]>
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); |
454be724 ML |
163 | if (!q->mq_ops) |
164 | blk_drain_queue(q); | |
f3af020b TH |
165 | blk_mq_freeze_queue_wait(q); |
166 | } | |
3ef28e83 DW |
167 | |
168 | void blk_mq_freeze_queue(struct request_queue *q) | |
169 | { | |
170 | /* | |
171 | * ...just an alias to keep freeze and unfreeze actions balanced | |
172 | * in the blk_mq_* namespace | |
173 | */ | |
174 | blk_freeze_queue(q); | |
175 | } | |
c761d96b | 176 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue); |
f3af020b | 177 | |
b4c6a028 | 178 | void blk_mq_unfreeze_queue(struct request_queue *q) |
320ae51f | 179 | { |
4ecd4fef | 180 | int freeze_depth; |
320ae51f | 181 | |
4ecd4fef CH |
182 | freeze_depth = atomic_dec_return(&q->mq_freeze_depth); |
183 | WARN_ON_ONCE(freeze_depth < 0); | |
184 | if (!freeze_depth) { | |
3ef28e83 | 185 | percpu_ref_reinit(&q->q_usage_counter); |
320ae51f | 186 | wake_up_all(&q->mq_freeze_wq); |
add703fd | 187 | } |
320ae51f | 188 | } |
b4c6a028 | 189 | EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue); |
320ae51f | 190 | |
852ec809 BVA |
191 | /* |
192 | * FIXME: replace the scsi_internal_device_*block_nowait() calls in the | |
193 | * mpt3sas driver such that this function can be removed. | |
194 | */ | |
195 | void blk_mq_quiesce_queue_nowait(struct request_queue *q) | |
196 | { | |
197 | unsigned long flags; | |
198 | ||
199 | spin_lock_irqsave(q->queue_lock, flags); | |
200 | queue_flag_set(QUEUE_FLAG_QUIESCED, q); | |
201 | spin_unlock_irqrestore(q->queue_lock, flags); | |
202 | } | |
203 | EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait); | |
204 | ||
6a83e74d | 205 | /** |
69e07c4a | 206 | * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished |
6a83e74d BVA |
207 | * @q: request queue. |
208 | * | |
209 | * Note: this function does not prevent that the struct request end_io() | |
69e07c4a ML |
210 | * callback function is invoked. Once this function is returned, we make |
211 | * sure no dispatch can happen until the queue is unquiesced via | |
212 | * blk_mq_unquiesce_queue(). | |
6a83e74d BVA |
213 | */ |
214 | void blk_mq_quiesce_queue(struct request_queue *q) | |
215 | { | |
216 | struct blk_mq_hw_ctx *hctx; | |
217 | unsigned int i; | |
218 | bool rcu = false; | |
219 | ||
1d9e9bc6 | 220 | blk_mq_quiesce_queue_nowait(q); |
f4560ffe | 221 | |
6a83e74d BVA |
222 | queue_for_each_hw_ctx(q, hctx, i) { |
223 | if (hctx->flags & BLK_MQ_F_BLOCKING) | |
05707b64 | 224 | synchronize_srcu(hctx->srcu); |
6a83e74d BVA |
225 | else |
226 | rcu = true; | |
227 | } | |
228 | if (rcu) | |
229 | synchronize_rcu(); | |
230 | } | |
231 | EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue); | |
232 | ||
e4e73913 ML |
233 | /* |
234 | * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue() | |
235 | * @q: request queue. | |
236 | * | |
237 | * This function recovers queue into the state before quiescing | |
238 | * which is done by blk_mq_quiesce_queue. | |
239 | */ | |
240 | void blk_mq_unquiesce_queue(struct request_queue *q) | |
241 | { | |
852ec809 BVA |
242 | unsigned long flags; |
243 | ||
244 | spin_lock_irqsave(q->queue_lock, flags); | |
f4560ffe | 245 | queue_flag_clear(QUEUE_FLAG_QUIESCED, q); |
852ec809 | 246 | spin_unlock_irqrestore(q->queue_lock, flags); |
f4560ffe | 247 | |
1d9e9bc6 ML |
248 | /* dispatch requests which are inserted during quiescing */ |
249 | blk_mq_run_hw_queues(q, true); | |
e4e73913 ML |
250 | } |
251 | EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue); | |
252 | ||
aed3ea94 JA |
253 | void blk_mq_wake_waiters(struct request_queue *q) |
254 | { | |
255 | struct blk_mq_hw_ctx *hctx; | |
256 | unsigned int i; | |
257 | ||
258 | queue_for_each_hw_ctx(q, hctx, i) | |
259 | if (blk_mq_hw_queue_mapped(hctx)) | |
260 | blk_mq_tag_wakeup_all(hctx->tags, true); | |
261 | } | |
262 | ||
320ae51f JA |
263 | bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx) |
264 | { | |
265 | return blk_mq_has_free_tags(hctx->tags); | |
266 | } | |
267 | EXPORT_SYMBOL(blk_mq_can_queue); | |
268 | ||
e4cdf1a1 CH |
269 | static struct request *blk_mq_rq_ctx_init(struct blk_mq_alloc_data *data, |
270 | unsigned int tag, unsigned int op) | |
320ae51f | 271 | { |
e4cdf1a1 CH |
272 | struct blk_mq_tags *tags = blk_mq_tags_from_data(data); |
273 | struct request *rq = tags->static_rqs[tag]; | |
bf9ae8c5 | 274 | req_flags_t rq_flags = 0; |
c3a148d2 | 275 | |
e4cdf1a1 CH |
276 | if (data->flags & BLK_MQ_REQ_INTERNAL) { |
277 | rq->tag = -1; | |
278 | rq->internal_tag = tag; | |
279 | } else { | |
280 | if (blk_mq_tag_busy(data->hctx)) { | |
bf9ae8c5 | 281 | rq_flags = RQF_MQ_INFLIGHT; |
e4cdf1a1 CH |
282 | atomic_inc(&data->hctx->nr_active); |
283 | } | |
284 | rq->tag = tag; | |
285 | rq->internal_tag = -1; | |
286 | data->hctx->tags->rqs[rq->tag] = rq; | |
287 | } | |
288 | ||
af76e555 | 289 | /* csd/requeue_work/fifo_time is initialized before use */ |
e4cdf1a1 CH |
290 | rq->q = data->q; |
291 | rq->mq_ctx = data->ctx; | |
bf9ae8c5 | 292 | rq->rq_flags = rq_flags; |
7c3fb70f | 293 | rq->cpu = -1; |
ef295ecf | 294 | rq->cmd_flags = op; |
1b6d65a0 BVA |
295 | if (data->flags & BLK_MQ_REQ_PREEMPT) |
296 | rq->rq_flags |= RQF_PREEMPT; | |
e4cdf1a1 | 297 | if (blk_queue_io_stat(data->q)) |
e8064021 | 298 | rq->rq_flags |= RQF_IO_STAT; |
7c3fb70f | 299 | INIT_LIST_HEAD(&rq->queuelist); |
af76e555 CH |
300 | INIT_HLIST_NODE(&rq->hash); |
301 | RB_CLEAR_NODE(&rq->rb_node); | |
af76e555 CH |
302 | rq->rq_disk = NULL; |
303 | rq->part = NULL; | |
3ee32372 | 304 | rq->start_time = jiffies; |
af76e555 CH |
305 | rq->nr_phys_segments = 0; |
306 | #if defined(CONFIG_BLK_DEV_INTEGRITY) | |
307 | rq->nr_integrity_segments = 0; | |
308 | #endif | |
af76e555 CH |
309 | rq->special = NULL; |
310 | /* tag was already set */ | |
af76e555 | 311 | rq->extra_len = 0; |
e14575b3 | 312 | rq->__deadline = 0; |
af76e555 | 313 | |
af76e555 | 314 | INIT_LIST_HEAD(&rq->timeout_list); |
f6be4fb4 JA |
315 | rq->timeout = 0; |
316 | ||
af76e555 CH |
317 | rq->end_io = NULL; |
318 | rq->end_io_data = NULL; | |
319 | rq->next_rq = NULL; | |
320 | ||
7c3fb70f JA |
321 | #ifdef CONFIG_BLK_CGROUP |
322 | rq->rl = NULL; | |
323 | set_start_time_ns(rq); | |
324 | rq->io_start_time_ns = 0; | |
325 | #endif | |
326 | ||
e4cdf1a1 CH |
327 | data->ctx->rq_dispatched[op_is_sync(op)]++; |
328 | return rq; | |
5dee8577 CH |
329 | } |
330 | ||
d2c0d383 CH |
331 | static struct request *blk_mq_get_request(struct request_queue *q, |
332 | struct bio *bio, unsigned int op, | |
333 | struct blk_mq_alloc_data *data) | |
334 | { | |
335 | struct elevator_queue *e = q->elevator; | |
336 | struct request *rq; | |
e4cdf1a1 | 337 | unsigned int tag; |
21e768b4 | 338 | bool put_ctx_on_error = false; |
d2c0d383 CH |
339 | |
340 | blk_queue_enter_live(q); | |
341 | data->q = q; | |
21e768b4 BVA |
342 | if (likely(!data->ctx)) { |
343 | data->ctx = blk_mq_get_ctx(q); | |
344 | put_ctx_on_error = true; | |
345 | } | |
d2c0d383 CH |
346 | if (likely(!data->hctx)) |
347 | data->hctx = blk_mq_map_queue(q, data->ctx->cpu); | |
03a07c92 GR |
348 | if (op & REQ_NOWAIT) |
349 | data->flags |= BLK_MQ_REQ_NOWAIT; | |
d2c0d383 CH |
350 | |
351 | if (e) { | |
352 | data->flags |= BLK_MQ_REQ_INTERNAL; | |
353 | ||
354 | /* | |
355 | * Flush requests are special and go directly to the | |
356 | * dispatch list. | |
357 | */ | |
5bbf4e5a CH |
358 | if (!op_is_flush(op) && e->type->ops.mq.limit_depth) |
359 | e->type->ops.mq.limit_depth(op, data); | |
d2c0d383 CH |
360 | } |
361 | ||
e4cdf1a1 CH |
362 | tag = blk_mq_get_tag(data); |
363 | if (tag == BLK_MQ_TAG_FAIL) { | |
21e768b4 BVA |
364 | if (put_ctx_on_error) { |
365 | blk_mq_put_ctx(data->ctx); | |
1ad43c00 ML |
366 | data->ctx = NULL; |
367 | } | |
037cebb8 CH |
368 | blk_queue_exit(q); |
369 | return NULL; | |
d2c0d383 CH |
370 | } |
371 | ||
e4cdf1a1 | 372 | rq = blk_mq_rq_ctx_init(data, tag, op); |
037cebb8 CH |
373 | if (!op_is_flush(op)) { |
374 | rq->elv.icq = NULL; | |
5bbf4e5a | 375 | if (e && e->type->ops.mq.prepare_request) { |
44e8c2bf CH |
376 | if (e->type->icq_cache && rq_ioc(bio)) |
377 | blk_mq_sched_assign_ioc(rq, bio); | |
378 | ||
5bbf4e5a CH |
379 | e->type->ops.mq.prepare_request(rq, bio); |
380 | rq->rq_flags |= RQF_ELVPRIV; | |
44e8c2bf | 381 | } |
037cebb8 CH |
382 | } |
383 | data->hctx->queued++; | |
384 | return rq; | |
d2c0d383 CH |
385 | } |
386 | ||
cd6ce148 | 387 | struct request *blk_mq_alloc_request(struct request_queue *q, unsigned int op, |
9a95e4ef | 388 | blk_mq_req_flags_t flags) |
320ae51f | 389 | { |
5a797e00 | 390 | struct blk_mq_alloc_data alloc_data = { .flags = flags }; |
bd166ef1 | 391 | struct request *rq; |
a492f075 | 392 | int ret; |
320ae51f | 393 | |
3a0a5299 | 394 | ret = blk_queue_enter(q, flags); |
a492f075 JL |
395 | if (ret) |
396 | return ERR_PTR(ret); | |
320ae51f | 397 | |
cd6ce148 | 398 | rq = blk_mq_get_request(q, NULL, op, &alloc_data); |
3280d66a | 399 | blk_queue_exit(q); |
841bac2c | 400 | |
bd166ef1 | 401 | if (!rq) |
a492f075 | 402 | return ERR_PTR(-EWOULDBLOCK); |
0c4de0f3 | 403 | |
1ad43c00 | 404 | blk_mq_put_ctx(alloc_data.ctx); |
1ad43c00 | 405 | |
0c4de0f3 CH |
406 | rq->__data_len = 0; |
407 | rq->__sector = (sector_t) -1; | |
408 | rq->bio = rq->biotail = NULL; | |
320ae51f JA |
409 | return rq; |
410 | } | |
4bb659b1 | 411 | EXPORT_SYMBOL(blk_mq_alloc_request); |
320ae51f | 412 | |
cd6ce148 | 413 | struct request *blk_mq_alloc_request_hctx(struct request_queue *q, |
9a95e4ef | 414 | unsigned int op, blk_mq_req_flags_t flags, unsigned int hctx_idx) |
1f5bd336 | 415 | { |
6d2809d5 | 416 | struct blk_mq_alloc_data alloc_data = { .flags = flags }; |
1f5bd336 | 417 | struct request *rq; |
6d2809d5 | 418 | unsigned int cpu; |
1f5bd336 ML |
419 | int ret; |
420 | ||
421 | /* | |
422 | * If the tag allocator sleeps we could get an allocation for a | |
423 | * different hardware context. No need to complicate the low level | |
424 | * allocator for this for the rare use case of a command tied to | |
425 | * a specific queue. | |
426 | */ | |
427 | if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT))) | |
428 | return ERR_PTR(-EINVAL); | |
429 | ||
430 | if (hctx_idx >= q->nr_hw_queues) | |
431 | return ERR_PTR(-EIO); | |
432 | ||
3a0a5299 | 433 | ret = blk_queue_enter(q, flags); |
1f5bd336 ML |
434 | if (ret) |
435 | return ERR_PTR(ret); | |
436 | ||
c8712c6a CH |
437 | /* |
438 | * Check if the hardware context is actually mapped to anything. | |
439 | * If not tell the caller that it should skip this queue. | |
440 | */ | |
6d2809d5 OS |
441 | alloc_data.hctx = q->queue_hw_ctx[hctx_idx]; |
442 | if (!blk_mq_hw_queue_mapped(alloc_data.hctx)) { | |
443 | blk_queue_exit(q); | |
444 | return ERR_PTR(-EXDEV); | |
c8712c6a | 445 | } |
20e4d813 | 446 | cpu = cpumask_first_and(alloc_data.hctx->cpumask, cpu_online_mask); |
6d2809d5 | 447 | alloc_data.ctx = __blk_mq_get_ctx(q, cpu); |
1f5bd336 | 448 | |
cd6ce148 | 449 | rq = blk_mq_get_request(q, NULL, op, &alloc_data); |
3280d66a | 450 | blk_queue_exit(q); |
c8712c6a | 451 | |
6d2809d5 OS |
452 | if (!rq) |
453 | return ERR_PTR(-EWOULDBLOCK); | |
454 | ||
455 | return rq; | |
1f5bd336 ML |
456 | } |
457 | EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx); | |
458 | ||
6af54051 | 459 | void blk_mq_free_request(struct request *rq) |
320ae51f | 460 | { |
320ae51f | 461 | struct request_queue *q = rq->q; |
6af54051 CH |
462 | struct elevator_queue *e = q->elevator; |
463 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
464 | struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, ctx->cpu); | |
465 | const int sched_tag = rq->internal_tag; | |
466 | ||
5bbf4e5a | 467 | if (rq->rq_flags & RQF_ELVPRIV) { |
6af54051 CH |
468 | if (e && e->type->ops.mq.finish_request) |
469 | e->type->ops.mq.finish_request(rq); | |
470 | if (rq->elv.icq) { | |
471 | put_io_context(rq->elv.icq->ioc); | |
472 | rq->elv.icq = NULL; | |
473 | } | |
474 | } | |
320ae51f | 475 | |
6af54051 | 476 | ctx->rq_completed[rq_is_sync(rq)]++; |
e8064021 | 477 | if (rq->rq_flags & RQF_MQ_INFLIGHT) |
0d2602ca | 478 | atomic_dec(&hctx->nr_active); |
87760e5e | 479 | |
7beb2f84 JA |
480 | if (unlikely(laptop_mode && !blk_rq_is_passthrough(rq))) |
481 | laptop_io_completion(q->backing_dev_info); | |
482 | ||
87760e5e | 483 | wbt_done(q->rq_wb, &rq->issue_stat); |
0d2602ca | 484 | |
85acb3ba SL |
485 | if (blk_rq_rl(rq)) |
486 | blk_put_rl(blk_rq_rl(rq)); | |
487 | ||
1d9bd516 | 488 | blk_mq_rq_update_state(rq, MQ_RQ_IDLE); |
bd166ef1 JA |
489 | if (rq->tag != -1) |
490 | blk_mq_put_tag(hctx, hctx->tags, ctx, rq->tag); | |
491 | if (sched_tag != -1) | |
c05f8525 | 492 | blk_mq_put_tag(hctx, hctx->sched_tags, ctx, sched_tag); |
6d8c6c0f | 493 | blk_mq_sched_restart(hctx); |
3ef28e83 | 494 | blk_queue_exit(q); |
320ae51f | 495 | } |
1a3b595a | 496 | EXPORT_SYMBOL_GPL(blk_mq_free_request); |
320ae51f | 497 | |
2a842aca | 498 | inline void __blk_mq_end_request(struct request *rq, blk_status_t error) |
320ae51f | 499 | { |
0d11e6ac ML |
500 | blk_account_io_done(rq); |
501 | ||
91b63639 | 502 | if (rq->end_io) { |
87760e5e | 503 | wbt_done(rq->q->rq_wb, &rq->issue_stat); |
320ae51f | 504 | rq->end_io(rq, error); |
91b63639 CH |
505 | } else { |
506 | if (unlikely(blk_bidi_rq(rq))) | |
507 | blk_mq_free_request(rq->next_rq); | |
320ae51f | 508 | blk_mq_free_request(rq); |
91b63639 | 509 | } |
320ae51f | 510 | } |
c8a446ad | 511 | EXPORT_SYMBOL(__blk_mq_end_request); |
63151a44 | 512 | |
2a842aca | 513 | void blk_mq_end_request(struct request *rq, blk_status_t error) |
63151a44 CH |
514 | { |
515 | if (blk_update_request(rq, error, blk_rq_bytes(rq))) | |
516 | BUG(); | |
c8a446ad | 517 | __blk_mq_end_request(rq, error); |
63151a44 | 518 | } |
c8a446ad | 519 | EXPORT_SYMBOL(blk_mq_end_request); |
320ae51f | 520 | |
30a91cb4 | 521 | static void __blk_mq_complete_request_remote(void *data) |
320ae51f | 522 | { |
3d6efbf6 | 523 | struct request *rq = data; |
320ae51f | 524 | |
30a91cb4 | 525 | rq->q->softirq_done_fn(rq); |
320ae51f | 526 | } |
320ae51f | 527 | |
453f8341 | 528 | static void __blk_mq_complete_request(struct request *rq) |
320ae51f JA |
529 | { |
530 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
38535201 | 531 | bool shared = false; |
320ae51f JA |
532 | int cpu; |
533 | ||
1d9bd516 | 534 | WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT); |
5a61c363 | 535 | blk_mq_rq_update_state(rq, MQ_RQ_COMPLETE); |
1d9bd516 | 536 | |
453f8341 CH |
537 | if (rq->internal_tag != -1) |
538 | blk_mq_sched_completed_request(rq); | |
539 | if (rq->rq_flags & RQF_STATS) { | |
