]>
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/delay.h> | |
24 | #include <linux/crash_dump.h> | |
25 | #include <linux/prefetch.h> | |
26 | ||
27 | #include <trace/events/block.h> | |
28 | ||
29 | #include <linux/blk-mq.h> | |
30 | #include "blk.h" | |
31 | #include "blk-mq.h" | |
32 | #include "blk-mq-tag.h" | |
33 | #include "blk-stat.h" | |
34 | #include "blk-wbt.h" | |
35 | ||
36 | static DEFINE_MUTEX(all_q_mutex); | |
37 | static LIST_HEAD(all_q_list); | |
38 | ||
39 | /* | |
40 | * Check if any of the ctx's have pending work in this hardware queue | |
41 | */ | |
42 | static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx) | |
43 | { | |
44 | return sbitmap_any_bit_set(&hctx->ctx_map); | |
45 | } | |
46 | ||
47 | /* | |
48 | * Mark this ctx as having pending work in this hardware queue | |
49 | */ | |
50 | static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx, | |
51 | struct blk_mq_ctx *ctx) | |
52 | { | |
53 | if (!sbitmap_test_bit(&hctx->ctx_map, ctx->index_hw)) | |
54 | sbitmap_set_bit(&hctx->ctx_map, ctx->index_hw); | |
55 | } | |
56 | ||
57 | static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx, | |
58 | struct blk_mq_ctx *ctx) | |
59 | { | |
60 | sbitmap_clear_bit(&hctx->ctx_map, ctx->index_hw); | |
61 | } | |
62 | ||
63 | void blk_mq_freeze_queue_start(struct request_queue *q) | |
64 | { | |
65 | int freeze_depth; | |
66 | ||
67 | freeze_depth = atomic_inc_return(&q->mq_freeze_depth); | |
68 | if (freeze_depth == 1) { | |
69 | percpu_ref_kill(&q->q_usage_counter); | |
70 | blk_mq_run_hw_queues(q, false); | |
71 | } | |
72 | } | |
73 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_start); | |
74 | ||
75 | static void blk_mq_freeze_queue_wait(struct request_queue *q) | |
76 | { | |
77 | wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter)); | |
78 | } | |
79 | ||
80 | /* | |
81 | * Guarantee no request is in use, so we can change any data structure of | |
82 | * the queue afterward. | |
83 | */ | |
84 | void blk_freeze_queue(struct request_queue *q) | |
85 | { | |
86 | /* | |
87 | * In the !blk_mq case we are only calling this to kill the | |
88 | * q_usage_counter, otherwise this increases the freeze depth | |
89 | * and waits for it to return to zero. For this reason there is | |
90 | * no blk_unfreeze_queue(), and blk_freeze_queue() is not | |
91 | * exported to drivers as the only user for unfreeze is blk_mq. | |
92 | */ | |
93 | blk_mq_freeze_queue_start(q); | |
94 | blk_mq_freeze_queue_wait(q); | |
95 | } | |
96 | ||
97 | void blk_mq_freeze_queue(struct request_queue *q) | |
98 | { | |
99 | /* | |
100 | * ...just an alias to keep freeze and unfreeze actions balanced | |
101 | * in the blk_mq_* namespace | |
102 | */ | |
103 | blk_freeze_queue(q); | |
104 | } | |
105 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue); | |
106 | ||
107 | void blk_mq_unfreeze_queue(struct request_queue *q) | |
108 | { | |
109 | int freeze_depth; | |
110 | ||
111 | freeze_depth = atomic_dec_return(&q->mq_freeze_depth); | |
112 | WARN_ON_ONCE(freeze_depth < 0); | |
113 | if (!freeze_depth) { | |
114 | percpu_ref_reinit(&q->q_usage_counter); | |
115 | wake_up_all(&q->mq_freeze_wq); | |
116 | } | |
117 | } | |
118 | EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue); | |
119 | ||
120 | /** | |
121 | * blk_mq_quiesce_queue() - wait until all ongoing queue_rq calls have finished | |
122 | * @q: request queue. | |
123 | * | |
124 | * Note: this function does not prevent that the struct request end_io() | |
125 | * callback function is invoked. Additionally, it is not prevented that | |
126 | * new queue_rq() calls occur unless the queue has been stopped first. | |
127 | */ | |
128 | void blk_mq_quiesce_queue(struct request_queue *q) | |
129 | { | |
130 | struct blk_mq_hw_ctx *hctx; | |
131 | unsigned int i; | |
132 | bool rcu = false; | |
133 | ||
134 | blk_mq_stop_hw_queues(q); | |
135 | ||
136 | queue_for_each_hw_ctx(q, hctx, i) { | |
137 | if (hctx->flags & BLK_MQ_F_BLOCKING) | |
138 | synchronize_srcu(&hctx->queue_rq_srcu); | |
139 | else | |
140 | rcu = true; | |
141 | } | |
142 | if (rcu) | |
143 | synchronize_rcu(); | |
144 | } | |
145 | EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue); | |
146 | ||
147 | void blk_mq_wake_waiters(struct request_queue *q) | |
148 | { | |
149 | struct blk_mq_hw_ctx *hctx; | |
150 | unsigned int i; | |
151 | ||
152 | queue_for_each_hw_ctx(q, hctx, i) | |
153 | if (blk_mq_hw_queue_mapped(hctx)) | |
154 | blk_mq_tag_wakeup_all(hctx->tags, true); | |
155 | ||
156 | /* | |
157 | * If we are called because the queue has now been marked as | |
158 | * dying, we need to ensure that processes currently waiting on | |
159 | * the queue are notified as well. | |
160 | */ | |
161 | wake_up_all(&q->mq_freeze_wq); | |
162 | } | |
163 | ||
164 | bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx) | |
165 | { | |
166 | return blk_mq_has_free_tags(hctx->tags); | |
167 | } | |
168 | EXPORT_SYMBOL(blk_mq_can_queue); | |
169 | ||
170 | static void blk_mq_rq_ctx_init(struct request_queue *q, struct blk_mq_ctx *ctx, | |
171 | struct request *rq, unsigned int op) | |
172 | { | |
173 | INIT_LIST_HEAD(&rq->queuelist); | |
174 | /* csd/requeue_work/fifo_time is initialized before use */ | |
175 | rq->q = q; | |
176 | rq->mq_ctx = ctx; | |
177 | rq->cmd_flags = op; | |
178 | if (blk_queue_io_stat(q)) | |
179 | rq->rq_flags |= RQF_IO_STAT; | |
180 | /* do not touch atomic flags, it needs atomic ops against the timer */ | |
181 | rq->cpu = -1; | |
182 | INIT_HLIST_NODE(&rq->hash); | |
183 | RB_CLEAR_NODE(&rq->rb_node); | |
184 | rq->rq_disk = NULL; | |
185 | rq->part = NULL; | |
186 | rq->start_time = jiffies; | |
187 | #ifdef CONFIG_BLK_CGROUP | |
188 | rq->rl = NULL; | |
189 | set_start_time_ns(rq); | |
190 | rq->io_start_time_ns = 0; | |
191 | #endif | |
192 | rq->nr_phys_segments = 0; | |
193 | #if defined(CONFIG_BLK_DEV_INTEGRITY) | |
194 | rq->nr_integrity_segments = 0; | |
195 | #endif | |
196 | rq->special = NULL; | |
197 | /* tag was already set */ | |
198 | rq->errors = 0; | |
199 | ||
200 | rq->cmd = rq->__cmd; | |
201 | ||
202 | rq->extra_len = 0; | |
203 | rq->sense_len = 0; | |
204 | rq->resid_len = 0; | |
205 | rq->sense = NULL; | |
206 | ||
207 | INIT_LIST_HEAD(&rq->timeout_list); | |
208 | rq->timeout = 0; | |
209 | ||
210 | rq->end_io = NULL; | |
211 | rq->end_io_data = NULL; | |
212 | rq->next_rq = NULL; | |
213 | ||
214 | ctx->rq_dispatched[op_is_sync(op)]++; | |
215 | } | |
216 | ||
217 | static struct request * | |
218 | __blk_mq_alloc_request(struct blk_mq_alloc_data *data, unsigned int op) | |
219 | { | |
220 | struct request *rq; | |
221 | unsigned int tag; | |
222 | ||
223 | tag = blk_mq_get_tag(data); | |
224 | if (tag != BLK_MQ_TAG_FAIL) { | |
225 | rq = data->hctx->tags->rqs[tag]; | |
226 | ||
227 | if (blk_mq_tag_busy(data->hctx)) { | |
228 | rq->rq_flags = RQF_MQ_INFLIGHT; | |
229 | atomic_inc(&data->hctx->nr_active); | |
230 | } | |
231 | ||
232 | rq->tag = tag; | |
233 | blk_mq_rq_ctx_init(data->q, data->ctx, rq, op); | |
234 | return rq; | |
235 | } | |
236 | ||
237 | return NULL; | |
238 | } | |
239 | ||
240 | struct request *blk_mq_alloc_request(struct request_queue *q, int rw, | |
241 | unsigned int flags) | |
242 | { | |
243 | struct blk_mq_ctx *ctx; | |
244 | struct blk_mq_hw_ctx *hctx; | |
245 | struct request *rq; | |
246 | struct blk_mq_alloc_data alloc_data; | |
247 | int ret; | |
248 | ||
249 | ret = blk_queue_enter(q, flags & BLK_MQ_REQ_NOWAIT); | |
250 | if (ret) | |
251 | return ERR_PTR(ret); | |
252 | ||
253 | ctx = blk_mq_get_ctx(q); | |
254 | hctx = blk_mq_map_queue(q, ctx->cpu); | |
255 | blk_mq_set_alloc_data(&alloc_data, q, flags, ctx, hctx); | |
256 | rq = __blk_mq_alloc_request(&alloc_data, rw); | |
257 | blk_mq_put_ctx(ctx); | |
258 | ||
259 | if (!rq) { | |
260 | blk_queue_exit(q); | |
261 | return ERR_PTR(-EWOULDBLOCK); | |
262 | } | |
263 | ||
264 | rq->__data_len = 0; | |
265 | rq->__sector = (sector_t) -1; | |
266 | rq->bio = rq->biotail = NULL; | |
267 | return rq; | |
268 | } | |
269 | EXPORT_SYMBOL(blk_mq_alloc_request); | |
270 | ||
271 | struct request *blk_mq_alloc_request_hctx(struct request_queue *q, int rw, | |
272 | unsigned int flags, unsigned int hctx_idx) | |
273 | { | |
274 | struct blk_mq_hw_ctx *hctx; | |
275 | struct blk_mq_ctx *ctx; | |
276 | struct request *rq; | |
277 | struct blk_mq_alloc_data alloc_data; | |
278 | int ret; | |
279 | ||
280 | /* | |
281 | * If the tag allocator sleeps we could get an allocation for a | |
282 | * different hardware context. No need to complicate the low level | |
283 | * allocator for this for the rare use case of a command tied to | |
284 | * a specific queue. | |
285 | */ | |
286 | if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT))) | |
287 | return ERR_PTR(-EINVAL); | |
288 | ||
289 | if (hctx_idx >= q->nr_hw_queues) | |
290 | return ERR_PTR(-EIO); | |
291 | ||
292 | ret = blk_queue_enter(q, true); | |
293 | if (ret) | |
294 | return ERR_PTR(ret); | |
295 | ||
296 | /* | |
297 | * Check if the hardware context is actually mapped to anything. | |
298 | * If not tell the caller that it should skip this queue. | |
299 | */ | |
300 | hctx = q->queue_hw_ctx[hctx_idx]; | |
301 | if (!blk_mq_hw_queue_mapped(hctx)) { | |
302 | ret = -EXDEV; | |
303 | goto out_queue_exit; | |
304 | } | |
305 | ctx = __blk_mq_get_ctx(q, cpumask_first(hctx->cpumask)); | |
306 | ||
307 | blk_mq_set_alloc_data(&alloc_data, q, flags, ctx, hctx); | |
308 | rq = __blk_mq_alloc_request(&alloc_data, rw); | |
309 | if (!rq) { | |
310 | ret = -EWOULDBLOCK; | |
311 | goto out_queue_exit; | |
312 | } | |
313 | ||
314 | return rq; | |
315 | ||
316 | out_queue_exit: | |
317 | blk_queue_exit(q); | |
318 | return ERR_PTR(ret); | |
319 | } | |
320 | EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx); | |
321 | ||
322 | static void __blk_mq_free_request(struct blk_mq_hw_ctx *hctx, | |
323 | struct blk_mq_ctx *ctx, struct request *rq) | |
324 | { | |
325 | const int tag = rq->tag; | |
326 | struct request_queue *q = rq->q; | |
327 | ||
328 | if (rq->rq_flags & RQF_MQ_INFLIGHT) | |
329 | atomic_dec(&hctx->nr_active); | |
330 | ||
331 | wbt_done(q->rq_wb, &rq->issue_stat); | |
332 | rq->rq_flags = 0; | |
333 | ||
334 | clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags); | |
335 | clear_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags); | |
336 | blk_mq_put_tag(hctx, ctx, tag); | |
337 | blk_queue_exit(q); | |
338 | } | |
339 | ||
340 | void blk_mq_free_hctx_request(struct blk_mq_hw_ctx *hctx, struct request *rq) | |
341 | { | |
342 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
343 | ||
344 | ctx->rq_completed[rq_is_sync(rq)]++; | |
345 | __blk_mq_free_request(hctx, ctx, rq); | |
346 | ||
347 | } | |
348 | EXPORT_SYMBOL_GPL(blk_mq_free_hctx_request); | |
349 | ||
350 | void blk_mq_free_request(struct request *rq) | |
351 | { | |
352 | blk_mq_free_hctx_request(blk_mq_map_queue(rq->q, rq->mq_ctx->cpu), rq); | |
