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" |
82b74cac | 45 | #include "blk-ioprio.h" |
320ae51f | 46 | |
f9ab4918 | 47 | static DEFINE_PER_CPU(struct llist_head, blk_cpu_done); |
c3077b5d | 48 | |
34dbad5d OS |
49 | static void blk_mq_poll_stats_start(struct request_queue *q); |
50 | static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb); | |
51 | ||
720b8ccc SB |
52 | static int blk_mq_poll_stats_bkt(const struct request *rq) |
53 | { | |
3d244306 | 54 | int ddir, sectors, bucket; |
720b8ccc | 55 | |
99c749a4 | 56 | ddir = rq_data_dir(rq); |
3d244306 | 57 | sectors = blk_rq_stats_sectors(rq); |
720b8ccc | 58 | |
3d244306 | 59 | bucket = ddir + 2 * ilog2(sectors); |
720b8ccc SB |
60 | |
61 | if (bucket < 0) | |
62 | return -1; | |
63 | else if (bucket >= BLK_MQ_POLL_STATS_BKTS) | |
64 | return ddir + BLK_MQ_POLL_STATS_BKTS - 2; | |
65 | ||
66 | return bucket; | |
67 | } | |
68 | ||
3e08773c CH |
69 | #define BLK_QC_T_SHIFT 16 |
70 | #define BLK_QC_T_INTERNAL (1U << 31) | |
71 | ||
f70299f0 CH |
72 | static inline struct blk_mq_hw_ctx *blk_qc_to_hctx(struct request_queue *q, |
73 | blk_qc_t qc) | |
74 | { | |
4e5cc99e ML |
75 | return xa_load(&q->hctx_table, |
76 | (qc & ~BLK_QC_T_INTERNAL) >> BLK_QC_T_SHIFT); | |
f70299f0 CH |
77 | } |
78 | ||
c6699d6f CH |
79 | static inline struct request *blk_qc_to_rq(struct blk_mq_hw_ctx *hctx, |
80 | blk_qc_t qc) | |
81 | { | |
efbabbe1 CH |
82 | unsigned int tag = qc & ((1U << BLK_QC_T_SHIFT) - 1); |
83 | ||
84 | if (qc & BLK_QC_T_INTERNAL) | |
85 | return blk_mq_tag_to_rq(hctx->sched_tags, tag); | |
86 | return blk_mq_tag_to_rq(hctx->tags, tag); | |
c6699d6f CH |
87 | } |
88 | ||
3e08773c CH |
89 | static inline blk_qc_t blk_rq_to_qc(struct request *rq) |
90 | { | |
91 | return (rq->mq_hctx->queue_num << BLK_QC_T_SHIFT) | | |
92 | (rq->tag != -1 ? | |
93 | rq->tag : (rq->internal_tag | BLK_QC_T_INTERNAL)); | |
94 | } | |
95 | ||
320ae51f | 96 | /* |
85fae294 YY |
97 | * Check if any of the ctx, dispatch list or elevator |
98 | * have pending work in this hardware queue. | |
320ae51f | 99 | */ |
79f720a7 | 100 | static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx) |
320ae51f | 101 | { |
79f720a7 JA |
102 | return !list_empty_careful(&hctx->dispatch) || |
103 | sbitmap_any_bit_set(&hctx->ctx_map) || | |
bd166ef1 | 104 | blk_mq_sched_has_work(hctx); |
1429d7c9 JA |
105 | } |
106 | ||
320ae51f JA |
107 | /* |
108 | * Mark this ctx as having pending work in this hardware queue | |
109 | */ | |
110 | static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx, | |
111 | struct blk_mq_ctx *ctx) | |
112 | { | |
f31967f0 JA |
113 | const int bit = ctx->index_hw[hctx->type]; |
114 | ||
115 | if (!sbitmap_test_bit(&hctx->ctx_map, bit)) | |
116 | sbitmap_set_bit(&hctx->ctx_map, bit); | |
1429d7c9 JA |
117 | } |
118 | ||
119 | static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx, | |
120 | struct blk_mq_ctx *ctx) | |
121 | { | |
f31967f0 JA |
122 | const int bit = ctx->index_hw[hctx->type]; |
123 | ||
124 | sbitmap_clear_bit(&hctx->ctx_map, bit); | |
320ae51f JA |
125 | } |
126 | ||
f299b7c7 | 127 | struct mq_inflight { |
8446fe92 | 128 | struct block_device *part; |
a2e80f6f | 129 | unsigned int inflight[2]; |
f299b7c7 JA |
130 | }; |
131 | ||
2dd6532e | 132 | static bool blk_mq_check_inflight(struct request *rq, void *priv) |
f299b7c7 JA |
133 | { |
134 | struct mq_inflight *mi = priv; | |
135 | ||
b81c14ca HW |
136 | if (rq->part && blk_do_io_stat(rq) && |
137 | (!mi->part->bd_partno || rq->part == mi->part) && | |
b0d97557 | 138 | blk_mq_rq_state(rq) == MQ_RQ_IN_FLIGHT) |
bb4e6b14 | 139 | mi->inflight[rq_data_dir(rq)]++; |
7baa8572 JA |
140 | |
141 | return true; | |
f299b7c7 JA |
142 | } |
143 | ||
8446fe92 CH |
144 | unsigned int blk_mq_in_flight(struct request_queue *q, |
145 | struct block_device *part) | |
f299b7c7 | 146 | { |
a2e80f6f | 147 | struct mq_inflight mi = { .part = part }; |
f299b7c7 | 148 | |
f299b7c7 | 149 | blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi); |
e016b782 | 150 | |
a2e80f6f | 151 | return mi.inflight[0] + mi.inflight[1]; |
bf0ddaba OS |
152 | } |
153 | ||
8446fe92 CH |
154 | void blk_mq_in_flight_rw(struct request_queue *q, struct block_device *part, |
155 | unsigned int inflight[2]) | |
bf0ddaba | 156 | { |
a2e80f6f | 157 | struct mq_inflight mi = { .part = part }; |
bf0ddaba | 158 | |
bb4e6b14 | 159 | blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi); |
a2e80f6f PB |
160 | inflight[0] = mi.inflight[0]; |
161 | inflight[1] = mi.inflight[1]; | |
bf0ddaba OS |
162 | } |
163 | ||
1671d522 | 164 | void blk_freeze_queue_start(struct request_queue *q) |
43a5e4e2 | 165 | { |
7996a8b5 BL |
166 | mutex_lock(&q->mq_freeze_lock); |
167 | if (++q->mq_freeze_depth == 1) { | |
3ef28e83 | 168 | percpu_ref_kill(&q->q_usage_counter); |
7996a8b5 | 169 | mutex_unlock(&q->mq_freeze_lock); |
344e9ffc | 170 | if (queue_is_mq(q)) |
055f6e18 | 171 | blk_mq_run_hw_queues(q, false); |
7996a8b5 BL |
172 | } else { |
173 | mutex_unlock(&q->mq_freeze_lock); | |
cddd5d17 | 174 | } |
f3af020b | 175 | } |
1671d522 | 176 | EXPORT_SYMBOL_GPL(blk_freeze_queue_start); |
f3af020b | 177 | |
6bae363e | 178 | void blk_mq_freeze_queue_wait(struct request_queue *q) |
f3af020b | 179 | { |
3ef28e83 | 180 | wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter)); |
43a5e4e2 | 181 | } |
6bae363e | 182 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait); |
43a5e4e2 | 183 | |
f91328c4 KB |
184 | int blk_mq_freeze_queue_wait_timeout(struct request_queue *q, |
185 | unsigned long timeout) | |
186 | { | |
187 | return wait_event_timeout(q->mq_freeze_wq, | |
188 | percpu_ref_is_zero(&q->q_usage_counter), | |
189 | timeout); | |
190 | } | |
191 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout); | |
43a5e4e2 | 192 | |
f3af020b TH |
193 | /* |
194 | * Guarantee no request is in use, so we can change any data structure of | |
195 | * the queue afterward. | |
196 | */ | |
3ef28e83 | 197 | void blk_freeze_queue(struct request_queue *q) |
f3af020b | 198 | { |
3ef28e83 DW |
199 | /* |
200 | * In the !blk_mq case we are only calling this to kill the | |
201 | * q_usage_counter, otherwise this increases the freeze depth | |
202 | * and waits for it to return to zero. For this reason there is | |
203 | * no blk_unfreeze_queue(), and blk_freeze_queue() is not | |
204 | * exported to drivers as the only user for unfreeze is blk_mq. | |
205 | */ | |
1671d522 | 206 | blk_freeze_queue_start(q); |
f3af020b TH |
207 | blk_mq_freeze_queue_wait(q); |
208 | } | |
3ef28e83 DW |
209 | |
210 | void blk_mq_freeze_queue(struct request_queue *q) | |
211 | { | |
212 | /* | |
213 | * ...just an alias to keep freeze and unfreeze actions balanced | |
214 | * in the blk_mq_* namespace | |
215 | */ | |
216 | blk_freeze_queue(q); | |
217 | } | |
c761d96b | 218 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue); |
f3af020b | 219 | |
aec89dc5 | 220 | void __blk_mq_unfreeze_queue(struct request_queue *q, bool force_atomic) |
320ae51f | 221 | { |
7996a8b5 | 222 | mutex_lock(&q->mq_freeze_lock); |
aec89dc5 CH |
223 | if (force_atomic) |
224 | q->q_usage_counter.data->force_atomic = true; | |
7996a8b5 BL |
225 | q->mq_freeze_depth--; |
226 | WARN_ON_ONCE(q->mq_freeze_depth < 0); | |
227 | if (!q->mq_freeze_depth) { | |
bdd63160 | 228 | percpu_ref_resurrect(&q->q_usage_counter); |
320ae51f | 229 | wake_up_all(&q->mq_freeze_wq); |
add703fd | 230 | } |
7996a8b5 | 231 | mutex_unlock(&q->mq_freeze_lock); |
320ae51f | 232 | } |
aec89dc5 CH |
233 | |
234 | void blk_mq_unfreeze_queue(struct request_queue *q) | |
235 | { | |
236 | __blk_mq_unfreeze_queue(q, false); | |
237 | } | |
b4c6a028 | 238 | EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue); |
320ae51f | 239 | |
852ec809 BVA |
240 | /* |
241 | * FIXME: replace the scsi_internal_device_*block_nowait() calls in the | |
242 | * mpt3sas driver such that this function can be removed. | |
243 | */ | |
244 | void blk_mq_quiesce_queue_nowait(struct request_queue *q) | |
245 | { | |
e70feb8b ML |
246 | unsigned long flags; |
247 | ||
248 | spin_lock_irqsave(&q->queue_lock, flags); | |
249 | if (!q->quiesce_depth++) | |
250 | blk_queue_flag_set(QUEUE_FLAG_QUIESCED, q); | |
251 | spin_unlock_irqrestore(&q->queue_lock, flags); | |
852ec809 BVA |
252 | } |
253 | EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait); | |
254 | ||
6a83e74d | 255 | /** |
9ef4d020 | 256 | * blk_mq_wait_quiesce_done() - wait until in-progress quiesce is done |
6a83e74d BVA |
257 | * @q: request queue. |
258 | * | |
9ef4d020 ML |
259 | * Note: it is driver's responsibility for making sure that quiesce has |
260 | * been started. | |
6a83e74d | 261 | */ |
9ef4d020 | 262 | void blk_mq_wait_quiesce_done(struct request_queue *q) |
6a83e74d | 263 | { |
704b914f ML |
264 | if (blk_queue_has_srcu(q)) |
265 | synchronize_srcu(q->srcu); | |
266 | else | |
6a83e74d BVA |
267 | synchronize_rcu(); |
268 | } | |
9ef4d020 ML |
269 | EXPORT_SYMBOL_GPL(blk_mq_wait_quiesce_done); |
270 | ||
271 | /** | |
272 | * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished | |
273 | * @q: request queue. | |
274 | * | |
275 | * Note: this function does not prevent that the struct request end_io() | |
276 | * callback function is invoked. Once this function is returned, we make | |
277 | * sure no dispatch can happen until the queue is unquiesced via | |
278 | * blk_mq_unquiesce_queue(). | |
279 | */ | |
280 | void blk_mq_quiesce_queue(struct request_queue *q) | |
281 | { | |
282 | blk_mq_quiesce_queue_nowait(q); | |
283 | blk_mq_wait_quiesce_done(q); | |
284 | } | |
6a83e74d BVA |
285 | EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue); |
286 | ||
e4e73913 ML |
287 | /* |
288 | * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue() | |
289 | * @q: request queue. | |
290 | * | |
291 | * This function recovers queue into the state before quiescing | |
292 | * which is done by blk_mq_quiesce_queue. | |
293 | */ | |
294 | void blk_mq_unquiesce_queue(struct request_queue *q) | |
295 | { | |
e70feb8b ML |
296 | unsigned long flags; |
297 | bool run_queue = false; | |
298 | ||
299 | spin_lock_irqsave(&q->queue_lock, flags); | |
300 | if (WARN_ON_ONCE(q->quiesce_depth <= 0)) { | |
301 | ; | |
302 | } else if (!--q->quiesce_depth) { | |
303 | blk_queue_flag_clear(QUEUE_FLAG_QUIESCED, q); | |
304 | run_queue = true; | |
305 | } | |
306 | spin_unlock_irqrestore(&q->queue_lock, flags); | |
f4560ffe | 307 | |
1d9e9bc6 | 308 | /* dispatch requests which are inserted during quiescing */ |
e70feb8b ML |
309 | if (run_queue) |
310 | blk_mq_run_hw_queues(q, true); | |
e4e73913 ML |
311 | } |
312 | EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue); | |
313 | ||
aed3ea94 JA |
314 | void blk_mq_wake_waiters(struct request_queue *q) |
315 | { | |
316 | struct blk_mq_hw_ctx *hctx; | |
4f481208 | 317 | unsigned long i; |
aed3ea94 JA |
318 | |
319 | queue_for_each_hw_ctx(q, hctx, i) | |
320 | if (blk_mq_hw_queue_mapped(hctx)) | |
321 | blk_mq_tag_wakeup_all(hctx->tags, true); | |
322 | } | |
323 | ||
52fdbbcc CH |
324 | void blk_rq_init(struct request_queue *q, struct request *rq) |
325 | { | |
326 | memset(rq, 0, sizeof(*rq)); | |
327 | ||
328 | INIT_LIST_HEAD(&rq->queuelist); | |
329 | rq->q = q; | |
330 | rq->__sector = (sector_t) -1; | |
331 | INIT_HLIST_NODE(&rq->hash); | |
332 | RB_CLEAR_NODE(&rq->rb_node); | |
333 | rq->tag = BLK_MQ_NO_TAG; | |
334 | rq->internal_tag = BLK_MQ_NO_TAG; | |
335 | rq->start_time_ns = ktime_get_ns(); | |
336 | rq->part = NULL; | |
337 | blk_crypto_rq_set_defaults(rq); | |
338 | } | |
339 | EXPORT_SYMBOL(blk_rq_init); | |
340 | ||
e4cdf1a1 | 341 | static struct request *blk_mq_rq_ctx_init(struct blk_mq_alloc_data *data, |
fe6134f6 | 342 | struct blk_mq_tags *tags, unsigned int tag, u64 alloc_time_ns) |
320ae51f | 343 | { |
605f784e PB |
344 | struct blk_mq_ctx *ctx = data->ctx; |
345 | struct blk_mq_hw_ctx *hctx = data->hctx; | |
346 | struct request_queue *q = data->q; | |
e4cdf1a1 | 347 | struct request *rq = tags->static_rqs[tag]; |
c3a148d2 | 348 | |
c7b84d42 JA |
349 | rq->q = q; |
350 | rq->mq_ctx = ctx; | |
351 | rq->mq_hctx = hctx; | |
352 | rq->cmd_flags = data->cmd_flags; | |
353 | ||
354 | if (data->flags & BLK_MQ_REQ_PM) | |
355 | data->rq_flags |= RQF_PM; | |
356 | if (blk_queue_io_stat(q)) | |
357 | data->rq_flags |= RQF_IO_STAT; | |
358 | rq->rq_flags = data->rq_flags; | |
359 | ||
56f8da64 | 360 | if (!(data->rq_flags & RQF_ELV)) { |
e4cdf1a1 | 361 | rq->tag = tag; |
76647368 | 362 | rq->internal_tag = BLK_MQ_NO_TAG; |
56f8da64 JA |
363 | } else { |
364 | rq->tag = BLK_MQ_NO_TAG; | |
365 | rq->internal_tag = tag; | |
e4cdf1a1 | 366 | } |
c7b84d42 | 367 | rq->timeout = 0; |
e4cdf1a1 | 368 | |
4f266f2b PB |
369 | if (blk_mq_need_time_stamp(rq)) |
370 | rq->start_time_ns = ktime_get_ns(); | |
371 | else | |
372 | rq->start_time_ns = 0; | |
af76e555 | 373 | rq->part = NULL; |
6f816b4b TH |
374 | #ifdef CONFIG_BLK_RQ_ALLOC_TIME |
375 | rq->alloc_time_ns = alloc_time_ns; | |
376 | #endif | |
544ccc8d | 377 | rq->io_start_time_ns = 0; |
3d244306 | 378 | rq->stats_sectors = 0; |
af76e555 CH |
379 | rq->nr_phys_segments = 0; |
380 | #if defined(CONFIG_BLK_DEV_INTEGRITY) | |
381 | rq->nr_integrity_segments = 0; | |
382 | #endif | |
af76e555 CH |
383 | rq->end_io = NULL; |
384 | rq->end_io_data = NULL; | |
af76e555 | 385 | |
4f266f2b PB |
386 | blk_crypto_rq_set_defaults(rq); |
387 | INIT_LIST_HEAD(&rq->queuelist); | |
388 | /* tag was already set */ | |
389 | WRITE_ONCE(rq->deadline, 0); | |
0a467d0f | 390 | req_ref_set(rq, 1); |
7ea4d8a4 | 391 | |
4f266f2b | 392 | if (rq->rq_flags & RQF_ELV) { |
7ea4d8a4 CH |
393 | struct elevator_queue *e = data->q->elevator; |
394 | ||
4f266f2b PB |
395 | INIT_HLIST_NODE(&rq->hash); |
396 | RB_CLEAR_NODE(&rq->rb_node); | |
397 | ||
398 | if (!op_is_flush(data->cmd_flags) && | |
399 | e->type->ops.prepare_request) { | |
7ea4d8a4 CH |
400 | e->type->ops.prepare_request(rq); |
401 | rq->rq_flags |= RQF_ELVPRIV; | |
402 | } | |
403 | } | |
404 | ||
e4cdf1a1 | 405 | return rq; |
5dee8577 CH |
406 | } |
407 | ||
349302da JA |
408 | static inline struct request * |
409 | __blk_mq_alloc_requests_batch(struct blk_mq_alloc_data *data, | |
410 | u64 alloc_time_ns) | |
411 | { | |
412 | unsigned int tag, tag_offset; | |
fe6134f6 | 413 | struct blk_mq_tags *tags; |
349302da | 414 | struct request *rq; |
fe6134f6 | 415 | unsigned long tag_mask; |
349302da JA |
416 | int i, nr = 0; |
417 | ||
fe6134f6 JA |
418 | tag_mask = blk_mq_get_tags(data, data->nr_tags, &tag_offset); |
419 | if (unlikely(!tag_mask)) | |
349302da JA |
420 | return NULL; |
421 | ||
fe6134f6 JA |
422 | tags = blk_mq_tags_from_data(data); |
423 | for (i = 0; tag_mask; i++) { | |
424 | if (!(tag_mask & (1UL << i))) | |
349302da JA |
425 | continue; |
426 | tag = tag_offset + i; | |
a22c00be | 427 | prefetch(tags->static_rqs[tag]); |
fe6134f6 JA |
428 | tag_mask &= ~(1UL << i); |
429 | rq = blk_mq_rq_ctx_init(data, tags, tag, alloc_time_ns); | |
013a7f95 | 430 | rq_list_add(data->cached_rq, rq); |
c5fc7b93 | 431 | nr++; |
349302da | 432 | } |
c5fc7b93 JA |
433 | /* caller already holds a reference, add for remainder */ |
434 | percpu_ref_get_many(&data->q->q_usage_counter, nr - 1); | |
349302da JA |
435 | data->nr_tags -= nr; |
436 | ||
013a7f95 | 437 | return rq_list_pop(data->cached_rq); |
349302da JA |
438 | } |
439 | ||
b90cfaed | 440 | static struct request *__blk_mq_alloc_requests(struct blk_mq_alloc_data *data) |
d2c0d383 | 441 | { |
e6e7abff | 442 | struct request_queue *q = data->q; |
6f816b4b | 443 | u64 alloc_time_ns = 0; |
47c122e3 | 444 | struct request *rq; |
600c3b0c | 445 | unsigned int tag; |
d2c0d383 | 446 | |
6f816b4b TH |
447 | /* alloc_time includes depth and tag waits */ |
448 | if (blk_queue_rq_alloc_time(q)) | |
449 | alloc_time_ns = ktime_get_ns(); | |
450 | ||
f9afca4d | 451 | if (data->cmd_flags & REQ_NOWAIT) |
03a07c92 | 452 | data->flags |= BLK_MQ_REQ_NOWAIT; |
d2c0d383 | 453 | |
781dd830 JA |
454 | if (q->elevator) { |
455 | struct elevator_queue *e = q->elevator; | |
456 | ||
457 | data->rq_flags |= RQF_ELV; | |
458 | ||
d2c0d383 | 459 | /* |
8d663f34 | 460 | * Flush/passthrough requests are special and go directly to the |
17a51199 JA |
461 | * dispatch list. Don't include reserved tags in the |
462 | * limiting, as it isn't useful. | |
d2c0d383 | 463 | */ |
f9afca4d | 464 | if (!op_is_flush(data->cmd_flags) && |
8d663f34 | 465 | !blk_op_is_passthrough(data->cmd_flags) && |
f9afca4d | 466 | e->type->ops.limit_depth && |
17a51199 | 467 | !(data->flags & BLK_MQ_REQ_RESERVED)) |
f9afca4d | 468 | e->type->ops.limit_depth(data->cmd_flags, data); |
d2c0d383 CH |
469 | } |
470 | ||
bf0beec0 | 471 | retry: |
600c3b0c CH |
472 | data->ctx = blk_mq_get_ctx(q); |
473 | data->hctx = blk_mq_map_queue(q, data->cmd_flags, data->ctx); | |
781dd830 | 474 | if (!(data->rq_flags & RQF_ELV)) |
600c3b0c CH |
475 | blk_mq_tag_busy(data->hctx); |
476 | ||
99e48cd6 JG |
477 | if (data->flags & BLK_MQ_REQ_RESERVED) |
478 | data->rq_flags |= RQF_RESV; | |
479 | ||
349302da JA |
480 | /* |
481 | * Try batched alloc if we want more than 1 tag. | |
482 | */ | |
483 | if (data->nr_tags > 1) { | |
484 | rq = __blk_mq_alloc_requests_batch(data, alloc_time_ns); | |
485 | if (rq) | |
486 | return rq; | |
487 | data->nr_tags = 1; | |
488 | } | |
489 | ||
bf0beec0 ML |
490 | /* |
491 | * Waiting allocations only fail because of an inactive hctx. In that | |
492 | * case just retry the hctx assignment and tag allocation as CPU hotplug | |
493 | * should have migrated us to an online CPU by now. | |
494 | */ | |
e4cdf1a1 | 495 | tag = blk_mq_get_tag(data); |
bf0beec0 ML |
496 | if (tag == BLK_MQ_NO_TAG) { |
497 | if (data->flags & BLK_MQ_REQ_NOWAIT) | |
498 | return NULL; | |
bf0beec0 | 499 | /* |
349302da JA |
500 | * Give up the CPU and sleep for a random short time to |
501 | * ensure that thread using a realtime scheduling class | |
502 | * are migrated off the CPU, and thus off the hctx that | |
503 | * is going away. | |
bf0beec0 ML |
504 | */ |
505 | msleep(3); | |
506 | goto retry; | |
507 | } | |
47c122e3 | 508 | |
fe6134f6 JA |
509 | return blk_mq_rq_ctx_init(data, blk_mq_tags_from_data(data), tag, |
510 | alloc_time_ns); | |
d2c0d383 CH |
511 | } |
512 | ||
4b6a5d9c JA |
513 | static struct request *blk_mq_rq_cache_fill(struct request_queue *q, |
514 | struct blk_plug *plug, | |
515 | blk_opf_t opf, | |
516 | blk_mq_req_flags_t flags) | |
320ae51f | 517 | { |
e6e7abff CH |
518 | struct blk_mq_alloc_data data = { |
519 | .q = q, | |
520 | .flags = flags, | |
16458cf3 | 521 | .cmd_flags = opf, |
4b6a5d9c JA |
522 | .nr_tags = plug->nr_ios, |
523 | .cached_rq = &plug->cached_rq, | |
e6e7abff | 524 | }; |
bd166ef1 | 525 | struct request *rq; |
320ae51f | 526 | |
4b6a5d9c JA |
527 | if (blk_queue_enter(q, flags)) |
528 | return NULL; | |
529 | ||
530 | plug->nr_ios = 1; | |
320ae51f | 531 | |
b90cfaed | 532 | rq = __blk_mq_alloc_requests(&data); |
4b6a5d9c JA |
533 | if (unlikely(!rq)) |
534 | blk_queue_exit(q); | |
535 | return rq; | |
536 | } | |
537 | ||
538 | static struct request *blk_mq_alloc_cached_request(struct request_queue *q, | |
539 | blk_opf_t opf, | |
540 | blk_mq_req_flags_t flags) | |
541 | { | |
542 | struct blk_plug *plug = current->plug; | |
543 | struct request *rq; | |
544 | ||
545 | if (!plug) | |
546 | return NULL; | |
547 | if (rq_list_empty(plug->cached_rq)) { | |
548 | if (plug->nr_ios == 1) | |
549 | return NULL; | |
550 | rq = blk_mq_rq_cache_fill(q, plug, opf, flags); | |
551 | if (rq) | |
552 | goto got_it; | |
553 | return NULL; | |
554 | } | |
555 | rq = rq_list_peek(&plug->cached_rq); | |
556 | if (!rq || rq->q != q) | |
557 | return NULL; | |
558 | ||
559 | if (blk_mq_get_hctx_type(opf) != rq->mq_hctx->type) | |
560 | return NULL; | |
561 | if (op_is_flush(rq->cmd_flags) != op_is_flush(opf)) | |
562 | return NULL; | |
563 | ||
564 | plug->cached_rq = rq_list_next(rq); | |
565 | got_it: | |
566 | rq->cmd_flags = opf; | |
567 | INIT_LIST_HEAD(&rq->queuelist); | |
568 | return rq; | |
569 | } | |
570 | ||
571 | struct request *blk_mq_alloc_request(struct request_queue *q, blk_opf_t opf, | |
572 | blk_mq_req_flags_t flags) | |
573 | { | |
574 | struct request *rq; | |
575 | ||
576 | rq = blk_mq_alloc_cached_request(q, opf, flags); | |
577 | if (!rq) { | |
578 | struct blk_mq_alloc_data data = { | |
579 | .q = q, | |
580 | .flags = flags, | |
581 | .cmd_flags = opf, | |
582 | .nr_tags = 1, | |
583 | }; | |
584 | int ret; | |
585 | ||
586 | ret = blk_queue_enter(q, flags); | |
587 | if (ret) | |
588 | return ERR_PTR(ret); | |
589 | ||
590 | rq = __blk_mq_alloc_requests(&data); | |
591 | if (!rq) | |
592 | goto out_queue_exit; | |
593 | } | |
0c4de0f3 CH |
594 | rq->__data_len = 0; |
595 | rq->__sector = (sector_t) -1; | |
596 | rq->bio = rq->biotail = NULL; | |
320ae51f | 597 | return rq; |
a5ea5811 CH |
598 | out_queue_exit: |
599 | blk_queue_exit(q); | |
600 | return ERR_PTR(-EWOULDBLOCK); | |
320ae51f | 601 | } |
4bb659b1 | 602 | EXPORT_SYMBOL(blk_mq_alloc_request); |
320ae51f | 603 | |
cd6ce148 | 604 | struct request *blk_mq_alloc_request_hctx(struct request_queue *q, |
16458cf3 | 605 | blk_opf_t opf, blk_mq_req_flags_t flags, unsigned int hctx_idx) |
1f5bd336 | 606 | { |
e6e7abff CH |
607 | struct blk_mq_alloc_data data = { |
608 | .q = q, | |
609 | .flags = flags, | |
16458cf3 | 610 | .cmd_flags = opf, |
47c122e3 | 611 | .nr_tags = 1, |
e6e7abff | 612 | }; |
600c3b0c | 613 | u64 alloc_time_ns = 0; |
6d2809d5 | 614 | unsigned int cpu; |
600c3b0c | 615 | unsigned int tag; |
1f5bd336 ML |
616 | int ret; |
617 | ||
600c3b0c CH |
618 | /* alloc_time includes depth and tag waits */ |
619 | if (blk_queue_rq_alloc_time(q)) | |
620 | alloc_time_ns = ktime_get_ns(); | |
621 | ||
1f5bd336 ML |
622 | /* |
623 | * If the tag allocator sleeps we could get an allocation for a | |
624 | * different hardware context. No need to complicate the low level | |
625 | * allocator for this for the rare use case of a command tied to | |
626 | * a specific queue. | |
627 | */ | |
600c3b0c | 628 | if (WARN_ON_ONCE(!(flags & (BLK_MQ_REQ_NOWAIT | BLK_MQ_REQ_RESERVED)))) |
1f5bd336 ML |
629 | return ERR_PTR(-EINVAL); |
630 | ||
631 | if (hctx_idx >= q->nr_hw_queues) | |
632 | return ERR_PTR(-EIO); | |
633 | ||
3a0a5299 | 634 | ret = blk_queue_enter(q, flags); |
1f5bd336 ML |
635 | if (ret) |
636 | return ERR_PTR(ret); | |
637 | ||
c8712c6a CH |
638 | /* |
639 | * Check if the hardware context is actually mapped to anything. | |
640 | * If not tell the caller that it should skip this queue. | |
641 | */ | |
a5ea5811 | 642 | ret = -EXDEV; |
4e5cc99e | 643 | data.hctx = xa_load(&q->hctx_table, hctx_idx); |
e6e7abff | 644 | if (!blk_mq_hw_queue_mapped(data.hctx)) |
a5ea5811 | 645 | goto out_queue_exit; |
e6e7abff | 646 | cpu = cpumask_first_and(data.hctx->cpumask, cpu_online_mask); |
14dc7a18 BVA |
647 | if (cpu >= nr_cpu_ids) |
648 | goto out_queue_exit; | |
e6e7abff | 649 | data.ctx = __blk_mq_get_ctx(q, cpu); |
1f5bd336 | 650 | |
42fdc5e4 | 651 | if (!q->elevator) |
600c3b0c | 652 | blk_mq_tag_busy(data.hctx); |
781dd830 JA |
653 | else |
654 | data.rq_flags |= RQF_ELV; | |
600c3b0c | 655 | |
99e48cd6 JG |
656 | if (flags & BLK_MQ_REQ_RESERVED) |
657 | data.rq_flags |= RQF_RESV; | |
658 | ||
a5ea5811 | 659 | ret = -EWOULDBLOCK; |
600c3b0c CH |
660 | tag = blk_mq_get_tag(&data); |
661 | if (tag == BLK_MQ_NO_TAG) | |
a5ea5811 | 662 | goto out_queue_exit; |
fe6134f6 JA |
