]>
Commit | Line | Data |
---|---|---|
1 | /* | |
2 | * Block multiqueue core code | |
3 | * | |
4 | * Copyright (C) 2013-2014 Jens Axboe | |
5 | * Copyright (C) 2013-2014 Christoph Hellwig | |
6 | */ | |
7 | #include <linux/kernel.h> | |
8 | #include <linux/module.h> | |
9 | #include <linux/backing-dev.h> | |
10 | #include <linux/bio.h> | |
11 | #include <linux/blkdev.h> | |
12 | #include <linux/kmemleak.h> | |
13 | #include <linux/mm.h> | |
14 | #include <linux/init.h> | |
15 | #include <linux/slab.h> | |
16 | #include <linux/workqueue.h> | |
17 | #include <linux/smp.h> | |
18 | #include <linux/llist.h> | |
19 | #include <linux/list_sort.h> | |
20 | #include <linux/cpu.h> | |
21 | #include <linux/cache.h> | |
22 | #include <linux/sched/sysctl.h> | |
23 | #include <linux/sched/topology.h> | |
24 | #include <linux/sched/signal.h> | |
25 | #include <linux/delay.h> | |
26 | #include <linux/crash_dump.h> | |
27 | #include <linux/prefetch.h> | |
28 | ||
29 | #include <trace/events/block.h> | |
30 | ||
31 | #include <linux/blk-mq.h> | |
32 | #include "blk.h" | |
33 | #include "blk-mq.h" | |
34 | #include "blk-mq-debugfs.h" | |
35 | #include "blk-mq-tag.h" | |
36 | #include "blk-stat.h" | |
37 | #include "blk-wbt.h" | |
38 | #include "blk-mq-sched.h" | |
39 | ||
40 | static bool blk_mq_poll(struct request_queue *q, blk_qc_t cookie); | |
41 | static void blk_mq_poll_stats_start(struct request_queue *q); | |
42 | static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb); | |
43 | ||
44 | static int blk_mq_poll_stats_bkt(const struct request *rq) | |
45 | { | |
46 | int ddir, bytes, bucket; | |
47 | ||
48 | ddir = rq_data_dir(rq); | |
49 | bytes = blk_rq_bytes(rq); | |
50 | ||
51 | bucket = ddir + 2*(ilog2(bytes) - 9); | |
52 | ||
53 | if (bucket < 0) | |
54 | return -1; | |
55 | else if (bucket >= BLK_MQ_POLL_STATS_BKTS) | |
56 | return ddir + BLK_MQ_POLL_STATS_BKTS - 2; | |
57 | ||
58 | return bucket; | |
59 | } | |
60 | ||
61 | /* | |
62 | * Check if any of the ctx's have pending work in this hardware queue | |
63 | */ | |
64 | static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx) | |
65 | { | |
66 | return !list_empty_careful(&hctx->dispatch) || | |
67 | sbitmap_any_bit_set(&hctx->ctx_map) || | |
68 | blk_mq_sched_has_work(hctx); | |
69 | } | |
70 | ||
71 | /* | |
72 | * Mark this ctx as having pending work in this hardware queue | |
73 | */ | |
74 | static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx, | |
75 | struct blk_mq_ctx *ctx) | |
76 | { | |
77 | if (!sbitmap_test_bit(&hctx->ctx_map, ctx->index_hw)) | |
78 | sbitmap_set_bit(&hctx->ctx_map, ctx->index_hw); | |
79 | } | |
80 | ||
81 | static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx, | |
82 | struct blk_mq_ctx *ctx) | |
83 | { | |
84 | sbitmap_clear_bit(&hctx->ctx_map, ctx->index_hw); | |
85 | } | |
86 | ||
87 | struct mq_inflight { | |
88 | struct hd_struct *part; | |
89 | unsigned int *inflight; | |
90 | }; | |
91 | ||
92 | static void blk_mq_check_inflight(struct blk_mq_hw_ctx *hctx, | |
93 | struct request *rq, void *priv, | |
94 | bool reserved) | |
95 | { | |
96 | struct mq_inflight *mi = priv; | |
97 | ||
98 | if (test_bit(REQ_ATOM_STARTED, &rq->atomic_flags) && | |
99 | !test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags)) { | |
100 | /* | |
101 | * index[0] counts the specific partition that was asked | |
102 | * for. index[1] counts the ones that are active on the | |
103 | * whole device, so increment that if mi->part is indeed | |
104 | * a partition, and not a whole device. | |
105 | */ | |
106 | if (rq->part == mi->part) | |
107 | mi->inflight[0]++; | |
108 | if (mi->part->partno) | |
109 | mi->inflight[1]++; | |
110 | } | |
111 | } | |
112 | ||
113 | void blk_mq_in_flight(struct request_queue *q, struct hd_struct *part, | |
114 | unsigned int inflight[2]) | |
115 | { | |
116 | struct mq_inflight mi = { .part = part, .inflight = inflight, }; | |
117 | ||
118 | inflight[0] = inflight[1] = 0; | |
119 | blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi); | |
120 | } | |
121 | ||
122 | void blk_freeze_queue_start(struct request_queue *q) | |
123 | { | |
124 | int freeze_depth; | |
125 | ||
126 | freeze_depth = atomic_inc_return(&q->mq_freeze_depth); | |
127 | if (freeze_depth == 1) { | |
128 | percpu_ref_kill(&q->q_usage_counter); | |
129 | if (q->mq_ops) | |
130 | blk_mq_run_hw_queues(q, false); | |
131 | } | |
132 | } | |
133 | EXPORT_SYMBOL_GPL(blk_freeze_queue_start); | |
134 | ||
135 | void blk_mq_freeze_queue_wait(struct request_queue *q) | |
136 | { | |
137 | wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter)); | |
138 | } | |
139 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait); | |
140 | ||
141 | int blk_mq_freeze_queue_wait_timeout(struct request_queue *q, | |
142 | unsigned long timeout) | |
143 | { | |
144 | return wait_event_timeout(q->mq_freeze_wq, | |
145 | percpu_ref_is_zero(&q->q_usage_counter), | |
146 | timeout); | |
147 | } | |
148 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout); | |
149 | ||
150 | /* | |
151 | * Guarantee no request is in use, so we can change any data structure of | |
152 | * the queue afterward. | |
153 | */ | |
154 | void blk_freeze_queue(struct request_queue *q) | |
155 | { | |
156 | /* | |
157 | * In the !blk_mq case we are only calling this to kill the | |
158 | * q_usage_counter, otherwise this increases the freeze depth | |
159 | * and waits for it to return to zero. For this reason there is | |
160 | * no blk_unfreeze_queue(), and blk_freeze_queue() is not | |
161 | * exported to drivers as the only user for unfreeze is blk_mq. | |
162 | */ | |
163 | blk_freeze_queue_start(q); | |
164 | if (!q->mq_ops) | |
165 | blk_drain_queue(q); | |
166 | blk_mq_freeze_queue_wait(q); | |
167 | } | |
168 | ||
169 | void blk_mq_freeze_queue(struct request_queue *q) | |
170 | { | |
171 | /* | |
172 | * ...just an alias to keep freeze and unfreeze actions balanced | |
173 | * in the blk_mq_* namespace | |
174 | */ | |
175 | blk_freeze_queue(q); | |
176 | } | |
177 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue); | |
178 | ||
179 | void blk_mq_unfreeze_queue(struct request_queue *q) | |
180 | { | |
181 | int freeze_depth; | |
182 | ||
183 | freeze_depth = atomic_dec_return(&q->mq_freeze_depth); | |
184 | WARN_ON_ONCE(freeze_depth < 0); | |
185 | if (!freeze_depth) { | |
186 | percpu_ref_reinit(&q->q_usage_counter); | |
187 | wake_up_all(&q->mq_freeze_wq); | |
188 | } | |
189 | } | |
190 | EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue); | |
191 | ||
192 | /* | |
193 | * FIXME: replace the scsi_internal_device_*block_nowait() calls in the | |
194 | * mpt3sas driver such that this function can be removed. | |
195 | */ | |
196 | void blk_mq_quiesce_queue_nowait(struct request_queue *q) | |
197 | { | |
198 | unsigned long flags; | |
199 | ||
200 | spin_lock_irqsave(q->queue_lock, flags); | |
201 | queue_flag_set(QUEUE_FLAG_QUIESCED, q); | |
202 | spin_unlock_irqrestore(q->queue_lock, flags); | |
203 | } | |
204 | EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait); | |
205 | ||
206 | /** | |
207 | * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished | |
208 | * @q: request queue. | |
209 | * | |
210 | * Note: this function does not prevent that the struct request end_io() | |
211 | * callback function is invoked. Once this function is returned, we make | |
212 | * sure no dispatch can happen until the queue is unquiesced via | |
213 | * blk_mq_unquiesce_queue(). | |
214 | */ | |
215 | void blk_mq_quiesce_queue(struct request_queue *q) | |
216 | { | |
217 | struct blk_mq_hw_ctx *hctx; | |
218 | unsigned int i; | |
219 | bool rcu = false; | |
220 | ||
221 | blk_mq_quiesce_queue_nowait(q); | |
222 | ||
223 | queue_for_each_hw_ctx(q, hctx, i) { | |
224 | if (hctx->flags & BLK_MQ_F_BLOCKING) | |
225 | synchronize_srcu(hctx->queue_rq_srcu); | |
226 | else | |
227 | rcu = true; | |
228 | } | |
229 | if (rcu) | |
230 | synchronize_rcu(); | |
231 | } | |
232 | EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue); | |
233 | ||
234 | /* | |
235 | * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue() | |
236 | * @q: request queue. | |
237 | * | |
238 | * This function recovers queue into the state before quiescing | |
239 | * which is done by blk_mq_quiesce_queue. | |
240 | */ | |
241 | void blk_mq_unquiesce_queue(struct request_queue *q) | |
242 | { | |
243 | unsigned long flags; | |
244 | ||
245 | spin_lock_irqsave(q->queue_lock, flags); | |
246 | queue_flag_clear(QUEUE_FLAG_QUIESCED, q); | |
247 | spin_unlock_irqrestore(q->queue_lock, flags); | |
248 | ||
249 | /* dispatch requests which are inserted during quiescing */ | |
250 | blk_mq_run_hw_queues(q, true); | |
251 | } | |
252 | EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue); | |
253 | ||
254 | void blk_mq_wake_waiters(struct request_queue *q) | |
255 | { | |
256 | struct blk_mq_hw_ctx *hctx; | |
257 | unsigned int i; | |
258 | ||
259 | queue_for_each_hw_ctx(q, hctx, i) | |
260 | if (blk_mq_hw_queue_mapped(hctx)) | |
261 | blk_mq_tag_wakeup_all(hctx->tags, true); | |
262 | } | |
263 | ||
264 | bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx) | |
265 | { | |
266 | return blk_mq_has_free_tags(hctx->tags); | |
267 | } | |
268 | EXPORT_SYMBOL(blk_mq_can_queue); | |
269 | ||
270 | static struct request *blk_mq_rq_ctx_init(struct blk_mq_alloc_data *data, | |
271 | unsigned int tag, unsigned int op) | |
272 | { | |
273 | struct blk_mq_tags *tags = blk_mq_tags_from_data(data); | |
274 | struct request *rq = tags->static_rqs[tag]; | |
275 | ||
276 | rq->rq_flags = 0; | |
277 | ||
278 | if (data->flags & BLK_MQ_REQ_INTERNAL) { | |
279 | rq->tag = -1; | |
280 | rq->internal_tag = tag; | |
281 | } else { | |
282 | if (blk_mq_tag_busy(data->hctx)) { | |
283 | rq->rq_flags = RQF_MQ_INFLIGHT; | |
284 | atomic_inc(&data->hctx->nr_active); | |
285 | } | |
286 | rq->tag = tag; | |
287 | rq->internal_tag = -1; | |
288 | data->hctx->tags->rqs[rq->tag] = rq; | |
289 | } | |
290 | ||
291 | INIT_LIST_HEAD(&rq->queuelist); | |
292 | /* csd/requeue_work/fifo_time is initialized before use */ | |
293 | rq->q = data->q; | |
294 | rq->mq_ctx = data->ctx; | |
295 | rq->cmd_flags = op; | |
296 | if (data->flags & BLK_MQ_REQ_PREEMPT) | |
297 | rq->rq_flags |= RQF_PREEMPT; | |
298 | if (blk_queue_io_stat(data->q)) | |
299 | rq->rq_flags |= RQF_IO_STAT; | |
300 | /* do not touch atomic flags, it needs atomic ops against the timer */ | |
301 | rq->cpu = -1; | |
302 | INIT_HLIST_NODE(&rq->hash); | |
303 | RB_CLEAR_NODE(&rq->rb_node); | |
304 | rq->rq_disk = NULL; | |
305 | rq->part = NULL; | |
306 | rq->start_time = jiffies; | |
307 | #ifdef CONFIG_BLK_CGROUP | |
308 | rq->rl = NULL; | |
309 | set_start_time_ns(rq); | |
310 | rq->io_start_time_ns = 0; | |
311 | #endif | |
312 | rq->nr_phys_segments = 0; | |
313 | #if defined(CONFIG_BLK_DEV_INTEGRITY) | |
314 | rq->nr_integrity_segments = 0; | |
315 | #endif | |
316 | rq->special = NULL; | |
317 | /* tag was already set */ | |
318 | rq->extra_len = 0; | |
319 | ||
320 | INIT_LIST_HEAD(&rq->timeout_list); | |
321 | rq->timeout = 0; | |
322 | ||
323 | rq->end_io = NULL; | |
324 | rq->end_io_data = NULL; | |
325 | rq->next_rq = NULL; | |
326 | ||
327 | data->ctx->rq_dispatched[op_is_sync(op)]++; | |
328 | return rq; | |
329 | } | |
330 | ||
331 | static struct request *blk_mq_get_request(struct request_queue *q, | |
332 | struct bio *bio, unsigned int op, | |
333 | struct blk_mq_alloc_data *data) | |
334 | { | |
335 | struct elevator_queue *e = q->elevator; | |
336 | struct request *rq; | |
337 | unsigned int tag; | |
338 | bool put_ctx_on_error = false; | |
339 | ||
340 | blk_queue_enter_live(q); | |
341 | data->q = q; | |
342 | if (likely(!data->ctx)) { | |
343 | data->ctx = blk_mq_get_ctx(q); | |
344 | put_ctx_on_error = true; | |
345 | } | |
346 | if (likely(!data->hctx)) | |
347 | data->hctx = blk_mq_map_queue(q, data->ctx->cpu); | |
348 | if (op & REQ_NOWAIT) | |
349 | data->flags |= BLK_MQ_REQ_NOWAIT; | |
350 | ||
351 | if (e) { | |
352 | data->flags |= BLK_MQ_REQ_INTERNAL; | |
353 | ||
354 | /* | |
355 | * Flush requests are special and go directly to the | |
356 | * dispatch list. | |
357 | */ | |
358 | if (!op_is_flush(op) && e->type->ops.mq.limit_depth) | |
359 | e->type->ops.mq.limit_depth(op, data); | |
360 | } | |
361 | ||
362 | tag = blk_mq_get_tag(data); | |
363 | if (tag == BLK_MQ_TAG_FAIL) { | |
364 | if (put_ctx_on_error) { | |
365 | blk_mq_put_ctx(data->ctx); | |
366 | data->ctx = NULL; | |
367 | } | |
368 | blk_queue_exit(q); | |
369 | return NULL; | |
370 | } | |
371 | ||
372 | rq = blk_mq_rq_ctx_init(data, tag, op); | |
373 | if (!op_is_flush(op)) { | |
374 | rq->elv.icq = NULL; | |
375 | if (e && e->type->ops.mq.prepare_request) { | |
376 | if (e->type->icq_cache && rq_ioc(bio)) | |
377 | blk_mq_sched_assign_ioc(rq, bio); | |
378 | ||
379 | e->type->ops.mq.prepare_request(rq, bio); | |
380 | rq->rq_flags |= RQF_ELVPRIV; | |
381 | } | |