353 | } | |
354 | EXPORT_SYMBOL_GPL(blk_mq_free_request); | |
355 | ||
356 | inline void __blk_mq_end_request(struct request *rq, int error) | |
357 | { | |
358 | blk_account_io_done(rq); | |
359 | ||
360 | if (rq->end_io) { | |
361 | wbt_done(rq->q->rq_wb, &rq->issue_stat); | |
362 | rq->end_io(rq, error); | |
363 | } else { | |
364 | if (unlikely(blk_bidi_rq(rq))) | |
365 | blk_mq_free_request(rq->next_rq); | |
366 | blk_mq_free_request(rq); | |
367 | } | |
368 | } | |
369 | EXPORT_SYMBOL(__blk_mq_end_request); | |
370 | ||
371 | void blk_mq_end_request(struct request *rq, int error) | |
372 | { | |
373 | if (blk_update_request(rq, error, blk_rq_bytes(rq))) | |
374 | BUG(); | |
375 | __blk_mq_end_request(rq, error); | |
376 | } | |
377 | EXPORT_SYMBOL(blk_mq_end_request); | |
378 | ||
379 | static void __blk_mq_complete_request_remote(void *data) | |
380 | { | |
381 | struct request *rq = data; | |
382 | ||
383 | rq->q->softirq_done_fn(rq); | |
384 | } | |
385 | ||
386 | static void blk_mq_ipi_complete_request(struct request *rq) | |
387 | { | |
388 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
389 | bool shared = false; | |
390 | int cpu; | |
391 | ||
392 | if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) { | |
393 | rq->q->softirq_done_fn(rq); | |
394 | return; | |
395 | } | |
396 | ||
397 | cpu = get_cpu(); | |
398 | if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags)) | |
399 | shared = cpus_share_cache(cpu, ctx->cpu); | |
400 | ||
401 | if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) { | |
402 | rq->csd.func = __blk_mq_complete_request_remote; | |
403 | rq->csd.info = rq; | |
404 | rq->csd.flags = 0; | |
405 | smp_call_function_single_async(ctx->cpu, &rq->csd); | |
406 | } else { | |
407 | rq->q->softirq_done_fn(rq); | |
408 | } | |
409 | put_cpu(); | |
410 | } | |
411 | ||
412 | static void blk_mq_stat_add(struct request *rq) | |
413 | { | |
414 | if (rq->rq_flags & RQF_STATS) { | |
415 | /* | |
416 | * We could rq->mq_ctx here, but there's less of a risk | |
417 | * of races if we have the completion event add the stats | |
418 | * to the local software queue. | |
419 | */ | |
420 | struct blk_mq_ctx *ctx; | |
421 | ||
422 | ctx = __blk_mq_get_ctx(rq->q, raw_smp_processor_id()); | |
423 | blk_stat_add(&ctx->stat[rq_data_dir(rq)], rq); | |
424 | } | |
425 | } | |
426 | ||
427 | static void __blk_mq_complete_request(struct request *rq) | |
428 | { | |
429 | struct request_queue *q = rq->q; | |
430 | ||
431 | blk_mq_stat_add(rq); | |
432 | ||
433 | if (!q->softirq_done_fn) | |
434 | blk_mq_end_request(rq, rq->errors); | |
435 | else | |
436 | blk_mq_ipi_complete_request(rq); | |
437 | } | |
438 | ||
439 | /** | |
440 | * blk_mq_complete_request - end I/O on a request | |
441 | * @rq: the request being processed | |
442 | * | |
443 | * Description: | |
444 | * Ends all I/O on a request. It does not handle partial completions. | |
445 | * The actual completion happens out-of-order, through a IPI handler. | |
446 | **/ | |
447 | void blk_mq_complete_request(struct request *rq, int error) | |
448 | { | |
449 | struct request_queue *q = rq->q; | |
450 | ||
451 | if (unlikely(blk_should_fake_timeout(q))) | |
452 | return; | |
453 | if (!blk_mark_rq_complete(rq)) { | |
454 | rq->errors = error; | |
455 | __blk_mq_complete_request(rq); | |
456 | } | |
457 | } | |
458 | EXPORT_SYMBOL(blk_mq_complete_request); | |
459 | ||
460 | int blk_mq_request_started(struct request *rq) | |
461 | { | |
462 | return test_bit(REQ_ATOM_STARTED, &rq->atomic_flags); | |
463 | } | |
464 | EXPORT_SYMBOL_GPL(blk_mq_request_started); | |
465 | ||
466 | void blk_mq_start_request(struct request *rq) | |
467 | { | |
468 | struct request_queue *q = rq->q; | |
469 | ||
470 | trace_block_rq_issue(q, rq); | |
471 | ||
472 | rq->resid_len = blk_rq_bytes(rq); | |
473 | if (unlikely(blk_bidi_rq(rq))) | |
474 | rq->next_rq->resid_len = blk_rq_bytes(rq->next_rq); | |
475 | ||
476 | if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) { | |
477 | blk_stat_set_issue_time(&rq->issue_stat); | |
478 | rq->rq_flags |= RQF_STATS; | |
479 | wbt_issue(q->rq_wb, &rq->issue_stat); | |
480 | } | |
481 | ||
482 | blk_add_timer(rq); | |
483 | ||
484 | /* | |
485 | * Ensure that ->deadline is visible before set the started | |
486 | * flag and clear the completed flag. | |
487 | */ | |
488 | smp_mb__before_atomic(); | |
489 | ||
490 | /* | |
491 | * Mark us as started and clear complete. Complete might have been | |
492 | * set if requeue raced with timeout, which then marked it as | |
493 | * complete. So be sure to clear complete again when we start | |
494 | * the request, otherwise we'll ignore the completion event. | |
495 | */ | |
496 | if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) | |
497 | set_bit(REQ_ATOM_STARTED, &rq->atomic_flags); | |
498 | if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags)) | |
499 | clear_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags); | |
500 | ||
501 | if (q->dma_drain_size && blk_rq_bytes(rq)) { | |
502 | /* | |
503 | * Make sure space for the drain appears. We know we can do | |
504 | * this because max_hw_segments has been adjusted to be one | |
505 | * fewer than the device can handle. | |
506 | */ | |
507 | rq->nr_phys_segments++; | |
508 | } | |
509 | } | |
510 | EXPORT_SYMBOL(blk_mq_start_request); | |
511 | ||
512 | static void __blk_mq_requeue_request(struct request *rq) | |
513 | { | |
514 | struct request_queue *q = rq->q; | |
515 | ||
516 | trace_block_rq_requeue(q, rq); | |
517 | wbt_requeue(q->rq_wb, &rq->issue_stat); | |
518 | ||
519 | if (test_and_clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) { | |
520 | if (q->dma_drain_size && blk_rq_bytes(rq)) | |
521 | rq->nr_phys_segments--; | |
522 | } | |
523 | } | |
524 | ||
525 | void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list) | |
526 | { | |
527 | __blk_mq_requeue_request(rq); | |
528 | ||
529 | BUG_ON(blk_queued_rq(rq)); | |
530 | blk_mq_add_to_requeue_list(rq, true, kick_requeue_list); | |
531 | } | |
532 | EXPORT_SYMBOL(blk_mq_requeue_request); | |
533 | ||
534 | static void blk_mq_requeue_work(struct work_struct *work) | |
535 | { | |
536 | struct request_queue *q = | |
537 | container_of(work, struct request_queue, requeue_work.work); | |
538 | LIST_HEAD(rq_list); | |
539 | struct request *rq, *next; | |
540 | unsigned long flags; | |
541 | ||
542 | spin_lock_irqsave(&q->requeue_lock, flags); | |
543 | list_splice_init(&q->requeue_list, &rq_list); | |
544 | spin_unlock_irqrestore(&q->requeue_lock, flags); | |
545 | ||
546 | list_for_each_entry_safe(rq, next, &rq_list, queuelist) { | |
547 | if (!(rq->rq_flags & RQF_SOFTBARRIER)) | |
548 | continue; | |
549 | ||
550 | rq->rq_flags &= ~RQF_SOFTBARRIER; | |
551 | list_del_init(&rq->queuelist); | |
552 | blk_mq_insert_request(rq, true, false, false); | |
553 | } | |
554 | ||
555 | while (!list_empty(&rq_list)) { | |
556 | rq = list_entry(rq_list.next, struct request, queuelist); | |
557 | list_del_init(&rq->queuelist); | |
558 | blk_mq_insert_request(rq, false, false, false); | |
559 | } | |
560 | ||
561 | blk_mq_run_hw_queues(q, false); | |
562 | } | |
563 | ||
564 | void blk_mq_add_to_requeue_list(struct request *rq, bool at_head, | |
565 | bool kick_requeue_list) | |
566 | { | |
567 | struct request_queue *q = rq->q; | |
568 | unsigned long flags; | |
569 | ||
570 | /* | |
571 | * We abuse this flag that is otherwise used by the I/O scheduler to | |
572 | * request head insertation from the workqueue. | |
573 | */ | |
574 | BUG_ON(rq->rq_flags & RQF_SOFTBARRIER); | |
575 | ||
576 | spin_lock_irqsave(&q->requeue_lock, flags); | |
577 | if (at_head) { | |
578 | rq->rq_flags |= RQF_SOFTBARRIER; | |
579 | list_add(&rq->queuelist, &q->requeue_list); | |
580 | } else { | |
581 | list_add_tail(&rq->queuelist, &q->requeue_list); | |
582 | } | |
583 | spin_unlock_irqrestore(&q->requeue_lock, flags); | |
584 | ||
585 | if (kick_requeue_list) | |
586 | blk_mq_kick_requeue_list(q); | |
587 | } | |
588 | EXPORT_SYMBOL(blk_mq_add_to_requeue_list); | |
589 | ||
590 | void blk_mq_kick_requeue_list(struct request_queue *q) | |
591 | { | |
592 | kblockd_schedule_delayed_work(&q->requeue_work, 0); | |
593 | } | |
594 | EXPORT_SYMBOL(blk_mq_kick_requeue_list); | |
595 | ||
596 | void blk_mq_delay_kick_requeue_list(struct request_queue *q, | |
597 | unsigned long msecs) | |
598 | { | |
599 | kblockd_schedule_delayed_work(&q->requeue_work, | |
600 | msecs_to_jiffies(msecs)); | |
601 | } | |
602 | EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list); | |
603 | ||
604 | void blk_mq_abort_requeue_list(struct request_queue *q) | |
605 | { | |
606 | unsigned long flags; | |
607 | LIST_HEAD(rq_list); | |
608 | ||
609 | spin_lock_irqsave(&q->requeue_lock, flags); | |
610 | list_splice_init(&q->requeue_list, &rq_list); | |
611 | spin_unlock_irqrestore(&q->requeue_lock, flags); | |
612 | ||
613 | while (!list_empty(&rq_list)) { | |
614 | struct request *rq; | |
615 | ||
616 | rq = list_first_entry(&rq_list, struct request, queuelist); | |
617 | list_del_init(&rq->queuelist); | |
618 | rq->errors = -EIO; | |
619 | blk_mq_end_request(rq, rq->errors); | |
620 | } | |
621 | } | |
622 | EXPORT_SYMBOL(blk_mq_abort_requeue_list); | |
623 | ||
624 | struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag) | |
625 | { | |
626 | if (tag < tags->nr_tags) { | |
627 | prefetch(tags->rqs[tag]); | |
628 | return tags->rqs[tag]; | |
629 | } | |
630 | ||
631 | return NULL; | |
632 | } | |
633 | EXPORT_SYMBOL(blk_mq_tag_to_rq); | |
634 | ||
635 | struct blk_mq_timeout_data { | |
636 | unsigned long next; | |
637 | unsigned int next_set; | |
638 | }; | |
639 | ||
640 | void blk_mq_rq_timed_out(struct request *req, bool reserved) | |
641 | { | |
642 | struct blk_mq_ops *ops = req->q->mq_ops; | |
643 | enum blk_eh_timer_return ret = BLK_EH_RESET_TIMER; | |
644 | ||
645 | /* | |
646 | * We know that complete is set at this point. If STARTED isn't set | |
647 | * anymore, then the request isn't active and the "timeout" should | |
648 | * just be ignored. This can happen due to the bitflag ordering. | |
649 | * Timeout first checks if STARTED is set, and if it is, assumes | |
650 | * the request is active. But if we race with completion, then | |
651 | * we both flags will get cleared. So check here again, and ignore | |
652 | * a timeout event with a request that isn't active. | |
653 | */ | |
654 | if (!test_bit(REQ_ATOM_STARTED, &req->atomic_flags)) | |
655 | return; | |
656 | ||
657 | if (ops->timeout) | |
658 | ret = ops->timeout(req, reserved); | |
659 | ||
660 | switch (ret) { | |
661 | case BLK_EH_HANDLED: | |
662 | __blk_mq_complete_request(req); | |
663 | break; | |
664 | case BLK_EH_RESET_TIMER: | |
665 | blk_add_timer(req); | |
666 | blk_clear_rq_complete(req); | |
667 | break; | |
668 | case BLK_EH_NOT_HANDLED: | |
669 | break; | |
670 | default: | |
671 | printk(KERN_ERR "block: bad eh return: %d\n", ret); | |
672 | break; | |
