]>
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 | static void blk_mq_check_inflight_rw(struct blk_mq_hw_ctx *hctx, | |
123 | struct request *rq, void *priv, | |
124 | bool reserved) | |
125 | { | |
126 | struct mq_inflight *mi = priv; | |
127 | ||
128 | if (rq->part == mi->part) | |
129 | mi->inflight[rq_data_dir(rq)]++; | |
130 | } | |
131 | ||
132 | void blk_mq_in_flight_rw(struct request_queue *q, struct hd_struct *part, | |
133 | unsigned int inflight[2]) | |
134 | { | |
135 | struct mq_inflight mi = { .part = part, .inflight = inflight, }; | |
136 | ||
137 | inflight[0] = inflight[1] = 0; | |
138 | blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight_rw, &mi); | |
139 | } | |
140 | ||
141 | void blk_freeze_queue_start(struct request_queue *q) | |
142 | { | |
143 | int freeze_depth; | |
144 | ||
145 | freeze_depth = atomic_inc_return(&q->mq_freeze_depth); | |
146 | if (freeze_depth == 1) { | |
147 | percpu_ref_kill(&q->q_usage_counter); | |
148 | if (q->mq_ops) | |
149 | blk_mq_run_hw_queues(q, false); | |
150 | } | |
151 | } | |
152 | EXPORT_SYMBOL_GPL(blk_freeze_queue_start); | |
153 | ||
154 | void blk_mq_freeze_queue_wait(struct request_queue *q) | |
155 | { | |
156 | wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter)); | |
157 | } | |
158 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait); | |
159 | ||
160 | int blk_mq_freeze_queue_wait_timeout(struct request_queue *q, | |
161 | unsigned long timeout) | |
162 | { | |
163 | return wait_event_timeout(q->mq_freeze_wq, | |
164 | percpu_ref_is_zero(&q->q_usage_counter), | |
165 | timeout); | |
166 | } | |
167 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout); | |
168 | ||
169 | /* | |
170 | * Guarantee no request is in use, so we can change any data structure of | |
171 | * the queue afterward. | |
172 | */ | |
173 | void blk_freeze_queue(struct request_queue *q) | |
174 | { | |
175 | /* | |
176 | * In the !blk_mq case we are only calling this to kill the | |
177 | * q_usage_counter, otherwise this increases the freeze depth | |
178 | * and waits for it to return to zero. For this reason there is | |
179 | * no blk_unfreeze_queue(), and blk_freeze_queue() is not | |
180 | * exported to drivers as the only user for unfreeze is blk_mq. | |
181 | */ | |
182 | blk_freeze_queue_start(q); | |
183 | if (!q->mq_ops) | |
184 | blk_drain_queue(q); | |
185 | blk_mq_freeze_queue_wait(q); | |
186 | } | |
187 | ||
188 | void blk_mq_freeze_queue(struct request_queue *q) | |
189 | { | |
190 | /* | |
191 | * ...just an alias to keep freeze and unfreeze actions balanced | |
192 | * in the blk_mq_* namespace | |
193 | */ | |
194 | blk_freeze_queue(q); | |
195 | } | |
196 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue); | |
197 | ||
198 | void blk_mq_unfreeze_queue(struct request_queue *q) | |
199 | { | |
200 | int freeze_depth; | |
201 | ||
202 | freeze_depth = atomic_dec_return(&q->mq_freeze_depth); | |
203 | WARN_ON_ONCE(freeze_depth < 0); | |
204 | if (!freeze_depth) { | |
205 | percpu_ref_reinit(&q->q_usage_counter); | |
206 | wake_up_all(&q->mq_freeze_wq); | |
207 | } | |
208 | } | |
209 | EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue); | |
210 | ||
211 | /* | |
212 | * FIXME: replace the scsi_internal_device_*block_nowait() calls in the | |
213 | * mpt3sas driver such that this function can be removed. | |
214 | */ | |
215 | void blk_mq_quiesce_queue_nowait(struct request_queue *q) | |
216 | { | |
217 | unsigned long flags; | |
218 | ||
219 | spin_lock_irqsave(q->queue_lock, flags); | |
220 | queue_flag_set(QUEUE_FLAG_QUIESCED, q); | |
221 | spin_unlock_irqrestore(q->queue_lock, flags); | |
222 | } | |
223 | EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait); | |
224 | ||
225 | /** | |
226 | * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished | |
227 | * @q: request queue. | |
228 | * | |
229 | * Note: this function does not prevent that the struct request end_io() | |
230 | * callback function is invoked. Once this function is returned, we make | |
231 | * sure no dispatch can happen until the queue is unquiesced via | |
232 | * blk_mq_unquiesce_queue(). | |
233 | */ | |
234 | void blk_mq_quiesce_queue(struct request_queue *q) | |
235 | { | |
236 | struct blk_mq_hw_ctx *hctx; | |
237 | unsigned int i; | |
238 | bool rcu = false; | |
239 | ||
240 | blk_mq_quiesce_queue_nowait(q); | |
241 | ||
242 | queue_for_each_hw_ctx(q, hctx, i) { | |
243 | if (hctx->flags & BLK_MQ_F_BLOCKING) | |
244 | synchronize_srcu(hctx->queue_rq_srcu); | |
245 | else | |
246 | rcu = true; | |
247 | } | |
248 | if (rcu) | |
249 | synchronize_rcu(); | |
250 | } | |
251 | EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue); | |
252 | ||
253 | /* | |
254 | * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue() | |
255 | * @q: request queue. | |
256 | * | |
257 | * This function recovers queue into the state before quiescing | |
258 | * which is done by blk_mq_quiesce_queue. | |
259 | */ | |
260 | void blk_mq_unquiesce_queue(struct request_queue *q) | |
261 | { | |
262 | unsigned long flags; | |
263 | ||
264 | spin_lock_irqsave(q->queue_lock, flags); | |
265 | queue_flag_clear(QUEUE_FLAG_QUIESCED, q); | |
266 | spin_unlock_irqrestore(q->queue_lock, flags); | |
267 | ||
268 | /* dispatch requests which are inserted during quiescing */ | |
269 | blk_mq_run_hw_queues(q, true); | |
270 | } | |
271 | EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue); | |
272 | ||
273 | void blk_mq_wake_waiters(struct request_queue *q) | |
274 | { | |
275 | struct blk_mq_hw_ctx *hctx; | |
276 | unsigned int i; | |
277 | ||
278 | queue_for_each_hw_ctx(q, hctx, i) | |
279 | if (blk_mq_hw_queue_mapped(hctx)) | |
280 | blk_mq_tag_wakeup_all(hctx->tags, true); | |
281 | } | |
282 | ||
283 | bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx) | |
284 | { | |
285 | return blk_mq_has_free_tags(hctx->tags); | |
286 | } | |
287 | EXPORT_SYMBOL(blk_mq_can_queue); | |
288 | ||
289 | static struct request *blk_mq_rq_ctx_init(struct blk_mq_alloc_data *data, | |
290 | unsigned int tag, unsigned int op) | |
291 | { | |
292 | struct blk_mq_tags *tags = blk_mq_tags_from_data(data); | |
293 | struct request *rq = tags->static_rqs[tag]; | |
294 | ||
295 | rq->rq_flags = 0; | |
296 | ||
297 | if (data->flags & BLK_MQ_REQ_INTERNAL) { | |
298 | rq->tag = -1; | |
299 | rq->internal_tag = tag; | |
300 | } else { | |
301 | if (blk_mq_tag_busy(data->hctx)) { | |
302 | rq->rq_flags = RQF_MQ_INFLIGHT; | |
303 | atomic_inc(&data->hctx->nr_active); | |
304 | } | |
305 | rq->tag = tag; | |
306 | rq->internal_tag = -1; | |
307 | data->hctx->tags->rqs[rq->tag] = rq; | |
308 | } | |
309 | ||
310 | INIT_LIST_HEAD(&rq->queuelist); | |
311 | /* csd/requeue_work/fifo_time is initialized before use */ | |
312 | rq->q = data->q; | |
313 | rq->mq_ctx = data->ctx; | |
314 | rq->cmd_flags = op; | |
315 | if (data->flags & BLK_MQ_REQ_PREEMPT) | |
316 | rq->rq_flags |= RQF_PREEMPT; | |
317 | if (blk_queue_io_stat(data->q)) | |
318 | rq->rq_flags |= RQF_IO_STAT; | |
319 | /* do not touch atomic flags, it needs atomic ops against the timer */ | |
320 | rq->cpu = -1; | |
321 | INIT_HLIST_NODE(&rq->hash); | |
322 | RB_CLEAR_NODE(&rq->rb_node); | |
323 | rq->rq_disk = NULL; | |
324 | rq->part = NULL; | |
325 | rq->start_time = jiffies; | |
326 | #ifdef CONFIG_BLK_CGROUP | |
327 | rq->rl = NULL; | |
328 | set_start_time_ns(rq); | |
329 | rq->io_start_time_ns = 0; | |
330 | #endif | |
331 | rq->nr_phys_segments = 0; | |
332 | #if defined(CONFIG_BLK_DEV_INTEGRITY) | |
333 | rq->nr_integrity_segments = 0; | |
334 | #endif | |
335 | rq->special = NULL; | |
336 | /* tag was already set */ | |
337 | rq->extra_len = 0; | |
338 | ||
339 | INIT_LIST_HEAD(&rq->timeout_list); | |
340 | rq->timeout = 0; | |
341 | ||
342 | rq->end_io = NULL; | |
343 | rq->end_io_data = NULL; | |
344 | rq->next_rq = NULL; | |
345 | ||
346 | data->ctx->rq_dispatched[op_is_sync(op)]++; | |
347 | return rq; | |
348 | } | |
349 | ||
350 | static struct request *blk_mq_get_request(struct request_queue *q, | |
351 | struct bio *bio, unsigned int op, | |
352 | struct blk_mq_alloc_data *data) | |
353 | { | |
354 | struct elevator_queue *e = q->elevator; | |
355 | struct request *rq; | |
356 | unsigned int tag; | |
357 | bool put_ctx_on_error = false; | |
358 | ||
359 | blk_queue_enter_live(q); | |
360 | data->q = q; | |
361 | if (likely(!data->ctx)) { | |
362 | data->ctx = blk_mq_get_ctx(q); | |
363 | put_ctx_on_error = true; | |
364 | } | |
365 | if (likely(!data->hctx)) | |
366 | data->hctx = blk_mq_map_queue(q, data->ctx->cpu); | |
367 | if (op & REQ_NOWAIT) | |
368 | data->flags |= BLK_MQ_REQ_NOWAIT; | |
369 | ||
370 | if (e) { | |
371 | data->flags |= BLK_MQ_REQ_INTERNAL; | |
372 | ||
373 | /* | |
374 | * Flush requests are special and go directly to the | |
375 | * dispatch list. | |
376 | */ | |
377 | if (!op_is_flush(op) && e->type->ops.mq.limit_depth) | |
378 | e->type->ops.mq.limit_depth(op, data); | |
379 | } | |
380 | ||
381 | tag = blk_mq_get_tag(data); | |
382 | if (tag == BLK_MQ_TAG_FAIL) { | |
383 | if (put_ctx_on_error) { | |
384 | blk_mq_put_ctx(data->ctx); | |
385 | data->ctx = NULL; | |
386 | } | |
387 | blk_queue_exit(q); | |
388 | return NULL; | |
389 | } | |
390 | ||
391 | rq = blk_mq_rq_ctx_init(data, tag, op); | |
392 | if (!op_is_flush(op)) { | |
393 | rq->elv.icq = NULL; | |
394 | if (e && e->type->ops.mq.prepare_request) { | |
395 | if (e->type->icq_cache && rq_ioc(bio)) | |
396 | blk_mq_sched_assign_ioc(rq, bio); | |
397 | ||
398 | e->type->ops.mq.prepare_request(rq, bio); | |
399 | rq->rq_flags |= RQF_ELVPRIV; | |
400 | } | |
401 | } | |
402 | data->hctx->queued++; | |
403 | return rq; | |
404 | } | |
405 | ||
406 | struct request *blk_mq_alloc_request(struct request_queue *q, unsigned int op, | |
407 | blk_mq_req_flags_t flags) | |
408 | { | |
409 | struct blk_mq_alloc_data alloc_data = { .flags = flags }; | |
410 | struct request *rq; | |
411 | int ret; | |
412 | ||
413 | ret = blk_queue_enter(q, flags); | |
414 | if (ret) | |
415 | return ERR_PTR(ret); | |
416 | ||
417 | rq = blk_mq_get_request(q, NULL, op, &alloc_data); | |
418 | blk_queue_exit(q); | |
419 | ||
420 | if (!rq) | |
421 | return ERR_PTR(-EWOULDBLOCK); | |
422 | ||
423 | blk_mq_put_ctx(alloc_data.ctx); | |
424 | ||
425 | rq->__data_len = 0; | |
426 | rq->__sector = (sector_t) -1; | |
427 | rq->bio = rq->biotail = NULL; | |
428 | return rq; | |
429 | } | |
430 | EXPORT_SYMBOL(blk_mq_alloc_request); | |
431 | ||
432 | struct request *blk_mq_alloc_request_hctx(struct request_queue *q, | |
433 | unsigned int op, blk_mq_req_flags_t flags, unsigned int hctx_idx) | |
434 | { | |
435 | struct blk_mq_alloc_data alloc_data = { .flags = flags }; | |
436 | struct request *rq; | |
437 | unsigned int cpu; | |
438 | int ret; | |
439 | ||
440 | /* | |
441 | * If the tag allocator sleeps we could get an allocation for a | |
442 | * different hardware context. No need to complicate the low level | |
443 | * allocator for this for the rare use case of a command tied to | |
444 | * a specific queue. | |
445 | */ | |
446 | if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT))) | |
447 | return ERR_PTR(-EINVAL); | |
448 | ||
449 | if (hctx_idx >= q->nr_hw_queues) | |
450 | return ERR_PTR(-EIO); | |
451 | ||
452 | ret = blk_queue_enter(q, flags); | |
453 | if (ret) | |
454 | return ERR_PTR(ret); | |
455 | ||
456 | /* | |
457 | * Check if the hardware context is actually mapped to anything. | |
458 | * If not tell the caller that it should skip this queue. | |
459 | */ | |
460 | alloc_data.hctx = q->queue_hw_ctx[hctx_idx]; | |
461 | if (!blk_mq_hw_queue_mapped(alloc_data.hctx)) { | |
462 | blk_queue_exit(q); | |
463 | return ERR_PTR(-EXDEV); | |
464 | } | |
465 | cpu = cpumask_first_and(alloc_data.hctx->cpumask, cpu_online_mask); | |
466 | alloc_data.ctx = __blk_mq_get_ctx(q, cpu); | |
467 | ||
468 | rq = blk_mq_get_request(q, NULL, op, &alloc_data); | |
469 | blk_queue_exit(q); | |
470 | ||
471 | if (!rq) | |
472 | return ERR_PTR(-EWOULDBLOCK); | |
473 | ||
474 | return rq; | |
475 | } | |
476 | EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx); | |
477 | ||
478 | void blk_mq_free_request(struct request *rq) | |
479 | { | |
480 | struct request_queue *q = rq->q; | |
481 | struct elevator_queue *e = q->elevator; | |
482 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
483 | struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, ctx->cpu); | |
484 | const int sched_tag = rq->internal_tag; | |
485 | ||
