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 | */ | |
6ee858a3 KS |
661 | if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT)) || |
662 | WARN_ON_ONCE(!(flags & BLK_MQ_REQ_RESERVED))) | |
1f5bd336 ML |
663 | return ERR_PTR(-EINVAL); |
664 | ||
665 | if (hctx_idx >= q->nr_hw_queues) | |
666 | return ERR_PTR(-EIO); | |
667 | ||
3a0a5299 | 668 | ret = blk_queue_enter(q, flags); |
1f5bd336 ML |
669 | if (ret) |
670 | return ERR_PTR(ret); | |
671 | ||
c8712c6a CH |
672 | /* |
673 | * Check if the hardware context is actually mapped to anything. | |
674 | * If not tell the caller that it should skip this queue. | |
675 | */ | |
a5ea5811 | 676 | ret = -EXDEV; |
4e5cc99e | 677 | data.hctx = xa_load(&q->hctx_table, hctx_idx); |
e6e7abff | 678 | if (!blk_mq_hw_queue_mapped(data.hctx)) |
a5ea5811 | 679 | goto out_queue_exit; |
e6e7abff | 680 | cpu = cpumask_first_and(data.hctx->cpumask, cpu_online_mask); |
14dc7a18 BVA |
681 | if (cpu >= nr_cpu_ids) |
682 | goto out_queue_exit; | |
e6e7abff | 683 | data.ctx = __blk_mq_get_ctx(q, cpu); |
1f5bd336 | 684 | |
42fdc5e4 | 685 | if (!q->elevator) |
600c3b0c | 686 | blk_mq_tag_busy(data.hctx); |
781dd830 JA |
687 | else |
688 | data.rq_flags |= RQF_ELV; | |
600c3b0c | 689 | |
99e48cd6 JG |
690 | if (flags & BLK_MQ_REQ_RESERVED) |
691 | data.rq_flags |= RQF_RESV; | |
692 | ||
a5ea5811 | 693 | ret = -EWOULDBLOCK; |
600c3b0c CH |
694 | tag = blk_mq_get_tag(&data); |
695 | if (tag == BLK_MQ_NO_TAG) | |
a5ea5811 | 696 | goto out_queue_exit; |
e3c5a78c | 697 | rq = blk_mq_rq_ctx_init(&data, blk_mq_tags_from_data(&data), tag, |
fe6134f6 | 698 | alloc_time_ns); |
e3c5a78c JG |
699 | rq->__data_len = 0; |
700 | rq->__sector = (sector_t) -1; | |
701 | rq->bio = rq->biotail = NULL; | |
702 | return rq; | |
600c3b0c | 703 | |
a5ea5811 CH |
704 | out_queue_exit: |
705 | blk_queue_exit(q); | |
706 | return ERR_PTR(ret); | |
1f5bd336 ML |
707 | } |
708 | EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx); | |
709 | ||
12f5b931 KB |
710 | static void __blk_mq_free_request(struct request *rq) |
711 | { | |
712 | struct request_queue *q = rq->q; | |
713 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
ea4f995e | 714 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; |
12f5b931 KB |
715 | const int sched_tag = rq->internal_tag; |
716 | ||
a892c8d5 | 717 | blk_crypto_free_request(rq); |
986d413b | 718 | blk_pm_mark_last_busy(rq); |
ea4f995e | 719 | rq->mq_hctx = NULL; |
76647368 | 720 | if (rq->tag != BLK_MQ_NO_TAG) |
cae740a0 | 721 | blk_mq_put_tag(hctx->tags, ctx, rq->tag); |
76647368 | 722 | if (sched_tag != BLK_MQ_NO_TAG) |
cae740a0 | 723 | blk_mq_put_tag(hctx->sched_tags, ctx, sched_tag); |
12f5b931 KB |
724 | blk_mq_sched_restart(hctx); |
725 | blk_queue_exit(q); | |
726 | } | |
727 | ||
6af54051 | 728 | void blk_mq_free_request(struct request *rq) |
320ae51f | 729 | { |
320ae51f | 730 | struct request_queue *q = rq->q; |
ea4f995e | 731 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; |
6af54051 | 732 | |
222ee581 CH |
733 | if ((rq->rq_flags & RQF_ELVPRIV) && |
734 | q->elevator->type->ops.finish_request) | |
735 | q->elevator->type->ops.finish_request(rq); | |
320ae51f | 736 | |
e8064021 | 737 | if (rq->rq_flags & RQF_MQ_INFLIGHT) |
bccf5e26 | 738 | __blk_mq_dec_active_requests(hctx); |
87760e5e | 739 | |
7beb2f84 | 740 | if (unlikely(laptop_mode && !blk_rq_is_passthrough(rq))) |
d152c682 | 741 | laptop_io_completion(q->disk->bdi); |
7beb2f84 | 742 | |
a7905043 | 743 | rq_qos_done(q, rq); |
0d2602ca | 744 | |
12f5b931 | 745 | WRITE_ONCE(rq->state, MQ_RQ_IDLE); |
0a467d0f | 746 | if (req_ref_put_and_test(rq)) |
12f5b931 | 747 | __blk_mq_free_request(rq); |
320ae51f | 748 | } |
1a3b595a | 749 | EXPORT_SYMBOL_GPL(blk_mq_free_request); |
320ae51f | 750 | |
47c122e3 | 751 | void blk_mq_free_plug_rqs(struct blk_plug *plug) |
320ae51f | 752 | { |
013a7f95 | 753 | struct request *rq; |
fe1f4526 | 754 | |
c5fc7b93 | 755 | while ((rq = rq_list_pop(&plug->cached_rq)) != NULL) |
47c122e3 | 756 | blk_mq_free_request(rq); |
47c122e3 | 757 | } |
522a7775 | 758 | |
22350ad7 CH |
759 | void blk_dump_rq_flags(struct request *rq, char *msg) |
760 | { | |
761 | printk(KERN_INFO "%s: dev %s: flags=%llx\n", msg, | |
f3fa33ac | 762 | rq->q->disk ? rq->q->disk->disk_name : "?", |
16458cf3 | 763 | (__force unsigned long long) rq->cmd_flags); |
22350ad7 CH |
764 | |
765 | printk(KERN_INFO " sector %llu, nr/cnr %u/%u\n", | |
766 | (unsigned long long)blk_rq_pos(rq), | |
767 | blk_rq_sectors(rq), blk_rq_cur_sectors(rq)); | |
768 | printk(KERN_INFO " bio %p, biotail %p, len %u\n", | |
769 | rq->bio, rq->biotail, blk_rq_bytes(rq)); | |
770 | } | |
771 | EXPORT_SYMBOL(blk_dump_rq_flags); | |
772 | ||
9be3e06f JA |
773 | static void req_bio_endio(struct request *rq, struct bio *bio, |
774 | unsigned int nbytes, blk_status_t error) | |
775 | { | |
478eb72b | 776 | if (unlikely(error)) { |
9be3e06f | 777 | bio->bi_status = error; |
478eb72b | 778 | } else if (req_op(rq) == REQ_OP_ZONE_APPEND) { |
9be3e06f JA |
779 | /* |
780 | * Partial zone append completions cannot be supported as the | |
781 | * BIO fragments may end up not being written sequentially. | |
782 | */ | |
297db731 | 783 | if (bio->bi_iter.bi_size != nbytes) |
9be3e06f JA |
784 | bio->bi_status = BLK_STS_IOERR; |
785 | else | |
786 | bio->bi_iter.bi_sector = rq->__sector; | |
787 | } | |
788 | ||
478eb72b PB |
789 | bio_advance(bio, nbytes); |
790 | ||
791 | if (unlikely(rq->rq_flags & RQF_QUIET)) | |
792 | bio_set_flag(bio, BIO_QUIET); | |
9be3e06f JA |
793 | /* don't actually finish bio if it's part of flush sequence */ |
794 | if (bio->bi_iter.bi_size == 0 && !(rq->rq_flags & RQF_FLUSH_SEQ)) | |
795 | bio_endio(bio); | |
796 | } | |
797 | ||
798 | static void blk_account_io_completion(struct request *req, unsigned int bytes) | |
799 | { | |
800 | if (req->part && blk_do_io_stat(req)) { | |
801 | const int sgrp = op_stat_group(req_op(req)); | |
802 | ||
803 | part_stat_lock(); | |
804 | part_stat_add(req->part, sectors[sgrp], bytes >> 9); | |
805 | part_stat_unlock(); | |
806 | } | |
807 | } | |
808 | ||
0d7a29a2 CH |
809 | static void blk_print_req_error(struct request *req, blk_status_t status) |
810 | { | |
811 | printk_ratelimited(KERN_ERR | |
812 | "%s error, dev %s, sector %llu op 0x%x:(%s) flags 0x%x " | |
813 | "phys_seg %u prio class %u\n", | |
814 | blk_status_to_str(status), | |
f3fa33ac | 815 | req->q->disk ? req->q->disk->disk_name : "?", |
16458cf3 BVA |
816 | blk_rq_pos(req), (__force u32)req_op(req), |
817 | blk_op_str(req_op(req)), | |
818 | (__force u32)(req->cmd_flags & ~REQ_OP_MASK), | |
0d7a29a2 CH |
819 | req->nr_phys_segments, |
820 | IOPRIO_PRIO_CLASS(req->ioprio)); | |
821 | } | |
822 | ||
5581a5dd JA |
823 | /* |
824 | * Fully end IO on a request. Does not support partial completions, or | |
825 | * errors. | |
826 | */ | |
827 | static void blk_complete_request(struct request *req) | |
828 | { | |
829 | const bool is_flush = (req->rq_flags & RQF_FLUSH_SEQ) != 0; | |
830 | int total_bytes = blk_rq_bytes(req); | |
831 | struct bio *bio = req->bio; | |
832 | ||
833 | trace_block_rq_complete(req, BLK_STS_OK, total_bytes); | |
834 | ||
835 | if (!bio) | |
836 | return; | |
837 | ||
838 | #ifdef CONFIG_BLK_DEV_INTEGRITY | |
839 | if (blk_integrity_rq(req) && req_op(req) == REQ_OP_READ) | |
840 | req->q->integrity.profile->complete_fn(req, total_bytes); | |
841 | #endif | |
842 | ||
843 | blk_account_io_completion(req, total_bytes); | |
844 | ||
845 | do { | |
846 | struct bio *next = bio->bi_next; | |
847 | ||
848 | /* Completion has already been traced */ | |
849 | bio_clear_flag(bio, BIO_TRACE_COMPLETION); | |
a12821d5 PR |
850 | |
851 | if (req_op(req) == REQ_OP_ZONE_APPEND) | |
852 | bio->bi_iter.bi_sector = req->__sector; | |
853 | ||
5581a5dd JA |
854 | if (!is_flush) |
855 | bio_endio(bio); | |
856 | bio = next; | |
857 | } while (bio); | |
858 | ||
859 | /* | |
860 | * Reset counters so that the request stacking driver | |
861 | * can find how many bytes remain in the request | |
862 | * later. | |
863 | */ | |
ab3e1d3b JA |
864 | if (!req->end_io) { |
865 | req->bio = NULL; | |
866 | req->__data_len = 0; | |
867 | } | |
5581a5dd JA |
868 | } |
869 | ||
9be3e06f JA |
870 | /** |
871 | * blk_update_request - Complete multiple bytes without completing the request | |
872 | * @req: the request being processed | |
873 | * @error: block status code | |
874 | * @nr_bytes: number of bytes to complete for @req | |
875 | * | |
876 | * Description: | |
877 | * Ends I/O on a number of bytes attached to @req, but doesn't complete | |
878 | * the request structure even if @req doesn't have leftover. | |
879 | * If @req has leftover, sets it up for the next range of segments. | |
880 | * | |
881 | * Passing the result of blk_rq_bytes() as @nr_bytes guarantees | |
882 | * %false return from this function. | |
883 | * | |
884 | * Note: | |
885 | * The RQF_SPECIAL_PAYLOAD flag is ignored on purpose in this function | |
886 | * except in the consistency check at the end of this function. | |
887 | * | |
888 | * Return: | |
889 | * %false - this request doesn't have any more data | |
890 | * %true - this request has more data | |
891 | **/ | |
892 | bool blk_update_request(struct request *req, blk_status_t error, | |
893 | unsigned int nr_bytes) | |
894 | { | |
895 | int total_bytes; | |
896 | ||
8a7d267b | 897 | trace_block_rq_complete(req, error, nr_bytes); |
9be3e06f JA |
898 | |
899 | if (!req->bio) | |
900 | return false; | |
901 | ||
902 | #ifdef CONFIG_BLK_DEV_INTEGRITY | |
903 | if (blk_integrity_rq(req) && req_op(req) == REQ_OP_READ && | |
904 | error == BLK_STS_OK) | |
905 | req->q->integrity.profile->complete_fn(req, nr_bytes); | |
906 | #endif | |
907 | ||
908 | if (unlikely(error && !blk_rq_is_passthrough(req) && | |
3d973a76 CH |
909 | !(req->rq_flags & RQF_QUIET)) && |
910 | !test_bit(GD_DEAD, &req->q->disk->state)) { | |
9be3e06f | 911 | blk_print_req_error(req, error); |
d5869fdc YS |
912 | trace_block_rq_error(req, error, nr_bytes); |
913 | } | |
9be3e06f JA |
914 | |
915 | blk_account_io_completion(req, nr_bytes); | |
916 | ||
917 | total_bytes = 0; | |
918 | while (req->bio) { | |
919 | struct bio *bio = req->bio; | |
920 | unsigned bio_bytes = min(bio->bi_iter.bi_size, nr_bytes); | |
921 | ||
922 | if (bio_bytes == bio->bi_iter.bi_size) | |
923 | req->bio = bio->bi_next; | |
924 | ||
925 | /* Completion has already been traced */ | |
926 | bio_clear_flag(bio, BIO_TRACE_COMPLETION); | |
927 | req_bio_endio(req, bio, bio_bytes, error); | |
928 | ||
929 | total_bytes += bio_bytes; | |
930 | nr_bytes -= bio_bytes; | |
931 | ||
932 | if (!nr_bytes) | |
933 | break; | |
934 | } | |
935 | ||
936 | /* | |
937 | * completely done | |
938 | */ | |
939 | if (!req->bio) { | |
940 | /* | |
941 | * Reset counters so that the request stacking driver | |
942 | * can find how many bytes remain in the request | |
943 | * later. | |
944 | */ | |
945 | req->__data_len = 0; | |
946 | return false; | |
947 | } | |
948 | ||
949 | req->__data_len -= total_bytes; | |
950 | ||
951 | /* update sector only for requests with clear definition of sector */ | |
952 | if (!blk_rq_is_passthrough(req)) | |
953 | req->__sector += total_bytes >> 9; | |
954 | ||
955 | /* mixed attributes always follow the first bio */ | |
956 | if (req->rq_flags & RQF_MIXED_MERGE) { | |
957 | req->cmd_flags &= ~REQ_FAILFAST_MASK; | |
958 | req->cmd_flags |= req->bio->bi_opf & REQ_FAILFAST_MASK; | |
959 | } | |
960 | ||
961 | if (!(req->rq_flags & RQF_SPECIAL_PAYLOAD)) { | |
962 | /* | |
963 | * If total number of sectors is less than the first segment | |
964 | * size, something has gone terribly wrong. | |
965 | */ | |
966 | if (blk_rq_bytes(req) < blk_rq_cur_bytes(req)) { | |
967 | blk_dump_rq_flags(req, "request botched"); | |
968 | req->__data_len = blk_rq_cur_bytes(req); | |
969 | } | |
970 | ||
971 | /* recalculate the number of segments */ | |
972 | req->nr_phys_segments = blk_recalc_rq_segments(req); | |
973 | } | |
974 | ||
975 | return true; | |
976 | } | |
977 | EXPORT_SYMBOL_GPL(blk_update_request); | |
978 | ||
450b7879 CH |
979 | static void __blk_account_io_done(struct request *req, u64 now) |
980 | { | |
981 | const int sgrp = op_stat_group(req_op(req)); | |
982 | ||
983 | part_stat_lock(); | |
984 | update_io_ticks(req->part, jiffies, true); | |
985 | part_stat_inc(req->part, ios[sgrp]); | |
986 | part_stat_add(req->part, nsecs[sgrp], now - req->start_time_ns); | |
987 | part_stat_unlock(); | |
988 | } | |
989 | ||
990 | static inline void blk_account_io_done(struct request *req, u64 now) | |
991 | { | |
992 | /* | |
993 | * Account IO completion. flush_rq isn't accounted as a | |
994 | * normal IO on queueing nor completion. Accounting the | |
995 | * containing request is enough. | |
996 | */ | |
997 | if (blk_do_io_stat(req) && req->part && | |
998 | !(req->rq_flags & RQF_FLUSH_SEQ)) | |
999 | __blk_account_io_done(req, now); | |
1000 | } | |
1001 | ||
1002 | static void __blk_account_io_start(struct request *rq) | |
1003 | { | |
41fa7222 CH |
1004 | /* |
1005 | * All non-passthrough requests are created from a bio with one | |
1006 | * exception: when a flush command that is part of a flush sequence | |
1007 | * generated by the state machine in blk-flush.c is cloned onto the | |
1008 | * lower device by dm-multipath we can get here without a bio. | |
1009 | */ | |
1010 | if (rq->bio) | |
450b7879 | 1011 | rq->part = rq->bio->bi_bdev; |
41fa7222 | 1012 | else |
f3fa33ac | 1013 | rq->part = rq->q->disk->part0; |
450b7879 CH |
1014 | |
1015 | part_stat_lock(); | |
1016 | update_io_ticks(rq->part, jiffies, false); | |
1017 | part_stat_unlock(); | |
1018 | } | |
1019 | ||
1020 | static inline void blk_account_io_start(struct request *req) | |
1021 | { | |
1022 | if (blk_do_io_stat(req)) | |
1023 | __blk_account_io_start(req); | |
1024 | } | |
1025 | ||
f794f335 | 1026 | static inline void __blk_mq_end_request_acct(struct request *rq, u64 now) |
320ae51f | 1027 | { |
4bc6339a OS |
1028 | if (rq->rq_flags & RQF_STATS) { |
1029 | blk_mq_poll_stats_start(rq->q); | |
522a7775 | 1030 | blk_stat_add(rq, now); |
4bc6339a OS |
1031 | } |
1032 | ||
87890092 | 1033 | blk_mq_sched_completed_request(rq, now); |
522a7775 | 1034 | blk_account_io_done(rq, now); |
f794f335 | 1035 | } |
522a7775 | 1036 | |
f794f335 JA |
1037 | inline void __blk_mq_end_request(struct request *rq, blk_status_t error) |
1038 | { | |
1039 | if (blk_mq_need_time_stamp(rq)) | |
1040 | __blk_mq_end_request_acct(rq, ktime_get_ns()); | |
0d11e6ac | 1041 | |
91b63639 | 1042 | if (rq->end_io) { |
a7905043 | 1043 | rq_qos_done(rq->q, rq); |
de671d61 JA |
1044 | if (rq->end_io(rq, error) == RQ_END_IO_FREE) |
1045 | blk_mq_free_request(rq); | |
91b63639 | 1046 | } else { |
320ae51f | 1047 | blk_mq_free_request(rq); |
91b63639 | 1048 | } |
320ae51f | 1049 | } |
c8a446ad | 1050 | EXPORT_SYMBOL(__blk_mq_end_request); |
63151a44 | 1051 | |
2a842aca | 1052 | void blk_mq_end_request(struct request *rq, blk_status_t error) |
63151a44 CH |
1053 | { |
1054 | if (blk_update_request(rq, error, blk_rq_bytes(rq))) | |
1055 | BUG(); | |
c8a446ad | 1056 | __blk_mq_end_request(rq, error); |
63151a44 | 1057 | } |
c8a446ad | 1058 | EXPORT_SYMBOL(blk_mq_end_request); |
320ae51f | 1059 | |
f794f335 JA |
1060 | #define TAG_COMP_BATCH 32 |
1061 | ||
1062 | static inline void blk_mq_flush_tag_batch(struct blk_mq_hw_ctx *hctx, | |
1063 | int *tag_array, int nr_tags) | |
1064 | { | |
1065 | struct request_queue *q = hctx->queue; | |
1066 | ||
3b87c6ea ML |
1067 | /* |
1068 | * All requests should have been marked as RQF_MQ_INFLIGHT, so | |
1069 | * update hctx->nr_active in batch | |
1070 | */ | |
1071 | if (hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) | |
1072 | __blk_mq_sub_active_requests(hctx, nr_tags); | |
1073 | ||
f794f335 JA |
1074 | blk_mq_put_tags(hctx->tags, tag_array, nr_tags); |
1075 | percpu_ref_put_many(&q->q_usage_counter, nr_tags); | |
1076 | } | |
1077 | ||
1078 | void blk_mq_end_request_batch(struct io_comp_batch *iob) | |
1079 | { | |
1080 | int tags[TAG_COMP_BATCH], nr_tags = 0; | |
02f7eab0 | 1081 | struct blk_mq_hw_ctx *cur_hctx = NULL; |
f794f335 JA |
1082 | struct request *rq; |
1083 | u64 now = 0; | |
1084 | ||
1085 | if (iob->need_ts) | |
1086 | now = ktime_get_ns(); | |
1087 | ||
1088 | while ((rq = rq_list_pop(&iob->req_list)) != NULL) { | |
1089 | prefetch(rq->bio); | |
1090 | prefetch(rq->rq_next); | |
1091 | ||
5581a5dd | 1092 | blk_complete_request(rq); |
f794f335 JA |
1093 | if (iob->need_ts) |
1094 | __blk_mq_end_request_acct(rq, now); | |
1095 | ||
98b26a0e JA |
1096 | rq_qos_done(rq->q, rq); |
1097 | ||
ab3e1d3b JA |
1098 | /* |
1099 | * If end_io handler returns NONE, then it still has | |
1100 | * ownership of the request. | |
1101 | */ | |
1102 | if (rq->end_io && rq->end_io(rq, 0) == RQ_END_IO_NONE) | |
1103 | continue; | |
1104 | ||
f794f335 | 1105 | WRITE_ONCE(rq->state, MQ_RQ_IDLE); |
0a467d0f | 1106 | if (!req_ref_put_and_test(rq)) |
f794f335 JA |
1107 | continue; |
1108 | ||
1109 | blk_crypto_free_request(rq); | |
1110 | blk_pm_mark_last_busy(rq); | |
f794f335 | 1111 | |
02f7eab0 JA |
1112 | if (nr_tags == TAG_COMP_BATCH || cur_hctx != rq->mq_hctx) { |
1113 | if (cur_hctx) | |
1114 | blk_mq_flush_tag_batch(cur_hctx, tags, nr_tags); | |
f794f335 | 1115 | nr_tags = 0; |
02f7eab0 | 1116 | cur_hctx = rq->mq_hctx; |
f794f335 JA |
1117 | } |
1118 | tags[nr_tags++] = rq->tag; | |
f794f335 JA |
1119 | } |
1120 | ||
1121 | if (nr_tags) | |
02f7eab0 | 1122 | blk_mq_flush_tag_batch(cur_hctx, tags, nr_tags); |
f794f335 JA |
1123 | } |
1124 | EXPORT_SYMBOL_GPL(blk_mq_end_request_batch); | |
1125 | ||
f9ab4918 | 1126 | static void blk_complete_reqs(struct llist_head *list) |
320ae51f | 1127 | { |
f9ab4918 SAS |
1128 | struct llist_node *entry = llist_reverse_order(llist_del_all(list)); |
1129 | struct request *rq, *next; | |
c3077b5d | 1130 | |
f9ab4918 | 1131 | llist_for_each_entry_safe(rq, next, entry, ipi_list) |
c3077b5d | 1132 | rq->q->mq_ops->complete(rq); |
320ae51f | 1133 | } |
320ae51f | 1134 | |
f9ab4918 | 1135 | static __latent_entropy void blk_done_softirq(struct softirq_action *h) |
320ae51f | 1136 | { |
f9ab4918 | 1137 | blk_complete_reqs(this_cpu_ptr(&blk_cpu_done)); |
115243f5 CH |
1138 | } |
1139 | ||
c3077b5d CH |
1140 | static int blk_softirq_cpu_dead(unsigned int cpu) |
1141 | { | |
f9ab4918 | 1142 | blk_complete_reqs(&per_cpu(blk_cpu_done, cpu)); |
c3077b5d CH |
1143 | return 0; |
1144 | } | |
1145 | ||
40d09b53 | 1146 | static void __blk_mq_complete_request_remote(void *data) |
c3077b5d | 1147 | { |
f9ab4918 | 1148 | __raise_softirq_irqoff(BLOCK_SOFTIRQ); |
c3077b5d CH |
1149 | } |
1150 | ||
96339526 CH |
1151 | static inline bool blk_mq_complete_need_ipi(struct request *rq) |
1152 | { | |
1153 | int cpu = raw_smp_processor_id(); | |
1154 | ||
1155 | if (!IS_ENABLED(CONFIG_SMP) || | |
1156 | !test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) | |
1157 | return false; | |
71425189 SAS |
1158 | /* |
1159 | * With force threaded interrupts enabled, raising softirq from an SMP | |
1160 | * function call will always result in waking the ksoftirqd thread. | |
1161 | * This is probably worse than completing the request on a different | |
1162 | * cache domain. | |
1163 | */ | |
91cc470e | 1164 | if (force_irqthreads()) |
71425189 | 1165 | return false; |
96339526 CH |
1166 | |
1167 | /* same CPU or cache domain? Complete locally */ | |
1168 | if (cpu == rq->mq_ctx->cpu || | |
1169 | (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags) && | |
1170 | cpus_share_cache(cpu, rq->mq_ctx->cpu))) | |
1171 | return false; | |
1172 | ||
1173 | /* don't try to IPI to an offline CPU */ | |
1174 | return cpu_online(rq->mq_ctx->cpu); | |
1175 | } | |
1176 | ||
f9ab4918 SAS |
1177 | static void blk_mq_complete_send_ipi(struct request *rq) |
1178 | { | |
1179 | struct llist_head *list; | |
1180 | unsigned int cpu; | |
1181 | ||
1182 | cpu = rq->mq_ctx->cpu; | |
1183 | list = &per_cpu(blk_cpu_done, cpu); | |
1184 | if (llist_add(&rq->ipi_list, list)) { | |
1185 | INIT_CSD(&rq->csd, __blk_mq_complete_request_remote, rq); | |
1186 | smp_call_function_single_async(cpu, &rq->csd); | |
1187 | } | |
1188 | } | |
1189 | ||
1190 | static void blk_mq_raise_softirq(struct request *rq) | |
1191 | { | |
1192 | struct llist_head *list; | |
1193 | ||
1194 | preempt_disable(); | |
1195 | list = this_cpu_ptr(&blk_cpu_done); | |
1196 | if (llist_add(&rq->ipi_list, list)) | |
1197 | raise_softirq(BLOCK_SOFTIRQ); | |
1198 | preempt_enable(); | |
1199 | } | |
1200 | ||
40d09b53 | 1201 | bool blk_mq_complete_request_remote(struct request *rq) |
320ae51f | 1202 | { |
af78ff7c | 1203 | WRITE_ONCE(rq->state, MQ_RQ_COMPLETE); |
36e76539 | 1204 | |
4ab32bf3 | 1205 | /* |
f168420c LS |
1206 | * For request which hctx has only one ctx mapping, |
1207 | * or a polled request, always complete locally, | |
1208 | * it's pointless to redirect the completion. | |
4ab32bf3 | 1209 | */ |
f168420c LS |
