Commit | Line | Data |
---|---|---|
3dcf60bc | 1 | // SPDX-License-Identifier: GPL-2.0 |
75bb4625 JA |
2 | /* |
3 | * Block multiqueue core code | |
4 | * | |
5 | * Copyright (C) 2013-2014 Jens Axboe | |
6 | * Copyright (C) 2013-2014 Christoph Hellwig | |
7 | */ | |
320ae51f JA |
8 | #include <linux/kernel.h> |
9 | #include <linux/module.h> | |
10 | #include <linux/backing-dev.h> | |
11 | #include <linux/bio.h> | |
12 | #include <linux/blkdev.h> | |
fe45e630 | 13 | #include <linux/blk-integrity.h> |
f75782e4 | 14 | #include <linux/kmemleak.h> |
320ae51f JA |
15 | #include <linux/mm.h> |
16 | #include <linux/init.h> | |
17 | #include <linux/slab.h> | |
18 | #include <linux/workqueue.h> | |
19 | #include <linux/smp.h> | |
e41d12f5 | 20 | #include <linux/interrupt.h> |
320ae51f | 21 | #include <linux/llist.h> |
320ae51f JA |
22 | #include <linux/cpu.h> |
23 | #include <linux/cache.h> | |
24 | #include <linux/sched/sysctl.h> | |
105ab3d8 | 25 | #include <linux/sched/topology.h> |
174cd4b1 | 26 | #include <linux/sched/signal.h> |
320ae51f | 27 | #include <linux/delay.h> |
aedcd72f | 28 | #include <linux/crash_dump.h> |
88c7b2b7 | 29 | #include <linux/prefetch.h> |
a892c8d5 | 30 | #include <linux/blk-crypto.h> |
82d981d4 | 31 | #include <linux/part_stat.h> |
320ae51f JA |
32 | |
33 | #include <trace/events/block.h> | |
34 | ||
35 | #include <linux/blk-mq.h> | |
54d4e6ab | 36 | #include <linux/t10-pi.h> |
320ae51f JA |
37 | #include "blk.h" |
38 | #include "blk-mq.h" | |
9c1051aa | 39 | #include "blk-mq-debugfs.h" |
320ae51f | 40 | #include "blk-mq-tag.h" |
986d413b | 41 | #include "blk-pm.h" |
cf43e6be | 42 | #include "blk-stat.h" |
bd166ef1 | 43 | #include "blk-mq-sched.h" |
c1c80384 | 44 | #include "blk-rq-qos.h" |
320ae51f | 45 | |
f9ab4918 | 46 | static DEFINE_PER_CPU(struct llist_head, blk_cpu_done); |
c3077b5d | 47 | |
34dbad5d OS |
48 | static void blk_mq_poll_stats_start(struct request_queue *q); |
49 | static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb); | |
50 | ||
720b8ccc SB |
51 | static int blk_mq_poll_stats_bkt(const struct request *rq) |
52 | { | |
3d244306 | 53 | int ddir, sectors, bucket; |
720b8ccc | 54 | |
99c749a4 | 55 | ddir = rq_data_dir(rq); |
3d244306 | 56 | sectors = blk_rq_stats_sectors(rq); |
720b8ccc | 57 | |
3d244306 | 58 | bucket = ddir + 2 * ilog2(sectors); |
720b8ccc SB |
59 | |
60 | if (bucket < 0) | |
61 | return -1; | |
62 | else if (bucket >= BLK_MQ_POLL_STATS_BKTS) | |
63 | return ddir + BLK_MQ_POLL_STATS_BKTS - 2; | |
64 | ||
65 | return bucket; | |
66 | } | |
67 | ||
3e08773c CH |
68 | #define BLK_QC_T_SHIFT 16 |
69 | #define BLK_QC_T_INTERNAL (1U << 31) | |
70 | ||
f70299f0 CH |
71 | static inline struct blk_mq_hw_ctx *blk_qc_to_hctx(struct request_queue *q, |
72 | blk_qc_t qc) | |
73 | { | |
74 | return q->queue_hw_ctx[(qc & ~BLK_QC_T_INTERNAL) >> BLK_QC_T_SHIFT]; | |
75 | } | |
76 | ||
c6699d6f CH |
77 | static inline struct request *blk_qc_to_rq(struct blk_mq_hw_ctx *hctx, |
78 | blk_qc_t qc) | |
79 | { | |
efbabbe1 CH |
80 | unsigned int tag = qc & ((1U << BLK_QC_T_SHIFT) - 1); |
81 | ||
82 | if (qc & BLK_QC_T_INTERNAL) | |
83 | return blk_mq_tag_to_rq(hctx->sched_tags, tag); | |
84 | return blk_mq_tag_to_rq(hctx->tags, tag); | |
c6699d6f CH |
85 | } |
86 | ||
3e08773c CH |
87 | static inline blk_qc_t blk_rq_to_qc(struct request *rq) |
88 | { | |
89 | return (rq->mq_hctx->queue_num << BLK_QC_T_SHIFT) | | |
90 | (rq->tag != -1 ? | |
91 | rq->tag : (rq->internal_tag | BLK_QC_T_INTERNAL)); | |
92 | } | |
93 | ||
320ae51f | 94 | /* |
85fae294 YY |
95 | * Check if any of the ctx, dispatch list or elevator |
96 | * have pending work in this hardware queue. | |
320ae51f | 97 | */ |
79f720a7 | 98 | static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx) |
320ae51f | 99 | { |
79f720a7 JA |
100 | return !list_empty_careful(&hctx->dispatch) || |
101 | sbitmap_any_bit_set(&hctx->ctx_map) || | |
bd166ef1 | 102 | blk_mq_sched_has_work(hctx); |
1429d7c9 JA |
103 | } |
104 | ||
320ae51f JA |
105 | /* |
106 | * Mark this ctx as having pending work in this hardware queue | |
107 | */ | |
108 | static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx, | |
109 | struct blk_mq_ctx *ctx) | |
110 | { | |
f31967f0 JA |
111 | const int bit = ctx->index_hw[hctx->type]; |
112 | ||
113 | if (!sbitmap_test_bit(&hctx->ctx_map, bit)) | |
114 | sbitmap_set_bit(&hctx->ctx_map, bit); | |
1429d7c9 JA |
115 | } |
116 | ||
117 | static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx, | |
118 | struct blk_mq_ctx *ctx) | |
119 | { | |
f31967f0 JA |
120 | const int bit = ctx->index_hw[hctx->type]; |
121 | ||
122 | sbitmap_clear_bit(&hctx->ctx_map, bit); | |
320ae51f JA |
123 | } |
124 | ||
f299b7c7 | 125 | struct mq_inflight { |
8446fe92 | 126 | struct block_device *part; |
a2e80f6f | 127 | unsigned int inflight[2]; |
f299b7c7 JA |
128 | }; |
129 | ||
8ab30a33 | 130 | static bool blk_mq_check_inflight(struct request *rq, void *priv, |
f299b7c7 JA |
131 | bool reserved) |
132 | { | |
133 | struct mq_inflight *mi = priv; | |
134 | ||
b0d97557 JX |
135 | if ((!mi->part->bd_partno || rq->part == mi->part) && |
136 | blk_mq_rq_state(rq) == MQ_RQ_IN_FLIGHT) | |
bb4e6b14 | 137 | mi->inflight[rq_data_dir(rq)]++; |
7baa8572 JA |
138 | |
139 | return true; | |
f299b7c7 JA |
140 | } |
141 | ||
8446fe92 CH |
142 | unsigned int blk_mq_in_flight(struct request_queue *q, |
143 | struct block_device *part) | |
f299b7c7 | 144 | { |
a2e80f6f | 145 | struct mq_inflight mi = { .part = part }; |
f299b7c7 | 146 | |
f299b7c7 | 147 | blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi); |
e016b782 | 148 | |
a2e80f6f | 149 | return mi.inflight[0] + mi.inflight[1]; |
bf0ddaba OS |
150 | } |
151 | ||
8446fe92 CH |
152 | void blk_mq_in_flight_rw(struct request_queue *q, struct block_device *part, |
153 | unsigned int inflight[2]) | |
bf0ddaba | 154 | { |
a2e80f6f | 155 | struct mq_inflight mi = { .part = part }; |
bf0ddaba | 156 | |
bb4e6b14 | 157 | blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi); |
a2e80f6f PB |
158 | inflight[0] = mi.inflight[0]; |
159 | inflight[1] = mi.inflight[1]; | |
bf0ddaba OS |
160 | } |
161 | ||
1671d522 | 162 | void blk_freeze_queue_start(struct request_queue *q) |
43a5e4e2 | 163 | { |
7996a8b5 BL |
164 | mutex_lock(&q->mq_freeze_lock); |
165 | if (++q->mq_freeze_depth == 1) { | |
3ef28e83 | 166 | percpu_ref_kill(&q->q_usage_counter); |
7996a8b5 | 167 | mutex_unlock(&q->mq_freeze_lock); |
344e9ffc | 168 | if (queue_is_mq(q)) |
055f6e18 | 169 | blk_mq_run_hw_queues(q, false); |
7996a8b5 BL |
170 | } else { |
171 | mutex_unlock(&q->mq_freeze_lock); | |
cddd5d17 | 172 | } |
f3af020b | 173 | } |
1671d522 | 174 | EXPORT_SYMBOL_GPL(blk_freeze_queue_start); |
f3af020b | 175 | |
6bae363e | 176 | void blk_mq_freeze_queue_wait(struct request_queue *q) |
f3af020b | 177 | { |
3ef28e83 | 178 | wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter)); |
43a5e4e2 | 179 | } |
6bae363e | 180 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait); |
43a5e4e2 | 181 | |
f91328c4 KB |
182 | int blk_mq_freeze_queue_wait_timeout(struct request_queue *q, |
183 | unsigned long timeout) | |
184 | { | |
185 | return wait_event_timeout(q->mq_freeze_wq, | |
186 | percpu_ref_is_zero(&q->q_usage_counter), | |
187 | timeout); | |
188 | } | |
189 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout); | |
43a5e4e2 | 190 | |
f3af020b TH |
191 | /* |
192 | * Guarantee no request is in use, so we can change any data structure of | |
193 | * the queue afterward. | |
194 | */ | |
3ef28e83 | 195 | void blk_freeze_queue(struct request_queue *q) |
f3af020b | 196 | { |
3ef28e83 DW |
197 | /* |
198 | * In the !blk_mq case we are only calling this to kill the | |
199 | * q_usage_counter, otherwise this increases the freeze depth | |
200 | * and waits for it to return to zero. For this reason there is | |
201 | * no blk_unfreeze_queue(), and blk_freeze_queue() is not | |
202 | * exported to drivers as the only user for unfreeze is blk_mq. | |
203 | */ | |
1671d522 | 204 | blk_freeze_queue_start(q); |
f3af020b TH |
205 | blk_mq_freeze_queue_wait(q); |
206 | } | |
3ef28e83 DW |
207 | |
208 | void blk_mq_freeze_queue(struct request_queue *q) | |
209 | { | |
210 | /* | |
211 | * ...just an alias to keep freeze and unfreeze actions balanced | |
212 | * in the blk_mq_* namespace | |
213 | */ | |
214 | blk_freeze_queue(q); | |
215 | } | |
c761d96b | 216 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue); |
f3af020b | 217 | |
aec89dc5 | 218 | void __blk_mq_unfreeze_queue(struct request_queue *q, bool force_atomic) |
320ae51f | 219 | { |
7996a8b5 | 220 | mutex_lock(&q->mq_freeze_lock); |
aec89dc5 CH |
221 | if (force_atomic) |
222 | q->q_usage_counter.data->force_atomic = true; | |
7996a8b5 BL |
223 | q->mq_freeze_depth--; |
224 | WARN_ON_ONCE(q->mq_freeze_depth < 0); | |
225 | if (!q->mq_freeze_depth) { | |
bdd63160 | 226 | percpu_ref_resurrect(&q->q_usage_counter); |
320ae51f | 227 | wake_up_all(&q->mq_freeze_wq); |
add703fd | 228 | } |
7996a8b5 | 229 | mutex_unlock(&q->mq_freeze_lock); |
320ae51f | 230 | } |
aec89dc5 CH |
231 | |
232 | void blk_mq_unfreeze_queue(struct request_queue *q) | |
233 | { | |
234 | __blk_mq_unfreeze_queue(q, false); | |
235 | } | |
b4c6a028 | 236 | EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue); |
320ae51f | 237 | |
852ec809 BVA |
238 | /* |
239 | * FIXME: replace the scsi_internal_device_*block_nowait() calls in the | |
240 | * mpt3sas driver such that this function can be removed. | |
241 | */ | |
242 | void blk_mq_quiesce_queue_nowait(struct request_queue *q) | |
243 | { | |
e70feb8b ML |
244 | unsigned long flags; |
245 | ||
246 | spin_lock_irqsave(&q->queue_lock, flags); | |
247 | if (!q->quiesce_depth++) | |
248 | blk_queue_flag_set(QUEUE_FLAG_QUIESCED, q); | |
249 | spin_unlock_irqrestore(&q->queue_lock, flags); | |
852ec809 BVA |
250 | } |
251 | EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait); | |
252 | ||
6a83e74d | 253 | /** |
9ef4d020 | 254 | * blk_mq_wait_quiesce_done() - wait until in-progress quiesce is done |
6a83e74d BVA |
255 | * @q: request queue. |
256 | * | |
9ef4d020 ML |
257 | * Note: it is driver's responsibility for making sure that quiesce has |
258 | * been started. | |
6a83e74d | 259 | */ |
9ef4d020 | 260 | void blk_mq_wait_quiesce_done(struct request_queue *q) |
6a83e74d | 261 | { |
704b914f ML |
262 | if (blk_queue_has_srcu(q)) |
263 | synchronize_srcu(q->srcu); | |
264 | else | |
6a83e74d BVA |
265 | synchronize_rcu(); |
266 | } | |
9ef4d020 ML |
267 | EXPORT_SYMBOL_GPL(blk_mq_wait_quiesce_done); |
268 | ||
269 | /** | |
270 | * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished | |
271 | * @q: request queue. | |
272 | * | |
273 | * Note: this function does not prevent that the struct request end_io() | |
274 | * callback function is invoked. Once this function is returned, we make | |
275 | * sure no dispatch can happen until the queue is unquiesced via | |
276 | * blk_mq_unquiesce_queue(). | |
277 | */ | |
278 | void blk_mq_quiesce_queue(struct request_queue *q) | |
279 | { | |
280 | blk_mq_quiesce_queue_nowait(q); | |
281 | blk_mq_wait_quiesce_done(q); | |
282 | } | |
6a83e74d BVA |
283 | EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue); |
284 | ||
e4e73913 ML |
285 | /* |
286 | * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue() | |
287 | * @q: request queue. | |
288 | * | |
289 | * This function recovers queue into the state before quiescing | |
290 | * which is done by blk_mq_quiesce_queue. | |
291 | */ | |
292 | void blk_mq_unquiesce_queue(struct request_queue *q) | |
293 | { | |
e70feb8b ML |
294 | unsigned long flags; |
295 | bool run_queue = false; | |
296 | ||
297 | spin_lock_irqsave(&q->queue_lock, flags); | |
298 | if (WARN_ON_ONCE(q->quiesce_depth <= 0)) { | |
299 | ; | |
300 | } else if (!--q->quiesce_depth) { | |
301 | blk_queue_flag_clear(QUEUE_FLAG_QUIESCED, q); | |
302 | run_queue = true; | |
303 | } | |
304 | spin_unlock_irqrestore(&q->queue_lock, flags); | |
f4560ffe | 305 | |
1d9e9bc6 | 306 | /* dispatch requests which are inserted during quiescing */ |
e70feb8b ML |
307 | if (run_queue) |
308 | blk_mq_run_hw_queues(q, true); | |
e4e73913 ML |
309 | } |
310 | EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue); | |
311 | ||
aed3ea94 JA |
312 | void blk_mq_wake_waiters(struct request_queue *q) |
313 | { | |
314 | struct blk_mq_hw_ctx *hctx; | |
315 | unsigned int i; | |
316 | ||
317 | queue_for_each_hw_ctx(q, hctx, i) | |
318 | if (blk_mq_hw_queue_mapped(hctx)) | |
319 | blk_mq_tag_wakeup_all(hctx->tags, true); | |
320 | } | |
321 | ||
52fdbbcc CH |
322 | void blk_rq_init(struct request_queue *q, struct request *rq) |
323 | { | |
324 | memset(rq, 0, sizeof(*rq)); | |
325 | ||
326 | INIT_LIST_HEAD(&rq->queuelist); | |
327 | rq->q = q; | |
328 | rq->__sector = (sector_t) -1; | |
329 | INIT_HLIST_NODE(&rq->hash); | |
330 | RB_CLEAR_NODE(&rq->rb_node); | |
331 | rq->tag = BLK_MQ_NO_TAG; | |
332 | rq->internal_tag = BLK_MQ_NO_TAG; | |
333 | rq->start_time_ns = ktime_get_ns(); | |
334 | rq->part = NULL; | |
335 | blk_crypto_rq_set_defaults(rq); | |
336 | } | |
337 | EXPORT_SYMBOL(blk_rq_init); | |
338 | ||
e4cdf1a1 | 339 | static struct request *blk_mq_rq_ctx_init(struct blk_mq_alloc_data *data, |
fe6134f6 | 340 | struct blk_mq_tags *tags, unsigned int tag, u64 alloc_time_ns) |
320ae51f | 341 | { |
605f784e PB |
342 | struct blk_mq_ctx *ctx = data->ctx; |
343 | struct blk_mq_hw_ctx *hctx = data->hctx; | |
344 | struct request_queue *q = data->q; | |
e4cdf1a1 | 345 | struct request *rq = tags->static_rqs[tag]; |
c3a148d2 | 346 | |
c7b84d42 JA |
347 | rq->q = q; |
348 | rq->mq_ctx = ctx; | |
349 | rq->mq_hctx = hctx; | |
350 | rq->cmd_flags = data->cmd_flags; | |
351 | ||
352 | if (data->flags & BLK_MQ_REQ_PM) | |
353 | data->rq_flags |= RQF_PM; | |
354 | if (blk_queue_io_stat(q)) | |
355 | data->rq_flags |= RQF_IO_STAT; | |
356 | rq->rq_flags = data->rq_flags; | |
357 | ||
56f8da64 | 358 | if (!(data->rq_flags & RQF_ELV)) { |
e4cdf1a1 | 359 | rq->tag = tag; |
76647368 | 360 | rq->internal_tag = BLK_MQ_NO_TAG; |
56f8da64 JA |
361 | } else { |
362 | rq->tag = BLK_MQ_NO_TAG; | |
363 | rq->internal_tag = tag; | |
e4cdf1a1 | 364 | } |
c7b84d42 | 365 | rq->timeout = 0; |
e4cdf1a1 | 366 | |
4f266f2b PB |
367 | if (blk_mq_need_time_stamp(rq)) |
368 | rq->start_time_ns = ktime_get_ns(); | |
369 | else | |
370 | rq->start_time_ns = 0; | |
af76e555 | 371 | rq->part = NULL; |
6f816b4b TH |
372 | #ifdef CONFIG_BLK_RQ_ALLOC_TIME |
373 | rq->alloc_time_ns = alloc_time_ns; | |
374 | #endif | |
544ccc8d | 375 | rq->io_start_time_ns = 0; |
3d244306 | 376 | rq->stats_sectors = 0; |
af76e555 CH |
377 | rq->nr_phys_segments = 0; |
378 | #if defined(CONFIG_BLK_DEV_INTEGRITY) | |
379 | rq->nr_integrity_segments = 0; | |
380 | #endif | |
af76e555 CH |
381 | rq->end_io = NULL; |
382 | rq->end_io_data = NULL; | |
af76e555 | 383 | |
4f266f2b PB |
384 | blk_crypto_rq_set_defaults(rq); |
385 | INIT_LIST_HEAD(&rq->queuelist); | |
386 | /* tag was already set */ | |
387 | WRITE_ONCE(rq->deadline, 0); | |
0a467d0f | 388 | req_ref_set(rq, 1); |
7ea4d8a4 | 389 | |
4f266f2b | 390 | if (rq->rq_flags & RQF_ELV) { |
7ea4d8a4 CH |
391 | struct elevator_queue *e = data->q->elevator; |
392 | ||
4f266f2b PB |
393 | INIT_HLIST_NODE(&rq->hash); |
394 | RB_CLEAR_NODE(&rq->rb_node); | |
395 | ||
396 | if (!op_is_flush(data->cmd_flags) && | |
397 | e->type->ops.prepare_request) { | |
7ea4d8a4 CH |
398 | e->type->ops.prepare_request(rq); |
399 | rq->rq_flags |= RQF_ELVPRIV; | |
400 | } | |
401 | } | |
402 | ||
e4cdf1a1 | 403 | return rq; |
5dee8577 CH |
404 | } |
405 | ||
349302da JA |
406 | static inline struct request * |
407 | __blk_mq_alloc_requests_batch(struct blk_mq_alloc_data *data, | |
408 | u64 alloc_time_ns) | |
409 | { | |
410 | unsigned int tag, tag_offset; | |
fe6134f6 | 411 | struct blk_mq_tags *tags; |
349302da | 412 | struct request *rq; |
fe6134f6 | 413 | unsigned long tag_mask; |
349302da JA |
414 | int i, nr = 0; |
415 | ||
fe6134f6 JA |
416 | tag_mask = blk_mq_get_tags(data, data->nr_tags, &tag_offset); |
417 | if (unlikely(!tag_mask)) | |
349302da JA |
418 | return NULL; |
419 | ||
fe6134f6 JA |
420 | tags = blk_mq_tags_from_data(data); |
421 | for (i = 0; tag_mask; i++) { | |
422 | if (!(tag_mask & (1UL << i))) | |
349302da JA |
423 | continue; |
424 | tag = tag_offset + i; | |
a22c00be | 425 | prefetch(tags->static_rqs[tag]); |
fe6134f6 JA |
426 | tag_mask &= ~(1UL << i); |
427 | rq = blk_mq_rq_ctx_init(data, tags, tag, alloc_time_ns); | |
013a7f95 | 428 | rq_list_add(data->cached_rq, rq); |
c5fc7b93 | 429 | nr++; |
349302da | 430 | } |
c5fc7b93 JA |
431 | /* caller already holds a reference, add for remainder */ |
432 | percpu_ref_get_many(&data->q->q_usage_counter, nr - 1); | |
349302da JA |
433 | data->nr_tags -= nr; |
434 | ||
013a7f95 | 435 | return rq_list_pop(data->cached_rq); |
349302da JA |
436 | } |
437 | ||
b90cfaed | 438 | static struct request *__blk_mq_alloc_requests(struct blk_mq_alloc_data *data) |
d2c0d383 | 439 | { |
e6e7abff | 440 | struct request_queue *q = data->q; |
6f816b4b | 441 | u64 alloc_time_ns = 0; |
47c122e3 | 442 | struct request *rq; |
600c3b0c | 443 | unsigned int tag; |
d2c0d383 | 444 | |
6f816b4b TH |
445 | /* alloc_time includes depth and tag waits */ |
446 | if (blk_queue_rq_alloc_time(q)) | |
447 | alloc_time_ns = ktime_get_ns(); | |
448 | ||
f9afca4d | 449 | if (data->cmd_flags & REQ_NOWAIT) |
03a07c92 | 450 | data->flags |= BLK_MQ_REQ_NOWAIT; |
d2c0d383 | 451 | |
781dd830 JA |
452 | if (q->elevator) { |
453 | struct elevator_queue *e = q->elevator; | |
454 | ||
455 | data->rq_flags |= RQF_ELV; | |
456 | ||
d2c0d383 | 457 | /* |
8d663f34 | 458 | * Flush/passthrough requests are special and go directly to the |
17a51199 JA |
459 | * dispatch list. Don't include reserved tags in the |
460 | * limiting, as it isn't useful. | |
d2c0d383 | 461 | */ |
f9afca4d | 462 | if (!op_is_flush(data->cmd_flags) && |
8d663f34 | 463 | !blk_op_is_passthrough(data->cmd_flags) && |
f9afca4d | 464 | e->type->ops.limit_depth && |
17a51199 | 465 | !(data->flags & BLK_MQ_REQ_RESERVED)) |
f9afca4d | 466 | e->type->ops.limit_depth(data->cmd_flags, data); |
d2c0d383 CH |
467 | } |
468 | ||
bf0beec0 | 469 | retry: |
600c3b0c CH |
470 | data->ctx = blk_mq_get_ctx(q); |
471 | data->hctx = blk_mq_map_queue(q, data->cmd_flags, data->ctx); | |
781dd830 | 472 | if (!(data->rq_flags & RQF_ELV)) |
600c3b0c CH |
473 | blk_mq_tag_busy(data->hctx); |
474 | ||
349302da JA |
475 | /* |
476 | * Try batched alloc if we want more than 1 tag. | |
477 | */ | |
478 | if (data->nr_tags > 1) { | |
479 | rq = __blk_mq_alloc_requests_batch(data, alloc_time_ns); | |
480 | if (rq) | |
481 | return rq; | |
482 | data->nr_tags = 1; | |
483 | } | |
484 | ||
bf0beec0 ML |
485 | /* |
486 | * Waiting allocations only fail because of an inactive hctx. In that | |
487 | * case just retry the hctx assignment and tag allocation as CPU hotplug | |
488 | * should have migrated us to an online CPU by now. | |
489 | */ | |
e4cdf1a1 | 490 | tag = blk_mq_get_tag(data); |
bf0beec0 ML |
491 | if (tag == BLK_MQ_NO_TAG) { |
492 | if (data->flags & BLK_MQ_REQ_NOWAIT) | |
493 | return NULL; | |
bf0beec0 | 494 | /* |
349302da JA |
495 | * Give up the CPU and sleep for a random short time to |
496 | * ensure that thread using a realtime scheduling class | |
497 | * are migrated off the CPU, and thus off the hctx that | |
498 | * is going away. | |
bf0beec0 ML |
499 | */ |
500 | msleep(3); | |
501 | goto retry; | |
502 | } | |
47c122e3 | 503 | |
fe6134f6 JA |
504 | return blk_mq_rq_ctx_init(data, blk_mq_tags_from_data(data), tag, |
505 | alloc_time_ns); | |
d2c0d383 CH |
506 | } |
507 | ||
cd6ce148 | 508 | struct request *blk_mq_alloc_request(struct request_queue *q, unsigned int op, |
9a95e4ef | 509 | blk_mq_req_flags_t flags) |
320ae51f | 510 | { |
e6e7abff CH |
511 | struct blk_mq_alloc_data data = { |
512 | .q = q, | |
513 | .flags = flags, | |
514 | .cmd_flags = op, | |
47c122e3 | 515 | .nr_tags = 1, |
e6e7abff | 516 | }; |
bd166ef1 | 517 | struct request *rq; |
a492f075 | 518 | int ret; |
320ae51f | 519 | |
3a0a5299 | 520 | ret = blk_queue_enter(q, flags); |
a492f075 JL |
521 | if (ret) |
522 | return ERR_PTR(ret); | |
320ae51f | 523 | |
b90cfaed | 524 | rq = __blk_mq_alloc_requests(&data); |
bd166ef1 | 525 | if (!rq) |
a5ea5811 | 526 | goto out_queue_exit; |
0c4de0f3 CH |
527 | rq->__data_len = 0; |
528 | rq->__sector = (sector_t) -1; | |
529 | rq->bio = rq->biotail = NULL; | |
320ae51f | 530 | return rq; |
a5ea5811 CH |
531 | out_queue_exit: |
532 | blk_queue_exit(q); | |
533 | return ERR_PTR(-EWOULDBLOCK); | |
320ae51f | 534 | } |
4bb659b1 | 535 | EXPORT_SYMBOL(blk_mq_alloc_request); |
320ae51f | 536 | |
cd6ce148 | 537 | struct request *blk_mq_alloc_request_hctx(struct request_queue *q, |
9a95e4ef | 538 | unsigned int op, blk_mq_req_flags_t flags, unsigned int hctx_idx) |
1f5bd336 | 539 | { |
e6e7abff CH |
540 | struct blk_mq_alloc_data data = { |
541 | .q = q, | |
542 | .flags = flags, | |
543 | .cmd_flags = op, | |
47c122e3 | 544 | .nr_tags = 1, |
e6e7abff | 545 | }; |
600c3b0c | 546 | u64 alloc_time_ns = 0; |
6d2809d5 | 547 | unsigned int cpu; |
600c3b0c | 548 | unsigned int tag; |
1f5bd336 ML |
549 | int ret; |
550 | ||
600c3b0c CH |
551 | /* alloc_time includes depth and tag waits */ |
552 | if (blk_queue_rq_alloc_time(q)) | |
553 | alloc_time_ns = ktime_get_ns(); | |
554 | ||
1f5bd336 ML |
555 | /* |
556 | * If the tag allocator sleeps we could get an allocation for a | |
557 | * different hardware context. No need to complicate the low level | |
558 | * allocator for this for the rare use case of a command tied to | |
559 | * a specific queue. | |
560 | */ | |
600c3b0c | 561 | if (WARN_ON_ONCE(!(flags & (BLK_MQ_REQ_NOWAIT | BLK_MQ_REQ_RESERVED)))) |
1f5bd336 ML |
562 | return ERR_PTR(-EINVAL); |
563 | ||
564 | if (hctx_idx >= q->nr_hw_queues) | |
565 | return ERR_PTR(-EIO); | |
566 | ||
3a0a5299 | 567 | ret = blk_queue_enter(q, flags); |
1f5bd336 ML |
568 | if (ret) |
569 | return ERR_PTR(ret); | |
570 | ||
c8712c6a CH |
571 | /* |
572 | * Check if the hardware context is actually mapped to anything. | |
573 | * If not tell the caller that it should skip this queue. | |
574 | */ | |
a5ea5811 | 575 | ret = -EXDEV; |
e6e7abff CH |
576 | data.hctx = q->queue_hw_ctx[hctx_idx]; |
577 | if (!blk_mq_hw_queue_mapped(data.hctx)) | |
a5ea5811 | 578 | goto out_queue_exit; |
e6e7abff CH |
579 | cpu = cpumask_first_and(data.hctx->cpumask, cpu_online_mask); |
580 | data.ctx = __blk_mq_get_ctx(q, cpu); | |
1f5bd336 | 581 | |
42fdc5e4 | 582 | if (!q->elevator) |
600c3b0c | 583 | blk_mq_tag_busy(data.hctx); |
781dd830 JA |
584 | else |
585 | data.rq_flags |= RQF_ELV; | |
600c3b0c | 586 | |
a5ea5811 | 587 | ret = -EWOULDBLOCK; |
600c3b0c CH |
588 | tag = blk_mq_get_tag(&data); |
589 | if (tag == BLK_MQ_NO_TAG) | |
a5ea5811 | 590 | goto out_queue_exit; |
fe6134f6 JA |
591 | return blk_mq_rq_ctx_init(&data, blk_mq_tags_from_data(&data), tag, |