486 | if (rq->rq_flags & RQF_ELVPRIV) { | |
487 | if (e && e->type->ops.mq.finish_request) | |
488 | e->type->ops.mq.finish_request(rq); | |
489 | if (rq->elv.icq) { | |
490 | put_io_context(rq->elv.icq->ioc); | |
491 | rq->elv.icq = NULL; | |
492 | } | |
493 | } | |
494 | ||
495 | ctx->rq_completed[rq_is_sync(rq)]++; | |
496 | if (rq->rq_flags & RQF_MQ_INFLIGHT) | |
497 | atomic_dec(&hctx->nr_active); | |
498 | ||
499 | if (unlikely(laptop_mode && !blk_rq_is_passthrough(rq))) | |
500 | laptop_io_completion(q->backing_dev_info); | |
501 | ||
502 | wbt_done(q->rq_wb, &rq->issue_stat); | |
503 | ||
504 | if (blk_rq_rl(rq)) | |
505 | blk_put_rl(blk_rq_rl(rq)); | |
506 | ||
507 | clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags); | |
508 | clear_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags); | |
509 | if (rq->tag != -1) | |
510 | blk_mq_put_tag(hctx, hctx->tags, ctx, rq->tag); | |
511 | if (sched_tag != -1) | |
512 | blk_mq_put_tag(hctx, hctx->sched_tags, ctx, sched_tag); | |
513 | blk_mq_sched_restart(hctx); | |
514 | blk_queue_exit(q); | |
515 | } | |
516 | EXPORT_SYMBOL_GPL(blk_mq_free_request); | |
517 | ||
518 | inline void __blk_mq_end_request(struct request *rq, blk_status_t error) | |
519 | { | |
520 | blk_account_io_done(rq); | |
521 | ||
522 | if (rq->end_io) { | |
523 | wbt_done(rq->q->rq_wb, &rq->issue_stat); | |
524 | rq->end_io(rq, error); | |
525 | } else { | |
526 | if (unlikely(blk_bidi_rq(rq))) | |
527 | blk_mq_free_request(rq->next_rq); | |
528 | blk_mq_free_request(rq); | |
529 | } | |
530 | } | |
531 | EXPORT_SYMBOL(__blk_mq_end_request); | |
532 | ||
533 | void blk_mq_end_request(struct request *rq, blk_status_t error) | |
534 | { | |
535 | if (blk_update_request(rq, error, blk_rq_bytes(rq))) | |
536 | BUG(); | |
537 | __blk_mq_end_request(rq, error); | |
538 | } | |
539 | EXPORT_SYMBOL(blk_mq_end_request); | |
540 | ||
541 | static void __blk_mq_complete_request_remote(void *data) | |
542 | { | |
543 | struct request *rq = data; | |
544 | ||
545 | rq->q->softirq_done_fn(rq); | |
546 | } | |
547 | ||
548 | static void __blk_mq_complete_request(struct request *rq) | |
549 | { | |
550 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
551 | bool shared = false; | |
552 | int cpu; | |
553 | ||
554 | if (rq->internal_tag != -1) | |
555 | blk_mq_sched_completed_request(rq); | |
556 | if (rq->rq_flags & RQF_STATS) { | |
557 | blk_mq_poll_stats_start(rq->q); | |
558 | blk_stat_add(rq); | |
559 | } | |
560 | ||
561 | if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) { | |
562 | rq->q->softirq_done_fn(rq); | |
563 | return; | |
564 | } | |
565 | ||
566 | cpu = get_cpu(); | |
567 | if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags)) | |
568 | shared = cpus_share_cache(cpu, ctx->cpu); | |
569 | ||
570 | if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) { | |
571 | rq->csd.func = __blk_mq_complete_request_remote; | |
572 | rq->csd.info = rq; | |
573 | rq->csd.flags = 0; | |
574 | smp_call_function_single_async(ctx->cpu, &rq->csd); | |
575 | } else { | |
576 | rq->q->softirq_done_fn(rq); | |
577 | } | |
578 | put_cpu(); | |
579 | } | |
580 | ||
581 | /** | |
582 | * blk_mq_complete_request - end I/O on a request | |
583 | * @rq: the request being processed | |
584 | * | |
585 | * Description: | |
586 | * Ends all I/O on a request. It does not handle partial completions. | |
587 | * The actual completion happens out-of-order, through a IPI handler. | |
588 | **/ | |
589 | void blk_mq_complete_request(struct request *rq) | |
590 | { | |
591 | struct request_queue *q = rq->q; | |
592 | ||
593 | if (unlikely(blk_should_fake_timeout(q))) | |
594 | return; | |
595 | if (!blk_mark_rq_complete(rq)) | |
596 | __blk_mq_complete_request(rq); | |
597 | } | |
598 | EXPORT_SYMBOL(blk_mq_complete_request); | |
599 | ||
600 | int blk_mq_request_started(struct request *rq) | |
601 | { | |
602 | return test_bit(REQ_ATOM_STARTED, &rq->atomic_flags); | |
603 | } | |
604 | EXPORT_SYMBOL_GPL(blk_mq_request_started); | |
605 | ||
606 | void blk_mq_start_request(struct request *rq) | |
607 | { | |
608 | struct request_queue *q = rq->q; | |
609 | ||
610 | blk_mq_sched_started_request(rq); | |
611 | ||
612 | trace_block_rq_issue(q, rq); | |
613 | ||
614 | if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) { | |
615 | blk_stat_set_issue(&rq->issue_stat, blk_rq_sectors(rq)); | |
616 | rq->rq_flags |= RQF_STATS; | |
617 | wbt_issue(q->rq_wb, &rq->issue_stat); | |
618 | } | |
619 | ||
620 | blk_add_timer(rq); | |
621 | ||
622 | WARN_ON_ONCE(test_bit(REQ_ATOM_STARTED, &rq->atomic_flags)); | |
623 | ||
624 | /* | |
625 | * Mark us as started and clear complete. Complete might have been | |
626 | * set if requeue raced with timeout, which then marked it as | |
627 | * complete. So be sure to clear complete again when we start | |
628 | * the request, otherwise we'll ignore the completion event. | |
629 | * | |
630 | * Ensure that ->deadline is visible before we set STARTED, such that | |
631 | * blk_mq_check_expired() is guaranteed to observe our ->deadline when | |
632 | * it observes STARTED. | |
633 | */ | |
634 | smp_wmb(); | |
635 | set_bit(REQ_ATOM_STARTED, &rq->atomic_flags); | |
636 | if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags)) { | |
637 | /* | |
638 | * Coherence order guarantees these consecutive stores to a | |
639 | * single variable propagate in the specified order. Thus the | |
640 | * clear_bit() is ordered _after_ the set bit. See | |
641 | * blk_mq_check_expired(). | |
642 | * | |
643 | * (the bits must be part of the same byte for this to be | |
644 | * true). | |
645 | */ | |
646 | clear_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags); | |
647 | } | |
648 | ||
649 | if (q->dma_drain_size && blk_rq_bytes(rq)) { | |
650 | /* | |
651 | * Make sure space for the drain appears. We know we can do | |
652 | * this because max_hw_segments has been adjusted to be one | |
653 | * fewer than the device can handle. | |
654 | */ | |
655 | rq->nr_phys_segments++; | |
656 | } | |
657 | } | |
658 | EXPORT_SYMBOL(blk_mq_start_request); | |
659 | ||
660 | /* | |
661 | * When we reach here because queue is busy, REQ_ATOM_COMPLETE | |
662 | * flag isn't set yet, so there may be race with timeout handler, | |
663 | * but given rq->deadline is just set in .queue_rq() under | |
664 | * this situation, the race won't be possible in reality because | |
665 | * rq->timeout should be set as big enough to cover the window | |
666 | * between blk_mq_start_request() called from .queue_rq() and | |
667 | * clearing REQ_ATOM_STARTED here. | |
668 | */ | |
669 | static void __blk_mq_requeue_request(struct request *rq) | |
670 | { | |
671 | struct request_queue *q = rq->q; | |
672 | ||
673 | blk_mq_put_driver_tag(rq); | |
674 | ||
675 | trace_block_rq_requeue(q, rq); | |
676 | wbt_requeue(q->rq_wb, &rq->issue_stat); | |
677 | ||
678 | if (test_and_clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) { | |
679 | if (q->dma_drain_size && blk_rq_bytes(rq)) | |
680 | rq->nr_phys_segments--; | |
681 | } | |
682 | } | |
683 | ||
684 | void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list) | |
685 | { | |
686 | __blk_mq_requeue_request(rq); | |
687 | ||
688 | /* this request will be re-inserted to io scheduler queue */ | |
689 | blk_mq_sched_requeue_request(rq); | |
690 | ||
691 | BUG_ON(blk_queued_rq(rq)); | |
692 | blk_mq_add_to_requeue_list(rq, true, kick_requeue_list); | |
693 | } | |
694 | EXPORT_SYMBOL(blk_mq_requeue_request); | |
695 | ||
696 | static void blk_mq_requeue_work(struct work_struct *work) | |
697 | { | |
698 | struct request_queue *q = | |
699 | container_of(work, struct request_queue, requeue_work.work); | |
700 | LIST_HEAD(rq_list); | |
701 | struct request *rq, *next; | |
702 | ||
703 | spin_lock_irq(&q->requeue_lock); | |
704 | list_splice_init(&q->requeue_list, &rq_list); | |
705 | spin_unlock_irq(&q->requeue_lock); | |
706 | ||
707 | list_for_each_entry_safe(rq, next, &rq_list, queuelist) { | |
708 | if (!(rq->rq_flags & RQF_SOFTBARRIER)) | |
709 | continue; | |
710 | ||
711 | rq->rq_flags &= ~RQF_SOFTBARRIER; | |
712 | list_del_init(&rq->queuelist); | |
713 | blk_mq_sched_insert_request(rq, true, false, false, true); | |
714 | } | |
715 | ||
716 | while (!list_empty(&rq_list)) { | |
717 | rq = list_entry(rq_list.next, struct request, queuelist); | |
718 | list_del_init(&rq->queuelist); | |
719 | blk_mq_sched_insert_request(rq, false, false, false, true); | |
720 | } | |
721 | ||
722 | blk_mq_run_hw_queues(q, false); | |
723 | } | |
724 | ||
725 | void blk_mq_add_to_requeue_list(struct request *rq, bool at_head, | |
726 | bool kick_requeue_list) | |
727 | { | |
728 | struct request_queue *q = rq->q; | |
729 | unsigned long flags; | |
730 | ||
731 | /* | |
732 | * We abuse this flag that is otherwise used by the I/O scheduler to | |
733 | * request head insertion from the workqueue. | |
734 | */ | |
735 | BUG_ON(rq->rq_flags & RQF_SOFTBARRIER); | |
736 | ||
737 | spin_lock_irqsave(&q->requeue_lock, flags); | |
738 | if (at_head) { | |
739 | rq->rq_flags |= RQF_SOFTBARRIER; | |
740 | list_add(&rq->queuelist, &q->requeue_list); | |
741 | } else { | |
742 | list_add_tail(&rq->queuelist, &q->requeue_list); | |
743 | } | |
744 | spin_unlock_irqrestore(&q->requeue_lock, flags); | |
745 | ||
746 | if (kick_requeue_list) | |
747 | blk_mq_kick_requeue_list(q); | |
748 | } | |
749 | EXPORT_SYMBOL(blk_mq_add_to_requeue_list); | |
750 | ||
751 | void blk_mq_kick_requeue_list(struct request_queue *q) | |
752 | { | |
753 | kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0); | |
754 | } | |
755 | EXPORT_SYMBOL(blk_mq_kick_requeue_list); | |
756 | ||
757 | void blk_mq_delay_kick_requeue_list(struct request_queue *q, | |
758 | unsigned long msecs) | |
759 | { | |
760 | kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, | |
761 | msecs_to_jiffies(msecs)); | |
762 | } | |
763 | EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list); | |
764 | ||
765 | struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag) | |
766 | { | |
767 | if (tag < tags->nr_tags) { | |
768 | prefetch(tags->rqs[tag]); | |
769 | return tags->rqs[tag]; | |
770 | } | |
771 | ||
772 | return NULL; | |
773 | } | |
774 | EXPORT_SYMBOL(blk_mq_tag_to_rq); | |
775 | ||
776 | struct blk_mq_timeout_data { | |
777 | unsigned long next; | |
778 | unsigned int next_set; | |
779 | }; | |
780 | ||
781 | void blk_mq_rq_timed_out(struct request *req, bool reserved) | |
782 | { | |
783 | const struct blk_mq_ops *ops = req->q->mq_ops; | |
784 | enum blk_eh_timer_return ret = BLK_EH_RESET_TIMER; | |
785 | ||
786 | /* | |
787 | * We know that complete is set at this point. If STARTED isn't set | |
788 | * anymore, then the request isn't active and the "timeout" should | |
789 | * just be ignored. This can happen due to the bitflag ordering. | |
790 | * Timeout first checks if STARTED is set, and if it is, assumes | |
791 | * the request is active. But if we race with completion, then | |
792 | * both flags will get cleared. So check here again, and ignore | |
793 | * a timeout event with a request that isn't active. | |
794 | */ | |
795 | if (!test_bit(REQ_ATOM_STARTED, &req->atomic_flags)) | |
796 | return; | |
797 | ||
798 | if (ops->timeout) | |
799 | ret = ops->timeout(req, reserved); | |
800 | ||
801 | switch (ret) { | |
802 | case BLK_EH_HANDLED: | |
803 | __blk_mq_complete_request(req); | |
804 | break; | |
805 | case BLK_EH_RESET_TIMER: | |
806 | blk_add_timer(req); | |
807 | blk_clear_rq_complete(req); | |
808 | break; | |
809 | case BLK_EH_NOT_HANDLED: | |
810 | break; | |
811 | default: | |
812 | printk(KERN_ERR "block: bad eh return: %d\n", ret); | |
813 | break; | |
814 | } | |
815 | } | |
816 | ||
817 | static void blk_mq_check_expired(struct blk_mq_hw_ctx *hctx, | |
818 | struct request *rq, void *priv, bool reserved) | |
819 | { | |
820 | struct blk_mq_timeout_data *data = priv; | |
821 | unsigned long deadline; | |
822 | ||
823 | if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) | |
824 | return; | |
825 | ||
826 | /* | |
827 | * Ensures that if we see STARTED we must also see our | |
828 | * up-to-date deadline, see blk_mq_start_request(). | |
829 | */ | |
830 | smp_rmb(); | |
831 | ||
832 | deadline = READ_ONCE(rq->deadline); | |
833 | ||
834 | /* | |
835 | * The rq being checked may have been freed and reallocated | |
836 | * out already here, we avoid this race by checking rq->deadline | |
837 | * and REQ_ATOM_COMPLETE flag together: | |
