queue(hctx, async, 0); |
1435 | return true; | |
1436 | } | |
1437 | ||
1438 | return false; | |
320ae51f | 1439 | } |
5b727272 | 1440 | EXPORT_SYMBOL(blk_mq_run_hw_queue); |
320ae51f | 1441 | |
b94ec296 | 1442 | void blk_mq_run_hw_queues(struct request_queue *q, bool async) |
320ae51f JA |
1443 | { |
1444 | struct blk_mq_hw_ctx *hctx; | |
1445 | int i; | |
1446 | ||
1447 | queue_for_each_hw_ctx(q, hctx, i) { | |
79f720a7 | 1448 | if (blk_mq_hctx_stopped(hctx)) |
320ae51f JA |
1449 | continue; |
1450 | ||
b94ec296 | 1451 | blk_mq_run_hw_queue(hctx, async); |
320ae51f JA |
1452 | } |
1453 | } | |
b94ec296 | 1454 | EXPORT_SYMBOL(blk_mq_run_hw_queues); |
320ae51f | 1455 | |
fd001443 BVA |
1456 | /** |
1457 | * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped | |
1458 | * @q: request queue. | |
1459 | * | |
1460 | * The caller is responsible for serializing this function against | |
1461 | * blk_mq_{start,stop}_hw_queue(). | |
1462 | */ | |
1463 | bool blk_mq_queue_stopped(struct request_queue *q) | |
1464 | { | |
1465 | struct blk_mq_hw_ctx *hctx; | |
1466 | int i; | |
1467 | ||
1468 | queue_for_each_hw_ctx(q, hctx, i) | |
1469 | if (blk_mq_hctx_stopped(hctx)) | |
1470 | return true; | |
1471 | ||
1472 | return false; | |
1473 | } | |
1474 | EXPORT_SYMBOL(blk_mq_queue_stopped); | |
1475 | ||
39a70c76 ML |
1476 | /* |
1477 | * This function is often used for pausing .queue_rq() by driver when | |
1478 | * there isn't enough resource or some conditions aren't satisfied, and | |
4d606219 | 1479 | * BLK_STS_RESOURCE is usually returned. |
39a70c76 ML |
1480 | * |
1481 | * We do not guarantee that dispatch can be drained or blocked | |
1482 | * after blk_mq_stop_hw_queue() returns. Please use | |
1483 | * blk_mq_quiesce_queue() for that requirement. | |
1484 | */ | |
2719aa21 JA |
1485 | void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx) |
1486 | { | |
641a9ed6 | 1487 | cancel_delayed_work(&hctx->run_work); |
280d45f6 | 1488 | |
641a9ed6 | 1489 | set_bit(BLK_MQ_S_STOPPED, &hctx->state); |
2719aa21 | 1490 | } |
641a9ed6 | 1491 | EXPORT_SYMBOL(blk_mq_stop_hw_queue); |
2719aa21 | 1492 | |
39a70c76 ML |
1493 | /* |
1494 | * This function is often used for pausing .queue_rq() by driver when | |
1495 | * there isn't enough resource or some conditions aren't satisfied, and | |
4d606219 | 1496 | * BLK_STS_RESOURCE is usually returned. |
39a70c76 ML |
1497 | * |
1498 | * We do not guarantee that dispatch can be drained or blocked | |
1499 | * after blk_mq_stop_hw_queues() returns. Please use | |
1500 | * blk_mq_quiesce_queue() for that requirement. | |
1501 | */ | |
2719aa21 JA |
1502 | void blk_mq_stop_hw_queues(struct request_queue *q) |
1503 | { | |
641a9ed6 ML |
1504 | struct blk_mq_hw_ctx *hctx; |
1505 | int i; | |
1506 | ||
1507 | queue_for_each_hw_ctx(q, hctx, i) | |
1508 | blk_mq_stop_hw_queue(hctx); | |
280d45f6 CH |
1509 | } |
1510 | EXPORT_SYMBOL(blk_mq_stop_hw_queues); | |
1511 | ||
320ae51f JA |
1512 | void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx) |
1513 | { | |
1514 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
e4043dcf | 1515 | |
0ffbce80 | 1516 | blk_mq_run_hw_queue(hctx, false); |
320ae51f JA |
1517 | } |
1518 | EXPORT_SYMBOL(blk_mq_start_hw_queue); | |
1519 | ||
2f268556 CH |
1520 | void blk_mq_start_hw_queues(struct request_queue *q) |
1521 | { | |
1522 | struct blk_mq_hw_ctx *hctx; | |
1523 | int i; | |
1524 | ||
1525 | queue_for_each_hw_ctx(q, hctx, i) | |
1526 | blk_mq_start_hw_queue(hctx); | |
1527 | } | |
1528 | EXPORT_SYMBOL(blk_mq_start_hw_queues); | |
1529 | ||
ae911c5e JA |
1530 | void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) |
1531 | { | |
1532 | if (!blk_mq_hctx_stopped(hctx)) | |
1533 | return; | |
1534 | ||
1535 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
1536 | blk_mq_run_hw_queue(hctx, async); | |
1537 | } | |
1538 | EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue); | |
1539 | ||
1b4a3258 | 1540 | void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async) |
320ae51f JA |
1541 | { |
1542 | struct blk_mq_hw_ctx *hctx; | |
1543 | int i; | |
1544 | ||
ae911c5e JA |
1545 | queue_for_each_hw_ctx(q, hctx, i) |
1546 | blk_mq_start_stopped_hw_queue(hctx, async); | |
320ae51f JA |
1547 | } |
1548 | EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues); | |
1549 | ||
70f4db63 | 1550 | static void blk_mq_run_work_fn(struct work_struct *work) |
320ae51f JA |
1551 | { |
1552 | struct blk_mq_hw_ctx *hctx; | |
1553 | ||
9f993737 | 1554 | hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work); |
320ae51f | 1555 | |
21c6e939 JA |
1556 | /* |
1557 | * If we are stopped, don't run the queue. The exception is if | |
1558 | * BLK_MQ_S_START_ON_RUN is set. For that case, we auto-clear | |
1559 | * the STOPPED bit and run it. | |
1560 | */ | |
1561 | if (test_bit(BLK_MQ_S_STOPPED, &hctx->state)) { | |
1562 | if (!test_bit(BLK_MQ_S_START_ON_RUN, &hctx->state)) | |
1563 | return; | |
7587a5ae | 1564 | |
21c6e939 JA |
1565 | clear_bit(BLK_MQ_S_START_ON_RUN, &hctx->state); |
1566 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
1567 | } | |
7587a5ae BVA |
1568 | |
1569 | __blk_mq_run_hw_queue(hctx); | |
1570 | } | |
1571 | ||
70f4db63 CH |
1572 | |
1573 | void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs) | |
1574 | { | |
5435c023 | 1575 | if (WARN_ON_ONCE(!blk_mq_hw_queue_mapped(hctx))) |
19c66e59 | 1576 | return; |
70f4db63 | 1577 | |
21c6e939 JA |
1578 | /* |
1579 | * Stop the hw queue, then modify currently delayed work. | |
1580 | * This should prevent us from running the queue prematurely. | |
1581 | * Mark the queue as auto-clearing STOPPED when it runs. | |
1582 | */ | |
7e79dadc | 1583 | blk_mq_stop_hw_queue(hctx); |
21c6e939 JA |
1584 | set_bit(BLK_MQ_S_START_ON_RUN, &hctx->state); |
1585 | kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), | |
1586 | &hctx->run_work, | |
1587 | msecs_to_jiffies(msecs)); | |
70f4db63 CH |
1588 | } |
1589 | EXPORT_SYMBOL(blk_mq_delay_queue); | |
1590 | ||
cfd0c552 | 1591 | static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx, |
cfd0c552 ML |
1592 | struct request *rq, |
1593 | bool at_head) | |
320ae51f | 1594 | { |
e57690fe JA |
1595 | struct blk_mq_ctx *ctx = rq->mq_ctx; |
1596 | ||
7b607814 BVA |
1597 | lockdep_assert_held(&ctx->lock); |
1598 | ||
01b983c9 JA |
1599 | trace_block_rq_insert(hctx->queue, rq); |
1600 | ||
72a0a36e CH |
1601 | if (at_head) |
1602 | list_add(&rq->queuelist, &ctx->rq_list); | |
1603 | else | |
1604 | list_add_tail(&rq->queuelist, &ctx->rq_list); | |
cfd0c552 | 1605 | } |
4bb659b1 | 1606 | |
2c3ad667 JA |
1607 | void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq, |
1608 | bool at_head) | |
cfd0c552 ML |
1609 | { |
1610 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
1611 | ||
7b607814 BVA |
1612 | lockdep_assert_held(&ctx->lock); |
1613 | ||
e57690fe | 1614 | __blk_mq_insert_req_list(hctx, rq, at_head); |
320ae51f | 1615 | blk_mq_hctx_mark_pending(hctx, ctx); |
320ae51f JA |
1616 | } |
1617 | ||
157f377b JA |
1618 | /* |
1619 | * Should only be used carefully, when the caller knows we want to | |
1620 | * bypass a potential IO scheduler on the target device. | |
1621 | */ | |
b0850297 | 1622 | void blk_mq_request_bypass_insert(struct request *rq, bool run_queue) |
157f377b JA |
1623 | { |
1624 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
1625 | struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(rq->q, ctx->cpu); | |
1626 | ||
1627 | spin_lock(&hctx->lock); | |
1628 | list_add_tail(&rq->queuelist, &hctx->dispatch); | |
1629 | spin_unlock(&hctx->lock); | |
1630 | ||
b0850297 ML |
1631 | if (run_queue) |
1632 | blk_mq_run_hw_queue(hctx, false); | |
157f377b JA |
1633 | } |
1634 | ||
bd166ef1 JA |
1635 | void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx, |
1636 | struct list_head *list) | |
320ae51f JA |
1637 | |
1638 | { | |
320ae51f JA |
1639 | /* |
1640 | * preemption doesn't flush plug list, so it's possible ctx->cpu is | |
1641 | * offline now | |
1642 | */ | |
1643 | spin_lock(&ctx->lock); | |
1644 | while (!list_empty(list)) { | |
1645 | struct request *rq; | |
1646 | ||
1647 | rq = list_first_entry(list, struct request, queuelist); | |
