]>
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" |
bd166ef1 | 35 | #include "blk-mq-sched.h" |
320ae51f JA |
36 | |
37 | static DEFINE_MUTEX(all_q_mutex); | |
38 | static LIST_HEAD(all_q_list); | |
39 | ||
320ae51f JA |
40 | /* |
41 | * Check if any of the ctx's have pending work in this hardware queue | |
42 | */ | |
43 | static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx) | |
44 | { | |
bd166ef1 JA |
45 | return sbitmap_any_bit_set(&hctx->ctx_map) || |
46 | !list_empty_careful(&hctx->dispatch) || | |
47 | blk_mq_sched_has_work(hctx); | |
1429d7c9 JA |
48 | } |
49 | ||
320ae51f JA |
50 | /* |
51 | * Mark this ctx as having pending work in this hardware queue | |
52 | */ | |
53 | static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx, | |
54 | struct blk_mq_ctx *ctx) | |
55 | { | |
88459642 OS |
56 | if (!sbitmap_test_bit(&hctx->ctx_map, ctx->index_hw)) |
57 | sbitmap_set_bit(&hctx->ctx_map, ctx->index_hw); | |
1429d7c9 JA |
58 | } |
59 | ||
60 | static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx, | |
61 | struct blk_mq_ctx *ctx) | |
62 | { | |
88459642 | 63 | sbitmap_clear_bit(&hctx->ctx_map, ctx->index_hw); |
320ae51f JA |
64 | } |
65 | ||
b4c6a028 | 66 | void blk_mq_freeze_queue_start(struct request_queue *q) |
43a5e4e2 | 67 | { |
4ecd4fef | 68 | int freeze_depth; |
cddd5d17 | 69 | |
4ecd4fef CH |
70 | freeze_depth = atomic_inc_return(&q->mq_freeze_depth); |
71 | if (freeze_depth == 1) { | |
3ef28e83 | 72 | percpu_ref_kill(&q->q_usage_counter); |
b94ec296 | 73 | blk_mq_run_hw_queues(q, false); |
cddd5d17 | 74 | } |
f3af020b | 75 | } |
b4c6a028 | 76 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_start); |
f3af020b TH |
77 | |
78 | static void blk_mq_freeze_queue_wait(struct request_queue *q) | |
79 | { | |
3ef28e83 | 80 | wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter)); |
43a5e4e2 ML |
81 | } |
82 | ||
f3af020b TH |
83 | /* |
84 | * Guarantee no request is in use, so we can change any data structure of | |
85 | * the queue afterward. | |
86 | */ | |
3ef28e83 | 87 | void blk_freeze_queue(struct request_queue *q) |
f3af020b | 88 | { |
3ef28e83 DW |
89 | /* |
90 | * In the !blk_mq case we are only calling this to kill the | |
91 | * q_usage_counter, otherwise this increases the freeze depth | |
92 | * and waits for it to return to zero. For this reason there is | |
93 | * no blk_unfreeze_queue(), and blk_freeze_queue() is not | |
94 | * exported to drivers as the only user for unfreeze is blk_mq. | |
95 | */ | |
f3af020b TH |
96 | blk_mq_freeze_queue_start(q); |
97 | blk_mq_freeze_queue_wait(q); | |
98 | } | |
3ef28e83 DW |
99 | |
100 | void blk_mq_freeze_queue(struct request_queue *q) | |
101 | { | |
102 | /* | |
103 | * ...just an alias to keep freeze and unfreeze actions balanced | |
104 | * in the blk_mq_* namespace | |
105 | */ | |
106 | blk_freeze_queue(q); | |
107 | } | |
c761d96b | 108 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue); |
f3af020b | 109 | |