663 | return blk_mq_rq_ctx_init(&data, blk_mq_tags_from_data(&data), tag, |
664 | alloc_time_ns); | |
600c3b0c | 665 | |
a5ea5811 CH |
666 | out_queue_exit: |
667 | blk_queue_exit(q); | |
668 | return ERR_PTR(ret); | |
1f5bd336 ML |
669 | } |
670 | EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx); | |
671 | ||
12f5b931 KB |
672 | static void __blk_mq_free_request(struct request *rq) |
673 | { | |
674 | struct request_queue *q = rq->q; | |
675 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
ea4f995e | 676 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; |
12f5b931 KB |
677 | const int sched_tag = rq->internal_tag; |
678 | ||
a892c8d5 | 679 | blk_crypto_free_request(rq); |
986d413b | 680 | blk_pm_mark_last_busy(rq); |
ea4f995e | 681 | rq->mq_hctx = NULL; |
76647368 | 682 | if (rq->tag != BLK_MQ_NO_TAG) |
cae740a0 | 683 | blk_mq_put_tag(hctx->tags, ctx, rq->tag); |
76647368 | 684 | if (sched_tag != BLK_MQ_NO_TAG) |
cae740a0 | 685 | blk_mq_put_tag(hctx->sched_tags, ctx, sched_tag); |
12f5b931 KB |
686 | blk_mq_sched_restart(hctx); |
687 | blk_queue_exit(q); | |
688 | } | |
689 | ||
6af54051 | 690 | void blk_mq_free_request(struct request *rq) |
320ae51f | 691 | { |
320ae51f | 692 | struct request_queue *q = rq->q; |
ea4f995e | 693 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; |
6af54051 | 694 | |
222ee581 CH |
695 | if ((rq->rq_flags & RQF_ELVPRIV) && |
696 | q->elevator->type->ops.finish_request) | |
697 | q->elevator->type->ops.finish_request(rq); | |
320ae51f | 698 | |
e8064021 | 699 | if (rq->rq_flags & RQF_MQ_INFLIGHT) |
bccf5e26 | 700 | __blk_mq_dec_active_requests(hctx); |
87760e5e | 701 | |
7beb2f84 | 702 | if (unlikely(laptop_mode && !blk_rq_is_passthrough(rq))) |
d152c682 | 703 | laptop_io_completion(q->disk->bdi); |
7beb2f84 | 704 | |
a7905043 | 705 | rq_qos_done(q, rq); |
0d2602ca | 706 | |
12f5b931 | 707 | WRITE_ONCE(rq->state, MQ_RQ_IDLE); |
0a467d0f | 708 | if (req_ref_put_and_test(rq)) |
12f5b931 | 709 | __blk_mq_free_request(rq); |
320ae51f | 710 | } |
1a3b595a | 711 | EXPORT_SYMBOL_GPL(blk_mq_free_request); |
320ae51f | 712 | |
47c122e3 | 713 | void blk_mq_free_plug_rqs(struct blk_plug *plug) |
320ae51f | 714 | { |
013a7f95 | 715 | struct request *rq; |
fe1f4526 | 716 | |
c5fc7b93 | 717 | while ((rq = rq_list_pop(&plug->cached_rq)) != NULL) |
47c122e3 | 718 | blk_mq_free_request(rq); |
47c122e3 | 719 | } |
522a7775 | 720 | |
22350ad7 CH |
721 | void blk_dump_rq_flags(struct request *rq, char *msg) |
722 | { | |
723 | printk(KERN_INFO "%s: dev %s: flags=%llx\n", msg, | |
f3fa33ac | 724 | rq->q->disk ? rq->q->disk->disk_name : "?", |
16458cf3 | 725 | (__force unsigned long long) rq->cmd_flags); |
22350ad7 CH |
726 | |
727 | printk(KERN_INFO " sector %llu, nr/cnr %u/%u\n", | |
728 | (unsigned long long)blk_rq_pos(rq), | |
729 | blk_rq_sectors(rq), blk_rq_cur_sectors(rq)); | |
730 | printk(KERN_INFO " bio %p, biotail %p, len %u\n", | |
731 | rq->bio, rq->biotail, blk_rq_bytes(rq)); | |
732 | } | |
733 | EXPORT_SYMBOL(blk_dump_rq_flags); | |
734 | ||
9be3e06f JA |
735 | static void req_bio_endio(struct request *rq, struct bio *bio, |
736 | unsigned int nbytes, blk_status_t error) | |
737 | { | |
478eb72b | 738 | if (unlikely(error)) { |
9be3e06f | 739 | bio->bi_status = error; |
478eb72b | 740 | } else if (req_op(rq) == REQ_OP_ZONE_APPEND) { |
9be3e06f JA |
741 | /* |
742 | * Partial zone append completions cannot be supported as the | |
743 | * BIO fragments may end up not being written sequentially. | |
744 | */ | |
297db731 | 745 | if (bio->bi_iter.bi_size != nbytes) |
9be3e06f JA |
746 | bio->bi_status = BLK_STS_IOERR; |
747 | else | |
748 | bio->bi_iter.bi_sector = rq->__sector; | |
749 | } | |
750 | ||
478eb72b PB |
751 | bio_advance(bio, nbytes); |
752 | ||
753 | if (unlikely(rq->rq_flags & RQF_QUIET)) | |
754 | bio_set_flag(bio, BIO_QUIET); | |
9be3e06f JA |
755 | /* don't actually finish bio if it's part of flush sequence */ |
756 | if (bio->bi_iter.bi_size == 0 && !(rq->rq_flags & RQF_FLUSH_SEQ)) | |
757 | bio_endio(bio); | |
758 | } | |
759 | ||
760 | static void blk_account_io_completion(struct request *req, unsigned int bytes) | |
761 | { | |
762 | if (req->part && blk_do_io_stat(req)) { | |
763 | const int sgrp = op_stat_group(req_op(req)); | |
764 | ||
765 | part_stat_lock(); | |
766 | part_stat_add(req->part, sectors[sgrp], bytes >> 9); | |
767 | part_stat_unlock(); | |
768 | } | |
769 | } | |
770 | ||
0d7a29a2 CH |
771 | static void blk_print_req_error(struct request *req, blk_status_t status) |
772 | { | |
773 | printk_ratelimited(KERN_ERR | |
774 | "%s error, dev %s, sector %llu op 0x%x:(%s) flags 0x%x " | |
775 | "phys_seg %u prio class %u\n", | |
776 | blk_status_to_str(status), | |
f3fa33ac | 777 | req->q->disk ? req->q->disk->disk_name : "?", |
16458cf3 BVA |
778 | blk_rq_pos(req), (__force u32)req_op(req), |
779 | blk_op_str(req_op(req)), | |
780 | (__force u32)(req->cmd_flags & ~REQ_OP_MASK), | |
0d7a29a2 CH |
781 | req->nr_phys_segments, |
782 | IOPRIO_PRIO_CLASS(req->ioprio)); | |
783 | } | |
784 | ||
5581a5dd JA |
785 | /* |
786 | * Fully end IO on a request. Does not support partial completions, or | |
787 | * errors. | |
788 | */ | |
789 | static void blk_complete_request(struct request *req) | |
790 | { | |
791 | const bool is_flush = (req->rq_flags & RQF_FLUSH_SEQ) != 0; | |
792 | int total_bytes = blk_rq_bytes(req); | |
793 | struct bio *bio = req->bio; | |
794 | ||
795 | trace_block_rq_complete(req, BLK_STS_OK, total_bytes); | |
796 | ||
797 | if (!bio) | |
798 | return; | |
799 | ||
800 | #ifdef CONFIG_BLK_DEV_INTEGRITY | |
801 | if (blk_integrity_rq(req) && req_op(req) == REQ_OP_READ) | |
802 | req->q->integrity.profile->complete_fn(req, total_bytes); | |
803 | #endif | |
804 | ||
805 | blk_account_io_completion(req, total_bytes); | |
806 | ||
807 | do { | |
808 | struct bio *next = bio->bi_next; | |
809 | ||
810 | /* Completion has already been traced */ | |
811 | bio_clear_flag(bio, BIO_TRACE_COMPLETION); | |
a12821d5 PR |
812 | |
813 | if (req_op(req) == REQ_OP_ZONE_APPEND) | |
814 | bio->bi_iter.bi_sector = req->__sector; | |
815 | ||
5581a5dd JA |
816 | if (!is_flush) |
817 | bio_endio(bio); | |
818 | bio = next; | |
819 | } while (bio); | |
820 | ||
821 | /* | |
822 | * Reset counters so that the request stacking driver | |
823 | * can find how many bytes remain in the request | |
824 | * later. | |
825 | */ | |
ab3e1d3b JA |
826 | if (!req->end_io) { |
827 | req->bio = NULL; | |
828 | req->__data_len = 0; | |
829 | } | |
5581a5dd JA |
830 | } |
831 | ||
9be3e06f JA |
832 | /** |
833 | * blk_update_request - Complete multiple bytes without completing the request | |
834 | * @req: the request being processed | |
835 | * @error: block status code | |
836 | * @nr_bytes: number of bytes to complete for @req | |
837 | * | |
838 | * Description: | |
839 | * Ends I/O on a number of bytes attached to @req, but doesn't complete | |
840 | * the request structure even if @req doesn't have leftover. | |
841 | * If @req has leftover, sets it up for the next range of segments. | |
842 | * | |
843 | * Passing the result of blk_rq_bytes() as @nr_bytes guarantees | |
844 | * %false return from this function. | |
845 | * | |
846 | * Note: | |
847 | * The RQF_SPECIAL_PAYLOAD flag is ignored on purpose in this function | |
848 | * except in the consistency check at the end of this function. | |
849 | * | |
850 | * Return: | |
851 | * %false - this request doesn't have any more data | |
852 | * %true - this request has more data | |
853 | **/ | |
854 | bool blk_update_request(struct request *req, blk_status_t error, | |
855 | unsigned int nr_bytes) | |
856 | { | |
857 | int total_bytes; | |
858 | ||
8a7d267b | 859 | trace_block_rq_complete(req, error, nr_bytes); |
9be3e06f JA |
860 | |
861 | if (!req->bio) | |
862 | return false; | |
863 | ||
864 | #ifdef CONFIG_BLK_DEV_INTEGRITY | |
865 | if (blk_integrity_rq(req) && req_op(req) == REQ_OP_READ && | |
866 | error == BLK_STS_OK) | |
867 | req->q->integrity.profile->complete_fn(req, nr_bytes); | |
868 | #endif | |
869 | ||
870 | if (unlikely(error && !blk_rq_is_passthrough(req) && | |
3d973a76 CH |
871 | !(req->rq_flags & RQF_QUIET)) && |
872 | !test_bit(GD_DEAD, &req->q->disk->state)) { | |
9be3e06f | 873 | blk_print_req_error(req, error); |
d5869fdc YS |
874 | trace_block_rq_error(req, error, nr_bytes); |
875 | } | |
9be3e06f JA |
876 | |
877 | blk_account_io_completion(req, nr_bytes); | |
878 | ||
879 | total_bytes = 0; | |
880 | while (req->bio) { | |
881 | struct bio *bio = req->bio; | |
882 | unsigned bio_bytes = min(bio->bi_iter.bi_size, nr_bytes); | |
883 | ||
884 | if (bio_bytes == bio->bi_iter.bi_size) | |
885 | req->bio = bio->bi_next; | |
886 | ||
887 | /* Completion has already been traced */ | |
888 | bio_clear_flag(bio, BIO_TRACE_COMPLETION); | |
889 | req_bio_endio(req, bio, bio_bytes, error); | |
890 | ||
891 | total_bytes += bio_bytes; | |
892 | nr_bytes -= bio_bytes; | |
893 | ||
894 | if (!nr_bytes) | |
895 | break; | |
896 | } | |
897 | ||
898 | /* | |
899 | * completely done | |
900 | */ | |
901 | if (!req->bio) { | |
902 | /* | |
903 | * Reset counters so that the request stacking driver | |
904 | * can find how many bytes remain in the request | |
905 | * later. | |
906 | */ | |
907 | req->__data_len = 0; | |
908 | return false; | |
909 | } | |
910 | ||
911 | req->__data_len -= total_bytes; | |
912 | ||
913 | /* update sector only for requests with clear definition of sector */ | |
914 | if (!blk_rq_is_passthrough(req)) | |
915 | req->__sector += total_bytes >> 9; | |
916 | ||
917 | /* mixed attributes always follow the first bio */ | |
918 | if (req->rq_flags & RQF_MIXED_MERGE) { | |
919 | req->cmd_flags &= ~REQ_FAILFAST_MASK; | |
920 | req->cmd_flags |= req->bio->bi_opf & REQ_FAILFAST_MASK; | |
921 | } | |
922 | ||
923 | if (!(req->rq_flags & RQF_SPECIAL_PAYLOAD)) { | |
924 | /* | |
925 | * If total number of sectors is less than the first segment | |
926 | * size, something has gone terribly wrong. | |
927 | */ | |
928 | if (blk_rq_bytes(req) < blk_rq_cur_bytes(req)) { | |
929 | blk_dump_rq_flags(req, "request botched"); | |
930 | req->__data_len = blk_rq_cur_bytes(req); | |
931 | } | |
932 | ||
933 | /* recalculate the number of segments */ | |
934 | req->nr_phys_segments = blk_recalc_rq_segments(req); | |
935 | } | |
936 | ||
937 | return true; | |
938 | } | |
939 | EXPORT_SYMBOL_GPL(blk_update_request); | |
940 | ||
450b7879 CH |
941 | static void __blk_account_io_done(struct request *req, u64 now) |
942 | { | |
943 | const int sgrp = op_stat_group(req_op(req)); | |
944 | ||
945 | part_stat_lock(); | |
946 | update_io_ticks(req->part, jiffies, true); | |
947 | part_stat_inc(req->part, ios[sgrp]); | |
948 | part_stat_add(req->part, nsecs[sgrp], now - req->start_time_ns); | |
949 | part_stat_unlock(); | |
950 | } | |
951 | ||
952 | static inline void blk_account_io_done(struct request *req, u64 now) | |
953 | { | |
954 | /* | |
955 | * Account IO completion. flush_rq isn't accounted as a | |
956 | * normal IO on queueing nor completion. Accounting the | |
957 | * containing request is enough. | |
958 | */ | |
959 | if (blk_do_io_stat(req) && req->part && | |
960 | !(req->rq_flags & RQF_FLUSH_SEQ)) | |
961 | __blk_account_io_done(req, now); | |
962 | } | |
963 | ||
964 | static void __blk_account_io_start(struct request *rq) | |
965 | { | |
41fa7222 CH |
966 | /* |
967 | * All non-passthrough requests are created from a bio with one | |
968 | * exception: when a flush command that is part of a flush sequence | |
969 | * generated by the state machine in blk-flush.c is cloned onto the | |
970 | * lower device by dm-multipath we can get here without a bio. | |
971 | */ | |
972 | if (rq->bio) | |
450b7879 | 973 | rq->part = rq->bio->bi_bdev; |
41fa7222 | 974 | else |
f3fa33ac | 975 | rq->part = rq->q->disk->part0; |
450b7879 CH |
976 | |
977 | part_stat_lock(); | |
978 | update_io_ticks(rq->part, jiffies, false); | |
979 | part_stat_unlock(); | |
980 | } | |
981 | ||
982 | static inline void blk_account_io_start(struct request *req) | |
983 | { | |
984 | if (blk_do_io_stat(req)) | |
985 | __blk_account_io_start(req); | |
986 | } | |
987 | ||
f794f335 | 988 | static inline void __blk_mq_end_request_acct(struct request *rq, u64 now) |
320ae51f | 989 | { |
4bc6339a OS |
990 | if (rq->rq_flags & RQF_STATS) { |
991 | blk_mq_poll_stats_start(rq->q); | |
522a7775 | 992 | blk_stat_add(rq, now); |
4bc6339a OS |
993 | } |
994 | ||
87890092 | 995 | blk_mq_sched_completed_request(rq, now); |
522a7775 | 996 | blk_account_io_done(rq, now); |
f794f335 | 997 | } |
522a7775 | 998 | |
f794f335 JA |
999 | inline void __blk_mq_end_request(struct request *rq, blk_status_t error) |
1000 | { | |
1001 | if (blk_mq_need_time_stamp(rq)) | |
1002 | __blk_mq_end_request_acct(rq, ktime_get_ns()); | |
0d11e6ac | 1003 | |
91b63639 | 1004 | if (rq->end_io) { |
a7905043 | 1005 | rq_qos_done(rq->q, rq); |
de671d61 JA |
1006 | if (rq->end_io(rq, error) == RQ_END_IO_FREE) |
1007 | blk_mq_free_request(rq); | |
91b63639 | 1008 | } else { |
320ae51f | 1009 | blk_mq_free_request(rq); |
91b63639 | 1010 | } |
320ae51f | 1011 | } |
c8a446ad | 1012 | EXPORT_SYMBOL(__blk_mq_end_request); |
63151a44 | 1013 | |
2a842aca | 1014 | void blk_mq_end_request(struct request *rq, blk_status_t error) |
63151a44 CH |
1015 | { |
1016 | if (blk_update_request(rq, error, blk_rq_bytes(rq))) | |
1017 | BUG(); | |
c8a446ad | 1018 | __blk_mq_end_request(rq, error); |
63151a44 | 1019 | } |
c8a446ad | 1020 | EXPORT_SYMBOL(blk_mq_end_request); |
320ae51f | 1021 | |
f794f335 JA |
1022 | #define TAG_COMP_BATCH 32 |
1023 | ||
1024 | static inline void blk_mq_flush_tag_batch(struct blk_mq_hw_ctx *hctx, | |
1025 | int *tag_array, int nr_tags) | |
1026 | { | |
1027 | struct request_queue *q = hctx->queue; | |
1028 | ||
3b87c6ea ML |
1029 | /* |
1030 | * All requests should have been marked as RQF_MQ_INFLIGHT, so | |
1031 | * update hctx->nr_active in batch | |
1032 | */ | |
1033 | if (hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) | |
1034 | __blk_mq_sub_active_requests(hctx, nr_tags); | |
1035 | ||
f794f335 JA |
1036 | blk_mq_put_tags(hctx->tags, tag_array, nr_tags); |
1037 | percpu_ref_put_many(&q->q_usage_counter, nr_tags); | |
1038 | } | |
1039 | ||
1040 | void blk_mq_end_request_batch(struct io_comp_batch *iob) | |
1041 | { | |
1042 | int tags[TAG_COMP_BATCH], nr_tags = 0; | |
02f7eab0 | 1043 | struct blk_mq_hw_ctx *cur_hctx = NULL; |
f794f335 JA |
1044 | struct request *rq; |
1045 | u64 now = 0; | |
1046 | ||
1047 | if (iob->need_ts) | |
1048 | now = ktime_get_ns(); | |
1049 | ||
1050 | while ((rq = rq_list_pop(&iob->req_list)) != NULL) { | |
1051 | prefetch(rq->bio); | |
1052 | prefetch(rq->rq_next); | |
1053 | ||
5581a5dd | 1054 | blk_complete_request(rq); |
f794f335 JA |
1055 | if (iob->need_ts) |
1056 | __blk_mq_end_request_acct(rq, now); | |
1057 | ||
98b26a0e JA |
1058 | rq_qos_done(rq->q, rq); |
1059 | ||
ab3e1d3b JA |
1060 | /* |
1061 | * If end_io handler returns NONE, then it still has | |
1062 | * ownership of the request. | |
1063 | */ | |
1064 | if (rq->end_io && rq->end_io(rq, 0) == RQ_END_IO_NONE) | |
1065 | continue; | |
1066 | ||
f794f335 | 1067 | WRITE_ONCE(rq->state, MQ_RQ_IDLE); |
0a467d0f | 1068 | if (!req_ref_put_and_test(rq)) |
f794f335 JA |
1069 | continue; |
1070 | ||
1071 | blk_crypto_free_request(rq); | |
1072 | blk_pm_mark_last_busy(rq); | |
f794f335 | 1073 | |
02f7eab0 JA |
1074 | if (nr_tags == TAG_COMP_BATCH || cur_hctx != rq->mq_hctx) { |
1075 | if (cur_hctx) | |
1076 | blk_mq_flush_tag_batch(cur_hctx, tags, nr_tags); | |
f794f335 | 1077 | nr_tags = 0; |
02f7eab0 | 1078 | cur_hctx = rq->mq_hctx; |
f794f335 JA |
1079 | } |
1080 | tags[nr_tags++] = rq->tag; | |
f794f335 JA |
1081 | } |
1082 | ||
1083 | if (nr_tags) | |
02f7eab0 | 1084 | blk_mq_flush_tag_batch(cur_hctx, tags, nr_tags); |
f794f335 JA |
1085 | } |
1086 | EXPORT_SYMBOL_GPL(blk_mq_end_request_batch); | |
1087 | ||
f9ab4918 | 1088 | static void blk_complete_reqs(struct llist_head *list) |
320ae51f | 1089 | { |
f9ab4918 SAS |
1090 | struct llist_node *entry = llist_reverse_order(llist_del_all(list)); |
1091 | struct request *rq, *next; | |
c3077b5d | 1092 | |
f9ab4918 | 1093 | llist_for_each_entry_safe(rq, next, entry, ipi_list) |
c3077b5d | 1094 | rq->q->mq_ops->complete(rq); |
320ae51f | 1095 | } |
320ae51f | 1096 | |
f9ab4918 | 1097 | static __latent_entropy void blk_done_softirq(struct softirq_action *h) |
320ae51f | 1098 | { |
f9ab4918 | 1099 | blk_complete_reqs(this_cpu_ptr(&blk_cpu_done)); |
115243f5 CH |
1100 | } |
1101 | ||
c3077b5d CH |
1102 | static int blk_softirq_cpu_dead(unsigned int cpu) |
1103 | { | |
f9ab4918 | 1104 | blk_complete_reqs(&per_cpu(blk_cpu_done, cpu)); |
c3077b5d CH |
1105 | return 0; |
1106 | } | |
1107 | ||
40d09b53 | 1108 | static void __blk_mq_complete_request_remote(void *data) |
c3077b5d | 1109 | { |
f9ab4918 | 1110 | __raise_softirq_irqoff(BLOCK_SOFTIRQ); |
c3077b5d CH |
1111 | } |
1112 | ||
96339526 CH |
1113 | static inline bool blk_mq_complete_need_ipi(struct request *rq) |
1114 | { | |
1115 | int cpu = raw_smp_processor_id(); | |
1116 | ||
1117 | if (!IS_ENABLED(CONFIG_SMP) || | |
1118 | !test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) | |
1119 | return false; | |
71425189 SAS |
1120 | /* |
1121 | * With force threaded interrupts enabled, raising softirq from an SMP | |
1122 | * function call will always result in waking the ksoftirqd thread. | |
1123 | * This is probably worse than completing the request on a different | |
1124 | * cache domain. | |
1125 | */ | |
91cc470e | 1126 | if (force_irqthreads()) |
71425189 | 1127 | return false; |
96339526 CH |
1128 | |
1129 | /* same CPU or cache domain? Complete locally */ | |
1130 | if (cpu == rq->mq_ctx->cpu || | |
1131 | (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags) && | |
1132 | cpus_share_cache(cpu, rq->mq_ctx->cpu))) | |
1133 | return false; | |
1134 | ||
1135 | /* don't try to IPI to an offline CPU */ | |
1136 | return cpu_online(rq->mq_ctx->cpu); | |
1137 | } | |
1138 | ||
f9ab4918 SAS |
1139 | static void blk_mq_complete_send_ipi(struct request *rq) |
1140 | { | |
1141 | struct llist_head *list; | |
1142 | unsigned int cpu; | |
1143 | ||
1144 | cpu = rq->mq_ctx->cpu; | |
1145 | list = &per_cpu(blk_cpu_done, cpu); | |
1146 | if (llist_add(&rq->ipi_list, list)) { | |
1147 | INIT_CSD(&rq->csd, __blk_mq_complete_request_remote, rq); | |
1148 | smp_call_function_single_async(cpu, &rq->csd); | |
1149 | } | |
1150 | } | |
1151 | ||
1152 | static void blk_mq_raise_softirq(struct request *rq) | |
1153 | { | |
1154 | struct llist_head *list; | |
1155 | ||
1156 | preempt_disable(); | |
1157 | list = this_cpu_ptr(&blk_cpu_done); | |
1158 | if (llist_add(&rq->ipi_list, list)) | |
1159 | raise_softirq(BLOCK_SOFTIRQ); | |
1160 | preempt_enable(); | |
1161 | } | |
1162 | ||
40d09b53 | 1163 | bool blk_mq_complete_request_remote(struct request *rq) |
320ae51f | 1164 | { |
af78ff7c | 1165 | WRITE_ONCE(rq->state, MQ_RQ_COMPLETE); |
36e76539 | 1166 | |
4ab32bf3 | 1167 | /* |
f168420c LS |
1168 | * For request which hctx has only one ctx mapping, |
1169 | * or a polled request, always complete locally, | |
1170 | * it's pointless to redirect the completion. | |
4ab32bf3 | 1171 | */ |
f168420c LS |
1172 | if (rq->mq_hctx->nr_ctx == 1 || |
1173 | rq->cmd_flags & REQ_POLLED) | |
40d09b53 | 1174 | return false; |
38535201 | 1175 | |
96339526 | 1176 | if (blk_mq_complete_need_ipi(rq)) { |
f9ab4918 SAS |
1177 | blk_mq_complete_send_ipi(rq); |
1178 | return true; | |
3d6efbf6 | 1179 | } |
40d09b53 | 1180 | |
f9ab4918 SAS |
1181 | if (rq->q->nr_hw_queues == 1) { |
1182 | blk_mq_raise_softirq(rq); | |
1183 | return true; | |
1184 | } | |
1185 | return false; | |
40d09b53 CH |
1186 | } |
1187 | EXPORT_SYMBOL_GPL(blk_mq_complete_request_remote); | |
1188 | ||
1189 | /** | |
1190 | * blk_mq_complete_request - end I/O on a request | |
1191 | * @rq: the request being processed | |
1192 | * | |
1193 | * Description: | |
1194 | * Complete a request by scheduling the ->complete_rq operation. | |
1195 | **/ | |
1196 | void blk_mq_complete_request(struct request *rq) | |
1197 | { | |
1198 | if (!blk_mq_complete_request_remote(rq)) | |
1199 | rq->q->mq_ops->complete(rq); | |
320ae51f | 1200 | } |
15f73f5b | 1201 | EXPORT_SYMBOL(blk_mq_complete_request); |
30a91cb4 | 1202 | |
105663f7 AA |
1203 | /** |
1204 | * blk_mq_start_request - Start processing a request | |
1205 | * @rq: Pointer to request to be started | |
1206 | * | |
1207 | * Function used by device drivers to notify the block layer that a request | |
1208 | * is going to be processed now, so blk layer can do proper initializations | |
1209 | * such as starting the timeout timer. | |
1210 | */ | |
e2490073 | 1211 | void blk_mq_start_request(struct request *rq) |
320ae51f JA |
1212 | { |
1213 | struct request_queue *q = rq->q; | |
1214 | ||
a54895fa | 1215 | trace_block_rq_issue(rq); |
320ae51f | 1216 | |
cf43e6be | 1217 | if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) { |
4cddeaca | 1218 | rq->io_start_time_ns = ktime_get_ns(); |
3d244306 | 1219 | rq->stats_sectors = blk_rq_sectors(rq); |
cf43e6be | 1220 | rq->rq_flags |= RQF_STATS; |
a7905043 | 1221 | rq_qos_issue(q, rq); |
cf43e6be JA |
1222 | } |
1223 | ||
1d9bd516 | 1224 | WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE); |
538b7534 | 1225 | |
1d9bd516 | 1226 | blk_add_timer(rq); |
12f5b931 | 1227 | WRITE_ONCE(rq->state, MQ_RQ_IN_FLIGHT); |
49f5baa5 | 1228 | |
54d4e6ab MG |
1229 | #ifdef CONFIG_BLK_DEV_INTEGRITY |
1230 | if (blk_integrity_rq(rq) && req_op(rq) == REQ_OP_WRITE) | |
1231 | q->integrity.profile->prepare_fn(rq); | |
1232 | #endif | |
3e08773c CH |
1233 | if (rq->bio && rq->bio->bi_opf & REQ_POLLED) |
1234 | WRITE_ONCE(rq->bio->bi_cookie, blk_rq_to_qc(rq)); | |
320ae51f | 1235 | } |
e2490073 | 1236 | EXPORT_SYMBOL(blk_mq_start_request); |
320ae51f | 1237 | |
a327c341 ML |
1238 | /* |
1239 | * Allow 2x BLK_MAX_REQUEST_COUNT requests on plug queue for multiple | |
1240 | * queues. This is important for md arrays to benefit from merging | |
1241 | * requests. | |
1242 | */ | |
1243 | static inline unsigned short blk_plug_max_rq_count(struct blk_plug *plug) | |
1244 | { | |
1245 | if (plug->multiple_queues) | |
1246 | return BLK_MAX_REQUEST_COUNT * 2; | |
1247 | return BLK_MAX_REQUEST_COUNT; | |
1248 | } | |
1249 | ||
1250 | static void blk_add_rq_to_plug(struct blk_plug *plug, struct request *rq) | |
1251 | { | |
1252 | struct request *last = rq_list_peek(&plug->mq_list); | |
1253 | ||
1254 | if (!plug->rq_count) { | |
1255 | trace_block_plug(rq->q); | |
1256 | } else if (plug->rq_count >= blk_plug_max_rq_count(plug) || | |
1257 | (!blk_queue_nomerges(rq->q) && | |
1258 | blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) { | |
1259 | blk_mq_flush_plug_list(plug, false); | |
1260 | trace_block_plug(rq->q); | |
1261 | } | |
1262 | ||
1263 | if (!plug->multiple_queues && last && last->q != rq->q) | |
1264 | plug->multiple_queues = true; | |
1265 | if (!plug->has_elevator && (rq->rq_flags & RQF_ELV)) | |
1266 | plug->has_elevator = true; | |
1267 | rq->rq_next = NULL; | |
1268 | rq_list_add(&plug->mq_list, rq); | |
1269 | plug->rq_count++; | |
1270 | } | |
1271 | ||
4054cff9 CH |
1272 | /** |
1273 | * blk_execute_rq_nowait - insert a request to I/O scheduler for execution | |
4054cff9 CH |
1274 | * @rq: request to insert |
1275 | * @at_head: insert request at head or tail of queue | |
4054cff9 CH |
1276 | * |
1277 | * Description: | |
