Commit | Line | Data |
---|---|---|
320ae51f JA |
1 | #include <linux/kernel.h> |
2 | #include <linux/module.h> | |
3 | #include <linux/backing-dev.h> | |
4 | #include <linux/bio.h> | |
5 | #include <linux/blkdev.h> | |
6 | #include <linux/mm.h> | |
7 | #include <linux/init.h> | |
8 | #include <linux/slab.h> | |
9 | #include <linux/workqueue.h> | |
10 | #include <linux/smp.h> | |
11 | #include <linux/llist.h> | |
12 | #include <linux/list_sort.h> | |
13 | #include <linux/cpu.h> | |
14 | #include <linux/cache.h> | |
15 | #include <linux/sched/sysctl.h> | |
16 | #include <linux/delay.h> | |
17 | ||
18 | #include <trace/events/block.h> | |
19 | ||
20 | #include <linux/blk-mq.h> | |
21 | #include "blk.h" | |
22 | #include "blk-mq.h" | |
23 | #include "blk-mq-tag.h" | |
24 | ||
25 | static DEFINE_MUTEX(all_q_mutex); | |
26 | static LIST_HEAD(all_q_list); | |
27 | ||
28 | static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx); | |
29 | ||
320ae51f JA |
30 | static struct blk_mq_ctx *__blk_mq_get_ctx(struct request_queue *q, |
31 | unsigned int cpu) | |
32 | { | |
33 | return per_cpu_ptr(q->queue_ctx, cpu); | |
34 | } | |
35 | ||
36 | /* | |
37 | * This assumes per-cpu software queueing queues. They could be per-node | |
38 | * as well, for instance. For now this is hardcoded as-is. Note that we don't | |
39 | * care about preemption, since we know the ctx's are persistent. This does | |
40 | * mean that we can't rely on ctx always matching the currently running CPU. | |
41 | */ | |
42 | static struct blk_mq_ctx *blk_mq_get_ctx(struct request_queue *q) | |
43 | { | |
44 | return __blk_mq_get_ctx(q, get_cpu()); | |
45 | } | |
46 | ||
47 | static void blk_mq_put_ctx(struct blk_mq_ctx *ctx) | |
48 | { | |
49 | put_cpu(); | |
50 | } | |
51 | ||
52 | /* | |
53 | * Check if any of the ctx's have pending work in this hardware queue | |
54 | */ | |
55 | static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx) | |
56 | { | |
57 | unsigned int i; | |
58 | ||
59 | for (i = 0; i < hctx->nr_ctx_map; i++) | |
60 | if (hctx->ctx_map[i]) | |
61 | return true; | |
62 | ||
63 | return false; | |
64 | } | |
65 | ||
66 | /* | |
67 | * Mark this ctx as having pending work in this hardware queue | |
68 | */ | |
69 | static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx, | |
70 | struct blk_mq_ctx *ctx) | |
71 | { | |
72 | if (!test_bit(ctx->index_hw, hctx->ctx_map)) | |
73 | set_bit(ctx->index_hw, hctx->ctx_map); | |
74 | } | |
75 | ||
081241e5 CH |
76 | static struct request *__blk_mq_alloc_request(struct blk_mq_hw_ctx *hctx, |
77 | gfp_t gfp, bool reserved) | |
320ae51f JA |
78 | { |
79 | struct request *rq; | |
80 | unsigned int tag; | |
81 | ||
82 | tag = blk_mq_get_tag(hctx->tags, gfp, reserved); | |
83 | if (tag != BLK_MQ_TAG_FAIL) { | |
84 | rq = hctx->rqs[tag]; | |
85 | rq->tag = tag; | |
86 | ||
87 | return rq; | |
88 | } | |
89 | ||
90 | return NULL; | |
91 | } | |
92 | ||
93 | static int blk_mq_queue_enter(struct request_queue *q) | |
94 | { | |
95 | int ret; | |
96 | ||
97 | __percpu_counter_add(&q->mq_usage_counter, 1, 1000000); | |
98 | smp_wmb(); | |
99 | /* we have problems to freeze the queue if it's initializing */ | |
100 | if (!blk_queue_bypass(q) || !blk_queue_init_done(q)) | |
101 | return 0; | |
102 | ||
103 | __percpu_counter_add(&q->mq_usage_counter, -1, 1000000); | |
104 | ||
105 | spin_lock_irq(q->queue_lock); | |
106 | ret = wait_event_interruptible_lock_irq(q->mq_freeze_wq, | |
43a5e4e2 ML |
107 | !blk_queue_bypass(q) || blk_queue_dying(q), |
108 | *q->queue_lock); | |
320ae51f | 109 | /* inc usage with lock hold to avoid freeze_queue runs here */ |
43a5e4e2 | 110 | if (!ret && !blk_queue_dying(q)) |
320ae51f | 111 | __percpu_counter_add(&q->mq_usage_counter, 1, 1000000); |
43a5e4e2 ML |
112 | else if (blk_queue_dying(q)) |
113 | ret = -ENODEV; | |
320ae51f JA |
114 | spin_unlock_irq(q->queue_lock); |
115 | ||
116 | return ret; | |
117 | } | |
118 | ||
119 | static void blk_mq_queue_exit(struct request_queue *q) | |
120 | { | |
121 | __percpu_counter_add(&q->mq_usage_counter, -1, 1000000); | |
122 | } | |
123 | ||
43a5e4e2 ML |
124 | static void __blk_mq_drain_queue(struct request_queue *q) |
125 | { | |
126 | while (true) { | |
127 | s64 count; | |
128 | ||
129 | spin_lock_irq(q->queue_lock); | |
130 | count = percpu_counter_sum(&q->mq_usage_counter); | |
131 | spin_unlock_irq(q->queue_lock); | |
132 | ||
133 | if (count == 0) | |
134 | break; | |
135 | blk_mq_run_queues(q, false); | |
136 | msleep(10); | |
137 | } | |
138 | } | |
139 | ||
320ae51f JA |
140 | /* |
141 | * Guarantee no request is in use, so we can change any data structure of | |
142 | * the queue afterward. | |
143 | */ | |
144 | static void blk_mq_freeze_queue(struct request_queue *q) | |
145 | { | |
146 | bool drain; | |
147 | ||
148 | spin_lock_irq(q->queue_lock); | |
149 | drain = !q->bypass_depth++; | |
150 | queue_flag_set(QUEUE_FLAG_BYPASS, q); | |
151 | spin_unlock_irq(q->queue_lock); | |
152 | ||
43a5e4e2 ML |
153 | if (drain) |
154 | __blk_mq_drain_queue(q); | |
155 | } | |
320ae51f | 156 | |
43a5e4e2 ML |
157 | void blk_mq_drain_queue(struct request_queue *q) |
158 | { | |
159 | __blk_mq_drain_queue(q); | |
320ae51f JA |
160 | } |
161 | ||
162 | static void blk_mq_unfreeze_queue(struct request_queue *q) | |
163 | { | |
164 | bool wake = false; | |
165 | ||
166 | spin_lock_irq(q->queue_lock); | |
167 | if (!--q->bypass_depth) { | |
168 | queue_flag_clear(QUEUE_FLAG_BYPASS, q); | |
169 | wake = true; | |
170 | } | |
171 | WARN_ON_ONCE(q->bypass_depth < 0); | |
172 | spin_unlock_irq(q->queue_lock); | |
173 | if (wake) | |
174 | wake_up_all(&q->mq_freeze_wq); | |
175 | } | |
176 | ||
177 | bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx) | |
178 | { | |
179 | return blk_mq_has_free_tags(hctx->tags); | |
180 | } | |
181 | EXPORT_SYMBOL(blk_mq_can_queue); | |
182 | ||
94eddfbe JA |
183 | static void blk_mq_rq_ctx_init(struct request_queue *q, struct blk_mq_ctx *ctx, |
184 | struct request *rq, unsigned int rw_flags) | |
320ae51f | 185 | { |
94eddfbe JA |
186 | if (blk_queue_io_stat(q)) |
187 | rw_flags |= REQ_IO_STAT; | |
188 | ||
320ae51f JA |
189 | rq->mq_ctx = ctx; |
190 | rq->cmd_flags = rw_flags; | |
0fec08b4 ML |
191 | rq->start_time = jiffies; |
192 | set_start_time_ns(rq); | |
320ae51f JA |
193 | ctx->rq_dispatched[rw_is_sync(rw_flags)]++; |
