Commit | Line | Data |
---|---|---|
75bb4625 JA |
1 | /* |
2 | * Block multiqueue core code | |
3 | * | |
4 | * Copyright (C) 2013-2014 Jens Axboe | |
5 | * Copyright (C) 2013-2014 Christoph Hellwig | |
6 | */ | |
320ae51f JA |
7 | #include <linux/kernel.h> |
8 | #include <linux/module.h> | |
9 | #include <linux/backing-dev.h> | |
10 | #include <linux/bio.h> | |
11 | #include <linux/blkdev.h> | |
12 | #include <linux/mm.h> | |
13 | #include <linux/init.h> | |
14 | #include <linux/slab.h> | |
15 | #include <linux/workqueue.h> | |
16 | #include <linux/smp.h> | |
17 | #include <linux/llist.h> | |
18 | #include <linux/list_sort.h> | |
19 | #include <linux/cpu.h> | |
20 | #include <linux/cache.h> | |
21 | #include <linux/sched/sysctl.h> | |
22 | #include <linux/delay.h> | |
23 | ||
24 | #include <trace/events/block.h> | |
25 | ||
26 | #include <linux/blk-mq.h> | |
27 | #include "blk.h" | |
28 | #include "blk-mq.h" | |
29 | #include "blk-mq-tag.h" | |
30 | ||
31 | static DEFINE_MUTEX(all_q_mutex); | |
32 | static LIST_HEAD(all_q_list); | |
33 | ||
34 | static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx); | |
35 | ||
320ae51f JA |
36 | /* |
37 | * Check if any of the ctx's have pending work in this hardware queue | |
38 | */ | |
39 | static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx) | |
40 | { | |
41 | unsigned int i; | |
42 | ||
1429d7c9 JA |
43 | for (i = 0; i < hctx->ctx_map.map_size; i++) |
44 | if (hctx->ctx_map.map[i].word) | |
320ae51f JA |
45 | return true; |
46 | ||
47 | return false; | |
48 | } | |
49 | ||
1429d7c9 JA |
50 | static inline struct blk_align_bitmap *get_bm(struct blk_mq_hw_ctx *hctx, |
51 | struct blk_mq_ctx *ctx) | |
52 | { | |
53 | return &hctx->ctx_map.map[ctx->index_hw / hctx->ctx_map.bits_per_word]; | |
54 | } | |
55 | ||
56 | #define CTX_TO_BIT(hctx, ctx) \ | |
57 | ((ctx)->index_hw & ((hctx)->ctx_map.bits_per_word - 1)) | |
58 | ||
320ae51f JA |
59 | /* |
60 | * Mark this ctx as having pending work in this hardware queue | |
61 | */ | |
62 | static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx, | |
63 | struct blk_mq_ctx *ctx) | |
64 | { | |
1429d7c9 JA |
65 | struct blk_align_bitmap *bm = get_bm(hctx, ctx); |
66 | ||
67 | if (!test_bit(CTX_TO_BIT(hctx, ctx), &bm->word)) | |
68 | set_bit(CTX_TO_BIT(hctx, ctx), &bm->word); | |
69 | } | |
70 | ||
71 | static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx, | |
72 | struct blk_mq_ctx *ctx) | |
73 | { | |
74 | struct blk_align_bitmap *bm = get_bm(hctx, ctx); | |
75 | ||
76 | clear_bit(CTX_TO_BIT(hctx, ctx), &bm->word); | |
320ae51f JA |
77 | } |
78 | ||
320ae51f JA |
79 | static int blk_mq_queue_enter(struct request_queue *q) |
80 | { | |
81 | int ret; | |
82 | ||
83 | __percpu_counter_add(&q->mq_usage_counter, 1, 1000000); | |
84 | smp_wmb(); | |
3b632cf0 KB |
85 | |
86 | /* we have problems freezing the queue if it's initializing */ | |
87 | if (!blk_queue_dying(q) && | |
88 | (!blk_queue_bypass(q) || !blk_queue_init_done(q))) | |
320ae51f JA |
89 | return 0; |
90 | ||
91 | __percpu_counter_add(&q->mq_usage_counter, -1, 1000000); | |
92 | ||
93 | spin_lock_irq(q->queue_lock); | |
94 | ret = wait_event_interruptible_lock_irq(q->mq_freeze_wq, | |
43a5e4e2 ML |
95 | !blk_queue_bypass(q) || blk_queue_dying(q), |
96 | *q->queue_lock); | |
320ae51f | 97 | /* inc usage with lock hold to avoid freeze_queue runs here */ |
43a5e4e2 | 98 | if (!ret && !blk_queue_dying(q)) |
320ae51f | 99 | __percpu_counter_add(&q->mq_usage_counter, 1, 1000000); |
43a5e4e2 ML |
100 | else if (blk_queue_dying(q)) |
101 | ret = -ENODEV; | |
320ae51f JA |
102 | spin_unlock_irq(q->queue_lock); |
103 | ||
104 | return ret; | |
105 | } | |
106 | ||
107 | static void blk_mq_queue_exit(struct request_queue *q) | |
108 | { | |
109 | __percpu_counter_add(&q->mq_usage_counter, -1, 1000000); | |
110 | } | |
111 | ||
43a5e4e2 ML |
112 | static void __blk_mq_drain_queue(struct request_queue *q) |
113 | { | |
114 | while (true) { | |
115 | s64 count; | |
116 | ||
117 | spin_lock_irq(q->queue_lock); | |
118 | count = percpu_counter_sum(&q->mq_usage_counter); | |
119 | spin_unlock_irq(q->queue_lock); | |
120 | ||
121 | if (count == 0) | |
122 | break; | |
123 | blk_mq_run_queues(q, false); | |
124 | msleep(10); | |
125 | } | |
126 | } | |
127 | ||
320ae51f JA |
128 | /* |
129 | * Guarantee no request is in use, so we can change any data structure of | |
130 | * the queue afterward. | |
131 | */ | |
132 | static void blk_mq_freeze_queue(struct request_queue *q) | |
133 | { | |
134 | bool drain; | |
135 | ||
136 | spin_lock_irq(q->queue_lock); | |
137 | drain = !q->bypass_depth++; | |
138 | queue_flag_set(QUEUE_FLAG_BYPASS, q); | |
139 | spin_unlock_irq(q->queue_lock); | |
140 | ||
43a5e4e2 ML |
141 | if (drain) |
142 | __blk_mq_drain_queue(q); | |
143 | } | |
320ae51f | 144 | |
43a5e4e2 ML |
145 | void blk_mq_drain_queue(struct request_queue *q) |
146 | { | |
147 | __blk_mq_drain_queue(q); | |
320ae51f JA |
148 | } |
149 | ||
150 | static void blk_mq_unfreeze_queue(struct request_queue *q) | |
151 | { | |
152 | bool wake = false; | |
153 | ||
154 | spin_lock_irq(q->queue_lock); | |
155 | if (!--q->bypass_depth) { | |
156 | queue_flag_clear(QUEUE_FLAG_BYPASS, q); | |
157 | wake = true; | |
158 | } | |
159 | WARN_ON_ONCE(q->bypass_depth < 0); | |
160 | spin_unlock_irq(q->queue_lock); | |
161 | if (wake) | |
162 | wake_up_all(&q->mq_freeze_wq); | |
163 | } | |
164 | ||
165 | bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx) | |
166 | { | |
167 | return blk_mq_has_free_tags(hctx->tags); | |
168 | } | |
169 | EXPORT_SYMBOL(blk_mq_can_queue); | |
170 | ||
94eddfbe JA |
171 | static void blk_mq_rq_ctx_init(struct request_queue *q, struct blk_mq_ctx *ctx, |
172 | struct request *rq, unsigned int rw_flags) | |
320ae51f | 173 | { |
94eddfbe JA |
174 | if (blk_queue_io_stat(q)) |
175 | rw_flags |= REQ_IO_STAT; | |
176 | ||
af76e555 CH |
177 | INIT_LIST_HEAD(&rq->queuelist); |
178 | /* csd/requeue_work/fifo_time is initialized before use */ | |
179 | rq->q = q; | |
320ae51f | 180 | rq->mq_ctx = ctx; |
0d2602ca | 181 | rq->cmd_flags |= rw_flags; |
af76e555 CH |
182 | /* do not touch atomic flags, it needs atomic ops against the timer */ |
183 | rq->cpu = -1; | |
af76e555 CH |
184 | INIT_HLIST_NODE(&rq->hash); |
185 | RB_CLEAR_NODE(&rq->rb_node); | |
af76e555 CH |
186 | rq->rq_disk = NULL; |
187 | rq->part = NULL; | |
3ee32372 | 188 | rq->start_time = jiffies; |
af76e555 CH |
189 | #ifdef CONFIG_BLK_CGROUP |
190 | rq->rl = NULL; | |
0fec08b4 | 191 | set_start_time_ns(rq); |
af76e555 CH |
192 | rq->io_start_time_ns = 0; |
193 | #endif | |
194 | rq->nr_phys_segments = 0; | |
195 | #if defined(CONFIG_BLK_DEV_INTEGRITY) | |
196 | rq->nr_integrity_segments = 0; | |
197 | #endif | |
af76e555 CH |
198 | rq->special = NULL; |
199 | /* tag was already set */ | |
200 | rq->errors = 0; | |
af76e555 CH |
201 | |
202 | rq->extra_len = 0; | |
203 | rq->sense_len = 0; | |
204 | rq->resid_len = 0; | |
205 | rq->sense = NULL; | |
206 | ||
af76e555 | 207 | INIT_LIST_HEAD(&rq->timeout_list); |
f6be4fb4 JA |
208 | rq->timeout = 0; |
209 | ||
af76e555 CH |
210 | rq->end_io = NULL; |
211 | rq->end_io_data = NULL; | |
212 | rq->next_rq = NULL; | |
213 | ||
320ae51f JA |
214 | ctx->rq_dispatched[rw_is_sync(rw_flags)]++; |
215 | } | |
216 | ||
5dee8577 | 217 | static struct request * |
cb96a42c | 218 | __blk_mq_alloc_request(struct blk_mq_alloc_data *data, int rw) |
5dee8577 CH |
219 | { |
220 | struct request *rq; | |
221 | unsigned int tag; | |
222 | ||
cb96a42c | 223 | tag = blk_mq_get_tag(data); |
5dee8577 | 224 | if (tag != BLK_MQ_TAG_FAIL) { |
cb96a42c | 225 | rq = data->hctx->tags->rqs[tag]; |
5dee8577 CH |
226 | |
227 | rq->cmd_flags = 0; | |
cb96a42c | 228 | if (blk_mq_tag_busy(data->hctx)) { |
5dee8577 | 229 | rq->cmd_flags = REQ_MQ_INFLIGHT; |
cb96a42c | 230 | atomic_inc(&data->hctx->nr_active); |
5dee8577 CH |
231 | } |
232 | ||
233 | rq->tag = tag; | |
cb96a42c | 234 | blk_mq_rq_ctx_init(data->q, data->ctx, rq, rw); |
5dee8577 CH |
235 | return rq; |
236 | } | |
237 | ||
238 | return NULL; | |
239 | } | |
240 | ||
4ce01dd1 CH |
241 | struct request *blk_mq_alloc_request(struct request_queue *q, int rw, gfp_t gfp, |
242 | bool reserved) | |
320ae51f | 243 | { |
d852564f CH |
244 | struct blk_mq_ctx *ctx; |
245 | struct blk_mq_hw_ctx *hctx; | |
320ae51f | 246 | struct request *rq; |
cb96a42c | 247 | struct blk_mq_alloc_data alloc_data; |
320ae51f JA |
248 | |
249 | if (blk_mq_queue_enter(q)) | |
250 | return NULL; | |
251 | ||
d852564f CH |
252 | ctx = blk_mq_get_ctx(q); |
253 | hctx = q->mq_ops->map_queue(q, ctx->cpu); | |
cb96a42c ML |
254 | blk_mq_set_alloc_data(&alloc_data, q, gfp & ~__GFP_WAIT, |
255 | reserved, ctx, hctx); | |
d852564f | 256 | |
cb96a42c | 257 | rq = __blk_mq_alloc_request(&alloc_data, rw); |
d852564f CH |
258 | if (!rq && (gfp & __GFP_WAIT)) { |
259 | __blk_mq_run_hw_queue(hctx); | |
260 | blk_mq_put_ctx(ctx); | |
261 | ||
262 | ctx = blk_mq_get_ctx(q); | |
263 | hctx = q->mq_ops->map_queue(q, ctx->cpu); | |
cb96a42c ML |