1278 | * Insert a fully prepared request at the back of the I/O scheduler queue | |
1279 | * for execution. Don't wait for completion. | |
1280 | * | |
1281 | * Note: | |
1282 | * This function will invoke @done directly if the queue is dead. | |
1283 | */ | |
e2e53086 | 1284 | void blk_execute_rq_nowait(struct request *rq, bool at_head) |
4054cff9 | 1285 | { |
ae948fd6 CH |
1286 | WARN_ON(irqs_disabled()); |
1287 | WARN_ON(!blk_rq_is_passthrough(rq)); | |
4054cff9 | 1288 | |
ae948fd6 | 1289 | blk_account_io_start(rq); |
110fdb44 PR |
1290 | |
1291 | /* | |
1292 | * As plugging can be enabled for passthrough requests on a zoned | |
1293 | * device, directly accessing the plug instead of using blk_mq_plug() | |
1294 | * should not have any consequences. | |
1295 | */ | |
ae948fd6 CH |
1296 | if (current->plug) |
1297 | blk_add_rq_to_plug(current->plug, rq); | |
1298 | else | |
1299 | blk_mq_sched_insert_request(rq, at_head, true, false); | |
4054cff9 CH |
1300 | } |
1301 | EXPORT_SYMBOL_GPL(blk_execute_rq_nowait); | |
1302 | ||
32ac5a9b CH |
1303 | struct blk_rq_wait { |
1304 | struct completion done; | |
1305 | blk_status_t ret; | |
1306 | }; | |
1307 | ||
de671d61 | 1308 | static enum rq_end_io_ret blk_end_sync_rq(struct request *rq, blk_status_t ret) |
32ac5a9b CH |
1309 | { |
1310 | struct blk_rq_wait *wait = rq->end_io_data; | |
1311 | ||
1312 | wait->ret = ret; | |
1313 | complete(&wait->done); | |
de671d61 | 1314 | return RQ_END_IO_NONE; |
32ac5a9b CH |
1315 | } |
1316 | ||
c6e99ea4 | 1317 | bool blk_rq_is_poll(struct request *rq) |
4054cff9 CH |
1318 | { |
1319 | if (!rq->mq_hctx) | |
1320 | return false; | |
1321 | if (rq->mq_hctx->type != HCTX_TYPE_POLL) | |
1322 | return false; | |
1323 | if (WARN_ON_ONCE(!rq->bio)) | |
1324 | return false; | |
1325 | return true; | |
1326 | } | |
c6e99ea4 | 1327 | EXPORT_SYMBOL_GPL(blk_rq_is_poll); |
4054cff9 CH |
1328 | |
1329 | static void blk_rq_poll_completion(struct request *rq, struct completion *wait) | |
1330 | { | |
1331 | do { | |
1332 | bio_poll(rq->bio, NULL, 0); | |
1333 | cond_resched(); | |
1334 | } while (!completion_done(wait)); | |
1335 | } | |
1336 | ||
1337 | /** | |
1338 | * blk_execute_rq - insert a request into queue for execution | |
4054cff9 CH |
1339 | * @rq: request to insert |
1340 | * @at_head: insert request at head or tail of queue | |
1341 | * | |
1342 | * Description: | |
1343 | * Insert a fully prepared request at the back of the I/O scheduler queue | |
1344 | * for execution and wait for completion. | |
1345 | * Return: The blk_status_t result provided to blk_mq_end_request(). | |
1346 | */ | |
b84ba30b | 1347 | blk_status_t blk_execute_rq(struct request *rq, bool at_head) |
4054cff9 | 1348 | { |
32ac5a9b CH |
1349 | struct blk_rq_wait wait = { |
1350 | .done = COMPLETION_INITIALIZER_ONSTACK(wait.done), | |
1351 | }; | |
4054cff9 | 1352 | |
ae948fd6 CH |
1353 | WARN_ON(irqs_disabled()); |
1354 | WARN_ON(!blk_rq_is_passthrough(rq)); | |
4054cff9 CH |
1355 | |
1356 | rq->end_io_data = &wait; | |
ae948fd6 | 1357 | rq->end_io = blk_end_sync_rq; |
4054cff9 | 1358 | |
ae948fd6 CH |
1359 | blk_account_io_start(rq); |
1360 | blk_mq_sched_insert_request(rq, at_head, true, false); | |
4054cff9 | 1361 | |
ae948fd6 | 1362 | if (blk_rq_is_poll(rq)) { |
32ac5a9b | 1363 | blk_rq_poll_completion(rq, &wait.done); |
ae948fd6 CH |
1364 | } else { |
1365 | /* | |
1366 | * Prevent hang_check timer from firing at us during very long | |
1367 | * I/O | |
1368 | */ | |
1369 | unsigned long hang_check = sysctl_hung_task_timeout_secs; | |
1370 | ||
1371 | if (hang_check) | |
32ac5a9b | 1372 | while (!wait_for_completion_io_timeout(&wait.done, |
ae948fd6 CH |
1373 | hang_check * (HZ/2))) |
1374 | ; | |
1375 | else | |
32ac5a9b | 1376 | wait_for_completion_io(&wait.done); |
ae948fd6 | 1377 | } |
4054cff9 | 1378 | |
32ac5a9b | 1379 | return wait.ret; |
4054cff9 CH |
1380 | } |
1381 | EXPORT_SYMBOL(blk_execute_rq); | |
1382 | ||
ed0791b2 | 1383 | static void __blk_mq_requeue_request(struct request *rq) |
320ae51f JA |
1384 | { |
1385 | struct request_queue *q = rq->q; | |
1386 | ||
923218f6 ML |
1387 | blk_mq_put_driver_tag(rq); |
1388 | ||
a54895fa | 1389 | trace_block_rq_requeue(rq); |
a7905043 | 1390 | rq_qos_requeue(q, rq); |
49f5baa5 | 1391 | |
12f5b931 KB |
1392 | if (blk_mq_request_started(rq)) { |
1393 | WRITE_ONCE(rq->state, MQ_RQ_IDLE); | |
da661267 | 1394 | rq->rq_flags &= ~RQF_TIMED_OUT; |
e2490073 | 1395 | } |
320ae51f JA |
1396 | } |
1397 | ||
2b053aca | 1398 | void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list) |
ed0791b2 | 1399 | { |
ed0791b2 | 1400 | __blk_mq_requeue_request(rq); |
ed0791b2 | 1401 | |
105976f5 ML |
1402 | /* this request will be re-inserted to io scheduler queue */ |
1403 | blk_mq_sched_requeue_request(rq); | |
1404 | ||
2b053aca | 1405 | blk_mq_add_to_requeue_list(rq, true, kick_requeue_list); |
ed0791b2 CH |
1406 | } |
1407 | EXPORT_SYMBOL(blk_mq_requeue_request); | |
1408 | ||
6fca6a61 CH |
1409 | static void blk_mq_requeue_work(struct work_struct *work) |
1410 | { | |
1411 | struct request_queue *q = | |
2849450a | 1412 | container_of(work, struct request_queue, requeue_work.work); |
6fca6a61 CH |
1413 | LIST_HEAD(rq_list); |
1414 | struct request *rq, *next; | |
6fca6a61 | 1415 | |
18e9781d | 1416 | spin_lock_irq(&q->requeue_lock); |
6fca6a61 | 1417 | list_splice_init(&q->requeue_list, &rq_list); |
18e9781d | 1418 | spin_unlock_irq(&q->requeue_lock); |
6fca6a61 CH |
1419 | |
1420 | list_for_each_entry_safe(rq, next, &rq_list, queuelist) { | |
aef1897c | 1421 | if (!(rq->rq_flags & (RQF_SOFTBARRIER | RQF_DONTPREP))) |
6fca6a61 CH |
1422 | continue; |
1423 | ||
e8064021 | 1424 | rq->rq_flags &= ~RQF_SOFTBARRIER; |
6fca6a61 | 1425 | list_del_init(&rq->queuelist); |
aef1897c JW |
1426 | /* |
1427 | * If RQF_DONTPREP, rq has contained some driver specific | |
1428 | * data, so insert it to hctx dispatch list to avoid any | |
1429 | * merge. | |
1430 | */ | |
1431 | if (rq->rq_flags & RQF_DONTPREP) | |
01e99aec | 1432 | blk_mq_request_bypass_insert(rq, false, false); |
aef1897c JW |
1433 | else |
1434 | blk_mq_sched_insert_request(rq, true, false, false); | |
6fca6a61 CH |
1435 | } |
1436 | ||
1437 | while (!list_empty(&rq_list)) { | |
1438 | rq = list_entry(rq_list.next, struct request, queuelist); | |
1439 | list_del_init(&rq->queuelist); | |
9e97d295 | 1440 | blk_mq_sched_insert_request(rq, false, false, false); |
6fca6a61 CH |
1441 | } |
1442 | ||
52d7f1b5 | 1443 | blk_mq_run_hw_queues(q, false); |
6fca6a61 CH |
1444 | } |
1445 | ||
2b053aca BVA |
1446 | void blk_mq_add_to_requeue_list(struct request *rq, bool at_head, |
1447 | bool kick_requeue_list) | |
6fca6a61 CH |
1448 | { |
1449 | struct request_queue *q = rq->q; | |
1450 | unsigned long flags; | |
1451 | ||
1452 | /* | |
1453 | * We abuse this flag that is otherwise used by the I/O scheduler to | |
ff821d27 | 1454 | * request head insertion from the workqueue. |
6fca6a61 | 1455 | */ |
e8064021 | 1456 | BUG_ON(rq->rq_flags & RQF_SOFTBARRIER); |
6fca6a61 CH |
1457 | |
1458 | spin_lock_irqsave(&q->requeue_lock, flags); | |
1459 | if (at_head) { | |
e8064021 | 1460 | rq->rq_flags |= RQF_SOFTBARRIER; |
6fca6a61 CH |
1461 | list_add(&rq->queuelist, &q->requeue_list); |
1462 | } else { | |
1463 | list_add_tail(&rq->queuelist, &q->requeue_list); | |
1464 | } | |
1465 | spin_unlock_irqrestore(&q->requeue_lock, flags); | |
2b053aca BVA |
1466 | |
1467 | if (kick_requeue_list) | |
1468 | blk_mq_kick_requeue_list(q); | |
6fca6a61 | 1469 | } |
6fca6a61 CH |
1470 | |
1471 | void blk_mq_kick_requeue_list(struct request_queue *q) | |
1472 | { | |
ae943d20 | 1473 | kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0); |
6fca6a61 CH |
1474 | } |
1475 | EXPORT_SYMBOL(blk_mq_kick_requeue_list); | |
1476 | ||
2849450a MS |
1477 | void blk_mq_delay_kick_requeue_list(struct request_queue *q, |
1478 | unsigned long msecs) | |
1479 | { | |
d4acf365 BVA |
1480 | kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, |
1481 | msecs_to_jiffies(msecs)); | |
2849450a MS |
1482 | } |
1483 | EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list); | |
1484 | ||
2dd6532e | 1485 | static bool blk_mq_rq_inflight(struct request *rq, void *priv) |
ae879912 JA |
1486 | { |
1487 | /* | |
8ab30a33 JG |
1488 | * If we find a request that isn't idle we know the queue is busy |
1489 | * as it's checked in the iter. | |
1490 | * Return false to stop the iteration. | |
ae879912 | 1491 | */ |
8ab30a33 | 1492 | if (blk_mq_request_started(rq)) { |
ae879912 JA |
1493 | bool *busy = priv; |
1494 | ||
1495 | *busy = true; | |
1496 | return false; | |
1497 | } | |
1498 | ||
1499 | return true; | |
1500 | } | |
1501 | ||
3c94d83c | 1502 | bool blk_mq_queue_inflight(struct request_queue *q) |
ae879912 JA |
1503 | { |
1504 | bool busy = false; | |
1505 | ||
3c94d83c | 1506 | blk_mq_queue_tag_busy_iter(q, blk_mq_rq_inflight, &busy); |
ae879912 JA |
1507 | return busy; |
1508 | } | |
3c94d83c | 1509 | EXPORT_SYMBOL_GPL(blk_mq_queue_inflight); |
ae879912 | 1510 | |
9bdb4833 | 1511 | static void blk_mq_rq_timed_out(struct request *req) |
320ae51f | 1512 | { |
da661267 | 1513 | req->rq_flags |= RQF_TIMED_OUT; |
d1210d5a CH |
1514 | if (req->q->mq_ops->timeout) { |
1515 | enum blk_eh_timer_return ret; | |
1516 | ||
9bdb4833 | 1517 | ret = req->q->mq_ops->timeout(req); |
d1210d5a CH |
1518 | if (ret == BLK_EH_DONE) |
1519 | return; | |
1520 | WARN_ON_ONCE(ret != BLK_EH_RESET_TIMER); | |
46f92d42 | 1521 | } |
d1210d5a CH |
1522 | |
1523 | blk_add_timer(req); | |
87ee7b11 | 1524 | } |
5b3f25fc | 1525 | |
12f5b931 | 1526 | static bool blk_mq_req_expired(struct request *rq, unsigned long *next) |
81481eb4 | 1527 | { |
12f5b931 | 1528 | unsigned long deadline; |
87ee7b11 | 1529 | |
12f5b931 KB |
1530 | if (blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT) |
1531 | return false; | |
da661267 CH |
1532 | if (rq->rq_flags & RQF_TIMED_OUT) |
1533 | return false; | |
a7af0af3 | 1534 | |
079076b3 | 1535 | deadline = READ_ONCE(rq->deadline); |
12f5b931 KB |
1536 | if (time_after_eq(jiffies, deadline)) |
1537 | return true; | |
a7af0af3 | 1538 | |
12f5b931 KB |
1539 | if (*next == 0) |
1540 | *next = deadline; | |
1541 | else if (time_after(*next, deadline)) | |
1542 | *next = deadline; | |
1543 | return false; | |
87ee7b11 JA |
1544 | } |
1545 | ||
2e315dc0 ML |
1546 | void blk_mq_put_rq_ref(struct request *rq) |
1547 | { | |
de671d61 JA |
1548 | if (is_flush_rq(rq)) { |
1549 | if (rq->end_io(rq, 0) == RQ_END_IO_FREE) | |
1550 | blk_mq_free_request(rq); | |
1551 | } else if (req_ref_put_and_test(rq)) { | |
2e315dc0 | 1552 | __blk_mq_free_request(rq); |
de671d61 | 1553 | } |
2e315dc0 ML |
1554 | } |
1555 | ||
2dd6532e | 1556 | static bool blk_mq_check_expired(struct request *rq, void *priv) |
1d9bd516 | 1557 | { |
12f5b931 KB |
1558 | unsigned long *next = priv; |
1559 | ||
1560 | /* | |
c797b40c ML |
1561 | * blk_mq_queue_tag_busy_iter() has locked the request, so it cannot |
1562 | * be reallocated underneath the timeout handler's processing, then | |
1563 | * the expire check is reliable. If the request is not expired, then | |
1564 | * it was completed and reallocated as a new request after returning | |
1565 | * from blk_mq_check_expired(). | |
1d9bd516 | 1566 | */ |
12f5b931 | 1567 | if (blk_mq_req_expired(rq, next)) |
9bdb4833 | 1568 | blk_mq_rq_timed_out(rq); |
7baa8572 | 1569 | return true; |
1d9bd516 TH |
1570 | } |
1571 | ||
287922eb | 1572 | static void blk_mq_timeout_work(struct work_struct *work) |
320ae51f | 1573 | { |
287922eb CH |
1574 | struct request_queue *q = |
1575 | container_of(work, struct request_queue, timeout_work); | |
12f5b931 | 1576 | unsigned long next = 0; |
1d9bd516 | 1577 | struct blk_mq_hw_ctx *hctx; |
4f481208 | 1578 | unsigned long i; |
320ae51f | 1579 | |
71f79fb3 GKB |
1580 | /* A deadlock might occur if a request is stuck requiring a |
1581 | * timeout at the same time a queue freeze is waiting | |
1582 | * completion, since the timeout code would not be able to | |
1583 | * acquire the queue reference here. | |
1584 | * | |
1585 | * That's why we don't use blk_queue_enter here; instead, we use | |
1586 | * percpu_ref_tryget directly, because we need to be able to | |
1587 | * obtain a reference even in the short window between the queue | |
1588 | * starting to freeze, by dropping the first reference in | |
1671d522 | 1589 | * blk_freeze_queue_start, and the moment the last request is |
71f79fb3 GKB |
1590 | * consumed, marked by the instant q_usage_counter reaches |
1591 | * zero. | |
1592 | */ | |
1593 | if (!percpu_ref_tryget(&q->q_usage_counter)) | |
287922eb CH |
1594 | return; |
1595 | ||
12f5b931 | 1596 | blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &next); |
320ae51f | 1597 | |
12f5b931 KB |
1598 | if (next != 0) { |
1599 | mod_timer(&q->timeout, next); | |
0d2602ca | 1600 | } else { |
fcd36c36 BVA |
1601 | /* |
1602 | * Request timeouts are handled as a forward rolling timer. If | |
1603 | * we end up here it means that no requests are pending and | |
1604 | * also that no request has been pending for a while. Mark | |
1605 | * each hctx as idle. | |
1606 | */ | |
f054b56c ML |
1607 | queue_for_each_hw_ctx(q, hctx, i) { |
1608 | /* the hctx may be unmapped, so check it here */ | |
1609 | if (blk_mq_hw_queue_mapped(hctx)) | |
1610 | blk_mq_tag_idle(hctx); | |
1611 | } | |
0d2602ca | 1612 | } |
287922eb | 1613 | blk_queue_exit(q); |
320ae51f JA |
1614 | } |
1615 | ||
88459642 OS |
1616 | struct flush_busy_ctx_data { |
1617 | struct blk_mq_hw_ctx *hctx; | |
1618 | struct list_head *list; | |
1619 | }; | |
1620 | ||
1621 | static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data) | |
1622 | { | |
1623 | struct flush_busy_ctx_data *flush_data = data; | |
1624 | struct blk_mq_hw_ctx *hctx = flush_data->hctx; | |
1625 | struct blk_mq_ctx *ctx = hctx->ctxs[bitnr]; | |
c16d6b5a | 1626 | enum hctx_type type = hctx->type; |
88459642 | 1627 | |
88459642 | 1628 | spin_lock(&ctx->lock); |
c16d6b5a | 1629 | list_splice_tail_init(&ctx->rq_lists[type], flush_data->list); |
e9a99a63 | 1630 | sbitmap_clear_bit(sb, bitnr); |
88459642 OS |
1631 | spin_unlock(&ctx->lock); |
1632 | return true; | |
1633 | } | |
1634 | ||
1429d7c9 JA |
1635 | /* |
1636 | * Process software queues that have been marked busy, splicing them | |
1637 | * to the for-dispatch | |
1638 | */ | |
2c3ad667 | 1639 | void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list) |
1429d7c9 | 1640 | { |
88459642 OS |
1641 | struct flush_busy_ctx_data data = { |
1642 | .hctx = hctx, | |
1643 | .list = list, | |
1644 | }; | |
1429d7c9 | 1645 | |
88459642 | 1646 | sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data); |
1429d7c9 | 1647 | } |
2c3ad667 | 1648 | EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs); |
1429d7c9 | 1649 | |
b347689f ML |
1650 | struct dispatch_rq_data { |
1651 | struct blk_mq_hw_ctx *hctx; | |
1652 | struct request *rq; | |
1653 | }; | |
1654 | ||
1655 | static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr, | |
1656 | void *data) | |
1657 | { | |
1658 | struct dispatch_rq_data *dispatch_data = data; | |
1659 | struct blk_mq_hw_ctx *hctx = dispatch_data->hctx; | |
1660 | struct blk_mq_ctx *ctx = hctx->ctxs[bitnr]; | |
c16d6b5a | 1661 | enum hctx_type type = hctx->type; |
b347689f ML |
1662 | |
1663 | spin_lock(&ctx->lock); | |
c16d6b5a ML |
1664 | if (!list_empty(&ctx->rq_lists[type])) { |
1665 | dispatch_data->rq = list_entry_rq(ctx->rq_lists[type].next); | |
b347689f | 1666 | list_del_init(&dispatch_data->rq->queuelist); |
c16d6b5a | 1667 | if (list_empty(&ctx->rq_lists[type])) |
b347689f ML |
1668 | sbitmap_clear_bit(sb, bitnr); |
1669 | } | |
1670 | spin_unlock(&ctx->lock); | |
1671 | ||
1672 | return !dispatch_data->rq; | |
1673 | } | |
1674 | ||
1675 | struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx, | |
1676 | struct blk_mq_ctx *start) | |
1677 | { | |
f31967f0 | 1678 | unsigned off = start ? start->index_hw[hctx->type] : 0; |
b347689f ML |
1679 | struct dispatch_rq_data data = { |
1680 | .hctx = hctx, | |
1681 | .rq = NULL, | |
1682 | }; | |
1683 | ||
1684 | __sbitmap_for_each_set(&hctx->ctx_map, off, | |
1685 | dispatch_rq_from_ctx, &data); | |
1686 | ||
1687 | return data.rq; | |
1688 | } | |
1689 | ||
a808a9d5 | 1690 | static bool __blk_mq_alloc_driver_tag(struct request *rq) |
570e9b73 | 1691 | { |
ae0f1a73 | 1692 | struct sbitmap_queue *bt = &rq->mq_hctx->tags->bitmap_tags; |
570e9b73 | 1693 | unsigned int tag_offset = rq->mq_hctx->tags->nr_reserved_tags; |
570e9b73 ML |
1694 | int tag; |
1695 | ||
568f2700 ML |
1696 | blk_mq_tag_busy(rq->mq_hctx); |
1697 | ||
570e9b73 | 1698 | if (blk_mq_tag_is_reserved(rq->mq_hctx->sched_tags, rq->internal_tag)) { |
ae0f1a73 | 1699 | bt = &rq->mq_hctx->tags->breserved_tags; |
570e9b73 | 1700 | tag_offset = 0; |
28500850 ML |
1701 | } else { |
1702 | if (!hctx_may_queue(rq->mq_hctx, bt)) | |
1703 | return false; | |
570e9b73 ML |
1704 | } |
1705 | ||
570e9b73 ML |
1706 | tag = __sbitmap_queue_get(bt); |
1707 | if (tag == BLK_MQ_NO_TAG) | |
1708 | return false; | |
1709 | ||
1710 | rq->tag = tag + tag_offset; | |
570e9b73 ML |
1711 | return true; |
1712 | } | |
1713 | ||
a808a9d5 | 1714 | bool __blk_mq_get_driver_tag(struct blk_mq_hw_ctx *hctx, struct request *rq) |
570e9b73 | 1715 | { |
a808a9d5 | 1716 | if (rq->tag == BLK_MQ_NO_TAG && !__blk_mq_alloc_driver_tag(rq)) |
568f2700 ML |
1717 | return false; |
1718 | ||
51db1c37 | 1719 | if ((hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) && |
568f2700 ML |
1720 | !(rq->rq_flags & RQF_MQ_INFLIGHT)) { |
1721 | rq->rq_flags |= RQF_MQ_INFLIGHT; | |
bccf5e26 | 1722 | __blk_mq_inc_active_requests(hctx); |
568f2700 ML |
1723 | } |
1724 | hctx->tags->rqs[rq->tag] = rq; | |
1725 | return true; | |
570e9b73 ML |
1726 | } |
1727 | ||
eb619fdb JA |
1728 | static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode, |
1729 | int flags, void *key) | |
da55f2cc OS |
1730 | { |
1731 | struct blk_mq_hw_ctx *hctx; | |
1732 | ||
1733 | hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait); | |
1734 | ||
5815839b | 1735 | spin_lock(&hctx->dispatch_wait_lock); |
e8618575 JA |
1736 | if (!list_empty(&wait->entry)) { |
1737 | struct sbitmap_queue *sbq; | |
1738 | ||
1739 | list_del_init(&wait->entry); | |
ae0f1a73 | 1740 | sbq = &hctx->tags->bitmap_tags; |
e8618575 JA |
1741 | atomic_dec(&sbq->ws_active); |
1742 | } | |
5815839b ML |
1743 | spin_unlock(&hctx->dispatch_wait_lock); |
1744 | ||
da55f2cc OS |
1745 | blk_mq_run_hw_queue(hctx, true); |
1746 | return 1; | |
1747 | } | |
1748 | ||
f906a6a0 JA |
1749 | /* |
1750 | * Mark us waiting for a tag. For shared tags, this involves hooking us into | |
ee3e4de5 BVA |
1751 | * the tag wakeups. For non-shared tags, we can simply mark us needing a |
1752 | * restart. For both cases, take care to check the condition again after | |
f906a6a0 JA |
1753 | * marking us as waiting. |
1754 | */ | |
2278d69f | 1755 | static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx *hctx, |
f906a6a0 | 1756 | struct request *rq) |
da55f2cc | 1757 | { |
ae0f1a73 | 1758 | struct sbitmap_queue *sbq = &hctx->tags->bitmap_tags; |
5815839b | 1759 | struct wait_queue_head *wq; |
f906a6a0 JA |
1760 | wait_queue_entry_t *wait; |
1761 | bool ret; | |
da55f2cc | 1762 | |
51db1c37 | 1763 | if (!(hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) { |
684b7324 | 1764 | blk_mq_sched_mark_restart_hctx(hctx); |
f906a6a0 | 1765 | |
c27d53fb BVA |
1766 | /* |
1767 | * It's possible that a tag was freed in the window between the | |
1768 | * allocation failure and adding the hardware queue to the wait | |
1769 | * queue. | |
1770 | * | |
1771 | * Don't clear RESTART here, someone else could have set it. | |
1772 | * At most this will cost an extra queue run. | |
1773 | */ | |
8ab6bb9e | 1774 | return blk_mq_get_driver_tag(rq); |
eb619fdb | 1775 | } |
eb619fdb | 1776 | |
2278d69f | 1777 | wait = &hctx->dispatch_wait; |
c27d53fb BVA |
1778 | if (!list_empty_careful(&wait->entry)) |
1779 | return false; | |
1780 | ||
e8618575 | 1781 | wq = &bt_wait_ptr(sbq, hctx)->wait; |
5815839b ML |
1782 | |
1783 | spin_lock_irq(&wq->lock); | |
1784 | spin_lock(&hctx->dispatch_wait_lock); | |
c27d53fb | 1785 | if (!list_empty(&wait->entry)) { |
5815839b ML |
1786 | spin_unlock(&hctx->dispatch_wait_lock); |
1787 | spin_unlock_irq(&wq->lock); | |
c27d53fb | 1788 | return false; |
eb619fdb JA |
1789 | } |
1790 | ||
e8618575 | 1791 | atomic_inc(&sbq->ws_active); |
5815839b ML |
1792 | wait->flags &= ~WQ_FLAG_EXCLUSIVE; |
1793 | __add_wait_queue(wq, wait); | |
c27d53fb | 1794 | |
da55f2cc | 1795 | /* |
eb619fdb JA |
1796 | * It's possible that a tag was freed in the window between the |
1797 | * allocation failure and adding the hardware queue to the wait | |
1798 | * queue. | |
da55f2cc | 1799 | */ |
8ab6bb9e | 1800 | ret = blk_mq_get_driver_tag(rq); |
c27d53fb | 1801 | if (!ret) { |
5815839b ML |
1802 | spin_unlock(&hctx->dispatch_wait_lock); |
1803 | spin_unlock_irq(&wq->lock); | |
c27d53fb | 1804 | return false; |
eb619fdb | 1805 | } |
c27d53fb BVA |
1806 | |
1807 | /* | |
1808 | * We got a tag, remove ourselves from the wait queue to ensure | |
1809 | * someone else gets the wakeup. | |
1810 | */ | |
c27d53fb | 1811 | list_del_init(&wait->entry); |
e8618575 | 1812 | atomic_dec(&sbq->ws_active); |
5815839b ML |
1813 | spin_unlock(&hctx->dispatch_wait_lock); |
1814 | spin_unlock_irq(&wq->lock); | |
c27d53fb BVA |
1815 | |
1816 | return true; | |
da55f2cc OS |
1817 | } |
1818 | ||
6e768717 ML |
1819 | #define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT 8 |
1820 | #define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR 4 | |
1821 | /* | |
1822 | * Update dispatch busy with the Exponential Weighted Moving Average(EWMA): | |
1823 | * - EWMA is one simple way to compute running average value | |
1824 | * - weight(7/8 and 1/8) is applied so that it can decrease exponentially | |
1825 | * - take 4 as factor for avoiding to get too small(0) result, and this | |
1826 | * factor doesn't matter because EWMA decreases exponentially | |
1827 | */ | |
1828 | static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx *hctx, bool busy) | |
1829 | { | |
1830 | unsigned int ewma; | |
1831 | ||
6e768717 ML |
1832 | ewma = hctx->dispatch_busy; |
1833 | ||
1834 | if (!ewma && !busy) | |
1835 | return; | |
1836 | ||
1837 | ewma *= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT - 1; | |
1838 | if (busy) | |
1839 | ewma += 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR; | |
1840 | ewma /= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT; | |
1841 | ||
1842 | hctx->dispatch_busy = ewma; | |
1843 | } | |
1844 | ||
86ff7c2a ML |
1845 | #define BLK_MQ_RESOURCE_DELAY 3 /* ms units */ |
1846 | ||
c92a4103 JT |
1847 | static void blk_mq_handle_dev_resource(struct request *rq, |
1848 | struct list_head *list) | |
1849 | { | |
1850 | struct request *next = | |
1851 | list_first_entry_or_null(list, struct request, queuelist); | |
1852 | ||
1853 | /* | |
1854 | * If an I/O scheduler has been configured and we got a driver tag for | |
1855 | * the next request already, free it. | |
1856 | */ | |
1857 | if (next) | |
1858 | blk_mq_put_driver_tag(next); | |
1859 | ||
1860 | list_add(&rq->queuelist, list); | |
1861 | __blk_mq_requeue_request(rq); | |
1862 | } | |
1863 | ||
0512a75b KB |
1864 | static void blk_mq_handle_zone_resource(struct request *rq, |
1865 | struct list_head *zone_list) | |
1866 | { | |
1867 | /* | |