1210 | if (rq->mq_hctx->nr_ctx == 1 || |
1211 | rq->cmd_flags & REQ_POLLED) | |
40d09b53 | 1212 | return false; |
38535201 | 1213 | |
96339526 | 1214 | if (blk_mq_complete_need_ipi(rq)) { |
f9ab4918 SAS |
1215 | blk_mq_complete_send_ipi(rq); |
1216 | return true; | |
3d6efbf6 | 1217 | } |
40d09b53 | 1218 | |
f9ab4918 SAS |
1219 | if (rq->q->nr_hw_queues == 1) { |
1220 | blk_mq_raise_softirq(rq); | |
1221 | return true; | |
1222 | } | |
1223 | return false; | |
40d09b53 CH |
1224 | } |
1225 | EXPORT_SYMBOL_GPL(blk_mq_complete_request_remote); | |
1226 | ||
1227 | /** | |
1228 | * blk_mq_complete_request - end I/O on a request | |
1229 | * @rq: the request being processed | |
1230 | * | |
1231 | * Description: | |
1232 | * Complete a request by scheduling the ->complete_rq operation. | |
1233 | **/ | |
1234 | void blk_mq_complete_request(struct request *rq) | |
1235 | { | |
1236 | if (!blk_mq_complete_request_remote(rq)) | |
1237 | rq->q->mq_ops->complete(rq); | |
320ae51f | 1238 | } |
15f73f5b | 1239 | EXPORT_SYMBOL(blk_mq_complete_request); |
30a91cb4 | 1240 | |
105663f7 AA |
1241 | /** |
1242 | * blk_mq_start_request - Start processing a request | |
1243 | * @rq: Pointer to request to be started | |
1244 | * | |
1245 | * Function used by device drivers to notify the block layer that a request | |
1246 | * is going to be processed now, so blk layer can do proper initializations | |
1247 | * such as starting the timeout timer. | |
1248 | */ | |
e2490073 | 1249 | void blk_mq_start_request(struct request *rq) |
320ae51f JA |
1250 | { |
1251 | struct request_queue *q = rq->q; | |
1252 | ||
a54895fa | 1253 | trace_block_rq_issue(rq); |
320ae51f | 1254 | |
cf43e6be | 1255 | if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) { |
4cddeaca | 1256 | rq->io_start_time_ns = ktime_get_ns(); |
3d244306 | 1257 | rq->stats_sectors = blk_rq_sectors(rq); |
cf43e6be | 1258 | rq->rq_flags |= RQF_STATS; |
a7905043 | 1259 | rq_qos_issue(q, rq); |
cf43e6be JA |
1260 | } |
1261 | ||
1d9bd516 | 1262 | WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE); |
538b7534 | 1263 | |
1d9bd516 | 1264 | blk_add_timer(rq); |
12f5b931 | 1265 | WRITE_ONCE(rq->state, MQ_RQ_IN_FLIGHT); |
49f5baa5 | 1266 | |
54d4e6ab MG |
1267 | #ifdef CONFIG_BLK_DEV_INTEGRITY |
1268 | if (blk_integrity_rq(rq) && req_op(rq) == REQ_OP_WRITE) | |
1269 | q->integrity.profile->prepare_fn(rq); | |
1270 | #endif | |
3e08773c CH |
1271 | if (rq->bio && rq->bio->bi_opf & REQ_POLLED) |
1272 | WRITE_ONCE(rq->bio->bi_cookie, blk_rq_to_qc(rq)); | |
320ae51f | 1273 | } |
e2490073 | 1274 | EXPORT_SYMBOL(blk_mq_start_request); |
320ae51f | 1275 | |
a327c341 ML |
1276 | /* |
1277 | * Allow 2x BLK_MAX_REQUEST_COUNT requests on plug queue for multiple | |
1278 | * queues. This is important for md arrays to benefit from merging | |
1279 | * requests. | |
1280 | */ | |
1281 | static inline unsigned short blk_plug_max_rq_count(struct blk_plug *plug) | |
1282 | { | |
1283 | if (plug->multiple_queues) | |
1284 | return BLK_MAX_REQUEST_COUNT * 2; | |
1285 | return BLK_MAX_REQUEST_COUNT; | |
1286 | } | |
1287 | ||
1288 | static void blk_add_rq_to_plug(struct blk_plug *plug, struct request *rq) | |
1289 | { | |
1290 | struct request *last = rq_list_peek(&plug->mq_list); | |
1291 | ||
1292 | if (!plug->rq_count) { | |
1293 | trace_block_plug(rq->q); | |
1294 | } else if (plug->rq_count >= blk_plug_max_rq_count(plug) || | |
1295 | (!blk_queue_nomerges(rq->q) && | |
1296 | blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) { | |
1297 | blk_mq_flush_plug_list(plug, false); | |
878eb6e4 | 1298 | last = NULL; |
a327c341 ML |
1299 | trace_block_plug(rq->q); |
1300 | } | |
1301 | ||
1302 | if (!plug->multiple_queues && last && last->q != rq->q) | |
1303 | plug->multiple_queues = true; | |
1304 | if (!plug->has_elevator && (rq->rq_flags & RQF_ELV)) | |
1305 | plug->has_elevator = true; | |
1306 | rq->rq_next = NULL; | |
1307 | rq_list_add(&plug->mq_list, rq); | |
1308 | plug->rq_count++; | |
1309 | } | |
1310 | ||
4054cff9 CH |
1311 | /** |
1312 | * blk_execute_rq_nowait - insert a request to I/O scheduler for execution | |
4054cff9 CH |
1313 | * @rq: request to insert |
1314 | * @at_head: insert request at head or tail of queue | |
4054cff9 CH |
1315 | * |
1316 | * Description: | |
1317 | * Insert a fully prepared request at the back of the I/O scheduler queue | |
1318 | * for execution. Don't wait for completion. | |
1319 | * | |
1320 | * Note: | |
1321 | * This function will invoke @done directly if the queue is dead. | |
1322 | */ | |
e2e53086 | 1323 | void blk_execute_rq_nowait(struct request *rq, bool at_head) |
4054cff9 | 1324 | { |
ae948fd6 CH |
1325 | WARN_ON(irqs_disabled()); |
1326 | WARN_ON(!blk_rq_is_passthrough(rq)); | |
4054cff9 | 1327 | |
ae948fd6 | 1328 | blk_account_io_start(rq); |
110fdb44 PR |
1329 | |
1330 | /* | |
1331 | * As plugging can be enabled for passthrough requests on a zoned | |
1332 | * device, directly accessing the plug instead of using blk_mq_plug() | |
1333 | * should not have any consequences. | |
1334 | */ | |
ae948fd6 CH |
1335 | if (current->plug) |
1336 | blk_add_rq_to_plug(current->plug, rq); | |
1337 | else | |
1338 | blk_mq_sched_insert_request(rq, at_head, true, false); | |
4054cff9 CH |
1339 | } |
1340 | EXPORT_SYMBOL_GPL(blk_execute_rq_nowait); | |
1341 | ||
32ac5a9b CH |
1342 | struct blk_rq_wait { |
1343 | struct completion done; | |
1344 | blk_status_t ret; | |
1345 | }; | |
1346 | ||
de671d61 | 1347 | static enum rq_end_io_ret blk_end_sync_rq(struct request *rq, blk_status_t ret) |
32ac5a9b CH |
1348 | { |
1349 | struct blk_rq_wait *wait = rq->end_io_data; | |
1350 | ||
1351 | wait->ret = ret; | |
1352 | complete(&wait->done); | |
de671d61 | 1353 | return RQ_END_IO_NONE; |
32ac5a9b CH |
1354 | } |
1355 | ||
c6e99ea4 | 1356 | bool blk_rq_is_poll(struct request *rq) |
4054cff9 CH |
1357 | { |
1358 | if (!rq->mq_hctx) | |
1359 | return false; | |
1360 | if (rq->mq_hctx->type != HCTX_TYPE_POLL) | |
1361 | return false; | |
1362 | if (WARN_ON_ONCE(!rq->bio)) | |
1363 | return false; | |
1364 | return true; | |
1365 | } | |
c6e99ea4 | 1366 | EXPORT_SYMBOL_GPL(blk_rq_is_poll); |
4054cff9 CH |
1367 | |
1368 | static void blk_rq_poll_completion(struct request *rq, struct completion *wait) | |
1369 | { | |
1370 | do { | |
1371 | bio_poll(rq->bio, NULL, 0); | |
1372 | cond_resched(); | |
1373 | } while (!completion_done(wait)); | |
1374 | } | |
1375 | ||
1376 | /** | |
1377 | * blk_execute_rq - insert a request into queue for execution | |
4054cff9 CH |
1378 | * @rq: request to insert |
1379 | * @at_head: insert request at head or tail of queue | |
1380 | * | |
1381 | * Description: | |
1382 | * Insert a fully prepared request at the back of the I/O scheduler queue | |
1383 | * for execution and wait for completion. | |
1384 | * Return: The blk_status_t result provided to blk_mq_end_request(). | |
1385 | */ | |
b84ba30b | 1386 | blk_status_t blk_execute_rq(struct request *rq, bool at_head) |
4054cff9 | 1387 | { |
32ac5a9b CH |
1388 | struct blk_rq_wait wait = { |
1389 | .done = COMPLETION_INITIALIZER_ONSTACK(wait.done), | |
1390 | }; | |
4054cff9 | 1391 | |
ae948fd6 CH |
1392 | WARN_ON(irqs_disabled()); |
1393 | WARN_ON(!blk_rq_is_passthrough(rq)); | |
4054cff9 CH |
1394 | |
1395 | rq->end_io_data = &wait; | |
ae948fd6 | 1396 | rq->end_io = blk_end_sync_rq; |
4054cff9 | 1397 | |
ae948fd6 CH |
1398 | blk_account_io_start(rq); |
1399 | blk_mq_sched_insert_request(rq, at_head, true, false); | |
4054cff9 | 1400 | |
ae948fd6 | 1401 | if (blk_rq_is_poll(rq)) { |
32ac5a9b | 1402 | blk_rq_poll_completion(rq, &wait.done); |
ae948fd6 CH |
1403 | } else { |
1404 | /* | |
1405 | * Prevent hang_check timer from firing at us during very long | |
1406 | * I/O | |
1407 | */ | |
1408 | unsigned long hang_check = sysctl_hung_task_timeout_secs; | |
1409 | ||
1410 | if (hang_check) | |
32ac5a9b | 1411 | while (!wait_for_completion_io_timeout(&wait.done, |
ae948fd6 CH |
1412 | hang_check * (HZ/2))) |
1413 | ; | |
1414 | else | |
32ac5a9b | 1415 | wait_for_completion_io(&wait.done); |
ae948fd6 | 1416 | } |
4054cff9 | 1417 | |
32ac5a9b | 1418 | return wait.ret; |
4054cff9 CH |
1419 | } |
1420 | EXPORT_SYMBOL(blk_execute_rq); | |
1421 | ||
ed0791b2 | 1422 | static void __blk_mq_requeue_request(struct request *rq) |
320ae51f JA |
1423 | { |
1424 | struct request_queue *q = rq->q; | |
1425 | ||
923218f6 ML |
1426 | blk_mq_put_driver_tag(rq); |
1427 | ||
a54895fa | 1428 | trace_block_rq_requeue(rq); |
a7905043 | 1429 | rq_qos_requeue(q, rq); |
49f5baa5 | 1430 | |
12f5b931 KB |
1431 | if (blk_mq_request_started(rq)) { |
1432 | WRITE_ONCE(rq->state, MQ_RQ_IDLE); | |
da661267 | 1433 | rq->rq_flags &= ~RQF_TIMED_OUT; |
e2490073 | 1434 | } |
320ae51f JA |
1435 | } |
1436 | ||
2b053aca | 1437 | void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list) |
ed0791b2 | 1438 | { |
ed0791b2 | 1439 | __blk_mq_requeue_request(rq); |
ed0791b2 | 1440 | |
105976f5 ML |
1441 | /* this request will be re-inserted to io scheduler queue */ |
1442 | blk_mq_sched_requeue_request(rq); | |
1443 | ||
2b053aca | 1444 | blk_mq_add_to_requeue_list(rq, true, kick_requeue_list); |
ed0791b2 CH |
1445 | } |
1446 | EXPORT_SYMBOL(blk_mq_requeue_request); | |
1447 | ||
6fca6a61 CH |
1448 | static void blk_mq_requeue_work(struct work_struct *work) |
1449 | { | |
1450 | struct request_queue *q = | |
2849450a | 1451 | container_of(work, struct request_queue, requeue_work.work); |
6fca6a61 CH |
1452 | LIST_HEAD(rq_list); |
1453 | struct request *rq, *next; | |
6fca6a61 | 1454 | |
18e9781d | 1455 | spin_lock_irq(&q->requeue_lock); |
6fca6a61 | 1456 | list_splice_init(&q->requeue_list, &rq_list); |
18e9781d | 1457 | spin_unlock_irq(&q->requeue_lock); |
6fca6a61 CH |
1458 | |
1459 | list_for_each_entry_safe(rq, next, &rq_list, queuelist) { | |
aef1897c | 1460 | if (!(rq->rq_flags & (RQF_SOFTBARRIER | RQF_DONTPREP))) |
6fca6a61 CH |
1461 | continue; |
1462 | ||
e8064021 | 1463 | rq->rq_flags &= ~RQF_SOFTBARRIER; |
6fca6a61 | 1464 | list_del_init(&rq->queuelist); |
aef1897c JW |
1465 | /* |
1466 | * If RQF_DONTPREP, rq has contained some driver specific | |
1467 | * data, so insert it to hctx dispatch list to avoid any | |
1468 | * merge. | |
1469 | */ | |
1470 | if (rq->rq_flags & RQF_DONTPREP) | |
01e99aec | 1471 | blk_mq_request_bypass_insert(rq, false, false); |
aef1897c JW |
1472 | else |
1473 | blk_mq_sched_insert_request(rq, true, false, false); | |
6fca6a61 CH |
1474 | } |
1475 | ||
1476 | while (!list_empty(&rq_list)) { | |
1477 | rq = list_entry(rq_list.next, struct request, queuelist); | |
1478 | list_del_init(&rq->queuelist); | |
9e97d295 | 1479 | blk_mq_sched_insert_request(rq, false, false, false); |
6fca6a61 CH |
1480 | } |
1481 | ||
52d7f1b5 | 1482 | blk_mq_run_hw_queues(q, false); |
6fca6a61 CH |
1483 | } |
1484 | ||
2b053aca BVA |
1485 | void blk_mq_add_to_requeue_list(struct request *rq, bool at_head, |
1486 | bool kick_requeue_list) | |
6fca6a61 CH |
1487 | { |
1488 | struct request_queue *q = rq->q; | |
1489 | unsigned long flags; | |
1490 | ||
1491 | /* | |
1492 | * We abuse this flag that is otherwise used by the I/O scheduler to | |
ff821d27 | 1493 | * request head insertion from the workqueue. |
6fca6a61 | 1494 | */ |
e8064021 | 1495 | BUG_ON(rq->rq_flags & RQF_SOFTBARRIER); |
6fca6a61 CH |
1496 | |
1497 | spin_lock_irqsave(&q->requeue_lock, flags); | |
1498 | if (at_head) { | |
e8064021 | 1499 | rq->rq_flags |= RQF_SOFTBARRIER; |
6fca6a61 CH |
1500 | list_add(&rq->queuelist, &q->requeue_list); |
1501 | } else { | |
1502 | list_add_tail(&rq->queuelist, &q->requeue_list); | |
1503 | } | |
1504 | spin_unlock_irqrestore(&q->requeue_lock, flags); | |
2b053aca BVA |
1505 | |
1506 | if (kick_requeue_list) | |
1507 | blk_mq_kick_requeue_list(q); | |
6fca6a61 | 1508 | } |
6fca6a61 CH |
1509 | |
1510 | void blk_mq_kick_requeue_list(struct request_queue *q) | |
1511 | { | |
ae943d20 | 1512 | kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0); |
6fca6a61 CH |
1513 | } |
1514 | EXPORT_SYMBOL(blk_mq_kick_requeue_list); | |
1515 | ||
2849450a MS |
1516 | void blk_mq_delay_kick_requeue_list(struct request_queue *q, |
1517 | unsigned long msecs) | |
1518 | { | |
d4acf365 BVA |
1519 | kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, |
1520 | msecs_to_jiffies(msecs)); | |
2849450a MS |
1521 | } |
1522 | EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list); | |
1523 | ||
2dd6532e | 1524 | static bool blk_mq_rq_inflight(struct request *rq, void *priv) |
ae879912 JA |
1525 | { |
1526 | /* | |
8ab30a33 JG |
1527 | * If we find a request that isn't idle we know the queue is busy |
1528 | * as it's checked in the iter. | |
1529 | * Return false to stop the iteration. | |
ae879912 | 1530 | */ |
8ab30a33 | 1531 | if (blk_mq_request_started(rq)) { |
ae879912 JA |
1532 | bool *busy = priv; |
1533 | ||
1534 | *busy = true; | |
1535 | return false; | |
1536 | } | |
1537 | ||
1538 | return true; | |
1539 | } | |
1540 | ||
3c94d83c | 1541 | bool blk_mq_queue_inflight(struct request_queue *q) |
ae879912 JA |
1542 | { |
1543 | bool busy = false; | |
1544 | ||
3c94d83c | 1545 | blk_mq_queue_tag_busy_iter(q, blk_mq_rq_inflight, &busy); |
ae879912 JA |
1546 | return busy; |
1547 | } | |
3c94d83c | 1548 | EXPORT_SYMBOL_GPL(blk_mq_queue_inflight); |
ae879912 | 1549 | |
9bdb4833 | 1550 | static void blk_mq_rq_timed_out(struct request *req) |
320ae51f | 1551 | { |
da661267 | 1552 | req->rq_flags |= RQF_TIMED_OUT; |
d1210d5a CH |
1553 | if (req->q->mq_ops->timeout) { |
1554 | enum blk_eh_timer_return ret; | |
1555 | ||
9bdb4833 | 1556 | ret = req->q->mq_ops->timeout(req); |
d1210d5a CH |
1557 | if (ret == BLK_EH_DONE) |
1558 | return; | |
1559 | WARN_ON_ONCE(ret != BLK_EH_RESET_TIMER); | |
46f92d42 | 1560 | } |
d1210d5a CH |
1561 | |
1562 | blk_add_timer(req); | |
87ee7b11 | 1563 | } |
5b3f25fc | 1564 | |
82c22947 DJ |
1565 | struct blk_expired_data { |
1566 | bool has_timedout_rq; | |
1567 | unsigned long next; | |
1568 | unsigned long timeout_start; | |
1569 | }; | |
1570 | ||
1571 | static bool blk_mq_req_expired(struct request *rq, struct blk_expired_data *expired) | |
81481eb4 | 1572 | { |
12f5b931 | 1573 | unsigned long deadline; |
87ee7b11 | 1574 | |
12f5b931 KB |
1575 | if (blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT) |
1576 | return false; | |
da661267 CH |
1577 | if (rq->rq_flags & RQF_TIMED_OUT) |
1578 | return false; | |
a7af0af3 | 1579 | |
079076b3 | 1580 | deadline = READ_ONCE(rq->deadline); |
82c22947 | 1581 | if (time_after_eq(expired->timeout_start, deadline)) |
12f5b931 | 1582 | return true; |
a7af0af3 | 1583 | |
82c22947 DJ |
1584 | if (expired->next == 0) |
1585 | expired->next = deadline; | |
1586 | else if (time_after(expired->next, deadline)) | |
1587 | expired->next = deadline; | |
12f5b931 | 1588 | return false; |
87ee7b11 JA |
1589 | } |
1590 | ||
2e315dc0 ML |
1591 | void blk_mq_put_rq_ref(struct request *rq) |
1592 | { | |
de671d61 JA |
1593 | if (is_flush_rq(rq)) { |
1594 | if (rq->end_io(rq, 0) == RQ_END_IO_FREE) | |
1595 | blk_mq_free_request(rq); | |
1596 | } else if (req_ref_put_and_test(rq)) { | |
2e315dc0 | 1597 | __blk_mq_free_request(rq); |
de671d61 | 1598 | } |
2e315dc0 ML |
1599 | } |
1600 | ||
2dd6532e | 1601 | static bool blk_mq_check_expired(struct request *rq, void *priv) |
1d9bd516 | 1602 | { |
82c22947 | 1603 | struct blk_expired_data *expired = priv; |
12f5b931 KB |
1604 | |
1605 | /* | |
c797b40c ML |
1606 | * blk_mq_queue_tag_busy_iter() has locked the request, so it cannot |
1607 | * be reallocated underneath the timeout handler's processing, then | |
1608 | * the expire check is reliable. If the request is not expired, then | |
1609 | * it was completed and reallocated as a new request after returning | |
1610 | * from blk_mq_check_expired(). | |
1d9bd516 | 1611 | */ |
82c22947 DJ |
1612 | if (blk_mq_req_expired(rq, expired)) { |
1613 | expired->has_timedout_rq = true; | |
1614 | return false; | |
1615 | } | |
1616 | return true; | |
1617 | } | |
1618 | ||
1619 | static bool blk_mq_handle_expired(struct request *rq, void *priv) | |
1620 | { | |
1621 | struct blk_expired_data *expired = priv; | |
1622 | ||
1623 | if (blk_mq_req_expired(rq, expired)) | |
9bdb4833 | 1624 | blk_mq_rq_timed_out(rq); |
7baa8572 | 1625 | return true; |
1d9bd516 TH |
1626 | } |
1627 | ||
287922eb | 1628 | static void blk_mq_timeout_work(struct work_struct *work) |
320ae51f | 1629 | { |
287922eb CH |
1630 | struct request_queue *q = |
1631 | container_of(work, struct request_queue, timeout_work); | |
82c22947 DJ |
1632 | struct blk_expired_data expired = { |
1633 | .timeout_start = jiffies, | |
1634 | }; | |
1d9bd516 | 1635 | struct blk_mq_hw_ctx *hctx; |
4f481208 | 1636 | unsigned long i; |
320ae51f | 1637 | |
71f79fb3 GKB |
1638 | /* A deadlock might occur if a request is stuck requiring a |
1639 | * timeout at the same time a queue freeze is waiting | |
1640 | * completion, since the timeout code would not be able to | |
1641 | * acquire the queue reference here. | |
1642 | * | |
1643 | * That's why we don't use blk_queue_enter here; instead, we use | |
1644 | * percpu_ref_tryget directly, because we need to be able to | |
1645 | * obtain a reference even in the short window between the queue | |
1646 | * starting to freeze, by dropping the first reference in | |
1671d522 | 1647 | * blk_freeze_queue_start, and the moment the last request is |
71f79fb3 GKB |
1648 | * consumed, marked by the instant q_usage_counter reaches |
1649 | * zero. | |
1650 | */ | |
1651 | if (!percpu_ref_tryget(&q->q_usage_counter)) | |
287922eb CH |
1652 | return; |
1653 | ||
82c22947 DJ |
1654 | /* check if there is any timed-out request */ |
1655 | blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &expired); | |
1656 | if (expired.has_timedout_rq) { | |
1657 | /* | |
1658 | * Before walking tags, we must ensure any submit started | |
1659 | * before the current time has finished. Since the submit | |
1660 | * uses srcu or rcu, wait for a synchronization point to | |
1661 | * ensure all running submits have finished | |
1662 | */ | |
483239c7 | 1663 | blk_mq_wait_quiesce_done(q->tag_set); |
82c22947 DJ |
1664 | |
1665 | expired.next = 0; | |
1666 | blk_mq_queue_tag_busy_iter(q, blk_mq_handle_expired, &expired); | |
1667 | } | |
320ae51f | 1668 | |
82c22947 DJ |
1669 | if (expired.next != 0) { |
1670 | mod_timer(&q->timeout, expired.next); | |
0d2602ca | 1671 | } else { |
fcd36c36 BVA |
1672 | /* |
1673 | * Request timeouts are handled as a forward rolling timer. If | |
1674 | * we end up here it means that no requests are pending and | |
1675 | * also that no request has been pending for a while. Mark | |
1676 | * each hctx as idle. | |
1677 | */ | |
f054b56c ML |
1678 | queue_for_each_hw_ctx(q, hctx, i) { |
1679 | /* the hctx may be unmapped, so check it here */ | |
1680 | if (blk_mq_hw_queue_mapped(hctx)) | |
1681 | blk_mq_tag_idle(hctx); | |
1682 | } | |
0d2602ca | 1683 | } |
287922eb | 1684 | blk_queue_exit(q); |
320ae51f JA |
1685 | } |
1686 | ||
88459642 OS |
1687 | struct flush_busy_ctx_data { |
1688 | struct blk_mq_hw_ctx *hctx; | |
1689 | struct list_head *list; | |
1690 | }; | |
1691 | ||
1692 | static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data) | |
1693 | { | |
1694 | struct flush_busy_ctx_data *flush_data = data; | |
1695 | struct blk_mq_hw_ctx *hctx = flush_data->hctx; | |
1696 | struct blk_mq_ctx *ctx = hctx->ctxs[bitnr]; | |
c16d6b5a | 1697 | enum hctx_type type = hctx->type; |
88459642 | 1698 | |
88459642 | 1699 | spin_lock(&ctx->lock); |
c16d6b5a | 1700 | list_splice_tail_init(&ctx->rq_lists[type], flush_data->list); |
e9a99a63 | 1701 | sbitmap_clear_bit(sb, bitnr); |
88459642 OS |
1702 | spin_unlock(&ctx->lock); |
1703 | return true; | |
1704 | } | |
1705 | ||
1429d7c9 JA |
1706 | /* |
1707 | * Process software queues that have been marked busy, splicing them | |
1708 | * to the for-dispatch | |
1709 | */ | |
2c3ad667 | 1710 | void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list) |
1429d7c9 | 1711 | { |
88459642 OS |
1712 | struct flush_busy_ctx_data data = { |
1713 | .hctx = hctx, | |
1714 | .list = list, | |
1715 | }; | |
1429d7c9 | 1716 | |
88459642 | 1717 | sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data); |
1429d7c9 | 1718 | } |
2c3ad667 | 1719 | EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs); |
1429d7c9 | 1720 | |
b347689f ML |
1721 | struct dispatch_rq_data { |
1722 | struct blk_mq_hw_ctx *hctx; | |
1723 | struct request *rq; | |
1724 | }; | |
1725 | ||
1726 | static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr, | |
1727 | void *data) | |
1728 | { | |
1729 | struct dispatch_rq_data *dispatch_data = data; | |
1730 | struct blk_mq_hw_ctx *hctx = dispatch_data->hctx; | |
1731 | struct blk_mq_ctx *ctx = hctx->ctxs[bitnr]; | |
c16d6b5a | 1732 | enum hctx_type type = hctx->type; |
b347689f ML |
1733 | |
1734 | spin_lock(&ctx->lock); | |
c16d6b5a ML |
1735 | if (!list_empty(&ctx->rq_lists[type])) { |