592 | alloc_time_ns); | |
600c3b0c | 593 | |
a5ea5811 CH |
594 | out_queue_exit: |
595 | blk_queue_exit(q); | |
596 | return ERR_PTR(ret); | |
1f5bd336 ML |
597 | } |
598 | EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx); | |
599 | ||
12f5b931 KB |
600 | static void __blk_mq_free_request(struct request *rq) |
601 | { | |
602 | struct request_queue *q = rq->q; | |
603 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
ea4f995e | 604 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; |
12f5b931 KB |
605 | const int sched_tag = rq->internal_tag; |
606 | ||
a892c8d5 | 607 | blk_crypto_free_request(rq); |
986d413b | 608 | blk_pm_mark_last_busy(rq); |
ea4f995e | 609 | rq->mq_hctx = NULL; |
76647368 | 610 | if (rq->tag != BLK_MQ_NO_TAG) |
cae740a0 | 611 | blk_mq_put_tag(hctx->tags, ctx, rq->tag); |
76647368 | 612 | if (sched_tag != BLK_MQ_NO_TAG) |
cae740a0 | 613 | blk_mq_put_tag(hctx->sched_tags, ctx, sched_tag); |
12f5b931 KB |
614 | blk_mq_sched_restart(hctx); |
615 | blk_queue_exit(q); | |
616 | } | |
617 | ||
6af54051 | 618 | void blk_mq_free_request(struct request *rq) |
320ae51f | 619 | { |
320ae51f | 620 | struct request_queue *q = rq->q; |
ea4f995e | 621 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; |
6af54051 | 622 | |
222ee581 CH |
623 | if ((rq->rq_flags & RQF_ELVPRIV) && |
624 | q->elevator->type->ops.finish_request) | |
625 | q->elevator->type->ops.finish_request(rq); | |
320ae51f | 626 | |
e8064021 | 627 | if (rq->rq_flags & RQF_MQ_INFLIGHT) |
bccf5e26 | 628 | __blk_mq_dec_active_requests(hctx); |
87760e5e | 629 | |
7beb2f84 | 630 | if (unlikely(laptop_mode && !blk_rq_is_passthrough(rq))) |
d152c682 | 631 | laptop_io_completion(q->disk->bdi); |
7beb2f84 | 632 | |
a7905043 | 633 | rq_qos_done(q, rq); |
0d2602ca | 634 | |
12f5b931 | 635 | WRITE_ONCE(rq->state, MQ_RQ_IDLE); |
0a467d0f | 636 | if (req_ref_put_and_test(rq)) |
12f5b931 | 637 | __blk_mq_free_request(rq); |
320ae51f | 638 | } |
1a3b595a | 639 | EXPORT_SYMBOL_GPL(blk_mq_free_request); |
320ae51f | 640 | |
47c122e3 | 641 | void blk_mq_free_plug_rqs(struct blk_plug *plug) |
320ae51f | 642 | { |
013a7f95 | 643 | struct request *rq; |
fe1f4526 | 644 | |
c5fc7b93 | 645 | while ((rq = rq_list_pop(&plug->cached_rq)) != NULL) |
47c122e3 | 646 | blk_mq_free_request(rq); |
47c122e3 | 647 | } |
522a7775 | 648 | |
22350ad7 CH |
649 | void blk_dump_rq_flags(struct request *rq, char *msg) |
650 | { | |
651 | printk(KERN_INFO "%s: dev %s: flags=%llx\n", msg, | |
f3fa33ac | 652 | rq->q->disk ? rq->q->disk->disk_name : "?", |
22350ad7 CH |
653 | (unsigned long long) rq->cmd_flags); |
654 | ||
655 | printk(KERN_INFO " sector %llu, nr/cnr %u/%u\n", | |
656 | (unsigned long long)blk_rq_pos(rq), | |
657 | blk_rq_sectors(rq), blk_rq_cur_sectors(rq)); | |
658 | printk(KERN_INFO " bio %p, biotail %p, len %u\n", | |
659 | rq->bio, rq->biotail, blk_rq_bytes(rq)); | |
660 | } | |
661 | EXPORT_SYMBOL(blk_dump_rq_flags); | |
662 | ||
9be3e06f JA |
663 | static void req_bio_endio(struct request *rq, struct bio *bio, |
664 | unsigned int nbytes, blk_status_t error) | |
665 | { | |
478eb72b | 666 | if (unlikely(error)) { |
9be3e06f | 667 | bio->bi_status = error; |
478eb72b | 668 | } else if (req_op(rq) == REQ_OP_ZONE_APPEND) { |
9be3e06f JA |
669 | /* |
670 | * Partial zone append completions cannot be supported as the | |
671 | * BIO fragments may end up not being written sequentially. | |
672 | */ | |
297db731 | 673 | if (bio->bi_iter.bi_size != nbytes) |
9be3e06f JA |
674 | bio->bi_status = BLK_STS_IOERR; |
675 | else | |
676 | bio->bi_iter.bi_sector = rq->__sector; | |
677 | } | |
678 | ||
478eb72b PB |
679 | bio_advance(bio, nbytes); |
680 | ||
681 | if (unlikely(rq->rq_flags & RQF_QUIET)) | |
682 | bio_set_flag(bio, BIO_QUIET); | |
9be3e06f JA |
683 | /* don't actually finish bio if it's part of flush sequence */ |
684 | if (bio->bi_iter.bi_size == 0 && !(rq->rq_flags & RQF_FLUSH_SEQ)) | |
685 | bio_endio(bio); | |
686 | } | |
687 | ||
688 | static void blk_account_io_completion(struct request *req, unsigned int bytes) | |
689 | { | |
690 | if (req->part && blk_do_io_stat(req)) { | |
691 | const int sgrp = op_stat_group(req_op(req)); | |
692 | ||
693 | part_stat_lock(); | |
694 | part_stat_add(req->part, sectors[sgrp], bytes >> 9); | |
695 | part_stat_unlock(); | |
696 | } | |
697 | } | |
698 | ||
0d7a29a2 CH |
699 | static void blk_print_req_error(struct request *req, blk_status_t status) |
700 | { | |
701 | printk_ratelimited(KERN_ERR | |
702 | "%s error, dev %s, sector %llu op 0x%x:(%s) flags 0x%x " | |
703 | "phys_seg %u prio class %u\n", | |
704 | blk_status_to_str(status), | |
f3fa33ac | 705 | req->q->disk ? req->q->disk->disk_name : "?", |
0d7a29a2 CH |
706 | blk_rq_pos(req), req_op(req), blk_op_str(req_op(req)), |
707 | req->cmd_flags & ~REQ_OP_MASK, | |
708 | req->nr_phys_segments, | |
709 | IOPRIO_PRIO_CLASS(req->ioprio)); | |
710 | } | |
711 | ||
5581a5dd JA |
712 | /* |
713 | * Fully end IO on a request. Does not support partial completions, or | |
714 | * errors. | |
715 | */ | |
716 | static void blk_complete_request(struct request *req) | |
717 | { | |
718 | const bool is_flush = (req->rq_flags & RQF_FLUSH_SEQ) != 0; | |
719 | int total_bytes = blk_rq_bytes(req); | |
720 | struct bio *bio = req->bio; | |
721 | ||
722 | trace_block_rq_complete(req, BLK_STS_OK, total_bytes); | |
723 | ||
724 | if (!bio) | |
725 | return; | |
726 | ||
727 | #ifdef CONFIG_BLK_DEV_INTEGRITY | |
728 | if (blk_integrity_rq(req) && req_op(req) == REQ_OP_READ) | |
729 | req->q->integrity.profile->complete_fn(req, total_bytes); | |
730 | #endif | |
731 | ||
732 | blk_account_io_completion(req, total_bytes); | |
733 | ||
734 | do { | |
735 | struct bio *next = bio->bi_next; | |
736 | ||
737 | /* Completion has already been traced */ | |
738 | bio_clear_flag(bio, BIO_TRACE_COMPLETION); | |
739 | if (!is_flush) | |
740 | bio_endio(bio); | |
741 | bio = next; | |
742 | } while (bio); | |
743 | ||
744 | /* | |
745 | * Reset counters so that the request stacking driver | |
746 | * can find how many bytes remain in the request | |
747 | * later. | |
748 | */ | |
749 | req->bio = NULL; | |
750 | req->__data_len = 0; | |
751 | } | |
752 | ||
9be3e06f JA |
753 | /** |
754 | * blk_update_request - Complete multiple bytes without completing the request | |
755 | * @req: the request being processed | |
756 | * @error: block status code | |
757 | * @nr_bytes: number of bytes to complete for @req | |
758 | * | |
759 | * Description: | |
760 | * Ends I/O on a number of bytes attached to @req, but doesn't complete | |
761 | * the request structure even if @req doesn't have leftover. | |
762 | * If @req has leftover, sets it up for the next range of segments. | |
763 | * | |
764 | * Passing the result of blk_rq_bytes() as @nr_bytes guarantees | |
765 | * %false return from this function. | |
766 | * | |
767 | * Note: | |
768 | * The RQF_SPECIAL_PAYLOAD flag is ignored on purpose in this function | |
769 | * except in the consistency check at the end of this function. | |
770 | * | |
771 | * Return: | |
772 | * %false - this request doesn't have any more data | |
773 | * %true - this request has more data | |
774 | **/ | |
775 | bool blk_update_request(struct request *req, blk_status_t error, | |
776 | unsigned int nr_bytes) | |
777 | { | |
778 | int total_bytes; | |
779 | ||
8a7d267b | 780 | trace_block_rq_complete(req, error, nr_bytes); |
9be3e06f JA |
781 | |
782 | if (!req->bio) | |
783 | return false; | |
784 | ||
785 | #ifdef CONFIG_BLK_DEV_INTEGRITY | |
786 | if (blk_integrity_rq(req) && req_op(req) == REQ_OP_READ && | |
787 | error == BLK_STS_OK) | |
788 | req->q->integrity.profile->complete_fn(req, nr_bytes); | |
789 | #endif | |
790 | ||
791 | if (unlikely(error && !blk_rq_is_passthrough(req) && | |
d5869fdc | 792 | !(req->rq_flags & RQF_QUIET))) { |
9be3e06f | 793 | blk_print_req_error(req, error); |
d5869fdc YS |
794 | trace_block_rq_error(req, error, nr_bytes); |
795 | } | |
9be3e06f JA |
796 | |
797 | blk_account_io_completion(req, nr_bytes); | |
798 | ||
799 | total_bytes = 0; | |
800 | while (req->bio) { | |
801 | struct bio *bio = req->bio; | |
802 | unsigned bio_bytes = min(bio->bi_iter.bi_size, nr_bytes); | |
803 | ||
804 | if (bio_bytes == bio->bi_iter.bi_size) | |
805 | req->bio = bio->bi_next; | |
806 | ||
807 | /* Completion has already been traced */ | |
808 | bio_clear_flag(bio, BIO_TRACE_COMPLETION); | |
809 | req_bio_endio(req, bio, bio_bytes, error); | |
810 | ||
811 | total_bytes += bio_bytes; | |
812 | nr_bytes -= bio_bytes; | |
813 | ||
814 | if (!nr_bytes) | |
815 | break; | |
816 | } | |
817 | ||
818 | /* | |
819 | * completely done | |
820 | */ | |
821 | if (!req->bio) { | |
822 | /* | |
823 | * Reset counters so that the request stacking driver | |
824 | * can find how many bytes remain in the request | |
825 | * later. | |
826 | */ | |
827 | req->__data_len = 0; | |
828 | return false; | |
829 | } | |
830 | ||
831 | req->__data_len -= total_bytes; | |
832 | ||
833 | /* update sector only for requests with clear definition of sector */ | |
834 | if (!blk_rq_is_passthrough(req)) | |
835 | req->__sector += total_bytes >> 9; | |
836 | ||
837 | /* mixed attributes always follow the first bio */ | |
838 | if (req->rq_flags & RQF_MIXED_MERGE) { | |
839 | req->cmd_flags &= ~REQ_FAILFAST_MASK; | |
840 | req->cmd_flags |= req->bio->bi_opf & REQ_FAILFAST_MASK; | |
841 | } | |
842 | ||
843 | if (!(req->rq_flags & RQF_SPECIAL_PAYLOAD)) { | |
844 | /* | |
845 | * If total number of sectors is less than the first segment | |
846 | * size, something has gone terribly wrong. | |
847 | */ | |
848 | if (blk_rq_bytes(req) < blk_rq_cur_bytes(req)) { | |
849 | blk_dump_rq_flags(req, "request botched"); | |
850 | req->__data_len = blk_rq_cur_bytes(req); | |
851 | } | |
852 | ||
853 | /* recalculate the number of segments */ | |
854 | req->nr_phys_segments = blk_recalc_rq_segments(req); | |
855 | } | |
856 | ||
857 | return true; | |
858 | } | |
859 | EXPORT_SYMBOL_GPL(blk_update_request); | |
860 | ||
450b7879 CH |
861 | static void __blk_account_io_done(struct request *req, u64 now) |
862 | { | |
863 | const int sgrp = op_stat_group(req_op(req)); | |
864 | ||
865 | part_stat_lock(); | |
866 | update_io_ticks(req->part, jiffies, true); | |
867 | part_stat_inc(req->part, ios[sgrp]); | |
868 | part_stat_add(req->part, nsecs[sgrp], now - req->start_time_ns); | |
869 | part_stat_unlock(); | |
870 | } | |
871 | ||
872 | static inline void blk_account_io_done(struct request *req, u64 now) | |
873 | { | |
874 | /* | |
875 | * Account IO completion. flush_rq isn't accounted as a | |
876 | * normal IO on queueing nor completion. Accounting the | |
877 | * containing request is enough. | |
878 | */ | |
879 | if (blk_do_io_stat(req) && req->part && | |
880 | !(req->rq_flags & RQF_FLUSH_SEQ)) | |
881 | __blk_account_io_done(req, now); | |
882 | } | |
883 | ||
884 | static void __blk_account_io_start(struct request *rq) | |
885 | { | |
886 | /* passthrough requests can hold bios that do not have ->bi_bdev set */ | |
887 | if (rq->bio && rq->bio->bi_bdev) | |
888 | rq->part = rq->bio->bi_bdev; | |
f3fa33ac CH |
889 | else if (rq->q->disk) |
890 | rq->part = rq->q->disk->part0; | |
450b7879 CH |
891 | |
892 | part_stat_lock(); | |
893 | update_io_ticks(rq->part, jiffies, false); | |
894 | part_stat_unlock(); | |
895 | } | |
896 | ||
897 | static inline void blk_account_io_start(struct request *req) | |
898 | { | |
899 | if (blk_do_io_stat(req)) | |
900 | __blk_account_io_start(req); | |
901 | } | |
902 | ||
f794f335 | 903 | static inline void __blk_mq_end_request_acct(struct request *rq, u64 now) |
320ae51f | 904 | { |
4bc6339a OS |
905 | if (rq->rq_flags & RQF_STATS) { |
906 | blk_mq_poll_stats_start(rq->q); | |
522a7775 | 907 | blk_stat_add(rq, now); |
4bc6339a OS |
908 | } |
909 | ||
87890092 | 910 | blk_mq_sched_completed_request(rq, now); |
522a7775 | 911 | blk_account_io_done(rq, now); |
f794f335 | 912 | } |
522a7775 | 913 | |
f794f335 JA |
914 | inline void __blk_mq_end_request(struct request *rq, blk_status_t error) |
915 | { | |
916 | if (blk_mq_need_time_stamp(rq)) | |
917 | __blk_mq_end_request_acct(rq, ktime_get_ns()); | |
0d11e6ac | 918 | |
91b63639 | 919 | if (rq->end_io) { |
a7905043 | 920 | rq_qos_done(rq->q, rq); |
320ae51f | 921 | rq->end_io(rq, error); |
91b63639 | 922 | } else { |
320ae51f | 923 | blk_mq_free_request(rq); |
91b63639 | 924 | } |
320ae51f | 925 | } |
c8a446ad | 926 | EXPORT_SYMBOL(__blk_mq_end_request); |
63151a44 | 927 | |
2a842aca | 928 | void blk_mq_end_request(struct request *rq, blk_status_t error) |
63151a44 CH |
929 | { |
930 | if (blk_update_request(rq, error, blk_rq_bytes(rq))) | |
931 | BUG(); | |
c8a446ad | 932 | __blk_mq_end_request(rq, error); |
63151a44 | 933 | } |
c8a446ad | 934 | EXPORT_SYMBOL(blk_mq_end_request); |
320ae51f | 935 | |
f794f335 JA |
936 | #define TAG_COMP_BATCH 32 |
937 | ||
938 | static inline void blk_mq_flush_tag_batch(struct blk_mq_hw_ctx *hctx, | |
939 | int *tag_array, int nr_tags) | |
940 | { | |
941 | struct request_queue *q = hctx->queue; | |
942 | ||
3b87c6ea ML |
943 | /* |
944 | * All requests should have been marked as RQF_MQ_INFLIGHT, so | |
945 | * update hctx->nr_active in batch | |
946 | */ | |
947 | if (hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) | |
948 | __blk_mq_sub_active_requests(hctx, nr_tags); | |
949 | ||
f794f335 JA |
950 | blk_mq_put_tags(hctx->tags, tag_array, nr_tags); |
951 | percpu_ref_put_many(&q->q_usage_counter, nr_tags); | |
952 | } | |
953 | ||
954 | void blk_mq_end_request_batch(struct io_comp_batch *iob) | |
955 | { | |
956 | int tags[TAG_COMP_BATCH], nr_tags = 0; | |
02f7eab0 | 957 | struct blk_mq_hw_ctx *cur_hctx = NULL; |
f794f335 JA |
958 | struct request *rq; |
959 | u64 now = 0; | |
960 | ||
961 | if (iob->need_ts) | |
962 | now = ktime_get_ns(); | |
963 | ||
964 | while ((rq = rq_list_pop(&iob->req_list)) != NULL) { | |
965 | prefetch(rq->bio); | |
966 | prefetch(rq->rq_next); | |
967 | ||
5581a5dd | 968 | blk_complete_request(rq); |
f794f335 JA |
969 | if (iob->need_ts) |
970 | __blk_mq_end_request_acct(rq, now); | |
971 | ||
98b26a0e JA |
972 | rq_qos_done(rq->q, rq); |
973 | ||
f794f335 | 974 | WRITE_ONCE(rq->state, MQ_RQ_IDLE); |
0a467d0f | 975 | if (!req_ref_put_and_test(rq)) |
f794f335 JA |
976 | continue; |
977 | ||
978 | blk_crypto_free_request(rq); | |
979 | blk_pm_mark_last_busy(rq); | |
f794f335 | 980 | |
02f7eab0 JA |
981 | if (nr_tags == TAG_COMP_BATCH || cur_hctx != rq->mq_hctx) { |
982 | if (cur_hctx) | |
983 | blk_mq_flush_tag_batch(cur_hctx, tags, nr_tags); | |
f794f335 | 984 | nr_tags = 0; |
02f7eab0 | 985 | cur_hctx = rq->mq_hctx; |
f794f335 JA |
986 | } |
987 | tags[nr_tags++] = rq->tag; | |
f794f335 JA |
988 | } |
989 | ||
990 | if (nr_tags) | |
02f7eab0 | 991 | blk_mq_flush_tag_batch(cur_hctx, tags, nr_tags); |
f794f335 JA |
992 | } |
993 | EXPORT_SYMBOL_GPL(blk_mq_end_request_batch); | |
994 | ||
f9ab4918 | 995 | static void blk_complete_reqs(struct llist_head *list) |
320ae51f | 996 | { |
f9ab4918 SAS |
997 | struct llist_node *entry = llist_reverse_order(llist_del_all(list)); |
998 | struct request *rq, *next; | |
c3077b5d | 999 | |
f9ab4918 | 1000 | llist_for_each_entry_safe(rq, next, entry, ipi_list) |
c3077b5d | 1001 | rq->q->mq_ops->complete(rq); |
320ae51f | 1002 | } |
320ae51f | 1003 | |
f9ab4918 | 1004 | static __latent_entropy void blk_done_softirq(struct softirq_action *h) |
320ae51f | 1005 | { |
f9ab4918 | 1006 | blk_complete_reqs(this_cpu_ptr(&blk_cpu_done)); |
115243f5 CH |
1007 | } |
1008 | ||
c3077b5d CH |
1009 | static int blk_softirq_cpu_dead(unsigned int cpu) |
1010 | { | |
f9ab4918 | 1011 | blk_complete_reqs(&per_cpu(blk_cpu_done, cpu)); |
c3077b5d CH |
1012 | return 0; |
1013 | } | |
1014 | ||
40d09b53 | 1015 | static void __blk_mq_complete_request_remote(void *data) |
c3077b5d | 1016 | { |
f9ab4918 | 1017 | __raise_softirq_irqoff(BLOCK_SOFTIRQ); |
c3077b5d CH |
1018 | } |
1019 | ||
96339526 CH |
1020 | static inline bool blk_mq_complete_need_ipi(struct request *rq) |
1021 | { | |
1022 | int cpu = raw_smp_processor_id(); | |
1023 | ||
1024 | if (!IS_ENABLED(CONFIG_SMP) || | |
1025 | !test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) | |
1026 | return false; | |
71425189 SAS |
1027 | /* |
1028 | * With force threaded interrupts enabled, raising softirq from an SMP | |
1029 | * function call will always result in waking the ksoftirqd thread. | |
1030 | * This is probably worse than completing the request on a different | |
1031 | * cache domain. | |
1032 | */ | |
91cc470e | 1033 | if (force_irqthreads()) |
71425189 | 1034 | return false; |
96339526 CH |
1035 | |
1036 | /* same CPU or cache domain? Complete locally */ | |
1037 | if (cpu == rq->mq_ctx->cpu || | |
1038 | (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags) && | |
1039 | cpus_share_cache(cpu, rq->mq_ctx->cpu))) | |
1040 | return false; | |
1041 | ||
1042 | /* don't try to IPI to an offline CPU */ | |
1043 | return cpu_online(rq->mq_ctx->cpu); | |
1044 | } | |
1045 | ||
f9ab4918 SAS |
1046 | static void blk_mq_complete_send_ipi(struct request *rq) |
1047 | { | |
1048 | struct llist_head *list; | |
1049 | unsigned int cpu; | |
1050 | ||
1051 | cpu = rq->mq_ctx->cpu; | |
1052 | list = &per_cpu(blk_cpu_done, cpu); | |
1053 | if (llist_add(&rq->ipi_list, list)) { | |
1054 | INIT_CSD(&rq->csd, __blk_mq_complete_request_remote, rq); | |
1055 | smp_call_function_single_async(cpu, &rq->csd); | |
1056 | } | |
1057 | } | |
1058 | ||
1059 | static void blk_mq_raise_softirq(struct request *rq) | |
1060 | { | |
1061 | struct llist_head *list; | |
1062 | ||
1063 | preempt_disable(); | |
1064 | list = this_cpu_ptr(&blk_cpu_done); | |
1065 | if (llist_add(&rq->ipi_list, list)) | |
1066 | raise_softirq(BLOCK_SOFTIRQ); | |
1067 | preempt_enable(); | |
1068 | } | |
1069 | ||
40d09b53 | 1070 | bool blk_mq_complete_request_remote(struct request *rq) |
320ae51f | 1071 | { |
af78ff7c | 1072 | WRITE_ONCE(rq->state, MQ_RQ_COMPLETE); |
36e76539 | 1073 | |
4ab32bf3 JA |
1074 | /* |
1075 | * For a polled request, always complete locallly, it's pointless | |
1076 | * to redirect the completion. | |
1077 | */ | |
6ce913fe | 1078 | if (rq->cmd_flags & REQ_POLLED) |
40d09b53 | 1079 | return false; |
38535201 | 1080 | |
96339526 | 1081 | if (blk_mq_complete_need_ipi(rq)) { |
f9ab4918 SAS |
1082 | blk_mq_complete_send_ipi(rq); |
1083 | return true; | |
3d6efbf6 | 1084 | } |
40d09b53 | 1085 | |
f9ab4918 SAS |
1086 | if (rq->q->nr_hw_queues == 1) { |
1087 | blk_mq_raise_softirq(rq); | |
1088 | return true; | |
1089 | } | |
1090 | return false; | |
40d09b53 CH |
1091 | } |
1092 | EXPORT_SYMBOL_GPL(blk_mq_complete_request_remote); | |
1093 | ||
1094 | /** | |
1095 | * blk_mq_complete_request - end I/O on a request | |
1096 | * @rq: the request being processed | |
1097 | * | |
1098 | * Description: | |
1099 | * Complete a request by scheduling the ->complete_rq operation. | |
1100 | **/ | |
1101 | void blk_mq_complete_request(struct request *rq) | |
1102 | { | |
1103 | if (!blk_mq_complete_request_remote(rq)) | |
1104 | rq->q->mq_ops->complete(rq); | |
320ae51f | 1105 | } |
15f73f5b | 1106 | EXPORT_SYMBOL(blk_mq_complete_request); |
30a91cb4 | 1107 | |
105663f7 AA |
1108 | /** |
1109 | * blk_mq_start_request - Start processing a request | |
1110 | * @rq: Pointer to request to be started | |
1111 | * | |
1112 | * Function used by device drivers to notify the block layer that a request | |
1113 | * is going to be processed now, so blk layer can do proper initializations | |
1114 | * such as starting the timeout timer. | |
1115 | */ | |
e2490073 | 1116 | void blk_mq_start_request(struct request *rq) |
320ae51f JA |
1117 | { |
1118 | struct request_queue *q = rq->q; | |
1119 | ||
a54895fa | 1120 | trace_block_rq_issue(rq); |
320ae51f | 1121 | |
cf43e6be | 1122 | if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) { |
00067077 JA |
1123 | u64 start_time; |
1124 | #ifdef CONFIG_BLK_CGROUP | |
1125 | if (rq->bio) | |
1126 | start_time = bio_issue_time(&rq->bio->bi_issue); | |
1127 | else | |
1128 | #endif | |
1129 | start_time = ktime_get_ns(); | |
1130 | rq->io_start_time_ns = start_time; | |
3d244306 | 1131 | rq->stats_sectors = blk_rq_sectors(rq); |
cf43e6be | 1132 | rq->rq_flags |= RQF_STATS; |
a7905043 | 1133 | rq_qos_issue(q, rq); |
cf43e6be JA |
1134 | } |
1135 | ||
1d9bd516 | 1136 | WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE); |
538b7534 | 1137 | |
1d9bd516 | 1138 | blk_add_timer(rq); |
12f5b931 | 1139 | WRITE_ONCE(rq->state, MQ_RQ_IN_FLIGHT); |
49f5baa5 | 1140 | |
54d4e6ab MG |
1141 | #ifdef CONFIG_BLK_DEV_INTEGRITY |
1142 | if (blk_integrity_rq(rq) && req_op(rq) == REQ_OP_WRITE) | |
1143 | q->integrity.profile->prepare_fn(rq); | |
1144 | #endif | |
3e08773c CH |
1145 | if (rq->bio && rq->bio->bi_opf & REQ_POLLED) |
1146 | WRITE_ONCE(rq->bio->bi_cookie, blk_rq_to_qc(rq)); | |
320ae51f | 1147 | } |
e2490073 | 1148 | EXPORT_SYMBOL(blk_mq_start_request); |
320ae51f | 1149 | |
4054cff9 CH |
1150 | /** |
1151 | * blk_end_sync_rq - executes a completion event on a request | |
1152 | * @rq: request to complete | |
1153 | * @error: end I/O status of the request | |
1154 | */ | |
1155 | static void blk_end_sync_rq(struct request *rq, blk_status_t error) | |
1156 | { | |
1157 | struct completion *waiting = rq->end_io_data; | |
1158 | ||
1159 | rq->end_io_data = (void *)(uintptr_t)error; | |
1160 | ||
1161 | /* | |
1162 | * complete last, if this is a stack request the process (and thus | |
1163 | * the rq pointer) could be invalid right after this complete() | |
1164 | */ | |
1165 | complete(waiting); | |
1166 | } | |
1167 | ||
1168 | /** | |
1169 | * blk_execute_rq_nowait - insert a request to I/O scheduler for execution | |
4054cff9 CH |
1170 | * @rq: request to insert |
1171 | * @at_head: insert request at head or tail of queue | |
1172 | * @done: I/O completion handler | |
1173 | * | |
1174 | * Description: | |
1175 | * Insert a fully prepared request at the back of the I/O scheduler queue | |
1176 | * for execution. Don't wait for completion. | |
1177 | * | |
1178 | * Note: | |
1179 | * This function will invoke @done directly if the queue is dead. | |
1180 | */ | |
b84ba30b | 1181 | void blk_execute_rq_nowait(struct request *rq, bool at_head, rq_end_io_fn *done) |
4054cff9 CH |
1182 | { |
1183 | WARN_ON(irqs_disabled()); | |
1184 | WARN_ON(!blk_rq_is_passthrough(rq)); | |
1185 | ||
4054cff9 CH |
1186 | rq->end_io = done; |
1187 | ||
1188 | blk_account_io_start(rq); | |
1189 | ||
1190 | /* | |
1191 | * don't check dying flag for MQ because the request won't | |
1192 | * be reused after dying flag is set | |
1193 | */ | |
1194 | blk_mq_sched_insert_request(rq, at_head, true, false); | |
1195 | } | |
1196 | EXPORT_SYMBOL_GPL(blk_execute_rq_nowait); | |
1197 | ||
1198 | static bool blk_rq_is_poll(struct request *rq) | |
1199 | { | |
1200 | if (!rq->mq_hctx) | |
1201 | return false; | |
1202 | if (rq->mq_hctx->type != HCTX_TYPE_POLL) | |
1203 | return false; | |