540 | blk_mq_poll_stats_start(rq->q); | |
541 | blk_stat_add(rq); | |
542 | } | |
543 | ||
38535201 | 544 | if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) { |
30a91cb4 CH |
545 | rq->q->softirq_done_fn(rq); |
546 | return; | |
547 | } | |
320ae51f JA |
548 | |
549 | cpu = get_cpu(); | |
38535201 CH |
550 | if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags)) |
551 | shared = cpus_share_cache(cpu, ctx->cpu); | |
552 | ||
553 | if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) { | |
30a91cb4 | 554 | rq->csd.func = __blk_mq_complete_request_remote; |
3d6efbf6 CH |
555 | rq->csd.info = rq; |
556 | rq->csd.flags = 0; | |
c46fff2a | 557 | smp_call_function_single_async(ctx->cpu, &rq->csd); |
3d6efbf6 | 558 | } else { |
30a91cb4 | 559 | rq->q->softirq_done_fn(rq); |
3d6efbf6 | 560 | } |
320ae51f JA |
561 | put_cpu(); |
562 | } | |
30a91cb4 | 563 | |
04ced159 | 564 | static void hctx_unlock(struct blk_mq_hw_ctx *hctx, int srcu_idx) |
b7435db8 | 565 | __releases(hctx->srcu) |
04ced159 JA |
566 | { |
567 | if (!(hctx->flags & BLK_MQ_F_BLOCKING)) | |
568 | rcu_read_unlock(); | |
569 | else | |
05707b64 | 570 | srcu_read_unlock(hctx->srcu, srcu_idx); |
04ced159 JA |
571 | } |
572 | ||
573 | static void hctx_lock(struct blk_mq_hw_ctx *hctx, int *srcu_idx) | |
b7435db8 | 574 | __acquires(hctx->srcu) |
04ced159 | 575 | { |
08b5a6e2 JA |
576 | if (!(hctx->flags & BLK_MQ_F_BLOCKING)) { |
577 | /* shut up gcc false positive */ | |
578 | *srcu_idx = 0; | |
04ced159 | 579 | rcu_read_lock(); |
08b5a6e2 | 580 | } else |
05707b64 | 581 | *srcu_idx = srcu_read_lock(hctx->srcu); |
04ced159 JA |
582 | } |
583 | ||
1d9bd516 TH |
584 | static void blk_mq_rq_update_aborted_gstate(struct request *rq, u64 gstate) |
585 | { | |
586 | unsigned long flags; | |
587 | ||
588 | /* | |
589 | * blk_mq_rq_aborted_gstate() is used from the completion path and | |
590 | * can thus be called from irq context. u64_stats_fetch in the | |
591 | * middle of update on the same CPU leads to lockup. Disable irq | |
592 | * while updating. | |
593 | */ | |
594 | local_irq_save(flags); | |
595 | u64_stats_update_begin(&rq->aborted_gstate_sync); | |
596 | rq->aborted_gstate = gstate; | |
597 | u64_stats_update_end(&rq->aborted_gstate_sync); | |
598 | local_irq_restore(flags); | |
599 | } | |
600 | ||
601 | static u64 blk_mq_rq_aborted_gstate(struct request *rq) | |
602 | { | |
603 | unsigned int start; | |
604 | u64 aborted_gstate; | |
605 | ||
606 | do { | |
607 | start = u64_stats_fetch_begin(&rq->aborted_gstate_sync); | |
608 | aborted_gstate = rq->aborted_gstate; | |
609 | } while (u64_stats_fetch_retry(&rq->aborted_gstate_sync, start)); | |
610 | ||
611 | return aborted_gstate; | |
612 | } | |
613 | ||
30a91cb4 CH |
614 | /** |
615 | * blk_mq_complete_request - end I/O on a request | |
616 | * @rq: the request being processed | |
617 | * | |
618 | * Description: | |
619 | * Ends all I/O on a request. It does not handle partial completions. | |
620 | * The actual completion happens out-of-order, through a IPI handler. | |
621 | **/ | |
08e0029a | 622 | void blk_mq_complete_request(struct request *rq) |
30a91cb4 | 623 | { |
95f09684 | 624 | struct request_queue *q = rq->q; |
5197c05e TH |
625 | struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, rq->mq_ctx->cpu); |
626 | int srcu_idx; | |
95f09684 JA |
627 | |
628 | if (unlikely(blk_should_fake_timeout(q))) | |
30a91cb4 | 629 | return; |
5197c05e | 630 | |
1d9bd516 TH |
631 | /* |
632 | * If @rq->aborted_gstate equals the current instance, timeout is | |
633 | * claiming @rq and we lost. This is synchronized through | |
634 | * hctx_lock(). See blk_mq_timeout_work() for details. | |
635 | * | |
636 | * Completion path never blocks and we can directly use RCU here | |
637 | * instead of hctx_lock() which can be either RCU or SRCU. | |
638 | * However, that would complicate paths which want to synchronize | |
639 | * against us. Let stay in sync with the issue path so that | |
640 | * hctx_lock() covers both issue and completion paths. | |
641 | */ | |
5197c05e | 642 | hctx_lock(hctx, &srcu_idx); |
634f9e46 | 643 | if (blk_mq_rq_aborted_gstate(rq) != rq->gstate) |
ed851860 | 644 | __blk_mq_complete_request(rq); |
5197c05e | 645 | hctx_unlock(hctx, srcu_idx); |
30a91cb4 CH |
646 | } |
647 | EXPORT_SYMBOL(blk_mq_complete_request); | |
320ae51f | 648 | |
973c0191 KB |
649 | int blk_mq_request_started(struct request *rq) |
650 | { | |
5a61c363 | 651 | return blk_mq_rq_state(rq) != MQ_RQ_IDLE; |
973c0191 KB |
652 | } |
653 | EXPORT_SYMBOL_GPL(blk_mq_request_started); | |
654 | ||
e2490073 | 655 | void blk_mq_start_request(struct request *rq) |
320ae51f JA |
656 | { |
657 | struct request_queue *q = rq->q; | |
658 | ||
bd166ef1 JA |
659 | blk_mq_sched_started_request(rq); |
660 | ||
320ae51f JA |
661 | trace_block_rq_issue(q, rq); |
662 | ||
cf43e6be | 663 | if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) { |
88eeca49 | 664 | blk_stat_set_issue(&rq->issue_stat, blk_rq_sectors(rq)); |
cf43e6be | 665 | rq->rq_flags |= RQF_STATS; |
87760e5e | 666 | wbt_issue(q->rq_wb, &rq->issue_stat); |
cf43e6be JA |
667 | } |
668 | ||
1d9bd516 | 669 | WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE); |
538b7534 | 670 | |
87ee7b11 | 671 | /* |
1d9bd516 TH |
672 | * Mark @rq in-flight which also advances the generation number, |
673 | * and register for timeout. Protect with a seqcount to allow the | |
674 | * timeout path to read both @rq->gstate and @rq->deadline | |
675 | * coherently. | |
a7af0af3 | 676 | * |
1d9bd516 TH |
677 | * This is the only place where a request is marked in-flight. If |
678 | * the timeout path reads an in-flight @rq->gstate, the | |
679 | * @rq->deadline it reads together under @rq->gstate_seq is | |
680 | * guaranteed to be the matching one. | |
87ee7b11 | 681 | */ |
1d9bd516 TH |
682 | preempt_disable(); |
683 | write_seqcount_begin(&rq->gstate_seq); | |
684 | ||
685 | blk_mq_rq_update_state(rq, MQ_RQ_IN_FLIGHT); | |
686 | blk_add_timer(rq); | |
687 | ||
688 | write_seqcount_end(&rq->gstate_seq); | |
689 | preempt_enable(); | |
49f5baa5 CH |
690 | |
691 | if (q->dma_drain_size && blk_rq_bytes(rq)) { | |
692 | /* | |
693 | * Make sure space for the drain appears. We know we can do | |
694 | * this because max_hw_segments has been adjusted to be one | |
695 | * fewer than the device can handle. | |
696 | */ | |
697 | rq->nr_phys_segments++; | |
698 | } | |
320ae51f | 699 | } |
e2490073 | 700 | EXPORT_SYMBOL(blk_mq_start_request); |
320ae51f | 701 | |
d9d149a3 | 702 | /* |
5a61c363 TH |
703 | * When we reach here because queue is busy, it's safe to change the state |
704 | * to IDLE without checking @rq->aborted_gstate because we should still be | |
705 | * holding the RCU read lock and thus protected against timeout. | |
d9d149a3 | 706 | */ |
ed0791b2 | 707 | static void __blk_mq_requeue_request(struct request *rq) |
320ae51f JA |
708 | { |
709 | struct request_queue *q = rq->q; | |
710 | ||
923218f6 ML |
711 | blk_mq_put_driver_tag(rq); |
712 | ||
320ae51f | 713 | trace_block_rq_requeue(q, rq); |
87760e5e | 714 | wbt_requeue(q->rq_wb, &rq->issue_stat); |
49f5baa5 | 715 | |
5a61c363 | 716 | if (blk_mq_rq_state(rq) != MQ_RQ_IDLE) { |
1d9bd516 | 717 | blk_mq_rq_update_state(rq, MQ_RQ_IDLE); |
e2490073 CH |
718 | if (q->dma_drain_size && blk_rq_bytes(rq)) |
719 | rq->nr_phys_segments--; | |
720 | } | |
320ae51f JA |
721 | } |
722 | ||
2b053aca | 723 | void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list) |
ed0791b2 | 724 | { |
ed0791b2 | 725 | __blk_mq_requeue_request(rq); |
ed0791b2 | 726 | |
105976f5 ML |
727 | /* this request will be re-inserted to io scheduler queue */ |
728 | blk_mq_sched_requeue_request(rq); | |
729 | ||
ed0791b2 | 730 | BUG_ON(blk_queued_rq(rq)); |
2b053aca | 731 | blk_mq_add_to_requeue_list(rq, true, kick_requeue_list); |
ed0791b2 CH |
732 | } |
733 | EXPORT_SYMBOL(blk_mq_requeue_request); | |
734 | ||
6fca6a61 CH |
735 | static void blk_mq_requeue_work(struct work_struct *work) |
736 | { | |
737 | struct request_queue *q = | |
2849450a | 738 | container_of(work, struct request_queue, requeue_work.work); |
6fca6a61 CH |
739 | LIST_HEAD(rq_list); |
740 | struct request *rq, *next; | |
6fca6a61 | 741 | |
18e9781d | 742 | spin_lock_irq(&q->requeue_lock); |
6fca6a61 | 743 | list_splice_init(&q->requeue_list, &rq_list); |
18e9781d | 744 | spin_unlock_irq(&q->requeue_lock); |
6fca6a61 CH |
745 | |
746 | list_for_each_entry_safe(rq, next, &rq_list, queuelist) { | |
e8064021 | 747 | if (!(rq->rq_flags & RQF_SOFTBARRIER)) |
6fca6a61 CH |
748 | continue; |
749 | ||
e8064021 | 750 | rq->rq_flags &= ~RQF_SOFTBARRIER; |
6fca6a61 | 751 | list_del_init(&rq->queuelist); |
9e97d295 | 752 | blk_mq_sched_insert_request(rq, true, false, false); |
6fca6a61 CH |
753 | } |
754 | ||
755 | while (!list_empty(&rq_list)) { | |
756 | rq = list_entry(rq_list.next, struct request, queuelist); | |
757 | list_del_init(&rq->queuelist); | |
9e97d295 | 758 | blk_mq_sched_insert_request(rq, false, false, false); |
6fca6a61 CH |
759 | } |
760 | ||
52d7f1b5 | 761 | blk_mq_run_hw_queues(q, false); |
6fca6a61 CH |
762 | } |
763 | ||
2b053aca BVA |
764 | void blk_mq_add_to_requeue_list(struct request *rq, bool at_head, |
765 | bool kick_requeue_list) | |
6fca6a61 CH |
766 | { |
767 | struct request_queue *q = rq->q; | |
768 | unsigned long flags; | |
769 | ||
770 | /* | |
771 | * We abuse this flag that is otherwise used by the I/O scheduler to | |
ff821d27 | 772 | * request head insertion from the workqueue. |
6fca6a61 | 773 | */ |
e8064021 | 774 | BUG_ON(rq->rq_flags & RQF_SOFTBARRIER); |
6fca6a61 CH |
775 | |
776 | spin_lock_irqsave(&q->requeue_lock, flags); | |
777 | if (at_head) { | |
e8064021 | 778 | rq->rq_flags |= RQF_SOFTBARRIER; |
6fca6a61 CH |
779 | list_add(&rq->queuelist, &q->requeue_list); |
780 | } else { | |
781 | list_add_tail(&rq->queuelist, &q->requeue_list); | |
782 | } | |
783 | spin_unlock_irqrestore(&q->requeue_lock, flags); | |
2b053aca BVA |
784 | |
785 | if (kick_requeue_list) | |
786 | blk_mq_kick_requeue_list(q); | |
6fca6a61 CH |
787 | } |
788 | EXPORT_SYMBOL(blk_mq_add_to_requeue_list); | |
789 | ||
790 | void blk_mq_kick_requeue_list(struct request_queue *q) | |
791 | { | |
ae943d20 | 792 | kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0); |
6fca6a61 CH |
793 | } |
794 | EXPORT_SYMBOL(blk_mq_kick_requeue_list); | |
795 | ||
2849450a MS |
796 | void blk_mq_delay_kick_requeue_list(struct request_queue *q, |
797 | unsigned long msecs) | |
798 | { | |
d4acf365 BVA |
799 | kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, |
800 | msecs_to_jiffies(msecs)); | |
2849450a MS |
801 | } |
802 | EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list); | |
803 | ||
0e62f51f JA |
804 | struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag) |
805 | { | |
88c7b2b7 JA |
806 | if (tag < tags->nr_tags) { |
807 | prefetch(tags->rqs[tag]); | |
4ee86bab | 808 | return tags->rqs[tag]; |
88c7b2b7 | 809 | } |
4ee86bab HR |
810 | |
811 | return NULL; | |
24d2f903 CH |
812 | } |
813 | EXPORT_SYMBOL(blk_mq_tag_to_rq); | |
814 | ||
320ae51f | 815 | struct blk_mq_timeout_data { |
46f92d42 CH |
816 | unsigned long next; |
817 | unsigned int next_set; | |
1d9bd516 | 818 | unsigned int nr_expired; |
320ae51f JA |
819 | }; |
820 | ||
358f70da | 821 | static void blk_mq_rq_timed_out(struct request *req, bool reserved) |
320ae51f | 822 | { |
f8a5b122 | 823 | const struct blk_mq_ops *ops = req->q->mq_ops; |
46f92d42 | 824 | enum blk_eh_timer_return ret = BLK_EH_RESET_TIMER; |
87ee7b11 | 825 | |
634f9e46 | 826 | req->rq_flags |= RQF_MQ_TIMEOUT_EXPIRED; |
87ee7b11 | 827 | |
46f92d42 | 828 | if (ops->timeout) |
0152fb6b | 829 | ret = ops->timeout(req, reserved); |
46f92d42 CH |
830 | |
831 | switch (ret) { | |
832 | case BLK_EH_HANDLED: | |
833 | __blk_mq_complete_request(req); | |
834 | break; | |
835 | case BLK_EH_RESET_TIMER: | |
1d9bd516 TH |
836 | /* |
837 | * As nothing prevents from completion happening while | |
838 | * ->aborted_gstate is set, this may lead to ignored | |
839 | * completions and further spurious timeouts. | |
840 | */ | |
841 | blk_mq_rq_update_aborted_gstate(req, 0); | |
46f92d42 | 842 | blk_add_timer(req); |
46f92d42 CH |
843 | break; |
844 | case BLK_EH_NOT_HANDLED: | |
845 | break; | |
846 | default: | |
847 | printk(KERN_ERR "block: bad eh return: %d\n", ret); | |
848 | break; | |
849 | } | |
87ee7b11 | 850 | } |
5b3f25fc | 851 | |
81481eb4 CH |
852 | static void blk_mq_check_expired(struct blk_mq_hw_ctx *hctx, |
853 | struct request *rq, void *priv, bool reserved) | |
854 | { | |
855 | struct blk_mq_timeout_data *data = priv; | |
1d9bd516 TH |
856 | unsigned long gstate, deadline; |
857 | int start; | |
87ee7b11 | 858 | |
1d9bd516 | 859 | might_sleep(); |
87ee7b11 | 860 | |
5a61c363 | 861 | if (rq->rq_flags & RQF_MQ_TIMEOUT_EXPIRED) |
46f92d42 | 862 | return; |
a7af0af3 | 863 | |
1d9bd516 TH |
864 | /* read coherent snapshots of @rq->state_gen and @rq->deadline */ |
865 | while (true) { | |
866 | start = read_seqcount_begin(&rq->gstate_seq); | |
867 | gstate = READ_ONCE(rq->gstate); | |
0a72e7f4 | 868 | deadline = blk_rq_deadline(rq); |
1d9bd516 TH |
869 | if (!read_seqcount_retry(&rq->gstate_seq, start)) |
870 | break; | |
871 | cond_resched(); | |
872 | } | |
a7af0af3 | 873 | |
1d9bd516 TH |
874 | /* if in-flight && overdue, mark for abortion */ |
875 | if ((gstate & MQ_RQ_STATE_MASK) == MQ_RQ_IN_FLIGHT && | |
876 | time_after_eq(jiffies, deadline)) { | |
877 | blk_mq_rq_update_aborted_gstate(rq, gstate); | |
878 | data->nr_expired++; | |
879 | hctx->nr_expired++; | |
a7af0af3 PZ |
880 | } else if (!data->next_set || time_after(data->next, deadline)) { |
881 | data->next = deadline; | |
46f92d42 CH |
882 | data->next_set = 1; |
883 | } | |
87ee7b11 JA |
884 | } |
885 | ||
1d9bd516 TH |
886 | static void blk_mq_terminate_expired(struct blk_mq_hw_ctx *hctx, |
887 | struct request *rq, void *priv, bool reserved) | |
888 | { | |
889 | /* | |
890 | * We marked @rq->aborted_gstate and waited for RCU. If there were | |
891 | * completions that we lost to, they would have finished and | |
892 | * updated @rq->gstate by now; otherwise, the completion path is | |
893 | * now guaranteed to see @rq->aborted_gstate and yield. If | |
894 | * @rq->aborted_gstate still matches @rq->gstate, @rq is ours. | |
895 | */ | |
634f9e46 TH |
896 | if (!(rq->rq_flags & RQF_MQ_TIMEOUT_EXPIRED) && |
897 | READ_ONCE(rq->gstate) == rq->aborted_gstate) | |
1d9bd516 TH |
898 | blk_mq_rq_timed_out(rq, reserved); |
899 | } | |
900 | ||
