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 |
483239c7 | 257 | * @set: tag_set to wait on |
6a83e74d | 258 | * |
9ef4d020 | 259 | * Note: it is driver's responsibility for making sure that quiesce has |
483239c7 CH |
260 | * been started on or more of the request_queues of the tag_set. This |
261 | * function only waits for the quiesce on those request_queues that had | |
262 | * the quiesce flag set using blk_mq_quiesce_queue_nowait. | |
6a83e74d | 263 | */ |
483239c7 | 264 | void blk_mq_wait_quiesce_done(struct blk_mq_tag_set *set) |
6a83e74d | 265 | { |
483239c7 CH |
266 | if (set->flags & BLK_MQ_F_BLOCKING) |
267 | synchronize_srcu(set->srcu); | |
704b914f | 268 | else |
6a83e74d BVA |
269 | synchronize_rcu(); |
270 | } | |
9ef4d020 ML |
271 | EXPORT_SYMBOL_GPL(blk_mq_wait_quiesce_done); |
272 | ||
273 | /** | |
274 | * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished | |
275 | * @q: request queue. | |
276 | * | |
277 | * Note: this function does not prevent that the struct request end_io() | |
278 | * callback function is invoked. Once this function is returned, we make | |
279 | * sure no dispatch can happen until the queue is unquiesced via | |
280 | * blk_mq_unquiesce_queue(). | |
281 | */ | |
282 | void blk_mq_quiesce_queue(struct request_queue *q) | |
283 | { | |
284 | blk_mq_quiesce_queue_nowait(q); | |
8537380b CH |
285 | /* nothing to wait for non-mq queues */ |
286 | if (queue_is_mq(q)) | |
483239c7 | 287 | blk_mq_wait_quiesce_done(q->tag_set); |
9ef4d020 | 288 | } |
6a83e74d BVA |
289 | EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue); |
290 | ||
e4e73913 ML |
291 | /* |
292 | * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue() | |
293 | * @q: request queue. | |
294 | * | |
295 | * This function recovers queue into the state before quiescing | |
296 | * which is done by blk_mq_quiesce_queue. | |
297 | */ | |
298 | void blk_mq_unquiesce_queue(struct request_queue *q) | |
299 | { | |
e70feb8b ML |
300 | unsigned long flags; |
301 | bool run_queue = false; | |
302 | ||
303 | spin_lock_irqsave(&q->queue_lock, flags); | |
304 | if (WARN_ON_ONCE(q->quiesce_depth <= 0)) { | |
305 | ; | |
306 | } else if (!--q->quiesce_depth) { | |
307 | blk_queue_flag_clear(QUEUE_FLAG_QUIESCED, q); | |
308 | run_queue = true; | |
309 | } | |
310 | spin_unlock_irqrestore(&q->queue_lock, flags); | |
f4560ffe | 311 | |
1d9e9bc6 | 312 | /* dispatch requests which are inserted during quiescing */ |
e70feb8b ML |
313 | if (run_queue) |
314 | blk_mq_run_hw_queues(q, true); | |
e4e73913 ML |
315 | } |
316 | EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue); | |
317 | ||
414dd48e CL |
318 | void blk_mq_quiesce_tagset(struct blk_mq_tag_set *set) |
319 | { | |
320 | struct request_queue *q; | |
321 | ||
322 | mutex_lock(&set->tag_list_lock); | |
323 | list_for_each_entry(q, &set->tag_list, tag_set_list) { | |
324 | if (!blk_queue_skip_tagset_quiesce(q)) | |
325 | blk_mq_quiesce_queue_nowait(q); | |
326 | } | |
327 | blk_mq_wait_quiesce_done(set); | |
328 | mutex_unlock(&set->tag_list_lock); | |
329 | } | |
330 | EXPORT_SYMBOL_GPL(blk_mq_quiesce_tagset); | |
331 | ||
332 | void blk_mq_unquiesce_tagset(struct blk_mq_tag_set *set) | |
333 | { | |
334 | struct request_queue *q; | |
335 | ||
336 | mutex_lock(&set->tag_list_lock); | |
337 | list_for_each_entry(q, &set->tag_list, tag_set_list) { | |
338 | if (!blk_queue_skip_tagset_quiesce(q)) | |
339 | blk_mq_unquiesce_queue(q); | |
340 | } | |
341 | mutex_unlock(&set->tag_list_lock); | |
342 | } | |
343 | EXPORT_SYMBOL_GPL(blk_mq_unquiesce_tagset); | |
344 | ||
aed3ea94 JA |
345 | void blk_mq_wake_waiters(struct request_queue *q) |
346 | { | |
347 | struct blk_mq_hw_ctx *hctx; | |
4f481208 | 348 | unsigned long i; |
aed3ea94 JA |
349 | |
350 | queue_for_each_hw_ctx(q, hctx, i) | |
351 | if (blk_mq_hw_queue_mapped(hctx)) | |
352 | blk_mq_tag_wakeup_all(hctx->tags, true); | |
353 | } | |
354 | ||
52fdbbcc CH |
355 | void blk_rq_init(struct request_queue *q, struct request *rq) |
356 | { | |
357 | memset(rq, 0, sizeof(*rq)); | |
358 | ||
359 | INIT_LIST_HEAD(&rq->queuelist); | |
360 | rq->q = q; | |
361 | rq->__sector = (sector_t) -1; | |
362 | INIT_HLIST_NODE(&rq->hash); | |
363 | RB_CLEAR_NODE(&rq->rb_node); | |
364 | rq->tag = BLK_MQ_NO_TAG; | |
365 | rq->internal_tag = BLK_MQ_NO_TAG; | |
366 | rq->start_time_ns = ktime_get_ns(); | |
367 | rq->part = NULL; | |
368 | blk_crypto_rq_set_defaults(rq); | |
369 | } | |
370 | EXPORT_SYMBOL(blk_rq_init); | |
371 | ||
e4cdf1a1 | 372 | static struct request *blk_mq_rq_ctx_init(struct blk_mq_alloc_data *data, |
fe6134f6 | 373 | struct blk_mq_tags *tags, unsigned int tag, u64 alloc_time_ns) |
320ae51f | 374 | { |
605f784e PB |
375 | struct blk_mq_ctx *ctx = data->ctx; |
376 | struct blk_mq_hw_ctx *hctx = data->hctx; | |
377 | struct request_queue *q = data->q; | |
e4cdf1a1 | 378 | struct request *rq = tags->static_rqs[tag]; |
c3a148d2 | 379 | |
c7b84d42 JA |
380 | rq->q = q; |
381 | rq->mq_ctx = ctx; | |
382 | rq->mq_hctx = hctx; | |
383 | rq->cmd_flags = data->cmd_flags; | |
384 | ||
385 | if (data->flags & BLK_MQ_REQ_PM) | |
386 | data->rq_flags |= RQF_PM; | |
387 | if (blk_queue_io_stat(q)) | |
388 | data->rq_flags |= RQF_IO_STAT; | |
389 | rq->rq_flags = data->rq_flags; | |
390 | ||
56f8da64 | 391 | if (!(data->rq_flags & RQF_ELV)) { |
e4cdf1a1 | 392 | rq->tag = tag; |
76647368 | 393 | rq->internal_tag = BLK_MQ_NO_TAG; |
56f8da64 JA |
394 | } else { |
395 | rq->tag = BLK_MQ_NO_TAG; | |
396 | rq->internal_tag = tag; | |
e4cdf1a1 | 397 | } |
c7b84d42 | 398 | rq->timeout = 0; |
e4cdf1a1 | 399 | |
4f266f2b PB |
400 | if (blk_mq_need_time_stamp(rq)) |
401 | rq->start_time_ns = ktime_get_ns(); | |
402 | else | |
403 | rq->start_time_ns = 0; | |
af76e555 | 404 | rq->part = NULL; |
6f816b4b TH |
405 | #ifdef CONFIG_BLK_RQ_ALLOC_TIME |
406 | rq->alloc_time_ns = alloc_time_ns; | |
407 | #endif | |
544ccc8d | 408 | rq->io_start_time_ns = 0; |
3d244306 | 409 | rq->stats_sectors = 0; |
af76e555 CH |
410 | rq->nr_phys_segments = 0; |
411 | #if defined(CONFIG_BLK_DEV_INTEGRITY) | |
412 | rq->nr_integrity_segments = 0; | |
413 | #endif | |
af76e555 CH |
414 | rq->end_io = NULL; |
415 | rq->end_io_data = NULL; | |
af76e555 | 416 | |
4f266f2b PB |
417 | blk_crypto_rq_set_defaults(rq); |
418 | INIT_LIST_HEAD(&rq->queuelist); | |
419 | /* tag was already set */ | |
420 | WRITE_ONCE(rq->deadline, 0); | |
0a467d0f | 421 | req_ref_set(rq, 1); |
7ea4d8a4 | 422 | |
4f266f2b | 423 | if (rq->rq_flags & RQF_ELV) { |
7ea4d8a4 CH |
424 | struct elevator_queue *e = data->q->elevator; |
425 | ||
4f266f2b PB |
426 | INIT_HLIST_NODE(&rq->hash); |
427 | RB_CLEAR_NODE(&rq->rb_node); | |
428 | ||
429 | if (!op_is_flush(data->cmd_flags) && | |
430 | e->type->ops.prepare_request) { | |
7ea4d8a4 CH |
431 | e->type->ops.prepare_request(rq); |
432 | rq->rq_flags |= RQF_ELVPRIV; | |
433 | } | |
434 | } | |
435 | ||
e4cdf1a1 | 436 | return rq; |
5dee8577 CH |
437 | } |
438 | ||
349302da JA |
439 | static inline struct request * |
440 | __blk_mq_alloc_requests_batch(struct blk_mq_alloc_data *data, | |
441 | u64 alloc_time_ns) | |
442 | { | |
443 | unsigned int tag, tag_offset; | |
fe6134f6 | 444 | struct blk_mq_tags *tags; |
349302da | 445 | struct request *rq; |
fe6134f6 | 446 | unsigned long tag_mask; |
349302da JA |
447 | int i, nr = 0; |
448 | ||
fe6134f6 JA |
449 | tag_mask = blk_mq_get_tags(data, data->nr_tags, &tag_offset); |
450 | if (unlikely(!tag_mask)) | |
349302da JA |
451 | return NULL; |
452 | ||
fe6134f6 JA |
453 | tags = blk_mq_tags_from_data(data); |
454 | for (i = 0; tag_mask; i++) { | |
455 | if (!(tag_mask & (1UL << i))) | |
349302da JA |
456 | continue; |
457 | tag = tag_offset + i; | |
a22c00be | 458 | prefetch(tags->static_rqs[tag]); |
fe6134f6 JA |
459 | tag_mask &= ~(1UL << i); |
460 | rq = blk_mq_rq_ctx_init(data, tags, tag, alloc_time_ns); | |
013a7f95 | 461 | rq_list_add(data->cached_rq, rq); |
c5fc7b93 | 462 | nr++; |
349302da | 463 | } |
c5fc7b93 JA |
464 | /* caller already holds a reference, add for remainder */ |
465 | percpu_ref_get_many(&data->q->q_usage_counter, nr - 1); | |
349302da JA |
466 | data->nr_tags -= nr; |
467 | ||
013a7f95 | 468 | return rq_list_pop(data->cached_rq); |
349302da JA |
469 | } |
470 | ||
b90cfaed | 471 | static struct request *__blk_mq_alloc_requests(struct blk_mq_alloc_data *data) |
d2c0d383 | 472 | { |
e6e7abff | 473 | struct request_queue *q = data->q; |
6f816b4b | 474 | u64 alloc_time_ns = 0; |
47c122e3 | 475 | struct request *rq; |
600c3b0c | 476 | unsigned int tag; |
d2c0d383 | 477 | |
6f816b4b TH |
478 | /* alloc_time includes depth and tag waits */ |
479 | if (blk_queue_rq_alloc_time(q)) | |
480 | alloc_time_ns = ktime_get_ns(); | |
481 | ||
f9afca4d | 482 | if (data->cmd_flags & REQ_NOWAIT) |
03a07c92 | 483 | data->flags |= BLK_MQ_REQ_NOWAIT; |
d2c0d383 | 484 | |
781dd830 JA |
485 | if (q->elevator) { |
486 | struct elevator_queue *e = q->elevator; | |
487 | ||
488 | data->rq_flags |= RQF_ELV; | |
489 | ||
d2c0d383 | 490 | /* |
8d663f34 | 491 | * Flush/passthrough requests are special and go directly to the |
17a51199 JA |
492 | * dispatch list. Don't include reserved tags in the |
493 | * limiting, as it isn't useful. | |
d2c0d383 | 494 | */ |
f9afca4d | 495 | if (!op_is_flush(data->cmd_flags) && |
8d663f34 | 496 | !blk_op_is_passthrough(data->cmd_flags) && |
f9afca4d | 497 | e->type->ops.limit_depth && |
17a51199 | 498 | !(data->flags & BLK_MQ_REQ_RESERVED)) |
f9afca4d | 499 | e->type->ops.limit_depth(data->cmd_flags, data); |
d2c0d383 CH |
500 | } |
501 | ||
bf0beec0 | 502 | retry: |
600c3b0c CH |
503 | data->ctx = blk_mq_get_ctx(q); |
504 | data->hctx = blk_mq_map_queue(q, data->cmd_flags, data->ctx); | |
781dd830 | 505 | if (!(data->rq_flags & RQF_ELV)) |
600c3b0c CH |
506 | blk_mq_tag_busy(data->hctx); |
507 | ||
99e48cd6 JG |
508 | if (data->flags & BLK_MQ_REQ_RESERVED) |
509 | data->rq_flags |= RQF_RESV; | |
510 | ||
349302da JA |
511 | /* |
512 | * Try batched alloc if we want more than 1 tag. | |
513 | */ | |
514 | if (data->nr_tags > 1) { | |
515 | rq = __blk_mq_alloc_requests_batch(data, alloc_time_ns); | |
516 | if (rq) | |
517 | return rq; | |
518 | data->nr_tags = 1; | |
519 | } | |
520 | ||
bf0beec0 ML |
521 | /* |
522 | * Waiting allocations only fail because of an inactive hctx. In that | |
523 | * case just retry the hctx assignment and tag allocation as CPU hotplug | |
524 | * should have migrated us to an online CPU by now. | |
525 | */ | |
e4cdf1a1 | 526 | tag = blk_mq_get_tag(data); |
bf0beec0 ML |
527 | if (tag == BLK_MQ_NO_TAG) { |
528 | if (data->flags & BLK_MQ_REQ_NOWAIT) | |
529 | return NULL; | |
bf0beec0 | 530 | /* |
349302da JA |
531 | * Give up the CPU and sleep for a random short time to |
532 | * ensure that thread using a realtime scheduling class | |
533 | * are migrated off the CPU, and thus off the hctx that | |
534 | * is going away. | |
bf0beec0 ML |
535 | */ |
536 | msleep(3); | |
537 | goto retry; | |
538 | } | |
47c122e3 | 539 | |
fe6134f6 JA |
540 | return blk_mq_rq_ctx_init(data, blk_mq_tags_from_data(data), tag, |
541 | alloc_time_ns); | |
d2c0d383 CH |
542 | } |
543 | ||
4b6a5d9c JA |
544 | static struct request *blk_mq_rq_cache_fill(struct request_queue *q, |
545 | struct blk_plug *plug, | |
546 | blk_opf_t opf, | |
547 | blk_mq_req_flags_t flags) | |
320ae51f | 548 | { |
e6e7abff CH |
549 | struct blk_mq_alloc_data data = { |
550 | .q = q, | |
551 | .flags = flags, | |
16458cf3 | 552 | .cmd_flags = opf, |
4b6a5d9c JA |
553 | .nr_tags = plug->nr_ios, |
554 | .cached_rq = &plug->cached_rq, | |
e6e7abff | 555 | }; |
bd166ef1 | 556 | struct request *rq; |
320ae51f | 557 | |
4b6a5d9c JA |
558 | if (blk_queue_enter(q, flags)) |
559 | return NULL; | |
560 | ||
561 | plug->nr_ios = 1; | |
320ae51f | 562 | |
b90cfaed | 563 | rq = __blk_mq_alloc_requests(&data); |
4b6a5d9c JA |
564 | if (unlikely(!rq)) |
565 | blk_queue_exit(q); | |
566 | return rq; | |
567 | } | |
568 | ||
569 | static struct request *blk_mq_alloc_cached_request(struct request_queue *q, | |
570 | blk_opf_t opf, | |
571 | blk_mq_req_flags_t flags) | |
572 | { | |
573 | struct blk_plug *plug = current->plug; | |
574 | struct request *rq; | |
575 | ||
576 | if (!plug) | |
577 | return NULL; | |
40467282 | 578 | |
4b6a5d9c JA |
579 | if (rq_list_empty(plug->cached_rq)) { |
580 | if (plug->nr_ios == 1) | |
581 | return NULL; | |
582 | rq = blk_mq_rq_cache_fill(q, plug, opf, flags); | |
40467282 JC |
583 | if (!rq) |
584 | return NULL; | |
585 | } else { | |
586 | rq = rq_list_peek(&plug->cached_rq); | |
587 | if (!rq || rq->q != q) | |
588 | return NULL; | |
4b6a5d9c | 589 | |
40467282 JC |
590 | if (blk_mq_get_hctx_type(opf) != rq->mq_hctx->type) |
591 | return NULL; | |
592 | if (op_is_flush(rq->cmd_flags) != op_is_flush(opf)) | |
593 | return NULL; | |
594 | ||
595 | plug->cached_rq = rq_list_next(rq); | |
596 | } | |
4b6a5d9c | 597 | |
4b6a5d9c JA |
598 | rq->cmd_flags = opf; |
599 | INIT_LIST_HEAD(&rq->queuelist); | |
600 | return rq; | |
601 | } | |
602 | ||
603 | struct request *blk_mq_alloc_request(struct request_queue *q, blk_opf_t opf, | |
604 | blk_mq_req_flags_t flags) | |
605 | { | |
606 | struct request *rq; | |
607 | ||
608 | rq = blk_mq_alloc_cached_request(q, opf, flags); | |
609 | if (!rq) { | |
610 | struct blk_mq_alloc_data data = { | |
611 | .q = q, | |
612 | .flags = flags, | |
613 | .cmd_flags = opf, | |
614 | .nr_tags = 1, | |
615 | }; | |
616 | int ret; | |
617 | ||
618 | ret = blk_queue_enter(q, flags); | |
619 | if (ret) | |
620 | return ERR_PTR(ret); | |
621 | ||
622 | rq = __blk_mq_alloc_requests(&data); | |
623 | if (!rq) | |
624 | goto out_queue_exit; | |
625 | } | |
0c4de0f3 CH |
626 | rq->__data_len = 0; |
627 | rq->__sector = (sector_t) -1; | |
628 | rq->bio = rq->biotail = NULL; | |
320ae51f | 629 | return rq; |
a5ea5811 CH |
630 | out_queue_exit: |
631 | blk_queue_exit(q); | |
632 | return ERR_PTR(-EWOULDBLOCK); | |
320ae51f | 633 | } |
4bb659b1 | 634 | EXPORT_SYMBOL(blk_mq_alloc_request); |
320ae51f | 635 | |
cd6ce148 | 636 | struct request *blk_mq_alloc_request_hctx(struct request_queue *q, |
16458cf3 | 637 | blk_opf_t opf, blk_mq_req_flags_t flags, unsigned int hctx_idx) |
1f5bd336 | 638 | { |
e6e7abff CH |
639 | struct blk_mq_alloc_data data = { |
640 | .q = q, | |
641 | .flags = flags, | |
16458cf3 | 642 | .cmd_flags = opf, |
47c122e3 | 643 | .nr_tags = 1, |
e6e7abff | 644 | }; |
600c3b0c | 645 | u64 alloc_time_ns = 0; |
e3c5a78c | 646 | struct request *rq; |
6d2809d5 | 647 | unsigned int cpu; |
600c3b0c | 648 | unsigned int tag; |
1f5bd336 ML |
649 | int ret; |
650 | ||
600c3b0c CH |
651 | /* alloc_time includes depth and tag waits */ |
652 | if (blk_queue_rq_alloc_time(q)) | |
653 | alloc_time_ns = ktime_get_ns(); | |
654 | ||
1f5bd336 ML |
655 | /* |
656 | * If the tag allocator sleeps we could get an allocation for a | |
657 | * different hardware context. No need to complicate the low level | |
658 | * allocator for this for the rare use case of a command tied to | |
659 | * a specific queue. | |
660 | */ | |
600c3b0c | 661 | if (WARN_ON_ONCE(!(flags & (BLK_MQ_REQ_NOWAIT | BLK_MQ_REQ_RESERVED)))) |
1f5bd336 ML |
662 | return ERR_PTR(-EINVAL); |
663 | ||
664 | if (hctx_idx >= q->nr_hw_queues) | |
665 | return ERR_PTR(-EIO); | |
666 | ||
3a0a5299 | 667 | ret = blk_queue_enter(q, flags); |
1f5bd336 ML |
668 | if (ret) |
669 | return ERR_PTR(ret); | |
670 | ||
c8712c6a CH |
671 | /* |
672 | * Check if the hardware context is actually mapped to anything. | |
673 | * If not tell the caller that it should skip this queue. | |
674 | */ | |
a5ea5811 | 675 | ret = -EXDEV; |
4e5cc99e | 676 | data.hctx = xa_load(&q->hctx_table, hctx_idx); |
e6e7abff | 677 | if (!blk_mq_hw_queue_mapped(data.hctx)) |
a5ea5811 | 678 | goto out_queue_exit; |
e6e7abff | 679 | cpu = cpumask_first_and(data.hctx->cpumask, cpu_online_mask); |
14dc7a18 BVA |
680 | if (cpu >= nr_cpu_ids) |
681 | goto out_queue_exit; | |
e6e7abff | 682 | data.ctx = __blk_mq_get_ctx(q, cpu); |
1f5bd336 | 683 | |
42fdc5e4 | 684 | if (!q->elevator) |
600c3b0c | 685 | blk_mq_tag_busy(data.hctx); |
781dd830 JA |
686 | else |
687 | data.rq_flags |= RQF_ELV; | |
600c3b0c | 688 | |
99e48cd6 JG |
689 | if (flags & BLK_MQ_REQ_RESERVED) |
690 | data.rq_flags |= RQF_RESV; | |
691 | ||
a5ea5811 | 692 | ret = -EWOULDBLOCK; |
600c3b0c CH |
693 | tag = blk_mq_get_tag(&data); |
694 | if (tag == BLK_MQ_NO_TAG) | |
a5ea5811 | 695 | goto out_queue_exit; |
e3c5a78c | 696 | rq = blk_mq_rq_ctx_init(&data, blk_mq_tags_from_data(&data), tag, |
fe6134f6 | 697 | alloc_time_ns); |
e3c5a78c JG |
698 | rq->__data_len = 0; |
699 | rq->__sector = (sector_t) -1; | |
700 | rq->bio = rq->biotail = NULL; | |
701 | return rq; | |
600c3b0c | 702 | |
a5ea5811 CH |
703 | out_queue_exit: |
704 | blk_queue_exit(q); | |
705 | return ERR_PTR(ret); | |
1f5bd336 ML |
706 | } |
707 | EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx); | |
708 | ||
12f5b931 KB |
709 | static void __blk_mq_free_request(struct request *rq) |
710 | { | |
711 | struct request_queue *q = rq->q; | |
712 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
ea4f995e | 713 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; |
12f5b931 KB |
714 | const int sched_tag = rq->internal_tag; |
715 | ||
a892c8d5 | 716 | blk_crypto_free_request(rq); |
986d413b | 717 | blk_pm_mark_last_busy(rq); |
ea4f995e | 718 | rq->mq_hctx = NULL; |
76647368 | 719 | if (rq->tag != BLK_MQ_NO_TAG) |
cae740a0 | 720 | blk_mq_put_tag(hctx->tags, ctx, rq->tag); |
76647368 | 721 | if (sched_tag != BLK_MQ_NO_TAG) |
cae740a0 | 722 | blk_mq_put_tag(hctx->sched_tags, ctx, sched_tag); |
12f5b931 KB |
723 | blk_mq_sched_restart(hctx); |
724 | blk_queue_exit(q); | |
725 | } | |
726 | ||
6af54051 | 727 | void blk_mq_free_request(struct request *rq) |
320ae51f | 728 | { |
320ae51f | 729 | struct request_queue *q = rq->q; |
ea4f995e | 730 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; |
6af54051 | 731 | |
222ee581 CH |
732 | if ((rq->rq_flags & RQF_ELVPRIV) && |
733 | q->elevator->type->ops.finish_request) | |
734 | q->elevator->type->ops.finish_request(rq); | |
320ae51f | 735 | |
e8064021 | 736 | if (rq->rq_flags & RQF_MQ_INFLIGHT) |
bccf5e26 | 737 | __blk_mq_dec_active_requests(hctx); |
87760e5e | 738 | |
7beb2f84 | 739 | if (unlikely(laptop_mode && !blk_rq_is_passthrough(rq))) |
d152c682 | 740 | laptop_io_completion(q->disk->bdi); |
7beb2f84 | 741 | |
a7905043 | 742 | rq_qos_done(q, rq); |
0d2602ca | 743 | |
12f5b931 | 744 | WRITE_ONCE(rq->state, MQ_RQ_IDLE); |
0a467d0f | 745 | if (req_ref_put_and_test(rq)) |
12f5b931 | 746 | __blk_mq_free_request(rq); |
320ae51f | 747 | } |
1a3b595a | 748 | EXPORT_SYMBOL_GPL(blk_mq_free_request); |
320ae51f | 749 | |
47c122e3 | 750 | void blk_mq_free_plug_rqs(struct blk_plug *plug) |
320ae51f | 751 | { |
013a7f95 | 752 | struct request *rq; |
fe1f4526 | 753 | |
c5fc7b93 | 754 | while ((rq = rq_list_pop(&plug->cached_rq)) != NULL) |
47c122e3 | 755 | blk_mq_free_request(rq); |
47c122e3 | 756 | } |
522a7775 | 757 | |
22350ad7 CH |
758 | void blk_dump_rq_flags(struct request *rq, char *msg) |
759 | { | |
760 | printk(KERN_INFO "%s: dev %s: flags=%llx\n", msg, | |
f3fa33ac | 761 | rq->q->disk ? rq->q->disk->disk_name : "?", |
16458cf3 | 762 | (__force unsigned long long) rq->cmd_flags); |
22350ad7 CH |
763 | |
764 | printk(KERN_INFO " sector %llu, nr/cnr %u/%u\n", | |
765 | (unsigned long long)blk_rq_pos(rq), | |
766 | blk_rq_sectors(rq), blk_rq_cur_sectors(rq)); | |
767 | printk(KERN_INFO " bio %p, biotail %p, len %u\n", | |
768 | rq->bio, rq->biotail, blk_rq_bytes(rq)); | |
769 | } | |
770 | EXPORT_SYMBOL(blk_dump_rq_flags); | |
771 | ||
9be3e06f JA |
772 | static void req_bio_endio(struct request *rq, struct bio *bio, |
773 | unsigned int nbytes, blk_status_t error) | |
774 | { | |
478eb72b | 775 | if (unlikely(error)) { |
9be3e06f | 776 | bio->bi_status = error; |
478eb72b | 777 | } else if (req_op(rq) == REQ_OP_ZONE_APPEND) { |
9be3e06f JA |
778 | /* |
779 | * Partial zone append completions cannot be supported as the | |
780 | * BIO fragments may end up not being written sequentially. | |
781 | */ | |
297db731 | 782 | if (bio->bi_iter.bi_size != nbytes) |
9be3e06f JA |
783 | bio->bi_status = BLK_STS_IOERR; |
784 | else | |
785 | bio->bi_iter.bi_sector = rq->__sector; | |
786 | } | |
787 | ||
478eb72b PB |
788 | bio_advance(bio, nbytes); |
789 | ||
790 | if (unlikely(rq->rq_flags & RQF_QUIET)) | |
791 | bio_set_flag(bio, BIO_QUIET); | |
9be3e06f JA |
792 | /* don't actually finish bio if it's part of flush sequence */ |
793 | if (bio->bi_iter.bi_size == 0 && !(rq->rq_flags & RQF_FLUSH_SEQ)) | |
794 | bio_endio(bio); | |
795 | } | |
796 | ||
797 | static void blk_account_io_completion(struct request *req, unsigned int bytes) | |
798 | { | |
799 | if (req->part && blk_do_io_stat(req)) { | |
800 | const int sgrp = op_stat_group(req_op(req)); | |
801 | ||
802 | part_stat_lock(); | |
803 | part_stat_add(req->part, sectors[sgrp], bytes >> 9); | |
804 | part_stat_unlock(); | |
805 | } | |
806 | } | |
807 | ||
0d7a29a2 CH |
808 | static void blk_print_req_error(struct request *req, blk_status_t status) |
809 | { | |
810 | printk_ratelimited(KERN_ERR | |
811 | "%s error, dev %s, sector %llu op 0x%x:(%s) flags 0x%x " | |
812 | "phys_seg %u prio class %u\n", | |
813 | blk_status_to_str(status), | |
f3fa33ac | 814 | req->q->disk ? req->q->disk->disk_name : "?", |
16458cf3 BVA |
815 | blk_rq_pos(req), (__force u32)req_op(req), |
816 | blk_op_str(req_op(req)), | |
817 | (__force u32)(req->cmd_flags & ~REQ_OP_MASK), | |
0d7a29a2 CH |
818 | req->nr_phys_segments, |
819 | IOPRIO_PRIO_CLASS(req->ioprio)); | |
820 | } | |
821 | ||
5581a5dd JA |
822 | /* |
823 | * Fully end IO on a request. Does not support partial completions, or | |
824 | * errors. | |
825 | */ | |
826 | static void blk_complete_request(struct request *req) | |
827 | { | |
828 | const bool is_flush = (req->rq_flags & RQF_FLUSH_SEQ) != 0; | |
829 | int total_bytes = blk_rq_bytes(req); | |
830 | struct bio *bio = req->bio; | |
831 | ||
832 | trace_block_rq_complete(req, BLK_STS_OK, total_bytes); | |
833 | ||
834 | if (!bio) | |
835 | return; | |
836 | ||
837 | #ifdef CONFIG_BLK_DEV_INTEGRITY | |
838 | if (blk_integrity_rq(req) && req_op(req) == REQ_OP_READ) | |
839 | req->q->integrity.profile->complete_fn(req, total_bytes); | |
840 | #endif | |
841 | ||
842 | blk_account_io_completion(req, total_bytes); | |
843 | ||
844 | do { | |
845 | struct bio *next = bio->bi_next; | |
846 | ||
847 | /* Completion has already been traced */ | |
848 | bio_clear_flag(bio, BIO_TRACE_COMPLETION); | |
a12821d5 PR |
849 | |
850 | if (req_op(req) == REQ_OP_ZONE_APPEND) | |
851 | bio->bi_iter.bi_sector = req->__sector; | |
852 | ||
5581a5dd JA |
853 | if (!is_flush) |
854 | bio_endio(bio); | |
855 | bio = next; | |
856 | } while (bio); | |
857 | ||
858 | /* | |
859 | * Reset counters so that the request stacking driver | |