b4c6a028 | 110 | void blk_mq_unfreeze_queue(struct request_queue *q) |
320ae51f | 111 | { |
4ecd4fef | 112 | int freeze_depth; |
320ae51f | 113 | |
4ecd4fef CH |
114 | freeze_depth = atomic_dec_return(&q->mq_freeze_depth); |
115 | WARN_ON_ONCE(freeze_depth < 0); | |
116 | if (!freeze_depth) { | |
3ef28e83 | 117 | percpu_ref_reinit(&q->q_usage_counter); |
320ae51f | 118 | wake_up_all(&q->mq_freeze_wq); |
add703fd | 119 | } |
320ae51f | 120 | } |
b4c6a028 | 121 | EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue); |
320ae51f | 122 | |
6a83e74d BVA |
123 | /** |
124 | * blk_mq_quiesce_queue() - wait until all ongoing queue_rq calls have finished | |
125 | * @q: request queue. | |
126 | * | |
127 | * Note: this function does not prevent that the struct request end_io() | |
128 | * callback function is invoked. Additionally, it is not prevented that | |
129 | * new queue_rq() calls occur unless the queue has been stopped first. | |
130 | */ | |
131 | void blk_mq_quiesce_queue(struct request_queue *q) | |
132 | { | |
133 | struct blk_mq_hw_ctx *hctx; | |
134 | unsigned int i; | |
135 | bool rcu = false; | |
136 | ||
137 | blk_mq_stop_hw_queues(q); | |
138 | ||
139 | queue_for_each_hw_ctx(q, hctx, i) { | |
140 | if (hctx->flags & BLK_MQ_F_BLOCKING) | |
141 | synchronize_srcu(&hctx->queue_rq_srcu); | |
142 | else | |
143 | rcu = true; | |
144 | } | |
145 | if (rcu) | |
146 | synchronize_rcu(); | |
147 | } | |
148 | EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue); | |
149 | ||
aed3ea94 JA |
150 | void blk_mq_wake_waiters(struct request_queue *q) |
151 | { | |
152 | struct blk_mq_hw_ctx *hctx; | |
153 | unsigned int i; | |
154 | ||
155 | queue_for_each_hw_ctx(q, hctx, i) | |
156 | if (blk_mq_hw_queue_mapped(hctx)) | |
157 | blk_mq_tag_wakeup_all(hctx->tags, true); | |
3fd5940c KB |
158 | |
159 | /* | |
160 | * If we are called because the queue has now been marked as | |
161 | * dying, we need to ensure that processes currently waiting on | |
162 | * the queue are notified as well. | |
163 | */ | |
164 | wake_up_all(&q->mq_freeze_wq); | |
aed3ea94 JA |
165 | } |
166 | ||
320ae51f JA |
167 | bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx) |
168 | { | |
169 | return blk_mq_has_free_tags(hctx->tags); | |
170 | } | |
171 | EXPORT_SYMBOL(blk_mq_can_queue); | |
172 | ||
2c3ad667 JA |
173 | void blk_mq_rq_ctx_init(struct request_queue *q, struct blk_mq_ctx *ctx, |
174 | struct request *rq, unsigned int op) | |
320ae51f | 175 | { |
af76e555 CH |
176 | INIT_LIST_HEAD(&rq->queuelist); |
177 | /* csd/requeue_work/fifo_time is initialized before use */ | |
178 | rq->q = q; | |
320ae51f | 179 | rq->mq_ctx = ctx; |
ef295ecf | 180 | rq->cmd_flags = op; |
e8064021 CH |
181 | if (blk_queue_io_stat(q)) |
182 | rq->rq_flags |= RQF_IO_STAT; | |
af76e555 CH |
183 | /* do not touch atomic flags, it needs atomic ops against the timer */ |
184 | rq->cpu = -1; | |
af76e555 CH |
185 | INIT_HLIST_NODE(&rq->hash); |
186 | RB_CLEAR_NODE(&rq->rb_node); | |