264 | blk_mq_set_alloc_data(&alloc_data, q, gfp, reserved, ctx, |
265 | hctx); | |
266 | rq = __blk_mq_alloc_request(&alloc_data, rw); | |
267 | ctx = alloc_data.ctx; | |
d852564f CH |
268 | } |
269 | blk_mq_put_ctx(ctx); | |
320ae51f JA |
270 | return rq; |
271 | } | |
4bb659b1 | 272 | EXPORT_SYMBOL(blk_mq_alloc_request); |
320ae51f | 273 | |
320ae51f JA |
274 | static void __blk_mq_free_request(struct blk_mq_hw_ctx *hctx, |
275 | struct blk_mq_ctx *ctx, struct request *rq) | |
276 | { | |
277 | const int tag = rq->tag; | |
278 | struct request_queue *q = rq->q; | |
279 | ||
0d2602ca JA |
280 | if (rq->cmd_flags & REQ_MQ_INFLIGHT) |
281 | atomic_dec(&hctx->nr_active); | |
282 | ||
af76e555 | 283 | clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags); |
0d2602ca | 284 | blk_mq_put_tag(hctx, tag, &ctx->last_tag); |
320ae51f JA |
285 | blk_mq_queue_exit(q); |
286 | } | |
287 | ||
288 | void blk_mq_free_request(struct request *rq) | |
289 | { | |
290 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
291 | struct blk_mq_hw_ctx *hctx; | |
292 | struct request_queue *q = rq->q; | |
293 | ||
294 | ctx->rq_completed[rq_is_sync(rq)]++; | |
295 | ||
296 | hctx = q->mq_ops->map_queue(q, ctx->cpu); | |
297 | __blk_mq_free_request(hctx, ctx, rq); | |
298 | } | |
299 | ||
8727af4b CH |
300 | /* |
301 | * Clone all relevant state from a request that has been put on hold in | |
302 | * the flush state machine into the preallocated flush request that hangs | |
303 | * off the request queue. | |
304 | * | |
305 | * For a driver the flush request should be invisible, that's why we are | |
306 | * impersonating the original request here. | |
307 | */ | |
308 | void blk_mq_clone_flush_request(struct request *flush_rq, | |
309 | struct request *orig_rq) | |
310 | { | |
311 | struct blk_mq_hw_ctx *hctx = | |
312 | orig_rq->q->mq_ops->map_queue(orig_rq->q, orig_rq->mq_ctx->cpu); | |
313 | ||
314 | flush_rq->mq_ctx = orig_rq->mq_ctx; | |
315 | flush_rq->tag = orig_rq->tag; | |
316 | memcpy(blk_mq_rq_to_pdu(flush_rq), blk_mq_rq_to_pdu(orig_rq), | |
317 | hctx->cmd_size); | |
318 | } | |
319 | ||
63151a44 | 320 | inline void __blk_mq_end_io(struct request *rq, int error) |
320ae51f | 321 | { |
0d11e6ac ML |
322 | blk_account_io_done(rq); |
323 | ||
91b63639 | 324 | if (rq->end_io) { |
320ae51f | 325 | rq->end_io(rq, error); |
91b63639 CH |
326 | } else { |
327 | if (unlikely(blk_bidi_rq(rq))) | |
328 | blk_mq_free_request(rq->next_rq); | |
320ae51f | 329 | blk_mq_free_request(rq); |
91b63639 | 330 | } |
320ae51f | 331 | } |
63151a44 CH |
332 | EXPORT_SYMBOL(__blk_mq_end_io); |
333 | ||
334 | void blk_mq_end_io(struct request *rq, int error) | |
335 | { | |
336 | if (blk_update_request(rq, error, blk_rq_bytes(rq))) | |
337 | BUG(); | |
338 | __blk_mq_end_io(rq, error); | |
339 | } | |
340 | EXPORT_SYMBOL(blk_mq_end_io); | |
320ae51f | 341 | |
30a91cb4 | 342 | static void __blk_mq_complete_request_remote(void *data) |
320ae51f | 343 | { |
3d6efbf6 | 344 | struct request *rq = data; |
320ae51f | 345 | |
30a91cb4 | 346 | rq->q->softirq_done_fn(rq); |
320ae51f | 347 | } |
320ae51f | 348 | |
ed851860 | 349 | static void blk_mq_ipi_complete_request(struct request *rq) |
320ae51f JA |
350 | { |
351 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
38535201 | 352 | bool shared = false; |
320ae51f JA |
353 | int cpu; |
354 | ||
38535201 | 355 | if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) { |
30a91cb4 CH |
356 | rq->q->softirq_done_fn(rq); |
357 | return; | |
358 | } | |
320ae51f JA |
359 | |
360 | cpu = get_cpu(); | |
38535201 CH |
361 | if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags)) |
362 | shared = cpus_share_cache(cpu, ctx->cpu); | |
363 | ||
364 | if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) { | |
30a91cb4 | 365 | rq->csd.func = __blk_mq_complete_request_remote; |
3d6efbf6 CH |
366 | rq->csd.info = rq; |
367 | rq->csd.flags = 0; | |
c46fff2a | 368 | smp_call_function_single_async(ctx->cpu, &rq->csd); |
3d6efbf6 | 369 | } else { |
30a91cb4 | 370 | rq->q->softirq_done_fn(rq); |
3d6efbf6 | 371 | } |
320ae51f JA |
372 | put_cpu(); |
373 | } | |
30a91cb4 | 374 | |
ed851860 JA |
375 | void __blk_mq_complete_request(struct request *rq) |
376 | { | |
377 | struct request_queue *q = rq->q; | |
378 | ||
379 | if (!q->softirq_done_fn) | |
380 | blk_mq_end_io(rq, rq->errors); | |
381 | else | |
382 | blk_mq_ipi_complete_request(rq); | |
383 | } | |
384 | ||
30a91cb4 CH |
385 | /** |
386 | * blk_mq_complete_request - end I/O on a request | |
387 | * @rq: the request being processed | |
388 | * | |
389 | * Description: | |
390 | * Ends all I/O on a request. It does not handle partial completions. | |
391 | * The actual completion happens out-of-order, through a IPI handler. | |
392 | **/ | |
393 | void blk_mq_complete_request(struct request *rq) | |
394 | { | |
95f09684 JA |
395 | struct request_queue *q = rq->q; |
396 | ||
397 | if (unlikely(blk_should_fake_timeout(q))) | |
30a91cb4 | 398 | return; |
ed851860 JA |
399 | if (!blk_mark_rq_complete(rq)) |
400 | __blk_mq_complete_request(rq); | |
30a91cb4 CH |
401 | } |
402 | EXPORT_SYMBOL(blk_mq_complete_request); | |
320ae51f | 403 | |
49f5baa5 | 404 | static void blk_mq_start_request(struct request *rq, bool last) |
320ae51f JA |
405 | { |
406 | struct request_queue *q = rq->q; | |
407 | ||
408 | trace_block_rq_issue(q, rq); | |
409 | ||
742ee69b | 410 | rq->resid_len = blk_rq_bytes(rq); |
91b63639 CH |
411 | if (unlikely(blk_bidi_rq(rq))) |
412 | rq->next_rq->resid_len = blk_rq_bytes(rq->next_rq); | |
742ee69b | 413 | |
320ae51f JA |
414 | /* |
415 | * Just mark start time and set the started bit. Due to memory | |
416 | * ordering, we know we'll see the correct deadline as long as | |
c22d9d8a JA |
417 | * REQ_ATOMIC_STARTED is seen. Use the default queue timeout, |
418 | * unless one has been set in the request. | |
320ae51f | 419 | */ |
c22d9d8a JA |
420 | if (!rq->timeout) |
421 | rq->deadline = jiffies + q->rq_timeout; | |
422 | else | |
423 | rq->deadline = jiffies + rq->timeout; | |
87ee7b11 JA |
424 | |
425 | /* | |
426 | * Mark us as started and clear complete. Complete might have been | |
427 | * set if requeue raced with timeout, which then marked it as | |
428 | * complete. So be sure to clear complete again when we start | |
429 | * the request, otherwise we'll ignore the completion event. | |
430 | */ | |
4b570521 JA |
431 | if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) |
432 | set_bit(REQ_ATOM_STARTED, &rq->atomic_flags); | |
433 | if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags)) | |
434 | clear_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags); | |
49f5baa5 CH |
435 | |
436 | if (q->dma_drain_size && blk_rq_bytes(rq)) { | |
437 | /* | |
438 | * Make sure space for the drain appears. We know we can do | |
439 | * this because max_hw_segments has been adjusted to be one | |
440 | * fewer than the device can handle. | |
441 | */ | |
442 | rq->nr_phys_segments++; | |
443 | } | |
444 | ||
445 | /* | |
446 | * Flag the last request in the series so that drivers know when IO | |
447 | * should be kicked off, if they don't do it on a per-request basis. | |
448 | * | |
449 | * Note: the flag isn't the only condition drivers should do kick off. | |
450 | * If drive is busy, the last request might not have the bit set. | |
451 | */ | |
452 | if (last) | |
453 | rq->cmd_flags |= REQ_END; | |
320ae51f JA |
454 | } |
455 | ||
ed0791b2 | 456 | static void __blk_mq_requeue_request(struct request *rq) |
320ae51f JA |
457 | { |
458 | struct request_queue *q = rq->q; | |
459 | ||
460 | trace_block_rq_requeue(q, rq); | |
461 | clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags); | |
49f5baa5 CH |
462 | |
463 | rq->cmd_flags &= ~REQ_END; | |
464 | ||
465 | if (q->dma_drain_size && blk_rq_bytes(rq)) | |
466 | rq->nr_phys_segments--; | |
320ae51f JA |
467 | } |
468 | ||
ed0791b2 CH |
469 | void blk_mq_requeue_request(struct request *rq) |
470 | { | |
ed0791b2 CH |
471 | __blk_mq_requeue_request(rq); |
472 | blk_clear_rq_complete(rq); | |
473 | ||
ed0791b2 | 474 | BUG_ON(blk_queued_rq(rq)); |
6fca6a61 | 475 | blk_mq_add_to_requeue_list(rq, true); |
ed0791b2 CH |
476 | } |
477 | EXPORT_SYMBOL(blk_mq_requeue_request); | |
478 | ||
6fca6a61 CH |
479 | static void blk_mq_requeue_work(struct work_struct *work) |
480 | { | |
481 | struct request_queue *q = | |
482 | container_of(work, struct request_queue, requeue_work); | |
483 | LIST_HEAD(rq_list); | |
484 | struct request *rq, *next; | |
485 | unsigned long flags; | |
486 | ||
487 | spin_lock_irqsave(&q->requeue_lock, flags); | |
488 | list_splice_init(&q->requeue_list, &rq_list); | |
489 | spin_unlock_irqrestore(&q->requeue_lock, flags); | |
490 | ||
491 | list_for_each_entry_safe(rq, next, &rq_list, queuelist) { | |
492 | if (!(rq->cmd_flags & REQ_SOFTBARRIER)) | |
493 | continue; | |
494 | ||
495 | rq->cmd_flags &= ~REQ_SOFTBARRIER; | |
496 | list_del_init(&rq->queuelist); | |
497 | blk_mq_insert_request(rq, true, false, false); | |
498 | } | |
499 | ||
500 | while (!list_empty(&rq_list)) { | |
