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 | ||
320ae51f JA |
254 | static void __blk_mq_free_request(struct blk_mq_hw_ctx *hctx, |
255 | struct blk_mq_ctx *ctx, struct request *rq) | |
256 | { | |
257 | const int tag = rq->tag; | |
258 | struct request_queue *q = rq->q; | |
259 | ||
9d74e257 | 260 | blk_rq_init(hctx->queue, rq); |
320ae51f JA |
261 | blk_mq_put_tag(hctx->tags, tag); |
262 | ||
263 | blk_mq_queue_exit(q); | |
264 | } | |
265 | ||
266 | void blk_mq_free_request(struct request *rq) | |
267 | { | |
268 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
269 | struct blk_mq_hw_ctx *hctx; | |
270 | struct request_queue *q = rq->q; | |
271 | ||
272 | ctx->rq_completed[rq_is_sync(rq)]++; | |
273 | ||
274 | hctx = q->mq_ops->map_queue(q, ctx->cpu); | |
275 | __blk_mq_free_request(hctx, ctx, rq); | |
276 | } | |
277 | ||
8727af4b CH |
278 | /* |
279 | * Clone all relevant state from a request that has been put on hold in | |
280 | * the flush state machine into the preallocated flush request that hangs | |
281 | * off the request queue. | |
282 | * | |
283 | * For a driver the flush request should be invisible, that's why we are | |
284 | * impersonating the original request here. | |
285 | */ | |
286 | void blk_mq_clone_flush_request(struct request *flush_rq, | |
287 | struct request *orig_rq) | |
288 | { | |
289 | struct blk_mq_hw_ctx *hctx = | |
290 | orig_rq->q->mq_ops->map_queue(orig_rq->q, orig_rq->mq_ctx->cpu); | |
291 | ||
292 | flush_rq->mq_ctx = orig_rq->mq_ctx; | |
293 | flush_rq->tag = orig_rq->tag; | |
294 | memcpy(blk_mq_rq_to_pdu(flush_rq), blk_mq_rq_to_pdu(orig_rq), | |
295 | hctx->cmd_size); | |
296 | } | |
297 | ||
7237c740 | 298 | bool blk_mq_end_io_partial(struct request *rq, int error, unsigned int nr_bytes) |
320ae51f | 299 | { |
7237c740 CH |
300 | if (blk_update_request(rq, error, blk_rq_bytes(rq))) |
301 | return true; | |
320ae51f | 302 | |
0d11e6ac ML |
303 | blk_account_io_done(rq); |
304 | ||
320ae51f JA |
305 | if (rq->end_io) |
306 | rq->end_io(rq, error); | |
307 | else | |
308 | blk_mq_free_request(rq); | |
7237c740 | 309 | return false; |
320ae51f | 310 | } |
7237c740 | 311 | EXPORT_SYMBOL(blk_mq_end_io_partial); |
320ae51f | 312 | |
30a91cb4 | 313 | static void __blk_mq_complete_request_remote(void *data) |
320ae51f | 314 | { |
3d6efbf6 | 315 | struct request *rq = data; |
320ae51f | 316 | |
30a91cb4 | 317 | rq->q->softirq_done_fn(rq); |
320ae51f | 318 | } |
320ae51f | 319 | |
30a91cb4 | 320 | void __blk_mq_complete_request(struct request *rq) |
320ae51f JA |
321 | { |
322 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
323 | int cpu; | |
324 | ||
30a91cb4 CH |
325 | if (!ctx->ipi_redirect) { |
326 | rq->q->softirq_done_fn(rq); | |
327 | return; | |
328 | } | |
320ae51f JA |
329 | |
330 | cpu = get_cpu(); | |
3d6efbf6 | 331 | if (cpu != ctx->cpu && cpu_online(ctx->cpu)) { |
30a91cb4 | 332 | rq->csd.func = __blk_mq_complete_request_remote; |
3d6efbf6 CH |
333 | rq->csd.info = rq; |
334 | rq->csd.flags = 0; | |
c46fff2a | 335 | smp_call_function_single_async(ctx->cpu, &rq->csd); |
3d6efbf6 | 336 | } else { |
30a91cb4 | 337 | rq->q->softirq_done_fn(rq); |
3d6efbf6 | 338 | } |
320ae51f JA |
339 | put_cpu(); |
340 | } | |
30a91cb4 CH |
341 | |
342 | /** | |
343 | * blk_mq_complete_request - end I/O on a request | |
344 | * @rq: the request being processed | |
345 | * | |
346 | * Description: | |
347 | * Ends all I/O on a request. It does not handle partial completions. | |
348 | * The actual completion happens out-of-order, through a IPI handler. | |
349 | **/ | |
350 | void blk_mq_complete_request(struct request *rq) | |
351 | { | |
352 | if (unlikely(blk_should_fake_timeout(rq->q))) | |
353 | return; | |
354 | if (!blk_mark_rq_complete(rq)) | |
355 | __blk_mq_complete_request(rq); | |
356 | } | |
357 | EXPORT_SYMBOL(blk_mq_complete_request); | |
320ae51f | 358 | |
49f5baa5 | 359 | static void blk_mq_start_request(struct request *rq, bool last) |
320ae51f JA |
360 | { |
361 | struct request_queue *q = rq->q; | |
362 | ||
363 | trace_block_rq_issue(q, rq); | |
364 | ||
742ee69b CH |
365 | rq->resid_len = blk_rq_bytes(rq); |
366 | ||
320ae51f JA |
367 | /* |
368 | * Just mark start time and set the started bit. Due to memory | |
369 | * ordering, we know we'll see the correct deadline as long as | |
370 | * REQ_ATOMIC_STARTED is seen. | |
371 | */ | |
372 | rq->deadline = jiffies + q->rq_timeout; | |
373 | set_bit(REQ_ATOM_STARTED, &rq->atomic_flags); | |
49f5baa5 CH |
374 | |
375 | if (q->dma_drain_size && blk_rq_bytes(rq)) { | |
376 | /* | |
377 | * Make sure space for the drain appears. We know we can do | |
378 | * this because max_hw_segments has been adjusted to be one | |
379 | * fewer than the device can handle. | |
380 | */ | |
381 | rq->nr_phys_segments++; | |
382 | } | |
383 | ||
384 | /* | |
385 | * Flag the last request in the series so that drivers know when IO | |
386 | * should be kicked off, if they don't do it on a per-request basis. | |
387 | * | |
388 | * Note: the flag isn't the only condition drivers should do kick off. | |
389 | * If drive is busy, the last request might not have the bit set. | |
390 | */ | |
391 | if (last) | |
392 | rq->cmd_flags |= REQ_END; | |
320ae51f JA |
393 | } |
394 | ||
395 | static void blk_mq_requeue_request(struct request *rq) | |
396 | { | |
397 | struct request_queue *q = rq->q; | |
398 | ||