860 | * can find how many bytes remain in the request | |
861 | * later. | |
862 | */ | |
ab3e1d3b JA |
863 | if (!req->end_io) { |
864 | req->bio = NULL; | |
865 | req->__data_len = 0; | |
866 | } | |
5581a5dd JA |
867 | } |
868 | ||
9be3e06f JA |
869 | /** |
870 | * blk_update_request - Complete multiple bytes without completing the request | |
871 | * @req: the request being processed | |
872 | * @error: block status code | |
873 | * @nr_bytes: number of bytes to complete for @req | |
874 | * | |
875 | * Description: | |
876 | * Ends I/O on a number of bytes attached to @req, but doesn't complete | |
877 | * the request structure even if @req doesn't have leftover. | |
878 | * If @req has leftover, sets it up for the next range of segments. | |
879 | * | |
880 | * Passing the result of blk_rq_bytes() as @nr_bytes guarantees | |
881 | * %false return from this function. | |
882 | * | |
883 | * Note: | |
884 | * The RQF_SPECIAL_PAYLOAD flag is ignored on purpose in this function | |
885 | * except in the consistency check at the end of this function. | |
886 | * | |
887 | * Return: | |
888 | * %false - this request doesn't have any more data | |
889 | * %true - this request has more data | |
890 | **/ | |
891 | bool blk_update_request(struct request *req, blk_status_t error, | |
892 | unsigned int nr_bytes) | |
893 | { | |
894 | int total_bytes; | |
895 | ||
8a7d267b | 896 | trace_block_rq_complete(req, error, nr_bytes); |
9be3e06f JA |
897 | |
898 | if (!req->bio) | |
899 | return false; | |
900 | ||
901 | #ifdef CONFIG_BLK_DEV_INTEGRITY | |
902 | if (blk_integrity_rq(req) && req_op(req) == REQ_OP_READ && | |
903 | error == BLK_STS_OK) | |
904 | req->q->integrity.profile->complete_fn(req, nr_bytes); | |
905 | #endif | |
906 | ||
907 | if (unlikely(error && !blk_rq_is_passthrough(req) && | |
3d973a76 CH |
908 | !(req->rq_flags & RQF_QUIET)) && |
909 | !test_bit(GD_DEAD, &req->q->disk->state)) { | |
9be3e06f | 910 | blk_print_req_error(req, error); |
d5869fdc YS |
911 | trace_block_rq_error(req, error, nr_bytes); |
912 | } | |
9be3e06f JA |
913 | |
914 | blk_account_io_completion(req, nr_bytes); | |
915 | ||
916 | total_bytes = 0; | |
917 | while (req->bio) { | |
918 | struct bio *bio = req->bio; | |
919 | unsigned bio_bytes = min(bio->bi_iter.bi_size, nr_bytes); | |
920 | ||
921 | if (bio_bytes == bio->bi_iter.bi_size) | |
922 | req->bio = bio->bi_next; | |
923 | ||
924 | /* Completion has already been traced */ | |
925 | bio_clear_flag(bio, BIO_TRACE_COMPLETION); | |
926 | req_bio_endio(req, bio, bio_bytes, error); | |
927 | ||
928 | total_bytes += bio_bytes; | |
929 | nr_bytes -= bio_bytes; | |
930 | ||
931 | if (!nr_bytes) | |
932 | break; | |
933 | } | |
934 | ||
935 | /* | |
936 | * completely done | |
937 | */ | |
938 | if (!req->bio) { | |
939 | /* | |
940 | * Reset counters so that the request stacking driver | |
941 | * can find how many bytes remain in the request | |
942 | * later. | |
943 | */ | |
944 | req->__data_len = 0; | |
945 | return false; | |
946 | } | |
947 | ||
948 | req->__data_len -= total_bytes; | |
949 | ||
950 | /* update sector only for requests with clear definition of sector */ | |
951 | if (!blk_rq_is_passthrough(req)) | |
952 | req->__sector += total_bytes >> 9; | |
953 | ||
954 | /* mixed attributes always follow the first bio */ | |
955 | if (req->rq_flags & RQF_MIXED_MERGE) { | |
956 | req->cmd_flags &= ~REQ_FAILFAST_MASK; | |
957 | req->cmd_flags |= req->bio->bi_opf & REQ_FAILFAST_MASK; | |
958 | } | |
959 | ||
960 | if (!(req->rq_flags & RQF_SPECIAL_PAYLOAD)) { | |
961 | /* | |
962 | * If total number of sectors is less than the first segment | |
963 | * size, something has gone terribly wrong. | |
964 | */ | |
965 | if (blk_rq_bytes(req) < blk_rq_cur_bytes(req)) { | |
966 | blk_dump_rq_flags(req, "request botched"); | |
967 | req->__data_len = blk_rq_cur_bytes(req); | |
968 | } | |
969 | ||
970 | /* recalculate the number of segments */ | |
971 | req->nr_phys_segments = blk_recalc_rq_segments(req); | |
972 | } | |
973 | ||
974 | return true; | |
975 | } | |
976 | EXPORT_SYMBOL_GPL(blk_update_request); | |
977 | ||
450b7879 CH |
978 | static void __blk_account_io_done(struct request *req, u64 now) |
979 | { | |
980 | const int sgrp = op_stat_group(req_op(req)); | |
981 | ||
982 | part_stat_lock(); | |
983 | update_io_ticks(req->part, jiffies, true); | |
984 | part_stat_inc(req->part, ios[sgrp]); | |
985 | part_stat_add(req->part, nsecs[sgrp], now - req->start_time_ns); | |
986 | part_stat_unlock(); | |
987 | } | |
988 | ||
989 | static inline void blk_account_io_done(struct request *req, u64 now) | |
990 | { | |
991 | /* | |
992 | * Account IO completion. flush_rq isn't accounted as a | |
993 | * normal IO on queueing nor completion. Accounting the | |
994 | * containing request is enough. | |
995 | */ | |
996 | if (blk_do_io_stat(req) && req->part && | |
997 | !(req->rq_flags & RQF_FLUSH_SEQ)) | |
998 | __blk_account_io_done(req, now); | |
999 | } | |
1000 | ||
1001 | static void __blk_account_io_start(struct request *rq) | |
1002 | { | |
41fa7222 CH |
1003 | /* |
1004 | * All non-passthrough requests are created from a bio with one | |
1005 | * exception: when a flush command that is part of a flush sequence | |
1006 | * generated by the state machine in blk-flush.c is cloned onto the | |
1007 | * lower device by dm-multipath we can get here without a bio. | |
1008 | */ | |
1009 | if (rq->bio) | |
450b7879 | 1010 | rq->part = rq->bio->bi_bdev; |
41fa7222 | 1011 | else |
f3fa33ac | 1012 | rq->part = rq->q->disk->part0; |
450b7879 CH |
1013 | |
1014 | part_stat_lock(); | |
1015 | update_io_ticks(rq->part, jiffies, false); | |
1016 | part_stat_unlock(); | |
1017 | } | |
1018 | ||
1019 | static inline void blk_account_io_start(struct request *req) | |
1020 | { | |
1021 | if (blk_do_io_stat(req)) | |
1022 | __blk_account_io_start(req); | |
1023 | } | |
1024 | ||
f794f335 | 1025 | static inline void __blk_mq_end_request_acct(struct request *rq, u64 now) |
320ae51f | 1026 | { |
4bc6339a OS |
1027 | if (rq->rq_flags & RQF_STATS) { |
1028 | blk_mq_poll_stats_start(rq->q); | |
522a7775 | 1029 | blk_stat_add(rq, now); |
4bc6339a OS |
1030 | } |
1031 | ||
87890092 | 1032 | blk_mq_sched_completed_request(rq, now); |
522a7775 | 1033 | blk_account_io_done(rq, now); |
f794f335 | 1034 | } |
522a7775 | 1035 | |
f794f335 JA |
1036 | inline void __blk_mq_end_request(struct request *rq, blk_status_t error) |
1037 | { | |
1038 | if (blk_mq_need_time_stamp(rq)) | |
1039 | __blk_mq_end_request_acct(rq, ktime_get_ns()); | |
0d11e6ac | 1040 | |
91b63639 | 1041 | if (rq->end_io) { |
a7905043 | 1042 | rq_qos_done(rq->q, rq); |
de671d61 JA |
1043 | if (rq->end_io(rq, error) == RQ_END_IO_FREE) |
1044 | blk_mq_free_request(rq); | |
91b63639 | 1045 | } else { |
320ae51f | 1046 | blk_mq_free_request(rq); |
91b63639 | 1047 | } |
320ae51f | 1048 | } |
c8a446ad | 1049 | EXPORT_SYMBOL(__blk_mq_end_request); |
63151a44 | 1050 | |
2a842aca | 1051 | void blk_mq_end_request(struct request *rq, blk_status_t error) |
63151a44 CH |
1052 | { |
1053 | if (blk_update_request(rq, error, blk_rq_bytes(rq))) | |
1054 | BUG(); | |
c8a446ad | 1055 | __blk_mq_end_request(rq, error); |
63151a44 | 1056 | } |
c8a446ad | 1057 | EXPORT_SYMBOL(blk_mq_end_request); |
320ae51f | 1058 | |
f794f335 JA |
1059 | #define TAG_COMP_BATCH 32 |
1060 | ||
1061 | static inline void blk_mq_flush_tag_batch(struct blk_mq_hw_ctx *hctx, | |
1062 | int *tag_array, int nr_tags) | |
1063 | { | |
1064 | struct request_queue *q = hctx->queue; | |
1065 | ||
3b87c6ea ML |
1066 | /* |
1067 | * All requests should have been marked as RQF_MQ_INFLIGHT, so | |
1068 | * update hctx->nr_active in batch | |
1069 | */ | |
1070 | if (hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) | |
1071 | __blk_mq_sub_active_requests(hctx, nr_tags); | |
1072 | ||
f794f335 JA |
1073 | blk_mq_put_tags(hctx->tags, tag_array, nr_tags); |
1074 | percpu_ref_put_many(&q->q_usage_counter, nr_tags); | |
1075 | } | |
1076 | ||
1077 | void blk_mq_end_request_batch(struct io_comp_batch *iob) | |
1078 | { | |
1079 | int tags[TAG_COMP_BATCH], nr_tags = 0; | |
02f7eab0 | 1080 | struct blk_mq_hw_ctx *cur_hctx = NULL; |
f794f335 JA |
1081 | struct request *rq; |
1082 | u64 now = 0; | |
1083 | ||
1084 | if (iob->need_ts) | |
1085 | now = ktime_get_ns(); | |
1086 | ||
1087 | while ((rq = rq_list_pop(&iob->req_list)) != NULL) { | |
1088 | prefetch(rq->bio); | |
1089 | prefetch(rq->rq_next); | |
1090 | ||
5581a5dd | 1091 | blk_complete_request(rq); |
f794f335 JA |
1092 | if (iob->need_ts) |
1093 | __blk_mq_end_request_acct(rq, now); | |
1094 | ||
98b26a0e JA |
1095 | rq_qos_done(rq->q, rq); |
1096 | ||
ab3e1d3b JA |
1097 | /* |
1098 | * If end_io handler returns NONE, then it still has | |
1099 | * ownership of the request. | |
1100 | */ | |
1101 | if (rq->end_io && rq->end_io(rq, 0) == RQ_END_IO_NONE) | |
1102 | continue; | |
1103 | ||
f794f335 | 1104 | WRITE_ONCE(rq->state, MQ_RQ_IDLE); |
0a467d0f | 1105 | if (!req_ref_put_and_test(rq)) |
f794f335 JA |
1106 | continue; |
1107 | ||
1108 | blk_crypto_free_request(rq); | |
1109 | blk_pm_mark_last_busy(rq); | |
f794f335 | 1110 | |
02f7eab0 JA |
1111 | if (nr_tags == TAG_COMP_BATCH || cur_hctx != rq->mq_hctx) { |
1112 | if (cur_hctx) | |
1113 | blk_mq_flush_tag_batch(cur_hctx, tags, nr_tags); | |
f794f335 | 1114 | nr_tags = 0; |
02f7eab0 | 1115 | cur_hctx = rq->mq_hctx; |
f794f335 JA |
1116 | } |
1117 | tags[nr_tags++] = rq->tag; | |
f794f335 JA |
1118 | } |
1119 | ||
1120 | if (nr_tags) | |
02f7eab0 | 1121 | blk_mq_flush_tag_batch(cur_hctx, tags, nr_tags); |
f794f335 JA |
1122 | } |
1123 | EXPORT_SYMBOL_GPL(blk_mq_end_request_batch); | |
1124 | ||
f9ab4918 | 1125 | static void blk_complete_reqs(struct llist_head *list) |
320ae51f | 1126 | { |
f9ab4918 SAS |
1127 | struct llist_node *entry = llist_reverse_order(llist_del_all(list)); |
1128 | struct request *rq, *next; | |
c3077b5d | 1129 | |
f9ab4918 | 1130 | llist_for_each_entry_safe(rq, next, entry, ipi_list) |
c3077b5d | 1131 | rq->q->mq_ops->complete(rq); |
320ae51f | 1132 | } |
320ae51f | 1133 | |
f9ab4918 | 1134 | static __latent_entropy void blk_done_softirq(struct softirq_action *h) |
320ae51f | 1135 | { |
f9ab4918 | 1136 | blk_complete_reqs(this_cpu_ptr(&blk_cpu_done)); |
115243f5 CH |
1137 | } |
1138 | ||
c3077b5d CH |
1139 | static int blk_softirq_cpu_dead(unsigned int cpu) |
1140 | { | |
f9ab4918 | 1141 | blk_complete_reqs(&per_cpu(blk_cpu_done, cpu)); |
c3077b5d CH |
1142 | return 0; |
1143 | } | |
1144 | ||
40d09b53 | 1145 | static void __blk_mq_complete_request_remote(void *data) |
c3077b5d | 1146 | { |
f9ab4918 | 1147 | __raise_softirq_irqoff(BLOCK_SOFTIRQ); |
c3077b5d CH |
1148 | } |
1149 | ||
96339526 CH |
1150 | static inline bool blk_mq_complete_need_ipi(struct request *rq) |
1151 | { | |
1152 | int cpu = raw_smp_processor_id(); | |
1153 | ||
1154 | if (!IS_ENABLED(CONFIG_SMP) || | |
1155 | !test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) | |
1156 | return false; | |
71425189 SAS |
1157 | /* |
1158 | * With force threaded interrupts enabled, raising softirq from an SMP | |
1159 | * function call will always result in waking the ksoftirqd thread. | |
1160 | * This is probably worse than completing the request on a different | |
1161 | * cache domain. | |
1162 | */ | |
91cc470e | 1163 | if (force_irqthreads()) |
71425189 | 1164 | return false; |
96339526 CH |
1165 | |
1166 | /* same CPU or cache domain? Complete locally */ | |
1167 | if (cpu == rq->mq_ctx->cpu || | |
1168 | (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags) && | |
1169 | cpus_share_cache(cpu, rq->mq_ctx->cpu))) | |
1170 | return false; | |
1171 | ||
1172 | /* don't try to IPI to an offline CPU */ | |
1173 | return cpu_online(rq->mq_ctx->cpu); | |
1174 | } | |
1175 | ||
f9ab4918 SAS |
1176 | static void blk_mq_complete_send_ipi(struct request *rq) |
1177 | { | |
1178 | struct llist_head *list; | |
1179 | unsigned int cpu; | |
1180 | ||
1181 | cpu = rq->mq_ctx->cpu; | |
1182 | list = &per_cpu(blk_cpu_done, cpu); | |
1183 | if (llist_add(&rq->ipi_list, list)) { | |
1184 | INIT_CSD(&rq->csd, __blk_mq_complete_request_remote, rq); | |
1185 | smp_call_function_single_async(cpu, &rq->csd); | |
1186 | } | |
1187 | } | |
1188 | ||
1189 | static void blk_mq_raise_softirq(struct request *rq) | |
1190 | { | |
1191 | struct llist_head *list; | |
1192 | ||
1193 | preempt_disable(); | |
1194 | list = this_cpu_ptr(&blk_cpu_done); | |
1195 | if (llist_add(&rq->ipi_list, list)) | |
1196 | raise_softirq(BLOCK_SOFTIRQ); | |
1197 | preempt_enable(); | |
1198 | } | |
1199 | ||
40d09b53 | 1200 | bool blk_mq_complete_request_remote(struct request *rq) |
320ae51f | 1201 | { |
af78ff7c | 1202 | WRITE_ONCE(rq->state, MQ_RQ_COMPLETE); |
36e76539 | 1203 | |
4ab32bf3 | 1204 | /* |
f168420c LS |
1205 | * For request which hctx has only one ctx mapping, |
1206 | * or a polled request, always complete locally, | |
1207 | * it's pointless to redirect the completion. | |
4ab32bf3 | 1208 | */ |
f168420c LS |
1209 | if (rq->mq_hctx->nr_ctx == 1 || |
1210 | rq->cmd_flags & REQ_POLLED) | |
40d09b53 | 1211 | return false; |
38535201 | 1212 | |
96339526 | 1213 | if (blk_mq_complete_need_ipi(rq)) { |
f9ab4918 SAS |
1214 | blk_mq_complete_send_ipi(rq); |
1215 | return true; | |
3d6efbf6 | 1216 | } |
40d09b53 | 1217 | |
f9ab4918 SAS |
1218 | if (rq->q->nr_hw_queues == 1) { |
1219 | blk_mq_raise_softirq(rq); | |
1220 | return true; | |
1221 | } | |
1222 | return false; | |
40d09b53 CH |
1223 | } |
1224 | EXPORT_SYMBOL_GPL(blk_mq_complete_request_remote); | |
1225 | ||
1226 | /** | |
1227 | * blk_mq_complete_request - end I/O on a request | |
1228 | * @rq: the request being processed | |
1229 | * | |
1230 | * Description: | |
1231 | * Complete a request by scheduling the ->complete_rq operation. | |
1232 | **/ | |
1233 | void blk_mq_complete_request(struct request *rq) | |
1234 | { | |
1235 | if (!blk_mq_complete_request_remote(rq)) | |
1236 | rq->q->mq_ops->complete(rq); | |
320ae51f | 1237 | } |
15f73f5b | 1238 | EXPORT_SYMBOL(blk_mq_complete_request); |
30a91cb4 | 1239 | |
105663f7 AA |
1240 | /** |
1241 | * blk_mq_start_request - Start processing a request | |
1242 | * @rq: Pointer to request to be started | |
1243 | * | |
1244 | * Function used by device drivers to notify the block layer that a request | |
1245 | * is going to be processed now, so blk layer can do proper initializations | |
1246 | * such as starting the timeout timer. | |
1247 | */ | |
e2490073 | 1248 | void blk_mq_start_request(struct request *rq) |
320ae51f JA |
1249 | { |
1250 | struct request_queue *q = rq->q; | |
1251 | ||
a54895fa | 1252 | trace_block_rq_issue(rq); |
320ae51f | 1253 | |
cf43e6be | 1254 | if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) { |
4cddeaca | 1255 | rq->io_start_time_ns = ktime_get_ns(); |
3d244306 | 1256 | rq->stats_sectors = blk_rq_sectors(rq); |
cf43e6be | 1257 | rq->rq_flags |= RQF_STATS; |
a7905043 | 1258 | rq_qos_issue(q, rq); |
cf43e6be JA |
1259 | } |
1260 | ||
1d9bd516 | 1261 | WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE); |
538b7534 | 1262 | |
1d9bd516 | 1263 | blk_add_timer(rq); |
12f5b931 | 1264 | WRITE_ONCE(rq->state, MQ_RQ_IN_FLIGHT); |
49f5baa5 | 1265 | |
54d4e6ab MG |
1266 | #ifdef CONFIG_BLK_DEV_INTEGRITY |
1267 | if (blk_integrity_rq(rq) && req_op(rq) == REQ_OP_WRITE) | |
1268 | q->integrity.profile->prepare_fn(rq); | |
1269 | #endif | |
3e08773c CH |
1270 | if (rq->bio && rq->bio->bi_opf & REQ_POLLED) |
1271 | WRITE_ONCE(rq->bio->bi_cookie, blk_rq_to_qc(rq)); | |
320ae51f | 1272 | } |
e2490073 | 1273 | EXPORT_SYMBOL(blk_mq_start_request); |
320ae51f | 1274 | |
a327c341 ML |
1275 | /* |
1276 | * Allow 2x BLK_MAX_REQUEST_COUNT requests on plug queue for multiple | |
1277 | * queues. This is important for md arrays to benefit from merging | |
1278 | * requests. | |
1279 | */ | |
1280 | static inline unsigned short blk_plug_max_rq_count(struct blk_plug *plug) | |
1281 | { | |
1282 | if (plug->multiple_queues) | |
1283 | return BLK_MAX_REQUEST_COUNT * 2; | |
1284 | return BLK_MAX_REQUEST_COUNT; | |
1285 | } | |
1286 | ||
1287 | static void blk_add_rq_to_plug(struct blk_plug *plug, struct request *rq) | |
1288 | { | |
1289 | struct request *last = rq_list_peek(&plug->mq_list); | |
1290 | ||
1291 | if (!plug->rq_count) { | |
1292 | trace_block_plug(rq->q); | |
1293 | } else if (plug->rq_count >= blk_plug_max_rq_count(plug) || | |
1294 | (!blk_queue_nomerges(rq->q) && | |
1295 | blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) { | |
1296 | blk_mq_flush_plug_list(plug, false); | |
878eb6e4 | 1297 | last = NULL; |
a327c341 ML |
1298 | trace_block_plug(rq->q); |
1299 | } | |
1300 | ||
1301 | if (!plug->multiple_queues && last && last->q != rq->q) | |
1302 | plug->multiple_queues = true; | |
1303 | if (!plug->has_elevator && (rq->rq_flags & RQF_ELV)) | |
1304 | plug->has_elevator = true; | |
1305 | rq->rq_next = NULL; | |
1306 | rq_list_add(&plug->mq_list, rq); | |
1307 | plug->rq_count++; | |
1308 | } | |
1309 | ||
4054cff9 CH |
1310 | /** |
1311 | * blk_execute_rq_nowait - insert a request to I/O scheduler for execution | |
4054cff9 CH |
1312 | * @rq: request to insert |
1313 | * @at_head: insert request at head or tail of queue | |
4054cff9 CH |
1314 | * |
1315 | * Description: | |
1316 | * Insert a fully prepared request at the back of the I/O scheduler queue | |
1317 | * for execution. Don't wait for completion. | |
1318 | * | |
1319 | * Note: | |
1320 | * This function will invoke @done directly if the queue is dead. | |
1321 | */ | |
e2e53086 | 1322 | void blk_execute_rq_nowait(struct request *rq, bool at_head) |
4054cff9 | 1323 | { |
ae948fd6 CH |
1324 | WARN_ON(irqs_disabled()); |
1325 | WARN_ON(!blk_rq_is_passthrough(rq)); | |
4054cff9 | 1326 | |
ae948fd6 | 1327 | blk_account_io_start(rq); |
110fdb44 PR |
1328 | |
1329 | /* | |
1330 | * As plugging can be enabled for passthrough requests on a zoned | |
1331 | * device, directly accessing the plug instead of using blk_mq_plug() | |
1332 | * should not have any consequences. | |
1333 | */ | |
ae948fd6 CH |
1334 | if (current->plug) |
1335 | blk_add_rq_to_plug(current->plug, rq); | |
1336 | else | |
1337 | blk_mq_sched_insert_request(rq, at_head, true, false); | |
4054cff9 CH |
1338 | } |
1339 | EXPORT_SYMBOL_GPL(blk_execute_rq_nowait); | |
1340 | ||
32ac5a9b CH |
1341 | struct blk_rq_wait { |
1342 | struct completion done; | |
1343 | blk_status_t ret; | |
1344 | }; | |
1345 | ||
de671d61 | 1346 | static enum rq_end_io_ret blk_end_sync_rq(struct request *rq, blk_status_t ret) |
32ac5a9b CH |
1347 | { |
1348 | struct blk_rq_wait *wait = rq->end_io_data; | |
1349 | ||
1350 | wait->ret = ret; | |
1351 | complete(&wait->done); | |
de671d61 | 1352 | return RQ_END_IO_NONE; |
32ac5a9b CH |
1353 | } |
1354 | ||
c6e99ea4 | 1355 | bool blk_rq_is_poll(struct request *rq) |
4054cff9 CH |
1356 | { |
1357 | if (!rq->mq_hctx) | |
1358 | return false; | |
1359 | if (rq->mq_hctx->type != HCTX_TYPE_POLL) | |
1360 | return false; | |
1361 | if (WARN_ON_ONCE(!rq->bio)) | |
1362 | return false; | |
1363 | return true; | |
1364 | } | |
c6e99ea4 | 1365 | EXPORT_SYMBOL_GPL(blk_rq_is_poll); |
4054cff9 CH |
1366 | |
1367 | static void blk_rq_poll_completion(struct request *rq, struct completion *wait) | |
1368 | { | |
1369 | do { | |
1370 | bio_poll(rq->bio, NULL, 0); | |
1371 | cond_resched(); | |
1372 | } while (!completion_done(wait)); | |
1373 | } | |
1374 | ||
1375 | /** | |
1376 | * blk_execute_rq - insert a request into queue for execution | |
4054cff9 CH |
1377 | * @rq: request to insert |
1378 | * @at_head: insert request at head or tail of queue | |
1379 | * | |
1380 | * Description: | |
1381 | * Insert a fully prepared request at the back of the I/O scheduler queue | |
1382 | * for execution and wait for completion. | |
1383 | * Return: The blk_status_t result provided to blk_mq_end_request(). | |
1384 | */ | |
b84ba30b | 1385 | blk_status_t blk_execute_rq(struct request *rq, bool at_head) |
4054cff9 | 1386 | { |
32ac5a9b CH |
1387 | struct blk_rq_wait wait = { |
1388 | .done = COMPLETION_INITIALIZER_ONSTACK(wait.done), | |
1389 | }; | |
4054cff9 | 1390 | |
ae948fd6 CH |
1391 | WARN_ON(irqs_disabled()); |
1392 | WARN_ON(!blk_rq_is_passthrough(rq)); | |
4054cff9 CH |
1393 | |
1394 | rq->end_io_data = &wait; | |
ae948fd6 | 1395 | rq->end_io = blk_end_sync_rq; |
4054cff9 | 1396 | |
ae948fd6 CH |
1397 | blk_account_io_start(rq); |
1398 | blk_mq_sched_insert_request(rq, at_head, true, false); | |
4054cff9 | 1399 | |
ae948fd6 | 1400 | if (blk_rq_is_poll(rq)) { |
32ac5a9b | 1401 | blk_rq_poll_completion(rq, &wait.done); |
ae948fd6 CH |
1402 | } else { |
1403 | /* | |
1404 | * Prevent hang_check timer from firing at us during very long | |
1405 | * I/O | |
1406 | */ | |
1407 | unsigned long hang_check = sysctl_hung_task_timeout_secs; | |
1408 | ||
1409 | if (hang_check) | |
32ac5a9b | 1410 | while (!wait_for_completion_io_timeout(&wait.done, |
ae948fd6 CH |
1411 | hang_check * (HZ/2))) |
1412 | ; | |
1413 | else | |
32ac5a9b | 1414 | wait_for_completion_io(&wait.done); |
ae948fd6 | 1415 | } |
4054cff9 | 1416 | |
32ac5a9b | 1417 | return wait.ret; |
4054cff9 CH |
1418 | } |
1419 | EXPORT_SYMBOL(blk_execute_rq); | |
1420 | ||
ed0791b2 | 1421 | static void __blk_mq_requeue_request(struct request *rq) |
320ae51f JA |
1422 | { |
1423 | struct request_queue *q = rq->q; | |
1424 | ||
923218f6 ML |
1425 | blk_mq_put_driver_tag(rq); |
1426 | ||
a54895fa | 1427 | trace_block_rq_requeue(rq); |
a7905043 | 1428 | rq_qos_requeue(q, rq); |
49f5baa5 | 1429 | |
12f5b931 KB |
1430 | if (blk_mq_request_started(rq)) { |
1431 | WRITE_ONCE(rq->state, MQ_RQ_IDLE); | |
da661267 | 1432 | rq->rq_flags &= ~RQF_TIMED_OUT; |
e2490073 | 1433 | } |
320ae51f JA |
1434 | } |
1435 | ||
2b053aca | 1436 | void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list) |
ed0791b2 | 1437 | { |
ed0791b2 | 1438 | __blk_mq_requeue_request(rq); |
ed0791b2 | 1439 | |
105976f5 ML |
1440 | /* this request will be re-inserted to io scheduler queue */ |
1441 | blk_mq_sched_requeue_request(rq); | |
1442 | ||
2b053aca | 1443 | blk_mq_add_to_requeue_list(rq, true, kick_requeue_list); |
ed0791b2 CH |
1444 | } |
1445 | EXPORT_SYMBOL(blk_mq_requeue_request); | |
1446 | ||
6fca6a61 CH |
1447 | static void blk_mq_requeue_work(struct work_struct *work) |
1448 | { | |
1449 | struct request_queue *q = | |
2849450a | 1450 | container_of(work, struct request_queue, requeue_work.work); |
6fca6a61 CH |
1451 | LIST_HEAD(rq_list); |
1452 | struct request *rq, *next; | |
6fca6a61 | 1453 | |
18e9781d | 1454 | spin_lock_irq(&q->requeue_lock); |
6fca6a61 | 1455 | list_splice_init(&q->requeue_list, &rq_list); |
18e9781d | 1456 | spin_unlock_irq(&q->requeue_lock); |
6fca6a61 CH |
1457 | |
1458 | list_for_each_entry_safe(rq, next, &rq_list, queuelist) { | |