287922eb | 901 | static void blk_mq_timeout_work(struct work_struct *work) |
320ae51f | 902 | { |
287922eb CH |
903 | struct request_queue *q = |
904 | container_of(work, struct request_queue, timeout_work); | |
81481eb4 CH |
905 | struct blk_mq_timeout_data data = { |
906 | .next = 0, | |
907 | .next_set = 0, | |
1d9bd516 | 908 | .nr_expired = 0, |
81481eb4 | 909 | }; |
1d9bd516 | 910 | struct blk_mq_hw_ctx *hctx; |
81481eb4 | 911 | int i; |
320ae51f | 912 | |
71f79fb3 GKB |
913 | /* A deadlock might occur if a request is stuck requiring a |
914 | * timeout at the same time a queue freeze is waiting | |
915 | * completion, since the timeout code would not be able to | |
916 | * acquire the queue reference here. | |
917 | * | |
918 | * That's why we don't use blk_queue_enter here; instead, we use | |
919 | * percpu_ref_tryget directly, because we need to be able to | |
920 | * obtain a reference even in the short window between the queue | |
921 | * starting to freeze, by dropping the first reference in | |
1671d522 | 922 | * blk_freeze_queue_start, and the moment the last request is |
71f79fb3 GKB |
923 | * consumed, marked by the instant q_usage_counter reaches |
924 | * zero. | |
925 | */ | |
926 | if (!percpu_ref_tryget(&q->q_usage_counter)) | |
287922eb CH |
927 | return; |
928 | ||
1d9bd516 | 929 | /* scan for the expired ones and set their ->aborted_gstate */ |
0bf6cd5b | 930 | blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &data); |
320ae51f | 931 | |
1d9bd516 TH |
932 | if (data.nr_expired) { |
933 | bool has_rcu = false; | |
934 | ||
935 | /* | |
936 | * Wait till everyone sees ->aborted_gstate. The | |
937 | * sequential waits for SRCUs aren't ideal. If this ever | |
938 | * becomes a problem, we can add per-hw_ctx rcu_head and | |
939 | * wait in parallel. | |
940 | */ | |
941 | queue_for_each_hw_ctx(q, hctx, i) { | |
942 | if (!hctx->nr_expired) | |
943 | continue; | |
944 | ||
945 | if (!(hctx->flags & BLK_MQ_F_BLOCKING)) | |
946 | has_rcu = true; | |
947 | else | |
05707b64 | 948 | synchronize_srcu(hctx->srcu); |
1d9bd516 TH |
949 | |
950 | hctx->nr_expired = 0; | |
951 | } | |
952 | if (has_rcu) | |
953 | synchronize_rcu(); | |
954 | ||
955 | /* terminate the ones we won */ | |
956 | blk_mq_queue_tag_busy_iter(q, blk_mq_terminate_expired, NULL); | |
957 | } | |
958 | ||
81481eb4 CH |
959 | if (data.next_set) { |
960 | data.next = blk_rq_timeout(round_jiffies_up(data.next)); | |
961 | mod_timer(&q->timeout, data.next); | |
0d2602ca | 962 | } else { |
fcd36c36 BVA |
963 | /* |
964 | * Request timeouts are handled as a forward rolling timer. If | |
965 | * we end up here it means that no requests are pending and | |
966 | * also that no request has been pending for a while. Mark | |
967 | * each hctx as idle. | |
968 | */ | |
f054b56c ML |
969 | queue_for_each_hw_ctx(q, hctx, i) { |
970 | /* the hctx may be unmapped, so check it here */ | |
971 | if (blk_mq_hw_queue_mapped(hctx)) | |
972 | blk_mq_tag_idle(hctx); | |
973 | } | |
0d2602ca | 974 | } |
287922eb | 975 | blk_queue_exit(q); |
320ae51f JA |
976 | } |
977 | ||
88459642 OS |
978 | struct flush_busy_ctx_data { |
979 | struct blk_mq_hw_ctx *hctx; | |
980 | struct list_head *list; | |
981 | }; | |
982 | ||
983 | static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data) | |
984 | { | |
985 | struct flush_busy_ctx_data *flush_data = data; | |
986 | struct blk_mq_hw_ctx *hctx = flush_data->hctx; | |
987 | struct blk_mq_ctx *ctx = hctx->ctxs[bitnr]; | |
988 | ||
989 | sbitmap_clear_bit(sb, bitnr); | |
990 | spin_lock(&ctx->lock); | |
991 | list_splice_tail_init(&ctx->rq_list, flush_data->list); | |
992 | spin_unlock(&ctx->lock); | |
993 | return true; | |
994 | } | |
995 | ||
1429d7c9 JA |
996 | /* |
997 | * Process software queues that have been marked busy, splicing them | |
998 | * to the for-dispatch | |
999 | */ | |
2c3ad667 | 1000 | void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list) |
1429d7c9 | 1001 | { |
88459642 OS |
1002 | struct flush_busy_ctx_data data = { |
1003 | .hctx = hctx, | |
1004 | .list = list, | |
1005 | }; | |
1429d7c9 | 1006 | |
88459642 | 1007 | sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data); |
1429d7c9 | 1008 | } |
2c3ad667 | 1009 | EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs); |
1429d7c9 | 1010 | |
b347689f ML |
1011 | struct dispatch_rq_data { |
1012 | struct blk_mq_hw_ctx *hctx; | |
1013 | struct request *rq; | |
1014 | }; | |
1015 | ||
1016 | static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr, | |
1017 | void *data) | |
1018 | { | |
1019 | struct dispatch_rq_data *dispatch_data = data; | |
1020 | struct blk_mq_hw_ctx *hctx = dispatch_data->hctx; | |
1021 | struct blk_mq_ctx *ctx = hctx->ctxs[bitnr]; | |
1022 | ||
1023 | spin_lock(&ctx->lock); | |
1024 | if (unlikely(!list_empty(&ctx->rq_list))) { | |
1025 | dispatch_data->rq = list_entry_rq(ctx->rq_list.next); | |
1026 | list_del_init(&dispatch_data->rq->queuelist); | |
1027 | if (list_empty(&ctx->rq_list)) | |
1028 | sbitmap_clear_bit(sb, bitnr); | |
1029 | } | |
1030 | spin_unlock(&ctx->lock); | |
1031 | ||
1032 | return !dispatch_data->rq; | |
1033 | } | |
1034 | ||
1035 | struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx, | |
1036 | struct blk_mq_ctx *start) | |
1037 | { | |
1038 | unsigned off = start ? start->index_hw : 0; | |
1039 | struct dispatch_rq_data data = { | |
1040 | .hctx = hctx, | |
1041 | .rq = NULL, | |
1042 | }; | |
1043 | ||
1044 | __sbitmap_for_each_set(&hctx->ctx_map, off, | |
1045 | dispatch_rq_from_ctx, &data); | |
1046 | ||
1047 | return data.rq; | |
1048 | } | |
1049 | ||
703fd1c0 JA |
1050 | static inline unsigned int queued_to_index(unsigned int queued) |
1051 | { | |
1052 | if (!queued) | |
1053 | return 0; | |
1429d7c9 | 1054 | |
703fd1c0 | 1055 | return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1); |
1429d7c9 JA |
1056 | } |
1057 | ||
bd6737f1 JA |
1058 | bool blk_mq_get_driver_tag(struct request *rq, struct blk_mq_hw_ctx **hctx, |
1059 | bool wait) | |
bd166ef1 JA |
1060 | { |
1061 | struct blk_mq_alloc_data data = { | |
1062 | .q = rq->q, | |
bd166ef1 JA |
1063 | .hctx = blk_mq_map_queue(rq->q, rq->mq_ctx->cpu), |
1064 | .flags = wait ? 0 : BLK_MQ_REQ_NOWAIT, | |
1065 | }; | |
1066 | ||
5feeacdd JA |
1067 | might_sleep_if(wait); |
1068 | ||
81380ca1 OS |
1069 | if (rq->tag != -1) |
1070 | goto done; | |
bd166ef1 | 1071 | |
415b806d SG |
1072 | if (blk_mq_tag_is_reserved(data.hctx->sched_tags, rq->internal_tag)) |
1073 | data.flags |= BLK_MQ_REQ_RESERVED; | |
1074 | ||
bd166ef1 JA |
1075 | rq->tag = blk_mq_get_tag(&data); |
1076 | if (rq->tag >= 0) { | |
200e86b3 JA |
1077 | if (blk_mq_tag_busy(data.hctx)) { |
1078 | rq->rq_flags |= RQF_MQ_INFLIGHT; | |
1079 | atomic_inc(&data.hctx->nr_active); | |
1080 | } | |
bd166ef1 | 1081 | data.hctx->tags->rqs[rq->tag] = rq; |
bd166ef1 JA |
1082 | } |
1083 | ||
81380ca1 OS |
1084 | done: |
1085 | if (hctx) | |
1086 | *hctx = data.hctx; | |
1087 | return rq->tag != -1; | |
bd166ef1 JA |
1088 | } |
1089 | ||
eb619fdb JA |
1090 | static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode, |
1091 | int flags, void *key) | |
da55f2cc OS |
1092 | { |
1093 | struct blk_mq_hw_ctx *hctx; | |
1094 | ||
1095 | hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait); | |
1096 | ||
eb619fdb | 1097 | list_del_init(&wait->entry); |
da55f2cc OS |
1098 | blk_mq_run_hw_queue(hctx, true); |
1099 | return 1; | |
1100 | } | |
1101 | ||
f906a6a0 JA |
1102 | /* |
1103 | * Mark us waiting for a tag. For shared tags, this involves hooking us into | |
ee3e4de5 BVA |
1104 | * the tag wakeups. For non-shared tags, we can simply mark us needing a |
1105 | * restart. For both cases, take care to check the condition again after | |
f906a6a0 JA |
1106 | * marking us as waiting. |
1107 | */ | |
1108 | static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx **hctx, | |
1109 | struct request *rq) | |
da55f2cc | 1110 | { |
eb619fdb | 1111 | struct blk_mq_hw_ctx *this_hctx = *hctx; |
da55f2cc | 1112 | struct sbq_wait_state *ws; |
f906a6a0 JA |
1113 | wait_queue_entry_t *wait; |
1114 | bool ret; | |
da55f2cc | 1115 | |
c27d53fb | 1116 | if (!(this_hctx->flags & BLK_MQ_F_TAG_SHARED)) { |
f906a6a0 JA |
1117 | if (!test_bit(BLK_MQ_S_SCHED_RESTART, &this_hctx->state)) |
1118 | set_bit(BLK_MQ_S_SCHED_RESTART, &this_hctx->state); | |
f906a6a0 | 1119 | |
c27d53fb BVA |
1120 | /* |
1121 | * It's possible that a tag was freed in the window between the | |
1122 | * allocation failure and adding the hardware queue to the wait | |
1123 | * queue. | |
1124 | * | |
1125 | * Don't clear RESTART here, someone else could have set it. | |
1126 | * At most this will cost an extra queue run. | |
1127 | */ | |
1128 | return blk_mq_get_driver_tag(rq, hctx, false); | |
eb619fdb | 1129 | } |
eb619fdb | 1130 | |
c27d53fb BVA |
1131 | wait = &this_hctx->dispatch_wait; |
1132 | if (!list_empty_careful(&wait->entry)) | |
1133 | return false; | |
1134 | ||
1135 | spin_lock(&this_hctx->lock); | |
1136 | if (!list_empty(&wait->entry)) { | |
1137 | spin_unlock(&this_hctx->lock); | |
1138 | return false; | |
eb619fdb JA |
1139 | } |
1140 | ||
c27d53fb BVA |
1141 | ws = bt_wait_ptr(&this_hctx->tags->bitmap_tags, this_hctx); |
1142 | add_wait_queue(&ws->wait, wait); | |
1143 | ||
da55f2cc | 1144 | /* |
eb619fdb JA |
1145 | * It's possible that a tag was freed in the window between the |
1146 | * allocation failure and adding the hardware queue to the wait | |
1147 | * queue. | |
da55f2cc | 1148 | */ |
f906a6a0 | 1149 | ret = blk_mq_get_driver_tag(rq, hctx, false); |
c27d53fb | 1150 | if (!ret) { |
eb619fdb | 1151 | spin_unlock(&this_hctx->lock); |
c27d53fb | 1152 | return false; |
eb619fdb | 1153 | } |
c27d53fb BVA |
1154 | |
1155 | /* | |
1156 | * We got a tag, remove ourselves from the wait queue to ensure | |
1157 | * someone else gets the wakeup. | |
1158 | */ | |
1159 | spin_lock_irq(&ws->wait.lock); | |
1160 | list_del_init(&wait->entry); | |
1161 | spin_unlock_irq(&ws->wait.lock); | |
1162 | spin_unlock(&this_hctx->lock); | |
1163 | ||
1164 | return true; | |
da55f2cc OS |
1165 | } |
1166 | ||
86ff7c2a ML |
1167 | #define BLK_MQ_RESOURCE_DELAY 3 /* ms units */ |
1168 | ||
de148297 | 1169 | bool blk_mq_dispatch_rq_list(struct request_queue *q, struct list_head *list, |
eb619fdb | 1170 | bool got_budget) |
320ae51f | 1171 | { |
81380ca1 | 1172 | struct blk_mq_hw_ctx *hctx; |
6d6f167c | 1173 | struct request *rq, *nxt; |
eb619fdb | 1174 | bool no_tag = false; |
fc17b653 | 1175 | int errors, queued; |
86ff7c2a | 1176 | blk_status_t ret = BLK_STS_OK; |
320ae51f | 1177 | |
81380ca1 OS |
1178 | if (list_empty(list)) |
1179 | return false; | |
1180 | ||
de148297 ML |
1181 | WARN_ON(!list_is_singular(list) && got_budget); |
1182 | ||
320ae51f JA |
1183 | /* |
1184 | * Now process all the entries, sending them to the driver. | |
1185 | */ | |
93efe981 | 1186 | errors = queued = 0; |
81380ca1 | 1187 | do { |
74c45052 | 1188 | struct blk_mq_queue_data bd; |
320ae51f | 1189 | |
f04c3df3 | 1190 | rq = list_first_entry(list, struct request, queuelist); |
bd166ef1 | 1191 | if (!blk_mq_get_driver_tag(rq, &hctx, false)) { |
3c782d67 | 1192 | /* |
da55f2cc | 1193 | * The initial allocation attempt failed, so we need to |
eb619fdb JA |
1194 | * rerun the hardware queue when a tag is freed. The |
1195 | * waitqueue takes care of that. If the queue is run | |
1196 | * before we add this entry back on the dispatch list, | |
1197 | * we'll re-run it below. | |
3c782d67 | 1198 | */ |
f906a6a0 | 1199 | if (!blk_mq_mark_tag_wait(&hctx, rq)) { |
de148297 ML |
1200 | if (got_budget) |
1201 | blk_mq_put_dispatch_budget(hctx); | |
f906a6a0 JA |
1202 | /* |
1203 | * For non-shared tags, the RESTART check | |
1204 | * will suffice. | |
1205 | */ | |
1206 | if (hctx->flags & BLK_MQ_F_TAG_SHARED) | |
1207 | no_tag = true; | |
de148297 ML |
1208 | break; |
1209 | } | |
1210 | } | |
1211 | ||
0c6af1cc ML |
1212 | if (!got_budget && !blk_mq_get_dispatch_budget(hctx)) { |
1213 | blk_mq_put_driver_tag(rq); | |
88022d72 | 1214 | break; |
0c6af1cc | 1215 | } |
da55f2cc | 1216 | |
320ae51f | 1217 | list_del_init(&rq->queuelist); |
320ae51f | 1218 | |
74c45052 | 1219 | bd.rq = rq; |
113285b4 JA |
1220 | |
1221 | /* | |
1222 | * Flag last if we have no more requests, or if we have more | |
1223 | * but can't assign a driver tag to it. | |
1224 | */ | |
1225 | if (list_empty(list)) | |
1226 | bd.last = true; | |
1227 | else { | |
113285b4 JA |
1228 | nxt = list_first_entry(list, struct request, queuelist); |
1229 | bd.last = !blk_mq_get_driver_tag(nxt, NULL, false); | |
1230 | } | |
74c45052 JA |
1231 | |
1232 | ret = q->mq_ops->queue_rq(hctx, &bd); | |
86ff7c2a | 1233 | if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE) { |
6d6f167c JW |
1234 | /* |
1235 | * If an I/O scheduler has been configured and we got a | |
ff821d27 JA |
1236 | * driver tag for the next request already, free it |
1237 | * again. | |
6d6f167c JW |
1238 | */ |
1239 | if (!list_empty(list)) { | |
1240 | nxt = list_first_entry(list, struct request, queuelist); | |
1241 | blk_mq_put_driver_tag(nxt); | |
1242 | } | |
f04c3df3 | 1243 | list_add(&rq->queuelist, list); |
ed0791b2 | 1244 | __blk_mq_requeue_request(rq); |
320ae51f | 1245 | break; |
fc17b653 CH |
1246 | } |
1247 | ||
1248 | if (unlikely(ret != BLK_STS_OK)) { | |
93efe981 | 1249 | errors++; |
2a842aca | 1250 | blk_mq_end_request(rq, BLK_STS_IOERR); |
fc17b653 | 1251 | continue; |
320ae51f JA |
1252 | } |
1253 | ||
fc17b653 | 1254 | queued++; |
81380ca1 | 1255 | } while (!list_empty(list)); |
320ae51f | 1256 | |
703fd1c0 | 1257 | hctx->dispatched[queued_to_index(queued)]++; |
320ae51f JA |
1258 | |
1259 | /* | |
1260 | * Any items that need requeuing? Stuff them into hctx->dispatch, | |
1261 | * that is where we will continue on next queue run. | |
1262 | */ | |
f04c3df3 | 1263 | if (!list_empty(list)) { |
86ff7c2a ML |
1264 | bool needs_restart; |
1265 | ||
320ae51f | 1266 | spin_lock(&hctx->lock); |
c13660a0 | 1267 | list_splice_init(list, &hctx->dispatch); |
320ae51f | 1268 | spin_unlock(&hctx->lock); |
f04c3df3 | 1269 | |
9ba52e58 | 1270 | /* |
710c785f BVA |
1271 | * If SCHED_RESTART was set by the caller of this function and |
1272 | * it is no longer set that means that it was cleared by another | |
1273 | * thread and hence that a queue rerun is needed. | |
9ba52e58 | 1274 | * |
eb619fdb JA |
1275 | * If 'no_tag' is set, that means that we failed getting |
1276 | * a driver tag with an I/O scheduler attached. If our dispatch | |
1277 | * waitqueue is no longer active, ensure that we run the queue | |
