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> | |
f75782e4 | 12 | #include <linux/kmemleak.h> |
320ae51f JA |
13 | #include <linux/mm.h> |
14 | #include <linux/init.h> | |
15 | #include <linux/slab.h> | |
16 | #include <linux/workqueue.h> | |
17 | #include <linux/smp.h> | |
18 | #include <linux/llist.h> | |
19 | #include <linux/list_sort.h> | |
20 | #include <linux/cpu.h> | |
21 | #include <linux/cache.h> | |
22 | #include <linux/sched/sysctl.h> | |
23 | #include <linux/delay.h> | |
aedcd72f | 24 | #include <linux/crash_dump.h> |
88c7b2b7 | 25 | #include <linux/prefetch.h> |
320ae51f JA |
26 | |
27 | #include <trace/events/block.h> | |
28 | ||
29 | #include <linux/blk-mq.h> | |
30 | #include "blk.h" | |
31 | #include "blk-mq.h" | |
32 | #include "blk-mq-tag.h" | |
cf43e6be | 33 | #include "blk-stat.h" |
87760e5e | 34 | #include "blk-wbt.h" |
320ae51f JA |
35 | |
36 | static DEFINE_MUTEX(all_q_mutex); | |
37 | static LIST_HEAD(all_q_list); | |
38 | ||
320ae51f JA |
39 | /* |
40 | * Check if any of the ctx's have pending work in this hardware queue | |
41 | */ | |
42 | static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx) | |
43 | { | |
88459642 | 44 | return sbitmap_any_bit_set(&hctx->ctx_map); |
1429d7c9 JA |
45 | } |
46 | ||
320ae51f JA |
47 | /* |
48 | * Mark this ctx as having pending work in this hardware queue | |
49 | */ | |
50 | static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx, | |
51 | struct blk_mq_ctx *ctx) | |
52 | { | |
88459642 OS |
53 | if (!sbitmap_test_bit(&hctx->ctx_map, ctx->index_hw)) |
54 | sbitmap_set_bit(&hctx->ctx_map, ctx->index_hw); | |
1429d7c9 JA |
55 | } |
56 | ||
57 | static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx, | |
58 | struct blk_mq_ctx *ctx) | |
59 | { | |
88459642 | 60 | sbitmap_clear_bit(&hctx->ctx_map, ctx->index_hw); |
320ae51f JA |
61 | } |
62 | ||
b4c6a028 | 63 | void blk_mq_freeze_queue_start(struct request_queue *q) |
43a5e4e2 | 64 | { |
4ecd4fef | 65 | int freeze_depth; |
cddd5d17 | 66 | |
4ecd4fef CH |
67 | freeze_depth = atomic_inc_return(&q->mq_freeze_depth); |
68 | if (freeze_depth == 1) { | |
3ef28e83 | 69 | percpu_ref_kill(&q->q_usage_counter); |
b94ec296 | 70 | blk_mq_run_hw_queues(q, false); |
cddd5d17 | 71 | } |
f3af020b | 72 | } |
b4c6a028 | 73 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_start); |
f3af020b TH |
74 | |
75 | static void blk_mq_freeze_queue_wait(struct request_queue *q) | |
76 | { | |
3ef28e83 | 77 | wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter)); |
43a5e4e2 ML |
78 | } |
79 | ||
f3af020b TH |
80 | /* |
81 | * Guarantee no request is in use, so we can change any data structure of | |
82 | * the queue afterward. | |
83 | */ | |
3ef28e83 | 84 | void blk_freeze_queue(struct request_queue *q) |
f3af020b | 85 | { |
3ef28e83 DW |
86 | /* |
87 | * In the !blk_mq case we are only calling this to kill the | |
88 | * q_usage_counter, otherwise this increases the freeze depth | |
89 | * and waits for it to return to zero. For this reason there is | |
90 | * no blk_unfreeze_queue(), and blk_freeze_queue() is not | |
91 | * exported to drivers as the only user for unfreeze is blk_mq. | |
92 | */ | |
f3af020b TH |
93 | blk_mq_freeze_queue_start(q); |
94 | blk_mq_freeze_queue_wait(q); | |
95 | } | |
3ef28e83 DW |
96 | |
97 | void blk_mq_freeze_queue(struct request_queue *q) | |
98 | { | |
99 | /* | |
100 | * ...just an alias to keep freeze and unfreeze actions balanced | |
101 | * in the blk_mq_* namespace | |
102 | */ | |
103 | blk_freeze_queue(q); | |
104 | } | |
c761d96b | 105 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue); |
f3af020b | 106 | |
b4c6a028 | 107 | void blk_mq_unfreeze_queue(struct request_queue *q) |
320ae51f | 108 | { |
4ecd4fef | 109 | int freeze_depth; |
320ae51f | 110 | |
4ecd4fef CH |
111 | freeze_depth = atomic_dec_return(&q->mq_freeze_depth); |
112 | WARN_ON_ONCE(freeze_depth < 0); | |
113 | if (!freeze_depth) { | |
3ef28e83 | 114 | percpu_ref_reinit(&q->q_usage_counter); |
320ae51f | 115 | wake_up_all(&q->mq_freeze_wq); |
add703fd | 116 | } |
320ae51f | 117 | } |
b4c6a028 | 118 | EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue); |
320ae51f | 119 | |
6a83e74d BVA |
120 | /** |
121 | * blk_mq_quiesce_queue() - wait until all ongoing queue_rq calls have finished | |
122 | * @q: request queue. | |
123 | * | |
124 | * Note: this function does not prevent that the struct request end_io() | |
125 | * callback function is invoked. Additionally, it is not prevented that | |
126 | * new queue_rq() calls occur unless the queue has been stopped first. | |
127 | */ | |
128 | void blk_mq_quiesce_queue(struct request_queue *q) | |
129 | { | |
130 | struct blk_mq_hw_ctx *hctx; | |
131 | unsigned int i; | |
132 | bool rcu = false; | |
133 | ||
134 | blk_mq_stop_hw_queues(q); | |
135 | ||
136 | queue_for_each_hw_ctx(q, hctx, i) { | |
137 | if (hctx->flags & BLK_MQ_F_BLOCKING) | |
138 | synchronize_srcu(&hctx->queue_rq_srcu); | |
139 | else | |
140 | rcu = true; | |
141 | } | |
142 | if (rcu) | |
143 | synchronize_rcu(); | |
144 | } | |
145 | EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue); | |
146 | ||
aed3ea94 JA |
147 | void blk_mq_wake_waiters(struct request_queue *q) |
148 | { | |
149 | struct blk_mq_hw_ctx *hctx; | |
150 | unsigned int i; | |
151 | ||
152 | queue_for_each_hw_ctx(q, hctx, i) | |
153 | if (blk_mq_hw_queue_mapped(hctx)) | |
154 | blk_mq_tag_wakeup_all(hctx->tags, true); | |
3fd5940c KB |
155 | |
156 | /* | |
157 | * If we are called because the queue has now been marked as | |
158 | * dying, we need to ensure that processes currently waiting on | |
159 | * the queue are notified as well. | |
160 | */ | |
161 | wake_up_all(&q->mq_freeze_wq); | |
aed3ea94 JA |
162 | } |
163 | ||
320ae51f JA |
164 | bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx) |
165 | { | |
166 | return blk_mq_has_free_tags(hctx->tags); | |
167 | } | |
168 | EXPORT_SYMBOL(blk_mq_can_queue); | |
169 | ||
2c3ad667 JA |
170 | void blk_mq_rq_ctx_init(struct request_queue *q, struct blk_mq_ctx *ctx, |
171 | struct request *rq, unsigned int op) | |
320ae51f | 172 | { |
af76e555 CH |
173 | INIT_LIST_HEAD(&rq->queuelist); |
174 | /* csd/requeue_work/fifo_time is initialized before use */ | |
175 | rq->q = q; | |
320ae51f | 176 | rq->mq_ctx = ctx; |
ef295ecf | 177 | rq->cmd_flags = op; |
e8064021 CH |
178 | if (blk_queue_io_stat(q)) |
179 | rq->rq_flags |= RQF_IO_STAT; | |
af76e555 CH |
180 | /* do not touch atomic flags, it needs atomic ops against the timer */ |
181 | rq->cpu = -1; | |
af76e555 CH |
182 | INIT_HLIST_NODE(&rq->hash); |
183 | RB_CLEAR_NODE(&rq->rb_node); | |
af76e555 CH |
184 | rq->rq_disk = NULL; |
185 | rq->part = NULL; | |
3ee32372 | 186 | rq->start_time = jiffies; |
af76e555 CH |
187 | #ifdef CONFIG_BLK_CGROUP |
188 | rq->rl = NULL; | |
0fec08b4 | 189 | set_start_time_ns(rq); |
af76e555 CH |
190 | rq->io_start_time_ns = 0; |
191 | #endif | |
192 | rq->nr_phys_segments = 0; | |
193 | #if defined(CONFIG_BLK_DEV_INTEGRITY) | |
194 | rq->nr_integrity_segments = 0; | |
195 | #endif | |
af76e555 CH |
196 | rq->special = NULL; |
197 | /* tag was already set */ | |
198 | rq->errors = 0; | |
af76e555 | 199 | |
6f4a1626 TB |
200 | rq->cmd = rq->__cmd; |
201 | ||
af76e555 CH |
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 | ||
ef295ecf | 214 | ctx->rq_dispatched[op_is_sync(op)]++; |
320ae51f | 215 | } |
2c3ad667 | 216 | EXPORT_SYMBOL_GPL(blk_mq_rq_ctx_init); |
320ae51f | 217 | |
2c3ad667 JA |
218 | struct request *__blk_mq_alloc_request(struct blk_mq_alloc_data *data, |
219 | unsigned int op) | |
5dee8577 CH |
220 | { |
221 | struct request *rq; | |
222 | unsigned int tag; | |
223 | ||
cb96a42c | 224 | tag = blk_mq_get_tag(data); |
5dee8577 | 225 | if (tag != BLK_MQ_TAG_FAIL) { |
cb96a42c | 226 | rq = data->hctx->tags->rqs[tag]; |
5dee8577 | 227 | |
cb96a42c | 228 | if (blk_mq_tag_busy(data->hctx)) { |
e8064021 | 229 | rq->rq_flags = RQF_MQ_INFLIGHT; |
cb96a42c | 230 | atomic_inc(&data->hctx->nr_active); |
5dee8577 CH |
231 | } |
232 | ||
233 | rq->tag = tag; | |
ef295ecf | 234 | blk_mq_rq_ctx_init(data->q, data->ctx, rq, op); |
5dee8577 CH |
235 | return rq; |
236 | } | |
237 | ||
238 | return NULL; | |
239 | } | |
2c3ad667 | 240 | EXPORT_SYMBOL_GPL(__blk_mq_alloc_request); |
5dee8577 | 241 | |
6f3b0e8b CH |
242 | struct request *blk_mq_alloc_request(struct request_queue *q, int rw, |
243 | unsigned int flags) | |
320ae51f | 244 | { |
d852564f CH |
245 | struct blk_mq_ctx *ctx; |
246 | struct blk_mq_hw_ctx *hctx; | |
320ae51f | 247 | struct request *rq; |
cb96a42c | 248 | struct blk_mq_alloc_data alloc_data; |
a492f075 | 249 | int ret; |
320ae51f | 250 | |
6f3b0e8b | 251 | ret = blk_queue_enter(q, flags & BLK_MQ_REQ_NOWAIT); |
a492f075 JL |
252 | if (ret) |
253 | return ERR_PTR(ret); | |
320ae51f | 254 | |
d852564f | 255 | ctx = blk_mq_get_ctx(q); |
7d7e0f90 | 256 | hctx = blk_mq_map_queue(q, ctx->cpu); |
6f3b0e8b | 257 | blk_mq_set_alloc_data(&alloc_data, q, flags, ctx, hctx); |
ef295ecf | 258 | rq = __blk_mq_alloc_request(&alloc_data, rw); |
d852564f | 259 | blk_mq_put_ctx(ctx); |
841bac2c | 260 | |
c76541a9 | 261 | if (!rq) { |
3ef28e83 | 262 | blk_queue_exit(q); |
a492f075 | 263 | return ERR_PTR(-EWOULDBLOCK); |
c76541a9 | 264 | } |
0c4de0f3 CH |
265 | |
266 | rq->__data_len = 0; | |
267 | rq->__sector = (sector_t) -1; | |
268 | rq->bio = rq->biotail = NULL; | |
320ae51f JA |
269 | return rq; |
270 | } | |
4bb659b1 | 271 | EXPORT_SYMBOL(blk_mq_alloc_request); |
320ae51f | 272 | |
1f5bd336 ML |
273 | struct request *blk_mq_alloc_request_hctx(struct request_queue *q, int rw, |
274 | unsigned int flags, unsigned int hctx_idx) | |
275 | { | |
276 | struct blk_mq_hw_ctx *hctx; | |
277 | struct blk_mq_ctx *ctx; | |
278 | struct request *rq; | |
279 | struct blk_mq_alloc_data alloc_data; | |
280 | int ret; | |
281 | ||
282 | /* | |
283 | * If the tag allocator sleeps we could get an allocation for a | |
284 | * different hardware context. No need to complicate the low level | |
285 | * allocator for this for the rare use case of a command tied to | |
286 | * a specific queue. | |
287 | */ | |
288 | if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT))) | |
289 | return ERR_PTR(-EINVAL); | |
290 | ||
291 | if (hctx_idx >= q->nr_hw_queues) | |
292 | return ERR_PTR(-EIO); | |
293 | ||
294 | ret = blk_queue_enter(q, true); | |
295 | if (ret) | |
296 | return ERR_PTR(ret); | |
297 | ||
c8712c6a CH |
298 | /* |
299 | * Check if the hardware context is actually mapped to anything. | |
300 | * If not tell the caller that it should skip this queue. | |
301 | */ | |
1f5bd336 | 302 | hctx = q->queue_hw_ctx[hctx_idx]; |
c8712c6a CH |
303 | if (!blk_mq_hw_queue_mapped(hctx)) { |
304 | ret = -EXDEV; | |
305 | goto out_queue_exit; | |
306 | } | |
1f5bd336 ML |
307 | ctx = __blk_mq_get_ctx(q, cpumask_first(hctx->cpumask)); |
308 | ||
309 | blk_mq_set_alloc_data(&alloc_data, q, flags, ctx, hctx); | |
ef295ecf | 310 | rq = __blk_mq_alloc_request(&alloc_data, rw); |
1f5bd336 | 311 | if (!rq) { |
c8712c6a CH |
312 | ret = -EWOULDBLOCK; |
313 | goto out_queue_exit; | |
1f5bd336 ML |
314 | } |
315 | ||
316 | return rq; | |
c8712c6a CH |
317 | |
318 | out_queue_exit: | |
319 | blk_queue_exit(q); | |
320 | return ERR_PTR(ret); | |
1f5bd336 ML |
321 | } |
322 | EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx); | |
323 | ||
2c3ad667 JA |
324 | void __blk_mq_free_request(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx, |
325 | struct request *rq) | |
320ae51f JA |
326 | { |
327 | const int tag = rq->tag; | |
328 | struct request_queue *q = rq->q; | |
329 | ||
e8064021 | 330 | if (rq->rq_flags & RQF_MQ_INFLIGHT) |
0d2602ca | 331 | atomic_dec(&hctx->nr_active); |
87760e5e JA |
332 | |
333 | wbt_done(q->rq_wb, &rq->issue_stat); | |
e8064021 | 334 | rq->rq_flags = 0; |
0d2602ca | 335 | |
af76e555 | 336 | clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags); |
06426adf | 337 | clear_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags); |
4941115b | 338 | blk_mq_put_tag(hctx, hctx->tags, ctx, tag); |
3ef28e83 | 339 | blk_queue_exit(q); |
320ae51f JA |
340 | } |
341 | ||
16a3c2a7 JA |
342 | static void blk_mq_free_hctx_request(struct blk_mq_hw_ctx *hctx, |
343 | struct request *rq) | |
320ae51f JA |
344 | { |
345 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
320ae51f JA |
346 | |
347 | ctx->rq_completed[rq_is_sync(rq)]++; | |