382 | } | |
383 | data->hctx->queued++; | |
384 | return rq; | |
385 | } | |
386 | ||
387 | struct request *blk_mq_alloc_request(struct request_queue *q, unsigned int op, | |
388 | blk_mq_req_flags_t flags) | |
389 | { | |
390 | struct blk_mq_alloc_data alloc_data = { .flags = flags }; | |
391 | struct request *rq; | |
392 | int ret; | |
393 | ||
394 | ret = blk_queue_enter(q, flags); | |
395 | if (ret) | |
396 | return ERR_PTR(ret); | |
397 | ||
398 | rq = blk_mq_get_request(q, NULL, op, &alloc_data); | |
399 | blk_queue_exit(q); | |
400 | ||
401 | if (!rq) | |
402 | return ERR_PTR(-EWOULDBLOCK); | |
403 | ||
404 | blk_mq_put_ctx(alloc_data.ctx); | |
405 | ||
406 | rq->__data_len = 0; | |
407 | rq->__sector = (sector_t) -1; | |
408 | rq->bio = rq->biotail = NULL; | |
409 | return rq; | |
410 | } | |
411 | EXPORT_SYMBOL(blk_mq_alloc_request); | |
412 | ||
413 | struct request *blk_mq_alloc_request_hctx(struct request_queue *q, | |
414 | unsigned int op, blk_mq_req_flags_t flags, unsigned int hctx_idx) | |
415 | { | |
416 | struct blk_mq_alloc_data alloc_data = { .flags = flags }; | |
417 | struct request *rq; | |
418 | unsigned int cpu; | |
419 | int ret; | |
420 | ||
421 | /* | |
422 | * If the tag allocator sleeps we could get an allocation for a | |
423 | * different hardware context. No need to complicate the low level | |
424 | * allocator for this for the rare use case of a command tied to | |
425 | * a specific queue. | |
426 | */ | |
427 | if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT))) | |
428 | return ERR_PTR(-EINVAL); | |
429 | ||
430 | if (hctx_idx >= q->nr_hw_queues) | |
431 | return ERR_PTR(-EIO); | |
432 | ||
433 | ret = blk_queue_enter(q, flags); | |
434 | if (ret) | |
435 | return ERR_PTR(ret); | |
436 | ||
437 | /* | |
438 | * Check if the hardware context is actually mapped to anything. | |
439 | * If not tell the caller that it should skip this queue. | |
440 | */ | |
441 | alloc_data.hctx = q->queue_hw_ctx[hctx_idx]; | |
442 | if (!blk_mq_hw_queue_mapped(alloc_data.hctx)) { | |
443 | blk_queue_exit(q); | |
444 | return ERR_PTR(-EXDEV); | |
445 | } | |
446 | cpu = cpumask_first_and(alloc_data.hctx->cpumask, cpu_online_mask); | |
447 | alloc_data.ctx = __blk_mq_get_ctx(q, cpu); | |
448 | ||
449 | rq = blk_mq_get_request(q, NULL, op, &alloc_data); | |
450 | blk_queue_exit(q); | |
451 | ||
452 | if (!rq) | |
453 | return ERR_PTR(-EWOULDBLOCK); | |
454 | ||
455 | return rq; | |
456 | } | |
457 | EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx); | |
458 | ||
459 | void blk_mq_free_request(struct request *rq) | |
460 | { | |
461 | struct request_queue *q = rq->q; | |
462 | struct elevator_queue *e = q->elevator; | |
463 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
464 | struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, ctx->cpu); | |
465 | const int sched_tag = rq->internal_tag; | |
466 | ||
467 | if (rq->rq_flags & RQF_ELVPRIV) { | |
468 | if (e && e->type->ops.mq.finish_request) | |
469 | e->type->ops.mq.finish_request(rq); | |
470 | if (rq->elv.icq) { | |
471 | put_io_context(rq->elv.icq->ioc); | |
472 | rq->elv.icq = NULL; | |
473 | } | |
474 | } | |
475 | ||
476 | ctx->rq_completed[rq_is_sync(rq)]++; | |
477 | if (rq->rq_flags & RQF_MQ_INFLIGHT) | |
478 | atomic_dec(&hctx->nr_active); | |
479 | ||
480 | if (unlikely(laptop_mode && !blk_rq_is_passthrough(rq))) | |
481 | laptop_io_completion(q->backing_dev_info); | |
482 | ||
483 | wbt_done(q->rq_wb, &rq->issue_stat); | |
484 | ||
485 | if (blk_rq_rl(rq)) | |
486 | blk_put_rl(blk_rq_rl(rq)); | |
487 | ||
488 | clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags); | |
489 | clear_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags); | |
490 | if (rq->tag != -1) | |
491 | blk_mq_put_tag(hctx, hctx->tags, ctx, rq->tag); | |
492 | if (sched_tag != -1) | |
493 | blk_mq_put_tag(hctx, hctx->sched_tags, ctx, sched_tag); | |
494 | blk_mq_sched_restart(hctx); | |
495 | blk_queue_exit(q); | |
496 | } | |
497 | EXPORT_SYMBOL_GPL(blk_mq_free_request); | |
498 | ||
499 | inline void __blk_mq_end_request(struct request *rq, blk_status_t error) | |
500 | { | |
501 | blk_account_io_done(rq); | |
502 | ||
503 | if (rq->end_io) { | |
504 | wbt_done(rq->q->rq_wb, &rq->issue_stat); | |
505 | rq->end_io(rq, error); | |
506 | } else { | |
507 | if (unlikely(blk_bidi_rq(rq))) | |
508 | blk_mq_free_request(rq->next_rq); | |
509 | blk_mq_free_request(rq); | |
510 | } | |
511 | } | |
512 | EXPORT_SYMBOL(__blk_mq_end_request); | |
513 | ||
514 | void blk_mq_end_request(struct request *rq, blk_status_t error) | |
515 | { | |
516 | if (blk_update_request(rq, error, blk_rq_bytes(rq))) | |
517 | BUG(); | |
518 | __blk_mq_end_request(rq, error); | |
519 | } | |
520 | EXPORT_SYMBOL(blk_mq_end_request); | |
521 | ||
522 | static void __blk_mq_complete_request_remote(void *data) | |
523 | { | |
524 | struct request *rq = data; | |
525 | ||
526 | rq->q->softirq_done_fn(rq); | |
527 | } | |
528 | ||
529 | static void __blk_mq_complete_request(struct request *rq) | |
530 | { | |
531 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
532 | bool shared = false; | |
533 | int cpu; | |
534 | ||
535 | if (rq->internal_tag != -1) | |
536 | blk_mq_sched_completed_request(rq); | |
537 | if (rq->rq_flags & RQF_STATS) { | |
538 | blk_mq_poll_stats_start(rq->q); | |
539 | blk_stat_add(rq); | |
540 | } | |
541 | ||
542 | if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) { | |
543 | rq->q->softirq_done_fn(rq); | |
544 | return; | |
545 | } | |
546 | ||
547 | cpu = get_cpu(); | |
548 | if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags)) | |
549 | shared = cpus_share_cache(cpu, ctx->cpu); | |
550 | ||
551 | if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) { | |
552 | rq->csd.func = __blk_mq_complete_request_remote; | |
553 | rq->csd.info = rq; | |
554 | rq->csd.flags = 0; | |
555 | smp_call_function_single_async(ctx->cpu, &rq->csd); | |
556 | } else { | |
557 | rq->q->softirq_done_fn(rq); | |
558 | } | |
559 | put_cpu(); | |
560 | } | |
561 | ||
562 | /** | |
563 | * blk_mq_complete_request - end I/O on a request | |
564 | * @rq: the request being processed | |
565 | * | |
566 | * Description: | |
567 | * Ends all I/O on a request. It does not handle partial completions. | |
568 | * The actual completion happens out-of-order, through a IPI handler. | |
569 | **/ | |
570 | void blk_mq_complete_request(struct request *rq) | |
571 | { | |
572 | struct request_queue *q = rq->q; | |
573 | ||
574 | if (unlikely(blk_should_fake_timeout(q))) | |
575 | return; | |
576 | if (!blk_mark_rq_complete(rq)) | |
577 | __blk_mq_complete_request(rq); | |
578 | } | |
579 | EXPORT_SYMBOL(blk_mq_complete_request); | |
580 | ||
581 | int blk_mq_request_started(struct request *rq) | |
582 | { | |
583 | return test_bit(REQ_ATOM_STARTED, &rq->atomic_flags); | |
584 | } | |
585 | EXPORT_SYMBOL_GPL(blk_mq_request_started); | |
586 | ||
587 | void blk_mq_start_request(struct request *rq) | |
588 | { | |
589 | struct request_queue *q = rq->q; | |
590 | ||
591 | blk_mq_sched_started_request(rq); | |
592 | ||
593 | trace_block_rq_issue(q, rq); | |
594 | ||
595 | if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) { | |
596 | blk_stat_set_issue(&rq->issue_stat, blk_rq_sectors(rq)); | |
597 | rq->rq_flags |= RQF_STATS; | |
598 | wbt_issue(q->rq_wb, &rq->issue_stat); | |
599 | } | |
600 | ||
601 | blk_add_timer(rq); | |
602 | ||
603 | WARN_ON_ONCE(test_bit(REQ_ATOM_STARTED, &rq->atomic_flags)); | |
604 | ||
605 | /* | |
606 | * Mark us as started and clear complete. Complete might have been | |
607 | * set if requeue raced with timeout, which then marked it as | |
608 | * complete. So be sure to clear complete again when we start | |
609 | * the request, otherwise we'll ignore the completion event. | |
610 | * | |
611 | * Ensure that ->deadline is visible before we set STARTED, such that | |
612 | * blk_mq_check_expired() is guaranteed to observe our ->deadline when | |
613 | * it observes STARTED. | |
614 | */ | |
615 | smp_wmb(); | |
616 | set_bit(REQ_ATOM_STARTED, &rq->atomic_flags); | |
617 | if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags)) { | |
618 | /* | |
619 | * Coherence order guarantees these consecutive stores to a | |
620 | * single variable propagate in the specified order. Thus the | |
621 | * clear_bit() is ordered _after_ the set bit. See | |
622 | * blk_mq_check_expired(). | |
623 | * | |
624 | * (the bits must be part of the same byte for this to be | |
625 | * true). | |
626 | */ | |
627 | clear_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags); | |
628 | } | |
629 | ||
630 | if (q->dma_drain_size && blk_rq_bytes(rq)) { | |
631 | /* | |
632 | * Make sure space for the drain appears. We know we can do | |
633 | * this because max_hw_segments has been adjusted to be one | |
634 | * fewer than the device can handle. | |
635 | */ | |
636 | rq->nr_phys_segments++; | |
637 | } | |
638 | } | |
639 | EXPORT_SYMBOL(blk_mq_start_request); | |
640 | ||
641 | /* | |
642 | * When we reach here because queue is busy, REQ_ATOM_COMPLETE | |
643 | * flag isn't set yet, so there may be race with timeout handler, | |
644 | * but given rq->deadline is just set in .queue_rq() under | |
645 | * this situation, the race won't be possible in reality because | |
646 | * rq->timeout should be set as big enough to cover the window | |
647 | * between blk_mq_start_request() called from .queue_rq() and | |
648 | * clearing REQ_ATOM_STARTED here. | |
649 | */ | |
650 | static void __blk_mq_requeue_request(struct request *rq) | |
651 | { | |
652 | struct request_queue *q = rq->q; | |
653 | ||
654 | blk_mq_put_driver_tag(rq); | |
655 | ||
656 | trace_block_rq_requeue(q, rq); | |
657 | wbt_requeue(q->rq_wb, &rq->issue_stat); | |
658 | ||
659 | if (test_and_clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) { | |
660 | if (q->dma_drain_size && blk_rq_bytes(rq)) | |
661 | rq->nr_phys_segments--; | |
662 | } | |
663 | } | |
664 | ||
665 | void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list) | |
666 | { | |
667 | __blk_mq_requeue_request(rq); | |
668 | ||
669 | /* this request will be re-inserted to io scheduler queue */ | |
670 | blk_mq_sched_requeue_request(rq); | |
671 | ||
672 | BUG_ON(blk_queued_rq(rq)); | |
673 | blk_mq_add_to_requeue_list(rq, true, kick_requeue_list); | |
674 | } | |
675 | EXPORT_SYMBOL(blk_mq_requeue_request); | |
676 | ||
677 | static void blk_mq_requeue_work(struct work_struct *work) | |
678 | { | |
679 | struct request_queue *q = | |
680 | container_of(work, struct request_queue, requeue_work.work); | |
681 | LIST_HEAD(rq_list); | |
682 | struct request *rq, *next; | |
683 | ||
684 | spin_lock_irq(&q->requeue_lock); | |
685 | list_splice_init(&q->requeue_list, &rq_list); | |
686 | spin_unlock_irq(&q->requeue_lock); | |
687 | ||
688 | list_for_each_entry_safe(rq, next, &rq_list, queuelist) { | |
689 | if (!(rq->rq_flags & RQF_SOFTBARRIER)) | |
690 | continue; | |
691 | ||
692 | rq->rq_flags &= ~RQF_SOFTBARRIER; | |
693 | list_del_init(&rq->queuelist); | |
694 | blk_mq_sched_insert_request(rq, true, false, false, true); | |
695 | } | |
696 | ||
697 | while (!list_empty(&rq_list)) { | |
698 | rq = list_entry(rq_list.next, struct request, queuelist); | |
699 | list_del_init(&rq->queuelist); | |
700 | blk_mq_sched_insert_request(rq, false, false, false, true); | |
701 | } | |
702 | ||
703 | blk_mq_run_hw_queues(q, false); | |
704 | } | |
705 | ||
706 | void blk_mq_add_to_requeue_list(struct request *rq, bool at_head, | |
707 | bool kick_requeue_list) | |
708 | { | |
709 | struct request_queue *q = rq->q; | |
710 | unsigned long flags; | |
711 | ||
712 | /* | |
713 | * We abuse this flag that is otherwise used by the I/O scheduler to | |
714 | * request head insertion from the workqueue. | |
715 | */ | |
716 | BUG_ON(rq->rq_flags & RQF_SOFTBARRIER); | |
717 | ||
718 | spin_lock_irqsave(&q->requeue_lock, flags); | |
719 | if (at_head) { | |
720 | rq->rq_flags |= RQF_SOFTBARRIER; | |
721 | list_add(&rq->queuelist, &q->requeue_list); | |
722 | } else { | |
723 | list_add_tail(&rq->queuelist, &q->requeue_list); | |
724 | } | |
725 | spin_unlock_irqrestore(&q->requeue_lock, flags); | |
726 | ||
727 | if (kick_requeue_list) | |
728 | blk_mq_kick_requeue_list(q); | |
729 | } | |
730 | EXPORT_SYMBOL(blk_mq_add_to_requeue_list); | |
731 | ||
732 | void blk_mq_kick_requeue_list(struct request_queue *q) | |
733 | { | |
734 | kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0); | |
735 | } | |
736 | EXPORT_SYMBOL(blk_mq_kick_requeue_list); | |
737 | ||
738 | void blk_mq_delay_kick_requeue_list(struct request_queue *q, | |
739 | unsigned long msecs) | |
740 | { | |
741 | kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, | |
742 | msecs_to_jiffies(msecs)); | |
743 | } | |
744 | EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list); | |
745 | ||
746 | struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag) | |
747 | { | |