194 | } | |
195 | ||
320ae51f JA |
196 | static struct request *blk_mq_alloc_request_pinned(struct request_queue *q, |
197 | int rw, gfp_t gfp, | |
198 | bool reserved) | |
199 | { | |
200 | struct request *rq; | |
201 | ||
202 | do { | |
203 | struct blk_mq_ctx *ctx = blk_mq_get_ctx(q); | |
204 | struct blk_mq_hw_ctx *hctx = q->mq_ops->map_queue(q, ctx->cpu); | |
205 | ||
18741986 | 206 | rq = __blk_mq_alloc_request(hctx, gfp & ~__GFP_WAIT, reserved); |
320ae51f | 207 | if (rq) { |
94eddfbe | 208 | blk_mq_rq_ctx_init(q, ctx, rq, rw); |
320ae51f | 209 | break; |
959a35f1 | 210 | } |
320ae51f | 211 | |
e4043dcf JA |
212 | if (gfp & __GFP_WAIT) { |
213 | __blk_mq_run_hw_queue(hctx); | |
214 | blk_mq_put_ctx(ctx); | |
215 | } else { | |
216 | blk_mq_put_ctx(ctx); | |
959a35f1 | 217 | break; |
e4043dcf | 218 | } |
959a35f1 | 219 | |
320ae51f JA |
220 | blk_mq_wait_for_tags(hctx->tags); |
221 | } while (1); | |
222 | ||
223 | return rq; | |
224 | } | |
225 | ||
18741986 | 226 | struct request *blk_mq_alloc_request(struct request_queue *q, int rw, gfp_t gfp) |
320ae51f JA |
227 | { |
228 | struct request *rq; | |
229 | ||
230 | if (blk_mq_queue_enter(q)) | |
231 | return NULL; | |
232 | ||
18741986 | 233 | rq = blk_mq_alloc_request_pinned(q, rw, gfp, false); |
959a35f1 JM |
234 | if (rq) |
235 | blk_mq_put_ctx(rq->mq_ctx); | |
320ae51f JA |
236 | return rq; |
237 | } | |
238 | ||
239 | struct request *blk_mq_alloc_reserved_request(struct request_queue *q, int rw, | |
240 | gfp_t gfp) | |
241 | { | |
242 | struct request *rq; | |
243 | ||
244 | if (blk_mq_queue_enter(q)) | |
245 | return NULL; | |
246 | ||
247 | rq = blk_mq_alloc_request_pinned(q, rw, gfp, true); | |
959a35f1 JM |
248 | if (rq) |
249 | blk_mq_put_ctx(rq->mq_ctx); | |
320ae51f JA |
250 | return rq; |
251 | } | |
252 | EXPORT_SYMBOL(blk_mq_alloc_reserved_request); | |
253 | ||
254 | /* | |
255 | * Re-init and set pdu, if we have it | |
256 | */ | |
18741986 | 257 | void blk_mq_rq_init(struct blk_mq_hw_ctx *hctx, struct request *rq) |
320ae51f JA |
258 | { |
259 | blk_rq_init(hctx->queue, rq); | |
260 | ||
261 | if (hctx->cmd_size) | |
262 | rq->special = blk_mq_rq_to_pdu(rq); | |
263 | } | |
264 | ||
265 | static void __blk_mq_free_request(struct blk_mq_hw_ctx *hctx, | |
266 | struct blk_mq_ctx *ctx, struct request *rq) | |
267 | { | |
268 | const int tag = rq->tag; | |
269 | struct request_queue *q = rq->q; | |
270 | ||
271 | blk_mq_rq_init(hctx, rq); | |
272 | blk_mq_put_tag(hctx->tags, tag); | |
273 | ||
274 | blk_mq_queue_exit(q); | |
275 | } | |
276 | ||
277 | void blk_mq_free_request(struct request *rq) | |
278 | { | |
279 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
280 | struct blk_mq_hw_ctx *hctx; | |
281 | struct request_queue *q = rq->q; | |
282 | ||
283 | ctx->rq_completed[rq_is_sync(rq)]++; | |
284 | ||
285 | hctx = q->mq_ops->map_queue(q, ctx->cpu); | |
286 | __blk_mq_free_request(hctx, ctx, rq); | |
287 | } | |
288 | ||
7237c740 | 289 | bool blk_mq_end_io_partial(struct request *rq, int error, unsigned int nr_bytes) |
320ae51f | 290 | { |
7237c740 CH |
291 | if (blk_update_request(rq, error, blk_rq_bytes(rq))) |
292 | return true; | |
320ae51f | 293 | |
0d11e6ac ML |
294 | blk_account_io_done(rq); |
295 | ||
320ae51f JA |
296 | if (rq->end_io) |
297 | rq->end_io(rq, error); | |
298 | else | |
299 | blk_mq_free_request(rq); | |
7237c740 | 300 | return false; |
320ae51f | 301 | } |
7237c740 | 302 | EXPORT_SYMBOL(blk_mq_end_io_partial); |
320ae51f | 303 | |
30a91cb4 | 304 | static void __blk_mq_complete_request_remote(void *data) |
320ae51f | 305 | { |
3d6efbf6 | 306 | struct request *rq = data; |
320ae51f | 307 | |
30a91cb4 | 308 | rq->q->softirq_done_fn(rq); |
320ae51f | 309 | } |
320ae51f | 310 | |
30a91cb4 | 311 | void __blk_mq_complete_request(struct request *rq) |
320ae51f JA |
312 | { |
313 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
314 | int cpu; | |
315 | ||
30a91cb4 CH |
316 | if (!ctx->ipi_redirect) { |
317 | rq->q->softirq_done_fn(rq); | |
318 | return; | |
319 | } | |
320ae51f JA |
320 | |
321 | cpu = get_cpu(); | |
3d6efbf6 | 322 | if (cpu != ctx->cpu && cpu_online(ctx->cpu)) { |
30a91cb4 | 323 | rq->csd.func = __blk_mq_complete_request_remote; |
3d6efbf6 CH |
324 | rq->csd.info = rq; |
325 | rq->csd.flags = 0; | |
c46fff2a | 326 | smp_call_function_single_async(ctx->cpu, &rq->csd); |
3d6efbf6 | 327 | } else { |
30a91cb4 | 328 | rq->q->softirq_done_fn(rq); |
3d6efbf6 | 329 | } |
320ae51f JA |
330 | put_cpu(); |
331 | } | |
30a91cb4 CH |
332 | |
333 | /** | |
334 | * blk_mq_complete_request - end I/O on a request | |
335 | * @rq: the request being processed | |
336 | * | |
337 | * Description: | |
338 | * Ends all I/O on a request. It does not handle partial completions. | |
339 | * The actual completion happens out-of-order, through a IPI handler. | |
340 | **/ | |
341 | void blk_mq_complete_request(struct request *rq) | |
342 | { | |
343 | if (unlikely(blk_should_fake_timeout(rq->q))) | |
344 | return; | |
345 | if (!blk_mark_rq_complete(rq)) | |
346 | __blk_mq_complete_request(rq); | |
347 | } | |
348 | EXPORT_SYMBOL(blk_mq_complete_request); | |
320ae51f | 349 | |
49f5baa5 | 350 | static void blk_mq_start_request(struct request *rq, bool last) |
320ae51f JA |
351 | { |
352 | struct request_queue *q = rq->q; | |
353 | ||
354 | trace_block_rq_issue(q, rq); | |
355 | ||
356 | /* | |
357 | * Just mark start time and set the started bit. Due to memory | |
358 | * ordering, we know we'll see the correct deadline as long as | |
359 | * REQ_ATOMIC_STARTED is seen. | |
360 | */ | |
361 | rq->deadline = jiffies + q->rq_timeout; | |
362 | set_bit(REQ_ATOM_STARTED, &rq->atomic_flags); | |
49f5baa5 CH |
363 | |
364 | if (q->dma_drain_size && blk_rq_bytes(rq)) { | |
365 | /* | |
366 | * Make sure space for the drain appears. We know we can do | |
367 | * this because max_hw_segments has been adjusted to be one | |
368 | * fewer than the device can handle. | |
369 | */ | |
370 | rq->nr_phys_segments++; | |
371 | } | |
372 | ||
373 | /* | |
374 | * Flag the last request in the series so that drivers know when IO | |
375 | * should be kicked off, if they don't do it on a per-request basis. | |
376 | * | |
377 | * Note: the flag isn't the only condition drivers should do kick off. | |
378 | * If drive is busy, the last request might not have the bit set. | |