e57690fe | 1648 | BUG_ON(rq->mq_ctx != ctx); |
320ae51f | 1649 | list_del_init(&rq->queuelist); |
e57690fe | 1650 | __blk_mq_insert_req_list(hctx, rq, false); |
320ae51f | 1651 | } |
cfd0c552 | 1652 | blk_mq_hctx_mark_pending(hctx, ctx); |
320ae51f | 1653 | spin_unlock(&ctx->lock); |
320ae51f JA |
1654 | } |
1655 | ||
1656 | static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b) | |
1657 | { | |
1658 | struct request *rqa = container_of(a, struct request, queuelist); | |
1659 | struct request *rqb = container_of(b, struct request, queuelist); | |
1660 | ||
1661 | return !(rqa->mq_ctx < rqb->mq_ctx || | |
1662 | (rqa->mq_ctx == rqb->mq_ctx && | |
1663 | blk_rq_pos(rqa) < blk_rq_pos(rqb))); | |
1664 | } | |
1665 | ||
1666 | void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule) | |
1667 | { | |
1668 | struct blk_mq_ctx *this_ctx; | |
1669 | struct request_queue *this_q; | |
1670 | struct request *rq; | |
1671 | LIST_HEAD(list); | |
1672 | LIST_HEAD(ctx_list); | |
1673 | unsigned int depth; | |
1674 | ||
1675 | list_splice_init(&plug->mq_list, &list); | |
1676 | ||
1677 | list_sort(NULL, &list, plug_ctx_cmp); | |
1678 | ||
1679 | this_q = NULL; | |
1680 | this_ctx = NULL; | |
1681 | depth = 0; | |
1682 | ||
1683 | while (!list_empty(&list)) { | |
1684 | rq = list_entry_rq(list.next); | |
1685 | list_del_init(&rq->queuelist); | |
1686 | BUG_ON(!rq->q); | |
1687 | if (rq->mq_ctx != this_ctx) { | |
1688 | if (this_ctx) { | |
bd166ef1 JA |
1689 | trace_block_unplug(this_q, depth, from_schedule); |
1690 | blk_mq_sched_insert_requests(this_q, this_ctx, | |
1691 | &ctx_list, | |
1692 | from_schedule); | |
320ae51f JA |
1693 | } |
1694 | ||
1695 | this_ctx = rq->mq_ctx; | |
1696 | this_q = rq->q; | |
1697 | depth = 0; | |
1698 | } | |
1699 | ||
1700 | depth++; | |
1701 | list_add_tail(&rq->queuelist, &ctx_list); | |
1702 | } | |
1703 | ||
1704 | /* | |
1705 | * If 'this_ctx' is set, we know we have entries to complete | |
1706 | * on 'ctx_list'. Do those. | |
1707 | */ | |
1708 | if (this_ctx) { | |
bd166ef1 JA |
1709 | trace_block_unplug(this_q, depth, from_schedule); |
1710 | blk_mq_sched_insert_requests(this_q, this_ctx, &ctx_list, | |
1711 | from_schedule); | |
320ae51f JA |
1712 | } |
1713 | } | |
1714 | ||
1715 | static void blk_mq_bio_to_request(struct request *rq, struct bio *bio) | |
1716 | { | |
da8d7f07 | 1717 | blk_init_request_from_bio(rq, bio); |
4b570521 | 1718 | |
85acb3ba SL |
1719 | blk_rq_set_rl(rq, blk_get_rl(rq->q, bio)); |
1720 | ||
6e85eaf3 | 1721 | blk_account_io_start(rq, true); |
320ae51f JA |
1722 | } |
1723 | ||
ab42f35d ML |
1724 | static inline void blk_mq_queue_io(struct blk_mq_hw_ctx *hctx, |
1725 | struct blk_mq_ctx *ctx, | |
1726 | struct request *rq) | |
1727 | { | |
1728 | spin_lock(&ctx->lock); | |
1729 | __blk_mq_insert_request(hctx, rq, false); | |
1730 | spin_unlock(&ctx->lock); | |
07068d5b | 1731 | } |
14ec77f3 | 1732 | |
fd2d3326 JA |
1733 | static blk_qc_t request_to_qc_t(struct blk_mq_hw_ctx *hctx, struct request *rq) |
1734 | { | |
bd166ef1 JA |
1735 | if (rq->tag != -1) |
1736 | return blk_tag_to_qc_t(rq->tag, hctx->queue_num, false); | |
1737 | ||
1738 | return blk_tag_to_qc_t(rq->internal_tag, hctx->queue_num, true); | |
fd2d3326 JA |
1739 | } |
1740 | ||
0f95549c MS |
1741 | static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx, |
1742 | struct request *rq, | |
1743 | blk_qc_t *cookie) | |
f984df1f | 1744 | { |
f984df1f | 1745 | struct request_queue *q = rq->q; |
f984df1f SL |
1746 | struct blk_mq_queue_data bd = { |
1747 | .rq = rq, | |
d945a365 | 1748 | .last = true, |
f984df1f | 1749 | }; |
bd166ef1 | 1750 | blk_qc_t new_cookie; |
f06345ad | 1751 | blk_status_t ret; |
0f95549c MS |
1752 | |
1753 | new_cookie = request_to_qc_t(hctx, rq); | |
1754 | ||
1755 | /* | |
1756 | * For OK queue, we are done. For error, caller may kill it. | |
1757 | * Any other error (busy), just add it to our list as we | |
1758 | * previously would have done. | |
1759 | */ | |
1760 | ret = q->mq_ops->queue_rq(hctx, &bd); | |
1761 | switch (ret) { | |
1762 | case BLK_STS_OK: | |
1763 | *cookie = new_cookie; | |
1764 | break; | |
1765 | case BLK_STS_RESOURCE: | |
1766 | __blk_mq_requeue_request(rq); | |
1767 | break; | |
1768 | default: | |
1769 | *cookie = BLK_QC_T_NONE; | |
1770 | break; | |
1771 | } | |
1772 | ||
1773 | return ret; | |
1774 | } | |
1775 | ||
0f95549c MS |
1776 | static blk_status_t __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx, |
1777 | struct request *rq, | |
396eaf21 ML |
1778 | blk_qc_t *cookie, |
1779 | bool bypass_insert) | |
0f95549c MS |
1780 | { |
1781 | struct request_queue *q = rq->q; | |
d964f04a ML |
1782 | bool run_queue = true; |
1783 | ||
23d4ee19 ML |
1784 | /* |
1785 | * RCU or SRCU read lock is needed before checking quiesced flag. | |
1786 | * | |
1787 | * When queue is stopped or quiesced, ignore 'bypass_insert' from | |
1788 | * blk_mq_request_direct_issue(), and return BLK_STS_OK to caller, | |
1789 | * and avoid driver to try to dispatch again. | |
1790 | */ | |
f4560ffe | 1791 | if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)) { |
d964f04a | 1792 | run_queue = false; |
23d4ee19 | 1793 | bypass_insert = false; |
d964f04a ML |
1794 | goto insert; |
1795 | } | |
f984df1f | 1796 | |
396eaf21 | 1797 | if (q->elevator && !bypass_insert) |
2253efc8 BVA |
1798 | goto insert; |
1799 | ||
d964f04a | 1800 | if (!blk_mq_get_driver_tag(rq, NULL, false)) |
bd166ef1 JA |
1801 | goto insert; |
1802 | ||
88022d72 | 1803 | if (!blk_mq_get_dispatch_budget(hctx)) { |
de148297 ML |
1804 | blk_mq_put_driver_tag(rq); |
1805 | goto insert; | |
88022d72 | 1806 | } |
de148297 | 1807 | |
0f95549c | 1808 | return __blk_mq_issue_directly(hctx, rq, cookie); |
2253efc8 | 1809 | insert: |
396eaf21 ML |
1810 | if (bypass_insert) |
1811 | return BLK_STS_RESOURCE; | |
0f95549c | 1812 | |
23d4ee19 | 1813 | blk_mq_sched_insert_request(rq, false, run_queue, false); |
0f95549c | 1814 | return BLK_STS_OK; |
f984df1f SL |
1815 | } |
1816 | ||
5eb6126e CH |
1817 | static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx, |
1818 | struct request *rq, blk_qc_t *cookie) | |
1819 | { | |
0f95549c | 1820 | blk_status_t ret; |
04ced159 | 1821 | int srcu_idx; |
bf4907c0 | 1822 | |
04ced159 | 1823 | might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING); |
bf4907c0 | 1824 | |
04ced159 | 1825 | hctx_lock(hctx, &srcu_idx); |
0f95549c | 1826 | |
396eaf21 | 1827 | ret = __blk_mq_try_issue_directly(hctx, rq, cookie, false); |
0f95549c | 1828 | if (ret == BLK_STS_RESOURCE) |
23d4ee19 | 1829 | blk_mq_sched_insert_request(rq, false, true, false); |
0f95549c MS |
1830 | else if (ret != BLK_STS_OK) |
1831 | blk_mq_end_request(rq, ret); | |
1832 | ||
04ced159 | 1833 | hctx_unlock(hctx, srcu_idx); |
5eb6126e CH |
1834 | } |
1835 | ||
396eaf21 ML |
1836 | blk_status_t blk_mq_request_direct_issue(struct request *rq) |
1837 | { | |
1838 | blk_status_t ret; | |
1839 | int srcu_idx; | |
1840 | blk_qc_t unused_cookie; | |
1841 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
1842 | struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(rq->q, ctx->cpu); | |
1843 | ||
1844 | hctx_lock(hctx, &srcu_idx); | |
1845 | ret = __blk_mq_try_issue_directly(hctx, rq, &unused_cookie, true); | |
1846 | hctx_unlock(hctx, srcu_idx); | |
1847 | ||
1848 | return ret; | |
1849 | } | |
1850 | ||
dece1635 | 1851 | static blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio) |
07068d5b | 1852 | { |
ef295ecf | 1853 | const int is_sync = op_is_sync(bio->bi_opf); |
f73f44eb | 1854 | const int is_flush_fua = op_is_flush(bio->bi_opf); |
5a797e00 | 1855 | struct blk_mq_alloc_data data = { .flags = 0 }; |
07068d5b | 1856 | struct request *rq; |
5eb6126e | 1857 | unsigned int request_count = 0; |
f984df1f | 1858 | struct blk_plug *plug; |
5b3f341f | 1859 | struct request *same_queue_rq = NULL; |
7b371636 | 1860 | blk_qc_t cookie; |
87760e5e | 1861 | unsigned int wb_acct; |
07068d5b JA |
1862 | |
1863 | blk_queue_bounce(q, &bio); | |
1864 | ||
af67c31f | 1865 | blk_queue_split(q, &bio); |
f36ea50c | 1866 | |
e23947bd | 1867 | if (!bio_integrity_prep(bio)) |
dece1635 | 1868 | return BLK_QC_T_NONE; |
07068d5b | 1869 | |
87c279e6 OS |