1204 | if (WARN_ON_ONCE(!rq->bio)) | |
1205 | return false; | |
1206 | return true; | |
1207 | } | |
1208 | ||
1209 | static void blk_rq_poll_completion(struct request *rq, struct completion *wait) | |
1210 | { | |
1211 | do { | |
1212 | bio_poll(rq->bio, NULL, 0); | |
1213 | cond_resched(); | |
1214 | } while (!completion_done(wait)); | |
1215 | } | |
1216 | ||
1217 | /** | |
1218 | * blk_execute_rq - insert a request into queue for execution | |
4054cff9 CH |
1219 | * @rq: request to insert |
1220 | * @at_head: insert request at head or tail of queue | |
1221 | * | |
1222 | * Description: | |
1223 | * Insert a fully prepared request at the back of the I/O scheduler queue | |
1224 | * for execution and wait for completion. | |
1225 | * Return: The blk_status_t result provided to blk_mq_end_request(). | |
1226 | */ | |
b84ba30b | 1227 | blk_status_t blk_execute_rq(struct request *rq, bool at_head) |
4054cff9 CH |
1228 | { |
1229 | DECLARE_COMPLETION_ONSTACK(wait); | |
1230 | unsigned long hang_check; | |
1231 | ||
1232 | rq->end_io_data = &wait; | |
b84ba30b | 1233 | blk_execute_rq_nowait(rq, at_head, blk_end_sync_rq); |
4054cff9 CH |
1234 | |
1235 | /* Prevent hang_check timer from firing at us during very long I/O */ | |
1236 | hang_check = sysctl_hung_task_timeout_secs; | |
1237 | ||
1238 | if (blk_rq_is_poll(rq)) | |
1239 | blk_rq_poll_completion(rq, &wait); | |
1240 | else if (hang_check) | |
1241 | while (!wait_for_completion_io_timeout(&wait, | |
1242 | hang_check * (HZ/2))) | |
1243 | ; | |
1244 | else | |
1245 | wait_for_completion_io(&wait); | |
1246 | ||
1247 | return (blk_status_t)(uintptr_t)rq->end_io_data; | |
1248 | } | |
1249 | EXPORT_SYMBOL(blk_execute_rq); | |
1250 | ||
ed0791b2 | 1251 | static void __blk_mq_requeue_request(struct request *rq) |
320ae51f JA |
1252 | { |
1253 | struct request_queue *q = rq->q; | |
1254 | ||
923218f6 ML |
1255 | blk_mq_put_driver_tag(rq); |
1256 | ||
a54895fa | 1257 | trace_block_rq_requeue(rq); |
a7905043 | 1258 | rq_qos_requeue(q, rq); |
49f5baa5 | 1259 | |
12f5b931 KB |
1260 | if (blk_mq_request_started(rq)) { |
1261 | WRITE_ONCE(rq->state, MQ_RQ_IDLE); | |
da661267 | 1262 | rq->rq_flags &= ~RQF_TIMED_OUT; |
e2490073 | 1263 | } |
320ae51f JA |
1264 | } |
1265 | ||
2b053aca | 1266 | void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list) |
ed0791b2 | 1267 | { |
ed0791b2 | 1268 | __blk_mq_requeue_request(rq); |
ed0791b2 | 1269 | |
105976f5 ML |
1270 | /* this request will be re-inserted to io scheduler queue */ |
1271 | blk_mq_sched_requeue_request(rq); | |
1272 | ||
2b053aca | 1273 | blk_mq_add_to_requeue_list(rq, true, kick_requeue_list); |
ed0791b2 CH |
1274 | } |
1275 | EXPORT_SYMBOL(blk_mq_requeue_request); | |
1276 | ||
6fca6a61 CH |
1277 | static void blk_mq_requeue_work(struct work_struct *work) |
1278 | { | |
1279 | struct request_queue *q = | |
2849450a | 1280 | container_of(work, struct request_queue, requeue_work.work); |
6fca6a61 CH |
1281 | LIST_HEAD(rq_list); |
1282 | struct request *rq, *next; | |
6fca6a61 | 1283 | |
18e9781d | 1284 | spin_lock_irq(&q->requeue_lock); |
6fca6a61 | 1285 | list_splice_init(&q->requeue_list, &rq_list); |
18e9781d | 1286 | spin_unlock_irq(&q->requeue_lock); |
6fca6a61 CH |
1287 | |
1288 | list_for_each_entry_safe(rq, next, &rq_list, queuelist) { | |
aef1897c | 1289 | if (!(rq->rq_flags & (RQF_SOFTBARRIER | RQF_DONTPREP))) |
6fca6a61 CH |
1290 | continue; |
1291 | ||
e8064021 | 1292 | rq->rq_flags &= ~RQF_SOFTBARRIER; |
6fca6a61 | 1293 | list_del_init(&rq->queuelist); |
aef1897c JW |
1294 | /* |
1295 | * If RQF_DONTPREP, rq has contained some driver specific | |
1296 | * data, so insert it to hctx dispatch list to avoid any | |
1297 | * merge. | |
1298 | */ | |
1299 | if (rq->rq_flags & RQF_DONTPREP) | |
01e99aec | 1300 | blk_mq_request_bypass_insert(rq, false, false); |
aef1897c JW |
1301 | else |
1302 | blk_mq_sched_insert_request(rq, true, false, false); | |
6fca6a61 CH |
1303 | } |
1304 | ||
1305 | while (!list_empty(&rq_list)) { | |
1306 | rq = list_entry(rq_list.next, struct request, queuelist); | |
1307 | list_del_init(&rq->queuelist); | |
9e97d295 | 1308 | blk_mq_sched_insert_request(rq, false, false, false); |
6fca6a61 CH |
1309 | } |
1310 | ||
52d7f1b5 | 1311 | blk_mq_run_hw_queues(q, false); |
6fca6a61 CH |
1312 | } |
1313 | ||
2b053aca BVA |
1314 | void blk_mq_add_to_requeue_list(struct request *rq, bool at_head, |
1315 | bool kick_requeue_list) | |
6fca6a61 CH |
1316 | { |
1317 | struct request_queue *q = rq->q; | |
1318 | unsigned long flags; | |
1319 | ||
1320 | /* | |
1321 | * We abuse this flag that is otherwise used by the I/O scheduler to | |
ff821d27 | 1322 | * request head insertion from the workqueue. |
6fca6a61 | 1323 | */ |
e8064021 | 1324 | BUG_ON(rq->rq_flags & RQF_SOFTBARRIER); |
6fca6a61 CH |
1325 | |
1326 | spin_lock_irqsave(&q->requeue_lock, flags); | |
1327 | if (at_head) { | |
e8064021 | 1328 | rq->rq_flags |= RQF_SOFTBARRIER; |
6fca6a61 CH |
1329 | list_add(&rq->queuelist, &q->requeue_list); |
1330 | } else { | |
1331 | list_add_tail(&rq->queuelist, &q->requeue_list); | |
1332 | } | |
1333 | spin_unlock_irqrestore(&q->requeue_lock, flags); | |
2b053aca BVA |
1334 | |
1335 | if (kick_requeue_list) | |
1336 | blk_mq_kick_requeue_list(q); | |
6fca6a61 | 1337 | } |
6fca6a61 CH |
1338 | |
1339 | void blk_mq_kick_requeue_list(struct request_queue *q) | |
1340 | { | |
ae943d20 | 1341 | kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0); |
6fca6a61 CH |
1342 | } |
1343 | EXPORT_SYMBOL(blk_mq_kick_requeue_list); | |
1344 | ||
2849450a MS |
1345 | void blk_mq_delay_kick_requeue_list(struct request_queue *q, |
1346 | unsigned long msecs) | |
1347 | { | |
d4acf365 BVA |
1348 | kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, |
1349 | msecs_to_jiffies(msecs)); | |
2849450a MS |
1350 | } |
1351 | EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list); | |
1352 | ||
8ab30a33 JG |
1353 | static bool blk_mq_rq_inflight(struct request *rq, void *priv, |
1354 | bool reserved) | |
ae879912 JA |
1355 | { |
1356 | /* | |
8ab30a33 JG |
1357 | * If we find a request that isn't idle we know the queue is busy |
1358 | * as it's checked in the iter. | |
1359 | * Return false to stop the iteration. | |
ae879912 | 1360 | */ |
8ab30a33 | 1361 | if (blk_mq_request_started(rq)) { |
ae879912 JA |
1362 | bool *busy = priv; |
1363 | ||
1364 | *busy = true; | |
1365 | return false; | |
1366 | } | |
1367 | ||
1368 | return true; | |
1369 | } | |
1370 | ||
3c94d83c | 1371 | bool blk_mq_queue_inflight(struct request_queue *q) |
ae879912 JA |
1372 | { |
1373 | bool busy = false; | |
1374 | ||
3c94d83c | 1375 | blk_mq_queue_tag_busy_iter(q, blk_mq_rq_inflight, &busy); |
ae879912 JA |
1376 | return busy; |
1377 | } | |
3c94d83c | 1378 | EXPORT_SYMBOL_GPL(blk_mq_queue_inflight); |
ae879912 | 1379 | |
358f70da | 1380 | static void blk_mq_rq_timed_out(struct request *req, bool reserved) |
320ae51f | 1381 | { |
da661267 | 1382 | req->rq_flags |= RQF_TIMED_OUT; |
d1210d5a CH |
1383 | if (req->q->mq_ops->timeout) { |
1384 | enum blk_eh_timer_return ret; | |
1385 | ||
1386 | ret = req->q->mq_ops->timeout(req, reserved); | |
1387 | if (ret == BLK_EH_DONE) | |
1388 | return; | |
1389 | WARN_ON_ONCE(ret != BLK_EH_RESET_TIMER); | |
46f92d42 | 1390 | } |
d1210d5a CH |
1391 | |
1392 | blk_add_timer(req); | |
87ee7b11 | 1393 | } |
5b3f25fc | 1394 | |
12f5b931 | 1395 | static bool blk_mq_req_expired(struct request *rq, unsigned long *next) |
81481eb4 | 1396 | { |
12f5b931 | 1397 | unsigned long deadline; |
87ee7b11 | 1398 | |
12f5b931 KB |
1399 | if (blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT) |
1400 | return false; | |
da661267 CH |
1401 | if (rq->rq_flags & RQF_TIMED_OUT) |
1402 | return false; | |
a7af0af3 | 1403 | |
079076b3 | 1404 | deadline = READ_ONCE(rq->deadline); |
12f5b931 KB |
1405 | if (time_after_eq(jiffies, deadline)) |
1406 | return true; | |
a7af0af3 | 1407 | |
12f5b931 KB |
1408 | if (*next == 0) |
1409 | *next = deadline; | |
1410 | else if (time_after(*next, deadline)) | |
1411 | *next = deadline; | |
1412 | return false; | |
87ee7b11 JA |
1413 | } |
1414 | ||
2e315dc0 ML |
1415 | void blk_mq_put_rq_ref(struct request *rq) |
1416 | { | |
a9ed27a7 | 1417 | if (is_flush_rq(rq)) |
2e315dc0 | 1418 | rq->end_io(rq, 0); |
0a467d0f | 1419 | else if (req_ref_put_and_test(rq)) |
2e315dc0 ML |
1420 | __blk_mq_free_request(rq); |
1421 | } | |
1422 | ||
8ab30a33 | 1423 | static bool blk_mq_check_expired(struct request *rq, void *priv, bool reserved) |
1d9bd516 | 1424 | { |
12f5b931 KB |
1425 | unsigned long *next = priv; |
1426 | ||
1427 | /* | |
c797b40c ML |
1428 | * blk_mq_queue_tag_busy_iter() has locked the request, so it cannot |
1429 | * be reallocated underneath the timeout handler's processing, then | |
1430 | * the expire check is reliable. If the request is not expired, then | |
1431 | * it was completed and reallocated as a new request after returning | |
1432 | * from blk_mq_check_expired(). | |
1d9bd516 | 1433 | */ |
12f5b931 | 1434 | if (blk_mq_req_expired(rq, next)) |
1d9bd516 | 1435 | blk_mq_rq_timed_out(rq, reserved); |
7baa8572 | 1436 | return true; |
1d9bd516 TH |
1437 | } |
1438 | ||
287922eb | 1439 | static void blk_mq_timeout_work(struct work_struct *work) |
320ae51f | 1440 | { |
287922eb CH |
1441 | struct request_queue *q = |
1442 | container_of(work, struct request_queue, timeout_work); | |
12f5b931 | 1443 | unsigned long next = 0; |
1d9bd516 | 1444 | struct blk_mq_hw_ctx *hctx; |
81481eb4 | 1445 | int i; |
320ae51f | 1446 | |
71f79fb3 GKB |
1447 | /* A deadlock might occur if a request is stuck requiring a |
1448 | * timeout at the same time a queue freeze is waiting | |
1449 | * completion, since the timeout code would not be able to | |
1450 | * acquire the queue reference here. | |
1451 | * | |
1452 | * That's why we don't use blk_queue_enter here; instead, we use | |
1453 | * percpu_ref_tryget directly, because we need to be able to | |
1454 | * obtain a reference even in the short window between the queue | |
1455 | * starting to freeze, by dropping the first reference in | |
1671d522 | 1456 | * blk_freeze_queue_start, and the moment the last request is |
71f79fb3 GKB |
1457 | * consumed, marked by the instant q_usage_counter reaches |
1458 | * zero. | |
1459 | */ | |
1460 | if (!percpu_ref_tryget(&q->q_usage_counter)) | |
287922eb CH |
1461 | return; |
1462 | ||
12f5b931 | 1463 | blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &next); |
320ae51f | 1464 | |
12f5b931 KB |
1465 | if (next != 0) { |
1466 | mod_timer(&q->timeout, next); | |
0d2602ca | 1467 | } else { |
fcd36c36 BVA |
1468 | /* |
1469 | * Request timeouts are handled as a forward rolling timer. If | |
1470 | * we end up here it means that no requests are pending and | |
1471 | * also that no request has been pending for a while. Mark | |
1472 | * each hctx as idle. | |
1473 | */ | |
f054b56c ML |
1474 | queue_for_each_hw_ctx(q, hctx, i) { |
1475 | /* the hctx may be unmapped, so check it here */ | |
1476 | if (blk_mq_hw_queue_mapped(hctx)) | |
1477 | blk_mq_tag_idle(hctx); | |
1478 | } | |
0d2602ca | 1479 | } |
287922eb | 1480 | blk_queue_exit(q); |
320ae51f JA |
1481 | } |
1482 | ||
88459642 OS |
1483 | struct flush_busy_ctx_data { |
1484 | struct blk_mq_hw_ctx *hctx; | |
1485 | struct list_head *list; | |
1486 | }; | |
1487 | ||
1488 | static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data) | |
1489 | { | |
1490 | struct flush_busy_ctx_data *flush_data = data; | |
1491 | struct blk_mq_hw_ctx *hctx = flush_data->hctx; | |
1492 | struct blk_mq_ctx *ctx = hctx->ctxs[bitnr]; | |
c16d6b5a | 1493 | enum hctx_type type = hctx->type; |
88459642 | 1494 | |
88459642 | 1495 | spin_lock(&ctx->lock); |
c16d6b5a | 1496 | list_splice_tail_init(&ctx->rq_lists[type], flush_data->list); |
e9a99a63 | 1497 | sbitmap_clear_bit(sb, bitnr); |
88459642 OS |
1498 | spin_unlock(&ctx->lock); |
1499 | return true; | |
1500 | } | |
1501 | ||
1429d7c9 JA |
1502 | /* |
1503 | * Process software queues that have been marked busy, splicing them | |
1504 | * to the for-dispatch | |
1505 | */ | |
2c3ad667 | 1506 | void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list) |
1429d7c9 | 1507 | { |
88459642 OS |
1508 | struct flush_busy_ctx_data data = { |
1509 | .hctx = hctx, | |
1510 | .list = list, | |
1511 | }; | |
1429d7c9 | 1512 | |
88459642 | 1513 | sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data); |
1429d7c9 | 1514 | } |
2c3ad667 | 1515 | EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs); |
1429d7c9 | 1516 | |
b347689f ML |
1517 | struct dispatch_rq_data { |
1518 | struct blk_mq_hw_ctx *hctx; | |
1519 | struct request *rq; | |
1520 | }; | |
1521 | ||
1522 | static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr, | |
1523 | void *data) | |
1524 | { | |
1525 | struct dispatch_rq_data *dispatch_data = data; | |
1526 | struct blk_mq_hw_ctx *hctx = dispatch_data->hctx; | |
1527 | struct blk_mq_ctx *ctx = hctx->ctxs[bitnr]; | |
c16d6b5a | 1528 | enum hctx_type type = hctx->type; |
b347689f ML |
1529 | |
1530 | spin_lock(&ctx->lock); | |
c16d6b5a ML |
1531 | if (!list_empty(&ctx->rq_lists[type])) { |
1532 | dispatch_data->rq = list_entry_rq(ctx->rq_lists[type].next); | |
b347689f | 1533 | list_del_init(&dispatch_data->rq->queuelist); |
c16d6b5a | 1534 | if (list_empty(&ctx->rq_lists[type])) |
b347689f ML |
1535 | sbitmap_clear_bit(sb, bitnr); |
1536 | } | |
1537 | spin_unlock(&ctx->lock); | |
1538 | ||
1539 | return !dispatch_data->rq; | |
1540 | } | |
1541 | ||
1542 | struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx, | |
1543 | struct blk_mq_ctx *start) | |
1544 | { | |
f31967f0 | 1545 | unsigned off = start ? start->index_hw[hctx->type] : 0; |
b347689f ML |
1546 | struct dispatch_rq_data data = { |
1547 | .hctx = hctx, | |
1548 | .rq = NULL, | |
1549 | }; | |
1550 | ||
1551 | __sbitmap_for_each_set(&hctx->ctx_map, off, | |
1552 | dispatch_rq_from_ctx, &data); | |
1553 | ||
1554 | return data.rq; | |
1555 | } | |
1556 | ||
a808a9d5 | 1557 | static bool __blk_mq_alloc_driver_tag(struct request *rq) |
570e9b73 | 1558 | { |
ae0f1a73 | 1559 | struct sbitmap_queue *bt = &rq->mq_hctx->tags->bitmap_tags; |
570e9b73 | 1560 | unsigned int tag_offset = rq->mq_hctx->tags->nr_reserved_tags; |
570e9b73 ML |
1561 | int tag; |
1562 | ||
568f2700 ML |
1563 | blk_mq_tag_busy(rq->mq_hctx); |
1564 | ||
570e9b73 | 1565 | if (blk_mq_tag_is_reserved(rq->mq_hctx->sched_tags, rq->internal_tag)) { |
ae0f1a73 | 1566 | bt = &rq->mq_hctx->tags->breserved_tags; |
570e9b73 | 1567 | tag_offset = 0; |
28500850 ML |
1568 | } else { |
1569 | if (!hctx_may_queue(rq->mq_hctx, bt)) | |
1570 | return false; | |
570e9b73 ML |
1571 | } |
1572 | ||
570e9b73 ML |
1573 | tag = __sbitmap_queue_get(bt); |
1574 | if (tag == BLK_MQ_NO_TAG) | |
1575 | return false; | |
1576 | ||
1577 | rq->tag = tag + tag_offset; | |
570e9b73 ML |
1578 | return true; |
1579 | } | |
1580 | ||
a808a9d5 | 1581 | bool __blk_mq_get_driver_tag(struct blk_mq_hw_ctx *hctx, struct request *rq) |
570e9b73 | 1582 | { |
a808a9d5 | 1583 | if (rq->tag == BLK_MQ_NO_TAG && !__blk_mq_alloc_driver_tag(rq)) |
568f2700 ML |
1584 | return false; |
1585 | ||
51db1c37 | 1586 | if ((hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) && |
568f2700 ML |
1587 | !(rq->rq_flags & RQF_MQ_INFLIGHT)) { |
1588 | rq->rq_flags |= RQF_MQ_INFLIGHT; | |
bccf5e26 | 1589 | __blk_mq_inc_active_requests(hctx); |
568f2700 ML |
1590 | } |
1591 | hctx->tags->rqs[rq->tag] = rq; | |
1592 | return true; | |
570e9b73 ML |
1593 | } |
1594 | ||
eb619fdb JA |
1595 | static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode, |
1596 | int flags, void *key) | |
da55f2cc OS |
1597 | { |
1598 | struct blk_mq_hw_ctx *hctx; | |
1599 | ||
1600 | hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait); | |
1601 | ||
5815839b | 1602 | spin_lock(&hctx->dispatch_wait_lock); |
e8618575 JA |
1603 | if (!list_empty(&wait->entry)) { |
1604 | struct sbitmap_queue *sbq; | |
1605 | ||
1606 | list_del_init(&wait->entry); | |
ae0f1a73 | 1607 | sbq = &hctx->tags->bitmap_tags; |
e8618575 JA |
1608 | atomic_dec(&sbq->ws_active); |
1609 | } | |
5815839b ML |
1610 | spin_unlock(&hctx->dispatch_wait_lock); |
1611 | ||
da55f2cc OS |
1612 | blk_mq_run_hw_queue(hctx, true); |
1613 | return 1; | |
1614 | } | |
1615 | ||
f906a6a0 JA |
1616 | /* |
1617 | * Mark us waiting for a tag. For shared tags, this involves hooking us into | |
ee3e4de5 BVA |
1618 | * the tag wakeups. For non-shared tags, we can simply mark us needing a |
1619 | * restart. For both cases, take care to check the condition again after | |
f906a6a0 JA |
1620 | * marking us as waiting. |
1621 | */ | |
2278d69f | 1622 | static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx *hctx, |
f906a6a0 | 1623 | struct request *rq) |
da55f2cc | 1624 | { |
ae0f1a73 | 1625 | struct sbitmap_queue *sbq = &hctx->tags->bitmap_tags; |
5815839b | 1626 | struct wait_queue_head *wq; |
f906a6a0 JA |
1627 | wait_queue_entry_t *wait; |
1628 | bool ret; | |
da55f2cc | 1629 | |
51db1c37 | 1630 | if (!(hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) { |
684b7324 | 1631 | blk_mq_sched_mark_restart_hctx(hctx); |
f906a6a0 | 1632 | |
c27d53fb BVA |
1633 | /* |
1634 | * It's possible that a tag was freed in the window between the | |
1635 | * allocation failure and adding the hardware queue to the wait | |
1636 | * queue. | |
1637 | * | |
1638 | * Don't clear RESTART here, someone else could have set it. | |
1639 | * At most this will cost an extra queue run. | |
1640 | */ | |
8ab6bb9e | 1641 | return blk_mq_get_driver_tag(rq); |
eb619fdb | 1642 | } |
eb619fdb | 1643 | |
2278d69f | 1644 | wait = &hctx->dispatch_wait; |
c27d53fb BVA |
1645 | if (!list_empty_careful(&wait->entry)) |
1646 | return false; | |
1647 | ||
e8618575 | 1648 | wq = &bt_wait_ptr(sbq, hctx)->wait; |
5815839b ML |
1649 | |
1650 | spin_lock_irq(&wq->lock); | |
1651 | spin_lock(&hctx->dispatch_wait_lock); | |
c27d53fb | 1652 | if (!list_empty(&wait->entry)) { |
5815839b ML |
1653 | spin_unlock(&hctx->dispatch_wait_lock); |
1654 | spin_unlock_irq(&wq->lock); | |
c27d53fb | 1655 | return false; |
eb619fdb JA |
1656 | } |
1657 | ||
e8618575 | 1658 | atomic_inc(&sbq->ws_active); |
5815839b ML |
1659 | wait->flags &= ~WQ_FLAG_EXCLUSIVE; |
1660 | __add_wait_queue(wq, wait); | |
c27d53fb | 1661 | |
da55f2cc | 1662 | /* |
eb619fdb JA |
1663 | * It's possible that a tag was freed in the window between the |
1664 | * allocation failure and adding the hardware queue to the wait | |
1665 | * queue. | |
da55f2cc | 1666 | */ |
8ab6bb9e | 1667 | ret = blk_mq_get_driver_tag(rq); |
c27d53fb | 1668 | if (!ret) { |
5815839b ML |
1669 | spin_unlock(&hctx->dispatch_wait_lock); |
1670 | spin_unlock_irq(&wq->lock); | |
c27d53fb | 1671 | return false; |
eb619fdb | 1672 | } |
c27d53fb BVA |
1673 | |
1674 | /* | |
1675 | * We got a tag, remove ourselves from the wait queue to ensure | |
1676 | * someone else gets the wakeup. | |
1677 | */ | |
c27d53fb | 1678 | list_del_init(&wait->entry); |
e8618575 | 1679 | atomic_dec(&sbq->ws_active); |
5815839b ML |
1680 | spin_unlock(&hctx->dispatch_wait_lock); |
1681 | spin_unlock_irq(&wq->lock); | |
c27d53fb BVA |
1682 | |
1683 | return true; | |
da55f2cc OS |
1684 | } |
1685 | ||
6e768717 ML |
1686 | #define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT 8 |
1687 | #define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR 4 | |
1688 | /* | |
1689 | * Update dispatch busy with the Exponential Weighted Moving Average(EWMA): | |
1690 | * - EWMA is one simple way to compute running average value | |
1691 | * - weight(7/8 and 1/8) is applied so that it can decrease exponentially | |
1692 | * - take 4 as factor for avoiding to get too small(0) result, and this | |
1693 | * factor doesn't matter because EWMA decreases exponentially | |
1694 | */ | |
1695 | static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx *hctx, bool busy) | |
1696 | { | |
1697 | unsigned int ewma; | |
1698 | ||
6e768717 ML |
1699 | ewma = hctx->dispatch_busy; |
1700 | ||
1701 | if (!ewma && !busy) | |
1702 | return; | |
1703 | ||
1704 | ewma *= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT - 1; | |
1705 | if (busy) | |
1706 | ewma += 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR; | |
1707 | ewma /= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT; | |
1708 | ||
1709 | hctx->dispatch_busy = ewma; | |
1710 | } | |
1711 | ||
86ff7c2a ML |
1712 | #define BLK_MQ_RESOURCE_DELAY 3 /* ms units */ |
1713 | ||
c92a4103 JT |
1714 | static void blk_mq_handle_dev_resource(struct request *rq, |
1715 | struct list_head *list) | |
1716 | { | |
1717 | struct request *next = | |
1718 | list_first_entry_or_null(list, struct request, queuelist); | |
1719 | ||
1720 | /* | |
1721 | * If an I/O scheduler has been configured and we got a driver tag for | |
1722 | * the next request already, free it. | |
1723 | */ | |
1724 | if (next) | |
1725 | blk_mq_put_driver_tag(next); | |
1726 | ||
1727 | list_add(&rq->queuelist, list); | |
1728 | __blk_mq_requeue_request(rq); | |
1729 | } | |
1730 | ||
0512a75b KB |
1731 | static void blk_mq_handle_zone_resource(struct request *rq, |
1732 | struct list_head *zone_list) | |
1733 | { | |
1734 | /* | |
1735 | * If we end up here it is because we cannot dispatch a request to a | |
1736 | * specific zone due to LLD level zone-write locking or other zone | |
1737 | * related resource not being available. In this case, set the request | |
1738 | * aside in zone_list for retrying it later. | |
1739 | */ | |
1740 | list_add(&rq->queuelist, zone_list); | |
1741 | __blk_mq_requeue_request(rq); | |
1742 | } | |
1743 | ||
75383524 ML |
1744 | enum prep_dispatch { |
1745 | PREP_DISPATCH_OK, | |
1746 | PREP_DISPATCH_NO_TAG, | |
1747 | PREP_DISPATCH_NO_BUDGET, | |
1748 | }; | |
1749 | ||
1750 | static enum prep_dispatch blk_mq_prep_dispatch_rq(struct request *rq, | |
1751 | bool need_budget) | |
1752 | { | |
1753 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; | |
2a5a24aa | 1754 | int budget_token = -1; |
75383524 | 1755 | |
2a5a24aa ML |
1756 | if (need_budget) { |
1757 | budget_token = blk_mq_get_dispatch_budget(rq->q); | |
1758 | if (budget_token < 0) { | |
1759 | blk_mq_put_driver_tag(rq); | |
1760 | return PREP_DISPATCH_NO_BUDGET; | |
1761 | } | |
1762 | blk_mq_set_rq_budget_token(rq, budget_token); | |
75383524 ML |
1763 | } |
1764 | ||
1765 | if (!blk_mq_get_driver_tag(rq)) { | |
1766 | /* | |
1767 | * The initial allocation attempt failed, so we need to | |
1768 | * rerun the hardware queue when a tag is freed. The | |
1769 | * waitqueue takes care of that. If the queue is run | |
1770 | * before we add this entry back on the dispatch list, | |
1771 | * we'll re-run it below. | |
1772 | */ | |
1773 | if (!blk_mq_mark_tag_wait(hctx, rq)) { | |
1fd40b5e ML |
1774 | /* |
1775 | * All budgets not got from this function will be put | |
1776 | * together during handling partial dispatch | |
1777 | */ | |
1778 | if (need_budget) | |
2a5a24aa | 1779 | blk_mq_put_dispatch_budget(rq->q, budget_token); |