1736 | dispatch_data->rq = list_entry_rq(ctx->rq_lists[type].next); | |
b347689f | 1737 | list_del_init(&dispatch_data->rq->queuelist); |
c16d6b5a | 1738 | if (list_empty(&ctx->rq_lists[type])) |
b347689f ML |
1739 | sbitmap_clear_bit(sb, bitnr); |
1740 | } | |
1741 | spin_unlock(&ctx->lock); | |
1742 | ||
1743 | return !dispatch_data->rq; | |
1744 | } | |
1745 | ||
1746 | struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx, | |
1747 | struct blk_mq_ctx *start) | |
1748 | { | |
f31967f0 | 1749 | unsigned off = start ? start->index_hw[hctx->type] : 0; |
b347689f ML |
1750 | struct dispatch_rq_data data = { |
1751 | .hctx = hctx, | |
1752 | .rq = NULL, | |
1753 | }; | |
1754 | ||
1755 | __sbitmap_for_each_set(&hctx->ctx_map, off, | |
1756 | dispatch_rq_from_ctx, &data); | |
1757 | ||
1758 | return data.rq; | |
1759 | } | |
1760 | ||
a808a9d5 | 1761 | static bool __blk_mq_alloc_driver_tag(struct request *rq) |
570e9b73 | 1762 | { |
ae0f1a73 | 1763 | struct sbitmap_queue *bt = &rq->mq_hctx->tags->bitmap_tags; |
570e9b73 | 1764 | unsigned int tag_offset = rq->mq_hctx->tags->nr_reserved_tags; |
570e9b73 ML |
1765 | int tag; |
1766 | ||
568f2700 ML |
1767 | blk_mq_tag_busy(rq->mq_hctx); |
1768 | ||
570e9b73 | 1769 | if (blk_mq_tag_is_reserved(rq->mq_hctx->sched_tags, rq->internal_tag)) { |
ae0f1a73 | 1770 | bt = &rq->mq_hctx->tags->breserved_tags; |
570e9b73 | 1771 | tag_offset = 0; |
28500850 ML |
1772 | } else { |
1773 | if (!hctx_may_queue(rq->mq_hctx, bt)) | |
1774 | return false; | |
570e9b73 ML |
1775 | } |
1776 | ||
570e9b73 ML |
1777 | tag = __sbitmap_queue_get(bt); |
1778 | if (tag == BLK_MQ_NO_TAG) | |
1779 | return false; | |
1780 | ||
1781 | rq->tag = tag + tag_offset; | |
570e9b73 ML |
1782 | return true; |
1783 | } | |
1784 | ||
a808a9d5 | 1785 | bool __blk_mq_get_driver_tag(struct blk_mq_hw_ctx *hctx, struct request *rq) |
570e9b73 | 1786 | { |
a808a9d5 | 1787 | if (rq->tag == BLK_MQ_NO_TAG && !__blk_mq_alloc_driver_tag(rq)) |
568f2700 ML |
1788 | return false; |
1789 | ||
51db1c37 | 1790 | if ((hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) && |
568f2700 ML |
1791 | !(rq->rq_flags & RQF_MQ_INFLIGHT)) { |
1792 | rq->rq_flags |= RQF_MQ_INFLIGHT; | |
bccf5e26 | 1793 | __blk_mq_inc_active_requests(hctx); |
568f2700 ML |
1794 | } |
1795 | hctx->tags->rqs[rq->tag] = rq; | |
1796 | return true; | |
570e9b73 ML |
1797 | } |
1798 | ||
eb619fdb JA |
1799 | static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode, |
1800 | int flags, void *key) | |
da55f2cc OS |
1801 | { |
1802 | struct blk_mq_hw_ctx *hctx; | |
1803 | ||
1804 | hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait); | |
1805 | ||
5815839b | 1806 | spin_lock(&hctx->dispatch_wait_lock); |
e8618575 JA |
1807 | if (!list_empty(&wait->entry)) { |
1808 | struct sbitmap_queue *sbq; | |
1809 | ||
1810 | list_del_init(&wait->entry); | |
ae0f1a73 | 1811 | sbq = &hctx->tags->bitmap_tags; |
e8618575 JA |
1812 | atomic_dec(&sbq->ws_active); |
1813 | } | |
5815839b ML |
1814 | spin_unlock(&hctx->dispatch_wait_lock); |
1815 | ||
da55f2cc OS |
1816 | blk_mq_run_hw_queue(hctx, true); |
1817 | return 1; | |
1818 | } | |
1819 | ||
f906a6a0 JA |
1820 | /* |
1821 | * Mark us waiting for a tag. For shared tags, this involves hooking us into | |
ee3e4de5 BVA |
1822 | * the tag wakeups. For non-shared tags, we can simply mark us needing a |
1823 | * restart. For both cases, take care to check the condition again after | |
f906a6a0 JA |
1824 | * marking us as waiting. |
1825 | */ | |
2278d69f | 1826 | static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx *hctx, |
f906a6a0 | 1827 | struct request *rq) |
da55f2cc | 1828 | { |
98b99e94 | 1829 | struct sbitmap_queue *sbq; |
5815839b | 1830 | struct wait_queue_head *wq; |
f906a6a0 JA |
1831 | wait_queue_entry_t *wait; |
1832 | bool ret; | |
da55f2cc | 1833 | |
47df9ce9 KS |
1834 | if (!(hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) && |
1835 | !(blk_mq_is_shared_tags(hctx->flags))) { | |
684b7324 | 1836 | blk_mq_sched_mark_restart_hctx(hctx); |
f906a6a0 | 1837 | |
c27d53fb BVA |
1838 | /* |
1839 | * It's possible that a tag was freed in the window between the | |
1840 | * allocation failure and adding the hardware queue to the wait | |
1841 | * queue. | |
1842 | * | |
1843 | * Don't clear RESTART here, someone else could have set it. | |
1844 | * At most this will cost an extra queue run. | |
1845 | */ | |
8ab6bb9e | 1846 | return blk_mq_get_driver_tag(rq); |
eb619fdb | 1847 | } |
eb619fdb | 1848 | |
2278d69f | 1849 | wait = &hctx->dispatch_wait; |
c27d53fb BVA |
1850 | if (!list_empty_careful(&wait->entry)) |
1851 | return false; | |
1852 | ||
98b99e94 KS |
1853 | if (blk_mq_tag_is_reserved(rq->mq_hctx->sched_tags, rq->internal_tag)) |
1854 | sbq = &hctx->tags->breserved_tags; | |
1855 | else | |
1856 | sbq = &hctx->tags->bitmap_tags; | |
e8618575 | 1857 | wq = &bt_wait_ptr(sbq, hctx)->wait; |
5815839b ML |
1858 | |
1859 | spin_lock_irq(&wq->lock); | |
1860 | spin_lock(&hctx->dispatch_wait_lock); | |
c27d53fb | 1861 | if (!list_empty(&wait->entry)) { |
5815839b ML |
1862 | spin_unlock(&hctx->dispatch_wait_lock); |
1863 | spin_unlock_irq(&wq->lock); | |
c27d53fb | 1864 | return false; |
eb619fdb JA |
1865 | } |
1866 | ||
e8618575 | 1867 | atomic_inc(&sbq->ws_active); |
5815839b ML |
1868 | wait->flags &= ~WQ_FLAG_EXCLUSIVE; |
1869 | __add_wait_queue(wq, wait); | |
c27d53fb | 1870 | |
da55f2cc | 1871 | /* |
eb619fdb JA |
1872 | * It's possible that a tag was freed in the window between the |
1873 | * allocation failure and adding the hardware queue to the wait | |
1874 | * queue. | |
da55f2cc | 1875 | */ |
8ab6bb9e | 1876 | ret = blk_mq_get_driver_tag(rq); |
c27d53fb | 1877 | if (!ret) { |
5815839b ML |
1878 | spin_unlock(&hctx->dispatch_wait_lock); |
1879 | spin_unlock_irq(&wq->lock); | |
c27d53fb | 1880 | return false; |
eb619fdb | 1881 | } |
c27d53fb BVA |
1882 | |
1883 | /* | |
1884 | * We got a tag, remove ourselves from the wait queue to ensure | |
1885 | * someone else gets the wakeup. | |
1886 | */ | |
c27d53fb | 1887 | list_del_init(&wait->entry); |
e8618575 | 1888 | atomic_dec(&sbq->ws_active); |
5815839b ML |
1889 | spin_unlock(&hctx->dispatch_wait_lock); |
1890 | spin_unlock_irq(&wq->lock); | |
c27d53fb BVA |
1891 | |
1892 | return true; | |
da55f2cc OS |
1893 | } |
1894 | ||
6e768717 ML |
1895 | #define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT 8 |
1896 | #define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR 4 | |
1897 | /* | |
1898 | * Update dispatch busy with the Exponential Weighted Moving Average(EWMA): | |
1899 | * - EWMA is one simple way to compute running average value | |
1900 | * - weight(7/8 and 1/8) is applied so that it can decrease exponentially | |
1901 | * - take 4 as factor for avoiding to get too small(0) result, and this | |
1902 | * factor doesn't matter because EWMA decreases exponentially | |
1903 | */ | |
1904 | static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx *hctx, bool busy) | |
1905 | { | |
1906 | unsigned int ewma; | |
1907 | ||
6e768717 ML |
1908 | ewma = hctx->dispatch_busy; |
1909 | ||
1910 | if (!ewma && !busy) | |
1911 | return; | |
1912 | ||
1913 | ewma *= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT - 1; | |
1914 | if (busy) | |
1915 | ewma += 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR; | |
1916 | ewma /= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT; | |
1917 | ||
1918 | hctx->dispatch_busy = ewma; | |
1919 | } | |
1920 | ||
86ff7c2a ML |
1921 | #define BLK_MQ_RESOURCE_DELAY 3 /* ms units */ |
1922 | ||
c92a4103 JT |
1923 | static void blk_mq_handle_dev_resource(struct request *rq, |
1924 | struct list_head *list) | |
1925 | { | |
c92a4103 JT |
1926 | list_add(&rq->queuelist, list); |
1927 | __blk_mq_requeue_request(rq); | |
1928 | } | |
1929 | ||
0512a75b KB |
1930 | static void blk_mq_handle_zone_resource(struct request *rq, |
1931 | struct list_head *zone_list) | |
1932 | { | |
1933 | /* | |
1934 | * If we end up here it is because we cannot dispatch a request to a | |
1935 | * specific zone due to LLD level zone-write locking or other zone | |
1936 | * related resource not being available. In this case, set the request | |
1937 | * aside in zone_list for retrying it later. | |
1938 | */ | |
1939 | list_add(&rq->queuelist, zone_list); | |
1940 | __blk_mq_requeue_request(rq); | |
1941 | } | |
1942 | ||
75383524 ML |
1943 | enum prep_dispatch { |
1944 | PREP_DISPATCH_OK, | |
1945 | PREP_DISPATCH_NO_TAG, | |
1946 | PREP_DISPATCH_NO_BUDGET, | |
1947 | }; | |
1948 | ||
1949 | static enum prep_dispatch blk_mq_prep_dispatch_rq(struct request *rq, | |
1950 | bool need_budget) | |
1951 | { | |
1952 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; | |
2a5a24aa | 1953 | int budget_token = -1; |
75383524 | 1954 | |
2a5a24aa ML |
1955 | if (need_budget) { |
1956 | budget_token = blk_mq_get_dispatch_budget(rq->q); | |
1957 | if (budget_token < 0) { | |
1958 | blk_mq_put_driver_tag(rq); | |
1959 | return PREP_DISPATCH_NO_BUDGET; | |
1960 | } | |
1961 | blk_mq_set_rq_budget_token(rq, budget_token); | |
75383524 ML |
1962 | } |
1963 | ||
1964 | if (!blk_mq_get_driver_tag(rq)) { | |
1965 | /* | |
1966 | * The initial allocation attempt failed, so we need to | |
1967 | * rerun the hardware queue when a tag is freed. The | |
1968 | * waitqueue takes care of that. If the queue is run | |
1969 | * before we add this entry back on the dispatch list, | |
1970 | * we'll re-run it below. | |
1971 | */ | |
1972 | if (!blk_mq_mark_tag_wait(hctx, rq)) { | |
1fd40b5e ML |
1973 | /* |
1974 | * All budgets not got from this function will be put | |
1975 | * together during handling partial dispatch | |
1976 | */ | |
1977 | if (need_budget) | |
2a5a24aa | 1978 | blk_mq_put_dispatch_budget(rq->q, budget_token); |
75383524 ML |
1979 | return PREP_DISPATCH_NO_TAG; |
1980 | } | |
1981 | } | |
1982 | ||
1983 | return PREP_DISPATCH_OK; | |
1984 | } | |
1985 | ||
1fd40b5e ML |
1986 | /* release all allocated budgets before calling to blk_mq_dispatch_rq_list */ |
1987 | static void blk_mq_release_budgets(struct request_queue *q, | |
2a5a24aa | 1988 | struct list_head *list) |
1fd40b5e | 1989 | { |
2a5a24aa | 1990 | struct request *rq; |
1fd40b5e | 1991 | |
2a5a24aa ML |
1992 | list_for_each_entry(rq, list, queuelist) { |
1993 | int budget_token = blk_mq_get_rq_budget_token(rq); | |
1fd40b5e | 1994 | |
2a5a24aa ML |
1995 | if (budget_token >= 0) |
1996 | blk_mq_put_dispatch_budget(q, budget_token); | |
1997 | } | |
1fd40b5e ML |
1998 | } |
1999 | ||
34c9f547 KS |
2000 | /* |
2001 | * blk_mq_commit_rqs will notify driver using bd->last that there is no | |
2002 | * more requests. (See comment in struct blk_mq_ops for commit_rqs for | |
2003 | * details) | |
2004 | * Attention, we should explicitly call this in unusual cases: | |
2005 | * 1) did not queue everything initially scheduled to queue | |
2006 | * 2) the last attempt to queue a request failed | |
2007 | */ | |
2008 | static void blk_mq_commit_rqs(struct blk_mq_hw_ctx *hctx, int queued, | |
2009 | bool from_schedule) | |
2010 | { | |
2011 | if (hctx->queue->mq_ops->commit_rqs && queued) { | |
2012 | trace_block_unplug(hctx->queue, queued, !from_schedule); | |
2013 | hctx->queue->mq_ops->commit_rqs(hctx); | |
2014 | } | |
2015 | } | |
2016 | ||
1f57f8d4 JA |
2017 | /* |
2018 | * Returns true if we did some work AND can potentially do more. | |
2019 | */ | |
445874e8 | 2020 | bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx *hctx, struct list_head *list, |
1fd40b5e | 2021 | unsigned int nr_budgets) |
320ae51f | 2022 | { |
75383524 | 2023 | enum prep_dispatch prep; |
445874e8 | 2024 | struct request_queue *q = hctx->queue; |
f1ce99f7 | 2025 | struct request *rq; |
4ea58fe4 | 2026 | int queued; |
86ff7c2a | 2027 | blk_status_t ret = BLK_STS_OK; |
0512a75b | 2028 | LIST_HEAD(zone_list); |
9586e67b | 2029 | bool needs_resource = false; |
320ae51f | 2030 | |
81380ca1 OS |
2031 | if (list_empty(list)) |
2032 | return false; | |
2033 | ||
320ae51f JA |
2034 | /* |
2035 | * Now process all the entries, sending them to the driver. | |
2036 | */ | |
4ea58fe4 | 2037 | queued = 0; |
81380ca1 | 2038 | do { |
74c45052 | 2039 | struct blk_mq_queue_data bd; |
320ae51f | 2040 | |
f04c3df3 | 2041 | rq = list_first_entry(list, struct request, queuelist); |
0bca799b | 2042 | |
445874e8 | 2043 | WARN_ON_ONCE(hctx != rq->mq_hctx); |
1fd40b5e | 2044 | prep = blk_mq_prep_dispatch_rq(rq, !nr_budgets); |
75383524 | 2045 | if (prep != PREP_DISPATCH_OK) |
0bca799b | 2046 | break; |
de148297 | 2047 | |
320ae51f | 2048 | list_del_init(&rq->queuelist); |
320ae51f | 2049 | |
74c45052 | 2050 | bd.rq = rq; |
f1ce99f7 | 2051 | bd.last = list_empty(list); |
74c45052 | 2052 | |
1fd40b5e ML |
2053 | /* |
2054 | * once the request is queued to lld, no need to cover the | |
2055 | * budget any more | |
2056 | */ | |
2057 | if (nr_budgets) | |
2058 | nr_budgets--; | |
74c45052 | 2059 | ret = q->mq_ops->queue_rq(hctx, &bd); |
7bf13729 ML |
2060 | switch (ret) { |
2061 | case BLK_STS_OK: | |
2062 | queued++; | |
320ae51f | 2063 | break; |
7bf13729 | 2064 | case BLK_STS_RESOURCE: |
9586e67b NA |
2065 | needs_resource = true; |
2066 | fallthrough; | |
7bf13729 ML |
2067 | case BLK_STS_DEV_RESOURCE: |
2068 | blk_mq_handle_dev_resource(rq, list); | |
2069 | goto out; | |
2070 | case BLK_STS_ZONE_RESOURCE: | |
0512a75b KB |
2071 | /* |
2072 | * Move the request to zone_list and keep going through | |
2073 | * the dispatch list to find more requests the drive can | |
2074 | * accept. | |
2075 | */ | |
2076 | blk_mq_handle_zone_resource(rq, &zone_list); | |
9586e67b | 2077 | needs_resource = true; |
7bf13729 ML |
2078 | break; |
2079 | default: | |
e21ee5a6 | 2080 | blk_mq_end_request(rq, ret); |
320ae51f | 2081 | } |
81380ca1 | 2082 | } while (!list_empty(list)); |
7bf13729 | 2083 | out: |
0512a75b KB |
2084 | if (!list_empty(&zone_list)) |
2085 | list_splice_tail_init(&zone_list, list); | |
2086 | ||
632bfb63 | 2087 | /* If we didn't flush the entire list, we could have told the driver |
2088 | * there was more coming, but that turned out to be a lie. | |
2089 | */ | |
e4ef2e05 KS |
2090 | if (!list_empty(list) || ret != BLK_STS_OK) |
2091 | blk_mq_commit_rqs(hctx, queued, false); | |
2092 | ||
320ae51f JA |
2093 | /* |
2094 | * Any items that need requeuing? Stuff them into hctx->dispatch, | |
2095 | * that is where we will continue on next queue run. | |
2096 | */ | |
f04c3df3 | 2097 | if (!list_empty(list)) { |
86ff7c2a | 2098 | bool needs_restart; |
75383524 ML |
2099 | /* For non-shared tags, the RESTART check will suffice */ |
2100 | bool no_tag = prep == PREP_DISPATCH_NO_TAG && | |
47df9ce9 KS |
2101 | ((hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) || |
2102 | blk_mq_is_shared_tags(hctx->flags)); | |
86ff7c2a | 2103 | |
2a5a24aa ML |
2104 | if (nr_budgets) |
2105 | blk_mq_release_budgets(q, list); | |
86ff7c2a | 2106 | |
320ae51f | 2107 | spin_lock(&hctx->lock); |
01e99aec | 2108 | list_splice_tail_init(list, &hctx->dispatch); |
320ae51f | 2109 | spin_unlock(&hctx->lock); |
f04c3df3 | 2110 | |
d7d8535f ML |
2111 | /* |
2112 | * Order adding requests to hctx->dispatch and checking | |
2113 | * SCHED_RESTART flag. The pair of this smp_mb() is the one | |
2114 | * in blk_mq_sched_restart(). Avoid restart code path to | |
2115 | * miss the new added requests to hctx->dispatch, meantime | |
2116 | * SCHED_RESTART is observed here. | |
2117 | */ | |
2118 | smp_mb(); | |
2119 | ||
9ba52e58 | 2120 | /* |
710c785f BVA |
2121 | * If SCHED_RESTART was set by the caller of this function and |
2122 | * it is no longer set that means that it was cleared by another | |
2123 | * thread and hence that a queue rerun is needed. | |
9ba52e58 | 2124 | * |
eb619fdb JA |
2125 | * If 'no_tag' is set, that means that we failed getting |
2126 | * a driver tag with an I/O scheduler attached. If our dispatch | |
2127 | * waitqueue is no longer active, ensure that we run the queue | |
2128 | * AFTER adding our entries back to the list. | |
bd166ef1 | 2129 | * |
710c785f BVA |
2130 | * If no I/O scheduler has been configured it is possible that |
2131 | * the hardware queue got stopped and restarted before requests | |
2132 | * were pushed back onto the dispatch list. Rerun the queue to | |
2133 | * avoid starvation. Notes: | |
2134 | * - blk_mq_run_hw_queue() checks whether or not a queue has | |
2135 | * been stopped before rerunning a queue. | |
2136 | * - Some but not all block drivers stop a queue before | |
fc17b653 | 2137 | * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq |
710c785f | 2138 | * and dm-rq. |
86ff7c2a ML |
2139 | * |
2140 | * If driver returns BLK_STS_RESOURCE and SCHED_RESTART | |
2141 | * bit is set, run queue after a delay to avoid IO stalls | |
ab3cee37 | 2142 | * that could otherwise occur if the queue is idle. We'll do |
9586e67b NA |
2143 | * similar if we couldn't get budget or couldn't lock a zone |
2144 | * and SCHED_RESTART is set. | |
bd166ef1 | 2145 | */ |
86ff7c2a | 2146 | needs_restart = blk_mq_sched_needs_restart(hctx); |
9586e67b NA |
2147 | if (prep == PREP_DISPATCH_NO_BUDGET) |
2148 | needs_resource = true; | |
86ff7c2a | 2149 | if (!needs_restart || |
eb619fdb | 2150 | (no_tag && list_empty_careful(&hctx->dispatch_wait.entry))) |
bd166ef1 | 2151 | blk_mq_run_hw_queue(hctx, true); |
6d5e8d21 | 2152 | else if (needs_resource) |
86ff7c2a | 2153 | blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY); |
1f57f8d4 | 2154 | |
6e768717 | 2155 | blk_mq_update_dispatch_busy(hctx, true); |
1f57f8d4 | 2156 | return false; |
4ea58fe4 | 2157 | } |
f04c3df3 | 2158 | |
4ea58fe4 KS |
2159 | blk_mq_update_dispatch_busy(hctx, false); |
2160 | return true; | |
f04c3df3 JA |
2161 | } |
2162 | ||
105663f7 AA |
2163 | /** |
2164 | * __blk_mq_run_hw_queue - Run a hardware queue. | |
2165 | * @hctx: Pointer to the hardware queue to run. | |
2166 | * | |
2167 | * Send pending requests to the hardware. | |
2168 | */ | |
6a83e74d BVA |
2169 | static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx) |
2170 | { | |
b7a71e66 JA |
2171 | /* |
2172 | * We can't run the queue inline with ints disabled. Ensure that | |
2173 | * we catch bad users of this early. | |
2174 | */ | |
2175 | WARN_ON_ONCE(in_interrupt()); | |
2176 | ||
bcc330f4 ML |
2177 | blk_mq_run_dispatch_ops(hctx->queue, |
2178 | blk_mq_sched_dispatch_requests(hctx)); | |
6a83e74d BVA |
2179 | } |
2180 | ||
f82ddf19 ML |
2181 | static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx) |
2182 | { | |
2183 | int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask); | |
2184 | ||
2185 | if (cpu >= nr_cpu_ids) | |
2186 | cpu = cpumask_first(hctx->cpumask); | |
2187 | return cpu; | |
2188 | } | |
2189 | ||
506e931f JA |
2190 | /* |
2191 | * It'd be great if the workqueue API had a way to pass | |
2192 | * in a mask and had some smarts for more clever placement. | |
2193 | * For now we just round-robin here, switching for every | |
2194 | * BLK_MQ_CPU_WORK_BATCH queued items. | |
2195 | */ | |
2196 | static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx) | |
2197 | { | |
7bed4595 | 2198 | bool tried = false; |
476f8c98 | 2199 | int next_cpu = hctx->next_cpu; |
7bed4595 | 2200 | |
b657d7e6 CH |
2201 | if (hctx->queue->nr_hw_queues == 1) |
2202 | return WORK_CPU_UNBOUND; | |
506e931f JA |
2203 | |
2204 | if (--hctx->next_cpu_batch <= 0) { | |
7bed4595 | 2205 | select_cpu: |
476f8c98 | 2206 | next_cpu = cpumask_next_and(next_cpu, hctx->cpumask, |
20e4d813 | 2207 | cpu_online_mask); |
506e931f | 2208 | if (next_cpu >= nr_cpu_ids) |
f82ddf19 | 2209 | next_cpu = blk_mq_first_mapped_cpu(hctx); |
506e931f JA |
2210 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; |
2211 | } | |
2212 | ||
7bed4595 ML |
2213 | /* |
2214 | * Do unbound schedule if we can't find a online CPU for this hctx, | |
2215 | * and it should only happen in the path of handling CPU DEAD. | |
2216 | */ | |
476f8c98 | 2217 | if (!cpu_online(next_cpu)) { |
7bed4595 ML |
2218 | if (!tried) { |
2219 | tried = true; | |
2220 | goto select_cpu; | |
2221 | } | |
2222 | ||
2223 | /* | |
2224 | * Make sure to re-select CPU next time once after CPUs | |
2225 | * in hctx->cpumask become online again. | |
2226 | */ | |
476f8c98 | 2227 | hctx->next_cpu = next_cpu; |
7bed4595 ML |
2228 | hctx->next_cpu_batch = 1; |
2229 | return WORK_CPU_UNBOUND; | |
2230 | } | |
476f8c98 ML |
2231 | |
2232 | hctx->next_cpu = next_cpu; | |
2233 | return next_cpu; | |
506e931f JA |
2234 | } |
2235 | ||
105663f7 AA |
2236 | /** |
2237 | * __blk_mq_delay_run_hw_queue - Run (or schedule to run) a hardware queue. | |
2238 | * @hctx: Pointer to the hardware queue to run. | |
2239 | * @async: If we want to run the queue asynchronously. | |
fa94ba8a | 2240 | * @msecs: Milliseconds of delay to wait before running the queue. |
105663f7 AA |
2241 | * |
2242 | * If !@async, try to run the queue now. Else, run the queue asynchronously and | |
2243 | * with a delay of @msecs. | |
2244 | */ | |
7587a5ae BVA |