aef1897c | 1459 | if (!(rq->rq_flags & (RQF_SOFTBARRIER | RQF_DONTPREP))) |
6fca6a61 CH |
1460 | continue; |
1461 | ||
e8064021 | 1462 | rq->rq_flags &= ~RQF_SOFTBARRIER; |
6fca6a61 | 1463 | list_del_init(&rq->queuelist); |
aef1897c JW |
1464 | /* |
1465 | * If RQF_DONTPREP, rq has contained some driver specific | |
1466 | * data, so insert it to hctx dispatch list to avoid any | |
1467 | * merge. | |
1468 | */ | |
1469 | if (rq->rq_flags & RQF_DONTPREP) | |
01e99aec | 1470 | blk_mq_request_bypass_insert(rq, false, false); |
aef1897c JW |
1471 | else |
1472 | blk_mq_sched_insert_request(rq, true, false, false); | |
6fca6a61 CH |
1473 | } |
1474 | ||
1475 | while (!list_empty(&rq_list)) { | |
1476 | rq = list_entry(rq_list.next, struct request, queuelist); | |
1477 | list_del_init(&rq->queuelist); | |
9e97d295 | 1478 | blk_mq_sched_insert_request(rq, false, false, false); |
6fca6a61 CH |
1479 | } |
1480 | ||
52d7f1b5 | 1481 | blk_mq_run_hw_queues(q, false); |
6fca6a61 CH |
1482 | } |
1483 | ||
2b053aca BVA |
1484 | void blk_mq_add_to_requeue_list(struct request *rq, bool at_head, |
1485 | bool kick_requeue_list) | |
6fca6a61 CH |
1486 | { |
1487 | struct request_queue *q = rq->q; | |
1488 | unsigned long flags; | |
1489 | ||
1490 | /* | |
1491 | * We abuse this flag that is otherwise used by the I/O scheduler to | |
ff821d27 | 1492 | * request head insertion from the workqueue. |
6fca6a61 | 1493 | */ |
e8064021 | 1494 | BUG_ON(rq->rq_flags & RQF_SOFTBARRIER); |
6fca6a61 CH |
1495 | |
1496 | spin_lock_irqsave(&q->requeue_lock, flags); | |
1497 | if (at_head) { | |
e8064021 | 1498 | rq->rq_flags |= RQF_SOFTBARRIER; |
6fca6a61 CH |
1499 | list_add(&rq->queuelist, &q->requeue_list); |
1500 | } else { | |
1501 | list_add_tail(&rq->queuelist, &q->requeue_list); | |
1502 | } | |
1503 | spin_unlock_irqrestore(&q->requeue_lock, flags); | |
2b053aca BVA |
1504 | |
1505 | if (kick_requeue_list) | |
1506 | blk_mq_kick_requeue_list(q); | |
6fca6a61 | 1507 | } |
6fca6a61 CH |
1508 | |
1509 | void blk_mq_kick_requeue_list(struct request_queue *q) | |
1510 | { | |
ae943d20 | 1511 | kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0); |
6fca6a61 CH |
1512 | } |
1513 | EXPORT_SYMBOL(blk_mq_kick_requeue_list); | |
1514 | ||
2849450a MS |
1515 | void blk_mq_delay_kick_requeue_list(struct request_queue *q, |
1516 | unsigned long msecs) | |
1517 | { | |
d4acf365 BVA |
1518 | kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, |
1519 | msecs_to_jiffies(msecs)); | |
2849450a MS |
1520 | } |
1521 | EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list); | |
1522 | ||
2dd6532e | 1523 | static bool blk_mq_rq_inflight(struct request *rq, void *priv) |
ae879912 JA |
1524 | { |
1525 | /* | |
8ab30a33 JG |
1526 | * If we find a request that isn't idle we know the queue is busy |
1527 | * as it's checked in the iter. | |
1528 | * Return false to stop the iteration. | |
ae879912 | 1529 | */ |
8ab30a33 | 1530 | if (blk_mq_request_started(rq)) { |
ae879912 JA |
1531 | bool *busy = priv; |
1532 | ||
1533 | *busy = true; | |
1534 | return false; | |
1535 | } | |
1536 | ||
1537 | return true; | |
1538 | } | |
1539 | ||
3c94d83c | 1540 | bool blk_mq_queue_inflight(struct request_queue *q) |
ae879912 JA |
1541 | { |
1542 | bool busy = false; | |
1543 | ||
3c94d83c | 1544 | blk_mq_queue_tag_busy_iter(q, blk_mq_rq_inflight, &busy); |
ae879912 JA |
1545 | return busy; |
1546 | } | |
3c94d83c | 1547 | EXPORT_SYMBOL_GPL(blk_mq_queue_inflight); |
ae879912 | 1548 | |
9bdb4833 | 1549 | static void blk_mq_rq_timed_out(struct request *req) |
320ae51f | 1550 | { |
da661267 | 1551 | req->rq_flags |= RQF_TIMED_OUT; |
d1210d5a CH |
1552 | if (req->q->mq_ops->timeout) { |
1553 | enum blk_eh_timer_return ret; | |
1554 | ||
9bdb4833 | 1555 | ret = req->q->mq_ops->timeout(req); |
d1210d5a CH |
1556 | if (ret == BLK_EH_DONE) |
1557 | return; | |
1558 | WARN_ON_ONCE(ret != BLK_EH_RESET_TIMER); | |
46f92d42 | 1559 | } |
d1210d5a CH |
1560 | |
1561 | blk_add_timer(req); | |
87ee7b11 | 1562 | } |
5b3f25fc | 1563 | |
82c22947 DJ |
1564 | struct blk_expired_data { |
1565 | bool has_timedout_rq; | |
1566 | unsigned long next; | |
1567 | unsigned long timeout_start; | |
1568 | }; | |
1569 | ||
1570 | static bool blk_mq_req_expired(struct request *rq, struct blk_expired_data *expired) | |
81481eb4 | 1571 | { |
12f5b931 | 1572 | unsigned long deadline; |
87ee7b11 | 1573 | |
12f5b931 KB |
1574 | if (blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT) |
1575 | return false; | |
da661267 CH |
1576 | if (rq->rq_flags & RQF_TIMED_OUT) |
1577 | return false; | |
a7af0af3 | 1578 | |
079076b3 | 1579 | deadline = READ_ONCE(rq->deadline); |
82c22947 | 1580 | if (time_after_eq(expired->timeout_start, deadline)) |
12f5b931 | 1581 | return true; |
a7af0af3 | 1582 | |
82c22947 DJ |
1583 | if (expired->next == 0) |
1584 | expired->next = deadline; | |
1585 | else if (time_after(expired->next, deadline)) | |
1586 | expired->next = deadline; | |
12f5b931 | 1587 | return false; |
87ee7b11 JA |
1588 | } |
1589 | ||
2e315dc0 ML |
1590 | void blk_mq_put_rq_ref(struct request *rq) |
1591 | { | |
de671d61 JA |
1592 | if (is_flush_rq(rq)) { |
1593 | if (rq->end_io(rq, 0) == RQ_END_IO_FREE) | |
1594 | blk_mq_free_request(rq); | |
1595 | } else if (req_ref_put_and_test(rq)) { | |
2e315dc0 | 1596 | __blk_mq_free_request(rq); |
de671d61 | 1597 | } |
2e315dc0 ML |
1598 | } |
1599 | ||
2dd6532e | 1600 | static bool blk_mq_check_expired(struct request *rq, void *priv) |
1d9bd516 | 1601 | { |
82c22947 | 1602 | struct blk_expired_data *expired = priv; |
12f5b931 KB |
1603 | |
1604 | /* | |
c797b40c ML |
1605 | * blk_mq_queue_tag_busy_iter() has locked the request, so it cannot |
1606 | * be reallocated underneath the timeout handler's processing, then | |
1607 | * the expire check is reliable. If the request is not expired, then | |
1608 | * it was completed and reallocated as a new request after returning | |
1609 | * from blk_mq_check_expired(). | |
1d9bd516 | 1610 | */ |
82c22947 DJ |
1611 | if (blk_mq_req_expired(rq, expired)) { |
1612 | expired->has_timedout_rq = true; | |
1613 | return false; | |
1614 | } | |
1615 | return true; | |
1616 | } | |
1617 | ||
1618 | static bool blk_mq_handle_expired(struct request *rq, void *priv) | |
1619 | { | |
1620 | struct blk_expired_data *expired = priv; | |
1621 | ||
1622 | if (blk_mq_req_expired(rq, expired)) | |
9bdb4833 | 1623 | blk_mq_rq_timed_out(rq); |
7baa8572 | 1624 | return true; |
1d9bd516 TH |
1625 | } |
1626 | ||
287922eb | 1627 | static void blk_mq_timeout_work(struct work_struct *work) |
320ae51f | 1628 | { |
287922eb CH |
1629 | struct request_queue *q = |
1630 | container_of(work, struct request_queue, timeout_work); | |
82c22947 DJ |
1631 | struct blk_expired_data expired = { |
1632 | .timeout_start = jiffies, | |
1633 | }; | |
1d9bd516 | 1634 | struct blk_mq_hw_ctx *hctx; |
4f481208 | 1635 | unsigned long i; |
320ae51f | 1636 | |
71f79fb3 GKB |
1637 | /* A deadlock might occur if a request is stuck requiring a |
1638 | * timeout at the same time a queue freeze is waiting | |
1639 | * completion, since the timeout code would not be able to | |
1640 | * acquire the queue reference here. | |
1641 | * | |
1642 | * That's why we don't use blk_queue_enter here; instead, we use | |
1643 | * percpu_ref_tryget directly, because we need to be able to | |
1644 | * obtain a reference even in the short window between the queue | |
1645 | * starting to freeze, by dropping the first reference in | |
1671d522 | 1646 | * blk_freeze_queue_start, and the moment the last request is |
71f79fb3 GKB |
1647 | * consumed, marked by the instant q_usage_counter reaches |
1648 | * zero. | |
1649 | */ | |
1650 | if (!percpu_ref_tryget(&q->q_usage_counter)) | |
287922eb CH |
1651 | return; |
1652 | ||
82c22947 DJ |
1653 | /* check if there is any timed-out request */ |
1654 | blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &expired); | |
1655 | if (expired.has_timedout_rq) { | |
1656 | /* | |
1657 | * Before walking tags, we must ensure any submit started | |
1658 | * before the current time has finished. Since the submit | |
1659 | * uses srcu or rcu, wait for a synchronization point to | |
1660 | * ensure all running submits have finished | |
1661 | */ | |
483239c7 | 1662 | blk_mq_wait_quiesce_done(q->tag_set); |
82c22947 DJ |
1663 | |
1664 | expired.next = 0; | |
1665 | blk_mq_queue_tag_busy_iter(q, blk_mq_handle_expired, &expired); | |
1666 | } | |
320ae51f | 1667 | |
82c22947 DJ |
1668 | if (expired.next != 0) { |
1669 | mod_timer(&q->timeout, expired.next); | |
0d2602ca | 1670 | } else { |
fcd36c36 BVA |
1671 | /* |
1672 | * Request timeouts are handled as a forward rolling timer. If | |
1673 | * we end up here it means that no requests are pending and | |
1674 | * also that no request has been pending for a while. Mark | |
1675 | * each hctx as idle. | |
1676 | */ | |
f054b56c ML |
1677 | queue_for_each_hw_ctx(q, hctx, i) { |
1678 | /* the hctx may be unmapped, so check it here */ | |
1679 | if (blk_mq_hw_queue_mapped(hctx)) | |
1680 | blk_mq_tag_idle(hctx); | |
1681 | } | |
0d2602ca | 1682 | } |
287922eb | 1683 | blk_queue_exit(q); |
320ae51f JA |
1684 | } |
1685 | ||
88459642 OS |
1686 | struct flush_busy_ctx_data { |
1687 | struct blk_mq_hw_ctx *hctx; | |
1688 | struct list_head *list; | |
1689 | }; | |
1690 | ||
1691 | static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data) | |
1692 | { | |
1693 | struct flush_busy_ctx_data *flush_data = data; | |
1694 | struct blk_mq_hw_ctx *hctx = flush_data->hctx; | |
1695 | struct blk_mq_ctx *ctx = hctx->ctxs[bitnr]; | |
c16d6b5a | 1696 | enum hctx_type type = hctx->type; |
88459642 | 1697 | |
88459642 | 1698 | spin_lock(&ctx->lock); |
c16d6b5a | 1699 | list_splice_tail_init(&ctx->rq_lists[type], flush_data->list); |
e9a99a63 | 1700 | sbitmap_clear_bit(sb, bitnr); |
88459642 OS |
1701 | spin_unlock(&ctx->lock); |
1702 | return true; | |
1703 | } | |
1704 | ||
1429d7c9 JA |
1705 | /* |
1706 | * Process software queues that have been marked busy, splicing them | |
1707 | * to the for-dispatch | |
1708 | */ | |
2c3ad667 | 1709 | void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list) |
1429d7c9 | 1710 | { |
88459642 OS |
1711 | struct flush_busy_ctx_data data = { |
1712 | .hctx = hctx, | |
1713 | .list = list, | |
1714 | }; | |
1429d7c9 | 1715 | |
88459642 | 1716 | sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data); |
1429d7c9 | 1717 | } |
2c3ad667 | 1718 | EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs); |
1429d7c9 | 1719 | |
b347689f ML |
1720 | struct dispatch_rq_data { |
1721 | struct blk_mq_hw_ctx *hctx; | |
1722 | struct request *rq; | |
1723 | }; | |
1724 | ||
1725 | static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr, | |
1726 | void *data) | |
1727 | { | |
1728 | struct dispatch_rq_data *dispatch_data = data; | |
1729 | struct blk_mq_hw_ctx *hctx = dispatch_data->hctx; | |
1730 | struct blk_mq_ctx *ctx = hctx->ctxs[bitnr]; | |
c16d6b5a | 1731 | enum hctx_type type = hctx->type; |
b347689f ML |
1732 | |
1733 | spin_lock(&ctx->lock); | |
c16d6b5a ML |
1734 | if (!list_empty(&ctx->rq_lists[type])) { |
1735 | dispatch_data->rq = list_entry_rq(ctx->rq_lists[type].next); | |
b347689f | 1736 | list_del_init(&dispatch_data->rq->queuelist); |
c16d6b5a | 1737 | if (list_empty(&ctx->rq_lists[type])) |
b347689f ML |
1738 | sbitmap_clear_bit(sb, bitnr); |
1739 | } | |
1740 | spin_unlock(&ctx->lock); | |
1741 | ||
1742 | return !dispatch_data->rq; | |
1743 | } | |
1744 | ||
1745 | struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx, | |
1746 | struct blk_mq_ctx *start) | |
1747 | { | |
f31967f0 | 1748 | unsigned off = start ? start->index_hw[hctx->type] : 0; |
b347689f ML |
1749 | struct dispatch_rq_data data = { |
1750 | .hctx = hctx, | |
1751 | .rq = NULL, | |
1752 | }; | |
1753 | ||
1754 | __sbitmap_for_each_set(&hctx->ctx_map, off, | |
1755 | dispatch_rq_from_ctx, &data); | |
1756 | ||
1757 | return data.rq; | |
1758 | } | |
1759 | ||
a808a9d5 | 1760 | static bool __blk_mq_alloc_driver_tag(struct request *rq) |
570e9b73 | 1761 | { |
ae0f1a73 | 1762 | struct sbitmap_queue *bt = &rq->mq_hctx->tags->bitmap_tags; |
570e9b73 | 1763 | unsigned int tag_offset = rq->mq_hctx->tags->nr_reserved_tags; |
570e9b73 ML |
1764 | int tag; |
1765 | ||
568f2700 ML |
1766 | blk_mq_tag_busy(rq->mq_hctx); |
1767 | ||
570e9b73 | 1768 | if (blk_mq_tag_is_reserved(rq->mq_hctx->sched_tags, rq->internal_tag)) { |
ae0f1a73 | 1769 | bt = &rq->mq_hctx->tags->breserved_tags; |
570e9b73 | 1770 | tag_offset = 0; |
28500850 ML |
1771 | } else { |
1772 | if (!hctx_may_queue(rq->mq_hctx, bt)) | |
1773 | return false; | |
570e9b73 ML |
1774 | } |
1775 | ||
570e9b73 ML |
1776 | tag = __sbitmap_queue_get(bt); |
1777 | if (tag == BLK_MQ_NO_TAG) | |
1778 | return false; | |
1779 | ||
1780 | rq->tag = tag + tag_offset; | |
570e9b73 ML |
1781 | return true; |
1782 | } | |
1783 | ||
a808a9d5 | 1784 | bool __blk_mq_get_driver_tag(struct blk_mq_hw_ctx *hctx, struct request *rq) |
570e9b73 | 1785 | { |
a808a9d5 | 1786 | if (rq->tag == BLK_MQ_NO_TAG && !__blk_mq_alloc_driver_tag(rq)) |
568f2700 ML |
1787 | return false; |
1788 | ||
51db1c37 | 1789 | if ((hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) && |
568f2700 ML |
1790 | !(rq->rq_flags & RQF_MQ_INFLIGHT)) { |
1791 | rq->rq_flags |= RQF_MQ_INFLIGHT; | |
bccf5e26 | 1792 | __blk_mq_inc_active_requests(hctx); |
568f2700 ML |
1793 | } |
1794 | hctx->tags->rqs[rq->tag] = rq; | |
1795 | return true; | |
570e9b73 ML |
1796 | } |
1797 | ||
eb619fdb JA |
1798 | static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode, |
1799 | int flags, void *key) | |
da55f2cc OS |
1800 | { |
1801 | struct blk_mq_hw_ctx *hctx; | |
1802 | ||
1803 | hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait); | |
1804 | ||
5815839b | 1805 | spin_lock(&hctx->dispatch_wait_lock); |
e8618575 JA |
1806 | if (!list_empty(&wait->entry)) { |
1807 | struct sbitmap_queue *sbq; | |
1808 | ||
1809 | list_del_init(&wait->entry); | |
ae0f1a73 | 1810 | sbq = &hctx->tags->bitmap_tags; |
e8618575 JA |
1811 | atomic_dec(&sbq->ws_active); |
1812 | } | |
5815839b ML |
1813 | spin_unlock(&hctx->dispatch_wait_lock); |
1814 | ||
da55f2cc OS |
1815 | blk_mq_run_hw_queue(hctx, true); |
1816 | return 1; | |
1817 | } | |
1818 | ||
f906a6a0 JA |
1819 | /* |
1820 | * Mark us waiting for a tag. For shared tags, this involves hooking us into | |
ee3e4de5 BVA |
1821 | * the tag wakeups. For non-shared tags, we can simply mark us needing a |
1822 | * restart. For both cases, take care to check the condition again after | |
f906a6a0 JA |
1823 | * marking us as waiting. |
1824 | */ | |
2278d69f | 1825 | static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx *hctx, |
f906a6a0 | 1826 | struct request *rq) |
da55f2cc | 1827 | { |
ae0f1a73 | 1828 | struct sbitmap_queue *sbq = &hctx->tags->bitmap_tags; |
5815839b | 1829 | struct wait_queue_head *wq; |
f906a6a0 JA |
1830 | wait_queue_entry_t *wait; |
1831 | bool ret; | |
da55f2cc | 1832 | |
51db1c37 | 1833 | if (!(hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) { |
684b7324 | 1834 | blk_mq_sched_mark_restart_hctx(hctx); |
f906a6a0 | 1835 | |
c27d53fb BVA |
1836 | /* |
1837 | * It's possible that a tag was freed in the window between the | |
1838 | * allocation failure and adding the hardware queue to the wait | |
1839 | * queue. | |
1840 | * | |
1841 | * Don't clear RESTART here, someone else could have set it. | |
1842 | * At most this will cost an extra queue run. | |
1843 | */ | |
8ab6bb9e | 1844 | return blk_mq_get_driver_tag(rq); |
eb619fdb | 1845 | } |
eb619fdb | 1846 | |
2278d69f | 1847 | wait = &hctx->dispatch_wait; |
c27d53fb BVA |
1848 | if (!list_empty_careful(&wait->entry)) |
1849 | return false; | |
1850 | ||
e8618575 | 1851 | wq = &bt_wait_ptr(sbq, hctx)->wait; |
5815839b ML |
1852 | |
1853 | spin_lock_irq(&wq->lock); | |
1854 | spin_lock(&hctx->dispatch_wait_lock); | |
c27d53fb | 1855 | if (!list_empty(&wait->entry)) { |
5815839b ML |
1856 | spin_unlock(&hctx->dispatch_wait_lock); |
1857 | spin_unlock_irq(&wq->lock); | |
c27d53fb | 1858 | return false; |
eb619fdb JA |
1859 | } |
1860 | ||
e8618575 | 1861 | atomic_inc(&sbq->ws_active); |
5815839b ML |
1862 | wait->flags &= ~WQ_FLAG_EXCLUSIVE; |
1863 | __add_wait_queue(wq, wait); | |
c27d53fb | 1864 | |
da55f2cc | 1865 | /* |
eb619fdb JA |
1866 | * It's possible that a tag was freed in the window between the |
1867 | * allocation failure and adding the hardware queue to the wait | |
1868 | * queue. | |
da55f2cc | 1869 | */ |
8ab6bb9e | 1870 | ret = blk_mq_get_driver_tag(rq); |
c27d53fb | 1871 | if (!ret) { |
5815839b ML |
1872 | spin_unlock(&hctx->dispatch_wait_lock); |
1873 | spin_unlock_irq(&wq->lock); | |
c27d53fb | 1874 | return false; |
eb619fdb | 1875 | } |
c27d53fb BVA |
1876 | |
1877 | /* | |
1878 | * We got a tag, remove ourselves from the wait queue to ensure | |
1879 | * someone else gets the wakeup. | |
1880 | */ | |
c27d53fb | 1881 | list_del_init(&wait->entry); |
e8618575 | 1882 | atomic_dec(&sbq->ws_active); |
5815839b ML |
1883 | spin_unlock(&hctx->dispatch_wait_lock); |
1884 | spin_unlock_irq(&wq->lock); | |
c27d53fb BVA |
1885 | |
1886 | return true; | |
da55f2cc OS |
1887 | } |
1888 | ||
6e768717 ML |
1889 | #define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT 8 |
1890 | #define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR 4 | |
1891 | /* | |
1892 | * Update dispatch busy with the Exponential Weighted Moving Average(EWMA): | |
1893 | * - EWMA is one simple way to compute running average value | |
1894 | * - weight(7/8 and 1/8) is applied so that it can decrease exponentially | |
1895 | * - take 4 as factor for avoiding to get too small(0) result, and this | |
1896 | * factor doesn't matter because EWMA decreases exponentially | |
1897 | */ | |
1898 | static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx *hctx, bool busy) | |
1899 | { | |
1900 | unsigned int ewma; | |
1901 | ||
6e768717 ML |
1902 | ewma = hctx->dispatch_busy; |
1903 | ||
1904 | if (!ewma && !busy) | |
1905 | return; | |
1906 | ||
1907 | ewma *= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT - 1; | |
1908 | if (busy) | |
1909 | ewma += 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR; | |
1910 | ewma /= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT; | |
1911 | ||
1912 | hctx->dispatch_busy = ewma; | |
1913 | } | |
1914 | ||
86ff7c2a ML |
1915 | #define BLK_MQ_RESOURCE_DELAY 3 /* ms units */ |
1916 | ||
c92a4103 JT |
1917 | static void blk_mq_handle_dev_resource(struct request *rq, |
1918 | struct list_head *list) | |
1919 | { | |
1920 | struct request *next = | |
1921 | list_first_entry_or_null(list, struct request, queuelist); | |
1922 | ||
1923 | /* | |
1924 | * If an I/O scheduler has been configured and we got a driver tag for | |
1925 | * the next request already, free it. | |
1926 | */ | |
1927 | if (next) | |
1928 | blk_mq_put_driver_tag(next); | |
1929 | ||
1930 | list_add(&rq->queuelist, list); | |
1931 | __blk_mq_requeue_request(rq); | |
1932 | } | |
1933 | ||
0512a75b KB |
1934 | static void blk_mq_handle_zone_resource(struct request *rq, |
1935 | struct list_head *zone_list) | |
1936 | { | |
1937 | /* | |
1938 | * If we end up here it is because we cannot dispatch a request to a | |
1939 | * specific zone due to LLD level zone-write locking or other zone | |
1940 | * related resource not being available. In this case, set the request | |
1941 | * aside in zone_list for retrying it later. | |
1942 | */ | |
1943 | list_add(&rq->queuelist, zone_list); | |
1944 | __blk_mq_requeue_request(rq); | |
1945 | } | |
1946 | ||
75383524 ML |
1947 | enum prep_dispatch { |
1948 | PREP_DISPATCH_OK, | |
1949 | PREP_DISPATCH_NO_TAG, | |
1950 | PREP_DISPATCH_NO_BUDGET, | |
1951 | }; | |
1952 | ||
1953 | static enum prep_dispatch blk_mq_prep_dispatch_rq(struct request *rq, | |
1954 | bool need_budget) | |
1955 | { | |
1956 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; | |
2a5a24aa | 1957 | int budget_token = -1; |
75383524 | 1958 | |
2a5a24aa ML |
1959 | if (need_budget) { |
1960 | budget_token = blk_mq_get_dispatch_budget(rq->q); | |
1961 | if (budget_token < 0) { | |
1962 | blk_mq_put_driver_tag(rq); | |
1963 | return PREP_DISPATCH_NO_BUDGET; | |
1964 | } | |
1965 | blk_mq_set_rq_budget_token(rq, budget_token); | |
75383524 ML |
1966 | } |
1967 | ||
1968 | if (!blk_mq_get_driver_tag(rq)) { | |
1969 | /* | |
1970 | * The initial allocation attempt failed, so we need to | |
1971 | * rerun the hardware queue when a tag is freed. The | |
1972 | * waitqueue takes care of that. If the queue is run | |
1973 | * before we add this entry back on the dispatch list, | |
1974 | * we'll re-run it below. | |
1975 | */ | |
1976 | if (!blk_mq_mark_tag_wait(hctx, rq)) { | |
1fd40b5e ML |
1977 | /* |
1978 | * All budgets not got from this function will be put | |
1979 | * together during handling partial dispatch | |
1980 | */ | |
1981 | if (need_budget) | |
2a5a24aa | 1982 | blk_mq_put_dispatch_budget(rq->q, budget_token); |
75383524 ML |
1983 | return PREP_DISPATCH_NO_TAG; |
1984 | } | |
1985 | } | |
1986 | ||
1987 | return PREP_DISPATCH_OK; | |
1988 | } | |
1989 | ||
1fd40b5e ML |
1990 | /* release all allocated budgets before calling to blk_mq_dispatch_rq_list */ |
1991 | static void blk_mq_release_budgets(struct request_queue *q, | |
2a5a24aa | 1992 | struct list_head *list) |
1fd40b5e | 1993 | { |
2a5a24aa | 1994 | struct request *rq; |
1fd40b5e | 1995 | |
2a5a24aa ML |
1996 | list_for_each_entry(rq, list, queuelist) { |
1997 | int budget_token = blk_mq_get_rq_budget_token(rq); | |
1fd40b5e | 1998 | |
2a5a24aa ML |
1999 | if (budget_token >= 0) |
2000 | blk_mq_put_dispatch_budget(q, budget_token); | |
2001 | } | |
1fd40b5e ML |
2002 | } |
2003 | ||
1f57f8d4 JA |
2004 | /* |
2005 | * Returns true if we did some work AND can potentially do more. | |
2006 | */ | |
445874e8 | 2007 | bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx *hctx, struct list_head *list, |
1fd40b5e | 2008 | unsigned int nr_budgets) |
320ae51f | 2009 | { |
75383524 | 2010 | enum prep_dispatch prep; |
445874e8 | 2011 | struct request_queue *q = hctx->queue; |
6d6f167c | 2012 | struct request *rq, *nxt; |
fc17b653 | 2013 | int errors, queued; |
86ff7c2a | 2014 | blk_status_t ret = BLK_STS_OK; |
0512a75b | 2015 | LIST_HEAD(zone_list); |
9586e67b | 2016 | bool needs_resource = false; |
320ae51f | 2017 | |
81380ca1 OS |
2018 | if (list_empty(list)) |
2019 | return false; | |
2020 | ||
320ae51f JA |
2021 | /* |
2022 | * Now process all the entries, sending them to the driver. | |
2023 | */ | |
93efe981 | 2024 | errors = queued = 0; |
81380ca1 | 2025 | do { |
74c45052 | 2026 | struct blk_mq_queue_data bd; |
320ae51f | 2027 | |
f04c3df3 | 2028 | rq = list_first_entry(list, struct request, queuelist); |
0bca799b | 2029 | |
445874e8 | 2030 | WARN_ON_ONCE(hctx != rq->mq_hctx); |