748 | if (tag < tags->nr_tags) { | |
749 | prefetch(tags->rqs[tag]); | |
750 | return tags->rqs[tag]; | |
751 | } | |
752 | ||
753 | return NULL; | |
754 | } | |
755 | EXPORT_SYMBOL(blk_mq_tag_to_rq); | |
756 | ||
757 | struct blk_mq_timeout_data { | |
758 | unsigned long next; | |
759 | unsigned int next_set; | |
760 | }; | |
761 | ||
762 | void blk_mq_rq_timed_out(struct request *req, bool reserved) | |
763 | { | |
764 | const struct blk_mq_ops *ops = req->q->mq_ops; | |
765 | enum blk_eh_timer_return ret = BLK_EH_RESET_TIMER; | |
766 | ||
767 | /* | |
768 | * We know that complete is set at this point. If STARTED isn't set | |
769 | * anymore, then the request isn't active and the "timeout" should | |
770 | * just be ignored. This can happen due to the bitflag ordering. | |
771 | * Timeout first checks if STARTED is set, and if it is, assumes | |
772 | * the request is active. But if we race with completion, then | |
773 | * both flags will get cleared. So check here again, and ignore | |
774 | * a timeout event with a request that isn't active. | |
775 | */ | |
776 | if (!test_bit(REQ_ATOM_STARTED, &req->atomic_flags)) | |
777 | return; | |
778 | ||
779 | if (ops->timeout) | |
780 | ret = ops->timeout(req, reserved); | |
781 | ||
782 | switch (ret) { | |
783 | case BLK_EH_HANDLED: | |
784 | __blk_mq_complete_request(req); | |
785 | break; | |
786 | case BLK_EH_RESET_TIMER: | |
787 | blk_add_timer(req); | |
788 | blk_clear_rq_complete(req); | |
789 | break; | |
790 | case BLK_EH_NOT_HANDLED: | |
791 | break; | |
792 | default: | |
793 | printk(KERN_ERR "block: bad eh return: %d\n", ret); | |
794 | break; | |
795 | } | |
796 | } | |
797 | ||
798 | static void blk_mq_check_expired(struct blk_mq_hw_ctx *hctx, | |
799 | struct request *rq, void *priv, bool reserved) | |
800 | { | |
801 | struct blk_mq_timeout_data *data = priv; | |
802 | unsigned long deadline; | |
803 | ||
804 | if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) | |
805 | return; | |
806 | ||
807 | /* | |
808 | * Ensures that if we see STARTED we must also see our | |
809 | * up-to-date deadline, see blk_mq_start_request(). | |
810 | */ | |
811 | smp_rmb(); | |
812 | ||
813 | deadline = READ_ONCE(rq->deadline); | |
814 | ||
815 | /* | |
816 | * The rq being checked may have been freed and reallocated | |
817 | * out already here, we avoid this race by checking rq->deadline | |
818 | * and REQ_ATOM_COMPLETE flag together: | |
819 | * | |
820 | * - if rq->deadline is observed as new value because of | |
821 | * reusing, the rq won't be timed out because of timing. | |
822 | * - if rq->deadline is observed as previous value, | |
823 | * REQ_ATOM_COMPLETE flag won't be cleared in reuse path | |
824 | * because we put a barrier between setting rq->deadline | |
825 | * and clearing the flag in blk_mq_start_request(), so | |
826 | * this rq won't be timed out too. | |
827 | */ | |
828 | if (time_after_eq(jiffies, deadline)) { | |
829 | if (!blk_mark_rq_complete(rq)) { | |
830 | /* | |
831 | * Again coherence order ensures that consecutive reads | |
832 | * from the same variable must be in that order. This | |
833 | * ensures that if we see COMPLETE clear, we must then | |
834 | * see STARTED set and we'll ignore this timeout. | |
835 | * | |
836 | * (There's also the MB implied by the test_and_clear()) | |
837 | */ | |
838 | blk_mq_rq_timed_out(rq, reserved); | |
839 | } | |
840 | } else if (!data->next_set || time_after(data->next, deadline)) { | |
841 | data->next = deadline; | |
842 | data->next_set = 1; | |
843 | } | |
844 | } | |
845 | ||
846 | static void blk_mq_timeout_work(struct work_struct *work) | |
847 | { | |
848 | struct request_queue *q = | |
849 | container_of(work, struct request_queue, timeout_work); | |
850 | struct blk_mq_timeout_data data = { | |
851 | .next = 0, | |
852 | .next_set = 0, | |
853 | }; | |
854 | int i; | |
855 | ||
856 | /* A deadlock might occur if a request is stuck requiring a | |
857 | * timeout at the same time a queue freeze is waiting | |
858 | * completion, since the timeout code would not be able to | |
859 | * acquire the queue reference here. | |
860 | * | |
861 | * That's why we don't use blk_queue_enter here; instead, we use | |
862 | * percpu_ref_tryget directly, because we need to be able to | |
863 | * obtain a reference even in the short window between the queue | |
864 | * starting to freeze, by dropping the first reference in | |
865 | * blk_freeze_queue_start, and the moment the last request is | |
866 | * consumed, marked by the instant q_usage_counter reaches | |
867 | * zero. | |
868 | */ | |
869 | if (!percpu_ref_tryget(&q->q_usage_counter)) | |
870 | return; | |
871 | ||
872 | blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &data); | |
873 | ||
874 | if (data.next_set) { | |
875 | data.next = blk_rq_timeout(round_jiffies_up(data.next)); | |
876 | mod_timer(&q->timeout, data.next); | |
877 | } else { | |
878 | struct blk_mq_hw_ctx *hctx; | |
879 | ||
880 | queue_for_each_hw_ctx(q, hctx, i) { | |
881 | /* the hctx may be unmapped, so check it here */ | |
882 | if (blk_mq_hw_queue_mapped(hctx)) | |
883 | blk_mq_tag_idle(hctx); | |
884 | } | |
885 | } | |
886 | blk_queue_exit(q); | |
887 | } | |
888 | ||
889 | struct flush_busy_ctx_data { | |
890 | struct blk_mq_hw_ctx *hctx; | |
891 | struct list_head *list; | |
892 | }; | |
893 | ||
894 | static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data) | |
895 | { | |
896 | struct flush_busy_ctx_data *flush_data = data; | |
897 | struct blk_mq_hw_ctx *hctx = flush_data->hctx; | |
898 | struct blk_mq_ctx *ctx = hctx->ctxs[bitnr]; | |
899 | ||
900 | sbitmap_clear_bit(sb, bitnr); | |
901 | spin_lock(&ctx->lock); | |
902 | list_splice_tail_init(&ctx->rq_list, flush_data->list); | |
903 | spin_unlock(&ctx->lock); | |
904 | return true; | |
905 | } | |
906 | ||
907 | /* | |
908 | * Process software queues that have been marked busy, splicing them | |
909 | * to the for-dispatch | |
910 | */ | |
911 | void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list) | |
912 | { | |
913 | struct flush_busy_ctx_data data = { | |
914 | .hctx = hctx, | |
915 | .list = list, | |
916 | }; | |
917 | ||
918 | sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data); | |
919 | } | |
920 | EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs); | |
921 | ||
922 | struct dispatch_rq_data { | |
923 | struct blk_mq_hw_ctx *hctx; | |
924 | struct request *rq; | |
925 | }; | |
926 | ||
927 | static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr, | |
928 | void *data) | |
929 | { | |
930 | struct dispatch_rq_data *dispatch_data = data; | |
931 | struct blk_mq_hw_ctx *hctx = dispatch_data->hctx; | |
932 | struct blk_mq_ctx *ctx = hctx->ctxs[bitnr]; | |
933 | ||
934 | spin_lock(&ctx->lock); | |
935 | if (unlikely(!list_empty(&ctx->rq_list))) { | |
936 | dispatch_data->rq = list_entry_rq(ctx->rq_list.next); | |
937 | list_del_init(&dispatch_data->rq->queuelist); | |
938 | if (list_empty(&ctx->rq_list)) | |
939 | sbitmap_clear_bit(sb, bitnr); | |
940 | } | |
941 | spin_unlock(&ctx->lock); | |
942 | ||
943 | return !dispatch_data->rq; | |
944 | } | |
945 | ||
946 | struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx, | |
947 | struct blk_mq_ctx *start) | |
948 | { | |
949 | unsigned off = start ? start->index_hw : 0; | |
950 | struct dispatch_rq_data data = { | |
951 | .hctx = hctx, | |
952 | .rq = NULL, | |
953 | }; | |
954 | ||
955 | __sbitmap_for_each_set(&hctx->ctx_map, off, | |
956 | dispatch_rq_from_ctx, &data); | |
957 | ||
958 | return data.rq; | |
959 | } | |
960 | ||
961 | static inline unsigned int queued_to_index(unsigned int queued) | |
962 | { | |
963 | if (!queued) | |
964 | return 0; | |
965 | ||
966 | return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1); | |
967 | } | |
968 | ||
969 | bool blk_mq_get_driver_tag(struct request *rq, struct blk_mq_hw_ctx **hctx, | |
970 | bool wait) | |
971 | { | |
972 | struct blk_mq_alloc_data data = { | |
973 | .q = rq->q, | |
974 | .hctx = blk_mq_map_queue(rq->q, rq->mq_ctx->cpu), | |
975 | .flags = wait ? 0 : BLK_MQ_REQ_NOWAIT, | |
976 | }; | |
977 | ||
978 | might_sleep_if(wait); | |
979 | ||
980 | if (rq->tag != -1) | |
981 | goto done; | |
982 | ||
983 | if (blk_mq_tag_is_reserved(data.hctx->sched_tags, rq->internal_tag)) | |
984 | data.flags |= BLK_MQ_REQ_RESERVED; | |
985 | ||
986 | rq->tag = blk_mq_get_tag(&data); | |
987 | if (rq->tag >= 0) { | |
988 | if (blk_mq_tag_busy(data.hctx)) { | |
989 | rq->rq_flags |= RQF_MQ_INFLIGHT; | |
990 | atomic_inc(&data.hctx->nr_active); | |
991 | } | |
992 | data.hctx->tags->rqs[rq->tag] = rq; | |
993 | } | |
994 | ||
995 | done: | |
996 | if (hctx) | |
997 | *hctx = data.hctx; | |
998 | return rq->tag != -1; | |
999 | } | |
1000 | ||
1001 | static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode, | |
1002 | int flags, void *key) | |
1003 | { | |
1004 | struct blk_mq_hw_ctx *hctx; | |
1005 | ||
1006 | hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait); | |
1007 | ||
1008 | list_del_init(&wait->entry); | |
1009 | blk_mq_run_hw_queue(hctx, true); | |
1010 | return 1; | |
1011 | } | |
1012 | ||
1013 | /* | |
1014 | * Mark us waiting for a tag. For shared tags, this involves hooking us into | |
1015 | * the tag wakeups. For non-shared tags, we can simply mark us nedeing a | |
1016 | * restart. For both caes, take care to check the condition again after | |
1017 | * marking us as waiting. | |
1018 | */ | |
1019 | static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx **hctx, | |
1020 | struct request *rq) | |
1021 | { | |
1022 | struct blk_mq_hw_ctx *this_hctx = *hctx; | |
1023 | bool shared_tags = (this_hctx->flags & BLK_MQ_F_TAG_SHARED) != 0; | |
1024 | struct sbq_wait_state *ws; | |
1025 | wait_queue_entry_t *wait; | |
1026 | bool ret; | |
1027 | ||
1028 | if (!shared_tags) { | |
1029 | if (!test_bit(BLK_MQ_S_SCHED_RESTART, &this_hctx->state)) | |
1030 | set_bit(BLK_MQ_S_SCHED_RESTART, &this_hctx->state); | |
1031 | } else { | |
1032 | wait = &this_hctx->dispatch_wait; | |
1033 | if (!list_empty_careful(&wait->entry)) | |
1034 | return false; | |
1035 | ||
1036 | spin_lock(&this_hctx->lock); | |
1037 | if (!list_empty(&wait->entry)) { | |
1038 | spin_unlock(&this_hctx->lock); | |
1039 | return false; | |
1040 | } | |
1041 | ||
1042 | ws = bt_wait_ptr(&this_hctx->tags->bitmap_tags, this_hctx); | |
1043 | add_wait_queue(&ws->wait, wait); | |
1044 | } | |
1045 | ||
1046 | /* | |
1047 | * It's possible that a tag was freed in the window between the | |
1048 | * allocation failure and adding the hardware queue to the wait | |
1049 | * queue. | |
1050 | */ | |
1051 | ret = blk_mq_get_driver_tag(rq, hctx, false); | |
1052 | ||
1053 | if (!shared_tags) { | |
1054 | /* | |
1055 | * Don't clear RESTART here, someone else could have set it. | |
1056 | * At most this will cost an extra queue run. | |
1057 | */ | |
1058 | return ret; | |
1059 | } else { | |
1060 | if (!ret) { | |
1061 | spin_unlock(&this_hctx->lock); | |
1062 | return false; | |
1063 | } | |
1064 | ||
1065 | /* | |
1066 | * We got a tag, remove ourselves from the wait queue to ensure | |
1067 | * someone else gets the wakeup. | |
1068 | */ | |
1069 | spin_lock_irq(&ws->wait.lock); | |
1070 | list_del_init(&wait->entry); | |
1071 | spin_unlock_irq(&ws->wait.lock); | |
1072 | spin_unlock(&this_hctx->lock); | |
1073 | return true; | |
1074 | } | |
1075 | } | |
1076 | ||
1077 | bool blk_mq_dispatch_rq_list(struct request_queue *q, struct list_head *list, | |
1078 | bool got_budget) | |
1079 | { | |
1080 | struct blk_mq_hw_ctx *hctx; | |
1081 | struct request *rq, *nxt; | |
1082 | bool no_tag = false; | |
1083 | int errors, queued; | |
1084 | ||
1085 | if (list_empty(list)) | |
1086 | return false; | |
1087 | ||
1088 | WARN_ON(!list_is_singular(list) && got_budget); | |
1089 | ||
1090 | /* | |
1091 | * Now process all the entries, sending them to the driver. | |
1092 | */ | |
1093 | errors = queued = 0; | |
1094 | do { | |
1095 | struct blk_mq_queue_data bd; | |
1096 | blk_status_t ret; | |
1097 | ||
1098 | rq = list_first_entry(list, struct request, queuelist); | |
1099 | ||
1100 | hctx = blk_mq_map_queue(rq->q, rq->mq_ctx->cpu); | |
1101 | if (!got_budget && !blk_mq_get_dispatch_budget(hctx)) | |
1102 | break; | |
1103 | ||
1104 | if (!blk_mq_get_driver_tag(rq, NULL, false)) { | |
1105 | /* | |
1106 | * The initial allocation attempt failed, so we need to | |
1107 | * rerun the hardware queue when a tag is freed. The | |
1108 | * waitqueue takes care of that. If the queue is run | |
1109 | * before we add this entry back on the dispatch list, | |
1110 | * we'll re-run it below. | |
1111 | */ | |
1112 | if (!blk_mq_mark_tag_wait(&hctx, rq)) { | |
1113 | blk_mq_put_dispatch_budget(hctx); | |
1114 | /* | |
1115 | * For non-shared tags, the RESTART check | |
1116 | * will suffice. | |
1117 | */ | |
1118 | if (hctx->flags & BLK_MQ_F_TAG_SHARED) | |
1119 | no_tag = true; | |
1120 | break; | |