1870 | if (!is_flush_fua && !blk_queue_nomerges(q) && |
1871 | blk_attempt_plug_merge(q, bio, &request_count, &same_queue_rq)) | |
1872 | return BLK_QC_T_NONE; | |
f984df1f | 1873 | |
bd166ef1 JA |
1874 | if (blk_mq_sched_bio_merge(q, bio)) |
1875 | return BLK_QC_T_NONE; | |
1876 | ||
87760e5e JA |
1877 | wb_acct = wbt_wait(q->rq_wb, bio, NULL); |
1878 | ||
bd166ef1 JA |
1879 | trace_block_getrq(q, bio, bio->bi_opf); |
1880 | ||
d2c0d383 | 1881 | rq = blk_mq_get_request(q, bio, bio->bi_opf, &data); |
87760e5e JA |
1882 | if (unlikely(!rq)) { |
1883 | __wbt_done(q->rq_wb, wb_acct); | |
03a07c92 GR |
1884 | if (bio->bi_opf & REQ_NOWAIT) |
1885 | bio_wouldblock_error(bio); | |
dece1635 | 1886 | return BLK_QC_T_NONE; |
87760e5e JA |
1887 | } |
1888 | ||
1889 | wbt_track(&rq->issue_stat, wb_acct); | |
07068d5b | 1890 | |
fd2d3326 | 1891 | cookie = request_to_qc_t(data.hctx, rq); |
07068d5b | 1892 | |
f984df1f | 1893 | plug = current->plug; |
07068d5b | 1894 | if (unlikely(is_flush_fua)) { |
f984df1f | 1895 | blk_mq_put_ctx(data.ctx); |
07068d5b | 1896 | blk_mq_bio_to_request(rq, bio); |
923218f6 ML |
1897 | |
1898 | /* bypass scheduler for flush rq */ | |
1899 | blk_insert_flush(rq); | |
1900 | blk_mq_run_hw_queue(data.hctx, true); | |
a4d907b6 | 1901 | } else if (plug && q->nr_hw_queues == 1) { |
600271d9 SL |
1902 | struct request *last = NULL; |
1903 | ||
b00c53e8 | 1904 | blk_mq_put_ctx(data.ctx); |
e6c4438b | 1905 | blk_mq_bio_to_request(rq, bio); |
0a6219a9 ML |
1906 | |
1907 | /* | |
1908 | * @request_count may become stale because of schedule | |
1909 | * out, so check the list again. | |
1910 | */ | |
1911 | if (list_empty(&plug->mq_list)) | |
1912 | request_count = 0; | |
254d259d CH |
1913 | else if (blk_queue_nomerges(q)) |
1914 | request_count = blk_plug_queued_count(q); | |
1915 | ||
676d0607 | 1916 | if (!request_count) |
e6c4438b | 1917 | trace_block_plug(q); |
600271d9 SL |
1918 | else |
1919 | last = list_entry_rq(plug->mq_list.prev); | |
b094f89c | 1920 | |
600271d9 SL |
1921 | if (request_count >= BLK_MAX_REQUEST_COUNT || (last && |
1922 | blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) { | |
e6c4438b JM |
1923 | blk_flush_plug_list(plug, false); |
1924 | trace_block_plug(q); | |
320ae51f | 1925 | } |
b094f89c | 1926 | |
e6c4438b | 1927 | list_add_tail(&rq->queuelist, &plug->mq_list); |
2299722c | 1928 | } else if (plug && !blk_queue_nomerges(q)) { |
bd166ef1 | 1929 | blk_mq_bio_to_request(rq, bio); |
07068d5b | 1930 | |
07068d5b | 1931 | /* |
6a83e74d | 1932 | * We do limited plugging. If the bio can be merged, do that. |
f984df1f SL |
1933 | * Otherwise the existing request in the plug list will be |
1934 | * issued. So the plug list will have one request at most | |
2299722c CH |
1935 | * The plug list might get flushed before this. If that happens, |
1936 | * the plug list is empty, and same_queue_rq is invalid. | |
07068d5b | 1937 | */ |
2299722c CH |
1938 | if (list_empty(&plug->mq_list)) |
1939 | same_queue_rq = NULL; | |
1940 | if (same_queue_rq) | |
1941 | list_del_init(&same_queue_rq->queuelist); | |
1942 | list_add_tail(&rq->queuelist, &plug->mq_list); | |
1943 | ||
bf4907c0 JA |
1944 | blk_mq_put_ctx(data.ctx); |
1945 | ||
dad7a3be ML |
1946 | if (same_queue_rq) { |
1947 | data.hctx = blk_mq_map_queue(q, | |
1948 | same_queue_rq->mq_ctx->cpu); | |
2299722c CH |
1949 | blk_mq_try_issue_directly(data.hctx, same_queue_rq, |
1950 | &cookie); | |
dad7a3be | 1951 | } |
a4d907b6 | 1952 | } else if (q->nr_hw_queues > 1 && is_sync) { |
bf4907c0 | 1953 | blk_mq_put_ctx(data.ctx); |
2299722c | 1954 | blk_mq_bio_to_request(rq, bio); |
2299722c | 1955 | blk_mq_try_issue_directly(data.hctx, rq, &cookie); |
a4d907b6 | 1956 | } else if (q->elevator) { |
b00c53e8 | 1957 | blk_mq_put_ctx(data.ctx); |
bd166ef1 | 1958 | blk_mq_bio_to_request(rq, bio); |
9e97d295 | 1959 | blk_mq_sched_insert_request(rq, false, true, true); |
ab42f35d | 1960 | } else { |
b00c53e8 | 1961 | blk_mq_put_ctx(data.ctx); |
ab42f35d ML |
1962 | blk_mq_bio_to_request(rq, bio); |
1963 | blk_mq_queue_io(data.hctx, data.ctx, rq); | |
a4d907b6 | 1964 | blk_mq_run_hw_queue(data.hctx, true); |
ab42f35d | 1965 | } |
320ae51f | 1966 | |
7b371636 | 1967 | return cookie; |
320ae51f JA |
1968 | } |
1969 | ||
cc71a6f4 JA |
1970 | void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags, |
1971 | unsigned int hctx_idx) | |
95363efd | 1972 | { |
e9b267d9 | 1973 | struct page *page; |
320ae51f | 1974 | |
24d2f903 | 1975 | if (tags->rqs && set->ops->exit_request) { |
e9b267d9 | 1976 | int i; |
320ae51f | 1977 | |
24d2f903 | 1978 | for (i = 0; i < tags->nr_tags; i++) { |
2af8cbe3 JA |
1979 | struct request *rq = tags->static_rqs[i]; |
1980 | ||
1981 | if (!rq) | |
e9b267d9 | 1982 | continue; |
d6296d39 | 1983 | set->ops->exit_request(set, rq, hctx_idx); |
2af8cbe3 | 1984 | tags->static_rqs[i] = NULL; |
e9b267d9 | 1985 | } |
320ae51f | 1986 | } |
320ae51f | 1987 | |
24d2f903 CH |
1988 | while (!list_empty(&tags->page_list)) { |
1989 | page = list_first_entry(&tags->page_list, struct page, lru); | |
6753471c | 1990 | list_del_init(&page->lru); |
f75782e4 CM |
1991 | /* |
1992 | * Remove kmemleak object previously allocated in | |
1993 | * blk_mq_init_rq_map(). | |
1994 | */ | |
1995 | kmemleak_free(page_address(page)); | |
320ae51f JA |
1996 | __free_pages(page, page->private); |
1997 | } | |
cc71a6f4 | 1998 | } |
320ae51f | 1999 | |
cc71a6f4 JA |
2000 | void blk_mq_free_rq_map(struct blk_mq_tags *tags) |
2001 | { | |
24d2f903 | 2002 | kfree(tags->rqs); |
cc71a6f4 | 2003 | tags->rqs = NULL; |
2af8cbe3 JA |
2004 | kfree(tags->static_rqs); |
2005 | tags->static_rqs = NULL; | |
320ae51f | 2006 | |
24d2f903 | 2007 | blk_mq_free_tags(tags); |
320ae51f JA |
2008 | } |
2009 | ||
cc71a6f4 JA |
2010 | struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, |
2011 | unsigned int hctx_idx, | |
2012 | unsigned int nr_tags, | |
2013 | unsigned int reserved_tags) | |
320ae51f | 2014 | { |
24d2f903 | 2015 | struct blk_mq_tags *tags; |
59f082e4 | 2016 | int node; |
320ae51f | 2017 | |
59f082e4 SL |
2018 | node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx); |
2019 | if (node == NUMA_NO_NODE) | |
2020 | node = set->numa_node; | |
2021 | ||
2022 | tags = blk_mq_init_tags(nr_tags, reserved_tags, node, | |
24391c0d | 2023 | BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags)); |
24d2f903 CH |
2024 | if (!tags) |
2025 | return NULL; | |
320ae51f | 2026 | |
cc71a6f4 | 2027 | tags->rqs = kzalloc_node(nr_tags * sizeof(struct request *), |
36e1f3d1 | 2028 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, |
59f082e4 | 2029 | node); |
24d2f903 CH |
2030 | if (!tags->rqs) { |
2031 | blk_mq_free_tags(tags); | |
2032 | return NULL; | |
2033 | } | |
320ae51f | 2034 | |
2af8cbe3 JA |
2035 | tags->static_rqs = kzalloc_node(nr_tags * sizeof(struct request *), |
2036 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, | |
59f082e4 | 2037 | node); |
2af8cbe3 JA |
2038 | if (!tags->static_rqs) { |
2039 | kfree(tags->rqs); | |
2040 | blk_mq_free_tags(tags); | |
2041 | return NULL; | |
2042 | } | |
2043 | ||
cc71a6f4 JA |
2044 | return tags; |
2045 | } | |
2046 | ||
2047 | static size_t order_to_size(unsigned int order) | |
2048 | { | |
2049 | return (size_t)PAGE_SIZE << order; | |
2050 | } | |
2051 | ||
1d9bd516 TH |
2052 | static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq, |
2053 | unsigned int hctx_idx, int node) | |
2054 | { | |
2055 | int ret; | |
2056 | ||
2057 | if (set->ops->init_request) { | |
2058 | ret = set->ops->init_request(set, rq, hctx_idx, node); | |
2059 | if (ret) | |
2060 | return ret; | |
2061 | } | |
2062 | ||
2063 | seqcount_init(&rq->gstate_seq); | |
2064 | u64_stats_init(&rq->aborted_gstate_sync); | |
2065 | return 0; | |
2066 | } | |
2067 | ||
cc71a6f4 JA |
2068 | int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags, |
2069 | unsigned int hctx_idx, unsigned int depth) | |
2070 | { | |
2071 | unsigned int i, j, entries_per_page, max_order = 4; | |
2072 | size_t rq_size, left; | |
59f082e4 SL |
2073 | int node; |
2074 | ||
2075 | node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx); | |
2076 | if (node == NUMA_NO_NODE) | |
2077 | node = set->numa_node; | |
cc71a6f4 JA |
2078 | |