1fd40b5e | 2031 | prep = blk_mq_prep_dispatch_rq(rq, !nr_budgets); |
75383524 | 2032 | if (prep != PREP_DISPATCH_OK) |
0bca799b | 2033 | break; |
de148297 | 2034 | |
320ae51f | 2035 | list_del_init(&rq->queuelist); |
320ae51f | 2036 | |
74c45052 | 2037 | bd.rq = rq; |
113285b4 JA |
2038 | |
2039 | /* | |
2040 | * Flag last if we have no more requests, or if we have more | |
2041 | * but can't assign a driver tag to it. | |
2042 | */ | |
2043 | if (list_empty(list)) | |
2044 | bd.last = true; | |
2045 | else { | |
113285b4 | 2046 | nxt = list_first_entry(list, struct request, queuelist); |
8ab6bb9e | 2047 | bd.last = !blk_mq_get_driver_tag(nxt); |
113285b4 | 2048 | } |
74c45052 | 2049 | |
1fd40b5e ML |
2050 | /* |
2051 | * once the request is queued to lld, no need to cover the | |
2052 | * budget any more | |
2053 | */ | |
2054 | if (nr_budgets) | |
2055 | nr_budgets--; | |
74c45052 | 2056 | ret = q->mq_ops->queue_rq(hctx, &bd); |
7bf13729 ML |
2057 | switch (ret) { |
2058 | case BLK_STS_OK: | |
2059 | queued++; | |
320ae51f | 2060 | break; |
7bf13729 | 2061 | case BLK_STS_RESOURCE: |
9586e67b NA |
2062 | needs_resource = true; |
2063 | fallthrough; | |
7bf13729 ML |
2064 | case BLK_STS_DEV_RESOURCE: |
2065 | blk_mq_handle_dev_resource(rq, list); | |
2066 | goto out; | |
2067 | case BLK_STS_ZONE_RESOURCE: | |
0512a75b KB |
2068 | /* |
2069 | * Move the request to zone_list and keep going through | |
2070 | * the dispatch list to find more requests the drive can | |
2071 | * accept. | |
2072 | */ | |
2073 | blk_mq_handle_zone_resource(rq, &zone_list); | |
9586e67b | 2074 | needs_resource = true; |
7bf13729 ML |
2075 | break; |
2076 | default: | |
93efe981 | 2077 | errors++; |
e21ee5a6 | 2078 | blk_mq_end_request(rq, ret); |
320ae51f | 2079 | } |
81380ca1 | 2080 | } while (!list_empty(list)); |
7bf13729 | 2081 | out: |
0512a75b KB |
2082 | if (!list_empty(&zone_list)) |
2083 | list_splice_tail_init(&zone_list, list); | |
2084 | ||
632bfb63 | 2085 | /* If we didn't flush the entire list, we could have told the driver |
2086 | * there was more coming, but that turned out to be a lie. | |
2087 | */ | |
65fac0d5 YK |
2088 | if ((!list_empty(list) || errors || needs_resource || |
2089 | ret == BLK_STS_DEV_RESOURCE) && q->mq_ops->commit_rqs && queued) | |
632bfb63 | 2090 | q->mq_ops->commit_rqs(hctx); |
320ae51f JA |
2091 | /* |
2092 | * Any items that need requeuing? Stuff them into hctx->dispatch, | |
2093 | * that is where we will continue on next queue run. | |
2094 | */ | |
f04c3df3 | 2095 | if (!list_empty(list)) { |
86ff7c2a | 2096 | bool needs_restart; |
75383524 ML |
2097 | /* For non-shared tags, the RESTART check will suffice */ |
2098 | bool no_tag = prep == PREP_DISPATCH_NO_TAG && | |
51db1c37 | 2099 | (hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED); |
86ff7c2a | 2100 | |
2a5a24aa ML |
2101 | if (nr_budgets) |
2102 | blk_mq_release_budgets(q, list); | |
86ff7c2a | 2103 | |
320ae51f | 2104 | spin_lock(&hctx->lock); |
01e99aec | 2105 | list_splice_tail_init(list, &hctx->dispatch); |
320ae51f | 2106 | spin_unlock(&hctx->lock); |
f04c3df3 | 2107 | |
d7d8535f ML |
2108 | /* |
2109 | * Order adding requests to hctx->dispatch and checking | |
2110 | * SCHED_RESTART flag. The pair of this smp_mb() is the one | |
2111 | * in blk_mq_sched_restart(). Avoid restart code path to | |
2112 | * miss the new added requests to hctx->dispatch, meantime | |
2113 | * SCHED_RESTART is observed here. | |
2114 | */ | |
2115 | smp_mb(); | |
2116 | ||
9ba52e58 | 2117 | /* |
710c785f BVA |
2118 | * If SCHED_RESTART was set by the caller of this function and |
2119 | * it is no longer set that means that it was cleared by another | |
2120 | * thread and hence that a queue rerun is needed. | |
9ba52e58 | 2121 | * |
eb619fdb JA |
2122 | * If 'no_tag' is set, that means that we failed getting |
2123 | * a driver tag with an I/O scheduler attached. If our dispatch | |
2124 | * waitqueue is no longer active, ensure that we run the queue | |
2125 | * AFTER adding our entries back to the list. | |
bd166ef1 | 2126 | * |
710c785f BVA |
2127 | * If no I/O scheduler has been configured it is possible that |
2128 | * the hardware queue got stopped and restarted before requests | |
2129 | * were pushed back onto the dispatch list. Rerun the queue to | |
2130 | * avoid starvation. Notes: | |
2131 | * - blk_mq_run_hw_queue() checks whether or not a queue has | |
2132 | * been stopped before rerunning a queue. | |
2133 | * - Some but not all block drivers stop a queue before | |
fc17b653 | 2134 | * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq |
710c785f | 2135 | * and dm-rq. |
86ff7c2a ML |
2136 | * |
2137 | * If driver returns BLK_STS_RESOURCE and SCHED_RESTART | |
2138 | * bit is set, run queue after a delay to avoid IO stalls | |
ab3cee37 | 2139 | * that could otherwise occur if the queue is idle. We'll do |
9586e67b NA |
2140 | * similar if we couldn't get budget or couldn't lock a zone |
2141 | * and SCHED_RESTART is set. | |
bd166ef1 | 2142 | */ |
86ff7c2a | 2143 | needs_restart = blk_mq_sched_needs_restart(hctx); |
9586e67b NA |
2144 | if (prep == PREP_DISPATCH_NO_BUDGET) |
2145 | needs_resource = true; | |
86ff7c2a | 2146 | if (!needs_restart || |
eb619fdb | 2147 | (no_tag && list_empty_careful(&hctx->dispatch_wait.entry))) |
bd166ef1 | 2148 | blk_mq_run_hw_queue(hctx, true); |
6d5e8d21 | 2149 | else if (needs_resource) |
86ff7c2a | 2150 | blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY); |
1f57f8d4 | 2151 | |
6e768717 | 2152 | blk_mq_update_dispatch_busy(hctx, true); |
1f57f8d4 | 2153 | return false; |
6e768717 ML |
2154 | } else |
2155 | blk_mq_update_dispatch_busy(hctx, false); | |
f04c3df3 | 2156 | |
93efe981 | 2157 | return (queued + errors) != 0; |
f04c3df3 JA |
2158 | } |
2159 | ||
105663f7 AA |
2160 | /** |
2161 | * __blk_mq_run_hw_queue - Run a hardware queue. | |
2162 | * @hctx: Pointer to the hardware queue to run. | |
2163 | * | |
2164 | * Send pending requests to the hardware. | |
2165 | */ | |
6a83e74d BVA |
2166 | static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx) |
2167 | { | |
b7a71e66 JA |
2168 | /* |
2169 | * We can't run the queue inline with ints disabled. Ensure that | |
2170 | * we catch bad users of this early. | |
2171 | */ | |
2172 | WARN_ON_ONCE(in_interrupt()); | |
2173 | ||
bcc330f4 ML |
2174 | blk_mq_run_dispatch_ops(hctx->queue, |
2175 | blk_mq_sched_dispatch_requests(hctx)); | |
6a83e74d BVA |
2176 | } |
2177 | ||
f82ddf19 ML |
2178 | static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx) |
2179 | { | |
2180 | int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask); | |
2181 | ||
2182 | if (cpu >= nr_cpu_ids) | |
2183 | cpu = cpumask_first(hctx->cpumask); | |
2184 | return cpu; | |
2185 | } | |
2186 | ||
506e931f JA |
2187 | /* |
2188 | * It'd be great if the workqueue API had a way to pass | |
2189 | * in a mask and had some smarts for more clever placement. | |
2190 | * For now we just round-robin here, switching for every | |
2191 | * BLK_MQ_CPU_WORK_BATCH queued items. | |
2192 | */ | |
2193 | static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx) | |
2194 | { | |
7bed4595 | 2195 | bool tried = false; |
476f8c98 | 2196 | int next_cpu = hctx->next_cpu; |
7bed4595 | 2197 | |
b657d7e6 CH |
2198 | if (hctx->queue->nr_hw_queues == 1) |
2199 | return WORK_CPU_UNBOUND; | |
506e931f JA |
2200 | |
2201 | if (--hctx->next_cpu_batch <= 0) { | |
7bed4595 | 2202 | select_cpu: |
476f8c98 | 2203 | next_cpu = cpumask_next_and(next_cpu, hctx->cpumask, |
20e4d813 | 2204 | cpu_online_mask); |
506e931f | 2205 | if (next_cpu >= nr_cpu_ids) |
f82ddf19 | 2206 | next_cpu = blk_mq_first_mapped_cpu(hctx); |
506e931f JA |
2207 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; |
2208 | } | |
2209 | ||
7bed4595 ML |
2210 | /* |
2211 | * Do unbound schedule if we can't find a online CPU for this hctx, | |
2212 | * and it should only happen in the path of handling CPU DEAD. | |
2213 | */ | |
476f8c98 | 2214 | if (!cpu_online(next_cpu)) { |
7bed4595 ML |
2215 | if (!tried) { |
2216 | tried = true; | |
2217 | goto select_cpu; | |
2218 | } | |
2219 | ||
2220 | /* | |
2221 | * Make sure to re-select CPU next time once after CPUs | |
2222 | * in hctx->cpumask become online again. | |
2223 | */ | |
476f8c98 | 2224 | hctx->next_cpu = next_cpu; |
7bed4595 ML |
2225 | hctx->next_cpu_batch = 1; |
2226 | return WORK_CPU_UNBOUND; | |
2227 | } | |
476f8c98 ML |
2228 | |
2229 | hctx->next_cpu = next_cpu; | |
2230 | return next_cpu; | |
506e931f JA |
2231 | } |
2232 | ||
105663f7 AA |
2233 | /** |
2234 | * __blk_mq_delay_run_hw_queue - Run (or schedule to run) a hardware queue. | |
2235 | * @hctx: Pointer to the hardware queue to run. | |
2236 | * @async: If we want to run the queue asynchronously. | |
fa94ba8a | 2237 | * @msecs: Milliseconds of delay to wait before running the queue. |
105663f7 AA |
2238 | * |
2239 | * If !@async, try to run the queue now. Else, run the queue asynchronously and | |
2240 | * with a delay of @msecs. | |
2241 | */ | |
7587a5ae BVA |
2242 | static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async, |
2243 | unsigned long msecs) | |
320ae51f | 2244 | { |
5435c023 | 2245 | if (unlikely(blk_mq_hctx_stopped(hctx))) |
320ae51f JA |
2246 | return; |
2247 | ||
1b792f2f | 2248 | if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) { |
3c8f9da4 | 2249 | if (cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask)) { |
398205b8 | 2250 | __blk_mq_run_hw_queue(hctx); |
398205b8 PB |
2251 | return; |
2252 | } | |
e4043dcf | 2253 | } |
398205b8 | 2254 | |
ae943d20 BVA |
2255 | kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work, |
2256 | msecs_to_jiffies(msecs)); | |
7587a5ae BVA |
2257 | } |
2258 | ||
105663f7 AA |
2259 | /** |
2260 | * blk_mq_delay_run_hw_queue - Run a hardware queue asynchronously. | |
2261 | * @hctx: Pointer to the hardware queue to run. | |
fa94ba8a | 2262 | * @msecs: Milliseconds of delay to wait before running the queue. |
105663f7 AA |
2263 | * |
2264 | * Run a hardware queue asynchronously with a delay of @msecs. | |
2265 | */ | |
7587a5ae BVA |
2266 | void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs) |
2267 | { | |
2268 | __blk_mq_delay_run_hw_queue(hctx, true, msecs); | |
2269 | } | |
2270 | EXPORT_SYMBOL(blk_mq_delay_run_hw_queue); | |
2271 | ||
105663f7 AA |
2272 | /** |
2273 | * blk_mq_run_hw_queue - Start to run a hardware queue. | |
2274 | * @hctx: Pointer to the hardware queue to run. | |
2275 | * @async: If we want to run the queue asynchronously. | |
2276 | * | |
2277 | * Check if the request queue is not in a quiesced state and if there are | |
2278 | * pending requests to be sent. If this is true, run the queue to send requests | |
2279 | * to hardware. | |
2280 | */ | |
626fb735 | 2281 | void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) |
7587a5ae | 2282 | { |
24f5a90f ML |
2283 | bool need_run; |
2284 | ||
2285 | /* | |
2286 | * When queue is quiesced, we may be switching io scheduler, or | |
2287 | * updating nr_hw_queues, or other things, and we can't run queue | |
2288 | * any more, even __blk_mq_hctx_has_pending() can't be called safely. | |
2289 | * | |
2290 | * And queue will be rerun in blk_mq_unquiesce_queue() if it is | |
2291 | * quiesced. | |
2292 | */ | |
41adf531 | 2293 | __blk_mq_run_dispatch_ops(hctx->queue, false, |
2a904d00 ML |
2294 | need_run = !blk_queue_quiesced(hctx->queue) && |
2295 | blk_mq_hctx_has_pending(hctx)); | |
24f5a90f | 2296 | |
626fb735 | 2297 | if (need_run) |
79f720a7 | 2298 | __blk_mq_delay_run_hw_queue(hctx, async, 0); |
320ae51f | 2299 | } |
5b727272 | 2300 | EXPORT_SYMBOL(blk_mq_run_hw_queue); |
320ae51f | 2301 | |
b6e68ee8 JK |
2302 | /* |
2303 | * Return prefered queue to dispatch from (if any) for non-mq aware IO | |
2304 | * scheduler. | |
2305 | */ | |
2306 | static struct blk_mq_hw_ctx *blk_mq_get_sq_hctx(struct request_queue *q) | |
2307 | { | |
5d05426e | 2308 | struct blk_mq_ctx *ctx = blk_mq_get_ctx(q); |
b6e68ee8 JK |
2309 | /* |
2310 | * If the IO scheduler does not respect hardware queues when | |
2311 | * dispatching, we just don't bother with multiple HW queues and | |
2312 | * dispatch from hctx for the current CPU since running multiple queues | |
2313 | * just causes lock contention inside the scheduler and pointless cache | |
2314 | * bouncing. | |
2315 | */ | |
51ab80f0 | 2316 | struct blk_mq_hw_ctx *hctx = ctx->hctxs[HCTX_TYPE_DEFAULT]; |
5d05426e | 2317 | |
b6e68ee8 JK |
2318 | if (!blk_mq_hctx_stopped(hctx)) |
2319 | return hctx; | |
2320 | return NULL; | |
2321 | } | |
2322 | ||
105663f7 | 2323 | /** |
24f7bb88 | 2324 | * blk_mq_run_hw_queues - Run all hardware queues in a request queue. |
105663f7 AA |
2325 | * @q: Pointer to the request queue to run. |
2326 | * @async: If we want to run the queue asynchronously. | |
2327 | */ | |
b94ec296 | 2328 | void blk_mq_run_hw_queues(struct request_queue *q, bool async) |
320ae51f | 2329 | { |
b6e68ee8 | 2330 | struct blk_mq_hw_ctx *hctx, *sq_hctx; |
4f481208 | 2331 | unsigned long i; |
320ae51f | 2332 | |
b6e68ee8 | 2333 | sq_hctx = NULL; |
4d337ceb | 2334 | if (blk_queue_sq_sched(q)) |
b6e68ee8 | 2335 | sq_hctx = blk_mq_get_sq_hctx(q); |
320ae51f | 2336 | queue_for_each_hw_ctx(q, hctx, i) { |
79f720a7 | 2337 | if (blk_mq_hctx_stopped(hctx)) |
320ae51f | 2338 | continue; |
b6e68ee8 JK |
2339 | /* |
2340 | * Dispatch from this hctx either if there's no hctx preferred | |
2341 | * by IO scheduler or if it has requests that bypass the | |
2342 | * scheduler. | |
2343 | */ | |
2344 | if (!sq_hctx || sq_hctx == hctx || | |
2345 | !list_empty_careful(&hctx->dispatch)) | |
2346 | blk_mq_run_hw_queue(hctx, async); | |
320ae51f JA |
2347 | } |
2348 | } | |
b94ec296 | 2349 | EXPORT_SYMBOL(blk_mq_run_hw_queues); |
320ae51f | 2350 | |
b9151e7b DA |
2351 | /** |
2352 | * blk_mq_delay_run_hw_queues - Run all hardware queues asynchronously. | |
2353 | * @q: Pointer to the request queue to run. | |
fa94ba8a | 2354 | * @msecs: Milliseconds of delay to wait before running the queues. |
b9151e7b DA |
2355 | */ |
2356 | void blk_mq_delay_run_hw_queues(struct request_queue *q, unsigned long msecs) | |
2357 | { | |
b6e68ee8 | 2358 | struct blk_mq_hw_ctx *hctx, *sq_hctx; |
4f481208 | 2359 | unsigned long i; |
b9151e7b | 2360 | |
b6e68ee8 | 2361 | sq_hctx = NULL; |
4d337ceb | 2362 | if (blk_queue_sq_sched(q)) |
b6e68ee8 | 2363 | sq_hctx = blk_mq_get_sq_hctx(q); |
b9151e7b DA |
2364 | queue_for_each_hw_ctx(q, hctx, i) { |
2365 | if (blk_mq_hctx_stopped(hctx)) | |
2366 | continue; | |
8f5fea65 DJ |
2367 | /* |
2368 | * If there is already a run_work pending, leave the | |
2369 | * pending delay untouched. Otherwise, a hctx can stall | |
2370 | * if another hctx is re-delaying the other's work | |
2371 | * before the work executes. | |
2372 | */ | |
2373 | if (delayed_work_pending(&hctx->run_work)) | |
2374 | continue; | |
b6e68ee8 JK |
2375 | /* |
2376 | * Dispatch from this hctx either if there's no hctx preferred | |
2377 | * by IO scheduler or if it has requests that bypass the | |
2378 | * scheduler. | |
2379 | */ | |
2380 | if (!sq_hctx || sq_hctx == hctx || | |
2381 | !list_empty_careful(&hctx->dispatch)) | |
2382 | blk_mq_delay_run_hw_queue(hctx, msecs); | |
b9151e7b DA |
2383 | } |
2384 | } | |
2385 | EXPORT_SYMBOL(blk_mq_delay_run_hw_queues); | |
2386 | ||
39a70c76 ML |
2387 | /* |
2388 | * This function is often used for pausing .queue_rq() by driver when | |
2389 | * there isn't enough resource or some conditions aren't satisfied, and | |
4d606219 | 2390 | * BLK_STS_RESOURCE is usually returned. |
39a70c76 ML |
2391 | * |
2392 | * We do not guarantee that dispatch can be drained or blocked | |
2393 | * after blk_mq_stop_hw_queue() returns. Please use | |
2394 | * blk_mq_quiesce_queue() for that requirement. | |
2395 | */ | |
2719aa21 JA |
2396 | void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx) |
2397 | { | |
641a9ed6 | 2398 | cancel_delayed_work(&hctx->run_work); |
280d45f6 | 2399 | |
641a9ed6 | 2400 | set_bit(BLK_MQ_S_STOPPED, &hctx->state); |
2719aa21 | 2401 | } |
641a9ed6 | 2402 | EXPORT_SYMBOL(blk_mq_stop_hw_queue); |
2719aa21 | 2403 | |
39a70c76 ML |
2404 | /* |
2405 | * This function is often used for pausing .queue_rq() by driver when | |
2406 | * there isn't enough resource or some conditions aren't satisfied, and | |
4d606219 | 2407 | * BLK_STS_RESOURCE is usually returned. |
39a70c76 ML |
2408 | * |
2409 | * We do not guarantee that dispatch can be drained or blocked | |
2410 | * after blk_mq_stop_hw_queues() returns. Please use | |
2411 | * blk_mq_quiesce_queue() for that requirement. | |
2412 | */ | |
2719aa21 JA |
2413 | void blk_mq_stop_hw_queues(struct request_queue *q) |
2414 | { | |
641a9ed6 | 2415 | struct blk_mq_hw_ctx *hctx; |
4f481208 | 2416 | unsigned long i; |
641a9ed6 ML |
2417 | |
2418 | queue_for_each_hw_ctx(q, hctx, i) | |
2419 | blk_mq_stop_hw_queue(hctx); | |
280d45f6 CH |
2420 | } |
2421 | EXPORT_SYMBOL(blk_mq_stop_hw_queues); | |
2422 | ||
320ae51f JA |
2423 | void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx) |
2424 | { | |
2425 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
e4043dcf | 2426 | |
0ffbce80 | 2427 | blk_mq_run_hw_queue(hctx, false); |
320ae51f JA |
2428 | } |
2429 | EXPORT_SYMBOL(blk_mq_start_hw_queue); | |
2430 | ||
2f268556 CH |
2431 | void blk_mq_start_hw_queues(struct request_queue *q) |
2432 | { | |
2433 | struct blk_mq_hw_ctx *hctx; | |
4f481208 | 2434 | unsigned long i; |
2f268556 CH |
2435 | |
2436 | queue_for_each_hw_ctx(q, hctx, i) | |
2437 | blk_mq_start_hw_queue(hctx); | |
2438 | } | |
2439 | EXPORT_SYMBOL(blk_mq_start_hw_queues); | |
2440 | ||
ae911c5e JA |
2441 | void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) |
2442 | { | |
2443 | if (!blk_mq_hctx_stopped(hctx)) | |
2444 | return; | |
2445 | ||
2446 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
2447 | blk_mq_run_hw_queue(hctx, async); | |
2448 | } | |
2449 | EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue); | |
2450 | ||
1b4a3258 | 2451 | void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async) |
320ae51f JA |
2452 | { |
2453 | struct blk_mq_hw_ctx *hctx; | |
4f481208 | 2454 | unsigned long i; |
320ae51f | 2455 | |
ae911c5e JA |
2456 | queue_for_each_hw_ctx(q, hctx, i) |
2457 | blk_mq_start_stopped_hw_queue(hctx, async); | |
320ae51f JA |
2458 | } |
2459 | EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues); | |
2460 | ||
70f4db63 | 2461 | static void blk_mq_run_work_fn(struct work_struct *work) |
320ae51f JA |
2462 | { |
2463 | struct blk_mq_hw_ctx *hctx; | |
2464 | ||
9f993737 | 2465 | hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work); |
320ae51f | 2466 | |
21c6e939 | 2467 | /* |
15fe8a90 | 2468 | * If we are stopped, don't run the queue. |
21c6e939 | 2469 | */ |
0841031a | 2470 | if (blk_mq_hctx_stopped(hctx)) |
0196d6b4 | 2471 | return; |
7587a5ae BVA |
2472 | |
2473 | __blk_mq_run_hw_queue(hctx); | |
2474 | } | |
2475 | ||
cfd0c552 | 2476 | static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx, |
cfd0c552 ML |
2477 | struct request *rq, |
2478 | bool at_head) | |
320ae51f | 2479 | { |
e57690fe | 2480 | struct blk_mq_ctx *ctx = rq->mq_ctx; |
c16d6b5a | 2481 | enum hctx_type type = hctx->type; |
e57690fe | 2482 | |
7b607814 BVA |
2483 | lockdep_assert_held(&ctx->lock); |
2484 | ||
a54895fa | 2485 | trace_block_rq_insert(rq); |
01b983c9 | 2486 | |
72a0a36e | 2487 | if (at_head) |
c16d6b5a | 2488 | list_add(&rq->queuelist, &ctx->rq_lists[type]); |
72a0a36e | 2489 | else |
c16d6b5a | 2490 | list_add_tail(&rq->queuelist, &ctx->rq_lists[type]); |
cfd0c552 | 2491 | } |
4bb659b1 | 2492 | |
2c3ad667 JA |
2493 | void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq, |
2494 | bool at_head) | |
cfd0c552 ML |
2495 | { |
2496 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
2497 | ||
7b607814 BVA |
2498 | lockdep_assert_held(&ctx->lock); |
2499 | ||
e57690fe | 2500 | __blk_mq_insert_req_list(hctx, rq, at_head); |
320ae51f | 2501 | blk_mq_hctx_mark_pending(hctx, ctx); |
320ae51f JA |
2502 | } |
2503 | ||
105663f7 AA |
2504 | /** |
2505 | * blk_mq_request_bypass_insert - Insert a request at dispatch list. | |
2506 | * @rq: Pointer to request to be inserted. | |
26bfeb26 | 2507 | * @at_head: true if the request should be inserted at the head of the list. |
105663f7 AA |
2508 | * @run_queue: If we should run the hardware queue after inserting the request. |
2509 | * | |
157f377b JA |
2510 | * Should only be used carefully, when the caller knows we want to |
2511 | * bypass a potential IO scheduler on the target device. | |
2512 | */ | |
01e99aec ML |
2513 | void blk_mq_request_bypass_insert(struct request *rq, bool at_head, |
2514 | bool run_queue) | |
157f377b | 2515 | { |
ea4f995e | 2516 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; |
157f377b JA |
2517 | |
2518 | spin_lock(&hctx->lock); | |
01e99aec ML |
2519 | if (at_head) |
2520 | list_add(&rq->queuelist, &hctx->dispatch); | |
2521 | else | |
2522 | list_add_tail(&rq->queuelist, &hctx->dispatch); | |
157f377b JA |
2523 | spin_unlock(&hctx->lock); |
2524 | ||
b0850297 ML |
2525 | if (run_queue) |
2526 | blk_mq_run_hw_queue(hctx, false); | |
157f377b JA |
2527 | } |
2528 | ||
bd166ef1 JA |
2529 | void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx, |
2530 | struct list_head *list) | |
320ae51f JA |
2531 | |
2532 | { | |
3f0cedc7 | 2533 | struct request *rq; |
c16d6b5a | 2534 | enum hctx_type type = hctx->type; |
3f0cedc7 | 2535 | |
320ae51f JA |
2536 | /* |
2537 | * preemption doesn't flush plug list, so it's possible ctx->cpu is | |
2538 | * offline now | |
2539 | */ | |
3f0cedc7 | 2540 | list_for_each_entry(rq, list, queuelist) { |
e57690fe | 2541 | BUG_ON(rq->mq_ctx != ctx); |
a54895fa | 2542 | trace_block_rq_insert(rq); |
320ae51f | 2543 | } |
3f0cedc7 ML |
2544 | |
2545 | spin_lock(&ctx->lock); | |
c16d6b5a | 2546 | list_splice_tail_init(list, &ctx->rq_lists[type]); |
cfd0c552 | 2547 | blk_mq_hctx_mark_pending(hctx, ctx); |
320ae51f | 2548 | spin_unlock(&ctx->lock); |
320ae51f JA |
2549 | } |
2550 | ||
dc5fc361 JA |
2551 | static void blk_mq_commit_rqs(struct blk_mq_hw_ctx *hctx, int *queued, |
2552 | bool from_schedule) | |
320ae51f | 2553 | { |
dc5fc361 JA |
2554 | if (hctx->queue->mq_ops->commit_rqs) { |
2555 | trace_block_unplug(hctx->queue, *queued, !from_schedule); | |
2556 | hctx->queue->mq_ops->commit_rqs(hctx); | |
2557 | } | |
2558 | *queued = 0; | |
2559 | } | |
320ae51f | 2560 | |
14ccb66b CH |
2561 | static void blk_mq_bio_to_request(struct request *rq, struct bio *bio, |
2562 | unsigned int nr_segs) | |
320ae51f | 2563 | { |
93f221ae EB |
2564 | int err; |
2565 | ||
f924cdde CH |
2566 | if (bio->bi_opf & REQ_RAHEAD) |
2567 | rq->cmd_flags |= REQ_FAILFAST_MASK; | |
2568 | ||
2569 | rq->__sector = bio->bi_iter.bi_sector; | |
14ccb66b | 2570 | blk_rq_bio_prep(rq, bio, nr_segs); |
93f221ae EB |
2571 | |