399 | trace_block_rq_requeue(q, rq); | |
400 | clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags); | |
49f5baa5 CH |
401 | |
402 | rq->cmd_flags &= ~REQ_END; | |
403 | ||
404 | if (q->dma_drain_size && blk_rq_bytes(rq)) | |
405 | rq->nr_phys_segments--; | |
320ae51f JA |
406 | } |
407 | ||
408 | struct blk_mq_timeout_data { | |
409 | struct blk_mq_hw_ctx *hctx; | |
410 | unsigned long *next; | |
411 | unsigned int *next_set; | |
412 | }; | |
413 | ||
414 | static void blk_mq_timeout_check(void *__data, unsigned long *free_tags) | |
415 | { | |
416 | struct blk_mq_timeout_data *data = __data; | |
417 | struct blk_mq_hw_ctx *hctx = data->hctx; | |
418 | unsigned int tag; | |
419 | ||
420 | /* It may not be in flight yet (this is where | |
421 | * the REQ_ATOMIC_STARTED flag comes in). The requests are | |
422 | * statically allocated, so we know it's always safe to access the | |
423 | * memory associated with a bit offset into ->rqs[]. | |
424 | */ | |
425 | tag = 0; | |
426 | do { | |
427 | struct request *rq; | |
428 | ||
429 | tag = find_next_zero_bit(free_tags, hctx->queue_depth, tag); | |
430 | if (tag >= hctx->queue_depth) | |
431 | break; | |
432 | ||
433 | rq = hctx->rqs[tag++]; | |
434 | ||
435 | if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) | |
436 | continue; | |
437 | ||
438 | blk_rq_check_expired(rq, data->next, data->next_set); | |
439 | } while (1); | |
440 | } | |
441 | ||
442 | static void blk_mq_hw_ctx_check_timeout(struct blk_mq_hw_ctx *hctx, | |
443 | unsigned long *next, | |
444 | unsigned int *next_set) | |
445 | { | |
446 | struct blk_mq_timeout_data data = { | |
447 | .hctx = hctx, | |
448 | .next = next, | |
449 | .next_set = next_set, | |
450 | }; | |
451 | ||
452 | /* | |
453 | * Ask the tagging code to iterate busy requests, so we can | |
454 | * check them for timeout. | |
455 | */ | |
456 | blk_mq_tag_busy_iter(hctx->tags, blk_mq_timeout_check, &data); | |
457 | } | |
458 | ||
459 | static void blk_mq_rq_timer(unsigned long data) | |
460 | { | |
461 | struct request_queue *q = (struct request_queue *) data; | |
462 | struct blk_mq_hw_ctx *hctx; | |
463 | unsigned long next = 0; | |
464 | int i, next_set = 0; | |
465 | ||
466 | queue_for_each_hw_ctx(q, hctx, i) | |
467 | blk_mq_hw_ctx_check_timeout(hctx, &next, &next_set); | |
468 | ||
469 | if (next_set) | |
470 | mod_timer(&q->timeout, round_jiffies_up(next)); | |
471 | } | |
472 | ||
473 | /* | |
474 | * Reverse check our software queue for entries that we could potentially | |
475 | * merge with. Currently includes a hand-wavy stop count of 8, to not spend | |
476 | * too much time checking for merges. | |
477 | */ | |
478 | static bool blk_mq_attempt_merge(struct request_queue *q, | |
479 | struct blk_mq_ctx *ctx, struct bio *bio) | |
480 | { | |
481 | struct request *rq; | |
482 | int checked = 8; | |
483 | ||
484 | list_for_each_entry_reverse(rq, &ctx->rq_list, queuelist) { | |
485 | int el_ret; | |
486 | ||
487 | if (!checked--) | |
488 | break; | |
489 | ||
490 | if (!blk_rq_merge_ok(rq, bio)) | |
491 | continue; | |
492 | ||
493 | el_ret = blk_try_merge(rq, bio); | |
494 | if (el_ret == ELEVATOR_BACK_MERGE) { | |
495 | if (bio_attempt_back_merge(q, rq, bio)) { | |
496 | ctx->rq_merged++; | |
497 | return true; | |
498 | } | |
499 | break; | |
500 | } else if (el_ret == ELEVATOR_FRONT_MERGE) { | |
501 | if (bio_attempt_front_merge(q, rq, bio)) { | |
502 | ctx->rq_merged++; | |
503 | return true; | |
504 | } | |
505 | break; | |
506 | } | |
507 | } | |
508 | ||
509 | return false; | |
510 | } | |
511 | ||
512 | void blk_mq_add_timer(struct request *rq) | |
513 | { | |
514 | __blk_add_timer(rq, NULL); | |
515 | } | |
516 | ||
517 | /* | |
518 | * Run this hardware queue, pulling any software queues mapped to it in. | |
519 | * Note that this function currently has various problems around ordering | |
520 | * of IO. In particular, we'd like FIFO behaviour on handling existing | |
521 | * items on the hctx->dispatch list. Ignore that for now. | |
522 | */ | |
523 | static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx) | |
524 | { | |
525 | struct request_queue *q = hctx->queue; | |
526 | struct blk_mq_ctx *ctx; | |
527 | struct request *rq; | |
528 | LIST_HEAD(rq_list); | |
529 | int bit, queued; | |
530 | ||
e4043dcf JA |
531 | WARN_ON(!preempt_count()); |
532 | ||
5d12f905 | 533 | if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state))) |
320ae51f JA |
534 | return; |
535 | ||
536 | hctx->run++; | |
537 | ||
538 | /* | |
539 | * Touch any software queue that has pending entries. | |
540 | */ | |
541 | for_each_set_bit(bit, hctx->ctx_map, hctx->nr_ctx) { | |
542 | clear_bit(bit, hctx->ctx_map); | |
543 | ctx = hctx->ctxs[bit]; | |
544 | BUG_ON(bit != ctx->index_hw); | |
545 | ||
546 | spin_lock(&ctx->lock); | |
547 | list_splice_tail_init(&ctx->rq_list, &rq_list); | |
548 | spin_unlock(&ctx->lock); | |
549 | } | |
550 | ||
551 | /* | |
552 | * If we have previous entries on our dispatch list, grab them | |
553 | * and stuff them at the front for more fair dispatch. | |
554 | */ | |
555 | if (!list_empty_careful(&hctx->dispatch)) { | |
556 | spin_lock(&hctx->lock); | |
557 | if (!list_empty(&hctx->dispatch)) | |
558 | list_splice_init(&hctx->dispatch, &rq_list); | |
559 | spin_unlock(&hctx->lock); | |
560 | } | |
561 | ||
562 | /* | |
563 | * Delete and return all entries from our dispatch list | |
564 | */ | |