75383524 ML |
1780 | return PREP_DISPATCH_NO_TAG; |
1781 | } | |
1782 | } | |
1783 | ||
1784 | return PREP_DISPATCH_OK; | |
1785 | } | |
1786 | ||
1fd40b5e ML |
1787 | /* release all allocated budgets before calling to blk_mq_dispatch_rq_list */ |
1788 | static void blk_mq_release_budgets(struct request_queue *q, | |
2a5a24aa | 1789 | struct list_head *list) |
1fd40b5e | 1790 | { |
2a5a24aa | 1791 | struct request *rq; |
1fd40b5e | 1792 | |
2a5a24aa ML |
1793 | list_for_each_entry(rq, list, queuelist) { |
1794 | int budget_token = blk_mq_get_rq_budget_token(rq); | |
1fd40b5e | 1795 | |
2a5a24aa ML |
1796 | if (budget_token >= 0) |
1797 | blk_mq_put_dispatch_budget(q, budget_token); | |
1798 | } | |
1fd40b5e ML |
1799 | } |
1800 | ||
1f57f8d4 JA |
1801 | /* |
1802 | * Returns true if we did some work AND can potentially do more. | |
1803 | */ | |
445874e8 | 1804 | bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx *hctx, struct list_head *list, |
1fd40b5e | 1805 | unsigned int nr_budgets) |
320ae51f | 1806 | { |
75383524 | 1807 | enum prep_dispatch prep; |
445874e8 | 1808 | struct request_queue *q = hctx->queue; |
6d6f167c | 1809 | struct request *rq, *nxt; |
fc17b653 | 1810 | int errors, queued; |
86ff7c2a | 1811 | blk_status_t ret = BLK_STS_OK; |
0512a75b | 1812 | LIST_HEAD(zone_list); |
9586e67b | 1813 | bool needs_resource = false; |
320ae51f | 1814 | |
81380ca1 OS |
1815 | if (list_empty(list)) |
1816 | return false; | |
1817 | ||
320ae51f JA |
1818 | /* |
1819 | * Now process all the entries, sending them to the driver. | |
1820 | */ | |
93efe981 | 1821 | errors = queued = 0; |
81380ca1 | 1822 | do { |
74c45052 | 1823 | struct blk_mq_queue_data bd; |
320ae51f | 1824 | |
f04c3df3 | 1825 | rq = list_first_entry(list, struct request, queuelist); |
0bca799b | 1826 | |
445874e8 | 1827 | WARN_ON_ONCE(hctx != rq->mq_hctx); |
1fd40b5e | 1828 | prep = blk_mq_prep_dispatch_rq(rq, !nr_budgets); |
75383524 | 1829 | if (prep != PREP_DISPATCH_OK) |
0bca799b | 1830 | break; |
de148297 | 1831 | |
320ae51f | 1832 | list_del_init(&rq->queuelist); |
320ae51f | 1833 | |
74c45052 | 1834 | bd.rq = rq; |
113285b4 JA |
1835 | |
1836 | /* | |
1837 | * Flag last if we have no more requests, or if we have more | |
1838 | * but can't assign a driver tag to it. | |
1839 | */ | |
1840 | if (list_empty(list)) | |
1841 | bd.last = true; | |
1842 | else { | |
113285b4 | 1843 | nxt = list_first_entry(list, struct request, queuelist); |
8ab6bb9e | 1844 | bd.last = !blk_mq_get_driver_tag(nxt); |
113285b4 | 1845 | } |
74c45052 | 1846 | |
1fd40b5e ML |
1847 | /* |
1848 | * once the request is queued to lld, no need to cover the | |
1849 | * budget any more | |
1850 | */ | |
1851 | if (nr_budgets) | |
1852 | nr_budgets--; | |
74c45052 | 1853 | ret = q->mq_ops->queue_rq(hctx, &bd); |
7bf13729 ML |
1854 | switch (ret) { |
1855 | case BLK_STS_OK: | |
1856 | queued++; | |
320ae51f | 1857 | break; |
7bf13729 | 1858 | case BLK_STS_RESOURCE: |
9586e67b NA |
1859 | needs_resource = true; |
1860 | fallthrough; | |
7bf13729 ML |
1861 | case BLK_STS_DEV_RESOURCE: |
1862 | blk_mq_handle_dev_resource(rq, list); | |
1863 | goto out; | |
1864 | case BLK_STS_ZONE_RESOURCE: | |
0512a75b KB |
1865 | /* |
1866 | * Move the request to zone_list and keep going through | |
1867 | * the dispatch list to find more requests the drive can | |
1868 | * accept. | |
1869 | */ | |
1870 | blk_mq_handle_zone_resource(rq, &zone_list); | |
9586e67b | 1871 | needs_resource = true; |
7bf13729 ML |
1872 | break; |
1873 | default: | |
93efe981 | 1874 | errors++; |
e21ee5a6 | 1875 | blk_mq_end_request(rq, ret); |
320ae51f | 1876 | } |
81380ca1 | 1877 | } while (!list_empty(list)); |
7bf13729 | 1878 | out: |
0512a75b KB |
1879 | if (!list_empty(&zone_list)) |
1880 | list_splice_tail_init(&zone_list, list); | |
1881 | ||
632bfb63 | 1882 | /* If we didn't flush the entire list, we could have told the driver |
1883 | * there was more coming, but that turned out to be a lie. | |
1884 | */ | |
1885 | if ((!list_empty(list) || errors) && q->mq_ops->commit_rqs && queued) | |
1886 | q->mq_ops->commit_rqs(hctx); | |
320ae51f JA |
1887 | /* |
1888 | * Any items that need requeuing? Stuff them into hctx->dispatch, | |
1889 | * that is where we will continue on next queue run. | |
1890 | */ | |
f04c3df3 | 1891 | if (!list_empty(list)) { |
86ff7c2a | 1892 | bool needs_restart; |
75383524 ML |
1893 | /* For non-shared tags, the RESTART check will suffice */ |
1894 | bool no_tag = prep == PREP_DISPATCH_NO_TAG && | |
51db1c37 | 1895 | (hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED); |
86ff7c2a | 1896 | |
2a5a24aa ML |
1897 | if (nr_budgets) |
1898 | blk_mq_release_budgets(q, list); | |
86ff7c2a | 1899 | |
320ae51f | 1900 | spin_lock(&hctx->lock); |
01e99aec | 1901 | list_splice_tail_init(list, &hctx->dispatch); |
320ae51f | 1902 | spin_unlock(&hctx->lock); |
f04c3df3 | 1903 | |
d7d8535f ML |
1904 | /* |
1905 | * Order adding requests to hctx->dispatch and checking | |
1906 | * SCHED_RESTART flag. The pair of this smp_mb() is the one | |
1907 | * in blk_mq_sched_restart(). Avoid restart code path to | |
1908 | * miss the new added requests to hctx->dispatch, meantime | |
1909 | * SCHED_RESTART is observed here. | |
1910 | */ | |
1911 | smp_mb(); | |
1912 | ||
9ba52e58 | 1913 | /* |
710c785f BVA |
1914 | * If SCHED_RESTART was set by the caller of this function and |
1915 | * it is no longer set that means that it was cleared by another | |
1916 | * thread and hence that a queue rerun is needed. | |
9ba52e58 | 1917 | * |
eb619fdb JA |
1918 | * If 'no_tag' is set, that means that we failed getting |
1919 | * a driver tag with an I/O scheduler attached. If our dispatch | |
1920 | * waitqueue is no longer active, ensure that we run the queue | |
1921 | * AFTER adding our entries back to the list. | |
bd166ef1 | 1922 | * |
710c785f BVA |
1923 | * If no I/O scheduler has been configured it is possible that |
1924 | * the hardware queue got stopped and restarted before requests | |
1925 | * were pushed back onto the dispatch list. Rerun the queue to | |
1926 | * avoid starvation. Notes: | |
1927 | * - blk_mq_run_hw_queue() checks whether or not a queue has | |
1928 | * been stopped before rerunning a queue. | |
1929 | * - Some but not all block drivers stop a queue before | |
fc17b653 | 1930 | * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq |
710c785f | 1931 | * and dm-rq. |
86ff7c2a ML |
1932 | * |
1933 | * If driver returns BLK_STS_RESOURCE and SCHED_RESTART | |
1934 | * bit is set, run queue after a delay to avoid IO stalls | |
ab3cee37 | 1935 | * that could otherwise occur if the queue is idle. We'll do |
9586e67b NA |
1936 | * similar if we couldn't get budget or couldn't lock a zone |
1937 | * and SCHED_RESTART is set. | |
bd166ef1 | 1938 | */ |
86ff7c2a | 1939 | needs_restart = blk_mq_sched_needs_restart(hctx); |
9586e67b NA |
1940 | if (prep == PREP_DISPATCH_NO_BUDGET) |
1941 | needs_resource = true; | |
86ff7c2a | 1942 | if (!needs_restart || |
eb619fdb | 1943 | (no_tag && list_empty_careful(&hctx->dispatch_wait.entry))) |
bd166ef1 | 1944 | blk_mq_run_hw_queue(hctx, true); |
9586e67b | 1945 | else if (needs_restart && needs_resource) |
86ff7c2a | 1946 | blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY); |
1f57f8d4 | 1947 | |
6e768717 | 1948 | blk_mq_update_dispatch_busy(hctx, true); |
1f57f8d4 | 1949 | return false; |
6e768717 ML |
1950 | } else |
1951 | blk_mq_update_dispatch_busy(hctx, false); | |
f04c3df3 | 1952 | |
93efe981 | 1953 | return (queued + errors) != 0; |
f04c3df3 JA |
1954 | } |
1955 | ||
105663f7 AA |
1956 | /** |
1957 | * __blk_mq_run_hw_queue - Run a hardware queue. | |
1958 | * @hctx: Pointer to the hardware queue to run. | |
1959 | * | |
1960 | * Send pending requests to the hardware. | |
1961 | */ | |
6a83e74d BVA |
1962 | static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx) |
1963 | { | |
b7a71e66 JA |
1964 | /* |
1965 | * We can't run the queue inline with ints disabled. Ensure that | |
1966 | * we catch bad users of this early. | |
1967 | */ | |
1968 | WARN_ON_ONCE(in_interrupt()); | |
1969 | ||
bcc330f4 ML |
1970 | blk_mq_run_dispatch_ops(hctx->queue, |
1971 | blk_mq_sched_dispatch_requests(hctx)); | |
6a83e74d BVA |
1972 | } |
1973 | ||
f82ddf19 ML |
1974 | static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx) |
1975 | { | |
1976 | int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask); | |
1977 | ||
1978 | if (cpu >= nr_cpu_ids) | |
1979 | cpu = cpumask_first(hctx->cpumask); | |
1980 | return cpu; | |
1981 | } | |
1982 | ||
506e931f JA |
1983 | /* |
1984 | * It'd be great if the workqueue API had a way to pass | |
1985 | * in a mask and had some smarts for more clever placement. | |
1986 | * For now we just round-robin here, switching for every | |
1987 | * BLK_MQ_CPU_WORK_BATCH queued items. | |
1988 | */ | |
1989 | static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx) | |
1990 | { | |
7bed4595 | 1991 | bool tried = false; |
476f8c98 | 1992 | int next_cpu = hctx->next_cpu; |
7bed4595 | 1993 | |
b657d7e6 CH |
1994 | if (hctx->queue->nr_hw_queues == 1) |
1995 | return WORK_CPU_UNBOUND; | |
506e931f JA |
1996 | |
1997 | if (--hctx->next_cpu_batch <= 0) { | |
7bed4595 | 1998 | select_cpu: |
476f8c98 | 1999 | next_cpu = cpumask_next_and(next_cpu, hctx->cpumask, |
20e4d813 | 2000 | cpu_online_mask); |
506e931f | 2001 | if (next_cpu >= nr_cpu_ids) |
f82ddf19 | 2002 | next_cpu = blk_mq_first_mapped_cpu(hctx); |
506e931f JA |
2003 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; |
2004 | } | |
2005 | ||
7bed4595 ML |
2006 | /* |
2007 | * Do unbound schedule if we can't find a online CPU for this hctx, | |
2008 | * and it should only happen in the path of handling CPU DEAD. | |
2009 | */ | |
476f8c98 | 2010 | if (!cpu_online(next_cpu)) { |
7bed4595 ML |
2011 | if (!tried) { |
2012 | tried = true; | |
2013 | goto select_cpu; | |
2014 | } | |
2015 | ||
2016 | /* | |
2017 | * Make sure to re-select CPU next time once after CPUs | |
2018 | * in hctx->cpumask become online again. | |
2019 | */ | |
476f8c98 | 2020 | hctx->next_cpu = next_cpu; |
7bed4595 ML |
2021 | hctx->next_cpu_batch = 1; |
2022 | return WORK_CPU_UNBOUND; | |
2023 | } | |
476f8c98 ML |
2024 | |
2025 | hctx->next_cpu = next_cpu; | |
2026 | return next_cpu; | |
506e931f JA |
2027 | } |
2028 | ||
105663f7 AA |
2029 | /** |
2030 | * __blk_mq_delay_run_hw_queue - Run (or schedule to run) a hardware queue. | |
2031 | * @hctx: Pointer to the hardware queue to run. | |
2032 | * @async: If we want to run the queue asynchronously. | |
fa94ba8a | 2033 | * @msecs: Milliseconds of delay to wait before running the queue. |
105663f7 AA |
2034 | * |
2035 | * If !@async, try to run the queue now. Else, run the queue asynchronously and | |
2036 | * with a delay of @msecs. | |
2037 | */ | |
7587a5ae BVA |
2038 | static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async, |
2039 | unsigned long msecs) | |
320ae51f | 2040 | { |
5435c023 | 2041 | if (unlikely(blk_mq_hctx_stopped(hctx))) |
320ae51f JA |
2042 | return; |
2043 | ||
1b792f2f | 2044 | if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) { |
2a90d4aa PB |
2045 | int cpu = get_cpu(); |
2046 | if (cpumask_test_cpu(cpu, hctx->cpumask)) { | |
398205b8 | 2047 | __blk_mq_run_hw_queue(hctx); |
2a90d4aa | 2048 | put_cpu(); |
398205b8 PB |
2049 | return; |
2050 | } | |
e4043dcf | 2051 | |
2a90d4aa | 2052 | put_cpu(); |
e4043dcf | 2053 | } |
398205b8 | 2054 | |
ae943d20 BVA |
2055 | kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work, |
2056 | msecs_to_jiffies(msecs)); | |
7587a5ae BVA |
2057 | } |
2058 | ||
105663f7 AA |
2059 | /** |
2060 | * blk_mq_delay_run_hw_queue - Run a hardware queue asynchronously. | |
2061 | * @hctx: Pointer to the hardware queue to run. | |
fa94ba8a | 2062 | * @msecs: Milliseconds of delay to wait before running the queue. |
105663f7 AA |
2063 | * |
2064 | * Run a hardware queue asynchronously with a delay of @msecs. | |
2065 | */ | |
7587a5ae BVA |
2066 | void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs) |
2067 | { | |
2068 | __blk_mq_delay_run_hw_queue(hctx, true, msecs); | |
2069 | } | |
2070 | EXPORT_SYMBOL(blk_mq_delay_run_hw_queue); | |
2071 | ||
105663f7 AA |
2072 | /** |
2073 | * blk_mq_run_hw_queue - Start to run a hardware queue. | |
2074 | * @hctx: Pointer to the hardware queue to run. | |
2075 | * @async: If we want to run the queue asynchronously. | |
2076 | * | |
2077 | * Check if the request queue is not in a quiesced state and if there are | |
2078 | * pending requests to be sent. If this is true, run the queue to send requests | |
2079 | * to hardware. | |
2080 | */ | |
626fb735 | 2081 | void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) |
7587a5ae | 2082 | { |
24f5a90f ML |
2083 | bool need_run; |
2084 | ||
2085 | /* | |
2086 | * When queue is quiesced, we may be switching io scheduler, or | |
2087 | * updating nr_hw_queues, or other things, and we can't run queue | |
2088 | * any more, even __blk_mq_hctx_has_pending() can't be called safely. | |
2089 | * | |
2090 | * And queue will be rerun in blk_mq_unquiesce_queue() if it is | |
2091 | * quiesced. | |
2092 | */ | |
41adf531 | 2093 | __blk_mq_run_dispatch_ops(hctx->queue, false, |
2a904d00 ML |
2094 | need_run = !blk_queue_quiesced(hctx->queue) && |
2095 | blk_mq_hctx_has_pending(hctx)); | |
24f5a90f | 2096 | |
626fb735 | 2097 | if (need_run) |
79f720a7 | 2098 | __blk_mq_delay_run_hw_queue(hctx, async, 0); |
320ae51f | 2099 | } |
5b727272 | 2100 | EXPORT_SYMBOL(blk_mq_run_hw_queue); |
320ae51f | 2101 | |
b6e68ee8 JK |
2102 | /* |
2103 | * Is the request queue handled by an IO scheduler that does not respect | |
2104 | * hardware queues when dispatching? | |
2105 | */ | |
2106 | static bool blk_mq_has_sqsched(struct request_queue *q) | |
2107 | { | |
2108 | struct elevator_queue *e = q->elevator; | |
2109 | ||
2110 | if (e && e->type->ops.dispatch_request && | |
2111 | !(e->type->elevator_features & ELEVATOR_F_MQ_AWARE)) | |
2112 | return true; | |
2113 | return false; | |
2114 | } | |
2115 | ||
2116 | /* | |
2117 | * Return prefered queue to dispatch from (if any) for non-mq aware IO | |
2118 | * scheduler. | |
2119 | */ | |
2120 | static struct blk_mq_hw_ctx *blk_mq_get_sq_hctx(struct request_queue *q) | |
2121 | { | |
2122 | struct blk_mq_hw_ctx *hctx; | |
2123 | ||
2124 | /* | |
2125 | * If the IO scheduler does not respect hardware queues when | |
2126 | * dispatching, we just don't bother with multiple HW queues and | |
2127 | * dispatch from hctx for the current CPU since running multiple queues | |
2128 | * just causes lock contention inside the scheduler and pointless cache | |
2129 | * bouncing. | |
2130 | */ | |
2131 | hctx = blk_mq_map_queue_type(q, HCTX_TYPE_DEFAULT, | |
2132 | raw_smp_processor_id()); | |
2133 | if (!blk_mq_hctx_stopped(hctx)) | |
2134 | return hctx; | |
2135 | return NULL; | |
2136 | } | |
2137 | ||
105663f7 | 2138 | /** |
24f7bb88 | 2139 | * blk_mq_run_hw_queues - Run all hardware queues in a request queue. |
105663f7 AA |
2140 | * @q: Pointer to the request queue to run. |
2141 | * @async: If we want to run the queue asynchronously. | |
2142 | */ | |
b94ec296 | 2143 | void blk_mq_run_hw_queues(struct request_queue *q, bool async) |
320ae51f | 2144 | { |
b6e68ee8 | 2145 | struct blk_mq_hw_ctx *hctx, *sq_hctx; |
320ae51f JA |
2146 | int i; |
2147 | ||
b6e68ee8 JK |
2148 | sq_hctx = NULL; |
2149 | if (blk_mq_has_sqsched(q)) | |
2150 | sq_hctx = blk_mq_get_sq_hctx(q); | |
320ae51f | 2151 | queue_for_each_hw_ctx(q, hctx, i) { |
79f720a7 | 2152 | if (blk_mq_hctx_stopped(hctx)) |
320ae51f | 2153 | continue; |
b6e68ee8 JK |
2154 | /* |
2155 | * Dispatch from this hctx either if there's no hctx preferred | |
2156 | * by IO scheduler or if it has requests that bypass the | |
2157 | * scheduler. | |
2158 | */ | |
2159 | if (!sq_hctx || sq_hctx == hctx || | |
2160 | !list_empty_careful(&hctx->dispatch)) | |
2161 | blk_mq_run_hw_queue(hctx, async); | |
320ae51f JA |
2162 | } |
2163 | } | |
b94ec296 | 2164 | EXPORT_SYMBOL(blk_mq_run_hw_queues); |
320ae51f | 2165 | |
b9151e7b DA |
2166 | /** |
2167 | * blk_mq_delay_run_hw_queues - Run all hardware queues asynchronously. | |
2168 | * @q: Pointer to the request queue to run. | |
fa94ba8a | 2169 | * @msecs: Milliseconds of delay to wait before running the queues. |
b9151e7b DA |
2170 | */ |
2171 | void blk_mq_delay_run_hw_queues(struct request_queue *q, unsigned long msecs) | |
2172 | { | |
b6e68ee8 | 2173 | struct blk_mq_hw_ctx *hctx, *sq_hctx; |
b9151e7b DA |
2174 | int i; |
2175 | ||
b6e68ee8 JK |
2176 | sq_hctx = NULL; |
2177 | if (blk_mq_has_sqsched(q)) | |
2178 | sq_hctx = blk_mq_get_sq_hctx(q); | |
b9151e7b DA |
2179 | queue_for_each_hw_ctx(q, hctx, i) { |
2180 | if (blk_mq_hctx_stopped(hctx)) | |
2181 | continue; | |
8f5fea65 DJ |
2182 | /* |
2183 | * If there is already a run_work pending, leave the | |
2184 | * pending delay untouched. Otherwise, a hctx can stall | |
2185 | * if another hctx is re-delaying the other's work | |
2186 | * before the work executes. | |
2187 | */ | |
2188 | if (delayed_work_pending(&hctx->run_work)) | |
2189 | continue; | |
b6e68ee8 JK |
2190 | /* |
2191 | * Dispatch from this hctx either if there's no hctx preferred | |
2192 | * by IO scheduler or if it has requests that bypass the | |
2193 | * scheduler. | |
2194 | */ | |
2195 | if (!sq_hctx || sq_hctx == hctx || | |
2196 | !list_empty_careful(&hctx->dispatch)) | |
2197 | blk_mq_delay_run_hw_queue(hctx, msecs); | |
b9151e7b DA |
2198 | } |
2199 | } | |
2200 | EXPORT_SYMBOL(blk_mq_delay_run_hw_queues); | |
2201 | ||
fd001443 BVA |
2202 | /** |
2203 | * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped | |
2204 | * @q: request queue. | |
2205 | * | |
2206 | * The caller is responsible for serializing this function against | |
2207 | * blk_mq_{start,stop}_hw_queue(). | |
2208 | */ | |
2209 | bool blk_mq_queue_stopped(struct request_queue *q) | |
2210 | { | |
2211 | struct blk_mq_hw_ctx *hctx; | |
2212 | int i; | |
2213 | ||
2214 | queue_for_each_hw_ctx(q, hctx, i) | |
2215 | if (blk_mq_hctx_stopped(hctx)) | |
2216 | return true; | |
2217 | ||
2218 | return false; | |
2219 | } | |
2220 | EXPORT_SYMBOL(blk_mq_queue_stopped); | |
2221 | ||
39a70c76 ML |
2222 | /* |
2223 | * This function is often used for pausing .queue_rq() by driver when | |
2224 | * there isn't enough resource or some conditions aren't satisfied, and | |
4d606219 | 2225 | * BLK_STS_RESOURCE is usually returned. |
39a70c76 ML |
2226 | * |
2227 | * We do not guarantee that dispatch can be drained or blocked | |
2228 | * after blk_mq_stop_hw_queue() returns. Please use | |
2229 | * blk_mq_quiesce_queue() for that requirement. | |
2230 | */ | |
2719aa21 JA |
2231 | void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx) |
2232 | { | |
641a9ed6 | 2233 | cancel_delayed_work(&hctx->run_work); |
280d45f6 | 2234 | |
641a9ed6 | 2235 | set_bit(BLK_MQ_S_STOPPED, &hctx->state); |
2719aa21 | 2236 | } |
641a9ed6 | 2237 | EXPORT_SYMBOL(blk_mq_stop_hw_queue); |
2719aa21 | 2238 | |
39a70c76 ML |
2239 | /* |
2240 | * This function is often used for pausing .queue_rq() by driver when | |
2241 | * there isn't enough resource or some conditions aren't satisfied, and | |
4d606219 | 2242 | * BLK_STS_RESOURCE is usually returned. |
39a70c76 ML |
2243 | * |
2244 | * We do not guarantee that dispatch can be drained or blocked | |
2245 | * after blk_mq_stop_hw_queues() returns. Please use | |
2246 | * blk_mq_quiesce_queue() for that requirement. | |
2247 | */ | |
2719aa21 JA |
2248 | void blk_mq_stop_hw_queues(struct request_queue *q) |
2249 | { | |
641a9ed6 ML |
2250 | struct blk_mq_hw_ctx *hctx; |
2251 | int i; | |
2252 | ||
2253 | queue_for_each_hw_ctx(q, hctx, i) | |
2254 | blk_mq_stop_hw_queue(hctx); | |
280d45f6 CH |
2255 | } |
2256 | EXPORT_SYMBOL(blk_mq_stop_hw_queues); | |
2257 | ||
320ae51f JA |
2258 | void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx) |
2259 | { | |
2260 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
e4043dcf | 2261 | |
0ffbce80 | 2262 | blk_mq_run_hw_queue(hctx, false); |
320ae51f JA |
2263 | } |
2264 | EXPORT_SYMBOL(blk_mq_start_hw_queue); | |
2265 | ||
2f268556 CH |
2266 | void blk_mq_start_hw_queues(struct request_queue *q) |
2267 | { | |
2268 | struct blk_mq_hw_ctx *hctx; | |
2269 | int i; | |
2270 | ||
2271 | queue_for_each_hw_ctx(q, hctx, i) | |
2272 | blk_mq_start_hw_queue(hctx); | |
2273 | } | |
2274 | EXPORT_SYMBOL(blk_mq_start_hw_queues); | |
2275 | ||
ae911c5e JA |
2276 | void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) |
2277 | { | |
2278 | if (!blk_mq_hctx_stopped(hctx)) | |
2279 | return; | |
2280 | ||
2281 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
2282 | blk_mq_run_hw_queue(hctx, async); | |
2283 | } | |
2284 | EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue); | |
2285 | ||
1b4a3258 | 2286 | void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async) |
320ae51f JA |
2287 | { |
2288 | struct blk_mq_hw_ctx *hctx; | |
2289 | int i; | |
2290 | ||
ae911c5e JA |
2291 | queue_for_each_hw_ctx(q, hctx, i) |
2292 | blk_mq_start_stopped_hw_queue(hctx, async); | |
320ae51f JA |
2293 | } |
2294 | EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues); | |
2295 | ||
70f4db63 | 2296 | static void blk_mq_run_work_fn(struct work_struct *work) |
320ae51f JA |
2297 | { |
2298 | struct blk_mq_hw_ctx *hctx; | |
2299 | ||
9f993737 | 2300 | hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work); |
320ae51f | 2301 | |
21c6e939 | 2302 | /* |
15fe8a90 | 2303 | * If we are stopped, don't run the queue. |
21c6e939 | 2304 | */ |
0841031a | 2305 | if (blk_mq_hctx_stopped(hctx)) |
0196d6b4 | 2306 | return; |
7587a5ae BVA |
2307 | |
2308 | __blk_mq_run_hw_queue(hctx); | |
2309 | } | |
2310 | ||
cfd0c552 | 2311 | static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx, |
cfd0c552 ML |
2312 | struct request *rq, |
2313 | bool at_head) | |
320ae51f | 2314 | { |
e57690fe | 2315 | struct blk_mq_ctx *ctx = rq->mq_ctx; |
c16d6b5a | 2316 | enum hctx_type type = hctx->type; |
e57690fe | 2317 | |
7b607814 BVA |
2318 | lockdep_assert_held(&ctx->lock); |
2319 | ||
a54895fa | 2320 | trace_block_rq_insert(rq); |
01b983c9 | 2321 | |
72a0a36e | 2322 | if (at_head) |
c16d6b5a | 2323 | list_add(&rq->queuelist, &ctx->rq_lists[type]); |
72a0a36e | 2324 | else |
c16d6b5a | 2325 | list_add_tail(&rq->queuelist, &ctx->rq_lists[type]); |
cfd0c552 | 2326 | } |
4bb659b1 | 2327 | |
2c3ad667 JA |
2328 | void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq, |
2329 | bool at_head) | |
cfd0c552 ML |
2330 | { |
2331 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
2332 | ||
7b607814 BVA |
2333 | lockdep_assert_held(&ctx->lock); |
2334 | ||
e57690fe | 2335 | __blk_mq_insert_req_list(hctx, rq, at_head); |
320ae51f | 2336 | blk_mq_hctx_mark_pending(hctx, ctx); |
320ae51f JA |
2337 | } |
2338 | ||
105663f7 AA |
2339 | /** |
2340 | * blk_mq_request_bypass_insert - Insert a request at dispatch list. | |