673 | } | |
674 | } | |
675 | ||
676 | static void blk_mq_check_expired(struct blk_mq_hw_ctx *hctx, | |
677 | struct request *rq, void *priv, bool reserved) | |
678 | { | |
679 | struct blk_mq_timeout_data *data = priv; | |
680 | ||
681 | if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) { | |
682 | /* | |
683 | * If a request wasn't started before the queue was | |
684 | * marked dying, kill it here or it'll go unnoticed. | |
685 | */ | |
686 | if (unlikely(blk_queue_dying(rq->q))) { | |
687 | rq->errors = -EIO; | |
688 | blk_mq_end_request(rq, rq->errors); | |
689 | } | |
690 | return; | |
691 | } | |
692 | ||
693 | if (time_after_eq(jiffies, rq->deadline)) { | |
694 | if (!blk_mark_rq_complete(rq)) | |
695 | blk_mq_rq_timed_out(rq, reserved); | |
696 | } else if (!data->next_set || time_after(data->next, rq->deadline)) { | |
697 | data->next = rq->deadline; | |
698 | data->next_set = 1; | |
699 | } | |
700 | } | |
701 | ||
702 | static void blk_mq_timeout_work(struct work_struct *work) | |
703 | { | |
704 | struct request_queue *q = | |
705 | container_of(work, struct request_queue, timeout_work); | |
706 | struct blk_mq_timeout_data data = { | |
707 | .next = 0, | |
708 | .next_set = 0, | |
709 | }; | |
710 | int i; | |
711 | ||
712 | /* A deadlock might occur if a request is stuck requiring a | |
713 | * timeout at the same time a queue freeze is waiting | |
714 | * completion, since the timeout code would not be able to | |
715 | * acquire the queue reference here. | |
716 | * | |
717 | * That's why we don't use blk_queue_enter here; instead, we use | |
718 | * percpu_ref_tryget directly, because we need to be able to | |
719 | * obtain a reference even in the short window between the queue | |
720 | * starting to freeze, by dropping the first reference in | |
721 | * blk_mq_freeze_queue_start, and the moment the last request is | |
722 | * consumed, marked by the instant q_usage_counter reaches | |
723 | * zero. | |
724 | */ | |
725 | if (!percpu_ref_tryget(&q->q_usage_counter)) | |
726 | return; | |
727 | ||
728 | blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &data); | |
729 | ||
730 | if (data.next_set) { | |
731 | data.next = blk_rq_timeout(round_jiffies_up(data.next)); | |
732 | mod_timer(&q->timeout, data.next); | |
733 | } else { | |
734 | struct blk_mq_hw_ctx *hctx; | |
735 | ||
736 | queue_for_each_hw_ctx(q, hctx, i) { | |
737 | /* the hctx may be unmapped, so check it here */ | |
738 | if (blk_mq_hw_queue_mapped(hctx)) | |
739 | blk_mq_tag_idle(hctx); | |
740 | } | |
741 | } | |
742 | blk_queue_exit(q); | |
743 | } | |
744 | ||
745 | /* | |
746 | * Reverse check our software queue for entries that we could potentially | |
747 | * merge with. Currently includes a hand-wavy stop count of 8, to not spend | |
748 | * too much time checking for merges. | |
749 | */ | |
750 | static bool blk_mq_attempt_merge(struct request_queue *q, | |
751 | struct blk_mq_ctx *ctx, struct bio *bio) | |
752 | { | |
753 | struct request *rq; | |
754 | int checked = 8; | |
755 | ||
756 | list_for_each_entry_reverse(rq, &ctx->rq_list, queuelist) { | |
757 | int el_ret; | |
758 | ||
759 | if (!checked--) | |
760 | break; | |
761 | ||
762 | if (!blk_rq_merge_ok(rq, bio)) | |
763 | continue; | |
764 | ||
765 | el_ret = blk_try_merge(rq, bio); | |
766 | if (el_ret == ELEVATOR_BACK_MERGE) { | |
767 | if (bio_attempt_back_merge(q, rq, bio)) { | |
768 | ctx->rq_merged++; | |
769 | return true; | |
770 | } | |
771 | break; | |
772 | } else if (el_ret == ELEVATOR_FRONT_MERGE) { | |
773 | if (bio_attempt_front_merge(q, rq, bio)) { | |
774 | ctx->rq_merged++; | |
775 | return true; | |
776 | } | |
777 | break; | |
778 | } | |
779 | } | |
780 | ||
781 | return false; | |
782 | } | |
783 | ||
784 | struct flush_busy_ctx_data { | |
785 | struct blk_mq_hw_ctx *hctx; | |
786 | struct list_head *list; | |
787 | }; | |
788 | ||
789 | static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data) | |
790 | { | |
791 | struct flush_busy_ctx_data *flush_data = data; | |
792 | struct blk_mq_hw_ctx *hctx = flush_data->hctx; | |
793 | struct blk_mq_ctx *ctx = hctx->ctxs[bitnr]; | |
794 | ||
795 | sbitmap_clear_bit(sb, bitnr); | |
796 | spin_lock(&ctx->lock); | |
797 | list_splice_tail_init(&ctx->rq_list, flush_data->list); | |
798 | spin_unlock(&ctx->lock); | |
799 | return true; | |
800 | } | |
801 | ||
802 | /* | |
803 | * Process software queues that have been marked busy, splicing them | |
804 | * to the for-dispatch | |
805 | */ | |
806 | static void flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list) | |
807 | { | |
808 | struct flush_busy_ctx_data data = { | |
809 | .hctx = hctx, | |
810 | .list = list, | |
811 | }; | |
812 | ||
813 | sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data); | |
814 | } | |
815 | ||
816 | static inline unsigned int queued_to_index(unsigned int queued) | |
817 | { | |
818 | if (!queued) | |
819 | return 0; | |
820 | ||
821 | return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1); | |
822 | } | |
823 | ||
824 | /* | |
825 | * Run this hardware queue, pulling any software queues mapped to it in. | |
826 | * Note that this function currently has various problems around ordering | |
827 | * of IO. In particular, we'd like FIFO behaviour on handling existing | |
828 | * items on the hctx->dispatch list. Ignore that for now. | |
829 | */ | |
830 | static void blk_mq_process_rq_list(struct blk_mq_hw_ctx *hctx) | |
831 | { | |
832 | struct request_queue *q = hctx->queue; | |
833 | struct request *rq; | |
834 | LIST_HEAD(rq_list); | |
835 | LIST_HEAD(driver_list); | |
836 | struct list_head *dptr; | |
837 | int queued; | |
838 | ||
839 | if (unlikely(blk_mq_hctx_stopped(hctx))) | |
840 | return; | |
841 | ||
842 | hctx->run++; | |
843 | ||
844 | /* | |
845 | * Touch any software queue that has pending entries. | |
846 | */ | |
847 | flush_busy_ctxs(hctx, &rq_list); | |
848 | ||
849 | /* | |
850 | * If we have previous entries on our dispatch list, grab them | |
851 | * and stuff them at the front for more fair dispatch. | |
852 | */ | |
853 | if (!list_empty_careful(&hctx->dispatch)) { | |
854 | spin_lock(&hctx->lock); | |
855 | if (!list_empty(&hctx->dispatch)) | |
856 | list_splice_init(&hctx->dispatch, &rq_list); | |
857 | spin_unlock(&hctx->lock); | |
858 | } | |
859 | ||
860 | /* | |
861 | * Start off with dptr being NULL, so we start the first request | |
862 | * immediately, even if we have more pending. | |
863 | */ | |
864 | dptr = NULL; | |
865 | ||
866 | /* | |
867 | * Now process all the entries, sending them to the driver. | |
868 | */ | |
869 | queued = 0; | |
870 | while (!list_empty(&rq_list)) { | |
871 | struct blk_mq_queue_data bd; | |
872 | int ret; | |
873 | ||
874 | rq = list_first_entry(&rq_list, struct request, queuelist); | |
875 | list_del_init(&rq->queuelist); | |
876 | ||
877 | bd.rq = rq; | |
878 | bd.list = dptr; | |
879 | bd.last = list_empty(&rq_list); | |
880 | ||
881 | ret = q->mq_ops->queue_rq(hctx, &bd); | |
882 | switch (ret) { | |
883 | case BLK_MQ_RQ_QUEUE_OK: | |
884 | queued++; | |
885 | break; | |
886 | case BLK_MQ_RQ_QUEUE_BUSY: | |
887 | list_add(&rq->queuelist, &rq_list); | |
888 | __blk_mq_requeue_request(rq); | |
889 | break; | |
890 | default: | |
891 | pr_err("blk-mq: bad return on queue: %d\n", ret); | |
892 | case BLK_MQ_RQ_QUEUE_ERROR: | |
893 | rq->errors = -EIO; | |
894 | blk_mq_end_request(rq, rq->errors); | |
895 | break; | |
896 | } | |
897 | ||
898 | if (ret == BLK_MQ_RQ_QUEUE_BUSY) | |
899 | break; | |
900 | ||
901 | /* | |
902 | * We've done the first request. If we have more than 1 | |
903 | * left in the list, set dptr to defer issue. | |
904 | */ | |
905 | if (!dptr && rq_list.next != rq_list.prev) | |
906 | dptr = &driver_list; | |
907 | } | |
908 | ||
909 | hctx->dispatched[queued_to_index(queued)]++; | |
910 | ||
911 | /* | |
912 | * Any items that need requeuing? Stuff them into hctx->dispatch, | |
913 | * that is where we will continue on next queue run. | |
914 | */ | |
915 | if (!list_empty(&rq_list)) { | |
916 | spin_lock(&hctx->lock); | |
917 | list_splice(&rq_list, &hctx->dispatch); | |
918 | spin_unlock(&hctx->lock); | |
919 | /* | |
920 | * the queue is expected stopped with BLK_MQ_RQ_QUEUE_BUSY, but | |
921 | * it's possible the queue is stopped and restarted again | |
922 | * before this. Queue restart will dispatch requests. And since | |
923 | * requests in rq_list aren't added into hctx->dispatch yet, | |
924 | * the requests in rq_list might get lost. | |
925 | * | |
926 | * blk_mq_run_hw_queue() already checks the STOPPED bit | |
927 | **/ | |
928 | blk_mq_run_hw_queue(hctx, true); | |
929 | } | |
930 | } | |
931 | ||
932 | static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx) | |
933 | { | |
934 | int srcu_idx; | |
935 | ||
936 | WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) && | |
937 | cpu_online(hctx->next_cpu)); | |
938 | ||
939 | if (!(hctx->flags & BLK_MQ_F_BLOCKING)) { | |
940 | rcu_read_lock(); | |
941 | blk_mq_process_rq_list(hctx); | |
942 | rcu_read_unlock(); | |
943 | } else { | |
944 | srcu_idx = srcu_read_lock(&hctx->queue_rq_srcu); | |
945 | blk_mq_process_rq_list(hctx); | |
946 | srcu_read_unlock(&hctx->queue_rq_srcu, srcu_idx); | |
947 | } | |
948 | } | |
949 | ||
950 | /* | |
951 | * It'd be great if the workqueue API had a way to pass | |
952 | * in a mask and had some smarts for more clever placement. | |
953 | * For now we just round-robin here, switching for every | |
954 | * BLK_MQ_CPU_WORK_BATCH queued items. | |
955 | */ | |
956 | static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx) | |
957 | { | |
958 | if (hctx->queue->nr_hw_queues == 1) | |
959 | return WORK_CPU_UNBOUND; | |
960 | ||
961 | if (--hctx->next_cpu_batch <= 0) { | |
962 | int next_cpu; | |
963 | ||
964 | next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask); | |
965 | if (next_cpu >= nr_cpu_ids) | |
966 | next_cpu = cpumask_first(hctx->cpumask); | |
967 | ||
968 | hctx->next_cpu = next_cpu; | |
969 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; | |
970 | } | |
971 | ||
972 | return hctx->next_cpu; | |
973 | } | |
974 | ||
975 | void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) | |
976 | { | |
977 | if (unlikely(blk_mq_hctx_stopped(hctx) || | |
978 | !blk_mq_hw_queue_mapped(hctx))) | |
979 | return; | |
980 | ||
981 | if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) { | |
982 | int cpu = get_cpu(); | |
983 | if (cpumask_test_cpu(cpu, hctx->cpumask)) { | |
984 | __blk_mq_run_hw_queue(hctx); | |
985 | put_cpu(); | |
986 | return; | |
987 | } | |
988 | ||
989 | put_cpu(); | |
990 | } | |
991 | ||
992 | kblockd_schedule_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work); | |
993 | } | |
994 | ||
995 | void blk_mq_run_hw_queues(struct request_queue *q, bool async) | |
996 | { | |
997 | struct blk_mq_hw_ctx *hctx; | |
998 | int i; | |
999 | ||
1000 | queue_for_each_hw_ctx(q, hctx, i) { | |
1001 | if ((!blk_mq_hctx_has_pending(hctx) && | |