565 | queued = 0; | |
566 | ||
567 | /* | |
568 | * Now process all the entries, sending them to the driver. | |
569 | */ | |
570 | while (!list_empty(&rq_list)) { | |
571 | int ret; | |
572 | ||
573 | rq = list_first_entry(&rq_list, struct request, queuelist); | |
574 | list_del_init(&rq->queuelist); | |
320ae51f | 575 | |
49f5baa5 | 576 | blk_mq_start_request(rq, list_empty(&rq_list)); |
320ae51f JA |
577 | |
578 | ret = q->mq_ops->queue_rq(hctx, rq); | |
579 | switch (ret) { | |
580 | case BLK_MQ_RQ_QUEUE_OK: | |
581 | queued++; | |
582 | continue; | |
583 | case BLK_MQ_RQ_QUEUE_BUSY: | |
584 | /* | |
585 | * FIXME: we should have a mechanism to stop the queue | |
586 | * like blk_stop_queue, otherwise we will waste cpu | |
587 | * time | |
588 | */ | |
589 | list_add(&rq->queuelist, &rq_list); | |
590 | blk_mq_requeue_request(rq); | |
591 | break; | |
592 | default: | |
593 | pr_err("blk-mq: bad return on queue: %d\n", ret); | |
320ae51f | 594 | case BLK_MQ_RQ_QUEUE_ERROR: |
1e93b8c2 | 595 | rq->errors = -EIO; |
320ae51f JA |
596 | blk_mq_end_io(rq, rq->errors); |
597 | break; | |
598 | } | |
599 | ||
600 | if (ret == BLK_MQ_RQ_QUEUE_BUSY) | |
601 | break; | |
602 | } | |
603 | ||
604 | if (!queued) | |
605 | hctx->dispatched[0]++; | |
606 | else if (queued < (1 << (BLK_MQ_MAX_DISPATCH_ORDER - 1))) | |
607 | hctx->dispatched[ilog2(queued) + 1]++; | |
608 | ||
609 | /* | |
610 | * Any items that need requeuing? Stuff them into hctx->dispatch, | |
611 | * that is where we will continue on next queue run. | |
612 | */ | |
613 | if (!list_empty(&rq_list)) { | |
614 | spin_lock(&hctx->lock); | |
615 | list_splice(&rq_list, &hctx->dispatch); | |
616 | spin_unlock(&hctx->lock); | |
617 | } | |
618 | } | |
619 | ||
620 | void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) | |
621 | { | |
5d12f905 | 622 | if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state))) |
320ae51f JA |
623 | return; |
624 | ||
e4043dcf | 625 | if (!async && cpumask_test_cpu(smp_processor_id(), hctx->cpumask)) |
320ae51f | 626 | __blk_mq_run_hw_queue(hctx); |
e4043dcf | 627 | else if (hctx->queue->nr_hw_queues == 1) |
59c3d45e | 628 | kblockd_schedule_delayed_work(&hctx->delayed_work, 0); |
e4043dcf JA |
629 | else { |
630 | unsigned int cpu; | |
631 | ||
632 | /* | |
633 | * It'd be great if the workqueue API had a way to pass | |
634 | * in a mask and had some smarts for more clever placement | |
635 | * than the first CPU. Or we could round-robin here. For now, | |
636 | * just queue on the first CPU. | |
637 | */ | |
638 | cpu = cpumask_first(hctx->cpumask); | |
639 | kblockd_schedule_delayed_work_on(cpu, &hctx->delayed_work, 0); | |
640 | } | |
320ae51f JA |
641 | } |
642 | ||
643 | void blk_mq_run_queues(struct request_queue *q, bool async) | |
644 | { | |
645 | struct blk_mq_hw_ctx *hctx; | |
646 | int i; | |
647 | ||
648 | queue_for_each_hw_ctx(q, hctx, i) { | |
649 | if ((!blk_mq_hctx_has_pending(hctx) && | |
650 | list_empty_careful(&hctx->dispatch)) || | |
5d12f905 | 651 | test_bit(BLK_MQ_S_STOPPED, &hctx->state)) |
320ae51f JA |
652 | continue; |
653 | ||
e4043dcf | 654 | preempt_disable(); |
320ae51f | 655 | blk_mq_run_hw_queue(hctx, async); |
e4043dcf | 656 | preempt_enable(); |
320ae51f JA |
657 | } |
658 | } | |
659 | EXPORT_SYMBOL(blk_mq_run_queues); | |
660 | ||
661 | void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx) | |
662 | { | |
663 | cancel_delayed_work(&hctx->delayed_work); | |
664 | set_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
665 | } | |
666 | EXPORT_SYMBOL(blk_mq_stop_hw_queue); | |
667 | ||
280d45f6 CH |
668 | void blk_mq_stop_hw_queues(struct request_queue *q) |
669 | { | |
670 | struct blk_mq_hw_ctx *hctx; | |
671 | int i; | |
672 | ||
673 | queue_for_each_hw_ctx(q, hctx, i) | |
674 | blk_mq_stop_hw_queue(hctx); | |
675 | } | |
676 | EXPORT_SYMBOL(blk_mq_stop_hw_queues); | |
677 | ||
320ae51f JA |
678 | void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx) |
679 | { | |
680 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
e4043dcf JA |
681 | |
682 | preempt_disable(); | |
320ae51f | 683 | __blk_mq_run_hw_queue(hctx); |
e4043dcf | 684 | preempt_enable(); |
320ae51f JA |
685 | } |
686 | EXPORT_SYMBOL(blk_mq_start_hw_queue); | |
687 | ||
688 | void blk_mq_start_stopped_hw_queues(struct request_queue *q) | |
689 | { | |
690 | struct blk_mq_hw_ctx *hctx; | |
691 | int i; | |
692 | ||
693 | queue_for_each_hw_ctx(q, hctx, i) { | |
694 | if (!test_bit(BLK_MQ_S_STOPPED, &hctx->state)) | |
695 | continue; | |
696 | ||
697 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
e4043dcf | 698 | preempt_disable(); |
320ae51f | 699 | blk_mq_run_hw_queue(hctx, true); |
e4043dcf | 700 | preempt_enable(); |
320ae51f JA |
701 | } |
702 | } | |
703 | EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues); | |
704 | ||
705 | static void blk_mq_work_fn(struct work_struct *work) | |
706 | { | |
707 | struct blk_mq_hw_ctx *hctx; | |
708 | ||
709 | hctx = container_of(work, struct blk_mq_hw_ctx, delayed_work.work); | |
e4043dcf JA |
710 | |
711 | preempt_disable(); | |
320ae51f | 712 | __blk_mq_run_hw_queue(hctx); |
e4043dcf | 713 | preempt_enable(); |
320ae51f JA |
714 | } |
715 | ||
716 | static void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, | |
72a0a36e | 717 | struct request *rq, bool at_head) |
320ae51f JA |
718 | { |
719 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