379 | */ | |
380 | if (last) | |
381 | rq->cmd_flags |= REQ_END; | |
320ae51f JA |
382 | } |
383 | ||
384 | static void blk_mq_requeue_request(struct request *rq) | |
385 | { | |
386 | struct request_queue *q = rq->q; | |
387 | ||
388 | trace_block_rq_requeue(q, rq); | |
389 | clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags); | |
49f5baa5 CH |
390 | |
391 | rq->cmd_flags &= ~REQ_END; | |
392 | ||
393 | if (q->dma_drain_size && blk_rq_bytes(rq)) | |
394 | rq->nr_phys_segments--; | |
320ae51f JA |
395 | } |
396 | ||
397 | struct blk_mq_timeout_data { | |
398 | struct blk_mq_hw_ctx *hctx; | |
399 | unsigned long *next; | |
400 | unsigned int *next_set; | |
401 | }; | |
402 | ||
403 | static void blk_mq_timeout_check(void *__data, unsigned long *free_tags) | |
404 | { | |
405 | struct blk_mq_timeout_data *data = __data; | |
406 | struct blk_mq_hw_ctx *hctx = data->hctx; | |
407 | unsigned int tag; | |
408 | ||
409 | /* It may not be in flight yet (this is where | |
410 | * the REQ_ATOMIC_STARTED flag comes in). The requests are | |
411 | * statically allocated, so we know it's always safe to access the | |
412 | * memory associated with a bit offset into ->rqs[]. | |
413 | */ | |
414 | tag = 0; | |
415 | do { | |
416 | struct request *rq; | |
417 | ||
418 | tag = find_next_zero_bit(free_tags, hctx->queue_depth, tag); | |
419 | if (tag >= hctx->queue_depth) | |
420 | break; | |
421 | ||
422 | rq = hctx->rqs[tag++]; | |
423 | ||
424 | if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) | |
425 | continue; | |
426 | ||
427 | blk_rq_check_expired(rq, data->next, data->next_set); | |
428 | } while (1); | |
429 | } | |
430 | ||
431 | static void blk_mq_hw_ctx_check_timeout(struct blk_mq_hw_ctx *hctx, | |
432 | unsigned long *next, | |
433 | unsigned int *next_set) | |
434 | { | |
435 | struct blk_mq_timeout_data data = { | |
436 | .hctx = hctx, | |
437 | .next = next, | |
438 | .next_set = next_set, | |
439 | }; | |
440 | ||
441 | /* | |
442 | * Ask the tagging code to iterate busy requests, so we can | |
443 | * check them for timeout. | |
444 | */ | |
445 | blk_mq_tag_busy_iter(hctx->tags, blk_mq_timeout_check, &data); | |
446 | } | |
447 | ||
448 | static void blk_mq_rq_timer(unsigned long data) | |
449 | { | |
450 | struct request_queue *q = (struct request_queue *) data; | |
451 | struct blk_mq_hw_ctx *hctx; | |
452 | unsigned long next = 0; | |
453 | int i, next_set = 0; | |
454 | ||
455 | queue_for_each_hw_ctx(q, hctx, i) | |
456 | blk_mq_hw_ctx_check_timeout(hctx, &next, &next_set); | |
457 | ||
458 | if (next_set) | |
459 | mod_timer(&q->timeout, round_jiffies_up(next)); | |
460 | } | |
461 | ||
462 | /* | |
463 | * Reverse check our software queue for entries that we could potentially | |
464 | * merge with. Currently includes a hand-wavy stop count of 8, to not spend | |
465 | * too much time checking for merges. | |
466 | */ | |
467 | static bool blk_mq_attempt_merge(struct request_queue *q, | |
468 | struct blk_mq_ctx *ctx, struct bio *bio) | |
469 | { | |
470 | struct request *rq; | |
471 | int checked = 8; | |
472 | ||
473 | list_for_each_entry_reverse(rq, &ctx->rq_list, queuelist) { | |
474 | int el_ret; | |
475 | ||
476 | if (!checked--) | |
477 | break; | |
478 | ||
479 | if (!blk_rq_merge_ok(rq, bio)) | |
480 | continue; | |
481 | ||
482 | el_ret = blk_try_merge(rq, bio); | |
483 | if (el_ret == ELEVATOR_BACK_MERGE) { | |
484 | if (bio_attempt_back_merge(q, rq, bio)) { | |
485 | ctx->rq_merged++; | |
486 | return true; | |
487 | } | |
488 | break; | |
489 | } else if (el_ret == ELEVATOR_FRONT_MERGE) { | |
490 | if (bio_attempt_front_merge(q, rq, bio)) { | |
491 | ctx->rq_merged++; | |
492 | return true; | |
493 | } | |
494 | break; | |
495 | } | |
496 | } | |
497 | ||
498 | return false; | |
499 | } | |
500 | ||
501 | void blk_mq_add_timer(struct request *rq) | |
502 | { | |
503 | __blk_add_timer(rq, NULL); | |
504 | } | |
505 | ||
506 | /* | |
507 | * Run this hardware queue, pulling any software queues mapped to it in. | |
508 | * Note that this function currently has various problems around ordering | |
509 | * of IO. In particular, we'd like FIFO behaviour on handling existing | |
510 | * items on the hctx->dispatch list. Ignore that for now. | |
511 | */ | |
512 | static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx) | |
513 | { | |
514 | struct request_queue *q = hctx->queue; | |
515 | struct blk_mq_ctx *ctx; | |
516 | struct request *rq; | |
517 | LIST_HEAD(rq_list); | |
518 | int bit, queued; | |
519 | ||
e4043dcf JA |
520 | WARN_ON(!preempt_count()); |
521 | ||
5d12f905 | 522 | if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state))) |
320ae51f JA |
523 | return; |
524 | ||
525 | hctx->run++; | |
526 | ||
527 | /* | |
528 | * Touch any software queue that has pending entries. | |
529 | */ | |
530 | for_each_set_bit(bit, hctx->ctx_map, hctx->nr_ctx) { | |
531 | clear_bit(bit, hctx->ctx_map); | |
532 | ctx = hctx->ctxs[bit]; | |
533 | BUG_ON(bit != ctx->index_hw); | |
534 | ||
535 | spin_lock(&ctx->lock); | |
536 | list_splice_tail_init(&ctx->rq_list, &rq_list); | |
537 | spin_unlock(&ctx->lock); | |
538 | } | |
539 | ||
540 | /* | |
541 | * If we have previous entries on our dispatch list, grab them | |
542 | * and stuff them at the front for more fair dispatch. | |
543 | */ | |
544 | if (!list_empty_careful(&hctx->dispatch)) { | |
545 | spin_lock(&hctx->lock); | |
546 | if (!list_empty(&hctx->dispatch)) | |
547 | list_splice_init(&hctx->dispatch, &rq_list); | |
548 | spin_unlock(&hctx->lock); | |
549 | } | |
550 | ||
551 | /* | |
552 | * Delete and return all entries from our dispatch list | |
553 | */ | |
554 | queued = 0; | |
555 | ||
556 | /* | |
557 | * Now process all the entries, sending them to the driver. | |
558 | */ | |
559 | while (!list_empty(&rq_list)) { | |
560 | int ret; | |
561 | ||
562 | rq = list_first_entry(&rq_list, struct request, queuelist); | |
563 | list_del_init(&rq->queuelist); | |
320ae51f | 564 | |
49f5baa5 | 565 | blk_mq_start_request(rq, list_empty(&rq_list)); |
320ae51f JA |
566 | |
567 | ret = q->mq_ops->queue_rq(hctx, rq); | |
568 | switch (ret) { | |
569 | case BLK_MQ_RQ_QUEUE_OK: | |
570 | queued++; | |
571 | continue; | |
572 | case BLK_MQ_RQ_QUEUE_BUSY: | |
573 | /* | |
574 | * FIXME: we should have a mechanism to stop the queue | |