1002 | list_empty_careful(&hctx->dispatch)) || | |
1003 | blk_mq_hctx_stopped(hctx)) | |
1004 | continue; | |
1005 | ||
1006 | blk_mq_run_hw_queue(hctx, async); | |
1007 | } | |
1008 | } | |
1009 | EXPORT_SYMBOL(blk_mq_run_hw_queues); | |
1010 | ||
1011 | /** | |
1012 | * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped | |
1013 | * @q: request queue. | |
1014 | * | |
1015 | * The caller is responsible for serializing this function against | |
1016 | * blk_mq_{start,stop}_hw_queue(). | |
1017 | */ | |
1018 | bool blk_mq_queue_stopped(struct request_queue *q) | |
1019 | { | |
1020 | struct blk_mq_hw_ctx *hctx; | |
1021 | int i; | |
1022 | ||
1023 | queue_for_each_hw_ctx(q, hctx, i) | |
1024 | if (blk_mq_hctx_stopped(hctx)) | |
1025 | return true; | |
1026 | ||
1027 | return false; | |
1028 | } | |
1029 | EXPORT_SYMBOL(blk_mq_queue_stopped); | |
1030 | ||
1031 | void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx) | |
1032 | { | |
1033 | cancel_work(&hctx->run_work); | |
1034 | cancel_delayed_work(&hctx->delay_work); | |
1035 | set_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
1036 | } | |
1037 | EXPORT_SYMBOL(blk_mq_stop_hw_queue); | |
1038 | ||
1039 | void blk_mq_stop_hw_queues(struct request_queue *q) | |
1040 | { | |
1041 | struct blk_mq_hw_ctx *hctx; | |
1042 | int i; | |
1043 | ||
1044 | queue_for_each_hw_ctx(q, hctx, i) | |
1045 | blk_mq_stop_hw_queue(hctx); | |
1046 | } | |
1047 | EXPORT_SYMBOL(blk_mq_stop_hw_queues); | |
1048 | ||
1049 | void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx) | |
1050 | { | |
1051 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
1052 | ||
1053 | blk_mq_run_hw_queue(hctx, false); | |
1054 | } | |
1055 | EXPORT_SYMBOL(blk_mq_start_hw_queue); | |
1056 | ||
1057 | void blk_mq_start_hw_queues(struct request_queue *q) | |
1058 | { | |
1059 | struct blk_mq_hw_ctx *hctx; | |
1060 | int i; | |
1061 | ||
1062 | queue_for_each_hw_ctx(q, hctx, i) | |
1063 | blk_mq_start_hw_queue(hctx); | |
1064 | } | |
1065 | EXPORT_SYMBOL(blk_mq_start_hw_queues); | |
1066 | ||
1067 | void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async) | |
1068 | { | |
1069 | struct blk_mq_hw_ctx *hctx; | |
1070 | int i; | |
1071 | ||
1072 | queue_for_each_hw_ctx(q, hctx, i) { | |
1073 | if (!blk_mq_hctx_stopped(hctx)) | |
1074 | continue; | |
1075 | ||
1076 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
1077 | blk_mq_run_hw_queue(hctx, async); | |
1078 | } | |
1079 | } | |
1080 | EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues); | |
1081 | ||
1082 | static void blk_mq_run_work_fn(struct work_struct *work) | |
1083 | { | |
1084 | struct blk_mq_hw_ctx *hctx; | |
1085 | ||
1086 | hctx = container_of(work, struct blk_mq_hw_ctx, run_work); | |
1087 | ||
1088 | __blk_mq_run_hw_queue(hctx); | |
1089 | } | |
1090 | ||
1091 | static void blk_mq_delay_work_fn(struct work_struct *work) | |
1092 | { | |
1093 | struct blk_mq_hw_ctx *hctx; | |
1094 | ||
1095 | hctx = container_of(work, struct blk_mq_hw_ctx, delay_work.work); | |
1096 | ||
1097 | if (test_and_clear_bit(BLK_MQ_S_STOPPED, &hctx->state)) | |
1098 | __blk_mq_run_hw_queue(hctx); | |
1099 | } | |
1100 | ||
1101 | void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs) | |
1102 | { | |
1103 | if (unlikely(!blk_mq_hw_queue_mapped(hctx))) | |
1104 | return; | |
1105 | ||
1106 | kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx), | |
1107 | &hctx->delay_work, msecs_to_jiffies(msecs)); | |
1108 | } | |
1109 | EXPORT_SYMBOL(blk_mq_delay_queue); | |
1110 | ||
1111 | static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx, | |
1112 | struct request *rq, | |
1113 | bool at_head) | |
1114 | { | |
1115 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
1116 | ||
1117 | trace_block_rq_insert(hctx->queue, rq); | |
1118 | ||
1119 | if (at_head) | |
1120 | list_add(&rq->queuelist, &ctx->rq_list); | |
1121 | else | |
1122 | list_add_tail(&rq->queuelist, &ctx->rq_list); | |
1123 | } | |
1124 | ||
1125 | static void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, | |
1126 | struct request *rq, bool at_head) | |
1127 | { | |
1128 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
1129 | ||
1130 | __blk_mq_insert_req_list(hctx, rq, at_head); | |
1131 | blk_mq_hctx_mark_pending(hctx, ctx); | |
1132 | } | |
1133 | ||
1134 | void blk_mq_insert_request(struct request *rq, bool at_head, bool run_queue, | |
1135 | bool async) | |
1136 | { | |
1137 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
1138 | struct request_queue *q = rq->q; | |
1139 | struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, ctx->cpu); | |
1140 | ||
1141 | spin_lock(&ctx->lock); | |
1142 | __blk_mq_insert_request(hctx, rq, at_head); | |
1143 | spin_unlock(&ctx->lock); | |
1144 | ||
1145 | if (run_queue) | |
1146 | blk_mq_run_hw_queue(hctx, async); | |
1147 | } | |
1148 | ||
1149 | static void blk_mq_insert_requests(struct request_queue *q, | |
1150 | struct blk_mq_ctx *ctx, | |
1151 | struct list_head *list, | |
1152 | int depth, | |
1153 | bool from_schedule) | |
1154 | ||
1155 | { | |
1156 | struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, ctx->cpu); | |
1157 | ||
1158 | trace_block_unplug(q, depth, !from_schedule); | |
1159 | ||
1160 | /* | |
1161 | * preemption doesn't flush plug list, so it's possible ctx->cpu is | |
1162 | * offline now | |
1163 | */ | |
1164 | spin_lock(&ctx->lock); | |
1165 | while (!list_empty(list)) { | |
1166 | struct request *rq; | |
1167 | ||
1168 | rq = list_first_entry(list, struct request, queuelist); | |
1169 | BUG_ON(rq->mq_ctx != ctx); | |
1170 | list_del_init(&rq->queuelist); | |
1171 | __blk_mq_insert_req_list(hctx, rq, false); | |
1172 | } | |
1173 | blk_mq_hctx_mark_pending(hctx, ctx); | |
1174 | spin_unlock(&ctx->lock); | |
1175 | ||
1176 | blk_mq_run_hw_queue(hctx, from_schedule); | |
1177 | } | |
1178 | ||
1179 | static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b) | |
1180 | { | |
1181 | struct request *rqa = container_of(a, struct request, queuelist); | |
1182 | struct request *rqb = container_of(b, struct request, queuelist); | |
1183 | ||
1184 | return !(rqa->mq_ctx < rqb->mq_ctx || | |
1185 | (rqa->mq_ctx == rqb->mq_ctx && | |
1186 | blk_rq_pos(rqa) < blk_rq_pos(rqb))); | |
1187 | } | |
1188 | ||
1189 | void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule) | |
1190 | { | |
1191 | struct blk_mq_ctx *this_ctx; | |
1192 | struct request_queue *this_q; | |
1193 | struct request *rq; | |
1194 | LIST_HEAD(list); | |
1195 | LIST_HEAD(ctx_list); | |
1196 | unsigned int depth; | |
1197 | ||
1198 | list_splice_init(&plug->mq_list, &list); | |
1199 | ||
1200 | list_sort(NULL, &list, plug_ctx_cmp); | |
1201 | ||
1202 | this_q = NULL; | |
1203 | this_ctx = NULL; | |
1204 | depth = 0; | |
1205 | ||
1206 | while (!list_empty(&list)) { | |
1207 | rq = list_entry_rq(list.next); | |
1208 | list_del_init(&rq->queuelist); | |
1209 | BUG_ON(!rq->q); | |
1210 | if (rq->mq_ctx != this_ctx) { | |
1211 | if (this_ctx) { | |
1212 | blk_mq_insert_requests(this_q, this_ctx, | |
1213 | &ctx_list, depth, | |
1214 | from_schedule); | |
1215 | } | |
1216 | ||
1217 | this_ctx = rq->mq_ctx; | |
1218 | this_q = rq->q; | |
1219 | depth = 0; | |
1220 | } | |
1221 | ||
1222 | depth++; | |
1223 | list_add_tail(&rq->queuelist, &ctx_list); | |
1224 | } | |
1225 | ||
1226 | /* | |
1227 | * If 'this_ctx' is set, we know we have entries to complete | |
1228 | * on 'ctx_list'. Do those. | |
1229 | */ | |
1230 | if (this_ctx) { | |
1231 | blk_mq_insert_requests(this_q, this_ctx, &ctx_list, depth, | |
1232 | from_schedule); | |
1233 | } | |
1234 | } | |
1235 | ||
1236 | static void blk_mq_bio_to_request(struct request *rq, struct bio *bio) | |
1237 | { | |
1238 | init_request_from_bio(rq, bio); | |
1239 | ||
1240 | blk_account_io_start(rq, 1); | |
1241 | } | |
1242 | ||
1243 | static inline bool hctx_allow_merges(struct blk_mq_hw_ctx *hctx) | |
1244 | { | |
1245 | return (hctx->flags & BLK_MQ_F_SHOULD_MERGE) && | |
1246 | !blk_queue_nomerges(hctx->queue); | |
1247 | } | |
1248 | ||
1249 | static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx *hctx, | |
1250 | struct blk_mq_ctx *ctx, | |
1251 | struct request *rq, struct bio *bio) | |
1252 | { | |
1253 | if (!hctx_allow_merges(hctx) || !bio_mergeable(bio)) { | |
1254 | blk_mq_bio_to_request(rq, bio); | |
1255 | spin_lock(&ctx->lock); | |
1256 | insert_rq: | |
1257 | __blk_mq_insert_request(hctx, rq, false); | |
1258 | spin_unlock(&ctx->lock); | |
1259 | return false; | |
1260 | } else { | |
1261 | struct request_queue *q = hctx->queue; | |
1262 | ||
1263 | spin_lock(&ctx->lock); | |
1264 | if (!blk_mq_attempt_merge(q, ctx, bio)) { | |
1265 | blk_mq_bio_to_request(rq, bio); | |
1266 | goto insert_rq; | |
1267 | } | |
1268 | ||
1269 | spin_unlock(&ctx->lock); | |
1270 | __blk_mq_free_request(hctx, ctx, rq); | |
1271 | return true; | |
1272 | } | |
1273 | } | |
1274 | ||
1275 | static struct request *blk_mq_map_request(struct request_queue *q, | |
1276 | struct bio *bio, | |
1277 | struct blk_mq_alloc_data *data) | |
1278 | { | |
1279 | struct blk_mq_hw_ctx *hctx; | |
1280 | struct blk_mq_ctx *ctx; | |
1281 | struct request *rq; | |
1282 | ||
1283 | blk_queue_enter_live(q); | |
1284 | ctx = blk_mq_get_ctx(q); | |
1285 | hctx = blk_mq_map_queue(q, ctx->cpu); | |
1286 | ||
1287 | trace_block_getrq(q, bio, bio->bi_opf); | |
1288 | blk_mq_set_alloc_data(data, q, 0, ctx, hctx); | |
1289 | rq = __blk_mq_alloc_request(data, bio->bi_opf); | |
1290 | ||
1291 | data->hctx->queued++; | |
1292 | return rq; | |
1293 | } | |
1294 | ||
1295 | static void blk_mq_try_issue_directly(struct request *rq, blk_qc_t *cookie) | |
1296 | { | |
1297 | int ret; | |
1298 | struct request_queue *q = rq->q; | |
1299 | struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, rq->mq_ctx->cpu); | |
1300 | struct blk_mq_queue_data bd = { | |
1301 | .rq = rq, | |
1302 | .list = NULL, | |
1303 | .last = 1 | |
1304 | }; | |
1305 | blk_qc_t new_cookie = blk_tag_to_qc_t(rq->tag, hctx->queue_num); | |
1306 | ||
1307 | if (blk_mq_hctx_stopped(hctx)) | |
1308 | goto insert; | |
1309 | ||
1310 | /* | |
1311 | * For OK queue, we are done. For error, kill it. Any other | |
1312 | * error (busy), just add it to our list as we previously | |
1313 | * would have done | |
1314 | */ | |
1315 | ret = q->mq_ops->queue_rq(hctx, &bd); | |
1316 | if (ret == BLK_MQ_RQ_QUEUE_OK) { | |
1317 | *cookie = new_cookie; | |
1318 | return; | |
1319 | } | |
1320 | ||
1321 | __blk_mq_requeue_request(rq); | |
1322 | ||
1323 | if (ret == BLK_MQ_RQ_QUEUE_ERROR) { | |
1324 | *cookie = BLK_QC_T_NONE; | |
1325 | rq->errors = -EIO; | |
1326 | blk_mq_end_request(rq, rq->errors); | |
1327 | return; | |
1328 | } | |
1329 | ||
1330 | insert: | |
1331 | blk_mq_insert_request(rq, false, true, true); | |
1332 | } | |
1333 | ||
1334 | /* | |
1335 | * Multiple hardware queue variant. This will not use per-process plugs, | |
1336 | * but will attempt to bypass the hctx queueing if we can go straight to | |