1278 | * AFTER adding our entries back to the list. | |
bd166ef1 | 1279 | * |
710c785f BVA |
1280 | * If no I/O scheduler has been configured it is possible that |
1281 | * the hardware queue got stopped and restarted before requests | |
1282 | * were pushed back onto the dispatch list. Rerun the queue to | |
1283 | * avoid starvation. Notes: | |
1284 | * - blk_mq_run_hw_queue() checks whether or not a queue has | |
1285 | * been stopped before rerunning a queue. | |
1286 | * - Some but not all block drivers stop a queue before | |
fc17b653 | 1287 | * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq |
710c785f | 1288 | * and dm-rq. |
86ff7c2a ML |
1289 | * |
1290 | * If driver returns BLK_STS_RESOURCE and SCHED_RESTART | |
1291 | * bit is set, run queue after a delay to avoid IO stalls | |
1292 | * that could otherwise occur if the queue is idle. | |
bd166ef1 | 1293 | */ |
86ff7c2a ML |
1294 | needs_restart = blk_mq_sched_needs_restart(hctx); |
1295 | if (!needs_restart || | |
eb619fdb | 1296 | (no_tag && list_empty_careful(&hctx->dispatch_wait.entry))) |
bd166ef1 | 1297 | blk_mq_run_hw_queue(hctx, true); |
86ff7c2a ML |
1298 | else if (needs_restart && (ret == BLK_STS_RESOURCE)) |
1299 | blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY); | |
320ae51f | 1300 | } |
f04c3df3 | 1301 | |
93efe981 | 1302 | return (queued + errors) != 0; |
f04c3df3 JA |
1303 | } |
1304 | ||
6a83e74d BVA |
1305 | static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx) |
1306 | { | |
1307 | int srcu_idx; | |
1308 | ||
b7a71e66 JA |
1309 | /* |
1310 | * We should be running this queue from one of the CPUs that | |
1311 | * are mapped to it. | |
7df938fb ML |
1312 | * |
1313 | * There are at least two related races now between setting | |
1314 | * hctx->next_cpu from blk_mq_hctx_next_cpu() and running | |
1315 | * __blk_mq_run_hw_queue(): | |
1316 | * | |
1317 | * - hctx->next_cpu is found offline in blk_mq_hctx_next_cpu(), | |
1318 | * but later it becomes online, then this warning is harmless | |
1319 | * at all | |
1320 | * | |
1321 | * - hctx->next_cpu is found online in blk_mq_hctx_next_cpu(), | |
1322 | * but later it becomes offline, then the warning can't be | |
1323 | * triggered, and we depend on blk-mq timeout handler to | |
1324 | * handle dispatched requests to this hctx | |
b7a71e66 | 1325 | */ |
7df938fb ML |
1326 | if (!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) && |
1327 | cpu_online(hctx->next_cpu)) { | |
1328 | printk(KERN_WARNING "run queue from wrong CPU %d, hctx %s\n", | |
1329 | raw_smp_processor_id(), | |
1330 | cpumask_empty(hctx->cpumask) ? "inactive": "active"); | |
1331 | dump_stack(); | |
1332 | } | |
6a83e74d | 1333 | |
b7a71e66 JA |
1334 | /* |
1335 | * We can't run the queue inline with ints disabled. Ensure that | |
1336 | * we catch bad users of this early. | |
1337 | */ | |
1338 | WARN_ON_ONCE(in_interrupt()); | |
1339 | ||
04ced159 | 1340 | might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING); |
bf4907c0 | 1341 | |
04ced159 JA |
1342 | hctx_lock(hctx, &srcu_idx); |
1343 | blk_mq_sched_dispatch_requests(hctx); | |
1344 | hctx_unlock(hctx, srcu_idx); | |
6a83e74d BVA |
1345 | } |
1346 | ||
506e931f JA |
1347 | /* |
1348 | * It'd be great if the workqueue API had a way to pass | |
1349 | * in a mask and had some smarts for more clever placement. | |
1350 | * For now we just round-robin here, switching for every | |
1351 | * BLK_MQ_CPU_WORK_BATCH queued items. | |
1352 | */ | |
1353 | static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx) | |
1354 | { | |
7bed4595 ML |
1355 | bool tried = false; |
1356 | ||
b657d7e6 CH |
1357 | if (hctx->queue->nr_hw_queues == 1) |
1358 | return WORK_CPU_UNBOUND; | |
506e931f JA |
1359 | |
1360 | if (--hctx->next_cpu_batch <= 0) { | |
c02ebfdd | 1361 | int next_cpu; |
7bed4595 | 1362 | select_cpu: |
20e4d813 CH |
1363 | next_cpu = cpumask_next_and(hctx->next_cpu, hctx->cpumask, |
1364 | cpu_online_mask); | |
506e931f | 1365 | if (next_cpu >= nr_cpu_ids) |
20e4d813 | 1366 | next_cpu = cpumask_first_and(hctx->cpumask,cpu_online_mask); |
506e931f | 1367 | |
7bed4595 ML |
1368 | /* |
1369 | * No online CPU is found, so have to make sure hctx->next_cpu | |
1370 | * is set correctly for not breaking workqueue. | |
1371 | */ | |
1372 | if (next_cpu >= nr_cpu_ids) | |
1373 | hctx->next_cpu = cpumask_first(hctx->cpumask); | |
1374 | else | |
1375 | hctx->next_cpu = next_cpu; | |
506e931f JA |
1376 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; |
1377 | } | |
1378 | ||
7bed4595 ML |
1379 | /* |
1380 | * Do unbound schedule if we can't find a online CPU for this hctx, | |
1381 | * and it should only happen in the path of handling CPU DEAD. | |
1382 | */ | |
1383 | if (!cpu_online(hctx->next_cpu)) { | |
1384 | if (!tried) { | |
1385 | tried = true; | |
1386 | goto select_cpu; | |
1387 | } | |
1388 | ||
1389 | /* | |
1390 | * Make sure to re-select CPU next time once after CPUs | |
1391 | * in hctx->cpumask become online again. | |
1392 | */ | |
1393 | hctx->next_cpu_batch = 1; | |
1394 | return WORK_CPU_UNBOUND; | |
1395 | } | |
b657d7e6 | 1396 | return hctx->next_cpu; |
506e931f JA |
1397 | } |
1398 | ||
7587a5ae BVA |
1399 | static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async, |
1400 | unsigned long msecs) | |
320ae51f | 1401 | { |
5435c023 BVA |
1402 | if (WARN_ON_ONCE(!blk_mq_hw_queue_mapped(hctx))) |
1403 | return; | |
1404 | ||
1405 | if (unlikely(blk_mq_hctx_stopped(hctx))) | |
320ae51f JA |
1406 | return; |
1407 | ||
1b792f2f | 1408 | if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) { |
2a90d4aa PB |
1409 | int cpu = get_cpu(); |
1410 | if (cpumask_test_cpu(cpu, hctx->cpumask)) { | |
398205b8 | 1411 | __blk_mq_run_hw_queue(hctx); |
2a90d4aa | 1412 | put_cpu(); |
398205b8 PB |
1413 | return; |
1414 | } | |
e4043dcf | 1415 | |
2a90d4aa | 1416 | put_cpu(); |
e4043dcf | 1417 | } |
398205b8 | 1418 | |
ae943d20 BVA |
1419 | kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work, |
1420 | msecs_to_jiffies(msecs)); | |
7587a5ae BVA |
1421 | } |
1422 | ||
1423 | void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs) | |
1424 | { | |
1425 | __blk_mq_delay_run_hw_queue(hctx, true, msecs); | |
1426 | } | |
1427 | EXPORT_SYMBOL(blk_mq_delay_run_hw_queue); | |
1428 | ||
79f720a7 | 1429 | bool blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) |
7587a5ae | 1430 | { |
24f5a90f ML |
1431 | int srcu_idx; |
1432 | bool need_run; | |
1433 | ||
1434 | /* | |
1435 | * When queue is quiesced, we may be switching io scheduler, or | |
1436 | * updating nr_hw_queues, or other things, and we can't run queue | |
1437 | * any more, even __blk_mq_hctx_has_pending() can't be called safely. | |
1438 | * | |
1439 | * And queue will be rerun in blk_mq_unquiesce_queue() if it is | |
1440 | * quiesced. | |
1441 | */ | |
04ced159 JA |
1442 | hctx_lock(hctx, &srcu_idx); |
1443 | need_run = !blk_queue_quiesced(hctx->queue) && | |
1444 | blk_mq_hctx_has_pending(hctx); | |
1445 | hctx_unlock(hctx, srcu_idx); | |
24f5a90f ML |
1446 | |
1447 | if (need_run) { | |
79f720a7 JA |
1448 | __blk_mq_delay_run_hw_queue(hctx, async, 0); |
1449 | return true; | |
1450 | } | |
1451 | ||
1452 | return false; | |
320ae51f | 1453 | } |
5b727272 | 1454 | EXPORT_SYMBOL(blk_mq_run_hw_queue); |
320ae51f | 1455 | |
b94ec296 | 1456 | void blk_mq_run_hw_queues(struct request_queue *q, bool async) |
320ae51f JA |
1457 | { |
1458 | struct blk_mq_hw_ctx *hctx; | |
1459 | int i; | |
1460 | ||
1461 | queue_for_each_hw_ctx(q, hctx, i) { | |
79f720a7 | 1462 | if (blk_mq_hctx_stopped(hctx)) |
320ae51f JA |
1463 | continue; |
1464 | ||
b94ec296 | 1465 | blk_mq_run_hw_queue(hctx, async); |
320ae51f JA |
1466 | } |
1467 | } | |
b94ec296 | 1468 | EXPORT_SYMBOL(blk_mq_run_hw_queues); |
320ae51f | 1469 | |
fd001443 BVA |
1470 | /** |
1471 | * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped | |
1472 | * @q: request queue. | |
1473 | * | |
1474 | * The caller is responsible for serializing this function against | |
1475 | * blk_mq_{start,stop}_hw_queue(). | |
1476 | */ | |
1477 | bool blk_mq_queue_stopped(struct request_queue *q) | |
1478 | { | |
1479 | struct blk_mq_hw_ctx *hctx; | |
1480 | int i; | |
1481 | ||
1482 | queue_for_each_hw_ctx(q, hctx, i) | |
1483 | if (blk_mq_hctx_stopped(hctx)) | |
1484 | return true; | |
1485 | ||
1486 | return false; | |
1487 | } | |
1488 | EXPORT_SYMBOL(blk_mq_queue_stopped); | |
1489 | ||
39a70c76 ML |
1490 | /* |
1491 | * This function is often used for pausing .queue_rq() by driver when | |
1492 | * there isn't enough resource or some conditions aren't satisfied, and | |
4d606219 | 1493 | * BLK_STS_RESOURCE is usually returned. |
39a70c76 ML |
1494 | * |
1495 | * We do not guarantee that dispatch can be drained or blocked | |
1496 | * after blk_mq_stop_hw_queue() returns. Please use | |
1497 | * blk_mq_quiesce_queue() for that requirement. | |
1498 | */ | |
2719aa21 JA |
1499 | void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx) |
1500 | { | |
641a9ed6 | 1501 | cancel_delayed_work(&hctx->run_work); |
280d45f6 | 1502 | |
641a9ed6 | 1503 | set_bit(BLK_MQ_S_STOPPED, &hctx->state); |
2719aa21 | 1504 | } |
641a9ed6 | 1505 | EXPORT_SYMBOL(blk_mq_stop_hw_queue); |
2719aa21 | 1506 | |
39a70c76 ML |
1507 | /* |
1508 | * This function is often used for pausing .queue_rq() by driver when | |
1509 | * there isn't enough resource or some conditions aren't satisfied, and | |
4d606219 | 1510 | * BLK_STS_RESOURCE is usually returned. |
39a70c76 ML |
1511 | * |
1512 | * We do not guarantee that dispatch can be drained or blocked | |
1513 | * after blk_mq_stop_hw_queues() returns. Please use | |
1514 | * blk_mq_quiesce_queue() for that requirement. | |
1515 | */ | |
2719aa21 JA |
1516 | void blk_mq_stop_hw_queues(struct request_queue *q) |
1517 | { | |
641a9ed6 ML |
1518 | struct blk_mq_hw_ctx *hctx; |
1519 | int i; | |
1520 | ||
1521 | queue_for_each_hw_ctx(q, hctx, i) | |
1522 | blk_mq_stop_hw_queue(hctx); | |
280d45f6 CH |
1523 | } |
1524 | EXPORT_SYMBOL(blk_mq_stop_hw_queues); | |
1525 | ||
320ae51f JA |
1526 | void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx) |
1527 | { | |
1528 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
e4043dcf | 1529 | |
0ffbce80 | 1530 | blk_mq_run_hw_queue(hctx, false); |
320ae51f JA |
1531 | } |
1532 | EXPORT_SYMBOL(blk_mq_start_hw_queue); | |
1533 | ||
2f268556 CH |
1534 | void blk_mq_start_hw_queues(struct request_queue *q) |
1535 | { | |
1536 | struct blk_mq_hw_ctx *hctx; | |
1537 | int i; | |
1538 | ||
1539 | queue_for_each_hw_ctx(q, hctx, i) | |
1540 | blk_mq_start_hw_queue(hctx); | |
1541 | } | |
1542 | EXPORT_SYMBOL(blk_mq_start_hw_queues); | |
1543 | ||
ae911c5e JA |
1544 | void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) |
1545 | { | |
1546 | if (!blk_mq_hctx_stopped(hctx)) | |
1547 | return; | |
1548 | ||
1549 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
1550 | blk_mq_run_hw_queue(hctx, async); | |
1551 | } | |
1552 | EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue); | |
1553 | ||
1b4a3258 | 1554 | void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async) |
320ae51f JA |
1555 | { |
1556 | struct blk_mq_hw_ctx *hctx; | |
1557 | int i; | |
1558 | ||
ae911c5e JA |
1559 | queue_for_each_hw_ctx(q, hctx, i) |
1560 | blk_mq_start_stopped_hw_queue(hctx, async); | |
320ae51f JA |
1561 | } |
1562 | EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues); | |
1563 | ||
70f4db63 | 1564 | static void blk_mq_run_work_fn(struct work_struct *work) |
320ae51f JA |
1565 | { |
1566 | struct blk_mq_hw_ctx *hctx; | |
1567 | ||
9f993737 | 1568 | hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work); |
320ae51f | 1569 | |
21c6e939 JA |
1570 | /* |
1571 | * If we are stopped, don't run the queue. The exception is if | |
1572 | * BLK_MQ_S_START_ON_RUN is set. For that case, we auto-clear | |
1573 | * the STOPPED bit and run it. | |
1574 | */ | |
1575 | if (test_bit(BLK_MQ_S_STOPPED, &hctx->state)) { | |
1576 | if (!test_bit(BLK_MQ_S_START_ON_RUN, &hctx->state)) | |
1577 | return; | |
7587a5ae | 1578 | |
21c6e939 JA |
1579 | clear_bit(BLK_MQ_S_START_ON_RUN, &hctx->state); |
1580 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
1581 | } | |
7587a5ae BVA |
1582 | |
1583 | __blk_mq_run_hw_queue(hctx); | |
1584 | } | |
1585 | ||
70f4db63 CH |
1586 | |
1587 | void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs) | |
1588 | { | |
5435c023 | 1589 | if (WARN_ON_ONCE(!blk_mq_hw_queue_mapped(hctx))) |
19c66e59 | 1590 | return; |
70f4db63 | 1591 | |
21c6e939 JA |
1592 | /* |
1593 | * Stop the hw queue, then modify currently delayed work. | |
1594 | * This should prevent us from running the queue prematurely. | |
1595 | * Mark the queue as auto-clearing STOPPED when it runs. | |
1596 | */ | |
7e79dadc | 1597 | blk_mq_stop_hw_queue(hctx); |
21c6e939 JA |
1598 | set_bit(BLK_MQ_S_START_ON_RUN, &hctx->state); |
1599 | kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), | |
1600 | &hctx->run_work, | |
1601 | msecs_to_jiffies(msecs)); | |
70f4db63 CH |
1602 | } |
1603 | EXPORT_SYMBOL(blk_mq_delay_queue); | |
1604 | ||
cfd0c552 | 1605 | static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx, |
cfd0c552 ML |
1606 | struct request *rq, |
1607 | bool at_head) | |
320ae51f | 1608 | { |
e57690fe JA |
1609 | struct blk_mq_ctx *ctx = rq->mq_ctx; |
1610 | ||
7b607814 BVA |
1611 | lockdep_assert_held(&ctx->lock); |
1612 | ||
01b983c9 JA |
1613 | trace_block_rq_insert(hctx->queue, rq); |
1614 | ||
72a0a36e CH |
1615 | if (at_head) |
1616 | list_add(&rq->queuelist, &ctx->rq_list); | |
1617 | else | |
1618 | list_add_tail(&rq->queuelist, &ctx->rq_list); | |
cfd0c552 | 1619 | } |
4bb659b1 | 1620 | |
2c3ad667 JA |
1621 | void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq, |
1622 | bool at_head) | |
cfd0c552 ML |
1623 | { |
1624 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
1625 | ||
7b607814 BVA |
1626 | lockdep_assert_held(&ctx->lock); |
1627 | ||
e57690fe | 1628 | __blk_mq_insert_req_list(hctx, rq, at_head); |
320ae51f | 1629 | blk_mq_hctx_mark_pending(hctx, ctx); |
320ae51f JA |
1630 | } |
1631 | ||
157f377b JA |
1632 | /* |
1633 | * Should only be used carefully, when the caller knows we want to | |
1634 | * bypass a potential IO scheduler on the target device. | |
1635 | */ | |
b0850297 | 1636 | void blk_mq_request_bypass_insert(struct request *rq, bool run_queue) |
157f377b JA |
1637 | { |
1638 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
1639 | struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(rq->q, ctx->cpu); | |
1640 | ||
1641 | spin_lock(&hctx->lock); | |