501 | rq = list_entry(rq_list.next, struct request, queuelist); | |
502 | list_del_init(&rq->queuelist); | |
503 | blk_mq_insert_request(rq, false, false, false); | |
504 | } | |
505 | ||
506 | blk_mq_run_queues(q, false); | |
507 | } | |
508 | ||
509 | void blk_mq_add_to_requeue_list(struct request *rq, bool at_head) | |
510 | { | |
511 | struct request_queue *q = rq->q; | |
512 | unsigned long flags; | |
513 | ||
514 | /* | |
515 | * We abuse this flag that is otherwise used by the I/O scheduler to | |
516 | * request head insertation from the workqueue. | |
517 | */ | |
518 | BUG_ON(rq->cmd_flags & REQ_SOFTBARRIER); | |
519 | ||
520 | spin_lock_irqsave(&q->requeue_lock, flags); | |
521 | if (at_head) { | |
522 | rq->cmd_flags |= REQ_SOFTBARRIER; | |
523 | list_add(&rq->queuelist, &q->requeue_list); | |
524 | } else { | |
525 | list_add_tail(&rq->queuelist, &q->requeue_list); | |
526 | } | |
527 | spin_unlock_irqrestore(&q->requeue_lock, flags); | |
528 | } | |
529 | EXPORT_SYMBOL(blk_mq_add_to_requeue_list); | |
530 | ||
531 | void blk_mq_kick_requeue_list(struct request_queue *q) | |
532 | { | |
533 | kblockd_schedule_work(&q->requeue_work); | |
534 | } | |
535 | EXPORT_SYMBOL(blk_mq_kick_requeue_list); | |
536 | ||
0e62f51f | 537 | static inline bool is_flush_request(struct request *rq, unsigned int tag) |
24d2f903 | 538 | { |
0e62f51f JA |
539 | return ((rq->cmd_flags & REQ_FLUSH_SEQ) && |
540 | rq->q->flush_rq->tag == tag); | |
541 | } | |
542 | ||
543 | struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag) | |
544 | { | |
545 | struct request *rq = tags->rqs[tag]; | |
22302375 | 546 | |
0e62f51f JA |
547 | if (!is_flush_request(rq, tag)) |
548 | return rq; | |
22302375 | 549 | |
0e62f51f | 550 | return rq->q->flush_rq; |
24d2f903 CH |
551 | } |
552 | EXPORT_SYMBOL(blk_mq_tag_to_rq); | |
553 | ||
320ae51f JA |
554 | struct blk_mq_timeout_data { |
555 | struct blk_mq_hw_ctx *hctx; | |
556 | unsigned long *next; | |
557 | unsigned int *next_set; | |
558 | }; | |
559 | ||
560 | static void blk_mq_timeout_check(void *__data, unsigned long *free_tags) | |
561 | { | |
562 | struct blk_mq_timeout_data *data = __data; | |
563 | struct blk_mq_hw_ctx *hctx = data->hctx; | |
564 | unsigned int tag; | |
565 | ||
566 | /* It may not be in flight yet (this is where | |
567 | * the REQ_ATOMIC_STARTED flag comes in). The requests are | |
568 | * statically allocated, so we know it's always safe to access the | |
569 | * memory associated with a bit offset into ->rqs[]. | |
570 | */ | |
571 | tag = 0; | |
572 | do { | |
573 | struct request *rq; | |
574 | ||
24d2f903 CH |
575 | tag = find_next_zero_bit(free_tags, hctx->tags->nr_tags, tag); |
576 | if (tag >= hctx->tags->nr_tags) | |
320ae51f JA |
577 | break; |
578 | ||
0e62f51f | 579 | rq = blk_mq_tag_to_rq(hctx->tags, tag++); |
24d2f903 CH |
580 | if (rq->q != hctx->queue) |
581 | continue; | |
320ae51f JA |
582 | if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) |
583 | continue; | |
584 | ||
585 | blk_rq_check_expired(rq, data->next, data->next_set); | |
586 | } while (1); | |
587 | } | |
588 | ||
589 | static void blk_mq_hw_ctx_check_timeout(struct blk_mq_hw_ctx *hctx, | |
590 | unsigned long *next, | |
591 | unsigned int *next_set) | |
592 | { | |
593 | struct blk_mq_timeout_data data = { | |
594 | .hctx = hctx, | |
595 | .next = next, | |
596 | .next_set = next_set, | |
597 | }; | |
598 | ||
599 | /* | |
600 | * Ask the tagging code to iterate busy requests, so we can | |
601 | * check them for timeout. | |
602 | */ | |
603 | blk_mq_tag_busy_iter(hctx->tags, blk_mq_timeout_check, &data); | |
604 | } | |
605 | ||
87ee7b11 JA |
606 | static enum blk_eh_timer_return blk_mq_rq_timed_out(struct request *rq) |
607 | { | |
608 | struct request_queue *q = rq->q; | |
609 | ||
610 | /* | |
611 | * We know that complete is set at this point. If STARTED isn't set | |
612 | * anymore, then the request isn't active and the "timeout" should | |
613 | * just be ignored. This can happen due to the bitflag ordering. | |
614 | * Timeout first checks if STARTED is set, and if it is, assumes | |
615 | * the request is active. But if we race with completion, then | |
616 | * we both flags will get cleared. So check here again, and ignore | |
617 | * a timeout event with a request that isn't active. | |
618 | */ | |
619 | if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) | |
620 | return BLK_EH_NOT_HANDLED; | |
621 | ||
622 | if (!q->mq_ops->timeout) | |
623 | return BLK_EH_RESET_TIMER; | |
624 | ||
625 | return q->mq_ops->timeout(rq); | |
626 | } | |
627 | ||
320ae51f JA |
628 | static void blk_mq_rq_timer(unsigned long data) |
629 | { | |
630 | struct request_queue *q = (struct request_queue *) data; | |
631 | struct blk_mq_hw_ctx *hctx; | |
632 | unsigned long next = 0; | |
633 | int i, next_set = 0; | |
634 | ||
484b4061 JA |
635 | queue_for_each_hw_ctx(q, hctx, i) { |
636 | /* | |
637 | * If not software queues are currently mapped to this | |
638 | * hardware queue, there's nothing to check | |
639 | */ | |
640 | if (!hctx->nr_ctx || !hctx->tags) | |
641 | continue; | |
642 | ||
320ae51f | 643 | blk_mq_hw_ctx_check_timeout(hctx, &next, &next_set); |
484b4061 | 644 | } |
320ae51f | 645 | |
0d2602ca JA |
646 | if (next_set) { |
647 | next = blk_rq_timeout(round_jiffies_up(next)); | |
648 | mod_timer(&q->timeout, next); | |
649 | } else { | |
650 | queue_for_each_hw_ctx(q, hctx, i) | |
651 | blk_mq_tag_idle(hctx); | |
652 | } | |
320ae51f JA |
653 | } |
654 | ||
655 | /* | |
656 | * Reverse check our software queue for entries that we could potentially | |
657 | * merge with. Currently includes a hand-wavy stop count of 8, to not spend | |
658 | * too much time checking for merges. | |
659 | */ | |
660 | static bool blk_mq_attempt_merge(struct request_queue *q, | |
661 | struct blk_mq_ctx *ctx, struct bio *bio) | |
662 | { | |
663 | struct request *rq; | |
664 | int checked = 8; | |
665 | ||
666 | list_for_each_entry_reverse(rq, &ctx->rq_list, queuelist) { | |
667 | int el_ret; | |
668 | ||
669 | if (!checked--) | |
670 | break; | |
671 | ||
672 | if (!blk_rq_merge_ok(rq, bio)) | |
673 | continue; | |
674 | ||
675 | el_ret = blk_try_merge(rq, bio); | |
676 | if (el_ret == ELEVATOR_BACK_MERGE) { | |
677 | if (bio_attempt_back_merge(q, rq, bio)) { | |
678 | ctx->rq_merged++; | |
679 | return true; | |
680 | } | |
681 | break; | |
682 | } else if (el_ret == ELEVATOR_FRONT_MERGE) { | |
683 | if (bio_attempt_front_merge(q, rq, bio)) { | |
684 | ctx->rq_merged++; | |
685 | return true; | |
686 | } | |
687 | break; | |
688 | } | |
689 | } | |
690 | ||
691 | return false; | |
692 | } | |
693 | ||
1429d7c9 JA |
694 | /* |
695 | * Process software queues that have been marked busy, splicing them | |
696 | * to the for-dispatch | |
697 | */ | |
698 | static void flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list) | |
699 | { | |
700 | struct blk_mq_ctx *ctx; | |
701 | int i; | |
702 | ||
703 | for (i = 0; i < hctx->ctx_map.map_size; i++) { | |
704 | struct blk_align_bitmap *bm = &hctx->ctx_map.map[i]; | |
705 | unsigned int off, bit; | |
706 | ||
707 | if (!bm->word) | |
708 | continue; | |
709 | ||
710 | bit = 0; | |
711 | off = i * hctx->ctx_map.bits_per_word; | |
712 | do { | |
713 | bit = find_next_bit(&bm->word, bm->depth, bit); | |
714 | if (bit >= bm->depth) | |
715 | break; | |
716 | ||
717 | ctx = hctx->ctxs[bit + off]; | |
718 | clear_bit(bit, &bm->word); | |
719 | spin_lock(&ctx->lock); | |
720 | list_splice_tail_init(&ctx->rq_list, list); | |
721 | spin_unlock(&ctx->lock); | |
722 | ||
723 | bit++; | |
724 | } while (1); | |
725 | } | |
726 | } | |
727 | ||
320ae51f JA |
728 | /* |
729 | * Run this hardware queue, pulling any software queues mapped to it in. | |
730 | * Note that this function currently has various problems around ordering | |
731 | * of IO. In particular, we'd like FIFO behaviour on handling existing | |
732 | * items on the hctx->dispatch list. Ignore that for now. | |
733 | */ | |
734 | static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx) | |
735 | { | |
736 | struct request_queue *q = hctx->queue; | |
320ae51f JA |
737 | struct request *rq; |
738 | LIST_HEAD(rq_list); | |
1429d7c9 | 739 | int queued; |
320ae51f | 740 | |
fd1270d5 | 741 | WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask)); |
e4043dcf | 742 | |
5d12f905 | 743 | if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state))) |
320ae51f JA |
744 | return; |
745 | ||
746 | hctx->run++; | |
747 | ||
748 | /* | |
749 | * Touch any software queue that has pending entries. | |
750 | */ | |
1429d7c9 | 751 | flush_busy_ctxs(hctx, &rq_list); |
320ae51f JA |
752 | |
753 | /* | |
754 | * If we have previous entries on our dispatch list, grab them | |
755 | * and stuff them at the front for more fair dispatch. | |
756 | */ | |
757 | if (!list_empty_careful(&hctx->dispatch)) { | |
758 | spin_lock(&hctx->lock); | |
759 | if (!list_empty(&hctx->dispatch)) | |
760 | list_splice_init(&hctx->dispatch, &rq_list); | |
761 | spin_unlock(&hctx->lock); | |