575 | * like blk_stop_queue, otherwise we will waste cpu | |
576 | * time | |
577 | */ | |
578 | list_add(&rq->queuelist, &rq_list); | |
579 | blk_mq_requeue_request(rq); | |
580 | break; | |
581 | default: | |
582 | pr_err("blk-mq: bad return on queue: %d\n", ret); | |
320ae51f | 583 | case BLK_MQ_RQ_QUEUE_ERROR: |
1e93b8c2 | 584 | rq->errors = -EIO; |
320ae51f JA |
585 | blk_mq_end_io(rq, rq->errors); |
586 | break; | |
587 | } | |
588 | ||
589 | if (ret == BLK_MQ_RQ_QUEUE_BUSY) | |
590 | break; | |
591 | } | |
592 | ||
593 | if (!queued) | |
594 | hctx->dispatched[0]++; | |
595 | else if (queued < (1 << (BLK_MQ_MAX_DISPATCH_ORDER - 1))) | |
596 | hctx->dispatched[ilog2(queued) + 1]++; | |
597 | ||
598 | /* | |
599 | * Any items that need requeuing? Stuff them into hctx->dispatch, | |
600 | * that is where we will continue on next queue run. | |
601 | */ | |
602 | if (!list_empty(&rq_list)) { | |
603 | spin_lock(&hctx->lock); | |
604 | list_splice(&rq_list, &hctx->dispatch); | |
605 | spin_unlock(&hctx->lock); | |
606 | } | |
607 | } | |
608 | ||
609 | void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) | |
610 | { | |
5d12f905 | 611 | if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state))) |
320ae51f JA |
612 | return; |
613 | ||
e4043dcf | 614 | if (!async && cpumask_test_cpu(smp_processor_id(), hctx->cpumask)) |
320ae51f | 615 | __blk_mq_run_hw_queue(hctx); |
e4043dcf | 616 | else if (hctx->queue->nr_hw_queues == 1) |
59c3d45e | 617 | kblockd_schedule_delayed_work(&hctx->delayed_work, 0); |
e4043dcf JA |
618 | else { |
619 | unsigned int cpu; | |
620 | ||
621 | /* | |
622 | * It'd be great if the workqueue API had a way to pass | |
623 | * in a mask and had some smarts for more clever placement | |
624 | * than the first CPU. Or we could round-robin here. For now, | |
625 | * just queue on the first CPU. | |
626 | */ | |
627 | cpu = cpumask_first(hctx->cpumask); | |
628 | kblockd_schedule_delayed_work_on(cpu, &hctx->delayed_work, 0); | |
629 | } | |
320ae51f JA |
630 | } |
631 | ||
632 | void blk_mq_run_queues(struct request_queue *q, bool async) | |
633 | { | |
634 | struct blk_mq_hw_ctx *hctx; | |
635 | int i; | |
636 | ||
637 | queue_for_each_hw_ctx(q, hctx, i) { | |
638 | if ((!blk_mq_hctx_has_pending(hctx) && | |
639 | list_empty_careful(&hctx->dispatch)) || | |
5d12f905 | 640 | test_bit(BLK_MQ_S_STOPPED, &hctx->state)) |
320ae51f JA |
641 | continue; |
642 | ||
e4043dcf | 643 | preempt_disable(); |
320ae51f | 644 | blk_mq_run_hw_queue(hctx, async); |
e4043dcf | 645 | preempt_enable(); |
320ae51f JA |
646 | } |
647 | } | |
648 | EXPORT_SYMBOL(blk_mq_run_queues); | |
649 | ||
650 | void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx) | |
651 | { | |
652 | cancel_delayed_work(&hctx->delayed_work); | |
653 | set_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
654 | } | |
655 | EXPORT_SYMBOL(blk_mq_stop_hw_queue); | |
656 | ||
280d45f6 CH |
657 | void blk_mq_stop_hw_queues(struct request_queue *q) |
658 | { | |
659 | struct blk_mq_hw_ctx *hctx; | |
660 | int i; | |
661 | ||
662 | queue_for_each_hw_ctx(q, hctx, i) | |
663 | blk_mq_stop_hw_queue(hctx); | |
664 | } | |
665 | EXPORT_SYMBOL(blk_mq_stop_hw_queues); | |
666 | ||
320ae51f JA |
667 | void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx) |
668 | { | |
669 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
e4043dcf JA |
670 | |
671 | preempt_disable(); | |
320ae51f | 672 | __blk_mq_run_hw_queue(hctx); |
e4043dcf | 673 | preempt_enable(); |
320ae51f JA |
674 | } |
675 | EXPORT_SYMBOL(blk_mq_start_hw_queue); | |
676 | ||
677 | void blk_mq_start_stopped_hw_queues(struct request_queue *q) | |
678 | { | |
679 | struct blk_mq_hw_ctx *hctx; | |
680 | int i; | |
681 | ||
682 | queue_for_each_hw_ctx(q, hctx, i) { | |
683 | if (!test_bit(BLK_MQ_S_STOPPED, &hctx->state)) | |
684 | continue; | |
685 | ||
686 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
e4043dcf | 687 | preempt_disable(); |
320ae51f | 688 | blk_mq_run_hw_queue(hctx, true); |
e4043dcf | 689 | preempt_enable(); |
320ae51f JA |
690 | } |
691 | } | |
692 | EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues); | |
693 | ||
694 | static void blk_mq_work_fn(struct work_struct *work) | |
695 | { | |
696 | struct blk_mq_hw_ctx *hctx; | |
697 | ||
698 | hctx = container_of(work, struct blk_mq_hw_ctx, delayed_work.work); | |
e4043dcf JA |
699 | |
700 | preempt_disable(); | |
320ae51f | 701 | __blk_mq_run_hw_queue(hctx); |
e4043dcf | 702 | preempt_enable(); |
320ae51f JA |
703 | } |
704 | ||
705 | static void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, | |
72a0a36e | 706 | struct request *rq, bool at_head) |
320ae51f JA |
707 | { |
708 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
709 | ||
01b983c9 JA |
710 | trace_block_rq_insert(hctx->queue, rq); |
711 | ||
72a0a36e CH |
712 | if (at_head) |
713 | list_add(&rq->queuelist, &ctx->rq_list); | |
714 | else | |
715 | list_add_tail(&rq->queuelist, &ctx->rq_list); | |
320ae51f JA |
716 | blk_mq_hctx_mark_pending(hctx, ctx); |
717 | ||
718 | /* | |
719 | * We do this early, to ensure we are on the right CPU. | |
720 | */ | |
721 | blk_mq_add_timer(rq); | |
722 | } | |
723 | ||
eeabc850 CH |
724 | void blk_mq_insert_request(struct request *rq, bool at_head, bool run_queue, |
725 | bool async) | |
320ae51f | 726 | { |
eeabc850 | 727 | struct request_queue *q = rq->q; |
320ae51f | 728 | struct blk_mq_hw_ctx *hctx; |
eeabc850 CH |
729 | struct blk_mq_ctx *ctx = rq->mq_ctx, *current_ctx; |
730 | ||
731 | current_ctx = blk_mq_get_ctx(q); | |
732 | if (!cpu_online(ctx->cpu)) | |
733 | rq->mq_ctx = ctx = current_ctx; | |
320ae51f | 734 | |
320ae51f JA |
735 | hctx = q->mq_ops->map_queue(q, ctx->cpu); |
736 | ||
eeabc850 CH |
737 | if (rq->cmd_flags & (REQ_FLUSH | REQ_FUA) && |
738 | !(rq->cmd_flags & (REQ_FLUSH_SEQ))) { | |
320ae51f JA |
739 | blk_insert_flush(rq); |
740 | } else { | |
320ae51f | 741 | spin_lock(&ctx->lock); |
72a0a36e | 742 | __blk_mq_insert_request(hctx, rq, at_head); |
320ae51f | 743 | spin_unlock(&ctx->lock); |
320ae51f JA |
744 | } |
745 | ||
320ae51f JA |
746 | if (run_queue) |
747 | blk_mq_run_hw_queue(hctx, async); | |
e4043dcf JA |
748 | |
749 | blk_mq_put_ctx(current_ctx); | |
320ae51f JA |
750 | } |
751 | ||
752 | static void blk_mq_insert_requests(struct request_queue *q, | |
753 | struct blk_mq_ctx *ctx, | |