320ae51f | 348 | __blk_mq_free_request(hctx, ctx, rq); |
7c7f2f2b | 349 | } |
7c7f2f2b JA |
350 | |
351 | void blk_mq_free_request(struct request *rq) | |
352 | { | |
7d7e0f90 | 353 | blk_mq_free_hctx_request(blk_mq_map_queue(rq->q, rq->mq_ctx->cpu), rq); |
320ae51f | 354 | } |
1a3b595a | 355 | EXPORT_SYMBOL_GPL(blk_mq_free_request); |
320ae51f | 356 | |
c8a446ad | 357 | inline void __blk_mq_end_request(struct request *rq, int error) |
320ae51f | 358 | { |
0d11e6ac ML |
359 | blk_account_io_done(rq); |
360 | ||
91b63639 | 361 | if (rq->end_io) { |
87760e5e | 362 | wbt_done(rq->q->rq_wb, &rq->issue_stat); |
320ae51f | 363 | rq->end_io(rq, error); |
91b63639 CH |
364 | } else { |
365 | if (unlikely(blk_bidi_rq(rq))) | |
366 | blk_mq_free_request(rq->next_rq); | |
320ae51f | 367 | blk_mq_free_request(rq); |
91b63639 | 368 | } |
320ae51f | 369 | } |
c8a446ad | 370 | EXPORT_SYMBOL(__blk_mq_end_request); |
63151a44 | 371 | |
c8a446ad | 372 | void blk_mq_end_request(struct request *rq, int error) |
63151a44 CH |
373 | { |
374 | if (blk_update_request(rq, error, blk_rq_bytes(rq))) | |
375 | BUG(); | |
c8a446ad | 376 | __blk_mq_end_request(rq, error); |
63151a44 | 377 | } |
c8a446ad | 378 | EXPORT_SYMBOL(blk_mq_end_request); |
320ae51f | 379 | |
30a91cb4 | 380 | static void __blk_mq_complete_request_remote(void *data) |
320ae51f | 381 | { |
3d6efbf6 | 382 | struct request *rq = data; |
320ae51f | 383 | |
30a91cb4 | 384 | rq->q->softirq_done_fn(rq); |
320ae51f | 385 | } |
320ae51f | 386 | |
ed851860 | 387 | static void blk_mq_ipi_complete_request(struct request *rq) |
320ae51f JA |
388 | { |
389 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
38535201 | 390 | bool shared = false; |
320ae51f JA |
391 | int cpu; |
392 | ||
38535201 | 393 | if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) { |
30a91cb4 CH |
394 | rq->q->softirq_done_fn(rq); |
395 | return; | |
396 | } | |
320ae51f JA |
397 | |
398 | cpu = get_cpu(); | |
38535201 CH |
399 | if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags)) |
400 | shared = cpus_share_cache(cpu, ctx->cpu); | |
401 | ||
402 | if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) { | |
30a91cb4 | 403 | rq->csd.func = __blk_mq_complete_request_remote; |
3d6efbf6 CH |
404 | rq->csd.info = rq; |
405 | rq->csd.flags = 0; | |
c46fff2a | 406 | smp_call_function_single_async(ctx->cpu, &rq->csd); |
3d6efbf6 | 407 | } else { |
30a91cb4 | 408 | rq->q->softirq_done_fn(rq); |
3d6efbf6 | 409 | } |
320ae51f JA |
410 | put_cpu(); |
411 | } | |
30a91cb4 | 412 | |
cf43e6be JA |
413 | static void blk_mq_stat_add(struct request *rq) |
414 | { | |
415 | if (rq->rq_flags & RQF_STATS) { | |
416 | /* | |
417 | * We could rq->mq_ctx here, but there's less of a risk | |
418 | * of races if we have the completion event add the stats | |
419 | * to the local software queue. | |
420 | */ | |
421 | struct blk_mq_ctx *ctx; | |
422 | ||
423 | ctx = __blk_mq_get_ctx(rq->q, raw_smp_processor_id()); | |
424 | blk_stat_add(&ctx->stat[rq_data_dir(rq)], rq); | |
425 | } | |
426 | } | |
427 | ||
1fa8cc52 | 428 | static void __blk_mq_complete_request(struct request *rq) |
ed851860 JA |
429 | { |
430 | struct request_queue *q = rq->q; | |
431 | ||
cf43e6be JA |
432 | blk_mq_stat_add(rq); |
433 | ||
ed851860 | 434 | if (!q->softirq_done_fn) |
c8a446ad | 435 | blk_mq_end_request(rq, rq->errors); |
ed851860 JA |
436 | else |
437 | blk_mq_ipi_complete_request(rq); | |
438 | } | |
439 | ||
30a91cb4 CH |
440 | /** |
441 | * blk_mq_complete_request - end I/O on a request | |
442 | * @rq: the request being processed | |
443 | * | |
444 | * Description: | |
445 | * Ends all I/O on a request. It does not handle partial completions. | |
446 | * The actual completion happens out-of-order, through a IPI handler. | |
447 | **/ | |
f4829a9b | 448 | void blk_mq_complete_request(struct request *rq, int error) |
30a91cb4 | 449 | { |
95f09684 JA |
450 | struct request_queue *q = rq->q; |
451 | ||
452 | if (unlikely(blk_should_fake_timeout(q))) | |
30a91cb4 | 453 | return; |
f4829a9b CH |
454 | if (!blk_mark_rq_complete(rq)) { |
455 | rq->errors = error; | |
ed851860 | 456 | __blk_mq_complete_request(rq); |
f4829a9b | 457 | } |
30a91cb4 CH |
458 | } |
459 | EXPORT_SYMBOL(blk_mq_complete_request); | |
320ae51f | 460 | |
973c0191 KB |
461 | int blk_mq_request_started(struct request *rq) |
462 | { | |
463 | return test_bit(REQ_ATOM_STARTED, &rq->atomic_flags); | |
464 | } | |
465 | EXPORT_SYMBOL_GPL(blk_mq_request_started); | |
466 | ||
e2490073 | 467 | void blk_mq_start_request(struct request *rq) |
320ae51f JA |
468 | { |
469 | struct request_queue *q = rq->q; | |
470 | ||
471 | trace_block_rq_issue(q, rq); | |
472 | ||
742ee69b | 473 | rq->resid_len = blk_rq_bytes(rq); |
91b63639 CH |
474 | if (unlikely(blk_bidi_rq(rq))) |
475 | rq->next_rq->resid_len = blk_rq_bytes(rq->next_rq); | |
742ee69b | 476 | |
cf43e6be JA |
477 | if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) { |
478 | blk_stat_set_issue_time(&rq->issue_stat); | |
479 | rq->rq_flags |= RQF_STATS; | |
87760e5e | 480 | wbt_issue(q->rq_wb, &rq->issue_stat); |
cf43e6be JA |
481 | } |
482 | ||
2b8393b4 | 483 | blk_add_timer(rq); |
87ee7b11 | 484 | |
538b7534 JA |
485 | /* |
486 | * Ensure that ->deadline is visible before set the started | |
487 | * flag and clear the completed flag. | |
488 | */ | |
489 | smp_mb__before_atomic(); | |
490 | ||
87ee7b11 JA |
491 | /* |
492 | * Mark us as started and clear complete. Complete might have been | |
493 | * set if requeue raced with timeout, which then marked it as | |
494 | * complete. So be sure to clear complete again when we start | |
495 | * the request, otherwise we'll ignore the completion event. | |
496 | */ | |
4b570521 JA |
497 | if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) |
498 | set_bit(REQ_ATOM_STARTED, &rq->atomic_flags); | |
499 | if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags)) | |
500 | clear_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags); | |
49f5baa5 CH |
501 | |
502 | if (q->dma_drain_size && blk_rq_bytes(rq)) { | |
503 | /* | |
504 | * Make sure space for the drain appears. We know we can do | |
505 | * this because max_hw_segments has been adjusted to be one | |
506 | * fewer than the device can handle. | |
507 | */ | |
508 | rq->nr_phys_segments++; | |
509 | } | |
320ae51f | 510 | } |
e2490073 | 511 | EXPORT_SYMBOL(blk_mq_start_request); |
320ae51f | 512 | |
ed0791b2 | 513 | static void __blk_mq_requeue_request(struct request *rq) |
320ae51f JA |
514 | { |
515 | struct request_queue *q = rq->q; | |
516 | ||
517 | trace_block_rq_requeue(q, rq); | |
87760e5e | 518 | wbt_requeue(q->rq_wb, &rq->issue_stat); |
49f5baa5 | 519 | |
e2490073 CH |
520 | if (test_and_clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) { |
521 | if (q->dma_drain_size && blk_rq_bytes(rq)) | |
522 | rq->nr_phys_segments--; | |
523 | } | |
320ae51f JA |
524 | } |
525 | ||
2b053aca | 526 | void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list) |
ed0791b2 | 527 | { |
ed0791b2 | 528 | __blk_mq_requeue_request(rq); |
ed0791b2 | 529 | |
ed0791b2 | 530 | BUG_ON(blk_queued_rq(rq)); |
2b053aca | 531 | blk_mq_add_to_requeue_list(rq, true, kick_requeue_list); |
ed0791b2 CH |
532 | } |
533 | EXPORT_SYMBOL(blk_mq_requeue_request); | |
534 | ||
6fca6a61 CH |
535 | static void blk_mq_requeue_work(struct work_struct *work) |
536 | { | |
537 | struct request_queue *q = | |
2849450a | 538 | container_of(work, struct request_queue, requeue_work.work); |
6fca6a61 CH |
539 | LIST_HEAD(rq_list); |
540 | struct request *rq, *next; | |
541 | unsigned long flags; | |
542 | ||
543 | spin_lock_irqsave(&q->requeue_lock, flags); | |
544 | list_splice_init(&q->requeue_list, &rq_list); | |
545 | spin_unlock_irqrestore(&q->requeue_lock, flags); | |
546 | ||
547 | list_for_each_entry_safe(rq, next, &rq_list, queuelist) { | |
e8064021 | 548 | if (!(rq->rq_flags & RQF_SOFTBARRIER)) |
6fca6a61 CH |
549 | continue; |
550 | ||
e8064021 | 551 | rq->rq_flags &= ~RQF_SOFTBARRIER; |
6fca6a61 CH |
552 | list_del_init(&rq->queuelist); |
553 | blk_mq_insert_request(rq, true, false, false); | |
554 | } | |
555 | ||
556 | while (!list_empty(&rq_list)) { | |
557 | rq = list_entry(rq_list.next, struct request, queuelist); | |
558 | list_del_init(&rq->queuelist); | |
559 | blk_mq_insert_request(rq, false, false, false); | |
560 | } | |
561 | ||
52d7f1b5 | 562 | blk_mq_run_hw_queues(q, false); |
6fca6a61 CH |
563 | } |
564 | ||
2b053aca BVA |
565 | void blk_mq_add_to_requeue_list(struct request *rq, bool at_head, |
566 | bool kick_requeue_list) | |
6fca6a61 CH |
567 | { |
568 | struct request_queue *q = rq->q; | |
569 | unsigned long flags; | |
570 | ||
571 | /* | |
572 | * We abuse this flag that is otherwise used by the I/O scheduler to | |
573 | * request head insertation from the workqueue. | |
574 | */ | |
e8064021 | 575 | BUG_ON(rq->rq_flags & RQF_SOFTBARRIER); |
6fca6a61 CH |
576 | |
577 | spin_lock_irqsave(&q->requeue_lock, flags); | |
578 | if (at_head) { | |
e8064021 | 579 | rq->rq_flags |= RQF_SOFTBARRIER; |
6fca6a61 CH |
580 | list_add(&rq->queuelist, &q->requeue_list); |
581 | } else { | |
582 | list_add_tail(&rq->queuelist, &q->requeue_list); | |
583 | } | |
584 | spin_unlock_irqrestore(&q->requeue_lock, flags); | |
2b053aca BVA |
585 | |
586 | if (kick_requeue_list) | |
587 | blk_mq_kick_requeue_list(q); | |
6fca6a61 CH |
588 | } |
589 | EXPORT_SYMBOL(blk_mq_add_to_requeue_list); | |
590 | ||
591 | void blk_mq_kick_requeue_list(struct request_queue *q) | |
592 | { | |
2849450a | 593 | kblockd_schedule_delayed_work(&q->requeue_work, 0); |
6fca6a61 CH |
594 | } |
595 | EXPORT_SYMBOL(blk_mq_kick_requeue_list); | |
596 | ||
2849450a MS |
597 | void blk_mq_delay_kick_requeue_list(struct request_queue *q, |
598 | unsigned long msecs) | |
599 | { | |
600 | kblockd_schedule_delayed_work(&q->requeue_work, | |
601 | msecs_to_jiffies(msecs)); | |
602 | } | |
603 | EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list); | |
604 | ||
1885b24d JA |
605 | void blk_mq_abort_requeue_list(struct request_queue *q) |
606 | { | |
607 | unsigned long flags; | |
608 | LIST_HEAD(rq_list); | |
609 | ||
610 | spin_lock_irqsave(&q->requeue_lock, flags); | |
611 | list_splice_init(&q->requeue_list, &rq_list); | |
612 | spin_unlock_irqrestore(&q->requeue_lock, flags); | |
613 | ||
614 | while (!list_empty(&rq_list)) { | |
615 | struct request *rq; | |
616 | ||
617 | rq = list_first_entry(&rq_list, struct request, queuelist); | |
618 | list_del_init(&rq->queuelist); | |
619 | rq->errors = -EIO; | |
620 | blk_mq_end_request(rq, rq->errors); | |
621 | } | |
622 | } | |
623 | EXPORT_SYMBOL(blk_mq_abort_requeue_list); | |
624 | ||
0e62f51f JA |
625 | struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag) |
626 | { | |
88c7b2b7 JA |
627 | if (tag < tags->nr_tags) { |
628 | prefetch(tags->rqs[tag]); | |
4ee86bab | 629 | return tags->rqs[tag]; |
88c7b2b7 | 630 | } |
4ee86bab HR |
631 | |
632 | return NULL; | |
24d2f903 CH |
633 | } |
634 | EXPORT_SYMBOL(blk_mq_tag_to_rq); | |
635 | ||
320ae51f | 636 | struct blk_mq_timeout_data { |
46f92d42 CH |
637 | unsigned long next; |
638 | unsigned int next_set; | |
320ae51f JA |
639 | }; |
640 | ||
90415837 | 641 | void blk_mq_rq_timed_out(struct request *req, bool reserved) |
320ae51f | 642 | { |
f8a5b122 | 643 | const struct blk_mq_ops *ops = req->q->mq_ops; |
46f92d42 | 644 | enum blk_eh_timer_return ret = BLK_EH_RESET_TIMER; |
87ee7b11 JA |
645 | |
646 | /* | |
647 | * We know that complete is set at this point. If STARTED isn't set | |
648 | * anymore, then the request isn't active and the "timeout" should | |
649 | * just be ignored. This can happen due to the bitflag ordering. | |
650 | * Timeout first checks if STARTED is set, and if it is, assumes | |
651 | * the request is active. But if we race with completion, then | |
652 | * we both flags will get cleared. So check here again, and ignore | |
653 | * a timeout event with a request that isn't active. | |
654 | */ | |
46f92d42 CH |
655 | if (!test_bit(REQ_ATOM_STARTED, &req->atomic_flags)) |
656 | return; | |
87ee7b11 | 657 | |
46f92d42 | 658 | if (ops->timeout) |
0152fb6b | 659 | ret = ops->timeout(req, reserved); |
46f92d42 CH |
660 | |
661 | switch (ret) { | |
662 | case BLK_EH_HANDLED: | |
663 | __blk_mq_complete_request(req); | |
664 | break; | |
665 | case BLK_EH_RESET_TIMER: | |
666 | blk_add_timer(req); | |
667 | blk_clear_rq_complete(req); | |