1868 | * If we end up here it is because we cannot dispatch a request to a | |
1869 | * specific zone due to LLD level zone-write locking or other zone | |
1870 | * related resource not being available. In this case, set the request | |
1871 | * aside in zone_list for retrying it later. | |
1872 | */ | |
1873 | list_add(&rq->queuelist, zone_list); | |
1874 | __blk_mq_requeue_request(rq); | |
1875 | } | |
1876 | ||
75383524 ML |
1877 | enum prep_dispatch { |
1878 | PREP_DISPATCH_OK, | |
1879 | PREP_DISPATCH_NO_TAG, | |
1880 | PREP_DISPATCH_NO_BUDGET, | |
1881 | }; | |
1882 | ||
1883 | static enum prep_dispatch blk_mq_prep_dispatch_rq(struct request *rq, | |
1884 | bool need_budget) | |
1885 | { | |
1886 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; | |
2a5a24aa | 1887 | int budget_token = -1; |
75383524 | 1888 | |
2a5a24aa ML |
1889 | if (need_budget) { |
1890 | budget_token = blk_mq_get_dispatch_budget(rq->q); | |
1891 | if (budget_token < 0) { | |
1892 | blk_mq_put_driver_tag(rq); | |
1893 | return PREP_DISPATCH_NO_BUDGET; | |
1894 | } | |
1895 | blk_mq_set_rq_budget_token(rq, budget_token); | |
75383524 ML |
1896 | } |
1897 | ||
1898 | if (!blk_mq_get_driver_tag(rq)) { | |
1899 | /* | |
1900 | * The initial allocation attempt failed, so we need to | |
1901 | * rerun the hardware queue when a tag is freed. The | |
1902 | * waitqueue takes care of that. If the queue is run | |
1903 | * before we add this entry back on the dispatch list, | |
1904 | * we'll re-run it below. | |
1905 | */ | |
1906 | if (!blk_mq_mark_tag_wait(hctx, rq)) { | |
1fd40b5e ML |
1907 | /* |
1908 | * All budgets not got from this function will be put | |
1909 | * together during handling partial dispatch | |
1910 | */ | |
1911 | if (need_budget) | |
2a5a24aa | 1912 | blk_mq_put_dispatch_budget(rq->q, budget_token); |
75383524 ML |
1913 | return PREP_DISPATCH_NO_TAG; |
1914 | } | |
1915 | } | |
1916 | ||
1917 | return PREP_DISPATCH_OK; | |
1918 | } | |
1919 | ||
1fd40b5e ML |
1920 | /* release all allocated budgets before calling to blk_mq_dispatch_rq_list */ |
1921 | static void blk_mq_release_budgets(struct request_queue *q, | |
2a5a24aa | 1922 | struct list_head *list) |
1fd40b5e | 1923 | { |
2a5a24aa | 1924 | struct request *rq; |
1fd40b5e | 1925 | |
2a5a24aa ML |
1926 | list_for_each_entry(rq, list, queuelist) { |
1927 | int budget_token = blk_mq_get_rq_budget_token(rq); | |
1fd40b5e | 1928 | |
2a5a24aa ML |
1929 | if (budget_token >= 0) |
1930 | blk_mq_put_dispatch_budget(q, budget_token); | |
1931 | } | |
1fd40b5e ML |
1932 | } |
1933 | ||
1f57f8d4 JA |
1934 | /* |
1935 | * Returns true if we did some work AND can potentially do more. | |
1936 | */ | |
445874e8 | 1937 | bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx *hctx, struct list_head *list, |
1fd40b5e | 1938 | unsigned int nr_budgets) |
320ae51f | 1939 | { |
75383524 | 1940 | enum prep_dispatch prep; |
445874e8 | 1941 | struct request_queue *q = hctx->queue; |
6d6f167c | 1942 | struct request *rq, *nxt; |
fc17b653 | 1943 | int errors, queued; |
86ff7c2a | 1944 | blk_status_t ret = BLK_STS_OK; |
0512a75b | 1945 | LIST_HEAD(zone_list); |
9586e67b | 1946 | bool needs_resource = false; |
320ae51f | 1947 | |
81380ca1 OS |
1948 | if (list_empty(list)) |
1949 | return false; | |
1950 | ||
320ae51f JA |
1951 | /* |
1952 | * Now process all the entries, sending them to the driver. | |
1953 | */ | |
93efe981 | 1954 | errors = queued = 0; |
81380ca1 | 1955 | do { |
74c45052 | 1956 | struct blk_mq_queue_data bd; |
320ae51f | 1957 | |
f04c3df3 | 1958 | rq = list_first_entry(list, struct request, queuelist); |
0bca799b | 1959 | |
445874e8 | 1960 | WARN_ON_ONCE(hctx != rq->mq_hctx); |
1fd40b5e | 1961 | prep = blk_mq_prep_dispatch_rq(rq, !nr_budgets); |
75383524 | 1962 | if (prep != PREP_DISPATCH_OK) |
0bca799b | 1963 | break; |
de148297 | 1964 | |
320ae51f | 1965 | list_del_init(&rq->queuelist); |
320ae51f | 1966 | |
74c45052 | 1967 | bd.rq = rq; |
113285b4 JA |
1968 | |
1969 | /* | |
1970 | * Flag last if we have no more requests, or if we have more | |
1971 | * but can't assign a driver tag to it. | |
1972 | */ | |
1973 | if (list_empty(list)) | |
1974 | bd.last = true; | |
1975 | else { | |
113285b4 | 1976 | nxt = list_first_entry(list, struct request, queuelist); |
8ab6bb9e | 1977 | bd.last = !blk_mq_get_driver_tag(nxt); |
113285b4 | 1978 | } |
74c45052 | 1979 | |
1fd40b5e ML |
1980 | /* |
1981 | * once the request is queued to lld, no need to cover the | |
1982 | * budget any more | |
1983 | */ | |
1984 | if (nr_budgets) | |
1985 | nr_budgets--; | |
74c45052 | 1986 | ret = q->mq_ops->queue_rq(hctx, &bd); |
7bf13729 ML |
1987 | switch (ret) { |
1988 | case BLK_STS_OK: | |
1989 | queued++; | |
320ae51f | 1990 | break; |
7bf13729 | 1991 | case BLK_STS_RESOURCE: |
9586e67b NA |
1992 | needs_resource = true; |
1993 | fallthrough; | |
7bf13729 ML |
1994 | case BLK_STS_DEV_RESOURCE: |
1995 | blk_mq_handle_dev_resource(rq, list); | |
1996 | goto out; | |
1997 | case BLK_STS_ZONE_RESOURCE: | |
0512a75b KB |
1998 | /* |
1999 | * Move the request to zone_list and keep going through | |
2000 | * the dispatch list to find more requests the drive can | |
2001 | * accept. | |
2002 | */ | |
2003 | blk_mq_handle_zone_resource(rq, &zone_list); | |
9586e67b | 2004 | needs_resource = true; |
7bf13729 ML |
2005 | break; |
2006 | default: | |
93efe981 | 2007 | errors++; |
e21ee5a6 | 2008 | blk_mq_end_request(rq, ret); |
320ae51f | 2009 | } |
81380ca1 | 2010 | } while (!list_empty(list)); |
7bf13729 | 2011 | out: |
0512a75b KB |
2012 | if (!list_empty(&zone_list)) |
2013 | list_splice_tail_init(&zone_list, list); | |
2014 | ||
632bfb63 | 2015 | /* If we didn't flush the entire list, we could have told the driver |
2016 | * there was more coming, but that turned out to be a lie. | |
2017 | */ | |
65fac0d5 YK |
2018 | if ((!list_empty(list) || errors || needs_resource || |
2019 | ret == BLK_STS_DEV_RESOURCE) && q->mq_ops->commit_rqs && queued) | |
632bfb63 | 2020 | q->mq_ops->commit_rqs(hctx); |
320ae51f JA |
2021 | /* |
2022 | * Any items that need requeuing? Stuff them into hctx->dispatch, | |
2023 | * that is where we will continue on next queue run. | |
2024 | */ | |
f04c3df3 | 2025 | if (!list_empty(list)) { |
86ff7c2a | 2026 | bool needs_restart; |
75383524 ML |
2027 | /* For non-shared tags, the RESTART check will suffice */ |
2028 | bool no_tag = prep == PREP_DISPATCH_NO_TAG && | |
51db1c37 | 2029 | (hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED); |
86ff7c2a | 2030 | |
2a5a24aa ML |
2031 | if (nr_budgets) |
2032 | blk_mq_release_budgets(q, list); | |
86ff7c2a | 2033 | |
320ae51f | 2034 | spin_lock(&hctx->lock); |
01e99aec | 2035 | list_splice_tail_init(list, &hctx->dispatch); |
320ae51f | 2036 | spin_unlock(&hctx->lock); |
f04c3df3 | 2037 | |
d7d8535f ML |
2038 | /* |
2039 | * Order adding requests to hctx->dispatch and checking | |
2040 | * SCHED_RESTART flag. The pair of this smp_mb() is the one | |
2041 | * in blk_mq_sched_restart(). Avoid restart code path to | |
2042 | * miss the new added requests to hctx->dispatch, meantime | |
2043 | * SCHED_RESTART is observed here. | |
2044 | */ | |
2045 | smp_mb(); | |
2046 | ||
9ba52e58 | 2047 | /* |
710c785f BVA |
2048 | * If SCHED_RESTART was set by the caller of this function and |
2049 | * it is no longer set that means that it was cleared by another | |
2050 | * thread and hence that a queue rerun is needed. | |
9ba52e58 | 2051 | * |
eb619fdb JA |
2052 | * If 'no_tag' is set, that means that we failed getting |
2053 | * a driver tag with an I/O scheduler attached. If our dispatch | |
2054 | * waitqueue is no longer active, ensure that we run the queue | |
2055 | * AFTER adding our entries back to the list. | |
bd166ef1 | 2056 | * |
710c785f BVA |
2057 | * If no I/O scheduler has been configured it is possible that |
2058 | * the hardware queue got stopped and restarted before requests | |
2059 | * were pushed back onto the dispatch list. Rerun the queue to | |
2060 | * avoid starvation. Notes: | |
2061 | * - blk_mq_run_hw_queue() checks whether or not a queue has | |
2062 | * been stopped before rerunning a queue. | |
2063 | * - Some but not all block drivers stop a queue before | |
fc17b653 | 2064 | * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq |
710c785f | 2065 | * and dm-rq. |
86ff7c2a ML |
2066 | * |
2067 | * If driver returns BLK_STS_RESOURCE and SCHED_RESTART | |
2068 | * bit is set, run queue after a delay to avoid IO stalls | |
ab3cee37 | 2069 | * that could otherwise occur if the queue is idle. We'll do |
9586e67b NA |
2070 | * similar if we couldn't get budget or couldn't lock a zone |
2071 | * and SCHED_RESTART is set. | |
bd166ef1 | 2072 | */ |
86ff7c2a | 2073 | needs_restart = blk_mq_sched_needs_restart(hctx); |
9586e67b NA |
2074 | if (prep == PREP_DISPATCH_NO_BUDGET) |
2075 | needs_resource = true; | |
86ff7c2a | 2076 | if (!needs_restart || |
eb619fdb | 2077 | (no_tag && list_empty_careful(&hctx->dispatch_wait.entry))) |
bd166ef1 | 2078 | blk_mq_run_hw_queue(hctx, true); |
6d5e8d21 | 2079 | else if (needs_resource) |
86ff7c2a | 2080 | blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY); |
1f57f8d4 | 2081 | |
6e768717 | 2082 | blk_mq_update_dispatch_busy(hctx, true); |
1f57f8d4 | 2083 | return false; |
6e768717 ML |
2084 | } else |
2085 | blk_mq_update_dispatch_busy(hctx, false); | |
f04c3df3 | 2086 | |
93efe981 | 2087 | return (queued + errors) != 0; |
f04c3df3 JA |
2088 | } |
2089 | ||
105663f7 AA |
2090 | /** |
2091 | * __blk_mq_run_hw_queue - Run a hardware queue. | |
2092 | * @hctx: Pointer to the hardware queue to run. | |
2093 | * | |
2094 | * Send pending requests to the hardware. | |
2095 | */ | |
6a83e74d BVA |
2096 | static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx) |
2097 | { | |
b7a71e66 JA |
2098 | /* |
2099 | * We can't run the queue inline with ints disabled. Ensure that | |
2100 | * we catch bad users of this early. | |
2101 | */ | |
2102 | WARN_ON_ONCE(in_interrupt()); | |
2103 | ||
bcc330f4 ML |
2104 | blk_mq_run_dispatch_ops(hctx->queue, |
2105 | blk_mq_sched_dispatch_requests(hctx)); | |
6a83e74d BVA |
2106 | } |
2107 | ||
f82ddf19 ML |
2108 | static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx) |
2109 | { | |
2110 | int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask); | |
2111 | ||
2112 | if (cpu >= nr_cpu_ids) | |
2113 | cpu = cpumask_first(hctx->cpumask); | |
2114 | return cpu; | |
2115 | } | |
2116 | ||
506e931f JA |
2117 | /* |
2118 | * It'd be great if the workqueue API had a way to pass | |
2119 | * in a mask and had some smarts for more clever placement. | |
2120 | * For now we just round-robin here, switching for every | |
2121 | * BLK_MQ_CPU_WORK_BATCH queued items. | |
2122 | */ | |
2123 | static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx) | |
2124 | { | |
7bed4595 | 2125 | bool tried = false; |
476f8c98 | 2126 | int next_cpu = hctx->next_cpu; |
7bed4595 | 2127 | |
b657d7e6 CH |
2128 | if (hctx->queue->nr_hw_queues == 1) |
2129 | return WORK_CPU_UNBOUND; | |
506e931f JA |
2130 | |
2131 | if (--hctx->next_cpu_batch <= 0) { | |
7bed4595 | 2132 | select_cpu: |
476f8c98 | 2133 | next_cpu = cpumask_next_and(next_cpu, hctx->cpumask, |
20e4d813 | 2134 | cpu_online_mask); |
506e931f | 2135 | if (next_cpu >= nr_cpu_ids) |
f82ddf19 | 2136 | next_cpu = blk_mq_first_mapped_cpu(hctx); |
506e931f JA |
2137 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; |
2138 | } | |
2139 | ||
7bed4595 ML |
2140 | /* |
2141 | * Do unbound schedule if we can't find a online CPU for this hctx, | |
2142 | * and it should only happen in the path of handling CPU DEAD. | |
2143 | */ | |
476f8c98 | 2144 | if (!cpu_online(next_cpu)) { |
7bed4595 ML |
2145 | if (!tried) { |
2146 | tried = true; | |
2147 | goto select_cpu; | |
2148 | } | |
2149 | ||
2150 | /* | |
2151 | * Make sure to re-select CPU next time once after CPUs | |
2152 | * in hctx->cpumask become online again. | |
2153 | */ | |
476f8c98 | 2154 | hctx->next_cpu = next_cpu; |
7bed4595 ML |
2155 | hctx->next_cpu_batch = 1; |
2156 | return WORK_CPU_UNBOUND; | |
2157 | } | |
476f8c98 ML |
2158 | |
2159 | hctx->next_cpu = next_cpu; | |
2160 | return next_cpu; | |
506e931f JA |
2161 | } |
2162 | ||
105663f7 AA |
2163 | /** |
2164 | * __blk_mq_delay_run_hw_queue - Run (or schedule to run) a hardware queue. | |
2165 | * @hctx: Pointer to the hardware queue to run. | |
2166 | * @async: If we want to run the queue asynchronously. | |
fa94ba8a | 2167 | * @msecs: Milliseconds of delay to wait before running the queue. |
105663f7 AA |
2168 | * |
2169 | * If !@async, try to run the queue now. Else, run the queue asynchronously and | |
2170 | * with a delay of @msecs. | |
2171 | */ | |
7587a5ae BVA |
2172 | static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async, |
2173 | unsigned long msecs) | |
320ae51f | 2174 | { |
5435c023 | 2175 | if (unlikely(blk_mq_hctx_stopped(hctx))) |
320ae51f JA |
2176 | return; |
2177 | ||
1b792f2f | 2178 | if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) { |
3c8f9da4 | 2179 | if (cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask)) { |
398205b8 | 2180 | __blk_mq_run_hw_queue(hctx); |
398205b8 PB |
2181 | return; |
2182 | } | |
e4043dcf | 2183 | } |
398205b8 | 2184 | |
ae943d20 BVA |
2185 | kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work, |
2186 | msecs_to_jiffies(msecs)); | |
7587a5ae BVA |
2187 | } |
2188 | ||
105663f7 AA |
2189 | /** |
2190 | * blk_mq_delay_run_hw_queue - Run a hardware queue asynchronously. | |
2191 | * @hctx: Pointer to the hardware queue to run. | |
fa94ba8a | 2192 | * @msecs: Milliseconds of delay to wait before running the queue. |
105663f7 AA |
2193 | * |
2194 | * Run a hardware queue asynchronously with a delay of @msecs. | |
2195 | */ | |
7587a5ae BVA |
2196 | void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs) |
2197 | { | |
2198 | __blk_mq_delay_run_hw_queue(hctx, true, msecs); | |
2199 | } | |
2200 | EXPORT_SYMBOL(blk_mq_delay_run_hw_queue); | |
2201 | ||
105663f7 AA |
2202 | /** |
2203 | * blk_mq_run_hw_queue - Start to run a hardware queue. | |
2204 | * @hctx: Pointer to the hardware queue to run. | |
2205 | * @async: If we want to run the queue asynchronously. | |
2206 | * | |
2207 | * Check if the request queue is not in a quiesced state and if there are | |
2208 | * pending requests to be sent. If this is true, run the queue to send requests | |
2209 | * to hardware. | |
2210 | */ | |
626fb735 | 2211 | void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) |
7587a5ae | 2212 | { |
24f5a90f ML |
2213 | bool need_run; |
2214 | ||
2215 | /* | |
2216 | * When queue is quiesced, we may be switching io scheduler, or | |
2217 | * updating nr_hw_queues, or other things, and we can't run queue | |
2218 | * any more, even __blk_mq_hctx_has_pending() can't be called safely. | |
2219 | * | |
2220 | * And queue will be rerun in blk_mq_unquiesce_queue() if it is | |
2221 | * quiesced. | |
2222 | */ | |
41adf531 | 2223 | __blk_mq_run_dispatch_ops(hctx->queue, false, |
2a904d00 ML |
2224 | need_run = !blk_queue_quiesced(hctx->queue) && |
2225 | blk_mq_hctx_has_pending(hctx)); | |
24f5a90f | 2226 | |
626fb735 | 2227 | if (need_run) |
79f720a7 | 2228 | __blk_mq_delay_run_hw_queue(hctx, async, 0); |
320ae51f | 2229 | } |
5b727272 | 2230 | EXPORT_SYMBOL(blk_mq_run_hw_queue); |
320ae51f | 2231 | |
b6e68ee8 JK |
2232 | /* |
2233 | * Return prefered queue to dispatch from (if any) for non-mq aware IO | |
2234 | * scheduler. | |
2235 | */ | |
2236 | static struct blk_mq_hw_ctx *blk_mq_get_sq_hctx(struct request_queue *q) | |
2237 | { | |
5d05426e | 2238 | struct blk_mq_ctx *ctx = blk_mq_get_ctx(q); |
b6e68ee8 JK |
2239 | /* |
2240 | * If the IO scheduler does not respect hardware queues when | |
2241 | * dispatching, we just don't bother with multiple HW queues and | |
2242 | * dispatch from hctx for the current CPU since running multiple queues | |
2243 | * just causes lock contention inside the scheduler and pointless cache | |
2244 | * bouncing. | |
2245 | */ | |
51ab80f0 | 2246 | struct blk_mq_hw_ctx *hctx = ctx->hctxs[HCTX_TYPE_DEFAULT]; |
5d05426e | 2247 | |
b6e68ee8 JK |
2248 | if (!blk_mq_hctx_stopped(hctx)) |
2249 | return hctx; | |
2250 | return NULL; | |
2251 | } | |
2252 | ||
105663f7 | 2253 | /** |
24f7bb88 | 2254 | * blk_mq_run_hw_queues - Run all hardware queues in a request queue. |
105663f7 AA |
2255 | * @q: Pointer to the request queue to run. |
2256 | * @async: If we want to run the queue asynchronously. | |
2257 | */ | |
b94ec296 | 2258 | void blk_mq_run_hw_queues(struct request_queue *q, bool async) |
320ae51f | 2259 | { |
b6e68ee8 | 2260 | struct blk_mq_hw_ctx *hctx, *sq_hctx; |
4f481208 | 2261 | unsigned long i; |
320ae51f | 2262 | |
b6e68ee8 | 2263 | sq_hctx = NULL; |
4d337ceb | 2264 | if (blk_queue_sq_sched(q)) |
b6e68ee8 | 2265 | sq_hctx = blk_mq_get_sq_hctx(q); |
320ae51f | 2266 | queue_for_each_hw_ctx(q, hctx, i) { |
79f720a7 | 2267 | if (blk_mq_hctx_stopped(hctx)) |
320ae51f | 2268 | continue; |
b6e68ee8 JK |
2269 | /* |
2270 | * Dispatch from this hctx either if there's no hctx preferred | |
2271 | * by IO scheduler or if it has requests that bypass the | |
2272 | * scheduler. | |
2273 | */ | |
2274 | if (!sq_hctx || sq_hctx == hctx || | |
2275 | !list_empty_careful(&hctx->dispatch)) | |
2276 | blk_mq_run_hw_queue(hctx, async); | |
320ae51f JA |
2277 | } |
2278 | } | |
b94ec296 | 2279 | EXPORT_SYMBOL(blk_mq_run_hw_queues); |
320ae51f | 2280 | |
b9151e7b DA |
2281 | /** |
2282 | * blk_mq_delay_run_hw_queues - Run all hardware queues asynchronously. | |
2283 | * @q: Pointer to the request queue to run. | |
fa94ba8a | 2284 | * @msecs: Milliseconds of delay to wait before running the queues. |
b9151e7b DA |
2285 | */ |
2286 | void blk_mq_delay_run_hw_queues(struct request_queue *q, unsigned long msecs) | |
2287 | { | |
b6e68ee8 | 2288 | struct blk_mq_hw_ctx *hctx, *sq_hctx; |
4f481208 | 2289 | unsigned long i; |
b9151e7b | 2290 | |
b6e68ee8 | 2291 | sq_hctx = NULL; |
4d337ceb | 2292 | if (blk_queue_sq_sched(q)) |
b6e68ee8 | 2293 | sq_hctx = blk_mq_get_sq_hctx(q); |
b9151e7b DA |
2294 | queue_for_each_hw_ctx(q, hctx, i) { |
2295 | if (blk_mq_hctx_stopped(hctx)) | |
2296 | continue; | |
8f5fea65 DJ |
2297 | /* |
2298 | * If there is already a run_work pending, leave the | |
2299 | * pending delay untouched. Otherwise, a hctx can stall | |
2300 | * if another hctx is re-delaying the other's work | |
2301 | * before the work executes. | |
2302 | */ | |
2303 | if (delayed_work_pending(&hctx->run_work)) | |
2304 | continue; | |
b6e68ee8 JK |
2305 | /* |
2306 | * Dispatch from this hctx either if there's no hctx preferred | |
2307 | * by IO scheduler or if it has requests that bypass the | |
2308 | * scheduler. | |
2309 | */ | |
2310 | if (!sq_hctx || sq_hctx == hctx || | |
2311 | !list_empty_careful(&hctx->dispatch)) | |
2312 | blk_mq_delay_run_hw_queue(hctx, msecs); | |
b9151e7b DA |
2313 | } |
2314 | } | |
2315 | EXPORT_SYMBOL(blk_mq_delay_run_hw_queues); | |
2316 | ||
39a70c76 ML |
2317 | /* |
2318 | * This function is often used for pausing .queue_rq() by driver when | |
2319 | * there isn't enough resource or some conditions aren't satisfied, and | |
4d606219 | 2320 | * BLK_STS_RESOURCE is usually returned. |
39a70c76 ML |
2321 | * |
2322 | * We do not guarantee that dispatch can be drained or blocked | |
2323 | * after blk_mq_stop_hw_queue() returns. Please use | |
2324 | * blk_mq_quiesce_queue() for that requirement. | |
2325 | */ | |
2719aa21 JA |
2326 | void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx) |
2327 | { | |
641a9ed6 | 2328 | cancel_delayed_work(&hctx->run_work); |
280d45f6 | 2329 | |
641a9ed6 | 2330 | set_bit(BLK_MQ_S_STOPPED, &hctx->state); |
2719aa21 | 2331 | } |
641a9ed6 | 2332 | EXPORT_SYMBOL(blk_mq_stop_hw_queue); |
2719aa21 | 2333 | |
39a70c76 ML |
2334 | /* |
2335 | * This function is often used for pausing .queue_rq() by driver when | |
2336 | * there isn't enough resource or some conditions aren't satisfied, and | |
4d606219 | 2337 | * BLK_STS_RESOURCE is usually returned. |
39a70c76 ML |
2338 | * |
2339 | * We do not guarantee that dispatch can be drained or blocked | |
2340 | * after blk_mq_stop_hw_queues() returns. Please use | |
2341 | * blk_mq_quiesce_queue() for that requirement. | |
2342 | */ | |
2719aa21 JA |
2343 | void blk_mq_stop_hw_queues(struct request_queue *q) |
2344 | { | |
641a9ed6 | 2345 | struct blk_mq_hw_ctx *hctx; |
4f481208 | 2346 | unsigned long i; |
641a9ed6 ML |
2347 | |
2348 | queue_for_each_hw_ctx(q, hctx, i) | |
2349 | blk_mq_stop_hw_queue(hctx); | |
280d45f6 CH |
2350 | } |
2351 | EXPORT_SYMBOL(blk_mq_stop_hw_queues); | |
2352 | ||
320ae51f JA |
2353 | void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx) |
2354 | { | |
2355 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
e4043dcf | 2356 | |
0ffbce80 | 2357 | blk_mq_run_hw_queue(hctx, false); |
320ae51f JA |
2358 | } |
2359 | EXPORT_SYMBOL(blk_mq_start_hw_queue); | |
2360 | ||
2f268556 CH |
2361 | void blk_mq_start_hw_queues(struct request_queue *q) |
2362 | { | |
2363 | struct blk_mq_hw_ctx *hctx; | |
4f481208 | 2364 | unsigned long i; |
2f268556 CH |
2365 | |
2366 | queue_for_each_hw_ctx(q, hctx, i) | |
2367 | blk_mq_start_hw_queue(hctx); | |
2368 | } | |
2369 | EXPORT_SYMBOL(blk_mq_start_hw_queues); | |
2370 | ||
ae911c5e JA |
2371 | void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) |
2372 | { | |
2373 | if (!blk_mq_hctx_stopped(hctx)) | |
2374 | return; | |
2375 | ||
2376 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
2377 | blk_mq_run_hw_queue(hctx, async); | |
2378 | } | |
2379 | EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue); | |
2380 | ||
1b4a3258 | 2381 | void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async) |
320ae51f JA |
2382 | { |
2383 | struct blk_mq_hw_ctx *hctx; | |
4f481208 | 2384 | unsigned long i; |
320ae51f | 2385 | |
ae911c5e JA |
2386 | queue_for_each_hw_ctx(q, hctx, i) |
2387 | blk_mq_start_stopped_hw_queue(hctx, async); | |
320ae51f JA |
2388 | } |
2389 | EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues); | |
2390 | ||
70f4db63 | 2391 | static void blk_mq_run_work_fn(struct work_struct *work) |
320ae51f JA |
2392 | { |
2393 | struct blk_mq_hw_ctx *hctx; | |
2394 | ||
9f993737 | 2395 | hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work); |
320ae51f | 2396 | |
21c6e939 | 2397 | /* |
15fe8a90 | 2398 | * If we are stopped, don't run the queue. |
21c6e939 | 2399 | */ |
0841031a | 2400 | if (blk_mq_hctx_stopped(hctx)) |
0196d6b4 | 2401 | return; |
7587a5ae BVA |
2402 | |
2403 | __blk_mq_run_hw_queue(hctx); | |
2404 | } | |
2405 | ||
cfd0c552 | 2406 | static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx, |
cfd0c552 ML |
2407 | struct request *rq, |
2408 | bool at_head) | |
320ae51f | 2409 | { |
e57690fe | 2410 | struct blk_mq_ctx *ctx = rq->mq_ctx; |
c16d6b5a | 2411 | enum hctx_type type = hctx->type; |
e57690fe | 2412 | |
7b607814 BVA |
2413 | lockdep_assert_held(&ctx->lock); |
2414 | ||
a54895fa | 2415 | trace_block_rq_insert(rq); |
01b983c9 | 2416 | |
72a0a36e | 2417 | if (at_head) |
c16d6b5a | 2418 | list_add(&rq->queuelist, &ctx->rq_lists[type]); |
72a0a36e | 2419 | else |
c16d6b5a | 2420 | list_add_tail(&rq->queuelist, &ctx->rq_lists[type]); |
cfd0c552 | 2421 | } |
4bb659b1 | 2422 | |
2c3ad667 JA |
2423 | void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq, |
2424 | bool at_head) | |
cfd0c552 ML |
2425 | { |
2426 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
2427 | ||