720 | ||
01b983c9 JA |
721 | trace_block_rq_insert(hctx->queue, rq); |
722 | ||
72a0a36e CH |
723 | if (at_head) |
724 | list_add(&rq->queuelist, &ctx->rq_list); | |
725 | else | |
726 | list_add_tail(&rq->queuelist, &ctx->rq_list); | |
320ae51f JA |
727 | blk_mq_hctx_mark_pending(hctx, ctx); |
728 | ||
729 | /* | |
730 | * We do this early, to ensure we are on the right CPU. | |
731 | */ | |
732 | blk_mq_add_timer(rq); | |
733 | } | |
734 | ||
eeabc850 CH |
735 | void blk_mq_insert_request(struct request *rq, bool at_head, bool run_queue, |
736 | bool async) | |
320ae51f | 737 | { |
eeabc850 | 738 | struct request_queue *q = rq->q; |
320ae51f | 739 | struct blk_mq_hw_ctx *hctx; |
eeabc850 CH |
740 | struct blk_mq_ctx *ctx = rq->mq_ctx, *current_ctx; |
741 | ||
742 | current_ctx = blk_mq_get_ctx(q); | |
743 | if (!cpu_online(ctx->cpu)) | |
744 | rq->mq_ctx = ctx = current_ctx; | |
320ae51f | 745 | |
320ae51f JA |
746 | hctx = q->mq_ops->map_queue(q, ctx->cpu); |
747 | ||
eeabc850 CH |
748 | if (rq->cmd_flags & (REQ_FLUSH | REQ_FUA) && |
749 | !(rq->cmd_flags & (REQ_FLUSH_SEQ))) { | |
320ae51f JA |
750 | blk_insert_flush(rq); |
751 | } else { | |
320ae51f | 752 | spin_lock(&ctx->lock); |
72a0a36e | 753 | __blk_mq_insert_request(hctx, rq, at_head); |
320ae51f | 754 | spin_unlock(&ctx->lock); |
320ae51f JA |
755 | } |
756 | ||
320ae51f JA |
757 | if (run_queue) |
758 | blk_mq_run_hw_queue(hctx, async); | |
e4043dcf JA |
759 | |
760 | blk_mq_put_ctx(current_ctx); | |
320ae51f JA |
761 | } |
762 | ||
763 | static void blk_mq_insert_requests(struct request_queue *q, | |
764 | struct blk_mq_ctx *ctx, | |
765 | struct list_head *list, | |
766 | int depth, | |
767 | bool from_schedule) | |
768 | ||
769 | { | |
770 | struct blk_mq_hw_ctx *hctx; | |
771 | struct blk_mq_ctx *current_ctx; | |
772 | ||
773 | trace_block_unplug(q, depth, !from_schedule); | |
774 | ||
775 | current_ctx = blk_mq_get_ctx(q); | |
776 | ||
777 | if (!cpu_online(ctx->cpu)) | |
778 | ctx = current_ctx; | |
779 | hctx = q->mq_ops->map_queue(q, ctx->cpu); | |
780 | ||
781 | /* | |
782 | * preemption doesn't flush plug list, so it's possible ctx->cpu is | |
783 | * offline now | |
784 | */ | |
785 | spin_lock(&ctx->lock); | |
786 | while (!list_empty(list)) { | |
787 | struct request *rq; | |
788 | ||
789 | rq = list_first_entry(list, struct request, queuelist); | |
790 | list_del_init(&rq->queuelist); | |
791 | rq->mq_ctx = ctx; | |
72a0a36e | 792 | __blk_mq_insert_request(hctx, rq, false); |
320ae51f JA |
793 | } |
794 | spin_unlock(&ctx->lock); | |
795 | ||
320ae51f | 796 | blk_mq_run_hw_queue(hctx, from_schedule); |
e4043dcf | 797 | blk_mq_put_ctx(current_ctx); |
320ae51f JA |
798 | } |
799 | ||
800 | static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b) | |
801 | { | |
802 | struct request *rqa = container_of(a, struct request, queuelist); | |
803 | struct request *rqb = container_of(b, struct request, queuelist); | |
804 | ||
805 | return !(rqa->mq_ctx < rqb->mq_ctx || | |
806 | (rqa->mq_ctx == rqb->mq_ctx && | |
807 | blk_rq_pos(rqa) < blk_rq_pos(rqb))); | |
808 | } | |
809 | ||
810 | void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule) | |
811 | { | |
812 | struct blk_mq_ctx *this_ctx; | |
813 | struct request_queue *this_q; | |
814 | struct request *rq; | |
815 | LIST_HEAD(list); | |
816 | LIST_HEAD(ctx_list); | |
817 | unsigned int depth; | |
818 | ||
819 | list_splice_init(&plug->mq_list, &list); | |
820 | ||
821 | list_sort(NULL, &list, plug_ctx_cmp); | |
822 | ||
823 | this_q = NULL; | |
824 | this_ctx = NULL; | |
825 | depth = 0; | |
826 | ||
827 | while (!list_empty(&list)) { | |
828 | rq = list_entry_rq(list.next); | |
829 | list_del_init(&rq->queuelist); | |
830 | BUG_ON(!rq->q); | |
831 | if (rq->mq_ctx != this_ctx) { | |
832 | if (this_ctx) { | |
833 | blk_mq_insert_requests(this_q, this_ctx, | |
834 | &ctx_list, depth, | |
835 | from_schedule); | |
836 | } | |
837 | ||
838 | this_ctx = rq->mq_ctx; | |
839 | this_q = rq->q; | |
840 | depth = 0; | |
841 | } | |
842 | ||
843 | depth++; | |
844 | list_add_tail(&rq->queuelist, &ctx_list); | |
845 | } | |
846 | ||
847 | /* | |
848 | * If 'this_ctx' is set, we know we have entries to complete | |
849 | * on 'ctx_list'. Do those. | |
850 | */ | |
851 | if (this_ctx) { | |
852 | blk_mq_insert_requests(this_q, this_ctx, &ctx_list, depth, | |
853 | from_schedule); | |
854 | } | |
855 | } | |
856 | ||
857 | static void blk_mq_bio_to_request(struct request *rq, struct bio *bio) | |
858 | { | |
859 | init_request_from_bio(rq, bio); | |
860 | blk_account_io_start(rq, 1); | |
861 | } | |
862 | ||
863 | static void blk_mq_make_request(struct request_queue *q, struct bio *bio) | |
864 | { | |
865 | struct blk_mq_hw_ctx *hctx; | |
866 | struct blk_mq_ctx *ctx; | |
867 | const int is_sync = rw_is_sync(bio->bi_rw); | |
868 | const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA); | |
869 | int rw = bio_data_dir(bio); | |
870 | struct request *rq; | |
871 | unsigned int use_plug, request_count = 0; | |
872 | ||
873 | /* | |
874 | * If we have multiple hardware queues, just go directly to | |
875 | * one of those for sync IO. | |
876 | */ | |
877 | use_plug = !is_flush_fua && ((q->nr_hw_queues == 1) || !is_sync); | |
878 | ||
879 | blk_queue_bounce(q, &bio); | |
880 | ||
14ec77f3 NB |
881 | if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) { |
882 | bio_endio(bio, -EIO); | |