838 | * | |
839 | * - if rq->deadline is observed as new value because of | |
840 | * reusing, the rq won't be timed out because of timing. | |
841 | * - if rq->deadline is observed as previous value, | |
842 | * REQ_ATOM_COMPLETE flag won't be cleared in reuse path | |
843 | * because we put a barrier between setting rq->deadline | |
844 | * and clearing the flag in blk_mq_start_request(), so | |
845 | * this rq won't be timed out too. | |
846 | */ | |
847 | if (time_after_eq(jiffies, deadline)) { | |
848 | if (!blk_mark_rq_complete(rq)) { | |
849 | /* | |
850 | * Again coherence order ensures that consecutive reads | |
851 | * from the same variable must be in that order. This | |
852 | * ensures that if we see COMPLETE clear, we must then | |
853 | * see STARTED set and we'll ignore this timeout. | |
854 | * | |
855 | * (There's also the MB implied by the test_and_clear()) | |
856 | */ | |
857 | blk_mq_rq_timed_out(rq, reserved); | |
858 | } | |
859 | } else if (!data->next_set || time_after(data->next, deadline)) { | |
860 | data->next = deadline; | |
861 | data->next_set = 1; | |
862 | } | |
863 | } | |
864 | ||
865 | static void blk_mq_timeout_work(struct work_struct *work) | |
866 | { | |
867 | struct request_queue *q = | |
868 | container_of(work, struct request_queue, timeout_work); | |
869 | struct blk_mq_timeout_data data = { | |
870 | .next = 0, | |
871 | .next_set = 0, | |
872 | }; | |
873 | int i; | |
874 | ||
875 | /* A deadlock might occur if a request is stuck requiring a | |
876 | * timeout at the same time a queue freeze is waiting | |
877 | * completion, since the timeout code would not be able to | |
878 | * acquire the queue reference here. | |
879 | * | |
880 | * That's why we don't use blk_queue_enter here; instead, we use | |
881 | * percpu_ref_tryget directly, because we need to be able to | |
882 | * obtain a reference even in the short window between the queue | |
883 | * starting to freeze, by dropping the first reference in | |
884 | * blk_freeze_queue_start, and the moment the last request is | |
885 | * consumed, marked by the instant q_usage_counter reaches | |
886 | * zero. | |
887 | */ | |
888 | if (!percpu_ref_tryget(&q->q_usage_counter)) | |
889 | return; | |
890 | ||
891 | blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &data); | |
892 | ||
893 | if (data.next_set) { | |
894 | data.next = blk_rq_timeout(round_jiffies_up(data.next)); | |
895 | mod_timer(&q->timeout, data.next); | |
896 | } else { | |
897 | struct blk_mq_hw_ctx *hctx; | |
898 | ||
899 | queue_for_each_hw_ctx(q, hctx, i) { | |
900 | /* the hctx may be unmapped, so check it here */ | |
901 | if (blk_mq_hw_queue_mapped(hctx)) | |
902 | blk_mq_tag_idle(hctx); | |
903 | } | |
904 | } | |
905 | blk_queue_exit(q); | |
906 | } | |
907 | ||
908 | struct flush_busy_ctx_data { | |
909 | struct blk_mq_hw_ctx *hctx; | |
910 | struct list_head *list; | |
911 | }; | |
912 | ||
913 | static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data) | |
914 | { | |
915 | struct flush_busy_ctx_data *flush_data = data; | |
916 | struct blk_mq_hw_ctx *hctx = flush_data->hctx; | |
917 | struct blk_mq_ctx *ctx = hctx->ctxs[bitnr]; | |
918 | ||
919 | sbitmap_clear_bit(sb, bitnr); | |
920 | spin_lock(&ctx->lock); | |
921 | list_splice_tail_init(&ctx->rq_list, flush_data->list); | |
922 | spin_unlock(&ctx->lock); | |
923 | return true; | |
924 | } | |
925 | ||
926 | /* | |
927 | * Process software queues that have been marked busy, splicing them | |
928 | * to the for-dispatch | |
929 | */ | |
930 | void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list) | |
931 | { | |
932 | struct flush_busy_ctx_data data = { | |
933 | .hctx = hctx, | |
934 | .list = list, | |
935 | }; | |
936 | ||
937 | sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data); | |
938 | } | |
939 | EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs); | |
940 | ||
941 | struct dispatch_rq_data { | |
942 | struct blk_mq_hw_ctx *hctx; | |
943 | struct request *rq; | |
944 | }; | |
945 | ||
946 | static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr, | |
947 | void *data) | |
948 | { | |
949 | struct dispatch_rq_data *dispatch_data = data; | |
950 | struct blk_mq_hw_ctx *hctx = dispatch_data->hctx; | |
951 | struct blk_mq_ctx *ctx = hctx->ctxs[bitnr]; | |
952 | ||
953 | spin_lock(&ctx->lock); | |
954 | if (unlikely(!list_empty(&ctx->rq_list))) { | |
955 | dispatch_data->rq = list_entry_rq(ctx->rq_list.next); | |
956 | list_del_init(&dispatch_data->rq->queuelist); | |
957 | if (list_empty(&ctx->rq_list)) | |
958 | sbitmap_clear_bit(sb, bitnr); | |
959 | } | |
960 | spin_unlock(&ctx->lock); | |
961 | ||
962 | return !dispatch_data->rq; | |
963 | } | |
964 | ||
965 | struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx, | |
966 | struct blk_mq_ctx *start) | |
967 | { | |
968 | unsigned off = start ? start->index_hw : 0; | |
969 | struct dispatch_rq_data data = { | |
970 | .hctx = hctx, | |
971 | .rq = NULL, | |
972 | }; | |
973 | ||
974 | __sbitmap_for_each_set(&hctx->ctx_map, off, | |
975 | dispatch_rq_from_ctx, &data); | |
976 | ||
977 | return data.rq; | |
978 | } | |
979 | ||
980 | static inline unsigned int queued_to_index(unsigned int queued) | |
981 | { | |
982 | if (!queued) | |
983 | return 0; | |
984 | ||
985 | return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1); | |
986 | } | |
987 | ||
988 | bool blk_mq_get_driver_tag(struct request *rq, struct blk_mq_hw_ctx **hctx, | |
989 | bool wait) | |
990 | { | |
991 | struct blk_mq_alloc_data data = { | |
992 | .q = rq->q, | |
993 | .hctx = blk_mq_map_queue(rq->q, rq->mq_ctx->cpu), | |
994 | .flags = wait ? 0 : BLK_MQ_REQ_NOWAIT, | |
995 | }; | |
996 | ||
997 | might_sleep_if(wait); | |
998 | ||
999 | if (rq->tag != -1) | |
1000 | goto done; | |
1001 | ||
1002 | if (blk_mq_tag_is_reserved(data.hctx->sched_tags, rq->internal_tag)) | |
1003 | data.flags |= BLK_MQ_REQ_RESERVED; | |
1004 | ||
1005 | rq->tag = blk_mq_get_tag(&data); | |
1006 | if (rq->tag >= 0) { | |
1007 | if (blk_mq_tag_busy(data.hctx)) { | |
1008 | rq->rq_flags |= RQF_MQ_INFLIGHT; | |
1009 | atomic_inc(&data.hctx->nr_active); | |
1010 | } | |
1011 | data.hctx->tags->rqs[rq->tag] = rq; | |
1012 | } | |
1013 | ||
1014 | done: | |
1015 | if (hctx) | |
1016 | *hctx = data.hctx; | |
1017 | return rq->tag != -1; | |
1018 | } | |
1019 | ||
1020 | static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode, | |
1021 | int flags, void *key) | |
1022 | { | |
1023 | struct blk_mq_hw_ctx *hctx; | |
1024 | ||
1025 | hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait); | |
1026 | ||
1027 | list_del_init(&wait->entry); | |
1028 | blk_mq_run_hw_queue(hctx, true); | |
1029 | return 1; | |
1030 | } | |
1031 | ||
1032 | /* | |
1033 | * Mark us waiting for a tag. For shared tags, this involves hooking us into | |
1034 | * the tag wakeups. For non-shared tags, we can simply mark us nedeing a | |
1035 | * restart. For both caes, take care to check the condition again after | |
1036 | * marking us as waiting. | |
1037 | */ | |
1038 | static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx **hctx, | |
1039 | struct request *rq) | |
1040 | { | |
1041 | struct blk_mq_hw_ctx *this_hctx = *hctx; | |
1042 | bool shared_tags = (this_hctx->flags & BLK_MQ_F_TAG_SHARED) != 0; | |
1043 | struct sbq_wait_state *ws; | |
1044 | wait_queue_entry_t *wait; | |
1045 | bool ret; | |
1046 | ||
1047 | if (!shared_tags) { | |
1048 | if (!test_bit(BLK_MQ_S_SCHED_RESTART, &this_hctx->state)) | |
1049 | set_bit(BLK_MQ_S_SCHED_RESTART, &this_hctx->state); | |
1050 | } else { | |
1051 | wait = &this_hctx->dispatch_wait; | |
1052 | if (!list_empty_careful(&wait->entry)) | |
1053 | return false; | |
1054 | ||
1055 | spin_lock(&this_hctx->lock); | |
1056 | if (!list_empty(&wait->entry)) { | |
1057 | spin_unlock(&this_hctx->lock); | |
1058 | return false; | |
1059 | } | |
1060 | ||
1061 | ws = bt_wait_ptr(&this_hctx->tags->bitmap_tags, this_hctx); | |
1062 | add_wait_queue(&ws->wait, wait); | |
1063 | } | |
1064 | ||
1065 | /* | |
1066 | * It's possible that a tag was freed in the window between the | |
1067 | * allocation failure and adding the hardware queue to the wait | |
1068 | * queue. | |
1069 | */ | |
1070 | ret = blk_mq_get_driver_tag(rq, hctx, false); | |
1071 | ||
1072 | if (!shared_tags) { | |
1073 | /* | |
1074 | * Don't clear RESTART here, someone else could have set it. | |
1075 | * At most this will cost an extra queue run. | |
1076 | */ | |
1077 | return ret; | |
1078 | } else { | |
1079 | if (!ret) { | |
1080 | spin_unlock(&this_hctx->lock); | |
1081 | return false; | |
1082 | } | |
1083 | ||
1084 | /* | |
1085 | * We got a tag, remove ourselves from the wait queue to ensure | |
1086 | * someone else gets the wakeup. | |
1087 | */ | |
1088 | spin_lock_irq(&ws->wait.lock); | |
1089 | list_del_init(&wait->entry); | |
1090 | spin_unlock_irq(&ws->wait.lock); | |
1091 | spin_unlock(&this_hctx->lock); | |
1092 | return true; | |
1093 | } | |
1094 | } | |
1095 | ||
1096 | bool blk_mq_dispatch_rq_list(struct request_queue *q, struct list_head *list, | |
1097 | bool got_budget) | |
1098 | { | |
1099 | struct blk_mq_hw_ctx *hctx; | |
1100 | struct request *rq, *nxt; | |
1101 | bool no_tag = false; | |
1102 | int errors, queued; | |
1103 | ||
1104 | if (list_empty(list)) | |
1105 | return false; | |
1106 | ||
1107 | WARN_ON(!list_is_singular(list) && got_budget); | |
1108 | ||
1109 | /* | |
1110 | * Now process all the entries, sending them to the driver. | |
1111 | */ | |
1112 | errors = queued = 0; | |
1113 | do { | |
1114 | struct blk_mq_queue_data bd; | |
1115 | blk_status_t ret; | |
1116 | ||
1117 | rq = list_first_entry(list, struct request, queuelist); | |
1118 | ||
1119 | hctx = blk_mq_map_queue(rq->q, rq->mq_ctx->cpu); | |
1120 | if (!got_budget && !blk_mq_get_dispatch_budget(hctx)) | |
1121 | break; | |
1122 | ||
1123 | if (!blk_mq_get_driver_tag(rq, NULL, false)) { | |
1124 | /* | |
1125 | * The initial allocation attempt failed, so we need to | |
1126 | * rerun the hardware queue when a tag is freed. The | |
1127 | * waitqueue takes care of that. If the queue is run | |
1128 | * before we add this entry back on the dispatch list, | |
1129 | * we'll re-run it below. | |
1130 | */ | |
1131 | if (!blk_mq_mark_tag_wait(&hctx, rq)) { | |
1132 | blk_mq_put_dispatch_budget(hctx); | |
1133 | /* | |
1134 | * For non-shared tags, the RESTART check | |
1135 | * will suffice. | |
1136 | */ | |
1137 | if (hctx->flags & BLK_MQ_F_TAG_SHARED) | |
1138 | no_tag = true; | |
1139 | break; | |
1140 | } | |
1141 | } | |
1142 | ||
1143 | list_del_init(&rq->queuelist); | |
1144 | ||
1145 | bd.rq = rq; | |
1146 | ||
1147 | /* | |
1148 | * Flag last if we have no more requests, or if we have more | |
1149 | * but can't assign a driver tag to it. | |
1150 | */ | |
1151 | if (list_empty(list)) | |
1152 | bd.last = true; | |
1153 | else { | |
1154 | nxt = list_first_entry(list, struct request, queuelist); | |
1155 | bd.last = !blk_mq_get_driver_tag(nxt, NULL, false); | |
1156 | } | |
1157 | ||
1158 | ret = q->mq_ops->queue_rq(hctx, &bd); | |
1159 | if (ret == BLK_STS_RESOURCE) { | |
1160 | /* | |
1161 | * If an I/O scheduler has been configured and we got a | |
1162 | * driver tag for the next request already, free it | |
1163 | * again. | |
1164 | */ | |
1165 | if (!list_empty(list)) { | |
1166 | nxt = list_first_entry(list, struct request, queuelist); | |
1167 | blk_mq_put_driver_tag(nxt); | |
1168 | } | |
1169 | list_add(&rq->queuelist, list); | |
1170 | __blk_mq_requeue_request(rq); | |
1171 | break; | |
1172 | } | |
1173 | ||
1174 | if (unlikely(ret != BLK_STS_OK)) { | |
1175 | errors++; | |
1176 | blk_mq_end_request(rq, BLK_STS_IOERR); | |
1177 | continue; | |
1178 | } | |
1179 | ||
1180 | queued++; | |
1181 | } while (!list_empty(list)); | |
1182 | ||
1183 | hctx->dispatched[queued_to_index(queued)]++; | |
1184 | ||
1185 | /* | |
1186 | * Any items that need requeuing? Stuff them into hctx->dispatch, | |
1187 | * that is where we will continue on next queue run. | |
1188 | */ | |
1189 | if (!list_empty(list)) { | |
1190 | spin_lock(&hctx->lock); | |
1191 | list_splice_init(list, &hctx->dispatch); | |
1192 | spin_unlock(&hctx->lock); | |
1193 | ||
1194 | /* | |
1195 | * If SCHED_RESTART was set by the caller of this function and | |
1196 | * it is no longer set that means that it was cleared by another | |
1197 | * thread and hence that a queue rerun is needed. | |
1198 | * | |
1199 | * If 'no_tag' is set, that means that we failed getting | |