7b607814 BVA |
2428 | lockdep_assert_held(&ctx->lock); |
2429 | ||
e57690fe | 2430 | __blk_mq_insert_req_list(hctx, rq, at_head); |
320ae51f | 2431 | blk_mq_hctx_mark_pending(hctx, ctx); |
320ae51f JA |
2432 | } |
2433 | ||
105663f7 AA |
2434 | /** |
2435 | * blk_mq_request_bypass_insert - Insert a request at dispatch list. | |
2436 | * @rq: Pointer to request to be inserted. | |
26bfeb26 | 2437 | * @at_head: true if the request should be inserted at the head of the list. |
105663f7 AA |
2438 | * @run_queue: If we should run the hardware queue after inserting the request. |
2439 | * | |
157f377b JA |
2440 | * Should only be used carefully, when the caller knows we want to |
2441 | * bypass a potential IO scheduler on the target device. | |
2442 | */ | |
01e99aec ML |
2443 | void blk_mq_request_bypass_insert(struct request *rq, bool at_head, |
2444 | bool run_queue) | |
157f377b | 2445 | { |
ea4f995e | 2446 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; |
157f377b JA |
2447 | |
2448 | spin_lock(&hctx->lock); | |
01e99aec ML |
2449 | if (at_head) |
2450 | list_add(&rq->queuelist, &hctx->dispatch); | |
2451 | else | |
2452 | list_add_tail(&rq->queuelist, &hctx->dispatch); | |
157f377b JA |
2453 | spin_unlock(&hctx->lock); |
2454 | ||
b0850297 ML |
2455 | if (run_queue) |
2456 | blk_mq_run_hw_queue(hctx, false); | |
157f377b JA |
2457 | } |
2458 | ||
bd166ef1 JA |
2459 | void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx, |
2460 | struct list_head *list) | |
320ae51f JA |
2461 | |
2462 | { | |
3f0cedc7 | 2463 | struct request *rq; |
c16d6b5a | 2464 | enum hctx_type type = hctx->type; |
3f0cedc7 | 2465 | |
320ae51f JA |
2466 | /* |
2467 | * preemption doesn't flush plug list, so it's possible ctx->cpu is | |
2468 | * offline now | |
2469 | */ | |
3f0cedc7 | 2470 | list_for_each_entry(rq, list, queuelist) { |
e57690fe | 2471 | BUG_ON(rq->mq_ctx != ctx); |
a54895fa | 2472 | trace_block_rq_insert(rq); |
320ae51f | 2473 | } |
3f0cedc7 ML |
2474 | |
2475 | spin_lock(&ctx->lock); | |
c16d6b5a | 2476 | list_splice_tail_init(list, &ctx->rq_lists[type]); |
cfd0c552 | 2477 | blk_mq_hctx_mark_pending(hctx, ctx); |
320ae51f | 2478 | spin_unlock(&ctx->lock); |
320ae51f JA |
2479 | } |
2480 | ||
dc5fc361 JA |
2481 | static void blk_mq_commit_rqs(struct blk_mq_hw_ctx *hctx, int *queued, |
2482 | bool from_schedule) | |
320ae51f | 2483 | { |
dc5fc361 JA |
2484 | if (hctx->queue->mq_ops->commit_rqs) { |
2485 | trace_block_unplug(hctx->queue, *queued, !from_schedule); | |
2486 | hctx->queue->mq_ops->commit_rqs(hctx); | |
2487 | } | |
2488 | *queued = 0; | |
2489 | } | |
320ae51f | 2490 | |
14ccb66b CH |
2491 | static void blk_mq_bio_to_request(struct request *rq, struct bio *bio, |
2492 | unsigned int nr_segs) | |
320ae51f | 2493 | { |
93f221ae EB |
2494 | int err; |
2495 | ||
f924cdde CH |
2496 | if (bio->bi_opf & REQ_RAHEAD) |
2497 | rq->cmd_flags |= REQ_FAILFAST_MASK; | |
2498 | ||
2499 | rq->__sector = bio->bi_iter.bi_sector; | |
14ccb66b | 2500 | blk_rq_bio_prep(rq, bio, nr_segs); |
93f221ae EB |
2501 | |
2502 | /* This can't fail, since GFP_NOIO includes __GFP_DIRECT_RECLAIM. */ | |
2503 | err = blk_crypto_rq_bio_prep(rq, bio, GFP_NOIO); | |
2504 | WARN_ON_ONCE(err); | |
4b570521 | 2505 | |
b5af37ab | 2506 | blk_account_io_start(rq); |
320ae51f JA |
2507 | } |
2508 | ||
0f95549c | 2509 | static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx, |
3e08773c | 2510 | struct request *rq, bool last) |
f984df1f | 2511 | { |
f984df1f | 2512 | struct request_queue *q = rq->q; |
f984df1f SL |
2513 | struct blk_mq_queue_data bd = { |
2514 | .rq = rq, | |
be94f058 | 2515 | .last = last, |
f984df1f | 2516 | }; |
f06345ad | 2517 | blk_status_t ret; |
0f95549c | 2518 | |
0f95549c MS |
2519 | /* |
2520 | * For OK queue, we are done. For error, caller may kill it. | |
2521 | * Any other error (busy), just add it to our list as we | |
2522 | * previously would have done. | |
2523 | */ | |
2524 | ret = q->mq_ops->queue_rq(hctx, &bd); | |
2525 | switch (ret) { | |
2526 | case BLK_STS_OK: | |
6ce3dd6e | 2527 | blk_mq_update_dispatch_busy(hctx, false); |
0f95549c MS |
2528 | break; |
2529 | case BLK_STS_RESOURCE: | |
86ff7c2a | 2530 | case BLK_STS_DEV_RESOURCE: |
6ce3dd6e | 2531 | blk_mq_update_dispatch_busy(hctx, true); |
0f95549c MS |
2532 | __blk_mq_requeue_request(rq); |
2533 | break; | |
2534 | default: | |
6ce3dd6e | 2535 | blk_mq_update_dispatch_busy(hctx, false); |
0f95549c MS |
2536 | break; |
2537 | } | |
2538 | ||
2539 | return ret; | |
2540 | } | |
2541 | ||
fd9c40f6 | 2542 | static blk_status_t __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx, |
0f95549c | 2543 | struct request *rq, |
fd9c40f6 | 2544 | bool bypass_insert, bool last) |
0f95549c MS |
2545 | { |
2546 | struct request_queue *q = rq->q; | |
d964f04a | 2547 | bool run_queue = true; |
2a5a24aa | 2548 | int budget_token; |
d964f04a | 2549 | |
23d4ee19 | 2550 | /* |
fd9c40f6 | 2551 | * RCU or SRCU read lock is needed before checking quiesced flag. |
23d4ee19 | 2552 | * |
fd9c40f6 BVA |
2553 | * When queue is stopped or quiesced, ignore 'bypass_insert' from |
2554 | * blk_mq_request_issue_directly(), and return BLK_STS_OK to caller, | |
2555 | * and avoid driver to try to dispatch again. | |
23d4ee19 | 2556 | */ |
fd9c40f6 | 2557 | if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)) { |
d964f04a | 2558 | run_queue = false; |
fd9c40f6 BVA |
2559 | bypass_insert = false; |
2560 | goto insert; | |
d964f04a | 2561 | } |
f984df1f | 2562 | |
2ff0682d | 2563 | if ((rq->rq_flags & RQF_ELV) && !bypass_insert) |
fd9c40f6 | 2564 | goto insert; |
2253efc8 | 2565 | |
2a5a24aa ML |
2566 | budget_token = blk_mq_get_dispatch_budget(q); |
2567 | if (budget_token < 0) | |
fd9c40f6 | 2568 | goto insert; |
bd166ef1 | 2569 | |
2a5a24aa ML |
2570 | blk_mq_set_rq_budget_token(rq, budget_token); |
2571 | ||
8ab6bb9e | 2572 | if (!blk_mq_get_driver_tag(rq)) { |
2a5a24aa | 2573 | blk_mq_put_dispatch_budget(q, budget_token); |
fd9c40f6 | 2574 | goto insert; |
88022d72 | 2575 | } |
de148297 | 2576 | |
3e08773c | 2577 | return __blk_mq_issue_directly(hctx, rq, last); |
fd9c40f6 BVA |
2578 | insert: |
2579 | if (bypass_insert) | |
2580 | return BLK_STS_RESOURCE; | |
2581 | ||
db03f88f ML |
2582 | blk_mq_sched_insert_request(rq, false, run_queue, false); |
2583 | ||
fd9c40f6 BVA |
2584 | return BLK_STS_OK; |
2585 | } | |
2586 | ||
105663f7 AA |
2587 | /** |
2588 | * blk_mq_try_issue_directly - Try to send a request directly to device driver. | |
2589 | * @hctx: Pointer of the associated hardware queue. | |
2590 | * @rq: Pointer to request to be sent. | |
105663f7 AA |
2591 | * |
2592 | * If the device has enough resources to accept a new request now, send the | |
2593 | * request directly to device driver. Else, insert at hctx->dispatch queue, so | |
2594 | * we can try send it another time in the future. Requests inserted at this | |
2595 | * queue have higher priority. | |
2596 | */ | |
fd9c40f6 | 2597 | static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx, |
3e08773c | 2598 | struct request *rq) |
fd9c40f6 | 2599 | { |
2a904d00 ML |
2600 | blk_status_t ret = |
2601 | __blk_mq_try_issue_directly(hctx, rq, false, true); | |
fd9c40f6 | 2602 | |
fd9c40f6 | 2603 | if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE) |
01e99aec | 2604 | blk_mq_request_bypass_insert(rq, false, true); |
fd9c40f6 BVA |
2605 | else if (ret != BLK_STS_OK) |
2606 | blk_mq_end_request(rq, ret); | |
fd9c40f6 BVA |
2607 | } |
2608 | ||
06c8c691 | 2609 | static blk_status_t blk_mq_request_issue_directly(struct request *rq, bool last) |
fd9c40f6 | 2610 | { |
4cafe86c | 2611 | return __blk_mq_try_issue_directly(rq->mq_hctx, rq, true, last); |
5eb6126e CH |
2612 | } |
2613 | ||
b84c5b50 CH |
2614 | static void blk_mq_plug_issue_direct(struct blk_plug *plug, bool from_schedule) |
2615 | { | |
2616 | struct blk_mq_hw_ctx *hctx = NULL; | |
2617 | struct request *rq; | |
2618 | int queued = 0; | |
2619 | int errors = 0; | |
2620 | ||
2621 | while ((rq = rq_list_pop(&plug->mq_list))) { | |
2622 | bool last = rq_list_empty(plug->mq_list); | |
2623 | blk_status_t ret; | |
2624 | ||
2625 | if (hctx != rq->mq_hctx) { | |
2626 | if (hctx) | |
2627 | blk_mq_commit_rqs(hctx, &queued, from_schedule); | |
2628 | hctx = rq->mq_hctx; | |
2629 | } | |
2630 | ||
2631 | ret = blk_mq_request_issue_directly(rq, last); | |
2632 | switch (ret) { | |
2633 | case BLK_STS_OK: | |
2634 | queued++; | |
2635 | break; | |
2636 | case BLK_STS_RESOURCE: | |
2637 | case BLK_STS_DEV_RESOURCE: | |
d3b38596 | 2638 | blk_mq_request_bypass_insert(rq, false, true); |
b84c5b50 CH |
2639 | blk_mq_commit_rqs(hctx, &queued, from_schedule); |
2640 | return; | |
2641 | default: | |
2642 | blk_mq_end_request(rq, ret); | |
2643 | errors++; | |
2644 | break; | |
2645 | } | |
2646 | } | |
2647 | ||
2648 | /* | |
2649 | * If we didn't flush the entire list, we could have told the driver | |
2650 | * there was more coming, but that turned out to be a lie. | |
2651 | */ | |
2652 | if (errors) | |
2653 | blk_mq_commit_rqs(hctx, &queued, from_schedule); | |
2654 | } | |
2655 | ||
518579a9 KB |
2656 | static void __blk_mq_flush_plug_list(struct request_queue *q, |
2657 | struct blk_plug *plug) | |
2658 | { | |
2659 | if (blk_queue_quiesced(q)) | |
2660 | return; | |
2661 | q->mq_ops->queue_rqs(&plug->mq_list); | |
2662 | } | |
2663 | ||
26fed4ac JA |
2664 | static void blk_mq_dispatch_plug_list(struct blk_plug *plug, bool from_sched) |
2665 | { | |
2666 | struct blk_mq_hw_ctx *this_hctx = NULL; | |
2667 | struct blk_mq_ctx *this_ctx = NULL; | |
2668 | struct request *requeue_list = NULL; | |
2669 | unsigned int depth = 0; | |
2670 | LIST_HEAD(list); | |
2671 | ||
2672 | do { | |
2673 | struct request *rq = rq_list_pop(&plug->mq_list); | |
2674 | ||
2675 | if (!this_hctx) { | |
2676 | this_hctx = rq->mq_hctx; | |
2677 | this_ctx = rq->mq_ctx; | |
2678 | } else if (this_hctx != rq->mq_hctx || this_ctx != rq->mq_ctx) { | |
2679 | rq_list_add(&requeue_list, rq); | |
2680 | continue; | |
2681 | } | |
2682 | list_add_tail(&rq->queuelist, &list); | |
2683 | depth++; | |
2684 | } while (!rq_list_empty(plug->mq_list)); | |
2685 | ||
2686 | plug->mq_list = requeue_list; | |
2687 | trace_block_unplug(this_hctx->queue, depth, !from_sched); | |
2688 | blk_mq_sched_insert_requests(this_hctx, this_ctx, &list, from_sched); | |
2689 | } | |
2690 | ||
b84c5b50 CH |
2691 | void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule) |
2692 | { | |
3c67d44d | 2693 | struct request *rq; |
b84c5b50 CH |
2694 | |
2695 | if (rq_list_empty(plug->mq_list)) | |
2696 | return; | |
2697 | plug->rq_count = 0; | |
2698 | ||
2699 | if (!plug->multiple_queues && !plug->has_elevator && !from_schedule) { | |
3c67d44d JA |
2700 | struct request_queue *q; |
2701 | ||
2702 | rq = rq_list_peek(&plug->mq_list); | |
2703 | q = rq->q; | |
2704 | ||
2705 | /* | |
2706 | * Peek first request and see if we have a ->queue_rqs() hook. | |
2707 | * If we do, we can dispatch the whole plug list in one go. We | |
2708 | * already know at this point that all requests belong to the | |
2709 | * same queue, caller must ensure that's the case. | |
2710 | * | |
2711 | * Since we pass off the full list to the driver at this point, | |
2712 | * we do not increment the active request count for the queue. | |
2713 | * Bypass shared tags for now because of that. | |
2714 | */ | |
2715 | if (q->mq_ops->queue_rqs && | |
2716 | !(rq->mq_hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) { | |
2717 | blk_mq_run_dispatch_ops(q, | |
518579a9 | 2718 | __blk_mq_flush_plug_list(q, plug)); |
3c67d44d JA |
2719 | if (rq_list_empty(plug->mq_list)) |
2720 | return; | |
2721 | } | |
73f3760e ML |
2722 | |
2723 | blk_mq_run_dispatch_ops(q, | |
4cafe86c | 2724 | blk_mq_plug_issue_direct(plug, false)); |
b84c5b50 CH |
2725 | if (rq_list_empty(plug->mq_list)) |
2726 | return; | |
2727 | } | |
2728 | ||
b84c5b50 | 2729 | do { |
26fed4ac | 2730 | blk_mq_dispatch_plug_list(plug, from_schedule); |
b84c5b50 | 2731 | } while (!rq_list_empty(plug->mq_list)); |
b84c5b50 CH |
2732 | } |
2733 | ||
6ce3dd6e ML |
2734 | void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx, |
2735 | struct list_head *list) | |
2736 | { | |
536167d4 | 2737 | int queued = 0; |
632bfb63 | 2738 | int errors = 0; |
536167d4 | 2739 | |
6ce3dd6e | 2740 | while (!list_empty(list)) { |
fd9c40f6 | 2741 | blk_status_t ret; |
6ce3dd6e ML |
2742 | struct request *rq = list_first_entry(list, struct request, |
2743 | queuelist); | |
2744 | ||
2745 | list_del_init(&rq->queuelist); | |
fd9c40f6 BVA |
2746 | ret = blk_mq_request_issue_directly(rq, list_empty(list)); |
2747 | if (ret != BLK_STS_OK) { | |
65fac0d5 | 2748 | errors++; |
fd9c40f6 BVA |
2749 | if (ret == BLK_STS_RESOURCE || |
2750 | ret == BLK_STS_DEV_RESOURCE) { | |
01e99aec | 2751 | blk_mq_request_bypass_insert(rq, false, |
c616cbee | 2752 | list_empty(list)); |
fd9c40f6 BVA |
2753 | break; |
2754 | } | |
2755 | blk_mq_end_request(rq, ret); | |
536167d4 KB |
2756 | } else |
2757 | queued++; | |
6ce3dd6e | 2758 | } |
d666ba98 JA |
2759 | |
2760 | /* | |
2761 | * If we didn't flush the entire list, we could have told | |
2762 | * the driver there was more coming, but that turned out to | |
2763 | * be a lie. | |
2764 | */ | |
632bfb63 | 2765 | if ((!list_empty(list) || errors) && |
2766 | hctx->queue->mq_ops->commit_rqs && queued) | |
d666ba98 | 2767 | hctx->queue->mq_ops->commit_rqs(hctx); |
6ce3dd6e ML |
2768 | } |
2769 | ||
b131f201 | 2770 | static bool blk_mq_attempt_bio_merge(struct request_queue *q, |
0c5bcc92 | 2771 | struct bio *bio, unsigned int nr_segs) |
900e0807 JA |
2772 | { |
2773 | if (!blk_queue_nomerges(q) && bio_mergeable(bio)) { | |
0c5bcc92 | 2774 | if (blk_attempt_plug_merge(q, bio, nr_segs)) |
900e0807 JA |
2775 | return true; |
2776 | if (blk_mq_sched_bio_merge(q, bio, nr_segs)) | |
2777 | return true; | |
2778 | } | |
2779 | return false; | |
2780 | } | |
2781 | ||
71539717 JA |
2782 | static struct request *blk_mq_get_new_requests(struct request_queue *q, |
2783 | struct blk_plug *plug, | |
0a5aa8d1 SK |
2784 | struct bio *bio, |
2785 | unsigned int nsegs) | |
71539717 JA |
2786 | { |
2787 | struct blk_mq_alloc_data data = { | |
2788 | .q = q, | |
2789 | .nr_tags = 1, | |
9d497e29 | 2790 | .cmd_flags = bio->bi_opf, |
71539717 JA |
2791 | }; |
2792 | struct request *rq; | |
2793 | ||
5b13bc8a | 2794 | if (unlikely(bio_queue_enter(bio))) |
b637108a | 2795 | return NULL; |
900e0807 | 2796 | |
0a5aa8d1 SK |
2797 | if (blk_mq_attempt_bio_merge(q, bio, nsegs)) |
2798 | goto queue_exit; | |
2799 | ||
2800 | rq_qos_throttle(q, bio); | |
2801 | ||
71539717 JA |
2802 | if (plug) { |
2803 | data.nr_tags = plug->nr_ios; | |
2804 | plug->nr_ios = 1; | |
2805 | data.cached_rq = &plug->cached_rq; | |
2806 | } | |
2807 | ||
2808 | rq = __blk_mq_alloc_requests(&data); | |
373b5416 JA |
2809 | if (rq) |
2810 | return rq; | |
71539717 JA |
2811 | rq_qos_cleanup(q, bio); |
2812 | if (bio->bi_opf & REQ_NOWAIT) | |
2813 | bio_wouldblock_error(bio); | |
0a5aa8d1 | 2814 | queue_exit: |
5b13bc8a | 2815 | blk_queue_exit(q); |
71539717 JA |
2816 | return NULL; |
2817 | } | |
2818 | ||
5b13bc8a | 2819 | static inline struct request *blk_mq_get_cached_request(struct request_queue *q, |
0a5aa8d1 | 2820 | struct blk_plug *plug, struct bio **bio, unsigned int nsegs) |
71539717 | 2821 | { |
b637108a | 2822 | struct request *rq; |
b637108a | 2823 | |
5b13bc8a CH |
2824 | if (!plug) |
2825 | return NULL; | |
2826 | rq = rq_list_peek(&plug->cached_rq); | |
2827 | if (!rq || rq->q != q) | |
2828 | return NULL; | |
71539717 | 2829 | |
0a5aa8d1 SK |
2830 | if (blk_mq_attempt_bio_merge(q, *bio, nsegs)) { |
2831 | *bio = NULL; | |
2832 | return NULL; | |
2833 | } | |
2834 | ||
0a5aa8d1 | 2835 | if (blk_mq_get_hctx_type((*bio)->bi_opf) != rq->mq_hctx->type) |
5b13bc8a | 2836 | return NULL; |
0a5aa8d1 | 2837 | if (op_is_flush(rq->cmd_flags) != op_is_flush((*bio)->bi_opf)) |
5b13bc8a CH |
2838 | return NULL; |
2839 | ||
2645672f JA |
2840 | /* |
2841 | * If any qos ->throttle() end up blocking, we will have flushed the | |
2842 | * plug and hence killed the cached_rq list as well. Pop this entry | |
2843 | * before we throttle. | |
2844 | */ | |
5b13bc8a | 2845 | plug->cached_rq = rq_list_next(rq); |
2645672f JA |
2846 | rq_qos_throttle(q, *bio); |
2847 | ||
2848 | rq->cmd_flags = (*bio)->bi_opf; | |
5b13bc8a | 2849 | INIT_LIST_HEAD(&rq->queuelist); |
5b13bc8a | 2850 | return rq; |
71539717 JA |
2851 | } |
2852 | ||
82b74cac JK |
2853 | static void bio_set_ioprio(struct bio *bio) |
2854 | { | |
a78418e6 JK |
2855 | /* Nobody set ioprio so far? Initialize it based on task's nice value */ |
2856 | if (IOPRIO_PRIO_CLASS(bio->bi_ioprio) == IOPRIO_CLASS_NONE) | |
2857 | bio->bi_ioprio = get_current_ioprio(); | |
82b74cac JK |
2858 | blkcg_set_ioprio(bio); |
2859 | } | |
2860 | ||
105663f7 | 2861 | /** |
c62b37d9 | 2862 | * blk_mq_submit_bio - Create and send a request to block device. |
105663f7 AA |
2863 | * @bio: Bio pointer. |
2864 | * | |
2865 | * Builds up a request structure from @q and @bio and send to the device. The | |
2866 | * request may not be queued directly to hardware if: | |
2867 | * * This request can be merged with another one | |
2868 | * * We want to place request at plug queue for possible future merging | |
2869 | * * There is an IO scheduler active at this queue | |
2870 | * | |
2871 | * It will not queue the request if there is an error with the bio, or at the | |
2872 | * request creation. | |
105663f7 | 2873 | */ |
3e08773c | 2874 | void blk_mq_submit_bio(struct bio *bio) |
07068d5b | 2875 | { |
ed6cddef | 2876 | struct request_queue *q = bdev_get_queue(bio->bi_bdev); |
6deacb3b | 2877 | struct blk_plug *plug = blk_mq_plug(bio); |
ef295ecf | 2878 | const int is_sync = op_is_sync(bio->bi_opf); |
07068d5b | 2879 | struct request *rq; |
abd45c15 | 2880 | unsigned int nr_segs = 1; |
a892c8d5 | 2881 | blk_status_t ret; |
07068d5b | 2882 | |
51d798cd | 2883 | bio = blk_queue_bounce(bio, q); |
c55ddd90 CH |
2884 | if (bio_may_exceed_limits(bio, &q->limits)) |
2885 | bio = __bio_split_to_limits(bio, &q->limits, &nr_segs); | |
f36ea50c | 2886 | |
e23947bd | 2887 | if (!bio_integrity_prep(bio)) |
900e0807 | 2888 | return; |
87760e5e | 2889 | |
9c6227e0 JK |
2890 | bio_set_ioprio(bio); |
2891 | ||
0a5aa8d1 | 2892 | rq = blk_mq_get_cached_request(q, plug, &bio, nr_segs); |
5b13bc8a | 2893 | if (!rq) { |
0a5aa8d1 SK |
2894 | if (!bio) |
2895 | return; | |
2896 | rq = blk_mq_get_new_requests(q, plug, bio, nr_segs); | |
5b13bc8a CH |
2897 | if (unlikely(!rq)) |
2898 | return; | |
2899 | } | |
87760e5e | 2900 | |
e8a676d6 | 2901 | trace_block_getrq(bio); |
d6f1dda2 | 2902 | |
c1c80384 | 2903 | rq_qos_track(q, rq, bio); |
07068d5b | 2904 | |
970d168d BVA |
2905 | blk_mq_bio_to_request(rq, bio, nr_segs); |
2906 | ||
a892c8d5 ST |
2907 | ret = blk_crypto_init_request(rq); |
2908 | if (ret != BLK_STS_OK) { | |
2909 | bio->bi_status = ret; | |
2910 | bio_endio(bio); | |
2911 | blk_mq_free_request(rq); | |
3e08773c | 2912 | return; |
a892c8d5 ST |
2913 | } |
2914 | ||
2b504bd4 ML |
2915 | if (op_is_flush(bio->bi_opf)) { |
2916 | blk_insert_flush(rq); | |
d92ca9d8 | 2917 | return; |
2b504bd4 | 2918 | } |
d92ca9d8 | 2919 | |
1e9c2303 | 2920 | if (plug) |
ce5b009c | 2921 | blk_add_rq_to_plug(plug, rq); |
1e9c2303 CH |
2922 | else if ((rq->rq_flags & RQF_ELV) || |
2923 | (rq->mq_hctx->dispatch_busy && | |
2924 | (q->nr_hw_queues == 1 || !is_sync))) | |
a12de1d4 | 2925 | blk_mq_sched_insert_request(rq, false, true, true); |
1e9c2303 | 2926 | else |
bcc330f4 | 2927 | blk_mq_run_dispatch_ops(rq->q, |
2a904d00 | 2928 | blk_mq_try_issue_directly(rq->mq_hctx, rq)); |
320ae51f JA |
2929 | } |
2930 | ||
248c7933 | 2931 | #ifdef CONFIG_BLK_MQ_STACKING |
06c8c691 | 2932 | /** |
a5efda3c | 2933 | * blk_insert_cloned_request - Helper for stacking drivers to submit a request |
a5efda3c | 2934 | * @rq: the request being queued |
06c8c691 | 2935 | */ |
28db4711 | 2936 | blk_status_t blk_insert_cloned_request(struct request *rq) |
06c8c691 | 2937 | { |
28db4711 | 2938 | struct request_queue *q = rq->q; |
06c8c691 | 2939 | unsigned int max_sectors = blk_queue_get_max_sectors(q, req_op(rq)); |
a5efda3c | 2940 | blk_status_t ret; |
06c8c691 CH |
2941 | |
2942 | if (blk_rq_sectors(rq) > max_sectors) { | |
2943 | /* | |
2944 | * SCSI device does not have a good way to return if | |
2945 | * Write Same/Zero is actually supported. If a device rejects | |
2946 | * a non-read/write command (discard, write same,etc.) the | |
2947 | * low-level device driver will set the relevant queue limit to | |
2948 | * 0 to prevent blk-lib from issuing more of the offending | |
2949 | * operations. Commands queued prior to the queue limit being | |
2950 | * reset need to be completed with BLK_STS_NOTSUPP to avoid I/O | |
2951 | * errors being propagated to upper layers. | |
2952 | */ | |
2953 | if (max_sectors == 0) | |
2954 | return BLK_STS_NOTSUPP; | |
2955 | ||
2956 | printk(KERN_ERR "%s: over max size limit. (%u > %u)\n", | |
2957 | __func__, blk_rq_sectors(rq), max_sectors); | |
2958 | return BLK_STS_IOERR; | |
2959 | } | |
2960 | ||
2961 | /* | |
2962 | * The queue settings related to segment counting may differ from the | |
2963 | * original queue. | |
2964 | */ | |
2965 | rq->nr_phys_segments = blk_recalc_rq_segments(rq); | |
2966 | if (rq->nr_phys_segments > queue_max_segments(q)) { | |
2967 | printk(KERN_ERR "%s: over max segments limit. (%hu > %hu)\n", | |
2968 | __func__, rq->nr_phys_segments, queue_max_segments(q)); | |
2969 | return BLK_STS_IOERR; | |
2970 | } | |
2971 | ||
28db4711 | 2972 | if (q->disk && should_fail_request(q->disk->part0, blk_rq_bytes(rq))) |
06c8c691 CH |
2973 | return BLK_STS_IOERR; |
2974 | ||
2975 | if (blk_crypto_insert_cloned_request(rq)) | |
2976 | return BLK_STS_IOERR; | |
2977 | ||
2978 | blk_account_io_start(rq); | |
2979 | ||
2980 | /* | |
2981 | * Since we have a scheduler attached on the top device, | |
2982 | * bypass a potential scheduler on the bottom device for | |
2983 | * insert. | |
2984 | */ | |
28db4711 | 2985 | blk_mq_run_dispatch_ops(q, |
4cafe86c | 2986 | ret = blk_mq_request_issue_directly(rq, true)); |
592ee119 YK |
2987 | if (ret) |
2988 | blk_account_io_done(rq, ktime_get_ns()); | |
4cafe86c | 2989 | return ret; |
06c8c691 CH |
2990 | } |
2991 | EXPORT_SYMBOL_GPL(blk_insert_cloned_request); | |
2992 | ||
2993 | /** | |
2994 | * blk_rq_unprep_clone - Helper function to free all bios in a cloned request | |
2995 | * @rq: the clone request to be cleaned up | |
2996 | * | |
2997 | * Description: | |
2998 | * Free all bios in @rq for a cloned request. | |
2999 | */ | |
3000 | void blk_rq_unprep_clone(struct request *rq) | |
3001 | { | |
3002 | struct bio *bio; | |
3003 | ||
3004 | while ((bio = rq->bio) != NULL) { | |
3005 | rq->bio = bio->bi_next; | |
3006 | ||
3007 | bio_put(bio); | |
3008 | } | |
3009 | } | |
3010 | EXPORT_SYMBOL_GPL(blk_rq_unprep_clone); | |
3011 | ||
3012 | /** | |
3013 | * blk_rq_prep_clone - Helper function to setup clone request | |
3014 | * @rq: the request to be setup | |
3015 | * @rq_src: original request to be cloned | |
3016 | * @bs: bio_set that bios for clone are allocated from | |
3017 | * @gfp_mask: memory allocation mask for bio | |
3018 | * @bio_ctr: setup function to be called for each clone bio. | |
3019 | * Returns %0 for success, non %0 for failure. | |
3020 | * @data: private data to be passed to @bio_ctr | |
3021 | * | |
3022 | * Description: | |