2245 | static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async, |
2246 | unsigned long msecs) | |
320ae51f | 2247 | { |
5435c023 | 2248 | if (unlikely(blk_mq_hctx_stopped(hctx))) |
320ae51f JA |
2249 | return; |
2250 | ||
1b792f2f | 2251 | if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) { |
3c8f9da4 | 2252 | if (cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask)) { |
398205b8 | 2253 | __blk_mq_run_hw_queue(hctx); |
398205b8 PB |
2254 | return; |
2255 | } | |
e4043dcf | 2256 | } |
398205b8 | 2257 | |
ae943d20 BVA |
2258 | kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work, |
2259 | msecs_to_jiffies(msecs)); | |
7587a5ae BVA |
2260 | } |
2261 | ||
105663f7 AA |
2262 | /** |
2263 | * blk_mq_delay_run_hw_queue - Run a hardware queue asynchronously. | |
2264 | * @hctx: Pointer to the hardware queue to run. | |
fa94ba8a | 2265 | * @msecs: Milliseconds of delay to wait before running the queue. |
105663f7 AA |
2266 | * |
2267 | * Run a hardware queue asynchronously with a delay of @msecs. | |
2268 | */ | |
7587a5ae BVA |
2269 | void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs) |
2270 | { | |
2271 | __blk_mq_delay_run_hw_queue(hctx, true, msecs); | |
2272 | } | |
2273 | EXPORT_SYMBOL(blk_mq_delay_run_hw_queue); | |
2274 | ||
105663f7 AA |
2275 | /** |
2276 | * blk_mq_run_hw_queue - Start to run a hardware queue. | |
2277 | * @hctx: Pointer to the hardware queue to run. | |
2278 | * @async: If we want to run the queue asynchronously. | |
2279 | * | |
2280 | * Check if the request queue is not in a quiesced state and if there are | |
2281 | * pending requests to be sent. If this is true, run the queue to send requests | |
2282 | * to hardware. | |
2283 | */ | |
626fb735 | 2284 | void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) |
7587a5ae | 2285 | { |
24f5a90f ML |
2286 | bool need_run; |
2287 | ||
2288 | /* | |
2289 | * When queue is quiesced, we may be switching io scheduler, or | |
2290 | * updating nr_hw_queues, or other things, and we can't run queue | |
2291 | * any more, even __blk_mq_hctx_has_pending() can't be called safely. | |
2292 | * | |
2293 | * And queue will be rerun in blk_mq_unquiesce_queue() if it is | |
2294 | * quiesced. | |
2295 | */ | |
41adf531 | 2296 | __blk_mq_run_dispatch_ops(hctx->queue, false, |
2a904d00 ML |
2297 | need_run = !blk_queue_quiesced(hctx->queue) && |
2298 | blk_mq_hctx_has_pending(hctx)); | |
24f5a90f | 2299 | |
626fb735 | 2300 | if (need_run) |
79f720a7 | 2301 | __blk_mq_delay_run_hw_queue(hctx, async, 0); |
320ae51f | 2302 | } |
5b727272 | 2303 | EXPORT_SYMBOL(blk_mq_run_hw_queue); |
320ae51f | 2304 | |
b6e68ee8 JK |
2305 | /* |
2306 | * Return prefered queue to dispatch from (if any) for non-mq aware IO | |
2307 | * scheduler. | |
2308 | */ | |
2309 | static struct blk_mq_hw_ctx *blk_mq_get_sq_hctx(struct request_queue *q) | |
2310 | { | |
5d05426e | 2311 | struct blk_mq_ctx *ctx = blk_mq_get_ctx(q); |
b6e68ee8 JK |
2312 | /* |
2313 | * If the IO scheduler does not respect hardware queues when | |
2314 | * dispatching, we just don't bother with multiple HW queues and | |
2315 | * dispatch from hctx for the current CPU since running multiple queues | |
2316 | * just causes lock contention inside the scheduler and pointless cache | |
2317 | * bouncing. | |
2318 | */ | |
51ab80f0 | 2319 | struct blk_mq_hw_ctx *hctx = ctx->hctxs[HCTX_TYPE_DEFAULT]; |
5d05426e | 2320 | |
b6e68ee8 JK |
2321 | if (!blk_mq_hctx_stopped(hctx)) |
2322 | return hctx; | |
2323 | return NULL; | |
2324 | } | |
2325 | ||
105663f7 | 2326 | /** |
24f7bb88 | 2327 | * blk_mq_run_hw_queues - Run all hardware queues in a request queue. |
105663f7 AA |
2328 | * @q: Pointer to the request queue to run. |
2329 | * @async: If we want to run the queue asynchronously. | |
2330 | */ | |
b94ec296 | 2331 | void blk_mq_run_hw_queues(struct request_queue *q, bool async) |
320ae51f | 2332 | { |
b6e68ee8 | 2333 | struct blk_mq_hw_ctx *hctx, *sq_hctx; |
4f481208 | 2334 | unsigned long i; |
320ae51f | 2335 | |
b6e68ee8 | 2336 | sq_hctx = NULL; |
4d337ceb | 2337 | if (blk_queue_sq_sched(q)) |
b6e68ee8 | 2338 | sq_hctx = blk_mq_get_sq_hctx(q); |
320ae51f | 2339 | queue_for_each_hw_ctx(q, hctx, i) { |
79f720a7 | 2340 | if (blk_mq_hctx_stopped(hctx)) |
320ae51f | 2341 | continue; |
b6e68ee8 JK |
2342 | /* |
2343 | * Dispatch from this hctx either if there's no hctx preferred | |
2344 | * by IO scheduler or if it has requests that bypass the | |
2345 | * scheduler. | |
2346 | */ | |
2347 | if (!sq_hctx || sq_hctx == hctx || | |
2348 | !list_empty_careful(&hctx->dispatch)) | |
2349 | blk_mq_run_hw_queue(hctx, async); | |
320ae51f JA |
2350 | } |
2351 | } | |
b94ec296 | 2352 | EXPORT_SYMBOL(blk_mq_run_hw_queues); |
320ae51f | 2353 | |
b9151e7b DA |
2354 | /** |
2355 | * blk_mq_delay_run_hw_queues - Run all hardware queues asynchronously. | |
2356 | * @q: Pointer to the request queue to run. | |
fa94ba8a | 2357 | * @msecs: Milliseconds of delay to wait before running the queues. |
b9151e7b DA |
2358 | */ |
2359 | void blk_mq_delay_run_hw_queues(struct request_queue *q, unsigned long msecs) | |
2360 | { | |
b6e68ee8 | 2361 | struct blk_mq_hw_ctx *hctx, *sq_hctx; |
4f481208 | 2362 | unsigned long i; |
b9151e7b | 2363 | |
b6e68ee8 | 2364 | sq_hctx = NULL; |
4d337ceb | 2365 | if (blk_queue_sq_sched(q)) |
b6e68ee8 | 2366 | sq_hctx = blk_mq_get_sq_hctx(q); |
b9151e7b DA |
2367 | queue_for_each_hw_ctx(q, hctx, i) { |
2368 | if (blk_mq_hctx_stopped(hctx)) | |
2369 | continue; | |
8f5fea65 DJ |
2370 | /* |
2371 | * If there is already a run_work pending, leave the | |
2372 | * pending delay untouched. Otherwise, a hctx can stall | |
2373 | * if another hctx is re-delaying the other's work | |
2374 | * before the work executes. | |
2375 | */ | |
2376 | if (delayed_work_pending(&hctx->run_work)) | |
2377 | continue; | |
b6e68ee8 JK |
2378 | /* |
2379 | * Dispatch from this hctx either if there's no hctx preferred | |
2380 | * by IO scheduler or if it has requests that bypass the | |
2381 | * scheduler. | |
2382 | */ | |
2383 | if (!sq_hctx || sq_hctx == hctx || | |
2384 | !list_empty_careful(&hctx->dispatch)) | |
2385 | blk_mq_delay_run_hw_queue(hctx, msecs); | |
b9151e7b DA |
2386 | } |
2387 | } | |
2388 | EXPORT_SYMBOL(blk_mq_delay_run_hw_queues); | |
2389 | ||
39a70c76 ML |
2390 | /* |
2391 | * This function is often used for pausing .queue_rq() by driver when | |
2392 | * there isn't enough resource or some conditions aren't satisfied, and | |
4d606219 | 2393 | * BLK_STS_RESOURCE is usually returned. |
39a70c76 ML |
2394 | * |
2395 | * We do not guarantee that dispatch can be drained or blocked | |
2396 | * after blk_mq_stop_hw_queue() returns. Please use | |
2397 | * blk_mq_quiesce_queue() for that requirement. | |
2398 | */ | |
2719aa21 JA |
2399 | void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx) |
2400 | { | |
641a9ed6 | 2401 | cancel_delayed_work(&hctx->run_work); |
280d45f6 | 2402 | |
641a9ed6 | 2403 | set_bit(BLK_MQ_S_STOPPED, &hctx->state); |
2719aa21 | 2404 | } |
641a9ed6 | 2405 | EXPORT_SYMBOL(blk_mq_stop_hw_queue); |
2719aa21 | 2406 | |
39a70c76 ML |
2407 | /* |
2408 | * This function is often used for pausing .queue_rq() by driver when | |
2409 | * there isn't enough resource or some conditions aren't satisfied, and | |
4d606219 | 2410 | * BLK_STS_RESOURCE is usually returned. |
39a70c76 ML |
2411 | * |
2412 | * We do not guarantee that dispatch can be drained or blocked | |
2413 | * after blk_mq_stop_hw_queues() returns. Please use | |
2414 | * blk_mq_quiesce_queue() for that requirement. | |
2415 | */ | |
2719aa21 JA |
2416 | void blk_mq_stop_hw_queues(struct request_queue *q) |
2417 | { | |
641a9ed6 | 2418 | struct blk_mq_hw_ctx *hctx; |
4f481208 | 2419 | unsigned long i; |
641a9ed6 ML |
2420 | |
2421 | queue_for_each_hw_ctx(q, hctx, i) | |
2422 | blk_mq_stop_hw_queue(hctx); | |
280d45f6 CH |
2423 | } |
2424 | EXPORT_SYMBOL(blk_mq_stop_hw_queues); | |
2425 | ||
320ae51f JA |
2426 | void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx) |
2427 | { | |
2428 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
e4043dcf | 2429 | |
0ffbce80 | 2430 | blk_mq_run_hw_queue(hctx, false); |
320ae51f JA |
2431 | } |
2432 | EXPORT_SYMBOL(blk_mq_start_hw_queue); | |
2433 | ||
2f268556 CH |
2434 | void blk_mq_start_hw_queues(struct request_queue *q) |
2435 | { | |
2436 | struct blk_mq_hw_ctx *hctx; | |
4f481208 | 2437 | unsigned long i; |
2f268556 CH |
2438 | |
2439 | queue_for_each_hw_ctx(q, hctx, i) | |
2440 | blk_mq_start_hw_queue(hctx); | |
2441 | } | |
2442 | EXPORT_SYMBOL(blk_mq_start_hw_queues); | |
2443 | ||
ae911c5e JA |
2444 | void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) |
2445 | { | |
2446 | if (!blk_mq_hctx_stopped(hctx)) | |
2447 | return; | |
2448 | ||
2449 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
2450 | blk_mq_run_hw_queue(hctx, async); | |
2451 | } | |
2452 | EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue); | |
2453 | ||
1b4a3258 | 2454 | void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async) |
320ae51f JA |
2455 | { |
2456 | struct blk_mq_hw_ctx *hctx; | |
4f481208 | 2457 | unsigned long i; |
320ae51f | 2458 | |
ae911c5e JA |
2459 | queue_for_each_hw_ctx(q, hctx, i) |
2460 | blk_mq_start_stopped_hw_queue(hctx, async); | |
320ae51f JA |
2461 | } |
2462 | EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues); | |
2463 | ||
70f4db63 | 2464 | static void blk_mq_run_work_fn(struct work_struct *work) |
320ae51f JA |
2465 | { |
2466 | struct blk_mq_hw_ctx *hctx; | |
2467 | ||
9f993737 | 2468 | hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work); |
320ae51f | 2469 | |
21c6e939 | 2470 | /* |
15fe8a90 | 2471 | * If we are stopped, don't run the queue. |
21c6e939 | 2472 | */ |
0841031a | 2473 | if (blk_mq_hctx_stopped(hctx)) |
0196d6b4 | 2474 | return; |
7587a5ae BVA |
2475 | |
2476 | __blk_mq_run_hw_queue(hctx); | |
2477 | } | |
2478 | ||
cfd0c552 | 2479 | static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx, |
cfd0c552 ML |
2480 | struct request *rq, |
2481 | bool at_head) | |
320ae51f | 2482 | { |
e57690fe | 2483 | struct blk_mq_ctx *ctx = rq->mq_ctx; |
c16d6b5a | 2484 | enum hctx_type type = hctx->type; |
e57690fe | 2485 | |
7b607814 BVA |
2486 | lockdep_assert_held(&ctx->lock); |
2487 | ||
a54895fa | 2488 | trace_block_rq_insert(rq); |
01b983c9 | 2489 | |
72a0a36e | 2490 | if (at_head) |
c16d6b5a | 2491 | list_add(&rq->queuelist, &ctx->rq_lists[type]); |
72a0a36e | 2492 | else |
c16d6b5a | 2493 | list_add_tail(&rq->queuelist, &ctx->rq_lists[type]); |
cfd0c552 | 2494 | } |
4bb659b1 | 2495 | |
2c3ad667 JA |
2496 | void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq, |
2497 | bool at_head) | |
cfd0c552 ML |
2498 | { |
2499 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
2500 | ||
7b607814 BVA |
2501 | lockdep_assert_held(&ctx->lock); |
2502 | ||
e57690fe | 2503 | __blk_mq_insert_req_list(hctx, rq, at_head); |
320ae51f | 2504 | blk_mq_hctx_mark_pending(hctx, ctx); |
320ae51f JA |
2505 | } |
2506 | ||
105663f7 AA |
2507 | /** |
2508 | * blk_mq_request_bypass_insert - Insert a request at dispatch list. | |
2509 | * @rq: Pointer to request to be inserted. | |
26bfeb26 | 2510 | * @at_head: true if the request should be inserted at the head of the list. |
105663f7 AA |
2511 | * @run_queue: If we should run the hardware queue after inserting the request. |
2512 | * | |
157f377b JA |
2513 | * Should only be used carefully, when the caller knows we want to |
2514 | * bypass a potential IO scheduler on the target device. | |
2515 | */ | |
01e99aec ML |
2516 | void blk_mq_request_bypass_insert(struct request *rq, bool at_head, |
2517 | bool run_queue) | |
157f377b | 2518 | { |
ea4f995e | 2519 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; |
157f377b JA |
2520 | |
2521 | spin_lock(&hctx->lock); | |
01e99aec ML |
2522 | if (at_head) |
2523 | list_add(&rq->queuelist, &hctx->dispatch); | |
2524 | else | |
2525 | list_add_tail(&rq->queuelist, &hctx->dispatch); | |
157f377b JA |
2526 | spin_unlock(&hctx->lock); |
2527 | ||
b0850297 ML |
2528 | if (run_queue) |
2529 | blk_mq_run_hw_queue(hctx, false); | |
157f377b JA |
2530 | } |
2531 | ||
bd166ef1 JA |
2532 | void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx, |
2533 | struct list_head *list) | |
320ae51f JA |
2534 | |
2535 | { | |
3f0cedc7 | 2536 | struct request *rq; |
c16d6b5a | 2537 | enum hctx_type type = hctx->type; |
3f0cedc7 | 2538 | |
320ae51f JA |
2539 | /* |
2540 | * preemption doesn't flush plug list, so it's possible ctx->cpu is | |
2541 | * offline now | |
2542 | */ | |
3f0cedc7 | 2543 | list_for_each_entry(rq, list, queuelist) { |
e57690fe | 2544 | BUG_ON(rq->mq_ctx != ctx); |
a54895fa | 2545 | trace_block_rq_insert(rq); |
320ae51f | 2546 | } |
3f0cedc7 ML |
2547 | |
2548 | spin_lock(&ctx->lock); | |
c16d6b5a | 2549 | list_splice_tail_init(list, &ctx->rq_lists[type]); |
cfd0c552 | 2550 | blk_mq_hctx_mark_pending(hctx, ctx); |
320ae51f | 2551 | spin_unlock(&ctx->lock); |
320ae51f JA |
2552 | } |
2553 | ||
14ccb66b CH |
2554 | static void blk_mq_bio_to_request(struct request *rq, struct bio *bio, |
2555 | unsigned int nr_segs) | |
320ae51f | 2556 | { |
93f221ae EB |
2557 | int err; |
2558 | ||
f924cdde CH |
2559 | if (bio->bi_opf & REQ_RAHEAD) |
2560 | rq->cmd_flags |= REQ_FAILFAST_MASK; | |
2561 | ||
2562 | rq->__sector = bio->bi_iter.bi_sector; | |
14ccb66b | 2563 | blk_rq_bio_prep(rq, bio, nr_segs); |
93f221ae EB |
2564 | |
2565 | /* This can't fail, since GFP_NOIO includes __GFP_DIRECT_RECLAIM. */ | |
2566 | err = blk_crypto_rq_bio_prep(rq, bio, GFP_NOIO); | |
2567 | WARN_ON_ONCE(err); | |
4b570521 | 2568 | |
b5af37ab | 2569 | blk_account_io_start(rq); |
320ae51f JA |
2570 | } |
2571 | ||
0f95549c | 2572 | static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx, |
3e08773c | 2573 | struct request *rq, bool last) |
f984df1f | 2574 | { |
f984df1f | 2575 | struct request_queue *q = rq->q; |
f984df1f SL |
2576 | struct blk_mq_queue_data bd = { |
2577 | .rq = rq, | |
be94f058 | 2578 | .last = last, |
f984df1f | 2579 | }; |
f06345ad | 2580 | blk_status_t ret; |
0f95549c | 2581 | |
0f95549c MS |
2582 | /* |
2583 | * For OK queue, we are done. For error, caller may kill it. | |
2584 | * Any other error (busy), just add it to our list as we | |
2585 | * previously would have done. | |
2586 | */ | |
2587 | ret = q->mq_ops->queue_rq(hctx, &bd); | |
2588 | switch (ret) { | |
2589 | case BLK_STS_OK: | |
6ce3dd6e | 2590 | blk_mq_update_dispatch_busy(hctx, false); |
0f95549c MS |
2591 | break; |
2592 | case BLK_STS_RESOURCE: | |
86ff7c2a | 2593 | case BLK_STS_DEV_RESOURCE: |
6ce3dd6e | 2594 | blk_mq_update_dispatch_busy(hctx, true); |
0f95549c MS |
2595 | __blk_mq_requeue_request(rq); |
2596 | break; | |
2597 | default: | |
6ce3dd6e | 2598 | blk_mq_update_dispatch_busy(hctx, false); |
0f95549c MS |
2599 | break; |
2600 | } | |
2601 | ||
2602 | return ret; | |
2603 | } | |
2604 | ||
fd9c40f6 | 2605 | static blk_status_t __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx, |
0f95549c | 2606 | struct request *rq, |
fd9c40f6 | 2607 | bool bypass_insert, bool last) |
0f95549c MS |
2608 | { |
2609 | struct request_queue *q = rq->q; | |
d964f04a | 2610 | bool run_queue = true; |
2a5a24aa | 2611 | int budget_token; |
d964f04a | 2612 | |
23d4ee19 | 2613 | /* |
fd9c40f6 | 2614 | * RCU or SRCU read lock is needed before checking quiesced flag. |
23d4ee19 | 2615 | * |
fd9c40f6 BVA |
2616 | * When queue is stopped or quiesced, ignore 'bypass_insert' from |
2617 | * blk_mq_request_issue_directly(), and return BLK_STS_OK to caller, | |
2618 | * and avoid driver to try to dispatch again. | |
23d4ee19 | 2619 | */ |
fd9c40f6 | 2620 | if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)) { |
d964f04a | 2621 | run_queue = false; |
fd9c40f6 BVA |
2622 | bypass_insert = false; |
2623 | goto insert; | |
d964f04a | 2624 | } |
f984df1f | 2625 | |
2ff0682d | 2626 | if ((rq->rq_flags & RQF_ELV) && !bypass_insert) |
fd9c40f6 | 2627 | goto insert; |
2253efc8 | 2628 | |
2a5a24aa ML |
2629 | budget_token = blk_mq_get_dispatch_budget(q); |
2630 | if (budget_token < 0) | |
fd9c40f6 | 2631 | goto insert; |
bd166ef1 | 2632 | |
2a5a24aa ML |
2633 | blk_mq_set_rq_budget_token(rq, budget_token); |
2634 | ||
8ab6bb9e | 2635 | if (!blk_mq_get_driver_tag(rq)) { |
2a5a24aa | 2636 | blk_mq_put_dispatch_budget(q, budget_token); |
fd9c40f6 | 2637 | goto insert; |
88022d72 | 2638 | } |
de148297 | 2639 | |
3e08773c | 2640 | return __blk_mq_issue_directly(hctx, rq, last); |
fd9c40f6 BVA |
2641 | insert: |
2642 | if (bypass_insert) | |
2643 | return BLK_STS_RESOURCE; | |
2644 | ||
db03f88f ML |
2645 | blk_mq_sched_insert_request(rq, false, run_queue, false); |
2646 | ||
fd9c40f6 BVA |
2647 | return BLK_STS_OK; |
2648 | } | |
2649 | ||
105663f7 AA |
2650 | /** |
2651 | * blk_mq_try_issue_directly - Try to send a request directly to device driver. | |
2652 | * @hctx: Pointer of the associated hardware queue. | |
2653 | * @rq: Pointer to request to be sent. | |
105663f7 AA |
2654 | * |
2655 | * If the device has enough resources to accept a new request now, send the | |
2656 | * request directly to device driver. Else, insert at hctx->dispatch queue, so | |
2657 | * we can try send it another time in the future. Requests inserted at this | |
2658 | * queue have higher priority. | |
2659 | */ | |
fd9c40f6 | 2660 | static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx, |
3e08773c | 2661 | struct request *rq) |
fd9c40f6 | 2662 | { |
2a904d00 ML |
2663 | blk_status_t ret = |
2664 | __blk_mq_try_issue_directly(hctx, rq, false, true); | |
fd9c40f6 | 2665 | |
fd9c40f6 | 2666 | if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE) |
01e99aec | 2667 | blk_mq_request_bypass_insert(rq, false, true); |
fd9c40f6 BVA |
2668 | else if (ret != BLK_STS_OK) |
2669 | blk_mq_end_request(rq, ret); | |
fd9c40f6 BVA |
2670 | } |
2671 | ||
06c8c691 | 2672 | static blk_status_t blk_mq_request_issue_directly(struct request *rq, bool last) |
fd9c40f6 | 2673 | { |
4cafe86c | 2674 | return __blk_mq_try_issue_directly(rq->mq_hctx, rq, true, last); |
5eb6126e CH |
2675 | } |
2676 | ||
3e368fb0 | 2677 | static void blk_mq_plug_issue_direct(struct blk_plug *plug) |
b84c5b50 CH |
2678 | { |
2679 | struct blk_mq_hw_ctx *hctx = NULL; | |
2680 | struct request *rq; | |
2681 | int queued = 0; | |
0d617a83 | 2682 | blk_status_t ret = BLK_STS_OK; |
b84c5b50 CH |
2683 | |
2684 | while ((rq = rq_list_pop(&plug->mq_list))) { | |
2685 | bool last = rq_list_empty(plug->mq_list); | |
b84c5b50 CH |
2686 | |
2687 | if (hctx != rq->mq_hctx) { | |
34c9f547 KS |
2688 | if (hctx) { |
2689 | blk_mq_commit_rqs(hctx, queued, false); | |
2690 | queued = 0; | |
2691 | } | |
b84c5b50 CH |
2692 | hctx = rq->mq_hctx; |
2693 | } | |
2694 | ||
2695 | ret = blk_mq_request_issue_directly(rq, last); | |
2696 | switch (ret) { | |
2697 | case BLK_STS_OK: | |
2698 | queued++; | |
2699 | break; | |
2700 | case BLK_STS_RESOURCE: | |
2701 | case BLK_STS_DEV_RESOURCE: | |
d3b38596 | 2702 | blk_mq_request_bypass_insert(rq, false, true); |
0d617a83 | 2703 | goto out; |
b84c5b50 CH |
2704 | default: |
2705 | blk_mq_end_request(rq, ret); | |
b84c5b50 CH |
2706 | break; |
2707 | } | |
2708 | } | |
2709 | ||
0d617a83 KS |
2710 | out: |
2711 | if (ret != BLK_STS_OK) | |
34c9f547 | 2712 | blk_mq_commit_rqs(hctx, queued, false); |
b84c5b50 CH |
2713 | } |
2714 | ||
518579a9 KB |
2715 | static void __blk_mq_flush_plug_list(struct request_queue *q, |
2716 | struct blk_plug *plug) | |
2717 | { | |
2718 | if (blk_queue_quiesced(q)) | |
2719 | return; | |
2720 | q->mq_ops->queue_rqs(&plug->mq_list); | |
2721 | } | |
2722 | ||
26fed4ac JA |
2723 | static void blk_mq_dispatch_plug_list(struct blk_plug *plug, bool from_sched) |
2724 | { | |
2725 | struct blk_mq_hw_ctx *this_hctx = NULL; | |
2726 | struct blk_mq_ctx *this_ctx = NULL; | |
2727 | struct request *requeue_list = NULL; | |
2728 | unsigned int depth = 0; | |
2729 | LIST_HEAD(list); | |
2730 | ||
2731 | do { | |
2732 | struct request *rq = rq_list_pop(&plug->mq_list); | |
2733 | ||
2734 | if (!this_hctx) { | |
2735 | this_hctx = rq->mq_hctx; | |
2736 | this_ctx = rq->mq_ctx; | |
2737 | } else if (this_hctx != rq->mq_hctx || this_ctx != rq->mq_ctx) { | |
2738 | rq_list_add(&requeue_list, rq); | |
2739 | continue; | |
2740 | } | |
2741 | list_add_tail(&rq->queuelist, &list); | |
2742 | depth++; | |
2743 | } while (!rq_list_empty(plug->mq_list)); | |
2744 | ||
2745 | plug->mq_list = requeue_list; | |
2746 | trace_block_unplug(this_hctx->queue, depth, !from_sched); | |
2747 | blk_mq_sched_insert_requests(this_hctx, this_ctx, &list, from_sched); | |
2748 | } | |
2749 | ||
b84c5b50 CH |
2750 | void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule) |
2751 | { | |
3c67d44d | 2752 | struct request *rq; |
b84c5b50 CH |
2753 | |
2754 | if (rq_list_empty(plug->mq_list)) | |
2755 | return; | |
2756 | plug->rq_count = 0; | |
2757 | ||
2758 | if (!plug->multiple_queues && !plug->has_elevator && !from_schedule) { | |
3c67d44d JA |