1200 | * a driver tag with an I/O scheduler attached. If our dispatch | |
1201 | * waitqueue is no longer active, ensure that we run the queue | |
1202 | * AFTER adding our entries back to the list. | |
1203 | * | |
1204 | * If no I/O scheduler has been configured it is possible that | |
1205 | * the hardware queue got stopped and restarted before requests | |
1206 | * were pushed back onto the dispatch list. Rerun the queue to | |
1207 | * avoid starvation. Notes: | |
1208 | * - blk_mq_run_hw_queue() checks whether or not a queue has | |
1209 | * been stopped before rerunning a queue. | |
1210 | * - Some but not all block drivers stop a queue before | |
1211 | * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq | |
1212 | * and dm-rq. | |
1213 | */ | |
1214 | if (!blk_mq_sched_needs_restart(hctx) || | |
1215 | (no_tag && list_empty_careful(&hctx->dispatch_wait.entry))) | |
1216 | blk_mq_run_hw_queue(hctx, true); | |
1217 | } | |
1218 | ||
1219 | return (queued + errors) != 0; | |
1220 | } | |
1221 | ||
1222 | static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx) | |
1223 | { | |
1224 | int srcu_idx; | |
1225 | ||
1226 | /* | |
1227 | * We should be running this queue from one of the CPUs that | |
1228 | * are mapped to it. | |
1229 | * | |
1230 | * There are at least two related races now between setting | |
1231 | * hctx->next_cpu from blk_mq_hctx_next_cpu() and running | |
1232 | * __blk_mq_run_hw_queue(): | |
1233 | * | |
1234 | * - hctx->next_cpu is found offline in blk_mq_hctx_next_cpu(), | |
1235 | * but later it becomes online, then this warning is harmless | |
1236 | * at all | |
1237 | * | |
1238 | * - hctx->next_cpu is found online in blk_mq_hctx_next_cpu(), | |
1239 | * but later it becomes offline, then the warning can't be | |
1240 | * triggered, and we depend on blk-mq timeout handler to | |
1241 | * handle dispatched requests to this hctx | |
1242 | */ | |
1243 | if (!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) && | |
1244 | cpu_online(hctx->next_cpu)) { | |
1245 | printk(KERN_WARNING "run queue from wrong CPU %d, hctx %s\n", | |
1246 | raw_smp_processor_id(), | |
1247 | cpumask_empty(hctx->cpumask) ? "inactive": "active"); | |
1248 | dump_stack(); | |
1249 | } | |
1250 | ||
1251 | /* | |
1252 | * We can't run the queue inline with ints disabled. Ensure that | |
1253 | * we catch bad users of this early. | |
1254 | */ | |
1255 | WARN_ON_ONCE(in_interrupt()); | |
1256 | ||
1257 | if (!(hctx->flags & BLK_MQ_F_BLOCKING)) { | |
1258 | rcu_read_lock(); | |
1259 | blk_mq_sched_dispatch_requests(hctx); | |
1260 | rcu_read_unlock(); | |
1261 | } else { | |
1262 | might_sleep(); | |
1263 | ||
1264 | srcu_idx = srcu_read_lock(hctx->queue_rq_srcu); | |
1265 | blk_mq_sched_dispatch_requests(hctx); | |
1266 | srcu_read_unlock(hctx->queue_rq_srcu, srcu_idx); | |
1267 | } | |
1268 | } | |
1269 | ||
1270 | static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx) | |
1271 | { | |
1272 | int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask); | |
1273 | ||
1274 | if (cpu >= nr_cpu_ids) | |
1275 | cpu = cpumask_first(hctx->cpumask); | |
1276 | return cpu; | |
1277 | } | |
1278 | ||
1279 | /* | |
1280 | * It'd be great if the workqueue API had a way to pass | |
1281 | * in a mask and had some smarts for more clever placement. | |
1282 | * For now we just round-robin here, switching for every | |
1283 | * BLK_MQ_CPU_WORK_BATCH queued items. | |
1284 | */ | |
1285 | static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx) | |
1286 | { | |
1287 | bool tried = false; | |
1288 | int next_cpu = hctx->next_cpu; | |
1289 | ||
1290 | if (hctx->queue->nr_hw_queues == 1) | |
1291 | return WORK_CPU_UNBOUND; | |
1292 | ||
1293 | if (--hctx->next_cpu_batch <= 0) { | |
1294 | select_cpu: | |
1295 | next_cpu = cpumask_next_and(next_cpu, hctx->cpumask, | |
1296 | cpu_online_mask); | |
1297 | if (next_cpu >= nr_cpu_ids) | |
1298 | next_cpu = blk_mq_first_mapped_cpu(hctx); | |
1299 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; | |
1300 | } | |
1301 | ||
1302 | /* | |
1303 | * Do unbound schedule if we can't find a online CPU for this hctx, | |
1304 | * and it should only happen in the path of handling CPU DEAD. | |
1305 | */ | |
1306 | if (!cpu_online(next_cpu)) { | |
1307 | if (!tried) { | |
1308 | tried = true; | |
1309 | goto select_cpu; | |
1310 | } | |
1311 | ||
1312 | /* | |
1313 | * Make sure to re-select CPU next time once after CPUs | |
1314 | * in hctx->cpumask become online again. | |
1315 | */ | |
1316 | hctx->next_cpu = next_cpu; | |
1317 | hctx->next_cpu_batch = 1; | |
1318 | return WORK_CPU_UNBOUND; | |
1319 | } | |
1320 | ||
1321 | hctx->next_cpu = next_cpu; | |
1322 | return next_cpu; | |
1323 | } | |
1324 | ||
1325 | static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async, | |
1326 | unsigned long msecs) | |
1327 | { | |
1328 | if (unlikely(blk_mq_hctx_stopped(hctx))) | |
1329 | return; | |
1330 | ||
1331 | if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) { | |
1332 | int cpu = get_cpu(); | |
1333 | if (cpumask_test_cpu(cpu, hctx->cpumask)) { | |
1334 | __blk_mq_run_hw_queue(hctx); | |
1335 | put_cpu(); | |
1336 | return; | |
1337 | } | |
1338 | ||
1339 | put_cpu(); | |
1340 | } | |
1341 | ||
1342 | kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work, | |
1343 | msecs_to_jiffies(msecs)); | |
1344 | } | |
1345 | ||
1346 | void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs) | |
1347 | { | |
1348 | __blk_mq_delay_run_hw_queue(hctx, true, msecs); | |
1349 | } | |
1350 | EXPORT_SYMBOL(blk_mq_delay_run_hw_queue); | |
1351 | ||
1352 | bool blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) | |
1353 | { | |
1354 | if (blk_mq_hctx_has_pending(hctx)) { | |
1355 | __blk_mq_delay_run_hw_queue(hctx, async, 0); | |
1356 | return true; | |
1357 | } | |
1358 | ||
1359 | return false; | |
1360 | } | |
1361 | EXPORT_SYMBOL(blk_mq_run_hw_queue); | |
1362 | ||
1363 | void blk_mq_run_hw_queues(struct request_queue *q, bool async) | |
1364 | { | |
1365 | struct blk_mq_hw_ctx *hctx; | |
1366 | int i; | |
1367 | ||
1368 | queue_for_each_hw_ctx(q, hctx, i) { | |
1369 | if (blk_mq_hctx_stopped(hctx)) | |
1370 | continue; | |
1371 | ||
1372 | blk_mq_run_hw_queue(hctx, async); | |
1373 | } | |
1374 | } | |
1375 | EXPORT_SYMBOL(blk_mq_run_hw_queues); | |
1376 | ||
1377 | /** | |
1378 | * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped | |
1379 | * @q: request queue. | |
1380 | * | |
1381 | * The caller is responsible for serializing this function against | |
1382 | * blk_mq_{start,stop}_hw_queue(). | |
1383 | */ | |
1384 | bool blk_mq_queue_stopped(struct request_queue *q) | |
1385 | { | |
1386 | struct blk_mq_hw_ctx *hctx; | |
1387 | int i; | |
1388 | ||
1389 | queue_for_each_hw_ctx(q, hctx, i) | |
1390 | if (blk_mq_hctx_stopped(hctx)) | |
1391 | return true; | |
1392 | ||
1393 | return false; | |
1394 | } | |
1395 | EXPORT_SYMBOL(blk_mq_queue_stopped); | |
1396 | ||
1397 | /* | |
1398 | * This function is often used for pausing .queue_rq() by driver when | |
1399 | * there isn't enough resource or some conditions aren't satisfied, and | |
1400 | * BLK_STS_RESOURCE is usually returned. | |
1401 | * | |
1402 | * We do not guarantee that dispatch can be drained or blocked | |
1403 | * after blk_mq_stop_hw_queue() returns. Please use | |
1404 | * blk_mq_quiesce_queue() for that requirement. | |
1405 | */ | |
1406 | void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx) | |
1407 | { | |
1408 | cancel_delayed_work(&hctx->run_work); | |
1409 | ||
1410 | set_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
1411 | } | |
1412 | EXPORT_SYMBOL(blk_mq_stop_hw_queue); | |
1413 | ||
1414 | /* | |
1415 | * This function is often used for pausing .queue_rq() by driver when | |
1416 | * there isn't enough resource or some conditions aren't satisfied, and | |
1417 | * BLK_STS_RESOURCE is usually returned. | |
1418 | * | |
1419 | * We do not guarantee that dispatch can be drained or blocked | |
1420 | * after blk_mq_stop_hw_queues() returns. Please use | |
1421 | * blk_mq_quiesce_queue() for that requirement. | |
1422 | */ | |
1423 | void blk_mq_stop_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_stop_hw_queue(hctx); | |
1430 | } | |
1431 | EXPORT_SYMBOL(blk_mq_stop_hw_queues); | |
1432 | ||
1433 | void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx) | |
1434 | { | |
1435 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
1436 | ||
1437 | blk_mq_run_hw_queue(hctx, false); | |
1438 | } | |
1439 | EXPORT_SYMBOL(blk_mq_start_hw_queue); | |
1440 | ||
1441 | void blk_mq_start_hw_queues(struct request_queue *q) | |
1442 | { | |
1443 | struct blk_mq_hw_ctx *hctx; | |
1444 | int i; | |
1445 | ||
1446 | queue_for_each_hw_ctx(q, hctx, i) | |
1447 | blk_mq_start_hw_queue(hctx); | |
1448 | } | |
1449 | EXPORT_SYMBOL(blk_mq_start_hw_queues); | |
1450 | ||
1451 | void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) | |
1452 | { | |
1453 | if (!blk_mq_hctx_stopped(hctx)) | |
1454 | return; | |
1455 | ||
1456 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
1457 | blk_mq_run_hw_queue(hctx, async); | |
1458 | } | |
1459 | EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue); | |
1460 | ||
1461 | void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async) | |
1462 | { | |
1463 | struct blk_mq_hw_ctx *hctx; | |
1464 | int i; | |
1465 | ||
1466 | queue_for_each_hw_ctx(q, hctx, i) | |
1467 | blk_mq_start_stopped_hw_queue(hctx, async); | |
1468 | } | |
1469 | EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues); | |
1470 | ||
1471 | static void blk_mq_run_work_fn(struct work_struct *work) | |
1472 | { | |
1473 | struct blk_mq_hw_ctx *hctx; | |
1474 | ||
1475 | hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work); | |
1476 | ||
1477 | /* | |
1478 | * If we are stopped, don't run the queue. The exception is if | |
1479 | * BLK_MQ_S_START_ON_RUN is set. For that case, we auto-clear | |
1480 | * the STOPPED bit and run it. | |
1481 | */ | |
1482 | if (test_bit(BLK_MQ_S_STOPPED, &hctx->state)) { | |
1483 | if (!test_bit(BLK_MQ_S_START_ON_RUN, &hctx->state)) | |
1484 | return; | |
1485 | ||
1486 | clear_bit(BLK_MQ_S_START_ON_RUN, &hctx->state); | |
1487 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
1488 | } | |
1489 | ||
1490 | __blk_mq_run_hw_queue(hctx); | |
1491 | } | |
1492 | ||
1493 | ||
1494 | void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs) | |
1495 | { | |
1496 | if (WARN_ON_ONCE(!blk_mq_hw_queue_mapped(hctx))) | |
1497 | return; | |
1498 | ||
1499 | /* | |
1500 | * Stop the hw queue, then modify currently delayed work. | |
1501 | * This should prevent us from running the queue prematurely. | |
1502 | * Mark the queue as auto-clearing STOPPED when it runs. | |
1503 | */ | |
1504 | blk_mq_stop_hw_queue(hctx); | |
1505 | set_bit(BLK_MQ_S_START_ON_RUN, &hctx->state); | |
1506 | kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), | |
1507 | &hctx->run_work, | |
1508 | msecs_to_jiffies(msecs)); | |
1509 | } | |
1510 | EXPORT_SYMBOL(blk_mq_delay_queue); | |
1511 | ||
1512 | static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx, | |
1513 | struct request *rq, | |
1514 | bool at_head) | |
1515 | { | |
1516 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
1517 | ||
1518 | lockdep_assert_held(&ctx->lock); | |
1519 | ||
1520 | trace_block_rq_insert(hctx->queue, rq); | |
1521 | ||
1522 | if (at_head) | |
1523 | list_add(&rq->queuelist, &ctx->rq_list); | |
1524 | else | |
1525 | list_add_tail(&rq->queuelist, &ctx->rq_list); | |
1526 | } | |
1527 | ||
1528 | void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq, | |
1529 | bool at_head) | |
1530 | { | |
1531 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
1532 | ||
1533 | lockdep_assert_held(&ctx->lock); | |
1534 | ||
1535 | __blk_mq_insert_req_list(hctx, rq, at_head); | |
1536 | blk_mq_hctx_mark_pending(hctx, ctx); | |
1537 | } | |
1538 | ||
1539 | /* | |
1540 | * Should only be used carefully, when the caller knows we want to | |
1541 | * bypass a potential IO scheduler on the target device. | |
1542 | */ | |
1543 | void blk_mq_request_bypass_insert(struct request *rq, bool run_queue) | |
1544 | { | |
1545 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
1546 | struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(rq->q, ctx->cpu); | |
1547 | ||
1548 | spin_lock(&hctx->lock); | |
1549 | list_add_tail(&rq->queuelist, &hctx->dispatch); | |
1550 | spin_unlock(&hctx->lock); | |
1551 | ||
1552 | if (run_queue) | |