3023 | * Clones bios in @rq_src to @rq, and copies attributes of @rq_src to @rq. | |
3024 | * Also, pages which the original bios are pointing to are not copied | |
3025 | * and the cloned bios just point same pages. | |
3026 | * So cloned bios must be completed before original bios, which means | |
3027 | * the caller must complete @rq before @rq_src. | |
3028 | */ | |
3029 | int blk_rq_prep_clone(struct request *rq, struct request *rq_src, | |
3030 | struct bio_set *bs, gfp_t gfp_mask, | |
3031 | int (*bio_ctr)(struct bio *, struct bio *, void *), | |
3032 | void *data) | |
3033 | { | |
3034 | struct bio *bio, *bio_src; | |
3035 | ||
3036 | if (!bs) | |
3037 | bs = &fs_bio_set; | |
3038 | ||
3039 | __rq_for_each_bio(bio_src, rq_src) { | |
abfc426d CH |
3040 | bio = bio_alloc_clone(rq->q->disk->part0, bio_src, gfp_mask, |
3041 | bs); | |
06c8c691 CH |
3042 | if (!bio) |
3043 | goto free_and_out; | |
3044 | ||
3045 | if (bio_ctr && bio_ctr(bio, bio_src, data)) | |
3046 | goto free_and_out; | |
3047 | ||
3048 | if (rq->bio) { | |
3049 | rq->biotail->bi_next = bio; | |
3050 | rq->biotail = bio; | |
3051 | } else { | |
3052 | rq->bio = rq->biotail = bio; | |
3053 | } | |
3054 | bio = NULL; | |
3055 | } | |
3056 | ||
3057 | /* Copy attributes of the original request to the clone request. */ | |
3058 | rq->__sector = blk_rq_pos(rq_src); | |
3059 | rq->__data_len = blk_rq_bytes(rq_src); | |
3060 | if (rq_src->rq_flags & RQF_SPECIAL_PAYLOAD) { | |
3061 | rq->rq_flags |= RQF_SPECIAL_PAYLOAD; | |
3062 | rq->special_vec = rq_src->special_vec; | |
3063 | } | |
3064 | rq->nr_phys_segments = rq_src->nr_phys_segments; | |
3065 | rq->ioprio = rq_src->ioprio; | |
3066 | ||
3067 | if (rq->bio && blk_crypto_rq_bio_prep(rq, rq->bio, gfp_mask) < 0) | |
3068 | goto free_and_out; | |
3069 | ||
3070 | return 0; | |
3071 | ||
3072 | free_and_out: | |
3073 | if (bio) | |
3074 | bio_put(bio); | |
3075 | blk_rq_unprep_clone(rq); | |
3076 | ||
3077 | return -ENOMEM; | |
3078 | } | |
3079 | EXPORT_SYMBOL_GPL(blk_rq_prep_clone); | |
248c7933 | 3080 | #endif /* CONFIG_BLK_MQ_STACKING */ |
06c8c691 | 3081 | |
f2b8f3ce CH |
3082 | /* |
3083 | * Steal bios from a request and add them to a bio list. | |
3084 | * The request must not have been partially completed before. | |
3085 | */ | |
3086 | void blk_steal_bios(struct bio_list *list, struct request *rq) | |
3087 | { | |
3088 | if (rq->bio) { | |
3089 | if (list->tail) | |
3090 | list->tail->bi_next = rq->bio; | |
3091 | else | |
3092 | list->head = rq->bio; | |
3093 | list->tail = rq->biotail; | |
3094 | ||
3095 | rq->bio = NULL; | |
3096 | rq->biotail = NULL; | |
3097 | } | |
3098 | ||
3099 | rq->__data_len = 0; | |
3100 | } | |
3101 | EXPORT_SYMBOL_GPL(blk_steal_bios); | |
3102 | ||
bd63141d ML |
3103 | static size_t order_to_size(unsigned int order) |
3104 | { | |
3105 | return (size_t)PAGE_SIZE << order; | |
3106 | } | |
3107 | ||
3108 | /* called before freeing request pool in @tags */ | |
f32e4eaf JG |
3109 | static void blk_mq_clear_rq_mapping(struct blk_mq_tags *drv_tags, |
3110 | struct blk_mq_tags *tags) | |
bd63141d | 3111 | { |
bd63141d ML |
3112 | struct page *page; |
3113 | unsigned long flags; | |
3114 | ||
76dd2980 YK |
3115 | /* |
3116 | * There is no need to clear mapping if driver tags is not initialized | |
3117 | * or the mapping belongs to the driver tags. | |
3118 | */ | |
3119 | if (!drv_tags || drv_tags == tags) | |
4f245d5b JG |
3120 | return; |
3121 | ||
bd63141d ML |
3122 | list_for_each_entry(page, &tags->page_list, lru) { |
3123 | unsigned long start = (unsigned long)page_address(page); | |
3124 | unsigned long end = start + order_to_size(page->private); | |
3125 | int i; | |
3126 | ||
f32e4eaf | 3127 | for (i = 0; i < drv_tags->nr_tags; i++) { |
bd63141d ML |
3128 | struct request *rq = drv_tags->rqs[i]; |
3129 | unsigned long rq_addr = (unsigned long)rq; | |
3130 | ||
3131 | if (rq_addr >= start && rq_addr < end) { | |
0a467d0f | 3132 | WARN_ON_ONCE(req_ref_read(rq) != 0); |
bd63141d ML |
3133 | cmpxchg(&drv_tags->rqs[i], rq, NULL); |
3134 | } | |
3135 | } | |
3136 | } | |
3137 | ||
3138 | /* | |
3139 | * Wait until all pending iteration is done. | |
3140 | * | |
3141 | * Request reference is cleared and it is guaranteed to be observed | |
3142 | * after the ->lock is released. | |
3143 | */ | |
3144 | spin_lock_irqsave(&drv_tags->lock, flags); | |
3145 | spin_unlock_irqrestore(&drv_tags->lock, flags); | |
3146 | } | |
3147 | ||
cc71a6f4 JA |
3148 | void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags, |
3149 | unsigned int hctx_idx) | |
95363efd | 3150 | { |
f32e4eaf | 3151 | struct blk_mq_tags *drv_tags; |
e9b267d9 | 3152 | struct page *page; |
320ae51f | 3153 | |
e02657ea ML |
3154 | if (list_empty(&tags->page_list)) |
3155 | return; | |
3156 | ||
079a2e3e JG |
3157 | if (blk_mq_is_shared_tags(set->flags)) |
3158 | drv_tags = set->shared_tags; | |
e155b0c2 JG |
3159 | else |
3160 | drv_tags = set->tags[hctx_idx]; | |
f32e4eaf | 3161 | |
65de57bb | 3162 | if (tags->static_rqs && set->ops->exit_request) { |
e9b267d9 | 3163 | int i; |
320ae51f | 3164 | |
24d2f903 | 3165 | for (i = 0; i < tags->nr_tags; i++) { |
2af8cbe3 JA |
3166 | struct request *rq = tags->static_rqs[i]; |
3167 | ||
3168 | if (!rq) | |
e9b267d9 | 3169 | continue; |
d6296d39 | 3170 | set->ops->exit_request(set, rq, hctx_idx); |
2af8cbe3 | 3171 | tags->static_rqs[i] = NULL; |
e9b267d9 | 3172 | } |
320ae51f | 3173 | } |
320ae51f | 3174 | |
f32e4eaf | 3175 | blk_mq_clear_rq_mapping(drv_tags, tags); |
bd63141d | 3176 | |
24d2f903 CH |
3177 | while (!list_empty(&tags->page_list)) { |
3178 | page = list_first_entry(&tags->page_list, struct page, lru); | |
6753471c | 3179 | list_del_init(&page->lru); |
f75782e4 CM |
3180 | /* |
3181 | * Remove kmemleak object previously allocated in | |
273938bf | 3182 | * blk_mq_alloc_rqs(). |
f75782e4 CM |
3183 | */ |
3184 | kmemleak_free(page_address(page)); | |
320ae51f JA |
3185 | __free_pages(page, page->private); |
3186 | } | |
cc71a6f4 | 3187 | } |
320ae51f | 3188 | |
e155b0c2 | 3189 | void blk_mq_free_rq_map(struct blk_mq_tags *tags) |
cc71a6f4 | 3190 | { |
24d2f903 | 3191 | kfree(tags->rqs); |
cc71a6f4 | 3192 | tags->rqs = NULL; |
2af8cbe3 JA |
3193 | kfree(tags->static_rqs); |
3194 | tags->static_rqs = NULL; | |
320ae51f | 3195 | |
e155b0c2 | 3196 | blk_mq_free_tags(tags); |
320ae51f JA |
3197 | } |
3198 | ||
4d805131 ML |
3199 | static enum hctx_type hctx_idx_to_type(struct blk_mq_tag_set *set, |
3200 | unsigned int hctx_idx) | |
3201 | { | |
3202 | int i; | |
3203 | ||
3204 | for (i = 0; i < set->nr_maps; i++) { | |
3205 | unsigned int start = set->map[i].queue_offset; | |
3206 | unsigned int end = start + set->map[i].nr_queues; | |
3207 | ||
3208 | if (hctx_idx >= start && hctx_idx < end) | |
3209 | break; | |
3210 | } | |
3211 | ||
3212 | if (i >= set->nr_maps) | |
3213 | i = HCTX_TYPE_DEFAULT; | |
3214 | ||
3215 | return i; | |
3216 | } | |
3217 | ||
3218 | static int blk_mq_get_hctx_node(struct blk_mq_tag_set *set, | |
3219 | unsigned int hctx_idx) | |
3220 | { | |
3221 | enum hctx_type type = hctx_idx_to_type(set, hctx_idx); | |
3222 | ||
3223 | return blk_mq_hw_queue_to_node(&set->map[type], hctx_idx); | |
3224 | } | |
3225 | ||
63064be1 JG |
3226 | static struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, |
3227 | unsigned int hctx_idx, | |
3228 | unsigned int nr_tags, | |
e155b0c2 | 3229 | unsigned int reserved_tags) |
320ae51f | 3230 | { |
4d805131 | 3231 | int node = blk_mq_get_hctx_node(set, hctx_idx); |
24d2f903 | 3232 | struct blk_mq_tags *tags; |
320ae51f | 3233 | |
59f082e4 SL |
3234 | if (node == NUMA_NO_NODE) |
3235 | node = set->numa_node; | |
3236 | ||
e155b0c2 JG |
3237 | tags = blk_mq_init_tags(nr_tags, reserved_tags, node, |
3238 | BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags)); | |
24d2f903 CH |
3239 | if (!tags) |
3240 | return NULL; | |
320ae51f | 3241 | |
590b5b7d | 3242 | tags->rqs = kcalloc_node(nr_tags, sizeof(struct request *), |
36e1f3d1 | 3243 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, |
59f082e4 | 3244 | node); |
24d2f903 | 3245 | if (!tags->rqs) { |
e155b0c2 | 3246 | blk_mq_free_tags(tags); |
24d2f903 CH |
3247 | return NULL; |
3248 | } | |
320ae51f | 3249 | |
590b5b7d KC |
3250 | tags->static_rqs = kcalloc_node(nr_tags, sizeof(struct request *), |
3251 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, | |
3252 | node); | |
2af8cbe3 JA |
3253 | if (!tags->static_rqs) { |
3254 | kfree(tags->rqs); | |
e155b0c2 | 3255 | blk_mq_free_tags(tags); |
2af8cbe3 JA |
3256 | return NULL; |
3257 | } | |
3258 | ||
cc71a6f4 JA |
3259 | return tags; |
3260 | } | |
3261 | ||
1d9bd516 TH |
3262 | static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq, |
3263 | unsigned int hctx_idx, int node) | |
3264 | { | |
3265 | int ret; | |
3266 | ||
3267 | if (set->ops->init_request) { | |
3268 | ret = set->ops->init_request(set, rq, hctx_idx, node); | |
3269 | if (ret) | |
3270 | return ret; | |
3271 | } | |
3272 | ||
12f5b931 | 3273 | WRITE_ONCE(rq->state, MQ_RQ_IDLE); |
1d9bd516 TH |
3274 | return 0; |
3275 | } | |
3276 | ||
63064be1 JG |
3277 | static int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, |
3278 | struct blk_mq_tags *tags, | |
3279 | unsigned int hctx_idx, unsigned int depth) | |
cc71a6f4 JA |
3280 | { |
3281 | unsigned int i, j, entries_per_page, max_order = 4; | |
4d805131 | 3282 | int node = blk_mq_get_hctx_node(set, hctx_idx); |
cc71a6f4 | 3283 | size_t rq_size, left; |
59f082e4 | 3284 | |
59f082e4 SL |
3285 | if (node == NUMA_NO_NODE) |
3286 | node = set->numa_node; | |
cc71a6f4 JA |
3287 | |
3288 | INIT_LIST_HEAD(&tags->page_list); | |
3289 | ||
320ae51f JA |
3290 | /* |
3291 | * rq_size is the size of the request plus driver payload, rounded | |
3292 | * to the cacheline size | |
3293 | */ | |
24d2f903 | 3294 | rq_size = round_up(sizeof(struct request) + set->cmd_size, |
320ae51f | 3295 | cache_line_size()); |
cc71a6f4 | 3296 | left = rq_size * depth; |
320ae51f | 3297 | |
cc71a6f4 | 3298 | for (i = 0; i < depth; ) { |
320ae51f JA |
3299 | int this_order = max_order; |
3300 | struct page *page; | |
3301 | int to_do; | |
3302 | void *p; | |
3303 | ||
b3a834b1 | 3304 | while (this_order && left < order_to_size(this_order - 1)) |
320ae51f JA |
3305 | this_order--; |
3306 | ||
3307 | do { | |
59f082e4 | 3308 | page = alloc_pages_node(node, |
36e1f3d1 | 3309 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO, |
a5164405 | 3310 | this_order); |
320ae51f JA |
3311 | if (page) |
3312 | break; | |
3313 | if (!this_order--) | |
3314 | break; | |
3315 | if (order_to_size(this_order) < rq_size) | |
3316 | break; | |
3317 | } while (1); | |
3318 | ||
3319 | if (!page) | |
24d2f903 | 3320 | goto fail; |
320ae51f JA |
3321 | |
3322 | page->private = this_order; | |
24d2f903 | 3323 | list_add_tail(&page->lru, &tags->page_list); |
320ae51f JA |
3324 | |
3325 | p = page_address(page); | |
f75782e4 CM |
3326 | /* |
3327 | * Allow kmemleak to scan these pages as they contain pointers | |
3328 | * to additional allocations like via ops->init_request(). | |
3329 | */ | |
36e1f3d1 | 3330 | kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO); |
320ae51f | 3331 | entries_per_page = order_to_size(this_order) / rq_size; |
cc71a6f4 | 3332 | to_do = min(entries_per_page, depth - i); |
320ae51f JA |
3333 | left -= to_do * rq_size; |
3334 | for (j = 0; j < to_do; j++) { | |
2af8cbe3 JA |
3335 | struct request *rq = p; |
3336 | ||
3337 | tags->static_rqs[i] = rq; | |
1d9bd516 TH |
3338 | if (blk_mq_init_request(set, rq, hctx_idx, node)) { |
3339 | tags->static_rqs[i] = NULL; | |
3340 | goto fail; | |
e9b267d9 CH |
3341 | } |
3342 | ||
320ae51f JA |
3343 | p += rq_size; |
3344 | i++; | |
3345 | } | |
3346 | } | |
cc71a6f4 | 3347 | return 0; |
320ae51f | 3348 | |
24d2f903 | 3349 | fail: |
cc71a6f4 JA |
3350 | blk_mq_free_rqs(set, tags, hctx_idx); |
3351 | return -ENOMEM; | |
320ae51f JA |
3352 | } |
3353 | ||
bf0beec0 ML |
3354 | struct rq_iter_data { |
3355 | struct blk_mq_hw_ctx *hctx; | |
3356 | bool has_rq; | |
3357 | }; | |
3358 | ||
2dd6532e | 3359 | static bool blk_mq_has_request(struct request *rq, void *data) |
bf0beec0 ML |
3360 | { |
3361 | struct rq_iter_data *iter_data = data; | |
3362 | ||
3363 | if (rq->mq_hctx != iter_data->hctx) | |
3364 | return true; | |
3365 | iter_data->has_rq = true; | |
3366 | return false; | |
3367 | } | |
3368 | ||
3369 | static bool blk_mq_hctx_has_requests(struct blk_mq_hw_ctx *hctx) | |
3370 | { | |
3371 | struct blk_mq_tags *tags = hctx->sched_tags ? | |
3372 | hctx->sched_tags : hctx->tags; | |
3373 | struct rq_iter_data data = { | |
3374 | .hctx = hctx, | |
3375 | }; | |
3376 | ||
3377 | blk_mq_all_tag_iter(tags, blk_mq_has_request, &data); | |
3378 | return data.has_rq; | |
3379 | } | |
3380 | ||
3381 | static inline bool blk_mq_last_cpu_in_hctx(unsigned int cpu, | |
3382 | struct blk_mq_hw_ctx *hctx) | |
3383 | { | |
9b51d9d8 | 3384 | if (cpumask_first_and(hctx->cpumask, cpu_online_mask) != cpu) |
bf0beec0 ML |
3385 | return false; |
3386 | if (cpumask_next_and(cpu, hctx->cpumask, cpu_online_mask) < nr_cpu_ids) | |
3387 | return false; | |
3388 | return true; | |
3389 | } | |
3390 | ||
3391 | static int blk_mq_hctx_notify_offline(unsigned int cpu, struct hlist_node *node) | |
3392 | { | |
3393 | struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node, | |
3394 | struct blk_mq_hw_ctx, cpuhp_online); | |
3395 | ||
3396 | if (!cpumask_test_cpu(cpu, hctx->cpumask) || | |
3397 | !blk_mq_last_cpu_in_hctx(cpu, hctx)) | |
3398 | return 0; | |
3399 | ||
3400 | /* | |
3401 | * Prevent new request from being allocated on the current hctx. | |
3402 | * | |
3403 | * The smp_mb__after_atomic() Pairs with the implied barrier in | |
3404 | * test_and_set_bit_lock in sbitmap_get(). Ensures the inactive flag is | |
3405 | * seen once we return from the tag allocator. | |
3406 | */ | |
3407 | set_bit(BLK_MQ_S_INACTIVE, &hctx->state); | |
3408 | smp_mb__after_atomic(); | |
3409 | ||
3410 | /* | |
3411 | * Try to grab a reference to the queue and wait for any outstanding | |
3412 | * requests. If we could not grab a reference the queue has been | |
3413 | * frozen and there are no requests. | |
3414 | */ | |
3415 | if (percpu_ref_tryget(&hctx->queue->q_usage_counter)) { | |
3416 | while (blk_mq_hctx_has_requests(hctx)) | |
3417 | msleep(5); | |
3418 | percpu_ref_put(&hctx->queue->q_usage_counter); | |
3419 | } | |
3420 | ||
3421 | return 0; | |
3422 | } | |
3423 | ||
3424 | static int blk_mq_hctx_notify_online(unsigned int cpu, struct hlist_node *node) | |
3425 | { | |
3426 | struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node, | |
3427 | struct blk_mq_hw_ctx, cpuhp_online); | |
3428 | ||
3429 | if (cpumask_test_cpu(cpu, hctx->cpumask)) | |
3430 | clear_bit(BLK_MQ_S_INACTIVE, &hctx->state); | |
3431 | return 0; | |
3432 | } | |
3433 | ||
e57690fe JA |
3434 | /* |
3435 | * 'cpu' is going away. splice any existing rq_list entries from this | |
3436 | * software queue to the hw queue dispatch list, and ensure that it | |
3437 | * gets run. | |
3438 | */ | |
9467f859 | 3439 | static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node) |
484b4061 | 3440 | { |
9467f859 | 3441 | struct blk_mq_hw_ctx *hctx; |
484b4061 JA |
3442 | struct blk_mq_ctx *ctx; |
3443 | LIST_HEAD(tmp); | |
c16d6b5a | 3444 | enum hctx_type type; |
484b4061 | 3445 | |
9467f859 | 3446 | hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead); |
bf0beec0 ML |
3447 | if (!cpumask_test_cpu(cpu, hctx->cpumask)) |
3448 | return 0; | |
3449 | ||
e57690fe | 3450 | ctx = __blk_mq_get_ctx(hctx->queue, cpu); |
c16d6b5a | 3451 | type = hctx->type; |
484b4061 JA |
3452 | |
3453 | spin_lock(&ctx->lock); | |
c16d6b5a ML |
3454 | if (!list_empty(&ctx->rq_lists[type])) { |
3455 | list_splice_init(&ctx->rq_lists[type], &tmp); | |
484b4061 JA |
3456 | blk_mq_hctx_clear_pending(hctx, ctx); |
3457 | } | |
3458 | spin_unlock(&ctx->lock); | |
3459 | ||
3460 | if (list_empty(&tmp)) | |
9467f859 | 3461 | return 0; |
484b4061 | 3462 | |
e57690fe JA |
3463 | spin_lock(&hctx->lock); |
3464 | list_splice_tail_init(&tmp, &hctx->dispatch); | |
3465 | spin_unlock(&hctx->lock); | |
484b4061 JA |
3466 | |
3467 | blk_mq_run_hw_queue(hctx, true); | |
9467f859 | 3468 | return 0; |
484b4061 JA |
3469 | } |
3470 | ||
9467f859 | 3471 | static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx) |
484b4061 | 3472 | { |
bf0beec0 ML |
3473 | if (!(hctx->flags & BLK_MQ_F_STACKING)) |
3474 | cpuhp_state_remove_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE, | |
3475 | &hctx->cpuhp_online); | |
9467f859 TG |
3476 | cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD, |
3477 | &hctx->cpuhp_dead); | |
484b4061 JA |
3478 | } |
3479 | ||
364b6181 ML |
3480 | /* |
3481 | * Before freeing hw queue, clearing the flush request reference in | |
3482 | * tags->rqs[] for avoiding potential UAF. | |
3483 | */ | |
3484 | static void blk_mq_clear_flush_rq_mapping(struct blk_mq_tags *tags, | |
3485 | unsigned int queue_depth, struct request *flush_rq) | |
3486 | { | |
3487 | int i; | |
3488 | unsigned long flags; | |
3489 | ||
3490 | /* The hw queue may not be mapped yet */ | |
3491 | if (!tags) | |
3492 | return; | |
3493 | ||
0a467d0f | 3494 | WARN_ON_ONCE(req_ref_read(flush_rq) != 0); |
364b6181 ML |
3495 | |
3496 | for (i = 0; i < queue_depth; i++) | |
3497 | cmpxchg(&tags->rqs[i], flush_rq, NULL); | |
3498 | ||
3499 | /* | |
3500 | * Wait until all pending iteration is done. | |
3501 | * | |
3502 | * Request reference is cleared and it is guaranteed to be observed | |
3503 | * after the ->lock is released. | |
3504 | */ | |
3505 | spin_lock_irqsave(&tags->lock, flags); | |
3506 | spin_unlock_irqrestore(&tags->lock, flags); | |
3507 | } | |
3508 | ||
c3b4afca | 3509 | /* hctx->ctxs will be freed in queue's release handler */ |
08e98fc6 ML |
3510 | static void blk_mq_exit_hctx(struct request_queue *q, |
3511 | struct blk_mq_tag_set *set, | |
3512 | struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx) | |
3513 | { | |
364b6181 ML |
3514 | struct request *flush_rq = hctx->fq->flush_rq; |
3515 | ||
8ab0b7dc ML |
3516 | if (blk_mq_hw_queue_mapped(hctx)) |
3517 | blk_mq_tag_idle(hctx); | |
08e98fc6 | 3518 | |
6cfeadbf ML |
3519 | if (blk_queue_init_done(q)) |
3520 | blk_mq_clear_flush_rq_mapping(set->tags[hctx_idx], | |
3521 | set->queue_depth, flush_rq); | |
f70ced09 | 3522 | if (set->ops->exit_request) |
364b6181 | 3523 | set->ops->exit_request(set, flush_rq, hctx_idx); |
f70ced09 | 3524 | |
08e98fc6 ML |
3525 | if (set->ops->exit_hctx) |
3526 | set->ops->exit_hctx(hctx, hctx_idx); | |
3527 | ||
9467f859 | 3528 | blk_mq_remove_cpuhp(hctx); |
2f8f1336 | 3529 | |
4e5cc99e ML |
3530 | xa_erase(&q->hctx_table, hctx_idx); |
3531 | ||
2f8f1336 ML |
3532 | spin_lock(&q->unused_hctx_lock); |
3533 | list_add(&hctx->hctx_list, &q->unused_hctx_list); | |
3534 | spin_unlock(&q->unused_hctx_lock); | |
08e98fc6 ML |
3535 | } |
3536 | ||
624dbe47 ML |
3537 | static void blk_mq_exit_hw_queues(struct request_queue *q, |
3538 | struct blk_mq_tag_set *set, int nr_queue) | |
3539 | { | |
3540 | struct blk_mq_hw_ctx *hctx; | |
4f481208 | 3541 | unsigned long i; |
624dbe47 ML |
3542 | |
3543 | queue_for_each_hw_ctx(q, hctx, i) { | |
3544 | if (i == nr_queue) | |
3545 | break; | |
08e98fc6 | 3546 | blk_mq_exit_hctx(q, set, hctx, i); |
624dbe47 | 3547 | } |
624dbe47 ML |
3548 | } |
3549 | ||
08e98fc6 ML |
3550 | static int blk_mq_init_hctx(struct request_queue *q, |
3551 | struct blk_mq_tag_set *set, | |
3552 | struct blk_mq_hw_ctx *hctx, unsigned hctx_idx) | |
320ae51f | 3553 | { |
7c6c5b7c ML |
3554 | hctx->queue_num = hctx_idx; |
3555 | ||
bf0beec0 ML |
3556 | if (!(hctx->flags & BLK_MQ_F_STACKING)) |
3557 | cpuhp_state_add_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE, | |
3558 | &hctx->cpuhp_online); | |
7c6c5b7c ML |
3559 | cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead); |
3560 | ||
3561 | hctx->tags = set->tags[hctx_idx]; | |
3562 | ||
3563 | if (set->ops->init_hctx && | |
3564 | set->ops->init_hctx(hctx, set->driver_data, hctx_idx)) | |
3565 | goto unregister_cpu_notifier; | |
08e98fc6 | 3566 | |
7c6c5b7c ML |
3567 | if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx, |
3568 | hctx->numa_node)) | |
3569 | goto exit_hctx; | |
4e5cc99e ML |
3570 | |
3571 | if (xa_insert(&q->hctx_table, hctx_idx, hctx, GFP_KERNEL)) | |
3572 | goto exit_flush_rq; | |
3573 | ||
7c6c5b7c ML |
3574 | return 0; |
3575 | ||
4e5cc99e ML |
3576 | exit_flush_rq: |
3577 | if (set->ops->exit_request) | |
3578 | set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx); | |
7c6c5b7c ML |
3579 | exit_hctx: |
3580 | if (set->ops->exit_hctx) | |
3581 | set->ops->exit_hctx(hctx, hctx_idx); | |
3582 | unregister_cpu_notifier: | |
3583 | blk_mq_remove_cpuhp(hctx); | |
3584 | return -1; | |
3585 | } | |
3586 | ||
3587 | static struct blk_mq_hw_ctx * | |
3588 | blk_mq_alloc_hctx(struct request_queue *q, struct blk_mq_tag_set *set, | |
3589 | int node) | |
3590 | { | |
3591 | struct blk_mq_hw_ctx *hctx; | |
3592 | gfp_t gfp = GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY; | |
3593 | ||
704b914f | 3594 | hctx = kzalloc_node(sizeof(struct blk_mq_hw_ctx), gfp, node); |
7c6c5b7c ML |
3595 | if (!hctx) |
3596 | goto fail_alloc_hctx; | |
3597 | ||
3598 | if (!zalloc_cpumask_var_node(&hctx->cpumask, gfp, node)) | |
3599 | goto free_hctx; | |
3600 | ||
3601 | atomic_set(&hctx->nr_active, 0); | |
08e98fc6 | 3602 | if (node == NUMA_NO_NODE) |
7c6c5b7c ML |
3603 | node = set->numa_node; |
3604 | hctx->numa_node = node; | |
08e98fc6 | 3605 | |
9f993737 | 3606 | INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn); |
08e98fc6 ML |
3607 | spin_lock_init(&hctx->lock); |
3608 | INIT_LIST_HEAD(&hctx->dispatch); | |
3609 | hctx->queue = q; | |
51db1c37 | 3610 | hctx->flags = set->flags & ~BLK_MQ_F_TAG_QUEUE_SHARED; |
08e98fc6 | 3611 | |
2f8f1336 ML |
3612 | INIT_LIST_HEAD(&hctx->hctx_list); |
3613 | ||
320ae51f | 3614 | /* |
08e98fc6 ML |
3615 | * Allocate space for all possible cpus to avoid allocation at |
3616 | * runtime | |
320ae51f | 3617 | */ |
d904bfa7 | 3618 | hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *), |
7c6c5b7c | 3619 | gfp, node); |
08e98fc6 | 3620 | if (!hctx->ctxs) |
7c6c5b7c | 3621 | goto free_cpumask; |
320ae51f | 3622 | |
5b202853 | 3623 | if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8), |
c548e62b | 3624 | gfp, node, false, false)) |
08e98fc6 | 3625 | goto free_ctxs; |
08e98fc6 | 3626 | hctx->nr_ctx = 0; |
320ae51f | 3627 | |
5815839b | 3628 | spin_lock_init(&hctx->dispatch_wait_lock); |
eb619fdb JA |
3629 | init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake); |
3630 | INIT_LIST_HEAD(&hctx->dispatch_wait.entry); | |
3631 | ||
754a1572 | 3632 | hctx->fq = blk_alloc_flush_queue(hctx->numa_node, set->cmd_size, gfp); |
f70ced09 | 3633 | if (!hctx->fq) |
7c6c5b7c | 3634 | goto free_bitmap; |
320ae51f | 3635 | |
7c6c5b7c | 3636 | blk_mq_hctx_kobj_init(hctx); |
6a83e74d | 3637 | |
7c6c5b7c | 3638 | return hctx; |
320ae51f | 3639 | |
08e98fc6 | 3640 | free_bitmap: |
88459642 | 3641 | sbitmap_free(&hctx->ctx_map); |
08e98fc6 ML |
3642 | free_ctxs: |
3643 | kfree(hctx->ctxs); | |
7c6c5b7c ML |
3644 | free_cpumask: |
3645 | free_cpumask_var(hctx->cpumask); | |
3646 | free_hctx: | |
3647 | kfree(hctx); | |
3648 | fail_alloc_hctx: | |
3649 | return NULL; | |
08e98fc6 | 3650 | } |
320ae51f | 3651 | |
320ae51f JA |
3652 | static void blk_mq_init_cpu_queues(struct request_queue *q, |