2341 | * @rq: Pointer to request to be inserted. | |
26bfeb26 | 2342 | * @at_head: true if the request should be inserted at the head of the list. |
105663f7 AA |
2343 | * @run_queue: If we should run the hardware queue after inserting the request. |
2344 | * | |
157f377b JA |
2345 | * Should only be used carefully, when the caller knows we want to |
2346 | * bypass a potential IO scheduler on the target device. | |
2347 | */ | |
01e99aec ML |
2348 | void blk_mq_request_bypass_insert(struct request *rq, bool at_head, |
2349 | bool run_queue) | |
157f377b | 2350 | { |
ea4f995e | 2351 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; |
157f377b JA |
2352 | |
2353 | spin_lock(&hctx->lock); | |
01e99aec ML |
2354 | if (at_head) |
2355 | list_add(&rq->queuelist, &hctx->dispatch); | |
2356 | else | |
2357 | list_add_tail(&rq->queuelist, &hctx->dispatch); | |
157f377b JA |
2358 | spin_unlock(&hctx->lock); |
2359 | ||
b0850297 ML |
2360 | if (run_queue) |
2361 | blk_mq_run_hw_queue(hctx, false); | |
157f377b JA |
2362 | } |
2363 | ||
bd166ef1 JA |
2364 | void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx, |
2365 | struct list_head *list) | |
320ae51f JA |
2366 | |
2367 | { | |
3f0cedc7 | 2368 | struct request *rq; |
c16d6b5a | 2369 | enum hctx_type type = hctx->type; |
3f0cedc7 | 2370 | |
320ae51f JA |
2371 | /* |
2372 | * preemption doesn't flush plug list, so it's possible ctx->cpu is | |
2373 | * offline now | |
2374 | */ | |
3f0cedc7 | 2375 | list_for_each_entry(rq, list, queuelist) { |
e57690fe | 2376 | BUG_ON(rq->mq_ctx != ctx); |
a54895fa | 2377 | trace_block_rq_insert(rq); |
320ae51f | 2378 | } |
3f0cedc7 ML |
2379 | |
2380 | spin_lock(&ctx->lock); | |
c16d6b5a | 2381 | list_splice_tail_init(list, &ctx->rq_lists[type]); |
cfd0c552 | 2382 | blk_mq_hctx_mark_pending(hctx, ctx); |
320ae51f | 2383 | spin_unlock(&ctx->lock); |
320ae51f JA |
2384 | } |
2385 | ||
dc5fc361 JA |
2386 | static void blk_mq_commit_rqs(struct blk_mq_hw_ctx *hctx, int *queued, |
2387 | bool from_schedule) | |
320ae51f | 2388 | { |
dc5fc361 JA |
2389 | if (hctx->queue->mq_ops->commit_rqs) { |
2390 | trace_block_unplug(hctx->queue, *queued, !from_schedule); | |
2391 | hctx->queue->mq_ops->commit_rqs(hctx); | |
2392 | } | |
2393 | *queued = 0; | |
2394 | } | |
320ae51f | 2395 | |
14ccb66b CH |
2396 | static void blk_mq_bio_to_request(struct request *rq, struct bio *bio, |
2397 | unsigned int nr_segs) | |
320ae51f | 2398 | { |
93f221ae EB |
2399 | int err; |
2400 | ||
f924cdde CH |
2401 | if (bio->bi_opf & REQ_RAHEAD) |
2402 | rq->cmd_flags |= REQ_FAILFAST_MASK; | |
2403 | ||
2404 | rq->__sector = bio->bi_iter.bi_sector; | |
2405 | rq->write_hint = bio->bi_write_hint; | |
14ccb66b | 2406 | blk_rq_bio_prep(rq, bio, nr_segs); |
93f221ae EB |
2407 | |
2408 | /* This can't fail, since GFP_NOIO includes __GFP_DIRECT_RECLAIM. */ | |
2409 | err = blk_crypto_rq_bio_prep(rq, bio, GFP_NOIO); | |
2410 | WARN_ON_ONCE(err); | |
4b570521 | 2411 | |
b5af37ab | 2412 | blk_account_io_start(rq); |
320ae51f JA |
2413 | } |
2414 | ||
0f95549c | 2415 | static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx, |
3e08773c | 2416 | struct request *rq, bool last) |
f984df1f | 2417 | { |
f984df1f | 2418 | struct request_queue *q = rq->q; |
f984df1f SL |
2419 | struct blk_mq_queue_data bd = { |
2420 | .rq = rq, | |
be94f058 | 2421 | .last = last, |
f984df1f | 2422 | }; |
f06345ad | 2423 | blk_status_t ret; |
0f95549c | 2424 | |
0f95549c MS |
2425 | /* |
2426 | * For OK queue, we are done. For error, caller may kill it. | |
2427 | * Any other error (busy), just add it to our list as we | |
2428 | * previously would have done. | |
2429 | */ | |
2430 | ret = q->mq_ops->queue_rq(hctx, &bd); | |
2431 | switch (ret) { | |
2432 | case BLK_STS_OK: | |
6ce3dd6e | 2433 | blk_mq_update_dispatch_busy(hctx, false); |
0f95549c MS |
2434 | break; |
2435 | case BLK_STS_RESOURCE: | |
86ff7c2a | 2436 | case BLK_STS_DEV_RESOURCE: |
6ce3dd6e | 2437 | blk_mq_update_dispatch_busy(hctx, true); |
0f95549c MS |
2438 | __blk_mq_requeue_request(rq); |
2439 | break; | |
2440 | default: | |
6ce3dd6e | 2441 | blk_mq_update_dispatch_busy(hctx, false); |
0f95549c MS |
2442 | break; |
2443 | } | |
2444 | ||
2445 | return ret; | |
2446 | } | |
2447 | ||
fd9c40f6 | 2448 | static blk_status_t __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx, |
0f95549c | 2449 | struct request *rq, |
fd9c40f6 | 2450 | bool bypass_insert, bool last) |
0f95549c MS |
2451 | { |
2452 | struct request_queue *q = rq->q; | |
d964f04a | 2453 | bool run_queue = true; |
2a5a24aa | 2454 | int budget_token; |
d964f04a | 2455 | |
23d4ee19 | 2456 | /* |
fd9c40f6 | 2457 | * RCU or SRCU read lock is needed before checking quiesced flag. |
23d4ee19 | 2458 | * |
fd9c40f6 BVA |
2459 | * When queue is stopped or quiesced, ignore 'bypass_insert' from |
2460 | * blk_mq_request_issue_directly(), and return BLK_STS_OK to caller, | |
2461 | * and avoid driver to try to dispatch again. | |
23d4ee19 | 2462 | */ |
fd9c40f6 | 2463 | if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)) { |
d964f04a | 2464 | run_queue = false; |
fd9c40f6 BVA |
2465 | bypass_insert = false; |
2466 | goto insert; | |
d964f04a | 2467 | } |
f984df1f | 2468 | |
2ff0682d | 2469 | if ((rq->rq_flags & RQF_ELV) && !bypass_insert) |
fd9c40f6 | 2470 | goto insert; |
2253efc8 | 2471 | |
2a5a24aa ML |
2472 | budget_token = blk_mq_get_dispatch_budget(q); |
2473 | if (budget_token < 0) | |
fd9c40f6 | 2474 | goto insert; |
bd166ef1 | 2475 | |
2a5a24aa ML |
2476 | blk_mq_set_rq_budget_token(rq, budget_token); |
2477 | ||
8ab6bb9e | 2478 | if (!blk_mq_get_driver_tag(rq)) { |
2a5a24aa | 2479 | blk_mq_put_dispatch_budget(q, budget_token); |
fd9c40f6 | 2480 | goto insert; |
88022d72 | 2481 | } |
de148297 | 2482 | |
3e08773c | 2483 | return __blk_mq_issue_directly(hctx, rq, last); |
fd9c40f6 BVA |
2484 | insert: |
2485 | if (bypass_insert) | |
2486 | return BLK_STS_RESOURCE; | |
2487 | ||
db03f88f ML |
2488 | blk_mq_sched_insert_request(rq, false, run_queue, false); |
2489 | ||
fd9c40f6 BVA |
2490 | return BLK_STS_OK; |
2491 | } | |
2492 | ||
105663f7 AA |
2493 | /** |
2494 | * blk_mq_try_issue_directly - Try to send a request directly to device driver. | |
2495 | * @hctx: Pointer of the associated hardware queue. | |
2496 | * @rq: Pointer to request to be sent. | |
105663f7 AA |
2497 | * |
2498 | * If the device has enough resources to accept a new request now, send the | |
2499 | * request directly to device driver. Else, insert at hctx->dispatch queue, so | |
2500 | * we can try send it another time in the future. Requests inserted at this | |
2501 | * queue have higher priority. | |
2502 | */ | |
fd9c40f6 | 2503 | static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx, |
3e08773c | 2504 | struct request *rq) |
fd9c40f6 | 2505 | { |
2a904d00 ML |
2506 | blk_status_t ret = |
2507 | __blk_mq_try_issue_directly(hctx, rq, false, true); | |
fd9c40f6 | 2508 | |
fd9c40f6 | 2509 | if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE) |
01e99aec | 2510 | blk_mq_request_bypass_insert(rq, false, true); |
fd9c40f6 BVA |
2511 | else if (ret != BLK_STS_OK) |
2512 | blk_mq_end_request(rq, ret); | |
fd9c40f6 BVA |
2513 | } |
2514 | ||
06c8c691 | 2515 | static blk_status_t blk_mq_request_issue_directly(struct request *rq, bool last) |
fd9c40f6 | 2516 | { |
4cafe86c | 2517 | return __blk_mq_try_issue_directly(rq->mq_hctx, rq, true, last); |
5eb6126e CH |
2518 | } |
2519 | ||
b84c5b50 CH |
2520 | static void blk_mq_plug_issue_direct(struct blk_plug *plug, bool from_schedule) |
2521 | { | |
2522 | struct blk_mq_hw_ctx *hctx = NULL; | |
2523 | struct request *rq; | |
2524 | int queued = 0; | |
2525 | int errors = 0; | |
2526 | ||
2527 | while ((rq = rq_list_pop(&plug->mq_list))) { | |
2528 | bool last = rq_list_empty(plug->mq_list); | |
2529 | blk_status_t ret; | |
2530 | ||
2531 | if (hctx != rq->mq_hctx) { | |
2532 | if (hctx) | |
2533 | blk_mq_commit_rqs(hctx, &queued, from_schedule); | |
2534 | hctx = rq->mq_hctx; | |
2535 | } | |
2536 | ||
2537 | ret = blk_mq_request_issue_directly(rq, last); | |
2538 | switch (ret) { | |
2539 | case BLK_STS_OK: | |
2540 | queued++; | |
2541 | break; | |
2542 | case BLK_STS_RESOURCE: | |
2543 | case BLK_STS_DEV_RESOURCE: | |
2544 | blk_mq_request_bypass_insert(rq, false, last); | |
2545 | blk_mq_commit_rqs(hctx, &queued, from_schedule); | |
2546 | return; | |
2547 | default: | |
2548 | blk_mq_end_request(rq, ret); | |
2549 | errors++; | |
2550 | break; | |
2551 | } | |
2552 | } | |
2553 | ||
2554 | /* | |
2555 | * If we didn't flush the entire list, we could have told the driver | |
2556 | * there was more coming, but that turned out to be a lie. | |
2557 | */ | |
2558 | if (errors) | |
2559 | blk_mq_commit_rqs(hctx, &queued, from_schedule); | |
2560 | } | |
2561 | ||
518579a9 KB |
2562 | static void __blk_mq_flush_plug_list(struct request_queue *q, |
2563 | struct blk_plug *plug) | |
2564 | { | |
2565 | if (blk_queue_quiesced(q)) | |
2566 | return; | |
2567 | q->mq_ops->queue_rqs(&plug->mq_list); | |
2568 | } | |
2569 | ||
b84c5b50 CH |
2570 | void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule) |
2571 | { | |
2572 | struct blk_mq_hw_ctx *this_hctx; | |
2573 | struct blk_mq_ctx *this_ctx; | |
3c67d44d | 2574 | struct request *rq; |
b84c5b50 CH |
2575 | unsigned int depth; |
2576 | LIST_HEAD(list); | |
2577 | ||
2578 | if (rq_list_empty(plug->mq_list)) | |
2579 | return; | |
2580 | plug->rq_count = 0; | |
2581 | ||
2582 | if (!plug->multiple_queues && !plug->has_elevator && !from_schedule) { | |
3c67d44d JA |
2583 | struct request_queue *q; |
2584 | ||
2585 | rq = rq_list_peek(&plug->mq_list); | |
2586 | q = rq->q; | |
2587 | ||
2588 | /* | |
2589 | * Peek first request and see if we have a ->queue_rqs() hook. | |
2590 | * If we do, we can dispatch the whole plug list in one go. We | |
2591 | * already know at this point that all requests belong to the | |
2592 | * same queue, caller must ensure that's the case. | |
2593 | * | |
2594 | * Since we pass off the full list to the driver at this point, | |
2595 | * we do not increment the active request count for the queue. | |
2596 | * Bypass shared tags for now because of that. | |
2597 | */ | |
2598 | if (q->mq_ops->queue_rqs && | |
2599 | !(rq->mq_hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) { | |
2600 | blk_mq_run_dispatch_ops(q, | |
518579a9 | 2601 | __blk_mq_flush_plug_list(q, plug)); |
3c67d44d JA |
2602 | if (rq_list_empty(plug->mq_list)) |
2603 | return; | |
2604 | } | |
73f3760e ML |
2605 | |
2606 | blk_mq_run_dispatch_ops(q, | |
4cafe86c | 2607 | blk_mq_plug_issue_direct(plug, false)); |
b84c5b50 CH |
2608 | if (rq_list_empty(plug->mq_list)) |
2609 | return; | |
2610 | } | |
2611 | ||
2612 | this_hctx = NULL; | |
2613 | this_ctx = NULL; | |
2614 | depth = 0; | |
2615 | do { | |
b84c5b50 CH |
2616 | rq = rq_list_pop(&plug->mq_list); |
2617 | ||
2618 | if (!this_hctx) { | |
2619 | this_hctx = rq->mq_hctx; | |
2620 | this_ctx = rq->mq_ctx; | |
2621 | } else if (this_hctx != rq->mq_hctx || this_ctx != rq->mq_ctx) { | |
2622 | trace_block_unplug(this_hctx->queue, depth, | |
2623 | !from_schedule); | |
2624 | blk_mq_sched_insert_requests(this_hctx, this_ctx, | |
2625 | &list, from_schedule); | |
2626 | depth = 0; | |
2627 | this_hctx = rq->mq_hctx; | |
2628 | this_ctx = rq->mq_ctx; | |
2629 | ||
2630 | } | |
2631 | ||
2632 | list_add(&rq->queuelist, &list); | |
2633 | depth++; | |
2634 | } while (!rq_list_empty(plug->mq_list)); | |
2635 | ||
2636 | if (!list_empty(&list)) { | |
2637 | trace_block_unplug(this_hctx->queue, depth, !from_schedule); | |
2638 | blk_mq_sched_insert_requests(this_hctx, this_ctx, &list, | |
2639 | from_schedule); | |
2640 | } | |
2641 | } | |
2642 | ||
6ce3dd6e ML |
2643 | void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx, |
2644 | struct list_head *list) | |
2645 | { | |
536167d4 | 2646 | int queued = 0; |
632bfb63 | 2647 | int errors = 0; |
536167d4 | 2648 | |
6ce3dd6e | 2649 | while (!list_empty(list)) { |
fd9c40f6 | 2650 | blk_status_t ret; |
6ce3dd6e ML |
2651 | struct request *rq = list_first_entry(list, struct request, |
2652 | queuelist); | |
2653 | ||
2654 | list_del_init(&rq->queuelist); | |
fd9c40f6 BVA |
2655 | ret = blk_mq_request_issue_directly(rq, list_empty(list)); |
2656 | if (ret != BLK_STS_OK) { | |
2657 | if (ret == BLK_STS_RESOURCE || | |
2658 | ret == BLK_STS_DEV_RESOURCE) { | |
01e99aec | 2659 | blk_mq_request_bypass_insert(rq, false, |
c616cbee | 2660 | list_empty(list)); |
fd9c40f6 BVA |
2661 | break; |
2662 | } | |
2663 | blk_mq_end_request(rq, ret); | |
632bfb63 | 2664 | errors++; |
536167d4 KB |
2665 | } else |
2666 | queued++; | |
6ce3dd6e | 2667 | } |
d666ba98 JA |
2668 | |
2669 | /* | |
2670 | * If we didn't flush the entire list, we could have told | |
2671 | * the driver there was more coming, but that turned out to | |
2672 | * be a lie. | |
2673 | */ | |
632bfb63 | 2674 | if ((!list_empty(list) || errors) && |
2675 | hctx->queue->mq_ops->commit_rqs && queued) | |
d666ba98 | 2676 | hctx->queue->mq_ops->commit_rqs(hctx); |
6ce3dd6e ML |
2677 | } |
2678 | ||
7f2a6a69 | 2679 | /* |
ba0ffdd8 | 2680 | * Allow 2x BLK_MAX_REQUEST_COUNT requests on plug queue for multiple |
7f2a6a69 SL |
2681 | * queues. This is important for md arrays to benefit from merging |
2682 | * requests. | |
2683 | */ | |
2684 | static inline unsigned short blk_plug_max_rq_count(struct blk_plug *plug) | |
2685 | { | |
2686 | if (plug->multiple_queues) | |
ba0ffdd8 | 2687 | return BLK_MAX_REQUEST_COUNT * 2; |
7f2a6a69 SL |
2688 | return BLK_MAX_REQUEST_COUNT; |
2689 | } | |
2690 | ||
1e9c2303 CH |
2691 | static void blk_add_rq_to_plug(struct blk_plug *plug, struct request *rq) |
2692 | { | |
2693 | struct request *last = rq_list_peek(&plug->mq_list); | |
2694 | ||
2695 | if (!plug->rq_count) { | |
2696 | trace_block_plug(rq->q); | |
2697 | } else if (plug->rq_count >= blk_plug_max_rq_count(plug) || | |
2698 | (!blk_queue_nomerges(rq->q) && | |
2699 | blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) { | |
2700 | blk_mq_flush_plug_list(plug, false); | |
2701 | trace_block_plug(rq->q); | |
2702 | } | |
2703 | ||
2704 | if (!plug->multiple_queues && last && last->q != rq->q) | |
2705 | plug->multiple_queues = true; | |
2706 | if (!plug->has_elevator && (rq->rq_flags & RQF_ELV)) | |
2707 | plug->has_elevator = true; | |
2708 | rq->rq_next = NULL; | |
2709 | rq_list_add(&plug->mq_list, rq); | |
2710 | plug->rq_count++; | |
2711 | } | |
2712 | ||
b131f201 | 2713 | static bool blk_mq_attempt_bio_merge(struct request_queue *q, |
0c5bcc92 | 2714 | struct bio *bio, unsigned int nr_segs) |
900e0807 JA |
2715 | { |
2716 | if (!blk_queue_nomerges(q) && bio_mergeable(bio)) { | |
0c5bcc92 | 2717 | if (blk_attempt_plug_merge(q, bio, nr_segs)) |
900e0807 JA |
2718 | return true; |
2719 | if (blk_mq_sched_bio_merge(q, bio, nr_segs)) | |
2720 | return true; | |
2721 | } | |
2722 | return false; | |
2723 | } | |
2724 | ||
71539717 JA |
2725 | static struct request *blk_mq_get_new_requests(struct request_queue *q, |
2726 | struct blk_plug *plug, | |
9d497e29 | 2727 | struct bio *bio) |
71539717 JA |
2728 | { |
2729 | struct blk_mq_alloc_data data = { | |
2730 | .q = q, | |
2731 | .nr_tags = 1, | |
9d497e29 | 2732 | .cmd_flags = bio->bi_opf, |
71539717 JA |
2733 | }; |
2734 | struct request *rq; | |
2735 | ||
5b13bc8a | 2736 | if (unlikely(bio_queue_enter(bio))) |
b637108a | 2737 | return NULL; |
900e0807 | 2738 | |
71539717 JA |
2739 | if (plug) { |
2740 | data.nr_tags = plug->nr_ios; | |
2741 | plug->nr_ios = 1; | |
2742 | data.cached_rq = &plug->cached_rq; | |
2743 | } | |
2744 | ||
2745 | rq = __blk_mq_alloc_requests(&data); | |
373b5416 JA |
2746 | if (rq) |
2747 | return rq; | |
71539717 JA |
2748 | rq_qos_cleanup(q, bio); |
2749 | if (bio->bi_opf & REQ_NOWAIT) | |
2750 | bio_wouldblock_error(bio); | |
5b13bc8a | 2751 | blk_queue_exit(q); |
71539717 JA |
2752 | return NULL; |
2753 | } | |
2754 | ||
5b13bc8a | 2755 | static inline struct request *blk_mq_get_cached_request(struct request_queue *q, |
9d497e29 | 2756 | struct blk_plug *plug, struct bio *bio) |
71539717 | 2757 | { |
b637108a | 2758 | struct request *rq; |
b637108a | 2759 | |
5b13bc8a CH |
2760 | if (!plug) |
2761 | return NULL; | |
2762 | rq = rq_list_peek(&plug->cached_rq); | |
2763 | if (!rq || rq->q != q) | |
2764 | return NULL; | |
71539717 | 2765 | |
9d497e29 | 2766 | if (blk_mq_get_hctx_type(bio->bi_opf) != rq->mq_hctx->type) |
5b13bc8a | 2767 | return NULL; |
9d497e29 | 2768 | if (op_is_flush(rq->cmd_flags) != op_is_flush(bio->bi_opf)) |
5b13bc8a CH |
2769 | return NULL; |
2770 | ||
9d497e29 | 2771 | rq->cmd_flags = bio->bi_opf; |
5b13bc8a CH |
2772 | plug->cached_rq = rq_list_next(rq); |
2773 | INIT_LIST_HEAD(&rq->queuelist); | |
5b13bc8a | 2774 | return rq; |
71539717 JA |
2775 | } |
2776 | ||
105663f7 | 2777 | /** |
c62b37d9 | 2778 | * blk_mq_submit_bio - Create and send a request to block device. |
105663f7 AA |
2779 | * @bio: Bio pointer. |
2780 | * | |
2781 | * Builds up a request structure from @q and @bio and send to the device. The | |
2782 | * request may not be queued directly to hardware if: | |
2783 | * * This request can be merged with another one | |
2784 | * * We want to place request at plug queue for possible future merging | |
2785 | * * There is an IO scheduler active at this queue | |
2786 | * | |
2787 | * It will not queue the request if there is an error with the bio, or at the | |
2788 | * request creation. | |
105663f7 | 2789 | */ |
3e08773c | 2790 | void blk_mq_submit_bio(struct bio *bio) |
07068d5b | 2791 | { |
ed6cddef | 2792 | struct request_queue *q = bdev_get_queue(bio->bi_bdev); |
5b13bc8a | 2793 | struct blk_plug *plug = blk_mq_plug(q, bio); |
ef295ecf | 2794 | const int is_sync = op_is_sync(bio->bi_opf); |
07068d5b | 2795 | struct request *rq; |
abd45c15 | 2796 | unsigned int nr_segs = 1; |
a892c8d5 | 2797 | blk_status_t ret; |
07068d5b JA |
2798 | |
2799 | blk_queue_bounce(q, &bio); | |
abd45c15 JA |
2800 | if (blk_may_split(q, bio)) |
2801 | __blk_queue_split(q, &bio, &nr_segs); | |
f36ea50c | 2802 | |
e23947bd | 2803 | if (!bio_integrity_prep(bio)) |
900e0807 | 2804 | return; |
87760e5e | 2805 | |
9d497e29 ML |
2806 | if (blk_mq_attempt_bio_merge(q, bio, nr_segs)) |
2807 | return; | |
2808 | ||
2809 | rq_qos_throttle(q, bio); | |
2810 | ||
2811 | rq = blk_mq_get_cached_request(q, plug, bio); | |
5b13bc8a | 2812 | if (!rq) { |
9d497e29 | 2813 | rq = blk_mq_get_new_requests(q, plug, bio); |
5b13bc8a CH |
2814 | if (unlikely(!rq)) |
2815 | return; | |
2816 | } | |
87760e5e | 2817 | |
e8a676d6 | 2818 | trace_block_getrq(bio); |
d6f1dda2 | 2819 | |
c1c80384 | 2820 | rq_qos_track(q, rq, bio); |
07068d5b | 2821 | |
970d168d BVA |
2822 | blk_mq_bio_to_request(rq, bio, nr_segs); |
2823 | ||
a892c8d5 ST |
2824 | ret = blk_crypto_init_request(rq); |
2825 | if (ret != BLK_STS_OK) { | |
2826 | bio->bi_status = ret; | |
2827 | bio_endio(bio); | |
2828 | blk_mq_free_request(rq); | |
3e08773c | 2829 | return; |
a892c8d5 ST |
2830 | } |
2831 | ||
2b504bd4 ML |
2832 | if (op_is_flush(bio->bi_opf)) { |
2833 | blk_insert_flush(rq); | |
d92ca9d8 | 2834 | return; |
2b504bd4 | 2835 | } |
d92ca9d8 | 2836 | |
1e9c2303 | 2837 | if (plug) |
ce5b009c | 2838 | blk_add_rq_to_plug(plug, rq); |
1e9c2303 CH |
2839 | else if ((rq->rq_flags & RQF_ELV) || |
2840 | (rq->mq_hctx->dispatch_busy && | |
2841 | (q->nr_hw_queues == 1 || !is_sync))) | |
a12de1d4 | 2842 | blk_mq_sched_insert_request(rq, false, true, true); |
1e9c2303 | 2843 | else |
bcc330f4 | 2844 | blk_mq_run_dispatch_ops(rq->q, |
2a904d00 | 2845 | blk_mq_try_issue_directly(rq->mq_hctx, rq)); |
320ae51f JA |
2846 | } |
2847 | ||
248c7933 | 2848 | #ifdef CONFIG_BLK_MQ_STACKING |
06c8c691 | 2849 | /** |
a5efda3c | 2850 | * blk_insert_cloned_request - Helper for stacking drivers to submit a request |
a5efda3c | 2851 | * @rq: the request being queued |
06c8c691 | 2852 | */ |
28db4711 | 2853 | blk_status_t blk_insert_cloned_request(struct request *rq) |
06c8c691 | 2854 | { |
28db4711 | 2855 | struct request_queue *q = rq->q; |
06c8c691 | 2856 | unsigned int max_sectors = blk_queue_get_max_sectors(q, req_op(rq)); |
a5efda3c | 2857 | blk_status_t ret; |
06c8c691 CH |
2858 | |
2859 | if (blk_rq_sectors(rq) > max_sectors) { | |
2860 | /* | |
2861 | * SCSI device does not have a good way to return if | |
2862 | * Write Same/Zero is actually supported. If a device rejects | |
2863 | * a non-read/write command (discard, write same,etc.) the | |
2864 | * low-level device driver will set the relevant queue limit to | |
2865 | * 0 to prevent blk-lib from issuing more of the offending | |
2866 | * operations. Commands queued prior to the queue limit being | |
2867 | * reset need to be completed with BLK_STS_NOTSUPP to avoid I/O | |
2868 | * errors being propagated to upper layers. | |
2869 | */ | |
2870 | if (max_sectors == 0) | |
2871 | return BLK_STS_NOTSUPP; | |
2872 | ||
2873 | printk(KERN_ERR "%s: over max size limit. (%u > %u)\n", | |
2874 | __func__, blk_rq_sectors(rq), max_sectors); | |
2875 | return BLK_STS_IOERR; | |
2876 | } | |
2877 | ||
2878 | /* | |
2879 | * The queue settings related to segment counting may differ from the | |
2880 | * original queue. | |
2881 | */ | |
2882 | rq->nr_phys_segments = blk_recalc_rq_segments(rq); | |
2883 | if (rq->nr_phys_segments > queue_max_segments(q)) { | |
2884 | printk(KERN_ERR "%s: over max segments limit. (%hu > %hu)\n", | |
2885 | __func__, rq->nr_phys_segments, queue_max_segments(q)); | |
2886 | return BLK_STS_IOERR; | |
2887 | } | |
2888 | ||
28db4711 | 2889 | if (q->disk && should_fail_request(q->disk->part0, blk_rq_bytes(rq))) |
06c8c691 CH |
2890 | return BLK_STS_IOERR; |
2891 | ||
2892 | if (blk_crypto_insert_cloned_request(rq)) | |
2893 | return BLK_STS_IOERR; | |
2894 | ||
2895 | blk_account_io_start(rq); | |
2896 | ||
2897 | /* | |
2898 | * Since we have a scheduler attached on the top device, | |
2899 | * bypass a potential scheduler on the bottom device for | |
2900 | * insert. | |
2901 | */ | |
28db4711 | 2902 | blk_mq_run_dispatch_ops(q, |
4cafe86c | 2903 | ret = blk_mq_request_issue_directly(rq, true)); |
592ee119 YK |
2904 | if (ret) |
2905 | blk_account_io_done(rq, ktime_get_ns()); | |
4cafe86c | 2906 | return ret; |
06c8c691 CH |
2907 | } |
2908 | EXPORT_SYMBOL_GPL(blk_insert_cloned_request); | |
2909 | ||
2910 | /** | |
2911 | * blk_rq_unprep_clone - Helper function to free all bios in a cloned request | |
2912 | * @rq: the clone request to be cleaned up | |
2913 | * | |
2914 | * Description: | |
2915 | * Free all bios in @rq for a cloned request. | |
2916 | */ | |
2917 | void blk_rq_unprep_clone(struct request *rq) | |
2918 | { | |
2919 | struct bio *bio; | |
2920 | ||
2921 | while ((bio = rq->bio) != NULL) { | |
2922 | rq->bio = bio->bi_next; | |
2923 | ||
2924 | bio_put(bio); | |
2925 | } | |
2926 | } | |
2927 | EXPORT_SYMBOL_GPL(blk_rq_unprep_clone); | |
2928 | ||
2929 | /** | |
2930 | * blk_rq_prep_clone - Helper function to setup clone request | |