754 | struct list_head *list, | |
755 | int depth, | |
756 | bool from_schedule) | |
757 | ||
758 | { | |
759 | struct blk_mq_hw_ctx *hctx; | |
760 | struct blk_mq_ctx *current_ctx; | |
761 | ||
762 | trace_block_unplug(q, depth, !from_schedule); | |
763 | ||
764 | current_ctx = blk_mq_get_ctx(q); | |
765 | ||
766 | if (!cpu_online(ctx->cpu)) | |
767 | ctx = current_ctx; | |
768 | hctx = q->mq_ops->map_queue(q, ctx->cpu); | |
769 | ||
770 | /* | |
771 | * preemption doesn't flush plug list, so it's possible ctx->cpu is | |
772 | * offline now | |
773 | */ | |
774 | spin_lock(&ctx->lock); | |
775 | while (!list_empty(list)) { | |
776 | struct request *rq; | |
777 | ||
778 | rq = list_first_entry(list, struct request, queuelist); | |
779 | list_del_init(&rq->queuelist); | |
780 | rq->mq_ctx = ctx; | |
72a0a36e | 781 | __blk_mq_insert_request(hctx, rq, false); |
320ae51f JA |
782 | } |
783 | spin_unlock(&ctx->lock); | |
784 | ||
320ae51f | 785 | blk_mq_run_hw_queue(hctx, from_schedule); |
e4043dcf | 786 | blk_mq_put_ctx(current_ctx); |
320ae51f JA |
787 | } |
788 | ||
789 | static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b) | |
790 | { | |
791 | struct request *rqa = container_of(a, struct request, queuelist); | |
792 | struct request *rqb = container_of(b, struct request, queuelist); | |
793 | ||
794 | return !(rqa->mq_ctx < rqb->mq_ctx || | |
795 | (rqa->mq_ctx == rqb->mq_ctx && | |
796 | blk_rq_pos(rqa) < blk_rq_pos(rqb))); | |
797 | } | |
798 | ||
799 | void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule) | |
800 | { | |
801 | struct blk_mq_ctx *this_ctx; | |
802 | struct request_queue *this_q; | |
803 | struct request *rq; | |
804 | LIST_HEAD(list); | |
805 | LIST_HEAD(ctx_list); | |
806 | unsigned int depth; | |
807 | ||
808 | list_splice_init(&plug->mq_list, &list); | |
809 | ||
810 | list_sort(NULL, &list, plug_ctx_cmp); | |
811 | ||
812 | this_q = NULL; | |
813 | this_ctx = NULL; | |
814 | depth = 0; | |
815 | ||
816 | while (!list_empty(&list)) { | |
817 | rq = list_entry_rq(list.next); | |
818 | list_del_init(&rq->queuelist); | |
819 | BUG_ON(!rq->q); | |
820 | if (rq->mq_ctx != this_ctx) { | |
821 | if (this_ctx) { | |
822 | blk_mq_insert_requests(this_q, this_ctx, | |
823 | &ctx_list, depth, | |
824 | from_schedule); | |
825 | } | |
826 | ||
827 | this_ctx = rq->mq_ctx; | |
828 | this_q = rq->q; | |
829 | depth = 0; | |
830 | } | |
831 | ||
832 | depth++; | |
833 | list_add_tail(&rq->queuelist, &ctx_list); | |
834 | } | |
835 | ||
836 | /* | |
837 | * If 'this_ctx' is set, we know we have entries to complete | |
838 | * on 'ctx_list'. Do those. | |
839 | */ | |
840 | if (this_ctx) { | |
841 | blk_mq_insert_requests(this_q, this_ctx, &ctx_list, depth, | |
842 | from_schedule); | |
843 | } | |
844 | } | |
845 | ||
846 | static void blk_mq_bio_to_request(struct request *rq, struct bio *bio) | |
847 | { | |
848 | init_request_from_bio(rq, bio); | |
849 | blk_account_io_start(rq, 1); | |
850 | } | |
851 | ||
852 | static void blk_mq_make_request(struct request_queue *q, struct bio *bio) | |
853 | { | |
854 | struct blk_mq_hw_ctx *hctx; | |
855 | struct blk_mq_ctx *ctx; | |
856 | const int is_sync = rw_is_sync(bio->bi_rw); | |
857 | const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA); | |
858 | int rw = bio_data_dir(bio); | |
859 | struct request *rq; | |
860 | unsigned int use_plug, request_count = 0; | |
861 | ||
862 | /* | |
863 | * If we have multiple hardware queues, just go directly to | |
864 | * one of those for sync IO. | |
865 | */ | |
866 | use_plug = !is_flush_fua && ((q->nr_hw_queues == 1) || !is_sync); | |
867 | ||
868 | blk_queue_bounce(q, &bio); | |
869 | ||
14ec77f3 NB |
870 | if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) { |
871 | bio_endio(bio, -EIO); | |
872 | return; | |
873 | } | |
874 | ||
320ae51f JA |
875 | if (use_plug && blk_attempt_plug_merge(q, bio, &request_count)) |
876 | return; | |
877 | ||
878 | if (blk_mq_queue_enter(q)) { | |
879 | bio_endio(bio, -EIO); | |
880 | return; | |
881 | } | |
882 | ||
883 | ctx = blk_mq_get_ctx(q); | |
884 | hctx = q->mq_ops->map_queue(q, ctx->cpu); | |
885 | ||
27fbf4e8 SL |
886 | if (is_sync) |
887 | rw |= REQ_SYNC; | |
320ae51f | 888 | trace_block_getrq(q, bio, rw); |
18741986 | 889 | rq = __blk_mq_alloc_request(hctx, GFP_ATOMIC, false); |
320ae51f | 890 | if (likely(rq)) |
18741986 | 891 | blk_mq_rq_ctx_init(q, ctx, rq, rw); |
320ae51f JA |
892 | else { |
893 | blk_mq_put_ctx(ctx); | |
894 | trace_block_sleeprq(q, bio, rw); | |
18741986 CH |
895 | rq = blk_mq_alloc_request_pinned(q, rw, __GFP_WAIT|GFP_ATOMIC, |
896 | false); | |
320ae51f JA |
897 | ctx = rq->mq_ctx; |
898 | hctx = q->mq_ops->map_queue(q, ctx->cpu); | |
899 | } | |
900 | ||
901 | hctx->queued++; | |
902 | ||
903 | if (unlikely(is_flush_fua)) { | |
904 | blk_mq_bio_to_request(rq, bio); | |
320ae51f JA |
905 | blk_insert_flush(rq); |
906 | goto run_queue; | |
907 | } | |
908 | ||
909 | /* | |
910 | * A task plug currently exists. Since this is completely lockless, | |
911 | * utilize that to temporarily store requests until the task is | |
912 | * either done or scheduled away. | |
913 | */ | |
914 | if (use_plug) { | |
915 | struct blk_plug *plug = current->plug; | |
916 | ||
917 | if (plug) { | |
918 | blk_mq_bio_to_request(rq, bio); | |
92f399c7 | 919 | if (list_empty(&plug->mq_list)) |
320ae51f JA |
920 | trace_block_plug(q); |
921 | else if (request_count >= BLK_MAX_REQUEST_COUNT) { | |
922 | blk_flush_plug_list(plug, false); | |
923 | trace_block_plug(q); | |
924 | } | |
925 | list_add_tail(&rq->queuelist, &plug->mq_list); | |
926 | blk_mq_put_ctx(ctx); | |
927 | return; | |
928 | } | |
929 | } | |
930 | ||
931 | spin_lock(&ctx->lock); | |
932 | ||
933 | if ((hctx->flags & BLK_MQ_F_SHOULD_MERGE) && | |
934 | blk_mq_attempt_merge(q, ctx, bio)) | |
935 | __blk_mq_free_request(hctx, ctx, rq); | |
936 | else { | |
937 | blk_mq_bio_to_request(rq, bio); | |
72a0a36e | 938 | __blk_mq_insert_request(hctx, rq, false); |
320ae51f JA |
939 | } |
940 | ||
941 | spin_unlock(&ctx->lock); | |
320ae51f JA |
942 | |
943 | /* | |
944 | * For a SYNC request, send it to the hardware immediately. For an | |
945 | * ASYNC request, just ensure that we run it later on. The latter | |
946 | * allows for merging opportunities and more efficient dispatching. | |
947 | */ | |