3653 | unsigned int nr_hw_queues) | |
3654 | { | |
b3c661b1 JA |
3655 | struct blk_mq_tag_set *set = q->tag_set; |
3656 | unsigned int i, j; | |
320ae51f JA |
3657 | |
3658 | for_each_possible_cpu(i) { | |
3659 | struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i); | |
3660 | struct blk_mq_hw_ctx *hctx; | |
c16d6b5a | 3661 | int k; |
320ae51f | 3662 | |
320ae51f JA |
3663 | __ctx->cpu = i; |
3664 | spin_lock_init(&__ctx->lock); | |
c16d6b5a ML |
3665 | for (k = HCTX_TYPE_DEFAULT; k < HCTX_MAX_TYPES; k++) |
3666 | INIT_LIST_HEAD(&__ctx->rq_lists[k]); | |
3667 | ||
320ae51f JA |
3668 | __ctx->queue = q; |
3669 | ||
320ae51f JA |
3670 | /* |
3671 | * Set local node, IFF we have more than one hw queue. If | |
3672 | * not, we remain on the home node of the device | |
3673 | */ | |
b3c661b1 JA |
3674 | for (j = 0; j < set->nr_maps; j++) { |
3675 | hctx = blk_mq_map_queue_type(q, j, i); | |
3676 | if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE) | |
576e85c5 | 3677 | hctx->numa_node = cpu_to_node(i); |
b3c661b1 | 3678 | } |
320ae51f JA |
3679 | } |
3680 | } | |
3681 | ||
63064be1 JG |
3682 | struct blk_mq_tags *blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set *set, |
3683 | unsigned int hctx_idx, | |
3684 | unsigned int depth) | |
cc71a6f4 | 3685 | { |
63064be1 JG |
3686 | struct blk_mq_tags *tags; |
3687 | int ret; | |
cc71a6f4 | 3688 | |
e155b0c2 | 3689 | tags = blk_mq_alloc_rq_map(set, hctx_idx, depth, set->reserved_tags); |
63064be1 JG |
3690 | if (!tags) |
3691 | return NULL; | |
cc71a6f4 | 3692 | |
63064be1 JG |
3693 | ret = blk_mq_alloc_rqs(set, tags, hctx_idx, depth); |
3694 | if (ret) { | |
e155b0c2 | 3695 | blk_mq_free_rq_map(tags); |
63064be1 JG |
3696 | return NULL; |
3697 | } | |
cc71a6f4 | 3698 | |
63064be1 | 3699 | return tags; |
cc71a6f4 JA |
3700 | } |
3701 | ||
63064be1 JG |
3702 | static bool __blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set *set, |
3703 | int hctx_idx) | |
cc71a6f4 | 3704 | { |
079a2e3e JG |
3705 | if (blk_mq_is_shared_tags(set->flags)) { |
3706 | set->tags[hctx_idx] = set->shared_tags; | |
1c0706a7 | 3707 | |
e155b0c2 | 3708 | return true; |
bd166ef1 | 3709 | } |
e155b0c2 | 3710 | |
63064be1 JG |
3711 | set->tags[hctx_idx] = blk_mq_alloc_map_and_rqs(set, hctx_idx, |
3712 | set->queue_depth); | |
3713 | ||
3714 | return set->tags[hctx_idx]; | |
cc71a6f4 JA |
3715 | } |
3716 | ||
645db34e JG |
3717 | void blk_mq_free_map_and_rqs(struct blk_mq_tag_set *set, |
3718 | struct blk_mq_tags *tags, | |
3719 | unsigned int hctx_idx) | |
cc71a6f4 | 3720 | { |
645db34e JG |
3721 | if (tags) { |
3722 | blk_mq_free_rqs(set, tags, hctx_idx); | |
e155b0c2 | 3723 | blk_mq_free_rq_map(tags); |
bd166ef1 | 3724 | } |
cc71a6f4 JA |
3725 | } |
3726 | ||
e155b0c2 JG |
3727 | static void __blk_mq_free_map_and_rqs(struct blk_mq_tag_set *set, |
3728 | unsigned int hctx_idx) | |
3729 | { | |
079a2e3e | 3730 | if (!blk_mq_is_shared_tags(set->flags)) |
e155b0c2 JG |
3731 | blk_mq_free_map_and_rqs(set, set->tags[hctx_idx], hctx_idx); |
3732 | ||
3733 | set->tags[hctx_idx] = NULL; | |
cc71a6f4 JA |
3734 | } |
3735 | ||
4b855ad3 | 3736 | static void blk_mq_map_swqueue(struct request_queue *q) |
320ae51f | 3737 | { |
4f481208 ML |
3738 | unsigned int j, hctx_idx; |
3739 | unsigned long i; | |
320ae51f JA |
3740 | struct blk_mq_hw_ctx *hctx; |
3741 | struct blk_mq_ctx *ctx; | |
2a34c087 | 3742 | struct blk_mq_tag_set *set = q->tag_set; |
320ae51f JA |
3743 | |
3744 | queue_for_each_hw_ctx(q, hctx, i) { | |
e4043dcf | 3745 | cpumask_clear(hctx->cpumask); |
320ae51f | 3746 | hctx->nr_ctx = 0; |
d416c92c | 3747 | hctx->dispatch_from = NULL; |
320ae51f JA |
3748 | } |
3749 | ||
3750 | /* | |
4b855ad3 | 3751 | * Map software to hardware queues. |
4412efec ML |
3752 | * |
3753 | * If the cpu isn't present, the cpu is mapped to first hctx. | |
320ae51f | 3754 | */ |
20e4d813 | 3755 | for_each_possible_cpu(i) { |
4412efec | 3756 | |
897bb0c7 | 3757 | ctx = per_cpu_ptr(q->queue_ctx, i); |
b3c661b1 | 3758 | for (j = 0; j < set->nr_maps; j++) { |
bb94aea1 JW |
3759 | if (!set->map[j].nr_queues) { |
3760 | ctx->hctxs[j] = blk_mq_map_queue_type(q, | |
3761 | HCTX_TYPE_DEFAULT, i); | |
e5edd5f2 | 3762 | continue; |
bb94aea1 | 3763 | } |
fd689871 ML |
3764 | hctx_idx = set->map[j].mq_map[i]; |
3765 | /* unmapped hw queue can be remapped after CPU topo changed */ | |
3766 | if (!set->tags[hctx_idx] && | |
63064be1 | 3767 | !__blk_mq_alloc_map_and_rqs(set, hctx_idx)) { |
fd689871 ML |
3768 | /* |
3769 | * If tags initialization fail for some hctx, | |
3770 | * that hctx won't be brought online. In this | |
3771 | * case, remap the current ctx to hctx[0] which | |
3772 | * is guaranteed to always have tags allocated | |
3773 | */ | |
3774 | set->map[j].mq_map[i] = 0; | |
3775 | } | |
e5edd5f2 | 3776 | |
b3c661b1 | 3777 | hctx = blk_mq_map_queue_type(q, j, i); |
8ccdf4a3 | 3778 | ctx->hctxs[j] = hctx; |
b3c661b1 JA |
3779 | /* |
3780 | * If the CPU is already set in the mask, then we've | |
3781 | * mapped this one already. This can happen if | |
3782 | * devices share queues across queue maps. | |
3783 | */ | |
3784 | if (cpumask_test_cpu(i, hctx->cpumask)) | |
3785 | continue; | |
3786 | ||
3787 | cpumask_set_cpu(i, hctx->cpumask); | |
3788 | hctx->type = j; | |
3789 | ctx->index_hw[hctx->type] = hctx->nr_ctx; | |
3790 | hctx->ctxs[hctx->nr_ctx++] = ctx; | |
3791 | ||
3792 | /* | |
3793 | * If the nr_ctx type overflows, we have exceeded the | |
3794 | * amount of sw queues we can support. | |
3795 | */ | |
3796 | BUG_ON(!hctx->nr_ctx); | |
3797 | } | |
bb94aea1 JW |
3798 | |
3799 | for (; j < HCTX_MAX_TYPES; j++) | |
3800 | ctx->hctxs[j] = blk_mq_map_queue_type(q, | |
3801 | HCTX_TYPE_DEFAULT, i); | |
320ae51f | 3802 | } |
506e931f JA |
3803 | |
3804 | queue_for_each_hw_ctx(q, hctx, i) { | |
4412efec ML |
3805 | /* |
3806 | * If no software queues are mapped to this hardware queue, | |
3807 | * disable it and free the request entries. | |
3808 | */ | |
3809 | if (!hctx->nr_ctx) { | |
3810 | /* Never unmap queue 0. We need it as a | |
3811 | * fallback in case of a new remap fails | |
3812 | * allocation | |
3813 | */ | |
e155b0c2 JG |
3814 | if (i) |
3815 | __blk_mq_free_map_and_rqs(set, i); | |
4412efec ML |
3816 | |
3817 | hctx->tags = NULL; | |
3818 | continue; | |
3819 | } | |
484b4061 | 3820 | |
2a34c087 ML |
3821 | hctx->tags = set->tags[i]; |
3822 | WARN_ON(!hctx->tags); | |
3823 | ||
889fa31f CY |
3824 | /* |
3825 | * Set the map size to the number of mapped software queues. | |
3826 | * This is more accurate and more efficient than looping | |
3827 | * over all possibly mapped software queues. | |
3828 | */ | |
88459642 | 3829 | sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx); |
889fa31f | 3830 | |
484b4061 JA |
3831 | /* |
3832 | * Initialize batch roundrobin counts | |
3833 | */ | |
f82ddf19 | 3834 | hctx->next_cpu = blk_mq_first_mapped_cpu(hctx); |
506e931f JA |
3835 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; |
3836 | } | |
320ae51f JA |
3837 | } |
3838 | ||
8e8320c9 JA |
3839 | /* |
3840 | * Caller needs to ensure that we're either frozen/quiesced, or that | |
3841 | * the queue isn't live yet. | |
3842 | */ | |
2404e607 | 3843 | static void queue_set_hctx_shared(struct request_queue *q, bool shared) |
0d2602ca JA |
3844 | { |
3845 | struct blk_mq_hw_ctx *hctx; | |
4f481208 | 3846 | unsigned long i; |
0d2602ca | 3847 | |
2404e607 | 3848 | queue_for_each_hw_ctx(q, hctx, i) { |
454bb677 | 3849 | if (shared) { |
51db1c37 | 3850 | hctx->flags |= BLK_MQ_F_TAG_QUEUE_SHARED; |
454bb677 YK |
3851 | } else { |
3852 | blk_mq_tag_idle(hctx); | |
51db1c37 | 3853 | hctx->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED; |
454bb677 | 3854 | } |
2404e607 JM |
3855 | } |
3856 | } | |
3857 | ||
655ac300 HR |
3858 | static void blk_mq_update_tag_set_shared(struct blk_mq_tag_set *set, |
3859 | bool shared) | |
2404e607 JM |
3860 | { |
3861 | struct request_queue *q; | |
0d2602ca | 3862 | |
705cda97 BVA |
3863 | lockdep_assert_held(&set->tag_list_lock); |
3864 | ||
0d2602ca JA |
3865 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
3866 | blk_mq_freeze_queue(q); | |
2404e607 | 3867 | queue_set_hctx_shared(q, shared); |
0d2602ca JA |
3868 | blk_mq_unfreeze_queue(q); |
3869 | } | |
3870 | } | |
3871 | ||
3872 | static void blk_mq_del_queue_tag_set(struct request_queue *q) | |
3873 | { | |
3874 | struct blk_mq_tag_set *set = q->tag_set; | |
3875 | ||
0d2602ca | 3876 | mutex_lock(&set->tag_list_lock); |
08c875cb | 3877 | list_del(&q->tag_set_list); |
2404e607 JM |
3878 | if (list_is_singular(&set->tag_list)) { |
3879 | /* just transitioned to unshared */ | |
51db1c37 | 3880 | set->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED; |
2404e607 | 3881 | /* update existing queue */ |
655ac300 | 3882 | blk_mq_update_tag_set_shared(set, false); |
2404e607 | 3883 | } |
0d2602ca | 3884 | mutex_unlock(&set->tag_list_lock); |
a347c7ad | 3885 | INIT_LIST_HEAD(&q->tag_set_list); |
0d2602ca JA |
3886 | } |
3887 | ||
3888 | static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set, | |
3889 | struct request_queue *q) | |
3890 | { | |
0d2602ca | 3891 | mutex_lock(&set->tag_list_lock); |
2404e607 | 3892 | |
ff821d27 JA |
3893 | /* |
3894 | * Check to see if we're transitioning to shared (from 1 to 2 queues). | |
3895 | */ | |
3896 | if (!list_empty(&set->tag_list) && | |
51db1c37 ML |
3897 | !(set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) { |
3898 | set->flags |= BLK_MQ_F_TAG_QUEUE_SHARED; | |
2404e607 | 3899 | /* update existing queue */ |
655ac300 | 3900 | blk_mq_update_tag_set_shared(set, true); |
2404e607 | 3901 | } |
51db1c37 | 3902 | if (set->flags & BLK_MQ_F_TAG_QUEUE_SHARED) |
2404e607 | 3903 | queue_set_hctx_shared(q, true); |
08c875cb | 3904 | list_add_tail(&q->tag_set_list, &set->tag_list); |
2404e607 | 3905 | |
0d2602ca JA |
3906 | mutex_unlock(&set->tag_list_lock); |
3907 | } | |
3908 | ||
1db4909e ML |
3909 | /* All allocations will be freed in release handler of q->mq_kobj */ |
3910 | static int blk_mq_alloc_ctxs(struct request_queue *q) | |
3911 | { | |
3912 | struct blk_mq_ctxs *ctxs; | |
3913 | int cpu; | |
3914 | ||
3915 | ctxs = kzalloc(sizeof(*ctxs), GFP_KERNEL); | |
3916 | if (!ctxs) | |
3917 | return -ENOMEM; | |
3918 | ||
3919 | ctxs->queue_ctx = alloc_percpu(struct blk_mq_ctx); | |
3920 | if (!ctxs->queue_ctx) | |
3921 | goto fail; | |
3922 | ||
3923 | for_each_possible_cpu(cpu) { | |
3924 | struct blk_mq_ctx *ctx = per_cpu_ptr(ctxs->queue_ctx, cpu); | |
3925 | ctx->ctxs = ctxs; | |
3926 | } | |
3927 | ||
3928 | q->mq_kobj = &ctxs->kobj; | |
3929 | q->queue_ctx = ctxs->queue_ctx; | |
3930 | ||
3931 | return 0; | |
3932 | fail: | |
3933 | kfree(ctxs); | |
3934 | return -ENOMEM; | |
3935 | } | |
3936 | ||
e09aae7e ML |
3937 | /* |
3938 | * It is the actual release handler for mq, but we do it from | |
3939 | * request queue's release handler for avoiding use-after-free | |
3940 | * and headache because q->mq_kobj shouldn't have been introduced, | |
3941 | * but we can't group ctx/kctx kobj without it. | |
3942 | */ | |
3943 | void blk_mq_release(struct request_queue *q) | |
3944 | { | |
2f8f1336 | 3945 | struct blk_mq_hw_ctx *hctx, *next; |
4f481208 | 3946 | unsigned long i; |
e09aae7e | 3947 | |
2f8f1336 ML |
3948 | queue_for_each_hw_ctx(q, hctx, i) |
3949 | WARN_ON_ONCE(hctx && list_empty(&hctx->hctx_list)); | |
3950 | ||
3951 | /* all hctx are in .unused_hctx_list now */ | |
3952 | list_for_each_entry_safe(hctx, next, &q->unused_hctx_list, hctx_list) { | |
3953 | list_del_init(&hctx->hctx_list); | |
6c8b232e | 3954 | kobject_put(&hctx->kobj); |
c3b4afca | 3955 | } |
e09aae7e | 3956 | |
4e5cc99e | 3957 | xa_destroy(&q->hctx_table); |
e09aae7e | 3958 | |
7ea5fe31 ML |
3959 | /* |
3960 | * release .mq_kobj and sw queue's kobject now because | |
3961 | * both share lifetime with request queue. | |
3962 | */ | |
3963 | blk_mq_sysfs_deinit(q); | |
e09aae7e ML |
3964 | } |
3965 | ||
5ec780a6 | 3966 | static struct request_queue *blk_mq_init_queue_data(struct blk_mq_tag_set *set, |
2f227bb9 | 3967 | void *queuedata) |
b62c21b7 | 3968 | { |
26a9750a CH |
3969 | struct request_queue *q; |
3970 | int ret; | |
b62c21b7 | 3971 | |
704b914f | 3972 | q = blk_alloc_queue(set->numa_node, set->flags & BLK_MQ_F_BLOCKING); |
26a9750a | 3973 | if (!q) |
b62c21b7 | 3974 | return ERR_PTR(-ENOMEM); |
26a9750a CH |
3975 | q->queuedata = queuedata; |
3976 | ret = blk_mq_init_allocated_queue(set, q); | |
3977 | if (ret) { | |
6f8191fd | 3978 | blk_put_queue(q); |
26a9750a CH |
3979 | return ERR_PTR(ret); |
3980 | } | |
b62c21b7 MS |
3981 | return q; |
3982 | } | |
2f227bb9 CH |
3983 | |
3984 | struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set) | |
3985 | { | |
3986 | return blk_mq_init_queue_data(set, NULL); | |
3987 | } | |
b62c21b7 MS |
3988 | EXPORT_SYMBOL(blk_mq_init_queue); |
3989 | ||
6f8191fd CH |
3990 | /** |
3991 | * blk_mq_destroy_queue - shutdown a request queue | |
3992 | * @q: request queue to shutdown | |
3993 | * | |
3994 | * This shuts down a request queue allocated by blk_mq_init_queue() and drops | |
3995 | * the initial reference. All future requests will failed with -ENODEV. | |
3996 | * | |
3997 | * Context: can sleep | |
3998 | */ | |
3999 | void blk_mq_destroy_queue(struct request_queue *q) | |
4000 | { | |
4001 | WARN_ON_ONCE(!queue_is_mq(q)); | |
4002 | WARN_ON_ONCE(blk_queue_registered(q)); | |
4003 | ||
4004 | might_sleep(); | |
4005 | ||
4006 | blk_queue_flag_set(QUEUE_FLAG_DYING, q); | |
4007 | blk_queue_start_drain(q); | |
4008 | blk_freeze_queue(q); | |
4009 | ||
4010 | blk_sync_queue(q); | |
4011 | blk_mq_cancel_work_sync(q); | |
4012 | blk_mq_exit_queue(q); | |
4013 | ||
4014 | /* @q is and will stay empty, shutdown and put */ | |
4015 | blk_put_queue(q); | |
4016 | } | |
4017 | EXPORT_SYMBOL(blk_mq_destroy_queue); | |
4018 | ||
4dcc4874 CH |
4019 | struct gendisk *__blk_mq_alloc_disk(struct blk_mq_tag_set *set, void *queuedata, |
4020 | struct lock_class_key *lkclass) | |
9316a9ed JA |
4021 | { |
4022 | struct request_queue *q; | |
b461dfc4 | 4023 | struct gendisk *disk; |
9316a9ed | 4024 | |
b461dfc4 CH |
4025 | q = blk_mq_init_queue_data(set, queuedata); |
4026 | if (IS_ERR(q)) | |
4027 | return ERR_CAST(q); | |
9316a9ed | 4028 | |
4a1fa41d | 4029 | disk = __alloc_disk_node(q, set->numa_node, lkclass); |
b461dfc4 | 4030 | if (!disk) { |
0a3e5cc7 | 4031 | blk_mq_destroy_queue(q); |
b461dfc4 | 4032 | return ERR_PTR(-ENOMEM); |
9316a9ed | 4033 | } |
6f8191fd | 4034 | set_bit(GD_OWNS_QUEUE, &disk->state); |
b461dfc4 | 4035 | return disk; |
9316a9ed | 4036 | } |
b461dfc4 | 4037 | EXPORT_SYMBOL(__blk_mq_alloc_disk); |
9316a9ed | 4038 | |
6f8191fd CH |
4039 | struct gendisk *blk_mq_alloc_disk_for_queue(struct request_queue *q, |
4040 | struct lock_class_key *lkclass) | |
4041 | { | |
4042 | if (!blk_get_queue(q)) | |
4043 | return NULL; | |
4044 | return __alloc_disk_node(q, NUMA_NO_NODE, lkclass); | |
4045 | } | |
4046 | EXPORT_SYMBOL(blk_mq_alloc_disk_for_queue); | |
4047 | ||
34d11ffa JW |
4048 | static struct blk_mq_hw_ctx *blk_mq_alloc_and_init_hctx( |
4049 | struct blk_mq_tag_set *set, struct request_queue *q, | |
4050 | int hctx_idx, int node) | |
4051 | { | |
2f8f1336 | 4052 | struct blk_mq_hw_ctx *hctx = NULL, *tmp; |
34d11ffa | 4053 | |
2f8f1336 ML |
4054 | /* reuse dead hctx first */ |
4055 | spin_lock(&q->unused_hctx_lock); | |
4056 | list_for_each_entry(tmp, &q->unused_hctx_list, hctx_list) { | |
4057 | if (tmp->numa_node == node) { | |
4058 | hctx = tmp; | |
4059 | break; | |
4060 | } | |
4061 | } | |
4062 | if (hctx) | |
4063 | list_del_init(&hctx->hctx_list); | |
4064 | spin_unlock(&q->unused_hctx_lock); | |
4065 | ||
4066 | if (!hctx) | |
4067 | hctx = blk_mq_alloc_hctx(q, set, node); | |
34d11ffa | 4068 | if (!hctx) |
7c6c5b7c | 4069 | goto fail; |
34d11ffa | 4070 | |
7c6c5b7c ML |
4071 | if (blk_mq_init_hctx(q, set, hctx, hctx_idx)) |
4072 | goto free_hctx; | |
34d11ffa JW |
4073 | |
4074 | return hctx; | |
7c6c5b7c ML |
4075 | |
4076 | free_hctx: | |
4077 | kobject_put(&hctx->kobj); | |
4078 | fail: | |
4079 | return NULL; | |
34d11ffa JW |
4080 | } |
4081 | ||
868f2f0b KB |
4082 | static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set, |
4083 | struct request_queue *q) | |
320ae51f | 4084 | { |
4e5cc99e ML |
4085 | struct blk_mq_hw_ctx *hctx; |
4086 | unsigned long i, j; | |
ac0d6b92 | 4087 | |
fb350e0a ML |
4088 | /* protect against switching io scheduler */ |
4089 | mutex_lock(&q->sysfs_lock); | |
24d2f903 | 4090 | for (i = 0; i < set->nr_hw_queues; i++) { |
306f13ee | 4091 | int old_node; |
4d805131 | 4092 | int node = blk_mq_get_hctx_node(set, i); |
4e5cc99e | 4093 | struct blk_mq_hw_ctx *old_hctx = xa_load(&q->hctx_table, i); |
868f2f0b | 4094 | |
306f13ee ML |
4095 | if (old_hctx) { |
4096 | old_node = old_hctx->numa_node; | |
4097 | blk_mq_exit_hctx(q, set, old_hctx, i); | |
4098 | } | |
868f2f0b | 4099 | |
4e5cc99e | 4100 | if (!blk_mq_alloc_and_init_hctx(set, q, i, node)) { |
306f13ee | 4101 | if (!old_hctx) |
34d11ffa | 4102 | break; |
306f13ee ML |
4103 | pr_warn("Allocate new hctx on node %d fails, fallback to previous one on node %d\n", |
4104 | node, old_node); | |
4e5cc99e ML |
4105 | hctx = blk_mq_alloc_and_init_hctx(set, q, i, old_node); |
4106 | WARN_ON_ONCE(!hctx); | |
868f2f0b | 4107 | } |
320ae51f | 4108 | } |
e01ad46d JW |
4109 | /* |
4110 | * Increasing nr_hw_queues fails. Free the newly allocated | |
4111 | * hctxs and keep the previous q->nr_hw_queues. | |
4112 | */ | |
4113 | if (i != set->nr_hw_queues) { | |
4114 | j = q->nr_hw_queues; | |
e01ad46d JW |
4115 | } else { |
4116 | j = i; | |
e01ad46d JW |
4117 | q->nr_hw_queues = set->nr_hw_queues; |
4118 | } | |
34d11ffa | 4119 | |
4e5cc99e ML |
4120 | xa_for_each_start(&q->hctx_table, j, hctx, j) |
4121 | blk_mq_exit_hctx(q, set, hctx, j); | |
fb350e0a | 4122 | mutex_unlock(&q->sysfs_lock); |
868f2f0b KB |
4123 | } |
4124 | ||
42ee3061 ML |
4125 | static void blk_mq_update_poll_flag(struct request_queue *q) |
4126 | { | |
4127 | struct blk_mq_tag_set *set = q->tag_set; | |
4128 | ||
4129 | if (set->nr_maps > HCTX_TYPE_POLL && | |
4130 | set->map[HCTX_TYPE_POLL].nr_queues) | |
4131 | blk_queue_flag_set(QUEUE_FLAG_POLL, q); | |
4132 | else | |
4133 | blk_queue_flag_clear(QUEUE_FLAG_POLL, q); | |
4134 | } | |
4135 | ||
26a9750a CH |
4136 | int blk_mq_init_allocated_queue(struct blk_mq_tag_set *set, |
4137 | struct request_queue *q) | |
868f2f0b | 4138 | { |
704b914f ML |
4139 | WARN_ON_ONCE(blk_queue_has_srcu(q) != |
4140 | !!(set->flags & BLK_MQ_F_BLOCKING)); | |
4141 | ||
66841672 ML |
4142 | /* mark the queue as mq asap */ |
4143 | q->mq_ops = set->ops; | |
4144 | ||
34dbad5d | 4145 | q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn, |
720b8ccc SB |
4146 | blk_mq_poll_stats_bkt, |
4147 | BLK_MQ_POLL_STATS_BKTS, q); | |
34dbad5d OS |
4148 | if (!q->poll_cb) |
4149 | goto err_exit; | |
4150 | ||
1db4909e | 4151 | if (blk_mq_alloc_ctxs(q)) |
41de54c6 | 4152 | goto err_poll; |
868f2f0b | 4153 | |
737f98cf ML |
4154 | /* init q->mq_kobj and sw queues' kobjects */ |
4155 | blk_mq_sysfs_init(q); | |
4156 | ||
2f8f1336 ML |
4157 | INIT_LIST_HEAD(&q->unused_hctx_list); |
4158 | spin_lock_init(&q->unused_hctx_lock); | |
4159 | ||
4e5cc99e ML |
4160 | xa_init(&q->hctx_table); |
4161 | ||
868f2f0b KB |
4162 | blk_mq_realloc_hw_ctxs(set, q); |
4163 | if (!q->nr_hw_queues) | |
4164 | goto err_hctxs; | |
320ae51f | 4165 | |
287922eb | 4166 | INIT_WORK(&q->timeout_work, blk_mq_timeout_work); |
e56f698b | 4167 | blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ); |
320ae51f | 4168 | |
a8908939 | 4169 | q->tag_set = set; |
320ae51f | 4170 | |
94eddfbe | 4171 | q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT; |
42ee3061 | 4172 | blk_mq_update_poll_flag(q); |
320ae51f | 4173 | |
2849450a | 4174 | INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work); |
6fca6a61 CH |
4175 | INIT_LIST_HEAD(&q->requeue_list); |
4176 | spin_lock_init(&q->requeue_lock); | |
4177 | ||
eba71768 JA |
4178 | q->nr_requests = set->queue_depth; |
4179 | ||
64f1c21e JA |
4180 | /* |
4181 | * Default to classic polling | |
4182 | */ | |
29ece8b4 | 4183 | q->poll_nsec = BLK_MQ_POLL_CLASSIC; |
64f1c21e | 4184 | |
24d2f903 | 4185 | blk_mq_init_cpu_queues(q, set->nr_hw_queues); |
0d2602ca | 4186 | blk_mq_add_queue_tag_set(set, q); |
4b855ad3 | 4187 | blk_mq_map_swqueue(q); |
26a9750a | 4188 | return 0; |
18741986 | 4189 | |
320ae51f | 4190 | err_hctxs: |
4e5cc99e | 4191 | xa_destroy(&q->hctx_table); |
73d9c8d4 | 4192 | q->nr_hw_queues = 0; |
1db4909e | 4193 | blk_mq_sysfs_deinit(q); |
41de54c6 JS |
4194 | err_poll: |
4195 | blk_stat_free_callback(q->poll_cb); | |
4196 | q->poll_cb = NULL; | |
c7de5726 ML |
4197 | err_exit: |
4198 | q->mq_ops = NULL; | |
26a9750a | 4199 | return -ENOMEM; |
320ae51f | 4200 | } |
b62c21b7 | 4201 | EXPORT_SYMBOL(blk_mq_init_allocated_queue); |
320ae51f | 4202 | |
c7e2d94b ML |
4203 | /* tags can _not_ be used after returning from blk_mq_exit_queue */ |
4204 | void blk_mq_exit_queue(struct request_queue *q) | |
320ae51f | 4205 | { |
630ef623 | 4206 | struct blk_mq_tag_set *set = q->tag_set; |
320ae51f | 4207 | |
630ef623 | 4208 | /* Checks hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED. */ |
624dbe47 | 4209 | blk_mq_exit_hw_queues(q, set, set->nr_hw_queues); |
630ef623 BVA |
4210 | /* May clear BLK_MQ_F_TAG_QUEUE_SHARED in hctx->flags. */ |
4211 | blk_mq_del_queue_tag_set(q); | |
320ae51f | 4212 | } |
320ae51f | 4213 | |
a5164405 JA |
4214 | static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set) |
4215 | { | |
4216 | int i; | |
4217 | ||
079a2e3e JG |
4218 | if (blk_mq_is_shared_tags(set->flags)) { |
4219 | set->shared_tags = blk_mq_alloc_map_and_rqs(set, | |
e155b0c2 JG |
4220 | BLK_MQ_NO_HCTX_IDX, |
4221 | set->queue_depth); | |
079a2e3e | 4222 | if (!set->shared_tags) |
e155b0c2 JG |
4223 | return -ENOMEM; |
4224 | } | |
4225 | ||
8229cca8 | 4226 | for (i = 0; i < set->nr_hw_queues; i++) { |
63064be1 | 4227 | if (!__blk_mq_alloc_map_and_rqs(set, i)) |
a5164405 | 4228 | goto out_unwind; |
8229cca8 XT |
4229 | cond_resched(); |
4230 | } | |
a5164405 JA |
4231 | |
4232 | return 0; | |
4233 | ||
4234 | out_unwind: | |
4235 | while (--i >= 0) | |
e155b0c2 JG |
4236 | __blk_mq_free_map_and_rqs(set, i); |
4237 | ||
079a2e3e JG |
4238 | if (blk_mq_is_shared_tags(set->flags)) { |
4239 | blk_mq_free_map_and_rqs(set, set->shared_tags, | |
e155b0c2 | 4240 | BLK_MQ_NO_HCTX_IDX); |
645db34e | 4241 | } |
a5164405 | 4242 | |
a5164405 JA |
4243 | return -ENOMEM; |
4244 | } | |
4245 | ||
4246 | /* | |
4247 | * Allocate the request maps associated with this tag_set. Note that this | |
4248 | * may reduce the depth asked for, if memory is tight. set->queue_depth | |
4249 | * will be updated to reflect the allocated depth. | |
4250 | */ | |
63064be1 | 4251 | static int blk_mq_alloc_set_map_and_rqs(struct blk_mq_tag_set *set) |
a5164405 JA |
4252 | { |
4253 | unsigned int depth; | |
4254 | int err; | |
4255 | ||
4256 | depth = set->queue_depth; | |
4257 | do { | |
4258 | err = __blk_mq_alloc_rq_maps(set); | |
4259 | if (!err) | |
4260 | break; | |
4261 | ||
4262 | set->queue_depth >>= 1; | |
4263 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) { | |
4264 | err = -ENOMEM; | |
4265 | break; | |
4266 | } | |
4267 | } while (set->queue_depth); | |
4268 | ||
4269 | if (!set->queue_depth || err) { | |
4270 | pr_err("blk-mq: failed to allocate request map\n"); | |
4271 | return -ENOMEM; | |