2079 | INIT_LIST_HEAD(&tags->page_list); | |
2080 | ||
320ae51f JA |
2081 | /* |
2082 | * rq_size is the size of the request plus driver payload, rounded | |
2083 | * to the cacheline size | |
2084 | */ | |
24d2f903 | 2085 | rq_size = round_up(sizeof(struct request) + set->cmd_size, |
320ae51f | 2086 | cache_line_size()); |
cc71a6f4 | 2087 | left = rq_size * depth; |
320ae51f | 2088 | |
cc71a6f4 | 2089 | for (i = 0; i < depth; ) { |
320ae51f JA |
2090 | int this_order = max_order; |
2091 | struct page *page; | |
2092 | int to_do; | |
2093 | void *p; | |
2094 | ||
b3a834b1 | 2095 | while (this_order && left < order_to_size(this_order - 1)) |
320ae51f JA |
2096 | this_order--; |
2097 | ||
2098 | do { | |
59f082e4 | 2099 | page = alloc_pages_node(node, |
36e1f3d1 | 2100 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO, |
a5164405 | 2101 | this_order); |
320ae51f JA |
2102 | if (page) |
2103 | break; | |
2104 | if (!this_order--) | |
2105 | break; | |
2106 | if (order_to_size(this_order) < rq_size) | |
2107 | break; | |
2108 | } while (1); | |
2109 | ||
2110 | if (!page) | |
24d2f903 | 2111 | goto fail; |
320ae51f JA |
2112 | |
2113 | page->private = this_order; | |
24d2f903 | 2114 | list_add_tail(&page->lru, &tags->page_list); |
320ae51f JA |
2115 | |
2116 | p = page_address(page); | |
f75782e4 CM |
2117 | /* |
2118 | * Allow kmemleak to scan these pages as they contain pointers | |
2119 | * to additional allocations like via ops->init_request(). | |
2120 | */ | |
36e1f3d1 | 2121 | kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO); |
320ae51f | 2122 | entries_per_page = order_to_size(this_order) / rq_size; |
cc71a6f4 | 2123 | to_do = min(entries_per_page, depth - i); |
320ae51f JA |
2124 | left -= to_do * rq_size; |
2125 | for (j = 0; j < to_do; j++) { | |
2af8cbe3 JA |
2126 | struct request *rq = p; |
2127 | ||
2128 | tags->static_rqs[i] = rq; | |
1d9bd516 TH |
2129 | if (blk_mq_init_request(set, rq, hctx_idx, node)) { |
2130 | tags->static_rqs[i] = NULL; | |
2131 | goto fail; | |
e9b267d9 CH |
2132 | } |
2133 | ||
320ae51f JA |
2134 | p += rq_size; |
2135 | i++; | |
2136 | } | |
2137 | } | |
cc71a6f4 | 2138 | return 0; |
320ae51f | 2139 | |
24d2f903 | 2140 | fail: |
cc71a6f4 JA |
2141 | blk_mq_free_rqs(set, tags, hctx_idx); |
2142 | return -ENOMEM; | |
320ae51f JA |
2143 | } |
2144 | ||
e57690fe JA |
2145 | /* |
2146 | * 'cpu' is going away. splice any existing rq_list entries from this | |
2147 | * software queue to the hw queue dispatch list, and ensure that it | |
2148 | * gets run. | |
2149 | */ | |
9467f859 | 2150 | static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node) |
484b4061 | 2151 | { |
9467f859 | 2152 | struct blk_mq_hw_ctx *hctx; |
484b4061 JA |
2153 | struct blk_mq_ctx *ctx; |
2154 | LIST_HEAD(tmp); | |
2155 | ||
9467f859 | 2156 | hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead); |
e57690fe | 2157 | ctx = __blk_mq_get_ctx(hctx->queue, cpu); |
484b4061 JA |
2158 | |
2159 | spin_lock(&ctx->lock); | |
2160 | if (!list_empty(&ctx->rq_list)) { | |
2161 | list_splice_init(&ctx->rq_list, &tmp); | |
2162 | blk_mq_hctx_clear_pending(hctx, ctx); | |
2163 | } | |
2164 | spin_unlock(&ctx->lock); | |
2165 | ||
2166 | if (list_empty(&tmp)) | |
9467f859 | 2167 | return 0; |
484b4061 | 2168 | |
e57690fe JA |
2169 | spin_lock(&hctx->lock); |
2170 | list_splice_tail_init(&tmp, &hctx->dispatch); | |
2171 | spin_unlock(&hctx->lock); | |
484b4061 JA |
2172 | |
2173 | blk_mq_run_hw_queue(hctx, true); | |
9467f859 | 2174 | return 0; |
484b4061 JA |
2175 | } |
2176 | ||
9467f859 | 2177 | static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx) |
484b4061 | 2178 | { |
9467f859 TG |
2179 | cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD, |
2180 | &hctx->cpuhp_dead); | |
484b4061 JA |
2181 | } |
2182 | ||
c3b4afca | 2183 | /* hctx->ctxs will be freed in queue's release handler */ |
08e98fc6 ML |
2184 | static void blk_mq_exit_hctx(struct request_queue *q, |
2185 | struct blk_mq_tag_set *set, | |
2186 | struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx) | |
2187 | { | |
9c1051aa OS |
2188 | blk_mq_debugfs_unregister_hctx(hctx); |
2189 | ||
8ab0b7dc ML |
2190 | if (blk_mq_hw_queue_mapped(hctx)) |
2191 | blk_mq_tag_idle(hctx); | |
08e98fc6 | 2192 | |
f70ced09 | 2193 | if (set->ops->exit_request) |
d6296d39 | 2194 | set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx); |
f70ced09 | 2195 | |
93252632 OS |
2196 | blk_mq_sched_exit_hctx(q, hctx, hctx_idx); |
2197 | ||
08e98fc6 ML |
2198 | if (set->ops->exit_hctx) |
2199 | set->ops->exit_hctx(hctx, hctx_idx); | |
2200 | ||
6a83e74d | 2201 | if (hctx->flags & BLK_MQ_F_BLOCKING) |
05707b64 | 2202 | cleanup_srcu_struct(hctx->srcu); |
6a83e74d | 2203 | |
9467f859 | 2204 | blk_mq_remove_cpuhp(hctx); |
f70ced09 | 2205 | blk_free_flush_queue(hctx->fq); |
88459642 | 2206 | sbitmap_free(&hctx->ctx_map); |
08e98fc6 ML |
2207 | } |
2208 | ||
624dbe47 ML |
2209 | static void blk_mq_exit_hw_queues(struct request_queue *q, |
2210 | struct blk_mq_tag_set *set, int nr_queue) | |
2211 | { | |
2212 | struct blk_mq_hw_ctx *hctx; | |
2213 | unsigned int i; | |
2214 | ||
2215 | queue_for_each_hw_ctx(q, hctx, i) { | |
2216 | if (i == nr_queue) | |
2217 | break; | |
08e98fc6 | 2218 | blk_mq_exit_hctx(q, set, hctx, i); |
624dbe47 | 2219 | } |
624dbe47 ML |
2220 | } |
2221 | ||
08e98fc6 ML |
2222 | static int blk_mq_init_hctx(struct request_queue *q, |
2223 | struct blk_mq_tag_set *set, | |
2224 | struct blk_mq_hw_ctx *hctx, unsigned hctx_idx) | |
320ae51f | 2225 | { |
08e98fc6 ML |
2226 | int node; |
2227 | ||
2228 | node = hctx->numa_node; | |
2229 | if (node == NUMA_NO_NODE) | |
2230 | node = hctx->numa_node = set->numa_node; | |
2231 | ||
9f993737 | 2232 | INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn); |
08e98fc6 ML |
2233 | spin_lock_init(&hctx->lock); |
2234 | INIT_LIST_HEAD(&hctx->dispatch); | |
2235 | hctx->queue = q; | |
2404e607 | 2236 | hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED; |
08e98fc6 | 2237 | |
9467f859 | 2238 | cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead); |
08e98fc6 ML |
2239 | |
2240 | hctx->tags = set->tags[hctx_idx]; | |
320ae51f JA |
2241 | |
2242 | /* | |
08e98fc6 ML |
2243 | * Allocate space for all possible cpus to avoid allocation at |
2244 | * runtime | |
320ae51f | 2245 | */ |
d904bfa7 | 2246 | hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *), |
08e98fc6 ML |
2247 | GFP_KERNEL, node); |
2248 | if (!hctx->ctxs) | |
2249 | goto unregister_cpu_notifier; | |
320ae51f | 2250 | |
88459642 OS |
2251 | if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8), GFP_KERNEL, |
2252 | node)) | |
08e98fc6 | 2253 | goto free_ctxs; |
320ae51f | 2254 | |
08e98fc6 | 2255 | hctx->nr_ctx = 0; |
320ae51f | 2256 | |
eb619fdb JA |
2257 | init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake); |
2258 | INIT_LIST_HEAD(&hctx->dispatch_wait.entry); | |
2259 | ||
08e98fc6 ML |
2260 | if (set->ops->init_hctx && |
2261 | set->ops->init_hctx(hctx, set->driver_data, hctx_idx)) | |
2262 | goto free_bitmap; | |
320ae51f | 2263 | |
93252632 OS |
2264 | if (blk_mq_sched_init_hctx(q, hctx, hctx_idx)) |
2265 | goto exit_hctx; | |
2266 | ||
f70ced09 ML |
2267 | hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size); |
2268 | if (!hctx->fq) | |
93252632 | 2269 | goto sched_exit_hctx; |
320ae51f | 2270 | |
1d9bd516 | 2271 | if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx, node)) |
f70ced09 | 2272 | goto free_fq; |
320ae51f | 2273 | |
6a83e74d | 2274 | if (hctx->flags & BLK_MQ_F_BLOCKING) |
05707b64 | 2275 | init_srcu_struct(hctx->srcu); |
6a83e74d | 2276 | |
9c1051aa OS |
2277 | blk_mq_debugfs_register_hctx(q, hctx); |
2278 | ||
08e98fc6 | 2279 | return 0; |
320ae51f | 2280 | |
f70ced09 ML |
2281 | free_fq: |
2282 | kfree(hctx->fq); | |
93252632 OS |
2283 | sched_exit_hctx: |
2284 | blk_mq_sched_exit_hctx(q, hctx, hctx_idx); | |
f70ced09 ML |
2285 | exit_hctx: |
2286 | if (set->ops->exit_hctx) | |
2287 | set->ops->exit_hctx(hctx, hctx_idx); | |
08e98fc6 | 2288 | free_bitmap: |
88459642 | 2289 | sbitmap_free(&hctx->ctx_map); |
08e98fc6 ML |
2290 | free_ctxs: |
2291 | kfree(hctx->ctxs); | |
2292 | unregister_cpu_notifier: | |