1337 | * hardware for SYNC IO. | |
1338 | */ | |
1339 | static blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio) | |
1340 | { | |
1341 | const int is_sync = op_is_sync(bio->bi_opf); | |
1342 | const int is_flush_fua = bio->bi_opf & (REQ_PREFLUSH | REQ_FUA); | |
1343 | struct blk_mq_alloc_data data; | |
1344 | struct request *rq; | |
1345 | unsigned int request_count = 0, srcu_idx; | |
1346 | struct blk_plug *plug; | |
1347 | struct request *same_queue_rq = NULL; | |
1348 | blk_qc_t cookie; | |
1349 | unsigned int wb_acct; | |
1350 | ||
1351 | blk_queue_bounce(q, &bio); | |
1352 | ||
1353 | if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) { | |
1354 | bio_io_error(bio); | |
1355 | return BLK_QC_T_NONE; | |
1356 | } | |
1357 | ||
1358 | blk_queue_split(q, &bio, q->bio_split); | |
1359 | ||
1360 | if (!is_flush_fua && !blk_queue_nomerges(q) && | |
1361 | blk_attempt_plug_merge(q, bio, &request_count, &same_queue_rq)) | |
1362 | return BLK_QC_T_NONE; | |
1363 | ||
1364 | wb_acct = wbt_wait(q->rq_wb, bio, NULL); | |
1365 | ||
1366 | rq = blk_mq_map_request(q, bio, &data); | |
1367 | if (unlikely(!rq)) { | |
1368 | __wbt_done(q->rq_wb, wb_acct); | |
1369 | return BLK_QC_T_NONE; | |
1370 | } | |
1371 | ||
1372 | wbt_track(&rq->issue_stat, wb_acct); | |
1373 | ||
1374 | cookie = blk_tag_to_qc_t(rq->tag, data.hctx->queue_num); | |
1375 | ||
1376 | if (unlikely(is_flush_fua)) { | |
1377 | blk_mq_bio_to_request(rq, bio); | |
1378 | blk_insert_flush(rq); | |
1379 | goto run_queue; | |
1380 | } | |
1381 | ||
1382 | plug = current->plug; | |
1383 | /* | |
1384 | * If the driver supports defer issued based on 'last', then | |
1385 | * queue it up like normal since we can potentially save some | |
1386 | * CPU this way. | |
1387 | */ | |
1388 | if (((plug && !blk_queue_nomerges(q)) || is_sync) && | |
1389 | !(data.hctx->flags & BLK_MQ_F_DEFER_ISSUE)) { | |
1390 | struct request *old_rq = NULL; | |
1391 | ||
1392 | blk_mq_bio_to_request(rq, bio); | |
1393 | ||
1394 | /* | |
1395 | * We do limited plugging. If the bio can be merged, do that. | |
1396 | * Otherwise the existing request in the plug list will be | |
1397 | * issued. So the plug list will have one request at most | |
1398 | */ | |
1399 | if (plug) { | |
1400 | /* | |
1401 | * The plug list might get flushed before this. If that | |
1402 | * happens, same_queue_rq is invalid and plug list is | |
1403 | * empty | |
1404 | */ | |
1405 | if (same_queue_rq && !list_empty(&plug->mq_list)) { | |
1406 | old_rq = same_queue_rq; | |
1407 | list_del_init(&old_rq->queuelist); | |
1408 | } | |
1409 | list_add_tail(&rq->queuelist, &plug->mq_list); | |
1410 | } else /* is_sync */ | |
1411 | old_rq = rq; | |
1412 | blk_mq_put_ctx(data.ctx); | |
1413 | if (!old_rq) | |
1414 | goto done; | |
1415 | ||
1416 | if (!(data.hctx->flags & BLK_MQ_F_BLOCKING)) { | |
1417 | rcu_read_lock(); | |
1418 | blk_mq_try_issue_directly(old_rq, &cookie); | |
1419 | rcu_read_unlock(); | |
1420 | } else { | |
1421 | srcu_idx = srcu_read_lock(&data.hctx->queue_rq_srcu); | |
1422 | blk_mq_try_issue_directly(old_rq, &cookie); | |
1423 | srcu_read_unlock(&data.hctx->queue_rq_srcu, srcu_idx); | |
1424 | } | |
1425 | goto done; | |
1426 | } | |
1427 | ||
1428 | if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) { | |
1429 | /* | |
1430 | * For a SYNC request, send it to the hardware immediately. For | |
1431 | * an ASYNC request, just ensure that we run it later on. The | |
1432 | * latter allows for merging opportunities and more efficient | |
1433 | * dispatching. | |
1434 | */ | |
1435 | run_queue: | |
1436 | blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua); | |
1437 | } | |
1438 | blk_mq_put_ctx(data.ctx); | |
1439 | done: | |
1440 | return cookie; | |
1441 | } | |
1442 | ||
1443 | /* | |
1444 | * Single hardware queue variant. This will attempt to use any per-process | |
1445 | * plug for merging and IO deferral. | |
1446 | */ | |
1447 | static blk_qc_t blk_sq_make_request(struct request_queue *q, struct bio *bio) | |
1448 | { | |
1449 | const int is_sync = op_is_sync(bio->bi_opf); | |
1450 | const int is_flush_fua = bio->bi_opf & (REQ_PREFLUSH | REQ_FUA); | |
1451 | struct blk_plug *plug; | |
1452 | unsigned int request_count = 0; | |
1453 | struct blk_mq_alloc_data data; | |
1454 | struct request *rq; | |
1455 | blk_qc_t cookie; | |
1456 | unsigned int wb_acct; | |
1457 | ||
1458 | blk_queue_bounce(q, &bio); | |
1459 | ||
1460 | if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) { | |
1461 | bio_io_error(bio); | |
1462 | return BLK_QC_T_NONE; | |
1463 | } | |
1464 | ||
1465 | blk_queue_split(q, &bio, q->bio_split); | |
1466 | ||
1467 | if (!is_flush_fua && !blk_queue_nomerges(q)) { | |
1468 | if (blk_attempt_plug_merge(q, bio, &request_count, NULL)) | |
1469 | return BLK_QC_T_NONE; | |
1470 | } else | |
1471 | request_count = blk_plug_queued_count(q); | |
1472 | ||
1473 | wb_acct = wbt_wait(q->rq_wb, bio, NULL); | |
1474 | ||
1475 | rq = blk_mq_map_request(q, bio, &data); | |
1476 | if (unlikely(!rq)) { | |
1477 | __wbt_done(q->rq_wb, wb_acct); | |
1478 | return BLK_QC_T_NONE; | |
1479 | } | |
1480 | ||
1481 | wbt_track(&rq->issue_stat, wb_acct); | |
1482 | ||
1483 | cookie = blk_tag_to_qc_t(rq->tag, data.hctx->queue_num); | |
1484 | ||
1485 | if (unlikely(is_flush_fua)) { | |
1486 | blk_mq_bio_to_request(rq, bio); | |
1487 | blk_insert_flush(rq); | |
1488 | goto run_queue; | |
1489 | } | |
1490 | ||
1491 | /* | |
1492 | * A task plug currently exists. Since this is completely lockless, | |
1493 | * utilize that to temporarily store requests until the task is | |
1494 | * either done or scheduled away. | |
1495 | */ | |
1496 | plug = current->plug; | |
1497 | if (plug) { | |
1498 | struct request *last = NULL; | |
1499 | ||
1500 | blk_mq_bio_to_request(rq, bio); | |
1501 | ||
1502 | /* | |
1503 | * @request_count may become stale because of schedule | |
1504 | * out, so check the list again. | |
1505 | */ | |
1506 | if (list_empty(&plug->mq_list)) | |
1507 | request_count = 0; | |
1508 | if (!request_count) | |
1509 | trace_block_plug(q); | |
1510 | else | |
1511 | last = list_entry_rq(plug->mq_list.prev); | |
1512 | ||
1513 | blk_mq_put_ctx(data.ctx); | |
1514 | ||
1515 | if (request_count >= BLK_MAX_REQUEST_COUNT || (last && | |
1516 | blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) { | |
1517 | blk_flush_plug_list(plug, false); | |
1518 | trace_block_plug(q); | |
1519 | } | |
1520 | ||
1521 | list_add_tail(&rq->queuelist, &plug->mq_list); | |
1522 | return cookie; | |
1523 | } | |
1524 | ||
1525 | if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) { | |
1526 | /* | |
1527 | * For a SYNC request, send it to the hardware immediately. For | |
1528 | * an ASYNC request, just ensure that we run it later on. The | |
1529 | * latter allows for merging opportunities and more efficient | |
1530 | * dispatching. | |
1531 | */ | |
1532 | run_queue: | |
1533 | blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua); | |
1534 | } | |
1535 | ||
1536 | blk_mq_put_ctx(data.ctx); | |
1537 | return cookie; | |
1538 | } | |
1539 | ||
1540 | static void blk_mq_free_rq_map(struct blk_mq_tag_set *set, | |
1541 | struct blk_mq_tags *tags, unsigned int hctx_idx) | |
1542 | { | |
1543 | struct page *page; | |
1544 | ||
1545 | if (tags->rqs && set->ops->exit_request) { | |
1546 | int i; | |
1547 | ||
1548 | for (i = 0; i < tags->nr_tags; i++) { | |
1549 | if (!tags->rqs[i]) | |
1550 | continue; | |
1551 | set->ops->exit_request(set->driver_data, tags->rqs[i], | |
1552 | hctx_idx, i); | |
1553 | tags->rqs[i] = NULL; | |
1554 | } | |
1555 | } | |
1556 | ||
1557 | while (!list_empty(&tags->page_list)) { | |
1558 | page = list_first_entry(&tags->page_list, struct page, lru); | |
1559 | list_del_init(&page->lru); | |
1560 | /* | |
1561 | * Remove kmemleak object previously allocated in | |
1562 | * blk_mq_init_rq_map(). | |
1563 | */ | |
1564 | kmemleak_free(page_address(page)); | |
1565 | __free_pages(page, page->private); | |
1566 | } | |
1567 | ||
1568 | kfree(tags->rqs); | |
1569 | ||
1570 | blk_mq_free_tags(tags); | |
1571 | } | |
1572 | ||
1573 | static size_t order_to_size(unsigned int order) | |
1574 | { | |
1575 | return (size_t)PAGE_SIZE << order; | |
1576 | } | |
1577 | ||
1578 | static struct blk_mq_tags *blk_mq_init_rq_map(struct blk_mq_tag_set *set, | |
1579 | unsigned int hctx_idx) | |
1580 | { | |
1581 | struct blk_mq_tags *tags; | |
1582 | unsigned int i, j, entries_per_page, max_order = 4; | |
1583 | size_t rq_size, left; | |
1584 | ||
1585 | tags = blk_mq_init_tags(set->queue_depth, set->reserved_tags, | |
1586 | set->numa_node, | |
1587 | BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags)); | |
1588 | if (!tags) | |
1589 | return NULL; | |
1590 | ||
1591 | INIT_LIST_HEAD(&tags->page_list); | |
1592 | ||
1593 | tags->rqs = kzalloc_node(set->queue_depth * sizeof(struct request *), | |
1594 | GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY, | |
1595 | set->numa_node); | |
1596 | if (!tags->rqs) { | |
1597 | blk_mq_free_tags(tags); | |
1598 | return NULL; | |
1599 | } | |
1600 | ||
1601 | /* | |
1602 | * rq_size is the size of the request plus driver payload, rounded | |
1603 | * to the cacheline size | |
1604 | */ | |
1605 | rq_size = round_up(sizeof(struct request) + set->cmd_size, | |
1606 | cache_line_size()); | |
1607 | left = rq_size * set->queue_depth; | |
1608 | ||
1609 | for (i = 0; i < set->queue_depth; ) { | |
1610 | int this_order = max_order; | |
1611 | struct page *page; | |
1612 | int to_do; | |
1613 | void *p; | |
1614 | ||
1615 | while (this_order && left < order_to_size(this_order - 1)) | |
1616 | this_order--; | |
1617 | ||
1618 | do { | |
1619 | page = alloc_pages_node(set->numa_node, | |
1620 | GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO, | |
1621 | this_order); | |
1622 | if (page) | |
1623 | break; | |
1624 | if (!this_order--) | |
1625 | break; | |
1626 | if (order_to_size(this_order) < rq_size) | |
1627 | break; | |
1628 | } while (1); | |
1629 | ||
1630 | if (!page) | |
1631 | goto fail; | |
1632 | ||
1633 | page->private = this_order; | |
1634 | list_add_tail(&page->lru, &tags->page_list); | |
1635 | ||
1636 | p = page_address(page); | |
1637 | /* | |
1638 | * Allow kmemleak to scan these pages as they contain pointers | |
1639 | * to additional allocations like via ops->init_request(). | |
1640 | */ | |
1641 | kmemleak_alloc(p, order_to_size(this_order), 1, GFP_KERNEL); | |
1642 | entries_per_page = order_to_size(this_order) / rq_size; | |
1643 | to_do = min(entries_per_page, set->queue_depth - i); | |
1644 | left -= to_do * rq_size; | |
1645 | for (j = 0; j < to_do; j++) { | |
1646 | tags->rqs[i] = p; | |
1647 | if (set->ops->init_request) { | |
1648 | if (set->ops->init_request(set->driver_data, | |
1649 | tags->rqs[i], hctx_idx, i, | |
1650 | set->numa_node)) { | |
1651 | tags->rqs[i] = NULL; | |
1652 | goto fail; | |
1653 | } | |
1654 | } | |
1655 | ||