1121 | } | |
1122 | } | |
1123 | ||
1124 | list_del_init(&rq->queuelist); | |
1125 | ||
1126 | bd.rq = rq; | |
1127 | ||
1128 | /* | |
1129 | * Flag last if we have no more requests, or if we have more | |
1130 | * but can't assign a driver tag to it. | |
1131 | */ | |
1132 | if (list_empty(list)) | |
1133 | bd.last = true; | |
1134 | else { | |
1135 | nxt = list_first_entry(list, struct request, queuelist); | |
1136 | bd.last = !blk_mq_get_driver_tag(nxt, NULL, false); | |
1137 | } | |
1138 | ||
1139 | ret = q->mq_ops->queue_rq(hctx, &bd); | |
1140 | if (ret == BLK_STS_RESOURCE) { | |
1141 | /* | |
1142 | * If an I/O scheduler has been configured and we got a | |
1143 | * driver tag for the next request already, free it | |
1144 | * again. | |
1145 | */ | |
1146 | if (!list_empty(list)) { | |
1147 | nxt = list_first_entry(list, struct request, queuelist); | |
1148 | blk_mq_put_driver_tag(nxt); | |
1149 | } | |
1150 | list_add(&rq->queuelist, list); | |
1151 | __blk_mq_requeue_request(rq); | |
1152 | break; | |
1153 | } | |
1154 | ||
1155 | if (unlikely(ret != BLK_STS_OK)) { | |
1156 | errors++; | |
1157 | blk_mq_end_request(rq, BLK_STS_IOERR); | |
1158 | continue; | |
1159 | } | |
1160 | ||
1161 | queued++; | |
1162 | } while (!list_empty(list)); | |
1163 | ||
1164 | hctx->dispatched[queued_to_index(queued)]++; | |
1165 | ||
1166 | /* | |
1167 | * Any items that need requeuing? Stuff them into hctx->dispatch, | |
1168 | * that is where we will continue on next queue run. | |
1169 | */ | |
1170 | if (!list_empty(list)) { | |
1171 | spin_lock(&hctx->lock); | |
1172 | list_splice_init(list, &hctx->dispatch); | |
1173 | spin_unlock(&hctx->lock); | |
1174 | ||
1175 | /* | |
1176 | * If SCHED_RESTART was set by the caller of this function and | |
1177 | * it is no longer set that means that it was cleared by another | |
1178 | * thread and hence that a queue rerun is needed. | |
1179 | * | |
1180 | * If 'no_tag' is set, that means that we failed getting | |
1181 | * a driver tag with an I/O scheduler attached. If our dispatch | |
1182 | * waitqueue is no longer active, ensure that we run the queue | |
1183 | * AFTER adding our entries back to the list. | |
1184 | * | |
1185 | * If no I/O scheduler has been configured it is possible that | |
1186 | * the hardware queue got stopped and restarted before requests | |
1187 | * were pushed back onto the dispatch list. Rerun the queue to | |
1188 | * avoid starvation. Notes: | |
1189 | * - blk_mq_run_hw_queue() checks whether or not a queue has | |
1190 | * been stopped before rerunning a queue. | |
1191 | * - Some but not all block drivers stop a queue before | |
1192 | * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq | |
1193 | * and dm-rq. | |
1194 | */ | |
1195 | if (!blk_mq_sched_needs_restart(hctx) || | |
1196 | (no_tag && list_empty_careful(&hctx->dispatch_wait.entry))) | |
1197 | blk_mq_run_hw_queue(hctx, true); | |
1198 | } | |
1199 | ||
1200 | return (queued + errors) != 0; | |
1201 | } | |
1202 | ||
1203 | static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx) | |
1204 | { | |
1205 | int srcu_idx; | |
1206 | ||
1207 | /* | |
1208 | * We should be running this queue from one of the CPUs that | |
1209 | * are mapped to it. | |
1210 | * | |
1211 | * There are at least two related races now between setting | |
1212 | * hctx->next_cpu from blk_mq_hctx_next_cpu() and running | |
1213 | * __blk_mq_run_hw_queue(): | |
1214 | * | |
1215 | * - hctx->next_cpu is found offline in blk_mq_hctx_next_cpu(), | |
1216 | * but later it becomes online, then this warning is harmless | |
1217 | * at all | |
1218 | * | |
1219 | * - hctx->next_cpu is found online in blk_mq_hctx_next_cpu(), | |
1220 | * but later it becomes offline, then the warning can't be | |
1221 | * triggered, and we depend on blk-mq timeout handler to | |
1222 | * handle dispatched requests to this hctx | |
1223 | */ | |
1224 | if (!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) && | |
1225 | cpu_online(hctx->next_cpu)) { | |
1226 | printk(KERN_WARNING "run queue from wrong CPU %d, hctx %s\n", | |
1227 | raw_smp_processor_id(), | |
1228 | cpumask_empty(hctx->cpumask) ? "inactive": "active"); | |
1229 | dump_stack(); | |
1230 | } | |
1231 | ||
1232 | /* | |
1233 | * We can't run the queue inline with ints disabled. Ensure that | |
1234 | * we catch bad users of this early. | |
1235 | */ | |
1236 | WARN_ON_ONCE(in_interrupt()); | |
1237 | ||
1238 | if (!(hctx->flags & BLK_MQ_F_BLOCKING)) { | |
1239 | rcu_read_lock(); | |
1240 | blk_mq_sched_dispatch_requests(hctx); | |
1241 | rcu_read_unlock(); | |
1242 | } else { | |
1243 | might_sleep(); | |
1244 | ||
1245 | srcu_idx = srcu_read_lock(hctx->queue_rq_srcu); | |
1246 | blk_mq_sched_dispatch_requests(hctx); | |
1247 | srcu_read_unlock(hctx->queue_rq_srcu, srcu_idx); | |
1248 | } | |
1249 | } | |
1250 | ||
1251 | /* | |
1252 | * It'd be great if the workqueue API had a way to pass | |
1253 | * in a mask and had some smarts for more clever placement. | |
1254 | * For now we just round-robin here, switching for every | |
1255 | * BLK_MQ_CPU_WORK_BATCH queued items. | |
1256 | */ | |
1257 | static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx) | |
1258 | { | |
1259 | bool tried = false; | |
1260 | int next_cpu = hctx->next_cpu; | |
1261 | ||
1262 | if (hctx->queue->nr_hw_queues == 1) | |
1263 | return WORK_CPU_UNBOUND; | |
1264 | ||
1265 | if (--hctx->next_cpu_batch <= 0) { | |
1266 | select_cpu: | |
1267 | next_cpu = cpumask_next_and(next_cpu, hctx->cpumask, | |
1268 | cpu_online_mask); | |
1269 | if (next_cpu >= nr_cpu_ids) | |
1270 | next_cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask); | |
1271 | ||
1272 | /* | |
1273 | * No online CPU is found, so have to make sure hctx->next_cpu | |
1274 | * is set correctly for not breaking workqueue. | |
1275 | */ | |
1276 | if (next_cpu >= nr_cpu_ids) | |
1277 | next_cpu = cpumask_first(hctx->cpumask); | |
1278 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; | |
1279 | } | |
1280 | ||
1281 | /* | |
1282 | * Do unbound schedule if we can't find a online CPU for this hctx, | |
1283 | * and it should only happen in the path of handling CPU DEAD. | |
1284 | */ | |
1285 | if (!cpu_online(next_cpu)) { | |
1286 | if (!tried) { | |
1287 | tried = true; | |
1288 | goto select_cpu; | |
1289 | } | |
1290 | ||
1291 | /* | |
1292 | * Make sure to re-select CPU next time once after CPUs | |
1293 | * in hctx->cpumask become online again. | |
1294 | */ | |
1295 | hctx->next_cpu = next_cpu; | |
1296 | hctx->next_cpu_batch = 1; | |
1297 | return WORK_CPU_UNBOUND; | |
1298 | } | |
1299 | ||
1300 | hctx->next_cpu = next_cpu; | |
1301 | return next_cpu; | |
1302 | } | |
1303 | ||
1304 | static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async, | |
1305 | unsigned long msecs) | |
1306 | { | |
1307 | if (WARN_ON_ONCE(!blk_mq_hw_queue_mapped(hctx))) | |
1308 | return; | |
1309 | ||
1310 | if (unlikely(blk_mq_hctx_stopped(hctx))) | |
1311 | return; | |
1312 | ||
1313 | if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) { | |
1314 | int cpu = get_cpu(); | |
1315 | if (cpumask_test_cpu(cpu, hctx->cpumask)) { | |
1316 | __blk_mq_run_hw_queue(hctx); | |
1317 | put_cpu(); | |
1318 | return; | |
1319 | } | |
1320 | ||
1321 | put_cpu(); | |
1322 | } | |
1323 | ||
1324 | kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work, | |
1325 | msecs_to_jiffies(msecs)); | |
1326 | } | |
1327 | ||
1328 | void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs) | |
1329 | { | |
1330 | __blk_mq_delay_run_hw_queue(hctx, true, msecs); | |
1331 | } | |
1332 | EXPORT_SYMBOL(blk_mq_delay_run_hw_queue); | |
1333 | ||
1334 | bool blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) | |
1335 | { | |
1336 | if (blk_mq_hctx_has_pending(hctx)) { | |
1337 | __blk_mq_delay_run_hw_queue(hctx, async, 0); | |
1338 | return true; | |
1339 | } | |
1340 | ||
1341 | return false; | |
1342 | } | |
1343 | EXPORT_SYMBOL(blk_mq_run_hw_queue); | |
1344 | ||
1345 | void blk_mq_run_hw_queues(struct request_queue *q, bool async) | |
1346 | { | |
1347 | struct blk_mq_hw_ctx *hctx; | |
1348 | int i; | |
1349 | ||
1350 | queue_for_each_hw_ctx(q, hctx, i) { | |
1351 | if (blk_mq_hctx_stopped(hctx)) | |
1352 | continue; | |
1353 | ||
1354 | blk_mq_run_hw_queue(hctx, async); | |
1355 | } | |
1356 | } | |
1357 | EXPORT_SYMBOL(blk_mq_run_hw_queues); | |
1358 | ||
1359 | /** | |
1360 | * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped | |
1361 | * @q: request queue. | |
1362 | * | |
1363 | * The caller is responsible for serializing this function against | |
1364 | * blk_mq_{start,stop}_hw_queue(). | |
1365 | */ | |
1366 | bool blk_mq_queue_stopped(struct request_queue *q) | |
1367 | { | |
1368 | struct blk_mq_hw_ctx *hctx; | |
1369 | int i; | |
1370 | ||
1371 | queue_for_each_hw_ctx(q, hctx, i) | |
1372 | if (blk_mq_hctx_stopped(hctx)) | |
1373 | return true; | |
1374 | ||
1375 | return false; | |
1376 | } | |
1377 | EXPORT_SYMBOL(blk_mq_queue_stopped); | |
1378 | ||
1379 | /* | |
1380 | * This function is often used for pausing .queue_rq() by driver when | |
1381 | * there isn't enough resource or some conditions aren't satisfied, and | |
1382 | * BLK_STS_RESOURCE is usually returned. | |
1383 | * | |
1384 | * We do not guarantee that dispatch can be drained or blocked | |
1385 | * after blk_mq_stop_hw_queue() returns. Please use | |
1386 | * blk_mq_quiesce_queue() for that requirement. | |
1387 | */ | |
1388 | void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx) | |
1389 | { | |
1390 | cancel_delayed_work(&hctx->run_work); | |
1391 | ||
1392 | set_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
1393 | } | |
1394 | EXPORT_SYMBOL(blk_mq_stop_hw_queue); | |
1395 | ||
1396 | /* | |
1397 | * This function is often used for pausing .queue_rq() by driver when | |
1398 | * there isn't enough resource or some conditions aren't satisfied, and | |
1399 | * BLK_STS_RESOURCE is usually returned. | |
1400 | * | |
1401 | * We do not guarantee that dispatch can be drained or blocked | |
1402 | * after blk_mq_stop_hw_queues() returns. Please use | |
1403 | * blk_mq_quiesce_queue() for that requirement. | |
1404 | */ | |
1405 | void blk_mq_stop_hw_queues(struct request_queue *q) | |
1406 | { | |
1407 | struct blk_mq_hw_ctx *hctx; | |
1408 | int i; | |
1409 | ||
1410 | queue_for_each_hw_ctx(q, hctx, i) | |
1411 | blk_mq_stop_hw_queue(hctx); | |
1412 | } | |
1413 | EXPORT_SYMBOL(blk_mq_stop_hw_queues); | |
1414 | ||
1415 | void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx) | |
1416 | { | |
1417 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
1418 | ||
1419 | blk_mq_run_hw_queue(hctx, false); | |
1420 | } | |
1421 | EXPORT_SYMBOL(blk_mq_start_hw_queue); | |
1422 | ||
1423 | void blk_mq_start_hw_queues(struct request_queue *q) | |
1424 | { | |
1425 | struct blk_mq_hw_ctx *hctx; | |
1426 | int i; | |
1427 | ||
1428 | queue_for_each_hw_ctx(q, hctx, i) | |
1429 | blk_mq_start_hw_queue(hctx); | |
1430 | } | |
1431 | EXPORT_SYMBOL(blk_mq_start_hw_queues); | |
1432 | ||
1433 | void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) | |
1434 | { | |
1435 | if (!blk_mq_hctx_stopped(hctx)) | |
1436 | return; | |
1437 | ||
1438 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
1439 | blk_mq_run_hw_queue(hctx, async); | |
1440 | } | |
1441 | EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue); | |
1442 | ||
1443 | void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async) | |
1444 | { | |
1445 | struct blk_mq_hw_ctx *hctx; | |
1446 | int i; | |
1447 | ||
1448 | queue_for_each_hw_ctx(q, hctx, i) | |
1449 | blk_mq_start_stopped_hw_queue(hctx, async); | |
1450 | } | |
1451 | EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues); | |
1452 | ||
1453 | static void blk_mq_run_work_fn(struct work_struct *work) | |
1454 | { | |
1455 | struct blk_mq_hw_ctx *hctx; | |
1456 | ||
1457 | hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work); | |
1458 | ||
1459 | /* | |
1460 | * If we are stopped, don't run the queue. The exception is if | |
1461 | * BLK_MQ_S_START_ON_RUN is set. For that case, we auto-clear | |
1462 | * the STOPPED bit and run it. | |
1463 | */ | |
1464 | if (test_bit(BLK_MQ_S_STOPPED, &hctx->state)) { | |
1465 | if (!test_bit(BLK_MQ_S_START_ON_RUN, &hctx->state)) | |
1466 | return; | |
1467 | ||
1468 | clear_bit(BLK_MQ_S_START_ON_RUN, &hctx->state); | |
1469 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
1470 | } | |
1471 | ||
1472 | __blk_mq_run_hw_queue(hctx); | |
1473 | } | |
1474 | ||
1475 | ||
1476 | void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs) | |
1477 | { | |
1478 | if (WARN_ON_ONCE(!blk_mq_hw_queue_mapped(hctx))) | |
1479 | return; | |
1480 | ||
1481 | /* | |
1482 | * Stop the hw queue, then modify currently delayed work. | |
1483 | * This should prevent us from running the queue prematurely. | |
1484 | * Mark the queue as auto-clearing STOPPED when it runs. | |
1485 | */ | |