2759 | struct request_queue *q; |
2760 | ||
2761 | rq = rq_list_peek(&plug->mq_list); | |
2762 | q = rq->q; | |
2763 | ||
2764 | /* | |
2765 | * Peek first request and see if we have a ->queue_rqs() hook. | |
2766 | * If we do, we can dispatch the whole plug list in one go. We | |
2767 | * already know at this point that all requests belong to the | |
2768 | * same queue, caller must ensure that's the case. | |
2769 | * | |
2770 | * Since we pass off the full list to the driver at this point, | |
2771 | * we do not increment the active request count for the queue. | |
2772 | * Bypass shared tags for now because of that. | |
2773 | */ | |
2774 | if (q->mq_ops->queue_rqs && | |
2775 | !(rq->mq_hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) { | |
2776 | blk_mq_run_dispatch_ops(q, | |
518579a9 | 2777 | __blk_mq_flush_plug_list(q, plug)); |
3c67d44d JA |
2778 | if (rq_list_empty(plug->mq_list)) |
2779 | return; | |
2780 | } | |
73f3760e ML |
2781 | |
2782 | blk_mq_run_dispatch_ops(q, | |
3e368fb0 | 2783 | blk_mq_plug_issue_direct(plug)); |
b84c5b50 CH |
2784 | if (rq_list_empty(plug->mq_list)) |
2785 | return; | |
2786 | } | |
2787 | ||
b84c5b50 | 2788 | do { |
26fed4ac | 2789 | blk_mq_dispatch_plug_list(plug, from_schedule); |
b84c5b50 | 2790 | } while (!rq_list_empty(plug->mq_list)); |
b84c5b50 CH |
2791 | } |
2792 | ||
6ce3dd6e ML |
2793 | void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx, |
2794 | struct list_head *list) | |
2795 | { | |
536167d4 | 2796 | int queued = 0; |
984ce0a7 | 2797 | blk_status_t ret = BLK_STS_OK; |
536167d4 | 2798 | |
6ce3dd6e | 2799 | while (!list_empty(list)) { |
6ce3dd6e ML |
2800 | struct request *rq = list_first_entry(list, struct request, |
2801 | queuelist); | |
2802 | ||
2803 | list_del_init(&rq->queuelist); | |
fd9c40f6 | 2804 | ret = blk_mq_request_issue_directly(rq, list_empty(list)); |
27e8b2bb KS |
2805 | switch (ret) { |
2806 | case BLK_STS_OK: | |
536167d4 | 2807 | queued++; |
27e8b2bb KS |
2808 | break; |
2809 | case BLK_STS_RESOURCE: | |
2810 | case BLK_STS_DEV_RESOURCE: | |
2811 | blk_mq_request_bypass_insert(rq, false, | |
2812 | list_empty(list)); | |
2813 | goto out; | |
2814 | default: | |
2815 | blk_mq_end_request(rq, ret); | |
2816 | break; | |
2817 | } | |
6ce3dd6e | 2818 | } |
d666ba98 | 2819 | |
27e8b2bb | 2820 | out: |
984ce0a7 KS |
2821 | if (ret != BLK_STS_OK) |
2822 | blk_mq_commit_rqs(hctx, queued, false); | |
6ce3dd6e ML |
2823 | } |
2824 | ||
b131f201 | 2825 | static bool blk_mq_attempt_bio_merge(struct request_queue *q, |
0c5bcc92 | 2826 | struct bio *bio, unsigned int nr_segs) |
900e0807 JA |
2827 | { |
2828 | if (!blk_queue_nomerges(q) && bio_mergeable(bio)) { | |
0c5bcc92 | 2829 | if (blk_attempt_plug_merge(q, bio, nr_segs)) |
900e0807 JA |
2830 | return true; |
2831 | if (blk_mq_sched_bio_merge(q, bio, nr_segs)) | |
2832 | return true; | |
2833 | } | |
2834 | return false; | |
2835 | } | |
2836 | ||
71539717 JA |
2837 | static struct request *blk_mq_get_new_requests(struct request_queue *q, |
2838 | struct blk_plug *plug, | |
0a5aa8d1 SK |
2839 | struct bio *bio, |
2840 | unsigned int nsegs) | |
71539717 JA |
2841 | { |
2842 | struct blk_mq_alloc_data data = { | |
2843 | .q = q, | |
2844 | .nr_tags = 1, | |
9d497e29 | 2845 | .cmd_flags = bio->bi_opf, |
71539717 JA |
2846 | }; |
2847 | struct request *rq; | |
2848 | ||
5b13bc8a | 2849 | if (unlikely(bio_queue_enter(bio))) |
b637108a | 2850 | return NULL; |
900e0807 | 2851 | |
0a5aa8d1 SK |
2852 | if (blk_mq_attempt_bio_merge(q, bio, nsegs)) |
2853 | goto queue_exit; | |
2854 | ||
2855 | rq_qos_throttle(q, bio); | |
2856 | ||
71539717 JA |
2857 | if (plug) { |
2858 | data.nr_tags = plug->nr_ios; | |
2859 | plug->nr_ios = 1; | |
2860 | data.cached_rq = &plug->cached_rq; | |
2861 | } | |
2862 | ||
2863 | rq = __blk_mq_alloc_requests(&data); | |
373b5416 JA |
2864 | if (rq) |
2865 | return rq; | |
71539717 JA |
2866 | rq_qos_cleanup(q, bio); |
2867 | if (bio->bi_opf & REQ_NOWAIT) | |
2868 | bio_wouldblock_error(bio); | |
0a5aa8d1 | 2869 | queue_exit: |
5b13bc8a | 2870 | blk_queue_exit(q); |
71539717 JA |
2871 | return NULL; |
2872 | } | |
2873 | ||
5b13bc8a | 2874 | static inline struct request *blk_mq_get_cached_request(struct request_queue *q, |
0a5aa8d1 | 2875 | struct blk_plug *plug, struct bio **bio, unsigned int nsegs) |
71539717 | 2876 | { |
b637108a | 2877 | struct request *rq; |
77465647 | 2878 | enum hctx_type type, hctx_type; |
b637108a | 2879 | |
5b13bc8a CH |
2880 | if (!plug) |
2881 | return NULL; | |
71539717 | 2882 | |
0a5aa8d1 SK |
2883 | if (blk_mq_attempt_bio_merge(q, *bio, nsegs)) { |
2884 | *bio = NULL; | |
2885 | return NULL; | |
2886 | } | |
2887 | ||
23f3e327 XN |
2888 | rq = rq_list_peek(&plug->cached_rq); |
2889 | if (!rq || rq->q != q) | |
2890 | return NULL; | |
2891 | ||
77465647 PB |
2892 | type = blk_mq_get_hctx_type((*bio)->bi_opf); |
2893 | hctx_type = rq->mq_hctx->type; | |
2894 | if (type != hctx_type && | |
2895 | !(type == HCTX_TYPE_READ && hctx_type == HCTX_TYPE_DEFAULT)) | |
5b13bc8a | 2896 | return NULL; |
0a5aa8d1 | 2897 | if (op_is_flush(rq->cmd_flags) != op_is_flush((*bio)->bi_opf)) |
5b13bc8a CH |
2898 | return NULL; |
2899 | ||
2645672f JA |
2900 | /* |
2901 | * If any qos ->throttle() end up blocking, we will have flushed the | |
2902 | * plug and hence killed the cached_rq list as well. Pop this entry | |
2903 | * before we throttle. | |
2904 | */ | |
5b13bc8a | 2905 | plug->cached_rq = rq_list_next(rq); |
2645672f JA |
2906 | rq_qos_throttle(q, *bio); |
2907 | ||
2908 | rq->cmd_flags = (*bio)->bi_opf; | |
5b13bc8a | 2909 | INIT_LIST_HEAD(&rq->queuelist); |
5b13bc8a | 2910 | return rq; |
71539717 JA |
2911 | } |
2912 | ||
82b74cac JK |
2913 | static void bio_set_ioprio(struct bio *bio) |
2914 | { | |
a78418e6 JK |
2915 | /* Nobody set ioprio so far? Initialize it based on task's nice value */ |
2916 | if (IOPRIO_PRIO_CLASS(bio->bi_ioprio) == IOPRIO_CLASS_NONE) | |
2917 | bio->bi_ioprio = get_current_ioprio(); | |
82b74cac JK |
2918 | blkcg_set_ioprio(bio); |
2919 | } | |
2920 | ||
105663f7 | 2921 | /** |
c62b37d9 | 2922 | * blk_mq_submit_bio - Create and send a request to block device. |
105663f7 AA |
2923 | * @bio: Bio pointer. |
2924 | * | |
2925 | * Builds up a request structure from @q and @bio and send to the device. The | |
2926 | * request may not be queued directly to hardware if: | |
2927 | * * This request can be merged with another one | |
2928 | * * We want to place request at plug queue for possible future merging | |
2929 | * * There is an IO scheduler active at this queue | |
2930 | * | |
2931 | * It will not queue the request if there is an error with the bio, or at the | |
2932 | * request creation. | |
105663f7 | 2933 | */ |
3e08773c | 2934 | void blk_mq_submit_bio(struct bio *bio) |
07068d5b | 2935 | { |
ed6cddef | 2936 | struct request_queue *q = bdev_get_queue(bio->bi_bdev); |
6deacb3b | 2937 | struct blk_plug *plug = blk_mq_plug(bio); |
ef295ecf | 2938 | const int is_sync = op_is_sync(bio->bi_opf); |
07068d5b | 2939 | struct request *rq; |
abd45c15 | 2940 | unsigned int nr_segs = 1; |
a892c8d5 | 2941 | blk_status_t ret; |
07068d5b | 2942 | |
51d798cd | 2943 | bio = blk_queue_bounce(bio, q); |
613b1488 | 2944 | if (bio_may_exceed_limits(bio, &q->limits)) { |
c55ddd90 | 2945 | bio = __bio_split_to_limits(bio, &q->limits, &nr_segs); |
613b1488 JA |
2946 | if (!bio) |
2947 | return; | |
2948 | } | |
f36ea50c | 2949 | |
e23947bd | 2950 | if (!bio_integrity_prep(bio)) |
900e0807 | 2951 | return; |
87760e5e | 2952 | |
9c6227e0 JK |
2953 | bio_set_ioprio(bio); |
2954 | ||
0a5aa8d1 | 2955 | rq = blk_mq_get_cached_request(q, plug, &bio, nr_segs); |
5b13bc8a | 2956 | if (!rq) { |
0a5aa8d1 SK |
2957 | if (!bio) |
2958 | return; | |
2959 | rq = blk_mq_get_new_requests(q, plug, bio, nr_segs); | |
5b13bc8a CH |
2960 | if (unlikely(!rq)) |
2961 | return; | |
2962 | } | |
87760e5e | 2963 | |
e8a676d6 | 2964 | trace_block_getrq(bio); |
d6f1dda2 | 2965 | |
c1c80384 | 2966 | rq_qos_track(q, rq, bio); |
07068d5b | 2967 | |
970d168d BVA |
2968 | blk_mq_bio_to_request(rq, bio, nr_segs); |
2969 | ||
a892c8d5 ST |
2970 | ret = blk_crypto_init_request(rq); |
2971 | if (ret != BLK_STS_OK) { | |
2972 | bio->bi_status = ret; | |
2973 | bio_endio(bio); | |
2974 | blk_mq_free_request(rq); | |
3e08773c | 2975 | return; |
a892c8d5 ST |
2976 | } |
2977 | ||
2b504bd4 ML |
2978 | if (op_is_flush(bio->bi_opf)) { |
2979 | blk_insert_flush(rq); | |
d92ca9d8 | 2980 | return; |
2b504bd4 | 2981 | } |
d92ca9d8 | 2982 | |
1e9c2303 | 2983 | if (plug) |
ce5b009c | 2984 | blk_add_rq_to_plug(plug, rq); |
1e9c2303 CH |
2985 | else if ((rq->rq_flags & RQF_ELV) || |
2986 | (rq->mq_hctx->dispatch_busy && | |
2987 | (q->nr_hw_queues == 1 || !is_sync))) | |
a12de1d4 | 2988 | blk_mq_sched_insert_request(rq, false, true, true); |
1e9c2303 | 2989 | else |
bcc330f4 | 2990 | blk_mq_run_dispatch_ops(rq->q, |
2a904d00 | 2991 | blk_mq_try_issue_directly(rq->mq_hctx, rq)); |
320ae51f JA |
2992 | } |
2993 | ||
248c7933 | 2994 | #ifdef CONFIG_BLK_MQ_STACKING |
06c8c691 | 2995 | /** |
a5efda3c | 2996 | * blk_insert_cloned_request - Helper for stacking drivers to submit a request |
a5efda3c | 2997 | * @rq: the request being queued |
06c8c691 | 2998 | */ |
28db4711 | 2999 | blk_status_t blk_insert_cloned_request(struct request *rq) |
06c8c691 | 3000 | { |
28db4711 | 3001 | struct request_queue *q = rq->q; |
06c8c691 | 3002 | unsigned int max_sectors = blk_queue_get_max_sectors(q, req_op(rq)); |
a5efda3c | 3003 | blk_status_t ret; |
06c8c691 CH |
3004 | |
3005 | if (blk_rq_sectors(rq) > max_sectors) { | |
3006 | /* | |
3007 | * SCSI device does not have a good way to return if | |
3008 | * Write Same/Zero is actually supported. If a device rejects | |
3009 | * a non-read/write command (discard, write same,etc.) the | |
3010 | * low-level device driver will set the relevant queue limit to | |
3011 | * 0 to prevent blk-lib from issuing more of the offending | |
3012 | * operations. Commands queued prior to the queue limit being | |
3013 | * reset need to be completed with BLK_STS_NOTSUPP to avoid I/O | |
3014 | * errors being propagated to upper layers. | |
3015 | */ | |
3016 | if (max_sectors == 0) | |
3017 | return BLK_STS_NOTSUPP; | |
3018 | ||
3019 | printk(KERN_ERR "%s: over max size limit. (%u > %u)\n", | |
3020 | __func__, blk_rq_sectors(rq), max_sectors); | |
3021 | return BLK_STS_IOERR; | |
3022 | } | |
3023 | ||
3024 | /* | |
3025 | * The queue settings related to segment counting may differ from the | |
3026 | * original queue. | |
3027 | */ | |
3028 | rq->nr_phys_segments = blk_recalc_rq_segments(rq); | |
3029 | if (rq->nr_phys_segments > queue_max_segments(q)) { | |
3030 | printk(KERN_ERR "%s: over max segments limit. (%hu > %hu)\n", | |
3031 | __func__, rq->nr_phys_segments, queue_max_segments(q)); | |
3032 | return BLK_STS_IOERR; | |
3033 | } | |
3034 | ||
28db4711 | 3035 | if (q->disk && should_fail_request(q->disk->part0, blk_rq_bytes(rq))) |
06c8c691 CH |
3036 | return BLK_STS_IOERR; |
3037 | ||
3038 | if (blk_crypto_insert_cloned_request(rq)) | |
3039 | return BLK_STS_IOERR; | |
3040 | ||
3041 | blk_account_io_start(rq); | |
3042 | ||
3043 | /* | |
3044 | * Since we have a scheduler attached on the top device, | |
3045 | * bypass a potential scheduler on the bottom device for | |
3046 | * insert. | |
3047 | */ | |
28db4711 | 3048 | blk_mq_run_dispatch_ops(q, |
4cafe86c | 3049 | ret = blk_mq_request_issue_directly(rq, true)); |
592ee119 YK |
3050 | if (ret) |
3051 | blk_account_io_done(rq, ktime_get_ns()); | |
4cafe86c | 3052 | return ret; |
06c8c691 CH |
3053 | } |
3054 | EXPORT_SYMBOL_GPL(blk_insert_cloned_request); | |
3055 | ||
3056 | /** | |
3057 | * blk_rq_unprep_clone - Helper function to free all bios in a cloned request | |
3058 | * @rq: the clone request to be cleaned up | |
3059 | * | |
3060 | * Description: | |
3061 | * Free all bios in @rq for a cloned request. | |
3062 | */ | |
3063 | void blk_rq_unprep_clone(struct request *rq) | |
3064 | { | |
3065 | struct bio *bio; | |
3066 | ||
3067 | while ((bio = rq->bio) != NULL) { | |
3068 | rq->bio = bio->bi_next; | |
3069 | ||
3070 | bio_put(bio); | |
3071 | } | |
3072 | } | |
3073 | EXPORT_SYMBOL_GPL(blk_rq_unprep_clone); | |
3074 | ||
3075 | /** | |
3076 | * blk_rq_prep_clone - Helper function to setup clone request | |
3077 | * @rq: the request to be setup | |
3078 | * @rq_src: original request to be cloned | |
3079 | * @bs: bio_set that bios for clone are allocated from | |
3080 | * @gfp_mask: memory allocation mask for bio | |
3081 | * @bio_ctr: setup function to be called for each clone bio. | |
3082 | * Returns %0 for success, non %0 for failure. | |
3083 | * @data: private data to be passed to @bio_ctr | |
3084 | * | |
3085 | * Description: | |
3086 | * Clones bios in @rq_src to @rq, and copies attributes of @rq_src to @rq. | |
3087 | * Also, pages which the original bios are pointing to are not copied | |
3088 | * and the cloned bios just point same pages. | |
3089 | * So cloned bios must be completed before original bios, which means | |
3090 | * the caller must complete @rq before @rq_src. | |
3091 | */ | |
3092 | int blk_rq_prep_clone(struct request *rq, struct request *rq_src, | |
3093 | struct bio_set *bs, gfp_t gfp_mask, | |
3094 | int (*bio_ctr)(struct bio *, struct bio *, void *), | |
3095 | void *data) | |
3096 | { | |
3097 | struct bio *bio, *bio_src; | |
3098 | ||
3099 | if (!bs) | |
3100 | bs = &fs_bio_set; | |
3101 | ||
3102 | __rq_for_each_bio(bio_src, rq_src) { | |
abfc426d CH |
3103 | bio = bio_alloc_clone(rq->q->disk->part0, bio_src, gfp_mask, |
3104 | bs); | |
06c8c691 CH |
3105 | if (!bio) |
3106 | goto free_and_out; | |
3107 | ||
3108 | if (bio_ctr && bio_ctr(bio, bio_src, data)) | |
3109 | goto free_and_out; | |
3110 | ||
3111 | if (rq->bio) { | |
3112 | rq->biotail->bi_next = bio; | |
3113 | rq->biotail = bio; | |
3114 | } else { | |
3115 | rq->bio = rq->biotail = bio; | |
3116 | } | |
3117 | bio = NULL; | |
3118 | } | |
3119 | ||
3120 | /* Copy attributes of the original request to the clone request. */ | |
3121 | rq->__sector = blk_rq_pos(rq_src); | |
3122 | rq->__data_len = blk_rq_bytes(rq_src); | |
3123 | if (rq_src->rq_flags & RQF_SPECIAL_PAYLOAD) { | |
3124 | rq->rq_flags |= RQF_SPECIAL_PAYLOAD; | |
3125 | rq->special_vec = rq_src->special_vec; | |
3126 | } | |
3127 | rq->nr_phys_segments = rq_src->nr_phys_segments; | |
3128 | rq->ioprio = rq_src->ioprio; | |
3129 | ||
3130 | if (rq->bio && blk_crypto_rq_bio_prep(rq, rq->bio, gfp_mask) < 0) | |
3131 | goto free_and_out; | |
3132 | ||
3133 | return 0; | |
3134 | ||
3135 | free_and_out: | |
3136 | if (bio) | |
3137 | bio_put(bio); | |
3138 | blk_rq_unprep_clone(rq); | |
3139 | ||
3140 | return -ENOMEM; | |
3141 | } | |
3142 | EXPORT_SYMBOL_GPL(blk_rq_prep_clone); | |
248c7933 | 3143 | #endif /* CONFIG_BLK_MQ_STACKING */ |
06c8c691 | 3144 | |
f2b8f3ce CH |
3145 | /* |
3146 | * Steal bios from a request and add them to a bio list. | |
3147 | * The request must not have been partially completed before. | |
3148 | */ | |
3149 | void blk_steal_bios(struct bio_list *list, struct request *rq) | |
3150 | { | |
3151 | if (rq->bio) { | |
3152 | if (list->tail) | |
3153 | list->tail->bi_next = rq->bio; | |
3154 | else | |
3155 | list->head = rq->bio; | |
3156 | list->tail = rq->biotail; | |
3157 | ||
3158 | rq->bio = NULL; | |
3159 | rq->biotail = NULL; | |
3160 | } | |
3161 | ||
3162 | rq->__data_len = 0; | |
3163 | } | |
3164 | EXPORT_SYMBOL_GPL(blk_steal_bios); | |
3165 | ||
bd63141d ML |
3166 | static size_t order_to_size(unsigned int order) |
3167 | { | |
3168 | return (size_t)PAGE_SIZE << order; | |
3169 | } | |
3170 | ||
3171 | /* called before freeing request pool in @tags */ | |
f32e4eaf JG |
3172 | static void blk_mq_clear_rq_mapping(struct blk_mq_tags *drv_tags, |
3173 | struct blk_mq_tags *tags) | |
bd63141d | 3174 | { |
bd63141d ML |
3175 | struct page *page; |
3176 | unsigned long flags; | |
3177 | ||
76dd2980 YK |
3178 | /* |
3179 | * There is no need to clear mapping if driver tags is not initialized | |
3180 | * or the mapping belongs to the driver tags. | |
3181 | */ | |
3182 | if (!drv_tags || drv_tags == tags) | |
4f245d5b JG |
3183 | return; |
3184 | ||
bd63141d ML |
3185 | list_for_each_entry(page, &tags->page_list, lru) { |
3186 | unsigned long start = (unsigned long)page_address(page); | |
3187 | unsigned long end = start + order_to_size(page->private); | |
3188 | int i; | |
3189 | ||
f32e4eaf | 3190 | for (i = 0; i < drv_tags->nr_tags; i++) { |
bd63141d ML |
3191 | struct request *rq = drv_tags->rqs[i]; |
3192 | unsigned long rq_addr = (unsigned long)rq; | |
3193 | ||
3194 | if (rq_addr >= start && rq_addr < end) { | |
0a467d0f | 3195 | WARN_ON_ONCE(req_ref_read(rq) != 0); |
bd63141d ML |
3196 | cmpxchg(&drv_tags->rqs[i], rq, NULL); |
3197 | } | |
3198 | } | |
3199 | } | |
3200 | ||
3201 | /* | |
3202 | * Wait until all pending iteration is done. | |
3203 | * | |
3204 | * Request reference is cleared and it is guaranteed to be observed | |
3205 | * after the ->lock is released. | |
3206 | */ | |
3207 | spin_lock_irqsave(&drv_tags->lock, flags); | |
3208 | spin_unlock_irqrestore(&drv_tags->lock, flags); | |
3209 | } | |
3210 | ||
cc71a6f4 JA |
3211 | void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags, |
3212 | unsigned int hctx_idx) | |
95363efd | 3213 | { |
f32e4eaf | 3214 | struct blk_mq_tags *drv_tags; |
e9b267d9 | 3215 | struct page *page; |
320ae51f | 3216 | |
e02657ea ML |
3217 | if (list_empty(&tags->page_list)) |
3218 | return; | |
3219 | ||
079a2e3e JG |
3220 | if (blk_mq_is_shared_tags(set->flags)) |
3221 | drv_tags = set->shared_tags; | |
e155b0c2 JG |
3222 | else |
3223 | drv_tags = set->tags[hctx_idx]; | |
f32e4eaf | 3224 | |
65de57bb | 3225 | if (tags->static_rqs && set->ops->exit_request) { |
e9b267d9 | 3226 | int i; |
320ae51f | 3227 | |
24d2f903 | 3228 | for (i = 0; i < tags->nr_tags; i++) { |
2af8cbe3 JA |
3229 | struct request *rq = tags->static_rqs[i]; |
3230 | ||
3231 | if (!rq) | |
e9b267d9 | 3232 | continue; |
d6296d39 | 3233 | set->ops->exit_request(set, rq, hctx_idx); |
2af8cbe3 | 3234 | tags->static_rqs[i] = NULL; |
e9b267d9 | 3235 | } |
320ae51f | 3236 | } |
320ae51f | 3237 | |
f32e4eaf | 3238 | blk_mq_clear_rq_mapping(drv_tags, tags); |
bd63141d | 3239 | |
24d2f903 CH |
3240 | while (!list_empty(&tags->page_list)) { |
3241 | page = list_first_entry(&tags->page_list, struct page, lru); | |
6753471c | 3242 | list_del_init(&page->lru); |
f75782e4 CM |
3243 | /* |
3244 | * Remove kmemleak object previously allocated in | |
273938bf | 3245 | * blk_mq_alloc_rqs(). |
f75782e4 CM |
3246 | */ |
3247 | kmemleak_free(page_address(page)); | |
320ae51f JA |
3248 | __free_pages(page, page->private); |
3249 | } | |
cc71a6f4 | 3250 | } |
320ae51f | 3251 | |
e155b0c2 | 3252 | void blk_mq_free_rq_map(struct blk_mq_tags *tags) |
cc71a6f4 | 3253 | { |
24d2f903 | 3254 | kfree(tags->rqs); |
cc71a6f4 | 3255 | tags->rqs = NULL; |
2af8cbe3 JA |
3256 | kfree(tags->static_rqs); |
3257 | tags->static_rqs = NULL; | |
320ae51f | 3258 | |
e155b0c2 | 3259 | blk_mq_free_tags(tags); |
320ae51f JA |
3260 | } |
3261 | ||
4d805131 ML |
3262 | static enum hctx_type hctx_idx_to_type(struct blk_mq_tag_set *set, |
3263 | unsigned int hctx_idx) | |
3264 | { | |
3265 | int i; | |
3266 | ||
3267 | for (i = 0; i < set->nr_maps; i++) { | |
3268 | unsigned int start = set->map[i].queue_offset; | |
3269 | unsigned int end = start + set->map[i].nr_queues; | |
3270 | ||
3271 | if (hctx_idx >= start && hctx_idx < end) | |
3272 | break; | |
3273 | } | |
3274 | ||
3275 | if (i >= set->nr_maps) | |
3276 | i = HCTX_TYPE_DEFAULT; | |
3277 | ||
3278 | return i; | |
3279 | } | |
3280 | ||
3281 | static int blk_mq_get_hctx_node(struct blk_mq_tag_set *set, | |
3282 | unsigned int hctx_idx) | |
3283 | { | |
3284 | enum hctx_type type = hctx_idx_to_type(set, hctx_idx); | |
3285 | ||
3286 | return blk_mq_hw_queue_to_node(&set->map[type], hctx_idx); | |
3287 | } | |
3288 | ||
63064be1 JG |
3289 | static struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, |
3290 | unsigned int hctx_idx, | |
3291 | unsigned int nr_tags, | |
e155b0c2 | 3292 | unsigned int reserved_tags) |
320ae51f | 3293 | { |
4d805131 | 3294 | int node = blk_mq_get_hctx_node(set, hctx_idx); |
24d2f903 | 3295 | struct blk_mq_tags *tags; |
320ae51f | 3296 | |
59f082e4 SL |
3297 | if (node == NUMA_NO_NODE) |
3298 | node = set->numa_node; | |
3299 | ||
e155b0c2 JG |
3300 | tags = blk_mq_init_tags(nr_tags, reserved_tags, node, |
3301 | BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags)); | |
24d2f903 CH |
3302 | if (!tags) |
3303 | return NULL; | |
320ae51f | 3304 | |
590b5b7d | 3305 | tags->rqs = kcalloc_node(nr_tags, sizeof(struct request *), |
36e1f3d1 | 3306 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, |
59f082e4 | 3307 | node); |
7edfd681 JC |
3308 | if (!tags->rqs) |
3309 | goto err_free_tags; | |
320ae51f | 3310 | |
590b5b7d KC |