1656 | p += rq_size; | |
1657 | i++; | |
1658 | } | |
1659 | } | |
1660 | return tags; | |
1661 | ||
1662 | fail: | |
1663 | blk_mq_free_rq_map(set, tags, hctx_idx); | |
1664 | return NULL; | |
1665 | } | |
1666 | ||
1667 | /* | |
1668 | * 'cpu' is going away. splice any existing rq_list entries from this | |
1669 | * software queue to the hw queue dispatch list, and ensure that it | |
1670 | * gets run. | |
1671 | */ | |
1672 | static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node) | |
1673 | { | |
1674 | struct blk_mq_hw_ctx *hctx; | |
1675 | struct blk_mq_ctx *ctx; | |
1676 | LIST_HEAD(tmp); | |
1677 | ||
1678 | hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead); | |
1679 | ctx = __blk_mq_get_ctx(hctx->queue, cpu); | |
1680 | ||
1681 | spin_lock(&ctx->lock); | |
1682 | if (!list_empty(&ctx->rq_list)) { | |
1683 | list_splice_init(&ctx->rq_list, &tmp); | |
1684 | blk_mq_hctx_clear_pending(hctx, ctx); | |
1685 | } | |
1686 | spin_unlock(&ctx->lock); | |
1687 | ||
1688 | if (list_empty(&tmp)) | |
1689 | return 0; | |
1690 | ||
1691 | spin_lock(&hctx->lock); | |
1692 | list_splice_tail_init(&tmp, &hctx->dispatch); | |
1693 | spin_unlock(&hctx->lock); | |
1694 | ||
1695 | blk_mq_run_hw_queue(hctx, true); | |
1696 | return 0; | |
1697 | } | |
1698 | ||
1699 | static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx) | |
1700 | { | |
1701 | cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD, | |
1702 | &hctx->cpuhp_dead); | |
1703 | } | |
1704 | ||
1705 | /* hctx->ctxs will be freed in queue's release handler */ | |
1706 | static void blk_mq_exit_hctx(struct request_queue *q, | |
1707 | struct blk_mq_tag_set *set, | |
1708 | struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx) | |
1709 | { | |
1710 | unsigned flush_start_tag = set->queue_depth; | |
1711 | ||
1712 | blk_mq_tag_idle(hctx); | |
1713 | ||
1714 | if (set->ops->exit_request) | |
1715 | set->ops->exit_request(set->driver_data, | |
1716 | hctx->fq->flush_rq, hctx_idx, | |
1717 | flush_start_tag + hctx_idx); | |
1718 | ||
1719 | if (set->ops->exit_hctx) | |
1720 | set->ops->exit_hctx(hctx, hctx_idx); | |
1721 | ||
1722 | if (hctx->flags & BLK_MQ_F_BLOCKING) | |
1723 | cleanup_srcu_struct(&hctx->queue_rq_srcu); | |
1724 | ||
1725 | blk_mq_remove_cpuhp(hctx); | |
1726 | blk_free_flush_queue(hctx->fq); | |
1727 | sbitmap_free(&hctx->ctx_map); | |
1728 | } | |
1729 | ||
1730 | static void blk_mq_exit_hw_queues(struct request_queue *q, | |
1731 | struct blk_mq_tag_set *set, int nr_queue) | |
1732 | { | |
1733 | struct blk_mq_hw_ctx *hctx; | |
1734 | unsigned int i; | |
1735 | ||
1736 | queue_for_each_hw_ctx(q, hctx, i) { | |
1737 | if (i == nr_queue) | |
1738 | break; | |
1739 | blk_mq_exit_hctx(q, set, hctx, i); | |
1740 | } | |
1741 | } | |
1742 | ||
1743 | static void blk_mq_free_hw_queues(struct request_queue *q, | |
1744 | struct blk_mq_tag_set *set) | |
1745 | { | |
1746 | struct blk_mq_hw_ctx *hctx; | |
1747 | unsigned int i; | |
1748 | ||
1749 | queue_for_each_hw_ctx(q, hctx, i) | |
1750 | free_cpumask_var(hctx->cpumask); | |
1751 | } | |
1752 | ||
1753 | static int blk_mq_init_hctx(struct request_queue *q, | |
1754 | struct blk_mq_tag_set *set, | |
1755 | struct blk_mq_hw_ctx *hctx, unsigned hctx_idx) | |
1756 | { | |
1757 | int node; | |
1758 | unsigned flush_start_tag = set->queue_depth; | |
1759 | ||
1760 | node = hctx->numa_node; | |
1761 | if (node == NUMA_NO_NODE) | |
1762 | node = hctx->numa_node = set->numa_node; | |
1763 | ||
1764 | INIT_WORK(&hctx->run_work, blk_mq_run_work_fn); | |
1765 | INIT_DELAYED_WORK(&hctx->delay_work, blk_mq_delay_work_fn); | |
1766 | spin_lock_init(&hctx->lock); | |
1767 | INIT_LIST_HEAD(&hctx->dispatch); | |
1768 | hctx->queue = q; | |
1769 | hctx->queue_num = hctx_idx; | |
1770 | hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED; | |
1771 | ||
1772 | cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead); | |
1773 | ||
1774 | hctx->tags = set->tags[hctx_idx]; | |
1775 | ||
1776 | /* | |
1777 | * Allocate space for all possible cpus to avoid allocation at | |
1778 | * runtime | |
1779 | */ | |
1780 | hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *), | |
1781 | GFP_KERNEL, node); | |
1782 | if (!hctx->ctxs) | |
1783 | goto unregister_cpu_notifier; | |
1784 | ||
1785 | if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8), GFP_KERNEL, | |
1786 | node)) | |
1787 | goto free_ctxs; | |
1788 | ||
1789 | hctx->nr_ctx = 0; | |
1790 | ||
1791 | if (set->ops->init_hctx && | |
1792 | set->ops->init_hctx(hctx, set->driver_data, hctx_idx)) | |
1793 | goto free_bitmap; | |
1794 | ||
1795 | hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size); | |
1796 | if (!hctx->fq) | |
1797 | goto exit_hctx; | |
1798 | ||
1799 | if (set->ops->init_request && | |
1800 | set->ops->init_request(set->driver_data, | |
1801 | hctx->fq->flush_rq, hctx_idx, | |
1802 | flush_start_tag + hctx_idx, node)) | |
1803 | goto free_fq; | |
1804 | ||
1805 | if (hctx->flags & BLK_MQ_F_BLOCKING) | |
1806 | init_srcu_struct(&hctx->queue_rq_srcu); | |
1807 | ||
1808 | return 0; | |
1809 | ||
1810 | free_fq: | |
1811 | kfree(hctx->fq); | |
1812 | exit_hctx: | |
1813 | if (set->ops->exit_hctx) | |
1814 | set->ops->exit_hctx(hctx, hctx_idx); | |
1815 | free_bitmap: | |
1816 | sbitmap_free(&hctx->ctx_map); | |
1817 | free_ctxs: | |
1818 | kfree(hctx->ctxs); | |
1819 | unregister_cpu_notifier: | |
1820 | blk_mq_remove_cpuhp(hctx); | |
1821 | return -1; | |
1822 | } | |
1823 | ||
1824 | static void blk_mq_init_cpu_queues(struct request_queue *q, | |
1825 | unsigned int nr_hw_queues) | |
1826 | { | |
1827 | unsigned int i; | |
1828 | ||
1829 | for_each_possible_cpu(i) { | |
1830 | struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i); | |
1831 | struct blk_mq_hw_ctx *hctx; | |
1832 | ||
1833 | memset(__ctx, 0, sizeof(*__ctx)); | |
1834 | __ctx->cpu = i; | |
1835 | spin_lock_init(&__ctx->lock); | |
1836 | INIT_LIST_HEAD(&__ctx->rq_list); | |
1837 | __ctx->queue = q; | |
1838 | blk_stat_init(&__ctx->stat[BLK_STAT_READ]); | |
1839 | blk_stat_init(&__ctx->stat[BLK_STAT_WRITE]); | |
1840 | ||
1841 | /* If the cpu isn't online, the cpu is mapped to first hctx */ | |
1842 | if (!cpu_online(i)) | |
1843 | continue; | |
1844 | ||
1845 | hctx = blk_mq_map_queue(q, i); | |
1846 | ||
1847 | /* | |
1848 | * Set local node, IFF we have more than one hw queue. If | |
1849 | * not, we remain on the home node of the device | |
1850 | */ | |
1851 | if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE) | |
1852 | hctx->numa_node = local_memory_node(cpu_to_node(i)); | |
1853 | } | |
1854 | } | |
1855 | ||
1856 | static void blk_mq_map_swqueue(struct request_queue *q, | |
1857 | const struct cpumask *online_mask) | |
1858 | { | |
1859 | unsigned int i; | |
1860 | struct blk_mq_hw_ctx *hctx; | |
1861 | struct blk_mq_ctx *ctx; | |
1862 | struct blk_mq_tag_set *set = q->tag_set; | |
1863 | ||
1864 | /* | |
1865 | * Avoid others reading imcomplete hctx->cpumask through sysfs | |
1866 | */ | |
1867 | mutex_lock(&q->sysfs_lock); | |
1868 | ||
1869 | queue_for_each_hw_ctx(q, hctx, i) { | |
1870 | cpumask_clear(hctx->cpumask); | |
1871 | hctx->nr_ctx = 0; | |
1872 | } | |
1873 | ||
1874 | /* | |
1875 | * Map software to hardware queues | |
1876 | */ | |
1877 | for_each_possible_cpu(i) { | |
1878 | /* If the cpu isn't online, the cpu is mapped to first hctx */ | |
1879 | if (!cpumask_test_cpu(i, online_mask)) | |
1880 | continue; | |
1881 | ||
1882 | ctx = per_cpu_ptr(q->queue_ctx, i); | |
1883 | hctx = blk_mq_map_queue(q, i); | |
1884 | ||
1885 | cpumask_set_cpu(i, hctx->cpumask); | |
1886 | ctx->index_hw = hctx->nr_ctx; | |
1887 | hctx->ctxs[hctx->nr_ctx++] = ctx; | |
1888 | } | |
1889 | ||
1890 | mutex_unlock(&q->sysfs_lock); | |
1891 | ||
1892 | queue_for_each_hw_ctx(q, hctx, i) { | |
1893 | /* | |
1894 | * If no software queues are mapped to this hardware queue, | |
1895 | * disable it and free the request entries. | |
1896 | */ | |
1897 | if (!hctx->nr_ctx) { | |
1898 | if (set->tags[i]) { | |
1899 | blk_mq_free_rq_map(set, set->tags[i], i); | |
1900 | set->tags[i] = NULL; | |
1901 | } | |
1902 | hctx->tags = NULL; | |
1903 | continue; | |
1904 | } | |
1905 | ||
1906 | /* unmapped hw queue can be remapped after CPU topo changed */ | |
1907 | if (!set->tags[i]) | |
1908 | set->tags[i] = blk_mq_init_rq_map(set, i); | |
1909 | hctx->tags = set->tags[i]; | |
1910 | WARN_ON(!hctx->tags); | |
1911 | ||
1912 | /* | |
1913 | * Set the map size to the number of mapped software queues. | |
1914 | * This is more accurate and more efficient than looping | |
1915 | * over all possibly mapped software queues. | |
1916 | */ | |
1917 | sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx); | |
1918 | ||
1919 | /* | |
1920 | * Initialize batch roundrobin counts | |
1921 | */ | |
1922 | hctx->next_cpu = cpumask_first(hctx->cpumask); | |
1923 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; | |
1924 | } | |
1925 | } | |
1926 | ||
1927 | static void queue_set_hctx_shared(struct request_queue *q, bool shared) | |
1928 | { | |
1929 | struct blk_mq_hw_ctx *hctx; | |
1930 | int i; | |
1931 | ||
1932 | queue_for_each_hw_ctx(q, hctx, i) { | |
1933 | if (shared) | |
1934 | hctx->flags |= BLK_MQ_F_TAG_SHARED; | |
1935 | else | |
1936 | hctx->flags &= ~BLK_MQ_F_TAG_SHARED; | |
1937 | } | |
1938 | } | |
1939 | ||
1940 | static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set, bool shared) | |
1941 | { | |
1942 | struct request_queue *q; | |
1943 | ||
1944 | list_for_each_entry(q, &set->tag_list, tag_set_list) { | |
1945 | blk_mq_freeze_queue(q); | |
1946 | queue_set_hctx_shared(q, shared); | |
1947 | blk_mq_unfreeze_queue(q); | |
1948 | } | |
1949 | } | |
1950 | ||
1951 | static void blk_mq_del_queue_tag_set(struct request_queue *q) | |
1952 | { | |
1953 | struct blk_mq_tag_set *set = q->tag_set; | |
1954 | ||
1955 | mutex_lock(&set->tag_list_lock); | |
1956 | list_del_init(&q->tag_set_list); | |
1957 | if (list_is_singular(&set->tag_list)) { | |
1958 | /* just transitioned to unshared */ | |
1959 | set->flags &= ~BLK_MQ_F_TAG_SHARED; | |
1960 | /* update existing queue */ | |
1961 | blk_mq_update_tag_set_depth(set, false); | |
1962 | } | |
1963 | mutex_unlock(&set->tag_list_lock); | |
1964 | } | |
1965 | ||
1966 | static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set, | |
1967 | struct request_queue *q) | |
1968 | { | |
1969 | q->tag_set = set; | |
1970 | ||
1971 | mutex_lock(&set->tag_list_lock); | |
1972 | ||
1973 | /* Check to see if we're transitioning to shared (from 1 to 2 queues). */ | |
1974 | if (!list_empty(&set->tag_list) && !(set->flags & BLK_MQ_F_TAG_SHARED)) { | |
1975 | set->flags |= BLK_MQ_F_TAG_SHARED; | |
1976 | /* update existing queue */ | |
1977 | blk_mq_update_tag_set_depth(set, true); | |
1978 | } | |
1979 | if (set->flags & BLK_MQ_F_TAG_SHARED) | |
1980 | queue_set_hctx_shared(q, true); | |
1981 | list_add_tail(&q->tag_set_list, &set->tag_list); | |
1982 | ||
1983 | mutex_unlock(&set->tag_list_lock); | |