762 | } | |
763 | ||
320ae51f JA |
764 | /* |
765 | * Now process all the entries, sending them to the driver. | |
766 | */ | |
1429d7c9 | 767 | queued = 0; |
320ae51f JA |
768 | while (!list_empty(&rq_list)) { |
769 | int ret; | |
770 | ||
771 | rq = list_first_entry(&rq_list, struct request, queuelist); | |
772 | list_del_init(&rq->queuelist); | |
320ae51f | 773 | |
49f5baa5 | 774 | blk_mq_start_request(rq, list_empty(&rq_list)); |
320ae51f JA |
775 | |
776 | ret = q->mq_ops->queue_rq(hctx, rq); | |
777 | switch (ret) { | |
778 | case BLK_MQ_RQ_QUEUE_OK: | |
779 | queued++; | |
780 | continue; | |
781 | case BLK_MQ_RQ_QUEUE_BUSY: | |
320ae51f | 782 | list_add(&rq->queuelist, &rq_list); |
ed0791b2 | 783 | __blk_mq_requeue_request(rq); |
320ae51f JA |
784 | break; |
785 | default: | |
786 | pr_err("blk-mq: bad return on queue: %d\n", ret); | |
320ae51f | 787 | case BLK_MQ_RQ_QUEUE_ERROR: |
1e93b8c2 | 788 | rq->errors = -EIO; |
320ae51f JA |
789 | blk_mq_end_io(rq, rq->errors); |
790 | break; | |
791 | } | |
792 | ||
793 | if (ret == BLK_MQ_RQ_QUEUE_BUSY) | |
794 | break; | |
795 | } | |
796 | ||
797 | if (!queued) | |
798 | hctx->dispatched[0]++; | |
799 | else if (queued < (1 << (BLK_MQ_MAX_DISPATCH_ORDER - 1))) | |
800 | hctx->dispatched[ilog2(queued) + 1]++; | |
801 | ||
802 | /* | |
803 | * Any items that need requeuing? Stuff them into hctx->dispatch, | |
804 | * that is where we will continue on next queue run. | |
805 | */ | |
806 | if (!list_empty(&rq_list)) { | |
807 | spin_lock(&hctx->lock); | |
808 | list_splice(&rq_list, &hctx->dispatch); | |
809 | spin_unlock(&hctx->lock); | |
810 | } | |
811 | } | |
812 | ||
506e931f JA |
813 | /* |
814 | * It'd be great if the workqueue API had a way to pass | |
815 | * in a mask and had some smarts for more clever placement. | |
816 | * For now we just round-robin here, switching for every | |
817 | * BLK_MQ_CPU_WORK_BATCH queued items. | |
818 | */ | |
819 | static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx) | |
820 | { | |
821 | int cpu = hctx->next_cpu; | |
822 | ||
823 | if (--hctx->next_cpu_batch <= 0) { | |
824 | int next_cpu; | |
825 | ||
826 | next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask); | |
827 | if (next_cpu >= nr_cpu_ids) | |
828 | next_cpu = cpumask_first(hctx->cpumask); | |
829 | ||
830 | hctx->next_cpu = next_cpu; | |
831 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; | |
832 | } | |
833 | ||
834 | return cpu; | |
835 | } | |
836 | ||
320ae51f JA |
837 | void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) |
838 | { | |
5d12f905 | 839 | if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state))) |
320ae51f JA |
840 | return; |
841 | ||
e4043dcf | 842 | if (!async && cpumask_test_cpu(smp_processor_id(), hctx->cpumask)) |
320ae51f | 843 | __blk_mq_run_hw_queue(hctx); |
e4043dcf | 844 | else if (hctx->queue->nr_hw_queues == 1) |
70f4db63 | 845 | kblockd_schedule_delayed_work(&hctx->run_work, 0); |
e4043dcf JA |
846 | else { |
847 | unsigned int cpu; | |
848 | ||
506e931f | 849 | cpu = blk_mq_hctx_next_cpu(hctx); |
70f4db63 | 850 | kblockd_schedule_delayed_work_on(cpu, &hctx->run_work, 0); |
e4043dcf | 851 | } |
320ae51f JA |
852 | } |
853 | ||
854 | void blk_mq_run_queues(struct request_queue *q, bool async) | |
855 | { | |
856 | struct blk_mq_hw_ctx *hctx; | |
857 | int i; | |
858 | ||
859 | queue_for_each_hw_ctx(q, hctx, i) { | |
860 | if ((!blk_mq_hctx_has_pending(hctx) && | |
861 | list_empty_careful(&hctx->dispatch)) || | |
5d12f905 | 862 | test_bit(BLK_MQ_S_STOPPED, &hctx->state)) |
320ae51f JA |
863 | continue; |
864 | ||
e4043dcf | 865 | preempt_disable(); |
320ae51f | 866 | blk_mq_run_hw_queue(hctx, async); |
e4043dcf | 867 | preempt_enable(); |
320ae51f JA |
868 | } |
869 | } | |
870 | EXPORT_SYMBOL(blk_mq_run_queues); | |
871 | ||
872 | void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx) | |
873 | { | |
70f4db63 CH |
874 | cancel_delayed_work(&hctx->run_work); |
875 | cancel_delayed_work(&hctx->delay_work); | |
320ae51f JA |
876 | set_bit(BLK_MQ_S_STOPPED, &hctx->state); |
877 | } | |
878 | EXPORT_SYMBOL(blk_mq_stop_hw_queue); | |
879 | ||
280d45f6 CH |
880 | void blk_mq_stop_hw_queues(struct request_queue *q) |
881 | { | |
882 | struct blk_mq_hw_ctx *hctx; | |
883 | int i; | |
884 | ||
885 | queue_for_each_hw_ctx(q, hctx, i) | |
886 | blk_mq_stop_hw_queue(hctx); | |
887 | } | |
888 | EXPORT_SYMBOL(blk_mq_stop_hw_queues); | |
889 | ||
320ae51f JA |
890 | void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx) |
891 | { | |
892 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
e4043dcf JA |
893 | |
894 | preempt_disable(); | |
320ae51f | 895 | __blk_mq_run_hw_queue(hctx); |
e4043dcf | 896 | preempt_enable(); |
320ae51f JA |
897 | } |
898 | EXPORT_SYMBOL(blk_mq_start_hw_queue); | |
899 | ||
2f268556 CH |
900 | void blk_mq_start_hw_queues(struct request_queue *q) |
901 | { | |
902 | struct blk_mq_hw_ctx *hctx; | |
903 | int i; | |
904 | ||
905 | queue_for_each_hw_ctx(q, hctx, i) | |
906 | blk_mq_start_hw_queue(hctx); | |
907 | } | |
908 | EXPORT_SYMBOL(blk_mq_start_hw_queues); | |
909 | ||
910 | ||
1b4a3258 | 911 | void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async) |
320ae51f JA |
912 | { |
913 | struct blk_mq_hw_ctx *hctx; | |
914 | int i; | |
915 | ||
916 | queue_for_each_hw_ctx(q, hctx, i) { | |
917 | if (!test_bit(BLK_MQ_S_STOPPED, &hctx->state)) | |
918 | continue; | |
919 | ||
920 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
e4043dcf | 921 | preempt_disable(); |
1b4a3258 | 922 | blk_mq_run_hw_queue(hctx, async); |
e4043dcf | 923 | preempt_enable(); |
320ae51f JA |
924 | } |
925 | } | |
926 | EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues); | |
927 | ||
70f4db63 | 928 | static void blk_mq_run_work_fn(struct work_struct *work) |
320ae51f JA |
929 | { |
930 | struct blk_mq_hw_ctx *hctx; | |
931 | ||
70f4db63 | 932 | hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work); |
e4043dcf | 933 | |
320ae51f JA |
934 | __blk_mq_run_hw_queue(hctx); |
935 | } | |
936 | ||
70f4db63 CH |
937 | static void blk_mq_delay_work_fn(struct work_struct *work) |
938 | { | |
939 | struct blk_mq_hw_ctx *hctx; | |
940 | ||
941 | hctx = container_of(work, struct blk_mq_hw_ctx, delay_work.work); | |
942 | ||
943 | if (test_and_clear_bit(BLK_MQ_S_STOPPED, &hctx->state)) | |
944 | __blk_mq_run_hw_queue(hctx); | |
945 | } | |
946 | ||
947 | void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs) | |
948 | { | |
949 | unsigned long tmo = msecs_to_jiffies(msecs); | |
950 | ||
951 | if (hctx->queue->nr_hw_queues == 1) | |
952 | kblockd_schedule_delayed_work(&hctx->delay_work, tmo); | |
953 | else { | |
954 | unsigned int cpu; | |
955 | ||
506e931f | 956 | cpu = blk_mq_hctx_next_cpu(hctx); |
70f4db63 CH |
957 | kblockd_schedule_delayed_work_on(cpu, &hctx->delay_work, tmo); |
958 | } | |
959 | } | |
960 | EXPORT_SYMBOL(blk_mq_delay_queue); | |
961 | ||
320ae51f | 962 | static void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, |
72a0a36e | 963 | struct request *rq, bool at_head) |
320ae51f JA |
964 | { |
965 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
966 | ||
01b983c9 JA |
967 | trace_block_rq_insert(hctx->queue, rq); |
968 | ||
72a0a36e CH |
969 | if (at_head) |
970 | list_add(&rq->queuelist, &ctx->rq_list); | |
971 | else | |
972 | list_add_tail(&rq->queuelist, &ctx->rq_list); | |
4bb659b1 | 973 | |
320ae51f JA |
974 | blk_mq_hctx_mark_pending(hctx, ctx); |
975 | ||
976 | /* | |
977 | * We do this early, to ensure we are on the right CPU. | |
978 | */ | |
87ee7b11 | 979 | blk_add_timer(rq); |
320ae51f JA |
980 | } |
981 | ||
eeabc850 CH |
982 | void blk_mq_insert_request(struct request *rq, bool at_head, bool run_queue, |
983 | bool async) | |
320ae51f | 984 | { |
eeabc850 | 985 | struct request_queue *q = rq->q; |
320ae51f | 986 | struct blk_mq_hw_ctx *hctx; |
eeabc850 CH |
987 | struct blk_mq_ctx *ctx = rq->mq_ctx, *current_ctx; |
988 | ||
989 | current_ctx = blk_mq_get_ctx(q); | |
990 | if (!cpu_online(ctx->cpu)) | |
991 | rq->mq_ctx = ctx = current_ctx; | |
320ae51f | 992 | |
320ae51f JA |
993 | hctx = q->mq_ops->map_queue(q, ctx->cpu); |
994 | ||
eeabc850 CH |
995 | if (rq->cmd_flags & (REQ_FLUSH | REQ_FUA) && |
996 | !(rq->cmd_flags & (REQ_FLUSH_SEQ))) { | |
320ae51f JA |
997 | blk_insert_flush(rq); |
998 | } else { | |
320ae51f | 999 | spin_lock(&ctx->lock); |
72a0a36e | 1000 | __blk_mq_insert_request(hctx, rq, at_head); |
320ae51f | 1001 | spin_unlock(&ctx->lock); |
320ae51f JA |
1002 | } |
1003 | ||
320ae51f JA |
1004 | if (run_queue) |
1005 | blk_mq_run_hw_queue(hctx, async); | |
e4043dcf JA |
1006 | |
1007 | blk_mq_put_ctx(current_ctx); | |
320ae51f JA |
1008 | } |
1009 | ||
1010 | static void blk_mq_insert_requests(struct request_queue *q, | |
1011 | struct blk_mq_ctx *ctx, | |
1012 | struct list_head *list, | |
1013 | int depth, | |
1014 | bool from_schedule) | |
1015 | ||
1016 | { | |
1017 | struct blk_mq_hw_ctx *hctx; | |
1018 | struct blk_mq_ctx *current_ctx; | |
1019 | ||
1020 | trace_block_unplug(q, depth, !from_schedule); | |