3311 | tags->static_rqs = kcalloc_node(nr_tags, sizeof(struct request *), |
3312 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, | |
3313 | node); | |
7edfd681 JC |
3314 | if (!tags->static_rqs) |
3315 | goto err_free_rqs; | |
2af8cbe3 | 3316 | |
cc71a6f4 | 3317 | return tags; |
7edfd681 JC |
3318 | |
3319 | err_free_rqs: | |
3320 | kfree(tags->rqs); | |
3321 | err_free_tags: | |
3322 | blk_mq_free_tags(tags); | |
3323 | return NULL; | |
cc71a6f4 JA |
3324 | } |
3325 | ||
1d9bd516 TH |
3326 | static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq, |
3327 | unsigned int hctx_idx, int node) | |
3328 | { | |
3329 | int ret; | |
3330 | ||
3331 | if (set->ops->init_request) { | |
3332 | ret = set->ops->init_request(set, rq, hctx_idx, node); | |
3333 | if (ret) | |
3334 | return ret; | |
3335 | } | |
3336 | ||
12f5b931 | 3337 | WRITE_ONCE(rq->state, MQ_RQ_IDLE); |
1d9bd516 TH |
3338 | return 0; |
3339 | } | |
3340 | ||
63064be1 JG |
3341 | static int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, |
3342 | struct blk_mq_tags *tags, | |
3343 | unsigned int hctx_idx, unsigned int depth) | |
cc71a6f4 JA |
3344 | { |
3345 | unsigned int i, j, entries_per_page, max_order = 4; | |
4d805131 | 3346 | int node = blk_mq_get_hctx_node(set, hctx_idx); |
cc71a6f4 | 3347 | size_t rq_size, left; |
59f082e4 | 3348 | |
59f082e4 SL |
3349 | if (node == NUMA_NO_NODE) |
3350 | node = set->numa_node; | |
cc71a6f4 JA |
3351 | |
3352 | INIT_LIST_HEAD(&tags->page_list); | |
3353 | ||
320ae51f JA |
3354 | /* |
3355 | * rq_size is the size of the request plus driver payload, rounded | |
3356 | * to the cacheline size | |
3357 | */ | |
24d2f903 | 3358 | rq_size = round_up(sizeof(struct request) + set->cmd_size, |
320ae51f | 3359 | cache_line_size()); |
cc71a6f4 | 3360 | left = rq_size * depth; |
320ae51f | 3361 | |
cc71a6f4 | 3362 | for (i = 0; i < depth; ) { |
320ae51f JA |
3363 | int this_order = max_order; |
3364 | struct page *page; | |
3365 | int to_do; | |
3366 | void *p; | |
3367 | ||
b3a834b1 | 3368 | while (this_order && left < order_to_size(this_order - 1)) |
320ae51f JA |
3369 | this_order--; |
3370 | ||
3371 | do { | |
59f082e4 | 3372 | page = alloc_pages_node(node, |
36e1f3d1 | 3373 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO, |
a5164405 | 3374 | this_order); |
320ae51f JA |
3375 | if (page) |
3376 | break; | |
3377 | if (!this_order--) | |
3378 | break; | |
3379 | if (order_to_size(this_order) < rq_size) | |
3380 | break; | |
3381 | } while (1); | |
3382 | ||
3383 | if (!page) | |
24d2f903 | 3384 | goto fail; |
320ae51f JA |
3385 | |
3386 | page->private = this_order; | |
24d2f903 | 3387 | list_add_tail(&page->lru, &tags->page_list); |
320ae51f JA |
3388 | |
3389 | p = page_address(page); | |
f75782e4 CM |
3390 | /* |
3391 | * Allow kmemleak to scan these pages as they contain pointers | |
3392 | * to additional allocations like via ops->init_request(). | |
3393 | */ | |
36e1f3d1 | 3394 | kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO); |
320ae51f | 3395 | entries_per_page = order_to_size(this_order) / rq_size; |
cc71a6f4 | 3396 | to_do = min(entries_per_page, depth - i); |
320ae51f JA |
3397 | left -= to_do * rq_size; |
3398 | for (j = 0; j < to_do; j++) { | |
2af8cbe3 JA |
3399 | struct request *rq = p; |
3400 | ||
3401 | tags->static_rqs[i] = rq; | |
1d9bd516 TH |
3402 | if (blk_mq_init_request(set, rq, hctx_idx, node)) { |
3403 | tags->static_rqs[i] = NULL; | |
3404 | goto fail; | |
e9b267d9 CH |
3405 | } |
3406 | ||
320ae51f JA |
3407 | p += rq_size; |
3408 | i++; | |
3409 | } | |
3410 | } | |
cc71a6f4 | 3411 | return 0; |
320ae51f | 3412 | |
24d2f903 | 3413 | fail: |
cc71a6f4 JA |
3414 | blk_mq_free_rqs(set, tags, hctx_idx); |
3415 | return -ENOMEM; | |
320ae51f JA |
3416 | } |
3417 | ||
bf0beec0 ML |
3418 | struct rq_iter_data { |
3419 | struct blk_mq_hw_ctx *hctx; | |
3420 | bool has_rq; | |
3421 | }; | |
3422 | ||
2dd6532e | 3423 | static bool blk_mq_has_request(struct request *rq, void *data) |
bf0beec0 ML |
3424 | { |
3425 | struct rq_iter_data *iter_data = data; | |
3426 | ||
3427 | if (rq->mq_hctx != iter_data->hctx) | |
3428 | return true; | |
3429 | iter_data->has_rq = true; | |
3430 | return false; | |
3431 | } | |
3432 | ||
3433 | static bool blk_mq_hctx_has_requests(struct blk_mq_hw_ctx *hctx) | |
3434 | { | |
3435 | struct blk_mq_tags *tags = hctx->sched_tags ? | |
3436 | hctx->sched_tags : hctx->tags; | |
3437 | struct rq_iter_data data = { | |
3438 | .hctx = hctx, | |
3439 | }; | |
3440 | ||
3441 | blk_mq_all_tag_iter(tags, blk_mq_has_request, &data); | |
3442 | return data.has_rq; | |
3443 | } | |
3444 | ||
3445 | static inline bool blk_mq_last_cpu_in_hctx(unsigned int cpu, | |
3446 | struct blk_mq_hw_ctx *hctx) | |
3447 | { | |
9b51d9d8 | 3448 | if (cpumask_first_and(hctx->cpumask, cpu_online_mask) != cpu) |
bf0beec0 ML |
3449 | return false; |
3450 | if (cpumask_next_and(cpu, hctx->cpumask, cpu_online_mask) < nr_cpu_ids) | |
3451 | return false; | |
3452 | return true; | |
3453 | } | |
3454 | ||
3455 | static int blk_mq_hctx_notify_offline(unsigned int cpu, struct hlist_node *node) | |
3456 | { | |
3457 | struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node, | |
3458 | struct blk_mq_hw_ctx, cpuhp_online); | |
3459 | ||
3460 | if (!cpumask_test_cpu(cpu, hctx->cpumask) || | |
3461 | !blk_mq_last_cpu_in_hctx(cpu, hctx)) | |
3462 | return 0; | |
3463 | ||
3464 | /* | |
3465 | * Prevent new request from being allocated on the current hctx. | |
3466 | * | |
3467 | * The smp_mb__after_atomic() Pairs with the implied barrier in | |
3468 | * test_and_set_bit_lock in sbitmap_get(). Ensures the inactive flag is | |
3469 | * seen once we return from the tag allocator. | |
3470 | */ | |
3471 | set_bit(BLK_MQ_S_INACTIVE, &hctx->state); | |
3472 | smp_mb__after_atomic(); | |
3473 | ||
3474 | /* | |
3475 | * Try to grab a reference to the queue and wait for any outstanding | |
3476 | * requests. If we could not grab a reference the queue has been | |
3477 | * frozen and there are no requests. | |
3478 | */ | |
3479 | if (percpu_ref_tryget(&hctx->queue->q_usage_counter)) { | |
3480 | while (blk_mq_hctx_has_requests(hctx)) | |
3481 | msleep(5); | |
3482 | percpu_ref_put(&hctx->queue->q_usage_counter); | |
3483 | } | |
3484 | ||
3485 | return 0; | |
3486 | } | |
3487 | ||
3488 | static int blk_mq_hctx_notify_online(unsigned int cpu, struct hlist_node *node) | |
3489 | { | |
3490 | struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node, | |
3491 | struct blk_mq_hw_ctx, cpuhp_online); | |
3492 | ||
3493 | if (cpumask_test_cpu(cpu, hctx->cpumask)) | |
3494 | clear_bit(BLK_MQ_S_INACTIVE, &hctx->state); | |
3495 | return 0; | |
3496 | } | |
3497 | ||
e57690fe JA |
3498 | /* |
3499 | * 'cpu' is going away. splice any existing rq_list entries from this | |
3500 | * software queue to the hw queue dispatch list, and ensure that it | |
3501 | * gets run. | |
3502 | */ | |
9467f859 | 3503 | static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node) |
484b4061 | 3504 | { |
9467f859 | 3505 | struct blk_mq_hw_ctx *hctx; |
484b4061 JA |
3506 | struct blk_mq_ctx *ctx; |
3507 | LIST_HEAD(tmp); | |
c16d6b5a | 3508 | enum hctx_type type; |
484b4061 | 3509 | |
9467f859 | 3510 | hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead); |
bf0beec0 ML |
3511 | if (!cpumask_test_cpu(cpu, hctx->cpumask)) |
3512 | return 0; | |
3513 | ||
e57690fe | 3514 | ctx = __blk_mq_get_ctx(hctx->queue, cpu); |
c16d6b5a | 3515 | type = hctx->type; |
484b4061 JA |
3516 | |
3517 | spin_lock(&ctx->lock); | |
c16d6b5a ML |
3518 | if (!list_empty(&ctx->rq_lists[type])) { |
3519 | list_splice_init(&ctx->rq_lists[type], &tmp); | |
484b4061 JA |
3520 | blk_mq_hctx_clear_pending(hctx, ctx); |
3521 | } | |
3522 | spin_unlock(&ctx->lock); | |
3523 | ||
3524 | if (list_empty(&tmp)) | |
9467f859 | 3525 | return 0; |
484b4061 | 3526 | |
e57690fe JA |
3527 | spin_lock(&hctx->lock); |
3528 | list_splice_tail_init(&tmp, &hctx->dispatch); | |
3529 | spin_unlock(&hctx->lock); | |
484b4061 JA |
3530 | |
3531 | blk_mq_run_hw_queue(hctx, true); | |
9467f859 | 3532 | return 0; |
484b4061 JA |
3533 | } |
3534 | ||
9467f859 | 3535 | static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx) |
484b4061 | 3536 | { |
bf0beec0 ML |
3537 | if (!(hctx->flags & BLK_MQ_F_STACKING)) |
3538 | cpuhp_state_remove_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE, | |
3539 | &hctx->cpuhp_online); | |
9467f859 TG |
3540 | cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD, |
3541 | &hctx->cpuhp_dead); | |
484b4061 JA |
3542 | } |
3543 | ||
364b6181 ML |
3544 | /* |
3545 | * Before freeing hw queue, clearing the flush request reference in | |
3546 | * tags->rqs[] for avoiding potential UAF. | |
3547 | */ | |
3548 | static void blk_mq_clear_flush_rq_mapping(struct blk_mq_tags *tags, | |
3549 | unsigned int queue_depth, struct request *flush_rq) | |
3550 | { | |
3551 | int i; | |
3552 | unsigned long flags; | |
3553 | ||
3554 | /* The hw queue may not be mapped yet */ | |
3555 | if (!tags) | |
3556 | return; | |
3557 | ||
0a467d0f | 3558 | WARN_ON_ONCE(req_ref_read(flush_rq) != 0); |
364b6181 ML |
3559 | |
3560 | for (i = 0; i < queue_depth; i++) | |
3561 | cmpxchg(&tags->rqs[i], flush_rq, NULL); | |
3562 | ||
3563 | /* | |
3564 | * Wait until all pending iteration is done. | |
3565 | * | |
3566 | * Request reference is cleared and it is guaranteed to be observed | |
3567 | * after the ->lock is released. | |
3568 | */ | |
3569 | spin_lock_irqsave(&tags->lock, flags); | |
3570 | spin_unlock_irqrestore(&tags->lock, flags); | |
3571 | } | |
3572 | ||
c3b4afca | 3573 | /* hctx->ctxs will be freed in queue's release handler */ |
08e98fc6 ML |
3574 | static void blk_mq_exit_hctx(struct request_queue *q, |
3575 | struct blk_mq_tag_set *set, | |
3576 | struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx) | |
3577 | { | |
364b6181 ML |
3578 | struct request *flush_rq = hctx->fq->flush_rq; |
3579 | ||
8ab0b7dc ML |
3580 | if (blk_mq_hw_queue_mapped(hctx)) |
3581 | blk_mq_tag_idle(hctx); | |
08e98fc6 | 3582 | |
6cfeadbf ML |
3583 | if (blk_queue_init_done(q)) |
3584 | blk_mq_clear_flush_rq_mapping(set->tags[hctx_idx], | |
3585 | set->queue_depth, flush_rq); | |
f70ced09 | 3586 | if (set->ops->exit_request) |
364b6181 | 3587 | set->ops->exit_request(set, flush_rq, hctx_idx); |
f70ced09 | 3588 | |
08e98fc6 ML |
3589 | if (set->ops->exit_hctx) |
3590 | set->ops->exit_hctx(hctx, hctx_idx); | |
3591 | ||
9467f859 | 3592 | blk_mq_remove_cpuhp(hctx); |
2f8f1336 | 3593 | |
4e5cc99e ML |
3594 | xa_erase(&q->hctx_table, hctx_idx); |
3595 | ||
2f8f1336 ML |
3596 | spin_lock(&q->unused_hctx_lock); |
3597 | list_add(&hctx->hctx_list, &q->unused_hctx_list); | |
3598 | spin_unlock(&q->unused_hctx_lock); | |
08e98fc6 ML |
3599 | } |
3600 | ||
624dbe47 ML |
3601 | static void blk_mq_exit_hw_queues(struct request_queue *q, |
3602 | struct blk_mq_tag_set *set, int nr_queue) | |
3603 | { | |
3604 | struct blk_mq_hw_ctx *hctx; | |
4f481208 | 3605 | unsigned long i; |
624dbe47 ML |
3606 | |
3607 | queue_for_each_hw_ctx(q, hctx, i) { | |
3608 | if (i == nr_queue) | |
3609 | break; | |
08e98fc6 | 3610 | blk_mq_exit_hctx(q, set, hctx, i); |
624dbe47 | 3611 | } |
624dbe47 ML |
3612 | } |
3613 | ||
08e98fc6 ML |
3614 | static int blk_mq_init_hctx(struct request_queue *q, |
3615 | struct blk_mq_tag_set *set, | |
3616 | struct blk_mq_hw_ctx *hctx, unsigned hctx_idx) | |
320ae51f | 3617 | { |
7c6c5b7c ML |
3618 | hctx->queue_num = hctx_idx; |
3619 | ||
bf0beec0 ML |
3620 | if (!(hctx->flags & BLK_MQ_F_STACKING)) |
3621 | cpuhp_state_add_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE, | |
3622 | &hctx->cpuhp_online); | |
7c6c5b7c ML |
3623 | cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead); |
3624 | ||
3625 | hctx->tags = set->tags[hctx_idx]; | |
3626 | ||
3627 | if (set->ops->init_hctx && | |
3628 | set->ops->init_hctx(hctx, set->driver_data, hctx_idx)) | |
3629 | goto unregister_cpu_notifier; | |
08e98fc6 | 3630 | |
7c6c5b7c ML |
3631 | if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx, |
3632 | hctx->numa_node)) | |
3633 | goto exit_hctx; | |
4e5cc99e ML |
3634 | |
3635 | if (xa_insert(&q->hctx_table, hctx_idx, hctx, GFP_KERNEL)) | |
3636 | goto exit_flush_rq; | |
3637 | ||
7c6c5b7c ML |
3638 | return 0; |
3639 | ||
4e5cc99e ML |
3640 | exit_flush_rq: |
3641 | if (set->ops->exit_request) | |
3642 | set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx); | |
7c6c5b7c ML |
3643 | exit_hctx: |
3644 | if (set->ops->exit_hctx) | |
3645 | set->ops->exit_hctx(hctx, hctx_idx); | |
3646 | unregister_cpu_notifier: | |
3647 | blk_mq_remove_cpuhp(hctx); | |
3648 | return -1; | |
3649 | } | |
3650 | ||
3651 | static struct blk_mq_hw_ctx * | |
3652 | blk_mq_alloc_hctx(struct request_queue *q, struct blk_mq_tag_set *set, | |
3653 | int node) | |
3654 | { | |
3655 | struct blk_mq_hw_ctx *hctx; | |
3656 | gfp_t gfp = GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY; | |
3657 | ||
704b914f | 3658 | hctx = kzalloc_node(sizeof(struct blk_mq_hw_ctx), gfp, node); |
7c6c5b7c ML |
3659 | if (!hctx) |
3660 | goto fail_alloc_hctx; | |
3661 | ||
3662 | if (!zalloc_cpumask_var_node(&hctx->cpumask, gfp, node)) | |
3663 | goto free_hctx; | |
3664 | ||
3665 | atomic_set(&hctx->nr_active, 0); | |
08e98fc6 | 3666 | if (node == NUMA_NO_NODE) |
7c6c5b7c ML |
3667 | node = set->numa_node; |
3668 | hctx->numa_node = node; | |
08e98fc6 | 3669 | |
9f993737 | 3670 | INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn); |
08e98fc6 ML |
3671 | spin_lock_init(&hctx->lock); |
3672 | INIT_LIST_HEAD(&hctx->dispatch); | |
3673 | hctx->queue = q; | |
51db1c37 | 3674 | hctx->flags = set->flags & ~BLK_MQ_F_TAG_QUEUE_SHARED; |
08e98fc6 | 3675 | |
2f8f1336 ML |
3676 | INIT_LIST_HEAD(&hctx->hctx_list); |
3677 | ||
320ae51f | 3678 | /* |
08e98fc6 ML |
3679 | * Allocate space for all possible cpus to avoid allocation at |
3680 | * runtime | |
320ae51f | 3681 | */ |
d904bfa7 | 3682 | hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *), |
7c6c5b7c | 3683 | gfp, node); |
08e98fc6 | 3684 | if (!hctx->ctxs) |
7c6c5b7c | 3685 | goto free_cpumask; |
320ae51f | 3686 | |
5b202853 | 3687 | if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8), |
c548e62b | 3688 | gfp, node, false, false)) |
08e98fc6 | 3689 | goto free_ctxs; |
08e98fc6 | 3690 | hctx->nr_ctx = 0; |
320ae51f | 3691 | |
5815839b | 3692 | spin_lock_init(&hctx->dispatch_wait_lock); |
eb619fdb JA |
3693 | init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake); |
3694 | INIT_LIST_HEAD(&hctx->dispatch_wait.entry); | |
3695 | ||
754a1572 | 3696 | hctx->fq = blk_alloc_flush_queue(hctx->numa_node, set->cmd_size, gfp); |
f70ced09 | 3697 | if (!hctx->fq) |
7c6c5b7c | 3698 | goto free_bitmap; |
320ae51f | 3699 | |
7c6c5b7c | 3700 | blk_mq_hctx_kobj_init(hctx); |
6a83e74d | 3701 | |
7c6c5b7c | 3702 | return hctx; |
320ae51f | 3703 | |
08e98fc6 | 3704 | free_bitmap: |
88459642 | 3705 | sbitmap_free(&hctx->ctx_map); |
08e98fc6 ML |
3706 | free_ctxs: |
3707 | kfree(hctx->ctxs); | |
7c6c5b7c ML |
3708 | free_cpumask: |
3709 | free_cpumask_var(hctx->cpumask); | |
3710 | free_hctx: | |
3711 | kfree(hctx); | |
3712 | fail_alloc_hctx: | |
3713 | return NULL; | |
08e98fc6 | 3714 | } |
320ae51f | 3715 | |
320ae51f JA |
3716 | static void blk_mq_init_cpu_queues(struct request_queue *q, |
3717 | unsigned int nr_hw_queues) | |
3718 | { | |
b3c661b1 JA |
3719 | struct blk_mq_tag_set *set = q->tag_set; |
3720 | unsigned int i, j; | |
320ae51f JA |
3721 | |
3722 | for_each_possible_cpu(i) { | |
3723 | struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i); | |
3724 | struct blk_mq_hw_ctx *hctx; | |
c16d6b5a | 3725 | int k; |
320ae51f | 3726 | |
320ae51f JA |
3727 | __ctx->cpu = i; |
3728 | spin_lock_init(&__ctx->lock); | |
c16d6b5a ML |
3729 | for (k = HCTX_TYPE_DEFAULT; k < HCTX_MAX_TYPES; k++) |
3730 | INIT_LIST_HEAD(&__ctx->rq_lists[k]); | |
3731 | ||
320ae51f JA |
3732 | __ctx->queue = q; |
3733 | ||
320ae51f JA |
3734 | /* |
3735 | * Set local node, IFF we have more than one hw queue. If | |
3736 | * not, we remain on the home node of the device | |
3737 | */ | |
b3c661b1 JA |
3738 | for (j = 0; j < set->nr_maps; j++) { |
3739 | hctx = blk_mq_map_queue_type(q, j, i); | |
3740 | if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE) | |
576e85c5 | 3741 | hctx->numa_node = cpu_to_node(i); |
b3c661b1 | 3742 | } |
320ae51f JA |
3743 | } |
3744 | } | |
3745 | ||
63064be1 JG |
3746 | struct blk_mq_tags *blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set *set, |
3747 | unsigned int hctx_idx, | |
3748 | unsigned int depth) | |
cc71a6f4 | 3749 | { |
63064be1 JG |
3750 | struct blk_mq_tags *tags; |
3751 | int ret; | |
cc71a6f4 | 3752 | |
e155b0c2 | 3753 | tags = blk_mq_alloc_rq_map(set, hctx_idx, depth, set->reserved_tags); |
63064be1 JG |
3754 | if (!tags) |
3755 | return NULL; | |
cc71a6f4 | 3756 | |
63064be1 JG |
3757 | ret = blk_mq_alloc_rqs(set, tags, hctx_idx, depth); |
3758 | if (ret) { | |
e155b0c2 | 3759 | blk_mq_free_rq_map(tags); |
63064be1 JG |
3760 | return NULL; |
3761 | } | |
cc71a6f4 | 3762 | |
63064be1 | 3763 | return tags; |
cc71a6f4 JA |
3764 | } |
3765 | ||
63064be1 JG |
3766 | static bool __blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set *set, |
3767 | int hctx_idx) | |
cc71a6f4 | 3768 | { |
079a2e3e JG |
3769 | if (blk_mq_is_shared_tags(set->flags)) { |
3770 | set->tags[hctx_idx] = set->shared_tags; | |
1c0706a7 | 3771 | |
e155b0c2 | 3772 | return true; |
bd166ef1 | 3773 | } |
e155b0c2 | 3774 | |
63064be1 JG |
3775 | set->tags[hctx_idx] = blk_mq_alloc_map_and_rqs(set, hctx_idx, |
3776 | set->queue_depth); | |
3777 | ||
3778 | return set->tags[hctx_idx]; | |
cc71a6f4 JA |
3779 | } |
3780 | ||
645db34e JG |
3781 | void blk_mq_free_map_and_rqs(struct blk_mq_tag_set *set, |
3782 | struct blk_mq_tags *tags, | |
3783 | unsigned int hctx_idx) | |
cc71a6f4 | 3784 | { |
645db34e JG |
3785 | if (tags) { |
3786 | blk_mq_free_rqs(set, tags, hctx_idx); | |
e155b0c2 | 3787 | blk_mq_free_rq_map(tags); |
bd166ef1 | 3788 | } |
cc71a6f4 JA |
3789 | } |
3790 | ||
e155b0c2 JG |
3791 | static void __blk_mq_free_map_and_rqs(struct blk_mq_tag_set *set, |
3792 | unsigned int hctx_idx) | |
3793 | { | |
079a2e3e | 3794 | if (!blk_mq_is_shared_tags(set->flags)) |
e155b0c2 JG |
3795 | blk_mq_free_map_and_rqs(set, set->tags[hctx_idx], hctx_idx); |
3796 | ||
3797 | set->tags[hctx_idx] = NULL; | |
cc71a6f4 JA |
3798 | } |
3799 | ||
4b855ad3 | 3800 | static void blk_mq_map_swqueue(struct request_queue *q) |
320ae51f | 3801 | { |
4f481208 ML |
3802 | unsigned int j, hctx_idx; |
3803 | unsigned long i; | |
320ae51f JA |
3804 | struct blk_mq_hw_ctx *hctx; |
3805 | struct blk_mq_ctx *ctx; | |
2a34c087 | 3806 | struct blk_mq_tag_set *set = q->tag_set; |
320ae51f JA |
3807 | |
3808 | queue_for_each_hw_ctx(q, hctx, i) { | |
e4043dcf | 3809 | cpumask_clear(hctx->cpumask); |
320ae51f | 3810 | hctx->nr_ctx = 0; |
d416c92c | 3811 | hctx->dispatch_from = NULL; |
320ae51f JA |
3812 | } |
3813 | ||
3814 | /* | |
4b855ad3 | 3815 | * Map software to hardware queues. |
4412efec ML |
3816 | * |
3817 | * If the cpu isn't present, the cpu is mapped to first hctx. | |
320ae51f | 3818 | */ |
20e4d813 | 3819 | for_each_possible_cpu(i) { |
4412efec | 3820 | |
897bb0c7 | 3821 | ctx = per_cpu_ptr(q->queue_ctx, i); |
b3c661b1 | 3822 | for (j = 0; j < set->nr_maps; j++) { |
bb94aea1 JW |
3823 | if (!set->map[j].nr_queues) { |
3824 | ctx->hctxs[j] = blk_mq_map_queue_type(q, | |
3825 | HCTX_TYPE_DEFAULT, i); | |
e5edd5f2 | 3826 | continue; |
bb94aea1 | 3827 | } |
fd689871 ML |
3828 | hctx_idx = set->map[j].mq_map[i]; |
3829 | /* unmapped hw queue can be remapped after CPU topo changed */ | |
3830 | if (!set->tags[hctx_idx] && | |
63064be1 | 3831 | !__blk_mq_alloc_map_and_rqs(set, hctx_idx)) { |
fd689871 ML |
3832 | /* |
3833 | * If tags initialization fail for some hctx, | |
3834 | * that hctx won't be brought online. In this | |
3835 | * case, remap the current ctx to hctx[0] which | |
3836 | * is guaranteed to always have tags allocated | |
3837 | */ | |
3838 | set->map[j].mq_map[i] = 0; | |
3839 | } | |
e5edd5f2 | 3840 | |
b3c661b1 | 3841 | hctx = blk_mq_map_queue_type(q, j, i); |
8ccdf4a3 | 3842 | ctx->hctxs[j] = hctx; |
b3c661b1 JA |
3843 | /* |
3844 | * If the CPU is already set in the mask, then we've | |
3845 | * mapped this one already. This can happen if | |
3846 | * devices share queues across queue maps. | |
3847 | */ | |
3848 | if (cpumask_test_cpu(i, hctx->cpumask)) | |
3849 | continue; | |
3850 | ||
3851 | cpumask_set_cpu(i, hctx->cpumask); | |
3852 | hctx->type = j; | |
3853 | ctx->index_hw[hctx->type] = hctx->nr_ctx; | |
3854 | hctx->ctxs[hctx->nr_ctx++] = ctx; | |
3855 | ||
3856 | /* | |
3857 | * If the nr_ctx type overflows, we have exceeded the | |
3858 | * amount of sw queues we can support. | |
3859 | */ | |
3860 | BUG_ON(!hctx->nr_ctx); | |
3861 | } | |
bb94aea1 JW |
3862 | |
3863 | for (; j < HCTX_MAX_TYPES; j++) | |
3864 | ctx->hctxs[j] = blk_mq_map_queue_type(q, | |