2931 | * @rq: the request to be setup | |
2932 | * @rq_src: original request to be cloned | |
2933 | * @bs: bio_set that bios for clone are allocated from | |
2934 | * @gfp_mask: memory allocation mask for bio | |
2935 | * @bio_ctr: setup function to be called for each clone bio. | |
2936 | * Returns %0 for success, non %0 for failure. | |
2937 | * @data: private data to be passed to @bio_ctr | |
2938 | * | |
2939 | * Description: | |
2940 | * Clones bios in @rq_src to @rq, and copies attributes of @rq_src to @rq. | |
2941 | * Also, pages which the original bios are pointing to are not copied | |
2942 | * and the cloned bios just point same pages. | |
2943 | * So cloned bios must be completed before original bios, which means | |
2944 | * the caller must complete @rq before @rq_src. | |
2945 | */ | |
2946 | int blk_rq_prep_clone(struct request *rq, struct request *rq_src, | |
2947 | struct bio_set *bs, gfp_t gfp_mask, | |
2948 | int (*bio_ctr)(struct bio *, struct bio *, void *), | |
2949 | void *data) | |
2950 | { | |
2951 | struct bio *bio, *bio_src; | |
2952 | ||
2953 | if (!bs) | |
2954 | bs = &fs_bio_set; | |
2955 | ||
2956 | __rq_for_each_bio(bio_src, rq_src) { | |
abfc426d CH |
2957 | bio = bio_alloc_clone(rq->q->disk->part0, bio_src, gfp_mask, |
2958 | bs); | |
06c8c691 CH |
2959 | if (!bio) |
2960 | goto free_and_out; | |
2961 | ||
2962 | if (bio_ctr && bio_ctr(bio, bio_src, data)) | |
2963 | goto free_and_out; | |
2964 | ||
2965 | if (rq->bio) { | |
2966 | rq->biotail->bi_next = bio; | |
2967 | rq->biotail = bio; | |
2968 | } else { | |
2969 | rq->bio = rq->biotail = bio; | |
2970 | } | |
2971 | bio = NULL; | |
2972 | } | |
2973 | ||
2974 | /* Copy attributes of the original request to the clone request. */ | |
2975 | rq->__sector = blk_rq_pos(rq_src); | |
2976 | rq->__data_len = blk_rq_bytes(rq_src); | |
2977 | if (rq_src->rq_flags & RQF_SPECIAL_PAYLOAD) { | |
2978 | rq->rq_flags |= RQF_SPECIAL_PAYLOAD; | |
2979 | rq->special_vec = rq_src->special_vec; | |
2980 | } | |
2981 | rq->nr_phys_segments = rq_src->nr_phys_segments; | |
2982 | rq->ioprio = rq_src->ioprio; | |
2983 | ||
2984 | if (rq->bio && blk_crypto_rq_bio_prep(rq, rq->bio, gfp_mask) < 0) | |
2985 | goto free_and_out; | |
2986 | ||
2987 | return 0; | |
2988 | ||
2989 | free_and_out: | |
2990 | if (bio) | |
2991 | bio_put(bio); | |
2992 | blk_rq_unprep_clone(rq); | |
2993 | ||
2994 | return -ENOMEM; | |
2995 | } | |
2996 | EXPORT_SYMBOL_GPL(blk_rq_prep_clone); | |
248c7933 | 2997 | #endif /* CONFIG_BLK_MQ_STACKING */ |
06c8c691 | 2998 | |
f2b8f3ce CH |
2999 | /* |
3000 | * Steal bios from a request and add them to a bio list. | |
3001 | * The request must not have been partially completed before. | |
3002 | */ | |
3003 | void blk_steal_bios(struct bio_list *list, struct request *rq) | |
3004 | { | |
3005 | if (rq->bio) { | |
3006 | if (list->tail) | |
3007 | list->tail->bi_next = rq->bio; | |
3008 | else | |
3009 | list->head = rq->bio; | |
3010 | list->tail = rq->biotail; | |
3011 | ||
3012 | rq->bio = NULL; | |
3013 | rq->biotail = NULL; | |
3014 | } | |
3015 | ||
3016 | rq->__data_len = 0; | |
3017 | } | |
3018 | EXPORT_SYMBOL_GPL(blk_steal_bios); | |
3019 | ||
bd63141d ML |
3020 | static size_t order_to_size(unsigned int order) |
3021 | { | |
3022 | return (size_t)PAGE_SIZE << order; | |
3023 | } | |
3024 | ||
3025 | /* called before freeing request pool in @tags */ | |
f32e4eaf JG |
3026 | static void blk_mq_clear_rq_mapping(struct blk_mq_tags *drv_tags, |
3027 | struct blk_mq_tags *tags) | |
bd63141d | 3028 | { |
bd63141d ML |
3029 | struct page *page; |
3030 | unsigned long flags; | |
3031 | ||
4f245d5b JG |
3032 | /* There is no need to clear a driver tags own mapping */ |
3033 | if (drv_tags == tags) | |
3034 | return; | |
3035 | ||
bd63141d ML |
3036 | list_for_each_entry(page, &tags->page_list, lru) { |
3037 | unsigned long start = (unsigned long)page_address(page); | |
3038 | unsigned long end = start + order_to_size(page->private); | |
3039 | int i; | |
3040 | ||
f32e4eaf | 3041 | for (i = 0; i < drv_tags->nr_tags; i++) { |
bd63141d ML |
3042 | struct request *rq = drv_tags->rqs[i]; |
3043 | unsigned long rq_addr = (unsigned long)rq; | |
3044 | ||
3045 | if (rq_addr >= start && rq_addr < end) { | |
0a467d0f | 3046 | WARN_ON_ONCE(req_ref_read(rq) != 0); |
bd63141d ML |
3047 | cmpxchg(&drv_tags->rqs[i], rq, NULL); |
3048 | } | |
3049 | } | |
3050 | } | |
3051 | ||
3052 | /* | |
3053 | * Wait until all pending iteration is done. | |
3054 | * | |
3055 | * Request reference is cleared and it is guaranteed to be observed | |
3056 | * after the ->lock is released. | |
3057 | */ | |
3058 | spin_lock_irqsave(&drv_tags->lock, flags); | |
3059 | spin_unlock_irqrestore(&drv_tags->lock, flags); | |
3060 | } | |
3061 | ||
cc71a6f4 JA |
3062 | void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags, |
3063 | unsigned int hctx_idx) | |
95363efd | 3064 | { |
f32e4eaf | 3065 | struct blk_mq_tags *drv_tags; |
e9b267d9 | 3066 | struct page *page; |
320ae51f | 3067 | |
079a2e3e JG |
3068 | if (blk_mq_is_shared_tags(set->flags)) |
3069 | drv_tags = set->shared_tags; | |
e155b0c2 JG |
3070 | else |
3071 | drv_tags = set->tags[hctx_idx]; | |
f32e4eaf | 3072 | |
65de57bb | 3073 | if (tags->static_rqs && set->ops->exit_request) { |
e9b267d9 | 3074 | int i; |
320ae51f | 3075 | |
24d2f903 | 3076 | for (i = 0; i < tags->nr_tags; i++) { |
2af8cbe3 JA |
3077 | struct request *rq = tags->static_rqs[i]; |
3078 | ||
3079 | if (!rq) | |
e9b267d9 | 3080 | continue; |
d6296d39 | 3081 | set->ops->exit_request(set, rq, hctx_idx); |
2af8cbe3 | 3082 | tags->static_rqs[i] = NULL; |
e9b267d9 | 3083 | } |
320ae51f | 3084 | } |
320ae51f | 3085 | |
f32e4eaf | 3086 | blk_mq_clear_rq_mapping(drv_tags, tags); |
bd63141d | 3087 | |
24d2f903 CH |
3088 | while (!list_empty(&tags->page_list)) { |
3089 | page = list_first_entry(&tags->page_list, struct page, lru); | |
6753471c | 3090 | list_del_init(&page->lru); |
f75782e4 CM |
3091 | /* |
3092 | * Remove kmemleak object previously allocated in | |
273938bf | 3093 | * blk_mq_alloc_rqs(). |
f75782e4 CM |
3094 | */ |
3095 | kmemleak_free(page_address(page)); | |
320ae51f JA |
3096 | __free_pages(page, page->private); |
3097 | } | |
cc71a6f4 | 3098 | } |
320ae51f | 3099 | |
e155b0c2 | 3100 | void blk_mq_free_rq_map(struct blk_mq_tags *tags) |
cc71a6f4 | 3101 | { |
24d2f903 | 3102 | kfree(tags->rqs); |
cc71a6f4 | 3103 | tags->rqs = NULL; |
2af8cbe3 JA |
3104 | kfree(tags->static_rqs); |
3105 | tags->static_rqs = NULL; | |
320ae51f | 3106 | |
e155b0c2 | 3107 | blk_mq_free_tags(tags); |
320ae51f JA |
3108 | } |
3109 | ||
4d805131 ML |
3110 | static enum hctx_type hctx_idx_to_type(struct blk_mq_tag_set *set, |
3111 | unsigned int hctx_idx) | |
3112 | { | |
3113 | int i; | |
3114 | ||
3115 | for (i = 0; i < set->nr_maps; i++) { | |
3116 | unsigned int start = set->map[i].queue_offset; | |
3117 | unsigned int end = start + set->map[i].nr_queues; | |
3118 | ||
3119 | if (hctx_idx >= start && hctx_idx < end) | |
3120 | break; | |
3121 | } | |
3122 | ||
3123 | if (i >= set->nr_maps) | |
3124 | i = HCTX_TYPE_DEFAULT; | |
3125 | ||
3126 | return i; | |
3127 | } | |
3128 | ||
3129 | static int blk_mq_get_hctx_node(struct blk_mq_tag_set *set, | |
3130 | unsigned int hctx_idx) | |
3131 | { | |
3132 | enum hctx_type type = hctx_idx_to_type(set, hctx_idx); | |
3133 | ||
3134 | return blk_mq_hw_queue_to_node(&set->map[type], hctx_idx); | |
3135 | } | |
3136 | ||
63064be1 JG |
3137 | static struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, |
3138 | unsigned int hctx_idx, | |
3139 | unsigned int nr_tags, | |
e155b0c2 | 3140 | unsigned int reserved_tags) |
320ae51f | 3141 | { |
4d805131 | 3142 | int node = blk_mq_get_hctx_node(set, hctx_idx); |
24d2f903 | 3143 | struct blk_mq_tags *tags; |
320ae51f | 3144 | |
59f082e4 SL |
3145 | if (node == NUMA_NO_NODE) |
3146 | node = set->numa_node; | |
3147 | ||
e155b0c2 JG |
3148 | tags = blk_mq_init_tags(nr_tags, reserved_tags, node, |
3149 | BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags)); | |
24d2f903 CH |
3150 | if (!tags) |
3151 | return NULL; | |
320ae51f | 3152 | |
590b5b7d | 3153 | tags->rqs = kcalloc_node(nr_tags, sizeof(struct request *), |
36e1f3d1 | 3154 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, |
59f082e4 | 3155 | node); |
24d2f903 | 3156 | if (!tags->rqs) { |
e155b0c2 | 3157 | blk_mq_free_tags(tags); |
24d2f903 CH |
3158 | return NULL; |
3159 | } | |
320ae51f | 3160 | |
590b5b7d KC |
3161 | tags->static_rqs = kcalloc_node(nr_tags, sizeof(struct request *), |
3162 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, | |
3163 | node); | |
2af8cbe3 JA |
3164 | if (!tags->static_rqs) { |
3165 | kfree(tags->rqs); | |
e155b0c2 | 3166 | blk_mq_free_tags(tags); |
2af8cbe3 JA |
3167 | return NULL; |
3168 | } | |
3169 | ||
cc71a6f4 JA |
3170 | return tags; |
3171 | } | |
3172 | ||
1d9bd516 TH |
3173 | static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq, |
3174 | unsigned int hctx_idx, int node) | |
3175 | { | |
3176 | int ret; | |
3177 | ||
3178 | if (set->ops->init_request) { | |
3179 | ret = set->ops->init_request(set, rq, hctx_idx, node); | |
3180 | if (ret) | |
3181 | return ret; | |
3182 | } | |
3183 | ||
12f5b931 | 3184 | WRITE_ONCE(rq->state, MQ_RQ_IDLE); |
1d9bd516 TH |
3185 | return 0; |
3186 | } | |
3187 | ||
63064be1 JG |
3188 | static int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, |
3189 | struct blk_mq_tags *tags, | |
3190 | unsigned int hctx_idx, unsigned int depth) | |
cc71a6f4 JA |
3191 | { |
3192 | unsigned int i, j, entries_per_page, max_order = 4; | |
4d805131 | 3193 | int node = blk_mq_get_hctx_node(set, hctx_idx); |
cc71a6f4 | 3194 | size_t rq_size, left; |
59f082e4 | 3195 | |
59f082e4 SL |
3196 | if (node == NUMA_NO_NODE) |
3197 | node = set->numa_node; | |
cc71a6f4 JA |
3198 | |
3199 | INIT_LIST_HEAD(&tags->page_list); | |
3200 | ||
320ae51f JA |
3201 | /* |
3202 | * rq_size is the size of the request plus driver payload, rounded | |
3203 | * to the cacheline size | |
3204 | */ | |
24d2f903 | 3205 | rq_size = round_up(sizeof(struct request) + set->cmd_size, |
320ae51f | 3206 | cache_line_size()); |
cc71a6f4 | 3207 | left = rq_size * depth; |
320ae51f | 3208 | |
cc71a6f4 | 3209 | for (i = 0; i < depth; ) { |
320ae51f JA |
3210 | int this_order = max_order; |
3211 | struct page *page; | |
3212 | int to_do; | |
3213 | void *p; | |
3214 | ||
b3a834b1 | 3215 | while (this_order && left < order_to_size(this_order - 1)) |
320ae51f JA |
3216 | this_order--; |
3217 | ||
3218 | do { | |
59f082e4 | 3219 | page = alloc_pages_node(node, |
36e1f3d1 | 3220 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO, |
a5164405 | 3221 | this_order); |
320ae51f JA |
3222 | if (page) |
3223 | break; | |
3224 | if (!this_order--) | |
3225 | break; | |
3226 | if (order_to_size(this_order) < rq_size) | |
3227 | break; | |
3228 | } while (1); | |
3229 | ||
3230 | if (!page) | |
24d2f903 | 3231 | goto fail; |
320ae51f JA |
3232 | |
3233 | page->private = this_order; | |
24d2f903 | 3234 | list_add_tail(&page->lru, &tags->page_list); |
320ae51f JA |
3235 | |
3236 | p = page_address(page); | |
f75782e4 CM |
3237 | /* |
3238 | * Allow kmemleak to scan these pages as they contain pointers | |
3239 | * to additional allocations like via ops->init_request(). | |
3240 | */ | |
36e1f3d1 | 3241 | kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO); |
320ae51f | 3242 | entries_per_page = order_to_size(this_order) / rq_size; |
cc71a6f4 | 3243 | to_do = min(entries_per_page, depth - i); |
320ae51f JA |
3244 | left -= to_do * rq_size; |
3245 | for (j = 0; j < to_do; j++) { | |
2af8cbe3 JA |
3246 | struct request *rq = p; |
3247 | ||
3248 | tags->static_rqs[i] = rq; | |
1d9bd516 TH |
3249 | if (blk_mq_init_request(set, rq, hctx_idx, node)) { |
3250 | tags->static_rqs[i] = NULL; | |
3251 | goto fail; | |
e9b267d9 CH |
3252 | } |
3253 | ||
320ae51f JA |
3254 | p += rq_size; |
3255 | i++; | |
3256 | } | |
3257 | } | |
cc71a6f4 | 3258 | return 0; |
320ae51f | 3259 | |
24d2f903 | 3260 | fail: |
cc71a6f4 JA |
3261 | blk_mq_free_rqs(set, tags, hctx_idx); |
3262 | return -ENOMEM; | |
320ae51f JA |
3263 | } |
3264 | ||
bf0beec0 ML |
3265 | struct rq_iter_data { |
3266 | struct blk_mq_hw_ctx *hctx; | |
3267 | bool has_rq; | |
3268 | }; | |
3269 | ||
3270 | static bool blk_mq_has_request(struct request *rq, void *data, bool reserved) | |
3271 | { | |
3272 | struct rq_iter_data *iter_data = data; | |
3273 | ||
3274 | if (rq->mq_hctx != iter_data->hctx) | |
3275 | return true; | |
3276 | iter_data->has_rq = true; | |
3277 | return false; | |
3278 | } | |
3279 | ||
3280 | static bool blk_mq_hctx_has_requests(struct blk_mq_hw_ctx *hctx) | |
3281 | { | |
3282 | struct blk_mq_tags *tags = hctx->sched_tags ? | |
3283 | hctx->sched_tags : hctx->tags; | |
3284 | struct rq_iter_data data = { | |
3285 | .hctx = hctx, | |
3286 | }; | |
3287 | ||
3288 | blk_mq_all_tag_iter(tags, blk_mq_has_request, &data); | |
3289 | return data.has_rq; | |
3290 | } | |
3291 | ||
3292 | static inline bool blk_mq_last_cpu_in_hctx(unsigned int cpu, | |
3293 | struct blk_mq_hw_ctx *hctx) | |
3294 | { | |
9b51d9d8 | 3295 | if (cpumask_first_and(hctx->cpumask, cpu_online_mask) != cpu) |
bf0beec0 ML |
3296 | return false; |
3297 | if (cpumask_next_and(cpu, hctx->cpumask, cpu_online_mask) < nr_cpu_ids) | |
3298 | return false; | |
3299 | return true; | |
3300 | } | |
3301 | ||
3302 | static int blk_mq_hctx_notify_offline(unsigned int cpu, struct hlist_node *node) | |
3303 | { | |
3304 | struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node, | |
3305 | struct blk_mq_hw_ctx, cpuhp_online); | |
3306 | ||
3307 | if (!cpumask_test_cpu(cpu, hctx->cpumask) || | |
3308 | !blk_mq_last_cpu_in_hctx(cpu, hctx)) | |
3309 | return 0; | |
3310 | ||
3311 | /* | |
3312 | * Prevent new request from being allocated on the current hctx. | |
3313 | * | |
3314 | * The smp_mb__after_atomic() Pairs with the implied barrier in | |
3315 | * test_and_set_bit_lock in sbitmap_get(). Ensures the inactive flag is | |
3316 | * seen once we return from the tag allocator. | |
3317 | */ | |
3318 | set_bit(BLK_MQ_S_INACTIVE, &hctx->state); | |
3319 | smp_mb__after_atomic(); | |
3320 | ||
3321 | /* | |
3322 | * Try to grab a reference to the queue and wait for any outstanding | |
3323 | * requests. If we could not grab a reference the queue has been | |
3324 | * frozen and there are no requests. | |
3325 | */ | |
3326 | if (percpu_ref_tryget(&hctx->queue->q_usage_counter)) { | |
3327 | while (blk_mq_hctx_has_requests(hctx)) | |
3328 | msleep(5); | |
3329 | percpu_ref_put(&hctx->queue->q_usage_counter); | |
3330 | } | |
3331 | ||
3332 | return 0; | |
3333 | } | |
3334 | ||
3335 | static int blk_mq_hctx_notify_online(unsigned int cpu, struct hlist_node *node) | |
3336 | { | |
3337 | struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node, | |
3338 | struct blk_mq_hw_ctx, cpuhp_online); | |
3339 | ||
3340 | if (cpumask_test_cpu(cpu, hctx->cpumask)) | |
3341 | clear_bit(BLK_MQ_S_INACTIVE, &hctx->state); | |
3342 | return 0; | |
3343 | } | |
3344 | ||
e57690fe JA |
3345 | /* |
3346 | * 'cpu' is going away. splice any existing rq_list entries from this | |
3347 | * software queue to the hw queue dispatch list, and ensure that it | |
3348 | * gets run. | |
3349 | */ | |
9467f859 | 3350 | static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node) |
484b4061 | 3351 | { |
9467f859 | 3352 | struct blk_mq_hw_ctx *hctx; |
484b4061 JA |
3353 | struct blk_mq_ctx *ctx; |
3354 | LIST_HEAD(tmp); | |
c16d6b5a | 3355 | enum hctx_type type; |
484b4061 | 3356 | |
9467f859 | 3357 | hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead); |
bf0beec0 ML |
3358 | if (!cpumask_test_cpu(cpu, hctx->cpumask)) |
3359 | return 0; | |
3360 | ||
e57690fe | 3361 | ctx = __blk_mq_get_ctx(hctx->queue, cpu); |
c16d6b5a | 3362 | type = hctx->type; |
484b4061 JA |
3363 | |
3364 | spin_lock(&ctx->lock); | |
c16d6b5a ML |
3365 | if (!list_empty(&ctx->rq_lists[type])) { |
3366 | list_splice_init(&ctx->rq_lists[type], &tmp); | |
484b4061 JA |
3367 | blk_mq_hctx_clear_pending(hctx, ctx); |
3368 | } | |
3369 | spin_unlock(&ctx->lock); | |
3370 | ||
3371 | if (list_empty(&tmp)) | |
9467f859 | 3372 | return 0; |
484b4061 | 3373 | |
e57690fe JA |
3374 | spin_lock(&hctx->lock); |
3375 | list_splice_tail_init(&tmp, &hctx->dispatch); | |
3376 | spin_unlock(&hctx->lock); | |
484b4061 JA |
3377 | |
3378 | blk_mq_run_hw_queue(hctx, true); | |
9467f859 | 3379 | return 0; |
484b4061 JA |
3380 | } |
3381 | ||
9467f859 | 3382 | static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx) |
484b4061 | 3383 | { |
bf0beec0 ML |
3384 | if (!(hctx->flags & BLK_MQ_F_STACKING)) |
3385 | cpuhp_state_remove_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE, | |
3386 | &hctx->cpuhp_online); | |
9467f859 TG |
3387 | cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD, |
3388 | &hctx->cpuhp_dead); | |
484b4061 JA |
3389 | } |
3390 | ||
364b6181 ML |
3391 | /* |
3392 | * Before freeing hw queue, clearing the flush request reference in | |
3393 | * tags->rqs[] for avoiding potential UAF. | |
3394 | */ | |
3395 | static void blk_mq_clear_flush_rq_mapping(struct blk_mq_tags *tags, | |
3396 | unsigned int queue_depth, struct request *flush_rq) | |
3397 | { | |
3398 | int i; | |
3399 | unsigned long flags; | |
3400 | ||
3401 | /* The hw queue may not be mapped yet */ | |
3402 | if (!tags) | |
3403 | return; | |
3404 | ||
0a467d0f | 3405 | WARN_ON_ONCE(req_ref_read(flush_rq) != 0); |
364b6181 ML |
3406 | |
3407 | for (i = 0; i < queue_depth; i++) | |
3408 | cmpxchg(&tags->rqs[i], flush_rq, NULL); | |
3409 | ||
3410 | /* | |
3411 | * Wait until all pending iteration is done. | |
3412 | * | |
3413 | * Request reference is cleared and it is guaranteed to be observed | |
3414 | * after the ->lock is released. | |
3415 | */ | |
3416 | spin_lock_irqsave(&tags->lock, flags); | |
3417 | spin_unlock_irqrestore(&tags->lock, flags); | |
3418 | } | |
3419 | ||
c3b4afca | 3420 | /* hctx->ctxs will be freed in queue's release handler */ |
08e98fc6 ML |
3421 | static void blk_mq_exit_hctx(struct request_queue *q, |
3422 | struct blk_mq_tag_set *set, | |
3423 | struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx) | |
3424 | { | |
364b6181 ML |
3425 | struct request *flush_rq = hctx->fq->flush_rq; |
3426 | ||
8ab0b7dc ML |
3427 | if (blk_mq_hw_queue_mapped(hctx)) |
3428 | blk_mq_tag_idle(hctx); | |
08e98fc6 | 3429 | |
364b6181 ML |
3430 | blk_mq_clear_flush_rq_mapping(set->tags[hctx_idx], |
3431 | set->queue_depth, flush_rq); | |
f70ced09 | 3432 | if (set->ops->exit_request) |
364b6181 | 3433 | set->ops->exit_request(set, flush_rq, hctx_idx); |
f70ced09 | 3434 | |
08e98fc6 ML |
3435 | if (set->ops->exit_hctx) |
3436 | set->ops->exit_hctx(hctx, hctx_idx); | |
3437 | ||
9467f859 | 3438 | blk_mq_remove_cpuhp(hctx); |
2f8f1336 ML |
3439 | |
3440 | spin_lock(&q->unused_hctx_lock); | |
3441 | list_add(&hctx->hctx_list, &q->unused_hctx_list); | |
3442 | spin_unlock(&q->unused_hctx_lock); | |
08e98fc6 ML |
3443 | } |
3444 | ||
624dbe47 ML |
3445 | static void blk_mq_exit_hw_queues(struct request_queue *q, |
3446 | struct blk_mq_tag_set *set, int nr_queue) | |
3447 | { | |
3448 | struct blk_mq_hw_ctx *hctx; | |
3449 | unsigned int i; | |
3450 | ||
3451 | queue_for_each_hw_ctx(q, hctx, i) { | |
3452 | if (i == nr_queue) | |
3453 | break; | |
477e19de | 3454 | blk_mq_debugfs_unregister_hctx(hctx); |
08e98fc6 | 3455 | blk_mq_exit_hctx(q, set, hctx, i); |
624dbe47 | 3456 | } |
624dbe47 ML |
3457 | } |
3458 | ||
08e98fc6 ML |
3459 | static int blk_mq_init_hctx(struct request_queue *q, |
3460 | struct blk_mq_tag_set *set, | |
3461 | struct blk_mq_hw_ctx *hctx, unsigned hctx_idx) | |
320ae51f | 3462 | { |
7c6c5b7c ML |
3463 | hctx->queue_num = hctx_idx; |
3464 | ||
bf0beec0 ML |
3465 | if (!(hctx->flags & BLK_MQ_F_STACKING)) |
3466 | cpuhp_state_add_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE, | |
3467 | &hctx->cpuhp_online); | |
7c6c5b7c ML |
3468 | cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead); |
3469 | ||
3470 | hctx->tags = set->tags[hctx_idx]; | |
3471 | ||
3472 | if (set->ops->init_hctx && | |
3473 | set->ops->init_hctx(hctx, set->driver_data, hctx_idx)) | |
3474 | goto unregister_cpu_notifier; | |
08e98fc6 | 3475 | |
7c6c5b7c ML |
3476 | if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx, |
3477 | hctx->numa_node)) | |
3478 | goto exit_hctx; | |
3479 | return 0; | |
3480 | ||
3481 | exit_hctx: | |
3482 | if (set->ops->exit_hctx) | |
3483 | set->ops->exit_hctx(hctx, hctx_idx); | |
3484 | unregister_cpu_notifier: | |
3485 | blk_mq_remove_cpuhp(hctx); | |
3486 | return -1; | |
3487 | } | |
3488 | ||
3489 | static struct blk_mq_hw_ctx * | |
3490 | blk_mq_alloc_hctx(struct request_queue *q, struct blk_mq_tag_set *set, | |
3491 | int node) | |
3492 | { | |
3493 | struct blk_mq_hw_ctx *hctx; | |
3494 | gfp_t gfp = GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY; | |
3495 | ||
704b914f | 3496 | hctx = kzalloc_node(sizeof(struct blk_mq_hw_ctx), gfp, node); |
7c6c5b7c ML |
3497 | if (!hctx) |
3498 | goto fail_alloc_hctx; | |
3499 | ||
3500 | if (!zalloc_cpumask_var_node(&hctx->cpumask, gfp, node)) | |
3501 | goto free_hctx; | |
3502 | ||
3503 | atomic_set(&hctx->nr_active, 0); | |
08e98fc6 | 3504 | if (node == NUMA_NO_NODE) |
7c6c5b7c ML |
3505 | node = set->numa_node; |
3506 | hctx->numa_node = node; | |
08e98fc6 | 3507 | |
9f993737 | 3508 | INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn); |
08e98fc6 ML |
3509 | spin_lock_init(&hctx->lock); |
3510 | INIT_LIST_HEAD(&hctx->dispatch); | |
3511 | hctx->queue = q; | |
51db1c37 | 3512 | hctx->flags = set->flags & ~BLK_MQ_F_TAG_QUEUE_SHARED; |
08e98fc6 | 3513 | |
2f8f1336 ML |
3514 | INIT_LIST_HEAD(&hctx->hctx_list); |
3515 | ||
320ae51f | 3516 | /* |
08e98fc6 ML |
3517 | * Allocate space for all possible cpus to avoid allocation at |
3518 | * runtime | |
320ae51f | 3519 | */ |
d904bfa7 | 3520 | hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *), |
7c6c5b7c | 3521 | gfp, node); |
08e98fc6 | 3522 | if (!hctx->ctxs) |
7c6c5b7c | 3523 | goto free_cpumask; |
320ae51f | 3524 | |
5b202853 | 3525 | if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8), |
c548e62b | 3526 | gfp, node, false, false)) |
08e98fc6 | 3527 | goto free_ctxs; |
08e98fc6 | 3528 | hctx->nr_ctx = 0; |
320ae51f | 3529 | |
5815839b | 3530 | spin_lock_init(&hctx->dispatch_wait_lock); |
eb619fdb JA |
3531 | init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake); |
3532 | INIT_LIST_HEAD(&hctx->dispatch_wait.entry); | |
3533 | ||
754a1572 | 3534 | hctx->fq = blk_alloc_flush_queue(hctx->numa_node, set->cmd_size, gfp); |
f70ced09 | 3535 | if (!hctx->fq) |
7c6c5b7c | 3536 | goto free_bitmap; |
320ae51f | 3537 | |
7c6c5b7c | 3538 | blk_mq_hctx_kobj_init(hctx); |
6a83e74d | 3539 | |
7c6c5b7c | 3540 | return hctx; |
320ae51f | 3541 | |
08e98fc6 | 3542 | free_bitmap: |
88459642 | 3543 | sbitmap_free(&hctx->ctx_map); |
08e98fc6 ML |
3544 | free_ctxs: |
3545 | kfree(hctx->ctxs); | |