2572 | /* This can't fail, since GFP_NOIO includes __GFP_DIRECT_RECLAIM. */ | |
2573 | err = blk_crypto_rq_bio_prep(rq, bio, GFP_NOIO); | |
2574 | WARN_ON_ONCE(err); | |
4b570521 | 2575 | |
b5af37ab | 2576 | blk_account_io_start(rq); |
320ae51f JA |
2577 | } |
2578 | ||
0f95549c | 2579 | static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx, |
3e08773c | 2580 | struct request *rq, bool last) |
f984df1f | 2581 | { |
f984df1f | 2582 | struct request_queue *q = rq->q; |
f984df1f SL |
2583 | struct blk_mq_queue_data bd = { |
2584 | .rq = rq, | |
be94f058 | 2585 | .last = last, |
f984df1f | 2586 | }; |
f06345ad | 2587 | blk_status_t ret; |
0f95549c | 2588 | |
0f95549c MS |
2589 | /* |
2590 | * For OK queue, we are done. For error, caller may kill it. | |
2591 | * Any other error (busy), just add it to our list as we | |
2592 | * previously would have done. | |
2593 | */ | |
2594 | ret = q->mq_ops->queue_rq(hctx, &bd); | |
2595 | switch (ret) { | |
2596 | case BLK_STS_OK: | |
6ce3dd6e | 2597 | blk_mq_update_dispatch_busy(hctx, false); |
0f95549c MS |
2598 | break; |
2599 | case BLK_STS_RESOURCE: | |
86ff7c2a | 2600 | case BLK_STS_DEV_RESOURCE: |
6ce3dd6e | 2601 | blk_mq_update_dispatch_busy(hctx, true); |
0f95549c MS |
2602 | __blk_mq_requeue_request(rq); |
2603 | break; | |
2604 | default: | |
6ce3dd6e | 2605 | blk_mq_update_dispatch_busy(hctx, false); |
0f95549c MS |
2606 | break; |
2607 | } | |
2608 | ||
2609 | return ret; | |
2610 | } | |
2611 | ||
fd9c40f6 | 2612 | static blk_status_t __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx, |
0f95549c | 2613 | struct request *rq, |
fd9c40f6 | 2614 | bool bypass_insert, bool last) |
0f95549c MS |
2615 | { |
2616 | struct request_queue *q = rq->q; | |
d964f04a | 2617 | bool run_queue = true; |
2a5a24aa | 2618 | int budget_token; |
d964f04a | 2619 | |
23d4ee19 | 2620 | /* |
fd9c40f6 | 2621 | * RCU or SRCU read lock is needed before checking quiesced flag. |
23d4ee19 | 2622 | * |
fd9c40f6 BVA |
2623 | * When queue is stopped or quiesced, ignore 'bypass_insert' from |
2624 | * blk_mq_request_issue_directly(), and return BLK_STS_OK to caller, | |
2625 | * and avoid driver to try to dispatch again. | |
23d4ee19 | 2626 | */ |
fd9c40f6 | 2627 | if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)) { |
d964f04a | 2628 | run_queue = false; |
fd9c40f6 BVA |
2629 | bypass_insert = false; |
2630 | goto insert; | |
d964f04a | 2631 | } |
f984df1f | 2632 | |
2ff0682d | 2633 | if ((rq->rq_flags & RQF_ELV) && !bypass_insert) |
fd9c40f6 | 2634 | goto insert; |
2253efc8 | 2635 | |
2a5a24aa ML |
2636 | budget_token = blk_mq_get_dispatch_budget(q); |
2637 | if (budget_token < 0) | |
fd9c40f6 | 2638 | goto insert; |
bd166ef1 | 2639 | |
2a5a24aa ML |
2640 | blk_mq_set_rq_budget_token(rq, budget_token); |
2641 | ||
8ab6bb9e | 2642 | if (!blk_mq_get_driver_tag(rq)) { |
2a5a24aa | 2643 | blk_mq_put_dispatch_budget(q, budget_token); |
fd9c40f6 | 2644 | goto insert; |
88022d72 | 2645 | } |
de148297 | 2646 | |
3e08773c | 2647 | return __blk_mq_issue_directly(hctx, rq, last); |
fd9c40f6 BVA |
2648 | insert: |
2649 | if (bypass_insert) | |
2650 | return BLK_STS_RESOURCE; | |
2651 | ||
db03f88f ML |
2652 | blk_mq_sched_insert_request(rq, false, run_queue, false); |
2653 | ||
fd9c40f6 BVA |
2654 | return BLK_STS_OK; |
2655 | } | |
2656 | ||
105663f7 AA |
2657 | /** |
2658 | * blk_mq_try_issue_directly - Try to send a request directly to device driver. | |
2659 | * @hctx: Pointer of the associated hardware queue. | |
2660 | * @rq: Pointer to request to be sent. | |
105663f7 AA |
2661 | * |
2662 | * If the device has enough resources to accept a new request now, send the | |
2663 | * request directly to device driver. Else, insert at hctx->dispatch queue, so | |
2664 | * we can try send it another time in the future. Requests inserted at this | |
2665 | * queue have higher priority. | |
2666 | */ | |
fd9c40f6 | 2667 | static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx, |
3e08773c | 2668 | struct request *rq) |
fd9c40f6 | 2669 | { |
2a904d00 ML |
2670 | blk_status_t ret = |
2671 | __blk_mq_try_issue_directly(hctx, rq, false, true); | |
fd9c40f6 | 2672 | |
fd9c40f6 | 2673 | if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE) |
01e99aec | 2674 | blk_mq_request_bypass_insert(rq, false, true); |
fd9c40f6 BVA |
2675 | else if (ret != BLK_STS_OK) |
2676 | blk_mq_end_request(rq, ret); | |
fd9c40f6 BVA |
2677 | } |
2678 | ||
06c8c691 | 2679 | static blk_status_t blk_mq_request_issue_directly(struct request *rq, bool last) |
fd9c40f6 | 2680 | { |
4cafe86c | 2681 | return __blk_mq_try_issue_directly(rq->mq_hctx, rq, true, last); |
5eb6126e CH |
2682 | } |
2683 | ||
b84c5b50 CH |
2684 | static void blk_mq_plug_issue_direct(struct blk_plug *plug, bool from_schedule) |
2685 | { | |
2686 | struct blk_mq_hw_ctx *hctx = NULL; | |
2687 | struct request *rq; | |
2688 | int queued = 0; | |
2689 | int errors = 0; | |
2690 | ||
2691 | while ((rq = rq_list_pop(&plug->mq_list))) { | |
2692 | bool last = rq_list_empty(plug->mq_list); | |
2693 | blk_status_t ret; | |
2694 | ||
2695 | if (hctx != rq->mq_hctx) { | |
2696 | if (hctx) | |
2697 | blk_mq_commit_rqs(hctx, &queued, from_schedule); | |
2698 | hctx = rq->mq_hctx; | |
2699 | } | |
2700 | ||
2701 | ret = blk_mq_request_issue_directly(rq, last); | |
2702 | switch (ret) { | |
2703 | case BLK_STS_OK: | |
2704 | queued++; | |
2705 | break; | |
2706 | case BLK_STS_RESOURCE: | |
2707 | case BLK_STS_DEV_RESOURCE: | |
d3b38596 | 2708 | blk_mq_request_bypass_insert(rq, false, true); |
b84c5b50 CH |
2709 | blk_mq_commit_rqs(hctx, &queued, from_schedule); |
2710 | return; | |
2711 | default: | |
2712 | blk_mq_end_request(rq, ret); | |
2713 | errors++; | |
2714 | break; | |
2715 | } | |
2716 | } | |
2717 | ||
2718 | /* | |
2719 | * If we didn't flush the entire list, we could have told the driver | |
2720 | * there was more coming, but that turned out to be a lie. | |
2721 | */ | |
2722 | if (errors) | |
2723 | blk_mq_commit_rqs(hctx, &queued, from_schedule); | |
2724 | } | |
2725 | ||
518579a9 KB |
2726 | static void __blk_mq_flush_plug_list(struct request_queue *q, |
2727 | struct blk_plug *plug) | |
2728 | { | |
2729 | if (blk_queue_quiesced(q)) | |
2730 | return; | |
2731 | q->mq_ops->queue_rqs(&plug->mq_list); | |
2732 | } | |
2733 | ||
26fed4ac JA |
2734 | static void blk_mq_dispatch_plug_list(struct blk_plug *plug, bool from_sched) |
2735 | { | |
2736 | struct blk_mq_hw_ctx *this_hctx = NULL; | |
2737 | struct blk_mq_ctx *this_ctx = NULL; | |
2738 | struct request *requeue_list = NULL; | |
2739 | unsigned int depth = 0; | |
2740 | LIST_HEAD(list); | |
2741 | ||
2742 | do { | |
2743 | struct request *rq = rq_list_pop(&plug->mq_list); | |
2744 | ||
2745 | if (!this_hctx) { | |
2746 | this_hctx = rq->mq_hctx; | |
2747 | this_ctx = rq->mq_ctx; | |
2748 | } else if (this_hctx != rq->mq_hctx || this_ctx != rq->mq_ctx) { | |
2749 | rq_list_add(&requeue_list, rq); | |
2750 | continue; | |
2751 | } | |
2752 | list_add_tail(&rq->queuelist, &list); | |
2753 | depth++; | |
2754 | } while (!rq_list_empty(plug->mq_list)); | |
2755 | ||
2756 | plug->mq_list = requeue_list; | |
2757 | trace_block_unplug(this_hctx->queue, depth, !from_sched); | |
2758 | blk_mq_sched_insert_requests(this_hctx, this_ctx, &list, from_sched); | |
2759 | } | |
2760 | ||
b84c5b50 CH |
2761 | void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule) |
2762 | { | |
3c67d44d | 2763 | struct request *rq; |
b84c5b50 CH |
2764 | |
2765 | if (rq_list_empty(plug->mq_list)) | |
2766 | return; | |
2767 | plug->rq_count = 0; | |
2768 | ||
2769 | if (!plug->multiple_queues && !plug->has_elevator && !from_schedule) { | |
3c67d44d JA |
2770 | struct request_queue *q; |
2771 | ||
2772 | rq = rq_list_peek(&plug->mq_list); | |
2773 | q = rq->q; | |
2774 | ||
2775 | /* | |
2776 | * Peek first request and see if we have a ->queue_rqs() hook. | |
2777 | * If we do, we can dispatch the whole plug list in one go. We | |
2778 | * already know at this point that all requests belong to the | |
2779 | * same queue, caller must ensure that's the case. | |
2780 | * | |
2781 | * Since we pass off the full list to the driver at this point, | |
2782 | * we do not increment the active request count for the queue. | |
2783 | * Bypass shared tags for now because of that. | |
2784 | */ | |
2785 | if (q->mq_ops->queue_rqs && | |
2786 | !(rq->mq_hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) { | |
2787 | blk_mq_run_dispatch_ops(q, | |
518579a9 | 2788 | __blk_mq_flush_plug_list(q, plug)); |
3c67d44d JA |
2789 | if (rq_list_empty(plug->mq_list)) |
2790 | return; | |
2791 | } | |
73f3760e ML |
2792 | |
2793 | blk_mq_run_dispatch_ops(q, | |
4cafe86c | 2794 | blk_mq_plug_issue_direct(plug, false)); |
b84c5b50 CH |
2795 | if (rq_list_empty(plug->mq_list)) |
2796 | return; | |
2797 | } | |
2798 | ||
b84c5b50 | 2799 | do { |
26fed4ac | 2800 | blk_mq_dispatch_plug_list(plug, from_schedule); |
b84c5b50 | 2801 | } while (!rq_list_empty(plug->mq_list)); |
b84c5b50 CH |
2802 | } |
2803 | ||
6ce3dd6e ML |
2804 | void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx, |
2805 | struct list_head *list) | |
2806 | { | |
536167d4 | 2807 | int queued = 0; |
632bfb63 | 2808 | int errors = 0; |
536167d4 | 2809 | |
6ce3dd6e | 2810 | while (!list_empty(list)) { |
fd9c40f6 | 2811 | blk_status_t ret; |
6ce3dd6e ML |
2812 | struct request *rq = list_first_entry(list, struct request, |
2813 | queuelist); | |
2814 | ||
2815 | list_del_init(&rq->queuelist); | |
fd9c40f6 BVA |
2816 | ret = blk_mq_request_issue_directly(rq, list_empty(list)); |
2817 | if (ret != BLK_STS_OK) { | |
65fac0d5 | 2818 | errors++; |
fd9c40f6 BVA |
2819 | if (ret == BLK_STS_RESOURCE || |
2820 | ret == BLK_STS_DEV_RESOURCE) { | |
01e99aec | 2821 | blk_mq_request_bypass_insert(rq, false, |
c616cbee | 2822 | list_empty(list)); |
fd9c40f6 BVA |
2823 | break; |
2824 | } | |
2825 | blk_mq_end_request(rq, ret); | |
536167d4 KB |
2826 | } else |
2827 | queued++; | |
6ce3dd6e | 2828 | } |
d666ba98 JA |
2829 | |
2830 | /* | |
2831 | * If we didn't flush the entire list, we could have told | |
2832 | * the driver there was more coming, but that turned out to | |
2833 | * be a lie. | |
2834 | */ | |
632bfb63 | 2835 | if ((!list_empty(list) || errors) && |
2836 | hctx->queue->mq_ops->commit_rqs && queued) | |
d666ba98 | 2837 | hctx->queue->mq_ops->commit_rqs(hctx); |
6ce3dd6e ML |
2838 | } |
2839 | ||
b131f201 | 2840 | static bool blk_mq_attempt_bio_merge(struct request_queue *q, |
0c5bcc92 | 2841 | struct bio *bio, unsigned int nr_segs) |
900e0807 JA |
2842 | { |
2843 | if (!blk_queue_nomerges(q) && bio_mergeable(bio)) { | |
0c5bcc92 | 2844 | if (blk_attempt_plug_merge(q, bio, nr_segs)) |
900e0807 JA |
2845 | return true; |
2846 | if (blk_mq_sched_bio_merge(q, bio, nr_segs)) | |
2847 | return true; | |
2848 | } | |
2849 | return false; | |
2850 | } | |
2851 | ||
71539717 JA |
2852 | static struct request *blk_mq_get_new_requests(struct request_queue *q, |
2853 | struct blk_plug *plug, | |
0a5aa8d1 SK |
2854 | struct bio *bio, |
2855 | unsigned int nsegs) | |
71539717 JA |
2856 | { |
2857 | struct blk_mq_alloc_data data = { | |
2858 | .q = q, | |
2859 | .nr_tags = 1, | |
9d497e29 | 2860 | .cmd_flags = bio->bi_opf, |
71539717 JA |
2861 | }; |
2862 | struct request *rq; | |
2863 | ||
5b13bc8a | 2864 | if (unlikely(bio_queue_enter(bio))) |
b637108a | 2865 | return NULL; |
900e0807 | 2866 | |
0a5aa8d1 SK |
2867 | if (blk_mq_attempt_bio_merge(q, bio, nsegs)) |
2868 | goto queue_exit; | |
2869 | ||
2870 | rq_qos_throttle(q, bio); | |
2871 | ||
71539717 JA |
2872 | if (plug) { |
2873 | data.nr_tags = plug->nr_ios; | |
2874 | plug->nr_ios = 1; | |
2875 | data.cached_rq = &plug->cached_rq; | |
2876 | } | |
2877 | ||
2878 | rq = __blk_mq_alloc_requests(&data); | |
373b5416 JA |
2879 | if (rq) |
2880 | return rq; | |
71539717 JA |
2881 | rq_qos_cleanup(q, bio); |
2882 | if (bio->bi_opf & REQ_NOWAIT) | |
2883 | bio_wouldblock_error(bio); | |
0a5aa8d1 | 2884 | queue_exit: |
5b13bc8a | 2885 | blk_queue_exit(q); |
71539717 JA |
2886 | return NULL; |
2887 | } | |
2888 | ||
5b13bc8a | 2889 | static inline struct request *blk_mq_get_cached_request(struct request_queue *q, |
0a5aa8d1 | 2890 | struct blk_plug *plug, struct bio **bio, unsigned int nsegs) |
71539717 | 2891 | { |
b637108a | 2892 | struct request *rq; |
b637108a | 2893 | |
5b13bc8a CH |
2894 | if (!plug) |
2895 | return NULL; | |
2896 | rq = rq_list_peek(&plug->cached_rq); | |
2897 | if (!rq || rq->q != q) | |
2898 | return NULL; | |
71539717 | 2899 | |
0a5aa8d1 SK |
2900 | if (blk_mq_attempt_bio_merge(q, *bio, nsegs)) { |
2901 | *bio = NULL; | |
2902 | return NULL; | |
2903 | } | |
2904 | ||
0a5aa8d1 | 2905 | if (blk_mq_get_hctx_type((*bio)->bi_opf) != rq->mq_hctx->type) |
5b13bc8a | 2906 | return NULL; |
0a5aa8d1 | 2907 | if (op_is_flush(rq->cmd_flags) != op_is_flush((*bio)->bi_opf)) |
5b13bc8a CH |
2908 | return NULL; |
2909 | ||
2645672f JA |
2910 | /* |
2911 | * If any qos ->throttle() end up blocking, we will have flushed the | |
2912 | * plug and hence killed the cached_rq list as well. Pop this entry | |
2913 | * before we throttle. | |
2914 | */ | |
5b13bc8a | 2915 | plug->cached_rq = rq_list_next(rq); |
2645672f JA |
2916 | rq_qos_throttle(q, *bio); |
2917 | ||
2918 | rq->cmd_flags = (*bio)->bi_opf; | |
5b13bc8a | 2919 | INIT_LIST_HEAD(&rq->queuelist); |
5b13bc8a | 2920 | return rq; |
71539717 JA |
2921 | } |
2922 | ||
82b74cac JK |
2923 | static void bio_set_ioprio(struct bio *bio) |
2924 | { | |
a78418e6 JK |
2925 | /* Nobody set ioprio so far? Initialize it based on task's nice value */ |
2926 | if (IOPRIO_PRIO_CLASS(bio->bi_ioprio) == IOPRIO_CLASS_NONE) | |
2927 | bio->bi_ioprio = get_current_ioprio(); | |
82b74cac JK |
2928 | blkcg_set_ioprio(bio); |
2929 | } | |
2930 | ||
105663f7 | 2931 | /** |
c62b37d9 | 2932 | * blk_mq_submit_bio - Create and send a request to block device. |
105663f7 AA |
2933 | * @bio: Bio pointer. |
2934 | * | |
2935 | * Builds up a request structure from @q and @bio and send to the device. The | |
2936 | * request may not be queued directly to hardware if: | |
2937 | * * This request can be merged with another one | |
2938 | * * We want to place request at plug queue for possible future merging | |
2939 | * * There is an IO scheduler active at this queue | |
2940 | * | |
2941 | * It will not queue the request if there is an error with the bio, or at the | |
2942 | * request creation. | |
105663f7 | 2943 | */ |
3e08773c | 2944 | void blk_mq_submit_bio(struct bio *bio) |
07068d5b | 2945 | { |
ed6cddef | 2946 | struct request_queue *q = bdev_get_queue(bio->bi_bdev); |
6deacb3b | 2947 | struct blk_plug *plug = blk_mq_plug(bio); |
ef295ecf | 2948 | const int is_sync = op_is_sync(bio->bi_opf); |
07068d5b | 2949 | struct request *rq; |
abd45c15 | 2950 | unsigned int nr_segs = 1; |
a892c8d5 | 2951 | blk_status_t ret; |
07068d5b | 2952 | |
51d798cd | 2953 | bio = blk_queue_bounce(bio, q); |
613b1488 | 2954 | if (bio_may_exceed_limits(bio, &q->limits)) { |
c55ddd90 | 2955 | bio = __bio_split_to_limits(bio, &q->limits, &nr_segs); |
613b1488 JA |
2956 | if (!bio) |
2957 | return; | |
2958 | } | |
f36ea50c | 2959 | |
e23947bd | 2960 | if (!bio_integrity_prep(bio)) |
900e0807 | 2961 | return; |
87760e5e | 2962 | |
9c6227e0 JK |
2963 | bio_set_ioprio(bio); |
2964 | ||
0a5aa8d1 | 2965 | rq = blk_mq_get_cached_request(q, plug, &bio, nr_segs); |
5b13bc8a | 2966 | if (!rq) { |
0a5aa8d1 SK |
2967 | if (!bio) |
2968 | return; | |
2969 | rq = blk_mq_get_new_requests(q, plug, bio, nr_segs); | |
5b13bc8a CH |
2970 | if (unlikely(!rq)) |
2971 | return; | |
2972 | } | |
87760e5e | 2973 | |
e8a676d6 | 2974 | trace_block_getrq(bio); |
d6f1dda2 | 2975 | |
c1c80384 | 2976 | rq_qos_track(q, rq, bio); |
07068d5b | 2977 | |
970d168d BVA |
2978 | blk_mq_bio_to_request(rq, bio, nr_segs); |
2979 | ||
a892c8d5 ST |
2980 | ret = blk_crypto_init_request(rq); |
2981 | if (ret != BLK_STS_OK) { | |
2982 | bio->bi_status = ret; | |
2983 | bio_endio(bio); | |
2984 | blk_mq_free_request(rq); | |
3e08773c | 2985 | return; |
a892c8d5 ST |
2986 | } |
2987 | ||
2b504bd4 ML |
2988 | if (op_is_flush(bio->bi_opf)) { |
2989 | blk_insert_flush(rq); | |
d92ca9d8 | 2990 | return; |
2b504bd4 | 2991 | } |
d92ca9d8 | 2992 | |
1e9c2303 | 2993 | if (plug) |
ce5b009c | 2994 | blk_add_rq_to_plug(plug, rq); |
1e9c2303 CH |
2995 | else if ((rq->rq_flags & RQF_ELV) || |
2996 | (rq->mq_hctx->dispatch_busy && | |
2997 | (q->nr_hw_queues == 1 || !is_sync))) | |
a12de1d4 | 2998 | blk_mq_sched_insert_request(rq, false, true, true); |
1e9c2303 | 2999 | else |
bcc330f4 | 3000 | blk_mq_run_dispatch_ops(rq->q, |
2a904d00 | 3001 | blk_mq_try_issue_directly(rq->mq_hctx, rq)); |
320ae51f JA |
3002 | } |
3003 | ||
248c7933 | 3004 | #ifdef CONFIG_BLK_MQ_STACKING |
06c8c691 | 3005 | /** |
a5efda3c | 3006 | * blk_insert_cloned_request - Helper for stacking drivers to submit a request |
a5efda3c | 3007 | * @rq: the request being queued |
06c8c691 | 3008 | */ |
28db4711 | 3009 | blk_status_t blk_insert_cloned_request(struct request *rq) |
06c8c691 | 3010 | { |
28db4711 | 3011 | struct request_queue *q = rq->q; |
06c8c691 | 3012 | unsigned int max_sectors = blk_queue_get_max_sectors(q, req_op(rq)); |
a5efda3c | 3013 | blk_status_t ret; |
06c8c691 CH |
3014 | |
3015 | if (blk_rq_sectors(rq) > max_sectors) { | |
3016 | /* | |
3017 | * SCSI device does not have a good way to return if | |
3018 | * Write Same/Zero is actually supported. If a device rejects | |
3019 | * a non-read/write command (discard, write same,etc.) the | |
3020 | * low-level device driver will set the relevant queue limit to | |
3021 | * 0 to prevent blk-lib from issuing more of the offending | |
3022 | * operations. Commands queued prior to the queue limit being | |
3023 | * reset need to be completed with BLK_STS_NOTSUPP to avoid I/O | |
3024 | * errors being propagated to upper layers. | |
3025 | */ | |
3026 | if (max_sectors == 0) | |
3027 | return BLK_STS_NOTSUPP; | |
3028 | ||
3029 | printk(KERN_ERR "%s: over max size limit. (%u > %u)\n", | |
3030 | __func__, blk_rq_sectors(rq), max_sectors); | |
3031 | return BLK_STS_IOERR; | |
3032 | } | |
3033 | ||
3034 | /* | |
3035 | * The queue settings related to segment counting may differ from the | |
3036 | * original queue. | |
3037 | */ | |
3038 | rq->nr_phys_segments = blk_recalc_rq_segments(rq); | |
3039 | if (rq->nr_phys_segments > queue_max_segments(q)) { | |
3040 | printk(KERN_ERR "%s: over max segments limit. (%hu > %hu)\n", | |
3041 | __func__, rq->nr_phys_segments, queue_max_segments(q)); | |
3042 | return BLK_STS_IOERR; | |
3043 | } | |
3044 | ||
28db4711 | 3045 | if (q->disk && should_fail_request(q->disk->part0, blk_rq_bytes(rq))) |
06c8c691 CH |
3046 | return BLK_STS_IOERR; |
3047 | ||
3048 | if (blk_crypto_insert_cloned_request(rq)) | |
3049 | return BLK_STS_IOERR; | |
3050 | ||
3051 | blk_account_io_start(rq); | |
3052 | ||
3053 | /* | |
3054 | * Since we have a scheduler attached on the top device, | |
3055 | * bypass a potential scheduler on the bottom device for | |
3056 | * insert. | |
3057 | */ | |
28db4711 | 3058 | blk_mq_run_dispatch_ops(q, |
4cafe86c | 3059 | ret = blk_mq_request_issue_directly(rq, true)); |
592ee119 YK |
3060 | if (ret) |
3061 | blk_account_io_done(rq, ktime_get_ns()); | |
4cafe86c | 3062 | return ret; |
06c8c691 CH |
3063 | } |
3064 | EXPORT_SYMBOL_GPL(blk_insert_cloned_request); | |
3065 | ||
3066 | /** | |
3067 | * blk_rq_unprep_clone - Helper function to free all bios in a cloned request | |
3068 | * @rq: the clone request to be cleaned up | |
3069 | * | |
3070 | * Description: | |
3071 | * Free all bios in @rq for a cloned request. | |
3072 | */ | |
3073 | void blk_rq_unprep_clone(struct request *rq) | |
3074 | { | |
3075 | struct bio *bio; | |
3076 | ||
3077 | while ((bio = rq->bio) != NULL) { | |
3078 | rq->bio = bio->bi_next; | |
3079 | ||
3080 | bio_put(bio); | |
3081 | } | |
3082 | } | |
3083 | EXPORT_SYMBOL_GPL(blk_rq_unprep_clone); | |
3084 | ||
3085 | /** | |
3086 | * blk_rq_prep_clone - Helper function to setup clone request | |
3087 | * @rq: the request to be setup | |
3088 | * @rq_src: original request to be cloned | |
3089 | * @bs: bio_set that bios for clone are allocated from | |
3090 | * @gfp_mask: memory allocation mask for bio | |
3091 | * @bio_ctr: setup function to be called for each clone bio. | |
3092 | * Returns %0 for success, non %0 for failure. | |
3093 | * @data: private data to be passed to @bio_ctr | |
3094 | * | |
3095 | * Description: | |
3096 | * Clones bios in @rq_src to @rq, and copies attributes of @rq_src to @rq. | |
3097 | * Also, pages which the original bios are pointing to are not copied | |
3098 | * and the cloned bios just point same pages. | |
3099 | * So cloned bios must be completed before original bios, which means | |
3100 | * the caller must complete @rq before @rq_src. | |
3101 | */ | |
3102 | int blk_rq_prep_clone(struct request *rq, struct request *rq_src, | |
3103 | struct bio_set *bs, gfp_t gfp_mask, | |
3104 | int (*bio_ctr)(struct bio *, struct bio *, void *), | |
3105 | void *data) | |
3106 | { | |
3107 | struct bio *bio, *bio_src; | |
3108 | ||
3109 | if (!bs) | |
3110 | bs = &fs_bio_set; | |
3111 | ||
3112 | __rq_for_each_bio(bio_src, rq_src) { | |
abfc426d CH |
3113 | bio = bio_alloc_clone(rq->q->disk->part0, bio_src, gfp_mask, |
3114 | bs); | |
06c8c691 CH |
3115 | if (!bio) |
3116 | goto free_and_out; | |
3117 | ||
3118 | if (bio_ctr && bio_ctr(bio, bio_src, data)) | |
3119 | goto free_and_out; | |
3120 | ||
3121 | if (rq->bio) { | |
3122 | rq->biotail->bi_next = bio; | |
3123 | rq->biotail = bio; | |
3124 | } else { | |
3125 | rq->bio = rq->biotail = bio; | |
3126 | } | |
3127 | bio = NULL; | |
3128 | } | |
3129 | ||
3130 | /* Copy attributes of the original request to the clone request. */ | |
3131 | rq->__sector = blk_rq_pos(rq_src); | |
3132 | rq->__data_len = blk_rq_bytes(rq_src); | |
3133 | if (rq_src->rq_flags & RQF_SPECIAL_PAYLOAD) { | |
3134 | rq->rq_flags |= RQF_SPECIAL_PAYLOAD; | |
3135 | rq->special_vec = rq_src->special_vec; | |
3136 | } | |
3137 | rq->nr_phys_segments = rq_src->nr_phys_segments; | |
3138 | rq->ioprio = rq_src->ioprio; | |
3139 | ||
3140 | if (rq->bio && blk_crypto_rq_bio_prep(rq, rq->bio, gfp_mask) < 0) | |
3141 | goto free_and_out; | |
3142 | ||
3143 | return 0; | |
3144 | ||
3145 | free_and_out: | |
3146 | if (bio) | |
3147 | bio_put(bio); | |
3148 | blk_rq_unprep_clone(rq); | |
3149 | ||
3150 | return -ENOMEM; | |
3151 | } | |
3152 | EXPORT_SYMBOL_GPL(blk_rq_prep_clone); | |
248c7933 | 3153 | #endif /* CONFIG_BLK_MQ_STACKING */ |
06c8c691 | 3154 | |
f2b8f3ce CH |
3155 | /* |
3156 | * Steal bios from a request and add them to a bio list. | |
3157 | * The request must not have been partially completed before. | |
3158 | */ | |
3159 | void blk_steal_bios(struct bio_list *list, struct request *rq) | |
3160 | { | |
3161 | if (rq->bio) { | |