3865 | HCTX_TYPE_DEFAULT, i); | |
320ae51f | 3866 | } |
506e931f JA |
3867 | |
3868 | queue_for_each_hw_ctx(q, hctx, i) { | |
4412efec ML |
3869 | /* |
3870 | * If no software queues are mapped to this hardware queue, | |
3871 | * disable it and free the request entries. | |
3872 | */ | |
3873 | if (!hctx->nr_ctx) { | |
3874 | /* Never unmap queue 0. We need it as a | |
3875 | * fallback in case of a new remap fails | |
3876 | * allocation | |
3877 | */ | |
e155b0c2 JG |
3878 | if (i) |
3879 | __blk_mq_free_map_and_rqs(set, i); | |
4412efec ML |
3880 | |
3881 | hctx->tags = NULL; | |
3882 | continue; | |
3883 | } | |
484b4061 | 3884 | |
2a34c087 ML |
3885 | hctx->tags = set->tags[i]; |
3886 | WARN_ON(!hctx->tags); | |
3887 | ||
889fa31f CY |
3888 | /* |
3889 | * Set the map size to the number of mapped software queues. | |
3890 | * This is more accurate and more efficient than looping | |
3891 | * over all possibly mapped software queues. | |
3892 | */ | |
88459642 | 3893 | sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx); |
889fa31f | 3894 | |
484b4061 JA |
3895 | /* |
3896 | * Initialize batch roundrobin counts | |
3897 | */ | |
f82ddf19 | 3898 | hctx->next_cpu = blk_mq_first_mapped_cpu(hctx); |
506e931f JA |
3899 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; |
3900 | } | |
320ae51f JA |
3901 | } |
3902 | ||
8e8320c9 JA |
3903 | /* |
3904 | * Caller needs to ensure that we're either frozen/quiesced, or that | |
3905 | * the queue isn't live yet. | |
3906 | */ | |
2404e607 | 3907 | static void queue_set_hctx_shared(struct request_queue *q, bool shared) |
0d2602ca JA |
3908 | { |
3909 | struct blk_mq_hw_ctx *hctx; | |
4f481208 | 3910 | unsigned long i; |
0d2602ca | 3911 | |
2404e607 | 3912 | queue_for_each_hw_ctx(q, hctx, i) { |
454bb677 | 3913 | if (shared) { |
51db1c37 | 3914 | hctx->flags |= BLK_MQ_F_TAG_QUEUE_SHARED; |
454bb677 YK |
3915 | } else { |
3916 | blk_mq_tag_idle(hctx); | |
51db1c37 | 3917 | hctx->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED; |
454bb677 | 3918 | } |
2404e607 JM |
3919 | } |
3920 | } | |
3921 | ||
655ac300 HR |
3922 | static void blk_mq_update_tag_set_shared(struct blk_mq_tag_set *set, |
3923 | bool shared) | |
2404e607 JM |
3924 | { |
3925 | struct request_queue *q; | |
0d2602ca | 3926 | |
705cda97 BVA |
3927 | lockdep_assert_held(&set->tag_list_lock); |
3928 | ||
0d2602ca JA |
3929 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
3930 | blk_mq_freeze_queue(q); | |
2404e607 | 3931 | queue_set_hctx_shared(q, shared); |
0d2602ca JA |
3932 | blk_mq_unfreeze_queue(q); |
3933 | } | |
3934 | } | |
3935 | ||
3936 | static void blk_mq_del_queue_tag_set(struct request_queue *q) | |
3937 | { | |
3938 | struct blk_mq_tag_set *set = q->tag_set; | |
3939 | ||
0d2602ca | 3940 | mutex_lock(&set->tag_list_lock); |
08c875cb | 3941 | list_del(&q->tag_set_list); |
2404e607 JM |
3942 | if (list_is_singular(&set->tag_list)) { |
3943 | /* just transitioned to unshared */ | |
51db1c37 | 3944 | set->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED; |
2404e607 | 3945 | /* update existing queue */ |
655ac300 | 3946 | blk_mq_update_tag_set_shared(set, false); |
2404e607 | 3947 | } |
0d2602ca | 3948 | mutex_unlock(&set->tag_list_lock); |
a347c7ad | 3949 | INIT_LIST_HEAD(&q->tag_set_list); |
0d2602ca JA |
3950 | } |
3951 | ||
3952 | static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set, | |
3953 | struct request_queue *q) | |
3954 | { | |
0d2602ca | 3955 | mutex_lock(&set->tag_list_lock); |
2404e607 | 3956 | |
ff821d27 JA |
3957 | /* |
3958 | * Check to see if we're transitioning to shared (from 1 to 2 queues). | |
3959 | */ | |
3960 | if (!list_empty(&set->tag_list) && | |
51db1c37 ML |
3961 | !(set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) { |
3962 | set->flags |= BLK_MQ_F_TAG_QUEUE_SHARED; | |
2404e607 | 3963 | /* update existing queue */ |
655ac300 | 3964 | blk_mq_update_tag_set_shared(set, true); |
2404e607 | 3965 | } |
51db1c37 | 3966 | if (set->flags & BLK_MQ_F_TAG_QUEUE_SHARED) |
2404e607 | 3967 | queue_set_hctx_shared(q, true); |
08c875cb | 3968 | list_add_tail(&q->tag_set_list, &set->tag_list); |
2404e607 | 3969 | |
0d2602ca JA |
3970 | mutex_unlock(&set->tag_list_lock); |
3971 | } | |
3972 | ||
1db4909e ML |
3973 | /* All allocations will be freed in release handler of q->mq_kobj */ |
3974 | static int blk_mq_alloc_ctxs(struct request_queue *q) | |
3975 | { | |
3976 | struct blk_mq_ctxs *ctxs; | |
3977 | int cpu; | |
3978 | ||
3979 | ctxs = kzalloc(sizeof(*ctxs), GFP_KERNEL); | |
3980 | if (!ctxs) | |
3981 | return -ENOMEM; | |
3982 | ||
3983 | ctxs->queue_ctx = alloc_percpu(struct blk_mq_ctx); | |
3984 | if (!ctxs->queue_ctx) | |
3985 | goto fail; | |
3986 | ||
3987 | for_each_possible_cpu(cpu) { | |
3988 | struct blk_mq_ctx *ctx = per_cpu_ptr(ctxs->queue_ctx, cpu); | |
3989 | ctx->ctxs = ctxs; | |
3990 | } | |
3991 | ||
3992 | q->mq_kobj = &ctxs->kobj; | |
3993 | q->queue_ctx = ctxs->queue_ctx; | |
3994 | ||
3995 | return 0; | |
3996 | fail: | |
3997 | kfree(ctxs); | |
3998 | return -ENOMEM; | |
3999 | } | |
4000 | ||
e09aae7e ML |
4001 | /* |
4002 | * It is the actual release handler for mq, but we do it from | |
4003 | * request queue's release handler for avoiding use-after-free | |
4004 | * and headache because q->mq_kobj shouldn't have been introduced, | |
4005 | * but we can't group ctx/kctx kobj without it. | |
4006 | */ | |
4007 | void blk_mq_release(struct request_queue *q) | |
4008 | { | |
2f8f1336 | 4009 | struct blk_mq_hw_ctx *hctx, *next; |
4f481208 | 4010 | unsigned long i; |
e09aae7e | 4011 | |
2f8f1336 ML |
4012 | queue_for_each_hw_ctx(q, hctx, i) |
4013 | WARN_ON_ONCE(hctx && list_empty(&hctx->hctx_list)); | |
4014 | ||
4015 | /* all hctx are in .unused_hctx_list now */ | |
4016 | list_for_each_entry_safe(hctx, next, &q->unused_hctx_list, hctx_list) { | |
4017 | list_del_init(&hctx->hctx_list); | |
6c8b232e | 4018 | kobject_put(&hctx->kobj); |
c3b4afca | 4019 | } |
e09aae7e | 4020 | |
4e5cc99e | 4021 | xa_destroy(&q->hctx_table); |
e09aae7e | 4022 | |
7ea5fe31 ML |
4023 | /* |
4024 | * release .mq_kobj and sw queue's kobject now because | |
4025 | * both share lifetime with request queue. | |
4026 | */ | |
4027 | blk_mq_sysfs_deinit(q); | |
e09aae7e ML |
4028 | } |
4029 | ||
5ec780a6 | 4030 | static struct request_queue *blk_mq_init_queue_data(struct blk_mq_tag_set *set, |
2f227bb9 | 4031 | void *queuedata) |
b62c21b7 | 4032 | { |
26a9750a CH |
4033 | struct request_queue *q; |
4034 | int ret; | |
b62c21b7 | 4035 | |
80bd4a7a | 4036 | q = blk_alloc_queue(set->numa_node); |
26a9750a | 4037 | if (!q) |
b62c21b7 | 4038 | return ERR_PTR(-ENOMEM); |
26a9750a CH |
4039 | q->queuedata = queuedata; |
4040 | ret = blk_mq_init_allocated_queue(set, q); | |
4041 | if (ret) { | |
6f8191fd | 4042 | blk_put_queue(q); |
26a9750a CH |
4043 | return ERR_PTR(ret); |
4044 | } | |
b62c21b7 MS |
4045 | return q; |
4046 | } | |
2f227bb9 CH |
4047 | |
4048 | struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set) | |
4049 | { | |
4050 | return blk_mq_init_queue_data(set, NULL); | |
4051 | } | |
b62c21b7 MS |
4052 | EXPORT_SYMBOL(blk_mq_init_queue); |
4053 | ||
6f8191fd CH |
4054 | /** |
4055 | * blk_mq_destroy_queue - shutdown a request queue | |
4056 | * @q: request queue to shutdown | |
4057 | * | |
81ea42b9 BVA |
4058 | * This shuts down a request queue allocated by blk_mq_init_queue(). All future |
4059 | * requests will be failed with -ENODEV. The caller is responsible for dropping | |
4060 | * the reference from blk_mq_init_queue() by calling blk_put_queue(). | |
6f8191fd CH |
4061 | * |
4062 | * Context: can sleep | |
4063 | */ | |
4064 | void blk_mq_destroy_queue(struct request_queue *q) | |
4065 | { | |
4066 | WARN_ON_ONCE(!queue_is_mq(q)); | |
4067 | WARN_ON_ONCE(blk_queue_registered(q)); | |
4068 | ||
4069 | might_sleep(); | |
4070 | ||
4071 | blk_queue_flag_set(QUEUE_FLAG_DYING, q); | |
4072 | blk_queue_start_drain(q); | |
56c1ee92 | 4073 | blk_mq_freeze_queue_wait(q); |
6f8191fd CH |
4074 | |
4075 | blk_sync_queue(q); | |
4076 | blk_mq_cancel_work_sync(q); | |
4077 | blk_mq_exit_queue(q); | |
6f8191fd CH |
4078 | } |
4079 | EXPORT_SYMBOL(blk_mq_destroy_queue); | |
4080 | ||
4dcc4874 CH |
4081 | struct gendisk *__blk_mq_alloc_disk(struct blk_mq_tag_set *set, void *queuedata, |
4082 | struct lock_class_key *lkclass) | |
9316a9ed JA |
4083 | { |
4084 | struct request_queue *q; | |
b461dfc4 | 4085 | struct gendisk *disk; |
9316a9ed | 4086 | |
b461dfc4 CH |
4087 | q = blk_mq_init_queue_data(set, queuedata); |
4088 | if (IS_ERR(q)) | |
4089 | return ERR_CAST(q); | |
9316a9ed | 4090 | |
4a1fa41d | 4091 | disk = __alloc_disk_node(q, set->numa_node, lkclass); |
b461dfc4 | 4092 | if (!disk) { |
0a3e5cc7 | 4093 | blk_mq_destroy_queue(q); |
2b3f056f | 4094 | blk_put_queue(q); |
b461dfc4 | 4095 | return ERR_PTR(-ENOMEM); |
9316a9ed | 4096 | } |
6f8191fd | 4097 | set_bit(GD_OWNS_QUEUE, &disk->state); |
b461dfc4 | 4098 | return disk; |
9316a9ed | 4099 | } |
b461dfc4 | 4100 | EXPORT_SYMBOL(__blk_mq_alloc_disk); |
9316a9ed | 4101 | |
6f8191fd CH |
4102 | struct gendisk *blk_mq_alloc_disk_for_queue(struct request_queue *q, |
4103 | struct lock_class_key *lkclass) | |
4104 | { | |
22c17e27 CH |
4105 | struct gendisk *disk; |
4106 | ||
6f8191fd CH |
4107 | if (!blk_get_queue(q)) |
4108 | return NULL; | |
22c17e27 CH |
4109 | disk = __alloc_disk_node(q, NUMA_NO_NODE, lkclass); |
4110 | if (!disk) | |
4111 | blk_put_queue(q); | |
4112 | return disk; | |
6f8191fd CH |
4113 | } |
4114 | EXPORT_SYMBOL(blk_mq_alloc_disk_for_queue); | |
4115 | ||
34d11ffa JW |
4116 | static struct blk_mq_hw_ctx *blk_mq_alloc_and_init_hctx( |
4117 | struct blk_mq_tag_set *set, struct request_queue *q, | |
4118 | int hctx_idx, int node) | |
4119 | { | |
2f8f1336 | 4120 | struct blk_mq_hw_ctx *hctx = NULL, *tmp; |
34d11ffa | 4121 | |
2f8f1336 ML |
4122 | /* reuse dead hctx first */ |
4123 | spin_lock(&q->unused_hctx_lock); | |
4124 | list_for_each_entry(tmp, &q->unused_hctx_list, hctx_list) { | |
4125 | if (tmp->numa_node == node) { | |
4126 | hctx = tmp; | |
4127 | break; | |
4128 | } | |
4129 | } | |
4130 | if (hctx) | |
4131 | list_del_init(&hctx->hctx_list); | |
4132 | spin_unlock(&q->unused_hctx_lock); | |
4133 | ||
4134 | if (!hctx) | |
4135 | hctx = blk_mq_alloc_hctx(q, set, node); | |
34d11ffa | 4136 | if (!hctx) |
7c6c5b7c | 4137 | goto fail; |
34d11ffa | 4138 | |
7c6c5b7c ML |
4139 | if (blk_mq_init_hctx(q, set, hctx, hctx_idx)) |
4140 | goto free_hctx; | |
34d11ffa JW |
4141 | |
4142 | return hctx; | |
7c6c5b7c ML |
4143 | |
4144 | free_hctx: | |
4145 | kobject_put(&hctx->kobj); | |
4146 | fail: | |
4147 | return NULL; | |
34d11ffa JW |
4148 | } |
4149 | ||
868f2f0b KB |
4150 | static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set, |
4151 | struct request_queue *q) | |
320ae51f | 4152 | { |
4e5cc99e ML |
4153 | struct blk_mq_hw_ctx *hctx; |
4154 | unsigned long i, j; | |
ac0d6b92 | 4155 | |
fb350e0a ML |
4156 | /* protect against switching io scheduler */ |
4157 | mutex_lock(&q->sysfs_lock); | |
24d2f903 | 4158 | for (i = 0; i < set->nr_hw_queues; i++) { |
306f13ee | 4159 | int old_node; |
4d805131 | 4160 | int node = blk_mq_get_hctx_node(set, i); |
4e5cc99e | 4161 | struct blk_mq_hw_ctx *old_hctx = xa_load(&q->hctx_table, i); |
868f2f0b | 4162 | |
306f13ee ML |
4163 | if (old_hctx) { |
4164 | old_node = old_hctx->numa_node; | |
4165 | blk_mq_exit_hctx(q, set, old_hctx, i); | |
4166 | } | |
868f2f0b | 4167 | |
4e5cc99e | 4168 | if (!blk_mq_alloc_and_init_hctx(set, q, i, node)) { |
306f13ee | 4169 | if (!old_hctx) |
34d11ffa | 4170 | break; |
306f13ee ML |
4171 | pr_warn("Allocate new hctx on node %d fails, fallback to previous one on node %d\n", |
4172 | node, old_node); | |
4e5cc99e ML |
4173 | hctx = blk_mq_alloc_and_init_hctx(set, q, i, old_node); |
4174 | WARN_ON_ONCE(!hctx); | |
868f2f0b | 4175 | } |
320ae51f | 4176 | } |
e01ad46d JW |
4177 | /* |
4178 | * Increasing nr_hw_queues fails. Free the newly allocated | |
4179 | * hctxs and keep the previous q->nr_hw_queues. | |
4180 | */ | |
4181 | if (i != set->nr_hw_queues) { | |
4182 | j = q->nr_hw_queues; | |
e01ad46d JW |
4183 | } else { |
4184 | j = i; | |
e01ad46d JW |
4185 | q->nr_hw_queues = set->nr_hw_queues; |
4186 | } | |
34d11ffa | 4187 | |
4e5cc99e ML |
4188 | xa_for_each_start(&q->hctx_table, j, hctx, j) |
4189 | blk_mq_exit_hctx(q, set, hctx, j); | |
fb350e0a | 4190 | mutex_unlock(&q->sysfs_lock); |
868f2f0b KB |
4191 | } |
4192 | ||
42ee3061 ML |
4193 | static void blk_mq_update_poll_flag(struct request_queue *q) |
4194 | { | |
4195 | struct blk_mq_tag_set *set = q->tag_set; | |
4196 | ||
4197 | if (set->nr_maps > HCTX_TYPE_POLL && | |
4198 | set->map[HCTX_TYPE_POLL].nr_queues) | |
4199 | blk_queue_flag_set(QUEUE_FLAG_POLL, q); | |
4200 | else | |
4201 | blk_queue_flag_clear(QUEUE_FLAG_POLL, q); | |
4202 | } | |
4203 | ||
26a9750a CH |
4204 | int blk_mq_init_allocated_queue(struct blk_mq_tag_set *set, |
4205 | struct request_queue *q) | |
868f2f0b | 4206 | { |
66841672 ML |
4207 | /* mark the queue as mq asap */ |
4208 | q->mq_ops = set->ops; | |
4209 | ||
34dbad5d | 4210 | q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn, |
720b8ccc SB |
4211 | blk_mq_poll_stats_bkt, |
4212 | BLK_MQ_POLL_STATS_BKTS, q); | |
34dbad5d OS |
4213 | if (!q->poll_cb) |
4214 | goto err_exit; | |
4215 | ||
1db4909e | 4216 | if (blk_mq_alloc_ctxs(q)) |
41de54c6 | 4217 | goto err_poll; |
868f2f0b | 4218 | |
737f98cf ML |
4219 | /* init q->mq_kobj and sw queues' kobjects */ |
4220 | blk_mq_sysfs_init(q); | |
4221 | ||
2f8f1336 ML |
4222 | INIT_LIST_HEAD(&q->unused_hctx_list); |
4223 | spin_lock_init(&q->unused_hctx_lock); | |
4224 | ||
4e5cc99e ML |
4225 | xa_init(&q->hctx_table); |
4226 | ||
868f2f0b KB |
4227 | blk_mq_realloc_hw_ctxs(set, q); |
4228 | if (!q->nr_hw_queues) | |
4229 | goto err_hctxs; | |
320ae51f | 4230 | |
287922eb | 4231 | INIT_WORK(&q->timeout_work, blk_mq_timeout_work); |
e56f698b | 4232 | blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ); |
320ae51f | 4233 | |
a8908939 | 4234 | q->tag_set = set; |
320ae51f | 4235 | |
94eddfbe | 4236 | q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT; |
42ee3061 | 4237 | blk_mq_update_poll_flag(q); |
320ae51f | 4238 | |
2849450a | 4239 | INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work); |
6fca6a61 CH |
4240 | INIT_LIST_HEAD(&q->requeue_list); |
4241 | spin_lock_init(&q->requeue_lock); | |
4242 | ||
eba71768 JA |
4243 | q->nr_requests = set->queue_depth; |
4244 | ||
64f1c21e JA |
4245 | /* |
4246 | * Default to classic polling | |
4247 | */ | |
29ece8b4 | 4248 | q->poll_nsec = BLK_MQ_POLL_CLASSIC; |
64f1c21e | 4249 | |
24d2f903 | 4250 | blk_mq_init_cpu_queues(q, set->nr_hw_queues); |
0d2602ca | 4251 | blk_mq_add_queue_tag_set(set, q); |
4b855ad3 | 4252 | blk_mq_map_swqueue(q); |
26a9750a | 4253 | return 0; |
18741986 | 4254 | |
320ae51f | 4255 | err_hctxs: |
943f45b9 | 4256 | blk_mq_release(q); |
41de54c6 JS |
4257 | err_poll: |
4258 | blk_stat_free_callback(q->poll_cb); | |
4259 | q->poll_cb = NULL; | |
c7de5726 ML |
4260 | err_exit: |
4261 | q->mq_ops = NULL; | |
26a9750a | 4262 | return -ENOMEM; |
320ae51f | 4263 | } |
b62c21b7 | 4264 | EXPORT_SYMBOL(blk_mq_init_allocated_queue); |
320ae51f | 4265 | |
c7e2d94b ML |
4266 | /* tags can _not_ be used after returning from blk_mq_exit_queue */ |
4267 | void blk_mq_exit_queue(struct request_queue *q) | |
320ae51f | 4268 | { |
630ef623 | 4269 | struct blk_mq_tag_set *set = q->tag_set; |
320ae51f | 4270 | |
630ef623 | 4271 | /* Checks hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED. */ |
624dbe47 | 4272 | blk_mq_exit_hw_queues(q, set, set->nr_hw_queues); |
630ef623 BVA |
4273 | /* May clear BLK_MQ_F_TAG_QUEUE_SHARED in hctx->flags. */ |
4274 | blk_mq_del_queue_tag_set(q); | |
320ae51f | 4275 | } |
320ae51f | 4276 | |
a5164405 JA |
4277 | static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set) |
4278 | { | |
4279 | int i; | |
4280 | ||
079a2e3e JG |
4281 | if (blk_mq_is_shared_tags(set->flags)) { |
4282 | set->shared_tags = blk_mq_alloc_map_and_rqs(set, | |
e155b0c2 JG |
4283 | BLK_MQ_NO_HCTX_IDX, |
4284 | set->queue_depth); | |
079a2e3e | 4285 | if (!set->shared_tags) |
e155b0c2 JG |
4286 | return -ENOMEM; |
4287 | } | |
4288 | ||
8229cca8 | 4289 | for (i = 0; i < set->nr_hw_queues; i++) { |
63064be1 | 4290 | if (!__blk_mq_alloc_map_and_rqs(set, i)) |
a5164405 | 4291 | goto out_unwind; |
8229cca8 XT |
4292 | cond_resched(); |
4293 | } | |
a5164405 JA |
4294 | |
4295 | return 0; | |
4296 | ||
4297 | out_unwind: | |
4298 | while (--i >= 0) | |
e155b0c2 JG |
4299 | __blk_mq_free_map_and_rqs(set, i); |
4300 | ||
079a2e3e JG |
4301 | if (blk_mq_is_shared_tags(set->flags)) { |
4302 | blk_mq_free_map_and_rqs(set, set->shared_tags, | |
e155b0c2 | 4303 | BLK_MQ_NO_HCTX_IDX); |
645db34e | 4304 | } |
a5164405 | 4305 | |
a5164405 JA |
4306 | return -ENOMEM; |
4307 | } | |
4308 | ||
4309 | /* | |
4310 | * Allocate the request maps associated with this tag_set. Note that this | |
4311 | * may reduce the depth asked for, if memory is tight. set->queue_depth | |
4312 | * will be updated to reflect the allocated depth. | |
4313 | */ | |
63064be1 | 4314 | static int blk_mq_alloc_set_map_and_rqs(struct blk_mq_tag_set *set) |
a5164405 JA |
4315 | { |
4316 | unsigned int depth; | |
4317 | int err; | |
4318 | ||
4319 | depth = set->queue_depth; | |
4320 | do { | |
4321 | err = __blk_mq_alloc_rq_maps(set); | |
4322 | if (!err) | |
4323 | break; | |
4324 | ||
4325 | set->queue_depth >>= 1; | |
4326 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) { | |
4327 | err = -ENOMEM; | |
4328 | break; | |
4329 | } | |
4330 | } while (set->queue_depth); | |
4331 | ||
4332 | if (!set->queue_depth || err) { | |
4333 | pr_err("blk-mq: failed to allocate request map\n"); | |
4334 | return -ENOMEM; | |
4335 | } | |
4336 | ||
4337 | if (depth != set->queue_depth) | |
4338 | pr_info("blk-mq: reduced tag depth (%u -> %u)\n", | |
4339 | depth, set->queue_depth); | |
4340 | ||
4341 | return 0; | |
4342 | } | |
4343 | ||
a4e1d0b7 | 4344 | static void blk_mq_update_queue_map(struct blk_mq_tag_set *set) |
ebe8bddb | 4345 | { |
6e66b493 BVA |
4346 | /* |
4347 | * blk_mq_map_queues() and multiple .map_queues() implementations | |
4348 | * expect that set->map[HCTX_TYPE_DEFAULT].nr_queues is set to the | |
4349 | * number of hardware queues. | |
4350 | */ | |
4351 | if (set->nr_maps == 1) | |
4352 | set->map[HCTX_TYPE_DEFAULT].nr_queues = set->nr_hw_queues; | |
4353 | ||
59388702 | 4354 | if (set->ops->map_queues && !is_kdump_kernel()) { |
b3c661b1 JA |
4355 | int i; |
4356 | ||
7d4901a9 ML |
4357 | /* |
4358 | * transport .map_queues is usually done in the following | |
4359 | * way: | |
4360 | * | |
4361 | * for (queue = 0; queue < set->nr_hw_queues; queue++) { | |
4362 | * mask = get_cpu_mask(queue) | |
4363 | * for_each_cpu(cpu, mask) | |
b3c661b1 | 4364 | * set->map[x].mq_map[cpu] = queue; |
7d4901a9 ML |
4365 | * } |
4366 | * | |
4367 | * When we need to remap, the table has to be cleared for | |
4368 | * killing stale mapping since one CPU may not be mapped | |
4369 | * to any hw queue. | |
4370 | */ | |
b3c661b1 JA |
4371 | for (i = 0; i < set->nr_maps; i++) |
4372 | blk_mq_clear_mq_map(&set->map[i]); | |
7d4901a9 | 4373 | |
a4e1d0b7 | 4374 | set->ops->map_queues(set); |
b3c661b1 JA |
4375 | } else { |
4376 | BUG_ON(set->nr_maps > 1); | |
a4e1d0b7 | 4377 | blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]); |
b3c661b1 | 4378 | } |
ebe8bddb OS |
4379 | } |
4380 | ||
f7e76dbc | 4381 | static int blk_mq_realloc_tag_set_tags(struct blk_mq_tag_set *set, |
ee9d5521 | 4382 | int new_nr_hw_queues) |
f7e76dbc BVA |
4383 | { |
4384 | struct blk_mq_tags **new_tags; | |
4385 | ||
ee9d5521 | 4386 | if (set->nr_hw_queues >= new_nr_hw_queues) |
d4b2e0d4 | 4387 | goto done; |
f7e76dbc BVA |
4388 | |
4389 | new_tags = kcalloc_node(new_nr_hw_queues, sizeof(struct blk_mq_tags *), | |
4390 | GFP_KERNEL, set->numa_node); | |
4391 | if (!new_tags) | |
4392 | return -ENOMEM; | |
4393 | ||
4394 | if (set->tags) | |
ee9d5521 | 4395 | memcpy(new_tags, set->tags, set->nr_hw_queues * |
f7e76dbc BVA |
4396 | sizeof(*set->tags)); |
4397 | kfree(set->tags); | |
4398 | set->tags = new_tags; | |
d4b2e0d4 | 4399 | done: |
f7e76dbc | 4400 | set->nr_hw_queues = new_nr_hw_queues; |
f7e76dbc BVA |
4401 | return 0; |
4402 | } | |
4403 | ||
a4391c64 JA |
4404 | /* |
4405 | * Alloc a tag set to be associated with one or more request queues. | |
4406 | * May fail with EINVAL for various error conditions. May adjust the | |
c018c84f | 4407 | * requested depth down, if it's too large. In that case, the set |
a4391c64 JA |
4408 | * value will be stored in set->queue_depth. |
4409 | */ | |
24d2f903 CH |
4410 | int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set) |
4411 | { | |
b3c661b1 | 4412 | int i, ret; |
da695ba2 | 4413 | |
205fb5f5 BVA |
4414 | BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS); |
4415 | ||
24d2f903 CH |