1486 | blk_mq_stop_hw_queue(hctx); | |
1487 | set_bit(BLK_MQ_S_START_ON_RUN, &hctx->state); | |
1488 | kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), | |
1489 | &hctx->run_work, | |
1490 | msecs_to_jiffies(msecs)); | |
1491 | } | |
1492 | EXPORT_SYMBOL(blk_mq_delay_queue); | |
1493 | ||
1494 | static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx, | |
1495 | struct request *rq, | |
1496 | bool at_head) | |
1497 | { | |
1498 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
1499 | ||
1500 | lockdep_assert_held(&ctx->lock); | |
1501 | ||
1502 | trace_block_rq_insert(hctx->queue, rq); | |
1503 | ||
1504 | if (at_head) | |
1505 | list_add(&rq->queuelist, &ctx->rq_list); | |
1506 | else | |
1507 | list_add_tail(&rq->queuelist, &ctx->rq_list); | |
1508 | } | |
1509 | ||
1510 | void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq, | |
1511 | bool at_head) | |
1512 | { | |
1513 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
1514 | ||
1515 | lockdep_assert_held(&ctx->lock); | |
1516 | ||
1517 | __blk_mq_insert_req_list(hctx, rq, at_head); | |
1518 | blk_mq_hctx_mark_pending(hctx, ctx); | |
1519 | } | |
1520 | ||
1521 | /* | |
1522 | * Should only be used carefully, when the caller knows we want to | |
1523 | * bypass a potential IO scheduler on the target device. | |
1524 | */ | |
1525 | void blk_mq_request_bypass_insert(struct request *rq, bool run_queue) | |
1526 | { | |
1527 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
1528 | struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(rq->q, ctx->cpu); | |
1529 | ||
1530 | spin_lock(&hctx->lock); | |
1531 | list_add_tail(&rq->queuelist, &hctx->dispatch); | |
1532 | spin_unlock(&hctx->lock); | |
1533 | ||
1534 | if (run_queue) | |
1535 | blk_mq_run_hw_queue(hctx, false); | |
1536 | } | |
1537 | ||
1538 | void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx, | |
1539 | struct list_head *list) | |
1540 | ||
1541 | { | |
1542 | /* | |
1543 | * preemption doesn't flush plug list, so it's possible ctx->cpu is | |
1544 | * offline now | |
1545 | */ | |
1546 | spin_lock(&ctx->lock); | |
1547 | while (!list_empty(list)) { | |
1548 | struct request *rq; | |
1549 | ||
1550 | rq = list_first_entry(list, struct request, queuelist); | |
1551 | BUG_ON(rq->mq_ctx != ctx); | |
1552 | list_del_init(&rq->queuelist); | |
1553 | __blk_mq_insert_req_list(hctx, rq, false); | |
1554 | } | |
1555 | blk_mq_hctx_mark_pending(hctx, ctx); | |
1556 | spin_unlock(&ctx->lock); | |
1557 | } | |
1558 | ||
1559 | static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b) | |
1560 | { | |
1561 | struct request *rqa = container_of(a, struct request, queuelist); | |
1562 | struct request *rqb = container_of(b, struct request, queuelist); | |
1563 | ||
1564 | return !(rqa->mq_ctx < rqb->mq_ctx || | |
1565 | (rqa->mq_ctx == rqb->mq_ctx && | |
1566 | blk_rq_pos(rqa) < blk_rq_pos(rqb))); | |
1567 | } | |
1568 | ||
1569 | void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule) | |
1570 | { | |
1571 | struct blk_mq_ctx *this_ctx; | |
1572 | struct request_queue *this_q; | |
1573 | struct request *rq; | |
1574 | LIST_HEAD(list); | |
1575 | LIST_HEAD(ctx_list); | |
1576 | unsigned int depth; | |
1577 | ||
1578 | list_splice_init(&plug->mq_list, &list); | |
1579 | ||
1580 | list_sort(NULL, &list, plug_ctx_cmp); | |
1581 | ||
1582 | this_q = NULL; | |
1583 | this_ctx = NULL; | |
1584 | depth = 0; | |
1585 | ||
1586 | while (!list_empty(&list)) { | |
1587 | rq = list_entry_rq(list.next); | |
1588 | list_del_init(&rq->queuelist); | |
1589 | BUG_ON(!rq->q); | |
1590 | if (rq->mq_ctx != this_ctx) { | |
1591 | if (this_ctx) { | |
1592 | trace_block_unplug(this_q, depth, from_schedule); | |
1593 | blk_mq_sched_insert_requests(this_q, this_ctx, | |
1594 | &ctx_list, | |
1595 | from_schedule); | |
1596 | } | |
1597 | ||
1598 | this_ctx = rq->mq_ctx; | |
1599 | this_q = rq->q; | |
1600 | depth = 0; | |
1601 | } | |
1602 | ||
1603 | depth++; | |
1604 | list_add_tail(&rq->queuelist, &ctx_list); | |
1605 | } | |
1606 | ||
1607 | /* | |
1608 | * If 'this_ctx' is set, we know we have entries to complete | |
1609 | * on 'ctx_list'. Do those. | |
1610 | */ | |
1611 | if (this_ctx) { | |
1612 | trace_block_unplug(this_q, depth, from_schedule); | |
1613 | blk_mq_sched_insert_requests(this_q, this_ctx, &ctx_list, | |
1614 | from_schedule); | |
1615 | } | |
1616 | } | |
1617 | ||
1618 | static void blk_mq_bio_to_request(struct request *rq, struct bio *bio) | |
1619 | { | |
1620 | blk_init_request_from_bio(rq, bio); | |
1621 | ||
1622 | blk_rq_set_rl(rq, blk_get_rl(rq->q, bio)); | |
1623 | ||
1624 | blk_account_io_start(rq, true); | |
1625 | } | |
1626 | ||
1627 | static inline void blk_mq_queue_io(struct blk_mq_hw_ctx *hctx, | |
1628 | struct blk_mq_ctx *ctx, | |
1629 | struct request *rq) | |
1630 | { | |
1631 | spin_lock(&ctx->lock); | |
1632 | __blk_mq_insert_request(hctx, rq, false); | |
1633 | spin_unlock(&ctx->lock); | |
1634 | } | |
1635 | ||
1636 | static blk_qc_t request_to_qc_t(struct blk_mq_hw_ctx *hctx, struct request *rq) | |
1637 | { | |
1638 | if (rq->tag != -1) | |
1639 | return blk_tag_to_qc_t(rq->tag, hctx->queue_num, false); | |
1640 | ||
1641 | return blk_tag_to_qc_t(rq->internal_tag, hctx->queue_num, true); | |
1642 | } | |
1643 | ||
1644 | static void __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx, | |
1645 | struct request *rq, | |
1646 | blk_qc_t *cookie, bool may_sleep) | |
1647 | { | |
1648 | struct request_queue *q = rq->q; | |
1649 | struct blk_mq_queue_data bd = { | |
1650 | .rq = rq, | |
1651 | .last = true, | |
1652 | }; | |
1653 | blk_qc_t new_cookie; | |
1654 | blk_status_t ret; | |
1655 | bool run_queue = true; | |
1656 | ||
1657 | /* RCU or SRCU read lock is needed before checking quiesced flag */ | |
1658 | if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)) { | |
1659 | run_queue = false; | |
1660 | goto insert; | |
1661 | } | |
1662 | ||
1663 | if (q->elevator) | |
1664 | goto insert; | |
1665 | ||
1666 | if (!blk_mq_get_dispatch_budget(hctx)) | |
1667 | goto insert; | |
1668 | ||
1669 | if (!blk_mq_get_driver_tag(rq, NULL, false)) { | |
1670 | blk_mq_put_dispatch_budget(hctx); | |
1671 | goto insert; | |
1672 | } | |
1673 | ||
1674 | new_cookie = request_to_qc_t(hctx, rq); | |
1675 | ||
1676 | /* | |
1677 | * For OK queue, we are done. For error, kill it. Any other | |
1678 | * error (busy), just add it to our list as we previously | |
1679 | * would have done | |
1680 | */ | |
1681 | ret = q->mq_ops->queue_rq(hctx, &bd); | |
1682 | switch (ret) { | |
1683 | case BLK_STS_OK: | |
1684 | *cookie = new_cookie; | |
1685 | return; | |
1686 | case BLK_STS_RESOURCE: | |
1687 | __blk_mq_requeue_request(rq); | |
1688 | goto insert; | |
1689 | default: | |
1690 | *cookie = BLK_QC_T_NONE; | |
1691 | blk_mq_end_request(rq, ret); | |
1692 | return; | |
1693 | } | |
1694 | ||
1695 | insert: | |
1696 | blk_mq_sched_insert_request(rq, false, run_queue, false, may_sleep); | |
1697 | } | |
1698 | ||
1699 | static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx, | |
1700 | struct request *rq, blk_qc_t *cookie) | |
1701 | { | |
1702 | if (!(hctx->flags & BLK_MQ_F_BLOCKING)) { | |
1703 | rcu_read_lock(); | |
1704 | __blk_mq_try_issue_directly(hctx, rq, cookie, false); | |
1705 | rcu_read_unlock(); | |
1706 | } else { | |
1707 | unsigned int srcu_idx; | |
1708 | ||
1709 | might_sleep(); | |
1710 | ||
1711 | srcu_idx = srcu_read_lock(hctx->queue_rq_srcu); | |
1712 | __blk_mq_try_issue_directly(hctx, rq, cookie, true); | |
1713 | srcu_read_unlock(hctx->queue_rq_srcu, srcu_idx); | |
1714 | } | |
1715 | } | |
1716 | ||
1717 | static blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio) | |
1718 | { | |
1719 | const int is_sync = op_is_sync(bio->bi_opf); | |
1720 | const int is_flush_fua = op_is_flush(bio->bi_opf); | |
1721 | struct blk_mq_alloc_data data = { .flags = 0 }; | |
1722 | struct request *rq; | |
1723 | unsigned int request_count = 0; | |
1724 | struct blk_plug *plug; | |
1725 | struct request *same_queue_rq = NULL; | |
1726 | blk_qc_t cookie; | |
1727 | unsigned int wb_acct; | |
1728 | ||
1729 | blk_queue_bounce(q, &bio); | |
1730 | ||
1731 | blk_queue_split(q, &bio); | |
1732 | ||
1733 | if (!bio_integrity_prep(bio)) | |
1734 | return BLK_QC_T_NONE; | |
1735 | ||
1736 | if (!is_flush_fua && !blk_queue_nomerges(q) && | |
1737 | blk_attempt_plug_merge(q, bio, &request_count, &same_queue_rq)) | |
1738 | return BLK_QC_T_NONE; | |
1739 | ||
1740 | if (blk_mq_sched_bio_merge(q, bio)) | |
1741 | return BLK_QC_T_NONE; | |
1742 | ||
1743 | wb_acct = wbt_wait(q->rq_wb, bio, NULL); | |
1744 | ||
1745 | trace_block_getrq(q, bio, bio->bi_opf); | |
1746 | ||
1747 | rq = blk_mq_get_request(q, bio, bio->bi_opf, &data); | |
1748 | if (unlikely(!rq)) { | |
1749 | __wbt_done(q->rq_wb, wb_acct); | |
1750 | if (bio->bi_opf & REQ_NOWAIT) | |
1751 | bio_wouldblock_error(bio); | |
1752 | return BLK_QC_T_NONE; | |
1753 | } | |
1754 | ||
1755 | wbt_track(&rq->issue_stat, wb_acct); | |
1756 | ||
1757 | cookie = request_to_qc_t(data.hctx, rq); | |
1758 | ||
1759 | plug = current->plug; | |
1760 | if (unlikely(is_flush_fua)) { | |
1761 | blk_mq_put_ctx(data.ctx); | |
1762 | blk_mq_bio_to_request(rq, bio); | |
1763 | ||
1764 | /* bypass scheduler for flush rq */ | |
1765 | blk_insert_flush(rq); | |
1766 | blk_mq_run_hw_queue(data.hctx, true); | |
1767 | } else if (plug && q->nr_hw_queues == 1) { | |
1768 | struct request *last = NULL; | |
1769 | ||
1770 | blk_mq_put_ctx(data.ctx); | |
1771 | blk_mq_bio_to_request(rq, bio); | |
1772 | ||
1773 | /* | |
1774 | * @request_count may become stale because of schedule | |
1775 | * out, so check the list again. | |
1776 | */ | |
1777 | if (list_empty(&plug->mq_list)) | |
1778 | request_count = 0; | |
1779 | else if (blk_queue_nomerges(q)) | |
1780 | request_count = blk_plug_queued_count(q); | |
1781 | ||
1782 | if (!request_count) | |
1783 | trace_block_plug(q); | |
1784 | else | |
1785 | last = list_entry_rq(plug->mq_list.prev); | |
1786 | ||
1787 | if (request_count >= BLK_MAX_REQUEST_COUNT || (last && | |
1788 | blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) { | |
1789 | blk_flush_plug_list(plug, false); | |
1790 | trace_block_plug(q); | |
1791 | } | |
1792 | ||
1793 | list_add_tail(&rq->queuelist, &plug->mq_list); | |
1794 | } else if (plug && !blk_queue_nomerges(q)) { | |
1795 | blk_mq_bio_to_request(rq, bio); | |
1796 | ||
1797 | /* | |
1798 | * We do limited plugging. If the bio can be merged, do that. | |
1799 | * Otherwise the existing request in the plug list will be | |
1800 | * issued. So the plug list will have one request at most | |
1801 | * The plug list might get flushed before this. If that happens, | |
1802 | * the plug list is empty, and same_queue_rq is invalid. | |
1803 | */ | |
1804 | if (list_empty(&plug->mq_list)) | |
1805 | same_queue_rq = NULL; | |
1806 | if (same_queue_rq) | |
1807 | list_del_init(&same_queue_rq->queuelist); | |
1808 | list_add_tail(&rq->queuelist, &plug->mq_list); | |
1809 | ||
1810 | blk_mq_put_ctx(data.ctx); | |
1811 | ||
1812 | if (same_queue_rq) { | |
1813 | data.hctx = blk_mq_map_queue(q, | |
1814 | same_queue_rq->mq_ctx->cpu); | |
1815 | blk_mq_try_issue_directly(data.hctx, same_queue_rq, | |
1816 | &cookie); | |
1817 | } | |
1818 | } else if (q->nr_hw_queues > 1 && is_sync) { | |
1819 | blk_mq_put_ctx(data.ctx); | |
1820 | blk_mq_bio_to_request(rq, bio); | |
1821 | blk_mq_try_issue_directly(data.hctx, rq, &cookie); | |
1822 | } else if (q->elevator) { | |
1823 | blk_mq_put_ctx(data.ctx); | |
1824 | blk_mq_bio_to_request(rq, bio); | |
1825 | blk_mq_sched_insert_request(rq, false, true, true, true); | |
1826 | } else { | |
1827 | blk_mq_put_ctx(data.ctx); | |
1828 | blk_mq_bio_to_request(rq, bio); | |
1829 | blk_mq_queue_io(data.hctx, data.ctx, rq); | |
1830 | blk_mq_run_hw_queue(data.hctx, true); | |
1831 | } | |
1832 | ||
1833 | return cookie; | |
1834 | } | |
1835 | ||
1836 | void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags, | |
1837 | unsigned int hctx_idx) | |
1838 | { | |
1839 | struct page *page; | |
1840 | ||
1841 | if (tags->rqs && set->ops->exit_request) { | |
1842 | int i; | |
1843 | ||
1844 | for (i = 0; i < tags->nr_tags; i++) { | |
1845 | struct request *rq = tags->static_rqs[i]; | |
1846 | ||
1847 | if (!rq) | |
1848 | continue; | |
1849 | set->ops->exit_request(set, rq, hctx_idx); | |
1850 | tags->static_rqs[i] = NULL; | |
1851 | } | |
1852 | } | |
1853 | ||
1854 | while (!list_empty(&tags->page_list)) { | |
1855 | page = list_first_entry(&tags->page_list, struct page, lru); | |
1856 | list_del_init(&page->lru); | |
1857 | /* | |
1858 | * Remove kmemleak object previously allocated in | |
1859 | * blk_mq_init_rq_map(). | |
1860 | */ | |
1861 | kmemleak_free(page_address(page)); | |
1862 | __free_pages(page, page->private); | |
1863 | } | |
1864 | } | |
1865 | ||
1866 | void blk_mq_free_rq_map(struct blk_mq_tags *tags) | |
1867 | { | |
1868 | kfree(tags->rqs); | |