3162 | if (list->tail) | |
3163 | list->tail->bi_next = rq->bio; | |
3164 | else | |
3165 | list->head = rq->bio; | |
3166 | list->tail = rq->biotail; | |
3167 | ||
3168 | rq->bio = NULL; | |
3169 | rq->biotail = NULL; | |
3170 | } | |
3171 | ||
3172 | rq->__data_len = 0; | |
3173 | } | |
3174 | EXPORT_SYMBOL_GPL(blk_steal_bios); | |
3175 | ||
bd63141d ML |
3176 | static size_t order_to_size(unsigned int order) |
3177 | { | |
3178 | return (size_t)PAGE_SIZE << order; | |
3179 | } | |
3180 | ||
3181 | /* called before freeing request pool in @tags */ | |
f32e4eaf JG |
3182 | static void blk_mq_clear_rq_mapping(struct blk_mq_tags *drv_tags, |
3183 | struct blk_mq_tags *tags) | |
bd63141d | 3184 | { |
bd63141d ML |
3185 | struct page *page; |
3186 | unsigned long flags; | |
3187 | ||
76dd2980 YK |
3188 | /* |
3189 | * There is no need to clear mapping if driver tags is not initialized | |
3190 | * or the mapping belongs to the driver tags. | |
3191 | */ | |
3192 | if (!drv_tags || drv_tags == tags) | |
4f245d5b JG |
3193 | return; |
3194 | ||
bd63141d ML |
3195 | list_for_each_entry(page, &tags->page_list, lru) { |
3196 | unsigned long start = (unsigned long)page_address(page); | |
3197 | unsigned long end = start + order_to_size(page->private); | |
3198 | int i; | |
3199 | ||
f32e4eaf | 3200 | for (i = 0; i < drv_tags->nr_tags; i++) { |
bd63141d ML |
3201 | struct request *rq = drv_tags->rqs[i]; |
3202 | unsigned long rq_addr = (unsigned long)rq; | |
3203 | ||
3204 | if (rq_addr >= start && rq_addr < end) { | |
0a467d0f | 3205 | WARN_ON_ONCE(req_ref_read(rq) != 0); |
bd63141d ML |
3206 | cmpxchg(&drv_tags->rqs[i], rq, NULL); |
3207 | } | |
3208 | } | |
3209 | } | |
3210 | ||
3211 | /* | |
3212 | * Wait until all pending iteration is done. | |
3213 | * | |
3214 | * Request reference is cleared and it is guaranteed to be observed | |
3215 | * after the ->lock is released. | |
3216 | */ | |
3217 | spin_lock_irqsave(&drv_tags->lock, flags); | |
3218 | spin_unlock_irqrestore(&drv_tags->lock, flags); | |
3219 | } | |
3220 | ||
cc71a6f4 JA |
3221 | void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags, |
3222 | unsigned int hctx_idx) | |
95363efd | 3223 | { |
f32e4eaf | 3224 | struct blk_mq_tags *drv_tags; |
e9b267d9 | 3225 | struct page *page; |
320ae51f | 3226 | |
e02657ea ML |
3227 | if (list_empty(&tags->page_list)) |
3228 | return; | |
3229 | ||
079a2e3e JG |
3230 | if (blk_mq_is_shared_tags(set->flags)) |
3231 | drv_tags = set->shared_tags; | |
e155b0c2 JG |
3232 | else |
3233 | drv_tags = set->tags[hctx_idx]; | |
f32e4eaf | 3234 | |
65de57bb | 3235 | if (tags->static_rqs && set->ops->exit_request) { |
e9b267d9 | 3236 | int i; |
320ae51f | 3237 | |
24d2f903 | 3238 | for (i = 0; i < tags->nr_tags; i++) { |
2af8cbe3 JA |
3239 | struct request *rq = tags->static_rqs[i]; |
3240 | ||
3241 | if (!rq) | |
e9b267d9 | 3242 | continue; |
d6296d39 | 3243 | set->ops->exit_request(set, rq, hctx_idx); |
2af8cbe3 | 3244 | tags->static_rqs[i] = NULL; |
e9b267d9 | 3245 | } |
320ae51f | 3246 | } |
320ae51f | 3247 | |
f32e4eaf | 3248 | blk_mq_clear_rq_mapping(drv_tags, tags); |
bd63141d | 3249 | |
24d2f903 CH |
3250 | while (!list_empty(&tags->page_list)) { |
3251 | page = list_first_entry(&tags->page_list, struct page, lru); | |
6753471c | 3252 | list_del_init(&page->lru); |
f75782e4 CM |
3253 | /* |
3254 | * Remove kmemleak object previously allocated in | |
273938bf | 3255 | * blk_mq_alloc_rqs(). |
f75782e4 CM |
3256 | */ |
3257 | kmemleak_free(page_address(page)); | |
320ae51f JA |
3258 | __free_pages(page, page->private); |
3259 | } | |
cc71a6f4 | 3260 | } |
320ae51f | 3261 | |
e155b0c2 | 3262 | void blk_mq_free_rq_map(struct blk_mq_tags *tags) |
cc71a6f4 | 3263 | { |
24d2f903 | 3264 | kfree(tags->rqs); |
cc71a6f4 | 3265 | tags->rqs = NULL; |
2af8cbe3 JA |
3266 | kfree(tags->static_rqs); |
3267 | tags->static_rqs = NULL; | |
320ae51f | 3268 | |
e155b0c2 | 3269 | blk_mq_free_tags(tags); |
320ae51f JA |
3270 | } |
3271 | ||
4d805131 ML |
3272 | static enum hctx_type hctx_idx_to_type(struct blk_mq_tag_set *set, |
3273 | unsigned int hctx_idx) | |
3274 | { | |
3275 | int i; | |
3276 | ||
3277 | for (i = 0; i < set->nr_maps; i++) { | |
3278 | unsigned int start = set->map[i].queue_offset; | |
3279 | unsigned int end = start + set->map[i].nr_queues; | |
3280 | ||
3281 | if (hctx_idx >= start && hctx_idx < end) | |
3282 | break; | |
3283 | } | |
3284 | ||
3285 | if (i >= set->nr_maps) | |
3286 | i = HCTX_TYPE_DEFAULT; | |
3287 | ||
3288 | return i; | |
3289 | } | |
3290 | ||
3291 | static int blk_mq_get_hctx_node(struct blk_mq_tag_set *set, | |
3292 | unsigned int hctx_idx) | |
3293 | { | |
3294 | enum hctx_type type = hctx_idx_to_type(set, hctx_idx); | |
3295 | ||
3296 | return blk_mq_hw_queue_to_node(&set->map[type], hctx_idx); | |
3297 | } | |
3298 | ||
63064be1 JG |
3299 | static struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, |
3300 | unsigned int hctx_idx, | |
3301 | unsigned int nr_tags, | |
e155b0c2 | 3302 | unsigned int reserved_tags) |
320ae51f | 3303 | { |
4d805131 | 3304 | int node = blk_mq_get_hctx_node(set, hctx_idx); |
24d2f903 | 3305 | struct blk_mq_tags *tags; |
320ae51f | 3306 | |
59f082e4 SL |
3307 | if (node == NUMA_NO_NODE) |
3308 | node = set->numa_node; | |
3309 | ||
e155b0c2 JG |
3310 | tags = blk_mq_init_tags(nr_tags, reserved_tags, node, |
3311 | BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags)); | |
24d2f903 CH |
3312 | if (!tags) |
3313 | return NULL; | |
320ae51f | 3314 | |
590b5b7d | 3315 | tags->rqs = kcalloc_node(nr_tags, sizeof(struct request *), |
36e1f3d1 | 3316 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, |
59f082e4 | 3317 | node); |
7edfd681 JC |
3318 | if (!tags->rqs) |
3319 | goto err_free_tags; | |
320ae51f | 3320 | |
590b5b7d KC |
3321 | tags->static_rqs = kcalloc_node(nr_tags, sizeof(struct request *), |
3322 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, | |
3323 | node); | |
7edfd681 JC |
3324 | if (!tags->static_rqs) |
3325 | goto err_free_rqs; | |
2af8cbe3 | 3326 | |
cc71a6f4 | 3327 | return tags; |
7edfd681 JC |
3328 | |
3329 | err_free_rqs: | |
3330 | kfree(tags->rqs); | |
3331 | err_free_tags: | |
3332 | blk_mq_free_tags(tags); | |
3333 | return NULL; | |
cc71a6f4 JA |
3334 | } |
3335 | ||
1d9bd516 TH |
3336 | static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq, |
3337 | unsigned int hctx_idx, int node) | |
3338 | { | |
3339 | int ret; | |
3340 | ||
3341 | if (set->ops->init_request) { | |
3342 | ret = set->ops->init_request(set, rq, hctx_idx, node); | |
3343 | if (ret) | |
3344 | return ret; | |
3345 | } | |
3346 | ||
12f5b931 | 3347 | WRITE_ONCE(rq->state, MQ_RQ_IDLE); |
1d9bd516 TH |
3348 | return 0; |
3349 | } | |
3350 | ||
63064be1 JG |
3351 | static int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, |
3352 | struct blk_mq_tags *tags, | |
3353 | unsigned int hctx_idx, unsigned int depth) | |
cc71a6f4 JA |
3354 | { |
3355 | unsigned int i, j, entries_per_page, max_order = 4; | |
4d805131 | 3356 | int node = blk_mq_get_hctx_node(set, hctx_idx); |
cc71a6f4 | 3357 | size_t rq_size, left; |
59f082e4 | 3358 | |
59f082e4 SL |
3359 | if (node == NUMA_NO_NODE) |
3360 | node = set->numa_node; | |
cc71a6f4 JA |
3361 | |
3362 | INIT_LIST_HEAD(&tags->page_list); | |
3363 | ||
320ae51f JA |
3364 | /* |
3365 | * rq_size is the size of the request plus driver payload, rounded | |
3366 | * to the cacheline size | |
3367 | */ | |
24d2f903 | 3368 | rq_size = round_up(sizeof(struct request) + set->cmd_size, |
320ae51f | 3369 | cache_line_size()); |
cc71a6f4 | 3370 | left = rq_size * depth; |
320ae51f | 3371 | |
cc71a6f4 | 3372 | for (i = 0; i < depth; ) { |
320ae51f JA |
3373 | int this_order = max_order; |
3374 | struct page *page; | |
3375 | int to_do; | |
3376 | void *p; | |
3377 | ||
b3a834b1 | 3378 | while (this_order && left < order_to_size(this_order - 1)) |
320ae51f JA |
3379 | this_order--; |
3380 | ||
3381 | do { | |
59f082e4 | 3382 | page = alloc_pages_node(node, |
36e1f3d1 | 3383 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO, |
a5164405 | 3384 | this_order); |
320ae51f JA |
3385 | if (page) |
3386 | break; | |
3387 | if (!this_order--) | |
3388 | break; | |
3389 | if (order_to_size(this_order) < rq_size) | |
3390 | break; | |
3391 | } while (1); | |
3392 | ||
3393 | if (!page) | |
24d2f903 | 3394 | goto fail; |
320ae51f JA |
3395 | |
3396 | page->private = this_order; | |
24d2f903 | 3397 | list_add_tail(&page->lru, &tags->page_list); |
320ae51f JA |
3398 | |
3399 | p = page_address(page); | |
f75782e4 CM |
3400 | /* |
3401 | * Allow kmemleak to scan these pages as they contain pointers | |
3402 | * to additional allocations like via ops->init_request(). | |
3403 | */ | |
36e1f3d1 | 3404 | kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO); |
320ae51f | 3405 | entries_per_page = order_to_size(this_order) / rq_size; |
cc71a6f4 | 3406 | to_do = min(entries_per_page, depth - i); |
320ae51f JA |
3407 | left -= to_do * rq_size; |
3408 | for (j = 0; j < to_do; j++) { | |
2af8cbe3 JA |
3409 | struct request *rq = p; |
3410 | ||
3411 | tags->static_rqs[i] = rq; | |
1d9bd516 TH |
3412 | if (blk_mq_init_request(set, rq, hctx_idx, node)) { |
3413 | tags->static_rqs[i] = NULL; | |
3414 | goto fail; | |
e9b267d9 CH |
3415 | } |
3416 | ||
320ae51f JA |
3417 | p += rq_size; |
3418 | i++; | |
3419 | } | |
3420 | } | |
cc71a6f4 | 3421 | return 0; |
320ae51f | 3422 | |
24d2f903 | 3423 | fail: |
cc71a6f4 JA |
3424 | blk_mq_free_rqs(set, tags, hctx_idx); |
3425 | return -ENOMEM; | |
320ae51f JA |
3426 | } |
3427 | ||
bf0beec0 ML |
3428 | struct rq_iter_data { |
3429 | struct blk_mq_hw_ctx *hctx; | |
3430 | bool has_rq; | |
3431 | }; | |
3432 | ||
2dd6532e | 3433 | static bool blk_mq_has_request(struct request *rq, void *data) |
bf0beec0 ML |
3434 | { |
3435 | struct rq_iter_data *iter_data = data; | |
3436 | ||
3437 | if (rq->mq_hctx != iter_data->hctx) | |
3438 | return true; | |
3439 | iter_data->has_rq = true; | |
3440 | return false; | |
3441 | } | |
3442 | ||
3443 | static bool blk_mq_hctx_has_requests(struct blk_mq_hw_ctx *hctx) | |
3444 | { | |
3445 | struct blk_mq_tags *tags = hctx->sched_tags ? | |
3446 | hctx->sched_tags : hctx->tags; | |
3447 | struct rq_iter_data data = { | |
3448 | .hctx = hctx, | |
3449 | }; | |
3450 | ||
3451 | blk_mq_all_tag_iter(tags, blk_mq_has_request, &data); | |
3452 | return data.has_rq; | |
3453 | } | |
3454 | ||
3455 | static inline bool blk_mq_last_cpu_in_hctx(unsigned int cpu, | |
3456 | struct blk_mq_hw_ctx *hctx) | |
3457 | { | |
9b51d9d8 | 3458 | if (cpumask_first_and(hctx->cpumask, cpu_online_mask) != cpu) |
bf0beec0 ML |
3459 | return false; |
3460 | if (cpumask_next_and(cpu, hctx->cpumask, cpu_online_mask) < nr_cpu_ids) | |
3461 | return false; | |
3462 | return true; | |
3463 | } | |
3464 | ||
3465 | static int blk_mq_hctx_notify_offline(unsigned int cpu, struct hlist_node *node) | |
3466 | { | |
3467 | struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node, | |
3468 | struct blk_mq_hw_ctx, cpuhp_online); | |
3469 | ||
3470 | if (!cpumask_test_cpu(cpu, hctx->cpumask) || | |
3471 | !blk_mq_last_cpu_in_hctx(cpu, hctx)) | |
3472 | return 0; | |
3473 | ||
3474 | /* | |
3475 | * Prevent new request from being allocated on the current hctx. | |
3476 | * | |
3477 | * The smp_mb__after_atomic() Pairs with the implied barrier in | |
3478 | * test_and_set_bit_lock in sbitmap_get(). Ensures the inactive flag is | |
3479 | * seen once we return from the tag allocator. | |
3480 | */ | |
3481 | set_bit(BLK_MQ_S_INACTIVE, &hctx->state); | |
3482 | smp_mb__after_atomic(); | |
3483 | ||
3484 | /* | |
3485 | * Try to grab a reference to the queue and wait for any outstanding | |
3486 | * requests. If we could not grab a reference the queue has been | |
3487 | * frozen and there are no requests. | |
3488 | */ | |
3489 | if (percpu_ref_tryget(&hctx->queue->q_usage_counter)) { | |
3490 | while (blk_mq_hctx_has_requests(hctx)) | |
3491 | msleep(5); | |
3492 | percpu_ref_put(&hctx->queue->q_usage_counter); | |
3493 | } | |
3494 | ||
3495 | return 0; | |
3496 | } | |
3497 | ||
3498 | static int blk_mq_hctx_notify_online(unsigned int cpu, struct hlist_node *node) | |
3499 | { | |
3500 | struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node, | |
3501 | struct blk_mq_hw_ctx, cpuhp_online); | |
3502 | ||
3503 | if (cpumask_test_cpu(cpu, hctx->cpumask)) | |
3504 | clear_bit(BLK_MQ_S_INACTIVE, &hctx->state); | |
3505 | return 0; | |
3506 | } | |
3507 | ||
e57690fe JA |
3508 | /* |
3509 | * 'cpu' is going away. splice any existing rq_list entries from this | |
3510 | * software queue to the hw queue dispatch list, and ensure that it | |
3511 | * gets run. | |
3512 | */ | |
9467f859 | 3513 | static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node) |
484b4061 | 3514 | { |
9467f859 | 3515 | struct blk_mq_hw_ctx *hctx; |
484b4061 JA |
3516 | struct blk_mq_ctx *ctx; |
3517 | LIST_HEAD(tmp); | |
c16d6b5a | 3518 | enum hctx_type type; |
484b4061 | 3519 | |
9467f859 | 3520 | hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead); |
bf0beec0 ML |
3521 | if (!cpumask_test_cpu(cpu, hctx->cpumask)) |
3522 | return 0; | |
3523 | ||
e57690fe | 3524 | ctx = __blk_mq_get_ctx(hctx->queue, cpu); |
c16d6b5a | 3525 | type = hctx->type; |
484b4061 JA |
3526 | |
3527 | spin_lock(&ctx->lock); | |
c16d6b5a ML |
3528 | if (!list_empty(&ctx->rq_lists[type])) { |
3529 | list_splice_init(&ctx->rq_lists[type], &tmp); | |
484b4061 JA |
3530 | blk_mq_hctx_clear_pending(hctx, ctx); |
3531 | } | |
3532 | spin_unlock(&ctx->lock); | |
3533 | ||
3534 | if (list_empty(&tmp)) | |
9467f859 | 3535 | return 0; |
484b4061 | 3536 | |
e57690fe JA |
3537 | spin_lock(&hctx->lock); |
3538 | list_splice_tail_init(&tmp, &hctx->dispatch); | |
3539 | spin_unlock(&hctx->lock); | |
484b4061 JA |
3540 | |
3541 | blk_mq_run_hw_queue(hctx, true); | |
9467f859 | 3542 | return 0; |
484b4061 JA |
3543 | } |
3544 | ||
9467f859 | 3545 | static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx) |
484b4061 | 3546 | { |
bf0beec0 ML |
3547 | if (!(hctx->flags & BLK_MQ_F_STACKING)) |
3548 | cpuhp_state_remove_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE, | |
3549 | &hctx->cpuhp_online); | |
9467f859 TG |
3550 | cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD, |
3551 | &hctx->cpuhp_dead); | |
484b4061 JA |
3552 | } |
3553 | ||
364b6181 ML |
3554 | /* |
3555 | * Before freeing hw queue, clearing the flush request reference in | |
3556 | * tags->rqs[] for avoiding potential UAF. | |
3557 | */ | |
3558 | static void blk_mq_clear_flush_rq_mapping(struct blk_mq_tags *tags, | |
3559 | unsigned int queue_depth, struct request *flush_rq) | |
3560 | { | |
3561 | int i; | |
3562 | unsigned long flags; | |
3563 | ||
3564 | /* The hw queue may not be mapped yet */ | |
3565 | if (!tags) | |
3566 | return; | |
3567 | ||
0a467d0f | 3568 | WARN_ON_ONCE(req_ref_read(flush_rq) != 0); |
364b6181 ML |
3569 | |
3570 | for (i = 0; i < queue_depth; i++) | |
3571 | cmpxchg(&tags->rqs[i], flush_rq, NULL); | |
3572 | ||
3573 | /* | |
3574 | * Wait until all pending iteration is done. | |
3575 | * | |
3576 | * Request reference is cleared and it is guaranteed to be observed | |
3577 | * after the ->lock is released. | |
3578 | */ | |
3579 | spin_lock_irqsave(&tags->lock, flags); | |
3580 | spin_unlock_irqrestore(&tags->lock, flags); | |
3581 | } | |
3582 | ||
c3b4afca | 3583 | /* hctx->ctxs will be freed in queue's release handler */ |
08e98fc6 ML |
3584 | static void blk_mq_exit_hctx(struct request_queue *q, |
3585 | struct blk_mq_tag_set *set, | |
3586 | struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx) | |
3587 | { | |
364b6181 ML |
3588 | struct request *flush_rq = hctx->fq->flush_rq; |
3589 | ||
8ab0b7dc ML |
3590 | if (blk_mq_hw_queue_mapped(hctx)) |
3591 | blk_mq_tag_idle(hctx); | |
08e98fc6 | 3592 | |
6cfeadbf ML |
3593 | if (blk_queue_init_done(q)) |
3594 | blk_mq_clear_flush_rq_mapping(set->tags[hctx_idx], | |
3595 | set->queue_depth, flush_rq); | |
f70ced09 | 3596 | if (set->ops->exit_request) |
364b6181 | 3597 | set->ops->exit_request(set, flush_rq, hctx_idx); |
f70ced09 | 3598 | |
08e98fc6 ML |
3599 | if (set->ops->exit_hctx) |
3600 | set->ops->exit_hctx(hctx, hctx_idx); | |
3601 | ||
9467f859 | 3602 | blk_mq_remove_cpuhp(hctx); |
2f8f1336 | 3603 | |
4e5cc99e ML |
3604 | xa_erase(&q->hctx_table, hctx_idx); |
3605 | ||
2f8f1336 ML |
3606 | spin_lock(&q->unused_hctx_lock); |
3607 | list_add(&hctx->hctx_list, &q->unused_hctx_list); | |
3608 | spin_unlock(&q->unused_hctx_lock); | |
08e98fc6 ML |
3609 | } |
3610 | ||
624dbe47 ML |
3611 | static void blk_mq_exit_hw_queues(struct request_queue *q, |
3612 | struct blk_mq_tag_set *set, int nr_queue) | |
3613 | { | |
3614 | struct blk_mq_hw_ctx *hctx; | |
4f481208 | 3615 | unsigned long i; |
624dbe47 ML |
3616 | |
3617 | queue_for_each_hw_ctx(q, hctx, i) { | |
3618 | if (i == nr_queue) | |
3619 | break; | |
08e98fc6 | 3620 | blk_mq_exit_hctx(q, set, hctx, i); |
624dbe47 | 3621 | } |
624dbe47 ML |
3622 | } |
3623 | ||
08e98fc6 ML |
3624 | static int blk_mq_init_hctx(struct request_queue *q, |
3625 | struct blk_mq_tag_set *set, | |
3626 | struct blk_mq_hw_ctx *hctx, unsigned hctx_idx) | |
320ae51f | 3627 | { |
7c6c5b7c ML |
3628 | hctx->queue_num = hctx_idx; |
3629 | ||
bf0beec0 ML |
3630 | if (!(hctx->flags & BLK_MQ_F_STACKING)) |
3631 | cpuhp_state_add_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE, | |
3632 | &hctx->cpuhp_online); | |
7c6c5b7c ML |
3633 | cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead); |
3634 | ||
3635 | hctx->tags = set->tags[hctx_idx]; | |
3636 | ||
3637 | if (set->ops->init_hctx && | |
3638 | set->ops->init_hctx(hctx, set->driver_data, hctx_idx)) | |
3639 | goto unregister_cpu_notifier; | |
08e98fc6 | 3640 | |
7c6c5b7c ML |
3641 | if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx, |
3642 | hctx->numa_node)) | |
3643 | goto exit_hctx; | |
4e5cc99e ML |
3644 | |
3645 | if (xa_insert(&q->hctx_table, hctx_idx, hctx, GFP_KERNEL)) | |
3646 | goto exit_flush_rq; | |
3647 | ||
7c6c5b7c ML |
3648 | return 0; |
3649 | ||
4e5cc99e ML |
3650 | exit_flush_rq: |
3651 | if (set->ops->exit_request) | |
3652 | set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx); | |
7c6c5b7c ML |
3653 | exit_hctx: |
3654 | if (set->ops->exit_hctx) | |
3655 | set->ops->exit_hctx(hctx, hctx_idx); | |
3656 | unregister_cpu_notifier: | |
3657 | blk_mq_remove_cpuhp(hctx); | |
3658 | return -1; | |
3659 | } | |
3660 | ||
3661 | static struct blk_mq_hw_ctx * | |
3662 | blk_mq_alloc_hctx(struct request_queue *q, struct blk_mq_tag_set *set, | |
3663 | int node) | |
3664 | { | |
3665 | struct blk_mq_hw_ctx *hctx; | |
3666 | gfp_t gfp = GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY; | |
3667 | ||
704b914f | 3668 | hctx = kzalloc_node(sizeof(struct blk_mq_hw_ctx), gfp, node); |
7c6c5b7c ML |
3669 | if (!hctx) |
3670 | goto fail_alloc_hctx; | |
3671 | ||
3672 | if (!zalloc_cpumask_var_node(&hctx->cpumask, gfp, node)) | |
3673 | goto free_hctx; | |
3674 | ||
3675 | atomic_set(&hctx->nr_active, 0); | |
08e98fc6 | 3676 | if (node == NUMA_NO_NODE) |
7c6c5b7c ML |
3677 | node = set->numa_node; |
3678 | hctx->numa_node = node; | |
08e98fc6 | 3679 | |
9f993737 | 3680 | INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn); |
08e98fc6 ML |
3681 | spin_lock_init(&hctx->lock); |
3682 | INIT_LIST_HEAD(&hctx->dispatch); | |
3683 | hctx->queue = q; | |
51db1c37 | 3684 | hctx->flags = set->flags & ~BLK_MQ_F_TAG_QUEUE_SHARED; |
08e98fc6 | 3685 | |
2f8f1336 ML |
3686 | INIT_LIST_HEAD(&hctx->hctx_list); |
3687 | ||
320ae51f | 3688 | /* |
08e98fc6 ML |
3689 | * Allocate space for all possible cpus to avoid allocation at |
3690 | * runtime | |
320ae51f | 3691 | */ |
d904bfa7 | 3692 | hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *), |
7c6c5b7c | 3693 | gfp, node); |
08e98fc6 | 3694 | if (!hctx->ctxs) |
7c6c5b7c | 3695 | goto free_cpumask; |
320ae51f | 3696 | |
5b202853 | 3697 | if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8), |
c548e62b | 3698 | gfp, node, false, false)) |
08e98fc6 | 3699 | goto free_ctxs; |
08e98fc6 | 3700 | hctx->nr_ctx = 0; |
320ae51f | 3701 | |
5815839b | 3702 | spin_lock_init(&hctx->dispatch_wait_lock); |
eb619fdb JA |
3703 | init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake); |
3704 | INIT_LIST_HEAD(&hctx->dispatch_wait.entry); | |
3705 | ||
754a1572 | 3706 | hctx->fq = blk_alloc_flush_queue(hctx->numa_node, set->cmd_size, gfp); |
f70ced09 | 3707 | if (!hctx->fq) |
7c6c5b7c | 3708 | goto free_bitmap; |
320ae51f | 3709 | |
7c6c5b7c | 3710 | blk_mq_hctx_kobj_init(hctx); |
6a83e74d | 3711 | |
7c6c5b7c | 3712 | return hctx; |
320ae51f | 3713 | |
08e98fc6 | 3714 | free_bitmap: |
88459642 | 3715 | sbitmap_free(&hctx->ctx_map); |
08e98fc6 ML |
3716 | free_ctxs: |
3717 | kfree(hctx->ctxs); | |
7c6c5b7c ML |
3718 | free_cpumask: |
3719 | free_cpumask_var(hctx->cpumask); | |
3720 | free_hctx: | |
3721 | kfree(hctx); | |
3722 | fail_alloc_hctx: | |
3723 | return NULL; | |
08e98fc6 | 3724 | } |
320ae51f | 3725 | |
320ae51f JA |
3726 | static void blk_mq_init_cpu_queues(struct request_queue *q, |
3727 | unsigned int nr_hw_queues) | |
3728 | { | |
b3c661b1 JA |
3729 | struct blk_mq_tag_set *set = q->tag_set; |
3730 | unsigned int i, j; | |
320ae51f JA |
3731 | |
3732 | for_each_possible_cpu(i) { | |
3733 | struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i); | |
3734 | struct blk_mq_hw_ctx *hctx; | |
c16d6b5a | 3735 | int k; |
320ae51f | 3736 | |
320ae51f JA |
3737 | __ctx->cpu = i; |
3738 | spin_lock_init(&__ctx->lock); | |
c16d6b5a ML |
3739 | for (k = HCTX_TYPE_DEFAULT; k < HCTX_MAX_TYPES; k++) |
3740 | INIT_LIST_HEAD(&__ctx->rq_lists[k]); | |
3741 | ||
320ae51f JA |
3742 | __ctx->queue = q; |
3743 | ||
320ae51f JA |
3744 | /* |
3745 | * Set local node, IFF we have more than one hw queue. If | |
3746 | * not, we remain on the home node of the device | |
3747 | */ | |
b3c661b1 JA |
3748 | for (j = 0; j < set->nr_maps; j++) { |
3749 | hctx = blk_mq_map_queue_type(q, j, i); | |
3750 | if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE) | |
576e85c5 | 3751 | hctx->numa_node = cpu_to_node(i); |
b3c661b1 | 3752 | } |
320ae51f JA |
3753 | } |
3754 | } | |
3755 | ||
63064be1 JG |
3756 | struct blk_mq_tags *blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set *set, |
3757 | unsigned int hctx_idx, | |
3758 | unsigned int depth) | |
cc71a6f4 | 3759 | { |
63064be1 JG |
3760 | struct blk_mq_tags *tags; |
3761 | int ret; | |
cc71a6f4 | 3762 | |
e155b0c2 | 3763 | tags = blk_mq_alloc_rq_map(set, hctx_idx, depth, set->reserved_tags); |
63064be1 JG |
3764 | if (!tags) |
3765 | return NULL; | |
cc71a6f4 | 3766 | |
63064be1 JG |
3767 | ret = blk_mq_alloc_rqs(set, tags, hctx_idx, depth); |
3768 | if (ret) { | |
e155b0c2 | 3769 | blk_mq_free_rq_map(tags); |
63064be1 JG |
3770 | return NULL; |
3771 | } | |
cc71a6f4 | 3772 | |