1642 | list_add_tail(&rq->queuelist, &hctx->dispatch); | |
1643 | spin_unlock(&hctx->lock); | |
1644 | ||
b0850297 ML |
1645 | if (run_queue) |
1646 | blk_mq_run_hw_queue(hctx, false); | |
157f377b JA |
1647 | } |
1648 | ||
bd166ef1 JA |
1649 | void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx, |
1650 | struct list_head *list) | |
320ae51f JA |
1651 | |
1652 | { | |
320ae51f JA |
1653 | /* |
1654 | * preemption doesn't flush plug list, so it's possible ctx->cpu is | |
1655 | * offline now | |
1656 | */ | |
1657 | spin_lock(&ctx->lock); | |
1658 | while (!list_empty(list)) { | |
1659 | struct request *rq; | |
1660 | ||
1661 | rq = list_first_entry(list, struct request, queuelist); | |
e57690fe | 1662 | BUG_ON(rq->mq_ctx != ctx); |
320ae51f | 1663 | list_del_init(&rq->queuelist); |
e57690fe | 1664 | __blk_mq_insert_req_list(hctx, rq, false); |
320ae51f | 1665 | } |
cfd0c552 | 1666 | blk_mq_hctx_mark_pending(hctx, ctx); |
320ae51f | 1667 | spin_unlock(&ctx->lock); |
320ae51f JA |
1668 | } |
1669 | ||
1670 | static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b) | |
1671 | { | |
1672 | struct request *rqa = container_of(a, struct request, queuelist); | |
1673 | struct request *rqb = container_of(b, struct request, queuelist); | |
1674 | ||
1675 | return !(rqa->mq_ctx < rqb->mq_ctx || | |
1676 | (rqa->mq_ctx == rqb->mq_ctx && | |
1677 | blk_rq_pos(rqa) < blk_rq_pos(rqb))); | |
1678 | } | |
1679 | ||
1680 | void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule) | |
1681 | { | |
1682 | struct blk_mq_ctx *this_ctx; | |
1683 | struct request_queue *this_q; | |
1684 | struct request *rq; | |
1685 | LIST_HEAD(list); | |
1686 | LIST_HEAD(ctx_list); | |
1687 | unsigned int depth; | |
1688 | ||
1689 | list_splice_init(&plug->mq_list, &list); | |
1690 | ||
1691 | list_sort(NULL, &list, plug_ctx_cmp); | |
1692 | ||
1693 | this_q = NULL; | |
1694 | this_ctx = NULL; | |
1695 | depth = 0; | |
1696 | ||
1697 | while (!list_empty(&list)) { | |
1698 | rq = list_entry_rq(list.next); | |
1699 | list_del_init(&rq->queuelist); | |
1700 | BUG_ON(!rq->q); | |
1701 | if (rq->mq_ctx != this_ctx) { | |
1702 | if (this_ctx) { | |
bd166ef1 JA |
1703 | trace_block_unplug(this_q, depth, from_schedule); |
1704 | blk_mq_sched_insert_requests(this_q, this_ctx, | |
1705 | &ctx_list, | |
1706 | from_schedule); | |
320ae51f JA |
1707 | } |
1708 | ||
1709 | this_ctx = rq->mq_ctx; | |
1710 | this_q = rq->q; | |
1711 | depth = 0; | |
1712 | } | |
1713 | ||
1714 | depth++; | |
1715 | list_add_tail(&rq->queuelist, &ctx_list); | |
1716 | } | |
1717 | ||
1718 | /* | |
1719 | * If 'this_ctx' is set, we know we have entries to complete | |
1720 | * on 'ctx_list'. Do those. | |
1721 | */ | |
1722 | if (this_ctx) { | |
bd166ef1 JA |
1723 | trace_block_unplug(this_q, depth, from_schedule); |
1724 | blk_mq_sched_insert_requests(this_q, this_ctx, &ctx_list, | |
1725 | from_schedule); | |
320ae51f JA |
1726 | } |
1727 | } | |
1728 | ||
1729 | static void blk_mq_bio_to_request(struct request *rq, struct bio *bio) | |
1730 | { | |
da8d7f07 | 1731 | blk_init_request_from_bio(rq, bio); |
4b570521 | 1732 | |
85acb3ba SL |
1733 | blk_rq_set_rl(rq, blk_get_rl(rq->q, bio)); |
1734 | ||
6e85eaf3 | 1735 | blk_account_io_start(rq, true); |
320ae51f JA |
1736 | } |
1737 | ||
ab42f35d ML |
1738 | static inline void blk_mq_queue_io(struct blk_mq_hw_ctx *hctx, |
1739 | struct blk_mq_ctx *ctx, | |
1740 | struct request *rq) | |
1741 | { | |
1742 | spin_lock(&ctx->lock); | |
1743 | __blk_mq_insert_request(hctx, rq, false); | |
1744 | spin_unlock(&ctx->lock); | |
07068d5b | 1745 | } |
14ec77f3 | 1746 | |
fd2d3326 JA |
1747 | static blk_qc_t request_to_qc_t(struct blk_mq_hw_ctx *hctx, struct request *rq) |
1748 | { | |
bd166ef1 JA |
1749 | if (rq->tag != -1) |
1750 | return blk_tag_to_qc_t(rq->tag, hctx->queue_num, false); | |
1751 | ||
1752 | return blk_tag_to_qc_t(rq->internal_tag, hctx->queue_num, true); | |
fd2d3326 JA |
1753 | } |
1754 | ||
0f95549c MS |
1755 | static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx, |
1756 | struct request *rq, | |
1757 | blk_qc_t *cookie) | |
f984df1f | 1758 | { |
f984df1f | 1759 | struct request_queue *q = rq->q; |
f984df1f SL |
1760 | struct blk_mq_queue_data bd = { |
1761 | .rq = rq, | |
d945a365 | 1762 | .last = true, |
f984df1f | 1763 | }; |
bd166ef1 | 1764 | blk_qc_t new_cookie; |
f06345ad | 1765 | blk_status_t ret; |
0f95549c MS |
1766 | |
1767 | new_cookie = request_to_qc_t(hctx, rq); | |
1768 | ||
1769 | /* | |
1770 | * For OK queue, we are done. For error, caller may kill it. | |
1771 | * Any other error (busy), just add it to our list as we | |
1772 | * previously would have done. | |
1773 | */ | |
1774 | ret = q->mq_ops->queue_rq(hctx, &bd); | |
1775 | switch (ret) { | |
1776 | case BLK_STS_OK: | |
1777 | *cookie = new_cookie; | |
1778 | break; | |
1779 | case BLK_STS_RESOURCE: | |
86ff7c2a | 1780 | case BLK_STS_DEV_RESOURCE: |
0f95549c MS |
1781 | __blk_mq_requeue_request(rq); |
1782 | break; | |
1783 | default: | |
1784 | *cookie = BLK_QC_T_NONE; | |
1785 | break; | |
1786 | } | |
1787 | ||
1788 | return ret; | |
1789 | } | |
1790 | ||
0f95549c MS |
1791 | static blk_status_t __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx, |
1792 | struct request *rq, | |
396eaf21 ML |
1793 | blk_qc_t *cookie, |
1794 | bool bypass_insert) | |
0f95549c MS |
1795 | { |
1796 | struct request_queue *q = rq->q; | |
d964f04a ML |
1797 | bool run_queue = true; |
1798 | ||
23d4ee19 ML |
1799 | /* |
1800 | * RCU or SRCU read lock is needed before checking quiesced flag. | |
1801 | * | |
1802 | * When queue is stopped or quiesced, ignore 'bypass_insert' from | |
c77ff7fd | 1803 | * blk_mq_request_issue_directly(), and return BLK_STS_OK to caller, |
23d4ee19 ML |
1804 | * and avoid driver to try to dispatch again. |
1805 | */ | |
f4560ffe | 1806 | if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)) { |
d964f04a | 1807 | run_queue = false; |
23d4ee19 | 1808 | bypass_insert = false; |
d964f04a ML |
1809 | goto insert; |
1810 | } | |
f984df1f | 1811 | |
396eaf21 | 1812 | if (q->elevator && !bypass_insert) |
2253efc8 BVA |
1813 | goto insert; |
1814 | ||
d964f04a | 1815 | if (!blk_mq_get_driver_tag(rq, NULL, false)) |
bd166ef1 JA |
1816 | goto insert; |
1817 | ||
88022d72 | 1818 | if (!blk_mq_get_dispatch_budget(hctx)) { |
de148297 ML |
1819 | blk_mq_put_driver_tag(rq); |
1820 | goto insert; | |
88022d72 | 1821 | } |
de148297 | 1822 | |
0f95549c | 1823 | return __blk_mq_issue_directly(hctx, rq, cookie); |
2253efc8 | 1824 | insert: |
396eaf21 ML |
1825 | if (bypass_insert) |
1826 | return BLK_STS_RESOURCE; | |
0f95549c | 1827 | |
23d4ee19 | 1828 | blk_mq_sched_insert_request(rq, false, run_queue, false); |
0f95549c | 1829 | return BLK_STS_OK; |
f984df1f SL |
1830 | } |
1831 | ||
5eb6126e CH |
1832 | static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx, |
1833 | struct request *rq, blk_qc_t *cookie) | |
1834 | { | |
0f95549c | 1835 | blk_status_t ret; |
04ced159 | 1836 | int srcu_idx; |
bf4907c0 | 1837 | |
04ced159 | 1838 | might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING); |
bf4907c0 | 1839 | |
04ced159 | 1840 | hctx_lock(hctx, &srcu_idx); |
0f95549c | 1841 | |
396eaf21 | 1842 | ret = __blk_mq_try_issue_directly(hctx, rq, cookie, false); |
86ff7c2a | 1843 | if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE) |
23d4ee19 | 1844 | blk_mq_sched_insert_request(rq, false, true, false); |
0f95549c MS |
1845 | else if (ret != BLK_STS_OK) |
1846 | blk_mq_end_request(rq, ret); | |
1847 | ||
04ced159 | 1848 | hctx_unlock(hctx, srcu_idx); |
5eb6126e CH |
1849 | } |
1850 | ||
c77ff7fd | 1851 | blk_status_t blk_mq_request_issue_directly(struct request *rq) |
396eaf21 ML |
1852 | { |
1853 | blk_status_t ret; | |
1854 | int srcu_idx; | |
1855 | blk_qc_t unused_cookie; | |
1856 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
1857 | struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(rq->q, ctx->cpu); | |
1858 | ||
1859 | hctx_lock(hctx, &srcu_idx); | |
1860 | ret = __blk_mq_try_issue_directly(hctx, rq, &unused_cookie, true); | |
1861 | hctx_unlock(hctx, srcu_idx); | |
1862 | ||
1863 | return ret; | |
5eb6126e CH |
1864 | } |
1865 | ||
dece1635 | 1866 | static blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio) |
07068d5b | 1867 | { |
ef295ecf | 1868 | const int is_sync = op_is_sync(bio->bi_opf); |
f73f44eb | 1869 | const int is_flush_fua = op_is_flush(bio->bi_opf); |
5a797e00 | 1870 | struct blk_mq_alloc_data data = { .flags = 0 }; |
07068d5b | 1871 | struct request *rq; |
5eb6126e | 1872 | unsigned int request_count = 0; |
f984df1f | 1873 | struct blk_plug *plug; |
5b3f341f | 1874 | struct request *same_queue_rq = NULL; |
7b371636 | 1875 | blk_qc_t cookie; |
87760e5e | 1876 | unsigned int wb_acct; |
07068d5b JA |
1877 | |
1878 | blk_queue_bounce(q, &bio); | |
1879 | ||
af67c31f | 1880 | blk_queue_split(q, &bio); |
f36ea50c | 1881 | |
e23947bd | 1882 | if (!bio_integrity_prep(bio)) |
dece1635 | 1883 | return BLK_QC_T_NONE; |
07068d5b | 1884 | |
87c279e6 OS |
1885 | if (!is_flush_fua && !blk_queue_nomerges(q) && |
1886 | blk_attempt_plug_merge(q, bio, &request_count, &same_queue_rq)) | |
1887 | return BLK_QC_T_NONE; | |
f984df1f | 1888 | |
bd166ef1 JA |
1889 | if (blk_mq_sched_bio_merge(q, bio)) |
1890 | return BLK_QC_T_NONE; | |
1891 | ||
87760e5e JA |
1892 | wb_acct = wbt_wait(q->rq_wb, bio, NULL); |
1893 | ||
bd166ef1 JA |
1894 | trace_block_getrq(q, bio, bio->bi_opf); |
1895 | ||
d2c0d383 | 1896 | rq = blk_mq_get_request(q, bio, bio->bi_opf, &data); |
87760e5e JA |
1897 | if (unlikely(!rq)) { |
1898 | __wbt_done(q->rq_wb, wb_acct); | |
03a07c92 GR |
1899 | if (bio->bi_opf & REQ_NOWAIT) |
1900 | bio_wouldblock_error(bio); | |
dece1635 | 1901 | return BLK_QC_T_NONE; |
87760e5e JA |
1902 | } |
1903 | ||
1904 | wbt_track(&rq->issue_stat, wb_acct); | |
07068d5b | 1905 | |
fd2d3326 | 1906 | cookie = request_to_qc_t(data.hctx, rq); |
07068d5b | 1907 | |
f984df1f | 1908 | plug = current->plug; |
07068d5b | 1909 | if (unlikely(is_flush_fua)) { |
f984df1f | 1910 | blk_mq_put_ctx(data.ctx); |
07068d5b | 1911 | blk_mq_bio_to_request(rq, bio); |
923218f6 ML |
1912 | |
1913 | /* bypass scheduler for flush rq */ | |
1914 | blk_insert_flush(rq); | |
1915 | blk_mq_run_hw_queue(data.hctx, true); | |
a4d907b6 | 1916 | } else if (plug && q->nr_hw_queues == 1) { |
600271d9 SL |
1917 | struct request *last = NULL; |
1918 | ||
b00c53e8 | 1919 | blk_mq_put_ctx(data.ctx); |
e6c4438b | 1920 | blk_mq_bio_to_request(rq, bio); |
0a6219a9 ML |
1921 | |
1922 | /* | |
1923 | * @request_count may become stale because of schedule | |
1924 | * out, so check the list again. | |
1925 | */ | |
1926 | if (list_empty(&plug->mq_list)) | |
1927 | request_count = 0; | |
254d259d CH |
1928 | else if (blk_queue_nomerges(q)) |
1929 | request_count = blk_plug_queued_count(q); | |
1930 | ||
676d0607 | 1931 | if (!request_count) |
e6c4438b | 1932 | trace_block_plug(q); |
600271d9 SL |
1933 | else |
1934 | last = list_entry_rq(plug->mq_list.prev); | |
b094f89c | 1935 | |
600271d9 SL |
1936 | if (request_count >= BLK_MAX_REQUEST_COUNT || (last && |
1937 | blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) { | |
e6c4438b JM |
1938 | blk_flush_plug_list(plug, false); |
1939 | trace_block_plug(q); | |
320ae51f | 1940 | } |
b094f89c | 1941 | |
e6c4438b | 1942 | list_add_tail(&rq->queuelist, &plug->mq_list); |
2299722c | 1943 | } else if (plug && !blk_queue_nomerges(q)) { |
bd166ef1 | 1944 | blk_mq_bio_to_request(rq, bio); |
07068d5b | 1945 | |
07068d5b | 1946 | /* |
6a83e74d | 1947 | * We do limited plugging. If the bio can be merged, do that. |
f984df1f SL |
1948 | * Otherwise the existing request in the plug list will be |
1949 | * issued. So the plug list will have one request at most | |
2299722c CH |
1950 | * The plug list might get flushed before this. If that happens, |
1951 | * the plug list is empty, and same_queue_rq is invalid. | |
07068d5b | 1952 | */ |
2299722c CH |
1953 | if (list_empty(&plug->mq_list)) |
1954 | same_queue_rq = NULL; | |
1955 | if (same_queue_rq) | |
1956 | list_del_init(&same_queue_rq->queuelist); | |
1957 | list_add_tail(&rq->queuelist, &plug->mq_list); | |
1958 | ||
bf4907c0 JA |
1959 | blk_mq_put_ctx(data.ctx); |
1960 | ||
dad7a3be ML |
1961 | if (same_queue_rq) { |
1962 | data.hctx = blk_mq_map_queue(q, | |
1963 | same_queue_rq->mq_ctx->cpu); | |
2299722c CH |
1964 | blk_mq_try_issue_directly(data.hctx, same_queue_rq, |
1965 | &cookie); | |
dad7a3be | 1966 | } |
a4d907b6 | 1967 | } else if (q->nr_hw_queues > 1 && is_sync) { |
bf4907c0 | 1968 | blk_mq_put_ctx(data.ctx); |
2299722c | 1969 | blk_mq_bio_to_request(rq, bio); |
2299722c | 1970 | blk_mq_try_issue_directly(data.hctx, rq, &cookie); |
a4d907b6 | 1971 | } else if (q->elevator) { |
b00c53e8 | 1972 | blk_mq_put_ctx(data.ctx); |
bd166ef1 | 1973 | blk_mq_bio_to_request(rq, bio); |
9e97d295 | 1974 | blk_mq_sched_insert_request(rq, false, true, true); |
ab42f35d | 1975 | } else { |
b00c53e8 | 1976 | blk_mq_put_ctx(data.ctx); |
ab42f35d ML |
1977 | blk_mq_bio_to_request(rq, bio); |
1978 | blk_mq_queue_io(data.hctx, data.ctx, rq); | |
a4d907b6 | 1979 | blk_mq_run_hw_queue(data.hctx, true); |
ab42f35d | 1980 | } |
320ae51f | 1981 | |
7b371636 | 1982 | return cookie; |
320ae51f JA |
1983 | } |
1984 | ||
cc71a6f4 JA |
1985 | void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags, |
1986 | unsigned int hctx_idx) | |
95363efd | 1987 | { |
e9b267d9 | 1988 | struct page *page; |
320ae51f | 1989 | |
24d2f903 | 1990 | if (tags->rqs && set->ops->exit_request) { |
e9b267d9 | 1991 | int i; |
320ae51f | 1992 | |
24d2f903 | 1993 | for (i = 0; i < tags->nr_tags; i++) { |
2af8cbe3 JA |
1994 | struct request *rq = tags->static_rqs[i]; |
1995 | ||
1996 | if (!rq) | |
e9b267d9 | 1997 | continue; |
d6296d39 | 1998 | set->ops->exit_request(set, rq, hctx_idx); |
2af8cbe3 | 1999 | tags->static_rqs[i] = NULL; |
e9b267d9 | 2000 | } |
320ae51f | 2001 | } |
320ae51f | 2002 | |
24d2f903 CH |
2003 | while (!list_empty(&tags->page_list)) { |
2004 | page = list_first_entry(&tags->page_list, struct page, lru); | |
6753471c | 2005 | list_del_init(&page->lru); |
f75782e4 CM |
2006 | /* |
2007 | * Remove kmemleak object previously allocated in | |
2008 | * blk_mq_init_rq_map(). | |
2009 | */ | |
2010 | kmemleak_free(page_address(page)); | |
320ae51f JA |
2011 | __free_pages(page, page->private); |
2012 | } | |
cc71a6f4 | 2013 | } |
320ae51f | 2014 | |
cc71a6f4 JA |
2015 | void blk_mq_free_rq_map(struct blk_mq_tags *tags) |
2016 | { | |
24d2f903 | 2017 | kfree(tags->rqs); |
cc71a6f4 | 2018 | tags->rqs = NULL; |