1869 | tags->rqs = NULL; | |
1870 | kfree(tags->static_rqs); | |
1871 | tags->static_rqs = NULL; | |
1872 | ||
1873 | blk_mq_free_tags(tags); | |
1874 | } | |
1875 | ||
1876 | struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, | |
1877 | unsigned int hctx_idx, | |
1878 | unsigned int nr_tags, | |
1879 | unsigned int reserved_tags) | |
1880 | { | |
1881 | struct blk_mq_tags *tags; | |
1882 | int node; | |
1883 | ||
1884 | node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx); | |
1885 | if (node == NUMA_NO_NODE) | |
1886 | node = set->numa_node; | |
1887 | ||
1888 | tags = blk_mq_init_tags(nr_tags, reserved_tags, node, | |
1889 | BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags)); | |
1890 | if (!tags) | |
1891 | return NULL; | |
1892 | ||
1893 | tags->rqs = kzalloc_node(nr_tags * sizeof(struct request *), | |
1894 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, | |
1895 | node); | |
1896 | if (!tags->rqs) { | |
1897 | blk_mq_free_tags(tags); | |
1898 | return NULL; | |
1899 | } | |
1900 | ||
1901 | tags->static_rqs = kzalloc_node(nr_tags * sizeof(struct request *), | |
1902 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, | |
1903 | node); | |
1904 | if (!tags->static_rqs) { | |
1905 | kfree(tags->rqs); | |
1906 | blk_mq_free_tags(tags); | |
1907 | return NULL; | |
1908 | } | |
1909 | ||
1910 | return tags; | |
1911 | } | |
1912 | ||
1913 | static size_t order_to_size(unsigned int order) | |
1914 | { | |
1915 | return (size_t)PAGE_SIZE << order; | |
1916 | } | |
1917 | ||
1918 | int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags, | |
1919 | unsigned int hctx_idx, unsigned int depth) | |
1920 | { | |
1921 | unsigned int i, j, entries_per_page, max_order = 4; | |
1922 | size_t rq_size, left; | |
1923 | int node; | |
1924 | ||
1925 | node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx); | |
1926 | if (node == NUMA_NO_NODE) | |
1927 | node = set->numa_node; | |
1928 | ||
1929 | INIT_LIST_HEAD(&tags->page_list); | |
1930 | ||
1931 | /* | |
1932 | * rq_size is the size of the request plus driver payload, rounded | |
1933 | * to the cacheline size | |
1934 | */ | |
1935 | rq_size = round_up(sizeof(struct request) + set->cmd_size, | |
1936 | cache_line_size()); | |
1937 | left = rq_size * depth; | |
1938 | ||
1939 | for (i = 0; i < depth; ) { | |
1940 | int this_order = max_order; | |
1941 | struct page *page; | |
1942 | int to_do; | |
1943 | void *p; | |
1944 | ||
1945 | while (this_order && left < order_to_size(this_order - 1)) | |
1946 | this_order--; | |
1947 | ||
1948 | do { | |
1949 | page = alloc_pages_node(node, | |
1950 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO, | |
1951 | this_order); | |
1952 | if (page) | |
1953 | break; | |
1954 | if (!this_order--) | |
1955 | break; | |
1956 | if (order_to_size(this_order) < rq_size) | |
1957 | break; | |
1958 | } while (1); | |
1959 | ||
1960 | if (!page) | |
1961 | goto fail; | |
1962 | ||
1963 | page->private = this_order; | |
1964 | list_add_tail(&page->lru, &tags->page_list); | |
1965 | ||
1966 | p = page_address(page); | |
1967 | /* | |
1968 | * Allow kmemleak to scan these pages as they contain pointers | |
1969 | * to additional allocations like via ops->init_request(). | |
1970 | */ | |
1971 | kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO); | |
1972 | entries_per_page = order_to_size(this_order) / rq_size; | |
1973 | to_do = min(entries_per_page, depth - i); | |
1974 | left -= to_do * rq_size; | |
1975 | for (j = 0; j < to_do; j++) { | |
1976 | struct request *rq = p; | |
1977 | ||
1978 | tags->static_rqs[i] = rq; | |
1979 | if (set->ops->init_request) { | |
1980 | if (set->ops->init_request(set, rq, hctx_idx, | |
1981 | node)) { | |
1982 | tags->static_rqs[i] = NULL; | |
1983 | goto fail; | |
1984 | } | |
1985 | } | |
1986 | ||
1987 | p += rq_size; | |
1988 | i++; | |
1989 | } | |
1990 | } | |
1991 | return 0; | |
1992 | ||
1993 | fail: | |
1994 | blk_mq_free_rqs(set, tags, hctx_idx); | |
1995 | return -ENOMEM; | |
1996 | } | |
1997 | ||
1998 | /* | |
1999 | * 'cpu' is going away. splice any existing rq_list entries from this | |
2000 | * software queue to the hw queue dispatch list, and ensure that it | |
2001 | * gets run. | |
2002 | */ | |
2003 | static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node) | |
2004 | { | |
2005 | struct blk_mq_hw_ctx *hctx; | |
2006 | struct blk_mq_ctx *ctx; | |
2007 | LIST_HEAD(tmp); | |
2008 | ||
2009 | hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead); | |
2010 | ctx = __blk_mq_get_ctx(hctx->queue, cpu); | |
2011 | ||
2012 | spin_lock(&ctx->lock); | |
2013 | if (!list_empty(&ctx->rq_list)) { | |
2014 | list_splice_init(&ctx->rq_list, &tmp); | |
2015 | blk_mq_hctx_clear_pending(hctx, ctx); | |
2016 | } | |
2017 | spin_unlock(&ctx->lock); | |
2018 | ||
2019 | if (list_empty(&tmp)) | |
2020 | return 0; | |
2021 | ||
2022 | spin_lock(&hctx->lock); | |
2023 | list_splice_tail_init(&tmp, &hctx->dispatch); | |
2024 | spin_unlock(&hctx->lock); | |
2025 | ||
2026 | blk_mq_run_hw_queue(hctx, true); | |
2027 | return 0; | |
2028 | } | |
2029 | ||
2030 | static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx) | |
2031 | { | |
2032 | cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD, | |
2033 | &hctx->cpuhp_dead); | |
2034 | } | |
2035 | ||
2036 | /* hctx->ctxs will be freed in queue's release handler */ | |
2037 | static void blk_mq_exit_hctx(struct request_queue *q, | |
2038 | struct blk_mq_tag_set *set, | |
2039 | struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx) | |
2040 | { | |
2041 | blk_mq_debugfs_unregister_hctx(hctx); | |
2042 | ||
2043 | if (blk_mq_hw_queue_mapped(hctx)) | |
2044 | blk_mq_tag_idle(hctx); | |
2045 | ||
2046 | if (set->ops->exit_request) | |
2047 | set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx); | |
2048 | ||
2049 | blk_mq_sched_exit_hctx(q, hctx, hctx_idx); | |
2050 | ||
2051 | if (set->ops->exit_hctx) | |
2052 | set->ops->exit_hctx(hctx, hctx_idx); | |
2053 | ||
2054 | if (hctx->flags & BLK_MQ_F_BLOCKING) | |
2055 | cleanup_srcu_struct(hctx->queue_rq_srcu); | |
2056 | ||
2057 | blk_mq_remove_cpuhp(hctx); | |
2058 | blk_free_flush_queue(hctx->fq); | |
2059 | sbitmap_free(&hctx->ctx_map); | |
2060 | } | |
2061 | ||
2062 | static void blk_mq_exit_hw_queues(struct request_queue *q, | |
2063 | struct blk_mq_tag_set *set, int nr_queue) | |
2064 | { | |
2065 | struct blk_mq_hw_ctx *hctx; | |
2066 | unsigned int i; | |
2067 | ||
2068 | queue_for_each_hw_ctx(q, hctx, i) { | |
2069 | if (i == nr_queue) | |
2070 | break; | |
2071 | blk_mq_exit_hctx(q, set, hctx, i); | |
2072 | } | |
2073 | } | |
2074 | ||
2075 | static int blk_mq_init_hctx(struct request_queue *q, | |
2076 | struct blk_mq_tag_set *set, | |
2077 | struct blk_mq_hw_ctx *hctx, unsigned hctx_idx) | |
2078 | { | |
2079 | int node; | |
2080 | ||
2081 | node = hctx->numa_node; | |
2082 | if (node == NUMA_NO_NODE) | |
2083 | node = hctx->numa_node = set->numa_node; | |
2084 | ||
2085 | INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn); | |
2086 | spin_lock_init(&hctx->lock); | |
2087 | INIT_LIST_HEAD(&hctx->dispatch); | |
2088 | hctx->queue = q; | |
2089 | hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED; | |
2090 | ||
2091 | cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead); | |
2092 | ||
2093 | hctx->tags = set->tags[hctx_idx]; | |
2094 | ||
2095 | /* | |
2096 | * Allocate space for all possible cpus to avoid allocation at | |
2097 | * runtime | |
2098 | */ | |
2099 | hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *), | |
2100 | GFP_KERNEL, node); | |
2101 | if (!hctx->ctxs) | |
2102 | goto unregister_cpu_notifier; | |
2103 | ||
2104 | if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8), GFP_KERNEL, | |
2105 | node)) | |
2106 | goto free_ctxs; | |
2107 | ||
2108 | hctx->nr_ctx = 0; | |
2109 | ||
2110 | init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake); | |
2111 | INIT_LIST_HEAD(&hctx->dispatch_wait.entry); | |
2112 | ||
2113 | if (set->ops->init_hctx && | |
2114 | set->ops->init_hctx(hctx, set->driver_data, hctx_idx)) | |
2115 | goto free_bitmap; | |
2116 | ||
2117 | if (blk_mq_sched_init_hctx(q, hctx, hctx_idx)) | |
2118 | goto exit_hctx; | |
2119 | ||
2120 | hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size); | |
2121 | if (!hctx->fq) | |
2122 | goto sched_exit_hctx; | |
2123 | ||
2124 | if (set->ops->init_request && | |
2125 | set->ops->init_request(set, hctx->fq->flush_rq, hctx_idx, | |
2126 | node)) | |
2127 | goto free_fq; | |
2128 | ||
2129 | if (hctx->flags & BLK_MQ_F_BLOCKING) | |
2130 | init_srcu_struct(hctx->queue_rq_srcu); | |
2131 | ||
2132 | blk_mq_debugfs_register_hctx(q, hctx); | |
2133 | ||
2134 | return 0; | |
2135 | ||
2136 | free_fq: | |
2137 | kfree(hctx->fq); | |
2138 | sched_exit_hctx: | |
2139 | blk_mq_sched_exit_hctx(q, hctx, hctx_idx); | |
2140 | exit_hctx: | |
2141 | if (set->ops->exit_hctx) | |
2142 | set->ops->exit_hctx(hctx, hctx_idx); | |
2143 | free_bitmap: | |
2144 | sbitmap_free(&hctx->ctx_map); | |
2145 | free_ctxs: | |
2146 | kfree(hctx->ctxs); | |
2147 | unregister_cpu_notifier: | |
2148 | blk_mq_remove_cpuhp(hctx); | |
2149 | return -1; | |
2150 | } | |
2151 | ||
2152 | static void blk_mq_init_cpu_queues(struct request_queue *q, | |
2153 | unsigned int nr_hw_queues) | |
2154 | { | |
2155 | unsigned int i; | |
2156 | ||
2157 | for_each_possible_cpu(i) { | |
2158 | struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i); | |
2159 | struct blk_mq_hw_ctx *hctx; | |
2160 | ||
2161 | __ctx->cpu = i; | |
2162 | spin_lock_init(&__ctx->lock); | |
2163 | INIT_LIST_HEAD(&__ctx->rq_list); | |
2164 | __ctx->queue = q; | |
2165 | ||
2166 | /* | |
2167 | * Set local node, IFF we have more than one hw queue. If | |
2168 | * not, we remain on the home node of the device | |
2169 | */ | |
2170 | hctx = blk_mq_map_queue(q, i); | |
2171 | if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE) | |
2172 | hctx->numa_node = local_memory_node(cpu_to_node(i)); | |
2173 | } | |
2174 | } | |
2175 | ||
2176 | static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, int hctx_idx) | |
2177 | { | |
2178 | int ret = 0; | |
2179 | ||
2180 | set->tags[hctx_idx] = blk_mq_alloc_rq_map(set, hctx_idx, | |
2181 | set->queue_depth, set->reserved_tags); | |
2182 | if (!set->tags[hctx_idx]) | |
2183 | return false; | |
2184 | ||
2185 | ret = blk_mq_alloc_rqs(set, set->tags[hctx_idx], hctx_idx, | |
2186 | set->queue_depth); | |
2187 | if (!ret) | |
2188 | return true; | |
2189 | ||
2190 | blk_mq_free_rq_map(set->tags[hctx_idx]); | |
2191 | set->tags[hctx_idx] = NULL; | |
2192 | return false; | |
2193 | } | |
2194 | ||
2195 | static void blk_mq_free_map_and_requests(struct blk_mq_tag_set *set, | |
2196 | unsigned int hctx_idx) | |
2197 | { | |
2198 | if (set->tags[hctx_idx]) { | |
2199 | blk_mq_free_rqs(set, set->tags[hctx_idx], hctx_idx); | |
2200 | blk_mq_free_rq_map(set->tags[hctx_idx]); | |
2201 | set->tags[hctx_idx] = NULL; | |
2202 | } | |
2203 | } | |
2204 | ||
2205 | static void blk_mq_map_swqueue(struct request_queue *q) | |
2206 | { | |
2207 | unsigned int i, hctx_idx; | |
2208 | struct blk_mq_hw_ctx *hctx; | |
2209 | struct blk_mq_ctx *ctx; | |
2210 | struct blk_mq_tag_set *set = q->tag_set; | |
2211 | ||
2212 | /* | |
2213 | * Avoid others reading imcomplete hctx->cpumask through sysfs | |
2214 | */ | |
2215 | mutex_lock(&q->sysfs_lock); | |
2216 | ||
2217 | queue_for_each_hw_ctx(q, hctx, i) { | |
2218 | cpumask_clear(hctx->cpumask); | |
2219 | hctx->nr_ctx = 0; | |
2220 | } | |
2221 | ||
2222 | /* | |
2223 | * Map software to hardware queues. | |
2224 | * | |
2225 | * If the cpu isn't present, the cpu is mapped to first hctx. | |
2226 | */ | |
2227 | for_each_possible_cpu(i) { | |
2228 | hctx_idx = q->mq_map[i]; | |
2229 | /* unmapped hw queue can be remapped after CPU topo changed */ | |
2230 | if (!set->tags[hctx_idx] && | |
2231 | !__blk_mq_alloc_rq_map(set, hctx_idx)) { | |
2232 | /* | |
2233 | * If tags initialization fail for some hctx, | |
2234 | * that hctx won't be brought online. In this | |
2235 | * case, remap the current ctx to hctx[0] which | |
2236 | * is guaranteed to always have tags allocated | |
2237 | */ | |
2238 | q->mq_map[i] = 0; | |
2239 | } | |
2240 | ||
2241 | ctx = per_cpu_ptr(q->queue_ctx, i); | |
2242 | hctx = blk_mq_map_queue(q, i); | |
2243 | ||
2244 | cpumask_set_cpu(i, hctx->cpumask); | |
2245 | ctx->index_hw = hctx->nr_ctx; | |
2246 | hctx->ctxs[hctx->nr_ctx++] = ctx; | |
2247 | } | |
2248 | ||
2249 | mutex_unlock(&q->sysfs_lock); | |
2250 | ||
2251 | queue_for_each_hw_ctx(q, hctx, i) { | |
2252 | /* | |
2253 | * If no software queues are mapped to this hardware queue, | |
2254 | * disable it and free the request entries. | |
2255 | */ | |
2256 | if (!hctx->nr_ctx) { | |
2257 | /* Never unmap queue 0. We need it as a | |
2258 | * fallback in case of a new remap fails | |
2259 | * allocation | |
2260 | */ | |
2261 | if (i && set->tags[i]) | |