7c6c5b7c ML |
3546 | free_cpumask: |
3547 | free_cpumask_var(hctx->cpumask); | |
3548 | free_hctx: | |
3549 | kfree(hctx); | |
3550 | fail_alloc_hctx: | |
3551 | return NULL; | |
08e98fc6 | 3552 | } |
320ae51f | 3553 | |
320ae51f JA |
3554 | static void blk_mq_init_cpu_queues(struct request_queue *q, |
3555 | unsigned int nr_hw_queues) | |
3556 | { | |
b3c661b1 JA |
3557 | struct blk_mq_tag_set *set = q->tag_set; |
3558 | unsigned int i, j; | |
320ae51f JA |
3559 | |
3560 | for_each_possible_cpu(i) { | |
3561 | struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i); | |
3562 | struct blk_mq_hw_ctx *hctx; | |
c16d6b5a | 3563 | int k; |
320ae51f | 3564 | |
320ae51f JA |
3565 | __ctx->cpu = i; |
3566 | spin_lock_init(&__ctx->lock); | |
c16d6b5a ML |
3567 | for (k = HCTX_TYPE_DEFAULT; k < HCTX_MAX_TYPES; k++) |
3568 | INIT_LIST_HEAD(&__ctx->rq_lists[k]); | |
3569 | ||
320ae51f JA |
3570 | __ctx->queue = q; |
3571 | ||
320ae51f JA |
3572 | /* |
3573 | * Set local node, IFF we have more than one hw queue. If | |
3574 | * not, we remain on the home node of the device | |
3575 | */ | |
b3c661b1 JA |
3576 | for (j = 0; j < set->nr_maps; j++) { |
3577 | hctx = blk_mq_map_queue_type(q, j, i); | |
3578 | if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE) | |
576e85c5 | 3579 | hctx->numa_node = cpu_to_node(i); |
b3c661b1 | 3580 | } |
320ae51f JA |
3581 | } |
3582 | } | |
3583 | ||
63064be1 JG |
3584 | struct blk_mq_tags *blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set *set, |
3585 | unsigned int hctx_idx, | |
3586 | unsigned int depth) | |
cc71a6f4 | 3587 | { |
63064be1 JG |
3588 | struct blk_mq_tags *tags; |
3589 | int ret; | |
cc71a6f4 | 3590 | |
e155b0c2 | 3591 | tags = blk_mq_alloc_rq_map(set, hctx_idx, depth, set->reserved_tags); |
63064be1 JG |
3592 | if (!tags) |
3593 | return NULL; | |
cc71a6f4 | 3594 | |
63064be1 JG |
3595 | ret = blk_mq_alloc_rqs(set, tags, hctx_idx, depth); |
3596 | if (ret) { | |
e155b0c2 | 3597 | blk_mq_free_rq_map(tags); |
63064be1 JG |
3598 | return NULL; |
3599 | } | |
cc71a6f4 | 3600 | |
63064be1 | 3601 | return tags; |
cc71a6f4 JA |
3602 | } |
3603 | ||
63064be1 JG |
3604 | static bool __blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set *set, |
3605 | int hctx_idx) | |
cc71a6f4 | 3606 | { |
079a2e3e JG |
3607 | if (blk_mq_is_shared_tags(set->flags)) { |
3608 | set->tags[hctx_idx] = set->shared_tags; | |
1c0706a7 | 3609 | |
e155b0c2 | 3610 | return true; |
bd166ef1 | 3611 | } |
e155b0c2 | 3612 | |
63064be1 JG |
3613 | set->tags[hctx_idx] = blk_mq_alloc_map_and_rqs(set, hctx_idx, |
3614 | set->queue_depth); | |
3615 | ||
3616 | return set->tags[hctx_idx]; | |
cc71a6f4 JA |
3617 | } |
3618 | ||
645db34e JG |
3619 | void blk_mq_free_map_and_rqs(struct blk_mq_tag_set *set, |
3620 | struct blk_mq_tags *tags, | |
3621 | unsigned int hctx_idx) | |
cc71a6f4 | 3622 | { |
645db34e JG |
3623 | if (tags) { |
3624 | blk_mq_free_rqs(set, tags, hctx_idx); | |
e155b0c2 | 3625 | blk_mq_free_rq_map(tags); |
bd166ef1 | 3626 | } |
cc71a6f4 JA |
3627 | } |
3628 | ||
e155b0c2 JG |
3629 | static void __blk_mq_free_map_and_rqs(struct blk_mq_tag_set *set, |
3630 | unsigned int hctx_idx) | |
3631 | { | |
079a2e3e | 3632 | if (!blk_mq_is_shared_tags(set->flags)) |
e155b0c2 JG |
3633 | blk_mq_free_map_and_rqs(set, set->tags[hctx_idx], hctx_idx); |
3634 | ||
3635 | set->tags[hctx_idx] = NULL; | |
cc71a6f4 JA |
3636 | } |
3637 | ||
4b855ad3 | 3638 | static void blk_mq_map_swqueue(struct request_queue *q) |
320ae51f | 3639 | { |
b3c661b1 | 3640 | unsigned int i, j, hctx_idx; |
320ae51f JA |
3641 | struct blk_mq_hw_ctx *hctx; |
3642 | struct blk_mq_ctx *ctx; | |
2a34c087 | 3643 | struct blk_mq_tag_set *set = q->tag_set; |
320ae51f JA |
3644 | |
3645 | queue_for_each_hw_ctx(q, hctx, i) { | |
e4043dcf | 3646 | cpumask_clear(hctx->cpumask); |
320ae51f | 3647 | hctx->nr_ctx = 0; |
d416c92c | 3648 | hctx->dispatch_from = NULL; |
320ae51f JA |
3649 | } |
3650 | ||
3651 | /* | |
4b855ad3 | 3652 | * Map software to hardware queues. |
4412efec ML |
3653 | * |
3654 | * If the cpu isn't present, the cpu is mapped to first hctx. | |
320ae51f | 3655 | */ |
20e4d813 | 3656 | for_each_possible_cpu(i) { |
4412efec | 3657 | |
897bb0c7 | 3658 | ctx = per_cpu_ptr(q->queue_ctx, i); |
b3c661b1 | 3659 | for (j = 0; j < set->nr_maps; j++) { |
bb94aea1 JW |
3660 | if (!set->map[j].nr_queues) { |
3661 | ctx->hctxs[j] = blk_mq_map_queue_type(q, | |
3662 | HCTX_TYPE_DEFAULT, i); | |
e5edd5f2 | 3663 | continue; |
bb94aea1 | 3664 | } |
fd689871 ML |
3665 | hctx_idx = set->map[j].mq_map[i]; |
3666 | /* unmapped hw queue can be remapped after CPU topo changed */ | |
3667 | if (!set->tags[hctx_idx] && | |
63064be1 | 3668 | !__blk_mq_alloc_map_and_rqs(set, hctx_idx)) { |
fd689871 ML |
3669 | /* |
3670 | * If tags initialization fail for some hctx, | |
3671 | * that hctx won't be brought online. In this | |
3672 | * case, remap the current ctx to hctx[0] which | |
3673 | * is guaranteed to always have tags allocated | |
3674 | */ | |
3675 | set->map[j].mq_map[i] = 0; | |
3676 | } | |
e5edd5f2 | 3677 | |
b3c661b1 | 3678 | hctx = blk_mq_map_queue_type(q, j, i); |
8ccdf4a3 | 3679 | ctx->hctxs[j] = hctx; |
b3c661b1 JA |
3680 | /* |
3681 | * If the CPU is already set in the mask, then we've | |
3682 | * mapped this one already. This can happen if | |
3683 | * devices share queues across queue maps. | |
3684 | */ | |
3685 | if (cpumask_test_cpu(i, hctx->cpumask)) | |
3686 | continue; | |
3687 | ||
3688 | cpumask_set_cpu(i, hctx->cpumask); | |
3689 | hctx->type = j; | |
3690 | ctx->index_hw[hctx->type] = hctx->nr_ctx; | |
3691 | hctx->ctxs[hctx->nr_ctx++] = ctx; | |
3692 | ||
3693 | /* | |
3694 | * If the nr_ctx type overflows, we have exceeded the | |
3695 | * amount of sw queues we can support. | |
3696 | */ | |
3697 | BUG_ON(!hctx->nr_ctx); | |
3698 | } | |
bb94aea1 JW |
3699 | |
3700 | for (; j < HCTX_MAX_TYPES; j++) | |
3701 | ctx->hctxs[j] = blk_mq_map_queue_type(q, | |
3702 | HCTX_TYPE_DEFAULT, i); | |
320ae51f | 3703 | } |
506e931f JA |
3704 | |
3705 | queue_for_each_hw_ctx(q, hctx, i) { | |
4412efec ML |
3706 | /* |
3707 | * If no software queues are mapped to this hardware queue, | |
3708 | * disable it and free the request entries. | |
3709 | */ | |
3710 | if (!hctx->nr_ctx) { | |
3711 | /* Never unmap queue 0. We need it as a | |
3712 | * fallback in case of a new remap fails | |
3713 | * allocation | |
3714 | */ | |
e155b0c2 JG |
3715 | if (i) |
3716 | __blk_mq_free_map_and_rqs(set, i); | |
4412efec ML |
3717 | |
3718 | hctx->tags = NULL; | |
3719 | continue; | |
3720 | } | |
484b4061 | 3721 | |
2a34c087 ML |
3722 | hctx->tags = set->tags[i]; |
3723 | WARN_ON(!hctx->tags); | |
3724 | ||
889fa31f CY |
3725 | /* |
3726 | * Set the map size to the number of mapped software queues. | |
3727 | * This is more accurate and more efficient than looping | |
3728 | * over all possibly mapped software queues. | |
3729 | */ | |
88459642 | 3730 | sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx); |
889fa31f | 3731 | |
484b4061 JA |
3732 | /* |
3733 | * Initialize batch roundrobin counts | |
3734 | */ | |
f82ddf19 | 3735 | hctx->next_cpu = blk_mq_first_mapped_cpu(hctx); |
506e931f JA |
3736 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; |
3737 | } | |
320ae51f JA |
3738 | } |
3739 | ||
8e8320c9 JA |
3740 | /* |
3741 | * Caller needs to ensure that we're either frozen/quiesced, or that | |
3742 | * the queue isn't live yet. | |
3743 | */ | |
2404e607 | 3744 | static void queue_set_hctx_shared(struct request_queue *q, bool shared) |
0d2602ca JA |
3745 | { |
3746 | struct blk_mq_hw_ctx *hctx; | |
0d2602ca JA |
3747 | int i; |
3748 | ||
2404e607 | 3749 | queue_for_each_hw_ctx(q, hctx, i) { |
454bb677 | 3750 | if (shared) { |
51db1c37 | 3751 | hctx->flags |= BLK_MQ_F_TAG_QUEUE_SHARED; |
454bb677 YK |
3752 | } else { |
3753 | blk_mq_tag_idle(hctx); | |
51db1c37 | 3754 | hctx->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED; |
454bb677 | 3755 | } |
2404e607 JM |
3756 | } |
3757 | } | |
3758 | ||
655ac300 HR |
3759 | static void blk_mq_update_tag_set_shared(struct blk_mq_tag_set *set, |
3760 | bool shared) | |
2404e607 JM |
3761 | { |
3762 | struct request_queue *q; | |
0d2602ca | 3763 | |
705cda97 BVA |
3764 | lockdep_assert_held(&set->tag_list_lock); |
3765 | ||
0d2602ca JA |
3766 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
3767 | blk_mq_freeze_queue(q); | |
2404e607 | 3768 | queue_set_hctx_shared(q, shared); |
0d2602ca JA |
3769 | blk_mq_unfreeze_queue(q); |
3770 | } | |
3771 | } | |
3772 | ||
3773 | static void blk_mq_del_queue_tag_set(struct request_queue *q) | |
3774 | { | |
3775 | struct blk_mq_tag_set *set = q->tag_set; | |
3776 | ||
0d2602ca | 3777 | mutex_lock(&set->tag_list_lock); |
08c875cb | 3778 | list_del(&q->tag_set_list); |
2404e607 JM |
3779 | if (list_is_singular(&set->tag_list)) { |
3780 | /* just transitioned to unshared */ | |
51db1c37 | 3781 | set->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED; |
2404e607 | 3782 | /* update existing queue */ |
655ac300 | 3783 | blk_mq_update_tag_set_shared(set, false); |
2404e607 | 3784 | } |
0d2602ca | 3785 | mutex_unlock(&set->tag_list_lock); |
a347c7ad | 3786 | INIT_LIST_HEAD(&q->tag_set_list); |
0d2602ca JA |
3787 | } |
3788 | ||
3789 | static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set, | |
3790 | struct request_queue *q) | |
3791 | { | |
0d2602ca | 3792 | mutex_lock(&set->tag_list_lock); |
2404e607 | 3793 | |
ff821d27 JA |
3794 | /* |
3795 | * Check to see if we're transitioning to shared (from 1 to 2 queues). | |
3796 | */ | |
3797 | if (!list_empty(&set->tag_list) && | |
51db1c37 ML |
3798 | !(set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) { |
3799 | set->flags |= BLK_MQ_F_TAG_QUEUE_SHARED; | |
2404e607 | 3800 | /* update existing queue */ |
655ac300 | 3801 | blk_mq_update_tag_set_shared(set, true); |
2404e607 | 3802 | } |
51db1c37 | 3803 | if (set->flags & BLK_MQ_F_TAG_QUEUE_SHARED) |
2404e607 | 3804 | queue_set_hctx_shared(q, true); |
08c875cb | 3805 | list_add_tail(&q->tag_set_list, &set->tag_list); |
2404e607 | 3806 | |
0d2602ca JA |
3807 | mutex_unlock(&set->tag_list_lock); |
3808 | } | |
3809 | ||
1db4909e ML |
3810 | /* All allocations will be freed in release handler of q->mq_kobj */ |
3811 | static int blk_mq_alloc_ctxs(struct request_queue *q) | |
3812 | { | |
3813 | struct blk_mq_ctxs *ctxs; | |
3814 | int cpu; | |
3815 | ||
3816 | ctxs = kzalloc(sizeof(*ctxs), GFP_KERNEL); | |
3817 | if (!ctxs) | |
3818 | return -ENOMEM; | |
3819 | ||
3820 | ctxs->queue_ctx = alloc_percpu(struct blk_mq_ctx); | |
3821 | if (!ctxs->queue_ctx) | |
3822 | goto fail; | |
3823 | ||
3824 | for_each_possible_cpu(cpu) { | |
3825 | struct blk_mq_ctx *ctx = per_cpu_ptr(ctxs->queue_ctx, cpu); | |
3826 | ctx->ctxs = ctxs; | |
3827 | } | |
3828 | ||
3829 | q->mq_kobj = &ctxs->kobj; | |
3830 | q->queue_ctx = ctxs->queue_ctx; | |
3831 | ||
3832 | return 0; | |
3833 | fail: | |
3834 | kfree(ctxs); | |
3835 | return -ENOMEM; | |
3836 | } | |
3837 | ||
e09aae7e ML |
3838 | /* |
3839 | * It is the actual release handler for mq, but we do it from | |
3840 | * request queue's release handler for avoiding use-after-free | |
3841 | * and headache because q->mq_kobj shouldn't have been introduced, | |
3842 | * but we can't group ctx/kctx kobj without it. | |
3843 | */ | |
3844 | void blk_mq_release(struct request_queue *q) | |
3845 | { | |
2f8f1336 ML |
3846 | struct blk_mq_hw_ctx *hctx, *next; |
3847 | int i; | |
e09aae7e | 3848 | |
2f8f1336 ML |
3849 | queue_for_each_hw_ctx(q, hctx, i) |
3850 | WARN_ON_ONCE(hctx && list_empty(&hctx->hctx_list)); | |
3851 | ||
3852 | /* all hctx are in .unused_hctx_list now */ | |
3853 | list_for_each_entry_safe(hctx, next, &q->unused_hctx_list, hctx_list) { | |
3854 | list_del_init(&hctx->hctx_list); | |
6c8b232e | 3855 | kobject_put(&hctx->kobj); |
c3b4afca | 3856 | } |
e09aae7e ML |
3857 | |
3858 | kfree(q->queue_hw_ctx); | |
3859 | ||
7ea5fe31 ML |
3860 | /* |
3861 | * release .mq_kobj and sw queue's kobject now because | |
3862 | * both share lifetime with request queue. | |
3863 | */ | |
3864 | blk_mq_sysfs_deinit(q); | |
e09aae7e ML |
3865 | } |
3866 | ||
5ec780a6 | 3867 | static struct request_queue *blk_mq_init_queue_data(struct blk_mq_tag_set *set, |
2f227bb9 | 3868 | void *queuedata) |
b62c21b7 | 3869 | { |
26a9750a CH |
3870 | struct request_queue *q; |
3871 | int ret; | |
b62c21b7 | 3872 | |
704b914f | 3873 | q = blk_alloc_queue(set->numa_node, set->flags & BLK_MQ_F_BLOCKING); |
26a9750a | 3874 | if (!q) |
b62c21b7 | 3875 | return ERR_PTR(-ENOMEM); |
26a9750a CH |
3876 | q->queuedata = queuedata; |
3877 | ret = blk_mq_init_allocated_queue(set, q); | |
3878 | if (ret) { | |
3879 | blk_cleanup_queue(q); | |
3880 | return ERR_PTR(ret); | |
3881 | } | |
b62c21b7 MS |
3882 | return q; |
3883 | } | |
2f227bb9 CH |
3884 | |
3885 | struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set) | |
3886 | { | |
3887 | return blk_mq_init_queue_data(set, NULL); | |
3888 | } | |
b62c21b7 MS |
3889 | EXPORT_SYMBOL(blk_mq_init_queue); |
3890 | ||
4dcc4874 CH |
3891 | struct gendisk *__blk_mq_alloc_disk(struct blk_mq_tag_set *set, void *queuedata, |
3892 | struct lock_class_key *lkclass) | |
9316a9ed JA |
3893 | { |
3894 | struct request_queue *q; | |
b461dfc4 | 3895 | struct gendisk *disk; |
9316a9ed | 3896 | |
b461dfc4 CH |
3897 | q = blk_mq_init_queue_data(set, queuedata); |
3898 | if (IS_ERR(q)) | |
3899 | return ERR_CAST(q); | |
9316a9ed | 3900 | |
4a1fa41d | 3901 | disk = __alloc_disk_node(q, set->numa_node, lkclass); |
b461dfc4 CH |
3902 | if (!disk) { |
3903 | blk_cleanup_queue(q); | |
3904 | return ERR_PTR(-ENOMEM); | |
9316a9ed | 3905 | } |
b461dfc4 | 3906 | return disk; |
9316a9ed | 3907 | } |
b461dfc4 | 3908 | EXPORT_SYMBOL(__blk_mq_alloc_disk); |
9316a9ed | 3909 | |
34d11ffa JW |
3910 | static struct blk_mq_hw_ctx *blk_mq_alloc_and_init_hctx( |
3911 | struct blk_mq_tag_set *set, struct request_queue *q, | |
3912 | int hctx_idx, int node) | |
3913 | { | |
2f8f1336 | 3914 | struct blk_mq_hw_ctx *hctx = NULL, *tmp; |
34d11ffa | 3915 | |
2f8f1336 ML |
3916 | /* reuse dead hctx first */ |
3917 | spin_lock(&q->unused_hctx_lock); | |
3918 | list_for_each_entry(tmp, &q->unused_hctx_list, hctx_list) { | |
3919 | if (tmp->numa_node == node) { | |
3920 | hctx = tmp; | |
3921 | break; | |
3922 | } | |
3923 | } | |
3924 | if (hctx) | |
3925 | list_del_init(&hctx->hctx_list); | |
3926 | spin_unlock(&q->unused_hctx_lock); | |
3927 | ||
3928 | if (!hctx) | |
3929 | hctx = blk_mq_alloc_hctx(q, set, node); | |
34d11ffa | 3930 | if (!hctx) |
7c6c5b7c | 3931 | goto fail; |
34d11ffa | 3932 | |
7c6c5b7c ML |
3933 | if (blk_mq_init_hctx(q, set, hctx, hctx_idx)) |
3934 | goto free_hctx; | |
34d11ffa JW |
3935 | |
3936 | return hctx; | |
7c6c5b7c ML |
3937 | |
3938 | free_hctx: | |
3939 | kobject_put(&hctx->kobj); | |
3940 | fail: | |
3941 | return NULL; | |
34d11ffa JW |
3942 | } |
3943 | ||
868f2f0b KB |
3944 | static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set, |
3945 | struct request_queue *q) | |
320ae51f | 3946 | { |
e01ad46d | 3947 | int i, j, end; |
868f2f0b | 3948 | struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx; |
f14bbe77 | 3949 | |
ac0d6b92 BVA |
3950 | if (q->nr_hw_queues < set->nr_hw_queues) { |
3951 | struct blk_mq_hw_ctx **new_hctxs; | |
3952 | ||
3953 | new_hctxs = kcalloc_node(set->nr_hw_queues, | |
3954 | sizeof(*new_hctxs), GFP_KERNEL, | |
3955 | set->numa_node); | |
3956 | if (!new_hctxs) | |
3957 | return; | |
3958 | if (hctxs) | |
3959 | memcpy(new_hctxs, hctxs, q->nr_hw_queues * | |
3960 | sizeof(*hctxs)); | |
3961 | q->queue_hw_ctx = new_hctxs; | |
ac0d6b92 BVA |
3962 | kfree(hctxs); |
3963 | hctxs = new_hctxs; | |
3964 | } | |
3965 | ||
fb350e0a ML |
3966 | /* protect against switching io scheduler */ |
3967 | mutex_lock(&q->sysfs_lock); | |
24d2f903 | 3968 | for (i = 0; i < set->nr_hw_queues; i++) { |
4d805131 | 3969 | int node = blk_mq_get_hctx_node(set, i); |
34d11ffa | 3970 | struct blk_mq_hw_ctx *hctx; |
868f2f0b | 3971 | |
34d11ffa JW |
3972 | /* |
3973 | * If the hw queue has been mapped to another numa node, | |
3974 | * we need to realloc the hctx. If allocation fails, fallback | |
3975 | * to use the previous one. | |
3976 | */ | |
3977 | if (hctxs[i] && (hctxs[i]->numa_node == node)) | |
3978 | continue; | |
868f2f0b | 3979 | |
34d11ffa JW |
3980 | hctx = blk_mq_alloc_and_init_hctx(set, q, i, node); |
3981 | if (hctx) { | |
2f8f1336 | 3982 | if (hctxs[i]) |
34d11ffa | 3983 | blk_mq_exit_hctx(q, set, hctxs[i], i); |
34d11ffa JW |
3984 | hctxs[i] = hctx; |
3985 | } else { | |
3986 | if (hctxs[i]) | |
3987 | pr_warn("Allocate new hctx on node %d fails,\ | |
3988 | fallback to previous one on node %d\n", | |
3989 | node, hctxs[i]->numa_node); | |
3990 | else | |
3991 | break; | |
868f2f0b | 3992 | } |
320ae51f | 3993 | } |
e01ad46d JW |
3994 | /* |
3995 | * Increasing nr_hw_queues fails. Free the newly allocated | |
3996 | * hctxs and keep the previous q->nr_hw_queues. | |
3997 | */ | |
3998 | if (i != set->nr_hw_queues) { | |
3999 | j = q->nr_hw_queues; | |
4000 | end = i; | |
4001 | } else { | |
4002 | j = i; | |
4003 | end = q->nr_hw_queues; | |
4004 | q->nr_hw_queues = set->nr_hw_queues; | |
4005 | } | |
34d11ffa | 4006 | |
e01ad46d | 4007 | for (; j < end; j++) { |
868f2f0b KB |
4008 | struct blk_mq_hw_ctx *hctx = hctxs[j]; |
4009 | ||
4010 | if (hctx) { | |
868f2f0b | 4011 | blk_mq_exit_hctx(q, set, hctx, j); |
868f2f0b | 4012 | hctxs[j] = NULL; |
868f2f0b KB |
4013 | } |
4014 | } | |
fb350e0a | 4015 | mutex_unlock(&q->sysfs_lock); |
868f2f0b KB |
4016 | } |
4017 | ||
26a9750a CH |
4018 | int blk_mq_init_allocated_queue(struct blk_mq_tag_set *set, |
4019 | struct request_queue *q) | |
868f2f0b | 4020 | { |
704b914f ML |
4021 | WARN_ON_ONCE(blk_queue_has_srcu(q) != |
4022 | !!(set->flags & BLK_MQ_F_BLOCKING)); | |
4023 | ||
66841672 ML |
4024 | /* mark the queue as mq asap */ |
4025 | q->mq_ops = set->ops; | |
4026 | ||
34dbad5d | 4027 | q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn, |
720b8ccc SB |
4028 | blk_mq_poll_stats_bkt, |
4029 | BLK_MQ_POLL_STATS_BKTS, q); | |
34dbad5d OS |
4030 | if (!q->poll_cb) |
4031 | goto err_exit; | |
4032 | ||
1db4909e | 4033 | if (blk_mq_alloc_ctxs(q)) |
41de54c6 | 4034 | goto err_poll; |
868f2f0b | 4035 | |
737f98cf ML |
4036 | /* init q->mq_kobj and sw queues' kobjects */ |
4037 | blk_mq_sysfs_init(q); | |
4038 | ||
2f8f1336 ML |
4039 | INIT_LIST_HEAD(&q->unused_hctx_list); |
4040 | spin_lock_init(&q->unused_hctx_lock); | |
4041 | ||
868f2f0b KB |
4042 | blk_mq_realloc_hw_ctxs(set, q); |
4043 | if (!q->nr_hw_queues) | |
4044 | goto err_hctxs; | |
320ae51f | 4045 | |
287922eb | 4046 | INIT_WORK(&q->timeout_work, blk_mq_timeout_work); |
e56f698b | 4047 | blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ); |
320ae51f | 4048 | |
a8908939 | 4049 | q->tag_set = set; |
320ae51f | 4050 | |
94eddfbe | 4051 | q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT; |
cd19181b ML |
4052 | if (set->nr_maps > HCTX_TYPE_POLL && |
4053 | set->map[HCTX_TYPE_POLL].nr_queues) | |
6544d229 | 4054 | blk_queue_flag_set(QUEUE_FLAG_POLL, q); |
320ae51f | 4055 | |
2849450a | 4056 | INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work); |
6fca6a61 CH |
4057 | INIT_LIST_HEAD(&q->requeue_list); |
4058 | spin_lock_init(&q->requeue_lock); | |
4059 | ||
eba71768 JA |
4060 | q->nr_requests = set->queue_depth; |
4061 | ||
64f1c21e JA |
4062 | /* |
4063 | * Default to classic polling | |
4064 | */ | |
29ece8b4 | 4065 | q->poll_nsec = BLK_MQ_POLL_CLASSIC; |
64f1c21e | 4066 | |
24d2f903 | 4067 | blk_mq_init_cpu_queues(q, set->nr_hw_queues); |
0d2602ca | 4068 | blk_mq_add_queue_tag_set(set, q); |
4b855ad3 | 4069 | blk_mq_map_swqueue(q); |
26a9750a | 4070 | return 0; |
18741986 | 4071 | |
320ae51f | 4072 | err_hctxs: |
868f2f0b | 4073 | kfree(q->queue_hw_ctx); |
73d9c8d4 | 4074 | q->nr_hw_queues = 0; |
1db4909e | 4075 | blk_mq_sysfs_deinit(q); |
41de54c6 JS |
4076 | err_poll: |
4077 | blk_stat_free_callback(q->poll_cb); | |
4078 | q->poll_cb = NULL; | |
c7de5726 ML |
4079 | err_exit: |
4080 | q->mq_ops = NULL; | |
26a9750a | 4081 | return -ENOMEM; |
320ae51f | 4082 | } |
b62c21b7 | 4083 | EXPORT_SYMBOL(blk_mq_init_allocated_queue); |
320ae51f | 4084 | |
c7e2d94b ML |
4085 | /* tags can _not_ be used after returning from blk_mq_exit_queue */ |
4086 | void blk_mq_exit_queue(struct request_queue *q) | |
320ae51f | 4087 | { |
630ef623 | 4088 | struct blk_mq_tag_set *set = q->tag_set; |
320ae51f | 4089 | |
630ef623 | 4090 | /* Checks hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED. */ |
624dbe47 | 4091 | blk_mq_exit_hw_queues(q, set, set->nr_hw_queues); |
630ef623 BVA |
4092 | /* May clear BLK_MQ_F_TAG_QUEUE_SHARED in hctx->flags. */ |
4093 | blk_mq_del_queue_tag_set(q); | |
320ae51f | 4094 | } |
320ae51f | 4095 | |
a5164405 JA |
4096 | static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set) |
4097 | { | |
4098 | int i; | |
4099 | ||
079a2e3e JG |
4100 | if (blk_mq_is_shared_tags(set->flags)) { |
4101 | set->shared_tags = blk_mq_alloc_map_and_rqs(set, | |
e155b0c2 JG |
4102 | BLK_MQ_NO_HCTX_IDX, |
4103 | set->queue_depth); | |
079a2e3e | 4104 | if (!set->shared_tags) |
e155b0c2 JG |
4105 | return -ENOMEM; |
4106 | } | |
4107 | ||
8229cca8 | 4108 | for (i = 0; i < set->nr_hw_queues; i++) { |
63064be1 | 4109 | if (!__blk_mq_alloc_map_and_rqs(set, i)) |
a5164405 | 4110 | goto out_unwind; |
8229cca8 XT |
4111 | cond_resched(); |
4112 | } | |
a5164405 JA |
4113 | |
4114 | return 0; | |
4115 | ||
4116 | out_unwind: | |
4117 | while (--i >= 0) | |
e155b0c2 JG |
4118 | __blk_mq_free_map_and_rqs(set, i); |
4119 | ||
079a2e3e JG |
4120 | if (blk_mq_is_shared_tags(set->flags)) { |
4121 | blk_mq_free_map_and_rqs(set, set->shared_tags, | |
e155b0c2 | 4122 | BLK_MQ_NO_HCTX_IDX); |
645db34e | 4123 | } |
a5164405 | 4124 | |
a5164405 JA |
4125 | return -ENOMEM; |
4126 | } | |
4127 | ||
4128 | /* | |
4129 | * Allocate the request maps associated with this tag_set. Note that this | |
4130 | * may reduce the depth asked for, if memory is tight. set->queue_depth | |
4131 | * will be updated to reflect the allocated depth. | |
4132 | */ | |
63064be1 | 4133 | static int blk_mq_alloc_set_map_and_rqs(struct blk_mq_tag_set *set) |
a5164405 JA |
4134 | { |
4135 | unsigned int depth; | |
4136 | int err; | |
4137 | ||
4138 | depth = set->queue_depth; | |
4139 | do { | |
4140 | err = __blk_mq_alloc_rq_maps(set); | |
4141 | if (!err) | |
4142 | break; | |
4143 | ||
4144 | set->queue_depth >>= 1; | |
4145 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) { | |
4146 | err = -ENOMEM; | |
4147 | break; | |
4148 | } | |
4149 | } while (set->queue_depth); | |
4150 | ||
4151 | if (!set->queue_depth || err) { | |
4152 | pr_err("blk-mq: failed to allocate request map\n"); | |
4153 | return -ENOMEM; | |
4154 | } | |
4155 | ||
4156 | if (depth != set->queue_depth) | |
4157 | pr_info("blk-mq: reduced tag depth (%u -> %u)\n", | |
4158 | depth, set->queue_depth); | |