1553 | blk_mq_run_hw_queue(hctx, false); | |
1554 | } | |
1555 | ||
1556 | void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx, | |
1557 | struct list_head *list) | |
1558 | ||
1559 | { | |
1560 | /* | |
1561 | * preemption doesn't flush plug list, so it's possible ctx->cpu is | |
1562 | * offline now | |
1563 | */ | |
1564 | spin_lock(&ctx->lock); | |
1565 | while (!list_empty(list)) { | |
1566 | struct request *rq; | |
1567 | ||
1568 | rq = list_first_entry(list, struct request, queuelist); | |
1569 | BUG_ON(rq->mq_ctx != ctx); | |
1570 | list_del_init(&rq->queuelist); | |
1571 | __blk_mq_insert_req_list(hctx, rq, false); | |
1572 | } | |
1573 | blk_mq_hctx_mark_pending(hctx, ctx); | |
1574 | spin_unlock(&ctx->lock); | |
1575 | } | |
1576 | ||
1577 | static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b) | |
1578 | { | |
1579 | struct request *rqa = container_of(a, struct request, queuelist); | |
1580 | struct request *rqb = container_of(b, struct request, queuelist); | |
1581 | ||
1582 | return !(rqa->mq_ctx < rqb->mq_ctx || | |
1583 | (rqa->mq_ctx == rqb->mq_ctx && | |
1584 | blk_rq_pos(rqa) < blk_rq_pos(rqb))); | |
1585 | } | |
1586 | ||
1587 | void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule) | |
1588 | { | |
1589 | struct blk_mq_ctx *this_ctx; | |
1590 | struct request_queue *this_q; | |
1591 | struct request *rq; | |
1592 | LIST_HEAD(list); | |
1593 | LIST_HEAD(ctx_list); | |
1594 | unsigned int depth; | |
1595 | ||
1596 | list_splice_init(&plug->mq_list, &list); | |
1597 | ||
1598 | list_sort(NULL, &list, plug_ctx_cmp); | |
1599 | ||
1600 | this_q = NULL; | |
1601 | this_ctx = NULL; | |
1602 | depth = 0; | |
1603 | ||
1604 | while (!list_empty(&list)) { | |
1605 | rq = list_entry_rq(list.next); | |
1606 | list_del_init(&rq->queuelist); | |
1607 | BUG_ON(!rq->q); | |
1608 | if (rq->mq_ctx != this_ctx) { | |
1609 | if (this_ctx) { | |
1610 | trace_block_unplug(this_q, depth, from_schedule); | |
1611 | blk_mq_sched_insert_requests(this_q, this_ctx, | |
1612 | &ctx_list, | |
1613 | from_schedule); | |
1614 | } | |
1615 | ||
1616 | this_ctx = rq->mq_ctx; | |
1617 | this_q = rq->q; | |
1618 | depth = 0; | |
1619 | } | |
1620 | ||
1621 | depth++; | |
1622 | list_add_tail(&rq->queuelist, &ctx_list); | |
1623 | } | |
1624 | ||
1625 | /* | |
1626 | * If 'this_ctx' is set, we know we have entries to complete | |
1627 | * on 'ctx_list'. Do those. | |
1628 | */ | |
1629 | if (this_ctx) { | |
1630 | trace_block_unplug(this_q, depth, from_schedule); | |
1631 | blk_mq_sched_insert_requests(this_q, this_ctx, &ctx_list, | |
1632 | from_schedule); | |
1633 | } | |
1634 | } | |
1635 | ||
1636 | static void blk_mq_bio_to_request(struct request *rq, struct bio *bio) | |
1637 | { | |
1638 | blk_init_request_from_bio(rq, bio); | |
1639 | ||
1640 | blk_rq_set_rl(rq, blk_get_rl(rq->q, bio)); | |
1641 | ||
1642 | blk_account_io_start(rq, true); | |
1643 | } | |
1644 | ||
1645 | static inline void blk_mq_queue_io(struct blk_mq_hw_ctx *hctx, | |
1646 | struct blk_mq_ctx *ctx, | |
1647 | struct request *rq) | |
1648 | { | |
1649 | spin_lock(&ctx->lock); | |
1650 | __blk_mq_insert_request(hctx, rq, false); | |
1651 | spin_unlock(&ctx->lock); | |
1652 | } | |
1653 | ||
1654 | static blk_qc_t request_to_qc_t(struct blk_mq_hw_ctx *hctx, struct request *rq) | |
1655 | { | |
1656 | if (rq->tag != -1) | |
1657 | return blk_tag_to_qc_t(rq->tag, hctx->queue_num, false); | |
1658 | ||
1659 | return blk_tag_to_qc_t(rq->internal_tag, hctx->queue_num, true); | |
1660 | } | |
1661 | ||
1662 | static void __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx, | |
1663 | struct request *rq, | |
1664 | blk_qc_t *cookie, bool may_sleep) | |
1665 | { | |
1666 | struct request_queue *q = rq->q; | |
1667 | struct blk_mq_queue_data bd = { | |
1668 | .rq = rq, | |
1669 | .last = true, | |
1670 | }; | |
1671 | blk_qc_t new_cookie; | |
1672 | blk_status_t ret; | |
1673 | bool run_queue = true; | |
1674 | ||
1675 | /* RCU or SRCU read lock is needed before checking quiesced flag */ | |
1676 | if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)) { | |
1677 | run_queue = false; | |
1678 | goto insert; | |
1679 | } | |
1680 | ||
1681 | if (q->elevator) | |
1682 | goto insert; | |
1683 | ||
1684 | if (!blk_mq_get_dispatch_budget(hctx)) | |
1685 | goto insert; | |
1686 | ||
1687 | if (!blk_mq_get_driver_tag(rq, NULL, false)) { | |
1688 | blk_mq_put_dispatch_budget(hctx); | |
1689 | goto insert; | |
1690 | } | |
1691 | ||
1692 | new_cookie = request_to_qc_t(hctx, rq); | |
1693 | ||
1694 | /* | |
1695 | * For OK queue, we are done. For error, kill it. Any other | |
1696 | * error (busy), just add it to our list as we previously | |
1697 | * would have done | |
1698 | */ | |
1699 | ret = q->mq_ops->queue_rq(hctx, &bd); | |
1700 | switch (ret) { | |
1701 | case BLK_STS_OK: | |
1702 | *cookie = new_cookie; | |
1703 | return; | |
1704 | case BLK_STS_RESOURCE: | |
1705 | __blk_mq_requeue_request(rq); | |
1706 | goto insert; | |
1707 | default: | |
1708 | *cookie = BLK_QC_T_NONE; | |
1709 | blk_mq_end_request(rq, ret); | |
1710 | return; | |
1711 | } | |
1712 | ||
1713 | insert: | |
1714 | blk_mq_sched_insert_request(rq, false, run_queue, false, may_sleep); | |
1715 | } | |
1716 | ||
1717 | static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx, | |
1718 | struct request *rq, blk_qc_t *cookie) | |
1719 | { | |
1720 | if (!(hctx->flags & BLK_MQ_F_BLOCKING)) { | |
1721 | rcu_read_lock(); | |
1722 | __blk_mq_try_issue_directly(hctx, rq, cookie, false); | |
1723 | rcu_read_unlock(); | |
1724 | } else { | |
1725 | unsigned int srcu_idx; | |
1726 | ||
1727 | might_sleep(); | |
1728 | ||
1729 | srcu_idx = srcu_read_lock(hctx->queue_rq_srcu); | |
1730 | __blk_mq_try_issue_directly(hctx, rq, cookie, true); | |
1731 | srcu_read_unlock(hctx->queue_rq_srcu, srcu_idx); | |
1732 | } | |
1733 | } | |
1734 | ||
1735 | static blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio) | |
1736 | { | |
1737 | const int is_sync = op_is_sync(bio->bi_opf); | |
1738 | const int is_flush_fua = op_is_flush(bio->bi_opf); | |
1739 | struct blk_mq_alloc_data data = { .flags = 0 }; | |
1740 | struct request *rq; | |
1741 | unsigned int request_count = 0; | |
1742 | struct blk_plug *plug; | |
1743 | struct request *same_queue_rq = NULL; | |
1744 | blk_qc_t cookie; | |
1745 | unsigned int wb_acct; | |
1746 | ||
1747 | blk_queue_bounce(q, &bio); | |
1748 | ||
1749 | blk_queue_split(q, &bio); | |
1750 | ||
1751 | if (!bio_integrity_prep(bio)) | |
1752 | return BLK_QC_T_NONE; | |
1753 | ||
1754 | if (!is_flush_fua && !blk_queue_nomerges(q) && | |
1755 | blk_attempt_plug_merge(q, bio, &request_count, &same_queue_rq)) | |
1756 | return BLK_QC_T_NONE; | |
1757 | ||
1758 | if (blk_mq_sched_bio_merge(q, bio)) | |
1759 | return BLK_QC_T_NONE; | |
1760 | ||
1761 | wb_acct = wbt_wait(q->rq_wb, bio, NULL); | |
1762 | ||
1763 | trace_block_getrq(q, bio, bio->bi_opf); | |
1764 | ||
1765 | rq = blk_mq_get_request(q, bio, bio->bi_opf, &data); | |
1766 | if (unlikely(!rq)) { | |
1767 | __wbt_done(q->rq_wb, wb_acct); | |
1768 | if (bio->bi_opf & REQ_NOWAIT) | |
1769 | bio_wouldblock_error(bio); | |
1770 | return BLK_QC_T_NONE; | |
1771 | } | |
1772 | ||
1773 | wbt_track(&rq->issue_stat, wb_acct); | |
1774 | ||
1775 | cookie = request_to_qc_t(data.hctx, rq); | |
1776 | ||
1777 | plug = current->plug; | |
1778 | if (unlikely(is_flush_fua)) { | |
1779 | blk_mq_put_ctx(data.ctx); | |
1780 | blk_mq_bio_to_request(rq, bio); | |
1781 | ||
1782 | /* bypass scheduler for flush rq */ | |
1783 | blk_insert_flush(rq); | |
1784 | blk_mq_run_hw_queue(data.hctx, true); | |
1785 | } else if (plug && q->nr_hw_queues == 1) { | |
1786 | struct request *last = NULL; | |
1787 | ||
1788 | blk_mq_put_ctx(data.ctx); | |
1789 | blk_mq_bio_to_request(rq, bio); | |
1790 | ||
1791 | /* | |
1792 | * @request_count may become stale because of schedule | |
1793 | * out, so check the list again. | |
1794 | */ | |
1795 | if (list_empty(&plug->mq_list)) | |
1796 | request_count = 0; | |
1797 | else if (blk_queue_nomerges(q)) | |
1798 | request_count = blk_plug_queued_count(q); | |
1799 | ||
1800 | if (!request_count) | |
1801 | trace_block_plug(q); | |
1802 | else | |
1803 | last = list_entry_rq(plug->mq_list.prev); | |
1804 | ||
1805 | if (request_count >= BLK_MAX_REQUEST_COUNT || (last && | |
1806 | blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) { | |
1807 | blk_flush_plug_list(plug, false); | |
1808 | trace_block_plug(q); | |
1809 | } | |
1810 | ||
1811 | list_add_tail(&rq->queuelist, &plug->mq_list); | |
1812 | } else if (plug && !blk_queue_nomerges(q)) { | |
1813 | blk_mq_bio_to_request(rq, bio); | |
1814 | ||
1815 | /* | |
1816 | * We do limited plugging. If the bio can be merged, do that. | |
1817 | * Otherwise the existing request in the plug list will be | |
1818 | * issued. So the plug list will have one request at most | |
1819 | * The plug list might get flushed before this. If that happens, | |
1820 | * the plug list is empty, and same_queue_rq is invalid. | |
1821 | */ | |
1822 | if (list_empty(&plug->mq_list)) | |
1823 | same_queue_rq = NULL; | |
1824 | if (same_queue_rq) | |
1825 | list_del_init(&same_queue_rq->queuelist); | |
1826 | list_add_tail(&rq->queuelist, &plug->mq_list); | |
1827 | ||
1828 | blk_mq_put_ctx(data.ctx); | |
1829 | ||
1830 | if (same_queue_rq) { | |
1831 | data.hctx = blk_mq_map_queue(q, | |
1832 | same_queue_rq->mq_ctx->cpu); | |
1833 | blk_mq_try_issue_directly(data.hctx, same_queue_rq, | |
1834 | &cookie); | |
1835 | } | |
1836 | } else if (q->nr_hw_queues > 1 && is_sync) { | |
1837 | blk_mq_put_ctx(data.ctx); | |
1838 | blk_mq_bio_to_request(rq, bio); | |
1839 | blk_mq_try_issue_directly(data.hctx, rq, &cookie); | |
1840 | } else if (q->elevator) { | |
1841 | blk_mq_put_ctx(data.ctx); | |
1842 | blk_mq_bio_to_request(rq, bio); | |
1843 | blk_mq_sched_insert_request(rq, false, true, true, true); | |
1844 | } else { | |
1845 | blk_mq_put_ctx(data.ctx); | |
1846 | blk_mq_bio_to_request(rq, bio); | |
1847 | blk_mq_queue_io(data.hctx, data.ctx, rq); | |
1848 | blk_mq_run_hw_queue(data.hctx, true); | |
1849 | } | |
1850 | ||
1851 | return cookie; | |
1852 | } | |
1853 | ||
1854 | void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags, | |
1855 | unsigned int hctx_idx) | |
1856 | { | |
1857 | struct page *page; | |
1858 | ||
1859 | if (tags->rqs && set->ops->exit_request) { | |
1860 | int i; | |
1861 | ||
1862 | for (i = 0; i < tags->nr_tags; i++) { | |
1863 | struct request *rq = tags->static_rqs[i]; | |
1864 | ||
1865 | if (!rq) | |
1866 | continue; | |
1867 | set->ops->exit_request(set, rq, hctx_idx); | |
1868 | tags->static_rqs[i] = NULL; | |
1869 | } | |
1870 | } | |
1871 | ||
1872 | while (!list_empty(&tags->page_list)) { | |
1873 | page = list_first_entry(&tags->page_list, struct page, lru); | |
1874 | list_del_init(&page->lru); | |
1875 | /* | |
1876 | * Remove kmemleak object previously allocated in | |
1877 | * blk_mq_init_rq_map(). | |
1878 | */ | |
1879 | kmemleak_free(page_address(page)); | |
1880 | __free_pages(page, page->private); | |
1881 | } | |
1882 | } | |
1883 | ||
1884 | void blk_mq_free_rq_map(struct blk_mq_tags *tags) | |
1885 | { | |
1886 | kfree(tags->rqs); | |
1887 | tags->rqs = NULL; | |
1888 | kfree(tags->static_rqs); | |
1889 | tags->static_rqs = NULL; | |
1890 | ||
1891 | blk_mq_free_tags(tags); | |
1892 | } | |
1893 | ||
1894 | struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, | |
1895 | unsigned int hctx_idx, | |
1896 | unsigned int nr_tags, | |
1897 | unsigned int reserved_tags) | |
1898 | { | |
1899 | struct blk_mq_tags *tags; | |
1900 | int node; | |
1901 | ||
1902 | node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx); | |
1903 | if (node == NUMA_NO_NODE) | |
1904 | node = set->numa_node; | |
1905 | ||
1906 | tags = blk_mq_init_tags(nr_tags, reserved_tags, node, | |
1907 | BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags)); | |
1908 | if (!tags) | |
1909 | return NULL; | |
1910 | ||
1911 | tags->rqs = kzalloc_node(nr_tags * sizeof(struct request *), | |
1912 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, | |
1913 | node); | |
1914 | if (!tags->rqs) { | |
1915 | blk_mq_free_tags(tags); | |
1916 | return NULL; | |
1917 | } | |
1918 | ||
1919 | tags->static_rqs = kzalloc_node(nr_tags * sizeof(struct request *), | |
1920 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, | |
1921 | node); | |
1922 | if (!tags->static_rqs) { | |
1923 | kfree(tags->rqs); | |
