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