883 | return; | |
884 | } | |
885 | ||
320ae51f JA |
886 | if (use_plug && blk_attempt_plug_merge(q, bio, &request_count)) |
887 | return; | |
888 | ||
889 | if (blk_mq_queue_enter(q)) { | |
890 | bio_endio(bio, -EIO); | |
891 | return; | |
892 | } | |
893 | ||
894 | ctx = blk_mq_get_ctx(q); | |
895 | hctx = q->mq_ops->map_queue(q, ctx->cpu); | |
896 | ||
27fbf4e8 SL |
897 | if (is_sync) |
898 | rw |= REQ_SYNC; | |
320ae51f | 899 | trace_block_getrq(q, bio, rw); |
18741986 | 900 | rq = __blk_mq_alloc_request(hctx, GFP_ATOMIC, false); |
320ae51f | 901 | if (likely(rq)) |
18741986 | 902 | blk_mq_rq_ctx_init(q, ctx, rq, rw); |
320ae51f JA |
903 | else { |
904 | blk_mq_put_ctx(ctx); | |
905 | trace_block_sleeprq(q, bio, rw); | |
18741986 CH |
906 | rq = blk_mq_alloc_request_pinned(q, rw, __GFP_WAIT|GFP_ATOMIC, |
907 | false); | |
320ae51f JA |
908 | ctx = rq->mq_ctx; |
909 | hctx = q->mq_ops->map_queue(q, ctx->cpu); | |
910 | } | |
911 | ||
912 | hctx->queued++; | |
913 | ||
914 | if (unlikely(is_flush_fua)) { | |
915 | blk_mq_bio_to_request(rq, bio); | |
320ae51f JA |
916 | blk_insert_flush(rq); |
917 | goto run_queue; | |
918 | } | |
919 | ||
920 | /* | |
921 | * A task plug currently exists. Since this is completely lockless, | |
922 | * utilize that to temporarily store requests until the task is | |
923 | * either done or scheduled away. | |
924 | */ | |
925 | if (use_plug) { | |
926 | struct blk_plug *plug = current->plug; | |
927 | ||
928 | if (plug) { | |
929 | blk_mq_bio_to_request(rq, bio); | |
92f399c7 | 930 | if (list_empty(&plug->mq_list)) |
320ae51f JA |
931 | trace_block_plug(q); |
932 | else if (request_count >= BLK_MAX_REQUEST_COUNT) { | |
933 | blk_flush_plug_list(plug, false); | |
934 | trace_block_plug(q); | |
935 | } | |
936 | list_add_tail(&rq->queuelist, &plug->mq_list); | |
937 | blk_mq_put_ctx(ctx); | |
938 | return; | |
939 | } | |
940 | } | |
941 | ||
942 | spin_lock(&ctx->lock); | |
943 | ||
944 | if ((hctx->flags & BLK_MQ_F_SHOULD_MERGE) && | |
945 | blk_mq_attempt_merge(q, ctx, bio)) | |
946 | __blk_mq_free_request(hctx, ctx, rq); | |
947 | else { | |
948 | blk_mq_bio_to_request(rq, bio); | |
72a0a36e | 949 | __blk_mq_insert_request(hctx, rq, false); |
320ae51f JA |
950 | } |
951 | ||
952 | spin_unlock(&ctx->lock); | |
320ae51f JA |
953 | |
954 | /* | |
955 | * For a SYNC request, send it to the hardware immediately. For an | |
956 | * ASYNC request, just ensure that we run it later on. The latter | |
957 | * allows for merging opportunities and more efficient dispatching. | |
958 | */ | |
959 | run_queue: | |
960 | blk_mq_run_hw_queue(hctx, !is_sync || is_flush_fua); | |
e4043dcf | 961 | blk_mq_put_ctx(ctx); |
320ae51f JA |
962 | } |
963 | ||
964 | /* | |
965 | * Default mapping to a software queue, since we use one per CPU. | |
966 | */ | |
967 | struct blk_mq_hw_ctx *blk_mq_map_queue(struct request_queue *q, const int cpu) | |
968 | { | |
969 | return q->queue_hw_ctx[q->mq_map[cpu]]; | |
970 | } | |
971 | EXPORT_SYMBOL(blk_mq_map_queue); | |
972 | ||
973 | struct blk_mq_hw_ctx *blk_mq_alloc_single_hw_queue(struct blk_mq_reg *reg, | |
974 | unsigned int hctx_index) | |
975 | { | |
976 | return kmalloc_node(sizeof(struct blk_mq_hw_ctx), | |
977 | GFP_KERNEL | __GFP_ZERO, reg->numa_node); | |
978 | } | |
979 | EXPORT_SYMBOL(blk_mq_alloc_single_hw_queue); | |
980 | ||
981 | void blk_mq_free_single_hw_queue(struct blk_mq_hw_ctx *hctx, | |
982 | unsigned int hctx_index) | |
983 | { | |
984 | kfree(hctx); | |
985 | } | |
986 | EXPORT_SYMBOL(blk_mq_free_single_hw_queue); | |
987 | ||
988 | static void blk_mq_hctx_notify(void *data, unsigned long action, | |
989 | unsigned int cpu) | |
990 | { | |
991 | struct blk_mq_hw_ctx *hctx = data; | |
bccb5f7c | 992 | struct request_queue *q = hctx->queue; |
320ae51f JA |
993 | struct blk_mq_ctx *ctx; |
994 | LIST_HEAD(tmp); | |
995 | ||
996 | if (action != CPU_DEAD && action != CPU_DEAD_FROZEN) | |
997 | return; | |
998 | ||
999 | /* | |
1000 | * Move ctx entries to new CPU, if this one is going away. | |
1001 | */ | |
bccb5f7c | 1002 | ctx = __blk_mq_get_ctx(q, cpu); |
320ae51f JA |
1003 | |
1004 | spin_lock(&ctx->lock); | |
1005 | if (!list_empty(&ctx->rq_list)) { | |
1006 | list_splice_init(&ctx->rq_list, &tmp); | |
1007 | clear_bit(ctx->index_hw, hctx->ctx_map); | |
1008 | } | |
1009 | spin_unlock(&ctx->lock); | |
1010 | ||
1011 | if (list_empty(&tmp)) | |
1012 | return; | |
1013 | ||
bccb5f7c | 1014 | ctx = blk_mq_get_ctx(q); |
320ae51f JA |
1015 | spin_lock(&ctx->lock); |
1016 | ||
1017 | while (!list_empty(&tmp)) { | |
1018 | struct request *rq; | |
1019 | ||
1020 | rq = list_first_entry(&tmp, struct request, queuelist); | |
1021 | rq->mq_ctx = ctx; | |
1022 | list_move_tail(&rq->queuelist, &ctx->rq_list); | |
1023 | } | |
1024 | ||
bccb5f7c | 1025 | hctx = q->mq_ops->map_queue(q, ctx->cpu); |
320ae51f JA |
1026 | blk_mq_hctx_mark_pending(hctx, ctx); |
1027 | ||
1028 | spin_unlock(&ctx->lock); | |
bccb5f7c JA |
1029 | |
1030 | blk_mq_run_hw_queue(hctx, true); | |
e4043dcf | 1031 | blk_mq_put_ctx(ctx); |
320ae51f JA |
1032 | } |
1033 | ||
95363efd JA |
1034 | static int blk_mq_init_hw_commands(struct blk_mq_hw_ctx *hctx, |
1035 | int (*init)(void *, struct blk_mq_hw_ctx *, | |
1036 | struct request *, unsigned int), | |
1037 | void *data) | |
1038 | { | |
1039 | unsigned int i; | |
1040 | int ret = 0; | |
1041 | ||
1042 | for (i = 0; i < hctx->queue_depth; i++) { | |