1924 | blk_mq_free_tags(tags); | |
1925 | return NULL; | |
1926 | } | |
1927 | ||
1928 | return tags; | |
1929 | } | |
1930 | ||
1931 | static size_t order_to_size(unsigned int order) | |
1932 | { | |
1933 | return (size_t)PAGE_SIZE << order; | |
1934 | } | |
1935 | ||
1936 | int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags, | |
1937 | unsigned int hctx_idx, unsigned int depth) | |
1938 | { | |
1939 | unsigned int i, j, entries_per_page, max_order = 4; | |
1940 | size_t rq_size, left; | |
1941 | int node; | |
1942 | ||
1943 | node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx); | |
1944 | if (node == NUMA_NO_NODE) | |
1945 | node = set->numa_node; | |
1946 | ||
1947 | INIT_LIST_HEAD(&tags->page_list); | |
1948 | ||
1949 | /* | |
1950 | * rq_size is the size of the request plus driver payload, rounded | |
1951 | * to the cacheline size | |
1952 | */ | |
1953 | rq_size = round_up(sizeof(struct request) + set->cmd_size, | |
1954 | cache_line_size()); | |
1955 | left = rq_size * depth; | |
1956 | ||
1957 | for (i = 0; i < depth; ) { | |
1958 | int this_order = max_order; | |
1959 | struct page *page; | |
1960 | int to_do; | |
1961 | void *p; | |
1962 | ||
1963 | while (this_order && left < order_to_size(this_order - 1)) | |
1964 | this_order--; | |
1965 | ||
1966 | do { | |
1967 | page = alloc_pages_node(node, | |
1968 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO, | |
1969 | this_order); | |
1970 | if (page) | |
1971 | break; | |
1972 | if (!this_order--) | |
1973 | break; | |
1974 | if (order_to_size(this_order) < rq_size) | |
1975 | break; | |
1976 | } while (1); | |
1977 | ||
1978 | if (!page) | |
1979 | goto fail; | |
1980 | ||
1981 | page->private = this_order; | |
1982 | list_add_tail(&page->lru, &tags->page_list); | |
1983 | ||
1984 | p = page_address(page); | |
1985 | /* | |
1986 | * Allow kmemleak to scan these pages as they contain pointers | |
1987 | * to additional allocations like via ops->init_request(). | |
1988 | */ | |
1989 | kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO); | |
1990 | entries_per_page = order_to_size(this_order) / rq_size; | |
1991 | to_do = min(entries_per_page, depth - i); | |
1992 | left -= to_do * rq_size; | |
1993 | for (j = 0; j < to_do; j++) { | |
1994 | struct request *rq = p; | |
1995 | ||
1996 | tags->static_rqs[i] = rq; | |
1997 | if (set->ops->init_request) { | |
1998 | if (set->ops->init_request(set, rq, hctx_idx, | |
1999 | node)) { | |
2000 | tags->static_rqs[i] = NULL; | |
2001 | goto fail; | |
2002 | } | |
2003 | } | |
2004 | ||
2005 | p += rq_size; | |
2006 | i++; | |
2007 | } | |
2008 | } | |
2009 | return 0; | |
2010 | ||
2011 | fail: | |
2012 | blk_mq_free_rqs(set, tags, hctx_idx); | |
2013 | return -ENOMEM; | |
2014 | } | |
2015 | ||
2016 | /* | |
2017 | * 'cpu' is going away. splice any existing rq_list entries from this | |
2018 | * software queue to the hw queue dispatch list, and ensure that it | |
2019 | * gets run. | |
2020 | */ | |
2021 | static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node) | |
2022 | { | |
2023 | struct blk_mq_hw_ctx *hctx; | |
2024 | struct blk_mq_ctx *ctx; | |
2025 | LIST_HEAD(tmp); | |
2026 | ||
2027 | hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead); | |
2028 | ctx = __blk_mq_get_ctx(hctx->queue, cpu); | |
2029 | ||
2030 | spin_lock(&ctx->lock); | |
2031 | if (!list_empty(&ctx->rq_list)) { | |
2032 | list_splice_init(&ctx->rq_list, &tmp); | |
2033 | blk_mq_hctx_clear_pending(hctx, ctx); | |
2034 | } | |
2035 | spin_unlock(&ctx->lock); | |
2036 | ||
2037 | if (list_empty(&tmp)) | |
2038 | return 0; | |
2039 | ||
2040 | spin_lock(&hctx->lock); | |
2041 | list_splice_tail_init(&tmp, &hctx->dispatch); | |
2042 | spin_unlock(&hctx->lock); | |
2043 | ||
2044 | blk_mq_run_hw_queue(hctx, true); | |
2045 | return 0; | |
2046 | } | |
2047 | ||
2048 | static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx) | |
2049 | { | |
2050 | cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD, | |
2051 | &hctx->cpuhp_dead); | |
2052 | } | |
2053 | ||
2054 | /* hctx->ctxs will be freed in queue's release handler */ | |
2055 | static void blk_mq_exit_hctx(struct request_queue *q, | |
2056 | struct blk_mq_tag_set *set, | |
2057 | struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx) | |
2058 | { | |
2059 | blk_mq_debugfs_unregister_hctx(hctx); | |
2060 | ||
2061 | if (blk_mq_hw_queue_mapped(hctx)) | |
2062 | blk_mq_tag_idle(hctx); | |
2063 | ||
2064 | if (set->ops->exit_request) | |
2065 | set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx); | |
2066 | ||
2067 | blk_mq_sched_exit_hctx(q, hctx, hctx_idx); | |
2068 | ||
2069 | if (set->ops->exit_hctx) | |
2070 | set->ops->exit_hctx(hctx, hctx_idx); | |
2071 | ||
2072 | if (hctx->flags & BLK_MQ_F_BLOCKING) | |
2073 | cleanup_srcu_struct(hctx->queue_rq_srcu); | |
2074 | ||
2075 | blk_mq_remove_cpuhp(hctx); | |
2076 | blk_free_flush_queue(hctx->fq); | |
2077 | sbitmap_free(&hctx->ctx_map); | |
2078 | } | |
2079 | ||
2080 | static void blk_mq_exit_hw_queues(struct request_queue *q, | |
2081 | struct blk_mq_tag_set *set, int nr_queue) | |
2082 | { | |
2083 | struct blk_mq_hw_ctx *hctx; | |
2084 | unsigned int i; | |
2085 | ||
2086 | queue_for_each_hw_ctx(q, hctx, i) { | |
2087 | if (i == nr_queue) | |
2088 | break; | |
2089 | blk_mq_exit_hctx(q, set, hctx, i); | |
2090 | } | |
2091 | } | |
2092 | ||
2093 | static int blk_mq_init_hctx(struct request_queue *q, | |
2094 | struct blk_mq_tag_set *set, | |
2095 | struct blk_mq_hw_ctx *hctx, unsigned hctx_idx) | |
2096 | { | |
2097 | int node; | |
2098 | ||
2099 | node = hctx->numa_node; | |
2100 | if (node == NUMA_NO_NODE) | |
2101 | node = hctx->numa_node = set->numa_node; | |
2102 | ||
2103 | INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn); | |
2104 | spin_lock_init(&hctx->lock); | |
2105 | INIT_LIST_HEAD(&hctx->dispatch); | |
2106 | hctx->queue = q; | |
2107 | hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED; | |
2108 | ||
2109 | cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead); | |
2110 | ||
2111 | hctx->tags = set->tags[hctx_idx]; | |
2112 | ||
2113 | /* | |
2114 | * Allocate space for all possible cpus to avoid allocation at | |
2115 | * runtime | |
2116 | */ | |
2117 | hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *), | |
2118 | GFP_KERNEL, node); | |
2119 | if (!hctx->ctxs) | |
2120 | goto unregister_cpu_notifier; | |
2121 | ||
2122 | if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8), GFP_KERNEL, | |
2123 | node)) | |
2124 | goto free_ctxs; | |
2125 | ||
2126 | hctx->nr_ctx = 0; | |
2127 | ||
2128 | init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake); | |
2129 | INIT_LIST_HEAD(&hctx->dispatch_wait.entry); | |
2130 | ||
2131 | if (set->ops->init_hctx && | |
2132 | set->ops->init_hctx(hctx, set->driver_data, hctx_idx)) | |
2133 | goto free_bitmap; | |
2134 | ||
2135 | if (blk_mq_sched_init_hctx(q, hctx, hctx_idx)) | |
2136 | goto exit_hctx; | |
2137 | ||
2138 | hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size); | |
2139 | if (!hctx->fq) | |
2140 | goto sched_exit_hctx; | |
2141 | ||
2142 | if (set->ops->init_request && | |
2143 | set->ops->init_request(set, hctx->fq->flush_rq, hctx_idx, | |
2144 | node)) | |
2145 | goto free_fq; | |
2146 | ||
2147 | if (hctx->flags & BLK_MQ_F_BLOCKING) | |
2148 | init_srcu_struct(hctx->queue_rq_srcu); | |
2149 | ||
2150 | blk_mq_debugfs_register_hctx(q, hctx); | |
2151 | ||
2152 | return 0; | |
2153 | ||
2154 | free_fq: | |
2155 | kfree(hctx->fq); | |
2156 | sched_exit_hctx: | |
2157 | blk_mq_sched_exit_hctx(q, hctx, hctx_idx); | |
2158 | exit_hctx: | |
2159 | if (set->ops->exit_hctx) | |
2160 | set->ops->exit_hctx(hctx, hctx_idx); | |
2161 | free_bitmap: | |
2162 | sbitmap_free(&hctx->ctx_map); | |
2163 | free_ctxs: | |
2164 | kfree(hctx->ctxs); | |
2165 | unregister_cpu_notifier: | |
2166 | blk_mq_remove_cpuhp(hctx); | |
2167 | return -1; | |
2168 | } | |
2169 | ||
2170 | static void blk_mq_init_cpu_queues(struct request_queue *q, | |
2171 | unsigned int nr_hw_queues) | |
2172 | { | |
2173 | unsigned int i; | |
2174 | ||
2175 | for_each_possible_cpu(i) { | |
2176 | struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i); | |
2177 | struct blk_mq_hw_ctx *hctx; | |
2178 | ||
2179 | __ctx->cpu = i; | |
2180 | spin_lock_init(&__ctx->lock); | |
2181 | INIT_LIST_HEAD(&__ctx->rq_list); | |
2182 | __ctx->queue = q; | |
2183 | ||
2184 | /* | |
2185 | * Set local node, IFF we have more than one hw queue. If | |
2186 | * not, we remain on the home node of the device | |
2187 | */ | |
2188 | hctx = blk_mq_map_queue(q, i); | |
2189 | if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE) | |
2190 | hctx->numa_node = local_memory_node(cpu_to_node(i)); | |
2191 | } | |
2192 | } | |
2193 | ||
2194 | static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, int hctx_idx) | |
2195 | { | |
2196 | int ret = 0; | |
2197 | ||
2198 | set->tags[hctx_idx] = blk_mq_alloc_rq_map(set, hctx_idx, | |
2199 | set->queue_depth, set->reserved_tags); | |
2200 | if (!set->tags[hctx_idx]) | |
2201 | return false; | |
2202 | ||
2203 | ret = blk_mq_alloc_rqs(set, set->tags[hctx_idx], hctx_idx, | |
2204 | set->queue_depth); | |
2205 | if (!ret) | |
2206 | return true; | |
2207 | ||
2208 | blk_mq_free_rq_map(set->tags[hctx_idx]); | |
2209 | set->tags[hctx_idx] = NULL; | |
2210 | return false; | |
2211 | } | |
2212 | ||
2213 | static void blk_mq_free_map_and_requests(struct blk_mq_tag_set *set, | |
2214 | unsigned int hctx_idx) | |
2215 | { | |
2216 | if (set->tags[hctx_idx]) { | |
2217 | blk_mq_free_rqs(set, set->tags[hctx_idx], hctx_idx); | |
2218 | blk_mq_free_rq_map(set->tags[hctx_idx]); | |
2219 | set->tags[hctx_idx] = NULL; | |
2220 | } | |
2221 | } | |
2222 | ||
2223 | static void blk_mq_map_swqueue(struct request_queue *q) | |
2224 | { | |
2225 | unsigned int i, hctx_idx; | |
2226 | struct blk_mq_hw_ctx *hctx; | |
2227 | struct blk_mq_ctx *ctx; | |
2228 | struct blk_mq_tag_set *set = q->tag_set; | |
2229 | ||
2230 | /* | |
2231 | * Avoid others reading imcomplete hctx->cpumask through sysfs | |
2232 | */ | |
2233 | mutex_lock(&q->sysfs_lock); | |
2234 | ||
2235 | queue_for_each_hw_ctx(q, hctx, i) { | |
2236 | cpumask_clear(hctx->cpumask); | |
2237 | hctx->nr_ctx = 0; | |
2238 | } | |
2239 | ||
2240 | /* | |
2241 | * Map software to hardware queues. | |
2242 | * | |
2243 | * If the cpu isn't present, the cpu is mapped to first hctx. | |
2244 | */ | |
2245 | for_each_possible_cpu(i) { | |
2246 | hctx_idx = q->mq_map[i]; | |
2247 | /* unmapped hw queue can be remapped after CPU topo changed */ | |
2248 | if (!set->tags[hctx_idx] && | |
2249 | !__blk_mq_alloc_rq_map(set, hctx_idx)) { | |
2250 | /* | |
2251 | * If tags initialization fail for some hctx, | |
2252 | * that hctx won't be brought online. In this | |
2253 | * case, remap the current ctx to hctx[0] which | |
2254 | * is guaranteed to always have tags allocated | |
2255 | */ | |
2256 | q->mq_map[i] = 0; | |
2257 | } | |
2258 | ||
2259 | ctx = per_cpu_ptr(q->queue_ctx, i); | |
2260 | hctx = blk_mq_map_queue(q, i); | |
2261 | ||
2262 | cpumask_set_cpu(i, hctx->cpumask); | |
2263 | ctx->index_hw = hctx->nr_ctx; | |
2264 | hctx->ctxs[hctx->nr_ctx++] = ctx; | |
2265 | } | |
2266 | ||
2267 | mutex_unlock(&q->sysfs_lock); | |
2268 | ||
2269 | queue_for_each_hw_ctx(q, hctx, i) { | |
2270 | /* | |
2271 | * If no software queues are mapped to this hardware queue, | |
2272 | * disable it and free the request entries. | |
2273 | */ | |
2274 | if (!hctx->nr_ctx) { | |
2275 | /* Never unmap queue 0. We need it as a | |
2276 | * fallback in case of a new remap fails | |
2277 | * allocation | |
2278 | */ | |
2279 | if (i && set->tags[i]) | |
2280 | blk_mq_free_map_and_requests(set, i); | |
2281 | ||
2282 | hctx->tags = NULL; | |
2283 | continue; | |
2284 | } | |
2285 | ||
2286 | hctx->tags = set->tags[i]; | |
2287 | WARN_ON(!hctx->tags); | |
2288 | ||
2289 | /* | |
2290 | * Set the map size to the number of mapped software queues. | |
2291 | * This is more accurate and more efficient than looping | |
2292 | * over all possibly mapped software queues. | |
2293 | */ | |
2294 | sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx); | |
2295 | ||
2296 | /* | |
2297 | * Initialize batch roundrobin counts | |
2298 | */ | |
2299 | hctx->next_cpu = blk_mq_first_mapped_cpu(hctx); | |
2300 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; | |
2301 | } | |
2302 | } | |
2303 | ||
2304 | /* | |
2305 | * Caller needs to ensure that we're either frozen/quiesced, or that | |