4159 | ||
4160 | return 0; | |
4161 | } | |
4162 | ||
ebe8bddb OS |
4163 | static int blk_mq_update_queue_map(struct blk_mq_tag_set *set) |
4164 | { | |
6e66b493 BVA |
4165 | /* |
4166 | * blk_mq_map_queues() and multiple .map_queues() implementations | |
4167 | * expect that set->map[HCTX_TYPE_DEFAULT].nr_queues is set to the | |
4168 | * number of hardware queues. | |
4169 | */ | |
4170 | if (set->nr_maps == 1) | |
4171 | set->map[HCTX_TYPE_DEFAULT].nr_queues = set->nr_hw_queues; | |
4172 | ||
59388702 | 4173 | if (set->ops->map_queues && !is_kdump_kernel()) { |
b3c661b1 JA |
4174 | int i; |
4175 | ||
7d4901a9 ML |
4176 | /* |
4177 | * transport .map_queues is usually done in the following | |
4178 | * way: | |
4179 | * | |
4180 | * for (queue = 0; queue < set->nr_hw_queues; queue++) { | |
4181 | * mask = get_cpu_mask(queue) | |
4182 | * for_each_cpu(cpu, mask) | |
b3c661b1 | 4183 | * set->map[x].mq_map[cpu] = queue; |
7d4901a9 ML |
4184 | * } |
4185 | * | |
4186 | * When we need to remap, the table has to be cleared for | |
4187 | * killing stale mapping since one CPU may not be mapped | |
4188 | * to any hw queue. | |
4189 | */ | |
b3c661b1 JA |
4190 | for (i = 0; i < set->nr_maps; i++) |
4191 | blk_mq_clear_mq_map(&set->map[i]); | |
7d4901a9 | 4192 | |
ebe8bddb | 4193 | return set->ops->map_queues(set); |
b3c661b1 JA |
4194 | } else { |
4195 | BUG_ON(set->nr_maps > 1); | |
7d76f856 | 4196 | return blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]); |
b3c661b1 | 4197 | } |
ebe8bddb OS |
4198 | } |
4199 | ||
f7e76dbc BVA |
4200 | static int blk_mq_realloc_tag_set_tags(struct blk_mq_tag_set *set, |
4201 | int cur_nr_hw_queues, int new_nr_hw_queues) | |
4202 | { | |
4203 | struct blk_mq_tags **new_tags; | |
4204 | ||
4205 | if (cur_nr_hw_queues >= new_nr_hw_queues) | |
4206 | return 0; | |
4207 | ||
4208 | new_tags = kcalloc_node(new_nr_hw_queues, sizeof(struct blk_mq_tags *), | |
4209 | GFP_KERNEL, set->numa_node); | |
4210 | if (!new_tags) | |
4211 | return -ENOMEM; | |
4212 | ||
4213 | if (set->tags) | |
4214 | memcpy(new_tags, set->tags, cur_nr_hw_queues * | |
4215 | sizeof(*set->tags)); | |
4216 | kfree(set->tags); | |
4217 | set->tags = new_tags; | |
4218 | set->nr_hw_queues = new_nr_hw_queues; | |
4219 | ||
4220 | return 0; | |
4221 | } | |
4222 | ||
91cdf265 MI |
4223 | static int blk_mq_alloc_tag_set_tags(struct blk_mq_tag_set *set, |
4224 | int new_nr_hw_queues) | |
4225 | { | |
4226 | return blk_mq_realloc_tag_set_tags(set, 0, new_nr_hw_queues); | |
4227 | } | |
4228 | ||
a4391c64 JA |
4229 | /* |
4230 | * Alloc a tag set to be associated with one or more request queues. | |
4231 | * May fail with EINVAL for various error conditions. May adjust the | |
c018c84f | 4232 | * requested depth down, if it's too large. In that case, the set |
a4391c64 JA |
4233 | * value will be stored in set->queue_depth. |
4234 | */ | |
24d2f903 CH |
4235 | int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set) |
4236 | { | |
b3c661b1 | 4237 | int i, ret; |
da695ba2 | 4238 | |
205fb5f5 BVA |
4239 | BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS); |
4240 | ||
24d2f903 CH |
4241 | if (!set->nr_hw_queues) |
4242 | return -EINVAL; | |
a4391c64 | 4243 | if (!set->queue_depth) |
24d2f903 CH |
4244 | return -EINVAL; |
4245 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) | |
4246 | return -EINVAL; | |
4247 | ||
7d7e0f90 | 4248 | if (!set->ops->queue_rq) |
24d2f903 CH |
4249 | return -EINVAL; |
4250 | ||
de148297 ML |
4251 | if (!set->ops->get_budget ^ !set->ops->put_budget) |
4252 | return -EINVAL; | |
4253 | ||
a4391c64 JA |
4254 | if (set->queue_depth > BLK_MQ_MAX_DEPTH) { |
4255 | pr_info("blk-mq: reduced tag depth to %u\n", | |
4256 | BLK_MQ_MAX_DEPTH); | |
4257 | set->queue_depth = BLK_MQ_MAX_DEPTH; | |
4258 | } | |
24d2f903 | 4259 | |
b3c661b1 JA |
4260 | if (!set->nr_maps) |
4261 | set->nr_maps = 1; | |
4262 | else if (set->nr_maps > HCTX_MAX_TYPES) | |
4263 | return -EINVAL; | |
4264 | ||
6637fadf SL |
4265 | /* |
4266 | * If a crashdump is active, then we are potentially in a very | |
4267 | * memory constrained environment. Limit us to 1 queue and | |
4268 | * 64 tags to prevent using too much memory. | |
4269 | */ | |
4270 | if (is_kdump_kernel()) { | |
4271 | set->nr_hw_queues = 1; | |
59388702 | 4272 | set->nr_maps = 1; |
6637fadf SL |
4273 | set->queue_depth = min(64U, set->queue_depth); |
4274 | } | |
868f2f0b | 4275 | /* |
392546ae JA |
4276 | * There is no use for more h/w queues than cpus if we just have |
4277 | * a single map | |
868f2f0b | 4278 | */ |
392546ae | 4279 | if (set->nr_maps == 1 && set->nr_hw_queues > nr_cpu_ids) |
868f2f0b | 4280 | set->nr_hw_queues = nr_cpu_ids; |
6637fadf | 4281 | |
91cdf265 | 4282 | if (blk_mq_alloc_tag_set_tags(set, set->nr_hw_queues) < 0) |
a5164405 | 4283 | return -ENOMEM; |
24d2f903 | 4284 | |
da695ba2 | 4285 | ret = -ENOMEM; |
b3c661b1 JA |
4286 | for (i = 0; i < set->nr_maps; i++) { |
4287 | set->map[i].mq_map = kcalloc_node(nr_cpu_ids, | |
07b35eb5 | 4288 | sizeof(set->map[i].mq_map[0]), |
b3c661b1 JA |
4289 | GFP_KERNEL, set->numa_node); |
4290 | if (!set->map[i].mq_map) | |
4291 | goto out_free_mq_map; | |
59388702 | 4292 | set->map[i].nr_queues = is_kdump_kernel() ? 1 : set->nr_hw_queues; |
b3c661b1 | 4293 | } |
bdd17e75 | 4294 | |
ebe8bddb | 4295 | ret = blk_mq_update_queue_map(set); |
da695ba2 CH |
4296 | if (ret) |
4297 | goto out_free_mq_map; | |
4298 | ||
63064be1 | 4299 | ret = blk_mq_alloc_set_map_and_rqs(set); |
da695ba2 | 4300 | if (ret) |
bdd17e75 | 4301 | goto out_free_mq_map; |
24d2f903 | 4302 | |
0d2602ca JA |
4303 | mutex_init(&set->tag_list_lock); |
4304 | INIT_LIST_HEAD(&set->tag_list); | |
4305 | ||
24d2f903 | 4306 | return 0; |
bdd17e75 CH |
4307 | |
4308 | out_free_mq_map: | |
b3c661b1 JA |
4309 | for (i = 0; i < set->nr_maps; i++) { |
4310 | kfree(set->map[i].mq_map); | |
4311 | set->map[i].mq_map = NULL; | |
4312 | } | |
5676e7b6 RE |
4313 | kfree(set->tags); |
4314 | set->tags = NULL; | |
da695ba2 | 4315 | return ret; |
24d2f903 CH |
4316 | } |
4317 | EXPORT_SYMBOL(blk_mq_alloc_tag_set); | |
4318 | ||
cdb14e0f CH |
4319 | /* allocate and initialize a tagset for a simple single-queue device */ |
4320 | int blk_mq_alloc_sq_tag_set(struct blk_mq_tag_set *set, | |
4321 | const struct blk_mq_ops *ops, unsigned int queue_depth, | |
4322 | unsigned int set_flags) | |
4323 | { | |
4324 | memset(set, 0, sizeof(*set)); | |
4325 | set->ops = ops; | |
4326 | set->nr_hw_queues = 1; | |
4327 | set->nr_maps = 1; | |
4328 | set->queue_depth = queue_depth; | |
4329 | set->numa_node = NUMA_NO_NODE; | |
4330 | set->flags = set_flags; | |
4331 | return blk_mq_alloc_tag_set(set); | |
4332 | } | |
4333 | EXPORT_SYMBOL_GPL(blk_mq_alloc_sq_tag_set); | |
4334 | ||
24d2f903 CH |
4335 | void blk_mq_free_tag_set(struct blk_mq_tag_set *set) |
4336 | { | |
b3c661b1 | 4337 | int i, j; |
24d2f903 | 4338 | |
f7e76dbc | 4339 | for (i = 0; i < set->nr_hw_queues; i++) |
e155b0c2 | 4340 | __blk_mq_free_map_and_rqs(set, i); |
484b4061 | 4341 | |
079a2e3e JG |
4342 | if (blk_mq_is_shared_tags(set->flags)) { |
4343 | blk_mq_free_map_and_rqs(set, set->shared_tags, | |
e155b0c2 JG |
4344 | BLK_MQ_NO_HCTX_IDX); |
4345 | } | |
32bc15af | 4346 | |
b3c661b1 JA |
4347 | for (j = 0; j < set->nr_maps; j++) { |
4348 | kfree(set->map[j].mq_map); | |
4349 | set->map[j].mq_map = NULL; | |
4350 | } | |
bdd17e75 | 4351 | |
981bd189 | 4352 | kfree(set->tags); |
5676e7b6 | 4353 | set->tags = NULL; |
24d2f903 CH |
4354 | } |
4355 | EXPORT_SYMBOL(blk_mq_free_tag_set); | |
4356 | ||
e3a2b3f9 JA |
4357 | int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr) |
4358 | { | |
4359 | struct blk_mq_tag_set *set = q->tag_set; | |
4360 | struct blk_mq_hw_ctx *hctx; | |
4361 | int i, ret; | |
4362 | ||
bd166ef1 | 4363 | if (!set) |
e3a2b3f9 JA |
4364 | return -EINVAL; |
4365 | ||
e5fa8140 AZ |
4366 | if (q->nr_requests == nr) |
4367 | return 0; | |
4368 | ||
70f36b60 | 4369 | blk_mq_freeze_queue(q); |
24f5a90f | 4370 | blk_mq_quiesce_queue(q); |
70f36b60 | 4371 | |
e3a2b3f9 JA |
4372 | ret = 0; |
4373 | queue_for_each_hw_ctx(q, hctx, i) { | |
e9137d4b KB |
4374 | if (!hctx->tags) |
4375 | continue; | |
bd166ef1 JA |
4376 | /* |
4377 | * If we're using an MQ scheduler, just update the scheduler | |
4378 | * queue depth. This is similar to what the old code would do. | |
4379 | */ | |
f6adcef5 | 4380 | if (hctx->sched_tags) { |
70f36b60 | 4381 | ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags, |
f6adcef5 | 4382 | nr, true); |
f6adcef5 JG |
4383 | } else { |
4384 | ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr, | |
4385 | false); | |
70f36b60 | 4386 | } |
e3a2b3f9 JA |
4387 | if (ret) |
4388 | break; | |
77f1e0a5 JA |
4389 | if (q->elevator && q->elevator->type->ops.depth_updated) |
4390 | q->elevator->type->ops.depth_updated(hctx); | |
e3a2b3f9 | 4391 | } |
d97e594c | 4392 | if (!ret) { |
e3a2b3f9 | 4393 | q->nr_requests = nr; |
079a2e3e | 4394 | if (blk_mq_is_shared_tags(set->flags)) { |
8fa04464 | 4395 | if (q->elevator) |
079a2e3e | 4396 | blk_mq_tag_update_sched_shared_tags(q); |
8fa04464 | 4397 | else |
079a2e3e | 4398 | blk_mq_tag_resize_shared_tags(set, nr); |
8fa04464 | 4399 | } |
d97e594c | 4400 | } |
e3a2b3f9 | 4401 | |
24f5a90f | 4402 | blk_mq_unquiesce_queue(q); |
70f36b60 | 4403 | blk_mq_unfreeze_queue(q); |
70f36b60 | 4404 | |
e3a2b3f9 JA |
4405 | return ret; |
4406 | } | |
4407 | ||
d48ece20 JW |
4408 | /* |
4409 | * request_queue and elevator_type pair. | |
4410 | * It is just used by __blk_mq_update_nr_hw_queues to cache | |
4411 | * the elevator_type associated with a request_queue. | |
4412 | */ | |
4413 | struct blk_mq_qe_pair { | |
4414 | struct list_head node; | |
4415 | struct request_queue *q; | |
4416 | struct elevator_type *type; | |
4417 | }; | |
4418 | ||
4419 | /* | |
4420 | * Cache the elevator_type in qe pair list and switch the | |
4421 | * io scheduler to 'none' | |
4422 | */ | |
4423 | static bool blk_mq_elv_switch_none(struct list_head *head, | |
4424 | struct request_queue *q) | |
4425 | { | |
4426 | struct blk_mq_qe_pair *qe; | |
4427 | ||
4428 | if (!q->elevator) | |
4429 | return true; | |
4430 | ||
4431 | qe = kmalloc(sizeof(*qe), GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY); | |
4432 | if (!qe) | |
4433 | return false; | |
4434 | ||
4435 | INIT_LIST_HEAD(&qe->node); | |
4436 | qe->q = q; | |
4437 | qe->type = q->elevator->type; | |
4438 | list_add(&qe->node, head); | |
4439 | ||
4440 | mutex_lock(&q->sysfs_lock); | |
4441 | /* | |
4442 | * After elevator_switch_mq, the previous elevator_queue will be | |
4443 | * released by elevator_release. The reference of the io scheduler | |
4444 | * module get by elevator_get will also be put. So we need to get | |
4445 | * a reference of the io scheduler module here to prevent it to be | |
4446 | * removed. | |
4447 | */ | |
4448 | __module_get(qe->type->elevator_owner); | |
4449 | elevator_switch_mq(q, NULL); | |
4450 | mutex_unlock(&q->sysfs_lock); | |
4451 | ||
4452 | return true; | |
4453 | } | |
4454 | ||
4455 | static void blk_mq_elv_switch_back(struct list_head *head, | |
4456 | struct request_queue *q) | |
4457 | { | |
4458 | struct blk_mq_qe_pair *qe; | |
4459 | struct elevator_type *t = NULL; | |
4460 | ||
4461 | list_for_each_entry(qe, head, node) | |
4462 | if (qe->q == q) { | |
4463 | t = qe->type; | |
4464 | break; | |
4465 | } | |
4466 | ||
4467 | if (!t) | |
4468 | return; | |
4469 | ||
4470 | list_del(&qe->node); | |
4471 | kfree(qe); | |
4472 | ||
4473 | mutex_lock(&q->sysfs_lock); | |
4474 | elevator_switch_mq(q, t); | |
4475 | mutex_unlock(&q->sysfs_lock); | |
4476 | } | |
4477 | ||
e4dc2b32 KB |
4478 | static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, |
4479 | int nr_hw_queues) | |
868f2f0b KB |
4480 | { |
4481 | struct request_queue *q; | |
d48ece20 | 4482 | LIST_HEAD(head); |
e01ad46d | 4483 | int prev_nr_hw_queues; |
868f2f0b | 4484 | |
705cda97 BVA |
4485 | lockdep_assert_held(&set->tag_list_lock); |
4486 | ||
392546ae | 4487 | if (set->nr_maps == 1 && nr_hw_queues > nr_cpu_ids) |
868f2f0b | 4488 | nr_hw_queues = nr_cpu_ids; |
fe35ec58 WZ |
4489 | if (nr_hw_queues < 1) |
4490 | return; | |
4491 | if (set->nr_maps == 1 && nr_hw_queues == set->nr_hw_queues) | |
868f2f0b KB |
4492 | return; |
4493 | ||
4494 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
4495 | blk_mq_freeze_queue(q); | |
d48ece20 JW |
4496 | /* |
4497 | * Switch IO scheduler to 'none', cleaning up the data associated | |
4498 | * with the previous scheduler. We will switch back once we are done | |
4499 | * updating the new sw to hw queue mappings. | |
4500 | */ | |
4501 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
4502 | if (!blk_mq_elv_switch_none(&head, q)) | |
4503 | goto switch_back; | |
868f2f0b | 4504 | |
477e19de JW |
4505 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
4506 | blk_mq_debugfs_unregister_hctxs(q); | |
4507 | blk_mq_sysfs_unregister(q); | |
4508 | } | |
4509 | ||
a2584e43 | 4510 | prev_nr_hw_queues = set->nr_hw_queues; |
f7e76dbc BVA |
4511 | if (blk_mq_realloc_tag_set_tags(set, set->nr_hw_queues, nr_hw_queues) < |
4512 | 0) | |
4513 | goto reregister; | |
4514 | ||
868f2f0b | 4515 | set->nr_hw_queues = nr_hw_queues; |
e01ad46d | 4516 | fallback: |
aa880ad6 | 4517 | blk_mq_update_queue_map(set); |
868f2f0b KB |
4518 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
4519 | blk_mq_realloc_hw_ctxs(set, q); | |
e01ad46d | 4520 | if (q->nr_hw_queues != set->nr_hw_queues) { |
a846a8e6 YB |
4521 | int i = prev_nr_hw_queues; |
4522 | ||
e01ad46d JW |
4523 | pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n", |
4524 | nr_hw_queues, prev_nr_hw_queues); | |
a846a8e6 YB |
4525 | for (; i < set->nr_hw_queues; i++) |
4526 | __blk_mq_free_map_and_rqs(set, i); | |
4527 | ||
e01ad46d | 4528 | set->nr_hw_queues = prev_nr_hw_queues; |
7d76f856 | 4529 | blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]); |
e01ad46d JW |
4530 | goto fallback; |
4531 | } | |
477e19de JW |
4532 | blk_mq_map_swqueue(q); |
4533 | } | |
4534 | ||
f7e76dbc | 4535 | reregister: |
477e19de JW |
4536 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
4537 | blk_mq_sysfs_register(q); | |
4538 | blk_mq_debugfs_register_hctxs(q); | |
868f2f0b KB |
4539 | } |
4540 | ||
d48ece20 JW |
4541 | switch_back: |
4542 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
4543 | blk_mq_elv_switch_back(&head, q); | |
4544 | ||
868f2f0b KB |
4545 | list_for_each_entry(q, &set->tag_list, tag_set_list) |
4546 | blk_mq_unfreeze_queue(q); | |
4547 | } | |
e4dc2b32 KB |
4548 | |
4549 | void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues) | |
4550 | { | |
4551 | mutex_lock(&set->tag_list_lock); | |
4552 | __blk_mq_update_nr_hw_queues(set, nr_hw_queues); | |
4553 | mutex_unlock(&set->tag_list_lock); | |
4554 | } | |
868f2f0b KB |
4555 | EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues); |
4556 | ||
34dbad5d OS |
4557 | /* Enable polling stats and return whether they were already enabled. */ |
4558 | static bool blk_poll_stats_enable(struct request_queue *q) | |
4559 | { | |
48b5c1fb | 4560 | if (q->poll_stat) |
34dbad5d | 4561 | return true; |
48b5c1fb JA |
4562 | |
4563 | return blk_stats_alloc_enable(q); | |
34dbad5d OS |
4564 | } |
4565 | ||
4566 | static void blk_mq_poll_stats_start(struct request_queue *q) | |
4567 | { | |
4568 | /* | |
4569 | * We don't arm the callback if polling stats are not enabled or the | |
4570 | * callback is already active. | |
4571 | */ | |
48b5c1fb | 4572 | if (!q->poll_stat || blk_stat_is_active(q->poll_cb)) |
34dbad5d OS |
4573 | return; |
4574 | ||
4575 | blk_stat_activate_msecs(q->poll_cb, 100); | |
4576 | } | |
4577 | ||
4578 | static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb) | |
4579 | { | |
4580 | struct request_queue *q = cb->data; | |
720b8ccc | 4581 | int bucket; |
34dbad5d | 4582 | |
720b8ccc SB |
4583 | for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) { |
4584 | if (cb->stat[bucket].nr_samples) | |
4585 | q->poll_stat[bucket] = cb->stat[bucket]; | |
4586 | } | |
34dbad5d OS |
4587 | } |
4588 | ||
64f1c21e | 4589 | static unsigned long blk_mq_poll_nsecs(struct request_queue *q, |
64f1c21e JA |
4590 | struct request *rq) |
4591 | { | |
64f1c21e | 4592 | unsigned long ret = 0; |
720b8ccc | 4593 | int bucket; |
64f1c21e JA |
4594 | |
4595 | /* | |
4596 | * If stats collection isn't on, don't sleep but turn it on for | |
4597 | * future users | |
4598 | */ | |
34dbad5d | 4599 | if (!blk_poll_stats_enable(q)) |
64f1c21e JA |
4600 | return 0; |
4601 | ||
64f1c21e JA |
4602 | /* |
4603 | * As an optimistic guess, use half of the mean service time | |
4604 | * for this type of request. We can (and should) make this smarter. | |
4605 | * For instance, if the completion latencies are tight, we can | |
4606 | * get closer than just half the mean. This is especially | |
4607 | * important on devices where the completion latencies are longer | |
720b8ccc SB |
4608 | * than ~10 usec. We do use the stats for the relevant IO size |
4609 | * if available which does lead to better estimates. | |
64f1c21e | 4610 | */ |
720b8ccc SB |
4611 | bucket = blk_mq_poll_stats_bkt(rq); |
4612 | if (bucket < 0) | |
4613 | return ret; | |
4614 | ||
4615 | if (q->poll_stat[bucket].nr_samples) | |
4616 | ret = (q->poll_stat[bucket].mean + 1) / 2; | |
64f1c21e JA |
4617 | |
4618 | return ret; | |
4619 | } | |
4620 | ||
c6699d6f | 4621 | static bool blk_mq_poll_hybrid(struct request_queue *q, blk_qc_t qc) |
06426adf | 4622 | { |
c6699d6f CH |
4623 | struct blk_mq_hw_ctx *hctx = blk_qc_to_hctx(q, qc); |
4624 | struct request *rq = blk_qc_to_rq(hctx, qc); | |
06426adf JA |
4625 | struct hrtimer_sleeper hs; |
4626 | enum hrtimer_mode mode; | |
64f1c21e | 4627 | unsigned int nsecs; |
06426adf JA |
4628 | ktime_t kt; |
4629 | ||
c6699d6f CH |
4630 | /* |
4631 | * If a request has completed on queue that uses an I/O scheduler, we | |
4632 | * won't get back a request from blk_qc_to_rq. | |
4633 | */ | |
4634 | if (!rq || (rq->rq_flags & RQF_MQ_POLL_SLEPT)) | |
64f1c21e JA |
4635 | return false; |
4636 | ||
4637 | /* | |
1052b8ac | 4638 | * If we get here, hybrid polling is enabled. Hence poll_nsec can be: |
64f1c21e | 4639 | * |
64f1c21e JA |
4640 | * 0: use half of prev avg |
4641 | * >0: use this specific value | |
4642 | */ | |
1052b8ac | 4643 | if (q->poll_nsec > 0) |
64f1c21e JA |
4644 | nsecs = q->poll_nsec; |
4645 | else | |
cae740a0 | 4646 | nsecs = blk_mq_poll_nsecs(q, rq); |
64f1c21e JA |
4647 | |
4648 | if (!nsecs) | |
06426adf JA |
4649 | return false; |
4650 | ||
76a86f9d | 4651 | rq->rq_flags |= RQF_MQ_POLL_SLEPT; |
06426adf JA |
4652 | |
4653 | /* | |
4654 | * This will be replaced with the stats tracking code, using | |
4655 | * 'avg_completion_time / 2' as the pre-sleep target. | |
4656 | */ | |
8b0e1953 | 4657 | kt = nsecs; |
06426adf JA |
4658 | |
4659 | mode = HRTIMER_MODE_REL; | |
dbc1625f | 4660 | hrtimer_init_sleeper_on_stack(&hs, CLOCK_MONOTONIC, mode); |
06426adf JA |
4661 | hrtimer_set_expires(&hs.timer, kt); |
4662 | ||
06426adf | 4663 | do { |
5a61c363 | 4664 | if (blk_mq_rq_state(rq) == MQ_RQ_COMPLETE) |
06426adf JA |
4665 | break; |
4666 | set_current_state(TASK_UNINTERRUPTIBLE); | |
9dd8813e | 4667 | hrtimer_sleeper_start_expires(&hs, mode); |
06426adf JA |
4668 | if (hs.task) |
4669 | io_schedule(); | |
4670 | hrtimer_cancel(&hs.timer); | |
4671 | mode = HRTIMER_MODE_ABS; | |
4672 | } while (hs.task && !signal_pending(current)); | |
4673 | ||
4674 | __set_current_state(TASK_RUNNING); | |
4675 | destroy_hrtimer_on_stack(&hs.timer); | |
1052b8ac | 4676 | |
06426adf | 4677 | /* |
c6699d6f CH |
4678 | * If we sleep, have the caller restart the poll loop to reset the |
4679 | * state. Like for the other success return cases, the caller is | |
4680 | * responsible for checking if the IO completed. If the IO isn't | |
4681 | * complete, we'll get called again and will go straight to the busy | |
4682 | * poll loop. | |
06426adf | 4683 | */ |
06426adf JA |
4684 | return true; |
4685 | } | |
06426adf | 4686 | |
c6699d6f | 4687 | static int blk_mq_poll_classic(struct request_queue *q, blk_qc_t cookie, |
5a72e899 | 4688 | struct io_comp_batch *iob, unsigned int flags) |
bbd7bb70 | 4689 | { |
c6699d6f CH |
4690 | struct blk_mq_hw_ctx *hctx = blk_qc_to_hctx(q, cookie); |
4691 | long state = get_current_state(); | |
4692 | int ret; | |
bbd7bb70 | 4693 | |
aa61bec3 | 4694 | do { |
5a72e899 | 4695 | ret = q->mq_ops->poll(hctx, iob); |
bbd7bb70 | 4696 | if (ret > 0) { |
849a3700 | 4697 | __set_current_state(TASK_RUNNING); |
85f4d4b6 | 4698 | return ret; |
bbd7bb70 JA |
4699 | } |
4700 | ||
4701 | if (signal_pending_state(state, current)) | |
849a3700 | 4702 | __set_current_state(TASK_RUNNING); |
b03fbd4f | 4703 | if (task_is_running(current)) |
85f4d4b6 | 4704 | return 1; |
c6699d6f | 4705 | |
ef99b2d3 | 4706 | if (ret < 0 || (flags & BLK_POLL_ONESHOT)) |
bbd7bb70 JA |
4707 | break; |
4708 | cpu_relax(); | |
aa61bec3 | 4709 | } while (!need_resched()); |
bbd7bb70 | 4710 | |
67b4110f | 4711 | __set_current_state(TASK_RUNNING); |
85f4d4b6 | 4712 | return 0; |
bbd7bb70 | 4713 | } |
1052b8ac | 4714 | |
5a72e899 JA |
4715 | int blk_mq_poll(struct request_queue *q, blk_qc_t cookie, struct io_comp_batch *iob, |
4716 | unsigned int flags) | |
1052b8ac | 4717 | { |
d729cf9a CH |
4718 | if (!(flags & BLK_POLL_NOSLEEP) && |
4719 | q->poll_nsec != BLK_MQ_POLL_CLASSIC) { | |
c6699d6f | 4720 | if (blk_mq_poll_hybrid(q, cookie)) |
85f4d4b6 | 4721 | return 1; |
c6699d6f | 4722 | } |
5a72e899 | 4723 | return blk_mq_poll_classic(q, cookie, iob, flags); |
bbd7bb70 JA |
4724 | } |
4725 | ||
9cf2bab6 JA |
4726 | unsigned int blk_mq_rq_cpu(struct request *rq) |
4727 | { | |
4728 | return rq->mq_ctx->cpu; | |
4729 | } | |
4730 | EXPORT_SYMBOL(blk_mq_rq_cpu); | |
4731 | ||
2a19b28f ML |
4732 | void blk_mq_cancel_work_sync(struct request_queue *q) |
4733 | { | |
4734 | if (queue_is_mq(q)) { | |
4735 | struct blk_mq_hw_ctx *hctx; | |
4736 | int i; | |
4737 | ||
4738 | cancel_delayed_work_sync(&q->requeue_work); | |
4739 | ||
4740 | queue_for_each_hw_ctx(q, hctx, i) | |
4741 | cancel_delayed_work_sync(&hctx->run_work); | |
4742 | } | |
4743 | } | |
4744 | ||
320ae51f JA |
4745 | static int __init blk_mq_init(void) |
4746 | { | |
c3077b5d CH |
4747 | int i; |
4748 | ||
4749 | for_each_possible_cpu(i) | |
f9ab4918 | 4750 | init_llist_head(&per_cpu(blk_cpu_done, i)); |
c3077b5d CH |
4751 | open_softirq(BLOCK_SOFTIRQ, blk_done_softirq); |
4752 | ||
4753 | cpuhp_setup_state_nocalls(CPUHP_BLOCK_SOFTIRQ_DEAD, | |
4754 | "block/softirq:dead", NULL, | |
4755 | blk_softirq_cpu_dead); | |
9467f859 TG |
4756 | cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL, |
4757 | blk_mq_hctx_notify_dead); | |
bf0beec0 ML |
4758 | cpuhp_setup_state_multi(CPUHP_AP_BLK_MQ_ONLINE, "block/mq:online", |
4759 | blk_mq_hctx_notify_online, | |
4760 | blk_mq_hctx_notify_offline); | |
320ae51f JA |
4761 | return 0; |
4762 | } | |
4763 | subsys_initcall(blk_mq_init); |