1043 | struct request *rq = hctx->rqs[i]; | |
1044 | ||
1045 | ret = init(data, hctx, rq, i); | |
1046 | if (ret) | |
1047 | break; | |
1048 | } | |
1049 | ||
1050 | return ret; | |
1051 | } | |
1052 | ||
1053 | int blk_mq_init_commands(struct request_queue *q, | |
1054 | int (*init)(void *, struct blk_mq_hw_ctx *, | |
1055 | struct request *, unsigned int), | |
1056 | void *data) | |
1057 | { | |
1058 | struct blk_mq_hw_ctx *hctx; | |
1059 | unsigned int i; | |
1060 | int ret = 0; | |
1061 | ||
1062 | queue_for_each_hw_ctx(q, hctx, i) { | |
1063 | ret = blk_mq_init_hw_commands(hctx, init, data); | |
1064 | if (ret) | |
1065 | break; | |
1066 | } | |
1067 | ||
1068 | return ret; | |
1069 | } | |
1070 | EXPORT_SYMBOL(blk_mq_init_commands); | |
1071 | ||
1072 | static void blk_mq_free_hw_commands(struct blk_mq_hw_ctx *hctx, | |
1073 | void (*free)(void *, struct blk_mq_hw_ctx *, | |
320ae51f JA |
1074 | struct request *, unsigned int), |
1075 | void *data) | |
1076 | { | |
1077 | unsigned int i; | |
1078 | ||
1079 | for (i = 0; i < hctx->queue_depth; i++) { | |
1080 | struct request *rq = hctx->rqs[i]; | |
1081 | ||
95363efd | 1082 | free(data, hctx, rq, i); |
320ae51f JA |
1083 | } |
1084 | } | |
1085 | ||
95363efd JA |
1086 | void blk_mq_free_commands(struct request_queue *q, |
1087 | void (*free)(void *, struct blk_mq_hw_ctx *, | |
320ae51f JA |
1088 | struct request *, unsigned int), |
1089 | void *data) | |
1090 | { | |
1091 | struct blk_mq_hw_ctx *hctx; | |
1092 | unsigned int i; | |
1093 | ||
1094 | queue_for_each_hw_ctx(q, hctx, i) | |
95363efd | 1095 | blk_mq_free_hw_commands(hctx, free, data); |
320ae51f | 1096 | } |
95363efd | 1097 | EXPORT_SYMBOL(blk_mq_free_commands); |
320ae51f JA |
1098 | |
1099 | static void blk_mq_free_rq_map(struct blk_mq_hw_ctx *hctx) | |
1100 | { | |
1101 | struct page *page; | |
1102 | ||
1103 | while (!list_empty(&hctx->page_list)) { | |
6753471c DH |
1104 | page = list_first_entry(&hctx->page_list, struct page, lru); |
1105 | list_del_init(&page->lru); | |
320ae51f JA |
1106 | __free_pages(page, page->private); |
1107 | } | |
1108 | ||
1109 | kfree(hctx->rqs); | |
1110 | ||
1111 | if (hctx->tags) | |
1112 | blk_mq_free_tags(hctx->tags); | |
1113 | } | |
1114 | ||
1115 | static size_t order_to_size(unsigned int order) | |
1116 | { | |
1117 | size_t ret = PAGE_SIZE; | |
1118 | ||
1119 | while (order--) | |
1120 | ret *= 2; | |
1121 | ||
1122 | return ret; | |
1123 | } | |
1124 | ||
1125 | static int blk_mq_init_rq_map(struct blk_mq_hw_ctx *hctx, | |
1126 | unsigned int reserved_tags, int node) | |
1127 | { | |
1128 | unsigned int i, j, entries_per_page, max_order = 4; | |
1129 | size_t rq_size, left; | |
1130 | ||
1131 | INIT_LIST_HEAD(&hctx->page_list); | |
1132 | ||
1133 | hctx->rqs = kmalloc_node(hctx->queue_depth * sizeof(struct request *), | |
1134 | GFP_KERNEL, node); | |
1135 | if (!hctx->rqs) | |
1136 | return -ENOMEM; | |
1137 | ||
1138 | /* | |
1139 | * rq_size is the size of the request plus driver payload, rounded | |
1140 | * to the cacheline size | |
1141 | */ | |
1142 | rq_size = round_up(sizeof(struct request) + hctx->cmd_size, | |
1143 | cache_line_size()); | |
1144 | left = rq_size * hctx->queue_depth; | |
1145 | ||
1146 | for (i = 0; i < hctx->queue_depth;) { | |
1147 | int this_order = max_order; | |
1148 | struct page *page; | |
1149 | int to_do; | |
1150 | void *p; | |
1151 | ||
1152 | while (left < order_to_size(this_order - 1) && this_order) | |
1153 | this_order--; | |
1154 | ||
1155 | do { | |
1156 | page = alloc_pages_node(node, GFP_KERNEL, this_order); | |
1157 | if (page) | |
1158 | break; | |
1159 | if (!this_order--) | |
1160 | break; | |
1161 | if (order_to_size(this_order) < rq_size) | |
1162 | break; | |
1163 | } while (1); | |
1164 | ||
1165 | if (!page) | |
1166 | break; | |
1167 | ||
1168 | page->private = this_order; | |
6753471c | 1169 | list_add_tail(&page->lru, &hctx->page_list); |
320ae51f JA |
1170 | |
1171 | p = page_address(page); | |
1172 | entries_per_page = order_to_size(this_order) / rq_size; | |
1173 | to_do = min(entries_per_page, hctx->queue_depth - i); | |
1174 | left -= to_do * rq_size; | |
1175 | for (j = 0; j < to_do; j++) { | |
1176 | hctx->rqs[i] = p; | |
9d74e257 | 1177 | blk_rq_init(hctx->queue, hctx->rqs[i]); |
320ae51f JA |
1178 | p += rq_size; |
1179 | i++; | |
1180 | } | |
1181 | } | |
1182 | ||
1183 | if (i < (reserved_tags + BLK_MQ_TAG_MIN)) | |
1184 | goto err_rq_map; | |
1185 | else if (i != hctx->queue_depth) { | |
1186 | hctx->queue_depth = i; | |
1187 | pr_warn("%s: queue depth set to %u because of low memory\n", | |
1188 | __func__, i); | |
1189 | } | |
1190 | ||
1191 | hctx->tags = blk_mq_init_tags(hctx->queue_depth, reserved_tags, node); | |
1192 | if (!hctx->tags) { | |
1193 | err_rq_map: | |
1194 | blk_mq_free_rq_map(hctx); | |
1195 | return -ENOMEM; | |
1196 | } | |
1197 | ||
1198 | return 0; | |
1199 | } | |
1200 | ||
1201 | static int blk_mq_init_hw_queues(struct request_queue *q, | |
1202 | struct blk_mq_reg *reg, void *driver_data) | |
1203 | { | |
1204 | struct blk_mq_hw_ctx *hctx; | |
1205 | unsigned int i, j; | |
1206 | ||
1207 | /* | |
1208 | * Initialize hardware queues | |
1209 | */ | |
1210 | queue_for_each_hw_ctx(q, hctx, i) { | |
1211 | unsigned int num_maps; | |
1212 | int node; | |
1213 | ||
1214 | node = hctx->numa_node; | |
1215 | if (node == NUMA_NO_NODE) | |