948 | run_queue: | |
949 | blk_mq_run_hw_queue(hctx, !is_sync || is_flush_fua); | |
e4043dcf | 950 | blk_mq_put_ctx(ctx); |
320ae51f JA |
951 | } |
952 | ||
953 | /* | |
954 | * Default mapping to a software queue, since we use one per CPU. | |
955 | */ | |
956 | struct blk_mq_hw_ctx *blk_mq_map_queue(struct request_queue *q, const int cpu) | |
957 | { | |
958 | return q->queue_hw_ctx[q->mq_map[cpu]]; | |
959 | } | |
960 | EXPORT_SYMBOL(blk_mq_map_queue); | |
961 | ||
962 | struct blk_mq_hw_ctx *blk_mq_alloc_single_hw_queue(struct blk_mq_reg *reg, | |
963 | unsigned int hctx_index) | |
964 | { | |
965 | return kmalloc_node(sizeof(struct blk_mq_hw_ctx), | |
966 | GFP_KERNEL | __GFP_ZERO, reg->numa_node); | |
967 | } | |
968 | EXPORT_SYMBOL(blk_mq_alloc_single_hw_queue); | |
969 | ||
970 | void blk_mq_free_single_hw_queue(struct blk_mq_hw_ctx *hctx, | |
971 | unsigned int hctx_index) | |
972 | { | |
973 | kfree(hctx); | |
974 | } | |
975 | EXPORT_SYMBOL(blk_mq_free_single_hw_queue); | |
976 | ||
977 | static void blk_mq_hctx_notify(void *data, unsigned long action, | |
978 | unsigned int cpu) | |
979 | { | |
980 | struct blk_mq_hw_ctx *hctx = data; | |
bccb5f7c | 981 | struct request_queue *q = hctx->queue; |
320ae51f JA |
982 | struct blk_mq_ctx *ctx; |
983 | LIST_HEAD(tmp); | |
984 | ||
985 | if (action != CPU_DEAD && action != CPU_DEAD_FROZEN) | |
986 | return; | |
987 | ||
988 | /* | |
989 | * Move ctx entries to new CPU, if this one is going away. | |
990 | */ | |
bccb5f7c | 991 | ctx = __blk_mq_get_ctx(q, cpu); |
320ae51f JA |
992 | |
993 | spin_lock(&ctx->lock); | |
994 | if (!list_empty(&ctx->rq_list)) { | |
995 | list_splice_init(&ctx->rq_list, &tmp); | |
996 | clear_bit(ctx->index_hw, hctx->ctx_map); | |
997 | } | |
998 | spin_unlock(&ctx->lock); | |
999 | ||
1000 | if (list_empty(&tmp)) | |
1001 | return; | |
1002 | ||
bccb5f7c | 1003 | ctx = blk_mq_get_ctx(q); |
320ae51f JA |
1004 | spin_lock(&ctx->lock); |
1005 | ||
1006 | while (!list_empty(&tmp)) { | |
1007 | struct request *rq; | |
1008 | ||
1009 | rq = list_first_entry(&tmp, struct request, queuelist); | |
1010 | rq->mq_ctx = ctx; | |
1011 | list_move_tail(&rq->queuelist, &ctx->rq_list); | |
1012 | } | |
1013 | ||
bccb5f7c | 1014 | hctx = q->mq_ops->map_queue(q, ctx->cpu); |
320ae51f JA |
1015 | blk_mq_hctx_mark_pending(hctx, ctx); |
1016 | ||
1017 | spin_unlock(&ctx->lock); | |
bccb5f7c JA |
1018 | |
1019 | blk_mq_run_hw_queue(hctx, true); | |
e4043dcf | 1020 | blk_mq_put_ctx(ctx); |
320ae51f JA |
1021 | } |
1022 | ||
95363efd JA |
1023 | static int blk_mq_init_hw_commands(struct blk_mq_hw_ctx *hctx, |
1024 | int (*init)(void *, struct blk_mq_hw_ctx *, | |
1025 | struct request *, unsigned int), | |
1026 | void *data) | |
1027 | { | |
1028 | unsigned int i; | |
1029 | int ret = 0; | |
1030 | ||
1031 | for (i = 0; i < hctx->queue_depth; i++) { | |
1032 | struct request *rq = hctx->rqs[i]; | |
1033 | ||
1034 | ret = init(data, hctx, rq, i); | |
1035 | if (ret) | |
1036 | break; | |
1037 | } | |
1038 | ||
1039 | return ret; | |
1040 | } | |
1041 | ||
1042 | int blk_mq_init_commands(struct request_queue *q, | |
1043 | int (*init)(void *, struct blk_mq_hw_ctx *, | |
1044 | struct request *, unsigned int), | |
1045 | void *data) | |
1046 | { | |
1047 | struct blk_mq_hw_ctx *hctx; | |
1048 | unsigned int i; | |
1049 | int ret = 0; | |
1050 | ||
1051 | queue_for_each_hw_ctx(q, hctx, i) { | |
1052 | ret = blk_mq_init_hw_commands(hctx, init, data); | |
1053 | if (ret) | |
1054 | break; | |
1055 | } | |
1056 | ||
1057 | return ret; | |
1058 | } | |
1059 | EXPORT_SYMBOL(blk_mq_init_commands); | |
1060 | ||
1061 | static void blk_mq_free_hw_commands(struct blk_mq_hw_ctx *hctx, | |
1062 | void (*free)(void *, struct blk_mq_hw_ctx *, | |
320ae51f JA |
1063 | struct request *, unsigned int), |
1064 | void *data) | |
1065 | { | |
1066 | unsigned int i; | |
1067 | ||
1068 | for (i = 0; i < hctx->queue_depth; i++) { | |
1069 | struct request *rq = hctx->rqs[i]; | |
1070 | ||
95363efd | 1071 | free(data, hctx, rq, i); |
320ae51f JA |
1072 | } |
1073 | } | |
1074 | ||
95363efd JA |
1075 | void blk_mq_free_commands(struct request_queue *q, |
1076 | void (*free)(void *, struct blk_mq_hw_ctx *, | |
320ae51f JA |
1077 | struct request *, unsigned int), |
1078 | void *data) | |
1079 | { | |
1080 | struct blk_mq_hw_ctx *hctx; | |
1081 | unsigned int i; | |
1082 | ||
1083 | queue_for_each_hw_ctx(q, hctx, i) | |
95363efd | 1084 | blk_mq_free_hw_commands(hctx, free, data); |
320ae51f | 1085 | } |
95363efd | 1086 | EXPORT_SYMBOL(blk_mq_free_commands); |
320ae51f JA |
1087 | |
1088 | static void blk_mq_free_rq_map(struct blk_mq_hw_ctx *hctx) | |
1089 | { | |
1090 | struct page *page; | |
1091 | ||
1092 | while (!list_empty(&hctx->page_list)) { | |
6753471c DH |
1093 | page = list_first_entry(&hctx->page_list, struct page, lru); |
1094 | list_del_init(&page->lru); | |
320ae51f JA |
1095 | __free_pages(page, page->private); |
1096 | } | |
1097 | ||
1098 | kfree(hctx->rqs); | |
1099 | ||
1100 | if (hctx->tags) | |
1101 | blk_mq_free_tags(hctx->tags); | |
1102 | } | |
1103 | ||
1104 | static size_t order_to_size(unsigned int order) | |
1105 | { | |
1106 | size_t ret = PAGE_SIZE; | |
1107 | ||
1108 | while (order--) | |
1109 | ret *= 2; | |
1110 | ||
1111 | return ret; | |
1112 | } | |
1113 | ||
1114 | static int blk_mq_init_rq_map(struct blk_mq_hw_ctx *hctx, | |
1115 | unsigned int reserved_tags, int node) | |
1116 | { | |
1117 | unsigned int i, j, entries_per_page, max_order = 4; | |
1118 | size_t rq_size, left; | |
1119 | ||
1120 | INIT_LIST_HEAD(&hctx->page_list); | |
1121 | ||
1122 | hctx->rqs = kmalloc_node(hctx->queue_depth * sizeof(struct request *), | |
1123 | GFP_KERNEL, node); | |
1124 | if (!hctx->rqs) | |
1125 | return -ENOMEM; | |
1126 | ||
1127 | /* | |
1128 | * rq_size is the size of the request plus driver payload, rounded | |
1129 | * to the cacheline size | |
1130 | */ | |
1131 | rq_size = round_up(sizeof(struct request) + hctx->cmd_size, | |
1132 | cache_line_size()); | |
1133 | left = rq_size * hctx->queue_depth; | |
1134 | ||
1135 | for (i = 0; i < hctx->queue_depth;) { | |
1136 | int this_order = max_order; | |
1137 | struct page *page; | |
1138 | int to_do; | |
1139 | void *p; | |
1140 | ||
1141 | while (left < order_to_size(this_order - 1) && this_order) | |
1142 | this_order--; | |
1143 | ||