2af8cbe3 JA |
2019 | kfree(tags->static_rqs); |
2020 | tags->static_rqs = NULL; | |
320ae51f | 2021 | |
24d2f903 | 2022 | blk_mq_free_tags(tags); |
320ae51f JA |
2023 | } |
2024 | ||
cc71a6f4 JA |
2025 | struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, |
2026 | unsigned int hctx_idx, | |
2027 | unsigned int nr_tags, | |
2028 | unsigned int reserved_tags) | |
320ae51f | 2029 | { |
24d2f903 | 2030 | struct blk_mq_tags *tags; |
59f082e4 | 2031 | int node; |
320ae51f | 2032 | |
59f082e4 SL |
2033 | node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx); |
2034 | if (node == NUMA_NO_NODE) | |
2035 | node = set->numa_node; | |
2036 | ||
2037 | tags = blk_mq_init_tags(nr_tags, reserved_tags, node, | |
24391c0d | 2038 | BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags)); |
24d2f903 CH |
2039 | if (!tags) |
2040 | return NULL; | |
320ae51f | 2041 | |
cc71a6f4 | 2042 | tags->rqs = kzalloc_node(nr_tags * sizeof(struct request *), |
36e1f3d1 | 2043 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, |
59f082e4 | 2044 | node); |
24d2f903 CH |
2045 | if (!tags->rqs) { |
2046 | blk_mq_free_tags(tags); | |
2047 | return NULL; | |
2048 | } | |
320ae51f | 2049 | |
2af8cbe3 JA |
2050 | tags->static_rqs = kzalloc_node(nr_tags * sizeof(struct request *), |
2051 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, | |
59f082e4 | 2052 | node); |
2af8cbe3 JA |
2053 | if (!tags->static_rqs) { |
2054 | kfree(tags->rqs); | |
2055 | blk_mq_free_tags(tags); | |
2056 | return NULL; | |
2057 | } | |
2058 | ||
cc71a6f4 JA |
2059 | return tags; |
2060 | } | |
2061 | ||
2062 | static size_t order_to_size(unsigned int order) | |
2063 | { | |
2064 | return (size_t)PAGE_SIZE << order; | |
2065 | } | |
2066 | ||
1d9bd516 TH |
2067 | static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq, |
2068 | unsigned int hctx_idx, int node) | |
2069 | { | |
2070 | int ret; | |
2071 | ||
2072 | if (set->ops->init_request) { | |
2073 | ret = set->ops->init_request(set, rq, hctx_idx, node); | |
2074 | if (ret) | |
2075 | return ret; | |
2076 | } | |
2077 | ||
2078 | seqcount_init(&rq->gstate_seq); | |
2079 | u64_stats_init(&rq->aborted_gstate_sync); | |
2080 | return 0; | |
2081 | } | |
2082 | ||
cc71a6f4 JA |
2083 | int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags, |
2084 | unsigned int hctx_idx, unsigned int depth) | |
2085 | { | |
2086 | unsigned int i, j, entries_per_page, max_order = 4; | |
2087 | size_t rq_size, left; | |
59f082e4 SL |
2088 | int node; |
2089 | ||
2090 | node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx); | |
2091 | if (node == NUMA_NO_NODE) | |
2092 | node = set->numa_node; | |
cc71a6f4 JA |
2093 | |
2094 | INIT_LIST_HEAD(&tags->page_list); | |
2095 | ||
320ae51f JA |
2096 | /* |
2097 | * rq_size is the size of the request plus driver payload, rounded | |
2098 | * to the cacheline size | |
2099 | */ | |
24d2f903 | 2100 | rq_size = round_up(sizeof(struct request) + set->cmd_size, |
320ae51f | 2101 | cache_line_size()); |
cc71a6f4 | 2102 | left = rq_size * depth; |
320ae51f | 2103 | |
cc71a6f4 | 2104 | for (i = 0; i < depth; ) { |
320ae51f JA |
2105 | int this_order = max_order; |
2106 | struct page *page; | |
2107 | int to_do; | |
2108 | void *p; | |
2109 | ||
b3a834b1 | 2110 | while (this_order && left < order_to_size(this_order - 1)) |
320ae51f JA |
2111 | this_order--; |
2112 | ||
2113 | do { | |
59f082e4 | 2114 | page = alloc_pages_node(node, |
36e1f3d1 | 2115 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO, |
a5164405 | 2116 | this_order); |
320ae51f JA |
2117 | if (page) |
2118 | break; | |
2119 | if (!this_order--) | |
2120 | break; | |
2121 | if (order_to_size(this_order) < rq_size) | |
2122 | break; | |
2123 | } while (1); | |
2124 | ||
2125 | if (!page) | |
24d2f903 | 2126 | goto fail; |
320ae51f JA |
2127 | |
2128 | page->private = this_order; | |
24d2f903 | 2129 | list_add_tail(&page->lru, &tags->page_list); |
320ae51f JA |
2130 | |
2131 | p = page_address(page); | |
f75782e4 CM |
2132 | /* |
2133 | * Allow kmemleak to scan these pages as they contain pointers | |
2134 | * to additional allocations like via ops->init_request(). | |
2135 | */ | |
36e1f3d1 | 2136 | kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO); |
320ae51f | 2137 | entries_per_page = order_to_size(this_order) / rq_size; |
cc71a6f4 | 2138 | to_do = min(entries_per_page, depth - i); |
320ae51f JA |
2139 | left -= to_do * rq_size; |
2140 | for (j = 0; j < to_do; j++) { | |
2af8cbe3 JA |
2141 | struct request *rq = p; |
2142 | ||
2143 | tags->static_rqs[i] = rq; | |
1d9bd516 TH |
2144 | if (blk_mq_init_request(set, rq, hctx_idx, node)) { |
2145 | tags->static_rqs[i] = NULL; | |
2146 | goto fail; | |
e9b267d9 CH |
2147 | } |
2148 | ||
320ae51f JA |
2149 | p += rq_size; |
2150 | i++; | |
2151 | } | |
2152 | } | |
cc71a6f4 | 2153 | return 0; |
320ae51f | 2154 | |
24d2f903 | 2155 | fail: |
cc71a6f4 JA |
2156 | blk_mq_free_rqs(set, tags, hctx_idx); |
2157 | return -ENOMEM; | |
320ae51f JA |
2158 | } |
2159 | ||
e57690fe JA |
2160 | /* |
2161 | * 'cpu' is going away. splice any existing rq_list entries from this | |
2162 | * software queue to the hw queue dispatch list, and ensure that it | |
2163 | * gets run. | |
2164 | */ | |
9467f859 | 2165 | static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node) |
484b4061 | 2166 | { |
9467f859 | 2167 | struct blk_mq_hw_ctx *hctx; |
484b4061 JA |
2168 | struct blk_mq_ctx *ctx; |
2169 | LIST_HEAD(tmp); | |
2170 | ||
9467f859 | 2171 | hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead); |
e57690fe | 2172 | ctx = __blk_mq_get_ctx(hctx->queue, cpu); |
484b4061 JA |
2173 | |
2174 | spin_lock(&ctx->lock); | |
2175 | if (!list_empty(&ctx->rq_list)) { | |
2176 | list_splice_init(&ctx->rq_list, &tmp); | |
2177 | blk_mq_hctx_clear_pending(hctx, ctx); | |
2178 | } | |
2179 | spin_unlock(&ctx->lock); | |
2180 | ||
2181 | if (list_empty(&tmp)) | |
9467f859 | 2182 | return 0; |
484b4061 | 2183 | |
e57690fe JA |
2184 | spin_lock(&hctx->lock); |
2185 | list_splice_tail_init(&tmp, &hctx->dispatch); | |
2186 | spin_unlock(&hctx->lock); | |
484b4061 JA |
2187 | |
2188 | blk_mq_run_hw_queue(hctx, true); | |
9467f859 | 2189 | return 0; |
484b4061 JA |
2190 | } |
2191 | ||
9467f859 | 2192 | static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx) |
484b4061 | 2193 | { |
9467f859 TG |
2194 | cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD, |
2195 | &hctx->cpuhp_dead); | |
484b4061 JA |
2196 | } |
2197 | ||
c3b4afca | 2198 | /* hctx->ctxs will be freed in queue's release handler */ |
08e98fc6 ML |
2199 | static void blk_mq_exit_hctx(struct request_queue *q, |
2200 | struct blk_mq_tag_set *set, | |
2201 | struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx) | |
2202 | { | |
9c1051aa OS |
2203 | blk_mq_debugfs_unregister_hctx(hctx); |
2204 | ||
8ab0b7dc ML |
2205 | if (blk_mq_hw_queue_mapped(hctx)) |
2206 | blk_mq_tag_idle(hctx); | |
08e98fc6 | 2207 | |
f70ced09 | 2208 | if (set->ops->exit_request) |
d6296d39 | 2209 | set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx); |
f70ced09 | 2210 | |
93252632 OS |
2211 | blk_mq_sched_exit_hctx(q, hctx, hctx_idx); |
2212 | ||
08e98fc6 ML |
2213 | if (set->ops->exit_hctx) |
2214 | set->ops->exit_hctx(hctx, hctx_idx); | |
2215 | ||
6a83e74d | 2216 | if (hctx->flags & BLK_MQ_F_BLOCKING) |
05707b64 | 2217 | cleanup_srcu_struct(hctx->srcu); |
6a83e74d | 2218 | |
9467f859 | 2219 | blk_mq_remove_cpuhp(hctx); |
f70ced09 | 2220 | blk_free_flush_queue(hctx->fq); |
88459642 | 2221 | sbitmap_free(&hctx->ctx_map); |
08e98fc6 ML |
2222 | } |
2223 | ||
624dbe47 ML |
2224 | static void blk_mq_exit_hw_queues(struct request_queue *q, |
2225 | struct blk_mq_tag_set *set, int nr_queue) | |
2226 | { | |
2227 | struct blk_mq_hw_ctx *hctx; | |
2228 | unsigned int i; | |
2229 | ||
2230 | queue_for_each_hw_ctx(q, hctx, i) { | |
2231 | if (i == nr_queue) | |
2232 | break; | |
08e98fc6 | 2233 | blk_mq_exit_hctx(q, set, hctx, i); |
624dbe47 | 2234 | } |
624dbe47 ML |
2235 | } |
2236 | ||
08e98fc6 ML |
2237 | static int blk_mq_init_hctx(struct request_queue *q, |
2238 | struct blk_mq_tag_set *set, | |
2239 | struct blk_mq_hw_ctx *hctx, unsigned hctx_idx) | |
320ae51f | 2240 | { |
08e98fc6 ML |
2241 | int node; |
2242 | ||
2243 | node = hctx->numa_node; | |
2244 | if (node == NUMA_NO_NODE) | |
2245 | node = hctx->numa_node = set->numa_node; | |
2246 | ||
9f993737 | 2247 | INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn); |
08e98fc6 ML |
2248 | spin_lock_init(&hctx->lock); |
2249 | INIT_LIST_HEAD(&hctx->dispatch); | |
2250 | hctx->queue = q; | |
2404e607 | 2251 | hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED; |
08e98fc6 | 2252 | |
9467f859 | 2253 | cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead); |
08e98fc6 ML |
2254 | |
2255 | hctx->tags = set->tags[hctx_idx]; | |
320ae51f JA |
2256 | |
2257 | /* | |
08e98fc6 ML |
2258 | * Allocate space for all possible cpus to avoid allocation at |
2259 | * runtime | |
320ae51f | 2260 | */ |
d904bfa7 | 2261 | hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *), |
08e98fc6 ML |
2262 | GFP_KERNEL, node); |
2263 | if (!hctx->ctxs) | |
2264 | goto unregister_cpu_notifier; | |
320ae51f | 2265 | |
88459642 OS |
2266 | if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8), GFP_KERNEL, |
2267 | node)) | |
08e98fc6 | 2268 | goto free_ctxs; |
320ae51f | 2269 | |
08e98fc6 | 2270 | hctx->nr_ctx = 0; |
320ae51f | 2271 | |
eb619fdb JA |
2272 | init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake); |
2273 | INIT_LIST_HEAD(&hctx->dispatch_wait.entry); | |
2274 | ||
08e98fc6 ML |
2275 | if (set->ops->init_hctx && |
2276 | set->ops->init_hctx(hctx, set->driver_data, hctx_idx)) | |
2277 | goto free_bitmap; | |
320ae51f | 2278 | |
93252632 OS |
2279 | if (blk_mq_sched_init_hctx(q, hctx, hctx_idx)) |
2280 | goto exit_hctx; | |
2281 | ||
f70ced09 ML |
2282 | hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size); |
2283 | if (!hctx->fq) | |
93252632 | 2284 | goto sched_exit_hctx; |
320ae51f | 2285 | |
1d9bd516 | 2286 | if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx, node)) |
f70ced09 | 2287 | goto free_fq; |
320ae51f | 2288 | |
6a83e74d | 2289 | if (hctx->flags & BLK_MQ_F_BLOCKING) |
05707b64 | 2290 | init_srcu_struct(hctx->srcu); |
6a83e74d | 2291 | |
9c1051aa OS |
2292 | blk_mq_debugfs_register_hctx(q, hctx); |
2293 | ||
08e98fc6 | 2294 | return 0; |
320ae51f | 2295 | |
f70ced09 ML |
2296 | free_fq: |
2297 | kfree(hctx->fq); | |
93252632 OS |
2298 | sched_exit_hctx: |
2299 | blk_mq_sched_exit_hctx(q, hctx, hctx_idx); | |
f70ced09 ML |
2300 | exit_hctx: |
2301 | if (set->ops->exit_hctx) | |
2302 | set->ops->exit_hctx(hctx, hctx_idx); | |
08e98fc6 | 2303 | free_bitmap: |
88459642 | 2304 | sbitmap_free(&hctx->ctx_map); |
08e98fc6 ML |
2305 | free_ctxs: |
2306 | kfree(hctx->ctxs); | |
2307 | unregister_cpu_notifier: | |
9467f859 | 2308 | blk_mq_remove_cpuhp(hctx); |
08e98fc6 ML |
2309 | return -1; |
2310 | } | |
320ae51f | 2311 | |
320ae51f JA |
2312 | static void blk_mq_init_cpu_queues(struct request_queue *q, |
2313 | unsigned int nr_hw_queues) | |
2314 | { | |
2315 | unsigned int i; | |
2316 | ||
2317 | for_each_possible_cpu(i) { | |
2318 | struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i); | |
2319 | struct blk_mq_hw_ctx *hctx; | |
2320 | ||
320ae51f JA |
2321 | __ctx->cpu = i; |
2322 | spin_lock_init(&__ctx->lock); | |
2323 | INIT_LIST_HEAD(&__ctx->rq_list); | |
2324 | __ctx->queue = q; | |
2325 | ||
320ae51f JA |
2326 | /* |
2327 | * Set local node, IFF we have more than one hw queue. If | |
2328 | * not, we remain on the home node of the device | |
2329 | */ | |
20e4d813 | 2330 | hctx = blk_mq_map_queue(q, i); |
320ae51f | 2331 | if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE) |
bffed457 | 2332 | hctx->numa_node = local_memory_node(cpu_to_node(i)); |
320ae51f JA |
2333 | } |
2334 | } | |
2335 | ||
cc71a6f4 JA |
2336 | static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, int hctx_idx) |
2337 | { | |
2338 | int ret = 0; | |
2339 | ||
2340 | set->tags[hctx_idx] = blk_mq_alloc_rq_map(set, hctx_idx, | |
2341 | set->queue_depth, set->reserved_tags); | |
2342 | if (!set->tags[hctx_idx]) | |
2343 | return false; | |
2344 | ||
2345 | ret = blk_mq_alloc_rqs(set, set->tags[hctx_idx], hctx_idx, | |
2346 | set->queue_depth); | |
2347 | if (!ret) | |
2348 | return true; | |
2349 | ||
2350 | blk_mq_free_rq_map(set->tags[hctx_idx]); | |
2351 | set->tags[hctx_idx] = NULL; | |
2352 | return false; | |
2353 | } | |
2354 | ||
2355 | static void blk_mq_free_map_and_requests(struct blk_mq_tag_set *set, | |
2356 | unsigned int hctx_idx) | |
2357 | { | |
bd166ef1 JA |
2358 | if (set->tags[hctx_idx]) { |
2359 | blk_mq_free_rqs(set, set->tags[hctx_idx], hctx_idx); | |
2360 | blk_mq_free_rq_map(set->tags[hctx_idx]); | |
2361 | set->tags[hctx_idx] = NULL; | |
2362 | } | |
cc71a6f4 JA |
2363 | } |
2364 | ||
4b855ad3 | 2365 | static void blk_mq_map_swqueue(struct request_queue *q) |
320ae51f | 2366 | { |
d1b1cea1 | 2367 | unsigned int i, hctx_idx; |
320ae51f JA |
2368 | struct blk_mq_hw_ctx *hctx; |
2369 | struct blk_mq_ctx *ctx; | |
2a34c087 | 2370 | struct blk_mq_tag_set *set = q->tag_set; |
320ae51f | 2371 | |
60de074b AM |
2372 | /* |
2373 | * Avoid others reading imcomplete hctx->cpumask through sysfs | |
2374 | */ | |
2375 | mutex_lock(&q->sysfs_lock); | |
2376 | ||
320ae51f | 2377 | queue_for_each_hw_ctx(q, hctx, i) { |
e4043dcf | 2378 | cpumask_clear(hctx->cpumask); |
320ae51f JA |
2379 | hctx->nr_ctx = 0; |
2380 | } | |
2381 | ||
2382 | /* | |
4b855ad3 CH |
2383 | * Map software to hardware queues. |
2384 | * | |
2385 | * If the cpu isn't present, the cpu is mapped to first hctx. | |
320ae51f | 2386 | */ |
20e4d813 | 2387 | for_each_possible_cpu(i) { |
d1b1cea1 GKB |
2388 | hctx_idx = q->mq_map[i]; |
2389 | /* unmapped hw queue can be remapped after CPU topo changed */ | |
cc71a6f4 JA |
2390 | if (!set->tags[hctx_idx] && |
2391 | !__blk_mq_alloc_rq_map(set, hctx_idx)) { | |
d1b1cea1 GKB |
2392 | /* |
2393 | * If tags initialization fail for some hctx, | |
2394 | * that hctx won't be brought online. In this | |
2395 | * case, remap the current ctx to hctx[0] which | |
2396 | * is guaranteed to always have tags allocated | |
2397 | */ | |
cc71a6f4 | 2398 | q->mq_map[i] = 0; |
d1b1cea1 GKB |
2399 | } |
2400 | ||
897bb0c7 | 2401 | ctx = per_cpu_ptr(q->queue_ctx, i); |
7d7e0f90 | 2402 | hctx = blk_mq_map_queue(q, i); |
868f2f0b | 2403 | |
e4043dcf | 2404 | cpumask_set_cpu(i, hctx->cpumask); |
320ae51f JA |
2405 | ctx->index_hw = hctx->nr_ctx; |
2406 | hctx->ctxs[hctx->nr_ctx++] = ctx; | |