2306 | * the queue isn't live yet. | |
2307 | */ | |
2308 | static void queue_set_hctx_shared(struct request_queue *q, bool shared) | |
2309 | { | |
2310 | struct blk_mq_hw_ctx *hctx; | |
2311 | int i; | |
2312 | ||
2313 | queue_for_each_hw_ctx(q, hctx, i) { | |
2314 | if (shared) { | |
2315 | if (test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state)) | |
2316 | atomic_inc(&q->shared_hctx_restart); | |
2317 | hctx->flags |= BLK_MQ_F_TAG_SHARED; | |
2318 | } else { | |
2319 | if (test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state)) | |
2320 | atomic_dec(&q->shared_hctx_restart); | |
2321 | hctx->flags &= ~BLK_MQ_F_TAG_SHARED; | |
2322 | } | |
2323 | } | |
2324 | } | |
2325 | ||
2326 | static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set, | |
2327 | bool shared) | |
2328 | { | |
2329 | struct request_queue *q; | |
2330 | ||
2331 | lockdep_assert_held(&set->tag_list_lock); | |
2332 | ||
2333 | list_for_each_entry(q, &set->tag_list, tag_set_list) { | |
2334 | blk_mq_freeze_queue(q); | |
2335 | queue_set_hctx_shared(q, shared); | |
2336 | blk_mq_unfreeze_queue(q); | |
2337 | } | |
2338 | } | |
2339 | ||
2340 | static void blk_mq_del_queue_tag_set(struct request_queue *q) | |
2341 | { | |
2342 | struct blk_mq_tag_set *set = q->tag_set; | |
2343 | ||
2344 | mutex_lock(&set->tag_list_lock); | |
2345 | list_del_rcu(&q->tag_set_list); | |
2346 | if (list_is_singular(&set->tag_list)) { | |
2347 | /* just transitioned to unshared */ | |
2348 | set->flags &= ~BLK_MQ_F_TAG_SHARED; | |
2349 | /* update existing queue */ | |
2350 | blk_mq_update_tag_set_depth(set, false); | |
2351 | } | |
2352 | mutex_unlock(&set->tag_list_lock); | |
2353 | synchronize_rcu(); | |
2354 | INIT_LIST_HEAD(&q->tag_set_list); | |
2355 | } | |
2356 | ||
2357 | static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set, | |
2358 | struct request_queue *q) | |
2359 | { | |
2360 | q->tag_set = set; | |
2361 | ||
2362 | mutex_lock(&set->tag_list_lock); | |
2363 | ||
2364 | /* | |
2365 | * Check to see if we're transitioning to shared (from 1 to 2 queues). | |
2366 | */ | |
2367 | if (!list_empty(&set->tag_list) && | |
2368 | !(set->flags & BLK_MQ_F_TAG_SHARED)) { | |
2369 | set->flags |= BLK_MQ_F_TAG_SHARED; | |
2370 | /* update existing queue */ | |
2371 | blk_mq_update_tag_set_depth(set, true); | |
2372 | } | |
2373 | if (set->flags & BLK_MQ_F_TAG_SHARED) | |
2374 | queue_set_hctx_shared(q, true); | |
2375 | list_add_tail_rcu(&q->tag_set_list, &set->tag_list); | |
2376 | ||
2377 | mutex_unlock(&set->tag_list_lock); | |
2378 | } | |
2379 | ||
2380 | /* | |
2381 | * It is the actual release handler for mq, but we do it from | |
2382 | * request queue's release handler for avoiding use-after-free | |
2383 | * and headache because q->mq_kobj shouldn't have been introduced, | |
2384 | * but we can't group ctx/kctx kobj without it. | |
2385 | */ | |
2386 | void blk_mq_release(struct request_queue *q) | |
2387 | { | |
2388 | struct blk_mq_hw_ctx *hctx; | |
2389 | unsigned int i; | |
2390 | ||
2391 | /* hctx kobj stays in hctx */ | |
2392 | queue_for_each_hw_ctx(q, hctx, i) { | |
2393 | if (!hctx) | |
2394 | continue; | |
2395 | kobject_put(&hctx->kobj); | |
2396 | } | |
2397 | ||
2398 | q->mq_map = NULL; | |
2399 | ||
2400 | kfree(q->queue_hw_ctx); | |
2401 | ||
2402 | /* | |
2403 | * release .mq_kobj and sw queue's kobject now because | |
2404 | * both share lifetime with request queue. | |
2405 | */ | |
2406 | blk_mq_sysfs_deinit(q); | |
2407 | ||
2408 | free_percpu(q->queue_ctx); | |
2409 | } | |
2410 | ||
2411 | struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set) | |
2412 | { | |
2413 | struct request_queue *uninit_q, *q; | |
2414 | ||
2415 | uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node); | |
2416 | if (!uninit_q) | |
2417 | return ERR_PTR(-ENOMEM); | |
2418 | ||
2419 | q = blk_mq_init_allocated_queue(set, uninit_q); | |
2420 | if (IS_ERR(q)) | |
2421 | blk_cleanup_queue(uninit_q); | |
2422 | ||
2423 | return q; | |
2424 | } | |
2425 | EXPORT_SYMBOL(blk_mq_init_queue); | |
2426 | ||
2427 | static int blk_mq_hw_ctx_size(struct blk_mq_tag_set *tag_set) | |
2428 | { | |
2429 | int hw_ctx_size = sizeof(struct blk_mq_hw_ctx); | |
2430 | ||
2431 | BUILD_BUG_ON(ALIGN(offsetof(struct blk_mq_hw_ctx, queue_rq_srcu), | |
2432 | __alignof__(struct blk_mq_hw_ctx)) != | |
2433 | sizeof(struct blk_mq_hw_ctx)); | |
2434 | ||
2435 | if (tag_set->flags & BLK_MQ_F_BLOCKING) | |
2436 | hw_ctx_size += sizeof(struct srcu_struct); | |
2437 | ||
2438 | return hw_ctx_size; | |
2439 | } | |
2440 | ||
2441 | static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set, | |
2442 | struct request_queue *q) | |
2443 | { | |
2444 | int i, j; | |
2445 | struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx; | |
2446 | ||
2447 | blk_mq_sysfs_unregister(q); | |
2448 | ||
2449 | /* protect against switching io scheduler */ | |
2450 | mutex_lock(&q->sysfs_lock); | |
2451 | for (i = 0; i < set->nr_hw_queues; i++) { | |
2452 | int node; | |
2453 | ||
2454 | if (hctxs[i]) | |
2455 | continue; | |
2456 | ||
2457 | node = blk_mq_hw_queue_to_node(q->mq_map, i); | |
2458 | hctxs[i] = kzalloc_node(blk_mq_hw_ctx_size(set), | |
2459 | GFP_KERNEL, node); | |
2460 | if (!hctxs[i]) | |
2461 | break; | |
2462 | ||
2463 | if (!zalloc_cpumask_var_node(&hctxs[i]->cpumask, GFP_KERNEL, | |
2464 | node)) { | |
2465 | kfree(hctxs[i]); | |
2466 | hctxs[i] = NULL; | |
2467 | break; | |
2468 | } | |
2469 | ||
2470 | atomic_set(&hctxs[i]->nr_active, 0); | |
2471 | hctxs[i]->numa_node = node; | |
2472 | hctxs[i]->queue_num = i; | |
2473 | ||
2474 | if (blk_mq_init_hctx(q, set, hctxs[i], i)) { | |
2475 | free_cpumask_var(hctxs[i]->cpumask); | |
2476 | kfree(hctxs[i]); | |
2477 | hctxs[i] = NULL; | |
2478 | break; | |
2479 | } | |
2480 | blk_mq_hctx_kobj_init(hctxs[i]); | |
2481 | } | |
2482 | for (j = i; j < q->nr_hw_queues; j++) { | |
2483 | struct blk_mq_hw_ctx *hctx = hctxs[j]; | |
2484 | ||
2485 | if (hctx) { | |
2486 | if (hctx->tags) | |
2487 | blk_mq_free_map_and_requests(set, j); | |
2488 | blk_mq_exit_hctx(q, set, hctx, j); | |
2489 | kobject_put(&hctx->kobj); | |
2490 | hctxs[j] = NULL; | |
2491 | ||
2492 | } | |
2493 | } | |
2494 | q->nr_hw_queues = i; | |
2495 | mutex_unlock(&q->sysfs_lock); | |
2496 | blk_mq_sysfs_register(q); | |
2497 | } | |
2498 | ||
2499 | struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set, | |
2500 | struct request_queue *q) | |
2501 | { | |
2502 | /* mark the queue as mq asap */ | |
2503 | q->mq_ops = set->ops; | |
2504 | ||
2505 | q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn, | |
2506 | blk_mq_poll_stats_bkt, | |
2507 | BLK_MQ_POLL_STATS_BKTS, q); | |
2508 | if (!q->poll_cb) | |
2509 | goto err_exit; | |
2510 | ||
2511 | q->queue_ctx = alloc_percpu(struct blk_mq_ctx); | |
2512 | if (!q->queue_ctx) | |
2513 | goto err_exit; | |
2514 | ||
2515 | /* init q->mq_kobj and sw queues' kobjects */ | |
2516 | blk_mq_sysfs_init(q); | |
2517 | ||
2518 | q->queue_hw_ctx = kzalloc_node(nr_cpu_ids * sizeof(*(q->queue_hw_ctx)), | |
2519 | GFP_KERNEL, set->numa_node); | |
2520 | if (!q->queue_hw_ctx) | |
2521 | goto err_percpu; | |
2522 | ||
2523 | q->mq_map = set->mq_map; | |
2524 | ||
2525 | blk_mq_realloc_hw_ctxs(set, q); | |
2526 | if (!q->nr_hw_queues) | |
2527 | goto err_hctxs; | |
2528 | ||
2529 | INIT_WORK(&q->timeout_work, blk_mq_timeout_work); | |
2530 | blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ); | |
2531 | ||
2532 | q->nr_queues = nr_cpu_ids; | |
2533 | ||
2534 | q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT; | |
2535 | ||
2536 | if (!(set->flags & BLK_MQ_F_SG_MERGE)) | |
2537 | q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE; | |
2538 | ||
2539 | q->sg_reserved_size = INT_MAX; | |
2540 | ||
2541 | INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work); | |
2542 | INIT_LIST_HEAD(&q->requeue_list); | |
2543 | spin_lock_init(&q->requeue_lock); | |
2544 | ||
2545 | blk_queue_make_request(q, blk_mq_make_request); | |
2546 | if (q->mq_ops->poll) | |
2547 | q->poll_fn = blk_mq_poll; | |
2548 | ||
2549 | /* | |
2550 | * Do this after blk_queue_make_request() overrides it... | |
2551 | */ | |
2552 | q->nr_requests = set->queue_depth; | |
2553 | ||
2554 | /* | |
2555 | * Default to classic polling | |
2556 | */ | |
2557 | q->poll_nsec = -1; | |
2558 | ||
2559 | if (set->ops->complete) | |
2560 | blk_queue_softirq_done(q, set->ops->complete); | |
2561 | ||
2562 | blk_mq_init_cpu_queues(q, set->nr_hw_queues); | |
2563 | blk_mq_add_queue_tag_set(set, q); | |
2564 | blk_mq_map_swqueue(q); | |
2565 | ||
2566 | if (!(set->flags & BLK_MQ_F_NO_SCHED)) { | |
2567 | int ret; | |
2568 | ||
2569 | ret = blk_mq_sched_init(q); | |
2570 | if (ret) | |
2571 | return ERR_PTR(ret); | |
2572 | } | |
2573 | ||
2574 | return q; | |
2575 | ||
2576 | err_hctxs: | |
2577 | kfree(q->queue_hw_ctx); | |
2578 | err_percpu: | |
2579 | free_percpu(q->queue_ctx); | |
2580 | err_exit: | |
2581 | q->mq_ops = NULL; | |
2582 | return ERR_PTR(-ENOMEM); | |
2583 | } | |
2584 | EXPORT_SYMBOL(blk_mq_init_allocated_queue); | |
2585 | ||
2586 | void blk_mq_free_queue(struct request_queue *q) | |
2587 | { | |
2588 | struct blk_mq_tag_set *set = q->tag_set; | |
2589 | ||
2590 | blk_mq_del_queue_tag_set(q); | |
2591 | blk_mq_exit_hw_queues(q, set, set->nr_hw_queues); | |
2592 | } | |
2593 | ||
2594 | /* Basically redo blk_mq_init_queue with queue frozen */ | |
2595 | static void blk_mq_queue_reinit(struct request_queue *q) | |
2596 | { | |
2597 | WARN_ON_ONCE(!atomic_read(&q->mq_freeze_depth)); | |
2598 | ||
2599 | blk_mq_debugfs_unregister_hctxs(q); | |
2600 | blk_mq_sysfs_unregister(q); | |
2601 | ||
2602 | /* | |
2603 | * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe | |
2604 | * we should change hctx numa_node according to the new topology (this | |
2605 | * involves freeing and re-allocating memory, worth doing?) | |
2606 | */ | |
2607 | blk_mq_map_swqueue(q); | |
2608 | ||
2609 | blk_mq_sysfs_register(q); | |
2610 | blk_mq_debugfs_register_hctxs(q); | |
2611 | } | |
2612 | ||
2613 | static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set) | |
2614 | { | |
2615 | int i; | |
2616 | ||
2617 | for (i = 0; i < set->nr_hw_queues; i++) | |
2618 | if (!__blk_mq_alloc_rq_map(set, i)) | |
2619 | goto out_unwind; | |
2620 | ||
2621 | return 0; | |
2622 | ||
2623 | out_unwind: | |
2624 | while (--i >= 0) | |
2625 | blk_mq_free_rq_map(set->tags[i]); | |
2626 | ||
2627 | return -ENOMEM; | |
2628 | } | |
2629 | ||
2630 | /* | |
2631 | * Allocate the request maps associated with this tag_set. Note that this | |
2632 | * may reduce the depth asked for, if memory is tight. set->queue_depth | |
2633 | * will be updated to reflect the allocated depth. | |
2634 | */ | |
2635 | static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set) | |
2636 | { | |
2637 | unsigned int depth; | |
2638 | int err; | |
2639 | ||
2640 | depth = set->queue_depth; | |
2641 | do { | |
2642 | err = __blk_mq_alloc_rq_maps(set); | |
2643 | if (!err) | |
2644 | break; | |
2645 | ||
2646 | set->queue_depth >>= 1; | |
2647 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) { | |
2648 | err = -ENOMEM; | |
2649 | break; | |
2650 | } | |
2651 | } while (set->queue_depth); | |
2652 | ||
2653 | if (!set->queue_depth || err) { | |
2654 | pr_err("blk-mq: failed to allocate request map\n"); | |
2655 | return -ENOMEM; | |
2656 | } | |
2657 | ||
2658 | if (depth != set->queue_depth) | |
2659 | pr_info("blk-mq: reduced tag depth (%u -> %u)\n", | |
2660 | depth, set->queue_depth); | |
2661 | ||
2662 | return 0; | |
2663 | } | |
2664 | ||
2665 | static int blk_mq_update_queue_map(struct blk_mq_tag_set *set) | |
2666 | { | |
2667 | if (set->ops->map_queues) { | |
2668 | int cpu; | |
2669 | /* | |
2670 | * transport .map_queues is usually done in the following | |
2671 | * way: | |
2672 | * | |
2673 | * for (queue = 0; queue < set->nr_hw_queues; queue++) { | |
2674 | * mask = get_cpu_mask(queue) | |
2675 | * for_each_cpu(cpu, mask) | |
2676 | * set->mq_map[cpu] = queue; | |
2677 | * } | |
2678 | * | |
2679 | * When we need to remap, the table has to be cleared for | |
2680 | * killing stale mapping since one CPU may not be mapped | |
2681 | * to any hw queue. | |
2682 | */ | |
2683 | for_each_possible_cpu(cpu) | |
2684 | set->mq_map[cpu] = 0; | |
2685 | ||
2686 | return set->ops->map_queues(set); | |
2687 | } else | |
2688 | return blk_mq_map_queues(set); | |
2689 | } | |
2690 | ||
2691 | /* | |