1021 | ||
1022 | current_ctx = blk_mq_get_ctx(q); | |
1023 | ||
1024 | if (!cpu_online(ctx->cpu)) | |
1025 | ctx = current_ctx; | |
1026 | hctx = q->mq_ops->map_queue(q, ctx->cpu); | |
1027 | ||
1028 | /* | |
1029 | * preemption doesn't flush plug list, so it's possible ctx->cpu is | |
1030 | * offline now | |
1031 | */ | |
1032 | spin_lock(&ctx->lock); | |
1033 | while (!list_empty(list)) { | |
1034 | struct request *rq; | |
1035 | ||
1036 | rq = list_first_entry(list, struct request, queuelist); | |
1037 | list_del_init(&rq->queuelist); | |
1038 | rq->mq_ctx = ctx; | |
72a0a36e | 1039 | __blk_mq_insert_request(hctx, rq, false); |
320ae51f JA |
1040 | } |
1041 | spin_unlock(&ctx->lock); | |
1042 | ||
320ae51f | 1043 | blk_mq_run_hw_queue(hctx, from_schedule); |
e4043dcf | 1044 | blk_mq_put_ctx(current_ctx); |
320ae51f JA |
1045 | } |
1046 | ||
1047 | static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b) | |
1048 | { | |
1049 | struct request *rqa = container_of(a, struct request, queuelist); | |
1050 | struct request *rqb = container_of(b, struct request, queuelist); | |
1051 | ||
1052 | return !(rqa->mq_ctx < rqb->mq_ctx || | |
1053 | (rqa->mq_ctx == rqb->mq_ctx && | |
1054 | blk_rq_pos(rqa) < blk_rq_pos(rqb))); | |
1055 | } | |
1056 | ||
1057 | void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule) | |
1058 | { | |
1059 | struct blk_mq_ctx *this_ctx; | |
1060 | struct request_queue *this_q; | |
1061 | struct request *rq; | |
1062 | LIST_HEAD(list); | |
1063 | LIST_HEAD(ctx_list); | |
1064 | unsigned int depth; | |
1065 | ||
1066 | list_splice_init(&plug->mq_list, &list); | |
1067 | ||
1068 | list_sort(NULL, &list, plug_ctx_cmp); | |
1069 | ||
1070 | this_q = NULL; | |
1071 | this_ctx = NULL; | |
1072 | depth = 0; | |
1073 | ||
1074 | while (!list_empty(&list)) { | |
1075 | rq = list_entry_rq(list.next); | |
1076 | list_del_init(&rq->queuelist); | |
1077 | BUG_ON(!rq->q); | |
1078 | if (rq->mq_ctx != this_ctx) { | |
1079 | if (this_ctx) { | |
1080 | blk_mq_insert_requests(this_q, this_ctx, | |
1081 | &ctx_list, depth, | |
1082 | from_schedule); | |
1083 | } | |
1084 | ||
1085 | this_ctx = rq->mq_ctx; | |
1086 | this_q = rq->q; | |
1087 | depth = 0; | |
1088 | } | |
1089 | ||
1090 | depth++; | |
1091 | list_add_tail(&rq->queuelist, &ctx_list); | |
1092 | } | |
1093 | ||
1094 | /* | |
1095 | * If 'this_ctx' is set, we know we have entries to complete | |
1096 | * on 'ctx_list'. Do those. | |
1097 | */ | |
1098 | if (this_ctx) { | |
1099 | blk_mq_insert_requests(this_q, this_ctx, &ctx_list, depth, | |
1100 | from_schedule); | |
1101 | } | |
1102 | } | |
1103 | ||
1104 | static void blk_mq_bio_to_request(struct request *rq, struct bio *bio) | |
1105 | { | |
1106 | init_request_from_bio(rq, bio); | |
4b570521 | 1107 | |
3ee32372 | 1108 | if (blk_do_io_stat(rq)) |
4b570521 | 1109 | blk_account_io_start(rq, 1); |
320ae51f JA |
1110 | } |
1111 | ||
07068d5b JA |
1112 | static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx *hctx, |
1113 | struct blk_mq_ctx *ctx, | |
1114 | struct request *rq, struct bio *bio) | |
320ae51f | 1115 | { |
07068d5b | 1116 | struct request_queue *q = hctx->queue; |
320ae51f | 1117 | |
07068d5b JA |
1118 | if (!(hctx->flags & BLK_MQ_F_SHOULD_MERGE)) { |
1119 | blk_mq_bio_to_request(rq, bio); | |
1120 | spin_lock(&ctx->lock); | |
1121 | insert_rq: | |
1122 | __blk_mq_insert_request(hctx, rq, false); | |
1123 | spin_unlock(&ctx->lock); | |
1124 | return false; | |
1125 | } else { | |
1126 | spin_lock(&ctx->lock); | |
1127 | if (!blk_mq_attempt_merge(q, ctx, bio)) { | |
1128 | blk_mq_bio_to_request(rq, bio); | |
1129 | goto insert_rq; | |
1130 | } | |
320ae51f | 1131 | |
07068d5b JA |
1132 | spin_unlock(&ctx->lock); |
1133 | __blk_mq_free_request(hctx, ctx, rq); | |
1134 | return true; | |
14ec77f3 | 1135 | } |
07068d5b | 1136 | } |
14ec77f3 | 1137 | |
07068d5b JA |
1138 | struct blk_map_ctx { |
1139 | struct blk_mq_hw_ctx *hctx; | |
1140 | struct blk_mq_ctx *ctx; | |
1141 | }; | |
1142 | ||
1143 | static struct request *blk_mq_map_request(struct request_queue *q, | |
1144 | struct bio *bio, | |
1145 | struct blk_map_ctx *data) | |
1146 | { | |
1147 | struct blk_mq_hw_ctx *hctx; | |
1148 | struct blk_mq_ctx *ctx; | |
1149 | struct request *rq; | |
1150 | int rw = bio_data_dir(bio); | |
cb96a42c | 1151 | struct blk_mq_alloc_data alloc_data; |
320ae51f | 1152 | |
07068d5b | 1153 | if (unlikely(blk_mq_queue_enter(q))) { |
320ae51f | 1154 | bio_endio(bio, -EIO); |
07068d5b | 1155 | return NULL; |
320ae51f JA |
1156 | } |
1157 | ||
1158 | ctx = blk_mq_get_ctx(q); | |
1159 | hctx = q->mq_ops->map_queue(q, ctx->cpu); | |
1160 | ||
07068d5b | 1161 | if (rw_is_sync(bio->bi_rw)) |
27fbf4e8 | 1162 | rw |= REQ_SYNC; |
07068d5b | 1163 | |
320ae51f | 1164 | trace_block_getrq(q, bio, rw); |
cb96a42c ML |
1165 | blk_mq_set_alloc_data(&alloc_data, q, GFP_ATOMIC, false, ctx, |
1166 | hctx); | |
1167 | rq = __blk_mq_alloc_request(&alloc_data, rw); | |
5dee8577 | 1168 | if (unlikely(!rq)) { |
793597a6 | 1169 | __blk_mq_run_hw_queue(hctx); |
320ae51f JA |
1170 | blk_mq_put_ctx(ctx); |
1171 | trace_block_sleeprq(q, bio, rw); | |
793597a6 CH |
1172 | |
1173 | ctx = blk_mq_get_ctx(q); | |
320ae51f | 1174 | hctx = q->mq_ops->map_queue(q, ctx->cpu); |
cb96a42c ML |
1175 | blk_mq_set_alloc_data(&alloc_data, q, |
1176 | __GFP_WAIT|GFP_ATOMIC, false, ctx, hctx); | |
1177 | rq = __blk_mq_alloc_request(&alloc_data, rw); | |
1178 | ctx = alloc_data.ctx; | |
1179 | hctx = alloc_data.hctx; | |
320ae51f JA |
1180 | } |
1181 | ||
1182 | hctx->queued++; | |
07068d5b JA |
1183 | data->hctx = hctx; |
1184 | data->ctx = ctx; | |
1185 | return rq; | |
1186 | } | |
1187 | ||
1188 | /* | |
1189 | * Multiple hardware queue variant. This will not use per-process plugs, | |
1190 | * but will attempt to bypass the hctx queueing if we can go straight to | |
1191 | * hardware for SYNC IO. | |
1192 | */ | |
1193 | static void blk_mq_make_request(struct request_queue *q, struct bio *bio) | |
1194 | { | |
1195 | const int is_sync = rw_is_sync(bio->bi_rw); | |
1196 | const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA); | |
1197 | struct blk_map_ctx data; | |
1198 | struct request *rq; | |
1199 | ||
1200 | blk_queue_bounce(q, &bio); | |
1201 | ||
1202 | if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) { | |
1203 | bio_endio(bio, -EIO); | |
1204 | return; | |
1205 | } | |
1206 | ||
1207 | rq = blk_mq_map_request(q, bio, &data); | |
1208 | if (unlikely(!rq)) | |
1209 | return; | |
1210 | ||
1211 | if (unlikely(is_flush_fua)) { | |
1212 | blk_mq_bio_to_request(rq, bio); | |
1213 | blk_insert_flush(rq); | |
1214 | goto run_queue; | |
1215 | } | |
1216 | ||
1217 | if (is_sync) { | |
1218 | int ret; | |
1219 | ||
1220 | blk_mq_bio_to_request(rq, bio); | |
1221 | blk_mq_start_request(rq, true); | |
feff6894 | 1222 | blk_add_timer(rq); |
07068d5b JA |
1223 | |
1224 | /* | |
1225 | * For OK queue, we are done. For error, kill it. Any other | |
1226 | * error (busy), just add it to our list as we previously | |
1227 | * would have done | |
1228 | */ | |
1229 | ret = q->mq_ops->queue_rq(data.hctx, rq); | |
1230 | if (ret == BLK_MQ_RQ_QUEUE_OK) | |
1231 | goto done; | |
1232 | else { | |
1233 | __blk_mq_requeue_request(rq); | |
1234 | ||
1235 | if (ret == BLK_MQ_RQ_QUEUE_ERROR) { | |
1236 | rq->errors = -EIO; | |
1237 | blk_mq_end_io(rq, rq->errors); | |
1238 | goto done; | |
1239 | } | |
1240 | } | |
1241 | } | |
1242 | ||
1243 | if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) { | |
1244 | /* | |
1245 | * For a SYNC request, send it to the hardware immediately. For | |
1246 | * an ASYNC request, just ensure that we run it later on. The | |
1247 | * latter allows for merging opportunities and more efficient | |
1248 | * dispatching. | |
1249 | */ | |
1250 | run_queue: | |
1251 | blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua); | |
1252 | } | |
1253 | done: | |
1254 | blk_mq_put_ctx(data.ctx); | |
1255 | } | |
1256 | ||
1257 | /* | |
1258 | * Single hardware queue variant. This will attempt to use any per-process | |
1259 | * plug for merging and IO deferral. | |
1260 | */ | |
1261 | static void blk_sq_make_request(struct request_queue *q, struct bio *bio) | |
1262 | { | |
1263 | const int is_sync = rw_is_sync(bio->bi_rw); | |
1264 | const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA); | |
1265 | unsigned int use_plug, request_count = 0; | |
1266 | struct blk_map_ctx data; | |
1267 | struct request *rq; | |
1268 | ||
1269 | /* | |
1270 | * If we have multiple hardware queues, just go directly to | |
1271 | * one of those for sync IO. | |
1272 | */ | |
1273 | use_plug = !is_flush_fua && !is_sync; | |
1274 | ||
1275 | blk_queue_bounce(q, &bio); | |
1276 | ||
1277 | if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) { | |
1278 | bio_endio(bio, -EIO); | |
1279 | return; | |
1280 | } | |
1281 | ||
1282 | if (use_plug && !blk_queue_nomerges(q) && | |
1283 | blk_attempt_plug_merge(q, bio, &request_count)) | |
1284 | return; | |
1285 | ||
1286 | rq = blk_mq_map_request(q, bio, &data); | |