668 | break; | |
669 | case BLK_EH_NOT_HANDLED: | |
670 | break; | |
671 | default: | |
672 | printk(KERN_ERR "block: bad eh return: %d\n", ret); | |
673 | break; | |
674 | } | |
87ee7b11 | 675 | } |
5b3f25fc | 676 | |
81481eb4 CH |
677 | static void blk_mq_check_expired(struct blk_mq_hw_ctx *hctx, |
678 | struct request *rq, void *priv, bool reserved) | |
679 | { | |
680 | struct blk_mq_timeout_data *data = priv; | |
87ee7b11 | 681 | |
eb130dbf KB |
682 | if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) { |
683 | /* | |
684 | * If a request wasn't started before the queue was | |
685 | * marked dying, kill it here or it'll go unnoticed. | |
686 | */ | |
a59e0f57 KB |
687 | if (unlikely(blk_queue_dying(rq->q))) { |
688 | rq->errors = -EIO; | |
689 | blk_mq_end_request(rq, rq->errors); | |
690 | } | |
46f92d42 | 691 | return; |
eb130dbf | 692 | } |
87ee7b11 | 693 | |
46f92d42 CH |
694 | if (time_after_eq(jiffies, rq->deadline)) { |
695 | if (!blk_mark_rq_complete(rq)) | |
0152fb6b | 696 | blk_mq_rq_timed_out(rq, reserved); |
46f92d42 CH |
697 | } else if (!data->next_set || time_after(data->next, rq->deadline)) { |
698 | data->next = rq->deadline; | |
699 | data->next_set = 1; | |
700 | } | |
87ee7b11 JA |
701 | } |
702 | ||
287922eb | 703 | static void blk_mq_timeout_work(struct work_struct *work) |
320ae51f | 704 | { |
287922eb CH |
705 | struct request_queue *q = |
706 | container_of(work, struct request_queue, timeout_work); | |
81481eb4 CH |
707 | struct blk_mq_timeout_data data = { |
708 | .next = 0, | |
709 | .next_set = 0, | |
710 | }; | |
81481eb4 | 711 | int i; |
320ae51f | 712 | |
71f79fb3 GKB |
713 | /* A deadlock might occur if a request is stuck requiring a |
714 | * timeout at the same time a queue freeze is waiting | |
715 | * completion, since the timeout code would not be able to | |
716 | * acquire the queue reference here. | |
717 | * | |
718 | * That's why we don't use blk_queue_enter here; instead, we use | |
719 | * percpu_ref_tryget directly, because we need to be able to | |
720 | * obtain a reference even in the short window between the queue | |
721 | * starting to freeze, by dropping the first reference in | |
722 | * blk_mq_freeze_queue_start, and the moment the last request is | |
723 | * consumed, marked by the instant q_usage_counter reaches | |
724 | * zero. | |
725 | */ | |
726 | if (!percpu_ref_tryget(&q->q_usage_counter)) | |
287922eb CH |
727 | return; |
728 | ||
0bf6cd5b | 729 | blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &data); |
320ae51f | 730 | |
81481eb4 CH |
731 | if (data.next_set) { |
732 | data.next = blk_rq_timeout(round_jiffies_up(data.next)); | |
733 | mod_timer(&q->timeout, data.next); | |
0d2602ca | 734 | } else { |
0bf6cd5b CH |
735 | struct blk_mq_hw_ctx *hctx; |
736 | ||
f054b56c ML |
737 | queue_for_each_hw_ctx(q, hctx, i) { |
738 | /* the hctx may be unmapped, so check it here */ | |
739 | if (blk_mq_hw_queue_mapped(hctx)) | |
740 | blk_mq_tag_idle(hctx); | |
741 | } | |
0d2602ca | 742 | } |
287922eb | 743 | blk_queue_exit(q); |
320ae51f JA |
744 | } |
745 | ||
746 | /* | |
747 | * Reverse check our software queue for entries that we could potentially | |
748 | * merge with. Currently includes a hand-wavy stop count of 8, to not spend | |
749 | * too much time checking for merges. | |
750 | */ | |
751 | static bool blk_mq_attempt_merge(struct request_queue *q, | |
752 | struct blk_mq_ctx *ctx, struct bio *bio) | |
753 | { | |
754 | struct request *rq; | |
755 | int checked = 8; | |
756 | ||
757 | list_for_each_entry_reverse(rq, &ctx->rq_list, queuelist) { | |
758 | int el_ret; | |
759 | ||
760 | if (!checked--) | |
761 | break; | |
762 | ||
763 | if (!blk_rq_merge_ok(rq, bio)) | |
764 | continue; | |
765 | ||
766 | el_ret = blk_try_merge(rq, bio); | |
767 | if (el_ret == ELEVATOR_BACK_MERGE) { | |
768 | if (bio_attempt_back_merge(q, rq, bio)) { | |
769 | ctx->rq_merged++; | |
770 | return true; | |
771 | } | |
772 | break; | |
773 | } else if (el_ret == ELEVATOR_FRONT_MERGE) { | |
774 | if (bio_attempt_front_merge(q, rq, bio)) { | |
775 | ctx->rq_merged++; | |
776 | return true; | |
777 | } | |
778 | break; | |
779 | } | |
780 | } | |
781 | ||
782 | return false; | |
783 | } | |
784 | ||
88459642 OS |
785 | struct flush_busy_ctx_data { |
786 | struct blk_mq_hw_ctx *hctx; | |
787 | struct list_head *list; | |
788 | }; | |
789 | ||
790 | static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data) | |
791 | { | |
792 | struct flush_busy_ctx_data *flush_data = data; | |
793 | struct blk_mq_hw_ctx *hctx = flush_data->hctx; | |
794 | struct blk_mq_ctx *ctx = hctx->ctxs[bitnr]; | |
795 | ||
796 | sbitmap_clear_bit(sb, bitnr); | |
797 | spin_lock(&ctx->lock); | |
798 | list_splice_tail_init(&ctx->rq_list, flush_data->list); | |
799 | spin_unlock(&ctx->lock); | |
800 | return true; | |
801 | } | |
802 | ||
1429d7c9 JA |
803 | /* |
804 | * Process software queues that have been marked busy, splicing them | |
805 | * to the for-dispatch | |
806 | */ | |
2c3ad667 | 807 | void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list) |
1429d7c9 | 808 | { |
88459642 OS |
809 | struct flush_busy_ctx_data data = { |
810 | .hctx = hctx, | |
811 | .list = list, | |
812 | }; | |
1429d7c9 | 813 | |
88459642 | 814 | sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data); |
1429d7c9 | 815 | } |
2c3ad667 | 816 | EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs); |
1429d7c9 | 817 | |
703fd1c0 JA |
818 | static inline unsigned int queued_to_index(unsigned int queued) |
819 | { | |
820 | if (!queued) | |
821 | return 0; | |
1429d7c9 | 822 | |
703fd1c0 | 823 | return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1); |
1429d7c9 JA |
824 | } |
825 | ||
f04c3df3 | 826 | bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx *hctx, struct list_head *list) |
320ae51f JA |
827 | { |
828 | struct request_queue *q = hctx->queue; | |
320ae51f | 829 | struct request *rq; |
74c45052 JA |
830 | LIST_HEAD(driver_list); |
831 | struct list_head *dptr; | |
f04c3df3 | 832 | int queued, ret = BLK_MQ_RQ_QUEUE_OK; |
320ae51f | 833 | |
74c45052 JA |
834 | /* |
835 | * Start off with dptr being NULL, so we start the first request | |
836 | * immediately, even if we have more pending. | |
837 | */ | |
838 | dptr = NULL; | |
839 | ||
320ae51f JA |
840 | /* |
841 | * Now process all the entries, sending them to the driver. | |
842 | */ | |
1429d7c9 | 843 | queued = 0; |
f04c3df3 | 844 | while (!list_empty(list)) { |
74c45052 | 845 | struct blk_mq_queue_data bd; |
320ae51f | 846 | |
f04c3df3 | 847 | rq = list_first_entry(list, struct request, queuelist); |
320ae51f | 848 | list_del_init(&rq->queuelist); |
320ae51f | 849 | |
74c45052 JA |
850 | bd.rq = rq; |
851 | bd.list = dptr; | |
f04c3df3 | 852 | bd.last = list_empty(list); |
74c45052 JA |
853 | |
854 | ret = q->mq_ops->queue_rq(hctx, &bd); | |
320ae51f JA |
855 | switch (ret) { |
856 | case BLK_MQ_RQ_QUEUE_OK: | |
857 | queued++; | |
52b9c330 | 858 | break; |
320ae51f | 859 | case BLK_MQ_RQ_QUEUE_BUSY: |
f04c3df3 | 860 | list_add(&rq->queuelist, list); |
ed0791b2 | 861 | __blk_mq_requeue_request(rq); |
320ae51f JA |
862 | break; |
863 | default: | |
864 | pr_err("blk-mq: bad return on queue: %d\n", ret); | |
320ae51f | 865 | case BLK_MQ_RQ_QUEUE_ERROR: |
1e93b8c2 | 866 | rq->errors = -EIO; |
c8a446ad | 867 | blk_mq_end_request(rq, rq->errors); |
320ae51f JA |
868 | break; |
869 | } | |
870 | ||
871 | if (ret == BLK_MQ_RQ_QUEUE_BUSY) | |
872 | break; | |
74c45052 JA |
873 | |
874 | /* | |
875 | * We've done the first request. If we have more than 1 | |
876 | * left in the list, set dptr to defer issue. | |
877 | */ | |
f04c3df3 | 878 | if (!dptr && list->next != list->prev) |
74c45052 | 879 | dptr = &driver_list; |
320ae51f JA |
880 | } |
881 | ||
703fd1c0 | 882 | hctx->dispatched[queued_to_index(queued)]++; |
320ae51f JA |
883 | |
884 | /* | |
885 | * Any items that need requeuing? Stuff them into hctx->dispatch, | |
886 | * that is where we will continue on next queue run. | |
887 | */ | |
f04c3df3 | 888 | if (!list_empty(list)) { |
320ae51f | 889 | spin_lock(&hctx->lock); |
f04c3df3 | 890 | list_splice(list, &hctx->dispatch); |
320ae51f | 891 | spin_unlock(&hctx->lock); |
f04c3df3 | 892 | |
9ba52e58 SL |
893 | /* |
894 | * the queue is expected stopped with BLK_MQ_RQ_QUEUE_BUSY, but | |
895 | * it's possible the queue is stopped and restarted again | |
896 | * before this. Queue restart will dispatch requests. And since | |
897 | * requests in rq_list aren't added into hctx->dispatch yet, | |
898 | * the requests in rq_list might get lost. | |
899 | * | |
900 | * blk_mq_run_hw_queue() already checks the STOPPED bit | |
901 | **/ | |
902 | blk_mq_run_hw_queue(hctx, true); | |
320ae51f | 903 | } |
f04c3df3 JA |
904 | |
905 | return ret != BLK_MQ_RQ_QUEUE_BUSY; | |
906 | } | |
907 | ||
908 | /* | |
909 | * Run this hardware queue, pulling any software queues mapped to it in. | |
910 | * Note that this function currently has various problems around ordering | |
911 | * of IO. In particular, we'd like FIFO behaviour on handling existing | |
912 | * items on the hctx->dispatch list. Ignore that for now. | |
913 | */ | |
914 | static void blk_mq_process_rq_list(struct blk_mq_hw_ctx *hctx) | |
915 | { | |
916 | LIST_HEAD(rq_list); | |
917 | LIST_HEAD(driver_list); | |
918 | ||
919 | if (unlikely(blk_mq_hctx_stopped(hctx))) | |
920 | return; | |
921 | ||
922 | hctx->run++; | |
923 | ||
924 | /* | |
925 | * Touch any software queue that has pending entries. | |
926 | */ | |
2c3ad667 | 927 | blk_mq_flush_busy_ctxs(hctx, &rq_list); |
f04c3df3 JA |
928 | |
929 | /* | |
930 | * If we have previous entries on our dispatch list, grab them | |
931 | * and stuff them at the front for more fair dispatch. | |
932 | */ | |
933 | if (!list_empty_careful(&hctx->dispatch)) { | |
934 | spin_lock(&hctx->lock); | |
935 | if (!list_empty(&hctx->dispatch)) | |
936 | list_splice_init(&hctx->dispatch, &rq_list); | |
937 | spin_unlock(&hctx->lock); | |
938 | } | |
939 | ||
940 | blk_mq_dispatch_rq_list(hctx, &rq_list); | |
320ae51f JA |
941 | } |
942 | ||
6a83e74d BVA |
943 | static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx) |
944 | { | |
945 | int srcu_idx; | |
946 | ||
947 | WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) && | |
948 | cpu_online(hctx->next_cpu)); | |
949 | ||
950 | if (!(hctx->flags & BLK_MQ_F_BLOCKING)) { | |
951 | rcu_read_lock(); | |
952 | blk_mq_process_rq_list(hctx); | |
953 | rcu_read_unlock(); | |
954 | } else { | |
955 | srcu_idx = srcu_read_lock(&hctx->queue_rq_srcu); | |
956 | blk_mq_process_rq_list(hctx); | |
957 | srcu_read_unlock(&hctx->queue_rq_srcu, srcu_idx); | |
958 | } | |
959 | } | |
960 | ||
506e931f JA |
961 | /* |
962 | * It'd be great if the workqueue API had a way to pass | |
963 | * in a mask and had some smarts for more clever placement. | |
964 | * For now we just round-robin here, switching for every | |
965 | * BLK_MQ_CPU_WORK_BATCH queued items. | |
966 | */ | |
967 | static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx) | |
968 | { | |
b657d7e6 CH |
969 | if (hctx->queue->nr_hw_queues == 1) |
970 | return WORK_CPU_UNBOUND; | |
506e931f JA |
971 | |
972 | if (--hctx->next_cpu_batch <= 0) { | |
c02ebfdd | 973 | int next_cpu; |
506e931f JA |
974 | |
975 | next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask); | |
976 | if (next_cpu >= nr_cpu_ids) | |
977 | next_cpu = cpumask_first(hctx->cpumask); | |
978 | ||
979 | hctx->next_cpu = next_cpu; | |
980 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; | |
981 | } | |
982 | ||
b657d7e6 | 983 | return hctx->next_cpu; |
506e931f JA |
984 | } |
985 | ||
320ae51f JA |
986 | void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) |
987 | { | |
5d1b25c1 BVA |
988 | if (unlikely(blk_mq_hctx_stopped(hctx) || |
989 | !blk_mq_hw_queue_mapped(hctx))) | |
320ae51f JA |
990 | return; |
991 | ||
1b792f2f | 992 | if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) { |
2a90d4aa PB |
993 | int cpu = get_cpu(); |
994 | if (cpumask_test_cpu(cpu, hctx->cpumask)) { | |
398205b8 | 995 | __blk_mq_run_hw_queue(hctx); |
2a90d4aa | 996 | put_cpu(); |
398205b8 PB |
997 | return; |
998 | } | |
e4043dcf | 999 | |
2a90d4aa | 1000 | put_cpu(); |
e4043dcf | 1001 | } |
398205b8 | 1002 | |
27489a3c | 1003 | kblockd_schedule_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work); |
320ae51f JA |
1004 | } |