2407 | } | |
506e931f | 2408 | |
60de074b AM |
2409 | mutex_unlock(&q->sysfs_lock); |
2410 | ||
506e931f | 2411 | queue_for_each_hw_ctx(q, hctx, i) { |
484b4061 | 2412 | /* |
a68aafa5 JA |
2413 | * If no software queues are mapped to this hardware queue, |
2414 | * disable it and free the request entries. | |
484b4061 JA |
2415 | */ |
2416 | if (!hctx->nr_ctx) { | |
d1b1cea1 GKB |
2417 | /* Never unmap queue 0. We need it as a |
2418 | * fallback in case of a new remap fails | |
2419 | * allocation | |
2420 | */ | |
cc71a6f4 JA |
2421 | if (i && set->tags[i]) |
2422 | blk_mq_free_map_and_requests(set, i); | |
2423 | ||
2a34c087 | 2424 | hctx->tags = NULL; |
484b4061 JA |
2425 | continue; |
2426 | } | |
2427 | ||
2a34c087 ML |
2428 | hctx->tags = set->tags[i]; |
2429 | WARN_ON(!hctx->tags); | |
2430 | ||
889fa31f CY |
2431 | /* |
2432 | * Set the map size to the number of mapped software queues. | |
2433 | * This is more accurate and more efficient than looping | |
2434 | * over all possibly mapped software queues. | |
2435 | */ | |
88459642 | 2436 | sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx); |
889fa31f | 2437 | |
484b4061 JA |
2438 | /* |
2439 | * Initialize batch roundrobin counts | |
2440 | */ | |
20e4d813 CH |
2441 | hctx->next_cpu = cpumask_first_and(hctx->cpumask, |
2442 | cpu_online_mask); | |
506e931f JA |
2443 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; |
2444 | } | |
320ae51f JA |
2445 | } |
2446 | ||
8e8320c9 JA |
2447 | /* |
2448 | * Caller needs to ensure that we're either frozen/quiesced, or that | |
2449 | * the queue isn't live yet. | |
2450 | */ | |
2404e607 | 2451 | static void queue_set_hctx_shared(struct request_queue *q, bool shared) |
0d2602ca JA |
2452 | { |
2453 | struct blk_mq_hw_ctx *hctx; | |
0d2602ca JA |
2454 | int i; |
2455 | ||
2404e607 | 2456 | queue_for_each_hw_ctx(q, hctx, i) { |
8e8320c9 JA |
2457 | if (shared) { |
2458 | if (test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state)) | |
2459 | atomic_inc(&q->shared_hctx_restart); | |
2404e607 | 2460 | hctx->flags |= BLK_MQ_F_TAG_SHARED; |
8e8320c9 JA |
2461 | } else { |
2462 | if (test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state)) | |
2463 | atomic_dec(&q->shared_hctx_restart); | |
2404e607 | 2464 | hctx->flags &= ~BLK_MQ_F_TAG_SHARED; |
8e8320c9 | 2465 | } |
2404e607 JM |
2466 | } |
2467 | } | |
2468 | ||
8e8320c9 JA |
2469 | static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set, |
2470 | bool shared) | |
2404e607 JM |
2471 | { |
2472 | struct request_queue *q; | |
0d2602ca | 2473 | |
705cda97 BVA |
2474 | lockdep_assert_held(&set->tag_list_lock); |
2475 | ||
0d2602ca JA |
2476 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
2477 | blk_mq_freeze_queue(q); | |
2404e607 | 2478 | queue_set_hctx_shared(q, shared); |
0d2602ca JA |
2479 | blk_mq_unfreeze_queue(q); |
2480 | } | |
2481 | } | |
2482 | ||
2483 | static void blk_mq_del_queue_tag_set(struct request_queue *q) | |
2484 | { | |
2485 | struct blk_mq_tag_set *set = q->tag_set; | |
2486 | ||
0d2602ca | 2487 | mutex_lock(&set->tag_list_lock); |
705cda97 BVA |
2488 | list_del_rcu(&q->tag_set_list); |
2489 | INIT_LIST_HEAD(&q->tag_set_list); | |
2404e607 JM |
2490 | if (list_is_singular(&set->tag_list)) { |
2491 | /* just transitioned to unshared */ | |
2492 | set->flags &= ~BLK_MQ_F_TAG_SHARED; | |
2493 | /* update existing queue */ | |
2494 | blk_mq_update_tag_set_depth(set, false); | |
2495 | } | |
0d2602ca | 2496 | mutex_unlock(&set->tag_list_lock); |
705cda97 BVA |
2497 | |
2498 | synchronize_rcu(); | |
0d2602ca JA |
2499 | } |
2500 | ||
2501 | static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set, | |
2502 | struct request_queue *q) | |
2503 | { | |
2504 | q->tag_set = set; | |
2505 | ||
2506 | mutex_lock(&set->tag_list_lock); | |
2404e607 | 2507 | |
ff821d27 JA |
2508 | /* |
2509 | * Check to see if we're transitioning to shared (from 1 to 2 queues). | |
2510 | */ | |
2511 | if (!list_empty(&set->tag_list) && | |
2512 | !(set->flags & BLK_MQ_F_TAG_SHARED)) { | |
2404e607 JM |
2513 | set->flags |= BLK_MQ_F_TAG_SHARED; |
2514 | /* update existing queue */ | |
2515 | blk_mq_update_tag_set_depth(set, true); | |
2516 | } | |
2517 | if (set->flags & BLK_MQ_F_TAG_SHARED) | |
2518 | queue_set_hctx_shared(q, true); | |
705cda97 | 2519 | list_add_tail_rcu(&q->tag_set_list, &set->tag_list); |
2404e607 | 2520 | |
0d2602ca JA |
2521 | mutex_unlock(&set->tag_list_lock); |
2522 | } | |
2523 | ||
e09aae7e ML |
2524 | /* |
2525 | * It is the actual release handler for mq, but we do it from | |
2526 | * request queue's release handler for avoiding use-after-free | |
2527 | * and headache because q->mq_kobj shouldn't have been introduced, | |
2528 | * but we can't group ctx/kctx kobj without it. | |
2529 | */ | |
2530 | void blk_mq_release(struct request_queue *q) | |
2531 | { | |
2532 | struct blk_mq_hw_ctx *hctx; | |
2533 | unsigned int i; | |
2534 | ||
2535 | /* hctx kobj stays in hctx */ | |
c3b4afca ML |
2536 | queue_for_each_hw_ctx(q, hctx, i) { |
2537 | if (!hctx) | |
2538 | continue; | |
6c8b232e | 2539 | kobject_put(&hctx->kobj); |
c3b4afca | 2540 | } |
e09aae7e | 2541 | |
a723bab3 AM |
2542 | q->mq_map = NULL; |
2543 | ||
e09aae7e ML |
2544 | kfree(q->queue_hw_ctx); |
2545 | ||
7ea5fe31 ML |
2546 | /* |
2547 | * release .mq_kobj and sw queue's kobject now because | |
2548 | * both share lifetime with request queue. | |
2549 | */ | |
2550 | blk_mq_sysfs_deinit(q); | |
2551 | ||
e09aae7e ML |
2552 | free_percpu(q->queue_ctx); |
2553 | } | |
2554 | ||
24d2f903 | 2555 | struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set) |
b62c21b7 MS |
2556 | { |
2557 | struct request_queue *uninit_q, *q; | |
2558 | ||
2559 | uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node); | |
2560 | if (!uninit_q) | |
2561 | return ERR_PTR(-ENOMEM); | |
2562 | ||
2563 | q = blk_mq_init_allocated_queue(set, uninit_q); | |
2564 | if (IS_ERR(q)) | |
2565 | blk_cleanup_queue(uninit_q); | |
2566 | ||
2567 | return q; | |
2568 | } | |
2569 | EXPORT_SYMBOL(blk_mq_init_queue); | |
2570 | ||
07319678 BVA |
2571 | static int blk_mq_hw_ctx_size(struct blk_mq_tag_set *tag_set) |
2572 | { | |
2573 | int hw_ctx_size = sizeof(struct blk_mq_hw_ctx); | |
2574 | ||
05707b64 | 2575 | BUILD_BUG_ON(ALIGN(offsetof(struct blk_mq_hw_ctx, srcu), |
07319678 BVA |
2576 | __alignof__(struct blk_mq_hw_ctx)) != |
2577 | sizeof(struct blk_mq_hw_ctx)); | |
2578 | ||
2579 | if (tag_set->flags & BLK_MQ_F_BLOCKING) | |
2580 | hw_ctx_size += sizeof(struct srcu_struct); | |
2581 | ||
2582 | return hw_ctx_size; | |
2583 | } | |
2584 | ||
868f2f0b KB |
2585 | static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set, |
2586 | struct request_queue *q) | |
320ae51f | 2587 | { |
868f2f0b KB |
2588 | int i, j; |
2589 | struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx; | |
f14bbe77 | 2590 | |
868f2f0b | 2591 | blk_mq_sysfs_unregister(q); |
fb350e0a ML |
2592 | |
2593 | /* protect against switching io scheduler */ | |
2594 | mutex_lock(&q->sysfs_lock); | |
24d2f903 | 2595 | for (i = 0; i < set->nr_hw_queues; i++) { |
868f2f0b | 2596 | int node; |
f14bbe77 | 2597 | |
868f2f0b KB |
2598 | if (hctxs[i]) |
2599 | continue; | |
2600 | ||
2601 | node = blk_mq_hw_queue_to_node(q->mq_map, i); | |
07319678 | 2602 | hctxs[i] = kzalloc_node(blk_mq_hw_ctx_size(set), |
cdef54dd | 2603 | GFP_KERNEL, node); |
320ae51f | 2604 | if (!hctxs[i]) |
868f2f0b | 2605 | break; |
320ae51f | 2606 | |
a86073e4 | 2607 | if (!zalloc_cpumask_var_node(&hctxs[i]->cpumask, GFP_KERNEL, |
868f2f0b KB |
2608 | node)) { |
2609 | kfree(hctxs[i]); | |
2610 | hctxs[i] = NULL; | |
2611 | break; | |
2612 | } | |
e4043dcf | 2613 | |
0d2602ca | 2614 | atomic_set(&hctxs[i]->nr_active, 0); |
f14bbe77 | 2615 | hctxs[i]->numa_node = node; |
320ae51f | 2616 | hctxs[i]->queue_num = i; |
868f2f0b KB |
2617 | |
2618 | if (blk_mq_init_hctx(q, set, hctxs[i], i)) { | |
2619 | free_cpumask_var(hctxs[i]->cpumask); | |
2620 | kfree(hctxs[i]); | |
2621 | hctxs[i] = NULL; | |
2622 | break; | |
2623 | } | |
2624 | blk_mq_hctx_kobj_init(hctxs[i]); | |
320ae51f | 2625 | } |
868f2f0b KB |
2626 | for (j = i; j < q->nr_hw_queues; j++) { |
2627 | struct blk_mq_hw_ctx *hctx = hctxs[j]; | |
2628 | ||
2629 | if (hctx) { | |
cc71a6f4 JA |
2630 | if (hctx->tags) |
2631 | blk_mq_free_map_and_requests(set, j); | |
868f2f0b | 2632 | blk_mq_exit_hctx(q, set, hctx, j); |
868f2f0b | 2633 | kobject_put(&hctx->kobj); |
868f2f0b KB |
2634 | hctxs[j] = NULL; |
2635 | ||
2636 | } | |
2637 | } | |
2638 | q->nr_hw_queues = i; | |
fb350e0a | 2639 | mutex_unlock(&q->sysfs_lock); |
868f2f0b KB |
2640 | blk_mq_sysfs_register(q); |
2641 | } | |
2642 | ||
2643 | struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set, | |
2644 | struct request_queue *q) | |
2645 | { | |
66841672 ML |
2646 | /* mark the queue as mq asap */ |
2647 | q->mq_ops = set->ops; | |
2648 | ||
34dbad5d | 2649 | q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn, |
720b8ccc SB |
2650 | blk_mq_poll_stats_bkt, |
2651 | BLK_MQ_POLL_STATS_BKTS, q); | |
34dbad5d OS |
2652 | if (!q->poll_cb) |
2653 | goto err_exit; | |
2654 | ||
868f2f0b KB |
2655 | q->queue_ctx = alloc_percpu(struct blk_mq_ctx); |
2656 | if (!q->queue_ctx) | |
c7de5726 | 2657 | goto err_exit; |
868f2f0b | 2658 | |
737f98cf ML |
2659 | /* init q->mq_kobj and sw queues' kobjects */ |
2660 | blk_mq_sysfs_init(q); | |
2661 | ||
868f2f0b KB |
2662 | q->queue_hw_ctx = kzalloc_node(nr_cpu_ids * sizeof(*(q->queue_hw_ctx)), |
2663 | GFP_KERNEL, set->numa_node); | |
2664 | if (!q->queue_hw_ctx) | |
2665 | goto err_percpu; | |
2666 | ||
bdd17e75 | 2667 | q->mq_map = set->mq_map; |
868f2f0b KB |
2668 | |
2669 | blk_mq_realloc_hw_ctxs(set, q); | |
2670 | if (!q->nr_hw_queues) | |
2671 | goto err_hctxs; | |
320ae51f | 2672 | |
287922eb | 2673 | INIT_WORK(&q->timeout_work, blk_mq_timeout_work); |
e56f698b | 2674 | blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ); |
320ae51f JA |
2675 | |
2676 | q->nr_queues = nr_cpu_ids; | |
320ae51f | 2677 | |
94eddfbe | 2678 | q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT; |
320ae51f | 2679 | |
05f1dd53 JA |
2680 | if (!(set->flags & BLK_MQ_F_SG_MERGE)) |
2681 | q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE; | |
2682 | ||
1be036e9 CH |
2683 | q->sg_reserved_size = INT_MAX; |
2684 | ||
2849450a | 2685 | INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work); |
6fca6a61 CH |
2686 | INIT_LIST_HEAD(&q->requeue_list); |
2687 | spin_lock_init(&q->requeue_lock); | |
2688 | ||
254d259d | 2689 | blk_queue_make_request(q, blk_mq_make_request); |
ea435e1b CH |
2690 | if (q->mq_ops->poll) |
2691 | q->poll_fn = blk_mq_poll; | |
07068d5b | 2692 | |
eba71768 JA |
2693 | /* |
2694 | * Do this after blk_queue_make_request() overrides it... | |
2695 | */ | |
2696 | q->nr_requests = set->queue_depth; | |
2697 | ||
64f1c21e JA |
2698 | /* |
2699 | * Default to classic polling | |
2700 | */ | |
2701 | q->poll_nsec = -1; | |
2702 | ||
24d2f903 CH |
2703 | if (set->ops->complete) |
2704 | blk_queue_softirq_done(q, set->ops->complete); | |
30a91cb4 | 2705 | |
24d2f903 | 2706 | blk_mq_init_cpu_queues(q, set->nr_hw_queues); |
0d2602ca | 2707 | blk_mq_add_queue_tag_set(set, q); |
4b855ad3 | 2708 | blk_mq_map_swqueue(q); |
4593fdbe | 2709 | |
d3484991 JA |
2710 | if (!(set->flags & BLK_MQ_F_NO_SCHED)) { |
2711 | int ret; | |
2712 | ||
2713 | ret = blk_mq_sched_init(q); | |
2714 | if (ret) | |
2715 | return ERR_PTR(ret); | |
2716 | } | |
2717 | ||
320ae51f | 2718 | return q; |
18741986 | 2719 | |
320ae51f | 2720 | err_hctxs: |
868f2f0b | 2721 | kfree(q->queue_hw_ctx); |
320ae51f | 2722 | err_percpu: |
868f2f0b | 2723 | free_percpu(q->queue_ctx); |
c7de5726 ML |
2724 | err_exit: |
2725 | q->mq_ops = NULL; | |
320ae51f JA |
2726 | return ERR_PTR(-ENOMEM); |
2727 | } | |
b62c21b7 | 2728 | EXPORT_SYMBOL(blk_mq_init_allocated_queue); |
320ae51f JA |
2729 | |
2730 | void blk_mq_free_queue(struct request_queue *q) | |
2731 | { | |
624dbe47 | 2732 | struct blk_mq_tag_set *set = q->tag_set; |
320ae51f | 2733 | |
0d2602ca | 2734 | blk_mq_del_queue_tag_set(q); |
624dbe47 | 2735 | blk_mq_exit_hw_queues(q, set, set->nr_hw_queues); |
320ae51f | 2736 | } |
320ae51f JA |
2737 | |
2738 | /* Basically redo blk_mq_init_queue with queue frozen */ | |
4b855ad3 | 2739 | static void blk_mq_queue_reinit(struct request_queue *q) |
320ae51f | 2740 | { |
4ecd4fef | 2741 | WARN_ON_ONCE(!atomic_read(&q->mq_freeze_depth)); |
320ae51f | 2742 | |
9c1051aa | 2743 | blk_mq_debugfs_unregister_hctxs(q); |
67aec14c JA |
2744 | blk_mq_sysfs_unregister(q); |
2745 | ||
320ae51f JA |
2746 | /* |
2747 | * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe | |
ff821d27 JA |
2748 | * we should change hctx numa_node according to the new topology (this |
2749 | * involves freeing and re-allocating memory, worth doing?) | |
320ae51f | 2750 | */ |
4b855ad3 | 2751 | blk_mq_map_swqueue(q); |
320ae51f | 2752 | |
67aec14c | 2753 | blk_mq_sysfs_register(q); |
9c1051aa | 2754 | blk_mq_debugfs_register_hctxs(q); |
320ae51f JA |
2755 | } |
2756 | ||
a5164405 JA |
2757 | static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set) |
2758 | { | |
2759 | int i; | |
2760 | ||
cc71a6f4 JA |
2761 | for (i = 0; i < set->nr_hw_queues; i++) |
2762 | if (!__blk_mq_alloc_rq_map(set, i)) | |
a5164405 | 2763 | goto out_unwind; |
a5164405 JA |
2764 | |
2765 | return 0; | |
2766 | ||
2767 | out_unwind: | |
2768 | while (--i >= 0) | |
cc71a6f4 | 2769 | blk_mq_free_rq_map(set->tags[i]); |
a5164405 | 2770 | |
a5164405 JA |
2771 | return -ENOMEM; |
2772 | } | |
2773 | ||
2774 | /* | |
2775 | * Allocate the request maps associated with this tag_set. Note that this | |
2776 | * may reduce the depth asked for, if memory is tight. set->queue_depth | |
2777 | * will be updated to reflect the allocated depth. | |
2778 | */ | |
2779 | static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set) | |
2780 | { | |
2781 | unsigned int depth; | |
2782 | int err; | |
2783 | ||
2784 | depth = set->queue_depth; | |
2785 | do { | |
2786 | err = __blk_mq_alloc_rq_maps(set); | |
2787 | if (!err) | |
2788 | break; | |
2789 | ||
2790 | set->queue_depth >>= 1; | |
2791 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) { | |
2792 | err = -ENOMEM; | |
2793 | break; | |
2794 | } | |
2795 | } while (set->queue_depth); | |
2796 | ||
2797 | if (!set->queue_depth || err) { | |
2798 | pr_err("blk-mq: failed to allocate request map\n"); | |
2799 | return -ENOMEM; | |
2800 | } | |
2801 | ||
2802 | if (depth != set->queue_depth) | |
2803 | pr_info("blk-mq: reduced tag depth (%u -> %u)\n", | |