2692 | * Alloc a tag set to be associated with one or more request queues. | |
2693 | * May fail with EINVAL for various error conditions. May adjust the | |
2694 | * requested depth down, if if it too large. In that case, the set | |
2695 | * value will be stored in set->queue_depth. | |
2696 | */ | |
2697 | int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set) | |
2698 | { | |
2699 | int ret; | |
2700 | ||
2701 | BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS); | |
2702 | ||
2703 | if (!set->nr_hw_queues) | |
2704 | return -EINVAL; | |
2705 | if (!set->queue_depth) | |
2706 | return -EINVAL; | |
2707 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) | |
2708 | return -EINVAL; | |
2709 | ||
2710 | if (!set->ops->queue_rq) | |
2711 | return -EINVAL; | |
2712 | ||
2713 | if (!set->ops->get_budget ^ !set->ops->put_budget) | |
2714 | return -EINVAL; | |
2715 | ||
2716 | if (set->queue_depth > BLK_MQ_MAX_DEPTH) { | |
2717 | pr_info("blk-mq: reduced tag depth to %u\n", | |
2718 | BLK_MQ_MAX_DEPTH); | |
2719 | set->queue_depth = BLK_MQ_MAX_DEPTH; | |
2720 | } | |
2721 | ||
2722 | /* | |
2723 | * If a crashdump is active, then we are potentially in a very | |
2724 | * memory constrained environment. Limit us to 1 queue and | |
2725 | * 64 tags to prevent using too much memory. | |
2726 | */ | |
2727 | if (is_kdump_kernel()) { | |
2728 | set->nr_hw_queues = 1; | |
2729 | set->queue_depth = min(64U, set->queue_depth); | |
2730 | } | |
2731 | /* | |
2732 | * There is no use for more h/w queues than cpus. | |
2733 | */ | |
2734 | if (set->nr_hw_queues > nr_cpu_ids) | |
2735 | set->nr_hw_queues = nr_cpu_ids; | |
2736 | ||
2737 | set->tags = kzalloc_node(nr_cpu_ids * sizeof(struct blk_mq_tags *), | |
2738 | GFP_KERNEL, set->numa_node); | |
2739 | if (!set->tags) | |
2740 | return -ENOMEM; | |
2741 | ||
2742 | ret = -ENOMEM; | |
2743 | set->mq_map = kzalloc_node(sizeof(*set->mq_map) * nr_cpu_ids, | |
2744 | GFP_KERNEL, set->numa_node); | |
2745 | if (!set->mq_map) | |
2746 | goto out_free_tags; | |
2747 | ||
2748 | ret = blk_mq_update_queue_map(set); | |
2749 | if (ret) | |
2750 | goto out_free_mq_map; | |
2751 | ||
2752 | ret = blk_mq_alloc_rq_maps(set); | |
2753 | if (ret) | |
2754 | goto out_free_mq_map; | |
2755 | ||
2756 | mutex_init(&set->tag_list_lock); | |
2757 | INIT_LIST_HEAD(&set->tag_list); | |
2758 | ||
2759 | return 0; | |
2760 | ||
2761 | out_free_mq_map: | |
2762 | kfree(set->mq_map); | |
2763 | set->mq_map = NULL; | |
2764 | out_free_tags: | |
2765 | kfree(set->tags); | |
2766 | set->tags = NULL; | |
2767 | return ret; | |
2768 | } | |
2769 | EXPORT_SYMBOL(blk_mq_alloc_tag_set); | |
2770 | ||
2771 | void blk_mq_free_tag_set(struct blk_mq_tag_set *set) | |
2772 | { | |
2773 | int i; | |
2774 | ||
2775 | for (i = 0; i < nr_cpu_ids; i++) | |
2776 | blk_mq_free_map_and_requests(set, i); | |
2777 | ||
2778 | kfree(set->mq_map); | |
2779 | set->mq_map = NULL; | |
2780 | ||
2781 | kfree(set->tags); | |
2782 | set->tags = NULL; | |
2783 | } | |
2784 | EXPORT_SYMBOL(blk_mq_free_tag_set); | |
2785 | ||
2786 | int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr) | |
2787 | { | |
2788 | struct blk_mq_tag_set *set = q->tag_set; | |
2789 | struct blk_mq_hw_ctx *hctx; | |
2790 | int i, ret; | |
2791 | ||
2792 | if (!set) | |
2793 | return -EINVAL; | |
2794 | ||
2795 | if (q->nr_requests == nr) | |
2796 | return 0; | |
2797 | ||
2798 | blk_mq_freeze_queue(q); | |
2799 | ||
2800 | ret = 0; | |
2801 | queue_for_each_hw_ctx(q, hctx, i) { | |
2802 | if (!hctx->tags) | |
2803 | continue; | |
2804 | /* | |
2805 | * If we're using an MQ scheduler, just update the scheduler | |
2806 | * queue depth. This is similar to what the old code would do. | |
2807 | */ | |
2808 | if (!hctx->sched_tags) { | |
2809 | ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr, | |
2810 | false); | |
2811 | } else { | |
2812 | ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags, | |
2813 | nr, true); | |
2814 | } | |
2815 | if (ret) | |
2816 | break; | |
2817 | } | |
2818 | ||
2819 | if (!ret) | |
2820 | q->nr_requests = nr; | |
2821 | ||
2822 | blk_mq_unfreeze_queue(q); | |
2823 | ||
2824 | return ret; | |
2825 | } | |
2826 | ||
2827 | static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, | |
2828 | int nr_hw_queues) | |
2829 | { | |
2830 | struct request_queue *q; | |
2831 | ||
2832 | lockdep_assert_held(&set->tag_list_lock); | |
2833 | ||
2834 | if (nr_hw_queues > nr_cpu_ids) | |
2835 | nr_hw_queues = nr_cpu_ids; | |
2836 | if (nr_hw_queues < 1 || nr_hw_queues == set->nr_hw_queues) | |
2837 | return; | |
2838 | ||
2839 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
2840 | blk_mq_freeze_queue(q); | |
2841 | ||
2842 | set->nr_hw_queues = nr_hw_queues; | |
2843 | blk_mq_update_queue_map(set); | |
2844 | list_for_each_entry(q, &set->tag_list, tag_set_list) { | |
2845 | blk_mq_realloc_hw_ctxs(set, q); | |
2846 | blk_mq_queue_reinit(q); | |
2847 | } | |
2848 | ||
2849 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
2850 | blk_mq_unfreeze_queue(q); | |
2851 | } | |
2852 | ||
2853 | void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues) | |
2854 | { | |
2855 | mutex_lock(&set->tag_list_lock); | |
2856 | __blk_mq_update_nr_hw_queues(set, nr_hw_queues); | |
2857 | mutex_unlock(&set->tag_list_lock); | |
2858 | } | |
2859 | EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues); | |
2860 | ||
2861 | /* Enable polling stats and return whether they were already enabled. */ | |
2862 | static bool blk_poll_stats_enable(struct request_queue *q) | |
2863 | { | |
2864 | if (test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) || | |
2865 | test_and_set_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags)) | |
2866 | return true; | |
2867 | blk_stat_add_callback(q, q->poll_cb); | |
2868 | return false; | |
2869 | } | |
2870 | ||
2871 | static void blk_mq_poll_stats_start(struct request_queue *q) | |
2872 | { | |
2873 | /* | |
2874 | * We don't arm the callback if polling stats are not enabled or the | |
2875 | * callback is already active. | |
2876 | */ | |
2877 | if (!test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) || | |
2878 | blk_stat_is_active(q->poll_cb)) | |
2879 | return; | |
2880 | ||
2881 | blk_stat_activate_msecs(q->poll_cb, 100); | |
2882 | } | |
2883 | ||
2884 | static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb) | |
2885 | { | |
2886 | struct request_queue *q = cb->data; | |
2887 | int bucket; | |
2888 | ||
2889 | for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) { | |
2890 | if (cb->stat[bucket].nr_samples) | |
2891 | q->poll_stat[bucket] = cb->stat[bucket]; | |
2892 | } | |
2893 | } | |
2894 | ||
2895 | static unsigned long blk_mq_poll_nsecs(struct request_queue *q, | |
2896 | struct blk_mq_hw_ctx *hctx, | |
2897 | struct request *rq) | |
2898 | { | |
2899 | unsigned long ret = 0; | |
2900 | int bucket; | |
2901 | ||
2902 | /* | |
2903 | * If stats collection isn't on, don't sleep but turn it on for | |
2904 | * future users | |
2905 | */ | |
2906 | if (!blk_poll_stats_enable(q)) | |
2907 | return 0; | |
2908 | ||
2909 | /* | |
2910 | * As an optimistic guess, use half of the mean service time | |
2911 | * for this type of request. We can (and should) make this smarter. | |
2912 | * For instance, if the completion latencies are tight, we can | |
2913 | * get closer than just half the mean. This is especially | |
2914 | * important on devices where the completion latencies are longer | |
2915 | * than ~10 usec. We do use the stats for the relevant IO size | |
2916 | * if available which does lead to better estimates. | |
2917 | */ | |
2918 | bucket = blk_mq_poll_stats_bkt(rq); | |
2919 | if (bucket < 0) | |
2920 | return ret; | |
2921 | ||
2922 | if (q->poll_stat[bucket].nr_samples) | |
2923 | ret = (q->poll_stat[bucket].mean + 1) / 2; | |
2924 | ||
2925 | return ret; | |
2926 | } | |
2927 | ||
2928 | static bool blk_mq_poll_hybrid_sleep(struct request_queue *q, | |
2929 | struct blk_mq_hw_ctx *hctx, | |
2930 | struct request *rq) | |
2931 | { | |
2932 | struct hrtimer_sleeper hs; | |
2933 | enum hrtimer_mode mode; | |
2934 | unsigned int nsecs; | |
2935 | ktime_t kt; | |
2936 | ||
2937 | if (test_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags)) | |
2938 | return false; | |
2939 | ||
2940 | /* | |
2941 | * poll_nsec can be: | |
2942 | * | |
2943 | * -1: don't ever hybrid sleep | |
2944 | * 0: use half of prev avg | |
2945 | * >0: use this specific value | |
2946 | */ | |
2947 | if (q->poll_nsec == -1) | |
2948 | return false; | |
2949 | else if (q->poll_nsec > 0) | |
2950 | nsecs = q->poll_nsec; | |
2951 | else | |
2952 | nsecs = blk_mq_poll_nsecs(q, hctx, rq); | |
2953 | ||
2954 | if (!nsecs) | |
2955 | return false; | |
2956 | ||
2957 | set_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags); | |
2958 | ||
2959 | /* | |
2960 | * This will be replaced with the stats tracking code, using | |
2961 | * 'avg_completion_time / 2' as the pre-sleep target. | |
2962 | */ | |
2963 | kt = nsecs; | |
2964 | ||
2965 | mode = HRTIMER_MODE_REL; | |
2966 | hrtimer_init_on_stack(&hs.timer, CLOCK_MONOTONIC, mode); | |
2967 | hrtimer_set_expires(&hs.timer, kt); | |
2968 | ||
2969 | hrtimer_init_sleeper(&hs, current); | |
2970 | do { | |
2971 | if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags)) | |
2972 | break; | |
2973 | set_current_state(TASK_UNINTERRUPTIBLE); | |
2974 | hrtimer_start_expires(&hs.timer, mode); | |
2975 | if (hs.task) | |
2976 | io_schedule(); | |
2977 | hrtimer_cancel(&hs.timer); | |
2978 | mode = HRTIMER_MODE_ABS; | |
2979 | } while (hs.task && !signal_pending(current)); | |
2980 | ||
2981 | __set_current_state(TASK_RUNNING); | |
2982 | destroy_hrtimer_on_stack(&hs.timer); | |
2983 | return true; | |
2984 | } | |
2985 | ||
2986 | static bool __blk_mq_poll(struct blk_mq_hw_ctx *hctx, struct request *rq) | |
2987 | { | |
2988 | struct request_queue *q = hctx->queue; | |
2989 | long state; | |
2990 | ||
2991 | /* | |
2992 | * If we sleep, have the caller restart the poll loop to reset | |
2993 | * the state. Like for the other success return cases, the | |
2994 | * caller is responsible for checking if the IO completed. If | |
2995 | * the IO isn't complete, we'll get called again and will go | |
2996 | * straight to the busy poll loop. | |
2997 | */ | |
2998 | if (blk_mq_poll_hybrid_sleep(q, hctx, rq)) | |
2999 | return true; | |
3000 | ||
3001 | hctx->poll_considered++; | |
3002 | ||
3003 | state = current->state; | |
3004 | while (!need_resched()) { | |
3005 | int ret; | |
3006 | ||
3007 | hctx->poll_invoked++; | |
3008 | ||
3009 | ret = q->mq_ops->poll(hctx, rq->tag); | |
3010 | if (ret > 0) { | |
3011 | hctx->poll_success++; | |
3012 | set_current_state(TASK_RUNNING); | |
3013 | return true; | |
3014 | } | |
3015 | ||
3016 | if (signal_pending_state(state, current)) | |
3017 | set_current_state(TASK_RUNNING); | |
3018 | ||
3019 | if (current->state == TASK_RUNNING) | |
3020 | return true; | |
3021 | if (ret < 0) | |
3022 | break; | |
3023 | cpu_relax(); | |
3024 | } | |
3025 | ||
3026 | return false; | |
3027 | } | |
3028 | ||
3029 | static bool blk_mq_poll(struct request_queue *q, blk_qc_t cookie) | |
3030 | { | |
3031 | struct blk_mq_hw_ctx *hctx; | |
3032 | struct request *rq; | |
3033 | ||
3034 | if (!test_bit(QUEUE_FLAG_POLL, &q->queue_flags)) | |
3035 | return false; | |
3036 | ||
3037 | hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)]; | |
3038 | if (!blk_qc_t_is_internal(cookie)) | |
3039 | rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie)); | |
3040 | else { | |
3041 | rq = blk_mq_tag_to_rq(hctx->sched_tags, blk_qc_t_to_tag(cookie)); | |
3042 | /* | |
3043 | * With scheduling, if the request has completed, we'll | |
3044 | * get a NULL return here, as we clear the sched tag when | |
3045 | * that happens. The request still remains valid, like always, | |
3046 | * so we should be safe with just the NULL check. | |
3047 | */ | |
3048 | if (!rq) | |
3049 | return false; | |
3050 | } | |
3051 | ||
3052 | return __blk_mq_poll(hctx, rq); | |
3053 | } | |
3054 | ||
3055 | static int __init blk_mq_init(void) | |
3056 | { | |
3057 | /* | |
3058 | * See comment in block/blk.h rq_atomic_flags enum | |
3059 | */ | |
3060 | BUILD_BUG_ON((REQ_ATOM_STARTED / BITS_PER_BYTE) != | |
3061 | (REQ_ATOM_COMPLETE / BITS_PER_BYTE)); | |
3062 | ||
3063 | cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL, | |
3064 | blk_mq_hctx_notify_dead); | |
3065 | return 0; | |
3066 | } | |
3067 | subsys_initcall(blk_mq_init); |