1005 | ||
b94ec296 | 1006 | void blk_mq_run_hw_queues(struct request_queue *q, bool async) |
320ae51f JA |
1007 | { |
1008 | struct blk_mq_hw_ctx *hctx; | |
1009 | int i; | |
1010 | ||
1011 | queue_for_each_hw_ctx(q, hctx, i) { | |
1012 | if ((!blk_mq_hctx_has_pending(hctx) && | |
1013 | list_empty_careful(&hctx->dispatch)) || | |
5d1b25c1 | 1014 | blk_mq_hctx_stopped(hctx)) |
320ae51f JA |
1015 | continue; |
1016 | ||
b94ec296 | 1017 | blk_mq_run_hw_queue(hctx, async); |
320ae51f JA |
1018 | } |
1019 | } | |
b94ec296 | 1020 | EXPORT_SYMBOL(blk_mq_run_hw_queues); |
320ae51f | 1021 | |
fd001443 BVA |
1022 | /** |
1023 | * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped | |
1024 | * @q: request queue. | |
1025 | * | |
1026 | * The caller is responsible for serializing this function against | |
1027 | * blk_mq_{start,stop}_hw_queue(). | |
1028 | */ | |
1029 | bool blk_mq_queue_stopped(struct request_queue *q) | |
1030 | { | |
1031 | struct blk_mq_hw_ctx *hctx; | |
1032 | int i; | |
1033 | ||
1034 | queue_for_each_hw_ctx(q, hctx, i) | |
1035 | if (blk_mq_hctx_stopped(hctx)) | |
1036 | return true; | |
1037 | ||
1038 | return false; | |
1039 | } | |
1040 | EXPORT_SYMBOL(blk_mq_queue_stopped); | |
1041 | ||
320ae51f JA |
1042 | void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx) |
1043 | { | |
27489a3c | 1044 | cancel_work(&hctx->run_work); |
70f4db63 | 1045 | cancel_delayed_work(&hctx->delay_work); |
320ae51f JA |
1046 | set_bit(BLK_MQ_S_STOPPED, &hctx->state); |
1047 | } | |
1048 | EXPORT_SYMBOL(blk_mq_stop_hw_queue); | |
1049 | ||
280d45f6 CH |
1050 | void blk_mq_stop_hw_queues(struct request_queue *q) |
1051 | { | |
1052 | struct blk_mq_hw_ctx *hctx; | |
1053 | int i; | |
1054 | ||
1055 | queue_for_each_hw_ctx(q, hctx, i) | |
1056 | blk_mq_stop_hw_queue(hctx); | |
1057 | } | |
1058 | EXPORT_SYMBOL(blk_mq_stop_hw_queues); | |
1059 | ||
320ae51f JA |
1060 | void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx) |
1061 | { | |
1062 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
e4043dcf | 1063 | |
0ffbce80 | 1064 | blk_mq_run_hw_queue(hctx, false); |
320ae51f JA |
1065 | } |
1066 | EXPORT_SYMBOL(blk_mq_start_hw_queue); | |
1067 | ||
2f268556 CH |
1068 | void blk_mq_start_hw_queues(struct request_queue *q) |
1069 | { | |
1070 | struct blk_mq_hw_ctx *hctx; | |
1071 | int i; | |
1072 | ||
1073 | queue_for_each_hw_ctx(q, hctx, i) | |
1074 | blk_mq_start_hw_queue(hctx); | |
1075 | } | |
1076 | EXPORT_SYMBOL(blk_mq_start_hw_queues); | |
1077 | ||
ae911c5e JA |
1078 | void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) |
1079 | { | |
1080 | if (!blk_mq_hctx_stopped(hctx)) | |
1081 | return; | |
1082 | ||
1083 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
1084 | blk_mq_run_hw_queue(hctx, async); | |
1085 | } | |
1086 | EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue); | |
1087 | ||
1b4a3258 | 1088 | void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async) |
320ae51f JA |
1089 | { |
1090 | struct blk_mq_hw_ctx *hctx; | |
1091 | int i; | |
1092 | ||
ae911c5e JA |
1093 | queue_for_each_hw_ctx(q, hctx, i) |
1094 | blk_mq_start_stopped_hw_queue(hctx, async); | |
320ae51f JA |
1095 | } |
1096 | EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues); | |
1097 | ||
70f4db63 | 1098 | static void blk_mq_run_work_fn(struct work_struct *work) |
320ae51f JA |
1099 | { |
1100 | struct blk_mq_hw_ctx *hctx; | |
1101 | ||
27489a3c | 1102 | hctx = container_of(work, struct blk_mq_hw_ctx, run_work); |
e4043dcf | 1103 | |
320ae51f JA |
1104 | __blk_mq_run_hw_queue(hctx); |
1105 | } | |
1106 | ||
70f4db63 CH |
1107 | static void blk_mq_delay_work_fn(struct work_struct *work) |
1108 | { | |
1109 | struct blk_mq_hw_ctx *hctx; | |
1110 | ||
1111 | hctx = container_of(work, struct blk_mq_hw_ctx, delay_work.work); | |
1112 | ||
1113 | if (test_and_clear_bit(BLK_MQ_S_STOPPED, &hctx->state)) | |
1114 | __blk_mq_run_hw_queue(hctx); | |
1115 | } | |
1116 | ||
1117 | void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs) | |
1118 | { | |
19c66e59 ML |
1119 | if (unlikely(!blk_mq_hw_queue_mapped(hctx))) |
1120 | return; | |
70f4db63 | 1121 | |
b657d7e6 CH |
1122 | kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx), |
1123 | &hctx->delay_work, msecs_to_jiffies(msecs)); | |
70f4db63 CH |
1124 | } |
1125 | EXPORT_SYMBOL(blk_mq_delay_queue); | |
1126 | ||
cfd0c552 | 1127 | static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx, |
cfd0c552 ML |
1128 | struct request *rq, |
1129 | bool at_head) | |
320ae51f | 1130 | { |
e57690fe JA |
1131 | struct blk_mq_ctx *ctx = rq->mq_ctx; |
1132 | ||
01b983c9 JA |
1133 | trace_block_rq_insert(hctx->queue, rq); |
1134 | ||
72a0a36e CH |
1135 | if (at_head) |
1136 | list_add(&rq->queuelist, &ctx->rq_list); | |
1137 | else | |
1138 | list_add_tail(&rq->queuelist, &ctx->rq_list); | |
cfd0c552 | 1139 | } |
4bb659b1 | 1140 | |
2c3ad667 JA |
1141 | void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq, |
1142 | bool at_head) | |
cfd0c552 ML |
1143 | { |
1144 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
1145 | ||
e57690fe | 1146 | __blk_mq_insert_req_list(hctx, rq, at_head); |
320ae51f | 1147 | blk_mq_hctx_mark_pending(hctx, ctx); |
320ae51f JA |
1148 | } |
1149 | ||
eeabc850 | 1150 | void blk_mq_insert_request(struct request *rq, bool at_head, bool run_queue, |
e57690fe | 1151 | bool async) |
320ae51f | 1152 | { |
e57690fe | 1153 | struct blk_mq_ctx *ctx = rq->mq_ctx; |
eeabc850 | 1154 | struct request_queue *q = rq->q; |
7d7e0f90 | 1155 | struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, ctx->cpu); |
320ae51f | 1156 | |
a57a178a CH |
1157 | spin_lock(&ctx->lock); |
1158 | __blk_mq_insert_request(hctx, rq, at_head); | |
1159 | spin_unlock(&ctx->lock); | |
320ae51f | 1160 | |
320ae51f JA |
1161 | if (run_queue) |
1162 | blk_mq_run_hw_queue(hctx, async); | |
1163 | } | |
1164 | ||
1165 | static void blk_mq_insert_requests(struct request_queue *q, | |
1166 | struct blk_mq_ctx *ctx, | |
1167 | struct list_head *list, | |
1168 | int depth, | |
1169 | bool from_schedule) | |
1170 | ||
1171 | { | |
7d7e0f90 | 1172 | struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, ctx->cpu); |
320ae51f JA |
1173 | |
1174 | trace_block_unplug(q, depth, !from_schedule); | |
1175 | ||
320ae51f JA |
1176 | /* |
1177 | * preemption doesn't flush plug list, so it's possible ctx->cpu is | |
1178 | * offline now | |
1179 | */ | |
1180 | spin_lock(&ctx->lock); | |
1181 | while (!list_empty(list)) { | |
1182 | struct request *rq; | |
1183 | ||
1184 | rq = list_first_entry(list, struct request, queuelist); | |
e57690fe | 1185 | BUG_ON(rq->mq_ctx != ctx); |
320ae51f | 1186 | list_del_init(&rq->queuelist); |
e57690fe | 1187 | __blk_mq_insert_req_list(hctx, rq, false); |
320ae51f | 1188 | } |
cfd0c552 | 1189 | blk_mq_hctx_mark_pending(hctx, ctx); |
320ae51f JA |
1190 | spin_unlock(&ctx->lock); |
1191 | ||
320ae51f JA |
1192 | blk_mq_run_hw_queue(hctx, from_schedule); |
1193 | } | |
1194 | ||
1195 | static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b) | |
1196 | { | |
1197 | struct request *rqa = container_of(a, struct request, queuelist); | |
1198 | struct request *rqb = container_of(b, struct request, queuelist); | |
1199 | ||
1200 | return !(rqa->mq_ctx < rqb->mq_ctx || | |
1201 | (rqa->mq_ctx == rqb->mq_ctx && | |
1202 | blk_rq_pos(rqa) < blk_rq_pos(rqb))); | |
1203 | } | |
1204 | ||
1205 | void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule) | |
1206 | { | |
1207 | struct blk_mq_ctx *this_ctx; | |
1208 | struct request_queue *this_q; | |
1209 | struct request *rq; | |
1210 | LIST_HEAD(list); | |
1211 | LIST_HEAD(ctx_list); | |
1212 | unsigned int depth; | |
1213 | ||
1214 | list_splice_init(&plug->mq_list, &list); | |
1215 | ||
1216 | list_sort(NULL, &list, plug_ctx_cmp); | |
1217 | ||
1218 | this_q = NULL; | |
1219 | this_ctx = NULL; | |
1220 | depth = 0; | |
1221 | ||
1222 | while (!list_empty(&list)) { | |
1223 | rq = list_entry_rq(list.next); | |
1224 | list_del_init(&rq->queuelist); | |
1225 | BUG_ON(!rq->q); | |
1226 | if (rq->mq_ctx != this_ctx) { | |
1227 | if (this_ctx) { | |
1228 | blk_mq_insert_requests(this_q, this_ctx, | |
1229 | &ctx_list, depth, | |
1230 | from_schedule); | |
1231 | } | |
1232 | ||
1233 | this_ctx = rq->mq_ctx; | |
1234 | this_q = rq->q; | |
1235 | depth = 0; | |
1236 | } | |
1237 | ||
1238 | depth++; | |
1239 | list_add_tail(&rq->queuelist, &ctx_list); | |
1240 | } | |
1241 | ||
1242 | /* | |
1243 | * If 'this_ctx' is set, we know we have entries to complete | |
1244 | * on 'ctx_list'. Do those. | |
1245 | */ | |
1246 | if (this_ctx) { | |
1247 | blk_mq_insert_requests(this_q, this_ctx, &ctx_list, depth, | |
1248 | from_schedule); | |
1249 | } | |
1250 | } | |
1251 | ||
1252 | static void blk_mq_bio_to_request(struct request *rq, struct bio *bio) | |
1253 | { | |
1254 | init_request_from_bio(rq, bio); | |
4b570521 | 1255 | |
6e85eaf3 | 1256 | blk_account_io_start(rq, true); |
320ae51f JA |
1257 | } |
1258 | ||
274a5843 JA |
1259 | static inline bool hctx_allow_merges(struct blk_mq_hw_ctx *hctx) |
1260 | { | |
1261 | return (hctx->flags & BLK_MQ_F_SHOULD_MERGE) && | |
1262 | !blk_queue_nomerges(hctx->queue); | |
1263 | } | |
1264 | ||
07068d5b JA |
1265 | static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx *hctx, |
1266 | struct blk_mq_ctx *ctx, | |
1267 | struct request *rq, struct bio *bio) | |
320ae51f | 1268 | { |
e18378a6 | 1269 | if (!hctx_allow_merges(hctx) || !bio_mergeable(bio)) { |
07068d5b JA |
1270 | blk_mq_bio_to_request(rq, bio); |
1271 | spin_lock(&ctx->lock); | |
1272 | insert_rq: | |
1273 | __blk_mq_insert_request(hctx, rq, false); | |
1274 | spin_unlock(&ctx->lock); | |
1275 | return false; | |
1276 | } else { | |
274a5843 JA |
1277 | struct request_queue *q = hctx->queue; |
1278 | ||
07068d5b JA |
1279 | spin_lock(&ctx->lock); |
1280 | if (!blk_mq_attempt_merge(q, ctx, bio)) { | |
1281 | blk_mq_bio_to_request(rq, bio); | |
1282 | goto insert_rq; | |
1283 | } | |
320ae51f | 1284 | |
07068d5b JA |
1285 | spin_unlock(&ctx->lock); |
1286 | __blk_mq_free_request(hctx, ctx, rq); | |
1287 | return true; | |
14ec77f3 | 1288 | } |
07068d5b | 1289 | } |
14ec77f3 | 1290 | |
07068d5b JA |
1291 | static struct request *blk_mq_map_request(struct request_queue *q, |
1292 | struct bio *bio, | |
2552e3f8 | 1293 | struct blk_mq_alloc_data *data) |
07068d5b JA |
1294 | { |
1295 | struct blk_mq_hw_ctx *hctx; | |
1296 | struct blk_mq_ctx *ctx; | |
1297 | struct request *rq; | |
320ae51f | 1298 | |
3ef28e83 | 1299 | blk_queue_enter_live(q); |
320ae51f | 1300 | ctx = blk_mq_get_ctx(q); |
7d7e0f90 | 1301 | hctx = blk_mq_map_queue(q, ctx->cpu); |
320ae51f | 1302 | |
ef295ecf | 1303 | trace_block_getrq(q, bio, bio->bi_opf); |
2552e3f8 | 1304 | blk_mq_set_alloc_data(data, q, 0, ctx, hctx); |
ef295ecf | 1305 | rq = __blk_mq_alloc_request(data, bio->bi_opf); |
320ae51f | 1306 | |
7dd2fb68 | 1307 | data->hctx->queued++; |
07068d5b JA |
1308 | return rq; |
1309 | } | |
1310 | ||
066a4a73 | 1311 | static void blk_mq_try_issue_directly(struct request *rq, blk_qc_t *cookie) |
f984df1f SL |
1312 | { |
1313 | int ret; | |
1314 | struct request_queue *q = rq->q; | |
066a4a73 | 1315 | struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, rq->mq_ctx->cpu); |
f984df1f SL |
1316 | struct blk_mq_queue_data bd = { |
1317 | .rq = rq, | |
1318 | .list = NULL, | |
1319 | .last = 1 | |
1320 | }; | |
7b371636 | 1321 | blk_qc_t new_cookie = blk_tag_to_qc_t(rq->tag, hctx->queue_num); |
f984df1f | 1322 | |
2253efc8 BVA |
1323 | if (blk_mq_hctx_stopped(hctx)) |
1324 | goto insert; | |
1325 | ||
f984df1f SL |
1326 | /* |
1327 | * For OK queue, we are done. For error, kill it. Any other | |
1328 | * error (busy), just add it to our list as we previously | |
1329 | * would have done | |
1330 | */ | |
1331 | ret = q->mq_ops->queue_rq(hctx, &bd); | |
7b371636 JA |
1332 | if (ret == BLK_MQ_RQ_QUEUE_OK) { |
1333 | *cookie = new_cookie; | |
2253efc8 | 1334 | return; |
7b371636 | 1335 | } |
f984df1f | 1336 | |
7b371636 JA |
1337 | __blk_mq_requeue_request(rq); |
1338 | ||
1339 | if (ret == BLK_MQ_RQ_QUEUE_ERROR) { | |
1340 | *cookie = BLK_QC_T_NONE; | |
1341 | rq->errors = -EIO; | |
1342 | blk_mq_end_request(rq, rq->errors); | |
2253efc8 | 1343 | return; |
f984df1f | 1344 | } |
7b371636 | 1345 | |
2253efc8 BVA |
1346 | insert: |
1347 | blk_mq_insert_request(rq, false, true, true); | |
f984df1f SL |
1348 | } |
1349 | ||