4272 | } | |
4273 | ||
4274 | if (depth != set->queue_depth) | |
4275 | pr_info("blk-mq: reduced tag depth (%u -> %u)\n", | |
4276 | depth, set->queue_depth); | |
4277 | ||
4278 | return 0; | |
4279 | } | |
4280 | ||
a4e1d0b7 | 4281 | static void blk_mq_update_queue_map(struct blk_mq_tag_set *set) |
ebe8bddb | 4282 | { |
6e66b493 BVA |
4283 | /* |
4284 | * blk_mq_map_queues() and multiple .map_queues() implementations | |
4285 | * expect that set->map[HCTX_TYPE_DEFAULT].nr_queues is set to the | |
4286 | * number of hardware queues. | |
4287 | */ | |
4288 | if (set->nr_maps == 1) | |
4289 | set->map[HCTX_TYPE_DEFAULT].nr_queues = set->nr_hw_queues; | |
4290 | ||
59388702 | 4291 | if (set->ops->map_queues && !is_kdump_kernel()) { |
b3c661b1 JA |
4292 | int i; |
4293 | ||
7d4901a9 ML |
4294 | /* |
4295 | * transport .map_queues is usually done in the following | |
4296 | * way: | |
4297 | * | |
4298 | * for (queue = 0; queue < set->nr_hw_queues; queue++) { | |
4299 | * mask = get_cpu_mask(queue) | |
4300 | * for_each_cpu(cpu, mask) | |
b3c661b1 | 4301 | * set->map[x].mq_map[cpu] = queue; |
7d4901a9 ML |
4302 | * } |
4303 | * | |
4304 | * When we need to remap, the table has to be cleared for | |
4305 | * killing stale mapping since one CPU may not be mapped | |
4306 | * to any hw queue. | |
4307 | */ | |
b3c661b1 JA |
4308 | for (i = 0; i < set->nr_maps; i++) |
4309 | blk_mq_clear_mq_map(&set->map[i]); | |
7d4901a9 | 4310 | |
a4e1d0b7 | 4311 | set->ops->map_queues(set); |
b3c661b1 JA |
4312 | } else { |
4313 | BUG_ON(set->nr_maps > 1); | |
a4e1d0b7 | 4314 | blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]); |
b3c661b1 | 4315 | } |
ebe8bddb OS |
4316 | } |
4317 | ||
f7e76dbc BVA |
4318 | static int blk_mq_realloc_tag_set_tags(struct blk_mq_tag_set *set, |
4319 | int cur_nr_hw_queues, int new_nr_hw_queues) | |
4320 | { | |
4321 | struct blk_mq_tags **new_tags; | |
4322 | ||
4323 | if (cur_nr_hw_queues >= new_nr_hw_queues) | |
4324 | return 0; | |
4325 | ||
4326 | new_tags = kcalloc_node(new_nr_hw_queues, sizeof(struct blk_mq_tags *), | |
4327 | GFP_KERNEL, set->numa_node); | |
4328 | if (!new_tags) | |
4329 | return -ENOMEM; | |
4330 | ||
4331 | if (set->tags) | |
4332 | memcpy(new_tags, set->tags, cur_nr_hw_queues * | |
4333 | sizeof(*set->tags)); | |
4334 | kfree(set->tags); | |
4335 | set->tags = new_tags; | |
4336 | set->nr_hw_queues = new_nr_hw_queues; | |
4337 | ||
4338 | return 0; | |
4339 | } | |
4340 | ||
91cdf265 MI |
4341 | static int blk_mq_alloc_tag_set_tags(struct blk_mq_tag_set *set, |
4342 | int new_nr_hw_queues) | |
4343 | { | |
4344 | return blk_mq_realloc_tag_set_tags(set, 0, new_nr_hw_queues); | |
4345 | } | |
4346 | ||
a4391c64 JA |
4347 | /* |
4348 | * Alloc a tag set to be associated with one or more request queues. | |
4349 | * May fail with EINVAL for various error conditions. May adjust the | |
c018c84f | 4350 | * requested depth down, if it's too large. In that case, the set |
a4391c64 JA |
4351 | * value will be stored in set->queue_depth. |
4352 | */ | |
24d2f903 CH |
4353 | int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set) |
4354 | { | |
b3c661b1 | 4355 | int i, ret; |
da695ba2 | 4356 | |
205fb5f5 BVA |
4357 | BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS); |
4358 | ||
24d2f903 CH |
4359 | if (!set->nr_hw_queues) |
4360 | return -EINVAL; | |
a4391c64 | 4361 | if (!set->queue_depth) |
24d2f903 CH |
4362 | return -EINVAL; |
4363 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) | |
4364 | return -EINVAL; | |
4365 | ||
7d7e0f90 | 4366 | if (!set->ops->queue_rq) |
24d2f903 CH |
4367 | return -EINVAL; |
4368 | ||
de148297 ML |
4369 | if (!set->ops->get_budget ^ !set->ops->put_budget) |
4370 | return -EINVAL; | |
4371 | ||
a4391c64 JA |
4372 | if (set->queue_depth > BLK_MQ_MAX_DEPTH) { |
4373 | pr_info("blk-mq: reduced tag depth to %u\n", | |
4374 | BLK_MQ_MAX_DEPTH); | |
4375 | set->queue_depth = BLK_MQ_MAX_DEPTH; | |
4376 | } | |
24d2f903 | 4377 | |
b3c661b1 JA |
4378 | if (!set->nr_maps) |
4379 | set->nr_maps = 1; | |
4380 | else if (set->nr_maps > HCTX_MAX_TYPES) | |
4381 | return -EINVAL; | |
4382 | ||
6637fadf SL |
4383 | /* |
4384 | * If a crashdump is active, then we are potentially in a very | |
4385 | * memory constrained environment. Limit us to 1 queue and | |
4386 | * 64 tags to prevent using too much memory. | |
4387 | */ | |
4388 | if (is_kdump_kernel()) { | |
4389 | set->nr_hw_queues = 1; | |
59388702 | 4390 | set->nr_maps = 1; |
6637fadf SL |
4391 | set->queue_depth = min(64U, set->queue_depth); |
4392 | } | |
868f2f0b | 4393 | /* |
392546ae JA |
4394 | * There is no use for more h/w queues than cpus if we just have |
4395 | * a single map | |
868f2f0b | 4396 | */ |
392546ae | 4397 | if (set->nr_maps == 1 && set->nr_hw_queues > nr_cpu_ids) |
868f2f0b | 4398 | set->nr_hw_queues = nr_cpu_ids; |
6637fadf | 4399 | |
91cdf265 | 4400 | if (blk_mq_alloc_tag_set_tags(set, set->nr_hw_queues) < 0) |
a5164405 | 4401 | return -ENOMEM; |
24d2f903 | 4402 | |
da695ba2 | 4403 | ret = -ENOMEM; |
b3c661b1 JA |
4404 | for (i = 0; i < set->nr_maps; i++) { |
4405 | set->map[i].mq_map = kcalloc_node(nr_cpu_ids, | |
07b35eb5 | 4406 | sizeof(set->map[i].mq_map[0]), |
b3c661b1 JA |
4407 | GFP_KERNEL, set->numa_node); |
4408 | if (!set->map[i].mq_map) | |
4409 | goto out_free_mq_map; | |
59388702 | 4410 | set->map[i].nr_queues = is_kdump_kernel() ? 1 : set->nr_hw_queues; |
b3c661b1 | 4411 | } |
bdd17e75 | 4412 | |
a4e1d0b7 | 4413 | blk_mq_update_queue_map(set); |
da695ba2 | 4414 | |
63064be1 | 4415 | ret = blk_mq_alloc_set_map_and_rqs(set); |
da695ba2 | 4416 | if (ret) |
bdd17e75 | 4417 | goto out_free_mq_map; |
24d2f903 | 4418 | |
0d2602ca JA |
4419 | mutex_init(&set->tag_list_lock); |
4420 | INIT_LIST_HEAD(&set->tag_list); | |
4421 | ||
24d2f903 | 4422 | return 0; |
bdd17e75 CH |
4423 | |
4424 | out_free_mq_map: | |
b3c661b1 JA |
4425 | for (i = 0; i < set->nr_maps; i++) { |
4426 | kfree(set->map[i].mq_map); | |
4427 | set->map[i].mq_map = NULL; | |
4428 | } | |
5676e7b6 RE |
4429 | kfree(set->tags); |
4430 | set->tags = NULL; | |
da695ba2 | 4431 | return ret; |
24d2f903 CH |
4432 | } |
4433 | EXPORT_SYMBOL(blk_mq_alloc_tag_set); | |
4434 | ||
cdb14e0f CH |
4435 | /* allocate and initialize a tagset for a simple single-queue device */ |
4436 | int blk_mq_alloc_sq_tag_set(struct blk_mq_tag_set *set, | |
4437 | const struct blk_mq_ops *ops, unsigned int queue_depth, | |
4438 | unsigned int set_flags) | |
4439 | { | |
4440 | memset(set, 0, sizeof(*set)); | |
4441 | set->ops = ops; | |
4442 | set->nr_hw_queues = 1; | |
4443 | set->nr_maps = 1; | |
4444 | set->queue_depth = queue_depth; | |
4445 | set->numa_node = NUMA_NO_NODE; | |
4446 | set->flags = set_flags; | |
4447 | return blk_mq_alloc_tag_set(set); | |
4448 | } | |
4449 | EXPORT_SYMBOL_GPL(blk_mq_alloc_sq_tag_set); | |
4450 | ||
24d2f903 CH |
4451 | void blk_mq_free_tag_set(struct blk_mq_tag_set *set) |
4452 | { | |
b3c661b1 | 4453 | int i, j; |
24d2f903 | 4454 | |
f7e76dbc | 4455 | for (i = 0; i < set->nr_hw_queues; i++) |
e155b0c2 | 4456 | __blk_mq_free_map_and_rqs(set, i); |
484b4061 | 4457 | |
079a2e3e JG |
4458 | if (blk_mq_is_shared_tags(set->flags)) { |
4459 | blk_mq_free_map_and_rqs(set, set->shared_tags, | |
e155b0c2 JG |
4460 | BLK_MQ_NO_HCTX_IDX); |
4461 | } | |
32bc15af | 4462 | |
b3c661b1 JA |
4463 | for (j = 0; j < set->nr_maps; j++) { |
4464 | kfree(set->map[j].mq_map); | |
4465 | set->map[j].mq_map = NULL; | |
4466 | } | |
bdd17e75 | 4467 | |
981bd189 | 4468 | kfree(set->tags); |
5676e7b6 | 4469 | set->tags = NULL; |
24d2f903 CH |
4470 | } |
4471 | EXPORT_SYMBOL(blk_mq_free_tag_set); | |
4472 | ||
e3a2b3f9 JA |
4473 | int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr) |
4474 | { | |
4475 | struct blk_mq_tag_set *set = q->tag_set; | |
4476 | struct blk_mq_hw_ctx *hctx; | |
4f481208 ML |
4477 | int ret; |
4478 | unsigned long i; | |
e3a2b3f9 | 4479 | |
bd166ef1 | 4480 | if (!set) |
e3a2b3f9 JA |
4481 | return -EINVAL; |
4482 | ||
e5fa8140 AZ |
4483 | if (q->nr_requests == nr) |
4484 | return 0; | |
4485 | ||
70f36b60 | 4486 | blk_mq_freeze_queue(q); |
24f5a90f | 4487 | blk_mq_quiesce_queue(q); |
70f36b60 | 4488 | |
e3a2b3f9 JA |
4489 | ret = 0; |
4490 | queue_for_each_hw_ctx(q, hctx, i) { | |
e9137d4b KB |
4491 | if (!hctx->tags) |
4492 | continue; | |
bd166ef1 JA |
4493 | /* |
4494 | * If we're using an MQ scheduler, just update the scheduler | |
4495 | * queue depth. This is similar to what the old code would do. | |
4496 | */ | |
f6adcef5 | 4497 | if (hctx->sched_tags) { |
70f36b60 | 4498 | ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags, |
f6adcef5 | 4499 | nr, true); |
f6adcef5 JG |
4500 | } else { |
4501 | ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr, | |
4502 | false); | |
70f36b60 | 4503 | } |
e3a2b3f9 JA |
4504 | if (ret) |
4505 | break; | |
77f1e0a5 JA |
4506 | if (q->elevator && q->elevator->type->ops.depth_updated) |
4507 | q->elevator->type->ops.depth_updated(hctx); | |
e3a2b3f9 | 4508 | } |
d97e594c | 4509 | if (!ret) { |
e3a2b3f9 | 4510 | q->nr_requests = nr; |
079a2e3e | 4511 | if (blk_mq_is_shared_tags(set->flags)) { |
8fa04464 | 4512 | if (q->elevator) |
079a2e3e | 4513 | blk_mq_tag_update_sched_shared_tags(q); |
8fa04464 | 4514 | else |
079a2e3e | 4515 | blk_mq_tag_resize_shared_tags(set, nr); |
8fa04464 | 4516 | } |
d97e594c | 4517 | } |
e3a2b3f9 | 4518 | |
24f5a90f | 4519 | blk_mq_unquiesce_queue(q); |
70f36b60 | 4520 | blk_mq_unfreeze_queue(q); |
70f36b60 | 4521 | |
e3a2b3f9 JA |
4522 | return ret; |
4523 | } | |
4524 | ||
d48ece20 JW |
4525 | /* |
4526 | * request_queue and elevator_type pair. | |
4527 | * It is just used by __blk_mq_update_nr_hw_queues to cache | |
4528 | * the elevator_type associated with a request_queue. | |
4529 | */ | |
4530 | struct blk_mq_qe_pair { | |
4531 | struct list_head node; | |
4532 | struct request_queue *q; | |
4533 | struct elevator_type *type; | |
4534 | }; | |
4535 | ||
4536 | /* | |
4537 | * Cache the elevator_type in qe pair list and switch the | |
4538 | * io scheduler to 'none' | |
4539 | */ | |
4540 | static bool blk_mq_elv_switch_none(struct list_head *head, | |
4541 | struct request_queue *q) | |
4542 | { | |
4543 | struct blk_mq_qe_pair *qe; | |
4544 | ||
4545 | if (!q->elevator) | |
4546 | return true; | |
4547 | ||
4548 | qe = kmalloc(sizeof(*qe), GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY); | |
4549 | if (!qe) | |
4550 | return false; | |
4551 | ||
5fd7a84a ML |
4552 | /* q->elevator needs protection from ->sysfs_lock */ |
4553 | mutex_lock(&q->sysfs_lock); | |
4554 | ||
d48ece20 JW |
4555 | INIT_LIST_HEAD(&qe->node); |
4556 | qe->q = q; | |
dd6f7f17 CH |
4557 | /* keep a reference to the elevator module as we'll switch back */ |
4558 | __elevator_get(qe->type); | |
d48ece20 JW |
4559 | qe->type = q->elevator->type; |
4560 | list_add(&qe->node, head); | |
8237c01f | 4561 | elevator_switch(q, NULL); |
d48ece20 JW |
4562 | mutex_unlock(&q->sysfs_lock); |
4563 | ||
4564 | return true; | |
4565 | } | |
4566 | ||
4a3b666e JK |
4567 | static struct blk_mq_qe_pair *blk_lookup_qe_pair(struct list_head *head, |
4568 | struct request_queue *q) | |
d48ece20 JW |
4569 | { |
4570 | struct blk_mq_qe_pair *qe; | |
d48ece20 JW |
4571 | |
4572 | list_for_each_entry(qe, head, node) | |
4a3b666e JK |
4573 | if (qe->q == q) |
4574 | return qe; | |
d48ece20 | 4575 | |
4a3b666e JK |
4576 | return NULL; |
4577 | } | |
d48ece20 | 4578 | |
4a3b666e JK |
4579 | static void blk_mq_elv_switch_back(struct list_head *head, |
4580 | struct request_queue *q) | |
4581 | { | |
4582 | struct blk_mq_qe_pair *qe; | |
4583 | struct elevator_type *t; | |
4584 | ||
4585 | qe = blk_lookup_qe_pair(head, q); | |
4586 | if (!qe) | |
4587 | return; | |
4588 | t = qe->type; | |
d48ece20 JW |
4589 | list_del(&qe->node); |
4590 | kfree(qe); | |
4591 | ||
4592 | mutex_lock(&q->sysfs_lock); | |
8237c01f | 4593 | elevator_switch(q, t); |
d48ece20 JW |
4594 | mutex_unlock(&q->sysfs_lock); |
4595 | } | |
4596 | ||
e4dc2b32 KB |
4597 | static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, |
4598 | int nr_hw_queues) | |
868f2f0b KB |
4599 | { |
4600 | struct request_queue *q; | |
d48ece20 | 4601 | LIST_HEAD(head); |
e01ad46d | 4602 | int prev_nr_hw_queues; |
868f2f0b | 4603 | |
705cda97 BVA |
4604 | lockdep_assert_held(&set->tag_list_lock); |
4605 | ||
392546ae | 4606 | if (set->nr_maps == 1 && nr_hw_queues > nr_cpu_ids) |
868f2f0b | 4607 | nr_hw_queues = nr_cpu_ids; |
fe35ec58 WZ |
4608 | if (nr_hw_queues < 1) |
4609 | return; | |
4610 | if (set->nr_maps == 1 && nr_hw_queues == set->nr_hw_queues) | |
868f2f0b KB |
4611 | return; |
4612 | ||
4613 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
4614 | blk_mq_freeze_queue(q); | |
d48ece20 JW |
4615 | /* |
4616 | * Switch IO scheduler to 'none', cleaning up the data associated | |
4617 | * with the previous scheduler. We will switch back once we are done | |
4618 | * updating the new sw to hw queue mappings. | |
4619 | */ | |
4620 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
4621 | if (!blk_mq_elv_switch_none(&head, q)) | |
4622 | goto switch_back; | |
868f2f0b | 4623 | |
477e19de JW |
4624 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
4625 | blk_mq_debugfs_unregister_hctxs(q); | |
eaa870f9 | 4626 | blk_mq_sysfs_unregister_hctxs(q); |
477e19de JW |
4627 | } |
4628 | ||
a2584e43 | 4629 | prev_nr_hw_queues = set->nr_hw_queues; |
f7e76dbc BVA |
4630 | if (blk_mq_realloc_tag_set_tags(set, set->nr_hw_queues, nr_hw_queues) < |
4631 | 0) | |
4632 | goto reregister; | |
4633 | ||
868f2f0b | 4634 | set->nr_hw_queues = nr_hw_queues; |
e01ad46d | 4635 | fallback: |
aa880ad6 | 4636 | blk_mq_update_queue_map(set); |
868f2f0b KB |
4637 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
4638 | blk_mq_realloc_hw_ctxs(set, q); | |
42ee3061 | 4639 | blk_mq_update_poll_flag(q); |
e01ad46d | 4640 | if (q->nr_hw_queues != set->nr_hw_queues) { |
a846a8e6 YB |
4641 | int i = prev_nr_hw_queues; |
4642 | ||
e01ad46d JW |
4643 | pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n", |
4644 | nr_hw_queues, prev_nr_hw_queues); | |
a846a8e6 YB |
4645 | for (; i < set->nr_hw_queues; i++) |
4646 | __blk_mq_free_map_and_rqs(set, i); | |
4647 | ||
e01ad46d | 4648 | set->nr_hw_queues = prev_nr_hw_queues; |
7d76f856 | 4649 | blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]); |
e01ad46d JW |
4650 | goto fallback; |
4651 | } | |
477e19de JW |
4652 | blk_mq_map_swqueue(q); |
4653 | } | |
4654 | ||
f7e76dbc | 4655 | reregister: |
477e19de | 4656 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
eaa870f9 | 4657 | blk_mq_sysfs_register_hctxs(q); |
477e19de | 4658 | blk_mq_debugfs_register_hctxs(q); |
868f2f0b KB |
4659 | } |
4660 | ||
d48ece20 JW |
4661 | switch_back: |
4662 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
4663 | blk_mq_elv_switch_back(&head, q); | |
4664 | ||
868f2f0b KB |
4665 | list_for_each_entry(q, &set->tag_list, tag_set_list) |
4666 | blk_mq_unfreeze_queue(q); | |
4667 | } | |
e4dc2b32 KB |
4668 | |
4669 | void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues) | |
4670 | { | |
4671 | mutex_lock(&set->tag_list_lock); | |
4672 | __blk_mq_update_nr_hw_queues(set, nr_hw_queues); | |
4673 | mutex_unlock(&set->tag_list_lock); | |
4674 | } | |
868f2f0b KB |
4675 | EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues); |
4676 | ||
34dbad5d OS |
4677 | /* Enable polling stats and return whether they were already enabled. */ |
4678 | static bool blk_poll_stats_enable(struct request_queue *q) | |
4679 | { | |
48b5c1fb | 4680 | if (q->poll_stat) |
34dbad5d | 4681 | return true; |
48b5c1fb JA |
4682 | |
4683 | return blk_stats_alloc_enable(q); | |
34dbad5d OS |
4684 | } |
4685 | ||
4686 | static void blk_mq_poll_stats_start(struct request_queue *q) | |
4687 | { | |
4688 | /* | |
4689 | * We don't arm the callback if polling stats are not enabled or the | |
4690 | * callback is already active. | |
4691 | */ | |
48b5c1fb | 4692 | if (!q->poll_stat || blk_stat_is_active(q->poll_cb)) |
34dbad5d OS |
4693 | return; |
4694 | ||
4695 | blk_stat_activate_msecs(q->poll_cb, 100); | |
4696 | } | |
4697 | ||
4698 | static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb) | |
4699 | { | |
4700 | struct request_queue *q = cb->data; | |
720b8ccc | 4701 | int bucket; |
34dbad5d | 4702 | |
720b8ccc SB |
4703 | for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) { |
4704 | if (cb->stat[bucket].nr_samples) | |
4705 | q->poll_stat[bucket] = cb->stat[bucket]; | |
4706 | } | |
34dbad5d OS |
4707 | } |
4708 | ||
64f1c21e | 4709 | static unsigned long blk_mq_poll_nsecs(struct request_queue *q, |
64f1c21e JA |
4710 | struct request *rq) |
4711 | { | |
64f1c21e | 4712 | unsigned long ret = 0; |
720b8ccc | 4713 | int bucket; |
64f1c21e JA |
4714 | |
4715 | /* | |
4716 | * If stats collection isn't on, don't sleep but turn it on for | |
4717 | * future users | |
4718 | */ | |
34dbad5d | 4719 | if (!blk_poll_stats_enable(q)) |
64f1c21e JA |
4720 | return 0; |
4721 | ||
64f1c21e JA |
4722 | /* |
4723 | * As an optimistic guess, use half of the mean service time | |
4724 | * for this type of request. We can (and should) make this smarter. | |
4725 | * For instance, if the completion latencies are tight, we can | |
4726 | * get closer than just half the mean. This is especially | |
4727 | * important on devices where the completion latencies are longer | |
720b8ccc SB |
4728 | * than ~10 usec. We do use the stats for the relevant IO size |
4729 | * if available which does lead to better estimates. | |
64f1c21e | 4730 | */ |
720b8ccc SB |
4731 | bucket = blk_mq_poll_stats_bkt(rq); |
4732 | if (bucket < 0) | |
4733 | return ret; | |
4734 | ||
4735 | if (q->poll_stat[bucket].nr_samples) | |
4736 | ret = (q->poll_stat[bucket].mean + 1) / 2; | |
64f1c21e JA |
4737 | |
4738 | return ret; | |
4739 | } | |
4740 | ||
c6699d6f | 4741 | static bool blk_mq_poll_hybrid(struct request_queue *q, blk_qc_t qc) |
06426adf | 4742 | { |
c6699d6f CH |
4743 | struct blk_mq_hw_ctx *hctx = blk_qc_to_hctx(q, qc); |
4744 | struct request *rq = blk_qc_to_rq(hctx, qc); | |
06426adf JA |
4745 | struct hrtimer_sleeper hs; |
4746 | enum hrtimer_mode mode; | |
64f1c21e | 4747 | unsigned int nsecs; |
06426adf JA |
4748 | ktime_t kt; |
4749 | ||
c6699d6f CH |
4750 | /* |
4751 | * If a request has completed on queue that uses an I/O scheduler, we | |
4752 | * won't get back a request from blk_qc_to_rq. | |
4753 | */ | |
4754 | if (!rq || (rq->rq_flags & RQF_MQ_POLL_SLEPT)) | |
64f1c21e JA |
4755 | return false; |
4756 | ||
4757 | /* | |
1052b8ac | 4758 | * If we get here, hybrid polling is enabled. Hence poll_nsec can be: |
64f1c21e | 4759 | * |
64f1c21e JA |
4760 | * 0: use half of prev avg |
4761 | * >0: use this specific value | |
4762 | */ | |
1052b8ac | 4763 | if (q->poll_nsec > 0) |
64f1c21e JA |
4764 | nsecs = q->poll_nsec; |
4765 | else | |
cae740a0 | 4766 | nsecs = blk_mq_poll_nsecs(q, rq); |
64f1c21e JA |
4767 | |
4768 | if (!nsecs) | |
06426adf JA |
4769 | return false; |
4770 | ||
76a86f9d | 4771 | rq->rq_flags |= RQF_MQ_POLL_SLEPT; |
06426adf JA |
4772 | |
4773 | /* | |
4774 | * This will be replaced with the stats tracking code, using | |
4775 | * 'avg_completion_time / 2' as the pre-sleep target. | |
4776 | */ | |
8b0e1953 | 4777 | kt = nsecs; |
06426adf JA |
4778 | |
4779 | mode = HRTIMER_MODE_REL; | |
dbc1625f | 4780 | hrtimer_init_sleeper_on_stack(&hs, CLOCK_MONOTONIC, mode); |
06426adf JA |
4781 | hrtimer_set_expires(&hs.timer, kt); |
4782 | ||
06426adf | 4783 | do { |
5a61c363 | 4784 | if (blk_mq_rq_state(rq) == MQ_RQ_COMPLETE) |
06426adf JA |
4785 | break; |
4786 | set_current_state(TASK_UNINTERRUPTIBLE); | |
9dd8813e | 4787 | hrtimer_sleeper_start_expires(&hs, mode); |
06426adf JA |
4788 | if (hs.task) |
4789 | io_schedule(); | |
4790 | hrtimer_cancel(&hs.timer); | |
4791 | mode = HRTIMER_MODE_ABS; | |
4792 | } while (hs.task && !signal_pending(current)); | |
4793 | ||
4794 | __set_current_state(TASK_RUNNING); | |
4795 | destroy_hrtimer_on_stack(&hs.timer); | |
1052b8ac | 4796 | |
06426adf | 4797 | /* |
c6699d6f CH |
4798 | * If we sleep, have the caller restart the poll loop to reset the |
4799 | * state. Like for the other success return cases, the caller is | |
4800 | * responsible for checking if the IO completed. If the IO isn't | |
4801 | * complete, we'll get called again and will go straight to the busy | |
4802 | * poll loop. | |
06426adf | 4803 | */ |
06426adf JA |
4804 | return true; |
4805 | } | |
06426adf | 4806 | |
c6699d6f | 4807 | static int blk_mq_poll_classic(struct request_queue *q, blk_qc_t cookie, |
5a72e899 | 4808 | struct io_comp_batch *iob, unsigned int flags) |
bbd7bb70 | 4809 | { |
c6699d6f CH |
4810 | struct blk_mq_hw_ctx *hctx = blk_qc_to_hctx(q, cookie); |
4811 | long state = get_current_state(); | |
4812 | int ret; | |
bbd7bb70 | 4813 | |
aa61bec3 | 4814 | do { |
5a72e899 | 4815 | ret = q->mq_ops->poll(hctx, iob); |
bbd7bb70 | 4816 | if (ret > 0) { |
849a3700 | 4817 | __set_current_state(TASK_RUNNING); |
85f4d4b6 | 4818 | return ret; |
bbd7bb70 JA |
4819 | } |
4820 | ||
4821 | if (signal_pending_state(state, current)) | |
849a3700 | 4822 | __set_current_state(TASK_RUNNING); |
b03fbd4f | 4823 | if (task_is_running(current)) |
85f4d4b6 | 4824 | return 1; |
c6699d6f | 4825 | |
ef99b2d3 | 4826 | if (ret < 0 || (flags & BLK_POLL_ONESHOT)) |
bbd7bb70 JA |
4827 | break; |
4828 | cpu_relax(); | |
aa61bec3 | 4829 | } while (!need_resched()); |
bbd7bb70 | 4830 | |
67b4110f | 4831 | __set_current_state(TASK_RUNNING); |
85f4d4b6 | 4832 | return 0; |
bbd7bb70 | 4833 | } |
1052b8ac | 4834 | |
5a72e899 JA |
4835 | int blk_mq_poll(struct request_queue *q, blk_qc_t cookie, struct io_comp_batch *iob, |
4836 | unsigned int flags) | |
1052b8ac | 4837 | { |
d729cf9a CH |
4838 | if (!(flags & BLK_POLL_NOSLEEP) && |
4839 | q->poll_nsec != BLK_MQ_POLL_CLASSIC) { | |
c6699d6f | 4840 | if (blk_mq_poll_hybrid(q, cookie)) |
85f4d4b6 | 4841 | return 1; |
c6699d6f | 4842 | } |
5a72e899 | 4843 | return blk_mq_poll_classic(q, cookie, iob, flags); |
bbd7bb70 JA |
4844 | } |
4845 | ||
9cf2bab6 JA |
4846 | unsigned int blk_mq_rq_cpu(struct request *rq) |
4847 | { | |
4848 | return rq->mq_ctx->cpu; | |
4849 | } | |
4850 | EXPORT_SYMBOL(blk_mq_rq_cpu); | |
4851 | ||
2a19b28f ML |
4852 | void blk_mq_cancel_work_sync(struct request_queue *q) |
4853 | { | |
4854 | if (queue_is_mq(q)) { | |
4855 | struct blk_mq_hw_ctx *hctx; | |
4f481208 | 4856 | unsigned long i; |
2a19b28f ML |
4857 | |
4858 | cancel_delayed_work_sync(&q->requeue_work); | |
4859 | ||
4860 | queue_for_each_hw_ctx(q, hctx, i) | |
4861 | cancel_delayed_work_sync(&hctx->run_work); | |
4862 | } | |
4863 | } | |
4864 | ||
320ae51f JA |
4865 | static int __init blk_mq_init(void) |
4866 | { | |
c3077b5d CH |
4867 | int i; |
4868 | ||
4869 | for_each_possible_cpu(i) | |
f9ab4918 | 4870 | init_llist_head(&per_cpu(blk_cpu_done, i)); |
c3077b5d CH |
4871 | open_softirq(BLOCK_SOFTIRQ, blk_done_softirq); |
4872 | ||
4873 | cpuhp_setup_state_nocalls(CPUHP_BLOCK_SOFTIRQ_DEAD, | |
4874 | "block/softirq:dead", NULL, | |
4875 | blk_softirq_cpu_dead); | |
9467f859 TG |
4876 | cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL, |
4877 | blk_mq_hctx_notify_dead); | |
bf0beec0 ML |
4878 | cpuhp_setup_state_multi(CPUHP_AP_BLK_MQ_ONLINE, "block/mq:online", |
4879 | blk_mq_hctx_notify_online, | |
4880 | blk_mq_hctx_notify_offline); | |
320ae51f JA |
4881 | return 0; |
4882 | } | |
4883 | subsys_initcall(blk_mq_init); |