ff87bcec JA |
1287 | if (unlikely(!rq)) |
1288 | return; | |
320ae51f JA |
1289 | |
1290 | if (unlikely(is_flush_fua)) { | |
1291 | blk_mq_bio_to_request(rq, bio); | |
320ae51f JA |
1292 | blk_insert_flush(rq); |
1293 | goto run_queue; | |
1294 | } | |
1295 | ||
1296 | /* | |
1297 | * A task plug currently exists. Since this is completely lockless, | |
1298 | * utilize that to temporarily store requests until the task is | |
1299 | * either done or scheduled away. | |
1300 | */ | |
1301 | if (use_plug) { | |
1302 | struct blk_plug *plug = current->plug; | |
1303 | ||
1304 | if (plug) { | |
1305 | blk_mq_bio_to_request(rq, bio); | |
92f399c7 | 1306 | if (list_empty(&plug->mq_list)) |
320ae51f JA |
1307 | trace_block_plug(q); |
1308 | else if (request_count >= BLK_MAX_REQUEST_COUNT) { | |
1309 | blk_flush_plug_list(plug, false); | |
1310 | trace_block_plug(q); | |
1311 | } | |
1312 | list_add_tail(&rq->queuelist, &plug->mq_list); | |
07068d5b | 1313 | blk_mq_put_ctx(data.ctx); |
320ae51f JA |
1314 | return; |
1315 | } | |
1316 | } | |
1317 | ||
07068d5b JA |
1318 | if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) { |
1319 | /* | |
1320 | * For a SYNC request, send it to the hardware immediately. For | |
1321 | * an ASYNC request, just ensure that we run it later on. The | |
1322 | * latter allows for merging opportunities and more efficient | |
1323 | * dispatching. | |
1324 | */ | |
1325 | run_queue: | |
1326 | blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua); | |
320ae51f JA |
1327 | } |
1328 | ||
07068d5b | 1329 | blk_mq_put_ctx(data.ctx); |
320ae51f JA |
1330 | } |
1331 | ||
1332 | /* | |
1333 | * Default mapping to a software queue, since we use one per CPU. | |
1334 | */ | |
1335 | struct blk_mq_hw_ctx *blk_mq_map_queue(struct request_queue *q, const int cpu) | |
1336 | { | |
1337 | return q->queue_hw_ctx[q->mq_map[cpu]]; | |
1338 | } | |
1339 | EXPORT_SYMBOL(blk_mq_map_queue); | |
1340 | ||
24d2f903 CH |
1341 | static void blk_mq_free_rq_map(struct blk_mq_tag_set *set, |
1342 | struct blk_mq_tags *tags, unsigned int hctx_idx) | |
95363efd | 1343 | { |
e9b267d9 | 1344 | struct page *page; |
320ae51f | 1345 | |
24d2f903 | 1346 | if (tags->rqs && set->ops->exit_request) { |
e9b267d9 | 1347 | int i; |
320ae51f | 1348 | |
24d2f903 CH |
1349 | for (i = 0; i < tags->nr_tags; i++) { |
1350 | if (!tags->rqs[i]) | |
e9b267d9 | 1351 | continue; |
24d2f903 CH |
1352 | set->ops->exit_request(set->driver_data, tags->rqs[i], |
1353 | hctx_idx, i); | |
e9b267d9 | 1354 | } |
320ae51f | 1355 | } |
320ae51f | 1356 | |
24d2f903 CH |
1357 | while (!list_empty(&tags->page_list)) { |
1358 | page = list_first_entry(&tags->page_list, struct page, lru); | |
6753471c | 1359 | list_del_init(&page->lru); |
320ae51f JA |
1360 | __free_pages(page, page->private); |
1361 | } | |
1362 | ||
24d2f903 | 1363 | kfree(tags->rqs); |
320ae51f | 1364 | |
24d2f903 | 1365 | blk_mq_free_tags(tags); |
320ae51f JA |
1366 | } |
1367 | ||
1368 | static size_t order_to_size(unsigned int order) | |
1369 | { | |
4ca08500 | 1370 | return (size_t)PAGE_SIZE << order; |
320ae51f JA |
1371 | } |
1372 | ||
24d2f903 CH |
1373 | static struct blk_mq_tags *blk_mq_init_rq_map(struct blk_mq_tag_set *set, |
1374 | unsigned int hctx_idx) | |
320ae51f | 1375 | { |
24d2f903 | 1376 | struct blk_mq_tags *tags; |
320ae51f JA |
1377 | unsigned int i, j, entries_per_page, max_order = 4; |
1378 | size_t rq_size, left; | |
1379 | ||
24d2f903 CH |
1380 | tags = blk_mq_init_tags(set->queue_depth, set->reserved_tags, |
1381 | set->numa_node); | |
1382 | if (!tags) | |
1383 | return NULL; | |
320ae51f | 1384 | |
24d2f903 CH |
1385 | INIT_LIST_HEAD(&tags->page_list); |
1386 | ||
1387 | tags->rqs = kmalloc_node(set->queue_depth * sizeof(struct request *), | |
1388 | GFP_KERNEL, set->numa_node); | |
1389 | if (!tags->rqs) { | |
1390 | blk_mq_free_tags(tags); | |
1391 | return NULL; | |
1392 | } | |
320ae51f JA |
1393 | |
1394 | /* | |
1395 | * rq_size is the size of the request plus driver payload, rounded | |
1396 | * to the cacheline size | |
1397 | */ | |
24d2f903 | 1398 | rq_size = round_up(sizeof(struct request) + set->cmd_size, |
320ae51f | 1399 | cache_line_size()); |
24d2f903 | 1400 | left = rq_size * set->queue_depth; |
320ae51f | 1401 | |
24d2f903 | 1402 | for (i = 0; i < set->queue_depth; ) { |
320ae51f JA |
1403 | int this_order = max_order; |
1404 | struct page *page; | |
1405 | int to_do; | |
1406 | void *p; | |
1407 | ||
1408 | while (left < order_to_size(this_order - 1) && this_order) | |
1409 | this_order--; | |
1410 | ||
1411 | do { | |
24d2f903 CH |
1412 | page = alloc_pages_node(set->numa_node, GFP_KERNEL, |
1413 | this_order); | |
320ae51f JA |
1414 | if (page) |
1415 | break; | |
1416 | if (!this_order--) | |
1417 | break; | |
1418 | if (order_to_size(this_order) < rq_size) | |
1419 | break; | |
1420 | } while (1); | |
1421 | ||
1422 | if (!page) | |
24d2f903 | 1423 | goto fail; |
320ae51f JA |
1424 | |
1425 | page->private = this_order; | |
24d2f903 | 1426 | list_add_tail(&page->lru, &tags->page_list); |
320ae51f JA |
1427 | |
1428 | p = page_address(page); | |
1429 | entries_per_page = order_to_size(this_order) / rq_size; | |
24d2f903 | 1430 | to_do = min(entries_per_page, set->queue_depth - i); |
320ae51f JA |
1431 | left -= to_do * rq_size; |
1432 | for (j = 0; j < to_do; j++) { | |
24d2f903 CH |
1433 | tags->rqs[i] = p; |
1434 | if (set->ops->init_request) { | |
1435 | if (set->ops->init_request(set->driver_data, | |
1436 | tags->rqs[i], hctx_idx, i, | |
1437 | set->numa_node)) | |
1438 | goto fail; | |
e9b267d9 CH |
1439 | } |
1440 | ||
320ae51f JA |
1441 | p += rq_size; |
1442 | i++; | |
1443 | } | |
1444 | } | |
1445 | ||
24d2f903 | 1446 | return tags; |
320ae51f | 1447 | |
24d2f903 CH |
1448 | fail: |
1449 | pr_warn("%s: failed to allocate requests\n", __func__); | |
1450 | blk_mq_free_rq_map(set, tags, hctx_idx); | |
1451 | return NULL; | |
320ae51f JA |
1452 | } |
1453 | ||
1429d7c9 JA |
1454 | static void blk_mq_free_bitmap(struct blk_mq_ctxmap *bitmap) |
1455 | { | |
1456 | kfree(bitmap->map); | |
1457 | } | |
1458 | ||
1459 | static int blk_mq_alloc_bitmap(struct blk_mq_ctxmap *bitmap, int node) | |
1460 | { | |
1461 | unsigned int bpw = 8, total, num_maps, i; | |
1462 | ||
1463 | bitmap->bits_per_word = bpw; | |
1464 | ||
1465 | num_maps = ALIGN(nr_cpu_ids, bpw) / bpw; | |
1466 | bitmap->map = kzalloc_node(num_maps * sizeof(struct blk_align_bitmap), | |
1467 | GFP_KERNEL, node); | |
1468 | if (!bitmap->map) | |
1469 | return -ENOMEM; | |
1470 | ||
1471 | bitmap->map_size = num_maps; | |
1472 | ||
1473 | total = nr_cpu_ids; | |
1474 | for (i = 0; i < num_maps; i++) { | |
1475 | bitmap->map[i].depth = min(total, bitmap->bits_per_word); | |
1476 | total -= bitmap->map[i].depth; | |
1477 | } | |
1478 | ||
1479 | return 0; | |
1480 | } | |
1481 | ||
484b4061 JA |
1482 | static int blk_mq_hctx_cpu_offline(struct blk_mq_hw_ctx *hctx, int cpu) |
1483 | { | |
1484 | struct request_queue *q = hctx->queue; | |
1485 | struct blk_mq_ctx *ctx; | |
1486 | LIST_HEAD(tmp); | |
1487 | ||
1488 | /* | |
1489 | * Move ctx entries to new CPU, if this one is going away. | |
1490 | */ | |
1491 | ctx = __blk_mq_get_ctx(q, cpu); | |
1492 | ||
1493 | spin_lock(&ctx->lock); | |
1494 | if (!list_empty(&ctx->rq_list)) { | |
1495 | list_splice_init(&ctx->rq_list, &tmp); | |
1496 | blk_mq_hctx_clear_pending(hctx, ctx); | |
1497 | } | |
1498 | spin_unlock(&ctx->lock); | |
1499 | ||
1500 | if (list_empty(&tmp)) | |
1501 | return NOTIFY_OK; | |
1502 | ||
1503 | ctx = blk_mq_get_ctx(q); | |
1504 | spin_lock(&ctx->lock); | |
1505 | ||
1506 | while (!list_empty(&tmp)) { | |
1507 | struct request *rq; | |
1508 | ||
1509 | rq = list_first_entry(&tmp, struct request, queuelist); | |
1510 | rq->mq_ctx = ctx; | |
1511 | list_move_tail(&rq->queuelist, &ctx->rq_list); | |
1512 | } | |
1513 | ||
1514 | hctx = q->mq_ops->map_queue(q, ctx->cpu); | |
1515 | blk_mq_hctx_mark_pending(hctx, ctx); | |
1516 | ||
1517 | spin_unlock(&ctx->lock); | |
1518 | ||
1519 | blk_mq_run_hw_queue(hctx, true); | |
1520 | blk_mq_put_ctx(ctx); | |
1521 | return NOTIFY_OK; | |
1522 | } | |
1523 | ||
1524 | static int blk_mq_hctx_cpu_online(struct blk_mq_hw_ctx *hctx, int cpu) | |
1525 | { | |
1526 | struct request_queue *q = hctx->queue; | |
1527 | struct blk_mq_tag_set *set = q->tag_set; | |
1528 | ||
1529 | if (set->tags[hctx->queue_num]) | |
1530 | return NOTIFY_OK; | |
1531 | ||
1532 | set->tags[hctx->queue_num] = blk_mq_init_rq_map(set, hctx->queue_num); | |
1533 | if (!set->tags[hctx->queue_num]) | |
1534 | return NOTIFY_STOP; | |
1535 | ||
1536 | hctx->tags = set->tags[hctx->queue_num]; | |
1537 | return NOTIFY_OK; | |
1538 | } | |
1539 | ||
1540 | static int blk_mq_hctx_notify(void *data, unsigned long action, | |
1541 | unsigned int cpu) | |
1542 | { | |
1543 | struct blk_mq_hw_ctx *hctx = data; | |
1544 | ||
1545 | if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) | |
1546 | return blk_mq_hctx_cpu_offline(hctx, cpu); | |