07068d5b JA |
1350 | /* |
1351 | * Multiple hardware queue variant. This will not use per-process plugs, | |
1352 | * but will attempt to bypass the hctx queueing if we can go straight to | |
1353 | * hardware for SYNC IO. | |
1354 | */ | |
dece1635 | 1355 | static blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio) |
07068d5b | 1356 | { |
ef295ecf | 1357 | const int is_sync = op_is_sync(bio->bi_opf); |
1eff9d32 | 1358 | const int is_flush_fua = bio->bi_opf & (REQ_PREFLUSH | REQ_FUA); |
2552e3f8 | 1359 | struct blk_mq_alloc_data data; |
07068d5b | 1360 | struct request *rq; |
6a83e74d | 1361 | unsigned int request_count = 0, srcu_idx; |
f984df1f | 1362 | struct blk_plug *plug; |
5b3f341f | 1363 | struct request *same_queue_rq = NULL; |
7b371636 | 1364 | blk_qc_t cookie; |
87760e5e | 1365 | unsigned int wb_acct; |
07068d5b JA |
1366 | |
1367 | blk_queue_bounce(q, &bio); | |
1368 | ||
1369 | if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) { | |
4246a0b6 | 1370 | bio_io_error(bio); |
dece1635 | 1371 | return BLK_QC_T_NONE; |
07068d5b JA |
1372 | } |
1373 | ||
54efd50b KO |
1374 | blk_queue_split(q, &bio, q->bio_split); |
1375 | ||
87c279e6 OS |
1376 | if (!is_flush_fua && !blk_queue_nomerges(q) && |
1377 | blk_attempt_plug_merge(q, bio, &request_count, &same_queue_rq)) | |
1378 | return BLK_QC_T_NONE; | |
f984df1f | 1379 | |
87760e5e JA |
1380 | wb_acct = wbt_wait(q->rq_wb, bio, NULL); |
1381 | ||
07068d5b | 1382 | rq = blk_mq_map_request(q, bio, &data); |
87760e5e JA |
1383 | if (unlikely(!rq)) { |
1384 | __wbt_done(q->rq_wb, wb_acct); | |
dece1635 | 1385 | return BLK_QC_T_NONE; |
87760e5e JA |
1386 | } |
1387 | ||
1388 | wbt_track(&rq->issue_stat, wb_acct); | |
07068d5b | 1389 | |
7b371636 | 1390 | cookie = blk_tag_to_qc_t(rq->tag, data.hctx->queue_num); |
07068d5b JA |
1391 | |
1392 | if (unlikely(is_flush_fua)) { | |
1393 | blk_mq_bio_to_request(rq, bio); | |
1394 | blk_insert_flush(rq); | |
1395 | goto run_queue; | |
1396 | } | |
1397 | ||
f984df1f | 1398 | plug = current->plug; |
e167dfb5 JA |
1399 | /* |
1400 | * If the driver supports defer issued based on 'last', then | |
1401 | * queue it up like normal since we can potentially save some | |
1402 | * CPU this way. | |
1403 | */ | |
f984df1f SL |
1404 | if (((plug && !blk_queue_nomerges(q)) || is_sync) && |
1405 | !(data.hctx->flags & BLK_MQ_F_DEFER_ISSUE)) { | |
1406 | struct request *old_rq = NULL; | |
07068d5b JA |
1407 | |
1408 | blk_mq_bio_to_request(rq, bio); | |
07068d5b JA |
1409 | |
1410 | /* | |
6a83e74d | 1411 | * We do limited plugging. If the bio can be merged, do that. |
f984df1f SL |
1412 | * Otherwise the existing request in the plug list will be |
1413 | * issued. So the plug list will have one request at most | |
07068d5b | 1414 | */ |
f984df1f | 1415 | if (plug) { |
5b3f341f SL |
1416 | /* |
1417 | * The plug list might get flushed before this. If that | |
b094f89c JA |
1418 | * happens, same_queue_rq is invalid and plug list is |
1419 | * empty | |
1420 | */ | |
5b3f341f SL |
1421 | if (same_queue_rq && !list_empty(&plug->mq_list)) { |
1422 | old_rq = same_queue_rq; | |
f984df1f | 1423 | list_del_init(&old_rq->queuelist); |
07068d5b | 1424 | } |
f984df1f SL |
1425 | list_add_tail(&rq->queuelist, &plug->mq_list); |
1426 | } else /* is_sync */ | |
1427 | old_rq = rq; | |
1428 | blk_mq_put_ctx(data.ctx); | |
1429 | if (!old_rq) | |
7b371636 | 1430 | goto done; |
6a83e74d BVA |
1431 | |
1432 | if (!(data.hctx->flags & BLK_MQ_F_BLOCKING)) { | |
1433 | rcu_read_lock(); | |
066a4a73 | 1434 | blk_mq_try_issue_directly(old_rq, &cookie); |
6a83e74d BVA |
1435 | rcu_read_unlock(); |
1436 | } else { | |
1437 | srcu_idx = srcu_read_lock(&data.hctx->queue_rq_srcu); | |
066a4a73 | 1438 | blk_mq_try_issue_directly(old_rq, &cookie); |
6a83e74d BVA |
1439 | srcu_read_unlock(&data.hctx->queue_rq_srcu, srcu_idx); |
1440 | } | |
7b371636 | 1441 | goto done; |
07068d5b JA |
1442 | } |
1443 | ||
1444 | if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) { | |
1445 | /* | |
1446 | * For a SYNC request, send it to the hardware immediately. For | |
1447 | * an ASYNC request, just ensure that we run it later on. The | |
1448 | * latter allows for merging opportunities and more efficient | |
1449 | * dispatching. | |
1450 | */ | |
1451 | run_queue: | |
1452 | blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua); | |
1453 | } | |
07068d5b | 1454 | blk_mq_put_ctx(data.ctx); |
7b371636 JA |
1455 | done: |
1456 | return cookie; | |
07068d5b JA |
1457 | } |
1458 | ||
1459 | /* | |
1460 | * Single hardware queue variant. This will attempt to use any per-process | |
1461 | * plug for merging and IO deferral. | |
1462 | */ | |
dece1635 | 1463 | static blk_qc_t blk_sq_make_request(struct request_queue *q, struct bio *bio) |
07068d5b | 1464 | { |
ef295ecf | 1465 | const int is_sync = op_is_sync(bio->bi_opf); |
1eff9d32 | 1466 | const int is_flush_fua = bio->bi_opf & (REQ_PREFLUSH | REQ_FUA); |
e6c4438b JM |
1467 | struct blk_plug *plug; |
1468 | unsigned int request_count = 0; | |
2552e3f8 | 1469 | struct blk_mq_alloc_data data; |
07068d5b | 1470 | struct request *rq; |
7b371636 | 1471 | blk_qc_t cookie; |
87760e5e | 1472 | unsigned int wb_acct; |
07068d5b | 1473 | |
07068d5b JA |
1474 | blk_queue_bounce(q, &bio); |
1475 | ||
1476 | if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) { | |
4246a0b6 | 1477 | bio_io_error(bio); |
dece1635 | 1478 | return BLK_QC_T_NONE; |
07068d5b JA |
1479 | } |
1480 | ||
54efd50b KO |
1481 | blk_queue_split(q, &bio, q->bio_split); |
1482 | ||
87c279e6 OS |
1483 | if (!is_flush_fua && !blk_queue_nomerges(q)) { |
1484 | if (blk_attempt_plug_merge(q, bio, &request_count, NULL)) | |
1485 | return BLK_QC_T_NONE; | |
1486 | } else | |
1487 | request_count = blk_plug_queued_count(q); | |
07068d5b | 1488 | |
87760e5e JA |
1489 | wb_acct = wbt_wait(q->rq_wb, bio, NULL); |
1490 | ||
07068d5b | 1491 | rq = blk_mq_map_request(q, bio, &data); |
87760e5e JA |
1492 | if (unlikely(!rq)) { |
1493 | __wbt_done(q->rq_wb, wb_acct); | |
dece1635 | 1494 | return BLK_QC_T_NONE; |
87760e5e JA |
1495 | } |
1496 | ||
1497 | wbt_track(&rq->issue_stat, wb_acct); | |
320ae51f | 1498 | |
7b371636 | 1499 | cookie = blk_tag_to_qc_t(rq->tag, data.hctx->queue_num); |
320ae51f JA |
1500 | |
1501 | if (unlikely(is_flush_fua)) { | |
1502 | blk_mq_bio_to_request(rq, bio); | |
320ae51f JA |
1503 | blk_insert_flush(rq); |
1504 | goto run_queue; | |
1505 | } | |
1506 | ||
1507 | /* | |
1508 | * A task plug currently exists. Since this is completely lockless, | |
1509 | * utilize that to temporarily store requests until the task is | |
1510 | * either done or scheduled away. | |
1511 | */ | |
e6c4438b JM |
1512 | plug = current->plug; |
1513 | if (plug) { | |
600271d9 SL |
1514 | struct request *last = NULL; |
1515 | ||
e6c4438b | 1516 | blk_mq_bio_to_request(rq, bio); |
0a6219a9 ML |
1517 | |
1518 | /* | |
1519 | * @request_count may become stale because of schedule | |
1520 | * out, so check the list again. | |
1521 | */ | |
1522 | if (list_empty(&plug->mq_list)) | |
1523 | request_count = 0; | |
676d0607 | 1524 | if (!request_count) |
e6c4438b | 1525 | trace_block_plug(q); |
600271d9 SL |
1526 | else |
1527 | last = list_entry_rq(plug->mq_list.prev); | |
b094f89c JA |
1528 | |
1529 | blk_mq_put_ctx(data.ctx); | |
1530 | ||
600271d9 SL |
1531 | if (request_count >= BLK_MAX_REQUEST_COUNT || (last && |
1532 | blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) { | |
e6c4438b JM |
1533 | blk_flush_plug_list(plug, false); |
1534 | trace_block_plug(q); | |
320ae51f | 1535 | } |
b094f89c | 1536 | |
e6c4438b | 1537 | list_add_tail(&rq->queuelist, &plug->mq_list); |
7b371636 | 1538 | return cookie; |
320ae51f JA |
1539 | } |
1540 | ||
07068d5b JA |
1541 | if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) { |
1542 | /* | |
1543 | * For a SYNC request, send it to the hardware immediately. For | |
1544 | * an ASYNC request, just ensure that we run it later on. The | |
1545 | * latter allows for merging opportunities and more efficient | |
1546 | * dispatching. | |
1547 | */ | |
1548 | run_queue: | |
1549 | blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua); | |
320ae51f JA |
1550 | } |
1551 | ||
07068d5b | 1552 | blk_mq_put_ctx(data.ctx); |
7b371636 | 1553 | return cookie; |
320ae51f JA |
1554 | } |
1555 | ||
2c3ad667 JA |
1556 | void blk_mq_free_rq_map(struct blk_mq_tag_set *set, struct blk_mq_tags *tags, |
1557 | unsigned int hctx_idx) | |
95363efd | 1558 | { |
e9b267d9 | 1559 | struct page *page; |
320ae51f | 1560 | |
24d2f903 | 1561 | if (tags->rqs && set->ops->exit_request) { |
e9b267d9 | 1562 | int i; |
320ae51f | 1563 | |
24d2f903 CH |
1564 | for (i = 0; i < tags->nr_tags; i++) { |
1565 | if (!tags->rqs[i]) | |
e9b267d9 | 1566 | continue; |
24d2f903 CH |
1567 | set->ops->exit_request(set->driver_data, tags->rqs[i], |
1568 | hctx_idx, i); | |
a5164405 | 1569 | tags->rqs[i] = NULL; |
e9b267d9 | 1570 | } |
320ae51f | 1571 | } |
320ae51f | 1572 | |
24d2f903 CH |
1573 | while (!list_empty(&tags->page_list)) { |
1574 | page = list_first_entry(&tags->page_list, struct page, lru); | |
6753471c | 1575 | list_del_init(&page->lru); |
f75782e4 CM |
1576 | /* |
1577 | * Remove kmemleak object previously allocated in | |
1578 | * blk_mq_init_rq_map(). | |
1579 | */ | |
1580 | kmemleak_free(page_address(page)); | |
320ae51f JA |
1581 | __free_pages(page, page->private); |
1582 | } | |
1583 | ||
24d2f903 | 1584 | kfree(tags->rqs); |
320ae51f | 1585 | |
24d2f903 | 1586 | blk_mq_free_tags(tags); |
320ae51f JA |
1587 | } |
1588 | ||
1589 | static size_t order_to_size(unsigned int order) | |
1590 | { | |
4ca08500 | 1591 | return (size_t)PAGE_SIZE << order; |
320ae51f JA |
1592 | } |
1593 | ||
2c3ad667 JA |
1594 | struct blk_mq_tags *blk_mq_init_rq_map(struct blk_mq_tag_set *set, |
1595 | unsigned int hctx_idx) | |
320ae51f | 1596 | { |
24d2f903 | 1597 | struct blk_mq_tags *tags; |
320ae51f JA |
1598 | unsigned int i, j, entries_per_page, max_order = 4; |
1599 | size_t rq_size, left; | |
1600 | ||
24d2f903 | 1601 | tags = blk_mq_init_tags(set->queue_depth, set->reserved_tags, |
24391c0d SL |
1602 | set->numa_node, |
1603 | BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags)); | |
24d2f903 CH |
1604 | if (!tags) |
1605 | return NULL; | |
320ae51f | 1606 | |
24d2f903 CH |
1607 | INIT_LIST_HEAD(&tags->page_list); |
1608 | ||
a5164405 | 1609 | tags->rqs = kzalloc_node(set->queue_depth * sizeof(struct request *), |
36e1f3d1 | 1610 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, |
a5164405 | 1611 | set->numa_node); |
24d2f903 CH |
1612 | if (!tags->rqs) { |
1613 | blk_mq_free_tags(tags); | |
1614 | return NULL; | |
1615 | } | |
320ae51f JA |
1616 | |
1617 | /* | |
1618 | * rq_size is the size of the request plus driver payload, rounded | |
1619 | * to the cacheline size | |
1620 | */ | |
24d2f903 | 1621 | rq_size = round_up(sizeof(struct request) + set->cmd_size, |
320ae51f | 1622 | cache_line_size()); |
24d2f903 | 1623 | left = rq_size * set->queue_depth; |
320ae51f | 1624 | |
24d2f903 | 1625 | for (i = 0; i < set->queue_depth; ) { |
320ae51f JA |
1626 | int this_order = max_order; |
1627 | struct page *page; | |
1628 | int to_do; | |
1629 | void *p; | |
1630 | ||
b3a834b1 | 1631 | while (this_order && left < order_to_size(this_order - 1)) |
320ae51f JA |
1632 | this_order--; |
1633 | ||
1634 | do { | |
a5164405 | 1635 | page = alloc_pages_node(set->numa_node, |
36e1f3d1 | 1636 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO, |
a5164405 | 1637 | this_order); |
320ae51f JA |
1638 | if (page) |
1639 | break; | |
1640 | if (!this_order--) | |
1641 | break; | |
1642 | if (order_to_size(this_order) < rq_size) | |
1643 | break; | |
1644 | } while (1); | |
1645 | ||
1646 | if (!page) | |
24d2f903 | 1647 | goto fail; |
320ae51f JA |
1648 | |
1649 | page->private = this_order; | |
24d2f903 | 1650 | list_add_tail(&page->lru, &tags->page_list); |
320ae51f JA |
1651 | |
1652 | p = page_address(page); | |
f75782e4 CM |
1653 | /* |
1654 | * Allow kmemleak to scan these pages as they contain pointers | |
1655 | * to additional allocations like via ops->init_request(). | |
1656 | */ | |
36e1f3d1 | 1657 | kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO); |
320ae51f | 1658 | entries_per_page = order_to_size(this_order) / rq_size; |
24d2f903 | 1659 | to_do = min(entries_per_page, set->queue_depth - i); |
320ae51f JA |
1660 | left -= to_do * rq_size; |