1144 | do { | |
1145 | page = alloc_pages_node(node, GFP_KERNEL, this_order); | |
1146 | if (page) | |
1147 | break; | |
1148 | if (!this_order--) | |
1149 | break; | |
1150 | if (order_to_size(this_order) < rq_size) | |
1151 | break; | |
1152 | } while (1); | |
1153 | ||
1154 | if (!page) | |
1155 | break; | |
1156 | ||
1157 | page->private = this_order; | |
6753471c | 1158 | list_add_tail(&page->lru, &hctx->page_list); |
320ae51f JA |
1159 | |
1160 | p = page_address(page); | |
1161 | entries_per_page = order_to_size(this_order) / rq_size; | |
1162 | to_do = min(entries_per_page, hctx->queue_depth - i); | |
1163 | left -= to_do * rq_size; | |
1164 | for (j = 0; j < to_do; j++) { | |
1165 | hctx->rqs[i] = p; | |
1166 | blk_mq_rq_init(hctx, hctx->rqs[i]); | |
1167 | p += rq_size; | |
1168 | i++; | |
1169 | } | |
1170 | } | |
1171 | ||
1172 | if (i < (reserved_tags + BLK_MQ_TAG_MIN)) | |
1173 | goto err_rq_map; | |
1174 | else if (i != hctx->queue_depth) { | |
1175 | hctx->queue_depth = i; | |
1176 | pr_warn("%s: queue depth set to %u because of low memory\n", | |
1177 | __func__, i); | |
1178 | } | |
1179 | ||
1180 | hctx->tags = blk_mq_init_tags(hctx->queue_depth, reserved_tags, node); | |
1181 | if (!hctx->tags) { | |
1182 | err_rq_map: | |
1183 | blk_mq_free_rq_map(hctx); | |
1184 | return -ENOMEM; | |
1185 | } | |
1186 | ||
1187 | return 0; | |
1188 | } | |
1189 | ||
1190 | static int blk_mq_init_hw_queues(struct request_queue *q, | |
1191 | struct blk_mq_reg *reg, void *driver_data) | |
1192 | { | |
1193 | struct blk_mq_hw_ctx *hctx; | |
1194 | unsigned int i, j; | |
1195 | ||
1196 | /* | |
1197 | * Initialize hardware queues | |
1198 | */ | |
1199 | queue_for_each_hw_ctx(q, hctx, i) { | |
1200 | unsigned int num_maps; | |
1201 | int node; | |
1202 | ||
1203 | node = hctx->numa_node; | |
1204 | if (node == NUMA_NO_NODE) | |
1205 | node = hctx->numa_node = reg->numa_node; | |
1206 | ||
1207 | INIT_DELAYED_WORK(&hctx->delayed_work, blk_mq_work_fn); | |
1208 | spin_lock_init(&hctx->lock); | |
1209 | INIT_LIST_HEAD(&hctx->dispatch); | |
1210 | hctx->queue = q; | |
1211 | hctx->queue_num = i; | |
1212 | hctx->flags = reg->flags; | |
1213 | hctx->queue_depth = reg->queue_depth; | |
1214 | hctx->cmd_size = reg->cmd_size; | |
1215 | ||
1216 | blk_mq_init_cpu_notifier(&hctx->cpu_notifier, | |
1217 | blk_mq_hctx_notify, hctx); | |
1218 | blk_mq_register_cpu_notifier(&hctx->cpu_notifier); | |
1219 | ||
1220 | if (blk_mq_init_rq_map(hctx, reg->reserved_tags, node)) | |
1221 | break; | |
1222 | ||
1223 | /* | |
1224 | * Allocate space for all possible cpus to avoid allocation in | |
1225 | * runtime | |
1226 | */ | |
1227 | hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *), | |
1228 | GFP_KERNEL, node); | |
1229 | if (!hctx->ctxs) | |
1230 | break; | |
1231 | ||
1232 | num_maps = ALIGN(nr_cpu_ids, BITS_PER_LONG) / BITS_PER_LONG; | |
1233 | hctx->ctx_map = kzalloc_node(num_maps * sizeof(unsigned long), | |
1234 | GFP_KERNEL, node); | |
1235 | if (!hctx->ctx_map) | |
1236 | break; | |
1237 | ||
1238 | hctx->nr_ctx_map = num_maps; | |
1239 | hctx->nr_ctx = 0; | |
1240 | ||
1241 | if (reg->ops->init_hctx && | |
1242 | reg->ops->init_hctx(hctx, driver_data, i)) | |
1243 | break; | |
1244 | } | |
1245 | ||
1246 | if (i == q->nr_hw_queues) | |
1247 | return 0; | |
1248 | ||
1249 | /* | |
1250 | * Init failed | |
1251 | */ | |
1252 | queue_for_each_hw_ctx(q, hctx, j) { | |
1253 | if (i == j) | |
1254 | break; | |
1255 | ||
1256 | if (reg->ops->exit_hctx) | |
1257 | reg->ops->exit_hctx(hctx, j); | |
1258 | ||
1259 | blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier); | |
1260 | blk_mq_free_rq_map(hctx); | |
1261 | kfree(hctx->ctxs); | |
1262 | } | |
1263 | ||
1264 | return 1; | |
1265 | } | |
1266 | ||
1267 | static void blk_mq_init_cpu_queues(struct request_queue *q, | |
1268 | unsigned int nr_hw_queues) | |
1269 | { | |
1270 | unsigned int i; | |
1271 | ||
1272 | for_each_possible_cpu(i) { | |
1273 | struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i); | |
1274 | struct blk_mq_hw_ctx *hctx; | |
1275 | ||
1276 | memset(__ctx, 0, sizeof(*__ctx)); | |
1277 | __ctx->cpu = i; | |
1278 | spin_lock_init(&__ctx->lock); | |
1279 | INIT_LIST_HEAD(&__ctx->rq_list); | |
1280 | __ctx->queue = q; | |
1281 | ||
1282 | /* If the cpu isn't online, the cpu is mapped to first hctx */ | |
320ae51f JA |
1283 | if (!cpu_online(i)) |
1284 | continue; | |
1285 | ||
e4043dcf JA |
1286 | hctx = q->mq_ops->map_queue(q, i); |
1287 | cpumask_set_cpu(i, hctx->cpumask); | |
1288 | hctx->nr_ctx++; | |
1289 | ||
320ae51f JA |
1290 | /* |
1291 | * Set local node, IFF we have more than one hw queue. If | |
1292 | * not, we remain on the home node of the device | |
1293 | */ | |
1294 | if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE) | |
1295 | hctx->numa_node = cpu_to_node(i); | |
1296 | } | |
1297 | } | |
1298 | ||
1299 | static void blk_mq_map_swqueue(struct request_queue *q) | |
1300 | { | |
1301 | unsigned int i; | |
1302 | struct blk_mq_hw_ctx *hctx; | |
1303 | struct blk_mq_ctx *ctx; | |
1304 | ||
1305 | queue_for_each_hw_ctx(q, hctx, i) { | |
e4043dcf | 1306 | cpumask_clear(hctx->cpumask); |
320ae51f JA |
1307 | hctx->nr_ctx = 0; |
1308 | } | |
1309 | ||
1310 | /* | |
1311 | * Map software to hardware queues | |
1312 | */ | |
1313 | queue_for_each_ctx(q, ctx, i) { | |
1314 | /* If the cpu isn't online, the cpu is mapped to first hctx */ | |
e4043dcf JA |
1315 | if (!cpu_online(i)) |
1316 | continue; | |
1317 | ||
320ae51f | 1318 | hctx = q->mq_ops->map_queue(q, i); |
e4043dcf | 1319 | cpumask_set_cpu(i, hctx->cpumask); |
320ae51f JA |
1320 | ctx->index_hw = hctx->nr_ctx; |
1321 | hctx->ctxs[hctx->nr_ctx++] = ctx; | |
1322 | } | |
1323 | } | |
1324 | ||
1325 | struct request_queue *blk_mq_init_queue(struct blk_mq_reg *reg, | |
1326 | void *driver_data) | |
1327 | { | |
1328 | struct blk_mq_hw_ctx **hctxs; | |
1329 | struct blk_mq_ctx *ctx; | |
1330 | struct request_queue *q; | |
1331 | int i; | |
1332 | ||
1333 | if (!reg->nr_hw_queues || | |
1334 | !reg->ops->queue_rq || !reg->ops->map_queue || | |
1335 | !reg->ops->alloc_hctx || !reg->ops->free_hctx) | |
1336 | return ERR_PTR(-EINVAL); | |
1337 | ||
1338 | if (!reg->queue_depth) | |
1339 | reg->queue_depth = BLK_MQ_MAX_DEPTH; | |
1340 | else if (reg->queue_depth > BLK_MQ_MAX_DEPTH) { | |