1547 | else if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) | |
1548 | return blk_mq_hctx_cpu_online(hctx, cpu); | |
1549 | ||
1550 | return NOTIFY_OK; | |
1551 | } | |
1552 | ||
624dbe47 ML |
1553 | static void blk_mq_exit_hw_queues(struct request_queue *q, |
1554 | struct blk_mq_tag_set *set, int nr_queue) | |
1555 | { | |
1556 | struct blk_mq_hw_ctx *hctx; | |
1557 | unsigned int i; | |
1558 | ||
1559 | queue_for_each_hw_ctx(q, hctx, i) { | |
1560 | if (i == nr_queue) | |
1561 | break; | |
1562 | ||
f899fed4 JA |
1563 | blk_mq_tag_idle(hctx); |
1564 | ||
624dbe47 ML |
1565 | if (set->ops->exit_hctx) |
1566 | set->ops->exit_hctx(hctx, i); | |
1567 | ||
1568 | blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier); | |
1569 | kfree(hctx->ctxs); | |
1570 | blk_mq_free_bitmap(&hctx->ctx_map); | |
1571 | } | |
1572 | ||
1573 | } | |
1574 | ||
1575 | static void blk_mq_free_hw_queues(struct request_queue *q, | |
1576 | struct blk_mq_tag_set *set) | |
1577 | { | |
1578 | struct blk_mq_hw_ctx *hctx; | |
1579 | unsigned int i; | |
1580 | ||
1581 | queue_for_each_hw_ctx(q, hctx, i) { | |
1582 | free_cpumask_var(hctx->cpumask); | |
cdef54dd | 1583 | kfree(hctx); |
624dbe47 ML |
1584 | } |
1585 | } | |
1586 | ||
320ae51f | 1587 | static int blk_mq_init_hw_queues(struct request_queue *q, |
24d2f903 | 1588 | struct blk_mq_tag_set *set) |
320ae51f JA |
1589 | { |
1590 | struct blk_mq_hw_ctx *hctx; | |
624dbe47 | 1591 | unsigned int i; |
320ae51f JA |
1592 | |
1593 | /* | |
1594 | * Initialize hardware queues | |
1595 | */ | |
1596 | queue_for_each_hw_ctx(q, hctx, i) { | |
320ae51f JA |
1597 | int node; |
1598 | ||
1599 | node = hctx->numa_node; | |
1600 | if (node == NUMA_NO_NODE) | |
24d2f903 | 1601 | node = hctx->numa_node = set->numa_node; |
320ae51f | 1602 | |
70f4db63 CH |
1603 | INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn); |
1604 | INIT_DELAYED_WORK(&hctx->delay_work, blk_mq_delay_work_fn); | |
320ae51f JA |
1605 | spin_lock_init(&hctx->lock); |
1606 | INIT_LIST_HEAD(&hctx->dispatch); | |
1607 | hctx->queue = q; | |
1608 | hctx->queue_num = i; | |
24d2f903 CH |
1609 | hctx->flags = set->flags; |
1610 | hctx->cmd_size = set->cmd_size; | |
320ae51f JA |
1611 | |
1612 | blk_mq_init_cpu_notifier(&hctx->cpu_notifier, | |
1613 | blk_mq_hctx_notify, hctx); | |
1614 | blk_mq_register_cpu_notifier(&hctx->cpu_notifier); | |
1615 | ||
24d2f903 | 1616 | hctx->tags = set->tags[i]; |
320ae51f JA |
1617 | |
1618 | /* | |
1619 | * Allocate space for all possible cpus to avoid allocation in | |
1620 | * runtime | |
1621 | */ | |
1622 | hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *), | |
1623 | GFP_KERNEL, node); | |
1624 | if (!hctx->ctxs) | |
1625 | break; | |
1626 | ||
1429d7c9 | 1627 | if (blk_mq_alloc_bitmap(&hctx->ctx_map, node)) |
320ae51f JA |
1628 | break; |
1629 | ||
320ae51f JA |
1630 | hctx->nr_ctx = 0; |
1631 | ||
24d2f903 CH |
1632 | if (set->ops->init_hctx && |
1633 | set->ops->init_hctx(hctx, set->driver_data, i)) | |
320ae51f JA |
1634 | break; |
1635 | } | |
1636 | ||
1637 | if (i == q->nr_hw_queues) | |
1638 | return 0; | |
1639 | ||
1640 | /* | |
1641 | * Init failed | |
1642 | */ | |
624dbe47 | 1643 | blk_mq_exit_hw_queues(q, set, i); |
320ae51f JA |
1644 | |
1645 | return 1; | |
1646 | } | |
1647 | ||
1648 | static void blk_mq_init_cpu_queues(struct request_queue *q, | |
1649 | unsigned int nr_hw_queues) | |
1650 | { | |
1651 | unsigned int i; | |
1652 | ||
1653 | for_each_possible_cpu(i) { | |
1654 | struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i); | |
1655 | struct blk_mq_hw_ctx *hctx; | |
1656 | ||
1657 | memset(__ctx, 0, sizeof(*__ctx)); | |
1658 | __ctx->cpu = i; | |
1659 | spin_lock_init(&__ctx->lock); | |
1660 | INIT_LIST_HEAD(&__ctx->rq_list); | |
1661 | __ctx->queue = q; | |
1662 | ||
1663 | /* If the cpu isn't online, the cpu is mapped to first hctx */ | |
320ae51f JA |
1664 | if (!cpu_online(i)) |
1665 | continue; | |
1666 | ||
e4043dcf JA |
1667 | hctx = q->mq_ops->map_queue(q, i); |
1668 | cpumask_set_cpu(i, hctx->cpumask); | |
1669 | hctx->nr_ctx++; | |
1670 | ||
320ae51f JA |
1671 | /* |
1672 | * Set local node, IFF we have more than one hw queue. If | |
1673 | * not, we remain on the home node of the device | |
1674 | */ | |
1675 | if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE) | |
1676 | hctx->numa_node = cpu_to_node(i); | |
1677 | } | |
1678 | } | |
1679 | ||
1680 | static void blk_mq_map_swqueue(struct request_queue *q) | |
1681 | { | |
1682 | unsigned int i; | |
1683 | struct blk_mq_hw_ctx *hctx; | |
1684 | struct blk_mq_ctx *ctx; | |
1685 | ||
1686 | queue_for_each_hw_ctx(q, hctx, i) { | |
e4043dcf | 1687 | cpumask_clear(hctx->cpumask); |
320ae51f JA |
1688 | hctx->nr_ctx = 0; |
1689 | } | |
1690 | ||
1691 | /* | |
1692 | * Map software to hardware queues | |
1693 | */ | |
1694 | queue_for_each_ctx(q, ctx, i) { | |
1695 | /* If the cpu isn't online, the cpu is mapped to first hctx */ | |
e4043dcf JA |
1696 | if (!cpu_online(i)) |
1697 | continue; | |
1698 | ||
320ae51f | 1699 | hctx = q->mq_ops->map_queue(q, i); |
e4043dcf | 1700 | cpumask_set_cpu(i, hctx->cpumask); |
320ae51f JA |
1701 | ctx->index_hw = hctx->nr_ctx; |
1702 | hctx->ctxs[hctx->nr_ctx++] = ctx; | |
1703 | } | |
506e931f JA |
1704 | |
1705 | queue_for_each_hw_ctx(q, hctx, i) { | |
484b4061 JA |
1706 | /* |
1707 | * If not software queues are mapped to this hardware queue, | |
1708 | * disable it and free the request entries | |
1709 | */ | |
1710 | if (!hctx->nr_ctx) { | |
1711 | struct blk_mq_tag_set *set = q->tag_set; | |
1712 | ||
1713 | if (set->tags[i]) { | |
1714 | blk_mq_free_rq_map(set, set->tags[i], i); | |
1715 | set->tags[i] = NULL; | |
1716 | hctx->tags = NULL; | |
1717 | } | |
1718 | continue; | |
1719 | } | |
1720 | ||
1721 | /* | |
1722 | * Initialize batch roundrobin counts | |
1723 | */ | |
506e931f JA |
1724 | hctx->next_cpu = cpumask_first(hctx->cpumask); |
1725 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; | |
1726 | } | |
320ae51f JA |
1727 | } |
1728 | ||
0d2602ca JA |
1729 | static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set) |
1730 | { | |
1731 | struct blk_mq_hw_ctx *hctx; | |
1732 | struct request_queue *q; | |
1733 | bool shared; | |
1734 | int i; | |
1735 | ||
1736 | if (set->tag_list.next == set->tag_list.prev) | |
1737 | shared = false; | |
1738 | else | |
1739 | shared = true; | |
1740 | ||
1741 | list_for_each_entry(q, &set->tag_list, tag_set_list) { | |
1742 | blk_mq_freeze_queue(q); | |
1743 | ||
1744 | queue_for_each_hw_ctx(q, hctx, i) { | |
1745 | if (shared) | |
1746 | hctx->flags |= BLK_MQ_F_TAG_SHARED; | |
1747 | else | |
1748 | hctx->flags &= ~BLK_MQ_F_TAG_SHARED; | |
1749 | } | |
1750 | blk_mq_unfreeze_queue(q); | |
1751 | } | |
1752 | } | |
1753 | ||
1754 | static void blk_mq_del_queue_tag_set(struct request_queue *q) | |
1755 | { | |
1756 | struct blk_mq_tag_set *set = q->tag_set; | |
1757 | ||
1758 | blk_mq_freeze_queue(q); | |
1759 | ||
1760 | mutex_lock(&set->tag_list_lock); | |
1761 | list_del_init(&q->tag_set_list); | |
1762 | blk_mq_update_tag_set_depth(set); | |
1763 | mutex_unlock(&set->tag_list_lock); | |
1764 | ||
1765 | blk_mq_unfreeze_queue(q); | |
1766 | } | |
1767 | ||
1768 | static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set, | |
1769 | struct request_queue *q) | |
1770 | { | |
1771 | q->tag_set = set; | |
1772 | ||
1773 | mutex_lock(&set->tag_list_lock); | |
1774 | list_add_tail(&q->tag_set_list, &set->tag_list); | |
1775 | blk_mq_update_tag_set_depth(set); | |
1776 | mutex_unlock(&set->tag_list_lock); | |
1777 | } | |
1778 | ||
24d2f903 | 1779 | struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set) |
320ae51f JA |
1780 | { |
1781 | struct blk_mq_hw_ctx **hctxs; | |
e6cdb092 | 1782 | struct blk_mq_ctx __percpu *ctx; |
320ae51f | 1783 | struct request_queue *q; |
f14bbe77 | 1784 | unsigned int *map; |
320ae51f JA |
1785 | int i; |
1786 | ||
320ae51f JA |
1787 | ctx = alloc_percpu(struct blk_mq_ctx); |
1788 | if (!ctx) | |
1789 | return ERR_PTR(-ENOMEM); | |
1790 | ||
24d2f903 CH |
1791 | hctxs = kmalloc_node(set->nr_hw_queues * sizeof(*hctxs), GFP_KERNEL, |
1792 | set->numa_node); | |
320ae51f JA |
1793 | |
1794 | if (!hctxs) | |
1795 | goto err_percpu; | |
1796 | ||
f14bbe77 JA |
1797 | map = blk_mq_make_queue_map(set); |
1798 | if (!map) | |
1799 | goto err_map; | |
1800 | ||
24d2f903 | 1801 | for (i = 0; i < set->nr_hw_queues; i++) { |
f14bbe77 JA |
1802 | int node = blk_mq_hw_queue_to_node(map, i); |
1803 | ||
cdef54dd CH |
1804 | hctxs[i] = kzalloc_node(sizeof(struct blk_mq_hw_ctx), |
1805 | GFP_KERNEL, node); | |
320ae51f JA |
1806 | if (!hctxs[i]) |
1807 | goto err_hctxs; | |
1808 | ||
e4043dcf JA |
1809 | if (!zalloc_cpumask_var(&hctxs[i]->cpumask, GFP_KERNEL)) |
1810 | goto err_hctxs; | |
1811 | ||
0d2602ca | 1812 | atomic_set(&hctxs[i]->nr_active, 0); |
f14bbe77 | 1813 | hctxs[i]->numa_node = node; |
320ae51f JA |
1814 | hctxs[i]->queue_num = i; |
1815 | } | |
1816 | ||
24d2f903 | 1817 | q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node); |
320ae51f JA |
1818 | if (!q) |