1661 | for (j = 0; j < to_do; j++) { | |
24d2f903 CH |
1662 | tags->rqs[i] = p; |
1663 | if (set->ops->init_request) { | |
1664 | if (set->ops->init_request(set->driver_data, | |
1665 | tags->rqs[i], hctx_idx, i, | |
a5164405 JA |
1666 | set->numa_node)) { |
1667 | tags->rqs[i] = NULL; | |
24d2f903 | 1668 | goto fail; |
a5164405 | 1669 | } |
e9b267d9 CH |
1670 | } |
1671 | ||
320ae51f JA |
1672 | p += rq_size; |
1673 | i++; | |
1674 | } | |
1675 | } | |
24d2f903 | 1676 | return tags; |
320ae51f | 1677 | |
24d2f903 | 1678 | fail: |
24d2f903 CH |
1679 | blk_mq_free_rq_map(set, tags, hctx_idx); |
1680 | return NULL; | |
320ae51f JA |
1681 | } |
1682 | ||
e57690fe JA |
1683 | /* |
1684 | * 'cpu' is going away. splice any existing rq_list entries from this | |
1685 | * software queue to the hw queue dispatch list, and ensure that it | |
1686 | * gets run. | |
1687 | */ | |
9467f859 | 1688 | static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node) |
484b4061 | 1689 | { |
9467f859 | 1690 | struct blk_mq_hw_ctx *hctx; |
484b4061 JA |
1691 | struct blk_mq_ctx *ctx; |
1692 | LIST_HEAD(tmp); | |
1693 | ||
9467f859 | 1694 | hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead); |
e57690fe | 1695 | ctx = __blk_mq_get_ctx(hctx->queue, cpu); |
484b4061 JA |
1696 | |
1697 | spin_lock(&ctx->lock); | |
1698 | if (!list_empty(&ctx->rq_list)) { | |
1699 | list_splice_init(&ctx->rq_list, &tmp); | |
1700 | blk_mq_hctx_clear_pending(hctx, ctx); | |
1701 | } | |
1702 | spin_unlock(&ctx->lock); | |
1703 | ||
1704 | if (list_empty(&tmp)) | |
9467f859 | 1705 | return 0; |
484b4061 | 1706 | |
e57690fe JA |
1707 | spin_lock(&hctx->lock); |
1708 | list_splice_tail_init(&tmp, &hctx->dispatch); | |
1709 | spin_unlock(&hctx->lock); | |
484b4061 JA |
1710 | |
1711 | blk_mq_run_hw_queue(hctx, true); | |
9467f859 | 1712 | return 0; |
484b4061 JA |
1713 | } |
1714 | ||
9467f859 | 1715 | static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx) |
484b4061 | 1716 | { |
9467f859 TG |
1717 | cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD, |
1718 | &hctx->cpuhp_dead); | |
484b4061 JA |
1719 | } |
1720 | ||
c3b4afca | 1721 | /* hctx->ctxs will be freed in queue's release handler */ |
08e98fc6 ML |
1722 | static void blk_mq_exit_hctx(struct request_queue *q, |
1723 | struct blk_mq_tag_set *set, | |
1724 | struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx) | |
1725 | { | |
f70ced09 ML |
1726 | unsigned flush_start_tag = set->queue_depth; |
1727 | ||
08e98fc6 ML |
1728 | blk_mq_tag_idle(hctx); |
1729 | ||
f70ced09 ML |
1730 | if (set->ops->exit_request) |
1731 | set->ops->exit_request(set->driver_data, | |
1732 | hctx->fq->flush_rq, hctx_idx, | |
1733 | flush_start_tag + hctx_idx); | |
1734 | ||
08e98fc6 ML |
1735 | if (set->ops->exit_hctx) |
1736 | set->ops->exit_hctx(hctx, hctx_idx); | |
1737 | ||
6a83e74d BVA |
1738 | if (hctx->flags & BLK_MQ_F_BLOCKING) |
1739 | cleanup_srcu_struct(&hctx->queue_rq_srcu); | |
1740 | ||
9467f859 | 1741 | blk_mq_remove_cpuhp(hctx); |
f70ced09 | 1742 | blk_free_flush_queue(hctx->fq); |
88459642 | 1743 | sbitmap_free(&hctx->ctx_map); |
08e98fc6 ML |
1744 | } |
1745 | ||
624dbe47 ML |
1746 | static void blk_mq_exit_hw_queues(struct request_queue *q, |
1747 | struct blk_mq_tag_set *set, int nr_queue) | |
1748 | { | |
1749 | struct blk_mq_hw_ctx *hctx; | |
1750 | unsigned int i; | |
1751 | ||
1752 | queue_for_each_hw_ctx(q, hctx, i) { | |
1753 | if (i == nr_queue) | |
1754 | break; | |
08e98fc6 | 1755 | blk_mq_exit_hctx(q, set, hctx, i); |
624dbe47 | 1756 | } |
624dbe47 ML |
1757 | } |
1758 | ||
1759 | static void blk_mq_free_hw_queues(struct request_queue *q, | |
1760 | struct blk_mq_tag_set *set) | |
1761 | { | |
1762 | struct blk_mq_hw_ctx *hctx; | |
1763 | unsigned int i; | |
1764 | ||
e09aae7e | 1765 | queue_for_each_hw_ctx(q, hctx, i) |
624dbe47 | 1766 | free_cpumask_var(hctx->cpumask); |
624dbe47 ML |
1767 | } |
1768 | ||
08e98fc6 ML |
1769 | static int blk_mq_init_hctx(struct request_queue *q, |
1770 | struct blk_mq_tag_set *set, | |
1771 | struct blk_mq_hw_ctx *hctx, unsigned hctx_idx) | |
320ae51f | 1772 | { |
08e98fc6 | 1773 | int node; |
f70ced09 | 1774 | unsigned flush_start_tag = set->queue_depth; |
08e98fc6 ML |
1775 | |
1776 | node = hctx->numa_node; | |
1777 | if (node == NUMA_NO_NODE) | |
1778 | node = hctx->numa_node = set->numa_node; | |
1779 | ||
27489a3c | 1780 | INIT_WORK(&hctx->run_work, blk_mq_run_work_fn); |
08e98fc6 ML |
1781 | INIT_DELAYED_WORK(&hctx->delay_work, blk_mq_delay_work_fn); |
1782 | spin_lock_init(&hctx->lock); | |
1783 | INIT_LIST_HEAD(&hctx->dispatch); | |
1784 | hctx->queue = q; | |
1785 | hctx->queue_num = hctx_idx; | |
2404e607 | 1786 | hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED; |
08e98fc6 | 1787 | |
9467f859 | 1788 | cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead); |
08e98fc6 ML |
1789 | |
1790 | hctx->tags = set->tags[hctx_idx]; | |
320ae51f JA |
1791 | |
1792 | /* | |
08e98fc6 ML |
1793 | * Allocate space for all possible cpus to avoid allocation at |
1794 | * runtime | |
320ae51f | 1795 | */ |
08e98fc6 ML |
1796 | hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *), |
1797 | GFP_KERNEL, node); | |
1798 | if (!hctx->ctxs) | |
1799 | goto unregister_cpu_notifier; | |
320ae51f | 1800 | |
88459642 OS |
1801 | if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8), GFP_KERNEL, |
1802 | node)) | |
08e98fc6 | 1803 | goto free_ctxs; |
320ae51f | 1804 | |
08e98fc6 | 1805 | hctx->nr_ctx = 0; |
320ae51f | 1806 | |
08e98fc6 ML |
1807 | if (set->ops->init_hctx && |
1808 | set->ops->init_hctx(hctx, set->driver_data, hctx_idx)) | |
1809 | goto free_bitmap; | |
320ae51f | 1810 | |
f70ced09 ML |
1811 | hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size); |
1812 | if (!hctx->fq) | |
1813 | goto exit_hctx; | |
320ae51f | 1814 | |
f70ced09 ML |
1815 | if (set->ops->init_request && |
1816 | set->ops->init_request(set->driver_data, | |
1817 | hctx->fq->flush_rq, hctx_idx, | |
1818 | flush_start_tag + hctx_idx, node)) | |
1819 | goto free_fq; | |
320ae51f | 1820 | |
6a83e74d BVA |
1821 | if (hctx->flags & BLK_MQ_F_BLOCKING) |
1822 | init_srcu_struct(&hctx->queue_rq_srcu); | |
1823 | ||
08e98fc6 | 1824 | return 0; |
320ae51f | 1825 | |
f70ced09 ML |
1826 | free_fq: |
1827 | kfree(hctx->fq); | |
1828 | exit_hctx: | |
1829 | if (set->ops->exit_hctx) | |
1830 | set->ops->exit_hctx(hctx, hctx_idx); | |
08e98fc6 | 1831 | free_bitmap: |
88459642 | 1832 | sbitmap_free(&hctx->ctx_map); |
08e98fc6 ML |
1833 | free_ctxs: |
1834 | kfree(hctx->ctxs); | |
1835 | unregister_cpu_notifier: | |
9467f859 | 1836 | blk_mq_remove_cpuhp(hctx); |
08e98fc6 ML |
1837 | return -1; |
1838 | } | |
320ae51f | 1839 | |
320ae51f JA |
1840 | static void blk_mq_init_cpu_queues(struct request_queue *q, |
1841 | unsigned int nr_hw_queues) | |
1842 | { | |
1843 | unsigned int i; | |
1844 | ||
1845 | for_each_possible_cpu(i) { | |
1846 | struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i); | |
1847 | struct blk_mq_hw_ctx *hctx; | |
1848 | ||
1849 | memset(__ctx, 0, sizeof(*__ctx)); | |
1850 | __ctx->cpu = i; | |
1851 | spin_lock_init(&__ctx->lock); | |
1852 | INIT_LIST_HEAD(&__ctx->rq_list); | |
1853 | __ctx->queue = q; | |
cf43e6be JA |
1854 | blk_stat_init(&__ctx->stat[BLK_STAT_READ]); |
1855 | blk_stat_init(&__ctx->stat[BLK_STAT_WRITE]); | |
320ae51f JA |
1856 | |
1857 | /* If the cpu isn't online, the cpu is mapped to first hctx */ | |
320ae51f JA |
1858 | if (!cpu_online(i)) |
1859 | continue; | |
1860 | ||
7d7e0f90 | 1861 | hctx = blk_mq_map_queue(q, i); |
e4043dcf | 1862 | |
320ae51f JA |
1863 | /* |
1864 | * Set local node, IFF we have more than one hw queue. If | |
1865 | * not, we remain on the home node of the device | |
1866 | */ | |
1867 | if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE) | |
bffed457 | 1868 | hctx->numa_node = local_memory_node(cpu_to_node(i)); |
320ae51f JA |
1869 | } |
1870 | } | |
1871 | ||
5778322e AM |
1872 | static void blk_mq_map_swqueue(struct request_queue *q, |
1873 | const struct cpumask *online_mask) | |
320ae51f | 1874 | { |
d1b1cea1 | 1875 | unsigned int i, hctx_idx; |
320ae51f JA |
1876 | struct blk_mq_hw_ctx *hctx; |
1877 | struct blk_mq_ctx *ctx; | |
2a34c087 | 1878 | struct blk_mq_tag_set *set = q->tag_set; |
320ae51f | 1879 | |
60de074b AM |
1880 | /* |
1881 | * Avoid others reading imcomplete hctx->cpumask through sysfs | |
1882 | */ | |
1883 | mutex_lock(&q->sysfs_lock); | |
1884 | ||
320ae51f | 1885 | queue_for_each_hw_ctx(q, hctx, i) { |
e4043dcf | 1886 | cpumask_clear(hctx->cpumask); |
320ae51f JA |
1887 | hctx->nr_ctx = 0; |
1888 | } | |
1889 | ||
1890 | /* | |
1891 | * Map software to hardware queues | |
1892 | */ | |
897bb0c7 | 1893 | for_each_possible_cpu(i) { |
320ae51f | 1894 | /* If the cpu isn't online, the cpu is mapped to first hctx */ |
5778322e | 1895 | if (!cpumask_test_cpu(i, online_mask)) |
e4043dcf JA |
1896 | continue; |
1897 | ||
d1b1cea1 GKB |
1898 | hctx_idx = q->mq_map[i]; |
1899 | /* unmapped hw queue can be remapped after CPU topo changed */ | |
1900 | if (!set->tags[hctx_idx]) { | |
1901 | set->tags[hctx_idx] = blk_mq_init_rq_map(set, hctx_idx); | |
1902 | ||
1903 | /* | |
1904 | * If tags initialization fail for some hctx, | |
1905 | * that hctx won't be brought online. In this | |
1906 | * case, remap the current ctx to hctx[0] which | |
1907 | * is guaranteed to always have tags allocated | |
1908 | */ | |
1909 | if (!set->tags[hctx_idx]) | |
1910 | q->mq_map[i] = 0; | |
1911 | } | |
1912 | ||
897bb0c7 | 1913 | ctx = per_cpu_ptr(q->queue_ctx, i); |
7d7e0f90 | 1914 | hctx = blk_mq_map_queue(q, i); |
868f2f0b | 1915 | |
e4043dcf | 1916 | cpumask_set_cpu(i, hctx->cpumask); |
320ae51f JA |
1917 | ctx->index_hw = hctx->nr_ctx; |
1918 | hctx->ctxs[hctx->nr_ctx++] = ctx; | |
1919 | } | |
506e931f | 1920 | |
60de074b AM |
1921 | mutex_unlock(&q->sysfs_lock); |
1922 | ||
506e931f | 1923 | queue_for_each_hw_ctx(q, hctx, i) { |
484b4061 | 1924 | /* |
a68aafa5 JA |
1925 | * If no software queues are mapped to this hardware queue, |
1926 | * disable it and free the request entries. | |
484b4061 JA |
1927 | */ |
1928 | if (!hctx->nr_ctx) { | |
d1b1cea1 GKB |
1929 | /* Never unmap queue 0. We need it as a |
1930 | * fallback in case of a new remap fails | |
1931 | * allocation | |
1932 | */ | |
1933 | if (i && set->tags[i]) { | |
484b4061 JA |
1934 | blk_mq_free_rq_map(set, set->tags[i], i); |
1935 | set->tags[i] = NULL; | |
484b4061 | 1936 | } |
2a34c087 | 1937 | hctx->tags = NULL; |
484b4061 JA |
1938 | continue; |
1939 | } | |
1940 | ||
2a34c087 ML |
1941 | hctx->tags = set->tags[i]; |
1942 | WARN_ON(!hctx->tags); | |
1943 | ||
889fa31f CY |
1944 | /* |
1945 | * Set the map size to the number of mapped software queues. | |
1946 | * This is more accurate and more efficient than looping | |
1947 | * over all possibly mapped software queues. | |
1948 | */ | |
88459642 | 1949 | sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx); |
889fa31f | 1950 | |
484b4061 JA |
1951 | /* |
1952 | * Initialize batch roundrobin counts | |
1953 | */ | |
506e931f JA |
1954 | hctx->next_cpu = cpumask_first(hctx->cpumask); |
1955 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; | |
1956 | } | |
320ae51f JA |
1957 | } |
1958 | ||
2404e607 | 1959 | static void queue_set_hctx_shared(struct request_queue *q, bool shared) |
0d2602ca JA |
1960 | { |
1961 | struct blk_mq_hw_ctx *hctx; | |
0d2602ca JA |
1962 | int i; |
1963 | ||
2404e607 JM |
1964 | queue_for_each_hw_ctx(q, hctx, i) { |
1965 | if (shared) | |
1966 | hctx->flags |= BLK_MQ_F_TAG_SHARED; | |
1967 | else | |
1968 | hctx->flags &= ~BLK_MQ_F_TAG_SHARED; | |
1969 | } | |
1970 | } | |
1971 | ||
1972 | static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set, bool shared) | |
1973 | { | |
1974 | struct request_queue *q; | |
0d2602ca JA |
1975 | |
1976 | list_for_each_entry(q, &set->tag_list, tag_set_list) { | |
1977 | blk_mq_freeze_queue(q); | |
2404e607 | 1978 | queue_set_hctx_shared(q, shared); |
0d2602ca JA |
1979 | blk_mq_unfreeze_queue(q); |
1980 | } | |
1981 | } | |
1982 | ||
1983 | static void blk_mq_del_queue_tag_set(struct request_queue *q) | |
1984 | { | |
1985 | struct blk_mq_tag_set *set = q->tag_set; | |
1986 | ||
0d2602ca JA |
1987 | mutex_lock(&set->tag_list_lock); |
1988 | list_del_init(&q->tag_set_list); | |
2404e607 JM |
1989 | if (list_is_singular(&set->tag_list)) { |