4416 | if (!set->nr_hw_queues) |
4417 | return -EINVAL; | |
a4391c64 | 4418 | if (!set->queue_depth) |
24d2f903 CH |
4419 | return -EINVAL; |
4420 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) | |
4421 | return -EINVAL; | |
4422 | ||
7d7e0f90 | 4423 | if (!set->ops->queue_rq) |
24d2f903 CH |
4424 | return -EINVAL; |
4425 | ||
de148297 ML |
4426 | if (!set->ops->get_budget ^ !set->ops->put_budget) |
4427 | return -EINVAL; | |
4428 | ||
a4391c64 JA |
4429 | if (set->queue_depth > BLK_MQ_MAX_DEPTH) { |
4430 | pr_info("blk-mq: reduced tag depth to %u\n", | |
4431 | BLK_MQ_MAX_DEPTH); | |
4432 | set->queue_depth = BLK_MQ_MAX_DEPTH; | |
4433 | } | |
24d2f903 | 4434 | |
b3c661b1 JA |
4435 | if (!set->nr_maps) |
4436 | set->nr_maps = 1; | |
4437 | else if (set->nr_maps > HCTX_MAX_TYPES) | |
4438 | return -EINVAL; | |
4439 | ||
6637fadf SL |
4440 | /* |
4441 | * If a crashdump is active, then we are potentially in a very | |
4442 | * memory constrained environment. Limit us to 1 queue and | |
4443 | * 64 tags to prevent using too much memory. | |
4444 | */ | |
4445 | if (is_kdump_kernel()) { | |
4446 | set->nr_hw_queues = 1; | |
59388702 | 4447 | set->nr_maps = 1; |
6637fadf SL |
4448 | set->queue_depth = min(64U, set->queue_depth); |
4449 | } | |
868f2f0b | 4450 | /* |
392546ae JA |
4451 | * There is no use for more h/w queues than cpus if we just have |
4452 | * a single map | |
868f2f0b | 4453 | */ |
392546ae | 4454 | if (set->nr_maps == 1 && set->nr_hw_queues > nr_cpu_ids) |
868f2f0b | 4455 | set->nr_hw_queues = nr_cpu_ids; |
6637fadf | 4456 | |
80bd4a7a CH |
4457 | if (set->flags & BLK_MQ_F_BLOCKING) { |
4458 | set->srcu = kmalloc(sizeof(*set->srcu), GFP_KERNEL); | |
4459 | if (!set->srcu) | |
4460 | return -ENOMEM; | |
4461 | ret = init_srcu_struct(set->srcu); | |
4462 | if (ret) | |
4463 | goto out_free_srcu; | |
4464 | } | |
24d2f903 | 4465 | |
da695ba2 | 4466 | ret = -ENOMEM; |
5ee20298 CH |
4467 | set->tags = kcalloc_node(set->nr_hw_queues, |
4468 | sizeof(struct blk_mq_tags *), GFP_KERNEL, | |
4469 | set->numa_node); | |
4470 | if (!set->tags) | |
80bd4a7a | 4471 | goto out_cleanup_srcu; |
24d2f903 | 4472 | |
b3c661b1 JA |
4473 | for (i = 0; i < set->nr_maps; i++) { |
4474 | set->map[i].mq_map = kcalloc_node(nr_cpu_ids, | |
07b35eb5 | 4475 | sizeof(set->map[i].mq_map[0]), |
b3c661b1 JA |
4476 | GFP_KERNEL, set->numa_node); |
4477 | if (!set->map[i].mq_map) | |
4478 | goto out_free_mq_map; | |
59388702 | 4479 | set->map[i].nr_queues = is_kdump_kernel() ? 1 : set->nr_hw_queues; |
b3c661b1 | 4480 | } |
bdd17e75 | 4481 | |
a4e1d0b7 | 4482 | blk_mq_update_queue_map(set); |
da695ba2 | 4483 | |
63064be1 | 4484 | ret = blk_mq_alloc_set_map_and_rqs(set); |
da695ba2 | 4485 | if (ret) |
bdd17e75 | 4486 | goto out_free_mq_map; |
24d2f903 | 4487 | |
0d2602ca JA |
4488 | mutex_init(&set->tag_list_lock); |
4489 | INIT_LIST_HEAD(&set->tag_list); | |
4490 | ||
24d2f903 | 4491 | return 0; |
bdd17e75 CH |
4492 | |
4493 | out_free_mq_map: | |
b3c661b1 JA |
4494 | for (i = 0; i < set->nr_maps; i++) { |
4495 | kfree(set->map[i].mq_map); | |
4496 | set->map[i].mq_map = NULL; | |
4497 | } | |
5676e7b6 RE |
4498 | kfree(set->tags); |
4499 | set->tags = NULL; | |
80bd4a7a CH |
4500 | out_cleanup_srcu: |
4501 | if (set->flags & BLK_MQ_F_BLOCKING) | |
4502 | cleanup_srcu_struct(set->srcu); | |
4503 | out_free_srcu: | |
4504 | if (set->flags & BLK_MQ_F_BLOCKING) | |
4505 | kfree(set->srcu); | |
da695ba2 | 4506 | return ret; |
24d2f903 CH |
4507 | } |
4508 | EXPORT_SYMBOL(blk_mq_alloc_tag_set); | |
4509 | ||
cdb14e0f CH |
4510 | /* allocate and initialize a tagset for a simple single-queue device */ |
4511 | int blk_mq_alloc_sq_tag_set(struct blk_mq_tag_set *set, | |
4512 | const struct blk_mq_ops *ops, unsigned int queue_depth, | |
4513 | unsigned int set_flags) | |
4514 | { | |
4515 | memset(set, 0, sizeof(*set)); | |
4516 | set->ops = ops; | |
4517 | set->nr_hw_queues = 1; | |
4518 | set->nr_maps = 1; | |
4519 | set->queue_depth = queue_depth; | |
4520 | set->numa_node = NUMA_NO_NODE; | |
4521 | set->flags = set_flags; | |
4522 | return blk_mq_alloc_tag_set(set); | |
4523 | } | |
4524 | EXPORT_SYMBOL_GPL(blk_mq_alloc_sq_tag_set); | |
4525 | ||
24d2f903 CH |
4526 | void blk_mq_free_tag_set(struct blk_mq_tag_set *set) |
4527 | { | |
b3c661b1 | 4528 | int i, j; |
24d2f903 | 4529 | |
f7e76dbc | 4530 | for (i = 0; i < set->nr_hw_queues; i++) |
e155b0c2 | 4531 | __blk_mq_free_map_and_rqs(set, i); |
484b4061 | 4532 | |
079a2e3e JG |
4533 | if (blk_mq_is_shared_tags(set->flags)) { |
4534 | blk_mq_free_map_and_rqs(set, set->shared_tags, | |
e155b0c2 JG |
4535 | BLK_MQ_NO_HCTX_IDX); |
4536 | } | |
32bc15af | 4537 | |
b3c661b1 JA |
4538 | for (j = 0; j < set->nr_maps; j++) { |
4539 | kfree(set->map[j].mq_map); | |
4540 | set->map[j].mq_map = NULL; | |
4541 | } | |
bdd17e75 | 4542 | |
981bd189 | 4543 | kfree(set->tags); |
5676e7b6 | 4544 | set->tags = NULL; |
80bd4a7a CH |
4545 | if (set->flags & BLK_MQ_F_BLOCKING) { |
4546 | cleanup_srcu_struct(set->srcu); | |
4547 | kfree(set->srcu); | |
4548 | } | |
24d2f903 CH |
4549 | } |
4550 | EXPORT_SYMBOL(blk_mq_free_tag_set); | |
4551 | ||
e3a2b3f9 JA |
4552 | int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr) |
4553 | { | |
4554 | struct blk_mq_tag_set *set = q->tag_set; | |
4555 | struct blk_mq_hw_ctx *hctx; | |
4f481208 ML |
4556 | int ret; |
4557 | unsigned long i; | |
e3a2b3f9 | 4558 | |
bd166ef1 | 4559 | if (!set) |
e3a2b3f9 JA |
4560 | return -EINVAL; |
4561 | ||
e5fa8140 AZ |
4562 | if (q->nr_requests == nr) |
4563 | return 0; | |
4564 | ||
70f36b60 | 4565 | blk_mq_freeze_queue(q); |
24f5a90f | 4566 | blk_mq_quiesce_queue(q); |
70f36b60 | 4567 | |
e3a2b3f9 JA |
4568 | ret = 0; |
4569 | queue_for_each_hw_ctx(q, hctx, i) { | |
e9137d4b KB |
4570 | if (!hctx->tags) |
4571 | continue; | |
bd166ef1 JA |
4572 | /* |
4573 | * If we're using an MQ scheduler, just update the scheduler | |
4574 | * queue depth. This is similar to what the old code would do. | |
4575 | */ | |
f6adcef5 | 4576 | if (hctx->sched_tags) { |
70f36b60 | 4577 | ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags, |
f6adcef5 | 4578 | nr, true); |
f6adcef5 JG |
4579 | } else { |
4580 | ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr, | |
4581 | false); | |
70f36b60 | 4582 | } |
e3a2b3f9 JA |
4583 | if (ret) |
4584 | break; | |
77f1e0a5 JA |
4585 | if (q->elevator && q->elevator->type->ops.depth_updated) |
4586 | q->elevator->type->ops.depth_updated(hctx); | |
e3a2b3f9 | 4587 | } |
d97e594c | 4588 | if (!ret) { |
e3a2b3f9 | 4589 | q->nr_requests = nr; |
079a2e3e | 4590 | if (blk_mq_is_shared_tags(set->flags)) { |
8fa04464 | 4591 | if (q->elevator) |
079a2e3e | 4592 | blk_mq_tag_update_sched_shared_tags(q); |
8fa04464 | 4593 | else |
079a2e3e | 4594 | blk_mq_tag_resize_shared_tags(set, nr); |
8fa04464 | 4595 | } |
d97e594c | 4596 | } |
e3a2b3f9 | 4597 | |
24f5a90f | 4598 | blk_mq_unquiesce_queue(q); |
70f36b60 | 4599 | blk_mq_unfreeze_queue(q); |
70f36b60 | 4600 | |
e3a2b3f9 JA |
4601 | return ret; |
4602 | } | |
4603 | ||
d48ece20 JW |
4604 | /* |
4605 | * request_queue and elevator_type pair. | |
4606 | * It is just used by __blk_mq_update_nr_hw_queues to cache | |
4607 | * the elevator_type associated with a request_queue. | |
4608 | */ | |
4609 | struct blk_mq_qe_pair { | |
4610 | struct list_head node; | |
4611 | struct request_queue *q; | |
4612 | struct elevator_type *type; | |
4613 | }; | |
4614 | ||
4615 | /* | |
4616 | * Cache the elevator_type in qe pair list and switch the | |
4617 | * io scheduler to 'none' | |
4618 | */ | |
4619 | static bool blk_mq_elv_switch_none(struct list_head *head, | |
4620 | struct request_queue *q) | |
4621 | { | |
4622 | struct blk_mq_qe_pair *qe; | |
4623 | ||
4624 | if (!q->elevator) | |
4625 | return true; | |
4626 | ||
4627 | qe = kmalloc(sizeof(*qe), GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY); | |
4628 | if (!qe) | |
4629 | return false; | |
4630 | ||
5fd7a84a ML |
4631 | /* q->elevator needs protection from ->sysfs_lock */ |
4632 | mutex_lock(&q->sysfs_lock); | |
4633 | ||
d48ece20 JW |
4634 | INIT_LIST_HEAD(&qe->node); |
4635 | qe->q = q; | |
4636 | qe->type = q->elevator->type; | |
dd6f7f17 CH |
4637 | /* keep a reference to the elevator module as we'll switch back */ |
4638 | __elevator_get(qe->type); | |
d48ece20 | 4639 | list_add(&qe->node, head); |
64b36075 | 4640 | elevator_disable(q); |
d48ece20 JW |
4641 | mutex_unlock(&q->sysfs_lock); |
4642 | ||
4643 | return true; | |
4644 | } | |
4645 | ||
4a3b666e JK |
4646 | static struct blk_mq_qe_pair *blk_lookup_qe_pair(struct list_head *head, |
4647 | struct request_queue *q) | |
d48ece20 JW |
4648 | { |
4649 | struct blk_mq_qe_pair *qe; | |
d48ece20 JW |
4650 | |
4651 | list_for_each_entry(qe, head, node) | |
4a3b666e JK |
4652 | if (qe->q == q) |
4653 | return qe; | |
d48ece20 | 4654 | |
4a3b666e JK |
4655 | return NULL; |
4656 | } | |
d48ece20 | 4657 | |
4a3b666e JK |
4658 | static void blk_mq_elv_switch_back(struct list_head *head, |
4659 | struct request_queue *q) | |
4660 | { | |
4661 | struct blk_mq_qe_pair *qe; | |
4662 | struct elevator_type *t; | |
4663 | ||
4664 | qe = blk_lookup_qe_pair(head, q); | |
4665 | if (!qe) | |
4666 | return; | |
4667 | t = qe->type; | |
d48ece20 JW |
4668 | list_del(&qe->node); |
4669 | kfree(qe); | |
4670 | ||
4671 | mutex_lock(&q->sysfs_lock); | |
8237c01f | 4672 | elevator_switch(q, t); |
8ed40ee3 JC |
4673 | /* drop the reference acquired in blk_mq_elv_switch_none */ |
4674 | elevator_put(t); | |
d48ece20 JW |
4675 | mutex_unlock(&q->sysfs_lock); |
4676 | } | |
4677 | ||
e4dc2b32 KB |
4678 | static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, |
4679 | int nr_hw_queues) | |
868f2f0b KB |
4680 | { |
4681 | struct request_queue *q; | |
d48ece20 | 4682 | LIST_HEAD(head); |
e01ad46d | 4683 | int prev_nr_hw_queues; |
868f2f0b | 4684 | |
705cda97 BVA |
4685 | lockdep_assert_held(&set->tag_list_lock); |
4686 | ||
392546ae | 4687 | if (set->nr_maps == 1 && nr_hw_queues > nr_cpu_ids) |
868f2f0b | 4688 | nr_hw_queues = nr_cpu_ids; |
fe35ec58 WZ |
4689 | if (nr_hw_queues < 1) |
4690 | return; | |
4691 | if (set->nr_maps == 1 && nr_hw_queues == set->nr_hw_queues) | |
868f2f0b KB |
4692 | return; |
4693 | ||
4694 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
4695 | blk_mq_freeze_queue(q); | |
d48ece20 JW |
4696 | /* |
4697 | * Switch IO scheduler to 'none', cleaning up the data associated | |
4698 | * with the previous scheduler. We will switch back once we are done | |
4699 | * updating the new sw to hw queue mappings. | |
4700 | */ | |
4701 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
4702 | if (!blk_mq_elv_switch_none(&head, q)) | |
4703 | goto switch_back; | |
868f2f0b | 4704 | |
477e19de JW |
4705 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
4706 | blk_mq_debugfs_unregister_hctxs(q); | |
eaa870f9 | 4707 | blk_mq_sysfs_unregister_hctxs(q); |
477e19de JW |
4708 | } |
4709 | ||
a2584e43 | 4710 | prev_nr_hw_queues = set->nr_hw_queues; |
ee9d5521 | 4711 | if (blk_mq_realloc_tag_set_tags(set, nr_hw_queues) < 0) |
f7e76dbc BVA |
4712 | goto reregister; |
4713 | ||
e01ad46d | 4714 | fallback: |
aa880ad6 | 4715 | blk_mq_update_queue_map(set); |
868f2f0b KB |
4716 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
4717 | blk_mq_realloc_hw_ctxs(set, q); | |
42ee3061 | 4718 | blk_mq_update_poll_flag(q); |
e01ad46d | 4719 | if (q->nr_hw_queues != set->nr_hw_queues) { |
a846a8e6 YB |
4720 | int i = prev_nr_hw_queues; |
4721 | ||
e01ad46d JW |
4722 | pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n", |
4723 | nr_hw_queues, prev_nr_hw_queues); | |
a846a8e6 YB |
4724 | for (; i < set->nr_hw_queues; i++) |
4725 | __blk_mq_free_map_and_rqs(set, i); | |
4726 | ||
e01ad46d | 4727 | set->nr_hw_queues = prev_nr_hw_queues; |
7d76f856 | 4728 | blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]); |
e01ad46d JW |
4729 | goto fallback; |
4730 | } | |
477e19de JW |
4731 | blk_mq_map_swqueue(q); |
4732 | } | |
4733 | ||
f7e76dbc | 4734 | reregister: |
477e19de | 4735 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
eaa870f9 | 4736 | blk_mq_sysfs_register_hctxs(q); |
477e19de | 4737 | blk_mq_debugfs_register_hctxs(q); |
868f2f0b KB |
4738 | } |
4739 | ||
d48ece20 JW |
4740 | switch_back: |
4741 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
4742 | blk_mq_elv_switch_back(&head, q); | |
4743 | ||
868f2f0b KB |
4744 | list_for_each_entry(q, &set->tag_list, tag_set_list) |
4745 | blk_mq_unfreeze_queue(q); | |
4746 | } | |
e4dc2b32 KB |
4747 | |
4748 | void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues) | |
4749 | { | |
4750 | mutex_lock(&set->tag_list_lock); | |
4751 | __blk_mq_update_nr_hw_queues(set, nr_hw_queues); | |
4752 | mutex_unlock(&set->tag_list_lock); | |
4753 | } | |
868f2f0b KB |
4754 | EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues); |
4755 | ||
34dbad5d OS |
4756 | /* Enable polling stats and return whether they were already enabled. */ |
4757 | static bool blk_poll_stats_enable(struct request_queue *q) | |
4758 | { | |
48b5c1fb | 4759 | if (q->poll_stat) |
34dbad5d | 4760 | return true; |
48b5c1fb JA |
4761 | |
4762 | return blk_stats_alloc_enable(q); | |
34dbad5d OS |
4763 | } |
4764 | ||
4765 | static void blk_mq_poll_stats_start(struct request_queue *q) | |
4766 | { | |
4767 | /* | |
4768 | * We don't arm the callback if polling stats are not enabled or the | |
4769 | * callback is already active. | |
4770 | */ | |
48b5c1fb | 4771 | if (!q->poll_stat || blk_stat_is_active(q->poll_cb)) |
34dbad5d OS |
4772 | return; |
4773 | ||
4774 | blk_stat_activate_msecs(q->poll_cb, 100); | |
4775 | } | |
4776 | ||
4777 | static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb) | |
4778 | { | |
4779 | struct request_queue *q = cb->data; | |
720b8ccc | 4780 | int bucket; |
34dbad5d | 4781 | |
720b8ccc SB |
4782 | for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) { |
4783 | if (cb->stat[bucket].nr_samples) | |
4784 | q->poll_stat[bucket] = cb->stat[bucket]; | |
4785 | } | |
34dbad5d OS |
4786 | } |
4787 | ||
64f1c21e | 4788 | static unsigned long blk_mq_poll_nsecs(struct request_queue *q, |
64f1c21e JA |
4789 | struct request *rq) |
4790 | { | |
64f1c21e | 4791 | unsigned long ret = 0; |
720b8ccc | 4792 | int bucket; |
64f1c21e JA |
4793 | |
4794 | /* | |
4795 | * If stats collection isn't on, don't sleep but turn it on for | |
4796 | * future users | |
4797 | */ | |
34dbad5d | 4798 | if (!blk_poll_stats_enable(q)) |
64f1c21e JA |
4799 | return 0; |
4800 | ||
64f1c21e JA |
4801 | /* |
4802 | * As an optimistic guess, use half of the mean service time | |
4803 | * for this type of request. We can (and should) make this smarter. | |
4804 | * For instance, if the completion latencies are tight, we can | |
4805 | * get closer than just half the mean. This is especially | |
4806 | * important on devices where the completion latencies are longer | |
720b8ccc SB |
4807 | * than ~10 usec. We do use the stats for the relevant IO size |
4808 | * if available which does lead to better estimates. | |
64f1c21e | 4809 | */ |
720b8ccc SB |
4810 | bucket = blk_mq_poll_stats_bkt(rq); |
4811 | if (bucket < 0) | |
4812 | return ret; | |
4813 | ||
4814 | if (q->poll_stat[bucket].nr_samples) | |
4815 | ret = (q->poll_stat[bucket].mean + 1) / 2; | |
64f1c21e JA |
4816 | |
4817 | return ret; | |
4818 | } | |
4819 | ||
c6699d6f | 4820 | static bool blk_mq_poll_hybrid(struct request_queue *q, blk_qc_t qc) |
06426adf | 4821 | { |
c6699d6f CH |
4822 | struct blk_mq_hw_ctx *hctx = blk_qc_to_hctx(q, qc); |
4823 | struct request *rq = blk_qc_to_rq(hctx, qc); | |
06426adf JA |
4824 | struct hrtimer_sleeper hs; |
4825 | enum hrtimer_mode mode; | |
64f1c21e | 4826 | unsigned int nsecs; |
06426adf JA |
4827 | ktime_t kt; |
4828 | ||
c6699d6f CH |
4829 | /* |
4830 | * If a request has completed on queue that uses an I/O scheduler, we | |
4831 | * won't get back a request from blk_qc_to_rq. | |
4832 | */ | |
4833 | if (!rq || (rq->rq_flags & RQF_MQ_POLL_SLEPT)) | |
64f1c21e JA |
4834 | return false; |
4835 | ||
4836 | /* | |
1052b8ac | 4837 | * If we get here, hybrid polling is enabled. Hence poll_nsec can be: |
64f1c21e | 4838 | * |
64f1c21e JA |
4839 | * 0: use half of prev avg |
4840 | * >0: use this specific value | |
4841 | */ | |
1052b8ac | 4842 | if (q->poll_nsec > 0) |
64f1c21e JA |
4843 | nsecs = q->poll_nsec; |
4844 | else | |
cae740a0 | 4845 | nsecs = blk_mq_poll_nsecs(q, rq); |
64f1c21e JA |
4846 | |
4847 | if (!nsecs) | |
06426adf JA |
4848 | return false; |
4849 | ||
76a86f9d | 4850 | rq->rq_flags |= RQF_MQ_POLL_SLEPT; |
06426adf JA |
4851 | |
4852 | /* | |
4853 | * This will be replaced with the stats tracking code, using | |
4854 | * 'avg_completion_time / 2' as the pre-sleep target. | |
4855 | */ | |
8b0e1953 | 4856 | kt = nsecs; |
06426adf JA |
4857 | |
4858 | mode = HRTIMER_MODE_REL; | |
dbc1625f | 4859 | hrtimer_init_sleeper_on_stack(&hs, CLOCK_MONOTONIC, mode); |
06426adf JA |
4860 | hrtimer_set_expires(&hs.timer, kt); |
4861 | ||
06426adf | 4862 | do { |
5a61c363 | 4863 | if (blk_mq_rq_state(rq) == MQ_RQ_COMPLETE) |
06426adf JA |
4864 | break; |
4865 | set_current_state(TASK_UNINTERRUPTIBLE); | |
9dd8813e | 4866 | hrtimer_sleeper_start_expires(&hs, mode); |
06426adf JA |
4867 | if (hs.task) |
4868 | io_schedule(); | |
4869 | hrtimer_cancel(&hs.timer); | |
4870 | mode = HRTIMER_MODE_ABS; | |
4871 | } while (hs.task && !signal_pending(current)); | |
4872 | ||
4873 | __set_current_state(TASK_RUNNING); | |
4874 | destroy_hrtimer_on_stack(&hs.timer); | |
1052b8ac | 4875 | |
06426adf | 4876 | /* |
c6699d6f CH |
4877 | * If we sleep, have the caller restart the poll loop to reset the |
4878 | * state. Like for the other success return cases, the caller is | |
4879 | * responsible for checking if the IO completed. If the IO isn't | |
4880 | * complete, we'll get called again and will go straight to the busy | |
4881 | * poll loop. | |
06426adf | 4882 | */ |
06426adf JA |
4883 | return true; |
4884 | } | |
06426adf | 4885 | |
c6699d6f | 4886 | static int blk_mq_poll_classic(struct request_queue *q, blk_qc_t cookie, |
5a72e899 | 4887 | struct io_comp_batch *iob, unsigned int flags) |
bbd7bb70 | 4888 | { |
c6699d6f CH |
4889 | struct blk_mq_hw_ctx *hctx = blk_qc_to_hctx(q, cookie); |
4890 | long state = get_current_state(); | |
4891 | int ret; | |
bbd7bb70 | 4892 | |
aa61bec3 | 4893 | do { |
5a72e899 | 4894 | ret = q->mq_ops->poll(hctx, iob); |
bbd7bb70 | 4895 | if (ret > 0) { |
849a3700 | 4896 | __set_current_state(TASK_RUNNING); |
85f4d4b6 | 4897 | return ret; |
bbd7bb70 JA |
4898 | } |
4899 | ||
4900 | if (signal_pending_state(state, current)) | |
849a3700 | 4901 | __set_current_state(TASK_RUNNING); |
b03fbd4f | 4902 | if (task_is_running(current)) |
85f4d4b6 | 4903 | return 1; |
c6699d6f | 4904 | |
ef99b2d3 | 4905 | if (ret < 0 || (flags & BLK_POLL_ONESHOT)) |
bbd7bb70 JA |
4906 | break; |
4907 | cpu_relax(); | |
aa61bec3 | 4908 | } while (!need_resched()); |
bbd7bb70 | 4909 | |
67b4110f | 4910 | __set_current_state(TASK_RUNNING); |
85f4d4b6 | 4911 | return 0; |
bbd7bb70 | 4912 | } |
1052b8ac | 4913 | |
5a72e899 JA |
4914 | int blk_mq_poll(struct request_queue *q, blk_qc_t cookie, struct io_comp_batch *iob, |
4915 | unsigned int flags) | |
1052b8ac | 4916 | { |
d729cf9a CH |
4917 | if (!(flags & BLK_POLL_NOSLEEP) && |
4918 | q->poll_nsec != BLK_MQ_POLL_CLASSIC) { | |
c6699d6f | 4919 | if (blk_mq_poll_hybrid(q, cookie)) |
85f4d4b6 | 4920 | return 1; |
c6699d6f | 4921 | } |
5a72e899 | 4922 | return blk_mq_poll_classic(q, cookie, iob, flags); |
bbd7bb70 JA |
4923 | } |
4924 | ||
9cf2bab6 JA |
4925 | unsigned int blk_mq_rq_cpu(struct request *rq) |
4926 | { | |
4927 | return rq->mq_ctx->cpu; | |
4928 | } | |
4929 | EXPORT_SYMBOL(blk_mq_rq_cpu); | |
4930 | ||
2a19b28f ML |
4931 | void blk_mq_cancel_work_sync(struct request_queue *q) |
4932 | { | |
219cf43c JC |
4933 | struct blk_mq_hw_ctx *hctx; |
4934 | unsigned long i; | |
2a19b28f | 4935 | |
219cf43c | 4936 | cancel_delayed_work_sync(&q->requeue_work); |
2a19b28f | 4937 | |
219cf43c JC |
4938 | queue_for_each_hw_ctx(q, hctx, i) |
4939 | cancel_delayed_work_sync(&hctx->run_work); | |
2a19b28f ML |
4940 | } |
4941 | ||
320ae51f JA |
4942 | static int __init blk_mq_init(void) |
4943 | { | |
c3077b5d CH |
4944 | int i; |
4945 | ||
4946 | for_each_possible_cpu(i) | |
f9ab4918 | 4947 | init_llist_head(&per_cpu(blk_cpu_done, i)); |
c3077b5d CH |
4948 | open_softirq(BLOCK_SOFTIRQ, blk_done_softirq); |
4949 | ||
4950 | cpuhp_setup_state_nocalls(CPUHP_BLOCK_SOFTIRQ_DEAD, | |
4951 | "block/softirq:dead", NULL, | |
4952 | blk_softirq_cpu_dead); | |
9467f859 TG |
4953 | cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL, |
4954 | blk_mq_hctx_notify_dead); | |
bf0beec0 ML |
4955 | cpuhp_setup_state_multi(CPUHP_AP_BLK_MQ_ONLINE, "block/mq:online", |
4956 | blk_mq_hctx_notify_online, | |
4957 | blk_mq_hctx_notify_offline); | |
320ae51f JA |
4958 | return 0; |
4959 | } | |
4960 | subsys_initcall(blk_mq_init); |