af76e555 CH |
187 | rq->rq_disk = NULL; |
188 | rq->part = NULL; | |
3ee32372 | 189 | rq->start_time = jiffies; |
af76e555 CH |
190 | #ifdef CONFIG_BLK_CGROUP |
191 | rq->rl = NULL; | |
0fec08b4 | 192 | set_start_time_ns(rq); |
af76e555 CH |
193 | rq->io_start_time_ns = 0; |
194 | #endif | |
195 | rq->nr_phys_segments = 0; | |
196 | #if defined(CONFIG_BLK_DEV_INTEGRITY) | |
197 | rq->nr_integrity_segments = 0; | |
198 | #endif | |
af76e555 CH |
199 | rq->special = NULL; |
200 | /* tag was already set */ | |
201 | rq->errors = 0; | |
af76e555 | 202 | |
6f4a1626 TB |
203 | rq->cmd = rq->__cmd; |
204 | ||
af76e555 CH |
205 | rq->extra_len = 0; |
206 | rq->sense_len = 0; | |
207 | rq->resid_len = 0; | |
208 | rq->sense = NULL; | |
209 | ||
af76e555 | 210 | INIT_LIST_HEAD(&rq->timeout_list); |
f6be4fb4 JA |
211 | rq->timeout = 0; |
212 | ||
af76e555 CH |
213 | rq->end_io = NULL; |
214 | rq->end_io_data = NULL; | |
215 | rq->next_rq = NULL; | |
216 | ||
ef295ecf | 217 | ctx->rq_dispatched[op_is_sync(op)]++; |
320ae51f | 218 | } |
2c3ad667 | 219 | EXPORT_SYMBOL_GPL(blk_mq_rq_ctx_init); |
320ae51f | 220 | |
2c3ad667 JA |
221 | struct request *__blk_mq_alloc_request(struct blk_mq_alloc_data *data, |
222 | unsigned int op) | |
5dee8577 CH |
223 | { |
224 | struct request *rq; | |
225 | unsigned int tag; | |
226 | ||
cb96a42c | 227 | tag = blk_mq_get_tag(data); |
5dee8577 | 228 | if (tag != BLK_MQ_TAG_FAIL) { |
bd166ef1 JA |
229 | struct blk_mq_tags *tags = blk_mq_tags_from_data(data); |
230 | ||
231 | rq = tags->static_rqs[tag]; | |
5dee8577 | 232 | |
cb96a42c | 233 | if (blk_mq_tag_busy(data->hctx)) { |
e8064021 | 234 | rq->rq_flags = RQF_MQ_INFLIGHT; |
cb96a42c | 235 | atomic_inc(&data->hctx->nr_active); |
5dee8577 CH |
236 | } |
237 | ||
bd166ef1 JA |
238 | if (data->flags & BLK_MQ_REQ_INTERNAL) { |
239 | rq->tag = -1; | |
240 | rq->internal_tag = tag; | |
241 | } else { | |
242 | rq->tag = tag; | |
243 | rq->internal_tag = -1; | |
244 | } | |
245 | ||
ef295ecf | 246 | blk_mq_rq_ctx_init(data->q, data->ctx, rq, op); |
5dee8577 CH |
247 | return rq; |
248 | } | |
249 | ||
250 | return NULL; | |
251 | } | |
2c3ad667 | 252 | EXPORT_SYMBOL_GPL(__blk_mq_alloc_request); |
5dee8577 | 253 | |
6f3b0e8b CH |
254 | struct request *blk_mq_alloc_request(struct request_queue *q, int rw, |
255 | unsigned int flags) | |
320ae51f | 256 | { |
cb96a42c | 257 | struct blk_mq_alloc_data alloc_data; |
bd166ef1 | 258 | struct request *rq; |
a492f075 | 259 | int ret; |
320ae51f | 260 | |
6f3b0e8b | 261 | ret = blk_queue_enter(q, flags & BLK_MQ_REQ_NOWAIT); |
a492f075 JL |
262 | if (ret) |
263 | return ERR_PTR(ret); | |
320ae51f | 264 | |
bd166ef1 | 265 | rq = blk_mq_sched_get_request(q, NULL, rw, &alloc_data); |
841bac2c | 266 | |
bd166ef1 JA |
267 | blk_mq_put_ctx(alloc_data.ctx); |
268 | blk_queue_exit(q); | |
269 | ||
270 | if (!rq) | |
a492f075 | 271 | return ERR_PTR(-EWOULDBLOCK); |
0c4de0f3 CH |
272 | |
273 | rq->__data_len = 0; | |
274 | rq->__sector = (sector_t) -1; | |