1990 | /* just transitioned to unshared */ | |
1991 | set->flags &= ~BLK_MQ_F_TAG_SHARED; | |
1992 | /* update existing queue */ | |
1993 | blk_mq_update_tag_set_depth(set, false); | |
1994 | } | |
0d2602ca | 1995 | mutex_unlock(&set->tag_list_lock); |
0d2602ca JA |
1996 | } |
1997 | ||
1998 | static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set, | |
1999 | struct request_queue *q) | |
2000 | { | |
2001 | q->tag_set = set; | |
2002 | ||
2003 | mutex_lock(&set->tag_list_lock); | |
2404e607 JM |
2004 | |
2005 | /* Check to see if we're transitioning to shared (from 1 to 2 queues). */ | |
2006 | if (!list_empty(&set->tag_list) && !(set->flags & BLK_MQ_F_TAG_SHARED)) { | |
2007 | set->flags |= BLK_MQ_F_TAG_SHARED; | |
2008 | /* update existing queue */ | |
2009 | blk_mq_update_tag_set_depth(set, true); | |
2010 | } | |
2011 | if (set->flags & BLK_MQ_F_TAG_SHARED) | |
2012 | queue_set_hctx_shared(q, true); | |
0d2602ca | 2013 | list_add_tail(&q->tag_set_list, &set->tag_list); |
2404e607 | 2014 | |
0d2602ca JA |
2015 | mutex_unlock(&set->tag_list_lock); |
2016 | } | |
2017 | ||
e09aae7e ML |
2018 | /* |
2019 | * It is the actual release handler for mq, but we do it from | |
2020 | * request queue's release handler for avoiding use-after-free | |
2021 | * and headache because q->mq_kobj shouldn't have been introduced, | |
2022 | * but we can't group ctx/kctx kobj without it. | |
2023 | */ | |
2024 | void blk_mq_release(struct request_queue *q) | |
2025 | { | |
2026 | struct blk_mq_hw_ctx *hctx; | |
2027 | unsigned int i; | |
2028 | ||
2029 | /* hctx kobj stays in hctx */ | |
c3b4afca ML |
2030 | queue_for_each_hw_ctx(q, hctx, i) { |
2031 | if (!hctx) | |
2032 | continue; | |
2033 | kfree(hctx->ctxs); | |
e09aae7e | 2034 | kfree(hctx); |
c3b4afca | 2035 | } |
e09aae7e | 2036 | |
a723bab3 AM |
2037 | q->mq_map = NULL; |
2038 | ||
e09aae7e ML |
2039 | kfree(q->queue_hw_ctx); |
2040 | ||
2041 | /* ctx kobj stays in queue_ctx */ | |
2042 | free_percpu(q->queue_ctx); | |
2043 | } | |
2044 | ||
24d2f903 | 2045 | struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set) |
b62c21b7 MS |
2046 | { |
2047 | struct request_queue *uninit_q, *q; | |
2048 | ||
2049 | uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node); | |
2050 | if (!uninit_q) | |
2051 | return ERR_PTR(-ENOMEM); | |
2052 | ||
2053 | q = blk_mq_init_allocated_queue(set, uninit_q); | |
2054 | if (IS_ERR(q)) | |
2055 | blk_cleanup_queue(uninit_q); | |
2056 | ||
2057 | return q; | |
2058 | } | |
2059 | EXPORT_SYMBOL(blk_mq_init_queue); | |
2060 | ||
868f2f0b KB |
2061 | static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set, |
2062 | struct request_queue *q) | |
320ae51f | 2063 | { |
868f2f0b KB |
2064 | int i, j; |
2065 | struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx; | |
f14bbe77 | 2066 | |
868f2f0b | 2067 | blk_mq_sysfs_unregister(q); |
24d2f903 | 2068 | for (i = 0; i < set->nr_hw_queues; i++) { |
868f2f0b | 2069 | int node; |
f14bbe77 | 2070 | |
868f2f0b KB |
2071 | if (hctxs[i]) |
2072 | continue; | |
2073 | ||
2074 | node = blk_mq_hw_queue_to_node(q->mq_map, i); | |
cdef54dd CH |
2075 | hctxs[i] = kzalloc_node(sizeof(struct blk_mq_hw_ctx), |
2076 | GFP_KERNEL, node); | |
320ae51f | 2077 | if (!hctxs[i]) |
868f2f0b | 2078 | break; |
320ae51f | 2079 | |
a86073e4 | 2080 | if (!zalloc_cpumask_var_node(&hctxs[i]->cpumask, GFP_KERNEL, |
868f2f0b KB |
2081 | node)) { |
2082 | kfree(hctxs[i]); | |
2083 | hctxs[i] = NULL; | |
2084 | break; | |
2085 | } | |
e4043dcf | 2086 | |
0d2602ca | 2087 | atomic_set(&hctxs[i]->nr_active, 0); |
f14bbe77 | 2088 | hctxs[i]->numa_node = node; |
320ae51f | 2089 | hctxs[i]->queue_num = i; |
868f2f0b KB |
2090 | |
2091 | if (blk_mq_init_hctx(q, set, hctxs[i], i)) { | |
2092 | free_cpumask_var(hctxs[i]->cpumask); | |
2093 | kfree(hctxs[i]); | |
2094 | hctxs[i] = NULL; | |
2095 | break; | |
2096 | } | |
2097 | blk_mq_hctx_kobj_init(hctxs[i]); | |
320ae51f | 2098 | } |
868f2f0b KB |
2099 | for (j = i; j < q->nr_hw_queues; j++) { |
2100 | struct blk_mq_hw_ctx *hctx = hctxs[j]; | |
2101 | ||
2102 | if (hctx) { | |
2103 | if (hctx->tags) { | |
2104 | blk_mq_free_rq_map(set, hctx->tags, j); | |
2105 | set->tags[j] = NULL; | |
2106 | } | |
2107 | blk_mq_exit_hctx(q, set, hctx, j); | |
2108 | free_cpumask_var(hctx->cpumask); | |
2109 | kobject_put(&hctx->kobj); | |
2110 | kfree(hctx->ctxs); | |
2111 | kfree(hctx); | |
2112 | hctxs[j] = NULL; | |
2113 | ||
2114 | } | |
2115 | } | |
2116 | q->nr_hw_queues = i; | |
2117 | blk_mq_sysfs_register(q); | |
2118 | } | |
2119 | ||
2120 | struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set, | |
2121 | struct request_queue *q) | |
2122 | { | |
66841672 ML |
2123 | /* mark the queue as mq asap */ |
2124 | q->mq_ops = set->ops; | |
2125 | ||
868f2f0b KB |
2126 | q->queue_ctx = alloc_percpu(struct blk_mq_ctx); |
2127 | if (!q->queue_ctx) | |
c7de5726 | 2128 | goto err_exit; |
868f2f0b KB |
2129 | |
2130 | q->queue_hw_ctx = kzalloc_node(nr_cpu_ids * sizeof(*(q->queue_hw_ctx)), | |
2131 | GFP_KERNEL, set->numa_node); | |
2132 | if (!q->queue_hw_ctx) | |
2133 | goto err_percpu; | |
2134 | ||
bdd17e75 | 2135 | q->mq_map = set->mq_map; |
868f2f0b KB |
2136 | |
2137 | blk_mq_realloc_hw_ctxs(set, q); | |
2138 | if (!q->nr_hw_queues) | |
2139 | goto err_hctxs; | |
320ae51f | 2140 | |
287922eb | 2141 | INIT_WORK(&q->timeout_work, blk_mq_timeout_work); |
e56f698b | 2142 | blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ); |
320ae51f JA |
2143 | |
2144 | q->nr_queues = nr_cpu_ids; | |
320ae51f | 2145 | |
94eddfbe | 2146 | q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT; |
320ae51f | 2147 | |
05f1dd53 JA |
2148 | if (!(set->flags & BLK_MQ_F_SG_MERGE)) |
2149 | q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE; | |
2150 | ||
1be036e9 CH |
2151 | q->sg_reserved_size = INT_MAX; |
2152 | ||
2849450a | 2153 | INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work); |
6fca6a61 CH |
2154 | INIT_LIST_HEAD(&q->requeue_list); |
2155 | spin_lock_init(&q->requeue_lock); | |
2156 | ||
07068d5b JA |
2157 | if (q->nr_hw_queues > 1) |
2158 | blk_queue_make_request(q, blk_mq_make_request); | |
2159 | else | |
2160 | blk_queue_make_request(q, blk_sq_make_request); | |
2161 | ||
eba71768 JA |
2162 | /* |
2163 | * Do this after blk_queue_make_request() overrides it... | |
2164 | */ | |
2165 | q->nr_requests = set->queue_depth; | |
2166 | ||
64f1c21e JA |
2167 | /* |
2168 | * Default to classic polling | |
2169 | */ | |
2170 | q->poll_nsec = -1; | |
2171 | ||
24d2f903 CH |
2172 | if (set->ops->complete) |
2173 | blk_queue_softirq_done(q, set->ops->complete); | |
30a91cb4 | 2174 | |
24d2f903 | 2175 | blk_mq_init_cpu_queues(q, set->nr_hw_queues); |
320ae51f | 2176 | |
5778322e | 2177 | get_online_cpus(); |
320ae51f | 2178 | mutex_lock(&all_q_mutex); |
320ae51f | 2179 | |
4593fdbe | 2180 | list_add_tail(&q->all_q_node, &all_q_list); |
0d2602ca | 2181 | blk_mq_add_queue_tag_set(set, q); |
5778322e | 2182 | blk_mq_map_swqueue(q, cpu_online_mask); |
484b4061 | 2183 | |
4593fdbe | 2184 | mutex_unlock(&all_q_mutex); |
5778322e | 2185 | put_online_cpus(); |
4593fdbe | 2186 | |
320ae51f | 2187 | return q; |
18741986 | 2188 | |
320ae51f | 2189 | err_hctxs: |
868f2f0b | 2190 | kfree(q->queue_hw_ctx); |
320ae51f | 2191 | err_percpu: |
868f2f0b | 2192 | free_percpu(q->queue_ctx); |
c7de5726 ML |
2193 | err_exit: |
2194 | q->mq_ops = NULL; | |
320ae51f JA |
2195 | return ERR_PTR(-ENOMEM); |
2196 | } | |
b62c21b7 | 2197 | EXPORT_SYMBOL(blk_mq_init_allocated_queue); |
320ae51f JA |
2198 | |
2199 | void blk_mq_free_queue(struct request_queue *q) | |
2200 | { | |
624dbe47 | 2201 | struct blk_mq_tag_set *set = q->tag_set; |
320ae51f | 2202 | |
0e626368 AM |
2203 | mutex_lock(&all_q_mutex); |
2204 | list_del_init(&q->all_q_node); | |
2205 | mutex_unlock(&all_q_mutex); | |
2206 | ||
87760e5e JA |
2207 | wbt_exit(q); |
2208 | ||
0d2602ca JA |
2209 | blk_mq_del_queue_tag_set(q); |
2210 | ||
624dbe47 ML |
2211 | blk_mq_exit_hw_queues(q, set, set->nr_hw_queues); |
2212 | blk_mq_free_hw_queues(q, set); | |
320ae51f | 2213 | } |
320ae51f JA |
2214 | |
2215 | /* Basically redo blk_mq_init_queue with queue frozen */ | |
5778322e AM |
2216 | static void blk_mq_queue_reinit(struct request_queue *q, |
2217 | const struct cpumask *online_mask) | |
320ae51f | 2218 | { |
4ecd4fef | 2219 | WARN_ON_ONCE(!atomic_read(&q->mq_freeze_depth)); |
320ae51f | 2220 | |
67aec14c JA |
2221 | blk_mq_sysfs_unregister(q); |
2222 | ||
320ae51f JA |
2223 | /* |
2224 | * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe | |
2225 | * we should change hctx numa_node according to new topology (this | |
2226 | * involves free and re-allocate memory, worthy doing?) | |
2227 | */ | |
2228 | ||
5778322e | 2229 | blk_mq_map_swqueue(q, online_mask); |
320ae51f | 2230 | |
67aec14c | 2231 | blk_mq_sysfs_register(q); |
320ae51f JA |
2232 | } |
2233 | ||
65d5291e SAS |
2234 | /* |
2235 | * New online cpumask which is going to be set in this hotplug event. | |
2236 | * Declare this cpumasks as global as cpu-hotplug operation is invoked | |
2237 | * one-by-one and dynamically allocating this could result in a failure. | |
2238 | */ | |
2239 | static struct cpumask cpuhp_online_new; | |
2240 | ||
2241 | static void blk_mq_queue_reinit_work(void) | |
320ae51f JA |
2242 | { |
2243 | struct request_queue *q; | |
320ae51f JA |
2244 | |
2245 | mutex_lock(&all_q_mutex); | |
f3af020b TH |
2246 | /* |
2247 | * We need to freeze and reinit all existing queues. Freezing | |
2248 | * involves synchronous wait for an RCU grace period and doing it | |
2249 | * one by one may take a long time. Start freezing all queues in | |
2250 | * one swoop and then wait for the completions so that freezing can | |
2251 | * take place in parallel. | |
2252 | */ | |
2253 | list_for_each_entry(q, &all_q_list, all_q_node) | |
2254 | blk_mq_freeze_queue_start(q); | |
415d3dab | 2255 | list_for_each_entry(q, &all_q_list, all_q_node) |
f3af020b TH |
2256 | blk_mq_freeze_queue_wait(q); |
2257 | ||
320ae51f | 2258 | list_for_each_entry(q, &all_q_list, all_q_node) |
65d5291e | 2259 | blk_mq_queue_reinit(q, &cpuhp_online_new); |
f3af020b TH |
2260 | |
2261 | list_for_each_entry(q, &all_q_list, all_q_node) | |
2262 | blk_mq_unfreeze_queue(q); | |
2263 | ||
320ae51f | 2264 | mutex_unlock(&all_q_mutex); |
65d5291e SAS |
2265 | } |
2266 | ||
2267 | static int blk_mq_queue_reinit_dead(unsigned int cpu) | |
2268 | { | |
97a32864 | 2269 | cpumask_copy(&cpuhp_online_new, cpu_online_mask); |
65d5291e SAS |
2270 | blk_mq_queue_reinit_work(); |
2271 | return 0; | |
2272 | } | |
2273 | ||
2274 | /* | |
2275 | * Before hotadded cpu starts handling requests, new mappings must be | |
2276 | * established. Otherwise, these requests in hw queue might never be | |
2277 | * dispatched. | |
2278 | * | |
2279 | * For example, there is a single hw queue (hctx) and two CPU queues (ctx0 | |
2280 | * for CPU0, and ctx1 for CPU1). | |
2281 | * | |
2282 | * Now CPU1 is just onlined and a request is inserted into ctx1->rq_list | |
2283 | * and set bit0 in pending bitmap as ctx1->index_hw is still zero. | |
2284 | * | |
2c3ad667 JA |
2285 | * And then while running hw queue, blk_mq_flush_busy_ctxs() finds bit0 is set |
2286 | * in pending bitmap and tries to retrieve requests in hctx->ctxs[0]->rq_list. | |
2287 | * But htx->ctxs[0] is a pointer to ctx0, so the request in ctx1->rq_list is | |
2288 | * ignored. | |
65d5291e SAS |
2289 | */ |
2290 | static int blk_mq_queue_reinit_prepare(unsigned int cpu) | |
2291 | { | |
2292 | cpumask_copy(&cpuhp_online_new, cpu_online_mask); | |
2293 | cpumask_set_cpu(cpu, &cpuhp_online_new); | |
2294 | blk_mq_queue_reinit_work(); | |
2295 | return 0; | |
320ae51f JA |
2296 | } |
2297 | ||
a5164405 JA |
2298 | static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set) |
2299 | { | |
2300 | int i; | |
2301 | ||
2302 | for (i = 0; i < set->nr_hw_queues; i++) { | |
2303 | set->tags[i] = blk_mq_init_rq_map(set, i); | |
2304 | if (!set->tags[i]) | |
2305 | goto out_unwind; | |
2306 | } | |
2307 | ||
2308 | return 0; | |
2309 | ||
2310 | out_unwind: | |
2311 | while (--i >= 0) | |
2312 | blk_mq_free_rq_map(set, set->tags[i], i); | |
2313 | ||
a5164405 JA |
2314 | return -ENOMEM; |
2315 | } | |
2316 | ||
2317 | /* | |
2318 | * Allocate the request maps associated with this tag_set. Note that this | |
2319 | * may reduce the depth asked for, if memory is tight. set->queue_depth | |