9467f859 | 2293 | blk_mq_remove_cpuhp(hctx); |
08e98fc6 ML |
2294 | return -1; |
2295 | } | |
320ae51f | 2296 | |
320ae51f JA |
2297 | static void blk_mq_init_cpu_queues(struct request_queue *q, |
2298 | unsigned int nr_hw_queues) | |
2299 | { | |
2300 | unsigned int i; | |
2301 | ||
2302 | for_each_possible_cpu(i) { | |
2303 | struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i); | |
2304 | struct blk_mq_hw_ctx *hctx; | |
2305 | ||
320ae51f JA |
2306 | __ctx->cpu = i; |
2307 | spin_lock_init(&__ctx->lock); | |
2308 | INIT_LIST_HEAD(&__ctx->rq_list); | |
2309 | __ctx->queue = q; | |
2310 | ||
320ae51f JA |
2311 | /* |
2312 | * Set local node, IFF we have more than one hw queue. If | |
2313 | * not, we remain on the home node of the device | |
2314 | */ | |
20e4d813 | 2315 | hctx = blk_mq_map_queue(q, i); |
320ae51f | 2316 | if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE) |
bffed457 | 2317 | hctx->numa_node = local_memory_node(cpu_to_node(i)); |
320ae51f JA |
2318 | } |
2319 | } | |
2320 | ||
cc71a6f4 JA |
2321 | static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, int hctx_idx) |
2322 | { | |
2323 | int ret = 0; | |
2324 | ||
2325 | set->tags[hctx_idx] = blk_mq_alloc_rq_map(set, hctx_idx, | |
2326 | set->queue_depth, set->reserved_tags); | |
2327 | if (!set->tags[hctx_idx]) | |
2328 | return false; | |
2329 | ||
2330 | ret = blk_mq_alloc_rqs(set, set->tags[hctx_idx], hctx_idx, | |
2331 | set->queue_depth); | |
2332 | if (!ret) | |
2333 | return true; | |
2334 | ||
2335 | blk_mq_free_rq_map(set->tags[hctx_idx]); | |
2336 | set->tags[hctx_idx] = NULL; | |
2337 | return false; | |
2338 | } | |
2339 | ||
2340 | static void blk_mq_free_map_and_requests(struct blk_mq_tag_set *set, | |
2341 | unsigned int hctx_idx) | |
2342 | { | |
bd166ef1 JA |
2343 | if (set->tags[hctx_idx]) { |
2344 | blk_mq_free_rqs(set, set->tags[hctx_idx], hctx_idx); | |
2345 | blk_mq_free_rq_map(set->tags[hctx_idx]); | |
2346 | set->tags[hctx_idx] = NULL; | |
2347 | } | |
cc71a6f4 JA |
2348 | } |
2349 | ||
4b855ad3 | 2350 | static void blk_mq_map_swqueue(struct request_queue *q) |
320ae51f | 2351 | { |
d1b1cea1 | 2352 | unsigned int i, hctx_idx; |
320ae51f JA |
2353 | struct blk_mq_hw_ctx *hctx; |
2354 | struct blk_mq_ctx *ctx; | |
2a34c087 | 2355 | struct blk_mq_tag_set *set = q->tag_set; |
320ae51f | 2356 | |
60de074b AM |
2357 | /* |
2358 | * Avoid others reading imcomplete hctx->cpumask through sysfs | |
2359 | */ | |
2360 | mutex_lock(&q->sysfs_lock); | |
2361 | ||
320ae51f | 2362 | queue_for_each_hw_ctx(q, hctx, i) { |
e4043dcf | 2363 | cpumask_clear(hctx->cpumask); |
320ae51f JA |
2364 | hctx->nr_ctx = 0; |
2365 | } | |
2366 | ||
2367 | /* | |
4b855ad3 CH |
2368 | * Map software to hardware queues. |
2369 | * | |
2370 | * If the cpu isn't present, the cpu is mapped to first hctx. | |
320ae51f | 2371 | */ |
20e4d813 | 2372 | for_each_possible_cpu(i) { |
d1b1cea1 GKB |
2373 | hctx_idx = q->mq_map[i]; |
2374 | /* unmapped hw queue can be remapped after CPU topo changed */ | |
cc71a6f4 JA |
2375 | if (!set->tags[hctx_idx] && |
2376 | !__blk_mq_alloc_rq_map(set, hctx_idx)) { | |
d1b1cea1 GKB |
2377 | /* |
2378 | * If tags initialization fail for some hctx, | |
2379 | * that hctx won't be brought online. In this | |
2380 | * case, remap the current ctx to hctx[0] which | |
2381 | * is guaranteed to always have tags allocated | |
2382 | */ | |
cc71a6f4 | 2383 | q->mq_map[i] = 0; |
d1b1cea1 GKB |
2384 | } |
2385 | ||
897bb0c7 | 2386 | ctx = per_cpu_ptr(q->queue_ctx, i); |
7d7e0f90 | 2387 | hctx = blk_mq_map_queue(q, i); |
868f2f0b | 2388 | |
e4043dcf | 2389 | cpumask_set_cpu(i, hctx->cpumask); |
320ae51f JA |
2390 | ctx->index_hw = hctx->nr_ctx; |
2391 | hctx->ctxs[hctx->nr_ctx++] = ctx; | |
2392 | } | |
506e931f | 2393 | |
60de074b AM |
2394 | mutex_unlock(&q->sysfs_lock); |
2395 | ||
506e931f | 2396 | queue_for_each_hw_ctx(q, hctx, i) { |
484b4061 | 2397 | /* |
a68aafa5 JA |
2398 | * If no software queues are mapped to this hardware queue, |
2399 | * disable it and free the request entries. | |
484b4061 JA |
2400 | */ |
2401 | if (!hctx->nr_ctx) { | |
d1b1cea1 GKB |
2402 | /* Never unmap queue 0. We need it as a |
2403 | * fallback in case of a new remap fails | |
2404 | * allocation | |
2405 | */ | |
cc71a6f4 JA |
2406 | if (i && set->tags[i]) |
2407 | blk_mq_free_map_and_requests(set, i); | |
2408 | ||
2a34c087 | 2409 | hctx->tags = NULL; |
484b4061 JA |
2410 | continue; |
2411 | } | |
2412 | ||
2a34c087 ML |
2413 | hctx->tags = set->tags[i]; |
2414 | WARN_ON(!hctx->tags); | |
2415 | ||
889fa31f CY |
2416 | /* |
2417 | * Set the map size to the number of mapped software queues. | |
2418 | * This is more accurate and more efficient than looping | |
2419 | * over all possibly mapped software queues. | |
2420 | */ | |
88459642 | 2421 | sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx); |
889fa31f | 2422 | |
484b4061 JA |
2423 | /* |
2424 | * Initialize batch roundrobin counts | |
2425 | */ | |
20e4d813 CH |
2426 | hctx->next_cpu = cpumask_first_and(hctx->cpumask, |
2427 | cpu_online_mask); | |
506e931f JA |
2428 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; |
2429 | } | |
320ae51f JA |
2430 | } |
2431 | ||
8e8320c9 JA |
2432 | /* |
2433 | * Caller needs to ensure that we're either frozen/quiesced, or that | |
2434 | * the queue isn't live yet. | |
2435 | */ | |
2404e607 | 2436 | static void queue_set_hctx_shared(struct request_queue *q, bool shared) |
0d2602ca JA |
2437 | { |
2438 | struct blk_mq_hw_ctx *hctx; | |
0d2602ca JA |
2439 | int i; |
2440 | ||
2404e607 | 2441 | queue_for_each_hw_ctx(q, hctx, i) { |
8e8320c9 JA |
2442 | if (shared) { |
2443 | if (test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state)) | |
2444 | atomic_inc(&q->shared_hctx_restart); | |
2404e607 | 2445 | hctx->flags |= BLK_MQ_F_TAG_SHARED; |
8e8320c9 JA |
2446 | } else { |
2447 | if (test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state)) | |
2448 | atomic_dec(&q->shared_hctx_restart); | |
2404e607 | 2449 | hctx->flags &= ~BLK_MQ_F_TAG_SHARED; |
8e8320c9 | 2450 | } |
2404e607 JM |
2451 | } |
2452 | } | |
2453 | ||
8e8320c9 JA |
2454 | static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set, |
2455 | bool shared) | |
2404e607 JM |
2456 | { |
2457 | struct request_queue *q; | |
0d2602ca | 2458 | |
705cda97 BVA |
2459 | lockdep_assert_held(&set->tag_list_lock); |
2460 | ||
0d2602ca JA |
2461 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
2462 | blk_mq_freeze_queue(q); | |
2404e607 | 2463 | queue_set_hctx_shared(q, shared); |
0d2602ca JA |
2464 | blk_mq_unfreeze_queue(q); |
2465 | } | |
2466 | } | |
2467 | ||
2468 | static void blk_mq_del_queue_tag_set(struct request_queue *q) | |
2469 | { | |
2470 | struct blk_mq_tag_set *set = q->tag_set; | |
2471 | ||
0d2602ca | 2472 | mutex_lock(&set->tag_list_lock); |
705cda97 BVA |
2473 | list_del_rcu(&q->tag_set_list); |
2474 | INIT_LIST_HEAD(&q->tag_set_list); | |
2404e607 JM |
2475 | if (list_is_singular(&set->tag_list)) { |
2476 | /* just transitioned to unshared */ | |
2477 | set->flags &= ~BLK_MQ_F_TAG_SHARED; | |
2478 | /* update existing queue */ | |
2479 | blk_mq_update_tag_set_depth(set, false); | |
2480 | } | |
0d2602ca | 2481 | mutex_unlock(&set->tag_list_lock); |
705cda97 BVA |
2482 | |
2483 | synchronize_rcu(); | |
0d2602ca JA |
2484 | } |
2485 | ||
2486 | static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set, | |
2487 | struct request_queue *q) | |
2488 | { | |
2489 | q->tag_set = set; | |
2490 | ||
2491 | mutex_lock(&set->tag_list_lock); | |
2404e607 | 2492 | |
ff821d27 JA |
2493 | /* |
2494 | * Check to see if we're transitioning to shared (from 1 to 2 queues). | |
2495 | */ | |
2496 | if (!list_empty(&set->tag_list) && | |
2497 | !(set->flags & BLK_MQ_F_TAG_SHARED)) { | |
2404e607 JM |
2498 | set->flags |= BLK_MQ_F_TAG_SHARED; |
2499 | /* update existing queue */ | |
2500 | blk_mq_update_tag_set_depth(set, true); | |
2501 | } | |
2502 | if (set->flags & BLK_MQ_F_TAG_SHARED) | |
2503 | queue_set_hctx_shared(q, true); | |
705cda97 | 2504 | list_add_tail_rcu(&q->tag_set_list, &set->tag_list); |
2404e607 | 2505 | |
0d2602ca JA |
2506 | mutex_unlock(&set->tag_list_lock); |
2507 | } | |
2508 | ||
e09aae7e ML |