1984 | } | |
1985 | ||
1986 | /* | |
1987 | * It is the actual release handler for mq, but we do it from | |
1988 | * request queue's release handler for avoiding use-after-free | |
1989 | * and headache because q->mq_kobj shouldn't have been introduced, | |
1990 | * but we can't group ctx/kctx kobj without it. | |
1991 | */ | |
1992 | void blk_mq_release(struct request_queue *q) | |
1993 | { | |
1994 | struct blk_mq_hw_ctx *hctx; | |
1995 | unsigned int i; | |
1996 | ||
1997 | /* hctx kobj stays in hctx */ | |
1998 | queue_for_each_hw_ctx(q, hctx, i) { | |
1999 | if (!hctx) | |
2000 | continue; | |
2001 | kfree(hctx->ctxs); | |
2002 | kfree(hctx); | |
2003 | } | |
2004 | ||
2005 | q->mq_map = NULL; | |
2006 | ||
2007 | kfree(q->queue_hw_ctx); | |
2008 | ||
2009 | /* ctx kobj stays in queue_ctx */ | |
2010 | free_percpu(q->queue_ctx); | |
2011 | } | |
2012 | ||
2013 | struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set) | |
2014 | { | |
2015 | struct request_queue *uninit_q, *q; | |
2016 | ||
2017 | uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node); | |
2018 | if (!uninit_q) | |
2019 | return ERR_PTR(-ENOMEM); | |
2020 | ||
2021 | q = blk_mq_init_allocated_queue(set, uninit_q); | |
2022 | if (IS_ERR(q)) | |
2023 | blk_cleanup_queue(uninit_q); | |
2024 | ||
2025 | return q; | |
2026 | } | |
2027 | EXPORT_SYMBOL(blk_mq_init_queue); | |
2028 | ||
2029 | static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set, | |
2030 | struct request_queue *q) | |
2031 | { | |
2032 | int i, j; | |
2033 | struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx; | |
2034 | ||
2035 | blk_mq_sysfs_unregister(q); | |
2036 | for (i = 0; i < set->nr_hw_queues; i++) { | |
2037 | int node; | |
2038 | ||
2039 | if (hctxs[i]) | |
2040 | continue; | |
2041 | ||
2042 | node = blk_mq_hw_queue_to_node(q->mq_map, i); | |
2043 | hctxs[i] = kzalloc_node(sizeof(struct blk_mq_hw_ctx), | |
2044 | GFP_KERNEL, node); | |
2045 | if (!hctxs[i]) | |
2046 | break; | |
2047 | ||
2048 | if (!zalloc_cpumask_var_node(&hctxs[i]->cpumask, GFP_KERNEL, | |
2049 | node)) { | |
2050 | kfree(hctxs[i]); | |
2051 | hctxs[i] = NULL; | |
2052 | break; | |
2053 | } | |
2054 | ||
2055 | atomic_set(&hctxs[i]->nr_active, 0); | |
2056 | hctxs[i]->numa_node = node; | |
2057 | hctxs[i]->queue_num = i; | |
2058 | ||
2059 | if (blk_mq_init_hctx(q, set, hctxs[i], i)) { | |
2060 | free_cpumask_var(hctxs[i]->cpumask); | |
2061 | kfree(hctxs[i]); | |
2062 | hctxs[i] = NULL; | |
2063 | break; | |
2064 | } | |
2065 | blk_mq_hctx_kobj_init(hctxs[i]); | |
2066 | } | |
2067 | for (j = i; j < q->nr_hw_queues; j++) { | |
2068 | struct blk_mq_hw_ctx *hctx = hctxs[j]; | |
2069 | ||
2070 | if (hctx) { | |
2071 | if (hctx->tags) { | |
2072 | blk_mq_free_rq_map(set, hctx->tags, j); | |
2073 | set->tags[j] = NULL; | |
2074 | } | |
2075 | blk_mq_exit_hctx(q, set, hctx, j); | |
2076 | free_cpumask_var(hctx->cpumask); | |
2077 | kobject_put(&hctx->kobj); | |
2078 | kfree(hctx->ctxs); | |
2079 | kfree(hctx); | |
2080 | hctxs[j] = NULL; | |
2081 | ||
2082 | } | |
2083 | } | |
2084 | q->nr_hw_queues = i; | |
2085 | blk_mq_sysfs_register(q); | |
2086 | } | |
2087 | ||
2088 | struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set, | |
2089 | struct request_queue *q) | |
2090 | { | |
2091 | /* mark the queue as mq asap */ | |
2092 | q->mq_ops = set->ops; | |
2093 | ||
2094 | q->queue_ctx = alloc_percpu(struct blk_mq_ctx); | |
2095 | if (!q->queue_ctx) | |
2096 | goto err_exit; | |
2097 | ||
2098 | q->queue_hw_ctx = kzalloc_node(nr_cpu_ids * sizeof(*(q->queue_hw_ctx)), | |
2099 | GFP_KERNEL, set->numa_node); | |
2100 | if (!q->queue_hw_ctx) | |
2101 | goto err_percpu; | |
2102 | ||
2103 | q->mq_map = set->mq_map; | |
2104 | ||
2105 | blk_mq_realloc_hw_ctxs(set, q); | |
2106 | if (!q->nr_hw_queues) | |
2107 | goto err_hctxs; | |
2108 | ||
2109 | INIT_WORK(&q->timeout_work, blk_mq_timeout_work); | |
2110 | blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ); | |
2111 | ||
2112 | q->nr_queues = nr_cpu_ids; | |
2113 | ||
2114 | q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT; | |
2115 | ||
2116 | if (!(set->flags & BLK_MQ_F_SG_MERGE)) | |
2117 | q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE; | |
2118 | ||
2119 | q->sg_reserved_size = INT_MAX; | |
2120 | ||
2121 | INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work); | |
2122 | INIT_LIST_HEAD(&q->requeue_list); | |
2123 | spin_lock_init(&q->requeue_lock); | |
2124 | ||
2125 | if (q->nr_hw_queues > 1) | |
2126 | blk_queue_make_request(q, blk_mq_make_request); | |
2127 | else | |
2128 | blk_queue_make_request(q, blk_sq_make_request); | |
2129 | ||
2130 | /* | |
2131 | * Do this after blk_queue_make_request() overrides it... | |
2132 | */ | |
2133 | q->nr_requests = set->queue_depth; | |
2134 | ||
2135 | /* | |
2136 | * Default to classic polling | |
2137 | */ | |
2138 | q->poll_nsec = -1; | |
2139 | ||
2140 | if (set->ops->complete) | |
2141 | blk_queue_softirq_done(q, set->ops->complete); | |
2142 | ||
2143 | blk_mq_init_cpu_queues(q, set->nr_hw_queues); | |
2144 | ||
2145 | get_online_cpus(); | |
2146 | mutex_lock(&all_q_mutex); | |
2147 | ||
2148 | list_add_tail(&q->all_q_node, &all_q_list); | |
2149 | blk_mq_add_queue_tag_set(set, q); | |
2150 | blk_mq_map_swqueue(q, cpu_online_mask); | |
2151 | ||
2152 | mutex_unlock(&all_q_mutex); | |
2153 | put_online_cpus(); | |
2154 | ||
2155 | return q; | |
2156 | ||
2157 | err_hctxs: | |
2158 | kfree(q->queue_hw_ctx); | |
2159 | err_percpu: | |
2160 | free_percpu(q->queue_ctx); | |
2161 | err_exit: | |
2162 | q->mq_ops = NULL; | |
2163 | return ERR_PTR(-ENOMEM); | |
2164 | } | |
2165 | EXPORT_SYMBOL(blk_mq_init_allocated_queue); | |
2166 | ||
2167 | void blk_mq_free_queue(struct request_queue *q) | |
2168 | { | |
2169 | struct blk_mq_tag_set *set = q->tag_set; | |
2170 | ||
2171 | mutex_lock(&all_q_mutex); | |
2172 | list_del_init(&q->all_q_node); | |
2173 | mutex_unlock(&all_q_mutex); | |
2174 | ||
2175 | wbt_exit(q); | |
2176 | ||
2177 | blk_mq_del_queue_tag_set(q); | |
2178 | ||
2179 | blk_mq_exit_hw_queues(q, set, set->nr_hw_queues); | |
2180 | blk_mq_free_hw_queues(q, set); | |
2181 | } | |
2182 | ||
2183 | /* Basically redo blk_mq_init_queue with queue frozen */ | |
2184 | static void blk_mq_queue_reinit(struct request_queue *q, | |
2185 | const struct cpumask *online_mask) | |
2186 | { | |
2187 | WARN_ON_ONCE(!atomic_read(&q->mq_freeze_depth)); | |
2188 | ||
2189 | blk_mq_sysfs_unregister(q); | |
2190 | ||
2191 | /* | |
2192 | * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe | |
2193 | * we should change hctx numa_node according to new topology (this | |
2194 | * involves free and re-allocate memory, worthy doing?) | |
2195 | */ | |
2196 | ||
2197 | blk_mq_map_swqueue(q, online_mask); | |
2198 | ||
2199 | blk_mq_sysfs_register(q); | |
2200 | } | |
2201 | ||
2202 | /* | |
2203 | * New online cpumask which is going to be set in this hotplug event. | |
2204 | * Declare this cpumasks as global as cpu-hotplug operation is invoked | |
2205 | * one-by-one and dynamically allocating this could result in a failure. | |
2206 | */ | |
2207 | static struct cpumask cpuhp_online_new; | |
2208 | ||
2209 | static void blk_mq_queue_reinit_work(void) | |
2210 | { | |
2211 | struct request_queue *q; | |
2212 | ||
2213 | mutex_lock(&all_q_mutex); | |
2214 | /* | |
2215 | * We need to freeze and reinit all existing queues. Freezing | |
2216 | * involves synchronous wait for an RCU grace period and doing it | |
2217 | * one by one may take a long time. Start freezing all queues in | |
2218 | * one swoop and then wait for the completions so that freezing can | |
2219 | * take place in parallel. | |
2220 | */ | |
2221 | list_for_each_entry(q, &all_q_list, all_q_node) | |
2222 | blk_mq_freeze_queue_start(q); | |
2223 | list_for_each_entry(q, &all_q_list, all_q_node) | |
2224 | blk_mq_freeze_queue_wait(q); | |
2225 | ||
2226 | list_for_each_entry(q, &all_q_list, all_q_node) | |
2227 | blk_mq_queue_reinit(q, &cpuhp_online_new); | |
2228 | ||
2229 | list_for_each_entry(q, &all_q_list, all_q_node) | |
2230 | blk_mq_unfreeze_queue(q); | |
2231 | ||
2232 | mutex_unlock(&all_q_mutex); | |
2233 | } | |
2234 | ||
2235 | static int blk_mq_queue_reinit_dead(unsigned int cpu) | |
2236 | { | |
2237 | cpumask_copy(&cpuhp_online_new, cpu_online_mask); | |
2238 | blk_mq_queue_reinit_work(); | |
2239 | return 0; | |
2240 | } | |
2241 | ||
2242 | /* | |
2243 | * Before hotadded cpu starts handling requests, new mappings must be | |
2244 | * established. Otherwise, these requests in hw queue might never be | |
2245 | * dispatched. | |
2246 | * | |
2247 | * For example, there is a single hw queue (hctx) and two CPU queues (ctx0 | |
2248 | * for CPU0, and ctx1 for CPU1). | |
2249 | * | |
2250 | * Now CPU1 is just onlined and a request is inserted into ctx1->rq_list | |
2251 | * and set bit0 in pending bitmap as ctx1->index_hw is still zero. | |
2252 | * | |
2253 | * And then while running hw queue, flush_busy_ctxs() finds bit0 is set in | |
2254 | * pending bitmap and tries to retrieve requests in hctx->ctxs[0]->rq_list. | |
2255 | * But htx->ctxs[0] is a pointer to ctx0, so the request in ctx1->rq_list | |
2256 | * is ignored. | |
2257 | */ | |
2258 | static int blk_mq_queue_reinit_prepare(unsigned int cpu) | |
2259 | { | |
2260 | cpumask_copy(&cpuhp_online_new, cpu_online_mask); | |
2261 | cpumask_set_cpu(cpu, &cpuhp_online_new); | |
2262 | blk_mq_queue_reinit_work(); | |
2263 | return 0; | |
2264 | } | |
2265 | ||
2266 | static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set) | |
2267 | { | |
2268 | int i; | |
2269 | ||
2270 | for (i = 0; i < set->nr_hw_queues; i++) { | |
2271 | set->tags[i] = blk_mq_init_rq_map(set, i); | |
2272 | if (!set->tags[i]) | |
2273 | goto out_unwind; | |
2274 | } | |
2275 | ||
2276 | return 0; | |
2277 | ||
2278 | out_unwind: | |
2279 | while (--i >= 0) | |
2280 | blk_mq_free_rq_map(set, set->tags[i], i); | |
2281 | ||
2282 | return -ENOMEM; | |
2283 | } | |
2284 | ||
2285 | /* | |
2286 | * Allocate the request maps associated with this tag_set. Note that this | |
2287 | * may reduce the depth asked for, if memory is tight. set->queue_depth | |
2288 | * will be updated to reflect the allocated depth. | |
2289 | */ | |
2290 | static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set) | |
2291 | { | |
2292 | unsigned int depth; | |
2293 | int err; | |
2294 | ||
2295 | depth = set->queue_depth; | |
2296 | do { | |
2297 | err = __blk_mq_alloc_rq_maps(set); | |
2298 | if (!err) | |
2299 | break; | |
2300 | ||
2301 | set->queue_depth >>= 1; | |
2302 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) { | |
2303 | err = -ENOMEM; | |
2304 | break; | |
2305 | } | |
2306 | } while (set->queue_depth); | |
2307 | ||
2308 | if (!set->queue_depth || err) { | |
2309 | pr_err("blk-mq: failed to allocate request map\n"); | |
2310 | return -ENOMEM; | |
2311 | } | |
2312 | ||
2313 | if (depth != set->queue_depth) | |