63064be1 | 3773 | return tags; |
cc71a6f4 JA |
3774 | } |
3775 | ||
63064be1 JG |
3776 | static bool __blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set *set, |
3777 | int hctx_idx) | |
cc71a6f4 | 3778 | { |
079a2e3e JG |
3779 | if (blk_mq_is_shared_tags(set->flags)) { |
3780 | set->tags[hctx_idx] = set->shared_tags; | |
1c0706a7 | 3781 | |
e155b0c2 | 3782 | return true; |
bd166ef1 | 3783 | } |
e155b0c2 | 3784 | |
63064be1 JG |
3785 | set->tags[hctx_idx] = blk_mq_alloc_map_and_rqs(set, hctx_idx, |
3786 | set->queue_depth); | |
3787 | ||
3788 | return set->tags[hctx_idx]; | |
cc71a6f4 JA |
3789 | } |
3790 | ||
645db34e JG |
3791 | void blk_mq_free_map_and_rqs(struct blk_mq_tag_set *set, |
3792 | struct blk_mq_tags *tags, | |
3793 | unsigned int hctx_idx) | |
cc71a6f4 | 3794 | { |
645db34e JG |
3795 | if (tags) { |
3796 | blk_mq_free_rqs(set, tags, hctx_idx); | |
e155b0c2 | 3797 | blk_mq_free_rq_map(tags); |
bd166ef1 | 3798 | } |
cc71a6f4 JA |
3799 | } |
3800 | ||
e155b0c2 JG |
3801 | static void __blk_mq_free_map_and_rqs(struct blk_mq_tag_set *set, |
3802 | unsigned int hctx_idx) | |
3803 | { | |
079a2e3e | 3804 | if (!blk_mq_is_shared_tags(set->flags)) |
e155b0c2 JG |
3805 | blk_mq_free_map_and_rqs(set, set->tags[hctx_idx], hctx_idx); |
3806 | ||
3807 | set->tags[hctx_idx] = NULL; | |
cc71a6f4 JA |
3808 | } |
3809 | ||
4b855ad3 | 3810 | static void blk_mq_map_swqueue(struct request_queue *q) |
320ae51f | 3811 | { |
4f481208 ML |
3812 | unsigned int j, hctx_idx; |
3813 | unsigned long i; | |
320ae51f JA |
3814 | struct blk_mq_hw_ctx *hctx; |
3815 | struct blk_mq_ctx *ctx; | |
2a34c087 | 3816 | struct blk_mq_tag_set *set = q->tag_set; |
320ae51f JA |
3817 | |
3818 | queue_for_each_hw_ctx(q, hctx, i) { | |
e4043dcf | 3819 | cpumask_clear(hctx->cpumask); |
320ae51f | 3820 | hctx->nr_ctx = 0; |
d416c92c | 3821 | hctx->dispatch_from = NULL; |
320ae51f JA |
3822 | } |
3823 | ||
3824 | /* | |
4b855ad3 | 3825 | * Map software to hardware queues. |
4412efec ML |
3826 | * |
3827 | * If the cpu isn't present, the cpu is mapped to first hctx. | |
320ae51f | 3828 | */ |
20e4d813 | 3829 | for_each_possible_cpu(i) { |
4412efec | 3830 | |
897bb0c7 | 3831 | ctx = per_cpu_ptr(q->queue_ctx, i); |
b3c661b1 | 3832 | for (j = 0; j < set->nr_maps; j++) { |
bb94aea1 JW |
3833 | if (!set->map[j].nr_queues) { |
3834 | ctx->hctxs[j] = blk_mq_map_queue_type(q, | |
3835 | HCTX_TYPE_DEFAULT, i); | |
e5edd5f2 | 3836 | continue; |
bb94aea1 | 3837 | } |
fd689871 ML |
3838 | hctx_idx = set->map[j].mq_map[i]; |
3839 | /* unmapped hw queue can be remapped after CPU topo changed */ | |
3840 | if (!set->tags[hctx_idx] && | |
63064be1 | 3841 | !__blk_mq_alloc_map_and_rqs(set, hctx_idx)) { |
fd689871 ML |
3842 | /* |
3843 | * If tags initialization fail for some hctx, | |
3844 | * that hctx won't be brought online. In this | |
3845 | * case, remap the current ctx to hctx[0] which | |
3846 | * is guaranteed to always have tags allocated | |
3847 | */ | |
3848 | set->map[j].mq_map[i] = 0; | |
3849 | } | |
e5edd5f2 | 3850 | |
b3c661b1 | 3851 | hctx = blk_mq_map_queue_type(q, j, i); |
8ccdf4a3 | 3852 | ctx->hctxs[j] = hctx; |
b3c661b1 JA |
3853 | /* |
3854 | * If the CPU is already set in the mask, then we've | |
3855 | * mapped this one already. This can happen if | |
3856 | * devices share queues across queue maps. | |
3857 | */ | |
3858 | if (cpumask_test_cpu(i, hctx->cpumask)) | |
3859 | continue; | |
3860 | ||
3861 | cpumask_set_cpu(i, hctx->cpumask); | |
3862 | hctx->type = j; | |
3863 | ctx->index_hw[hctx->type] = hctx->nr_ctx; | |
3864 | hctx->ctxs[hctx->nr_ctx++] = ctx; | |
3865 | ||
3866 | /* | |
3867 | * If the nr_ctx type overflows, we have exceeded the | |
3868 | * amount of sw queues we can support. | |
3869 | */ | |
3870 | BUG_ON(!hctx->nr_ctx); | |
3871 | } | |
bb94aea1 JW |
3872 | |
3873 | for (; j < HCTX_MAX_TYPES; j++) | |
3874 | ctx->hctxs[j] = blk_mq_map_queue_type(q, | |
3875 | HCTX_TYPE_DEFAULT, i); | |
320ae51f | 3876 | } |
506e931f JA |
3877 | |
3878 | queue_for_each_hw_ctx(q, hctx, i) { | |
4412efec ML |
3879 | /* |
3880 | * If no software queues are mapped to this hardware queue, | |
3881 | * disable it and free the request entries. | |
3882 | */ | |
3883 | if (!hctx->nr_ctx) { | |
3884 | /* Never unmap queue 0. We need it as a | |
3885 | * fallback in case of a new remap fails | |
3886 | * allocation | |
3887 | */ | |
e155b0c2 JG |
3888 | if (i) |
3889 | __blk_mq_free_map_and_rqs(set, i); | |
4412efec ML |
3890 | |
3891 | hctx->tags = NULL; | |
3892 | continue; | |
3893 | } | |
484b4061 | 3894 | |
2a34c087 ML |
3895 | hctx->tags = set->tags[i]; |
3896 | WARN_ON(!hctx->tags); | |
3897 | ||
889fa31f CY |
3898 | /* |
3899 | * Set the map size to the number of mapped software queues. | |
3900 | * This is more accurate and more efficient than looping | |
3901 | * over all possibly mapped software queues. | |
3902 | */ | |
88459642 | 3903 | sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx); |
889fa31f | 3904 | |
484b4061 JA |
3905 | /* |
3906 | * Initialize batch roundrobin counts | |
3907 | */ | |
f82ddf19 | 3908 | hctx->next_cpu = blk_mq_first_mapped_cpu(hctx); |
506e931f JA |
3909 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; |
3910 | } | |
320ae51f JA |
3911 | } |
3912 | ||
8e8320c9 JA |
3913 | /* |
3914 | * Caller needs to ensure that we're either frozen/quiesced, or that | |
3915 | * the queue isn't live yet. | |
3916 | */ | |
2404e607 | 3917 | static void queue_set_hctx_shared(struct request_queue *q, bool shared) |
0d2602ca JA |
3918 | { |
3919 | struct blk_mq_hw_ctx *hctx; | |
4f481208 | 3920 | unsigned long i; |
0d2602ca | 3921 | |
2404e607 | 3922 | queue_for_each_hw_ctx(q, hctx, i) { |
454bb677 | 3923 | if (shared) { |
51db1c37 | 3924 | hctx->flags |= BLK_MQ_F_TAG_QUEUE_SHARED; |
454bb677 YK |
3925 | } else { |
3926 | blk_mq_tag_idle(hctx); | |
51db1c37 | 3927 | hctx->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED; |
454bb677 | 3928 | } |
2404e607 JM |
3929 | } |
3930 | } | |
3931 | ||
655ac300 HR |
3932 | static void blk_mq_update_tag_set_shared(struct blk_mq_tag_set *set, |
3933 | bool shared) | |
2404e607 JM |
3934 | { |
3935 | struct request_queue *q; | |
0d2602ca | 3936 | |
705cda97 BVA |
3937 | lockdep_assert_held(&set->tag_list_lock); |
3938 | ||
0d2602ca JA |
3939 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
3940 | blk_mq_freeze_queue(q); | |
2404e607 | 3941 | queue_set_hctx_shared(q, shared); |
0d2602ca JA |
3942 | blk_mq_unfreeze_queue(q); |
3943 | } | |
3944 | } | |
3945 | ||
3946 | static void blk_mq_del_queue_tag_set(struct request_queue *q) | |
3947 | { | |
3948 | struct blk_mq_tag_set *set = q->tag_set; | |
3949 | ||
0d2602ca | 3950 | mutex_lock(&set->tag_list_lock); |
08c875cb | 3951 | list_del(&q->tag_set_list); |
2404e607 JM |
3952 | if (list_is_singular(&set->tag_list)) { |
3953 | /* just transitioned to unshared */ | |
51db1c37 | 3954 | set->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED; |
2404e607 | 3955 | /* update existing queue */ |
655ac300 | 3956 | blk_mq_update_tag_set_shared(set, false); |
2404e607 | 3957 | } |
0d2602ca | 3958 | mutex_unlock(&set->tag_list_lock); |
a347c7ad | 3959 | INIT_LIST_HEAD(&q->tag_set_list); |
0d2602ca JA |
3960 | } |
3961 | ||
3962 | static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set, | |
3963 | struct request_queue *q) | |
3964 | { | |
0d2602ca | 3965 | mutex_lock(&set->tag_list_lock); |
2404e607 | 3966 | |
ff821d27 JA |
3967 | /* |
3968 | * Check to see if we're transitioning to shared (from 1 to 2 queues). | |
3969 | */ | |
3970 | if (!list_empty(&set->tag_list) && | |
51db1c37 ML |
3971 | !(set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) { |
3972 | set->flags |= BLK_MQ_F_TAG_QUEUE_SHARED; | |
2404e607 | 3973 | /* update existing queue */ |
655ac300 | 3974 | blk_mq_update_tag_set_shared(set, true); |
2404e607 | 3975 | } |
51db1c37 | 3976 | if (set->flags & BLK_MQ_F_TAG_QUEUE_SHARED) |
2404e607 | 3977 | queue_set_hctx_shared(q, true); |
08c875cb | 3978 | list_add_tail(&q->tag_set_list, &set->tag_list); |
2404e607 | 3979 | |
0d2602ca JA |
3980 | mutex_unlock(&set->tag_list_lock); |
3981 | } | |
3982 | ||
1db4909e ML |
3983 | /* All allocations will be freed in release handler of q->mq_kobj */ |
3984 | static int blk_mq_alloc_ctxs(struct request_queue *q) | |
3985 | { | |
3986 | struct blk_mq_ctxs *ctxs; | |
3987 | int cpu; | |
3988 | ||
3989 | ctxs = kzalloc(sizeof(*ctxs), GFP_KERNEL); | |
3990 | if (!ctxs) | |
3991 | return -ENOMEM; | |
3992 | ||
3993 | ctxs->queue_ctx = alloc_percpu(struct blk_mq_ctx); | |
3994 | if (!ctxs->queue_ctx) | |
3995 | goto fail; | |
3996 | ||
3997 | for_each_possible_cpu(cpu) { | |
3998 | struct blk_mq_ctx *ctx = per_cpu_ptr(ctxs->queue_ctx, cpu); | |
3999 | ctx->ctxs = ctxs; | |
4000 | } | |
4001 | ||
4002 | q->mq_kobj = &ctxs->kobj; | |
4003 | q->queue_ctx = ctxs->queue_ctx; | |
4004 | ||
4005 | return 0; | |
4006 | fail: | |
4007 | kfree(ctxs); | |
4008 | return -ENOMEM; | |
4009 | } | |
4010 | ||
e09aae7e ML |
4011 | /* |
4012 | * It is the actual release handler for mq, but we do it from | |
4013 | * request queue's release handler for avoiding use-after-free | |
4014 | * and headache because q->mq_kobj shouldn't have been introduced, | |
4015 | * but we can't group ctx/kctx kobj without it. | |
4016 | */ | |
4017 | void blk_mq_release(struct request_queue *q) | |
4018 | { | |
2f8f1336 | 4019 | struct blk_mq_hw_ctx *hctx, *next; |
4f481208 | 4020 | unsigned long i; |
e09aae7e | 4021 | |
2f8f1336 ML |
4022 | queue_for_each_hw_ctx(q, hctx, i) |
4023 | WARN_ON_ONCE(hctx && list_empty(&hctx->hctx_list)); | |
4024 | ||
4025 | /* all hctx are in .unused_hctx_list now */ | |
4026 | list_for_each_entry_safe(hctx, next, &q->unused_hctx_list, hctx_list) { | |
4027 | list_del_init(&hctx->hctx_list); | |
6c8b232e | 4028 | kobject_put(&hctx->kobj); |
c3b4afca | 4029 | } |
e09aae7e | 4030 | |
4e5cc99e | 4031 | xa_destroy(&q->hctx_table); |
e09aae7e | 4032 | |
7ea5fe31 ML |
4033 | /* |
4034 | * release .mq_kobj and sw queue's kobject now because | |
4035 | * both share lifetime with request queue. | |
4036 | */ | |
4037 | blk_mq_sysfs_deinit(q); | |
e09aae7e ML |
4038 | } |
4039 | ||
5ec780a6 | 4040 | static struct request_queue *blk_mq_init_queue_data(struct blk_mq_tag_set *set, |
2f227bb9 | 4041 | void *queuedata) |
b62c21b7 | 4042 | { |
26a9750a CH |
4043 | struct request_queue *q; |
4044 | int ret; | |
b62c21b7 | 4045 | |
80bd4a7a | 4046 | q = blk_alloc_queue(set->numa_node); |
26a9750a | 4047 | if (!q) |
b62c21b7 | 4048 | return ERR_PTR(-ENOMEM); |
26a9750a CH |
4049 | q->queuedata = queuedata; |
4050 | ret = blk_mq_init_allocated_queue(set, q); | |
4051 | if (ret) { | |
6f8191fd | 4052 | blk_put_queue(q); |
26a9750a CH |
4053 | return ERR_PTR(ret); |
4054 | } | |
b62c21b7 MS |
4055 | return q; |
4056 | } | |
2f227bb9 CH |
4057 | |
4058 | struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set) | |
4059 | { | |
4060 | return blk_mq_init_queue_data(set, NULL); | |
4061 | } | |
b62c21b7 MS |
4062 | EXPORT_SYMBOL(blk_mq_init_queue); |
4063 | ||
6f8191fd CH |
4064 | /** |
4065 | * blk_mq_destroy_queue - shutdown a request queue | |
4066 | * @q: request queue to shutdown | |
4067 | * | |
4068 | * This shuts down a request queue allocated by blk_mq_init_queue() and drops | |
4069 | * the initial reference. All future requests will failed with -ENODEV. | |
4070 | * | |
4071 | * Context: can sleep | |
4072 | */ | |
4073 | void blk_mq_destroy_queue(struct request_queue *q) | |
4074 | { | |
4075 | WARN_ON_ONCE(!queue_is_mq(q)); | |
4076 | WARN_ON_ONCE(blk_queue_registered(q)); | |
4077 | ||
4078 | might_sleep(); | |
4079 | ||
4080 | blk_queue_flag_set(QUEUE_FLAG_DYING, q); | |
4081 | blk_queue_start_drain(q); | |
56c1ee92 | 4082 | blk_mq_freeze_queue_wait(q); |
6f8191fd CH |
4083 | |
4084 | blk_sync_queue(q); | |
4085 | blk_mq_cancel_work_sync(q); | |
4086 | blk_mq_exit_queue(q); | |
6f8191fd CH |
4087 | } |
4088 | EXPORT_SYMBOL(blk_mq_destroy_queue); | |
4089 | ||
4dcc4874 CH |
4090 | struct gendisk *__blk_mq_alloc_disk(struct blk_mq_tag_set *set, void *queuedata, |
4091 | struct lock_class_key *lkclass) | |
9316a9ed JA |
4092 | { |
4093 | struct request_queue *q; | |
b461dfc4 | 4094 | struct gendisk *disk; |
9316a9ed | 4095 | |
b461dfc4 CH |
4096 | q = blk_mq_init_queue_data(set, queuedata); |
4097 | if (IS_ERR(q)) | |
4098 | return ERR_CAST(q); | |
9316a9ed | 4099 | |
4a1fa41d | 4100 | disk = __alloc_disk_node(q, set->numa_node, lkclass); |
b461dfc4 | 4101 | if (!disk) { |
0a3e5cc7 | 4102 | blk_mq_destroy_queue(q); |
2b3f056f | 4103 | blk_put_queue(q); |
b461dfc4 | 4104 | return ERR_PTR(-ENOMEM); |
9316a9ed | 4105 | } |
6f8191fd | 4106 | set_bit(GD_OWNS_QUEUE, &disk->state); |
b461dfc4 | 4107 | return disk; |
9316a9ed | 4108 | } |
b461dfc4 | 4109 | EXPORT_SYMBOL(__blk_mq_alloc_disk); |
9316a9ed | 4110 | |
6f8191fd CH |
4111 | struct gendisk *blk_mq_alloc_disk_for_queue(struct request_queue *q, |
4112 | struct lock_class_key *lkclass) | |
4113 | { | |
22c17e27 CH |
4114 | struct gendisk *disk; |
4115 | ||
6f8191fd CH |
4116 | if (!blk_get_queue(q)) |
4117 | return NULL; | |
22c17e27 CH |
4118 | disk = __alloc_disk_node(q, NUMA_NO_NODE, lkclass); |
4119 | if (!disk) | |
4120 | blk_put_queue(q); | |
4121 | return disk; | |
6f8191fd CH |
4122 | } |
4123 | EXPORT_SYMBOL(blk_mq_alloc_disk_for_queue); | |
4124 | ||
34d11ffa JW |
4125 | static struct blk_mq_hw_ctx *blk_mq_alloc_and_init_hctx( |
4126 | struct blk_mq_tag_set *set, struct request_queue *q, | |
4127 | int hctx_idx, int node) | |
4128 | { | |
2f8f1336 | 4129 | struct blk_mq_hw_ctx *hctx = NULL, *tmp; |
34d11ffa | 4130 | |
2f8f1336 ML |
4131 | /* reuse dead hctx first */ |
4132 | spin_lock(&q->unused_hctx_lock); | |
4133 | list_for_each_entry(tmp, &q->unused_hctx_list, hctx_list) { | |
4134 | if (tmp->numa_node == node) { | |
4135 | hctx = tmp; | |
4136 | break; | |
4137 | } | |
4138 | } | |
4139 | if (hctx) | |
4140 | list_del_init(&hctx->hctx_list); | |
4141 | spin_unlock(&q->unused_hctx_lock); | |
4142 | ||
4143 | if (!hctx) | |
4144 | hctx = blk_mq_alloc_hctx(q, set, node); | |
34d11ffa | 4145 | if (!hctx) |
7c6c5b7c | 4146 | goto fail; |
34d11ffa | 4147 | |
7c6c5b7c ML |
4148 | if (blk_mq_init_hctx(q, set, hctx, hctx_idx)) |
4149 | goto free_hctx; | |
34d11ffa JW |
4150 | |
4151 | return hctx; | |
7c6c5b7c ML |
4152 | |
4153 | free_hctx: | |
4154 | kobject_put(&hctx->kobj); | |
4155 | fail: | |
4156 | return NULL; | |
34d11ffa JW |
4157 | } |
4158 | ||
868f2f0b KB |
4159 | static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set, |
4160 | struct request_queue *q) | |
320ae51f | 4161 | { |
4e5cc99e ML |
4162 | struct blk_mq_hw_ctx *hctx; |
4163 | unsigned long i, j; | |
ac0d6b92 | 4164 | |
fb350e0a ML |
4165 | /* protect against switching io scheduler */ |
4166 | mutex_lock(&q->sysfs_lock); | |
24d2f903 | 4167 | for (i = 0; i < set->nr_hw_queues; i++) { |
306f13ee | 4168 | int old_node; |
4d805131 | 4169 | int node = blk_mq_get_hctx_node(set, i); |
4e5cc99e | 4170 | struct blk_mq_hw_ctx *old_hctx = xa_load(&q->hctx_table, i); |
868f2f0b | 4171 | |
306f13ee ML |
4172 | if (old_hctx) { |
4173 | old_node = old_hctx->numa_node; | |
4174 | blk_mq_exit_hctx(q, set, old_hctx, i); | |
4175 | } | |
868f2f0b | 4176 | |
4e5cc99e | 4177 | if (!blk_mq_alloc_and_init_hctx(set, q, i, node)) { |
306f13ee | 4178 | if (!old_hctx) |
34d11ffa | 4179 | break; |
306f13ee ML |
4180 | pr_warn("Allocate new hctx on node %d fails, fallback to previous one on node %d\n", |
4181 | node, old_node); | |
4e5cc99e ML |
4182 | hctx = blk_mq_alloc_and_init_hctx(set, q, i, old_node); |
4183 | WARN_ON_ONCE(!hctx); | |
868f2f0b | 4184 | } |
320ae51f | 4185 | } |
e01ad46d JW |
4186 | /* |
4187 | * Increasing nr_hw_queues fails. Free the newly allocated | |
4188 | * hctxs and keep the previous q->nr_hw_queues. | |
4189 | */ | |
4190 | if (i != set->nr_hw_queues) { | |
4191 | j = q->nr_hw_queues; | |
e01ad46d JW |
4192 | } else { |
4193 | j = i; | |
e01ad46d JW |
4194 | q->nr_hw_queues = set->nr_hw_queues; |
4195 | } | |
34d11ffa | 4196 | |
4e5cc99e ML |
4197 | xa_for_each_start(&q->hctx_table, j, hctx, j) |
4198 | blk_mq_exit_hctx(q, set, hctx, j); | |
fb350e0a | 4199 | mutex_unlock(&q->sysfs_lock); |
868f2f0b KB |
4200 | } |
4201 | ||
42ee3061 ML |
4202 | static void blk_mq_update_poll_flag(struct request_queue *q) |
4203 | { | |
4204 | struct blk_mq_tag_set *set = q->tag_set; | |
4205 | ||
4206 | if (set->nr_maps > HCTX_TYPE_POLL && | |
4207 | set->map[HCTX_TYPE_POLL].nr_queues) | |
4208 | blk_queue_flag_set(QUEUE_FLAG_POLL, q); | |
4209 | else | |
4210 | blk_queue_flag_clear(QUEUE_FLAG_POLL, q); | |
4211 | } | |
4212 | ||
26a9750a CH |
4213 | int blk_mq_init_allocated_queue(struct blk_mq_tag_set *set, |
4214 | struct request_queue *q) | |
868f2f0b | 4215 | { |
66841672 ML |
4216 | /* mark the queue as mq asap */ |
4217 | q->mq_ops = set->ops; | |
4218 | ||
34dbad5d | 4219 | q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn, |
720b8ccc SB |
4220 | blk_mq_poll_stats_bkt, |
4221 | BLK_MQ_POLL_STATS_BKTS, q); | |
34dbad5d OS |
4222 | if (!q->poll_cb) |
4223 | goto err_exit; | |
4224 | ||
1db4909e | 4225 | if (blk_mq_alloc_ctxs(q)) |
41de54c6 | 4226 | goto err_poll; |
868f2f0b | 4227 | |
737f98cf ML |
4228 | /* init q->mq_kobj and sw queues' kobjects */ |
4229 | blk_mq_sysfs_init(q); | |
4230 | ||
2f8f1336 ML |
4231 | INIT_LIST_HEAD(&q->unused_hctx_list); |
4232 | spin_lock_init(&q->unused_hctx_lock); | |
4233 | ||
4e5cc99e ML |
4234 | xa_init(&q->hctx_table); |
4235 | ||
868f2f0b KB |
4236 | blk_mq_realloc_hw_ctxs(set, q); |
4237 | if (!q->nr_hw_queues) | |
4238 | goto err_hctxs; | |
320ae51f | 4239 | |
287922eb | 4240 | INIT_WORK(&q->timeout_work, blk_mq_timeout_work); |
e56f698b | 4241 | blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ); |
320ae51f | 4242 | |
a8908939 | 4243 | q->tag_set = set; |
320ae51f | 4244 | |
94eddfbe | 4245 | q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT; |
42ee3061 | 4246 | blk_mq_update_poll_flag(q); |
320ae51f | 4247 | |
2849450a | 4248 | INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work); |
6fca6a61 CH |
4249 | INIT_LIST_HEAD(&q->requeue_list); |
4250 | spin_lock_init(&q->requeue_lock); | |
4251 | ||
eba71768 JA |
4252 | q->nr_requests = set->queue_depth; |
4253 | ||
64f1c21e JA |
4254 | /* |
4255 | * Default to classic polling | |
4256 | */ | |
29ece8b4 | 4257 | q->poll_nsec = BLK_MQ_POLL_CLASSIC; |
64f1c21e | 4258 | |
24d2f903 | 4259 | blk_mq_init_cpu_queues(q, set->nr_hw_queues); |
0d2602ca | 4260 | blk_mq_add_queue_tag_set(set, q); |
4b855ad3 | 4261 | blk_mq_map_swqueue(q); |
26a9750a | 4262 | return 0; |
18741986 | 4263 | |
320ae51f | 4264 | err_hctxs: |
943f45b9 | 4265 | blk_mq_release(q); |
41de54c6 JS |
4266 | err_poll: |
4267 | blk_stat_free_callback(q->poll_cb); | |
4268 | q->poll_cb = NULL; | |
c7de5726 ML |
4269 | err_exit: |
4270 | q->mq_ops = NULL; | |
26a9750a | 4271 | return -ENOMEM; |
320ae51f | 4272 | } |
b62c21b7 | 4273 | EXPORT_SYMBOL(blk_mq_init_allocated_queue); |
320ae51f | 4274 | |
c7e2d94b ML |
4275 | /* tags can _not_ be used after returning from blk_mq_exit_queue */ |
4276 | void blk_mq_exit_queue(struct request_queue *q) | |
320ae51f | 4277 | { |
630ef623 | 4278 | struct blk_mq_tag_set *set = q->tag_set; |
320ae51f | 4279 | |
630ef623 | 4280 | /* Checks hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED. */ |
624dbe47 | 4281 | blk_mq_exit_hw_queues(q, set, set->nr_hw_queues); |
630ef623 BVA |
4282 | /* May clear BLK_MQ_F_TAG_QUEUE_SHARED in hctx->flags. */ |
4283 | blk_mq_del_queue_tag_set(q); | |
320ae51f | 4284 | } |
320ae51f | 4285 | |
a5164405 JA |
4286 | static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set) |
4287 | { | |
4288 | int i; | |
4289 | ||
079a2e3e JG |
4290 | if (blk_mq_is_shared_tags(set->flags)) { |
4291 | set->shared_tags = blk_mq_alloc_map_and_rqs(set, | |
e155b0c2 JG |
4292 | BLK_MQ_NO_HCTX_IDX, |
4293 | set->queue_depth); | |
079a2e3e | 4294 | if (!set->shared_tags) |
e155b0c2 JG |
4295 | return -ENOMEM; |
4296 | } | |
4297 | ||
8229cca8 | 4298 | for (i = 0; i < set->nr_hw_queues; i++) { |
63064be1 | 4299 | if (!__blk_mq_alloc_map_and_rqs(set, i)) |
a5164405 | 4300 | goto out_unwind; |
8229cca8 XT |
4301 | cond_resched(); |
4302 | } | |
a5164405 JA |
4303 | |
4304 | return 0; | |
4305 | ||
4306 | out_unwind: | |
4307 | while (--i >= 0) | |
e155b0c2 JG |
4308 | __blk_mq_free_map_and_rqs(set, i); |
4309 | ||
079a2e3e JG |
4310 | if (blk_mq_is_shared_tags(set->flags)) { |
4311 | blk_mq_free_map_and_rqs(set, set->shared_tags, | |
e155b0c2 | 4312 | BLK_MQ_NO_HCTX_IDX); |
645db34e | 4313 | } |
a5164405 | 4314 | |
a5164405 JA |
4315 | return -ENOMEM; |
4316 | } | |
4317 | ||
4318 | /* | |
4319 | * Allocate the request maps associated with this tag_set. Note that this | |
4320 | * may reduce the depth asked for, if memory is tight. set->queue_depth | |
4321 | * will be updated to reflect the allocated depth. | |
4322 | */ | |
63064be1 | 4323 | static int blk_mq_alloc_set_map_and_rqs(struct blk_mq_tag_set *set) |
a5164405 JA |
4324 | { |
4325 | unsigned int depth; | |
4326 | int err; | |
4327 | ||
4328 | depth = set->queue_depth; | |
4329 | do { | |
4330 | err = __blk_mq_alloc_rq_maps(set); | |
4331 | if (!err) | |
4332 | break; | |
4333 | ||
4334 | set->queue_depth >>= 1; | |
4335 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) { | |
4336 | err = -ENOMEM; | |
4337 | break; | |
4338 | } | |
4339 | } while (set->queue_depth); | |
4340 | ||
4341 | if (!set->queue_depth || err) { | |
4342 | pr_err("blk-mq: failed to allocate request map\n"); | |
4343 | return -ENOMEM; | |
4344 | } | |
4345 | ||
4346 | if (depth != set->queue_depth) | |
4347 | pr_info("blk-mq: reduced tag depth (%u -> %u)\n", | |
4348 | depth, set->queue_depth); | |
4349 | ||
4350 | return 0; | |
4351 | } | |
4352 | ||
a4e1d0b7 | 4353 | static void blk_mq_update_queue_map(struct blk_mq_tag_set *set) |
ebe8bddb | 4354 | { |
6e66b493 BVA |
4355 | /* |
4356 | * blk_mq_map_queues() and multiple .map_queues() implementations | |
4357 | * expect that set->map[HCTX_TYPE_DEFAULT].nr_queues is set to the | |
4358 | * number of hardware queues. | |
4359 | */ | |
4360 | if (set->nr_maps == 1) | |
4361 | set->map[HCTX_TYPE_DEFAULT].nr_queues = set->nr_hw_queues; | |
4362 | ||
59388702 | 4363 | if (set->ops->map_queues && !is_kdump_kernel()) { |
b3c661b1 JA |
4364 | int i; |
4365 | ||
7d4901a9 ML |
4366 | /* |
4367 | * transport .map_queues is usually done in the following | |
4368 | * way: | |
4369 | * | |
4370 | * for (queue = 0; queue < set->nr_hw_queues; queue++) { | |
4371 | * mask = get_cpu_mask(queue) | |
4372 | * for_each_cpu(cpu, mask) | |
b3c661b1 | 4373 | * set->map[x].mq_map[cpu] = queue; |
7d4901a9 ML |
4374 | * } |
4375 | * | |
4376 | * When we need to remap, the table has to be cleared for | |
4377 | * killing stale mapping since one CPU may not be mapped | |