2804 | depth, set->queue_depth); | |
2805 | ||
2806 | return 0; | |
2807 | } | |
2808 | ||
ebe8bddb OS |
2809 | static int blk_mq_update_queue_map(struct blk_mq_tag_set *set) |
2810 | { | |
7d4901a9 ML |
2811 | if (set->ops->map_queues) { |
2812 | int cpu; | |
2813 | /* | |
2814 | * transport .map_queues is usually done in the following | |
2815 | * way: | |
2816 | * | |
2817 | * for (queue = 0; queue < set->nr_hw_queues; queue++) { | |
2818 | * mask = get_cpu_mask(queue) | |
2819 | * for_each_cpu(cpu, mask) | |
2820 | * set->mq_map[cpu] = queue; | |
2821 | * } | |
2822 | * | |
2823 | * When we need to remap, the table has to be cleared for | |
2824 | * killing stale mapping since one CPU may not be mapped | |
2825 | * to any hw queue. | |
2826 | */ | |
2827 | for_each_possible_cpu(cpu) | |
2828 | set->mq_map[cpu] = 0; | |
2829 | ||
ebe8bddb | 2830 | return set->ops->map_queues(set); |
7d4901a9 | 2831 | } else |
ebe8bddb OS |
2832 | return blk_mq_map_queues(set); |
2833 | } | |
2834 | ||
a4391c64 JA |
2835 | /* |
2836 | * Alloc a tag set to be associated with one or more request queues. | |
2837 | * May fail with EINVAL for various error conditions. May adjust the | |
2838 | * requested depth down, if if it too large. In that case, the set | |
2839 | * value will be stored in set->queue_depth. | |
2840 | */ | |
24d2f903 CH |
2841 | int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set) |
2842 | { | |
da695ba2 CH |
2843 | int ret; |
2844 | ||
205fb5f5 BVA |
2845 | BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS); |
2846 | ||
24d2f903 CH |
2847 | if (!set->nr_hw_queues) |
2848 | return -EINVAL; | |
a4391c64 | 2849 | if (!set->queue_depth) |
24d2f903 CH |
2850 | return -EINVAL; |
2851 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) | |
2852 | return -EINVAL; | |
2853 | ||
7d7e0f90 | 2854 | if (!set->ops->queue_rq) |
24d2f903 CH |
2855 | return -EINVAL; |
2856 | ||
de148297 ML |
2857 | if (!set->ops->get_budget ^ !set->ops->put_budget) |
2858 | return -EINVAL; | |
2859 | ||
a4391c64 JA |
2860 | if (set->queue_depth > BLK_MQ_MAX_DEPTH) { |
2861 | pr_info("blk-mq: reduced tag depth to %u\n", | |
2862 | BLK_MQ_MAX_DEPTH); | |
2863 | set->queue_depth = BLK_MQ_MAX_DEPTH; | |
2864 | } | |
24d2f903 | 2865 | |
6637fadf SL |
2866 | /* |
2867 | * If a crashdump is active, then we are potentially in a very | |
2868 | * memory constrained environment. Limit us to 1 queue and | |
2869 | * 64 tags to prevent using too much memory. | |
2870 | */ | |
2871 | if (is_kdump_kernel()) { | |
2872 | set->nr_hw_queues = 1; | |
2873 | set->queue_depth = min(64U, set->queue_depth); | |
2874 | } | |
868f2f0b KB |
2875 | /* |
2876 | * There is no use for more h/w queues than cpus. | |
2877 | */ | |
2878 | if (set->nr_hw_queues > nr_cpu_ids) | |
2879 | set->nr_hw_queues = nr_cpu_ids; | |
6637fadf | 2880 | |
868f2f0b | 2881 | set->tags = kzalloc_node(nr_cpu_ids * sizeof(struct blk_mq_tags *), |
24d2f903 CH |
2882 | GFP_KERNEL, set->numa_node); |
2883 | if (!set->tags) | |
a5164405 | 2884 | return -ENOMEM; |
24d2f903 | 2885 | |
da695ba2 CH |
2886 | ret = -ENOMEM; |
2887 | set->mq_map = kzalloc_node(sizeof(*set->mq_map) * nr_cpu_ids, | |
2888 | GFP_KERNEL, set->numa_node); | |
bdd17e75 CH |
2889 | if (!set->mq_map) |
2890 | goto out_free_tags; | |
2891 | ||
ebe8bddb | 2892 | ret = blk_mq_update_queue_map(set); |
da695ba2 CH |
2893 | if (ret) |
2894 | goto out_free_mq_map; | |
2895 | ||
2896 | ret = blk_mq_alloc_rq_maps(set); | |
2897 | if (ret) | |
bdd17e75 | 2898 | goto out_free_mq_map; |
24d2f903 | 2899 | |
0d2602ca JA |
2900 | mutex_init(&set->tag_list_lock); |
2901 | INIT_LIST_HEAD(&set->tag_list); | |
2902 | ||
24d2f903 | 2903 | return 0; |
bdd17e75 CH |
2904 | |
2905 | out_free_mq_map: | |
2906 | kfree(set->mq_map); | |
2907 | set->mq_map = NULL; | |
2908 | out_free_tags: | |
5676e7b6 RE |
2909 | kfree(set->tags); |
2910 | set->tags = NULL; | |
da695ba2 | 2911 | return ret; |
24d2f903 CH |
2912 | } |
2913 | EXPORT_SYMBOL(blk_mq_alloc_tag_set); | |
2914 | ||
2915 | void blk_mq_free_tag_set(struct blk_mq_tag_set *set) | |
2916 | { | |
2917 | int i; | |
2918 | ||
cc71a6f4 JA |
2919 | for (i = 0; i < nr_cpu_ids; i++) |
2920 | blk_mq_free_map_and_requests(set, i); | |
484b4061 | 2921 | |
bdd17e75 CH |
2922 | kfree(set->mq_map); |
2923 | set->mq_map = NULL; | |
2924 | ||
981bd189 | 2925 | kfree(set->tags); |
5676e7b6 | 2926 | set->tags = NULL; |
24d2f903 CH |
2927 | } |
2928 | EXPORT_SYMBOL(blk_mq_free_tag_set); | |
2929 | ||
e3a2b3f9 JA |
2930 | int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr) |
2931 | { | |
2932 | struct blk_mq_tag_set *set = q->tag_set; | |
2933 | struct blk_mq_hw_ctx *hctx; | |
2934 | int i, ret; | |
2935 | ||
bd166ef1 | 2936 | if (!set) |
e3a2b3f9 JA |
2937 | return -EINVAL; |
2938 | ||
70f36b60 | 2939 | blk_mq_freeze_queue(q); |
24f5a90f | 2940 | blk_mq_quiesce_queue(q); |
70f36b60 | 2941 | |
e3a2b3f9 JA |
2942 | ret = 0; |
2943 | queue_for_each_hw_ctx(q, hctx, i) { | |
e9137d4b KB |
2944 | if (!hctx->tags) |
2945 | continue; | |
bd166ef1 JA |
2946 | /* |
2947 | * If we're using an MQ scheduler, just update the scheduler | |
2948 | * queue depth. This is similar to what the old code would do. | |
2949 | */ | |
70f36b60 | 2950 | if (!hctx->sched_tags) { |
c2e82a23 | 2951 | ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr, |
70f36b60 JA |
2952 | false); |
2953 | } else { | |
2954 | ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags, | |
2955 | nr, true); | |
2956 | } | |
e3a2b3f9 JA |
2957 | if (ret) |
2958 | break; | |
2959 | } | |
2960 | ||
2961 | if (!ret) | |
2962 | q->nr_requests = nr; | |
2963 | ||
24f5a90f | 2964 | blk_mq_unquiesce_queue(q); |
70f36b60 | 2965 | blk_mq_unfreeze_queue(q); |
70f36b60 | 2966 | |
e3a2b3f9 JA |
2967 | return ret; |
2968 | } | |
2969 | ||
e4dc2b32 KB |
2970 | static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, |
2971 | int nr_hw_queues) | |
868f2f0b KB |
2972 | { |
2973 | struct request_queue *q; | |
2974 | ||
705cda97 BVA |
2975 | lockdep_assert_held(&set->tag_list_lock); |
2976 | ||
868f2f0b KB |
2977 | if (nr_hw_queues > nr_cpu_ids) |
2978 | nr_hw_queues = nr_cpu_ids; | |
2979 | if (nr_hw_queues < 1 || nr_hw_queues == set->nr_hw_queues) | |
2980 | return; | |
2981 | ||
2982 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
2983 | blk_mq_freeze_queue(q); | |
2984 | ||
2985 | set->nr_hw_queues = nr_hw_queues; | |
ebe8bddb | 2986 | blk_mq_update_queue_map(set); |
868f2f0b KB |
2987 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
2988 | blk_mq_realloc_hw_ctxs(set, q); | |
4b855ad3 | 2989 | blk_mq_queue_reinit(q); |
868f2f0b KB |
2990 | } |
2991 | ||
2992 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
2993 | blk_mq_unfreeze_queue(q); | |
2994 | } | |
e4dc2b32 KB |
2995 | |
2996 | void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues) | |
2997 | { | |
2998 | mutex_lock(&set->tag_list_lock); | |
2999 | __blk_mq_update_nr_hw_queues(set, nr_hw_queues); | |
3000 | mutex_unlock(&set->tag_list_lock); | |
3001 | } | |
868f2f0b KB |
3002 | EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues); |
3003 | ||
34dbad5d OS |
3004 | /* Enable polling stats and return whether they were already enabled. */ |
3005 | static bool blk_poll_stats_enable(struct request_queue *q) | |
3006 | { | |
3007 | if (test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) || | |
3008 | test_and_set_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags)) | |
3009 | return true; | |
3010 | blk_stat_add_callback(q, q->poll_cb); | |
3011 | return false; | |
3012 | } | |
3013 | ||
3014 | static void blk_mq_poll_stats_start(struct request_queue *q) | |
3015 | { | |
3016 | /* | |
3017 | * We don't arm the callback if polling stats are not enabled or the | |
3018 | * callback is already active. | |
3019 | */ | |
3020 | if (!test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) || | |
3021 | blk_stat_is_active(q->poll_cb)) | |
3022 | return; | |
3023 | ||
3024 | blk_stat_activate_msecs(q->poll_cb, 100); | |
3025 | } | |
3026 | ||
3027 | static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb) | |
3028 | { | |
3029 | struct request_queue *q = cb->data; | |
720b8ccc | 3030 | int bucket; |
34dbad5d | 3031 | |
720b8ccc SB |
3032 | for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) { |
3033 | if (cb->stat[bucket].nr_samples) | |
3034 | q->poll_stat[bucket] = cb->stat[bucket]; | |
3035 | } | |
34dbad5d OS |
3036 | } |
3037 | ||
64f1c21e JA |
3038 | static unsigned long blk_mq_poll_nsecs(struct request_queue *q, |
3039 | struct blk_mq_hw_ctx *hctx, | |
3040 | struct request *rq) | |
3041 | { | |
64f1c21e | 3042 | unsigned long ret = 0; |
720b8ccc | 3043 | int bucket; |
64f1c21e JA |
3044 | |
3045 | /* | |
3046 | * If stats collection isn't on, don't sleep but turn it on for | |
3047 | * future users | |
3048 | */ | |
34dbad5d | 3049 | if (!blk_poll_stats_enable(q)) |
64f1c21e JA |
3050 | return 0; |
3051 | ||
64f1c21e JA |
3052 | /* |
3053 | * As an optimistic guess, use half of the mean service time | |
3054 | * for this type of request. We can (and should) make this smarter. | |
3055 | * For instance, if the completion latencies are tight, we can | |
3056 | * get closer than just half the mean. This is especially | |
3057 | * important on devices where the completion latencies are longer | |
720b8ccc SB |
3058 | * than ~10 usec. We do use the stats for the relevant IO size |
3059 | * if available which does lead to better estimates. | |
64f1c21e | 3060 | */ |
720b8ccc SB |
3061 | bucket = blk_mq_poll_stats_bkt(rq); |
3062 | if (bucket < 0) | |
3063 | return ret; | |
3064 | ||
3065 | if (q->poll_stat[bucket].nr_samples) | |
3066 | ret = (q->poll_stat[bucket].mean + 1) / 2; | |
64f1c21e JA |
3067 | |
3068 | return ret; | |
3069 | } | |
3070 | ||
06426adf | 3071 | static bool blk_mq_poll_hybrid_sleep(struct request_queue *q, |
64f1c21e | 3072 | struct blk_mq_hw_ctx *hctx, |
06426adf JA |
3073 | struct request *rq) |
3074 | { | |
3075 | struct hrtimer_sleeper hs; | |
3076 | enum hrtimer_mode mode; | |
64f1c21e | 3077 | unsigned int nsecs; |
06426adf JA |
3078 | ktime_t kt; |
3079 | ||
76a86f9d | 3080 | if (rq->rq_flags & RQF_MQ_POLL_SLEPT) |
64f1c21e JA |
3081 | return false; |
3082 | ||
3083 | /* | |
3084 | * poll_nsec can be: | |
3085 | * | |
3086 | * -1: don't ever hybrid sleep | |
3087 | * 0: use half of prev avg | |
3088 | * >0: use this specific value | |
3089 | */ | |
3090 | if (q->poll_nsec == -1) | |
3091 | return false; | |
3092 | else if (q->poll_nsec > 0) | |
3093 | nsecs = q->poll_nsec; | |
3094 | else | |
3095 | nsecs = blk_mq_poll_nsecs(q, hctx, rq); | |
3096 | ||
3097 | if (!nsecs) | |
06426adf JA |
3098 | return false; |
3099 | ||
76a86f9d | 3100 | rq->rq_flags |= RQF_MQ_POLL_SLEPT; |
06426adf JA |
3101 | |
3102 | /* | |
3103 | * This will be replaced with the stats tracking code, using | |
3104 | * 'avg_completion_time / 2' as the pre-sleep target. | |
3105 | */ | |
8b0e1953 | 3106 | kt = nsecs; |
06426adf JA |
3107 | |
3108 | mode = HRTIMER_MODE_REL; | |
3109 | hrtimer_init_on_stack(&hs.timer, CLOCK_MONOTONIC, mode); | |
3110 | hrtimer_set_expires(&hs.timer, kt); | |
3111 | ||
3112 | hrtimer_init_sleeper(&hs, current); | |
3113 | do { | |
5a61c363 | 3114 | if (blk_mq_rq_state(rq) == MQ_RQ_COMPLETE) |
06426adf JA |
3115 | break; |
3116 | set_current_state(TASK_UNINTERRUPTIBLE); | |
3117 | hrtimer_start_expires(&hs.timer, mode); | |
3118 | if (hs.task) | |
3119 | io_schedule(); | |
3120 | hrtimer_cancel(&hs.timer); | |
3121 | mode = HRTIMER_MODE_ABS; | |
3122 | } while (hs.task && !signal_pending(current)); | |
3123 | ||
3124 | __set_current_state(TASK_RUNNING); | |
3125 | destroy_hrtimer_on_stack(&hs.timer); | |
3126 | return true; | |
3127 | } | |
3128 | ||
bbd7bb70 JA |
3129 | static bool __blk_mq_poll(struct blk_mq_hw_ctx *hctx, struct request *rq) |
3130 | { | |
3131 | struct request_queue *q = hctx->queue; | |
3132 | long state; | |
3133 | ||
06426adf JA |
3134 | /* |
3135 | * If we sleep, have the caller restart the poll loop to reset | |
3136 | * the state. Like for the other success return cases, the | |
3137 | * caller is responsible for checking if the IO completed. If | |
3138 | * the IO isn't complete, we'll get called again and will go | |
3139 | * straight to the busy poll loop. | |
3140 | */ | |
64f1c21e | 3141 | if (blk_mq_poll_hybrid_sleep(q, hctx, rq)) |
06426adf JA |
3142 | return true; |
3143 | ||
bbd7bb70 JA |
3144 | hctx->poll_considered++; |
3145 | ||
3146 | state = current->state; | |
3147 | while (!need_resched()) { | |
3148 | int ret; | |
3149 | ||
3150 | hctx->poll_invoked++; | |
3151 | ||
3152 | ret = q->mq_ops->poll(hctx, rq->tag); | |
3153 | if (ret > 0) { | |
3154 | hctx->poll_success++; | |
3155 | set_current_state(TASK_RUNNING); | |
3156 | return true; | |
3157 | } | |
3158 | ||
3159 | if (signal_pending_state(state, current)) | |
3160 | set_current_state(TASK_RUNNING); | |
3161 | ||
3162 | if (current->state == TASK_RUNNING) | |
3163 | return true; | |
3164 | if (ret < 0) | |
3165 | break; | |
3166 | cpu_relax(); | |
3167 | } | |
3168 | ||
67b4110f | 3169 | __set_current_state(TASK_RUNNING); |
bbd7bb70 JA |
3170 | return false; |
3171 | } | |
3172 | ||
ea435e1b | 3173 | static bool blk_mq_poll(struct request_queue *q, blk_qc_t cookie) |
bbd7bb70 JA |
3174 | { |
3175 | struct blk_mq_hw_ctx *hctx; | |
bbd7bb70 JA |
3176 | struct request *rq; |
3177 | ||
ea435e1b | 3178 | if (!test_bit(QUEUE_FLAG_POLL, &q->queue_flags)) |
bbd7bb70 JA |
3179 | return false; |
3180 | ||
bbd7bb70 | 3181 | hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)]; |
bd166ef1 JA |
3182 | if (!blk_qc_t_is_internal(cookie)) |
3183 | rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie)); | |
3a07bb1d | 3184 | else { |
bd166ef1 | 3185 | rq = blk_mq_tag_to_rq(hctx->sched_tags, blk_qc_t_to_tag(cookie)); |
3a07bb1d JA |
3186 | /* |
3187 | * With scheduling, if the request has completed, we'll | |
3188 | * get a NULL return here, as we clear the sched tag when | |
3189 | * that happens. The request still remains valid, like always, | |
3190 | * so we should be safe with just the NULL check. | |
3191 | */ | |
3192 | if (!rq) | |
3193 | return false; | |
3194 | } | |
bbd7bb70 JA |
3195 | |
3196 | return __blk_mq_poll(hctx, rq); | |
3197 | } | |
bbd7bb70 | 3198 | |
320ae51f JA |
3199 | static int __init blk_mq_init(void) |
3200 | { | |
9467f859 TG |
3201 | cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL, |
3202 | blk_mq_hctx_notify_dead); | |
320ae51f JA |
3203 | return 0; |
3204 | } | |
3205 | subsys_initcall(blk_mq_init); |