2509 | /* |
2510 | * It is the actual release handler for mq, but we do it from | |
2511 | * request queue's release handler for avoiding use-after-free | |
2512 | * and headache because q->mq_kobj shouldn't have been introduced, | |
2513 | * but we can't group ctx/kctx kobj without it. | |
2514 | */ | |
2515 | void blk_mq_release(struct request_queue *q) | |
2516 | { | |
2517 | struct blk_mq_hw_ctx *hctx; | |
2518 | unsigned int i; | |
2519 | ||
2520 | /* hctx kobj stays in hctx */ | |
c3b4afca ML |
2521 | queue_for_each_hw_ctx(q, hctx, i) { |
2522 | if (!hctx) | |
2523 | continue; | |
6c8b232e | 2524 | kobject_put(&hctx->kobj); |
c3b4afca | 2525 | } |
e09aae7e | 2526 | |
a723bab3 AM |
2527 | q->mq_map = NULL; |
2528 | ||
e09aae7e ML |
2529 | kfree(q->queue_hw_ctx); |
2530 | ||
7ea5fe31 ML |
2531 | /* |
2532 | * release .mq_kobj and sw queue's kobject now because | |
2533 | * both share lifetime with request queue. | |
2534 | */ | |
2535 | blk_mq_sysfs_deinit(q); | |
2536 | ||
e09aae7e ML |
2537 | free_percpu(q->queue_ctx); |
2538 | } | |
2539 | ||
24d2f903 | 2540 | struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set) |
b62c21b7 MS |
2541 | { |
2542 | struct request_queue *uninit_q, *q; | |
2543 | ||
2544 | uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node); | |
2545 | if (!uninit_q) | |
2546 | return ERR_PTR(-ENOMEM); | |
2547 | ||
2548 | q = blk_mq_init_allocated_queue(set, uninit_q); | |
2549 | if (IS_ERR(q)) | |
2550 | blk_cleanup_queue(uninit_q); | |
2551 | ||
2552 | return q; | |
2553 | } | |
2554 | EXPORT_SYMBOL(blk_mq_init_queue); | |
2555 | ||
07319678 BVA |
2556 | static int blk_mq_hw_ctx_size(struct blk_mq_tag_set *tag_set) |
2557 | { | |
2558 | int hw_ctx_size = sizeof(struct blk_mq_hw_ctx); | |
2559 | ||
05707b64 | 2560 | BUILD_BUG_ON(ALIGN(offsetof(struct blk_mq_hw_ctx, srcu), |
07319678 BVA |
2561 | __alignof__(struct blk_mq_hw_ctx)) != |
2562 | sizeof(struct blk_mq_hw_ctx)); | |
2563 | ||
2564 | if (tag_set->flags & BLK_MQ_F_BLOCKING) | |
2565 | hw_ctx_size += sizeof(struct srcu_struct); | |
2566 | ||
2567 | return hw_ctx_size; | |
2568 | } | |
2569 | ||
868f2f0b KB |
2570 | static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set, |
2571 | struct request_queue *q) | |
320ae51f | 2572 | { |
868f2f0b KB |
2573 | int i, j; |
2574 | struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx; | |
f14bbe77 | 2575 | |
868f2f0b | 2576 | blk_mq_sysfs_unregister(q); |
fb350e0a ML |
2577 | |
2578 | /* protect against switching io scheduler */ | |
2579 | mutex_lock(&q->sysfs_lock); | |
24d2f903 | 2580 | for (i = 0; i < set->nr_hw_queues; i++) { |
868f2f0b | 2581 | int node; |
f14bbe77 | 2582 | |
868f2f0b KB |
2583 | if (hctxs[i]) |
2584 | continue; | |
2585 | ||
2586 | node = blk_mq_hw_queue_to_node(q->mq_map, i); | |
07319678 | 2587 | hctxs[i] = kzalloc_node(blk_mq_hw_ctx_size(set), |
cdef54dd | 2588 | GFP_KERNEL, node); |
320ae51f | 2589 | if (!hctxs[i]) |
868f2f0b | 2590 | break; |
320ae51f | 2591 | |
a86073e4 | 2592 | if (!zalloc_cpumask_var_node(&hctxs[i]->cpumask, GFP_KERNEL, |
868f2f0b KB |
2593 | node)) { |
2594 | kfree(hctxs[i]); | |
2595 | hctxs[i] = NULL; | |
2596 | break; | |
2597 | } | |
e4043dcf | 2598 | |
0d2602ca | 2599 | atomic_set(&hctxs[i]->nr_active, 0); |
f14bbe77 | 2600 | hctxs[i]->numa_node = node; |
320ae51f | 2601 | hctxs[i]->queue_num = i; |
868f2f0b KB |
2602 | |
2603 | if (blk_mq_init_hctx(q, set, hctxs[i], i)) { | |
2604 | free_cpumask_var(hctxs[i]->cpumask); | |
2605 | kfree(hctxs[i]); | |
2606 | hctxs[i] = NULL; | |
2607 | break; | |
2608 | } | |
2609 | blk_mq_hctx_kobj_init(hctxs[i]); | |
320ae51f | 2610 | } |
868f2f0b KB |
2611 | for (j = i; j < q->nr_hw_queues; j++) { |
2612 | struct blk_mq_hw_ctx *hctx = hctxs[j]; | |
2613 | ||
2614 | if (hctx) { | |
cc71a6f4 JA |
2615 | if (hctx->tags) |
2616 | blk_mq_free_map_and_requests(set, j); | |
868f2f0b | 2617 | blk_mq_exit_hctx(q, set, hctx, j); |
868f2f0b | 2618 | kobject_put(&hctx->kobj); |
868f2f0b KB |
2619 | hctxs[j] = NULL; |
2620 | ||
2621 | } | |
2622 | } | |
2623 | q->nr_hw_queues = i; | |
fb350e0a | 2624 | mutex_unlock(&q->sysfs_lock); |
868f2f0b KB |
2625 | blk_mq_sysfs_register(q); |
2626 | } | |
2627 | ||
2628 | struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set, | |
2629 | struct request_queue *q) | |
2630 | { | |
66841672 ML |
2631 | /* mark the queue as mq asap */ |
2632 | q->mq_ops = set->ops; | |
2633 | ||
34dbad5d | 2634 | q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn, |
720b8ccc SB |
2635 | blk_mq_poll_stats_bkt, |
2636 | BLK_MQ_POLL_STATS_BKTS, q); | |
34dbad5d OS |
2637 | if (!q->poll_cb) |
2638 | goto err_exit; | |
2639 | ||
868f2f0b KB |
2640 | q->queue_ctx = alloc_percpu(struct blk_mq_ctx); |
2641 | if (!q->queue_ctx) | |
c7de5726 | 2642 | goto err_exit; |
868f2f0b | 2643 | |
737f98cf ML |
2644 | /* init q->mq_kobj and sw queues' kobjects */ |
2645 | blk_mq_sysfs_init(q); | |
2646 | ||
868f2f0b KB |
2647 | q->queue_hw_ctx = kzalloc_node(nr_cpu_ids * sizeof(*(q->queue_hw_ctx)), |
2648 | GFP_KERNEL, set->numa_node); | |
2649 | if (!q->queue_hw_ctx) | |
2650 | goto err_percpu; | |
2651 | ||
bdd17e75 | 2652 | q->mq_map = set->mq_map; |
868f2f0b KB |
2653 | |
2654 | blk_mq_realloc_hw_ctxs(set, q); | |
2655 | if (!q->nr_hw_queues) | |
2656 | goto err_hctxs; | |
320ae51f | 2657 | |
287922eb | 2658 | INIT_WORK(&q->timeout_work, blk_mq_timeout_work); |
e56f698b | 2659 | blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ); |
320ae51f JA |
2660 | |
2661 | q->nr_queues = nr_cpu_ids; | |
320ae51f | 2662 | |
94eddfbe | 2663 | q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT; |
320ae51f | 2664 | |
05f1dd53 JA |
2665 | if (!(set->flags & BLK_MQ_F_SG_MERGE)) |
2666 | q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE; | |
2667 | ||
1be036e9 CH |
2668 | q->sg_reserved_size = INT_MAX; |
2669 | ||
2849450a | 2670 | INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work); |
6fca6a61 CH |
2671 | INIT_LIST_HEAD(&q->requeue_list); |
2672 | spin_lock_init(&q->requeue_lock); | |
2673 | ||
254d259d | 2674 | blk_queue_make_request(q, blk_mq_make_request); |
ea435e1b CH |
2675 | if (q->mq_ops->poll) |
2676 | q->poll_fn = blk_mq_poll; | |
07068d5b | 2677 | |
eba71768 JA |
2678 | /* |
2679 | * Do this after blk_queue_make_request() overrides it... | |
2680 | */ | |
2681 | q->nr_requests = set->queue_depth; | |
2682 | ||
64f1c21e JA |
2683 | /* |
2684 | * Default to classic polling | |
2685 | */ | |
2686 | q->poll_nsec = -1; | |
2687 | ||
24d2f903 CH |
2688 | if (set->ops->complete) |
2689 | blk_queue_softirq_done(q, set->ops->complete); | |
30a91cb4 | 2690 | |
24d2f903 | 2691 | blk_mq_init_cpu_queues(q, set->nr_hw_queues); |
0d2602ca | 2692 | blk_mq_add_queue_tag_set(set, q); |
4b855ad3 | 2693 | blk_mq_map_swqueue(q); |
4593fdbe | 2694 | |
d3484991 JA |
2695 | if (!(set->flags & BLK_MQ_F_NO_SCHED)) { |
2696 | int ret; | |
2697 | ||
2698 | ret = blk_mq_sched_init(q); | |
2699 | if (ret) | |
2700 | return ERR_PTR(ret); | |
2701 | } | |
2702 | ||
320ae51f | 2703 | return q; |
18741986 | 2704 | |
320ae51f | 2705 | err_hctxs: |
868f2f0b | 2706 | kfree(q->queue_hw_ctx); |
320ae51f | 2707 | err_percpu: |
868f2f0b | 2708 | free_percpu(q->queue_ctx); |
c7de5726 ML |
2709 | err_exit: |
2710 | q->mq_ops = NULL; | |
320ae51f JA |
2711 | return ERR_PTR(-ENOMEM); |
2712 | } | |
b62c21b7 | 2713 | EXPORT_SYMBOL(blk_mq_init_allocated_queue); |
320ae51f JA |
2714 | |
2715 | void blk_mq_free_queue(struct request_queue *q) | |
2716 | { | |
624dbe47 | 2717 | struct blk_mq_tag_set *set = q->tag_set; |
320ae51f | 2718 | |
0d2602ca | 2719 | blk_mq_del_queue_tag_set(q); |
624dbe47 | 2720 | blk_mq_exit_hw_queues(q, set, set->nr_hw_queues); |
320ae51f | 2721 | } |
320ae51f JA |
2722 | |
2723 | /* Basically redo blk_mq_init_queue with queue frozen */ | |
4b855ad3 | 2724 | static void blk_mq_queue_reinit(struct request_queue *q) |
320ae51f | 2725 | { |
4ecd4fef | 2726 | WARN_ON_ONCE(!atomic_read(&q->mq_freeze_depth)); |
320ae51f | 2727 | |
9c1051aa | 2728 | blk_mq_debugfs_unregister_hctxs(q); |
67aec14c JA |
2729 | blk_mq_sysfs_unregister(q); |
2730 | ||
320ae51f JA |
2731 | /* |
2732 | * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe | |
ff821d27 JA |