4378 | * to any hw queue. | |
4379 | */ | |
b3c661b1 JA |
4380 | for (i = 0; i < set->nr_maps; i++) |
4381 | blk_mq_clear_mq_map(&set->map[i]); | |
7d4901a9 | 4382 | |
a4e1d0b7 | 4383 | set->ops->map_queues(set); |
b3c661b1 JA |
4384 | } else { |
4385 | BUG_ON(set->nr_maps > 1); | |
a4e1d0b7 | 4386 | blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]); |
b3c661b1 | 4387 | } |
ebe8bddb OS |
4388 | } |
4389 | ||
f7e76dbc | 4390 | static int blk_mq_realloc_tag_set_tags(struct blk_mq_tag_set *set, |
ee9d5521 | 4391 | int new_nr_hw_queues) |
f7e76dbc BVA |
4392 | { |
4393 | struct blk_mq_tags **new_tags; | |
4394 | ||
ee9d5521 | 4395 | if (set->nr_hw_queues >= new_nr_hw_queues) |
d4b2e0d4 | 4396 | goto done; |
f7e76dbc BVA |
4397 | |
4398 | new_tags = kcalloc_node(new_nr_hw_queues, sizeof(struct blk_mq_tags *), | |
4399 | GFP_KERNEL, set->numa_node); | |
4400 | if (!new_tags) | |
4401 | return -ENOMEM; | |
4402 | ||
4403 | if (set->tags) | |
ee9d5521 | 4404 | memcpy(new_tags, set->tags, set->nr_hw_queues * |
f7e76dbc BVA |
4405 | sizeof(*set->tags)); |
4406 | kfree(set->tags); | |
4407 | set->tags = new_tags; | |
d4b2e0d4 | 4408 | done: |
f7e76dbc | 4409 | set->nr_hw_queues = new_nr_hw_queues; |
f7e76dbc BVA |
4410 | return 0; |
4411 | } | |
4412 | ||
a4391c64 JA |
4413 | /* |
4414 | * Alloc a tag set to be associated with one or more request queues. | |
4415 | * May fail with EINVAL for various error conditions. May adjust the | |
c018c84f | 4416 | * requested depth down, if it's too large. In that case, the set |
a4391c64 JA |
4417 | * value will be stored in set->queue_depth. |
4418 | */ | |
24d2f903 CH |
4419 | int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set) |
4420 | { | |
b3c661b1 | 4421 | int i, ret; |
da695ba2 | 4422 | |
205fb5f5 BVA |
4423 | BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS); |
4424 | ||
24d2f903 CH |
4425 | if (!set->nr_hw_queues) |
4426 | return -EINVAL; | |
a4391c64 | 4427 | if (!set->queue_depth) |
24d2f903 CH |
4428 | return -EINVAL; |
4429 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) | |
4430 | return -EINVAL; | |
4431 | ||
7d7e0f90 | 4432 | if (!set->ops->queue_rq) |
24d2f903 CH |
4433 | return -EINVAL; |
4434 | ||
de148297 ML |
4435 | if (!set->ops->get_budget ^ !set->ops->put_budget) |
4436 | return -EINVAL; | |
4437 | ||
a4391c64 JA |
4438 | if (set->queue_depth > BLK_MQ_MAX_DEPTH) { |
4439 | pr_info("blk-mq: reduced tag depth to %u\n", | |
4440 | BLK_MQ_MAX_DEPTH); | |
4441 | set->queue_depth = BLK_MQ_MAX_DEPTH; | |
4442 | } | |
24d2f903 | 4443 | |
b3c661b1 JA |
4444 | if (!set->nr_maps) |
4445 | set->nr_maps = 1; | |
4446 | else if (set->nr_maps > HCTX_MAX_TYPES) | |
4447 | return -EINVAL; | |
4448 | ||
6637fadf SL |
4449 | /* |
4450 | * If a crashdump is active, then we are potentially in a very | |
4451 | * memory constrained environment. Limit us to 1 queue and | |
4452 | * 64 tags to prevent using too much memory. | |
4453 | */ | |
4454 | if (is_kdump_kernel()) { | |
4455 | set->nr_hw_queues = 1; | |
59388702 | 4456 | set->nr_maps = 1; |
6637fadf SL |
4457 | set->queue_depth = min(64U, set->queue_depth); |
4458 | } | |
868f2f0b | 4459 | /* |
392546ae JA |
4460 | * There is no use for more h/w queues than cpus if we just have |
4461 | * a single map | |
868f2f0b | 4462 | */ |
392546ae | 4463 | if (set->nr_maps == 1 && set->nr_hw_queues > nr_cpu_ids) |
868f2f0b | 4464 | set->nr_hw_queues = nr_cpu_ids; |
6637fadf | 4465 | |
80bd4a7a CH |
4466 | if (set->flags & BLK_MQ_F_BLOCKING) { |
4467 | set->srcu = kmalloc(sizeof(*set->srcu), GFP_KERNEL); | |
4468 | if (!set->srcu) | |
4469 | return -ENOMEM; | |
4470 | ret = init_srcu_struct(set->srcu); | |
4471 | if (ret) | |
4472 | goto out_free_srcu; | |
4473 | } | |
24d2f903 | 4474 | |
da695ba2 | 4475 | ret = -ENOMEM; |
5ee20298 CH |
4476 | set->tags = kcalloc_node(set->nr_hw_queues, |
4477 | sizeof(struct blk_mq_tags *), GFP_KERNEL, | |
4478 | set->numa_node); | |
4479 | if (!set->tags) | |
80bd4a7a | 4480 | goto out_cleanup_srcu; |
24d2f903 | 4481 | |
b3c661b1 JA |
4482 | for (i = 0; i < set->nr_maps; i++) { |
4483 | set->map[i].mq_map = kcalloc_node(nr_cpu_ids, | |
07b35eb5 | 4484 | sizeof(set->map[i].mq_map[0]), |
b3c661b1 JA |
4485 | GFP_KERNEL, set->numa_node); |
4486 | if (!set->map[i].mq_map) | |
4487 | goto out_free_mq_map; | |
59388702 | 4488 | set->map[i].nr_queues = is_kdump_kernel() ? 1 : set->nr_hw_queues; |
b3c661b1 | 4489 | } |
bdd17e75 | 4490 | |
a4e1d0b7 | 4491 | blk_mq_update_queue_map(set); |
da695ba2 | 4492 | |
63064be1 | 4493 | ret = blk_mq_alloc_set_map_and_rqs(set); |
da695ba2 | 4494 | if (ret) |
bdd17e75 | 4495 | goto out_free_mq_map; |
24d2f903 | 4496 | |
0d2602ca JA |
4497 | mutex_init(&set->tag_list_lock); |
4498 | INIT_LIST_HEAD(&set->tag_list); | |
4499 | ||
24d2f903 | 4500 | return 0; |
bdd17e75 CH |
4501 | |
4502 | out_free_mq_map: | |
b3c661b1 JA |
4503 | for (i = 0; i < set->nr_maps; i++) { |
4504 | kfree(set->map[i].mq_map); | |
4505 | set->map[i].mq_map = NULL; | |
4506 | } | |
5676e7b6 RE |
4507 | kfree(set->tags); |
4508 | set->tags = NULL; | |
80bd4a7a CH |
4509 | out_cleanup_srcu: |
4510 | if (set->flags & BLK_MQ_F_BLOCKING) | |
4511 | cleanup_srcu_struct(set->srcu); | |
4512 | out_free_srcu: | |
4513 | if (set->flags & BLK_MQ_F_BLOCKING) | |
4514 | kfree(set->srcu); | |
da695ba2 | 4515 | return ret; |
24d2f903 CH |
4516 | } |
4517 | EXPORT_SYMBOL(blk_mq_alloc_tag_set); | |
4518 | ||
cdb14e0f CH |
4519 | /* allocate and initialize a tagset for a simple single-queue device */ |
4520 | int blk_mq_alloc_sq_tag_set(struct blk_mq_tag_set *set, | |
4521 | const struct blk_mq_ops *ops, unsigned int queue_depth, | |
4522 | unsigned int set_flags) | |
4523 | { | |
4524 | memset(set, 0, sizeof(*set)); | |
4525 | set->ops = ops; | |
4526 | set->nr_hw_queues = 1; | |
4527 | set->nr_maps = 1; | |
4528 | set->queue_depth = queue_depth; | |
4529 | set->numa_node = NUMA_NO_NODE; | |
4530 | set->flags = set_flags; | |
4531 | return blk_mq_alloc_tag_set(set); | |
4532 | } | |
4533 | EXPORT_SYMBOL_GPL(blk_mq_alloc_sq_tag_set); | |
4534 | ||
24d2f903 CH |
4535 | void blk_mq_free_tag_set(struct blk_mq_tag_set *set) |
4536 | { | |
b3c661b1 | 4537 | int i, j; |
24d2f903 | 4538 | |
f7e76dbc | 4539 | for (i = 0; i < set->nr_hw_queues; i++) |
e155b0c2 | 4540 | __blk_mq_free_map_and_rqs(set, i); |
484b4061 | 4541 | |
079a2e3e JG |
4542 | if (blk_mq_is_shared_tags(set->flags)) { |
4543 | blk_mq_free_map_and_rqs(set, set->shared_tags, | |
e155b0c2 JG |
4544 | BLK_MQ_NO_HCTX_IDX); |
4545 | } | |
32bc15af | 4546 | |
b3c661b1 JA |
4547 | for (j = 0; j < set->nr_maps; j++) { |
4548 | kfree(set->map[j].mq_map); | |
4549 | set->map[j].mq_map = NULL; | |
4550 | } | |
bdd17e75 | 4551 | |
981bd189 | 4552 | kfree(set->tags); |
5676e7b6 | 4553 | set->tags = NULL; |
80bd4a7a CH |
4554 | if (set->flags & BLK_MQ_F_BLOCKING) { |
4555 | cleanup_srcu_struct(set->srcu); | |
4556 | kfree(set->srcu); | |
4557 | } | |
24d2f903 CH |
4558 | } |
4559 | EXPORT_SYMBOL(blk_mq_free_tag_set); | |
4560 | ||
e3a2b3f9 JA |
4561 | int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr) |
4562 | { | |
4563 | struct blk_mq_tag_set *set = q->tag_set; | |
4564 | struct blk_mq_hw_ctx *hctx; | |
4f481208 ML |
4565 | int ret; |
4566 | unsigned long i; | |
e3a2b3f9 | 4567 | |
bd166ef1 | 4568 | if (!set) |
e3a2b3f9 JA |
4569 | return -EINVAL; |
4570 | ||
e5fa8140 AZ |
4571 | if (q->nr_requests == nr) |
4572 | return 0; | |
4573 | ||
70f36b60 | 4574 | blk_mq_freeze_queue(q); |
24f5a90f | 4575 | blk_mq_quiesce_queue(q); |
70f36b60 | 4576 | |
e3a2b3f9 JA |
4577 | ret = 0; |
4578 | queue_for_each_hw_ctx(q, hctx, i) { | |
e9137d4b KB |
4579 | if (!hctx->tags) |
4580 | continue; | |
bd166ef1 JA |
4581 | /* |
4582 | * If we're using an MQ scheduler, just update the scheduler | |
4583 | * queue depth. This is similar to what the old code would do. | |
4584 | */ | |
f6adcef5 | 4585 | if (hctx->sched_tags) { |
70f36b60 | 4586 | ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags, |
f6adcef5 | 4587 | nr, true); |
f6adcef5 JG |
4588 | } else { |
4589 | ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr, | |
4590 | false); | |
70f36b60 | 4591 | } |
e3a2b3f9 JA |
4592 | if (ret) |
4593 | break; | |
77f1e0a5 JA |
4594 | if (q->elevator && q->elevator->type->ops.depth_updated) |
4595 | q->elevator->type->ops.depth_updated(hctx); | |
e3a2b3f9 | 4596 | } |
d97e594c | 4597 | if (!ret) { |
e3a2b3f9 | 4598 | q->nr_requests = nr; |
079a2e3e | 4599 | if (blk_mq_is_shared_tags(set->flags)) { |
8fa04464 | 4600 | if (q->elevator) |
079a2e3e | 4601 | blk_mq_tag_update_sched_shared_tags(q); |
8fa04464 | 4602 | else |
079a2e3e | 4603 | blk_mq_tag_resize_shared_tags(set, nr); |
8fa04464 | 4604 | } |
d97e594c | 4605 | } |
e3a2b3f9 | 4606 | |
24f5a90f | 4607 | blk_mq_unquiesce_queue(q); |
70f36b60 | 4608 | blk_mq_unfreeze_queue(q); |
70f36b60 | 4609 | |
e3a2b3f9 JA |
4610 | return ret; |
4611 | } | |
4612 | ||
d48ece20 JW |
4613 | /* |
4614 | * request_queue and elevator_type pair. | |
4615 | * It is just used by __blk_mq_update_nr_hw_queues to cache | |
4616 | * the elevator_type associated with a request_queue. | |
4617 | */ | |
4618 | struct blk_mq_qe_pair { | |
4619 | struct list_head node; | |
4620 | struct request_queue *q; | |
4621 | struct elevator_type *type; | |
4622 | }; | |
4623 | ||
4624 | /* | |
4625 | * Cache the elevator_type in qe pair list and switch the | |
4626 | * io scheduler to 'none' | |
4627 | */ | |
4628 | static bool blk_mq_elv_switch_none(struct list_head *head, | |
4629 | struct request_queue *q) | |
4630 | { | |
4631 | struct blk_mq_qe_pair *qe; | |
4632 | ||
4633 | if (!q->elevator) | |
4634 | return true; | |
4635 | ||
4636 | qe = kmalloc(sizeof(*qe), GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY); | |
4637 | if (!qe) | |
4638 | return false; | |
4639 | ||
5fd7a84a ML |
4640 | /* q->elevator needs protection from ->sysfs_lock */ |
4641 | mutex_lock(&q->sysfs_lock); | |
4642 | ||
d48ece20 JW |
4643 | INIT_LIST_HEAD(&qe->node); |
4644 | qe->q = q; | |
4645 | qe->type = q->elevator->type; | |
dd6f7f17 CH |
4646 | /* keep a reference to the elevator module as we'll switch back */ |
4647 | __elevator_get(qe->type); | |
d48ece20 | 4648 | list_add(&qe->node, head); |
64b36075 | 4649 | elevator_disable(q); |
d48ece20 JW |
4650 | mutex_unlock(&q->sysfs_lock); |
4651 | ||
4652 | return true; | |
4653 | } | |
4654 | ||
4a3b666e JK |
4655 | static struct blk_mq_qe_pair *blk_lookup_qe_pair(struct list_head *head, |
4656 | struct request_queue *q) | |
d48ece20 JW |
4657 | { |
4658 | struct blk_mq_qe_pair *qe; | |
d48ece20 JW |
4659 | |
4660 | list_for_each_entry(qe, head, node) | |
4a3b666e JK |
4661 | if (qe->q == q) |
4662 | return qe; | |
d48ece20 | 4663 | |
4a3b666e JK |
4664 | return NULL; |
4665 | } | |
d48ece20 | 4666 | |
4a3b666e JK |
4667 | static void blk_mq_elv_switch_back(struct list_head *head, |
4668 | struct request_queue *q) | |
4669 | { | |
4670 | struct blk_mq_qe_pair *qe; | |
4671 | struct elevator_type *t; | |
4672 | ||
4673 | qe = blk_lookup_qe_pair(head, q); | |
4674 | if (!qe) | |
4675 | return; | |
4676 | t = qe->type; | |
d48ece20 JW |
4677 | list_del(&qe->node); |
4678 | kfree(qe); | |
4679 | ||
4680 | mutex_lock(&q->sysfs_lock); | |
8237c01f | 4681 | elevator_switch(q, t); |
8ed40ee3 JC |
4682 | /* drop the reference acquired in blk_mq_elv_switch_none */ |
4683 | elevator_put(t); | |
d48ece20 JW |
4684 | mutex_unlock(&q->sysfs_lock); |
4685 | } | |
4686 | ||
e4dc2b32 KB |
4687 | static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, |
4688 | int nr_hw_queues) | |
868f2f0b KB |
4689 | { |
4690 | struct request_queue *q; | |
d48ece20 | 4691 | LIST_HEAD(head); |
e01ad46d | 4692 | int prev_nr_hw_queues; |
868f2f0b | 4693 | |
705cda97 BVA |
4694 | lockdep_assert_held(&set->tag_list_lock); |
4695 | ||
392546ae | 4696 | if (set->nr_maps == 1 && nr_hw_queues > nr_cpu_ids) |
868f2f0b | 4697 | nr_hw_queues = nr_cpu_ids; |
fe35ec58 WZ |
4698 | if (nr_hw_queues < 1) |
4699 | return; | |
4700 | if (set->nr_maps == 1 && nr_hw_queues == set->nr_hw_queues) | |
868f2f0b KB |
4701 | return; |
4702 | ||
4703 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
4704 | blk_mq_freeze_queue(q); | |
d48ece20 JW |
4705 | /* |
4706 | * Switch IO scheduler to 'none', cleaning up the data associated | |
4707 | * with the previous scheduler. We will switch back once we are done | |
4708 | * updating the new sw to hw queue mappings. | |
4709 | */ | |
4710 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
4711 | if (!blk_mq_elv_switch_none(&head, q)) | |
4712 | goto switch_back; | |
868f2f0b | 4713 | |
477e19de JW |
4714 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
4715 | blk_mq_debugfs_unregister_hctxs(q); | |
eaa870f9 | 4716 | blk_mq_sysfs_unregister_hctxs(q); |
477e19de JW |
4717 | } |
4718 | ||
a2584e43 | 4719 | prev_nr_hw_queues = set->nr_hw_queues; |
ee9d5521 | 4720 | if (blk_mq_realloc_tag_set_tags(set, nr_hw_queues) < 0) |
f7e76dbc BVA |
4721 | goto reregister; |
4722 | ||
e01ad46d | 4723 | fallback: |
aa880ad6 | 4724 | blk_mq_update_queue_map(set); |
868f2f0b KB |
4725 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
4726 | blk_mq_realloc_hw_ctxs(set, q); | |
42ee3061 | 4727 | blk_mq_update_poll_flag(q); |
e01ad46d | 4728 | if (q->nr_hw_queues != set->nr_hw_queues) { |
a846a8e6 YB |
4729 | int i = prev_nr_hw_queues; |
4730 | ||
e01ad46d JW |
4731 | pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n", |
4732 | nr_hw_queues, prev_nr_hw_queues); | |
a846a8e6 YB |
4733 | for (; i < set->nr_hw_queues; i++) |
4734 | __blk_mq_free_map_and_rqs(set, i); | |
4735 | ||
e01ad46d | 4736 | set->nr_hw_queues = prev_nr_hw_queues; |
7d76f856 | 4737 | blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]); |
e01ad46d JW |
4738 | goto fallback; |
4739 | } | |
477e19de JW |
4740 | blk_mq_map_swqueue(q); |
4741 | } | |
4742 | ||
f7e76dbc | 4743 | reregister: |
477e19de | 4744 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
eaa870f9 | 4745 | blk_mq_sysfs_register_hctxs(q); |
477e19de | 4746 | blk_mq_debugfs_register_hctxs(q); |
868f2f0b KB |
4747 | } |
4748 | ||
d48ece20 JW |
4749 | switch_back: |
4750 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
4751 | blk_mq_elv_switch_back(&head, q); | |
4752 | ||
868f2f0b KB |
4753 | list_for_each_entry(q, &set->tag_list, tag_set_list) |
4754 | blk_mq_unfreeze_queue(q); | |
4755 | } | |
e4dc2b32 KB |
4756 | |
4757 | void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues) | |
4758 | { | |
4759 | mutex_lock(&set->tag_list_lock); | |
4760 | __blk_mq_update_nr_hw_queues(set, nr_hw_queues); | |
4761 | mutex_unlock(&set->tag_list_lock); | |
4762 | } | |
868f2f0b KB |
4763 | EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues); |
4764 | ||
34dbad5d OS |
4765 | /* Enable polling stats and return whether they were already enabled. */ |
4766 | static bool blk_poll_stats_enable(struct request_queue *q) | |
4767 | { | |
48b5c1fb | 4768 | if (q->poll_stat) |
34dbad5d | 4769 | return true; |
48b5c1fb JA |
4770 | |
4771 | return blk_stats_alloc_enable(q); | |
34dbad5d OS |
4772 | } |
4773 | ||
4774 | static void blk_mq_poll_stats_start(struct request_queue *q) | |
4775 | { | |
4776 | /* | |
4777 | * We don't arm the callback if polling stats are not enabled or the | |
4778 | * callback is already active. | |
4779 | */ | |
48b5c1fb | 4780 | if (!q->poll_stat || blk_stat_is_active(q->poll_cb)) |
34dbad5d OS |
4781 | return; |
4782 | ||
4783 | blk_stat_activate_msecs(q->poll_cb, 100); | |
4784 | } | |
4785 | ||
4786 | static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb) | |
4787 | { | |
4788 | struct request_queue *q = cb->data; | |
720b8ccc | 4789 | int bucket; |
34dbad5d | 4790 | |
720b8ccc SB |
4791 | for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) { |
4792 | if (cb->stat[bucket].nr_samples) | |
4793 | q->poll_stat[bucket] = cb->stat[bucket]; | |
4794 | } | |
34dbad5d OS |
4795 | } |
4796 | ||
64f1c21e | 4797 | static unsigned long blk_mq_poll_nsecs(struct request_queue *q, |
64f1c21e JA |
4798 | struct request *rq) |
4799 | { | |
64f1c21e | 4800 | unsigned long ret = 0; |
720b8ccc | 4801 | int bucket; |
64f1c21e JA |
4802 | |
4803 | /* | |
4804 | * If stats collection isn't on, don't sleep but turn it on for | |
4805 | * future users | |
4806 | */ | |
34dbad5d | 4807 | if (!blk_poll_stats_enable(q)) |
64f1c21e JA |
4808 | return 0; |
4809 | ||
64f1c21e JA |
4810 | /* |
4811 | * As an optimistic guess, use half of the mean service time | |
4812 | * for this type of request. We can (and should) make this smarter. | |
4813 | * For instance, if the completion latencies are tight, we can | |
4814 | * get closer than just half the mean. This is especially | |
4815 | * important on devices where the completion latencies are longer | |
720b8ccc SB |
4816 | * than ~10 usec. We do use the stats for the relevant IO size |
4817 | * if available which does lead to better estimates. | |
64f1c21e | 4818 | */ |
720b8ccc SB |
4819 | bucket = blk_mq_poll_stats_bkt(rq); |
4820 | if (bucket < 0) | |
4821 | return ret; | |
4822 | ||
4823 | if (q->poll_stat[bucket].nr_samples) | |
4824 | ret = (q->poll_stat[bucket].mean + 1) / 2; | |
64f1c21e JA |
4825 | |
4826 | return ret; | |
4827 | } | |
4828 | ||
c6699d6f | 4829 | static bool blk_mq_poll_hybrid(struct request_queue *q, blk_qc_t qc) |
06426adf | 4830 | { |
c6699d6f CH |
4831 | struct blk_mq_hw_ctx *hctx = blk_qc_to_hctx(q, qc); |
4832 | struct request *rq = blk_qc_to_rq(hctx, qc); | |
06426adf JA |
4833 | struct hrtimer_sleeper hs; |
4834 | enum hrtimer_mode mode; | |
64f1c21e | 4835 | unsigned int nsecs; |
06426adf JA |
4836 | ktime_t kt; |
4837 | ||
c6699d6f CH |
4838 | /* |
4839 | * If a request has completed on queue that uses an I/O scheduler, we | |
4840 | * won't get back a request from blk_qc_to_rq. | |
4841 | */ | |
4842 | if (!rq || (rq->rq_flags & RQF_MQ_POLL_SLEPT)) | |
64f1c21e JA |
4843 | return false; |
4844 | ||
4845 | /* | |
1052b8ac | 4846 | * If we get here, hybrid polling is enabled. Hence poll_nsec can be: |
64f1c21e | 4847 | * |
64f1c21e JA |
4848 | * 0: use half of prev avg |
4849 | * >0: use this specific value | |
4850 | */ | |
1052b8ac | 4851 | if (q->poll_nsec > 0) |
64f1c21e JA |
4852 | nsecs = q->poll_nsec; |
4853 | else | |
cae740a0 | 4854 | nsecs = blk_mq_poll_nsecs(q, rq); |
64f1c21e JA |
4855 | |
4856 | if (!nsecs) | |
06426adf JA |
4857 | return false; |
4858 | ||
76a86f9d | 4859 | rq->rq_flags |= RQF_MQ_POLL_SLEPT; |
06426adf JA |
4860 | |
4861 | /* | |
4862 | * This will be replaced with the stats tracking code, using | |
4863 | * 'avg_completion_time / 2' as the pre-sleep target. | |
4864 | */ | |
8b0e1953 | 4865 | kt = nsecs; |
06426adf JA |
4866 | |
4867 | mode = HRTIMER_MODE_REL; | |
dbc1625f | 4868 | hrtimer_init_sleeper_on_stack(&hs, CLOCK_MONOTONIC, mode); |
06426adf JA |
4869 | hrtimer_set_expires(&hs.timer, kt); |
4870 | ||
06426adf | 4871 | do { |
5a61c363 | 4872 | if (blk_mq_rq_state(rq) == MQ_RQ_COMPLETE) |
06426adf JA |
4873 | break; |
4874 | set_current_state(TASK_UNINTERRUPTIBLE); | |
9dd8813e | 4875 | hrtimer_sleeper_start_expires(&hs, mode); |
06426adf JA |
4876 | if (hs.task) |
4877 | io_schedule(); | |
4878 | hrtimer_cancel(&hs.timer); | |
4879 | mode = HRTIMER_MODE_ABS; | |
4880 | } while (hs.task && !signal_pending(current)); | |
4881 | ||
4882 | __set_current_state(TASK_RUNNING); | |
4883 | destroy_hrtimer_on_stack(&hs.timer); | |
1052b8ac | 4884 | |
06426adf | 4885 | /* |
c6699d6f CH |
4886 | * If we sleep, have the caller restart the poll loop to reset the |
4887 | * state. Like for the other success return cases, the caller is | |
4888 | * responsible for checking if the IO completed. If the IO isn't | |
4889 | * complete, we'll get called again and will go straight to the busy | |
4890 | * poll loop. | |
06426adf | 4891 | */ |
06426adf JA |
4892 | return true; |
4893 | } | |
06426adf | 4894 | |
c6699d6f | 4895 | static int blk_mq_poll_classic(struct request_queue *q, blk_qc_t cookie, |
5a72e899 | 4896 | struct io_comp_batch *iob, unsigned int flags) |
bbd7bb70 | 4897 | { |
c6699d6f CH |
4898 | struct blk_mq_hw_ctx *hctx = blk_qc_to_hctx(q, cookie); |
4899 | long state = get_current_state(); | |
4900 | int ret; | |
bbd7bb70 | 4901 | |
aa61bec3 | 4902 | do { |
5a72e899 | 4903 | ret = q->mq_ops->poll(hctx, iob); |
bbd7bb70 | 4904 | if (ret > 0) { |
849a3700 | 4905 | __set_current_state(TASK_RUNNING); |
85f4d4b6 | 4906 | return ret; |
bbd7bb70 JA |
4907 | } |
4908 | ||
4909 | if (signal_pending_state(state, current)) | |
849a3700 | 4910 | __set_current_state(TASK_RUNNING); |
b03fbd4f | 4911 | if (task_is_running(current)) |
85f4d4b6 | 4912 | return 1; |
c6699d6f | 4913 | |
ef99b2d3 | 4914 | if (ret < 0 || (flags & BLK_POLL_ONESHOT)) |
bbd7bb70 JA |
4915 | break; |
4916 | cpu_relax(); | |
aa61bec3 | 4917 | } while (!need_resched()); |
bbd7bb70 | 4918 | |
67b4110f | 4919 | __set_current_state(TASK_RUNNING); |
85f4d4b6 | 4920 | return 0; |
bbd7bb70 | 4921 | } |
1052b8ac | 4922 | |
5a72e899 JA |
4923 | int blk_mq_poll(struct request_queue *q, blk_qc_t cookie, struct io_comp_batch *iob, |
4924 | unsigned int flags) | |
1052b8ac | 4925 | { |
d729cf9a CH |
4926 | if (!(flags & BLK_POLL_NOSLEEP) && |
4927 | q->poll_nsec != BLK_MQ_POLL_CLASSIC) { | |
c6699d6f | 4928 | if (blk_mq_poll_hybrid(q, cookie)) |
85f4d4b6 | 4929 | return 1; |
c6699d6f | 4930 | } |
5a72e899 | 4931 | return blk_mq_poll_classic(q, cookie, iob, flags); |
bbd7bb70 JA |
4932 | } |
4933 | ||
9cf2bab6 JA |
4934 | unsigned int blk_mq_rq_cpu(struct request *rq) |
4935 | { | |
4936 | return rq->mq_ctx->cpu; | |
4937 | } | |
4938 | EXPORT_SYMBOL(blk_mq_rq_cpu); | |
4939 | ||
2a19b28f ML |
4940 | void blk_mq_cancel_work_sync(struct request_queue *q) |
4941 | { | |
219cf43c JC |
4942 | struct blk_mq_hw_ctx *hctx; |
4943 | unsigned long i; | |
2a19b28f | 4944 | |
219cf43c | 4945 | cancel_delayed_work_sync(&q->requeue_work); |
2a19b28f | 4946 | |
219cf43c JC |
4947 | queue_for_each_hw_ctx(q, hctx, i) |
4948 | cancel_delayed_work_sync(&hctx->run_work); | |
2a19b28f ML |
4949 | } |
4950 | ||
320ae51f JA |
4951 | static int __init blk_mq_init(void) |
4952 | { | |
c3077b5d CH |
4953 | int i; |
4954 | ||
4955 | for_each_possible_cpu(i) | |
f9ab4918 | 4956 | init_llist_head(&per_cpu(blk_cpu_done, i)); |
c3077b5d CH |
4957 | open_softirq(BLOCK_SOFTIRQ, blk_done_softirq); |
4958 | ||
4959 | cpuhp_setup_state_nocalls(CPUHP_BLOCK_SOFTIRQ_DEAD, | |
4960 | "block/softirq:dead", NULL, | |
4961 | blk_softirq_cpu_dead); | |
9467f859 TG |
4962 | cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL, |
4963 | blk_mq_hctx_notify_dead); | |
bf0beec0 ML |
4964 | cpuhp_setup_state_multi(CPUHP_AP_BLK_MQ_ONLINE, "block/mq:online", |
4965 | blk_mq_hctx_notify_online, | |
4966 | blk_mq_hctx_notify_offline); | |
320ae51f JA |
4967 | return 0; |
4968 | } | |
4969 | subsys_initcall(blk_mq_init); |