2314 | pr_info("blk-mq: reduced tag depth (%u -> %u)\n", | |
2315 | depth, set->queue_depth); | |
2316 | ||
2317 | return 0; | |
2318 | } | |
2319 | ||
2320 | /* | |
2321 | * Alloc a tag set to be associated with one or more request queues. | |
2322 | * May fail with EINVAL for various error conditions. May adjust the | |
2323 | * requested depth down, if if it too large. In that case, the set | |
2324 | * value will be stored in set->queue_depth. | |
2325 | */ | |
2326 | int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set) | |
2327 | { | |
2328 | int ret; | |
2329 | ||
2330 | BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS); | |
2331 | ||
2332 | if (!set->nr_hw_queues) | |
2333 | return -EINVAL; | |
2334 | if (!set->queue_depth) | |
2335 | return -EINVAL; | |
2336 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) | |
2337 | return -EINVAL; | |
2338 | ||
2339 | if (!set->ops->queue_rq) | |
2340 | return -EINVAL; | |
2341 | ||
2342 | if (set->queue_depth > BLK_MQ_MAX_DEPTH) { | |
2343 | pr_info("blk-mq: reduced tag depth to %u\n", | |
2344 | BLK_MQ_MAX_DEPTH); | |
2345 | set->queue_depth = BLK_MQ_MAX_DEPTH; | |
2346 | } | |
2347 | ||
2348 | /* | |
2349 | * If a crashdump is active, then we are potentially in a very | |
2350 | * memory constrained environment. Limit us to 1 queue and | |
2351 | * 64 tags to prevent using too much memory. | |
2352 | */ | |
2353 | if (is_kdump_kernel()) { | |
2354 | set->nr_hw_queues = 1; | |
2355 | set->queue_depth = min(64U, set->queue_depth); | |
2356 | } | |
2357 | /* | |
2358 | * There is no use for more h/w queues than cpus. | |
2359 | */ | |
2360 | if (set->nr_hw_queues > nr_cpu_ids) | |
2361 | set->nr_hw_queues = nr_cpu_ids; | |
2362 | ||
2363 | set->tags = kzalloc_node(nr_cpu_ids * sizeof(struct blk_mq_tags *), | |
2364 | GFP_KERNEL, set->numa_node); | |
2365 | if (!set->tags) | |
2366 | return -ENOMEM; | |
2367 | ||
2368 | ret = -ENOMEM; | |
2369 | set->mq_map = kzalloc_node(sizeof(*set->mq_map) * nr_cpu_ids, | |
2370 | GFP_KERNEL, set->numa_node); | |
2371 | if (!set->mq_map) | |
2372 | goto out_free_tags; | |
2373 | ||
2374 | if (set->ops->map_queues) | |
2375 | ret = set->ops->map_queues(set); | |
2376 | else | |
2377 | ret = blk_mq_map_queues(set); | |
2378 | if (ret) | |
2379 | goto out_free_mq_map; | |
2380 | ||
2381 | ret = blk_mq_alloc_rq_maps(set); | |
2382 | if (ret) | |
2383 | goto out_free_mq_map; | |
2384 | ||
2385 | mutex_init(&set->tag_list_lock); | |
2386 | INIT_LIST_HEAD(&set->tag_list); | |
2387 | ||
2388 | return 0; | |
2389 | ||
2390 | out_free_mq_map: | |
2391 | kfree(set->mq_map); | |
2392 | set->mq_map = NULL; | |
2393 | out_free_tags: | |
2394 | kfree(set->tags); | |
2395 | set->tags = NULL; | |
2396 | return ret; | |
2397 | } | |
2398 | EXPORT_SYMBOL(blk_mq_alloc_tag_set); | |
2399 | ||
2400 | void blk_mq_free_tag_set(struct blk_mq_tag_set *set) | |
2401 | { | |
2402 | int i; | |
2403 | ||
2404 | for (i = 0; i < nr_cpu_ids; i++) { | |
2405 | if (set->tags[i]) | |
2406 | blk_mq_free_rq_map(set, set->tags[i], i); | |
2407 | } | |
2408 | ||
2409 | kfree(set->mq_map); | |
2410 | set->mq_map = NULL; | |
2411 | ||
2412 | kfree(set->tags); | |
2413 | set->tags = NULL; | |
2414 | } | |
2415 | EXPORT_SYMBOL(blk_mq_free_tag_set); | |
2416 | ||
2417 | int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr) | |
2418 | { | |
2419 | struct blk_mq_tag_set *set = q->tag_set; | |
2420 | struct blk_mq_hw_ctx *hctx; | |
2421 | int i, ret; | |
2422 | ||
2423 | if (!set || nr > set->queue_depth) | |
2424 | return -EINVAL; | |
2425 | ||
2426 | ret = 0; | |
2427 | queue_for_each_hw_ctx(q, hctx, i) { | |
2428 | if (!hctx->tags) | |
2429 | continue; | |
2430 | ret = blk_mq_tag_update_depth(hctx->tags, nr); | |
2431 | if (ret) | |
2432 | break; | |
2433 | } | |
2434 | ||
2435 | if (!ret) | |
2436 | q->nr_requests = nr; | |
2437 | ||
2438 | return ret; | |
2439 | } | |
2440 | ||
2441 | void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues) | |
2442 | { | |
2443 | struct request_queue *q; | |
2444 | ||
2445 | if (nr_hw_queues > nr_cpu_ids) | |
2446 | nr_hw_queues = nr_cpu_ids; | |
2447 | if (nr_hw_queues < 1 || nr_hw_queues == set->nr_hw_queues) | |
2448 | return; | |
2449 | ||
2450 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
2451 | blk_mq_freeze_queue(q); | |
2452 | ||
2453 | set->nr_hw_queues = nr_hw_queues; | |
2454 | list_for_each_entry(q, &set->tag_list, tag_set_list) { | |
2455 | blk_mq_realloc_hw_ctxs(set, q); | |
2456 | ||
2457 | if (q->nr_hw_queues > 1) | |
2458 | blk_queue_make_request(q, blk_mq_make_request); | |
2459 | else | |
2460 | blk_queue_make_request(q, blk_sq_make_request); | |
2461 | ||
2462 | blk_mq_queue_reinit(q, cpu_online_mask); | |
2463 | } | |
2464 | ||
2465 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
2466 | blk_mq_unfreeze_queue(q); | |
2467 | } | |
2468 | EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues); | |
2469 | ||
2470 | static unsigned long blk_mq_poll_nsecs(struct request_queue *q, | |
2471 | struct blk_mq_hw_ctx *hctx, | |
2472 | struct request *rq) | |
2473 | { | |
2474 | struct blk_rq_stat stat[2]; | |
2475 | unsigned long ret = 0; | |
2476 | ||
2477 | /* | |
2478 | * If stats collection isn't on, don't sleep but turn it on for | |
2479 | * future users | |
2480 | */ | |
2481 | if (!blk_stat_enable(q)) | |
2482 | return 0; | |
2483 | ||
2484 | /* | |
2485 | * We don't have to do this once per IO, should optimize this | |
2486 | * to just use the current window of stats until it changes | |
2487 | */ | |
2488 | memset(&stat, 0, sizeof(stat)); | |
2489 | blk_hctx_stat_get(hctx, stat); | |
2490 | ||
2491 | /* | |
2492 | * As an optimistic guess, use half of the mean service time | |
2493 | * for this type of request. We can (and should) make this smarter. | |
2494 | * For instance, if the completion latencies are tight, we can | |
2495 | * get closer than just half the mean. This is especially | |
2496 | * important on devices where the completion latencies are longer | |
2497 | * than ~10 usec. | |
2498 | */ | |
2499 | if (req_op(rq) == REQ_OP_READ && stat[BLK_STAT_READ].nr_samples) | |
2500 | ret = (stat[BLK_STAT_READ].mean + 1) / 2; | |
2501 | else if (req_op(rq) == REQ_OP_WRITE && stat[BLK_STAT_WRITE].nr_samples) | |
2502 | ret = (stat[BLK_STAT_WRITE].mean + 1) / 2; | |
2503 | ||
2504 | return ret; | |
2505 | } | |
2506 | ||
2507 | static bool blk_mq_poll_hybrid_sleep(struct request_queue *q, | |
2508 | struct blk_mq_hw_ctx *hctx, | |
2509 | struct request *rq) | |
2510 | { | |
2511 | struct hrtimer_sleeper hs; | |
2512 | enum hrtimer_mode mode; | |
2513 | unsigned int nsecs; | |
2514 | ktime_t kt; | |
2515 | ||
2516 | if (test_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags)) | |
2517 | return false; | |
2518 | ||
2519 | /* | |
2520 | * poll_nsec can be: | |
2521 | * | |
2522 | * -1: don't ever hybrid sleep | |
2523 | * 0: use half of prev avg | |
2524 | * >0: use this specific value | |
2525 | */ | |
2526 | if (q->poll_nsec == -1) | |
2527 | return false; | |
2528 | else if (q->poll_nsec > 0) | |
2529 | nsecs = q->poll_nsec; | |
2530 | else | |
2531 | nsecs = blk_mq_poll_nsecs(q, hctx, rq); | |
2532 | ||
2533 | if (!nsecs) | |
2534 | return false; | |
2535 | ||
2536 | set_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags); | |
2537 | ||
2538 | /* | |
2539 | * This will be replaced with the stats tracking code, using | |
2540 | * 'avg_completion_time / 2' as the pre-sleep target. | |
2541 | */ | |
2542 | kt = ktime_set(0, nsecs); | |
2543 | ||
2544 | mode = HRTIMER_MODE_REL; | |
2545 | hrtimer_init_on_stack(&hs.timer, CLOCK_MONOTONIC, mode); | |
2546 | hrtimer_set_expires(&hs.timer, kt); | |
2547 | ||
2548 | hrtimer_init_sleeper(&hs, current); | |
2549 | do { | |
2550 | if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags)) | |
2551 | break; | |
2552 | set_current_state(TASK_UNINTERRUPTIBLE); | |
2553 | hrtimer_start_expires(&hs.timer, mode); | |
2554 | if (hs.task) | |
2555 | io_schedule(); | |
2556 | hrtimer_cancel(&hs.timer); | |
2557 | mode = HRTIMER_MODE_ABS; | |
2558 | } while (hs.task && !signal_pending(current)); | |
2559 | ||
2560 | __set_current_state(TASK_RUNNING); | |
2561 | destroy_hrtimer_on_stack(&hs.timer); | |
2562 | return true; | |
2563 | } | |
2564 | ||
2565 | static bool __blk_mq_poll(struct blk_mq_hw_ctx *hctx, struct request *rq) | |
2566 | { | |
2567 | struct request_queue *q = hctx->queue; | |
2568 | long state; | |
2569 | ||
2570 | /* | |
2571 | * If we sleep, have the caller restart the poll loop to reset | |
2572 | * the state. Like for the other success return cases, the | |
2573 | * caller is responsible for checking if the IO completed. If | |
2574 | * the IO isn't complete, we'll get called again and will go | |
2575 | * straight to the busy poll loop. | |
2576 | */ | |
2577 | if (blk_mq_poll_hybrid_sleep(q, hctx, rq)) | |
2578 | return true; | |
2579 | ||
2580 | hctx->poll_considered++; | |
2581 | ||
2582 | state = current->state; | |
2583 | while (!need_resched()) { | |
2584 | int ret; | |
2585 | ||
2586 | hctx->poll_invoked++; | |
2587 | ||
2588 | ret = q->mq_ops->poll(hctx, rq->tag); | |
2589 | if (ret > 0) { | |
2590 | hctx->poll_success++; | |
2591 | set_current_state(TASK_RUNNING); | |
2592 | return true; | |
2593 | } | |
2594 | ||
2595 | if (signal_pending_state(state, current)) | |
2596 | set_current_state(TASK_RUNNING); | |
2597 | ||
2598 | if (current->state == TASK_RUNNING) | |
2599 | return true; | |
2600 | if (ret < 0) | |
2601 | break; | |
2602 | cpu_relax(); | |
2603 | } | |
2604 | ||
2605 | return false; | |
2606 | } | |
2607 | ||
2608 | bool blk_mq_poll(struct request_queue *q, blk_qc_t cookie) | |
2609 | { | |
2610 | struct blk_mq_hw_ctx *hctx; | |
2611 | struct blk_plug *plug; | |
2612 | struct request *rq; | |
2613 | ||
2614 | if (!q->mq_ops || !q->mq_ops->poll || !blk_qc_t_valid(cookie) || | |
2615 | !test_bit(QUEUE_FLAG_POLL, &q->queue_flags)) | |
2616 | return false; | |
2617 | ||
2618 | plug = current->plug; | |
2619 | if (plug) | |
2620 | blk_flush_plug_list(plug, false); | |
2621 | ||
2622 | hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)]; | |
2623 | rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie)); | |
2624 | ||
2625 | return __blk_mq_poll(hctx, rq); | |
2626 | } | |
2627 | EXPORT_SYMBOL_GPL(blk_mq_poll); | |
2628 | ||
2629 | void blk_mq_disable_hotplug(void) | |
2630 | { | |
2631 | mutex_lock(&all_q_mutex); | |
2632 | } | |
2633 | ||
2634 | void blk_mq_enable_hotplug(void) | |
2635 | { | |
2636 | mutex_unlock(&all_q_mutex); | |
2637 | } | |
2638 | ||
2639 | static int __init blk_mq_init(void) | |
2640 | { | |
2641 | cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL, | |
2642 | blk_mq_hctx_notify_dead); | |
2643 | ||
2644 | cpuhp_setup_state_nocalls(CPUHP_BLK_MQ_PREPARE, "block/mq:prepare", | |
2645 | blk_mq_queue_reinit_prepare, | |
2646 | blk_mq_queue_reinit_dead); | |
2647 | return 0; | |
2648 | } | |
2649 | subsys_initcall(blk_mq_init); |