2320 | * will be updated to reflect the allocated depth. | |
2321 | */ | |
2322 | static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set) | |
2323 | { | |
2324 | unsigned int depth; | |
2325 | int err; | |
2326 | ||
2327 | depth = set->queue_depth; | |
2328 | do { | |
2329 | err = __blk_mq_alloc_rq_maps(set); | |
2330 | if (!err) | |
2331 | break; | |
2332 | ||
2333 | set->queue_depth >>= 1; | |
2334 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) { | |
2335 | err = -ENOMEM; | |
2336 | break; | |
2337 | } | |
2338 | } while (set->queue_depth); | |
2339 | ||
2340 | if (!set->queue_depth || err) { | |
2341 | pr_err("blk-mq: failed to allocate request map\n"); | |
2342 | return -ENOMEM; | |
2343 | } | |
2344 | ||
2345 | if (depth != set->queue_depth) | |
2346 | pr_info("blk-mq: reduced tag depth (%u -> %u)\n", | |
2347 | depth, set->queue_depth); | |
2348 | ||
2349 | return 0; | |
2350 | } | |
2351 | ||
a4391c64 JA |
2352 | /* |
2353 | * Alloc a tag set to be associated with one or more request queues. | |
2354 | * May fail with EINVAL for various error conditions. May adjust the | |
2355 | * requested depth down, if if it too large. In that case, the set | |
2356 | * value will be stored in set->queue_depth. | |
2357 | */ | |
24d2f903 CH |
2358 | int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set) |
2359 | { | |
da695ba2 CH |
2360 | int ret; |
2361 | ||
205fb5f5 BVA |
2362 | BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS); |
2363 | ||
24d2f903 CH |
2364 | if (!set->nr_hw_queues) |
2365 | return -EINVAL; | |
a4391c64 | 2366 | if (!set->queue_depth) |
24d2f903 CH |
2367 | return -EINVAL; |
2368 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) | |
2369 | return -EINVAL; | |
2370 | ||
7d7e0f90 | 2371 | if (!set->ops->queue_rq) |
24d2f903 CH |
2372 | return -EINVAL; |
2373 | ||
a4391c64 JA |
2374 | if (set->queue_depth > BLK_MQ_MAX_DEPTH) { |
2375 | pr_info("blk-mq: reduced tag depth to %u\n", | |
2376 | BLK_MQ_MAX_DEPTH); | |
2377 | set->queue_depth = BLK_MQ_MAX_DEPTH; | |
2378 | } | |
24d2f903 | 2379 | |
6637fadf SL |
2380 | /* |
2381 | * If a crashdump is active, then we are potentially in a very | |
2382 | * memory constrained environment. Limit us to 1 queue and | |
2383 | * 64 tags to prevent using too much memory. | |
2384 | */ | |
2385 | if (is_kdump_kernel()) { | |
2386 | set->nr_hw_queues = 1; | |
2387 | set->queue_depth = min(64U, set->queue_depth); | |
2388 | } | |
868f2f0b KB |
2389 | /* |
2390 | * There is no use for more h/w queues than cpus. | |
2391 | */ | |
2392 | if (set->nr_hw_queues > nr_cpu_ids) | |
2393 | set->nr_hw_queues = nr_cpu_ids; | |
6637fadf | 2394 | |
868f2f0b | 2395 | set->tags = kzalloc_node(nr_cpu_ids * sizeof(struct blk_mq_tags *), |
24d2f903 CH |
2396 | GFP_KERNEL, set->numa_node); |
2397 | if (!set->tags) | |
a5164405 | 2398 | return -ENOMEM; |
24d2f903 | 2399 | |
da695ba2 CH |
2400 | ret = -ENOMEM; |
2401 | set->mq_map = kzalloc_node(sizeof(*set->mq_map) * nr_cpu_ids, | |
2402 | GFP_KERNEL, set->numa_node); | |
bdd17e75 CH |
2403 | if (!set->mq_map) |
2404 | goto out_free_tags; | |
2405 | ||
da695ba2 CH |
2406 | if (set->ops->map_queues) |
2407 | ret = set->ops->map_queues(set); | |
2408 | else | |
2409 | ret = blk_mq_map_queues(set); | |
2410 | if (ret) | |
2411 | goto out_free_mq_map; | |
2412 | ||
2413 | ret = blk_mq_alloc_rq_maps(set); | |
2414 | if (ret) | |
bdd17e75 | 2415 | goto out_free_mq_map; |
24d2f903 | 2416 | |
0d2602ca JA |
2417 | mutex_init(&set->tag_list_lock); |
2418 | INIT_LIST_HEAD(&set->tag_list); | |
2419 | ||
24d2f903 | 2420 | return 0; |
bdd17e75 CH |
2421 | |
2422 | out_free_mq_map: | |
2423 | kfree(set->mq_map); | |
2424 | set->mq_map = NULL; | |
2425 | out_free_tags: | |
5676e7b6 RE |
2426 | kfree(set->tags); |
2427 | set->tags = NULL; | |
da695ba2 | 2428 | return ret; |
24d2f903 CH |
2429 | } |
2430 | EXPORT_SYMBOL(blk_mq_alloc_tag_set); | |
2431 | ||
2432 | void blk_mq_free_tag_set(struct blk_mq_tag_set *set) | |
2433 | { | |
2434 | int i; | |
2435 | ||
868f2f0b | 2436 | for (i = 0; i < nr_cpu_ids; i++) { |
f42d79ab | 2437 | if (set->tags[i]) |
484b4061 JA |
2438 | blk_mq_free_rq_map(set, set->tags[i], i); |
2439 | } | |
2440 | ||
bdd17e75 CH |
2441 | kfree(set->mq_map); |
2442 | set->mq_map = NULL; | |
2443 | ||
981bd189 | 2444 | kfree(set->tags); |
5676e7b6 | 2445 | set->tags = NULL; |
24d2f903 CH |
2446 | } |
2447 | EXPORT_SYMBOL(blk_mq_free_tag_set); | |
2448 | ||
e3a2b3f9 JA |
2449 | int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr) |
2450 | { | |
2451 | struct blk_mq_tag_set *set = q->tag_set; | |
2452 | struct blk_mq_hw_ctx *hctx; | |
2453 | int i, ret; | |
2454 | ||
2455 | if (!set || nr > set->queue_depth) | |
2456 | return -EINVAL; | |
2457 | ||
2458 | ret = 0; | |
2459 | queue_for_each_hw_ctx(q, hctx, i) { | |
e9137d4b KB |
2460 | if (!hctx->tags) |
2461 | continue; | |
e3a2b3f9 JA |
2462 | ret = blk_mq_tag_update_depth(hctx->tags, nr); |
2463 | if (ret) | |
2464 | break; | |
2465 | } | |
2466 | ||
2467 | if (!ret) | |
2468 | q->nr_requests = nr; | |
2469 | ||
2470 | return ret; | |
2471 | } | |
2472 | ||
868f2f0b KB |
2473 | void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues) |
2474 | { | |
2475 | struct request_queue *q; | |
2476 | ||
2477 | if (nr_hw_queues > nr_cpu_ids) | |
2478 | nr_hw_queues = nr_cpu_ids; | |
2479 | if (nr_hw_queues < 1 || nr_hw_queues == set->nr_hw_queues) | |
2480 | return; | |
2481 | ||
2482 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
2483 | blk_mq_freeze_queue(q); | |
2484 | ||
2485 | set->nr_hw_queues = nr_hw_queues; | |
2486 | list_for_each_entry(q, &set->tag_list, tag_set_list) { | |
2487 | blk_mq_realloc_hw_ctxs(set, q); | |
2488 | ||
2489 | if (q->nr_hw_queues > 1) | |
2490 | blk_queue_make_request(q, blk_mq_make_request); | |
2491 | else | |
2492 | blk_queue_make_request(q, blk_sq_make_request); | |
2493 | ||
2494 | blk_mq_queue_reinit(q, cpu_online_mask); | |
2495 | } | |
2496 | ||
2497 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
2498 | blk_mq_unfreeze_queue(q); | |
2499 | } | |
2500 | EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues); | |
2501 | ||
64f1c21e JA |
2502 | static unsigned long blk_mq_poll_nsecs(struct request_queue *q, |
2503 | struct blk_mq_hw_ctx *hctx, | |
2504 | struct request *rq) | |
2505 | { | |
2506 | struct blk_rq_stat stat[2]; | |
2507 | unsigned long ret = 0; | |
2508 | ||
2509 | /* | |
2510 | * If stats collection isn't on, don't sleep but turn it on for | |
2511 | * future users | |
2512 | */ | |
2513 | if (!blk_stat_enable(q)) | |
2514 | return 0; | |
2515 | ||
2516 | /* | |
2517 | * We don't have to do this once per IO, should optimize this | |
2518 | * to just use the current window of stats until it changes | |
2519 | */ | |
2520 | memset(&stat, 0, sizeof(stat)); | |
2521 | blk_hctx_stat_get(hctx, stat); | |
2522 | ||
2523 | /* | |
2524 | * As an optimistic guess, use half of the mean service time | |
2525 | * for this type of request. We can (and should) make this smarter. | |
2526 | * For instance, if the completion latencies are tight, we can | |
2527 | * get closer than just half the mean. This is especially | |
2528 | * important on devices where the completion latencies are longer | |
2529 | * than ~10 usec. | |
2530 | */ | |
2531 | if (req_op(rq) == REQ_OP_READ && stat[BLK_STAT_READ].nr_samples) | |
2532 | ret = (stat[BLK_STAT_READ].mean + 1) / 2; | |
2533 | else if (req_op(rq) == REQ_OP_WRITE && stat[BLK_STAT_WRITE].nr_samples) | |
2534 | ret = (stat[BLK_STAT_WRITE].mean + 1) / 2; | |
2535 | ||
2536 | return ret; | |
2537 | } | |
2538 | ||
06426adf | 2539 | static bool blk_mq_poll_hybrid_sleep(struct request_queue *q, |
64f1c21e | 2540 | struct blk_mq_hw_ctx *hctx, |
06426adf JA |
2541 | struct request *rq) |
2542 | { | |
2543 | struct hrtimer_sleeper hs; | |
2544 | enum hrtimer_mode mode; | |
64f1c21e | 2545 | unsigned int nsecs; |
06426adf JA |
2546 | ktime_t kt; |
2547 | ||
64f1c21e JA |
2548 | if (test_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags)) |
2549 | return false; | |
2550 | ||
2551 | /* | |
2552 | * poll_nsec can be: | |
2553 | * | |
2554 | * -1: don't ever hybrid sleep | |
2555 | * 0: use half of prev avg | |
2556 | * >0: use this specific value | |
2557 | */ | |
2558 | if (q->poll_nsec == -1) | |
2559 | return false; | |
2560 | else if (q->poll_nsec > 0) | |
2561 | nsecs = q->poll_nsec; | |
2562 | else | |
2563 | nsecs = blk_mq_poll_nsecs(q, hctx, rq); | |
2564 | ||
2565 | if (!nsecs) | |
06426adf JA |
2566 | return false; |
2567 | ||
2568 | set_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags); | |
2569 | ||
2570 | /* | |
2571 | * This will be replaced with the stats tracking code, using | |
2572 | * 'avg_completion_time / 2' as the pre-sleep target. | |
2573 | */ | |
8b0e1953 | 2574 | kt = nsecs; |
06426adf JA |
2575 | |
2576 | mode = HRTIMER_MODE_REL; | |
2577 | hrtimer_init_on_stack(&hs.timer, CLOCK_MONOTONIC, mode); | |
2578 | hrtimer_set_expires(&hs.timer, kt); | |
2579 | ||
2580 | hrtimer_init_sleeper(&hs, current); | |
2581 | do { | |
2582 | if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags)) | |
2583 | break; | |
2584 | set_current_state(TASK_UNINTERRUPTIBLE); | |
2585 | hrtimer_start_expires(&hs.timer, mode); | |
2586 | if (hs.task) | |
2587 | io_schedule(); | |
2588 | hrtimer_cancel(&hs.timer); | |
2589 | mode = HRTIMER_MODE_ABS; | |
2590 | } while (hs.task && !signal_pending(current)); | |
2591 | ||
2592 | __set_current_state(TASK_RUNNING); | |
2593 | destroy_hrtimer_on_stack(&hs.timer); | |
2594 | return true; | |
2595 | } | |
2596 | ||
bbd7bb70 JA |
2597 | static bool __blk_mq_poll(struct blk_mq_hw_ctx *hctx, struct request *rq) |
2598 | { | |
2599 | struct request_queue *q = hctx->queue; | |
2600 | long state; | |
2601 | ||
06426adf JA |
2602 | /* |
2603 | * If we sleep, have the caller restart the poll loop to reset | |
2604 | * the state. Like for the other success return cases, the | |
2605 | * caller is responsible for checking if the IO completed. If | |
2606 | * the IO isn't complete, we'll get called again and will go | |
2607 | * straight to the busy poll loop. | |
2608 | */ | |
64f1c21e | 2609 | if (blk_mq_poll_hybrid_sleep(q, hctx, rq)) |
06426adf JA |
2610 | return true; |
2611 | ||
bbd7bb70 JA |
2612 | hctx->poll_considered++; |
2613 | ||
2614 | state = current->state; | |
2615 | while (!need_resched()) { | |
2616 | int ret; | |
2617 | ||
2618 | hctx->poll_invoked++; | |
2619 | ||
2620 | ret = q->mq_ops->poll(hctx, rq->tag); | |
2621 | if (ret > 0) { | |
2622 | hctx->poll_success++; | |
2623 | set_current_state(TASK_RUNNING); | |
2624 | return true; | |
2625 | } | |
2626 | ||
2627 | if (signal_pending_state(state, current)) | |
2628 | set_current_state(TASK_RUNNING); | |
2629 | ||
2630 | if (current->state == TASK_RUNNING) | |
2631 | return true; | |
2632 | if (ret < 0) | |
2633 | break; | |
2634 | cpu_relax(); | |
2635 | } | |
2636 | ||
2637 | return false; | |
2638 | } | |
2639 | ||
2640 | bool blk_mq_poll(struct request_queue *q, blk_qc_t cookie) | |
2641 | { | |
2642 | struct blk_mq_hw_ctx *hctx; | |
2643 | struct blk_plug *plug; | |
2644 | struct request *rq; | |
2645 | ||
2646 | if (!q->mq_ops || !q->mq_ops->poll || !blk_qc_t_valid(cookie) || | |
2647 | !test_bit(QUEUE_FLAG_POLL, &q->queue_flags)) | |
2648 | return false; | |
2649 | ||
2650 | plug = current->plug; | |
2651 | if (plug) | |
2652 | blk_flush_plug_list(plug, false); | |
2653 | ||
2654 | hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)]; | |
2655 | rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie)); | |
2656 | ||
2657 | return __blk_mq_poll(hctx, rq); | |
2658 | } | |
2659 | EXPORT_SYMBOL_GPL(blk_mq_poll); | |
2660 | ||
676141e4 JA |
2661 | void blk_mq_disable_hotplug(void) |
2662 | { | |
2663 | mutex_lock(&all_q_mutex); | |
2664 | } | |
2665 | ||
2666 | void blk_mq_enable_hotplug(void) | |
2667 | { | |
2668 | mutex_unlock(&all_q_mutex); | |
2669 | } | |
2670 | ||
320ae51f JA |
2671 | static int __init blk_mq_init(void) |
2672 | { | |
9467f859 TG |
2673 | cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL, |
2674 | blk_mq_hctx_notify_dead); | |
320ae51f | 2675 | |
65d5291e SAS |
2676 | cpuhp_setup_state_nocalls(CPUHP_BLK_MQ_PREPARE, "block/mq:prepare", |
2677 | blk_mq_queue_reinit_prepare, | |
2678 | blk_mq_queue_reinit_dead); | |
320ae51f JA |
2679 | return 0; |
2680 | } | |
2681 | subsys_initcall(blk_mq_init); |