2733 | * we should change hctx numa_node according to the new topology (this |
2734 | * involves freeing and re-allocating memory, worth doing?) | |
320ae51f | 2735 | */ |
4b855ad3 | 2736 | blk_mq_map_swqueue(q); |
320ae51f | 2737 | |
67aec14c | 2738 | blk_mq_sysfs_register(q); |
9c1051aa | 2739 | blk_mq_debugfs_register_hctxs(q); |
320ae51f JA |
2740 | } |
2741 | ||
a5164405 JA |
2742 | static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set) |
2743 | { | |
2744 | int i; | |
2745 | ||
cc71a6f4 JA |
2746 | for (i = 0; i < set->nr_hw_queues; i++) |
2747 | if (!__blk_mq_alloc_rq_map(set, i)) | |
a5164405 | 2748 | goto out_unwind; |
a5164405 JA |
2749 | |
2750 | return 0; | |
2751 | ||
2752 | out_unwind: | |
2753 | while (--i >= 0) | |
cc71a6f4 | 2754 | blk_mq_free_rq_map(set->tags[i]); |
a5164405 | 2755 | |
a5164405 JA |
2756 | return -ENOMEM; |
2757 | } | |
2758 | ||
2759 | /* | |
2760 | * Allocate the request maps associated with this tag_set. Note that this | |
2761 | * may reduce the depth asked for, if memory is tight. set->queue_depth | |
2762 | * will be updated to reflect the allocated depth. | |
2763 | */ | |
2764 | static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set) | |
2765 | { | |
2766 | unsigned int depth; | |
2767 | int err; | |
2768 | ||
2769 | depth = set->queue_depth; | |
2770 | do { | |
2771 | err = __blk_mq_alloc_rq_maps(set); | |
2772 | if (!err) | |
2773 | break; | |
2774 | ||
2775 | set->queue_depth >>= 1; | |
2776 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) { | |
2777 | err = -ENOMEM; | |
2778 | break; | |
2779 | } | |
2780 | } while (set->queue_depth); | |
2781 | ||
2782 | if (!set->queue_depth || err) { | |
2783 | pr_err("blk-mq: failed to allocate request map\n"); | |
2784 | return -ENOMEM; | |
2785 | } | |
2786 | ||
2787 | if (depth != set->queue_depth) | |
2788 | pr_info("blk-mq: reduced tag depth (%u -> %u)\n", | |
2789 | depth, set->queue_depth); | |
2790 | ||
2791 | return 0; | |
2792 | } | |
2793 | ||
ebe8bddb OS |
2794 | static int blk_mq_update_queue_map(struct blk_mq_tag_set *set) |
2795 | { | |
7d4901a9 ML |
2796 | if (set->ops->map_queues) { |
2797 | int cpu; | |
2798 | /* | |
2799 | * transport .map_queues is usually done in the following | |
2800 | * way: | |
2801 | * | |
2802 | * for (queue = 0; queue < set->nr_hw_queues; queue++) { | |
2803 | * mask = get_cpu_mask(queue) | |
2804 | * for_each_cpu(cpu, mask) | |
2805 | * set->mq_map[cpu] = queue; | |
2806 | * } | |
2807 | * | |
2808 | * When we need to remap, the table has to be cleared for | |
2809 | * killing stale mapping since one CPU may not be mapped | |
2810 | * to any hw queue. | |
2811 | */ | |
2812 | for_each_possible_cpu(cpu) | |
2813 | set->mq_map[cpu] = 0; | |
2814 | ||
ebe8bddb | 2815 | return set->ops->map_queues(set); |
7d4901a9 | 2816 | } else |
ebe8bddb OS |
2817 | return blk_mq_map_queues(set); |
2818 | } | |
2819 | ||
a4391c64 JA |
2820 | /* |
2821 | * Alloc a tag set to be associated with one or more request queues. | |
2822 | * May fail with EINVAL for various error conditions. May adjust the | |
2823 | * requested depth down, if if it too large. In that case, the set | |
2824 | * value will be stored in set->queue_depth. | |
2825 | */ | |
24d2f903 CH |
2826 | int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set) |
2827 | { | |
da695ba2 CH |
2828 | int ret; |
2829 | ||
205fb5f5 BVA |
2830 | BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS); |
2831 | ||
24d2f903 CH |
2832 | if (!set->nr_hw_queues) |
2833 | return -EINVAL; | |
a4391c64 | 2834 | if (!set->queue_depth) |
24d2f903 CH |
2835 | return -EINVAL; |
2836 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) | |
2837 | return -EINVAL; | |
2838 | ||
7d7e0f90 | 2839 | if (!set->ops->queue_rq) |
24d2f903 CH |
2840 | return -EINVAL; |
2841 | ||
de148297 ML |
2842 | if (!set->ops->get_budget ^ !set->ops->put_budget) |
2843 | return -EINVAL; | |
2844 | ||
a4391c64 JA |
2845 | if (set->queue_depth > BLK_MQ_MAX_DEPTH) { |
2846 | pr_info("blk-mq: reduced tag depth to %u\n", | |
2847 | BLK_MQ_MAX_DEPTH); | |
2848 | set->queue_depth = BLK_MQ_MAX_DEPTH; | |
2849 | } | |
24d2f903 | 2850 | |
6637fadf SL |
2851 | /* |
2852 | * If a crashdump is active, then we are potentially in a very | |
2853 | * memory constrained environment. Limit us to 1 queue and | |
2854 | * 64 tags to prevent using too much memory. | |
2855 | */ | |
2856 | if (is_kdump_kernel()) { | |
2857 | set->nr_hw_queues = 1; | |
2858 | set->queue_depth = min(64U, set->queue_depth); | |
2859 | } | |
868f2f0b KB |
2860 | /* |
2861 | * There is no use for more h/w queues than cpus. | |
2862 | */ | |
2863 | if (set->nr_hw_queues > nr_cpu_ids) | |
2864 | set->nr_hw_queues = nr_cpu_ids; | |
6637fadf | 2865 | |
868f2f0b | 2866 | set->tags = kzalloc_node(nr_cpu_ids * sizeof(struct blk_mq_tags *), |
24d2f903 CH |
2867 | GFP_KERNEL, set->numa_node); |
2868 | if (!set->tags) | |
a5164405 | 2869 | return -ENOMEM; |
24d2f903 | 2870 | |
da695ba2 CH |
2871 | ret = -ENOMEM; |
2872 | set->mq_map = kzalloc_node(sizeof(*set->mq_map) * nr_cpu_ids, | |
2873 | GFP_KERNEL, set->numa_node); | |
bdd17e75 CH |
2874 | if (!set->mq_map) |
2875 | goto out_free_tags; | |
2876 | ||
ebe8bddb | 2877 | ret = blk_mq_update_queue_map(set); |
da695ba2 CH |
2878 | if (ret) |
2879 | goto out_free_mq_map; | |
2880 | ||
2881 | ret = blk_mq_alloc_rq_maps(set); | |
2882 | if (ret) | |
bdd17e75 | 2883 | goto out_free_mq_map; |
24d2f903 | 2884 | |
0d2602ca JA |
2885 | mutex_init(&set->tag_list_lock); |
2886 | INIT_LIST_HEAD(&set->tag_list); | |
2887 | ||
24d2f903 | 2888 | return 0; |
bdd17e75 CH |
2889 | |
2890 | out_free_mq_map: | |
2891 | kfree(set->mq_map); | |
2892 | set->mq_map = NULL; | |
2893 | out_free_tags: | |
5676e7b6 RE |
2894 | kfree(set->tags); |
2895 | set->tags = NULL; | |
da695ba2 | 2896 | return ret; |
24d2f903 CH |
2897 | } |
2898 | EXPORT_SYMBOL(blk_mq_alloc_tag_set); | |
2899 | ||
2900 | void blk_mq_free_tag_set(struct blk_mq_tag_set *set) | |
2901 | { | |
2902 | int i; | |
2903 | ||
cc71a6f4 JA |
2904 | for (i = 0; i < nr_cpu_ids; i++) |
2905 | blk_mq_free_map_and_requests(set, i); | |
484b4061 | 2906 | |
bdd17e75 CH |
2907 | kfree(set->mq_map); |
2908 | set->mq_map = NULL; | |
2909 | ||
981bd189 | 2910 | kfree(set->tags); |
5676e7b6 | 2911 | set->tags = NULL; |
24d2f903 CH |
2912 | } |
2913 | EXPORT_SYMBOL(blk_mq_free_tag_set); | |
2914 | ||
e3a2b3f9 JA |
2915 | int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr) |
2916 | { | |
2917 | struct blk_mq_tag_set *set = q->tag_set; | |
2918 | struct blk_mq_hw_ctx *hctx; | |
2919 | int i, ret; | |
2920 | ||
bd166ef1 | 2921 | if (!set) |
e3a2b3f9 JA |
2922 | return -EINVAL; |
2923 | ||
70f36b60 | 2924 | blk_mq_freeze_queue(q); |
24f5a90f | 2925 | blk_mq_quiesce_queue(q); |
70f36b60 | 2926 | |
e3a2b3f9 JA |
2927 | ret = 0; |
2928 | queue_for_each_hw_ctx(q, hctx, i) { | |
e9137d4b KB |
2929 | if (!hctx->tags) |
2930 | continue; | |
bd166ef1 JA |
2931 | /* |
2932 | * If we're using an MQ scheduler, just update the scheduler | |
2933 | * queue depth. This is similar to what the old code would do. | |
2934 | */ | |
70f36b60 | 2935 | if (!hctx->sched_tags) { |
c2e82a23 | 2936 | ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr, |
70f36b60 JA |
2937 | false); |
2938 | } else { | |
2939 | ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags, | |
2940 | nr, true); | |
2941 | } | |
e3a2b3f9 JA |
2942 | if (ret) |
2943 | break; | |
2944 | } | |
2945 | ||
2946 | if (!ret) | |
2947 | q->nr_requests = nr; | |
2948 | ||
24f5a90f | 2949 | blk_mq_unquiesce_queue(q); |
70f36b60 | 2950 | blk_mq_unfreeze_queue(q); |
70f36b60 | 2951 | |
e3a2b3f9 JA |
2952 | return ret; |
2953 | } | |
2954 | ||
e4dc2b32 KB |
2955 | static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, |
2956 | int nr_hw_queues) | |
868f2f0b KB |
2957 | { |
2958 | struct request_queue *q; | |
2959 | ||
705cda97 BVA |
2960 | lockdep_assert_held(&set->tag_list_lock); |
2961 | ||
868f2f0b KB |
2962 | if (nr_hw_queues > nr_cpu_ids) |
2963 | nr_hw_queues = nr_cpu_ids; | |
2964 | if (nr_hw_queues < 1 || nr_hw_queues == set->nr_hw_queues) | |
2965 | return; | |
2966 | ||
2967 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