1819 | goto err_hctxs; | |
1820 | ||
3d2936f4 ML |
1821 | if (percpu_counter_init(&q->mq_usage_counter, 0)) |
1822 | goto err_map; | |
1823 | ||
320ae51f JA |
1824 | setup_timer(&q->timeout, blk_mq_rq_timer, (unsigned long) q); |
1825 | blk_queue_rq_timeout(q, 30000); | |
1826 | ||
1827 | q->nr_queues = nr_cpu_ids; | |
24d2f903 | 1828 | q->nr_hw_queues = set->nr_hw_queues; |
f14bbe77 | 1829 | q->mq_map = map; |
320ae51f JA |
1830 | |
1831 | q->queue_ctx = ctx; | |
1832 | q->queue_hw_ctx = hctxs; | |
1833 | ||
24d2f903 | 1834 | q->mq_ops = set->ops; |
94eddfbe | 1835 | q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT; |
320ae51f | 1836 | |
05f1dd53 JA |
1837 | if (!(set->flags & BLK_MQ_F_SG_MERGE)) |
1838 | q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE; | |
1839 | ||
1be036e9 CH |
1840 | q->sg_reserved_size = INT_MAX; |
1841 | ||
6fca6a61 CH |
1842 | INIT_WORK(&q->requeue_work, blk_mq_requeue_work); |
1843 | INIT_LIST_HEAD(&q->requeue_list); | |
1844 | spin_lock_init(&q->requeue_lock); | |
1845 | ||
07068d5b JA |
1846 | if (q->nr_hw_queues > 1) |
1847 | blk_queue_make_request(q, blk_mq_make_request); | |
1848 | else | |
1849 | blk_queue_make_request(q, blk_sq_make_request); | |
1850 | ||
87ee7b11 | 1851 | blk_queue_rq_timed_out(q, blk_mq_rq_timed_out); |
24d2f903 CH |
1852 | if (set->timeout) |
1853 | blk_queue_rq_timeout(q, set->timeout); | |
320ae51f | 1854 | |
eba71768 JA |
1855 | /* |
1856 | * Do this after blk_queue_make_request() overrides it... | |
1857 | */ | |
1858 | q->nr_requests = set->queue_depth; | |
1859 | ||
24d2f903 CH |
1860 | if (set->ops->complete) |
1861 | blk_queue_softirq_done(q, set->ops->complete); | |
30a91cb4 | 1862 | |
320ae51f | 1863 | blk_mq_init_flush(q); |
24d2f903 | 1864 | blk_mq_init_cpu_queues(q, set->nr_hw_queues); |
320ae51f | 1865 | |
24d2f903 CH |
1866 | q->flush_rq = kzalloc(round_up(sizeof(struct request) + |
1867 | set->cmd_size, cache_line_size()), | |
1868 | GFP_KERNEL); | |
18741986 | 1869 | if (!q->flush_rq) |
320ae51f JA |
1870 | goto err_hw; |
1871 | ||
24d2f903 | 1872 | if (blk_mq_init_hw_queues(q, set)) |
18741986 CH |
1873 | goto err_flush_rq; |
1874 | ||
320ae51f JA |
1875 | mutex_lock(&all_q_mutex); |
1876 | list_add_tail(&q->all_q_node, &all_q_list); | |
1877 | mutex_unlock(&all_q_mutex); | |
1878 | ||
0d2602ca JA |
1879 | blk_mq_add_queue_tag_set(set, q); |
1880 | ||
484b4061 JA |
1881 | blk_mq_map_swqueue(q); |
1882 | ||
320ae51f | 1883 | return q; |
18741986 CH |
1884 | |
1885 | err_flush_rq: | |
1886 | kfree(q->flush_rq); | |
320ae51f | 1887 | err_hw: |
320ae51f JA |
1888 | blk_cleanup_queue(q); |
1889 | err_hctxs: | |
f14bbe77 | 1890 | kfree(map); |
24d2f903 | 1891 | for (i = 0; i < set->nr_hw_queues; i++) { |
320ae51f JA |
1892 | if (!hctxs[i]) |
1893 | break; | |
e4043dcf | 1894 | free_cpumask_var(hctxs[i]->cpumask); |
cdef54dd | 1895 | kfree(hctxs[i]); |
320ae51f | 1896 | } |
f14bbe77 | 1897 | err_map: |
320ae51f JA |
1898 | kfree(hctxs); |
1899 | err_percpu: | |
1900 | free_percpu(ctx); | |
1901 | return ERR_PTR(-ENOMEM); | |
1902 | } | |
1903 | EXPORT_SYMBOL(blk_mq_init_queue); | |
1904 | ||
1905 | void blk_mq_free_queue(struct request_queue *q) | |
1906 | { | |
624dbe47 | 1907 | struct blk_mq_tag_set *set = q->tag_set; |
320ae51f | 1908 | |
0d2602ca JA |
1909 | blk_mq_del_queue_tag_set(q); |
1910 | ||
624dbe47 ML |
1911 | blk_mq_exit_hw_queues(q, set, set->nr_hw_queues); |
1912 | blk_mq_free_hw_queues(q, set); | |
320ae51f | 1913 | |
3d2936f4 ML |
1914 | percpu_counter_destroy(&q->mq_usage_counter); |
1915 | ||
320ae51f JA |
1916 | free_percpu(q->queue_ctx); |
1917 | kfree(q->queue_hw_ctx); | |
1918 | kfree(q->mq_map); | |
1919 | ||
1920 | q->queue_ctx = NULL; | |
1921 | q->queue_hw_ctx = NULL; | |
1922 | q->mq_map = NULL; | |
1923 | ||
1924 | mutex_lock(&all_q_mutex); | |
1925 | list_del_init(&q->all_q_node); | |
1926 | mutex_unlock(&all_q_mutex); | |
1927 | } | |
320ae51f JA |
1928 | |
1929 | /* Basically redo blk_mq_init_queue with queue frozen */ | |
f618ef7c | 1930 | static void blk_mq_queue_reinit(struct request_queue *q) |
320ae51f JA |
1931 | { |
1932 | blk_mq_freeze_queue(q); | |
1933 | ||
67aec14c JA |
1934 | blk_mq_sysfs_unregister(q); |
1935 | ||
320ae51f JA |
1936 | blk_mq_update_queue_map(q->mq_map, q->nr_hw_queues); |
1937 | ||
1938 | /* | |
1939 | * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe | |
1940 | * we should change hctx numa_node according to new topology (this | |
1941 | * involves free and re-allocate memory, worthy doing?) | |
1942 | */ | |
1943 | ||
1944 | blk_mq_map_swqueue(q); | |
1945 | ||
67aec14c JA |
1946 | blk_mq_sysfs_register(q); |
1947 | ||
320ae51f JA |
1948 | blk_mq_unfreeze_queue(q); |
1949 | } | |
1950 | ||
f618ef7c PG |
1951 | static int blk_mq_queue_reinit_notify(struct notifier_block *nb, |
1952 | unsigned long action, void *hcpu) | |
320ae51f JA |
1953 | { |
1954 | struct request_queue *q; | |
1955 | ||
1956 | /* | |
9fccfed8 JA |
1957 | * Before new mappings are established, hotadded cpu might already |
1958 | * start handling requests. This doesn't break anything as we map | |
1959 | * offline CPUs to first hardware queue. We will re-init the queue | |
1960 | * below to get optimal settings. | |
320ae51f JA |
1961 | */ |
1962 | if (action != CPU_DEAD && action != CPU_DEAD_FROZEN && | |
1963 | action != CPU_ONLINE && action != CPU_ONLINE_FROZEN) | |
1964 | return NOTIFY_OK; | |
1965 | ||
1966 | mutex_lock(&all_q_mutex); | |
1967 | list_for_each_entry(q, &all_q_list, all_q_node) | |
1968 | blk_mq_queue_reinit(q); | |
1969 | mutex_unlock(&all_q_mutex); | |
1970 | return NOTIFY_OK; | |
1971 | } | |
1972 | ||
a4391c64 JA |
1973 | /* |
1974 | * Alloc a tag set to be associated with one or more request queues. | |
1975 | * May fail with EINVAL for various error conditions. May adjust the | |
1976 | * requested depth down, if if it too large. In that case, the set | |
1977 | * value will be stored in set->queue_depth. | |
1978 | */ | |
24d2f903 CH |
1979 | int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set) |
1980 | { | |
1981 | int i; | |
1982 | ||
1983 | if (!set->nr_hw_queues) | |
1984 | return -EINVAL; | |
a4391c64 | 1985 | if (!set->queue_depth) |
24d2f903 CH |
1986 | return -EINVAL; |
1987 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) | |
1988 | return -EINVAL; | |
1989 | ||
cdef54dd | 1990 | if (!set->nr_hw_queues || !set->ops->queue_rq || !set->ops->map_queue) |
24d2f903 CH |
1991 | return -EINVAL; |
1992 | ||
a4391c64 JA |
1993 | if (set->queue_depth > BLK_MQ_MAX_DEPTH) { |
1994 | pr_info("blk-mq: reduced tag depth to %u\n", | |
1995 | BLK_MQ_MAX_DEPTH); | |
1996 | set->queue_depth = BLK_MQ_MAX_DEPTH; | |
1997 | } | |
24d2f903 | 1998 | |
48479005 ML |
1999 | set->tags = kmalloc_node(set->nr_hw_queues * |
2000 | sizeof(struct blk_mq_tags *), | |
24d2f903 CH |
2001 | GFP_KERNEL, set->numa_node); |
2002 | if (!set->tags) | |
2003 | goto out; | |
2004 | ||
2005 | for (i = 0; i < set->nr_hw_queues; i++) { | |
2006 | set->tags[i] = blk_mq_init_rq_map(set, i); | |
2007 | if (!set->tags[i]) | |
2008 | goto out_unwind; | |
2009 | } | |
2010 | ||
0d2602ca JA |
2011 | mutex_init(&set->tag_list_lock); |
2012 | INIT_LIST_HEAD(&set->tag_list); | |
2013 | ||
24d2f903 CH |
2014 | return 0; |
2015 | ||
2016 | out_unwind: | |
2017 | while (--i >= 0) | |
2018 | blk_mq_free_rq_map(set, set->tags[i], i); | |
2019 | out: | |
2020 | return -ENOMEM; | |
2021 | } | |
2022 | EXPORT_SYMBOL(blk_mq_alloc_tag_set); | |
2023 | ||
2024 | void blk_mq_free_tag_set(struct blk_mq_tag_set *set) | |
2025 | { | |
2026 | int i; | |
2027 | ||
484b4061 JA |
2028 | for (i = 0; i < set->nr_hw_queues; i++) { |
2029 | if (set->tags[i]) | |
2030 | blk_mq_free_rq_map(set, set->tags[i], i); | |
2031 | } | |
2032 | ||
981bd189 | 2033 | kfree(set->tags); |
24d2f903 CH |
2034 | } |
2035 | EXPORT_SYMBOL(blk_mq_free_tag_set); | |
2036 | ||
e3a2b3f9 JA |
2037 | int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr) |
2038 | { | |
2039 | struct blk_mq_tag_set *set = q->tag_set; | |
2040 | struct blk_mq_hw_ctx *hctx; | |
2041 | int i, ret; | |
2042 | ||
2043 | if (!set || nr > set->queue_depth) | |
2044 | return -EINVAL; | |
2045 | ||
2046 | ret = 0; | |
2047 | queue_for_each_hw_ctx(q, hctx, i) { | |
2048 | ret = blk_mq_tag_update_depth(hctx->tags, nr); | |
2049 | if (ret) | |
2050 | break; | |
2051 | } | |
2052 | ||
2053 | if (!ret) | |
2054 | q->nr_requests = nr; | |
2055 | ||
2056 | return ret; | |
2057 | } | |
2058 | ||
676141e4 JA |
2059 | void blk_mq_disable_hotplug(void) |
2060 | { | |
2061 | mutex_lock(&all_q_mutex); | |
2062 | } | |
2063 | ||
2064 | void blk_mq_enable_hotplug(void) | |
2065 | { | |
2066 | mutex_unlock(&all_q_mutex); | |
2067 | } | |
2068 | ||
320ae51f JA |
2069 | static int __init blk_mq_init(void) |
2070 | { | |
320ae51f JA |
2071 | blk_mq_cpu_init(); |
2072 | ||
2073 | /* Must be called after percpu_counter_hotcpu_callback() */ | |
2074 | hotcpu_notifier(blk_mq_queue_reinit_notify, -10); | |
2075 | ||
2076 | return 0; | |
2077 | } | |
2078 | subsys_initcall(blk_mq_init); |