2262 | blk_mq_free_map_and_requests(set, i); | |
2263 | ||
2264 | hctx->tags = NULL; | |
2265 | continue; | |
2266 | } | |
2267 | ||
2268 | hctx->tags = set->tags[i]; | |
2269 | WARN_ON(!hctx->tags); | |
2270 | ||
2271 | /* | |
2272 | * Set the map size to the number of mapped software queues. | |
2273 | * This is more accurate and more efficient than looping | |
2274 | * over all possibly mapped software queues. | |
2275 | */ | |
2276 | sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx); | |
2277 | ||
2278 | /* | |
2279 | * Initialize batch roundrobin counts | |
2280 | */ | |
2281 | hctx->next_cpu = cpumask_first_and(hctx->cpumask, | |
2282 | cpu_online_mask); | |
2283 | if (hctx->next_cpu >= nr_cpu_ids) | |
2284 | hctx->next_cpu = cpumask_first(hctx->cpumask); | |
2285 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; | |
2286 | } | |
2287 | } | |
2288 | ||
2289 | /* | |
2290 | * Caller needs to ensure that we're either frozen/quiesced, or that | |
2291 | * the queue isn't live yet. | |
2292 | */ | |
2293 | static void queue_set_hctx_shared(struct request_queue *q, bool shared) | |
2294 | { | |
2295 | struct blk_mq_hw_ctx *hctx; | |
2296 | int i; | |
2297 | ||
2298 | queue_for_each_hw_ctx(q, hctx, i) { | |
2299 | if (shared) { | |
2300 | if (test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state)) | |
2301 | atomic_inc(&q->shared_hctx_restart); | |
2302 | hctx->flags |= BLK_MQ_F_TAG_SHARED; | |
2303 | } else { | |
2304 | if (test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state)) | |
2305 | atomic_dec(&q->shared_hctx_restart); | |
2306 | hctx->flags &= ~BLK_MQ_F_TAG_SHARED; | |
2307 | } | |
2308 | } | |
2309 | } | |
2310 | ||
2311 | static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set, | |
2312 | bool shared) | |
2313 | { | |
2314 | struct request_queue *q; | |
2315 | ||
2316 | lockdep_assert_held(&set->tag_list_lock); | |
2317 | ||
2318 | list_for_each_entry(q, &set->tag_list, tag_set_list) { | |
2319 | blk_mq_freeze_queue(q); | |
2320 | queue_set_hctx_shared(q, shared); | |
2321 | blk_mq_unfreeze_queue(q); | |
2322 | } | |
2323 | } | |
2324 | ||
2325 | static void blk_mq_del_queue_tag_set(struct request_queue *q) | |
2326 | { | |
2327 | struct blk_mq_tag_set *set = q->tag_set; | |
2328 | ||
2329 | mutex_lock(&set->tag_list_lock); | |
2330 | list_del_rcu(&q->tag_set_list); | |
2331 | INIT_LIST_HEAD(&q->tag_set_list); | |
2332 | if (list_is_singular(&set->tag_list)) { | |
2333 | /* just transitioned to unshared */ | |
2334 | set->flags &= ~BLK_MQ_F_TAG_SHARED; | |
2335 | /* update existing queue */ | |
2336 | blk_mq_update_tag_set_depth(set, false); | |
2337 | } | |
2338 | mutex_unlock(&set->tag_list_lock); | |
2339 | ||
2340 | synchronize_rcu(); | |
2341 | } | |
2342 | ||
2343 | static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set, | |
2344 | struct request_queue *q) | |
2345 | { | |
2346 | q->tag_set = set; | |
2347 | ||
2348 | mutex_lock(&set->tag_list_lock); | |
2349 | ||
2350 | /* | |
2351 | * Check to see if we're transitioning to shared (from 1 to 2 queues). | |
2352 | */ | |
2353 | if (!list_empty(&set->tag_list) && | |
2354 | !(set->flags & BLK_MQ_F_TAG_SHARED)) { | |
2355 | set->flags |= BLK_MQ_F_TAG_SHARED; | |
2356 | /* update existing queue */ | |
2357 | blk_mq_update_tag_set_depth(set, true); | |
2358 | } | |
2359 | if (set->flags & BLK_MQ_F_TAG_SHARED) | |
2360 | queue_set_hctx_shared(q, true); | |
2361 | list_add_tail_rcu(&q->tag_set_list, &set->tag_list); | |
2362 | ||
2363 | mutex_unlock(&set->tag_list_lock); | |
2364 | } | |
2365 | ||
2366 | /* | |
2367 | * It is the actual release handler for mq, but we do it from | |
2368 | * request queue's release handler for avoiding use-after-free | |
2369 | * and headache because q->mq_kobj shouldn't have been introduced, | |
2370 | * but we can't group ctx/kctx kobj without it. | |
2371 | */ | |
2372 | void blk_mq_release(struct request_queue *q) | |
2373 | { | |
2374 | struct blk_mq_hw_ctx *hctx; | |
2375 | unsigned int i; | |
2376 | ||
2377 | /* hctx kobj stays in hctx */ | |
2378 | queue_for_each_hw_ctx(q, hctx, i) { | |
2379 | if (!hctx) | |
2380 | continue; | |
2381 | kobject_put(&hctx->kobj); | |
2382 | } | |
2383 | ||
2384 | q->mq_map = NULL; | |
2385 | ||
2386 | kfree(q->queue_hw_ctx); | |
2387 | ||
2388 | /* | |
2389 | * release .mq_kobj and sw queue's kobject now because | |
2390 | * both share lifetime with request queue. | |
2391 | */ | |
2392 | blk_mq_sysfs_deinit(q); | |
2393 | ||
2394 | free_percpu(q->queue_ctx); | |
2395 | } | |
2396 | ||
2397 | struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set) | |
2398 | { | |
2399 | struct request_queue *uninit_q, *q; | |
2400 | ||
2401 | uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node); | |
2402 | if (!uninit_q) | |
2403 | return ERR_PTR(-ENOMEM); | |
2404 | ||
2405 | q = blk_mq_init_allocated_queue(set, uninit_q); | |
2406 | if (IS_ERR(q)) | |
2407 | blk_cleanup_queue(uninit_q); | |
2408 | ||
2409 | return q; | |
2410 | } | |
2411 | EXPORT_SYMBOL(blk_mq_init_queue); | |
2412 | ||
2413 | static int blk_mq_hw_ctx_size(struct blk_mq_tag_set *tag_set) | |
2414 | { | |
2415 | int hw_ctx_size = sizeof(struct blk_mq_hw_ctx); | |
2416 | ||
2417 | BUILD_BUG_ON(ALIGN(offsetof(struct blk_mq_hw_ctx, queue_rq_srcu), | |
2418 | __alignof__(struct blk_mq_hw_ctx)) != | |
2419 | sizeof(struct blk_mq_hw_ctx)); | |
2420 | ||
2421 | if (tag_set->flags & BLK_MQ_F_BLOCKING) | |
2422 | hw_ctx_size += sizeof(struct srcu_struct); | |
2423 | ||
2424 | return hw_ctx_size; | |
2425 | } | |
2426 | ||
2427 | static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set, | |
2428 | struct request_queue *q) | |
2429 | { | |
2430 | int i, j; | |
2431 | struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx; | |
2432 | ||
2433 | blk_mq_sysfs_unregister(q); | |
2434 | ||
2435 | /* protect against switching io scheduler */ | |
2436 | mutex_lock(&q->sysfs_lock); | |
2437 | for (i = 0; i < set->nr_hw_queues; i++) { | |
2438 | int node; | |
2439 | ||
2440 | if (hctxs[i]) | |
2441 | continue; | |
2442 | ||
2443 | node = blk_mq_hw_queue_to_node(q->mq_map, i); | |
2444 | hctxs[i] = kzalloc_node(blk_mq_hw_ctx_size(set), | |
2445 | GFP_KERNEL, node); | |
2446 | if (!hctxs[i]) | |
2447 | break; | |
2448 | ||
2449 | if (!zalloc_cpumask_var_node(&hctxs[i]->cpumask, GFP_KERNEL, | |
2450 | node)) { | |
2451 | kfree(hctxs[i]); | |
2452 | hctxs[i] = NULL; | |
2453 | break; | |
2454 | } | |
2455 | ||
2456 | atomic_set(&hctxs[i]->nr_active, 0); | |
2457 | hctxs[i]->numa_node = node; | |
2458 | hctxs[i]->queue_num = i; | |
2459 | ||
2460 | if (blk_mq_init_hctx(q, set, hctxs[i], i)) { | |
2461 | free_cpumask_var(hctxs[i]->cpumask); | |
2462 | kfree(hctxs[i]); | |
2463 | hctxs[i] = NULL; | |
2464 | break; | |
2465 | } | |
2466 | blk_mq_hctx_kobj_init(hctxs[i]); | |
2467 | } | |
2468 | for (j = i; j < q->nr_hw_queues; j++) { | |
2469 | struct blk_mq_hw_ctx *hctx = hctxs[j]; | |
2470 | ||
2471 | if (hctx) { | |
2472 | if (hctx->tags) | |
2473 | blk_mq_free_map_and_requests(set, j); | |
2474 | blk_mq_exit_hctx(q, set, hctx, j); | |
2475 | kobject_put(&hctx->kobj); | |
2476 | hctxs[j] = NULL; | |
2477 | ||
2478 | } | |
2479 | } | |
2480 | q->nr_hw_queues = i; | |
2481 | mutex_unlock(&q->sysfs_lock); | |
2482 | blk_mq_sysfs_register(q); | |
2483 | } | |
2484 | ||
2485 | struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set, | |
2486 | struct request_queue *q) | |
2487 | { | |
2488 | /* mark the queue as mq asap */ | |
2489 | q->mq_ops = set->ops; | |
2490 | ||
2491 | q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn, | |
2492 | blk_mq_poll_stats_bkt, | |
2493 | BLK_MQ_POLL_STATS_BKTS, q); | |
2494 | if (!q->poll_cb) | |
2495 | goto err_exit; | |
2496 | ||
2497 | q->queue_ctx = alloc_percpu(struct blk_mq_ctx); | |
2498 | if (!q->queue_ctx) | |
2499 | goto err_exit; | |
2500 | ||
2501 | /* init q->mq_kobj and sw queues' kobjects */ | |
2502 | blk_mq_sysfs_init(q); | |
2503 | ||
2504 | q->queue_hw_ctx = kzalloc_node(nr_cpu_ids * sizeof(*(q->queue_hw_ctx)), | |
2505 | GFP_KERNEL, set->numa_node); | |
2506 | if (!q->queue_hw_ctx) | |
2507 | goto err_percpu; | |
2508 | ||
2509 | q->mq_map = set->mq_map; | |
2510 | ||
2511 | blk_mq_realloc_hw_ctxs(set, q); | |
2512 | if (!q->nr_hw_queues) | |
2513 | goto err_hctxs; | |
2514 | ||
2515 | INIT_WORK(&q->timeout_work, blk_mq_timeout_work); | |
2516 | blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ); | |
2517 | ||
2518 | q->nr_queues = nr_cpu_ids; | |
2519 | ||
2520 | q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT; | |
2521 | ||
2522 | if (!(set->flags & BLK_MQ_F_SG_MERGE)) | |
2523 | q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE; | |
2524 | ||
2525 | q->sg_reserved_size = INT_MAX; | |
2526 | ||
2527 | INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work); | |
2528 | INIT_LIST_HEAD(&q->requeue_list); | |
2529 | spin_lock_init(&q->requeue_lock); | |
2530 | ||
2531 | blk_queue_make_request(q, blk_mq_make_request); | |
2532 | if (q->mq_ops->poll) | |
2533 | q->poll_fn = blk_mq_poll; | |
2534 | ||
2535 | /* | |
2536 | * Do this after blk_queue_make_request() overrides it... | |
2537 | */ | |
2538 | q->nr_requests = set->queue_depth; | |
2539 | ||
2540 | /* | |
2541 | * Default to classic polling | |
2542 | */ | |
2543 | q->poll_nsec = -1; | |
2544 | ||
2545 | if (set->ops->complete) | |
2546 | blk_queue_softirq_done(q, set->ops->complete); | |
2547 | ||
2548 | blk_mq_init_cpu_queues(q, set->nr_hw_queues); | |
2549 | blk_mq_add_queue_tag_set(set, q); | |
2550 | blk_mq_map_swqueue(q); | |
2551 | ||
2552 | if (!(set->flags & BLK_MQ_F_NO_SCHED)) { | |
2553 | int ret; | |
2554 | ||
2555 | ret = blk_mq_sched_init(q); | |
2556 | if (ret) | |
2557 | return ERR_PTR(ret); | |
2558 | } | |
2559 | ||
2560 | return q; | |
2561 | ||
2562 | err_hctxs: | |
2563 | kfree(q->queue_hw_ctx); | |
2564 | err_percpu: | |
2565 | free_percpu(q->queue_ctx); | |
2566 | err_exit: | |
2567 | q->mq_ops = NULL; | |
2568 | return ERR_PTR(-ENOMEM); | |
2569 | } | |
2570 | EXPORT_SYMBOL(blk_mq_init_allocated_queue); | |
2571 | ||
2572 | void blk_mq_free_queue(struct request_queue *q) | |
2573 | { | |
2574 | struct blk_mq_tag_set *set = q->tag_set; | |
2575 | ||
2576 | blk_mq_del_queue_tag_set(q); | |
2577 | blk_mq_exit_hw_queues(q, set, set->nr_hw_queues); | |
2578 | } | |
2579 | ||
2580 | /* Basically redo blk_mq_init_queue with queue frozen */ | |
2581 | static void blk_mq_queue_reinit(struct request_queue *q) | |
2582 | { | |
2583 | WARN_ON_ONCE(!atomic_read(&q->mq_freeze_depth)); | |
2584 | ||
2585 | blk_mq_debugfs_unregister_hctxs(q); | |
2586 | blk_mq_sysfs_unregister(q); | |
2587 | ||
2588 | /* | |
2589 | * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe | |
2590 | * we should change hctx numa_node according to the new topology (this | |
2591 | * involves freeing and re-allocating memory, worth doing?) | |
2592 | */ | |
2593 | blk_mq_map_swqueue(q); | |
2594 | ||
2595 | blk_mq_sysfs_register(q); | |
2596 | blk_mq_debugfs_register_hctxs(q); | |
2597 | } | |
2598 | ||
2599 | static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set) | |
2600 | { | |
2601 | int i; | |
2602 | ||
2603 | for (i = 0; i < set->nr_hw_queues; i++) | |
2604 | if (!__blk_mq_alloc_rq_map(set, i)) | |
2605 | goto out_unwind; | |
2606 | ||
2607 | return 0; | |
2608 | ||
2609 | out_unwind: | |
2610 | while (--i >= 0) | |
2611 | blk_mq_free_rq_map(set->tags[i]); | |
2612 | ||
2613 | return -ENOMEM; | |
2614 | } | |
2615 | ||
2616 | /* | |
2617 | * Allocate the request maps associated with this tag_set. Note that this | |
2618 | * may reduce the depth asked for, if memory is tight. set->queue_depth | |
2619 | * will be updated to reflect the allocated depth. | |
2620 | */ | |
2621 | static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set) | |
2622 | { | |
2623 | unsigned int depth; | |
2624 | int err; | |
2625 | ||
2626 | depth = set->queue_depth; | |
2627 | do { | |
2628 | err = __blk_mq_alloc_rq_maps(set); | |
2629 | if (!err) | |
2630 | break; | |
2631 | ||
2632 | set->queue_depth >>= 1; | |
2633 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) { | |
2634 | err = -ENOMEM; | |
2635 | break; | |
2636 | } | |
2637 | } while (set->queue_depth); | |
2638 | ||
2639 | if (!set->queue_depth || err) { | |
2640 | pr_err("blk-mq: failed to allocate request map\n"); | |
2641 | return -ENOMEM; | |
2642 | } | |
2643 | ||
2644 | if (depth != set->queue_depth) | |
2645 | pr_info("blk-mq: reduced tag depth (%u -> %u)\n", | |
2646 | depth, set->queue_depth); | |
2647 | ||
2648 | return 0; | |
2649 | } | |
2650 | ||
2651 | static int blk_mq_update_queue_map(struct blk_mq_tag_set *set) | |
2652 | { | |
2653 | if (set->ops->map_queues) { | |
2654 | int cpu; | |
2655 | /* | |
2656 | * transport .map_queues is usually done in the following | |
2657 | * way: | |
2658 | * | |
2659 | * for (queue = 0; queue < set->nr_hw_queues; queue++) { | |
2660 | * mask = get_cpu_mask(queue) | |