1216 | node = hctx->numa_node = reg->numa_node; | |
1217 | ||
1218 | INIT_DELAYED_WORK(&hctx->delayed_work, blk_mq_work_fn); | |
1219 | spin_lock_init(&hctx->lock); | |
1220 | INIT_LIST_HEAD(&hctx->dispatch); | |
1221 | hctx->queue = q; | |
1222 | hctx->queue_num = i; | |
1223 | hctx->flags = reg->flags; | |
1224 | hctx->queue_depth = reg->queue_depth; | |
1225 | hctx->cmd_size = reg->cmd_size; | |
1226 | ||
1227 | blk_mq_init_cpu_notifier(&hctx->cpu_notifier, | |
1228 | blk_mq_hctx_notify, hctx); | |
1229 | blk_mq_register_cpu_notifier(&hctx->cpu_notifier); | |
1230 | ||
1231 | if (blk_mq_init_rq_map(hctx, reg->reserved_tags, node)) | |
1232 | break; | |
1233 | ||
1234 | /* | |
1235 | * Allocate space for all possible cpus to avoid allocation in | |
1236 | * runtime | |
1237 | */ | |
1238 | hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *), | |
1239 | GFP_KERNEL, node); | |
1240 | if (!hctx->ctxs) | |
1241 | break; | |
1242 | ||
1243 | num_maps = ALIGN(nr_cpu_ids, BITS_PER_LONG) / BITS_PER_LONG; | |
1244 | hctx->ctx_map = kzalloc_node(num_maps * sizeof(unsigned long), | |
1245 | GFP_KERNEL, node); | |
1246 | if (!hctx->ctx_map) | |
1247 | break; | |
1248 | ||
1249 | hctx->nr_ctx_map = num_maps; | |
1250 | hctx->nr_ctx = 0; | |
1251 | ||
1252 | if (reg->ops->init_hctx && | |
1253 | reg->ops->init_hctx(hctx, driver_data, i)) | |
1254 | break; | |
1255 | } | |
1256 | ||
1257 | if (i == q->nr_hw_queues) | |
1258 | return 0; | |
1259 | ||
1260 | /* | |
1261 | * Init failed | |
1262 | */ | |
1263 | queue_for_each_hw_ctx(q, hctx, j) { | |
1264 | if (i == j) | |
1265 | break; | |
1266 | ||
1267 | if (reg->ops->exit_hctx) | |
1268 | reg->ops->exit_hctx(hctx, j); | |
1269 | ||
1270 | blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier); | |
1271 | blk_mq_free_rq_map(hctx); | |
1272 | kfree(hctx->ctxs); | |
1273 | } | |
1274 | ||
1275 | return 1; | |
1276 | } | |
1277 | ||
1278 | static void blk_mq_init_cpu_queues(struct request_queue *q, | |
1279 | unsigned int nr_hw_queues) | |
1280 | { | |
1281 | unsigned int i; | |
1282 | ||
1283 | for_each_possible_cpu(i) { | |
1284 | struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i); | |
1285 | struct blk_mq_hw_ctx *hctx; | |
1286 | ||
1287 | memset(__ctx, 0, sizeof(*__ctx)); | |
1288 | __ctx->cpu = i; | |
1289 | spin_lock_init(&__ctx->lock); | |
1290 | INIT_LIST_HEAD(&__ctx->rq_list); | |
1291 | __ctx->queue = q; | |
1292 | ||
1293 | /* If the cpu isn't online, the cpu is mapped to first hctx */ | |
320ae51f JA |
1294 | if (!cpu_online(i)) |
1295 | continue; | |
1296 | ||
e4043dcf JA |
1297 | hctx = q->mq_ops->map_queue(q, i); |
1298 | cpumask_set_cpu(i, hctx->cpumask); | |
1299 | hctx->nr_ctx++; | |
1300 | ||
320ae51f JA |
1301 | /* |
1302 | * Set local node, IFF we have more than one hw queue. If | |
1303 | * not, we remain on the home node of the device | |
1304 | */ | |
1305 | if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE) | |
1306 | hctx->numa_node = cpu_to_node(i); | |
1307 | } | |
1308 | } | |
1309 | ||
1310 | static void blk_mq_map_swqueue(struct request_queue *q) | |
1311 | { | |
1312 | unsigned int i; | |
1313 | struct blk_mq_hw_ctx *hctx; | |
1314 | struct blk_mq_ctx *ctx; | |
1315 | ||
1316 | queue_for_each_hw_ctx(q, hctx, i) { | |
e4043dcf | 1317 | cpumask_clear(hctx->cpumask); |
320ae51f JA |
1318 | hctx->nr_ctx = 0; |
1319 | } | |
1320 | ||
1321 | /* | |
1322 | * Map software to hardware queues | |
1323 | */ | |
1324 | queue_for_each_ctx(q, ctx, i) { | |
1325 | /* If the cpu isn't online, the cpu is mapped to first hctx */ | |
e4043dcf JA |
1326 | if (!cpu_online(i)) |
1327 | continue; | |
1328 | ||
320ae51f | 1329 | hctx = q->mq_ops->map_queue(q, i); |
e4043dcf | 1330 | cpumask_set_cpu(i, hctx->cpumask); |
320ae51f JA |
1331 | ctx->index_hw = hctx->nr_ctx; |
1332 | hctx->ctxs[hctx->nr_ctx++] = ctx; | |
1333 | } | |
1334 | } | |
1335 | ||
1336 | struct request_queue *blk_mq_init_queue(struct blk_mq_reg *reg, | |
1337 | void *driver_data) | |
1338 | { | |
1339 | struct blk_mq_hw_ctx **hctxs; | |
1340 | struct blk_mq_ctx *ctx; | |
1341 | struct request_queue *q; | |
1342 | int i; | |
1343 | ||
1344 | if (!reg->nr_hw_queues || | |
1345 | !reg->ops->queue_rq || !reg->ops->map_queue || | |
1346 | !reg->ops->alloc_hctx || !reg->ops->free_hctx) | |
1347 | return ERR_PTR(-EINVAL); | |
1348 | ||
1349 | if (!reg->queue_depth) | |
1350 | reg->queue_depth = BLK_MQ_MAX_DEPTH; | |
1351 | else if (reg->queue_depth > BLK_MQ_MAX_DEPTH) { | |
1352 | pr_err("blk-mq: queuedepth too large (%u)\n", reg->queue_depth); | |
1353 | reg->queue_depth = BLK_MQ_MAX_DEPTH; | |
1354 | } | |
1355 | ||
1356 | if (reg->queue_depth < (reg->reserved_tags + BLK_MQ_TAG_MIN)) | |
1357 | return ERR_PTR(-EINVAL); | |
1358 | ||
1359 | ctx = alloc_percpu(struct blk_mq_ctx); | |
1360 | if (!ctx) | |
1361 | return ERR_PTR(-ENOMEM); | |
1362 | ||
1363 | hctxs = kmalloc_node(reg->nr_hw_queues * sizeof(*hctxs), GFP_KERNEL, | |
1364 | reg->numa_node); | |
1365 | ||
1366 | if (!hctxs) | |
1367 | goto err_percpu; | |
1368 | ||
1369 | for (i = 0; i < reg->nr_hw_queues; i++) { | |
1370 | hctxs[i] = reg->ops->alloc_hctx(reg, i); | |
1371 | if (!hctxs[i]) | |
1372 | goto err_hctxs; | |
1373 | ||
e4043dcf JA |
1374 | if (!zalloc_cpumask_var(&hctxs[i]->cpumask, GFP_KERNEL)) |
1375 | goto err_hctxs; | |
1376 | ||
320ae51f JA |
1377 | hctxs[i]->numa_node = NUMA_NO_NODE; |
1378 | hctxs[i]->queue_num = i; | |
1379 | } | |