275 | rq->bio = rq->biotail = NULL; | |
320ae51f JA |
276 | return rq; |
277 | } | |
4bb659b1 | 278 | EXPORT_SYMBOL(blk_mq_alloc_request); |
320ae51f | 279 | |
1f5bd336 ML |
280 | struct request *blk_mq_alloc_request_hctx(struct request_queue *q, int rw, |
281 | unsigned int flags, unsigned int hctx_idx) | |
282 | { | |
283 | struct blk_mq_hw_ctx *hctx; | |
284 | struct blk_mq_ctx *ctx; | |
285 | struct request *rq; | |
286 | struct blk_mq_alloc_data alloc_data; | |
287 | int ret; | |
288 | ||
289 | /* | |
290 | * If the tag allocator sleeps we could get an allocation for a | |
291 | * different hardware context. No need to complicate the low level | |
292 | * allocator for this for the rare use case of a command tied to | |
293 | * a specific queue. | |
294 | */ | |
295 | if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT))) | |
296 | return ERR_PTR(-EINVAL); | |
297 | ||
298 | if (hctx_idx >= q->nr_hw_queues) | |
299 | return ERR_PTR(-EIO); | |
300 | ||
301 | ret = blk_queue_enter(q, true); | |
302 | if (ret) | |
303 | return ERR_PTR(ret); | |
304 | ||
c8712c6a CH |
305 | /* |
306 | * Check if the hardware context is actually mapped to anything. | |
307 | * If not tell the caller that it should skip this queue. | |
308 | */ | |
1f5bd336 | 309 | hctx = q->queue_hw_ctx[hctx_idx]; |
c8712c6a CH |
310 | if (!blk_mq_hw_queue_mapped(hctx)) { |
311 | ret = -EXDEV; | |
312 | goto out_queue_exit; | |
313 | } | |
1f5bd336 ML |
314 | ctx = __blk_mq_get_ctx(q, cpumask_first(hctx->cpumask)); |
315 | ||
316 | blk_mq_set_alloc_data(&alloc_data, q, flags, ctx, hctx); | |
ef295ecf | 317 | rq = __blk_mq_alloc_request(&alloc_data, rw); |
1f5bd336 | 318 | if (!rq) { |
c8712c6a CH |
319 | ret = -EWOULDBLOCK; |
320 | goto out_queue_exit; | |
1f5bd336 ML |
321 | } |
322 | ||
323 | return rq; | |
c8712c6a CH |
324 | |
325 | out_queue_exit: | |
326 | blk_queue_exit(q); | |
327 | return ERR_PTR(ret); | |
1f5bd336 ML |
328 | } |
329 | EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx); | |
330 | ||
bd166ef1 JA |
331 | void __blk_mq_finish_request(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx, |
332 | struct request *rq) | |
320ae51f | 333 | { |
bd166ef1 | 334 | const int sched_tag = rq->internal_tag; |
320ae51f JA |
335 | struct request_queue *q = rq->q; |
336 | ||
e8064021 | 337 | if (rq->rq_flags & RQF_MQ_INFLIGHT) |
0d2602ca | 338 | atomic_dec(&hctx->nr_active); |
87760e5e JA |
339 | |
340 | wbt_done(q->rq_wb, &rq->issue_stat); | |
e8064021 | 341 | rq->rq_flags = 0; |
0d2602ca | 342 | |
af76e555 | 343 | clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags); |
06426adf | 344 | clear_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags); |
bd166ef1 JA |
345 | if (rq->tag != -1) |
346 | blk_mq_put_tag(hctx, hctx->tags, ctx, rq->tag); | |
347 | if (sched_tag != -1) | |
348 | blk_mq_sched_completed_request(hctx, rq); | |
3ef28e83 | 349 | blk_queue_exit(q); |
320ae51f JA |
350 | } |
351 | ||