1341 | pr_err("blk-mq: queuedepth too large (%u)\n", reg->queue_depth); | |
1342 | reg->queue_depth = BLK_MQ_MAX_DEPTH; | |
1343 | } | |
1344 | ||
1345 | if (reg->queue_depth < (reg->reserved_tags + BLK_MQ_TAG_MIN)) | |
1346 | return ERR_PTR(-EINVAL); | |
1347 | ||
1348 | ctx = alloc_percpu(struct blk_mq_ctx); | |
1349 | if (!ctx) | |
1350 | return ERR_PTR(-ENOMEM); | |
1351 | ||
1352 | hctxs = kmalloc_node(reg->nr_hw_queues * sizeof(*hctxs), GFP_KERNEL, | |
1353 | reg->numa_node); | |
1354 | ||
1355 | if (!hctxs) | |
1356 | goto err_percpu; | |
1357 | ||
1358 | for (i = 0; i < reg->nr_hw_queues; i++) { | |
1359 | hctxs[i] = reg->ops->alloc_hctx(reg, i); | |
1360 | if (!hctxs[i]) | |
1361 | goto err_hctxs; | |
1362 | ||
e4043dcf JA |
1363 | if (!zalloc_cpumask_var(&hctxs[i]->cpumask, GFP_KERNEL)) |
1364 | goto err_hctxs; | |
1365 | ||
320ae51f JA |
1366 | hctxs[i]->numa_node = NUMA_NO_NODE; |
1367 | hctxs[i]->queue_num = i; | |
1368 | } | |
1369 | ||
1370 | q = blk_alloc_queue_node(GFP_KERNEL, reg->numa_node); | |
1371 | if (!q) | |
1372 | goto err_hctxs; | |
1373 | ||
1374 | q->mq_map = blk_mq_make_queue_map(reg); | |
1375 | if (!q->mq_map) | |
1376 | goto err_map; | |
1377 | ||
1378 | setup_timer(&q->timeout, blk_mq_rq_timer, (unsigned long) q); | |
1379 | blk_queue_rq_timeout(q, 30000); | |
1380 | ||
1381 | q->nr_queues = nr_cpu_ids; | |
1382 | q->nr_hw_queues = reg->nr_hw_queues; | |
1383 | ||
1384 | q->queue_ctx = ctx; | |
1385 | q->queue_hw_ctx = hctxs; | |
1386 | ||
1387 | q->mq_ops = reg->ops; | |
94eddfbe | 1388 | q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT; |
320ae51f | 1389 | |
1be036e9 CH |
1390 | q->sg_reserved_size = INT_MAX; |
1391 | ||
320ae51f JA |
1392 | blk_queue_make_request(q, blk_mq_make_request); |
1393 | blk_queue_rq_timed_out(q, reg->ops->timeout); | |
1394 | if (reg->timeout) | |
1395 | blk_queue_rq_timeout(q, reg->timeout); | |
1396 | ||
30a91cb4 CH |
1397 | if (reg->ops->complete) |
1398 | blk_queue_softirq_done(q, reg->ops->complete); | |
1399 | ||
320ae51f JA |
1400 | blk_mq_init_flush(q); |
1401 | blk_mq_init_cpu_queues(q, reg->nr_hw_queues); | |
1402 | ||
18741986 CH |
1403 | q->flush_rq = kzalloc(round_up(sizeof(struct request) + reg->cmd_size, |
1404 | cache_line_size()), GFP_KERNEL); | |
1405 | if (!q->flush_rq) | |
320ae51f JA |
1406 | goto err_hw; |
1407 | ||
18741986 CH |
1408 | if (blk_mq_init_hw_queues(q, reg, driver_data)) |
1409 | goto err_flush_rq; | |
1410 | ||
320ae51f JA |
1411 | blk_mq_map_swqueue(q); |
1412 | ||
1413 | mutex_lock(&all_q_mutex); | |
1414 | list_add_tail(&q->all_q_node, &all_q_list); | |
1415 | mutex_unlock(&all_q_mutex); | |
1416 | ||
1417 | return q; | |
18741986 CH |
1418 | |
1419 | err_flush_rq: | |
1420 | kfree(q->flush_rq); | |
320ae51f JA |
1421 | err_hw: |
1422 | kfree(q->mq_map); | |
1423 | err_map: | |
1424 | blk_cleanup_queue(q); | |
1425 | err_hctxs: | |
1426 | for (i = 0; i < reg->nr_hw_queues; i++) { | |
1427 | if (!hctxs[i]) | |
1428 | break; | |
e4043dcf | 1429 | free_cpumask_var(hctxs[i]->cpumask); |
320ae51f JA |
1430 | reg->ops->free_hctx(hctxs[i], i); |
1431 | } | |
1432 | kfree(hctxs); | |
1433 | err_percpu: | |
1434 | free_percpu(ctx); | |
1435 | return ERR_PTR(-ENOMEM); | |
1436 | } | |
1437 | EXPORT_SYMBOL(blk_mq_init_queue); | |
1438 | ||
1439 | void blk_mq_free_queue(struct request_queue *q) | |
1440 | { | |
1441 | struct blk_mq_hw_ctx *hctx; | |
1442 | int i; | |
1443 | ||
1444 | queue_for_each_hw_ctx(q, hctx, i) { | |
320ae51f JA |
1445 | kfree(hctx->ctx_map); |
1446 | kfree(hctx->ctxs); | |
1447 | blk_mq_free_rq_map(hctx); | |
1448 | blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier); | |
1449 | if (q->mq_ops->exit_hctx) | |
1450 | q->mq_ops->exit_hctx(hctx, i); | |
e4043dcf | 1451 | free_cpumask_var(hctx->cpumask); |
320ae51f JA |
1452 | q->mq_ops->free_hctx(hctx, i); |
1453 | } | |
1454 | ||
1455 | free_percpu(q->queue_ctx); | |
1456 | kfree(q->queue_hw_ctx); | |
1457 | kfree(q->mq_map); | |
1458 | ||
1459 | q->queue_ctx = NULL; | |
1460 | q->queue_hw_ctx = NULL; | |
1461 | q->mq_map = NULL; | |
1462 | ||
1463 | mutex_lock(&all_q_mutex); | |
1464 | list_del_init(&q->all_q_node); | |
1465 | mutex_unlock(&all_q_mutex); | |
1466 | } | |
320ae51f JA |
1467 | |
1468 | /* Basically redo blk_mq_init_queue with queue frozen */ | |
f618ef7c | 1469 | static void blk_mq_queue_reinit(struct request_queue *q) |
320ae51f JA |
1470 | { |
1471 | blk_mq_freeze_queue(q); | |
1472 | ||
1473 | blk_mq_update_queue_map(q->mq_map, q->nr_hw_queues); | |
1474 | ||
1475 | /* | |
1476 | * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe | |
1477 | * we should change hctx numa_node according to new topology (this | |
1478 | * involves free and re-allocate memory, worthy doing?) | |
1479 | */ | |
1480 | ||
1481 | blk_mq_map_swqueue(q); | |
1482 | ||
1483 | blk_mq_unfreeze_queue(q); | |
1484 | } | |
1485 | ||
f618ef7c PG |
1486 | static int blk_mq_queue_reinit_notify(struct notifier_block *nb, |
1487 | unsigned long action, void *hcpu) | |
320ae51f JA |
1488 | { |
1489 | struct request_queue *q; | |
1490 | ||
1491 | /* | |
1492 | * Before new mapping is established, hotadded cpu might already start | |
1493 | * handling requests. This doesn't break anything as we map offline | |
1494 | * CPUs to first hardware queue. We will re-init queue below to get | |
1495 | * optimal settings. | |
1496 | */ | |
1497 | if (action != CPU_DEAD && action != CPU_DEAD_FROZEN && | |
1498 | action != CPU_ONLINE && action != CPU_ONLINE_FROZEN) | |
1499 | return NOTIFY_OK; | |
1500 | ||
1501 | mutex_lock(&all_q_mutex); | |
1502 | list_for_each_entry(q, &all_q_list, all_q_node) | |
1503 | blk_mq_queue_reinit(q); | |
1504 | mutex_unlock(&all_q_mutex); | |
1505 | return NOTIFY_OK; | |
1506 | } | |
1507 | ||
676141e4 JA |
1508 | void blk_mq_disable_hotplug(void) |
1509 | { | |
1510 | mutex_lock(&all_q_mutex); | |
1511 | } | |
1512 | ||
1513 | void blk_mq_enable_hotplug(void) | |
1514 | { | |
1515 | mutex_unlock(&all_q_mutex); | |
1516 | } | |
1517 | ||
320ae51f JA |
1518 | static int __init blk_mq_init(void) |
1519 | { | |
320ae51f JA |
1520 | blk_mq_cpu_init(); |
1521 | ||
1522 | /* Must be called after percpu_counter_hotcpu_callback() */ | |
1523 | hotcpu_notifier(blk_mq_queue_reinit_notify, -10); | |
1524 | ||
1525 | return 0; | |
1526 | } | |
1527 | subsys_initcall(blk_mq_init); |