2968 | blk_mq_freeze_queue(q); | |
2969 | ||
2970 | set->nr_hw_queues = nr_hw_queues; | |
ebe8bddb | 2971 | blk_mq_update_queue_map(set); |
868f2f0b KB |
2972 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
2973 | blk_mq_realloc_hw_ctxs(set, q); | |
4b855ad3 | 2974 | blk_mq_queue_reinit(q); |
868f2f0b KB |
2975 | } |
2976 | ||
2977 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
2978 | blk_mq_unfreeze_queue(q); | |
2979 | } | |
e4dc2b32 KB |
2980 | |
2981 | void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues) | |
2982 | { | |
2983 | mutex_lock(&set->tag_list_lock); | |
2984 | __blk_mq_update_nr_hw_queues(set, nr_hw_queues); | |
2985 | mutex_unlock(&set->tag_list_lock); | |
2986 | } | |
868f2f0b KB |
2987 | EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues); |
2988 | ||
34dbad5d OS |
2989 | /* Enable polling stats and return whether they were already enabled. */ |
2990 | static bool blk_poll_stats_enable(struct request_queue *q) | |
2991 | { | |
2992 | if (test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) || | |
2993 | test_and_set_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags)) | |
2994 | return true; | |
2995 | blk_stat_add_callback(q, q->poll_cb); | |
2996 | return false; | |
2997 | } | |
2998 | ||
2999 | static void blk_mq_poll_stats_start(struct request_queue *q) | |
3000 | { | |
3001 | /* | |
3002 | * We don't arm the callback if polling stats are not enabled or the | |
3003 | * callback is already active. | |
3004 | */ | |
3005 | if (!test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) || | |
3006 | blk_stat_is_active(q->poll_cb)) | |
3007 | return; | |
3008 | ||
3009 | blk_stat_activate_msecs(q->poll_cb, 100); | |
3010 | } | |
3011 | ||
3012 | static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb) | |
3013 | { | |
3014 | struct request_queue *q = cb->data; | |
720b8ccc | 3015 | int bucket; |
34dbad5d | 3016 | |
720b8ccc SB |
3017 | for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) { |
3018 | if (cb->stat[bucket].nr_samples) | |
3019 | q->poll_stat[bucket] = cb->stat[bucket]; | |
3020 | } | |
34dbad5d OS |
3021 | } |
3022 | ||
64f1c21e JA |
3023 | static unsigned long blk_mq_poll_nsecs(struct request_queue *q, |
3024 | struct blk_mq_hw_ctx *hctx, | |
3025 | struct request *rq) | |
3026 | { | |
64f1c21e | 3027 | unsigned long ret = 0; |
720b8ccc | 3028 | int bucket; |
64f1c21e JA |
3029 | |
3030 | /* | |
3031 | * If stats collection isn't on, don't sleep but turn it on for | |
3032 | * future users | |
3033 | */ | |
34dbad5d | 3034 | if (!blk_poll_stats_enable(q)) |
64f1c21e JA |
3035 | return 0; |
3036 | ||
64f1c21e JA |
3037 | /* |
3038 | * As an optimistic guess, use half of the mean service time | |
3039 | * for this type of request. We can (and should) make this smarter. | |
3040 | * For instance, if the completion latencies are tight, we can | |
3041 | * get closer than just half the mean. This is especially | |
3042 | * important on devices where the completion latencies are longer | |
720b8ccc SB |
3043 | * than ~10 usec. We do use the stats for the relevant IO size |
3044 | * if available which does lead to better estimates. | |
64f1c21e | 3045 | */ |
720b8ccc SB |
3046 | bucket = blk_mq_poll_stats_bkt(rq); |
3047 | if (bucket < 0) | |
3048 | return ret; | |
3049 | ||
3050 | if (q->poll_stat[bucket].nr_samples) | |
3051 | ret = (q->poll_stat[bucket].mean + 1) / 2; | |
64f1c21e JA |
3052 | |
3053 | return ret; | |
3054 | } | |
3055 | ||
06426adf | 3056 | static bool blk_mq_poll_hybrid_sleep(struct request_queue *q, |
64f1c21e | 3057 | struct blk_mq_hw_ctx *hctx, |
06426adf JA |
3058 | struct request *rq) |
3059 | { | |
3060 | struct hrtimer_sleeper hs; | |
3061 | enum hrtimer_mode mode; | |
64f1c21e | 3062 | unsigned int nsecs; |
06426adf JA |
3063 | ktime_t kt; |
3064 | ||
76a86f9d | 3065 | if (rq->rq_flags & RQF_MQ_POLL_SLEPT) |
64f1c21e JA |
3066 | return false; |
3067 | ||
3068 | /* | |
3069 | * poll_nsec can be: | |
3070 | * | |
3071 | * -1: don't ever hybrid sleep | |
3072 | * 0: use half of prev avg | |
3073 | * >0: use this specific value | |
3074 | */ | |
3075 | if (q->poll_nsec == -1) | |
3076 | return false; | |
3077 | else if (q->poll_nsec > 0) | |
3078 | nsecs = q->poll_nsec; | |
3079 | else | |
3080 | nsecs = blk_mq_poll_nsecs(q, hctx, rq); | |
3081 | ||
3082 | if (!nsecs) | |
06426adf JA |
3083 | return false; |
3084 | ||
76a86f9d | 3085 | rq->rq_flags |= RQF_MQ_POLL_SLEPT; |
06426adf JA |
3086 | |
3087 | /* | |
3088 | * This will be replaced with the stats tracking code, using | |
3089 | * 'avg_completion_time / 2' as the pre-sleep target. | |
3090 | */ | |
8b0e1953 | 3091 | kt = nsecs; |
06426adf JA |
3092 | |
3093 | mode = HRTIMER_MODE_REL; | |
3094 | hrtimer_init_on_stack(&hs.timer, CLOCK_MONOTONIC, mode); | |
3095 | hrtimer_set_expires(&hs.timer, kt); | |
3096 | ||
3097 | hrtimer_init_sleeper(&hs, current); | |
3098 | do { | |
5a61c363 | 3099 | if (blk_mq_rq_state(rq) == MQ_RQ_COMPLETE) |
06426adf JA |
3100 | break; |
3101 | set_current_state(TASK_UNINTERRUPTIBLE); | |
3102 | hrtimer_start_expires(&hs.timer, mode); | |
3103 | if (hs.task) | |
3104 | io_schedule(); | |
3105 | hrtimer_cancel(&hs.timer); | |
3106 | mode = HRTIMER_MODE_ABS; | |
3107 | } while (hs.task && !signal_pending(current)); | |
3108 | ||
3109 | __set_current_state(TASK_RUNNING); | |
3110 | destroy_hrtimer_on_stack(&hs.timer); | |
3111 | return true; | |
3112 | } | |
3113 | ||
bbd7bb70 JA |
3114 | static bool __blk_mq_poll(struct blk_mq_hw_ctx *hctx, struct request *rq) |
3115 | { | |
3116 | struct request_queue *q = hctx->queue; | |
3117 | long state; | |
3118 | ||
06426adf JA |
3119 | /* |
3120 | * If we sleep, have the caller restart the poll loop to reset | |
3121 | * the state. Like for the other success return cases, the | |
3122 | * caller is responsible for checking if the IO completed. If | |
3123 | * the IO isn't complete, we'll get called again and will go | |
3124 | * straight to the busy poll loop. | |
3125 | */ | |
64f1c21e | 3126 | if (blk_mq_poll_hybrid_sleep(q, hctx, rq)) |
06426adf JA |
3127 | return true; |
3128 | ||
bbd7bb70 JA |
3129 | hctx->poll_considered++; |
3130 | ||
3131 | state = current->state; | |
3132 | while (!need_resched()) { | |
3133 | int ret; | |
3134 | ||
3135 | hctx->poll_invoked++; | |
3136 | ||
3137 | ret = q->mq_ops->poll(hctx, rq->tag); | |
3138 | if (ret > 0) { | |
3139 | hctx->poll_success++; | |
3140 | set_current_state(TASK_RUNNING); | |
3141 | return true; | |
3142 | } | |
3143 | ||
3144 | if (signal_pending_state(state, current)) | |
3145 | set_current_state(TASK_RUNNING); | |
3146 | ||
3147 | if (current->state == TASK_RUNNING) | |
3148 | return true; | |
3149 | if (ret < 0) | |
3150 | break; | |
3151 | cpu_relax(); | |
3152 | } | |
3153 | ||
3154 | return false; | |
3155 | } | |
3156 | ||
ea435e1b | 3157 | static bool blk_mq_poll(struct request_queue *q, blk_qc_t cookie) |
bbd7bb70 JA |
3158 | { |
3159 | struct blk_mq_hw_ctx *hctx; | |
bbd7bb70 JA |
3160 | struct request *rq; |
3161 | ||
ea435e1b | 3162 | if (!test_bit(QUEUE_FLAG_POLL, &q->queue_flags)) |
bbd7bb70 JA |
3163 | return false; |
3164 | ||
bbd7bb70 | 3165 | hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)]; |
bd166ef1 JA |
3166 | if (!blk_qc_t_is_internal(cookie)) |
3167 | rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie)); | |
3a07bb1d | 3168 | else { |
bd166ef1 | 3169 | rq = blk_mq_tag_to_rq(hctx->sched_tags, blk_qc_t_to_tag(cookie)); |
3a07bb1d JA |
3170 | /* |
3171 | * With scheduling, if the request has completed, we'll | |
3172 | * get a NULL return here, as we clear the sched tag when | |
3173 | * that happens. The request still remains valid, like always, | |
3174 | * so we should be safe with just the NULL check. | |
3175 | */ | |
3176 | if (!rq) | |
3177 | return false; | |
3178 | } | |
bbd7bb70 JA |
3179 | |
3180 | return __blk_mq_poll(hctx, rq); | |
3181 | } | |
bbd7bb70 | 3182 | |
320ae51f JA |
3183 | static int __init blk_mq_init(void) |
3184 | { | |
9467f859 TG |
3185 | cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL, |
3186 | blk_mq_hctx_notify_dead); | |
320ae51f JA |
3187 | return 0; |
3188 | } | |
3189 | subsys_initcall(blk_mq_init); |