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