1380 | ||
1381 | q = blk_alloc_queue_node(GFP_KERNEL, reg->numa_node); | |
1382 | if (!q) | |
1383 | goto err_hctxs; | |
1384 | ||
1385 | q->mq_map = blk_mq_make_queue_map(reg); | |
1386 | if (!q->mq_map) | |
1387 | goto err_map; | |
1388 | ||
1389 | setup_timer(&q->timeout, blk_mq_rq_timer, (unsigned long) q); | |
1390 | blk_queue_rq_timeout(q, 30000); | |
1391 | ||
1392 | q->nr_queues = nr_cpu_ids; | |
1393 | q->nr_hw_queues = reg->nr_hw_queues; | |
1394 | ||
1395 | q->queue_ctx = ctx; | |
1396 | q->queue_hw_ctx = hctxs; | |
1397 | ||
1398 | q->mq_ops = reg->ops; | |
94eddfbe | 1399 | q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT; |
320ae51f | 1400 | |
1be036e9 CH |
1401 | q->sg_reserved_size = INT_MAX; |
1402 | ||
320ae51f JA |
1403 | blk_queue_make_request(q, blk_mq_make_request); |
1404 | blk_queue_rq_timed_out(q, reg->ops->timeout); | |
1405 | if (reg->timeout) | |
1406 | blk_queue_rq_timeout(q, reg->timeout); | |
1407 | ||
30a91cb4 CH |
1408 | if (reg->ops->complete) |
1409 | blk_queue_softirq_done(q, reg->ops->complete); | |
1410 | ||
320ae51f JA |
1411 | blk_mq_init_flush(q); |
1412 | blk_mq_init_cpu_queues(q, reg->nr_hw_queues); | |
1413 | ||
18741986 CH |
1414 | q->flush_rq = kzalloc(round_up(sizeof(struct request) + reg->cmd_size, |
1415 | cache_line_size()), GFP_KERNEL); | |
1416 | if (!q->flush_rq) | |
320ae51f JA |
1417 | goto err_hw; |
1418 | ||
18741986 CH |
1419 | if (blk_mq_init_hw_queues(q, reg, driver_data)) |
1420 | goto err_flush_rq; | |
1421 | ||
320ae51f JA |
1422 | blk_mq_map_swqueue(q); |
1423 | ||
1424 | mutex_lock(&all_q_mutex); | |
1425 | list_add_tail(&q->all_q_node, &all_q_list); | |
1426 | mutex_unlock(&all_q_mutex); | |
1427 | ||
1428 | return q; | |
18741986 CH |
1429 | |
1430 | err_flush_rq: | |
1431 | kfree(q->flush_rq); | |
320ae51f JA |
1432 | err_hw: |
1433 | kfree(q->mq_map); | |
1434 | err_map: | |
1435 | blk_cleanup_queue(q); | |
1436 | err_hctxs: | |
1437 | for (i = 0; i < reg->nr_hw_queues; i++) { | |
1438 | if (!hctxs[i]) | |
1439 | break; | |
e4043dcf | 1440 | free_cpumask_var(hctxs[i]->cpumask); |
320ae51f JA |
1441 | reg->ops->free_hctx(hctxs[i], i); |
1442 | } | |
1443 | kfree(hctxs); | |
1444 | err_percpu: | |
1445 | free_percpu(ctx); | |
1446 | return ERR_PTR(-ENOMEM); | |
1447 | } | |
1448 | EXPORT_SYMBOL(blk_mq_init_queue); | |
1449 | ||
1450 | void blk_mq_free_queue(struct request_queue *q) | |
1451 | { | |
1452 | struct blk_mq_hw_ctx *hctx; | |
1453 | int i; | |
1454 | ||
1455 | queue_for_each_hw_ctx(q, hctx, i) { | |
320ae51f JA |
1456 | kfree(hctx->ctx_map); |
1457 | kfree(hctx->ctxs); | |
1458 | blk_mq_free_rq_map(hctx); | |
1459 | blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier); | |
1460 | if (q->mq_ops->exit_hctx) | |
1461 | q->mq_ops->exit_hctx(hctx, i); | |
e4043dcf | 1462 | free_cpumask_var(hctx->cpumask); |
320ae51f JA |
1463 | q->mq_ops->free_hctx(hctx, i); |
1464 | } | |
1465 | ||
1466 | free_percpu(q->queue_ctx); | |
1467 | kfree(q->queue_hw_ctx); | |
1468 | kfree(q->mq_map); | |
1469 | ||
1470 | q->queue_ctx = NULL; | |
1471 | q->queue_hw_ctx = NULL; | |
1472 | q->mq_map = NULL; | |
1473 | ||
1474 | mutex_lock(&all_q_mutex); | |
1475 | list_del_init(&q->all_q_node); | |
1476 | mutex_unlock(&all_q_mutex); | |
1477 | } | |
320ae51f JA |
1478 | |
1479 | /* Basically redo blk_mq_init_queue with queue frozen */ | |
f618ef7c | 1480 | static void blk_mq_queue_reinit(struct request_queue *q) |
320ae51f JA |
1481 | { |
1482 | blk_mq_freeze_queue(q); | |
1483 | ||
1484 | blk_mq_update_queue_map(q->mq_map, q->nr_hw_queues); | |
1485 | ||
1486 | /* | |
1487 | * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe | |
1488 | * we should change hctx numa_node according to new topology (this | |
1489 | * involves free and re-allocate memory, worthy doing?) | |
1490 | */ | |
1491 | ||
1492 | blk_mq_map_swqueue(q); | |
1493 | ||
1494 | blk_mq_unfreeze_queue(q); | |
1495 | } | |
1496 | ||
f618ef7c PG |
1497 | static int blk_mq_queue_reinit_notify(struct notifier_block *nb, |
1498 | unsigned long action, void *hcpu) | |
320ae51f JA |
1499 | { |
1500 | struct request_queue *q; | |
1501 | ||
1502 | /* | |
1503 | * Before new mapping is established, hotadded cpu might already start | |
1504 | * handling requests. This doesn't break anything as we map offline | |
1505 | * CPUs to first hardware queue. We will re-init queue below to get | |
1506 | * optimal settings. | |
1507 | */ | |
1508 | if (action != CPU_DEAD && action != CPU_DEAD_FROZEN && | |
1509 | action != CPU_ONLINE && action != CPU_ONLINE_FROZEN) | |
1510 | return NOTIFY_OK; | |
1511 | ||
1512 | mutex_lock(&all_q_mutex); | |
1513 | list_for_each_entry(q, &all_q_list, all_q_node) | |
1514 | blk_mq_queue_reinit(q); | |
1515 | mutex_unlock(&all_q_mutex); | |
1516 | return NOTIFY_OK; | |
1517 | } | |
1518 | ||
676141e4 JA |
1519 | void blk_mq_disable_hotplug(void) |
1520 | { | |
1521 | mutex_lock(&all_q_mutex); | |
1522 | } | |
1523 | ||
1524 | void blk_mq_enable_hotplug(void) | |
1525 | { | |
1526 | mutex_unlock(&all_q_mutex); | |
1527 | } | |
1528 | ||
320ae51f JA |
1529 | static int __init blk_mq_init(void) |
1530 | { | |
320ae51f JA |
1531 | blk_mq_cpu_init(); |
1532 | ||
1533 | /* Must be called after percpu_counter_hotcpu_callback() */ | |
1534 | hotcpu_notifier(blk_mq_queue_reinit_notify, -10); | |
1535 | ||
1536 | return 0; | |
1537 | } | |
1538 | subsys_initcall(blk_mq_init); |