2661 | * for_each_cpu(cpu, mask) | |
2662 | * set->mq_map[cpu] = queue; | |
2663 | * } | |
2664 | * | |
2665 | * When we need to remap, the table has to be cleared for | |
2666 | * killing stale mapping since one CPU may not be mapped | |
2667 | * to any hw queue. | |
2668 | */ | |
2669 | for_each_possible_cpu(cpu) | |
2670 | set->mq_map[cpu] = 0; | |
2671 | ||
2672 | return set->ops->map_queues(set); | |
2673 | } else | |
2674 | return blk_mq_map_queues(set); | |
2675 | } | |
2676 | ||
2677 | /* | |
2678 | * Alloc a tag set to be associated with one or more request queues. | |
2679 | * May fail with EINVAL for various error conditions. May adjust the | |
2680 | * requested depth down, if if it too large. In that case, the set | |
2681 | * value will be stored in set->queue_depth. | |
2682 | */ | |
2683 | int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set) | |
2684 | { | |
2685 | int ret; | |
2686 | ||
2687 | BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS); | |
2688 | ||
2689 | if (!set->nr_hw_queues) | |
2690 | return -EINVAL; | |
2691 | if (!set->queue_depth) | |
2692 | return -EINVAL; | |
2693 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) | |
2694 | return -EINVAL; | |
2695 | ||
2696 | if (!set->ops->queue_rq) | |
2697 | return -EINVAL; | |
2698 | ||
2699 | if (!set->ops->get_budget ^ !set->ops->put_budget) | |
2700 | return -EINVAL; | |
2701 | ||
2702 | if (set->queue_depth > BLK_MQ_MAX_DEPTH) { | |
2703 | pr_info("blk-mq: reduced tag depth to %u\n", | |
2704 | BLK_MQ_MAX_DEPTH); | |
2705 | set->queue_depth = BLK_MQ_MAX_DEPTH; | |
2706 | } | |
2707 | ||
2708 | /* | |
2709 | * If a crashdump is active, then we are potentially in a very | |
2710 | * memory constrained environment. Limit us to 1 queue and | |
2711 | * 64 tags to prevent using too much memory. | |
2712 | */ | |
2713 | if (is_kdump_kernel()) { | |
2714 | set->nr_hw_queues = 1; | |
2715 | set->queue_depth = min(64U, set->queue_depth); | |
2716 | } | |
2717 | /* | |
2718 | * There is no use for more h/w queues than cpus. | |
2719 | */ | |
2720 | if (set->nr_hw_queues > nr_cpu_ids) | |
2721 | set->nr_hw_queues = nr_cpu_ids; | |
2722 | ||
2723 | set->tags = kzalloc_node(nr_cpu_ids * sizeof(struct blk_mq_tags *), | |
2724 | GFP_KERNEL, set->numa_node); | |
2725 | if (!set->tags) | |
2726 | return -ENOMEM; | |
2727 | ||
2728 | ret = -ENOMEM; | |
2729 | set->mq_map = kzalloc_node(sizeof(*set->mq_map) * nr_cpu_ids, | |
2730 | GFP_KERNEL, set->numa_node); | |
2731 | if (!set->mq_map) | |
2732 | goto out_free_tags; | |
2733 | ||
2734 | ret = blk_mq_update_queue_map(set); | |
2735 | if (ret) | |
2736 | goto out_free_mq_map; | |
2737 | ||
2738 | ret = blk_mq_alloc_rq_maps(set); | |
2739 | if (ret) | |
2740 | goto out_free_mq_map; | |
2741 | ||
2742 | mutex_init(&set->tag_list_lock); | |
2743 | INIT_LIST_HEAD(&set->tag_list); | |
2744 | ||
2745 | return 0; | |
2746 | ||
2747 | out_free_mq_map: | |
2748 | kfree(set->mq_map); | |
2749 | set->mq_map = NULL; | |
2750 | out_free_tags: | |
2751 | kfree(set->tags); | |
2752 | set->tags = NULL; | |
2753 | return ret; | |
2754 | } | |
2755 | EXPORT_SYMBOL(blk_mq_alloc_tag_set); | |
2756 | ||
2757 | void blk_mq_free_tag_set(struct blk_mq_tag_set *set) | |
2758 | { | |
2759 | int i; | |
2760 | ||
2761 | for (i = 0; i < nr_cpu_ids; i++) | |
2762 | blk_mq_free_map_and_requests(set, i); | |
2763 | ||
2764 | kfree(set->mq_map); | |
2765 | set->mq_map = NULL; | |
2766 | ||
2767 | kfree(set->tags); | |
2768 | set->tags = NULL; | |
2769 | } | |
2770 | EXPORT_SYMBOL(blk_mq_free_tag_set); | |
2771 | ||
2772 | int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr) | |
2773 | { | |
2774 | struct blk_mq_tag_set *set = q->tag_set; | |
2775 | struct blk_mq_hw_ctx *hctx; | |
2776 | int i, ret; | |
2777 | ||
2778 | if (!set) | |
2779 | return -EINVAL; | |
2780 | ||
2781 | blk_mq_freeze_queue(q); | |
2782 | ||
2783 | ret = 0; | |
2784 | queue_for_each_hw_ctx(q, hctx, i) { | |
2785 | if (!hctx->tags) | |
2786 | continue; | |
2787 | /* | |
2788 | * If we're using an MQ scheduler, just update the scheduler | |
2789 | * queue depth. This is similar to what the old code would do. | |
2790 | */ | |
2791 | if (!hctx->sched_tags) { | |
2792 | ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr, | |
2793 | false); | |
2794 | } else { | |
2795 | ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags, | |
2796 | nr, true); | |
2797 | } | |
2798 | if (ret) | |
2799 | break; | |
2800 | } | |
2801 | ||
2802 | if (!ret) | |
2803 | q->nr_requests = nr; | |
2804 | ||
2805 | blk_mq_unfreeze_queue(q); | |
2806 | ||
2807 | return ret; | |
2808 | } | |
2809 | ||
2810 | static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, | |
2811 | int nr_hw_queues) | |
2812 | { | |
2813 | struct request_queue *q; | |
2814 | ||
2815 | lockdep_assert_held(&set->tag_list_lock); | |
2816 | ||
2817 | if (nr_hw_queues > nr_cpu_ids) | |
2818 | nr_hw_queues = nr_cpu_ids; | |
2819 | if (nr_hw_queues < 1 || nr_hw_queues == set->nr_hw_queues) | |
2820 | return; | |
2821 | ||
2822 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
2823 | blk_mq_freeze_queue(q); | |
2824 | ||
2825 | set->nr_hw_queues = nr_hw_queues; | |
2826 | blk_mq_update_queue_map(set); | |
2827 | list_for_each_entry(q, &set->tag_list, tag_set_list) { | |
2828 | blk_mq_realloc_hw_ctxs(set, q); | |
2829 | blk_mq_queue_reinit(q); | |
2830 | } | |
2831 | ||
2832 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
2833 | blk_mq_unfreeze_queue(q); | |
2834 | } | |
2835 | ||
2836 | void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues) | |
2837 | { | |
2838 | mutex_lock(&set->tag_list_lock); | |
2839 | __blk_mq_update_nr_hw_queues(set, nr_hw_queues); | |
2840 | mutex_unlock(&set->tag_list_lock); | |
2841 | } | |
2842 | EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues); | |
2843 | ||
2844 | /* Enable polling stats and return whether they were already enabled. */ | |
2845 | static bool blk_poll_stats_enable(struct request_queue *q) | |
2846 | { | |
2847 | if (test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) || | |
2848 | test_and_set_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags)) | |
2849 | return true; | |
2850 | blk_stat_add_callback(q, q->poll_cb); | |
2851 | return false; | |
2852 | } | |
2853 | ||
2854 | static void blk_mq_poll_stats_start(struct request_queue *q) | |
2855 | { | |
2856 | /* | |
2857 | * We don't arm the callback if polling stats are not enabled or the | |
2858 | * callback is already active. | |
2859 | */ | |
2860 | if (!test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) || | |
2861 | blk_stat_is_active(q->poll_cb)) | |
2862 | return; | |
2863 | ||
2864 | blk_stat_activate_msecs(q->poll_cb, 100); | |
2865 | } | |
2866 | ||
2867 | static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb) | |
2868 | { | |
2869 | struct request_queue *q = cb->data; | |
2870 | int bucket; | |
2871 | ||
2872 | for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) { | |
2873 | if (cb->stat[bucket].nr_samples) | |
2874 | q->poll_stat[bucket] = cb->stat[bucket]; | |
2875 | } | |
2876 | } | |
2877 | ||
2878 | static unsigned long blk_mq_poll_nsecs(struct request_queue *q, | |
2879 | struct blk_mq_hw_ctx *hctx, | |
2880 | struct request *rq) | |
2881 | { | |
2882 | unsigned long ret = 0; | |
2883 | int bucket; | |
2884 | ||
2885 | /* | |
2886 | * If stats collection isn't on, don't sleep but turn it on for | |
2887 | * future users | |
2888 | */ | |
2889 | if (!blk_poll_stats_enable(q)) | |
2890 | return 0; | |
2891 | ||
2892 | /* | |
2893 | * As an optimistic guess, use half of the mean service time | |
2894 | * for this type of request. We can (and should) make this smarter. | |
2895 | * For instance, if the completion latencies are tight, we can | |
2896 | * get closer than just half the mean. This is especially | |
2897 | * important on devices where the completion latencies are longer | |
2898 | * than ~10 usec. We do use the stats for the relevant IO size | |
2899 | * if available which does lead to better estimates. | |
2900 | */ | |
2901 | bucket = blk_mq_poll_stats_bkt(rq); | |
2902 | if (bucket < 0) | |
2903 | return ret; | |
2904 | ||
2905 | if (q->poll_stat[bucket].nr_samples) | |
2906 | ret = (q->poll_stat[bucket].mean + 1) / 2; | |
2907 | ||
2908 | return ret; | |
2909 | } | |
2910 | ||
2911 | static bool blk_mq_poll_hybrid_sleep(struct request_queue *q, | |
2912 | struct blk_mq_hw_ctx *hctx, | |
2913 | struct request *rq) | |
2914 | { | |
2915 | struct hrtimer_sleeper hs; | |
2916 | enum hrtimer_mode mode; | |
2917 | unsigned int nsecs; | |
2918 | ktime_t kt; | |
2919 | ||
2920 | if (test_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags)) | |
2921 | return false; | |
2922 | ||
2923 | /* | |
2924 | * poll_nsec can be: | |
2925 | * | |
2926 | * -1: don't ever hybrid sleep | |
2927 | * 0: use half of prev avg | |
2928 | * >0: use this specific value | |
2929 | */ | |
2930 | if (q->poll_nsec == -1) | |
2931 | return false; | |
2932 | else if (q->poll_nsec > 0) | |
2933 | nsecs = q->poll_nsec; | |
2934 | else | |
2935 | nsecs = blk_mq_poll_nsecs(q, hctx, rq); | |
2936 | ||
2937 | if (!nsecs) | |
2938 | return false; | |
2939 | ||
2940 | set_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags); | |
2941 | ||
2942 | /* | |
2943 | * This will be replaced with the stats tracking code, using | |
2944 | * 'avg_completion_time / 2' as the pre-sleep target. | |
2945 | */ | |
2946 | kt = nsecs; | |
2947 | ||
2948 | mode = HRTIMER_MODE_REL; | |
2949 | hrtimer_init_on_stack(&hs.timer, CLOCK_MONOTONIC, mode); | |
2950 | hrtimer_set_expires(&hs.timer, kt); | |
2951 | ||
2952 | hrtimer_init_sleeper(&hs, current); | |
2953 | do { | |
2954 | if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags)) | |
2955 | break; | |
2956 | set_current_state(TASK_UNINTERRUPTIBLE); | |
2957 | hrtimer_start_expires(&hs.timer, mode); | |
2958 | if (hs.task) | |
2959 | io_schedule(); | |
2960 | hrtimer_cancel(&hs.timer); | |
2961 | mode = HRTIMER_MODE_ABS; | |
2962 | } while (hs.task && !signal_pending(current)); | |
2963 | ||
2964 | __set_current_state(TASK_RUNNING); | |
2965 | destroy_hrtimer_on_stack(&hs.timer); | |
2966 | return true; | |
2967 | } | |
2968 | ||
2969 | static bool __blk_mq_poll(struct blk_mq_hw_ctx *hctx, struct request *rq) | |
2970 | { | |
2971 | struct request_queue *q = hctx->queue; | |
2972 | long state; | |
2973 | ||
2974 | /* | |
2975 | * If we sleep, have the caller restart the poll loop to reset | |
2976 | * the state. Like for the other success return cases, the | |
2977 | * caller is responsible for checking if the IO completed. If | |
2978 | * the IO isn't complete, we'll get called again and will go | |
2979 | * straight to the busy poll loop. | |
2980 | */ | |
2981 | if (blk_mq_poll_hybrid_sleep(q, hctx, rq)) | |
2982 | return true; | |
2983 | ||
2984 | hctx->poll_considered++; | |
2985 | ||
2986 | state = current->state; | |
2987 | while (!need_resched()) { | |
2988 | int ret; | |
2989 | ||
2990 | hctx->poll_invoked++; | |
2991 | ||
2992 | ret = q->mq_ops->poll(hctx, rq->tag); | |
2993 | if (ret > 0) { | |
2994 | hctx->poll_success++; | |
2995 | set_current_state(TASK_RUNNING); | |
2996 | return true; | |
2997 | } | |
2998 | ||
2999 | if (signal_pending_state(state, current)) | |
3000 | set_current_state(TASK_RUNNING); | |
3001 | ||
3002 | if (current->state == TASK_RUNNING) | |
3003 | return true; | |
3004 | if (ret < 0) | |
3005 | break; | |
3006 | cpu_relax(); | |
3007 | } | |
3008 | ||
3009 | return false; | |
3010 | } | |
3011 | ||
3012 | static bool blk_mq_poll(struct request_queue *q, blk_qc_t cookie) | |
3013 | { | |
3014 | struct blk_mq_hw_ctx *hctx; | |
3015 | struct request *rq; | |
3016 | ||
3017 | if (!test_bit(QUEUE_FLAG_POLL, &q->queue_flags)) | |
3018 | return false; | |
3019 | ||
3020 | hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)]; | |
3021 | if (!blk_qc_t_is_internal(cookie)) | |
3022 | rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie)); | |
3023 | else { | |
3024 | rq = blk_mq_tag_to_rq(hctx->sched_tags, blk_qc_t_to_tag(cookie)); | |
3025 | /* | |
3026 | * With scheduling, if the request has completed, we'll | |
3027 | * get a NULL return here, as we clear the sched tag when | |
3028 | * that happens. The request still remains valid, like always, | |
3029 | * so we should be safe with just the NULL check. | |
3030 | */ | |
3031 | if (!rq) | |
3032 | return false; | |
3033 | } | |
3034 | ||
3035 | return __blk_mq_poll(hctx, rq); | |
3036 | } | |
3037 | ||
3038 | static int __init blk_mq_init(void) | |
3039 | { | |
3040 | /* | |
3041 | * See comment in block/blk.h rq_atomic_flags enum | |
3042 | */ | |
3043 | BUILD_BUG_ON((REQ_ATOM_STARTED / BITS_PER_BYTE) != | |
3044 | (REQ_ATOM_COMPLETE / BITS_PER_BYTE)); | |
3045 | ||
3046 | cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL, | |
3047 | blk_mq_hctx_notify_dead); | |
3048 | return 0; | |
3049 | } | |
3050 | subsys_initcall(blk_mq_init); |