bd166ef1 | 352 | static void blk_mq_finish_hctx_request(struct blk_mq_hw_ctx *hctx, |
16a3c2a7 | 353 | struct request *rq) |
320ae51f JA |
354 | { |
355 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
320ae51f JA |
356 | |
357 | ctx->rq_completed[rq_is_sync(rq)]++; | |
bd166ef1 JA |
358 | __blk_mq_finish_request(hctx, ctx, rq); |
359 | } | |
360 | ||
361 | void blk_mq_finish_request(struct request *rq) | |
362 | { | |
363 | blk_mq_finish_hctx_request(blk_mq_map_queue(rq->q, rq->mq_ctx->cpu), rq); | |
7c7f2f2b | 364 | } |
7c7f2f2b JA |
365 | |
366 | void blk_mq_free_request(struct request *rq) | |
367 | { | |
bd166ef1 | 368 | blk_mq_sched_put_request(rq); |
320ae51f | 369 | } |
1a3b595a | 370 | EXPORT_SYMBOL_GPL(blk_mq_free_request); |
320ae51f | 371 | |
c8a446ad | 372 | inline void __blk_mq_end_request(struct request *rq, int error) |
320ae51f | 373 | { |
0d11e6ac ML |
374 | blk_account_io_done(rq); |
375 | ||
91b63639 | 376 | if (rq->end_io) { |
87760e5e | 377 | wbt_done(rq->q->rq_wb, &rq->issue_stat); |
320ae51f | 378 | rq->end_io(rq, error); |
91b63639 CH |
379 | } else { |
380 | if (unlikely(blk_bidi_rq(rq))) | |
381 | blk_mq_free_request(rq->next_rq); | |
320ae51f | 382 | blk_mq_free_request(rq); |
91b63639 | 383 | } |
320ae51f | 384 | } |
c8a446ad | 385 | EXPORT_SYMBOL(__blk_mq_end_request); |
63151a44 | 386 | |
c8a446ad | 387 | void blk_mq_end_request(struct request *rq, int error) |
63151a44 CH |
388 | { |
389 | if (blk_update_request(rq, error, blk_rq_bytes(rq))) | |
390 | BUG(); | |
c8a446ad | 391 | __blk_mq_end_request(rq, error); |
63151a44 | 392 | } |
c8a446ad | 393 | EXPORT_SYMBOL(blk_mq_end_request); |
320ae51f | 394 | |
30a91cb4 | 395 | static void __blk_mq_complete_request_remote(void *data) |
320ae51f | 396 | { |
3d6efbf6 | 397 | struct request *rq = data; |
320ae51f | 398 | |
30a91cb4 | 399 | rq->q->softirq_done_fn(rq); |
320ae51f | 400 | } |
320ae51f | 401 | |
ed851860 | 402 | static void blk_mq_ipi_complete_request(struct request *rq) |
320ae51f JA |
403 | { |
404 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
38535201 | 405 | bool shared = false; |
320ae51f JA |
406 | int cpu; |
407 | ||
38535201 | 408 | if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) { |
30a91cb4 CH |
409 | rq->q->softirq_done_fn(rq); |
410 | return; | |
411 | } | |
320ae51f JA |
412 | |
413 | cpu = get_cpu(); | |
38535201 CH |
414 | if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags)) |
415 | shared = cpus_share_cache(cpu, ctx->cpu); | |
416 | ||
417 | if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) { | |
30a91cb4 | 418 | rq->csd.func = __blk_mq_complete_request_remote; |
3d6efbf6 CH |
419 | rq->csd.info = rq; |
420 | rq->csd.flags = 0; | |
c46fff2a | 421 | smp_call_function_single_async(ctx->cpu, &rq->csd); |
3d6efbf6 | 422 | } else { |
30a91cb4 | 423 | rq->q->softirq_done_fn(rq); |
3d6efbf6 | 424 | } |
320ae51f JA |
425 | put_cpu(); |
426 | } | |
30a91cb4 | 427 | |
cf43e6be JA |
428 | static void blk_mq_stat_add(struct request *rq) |
429 | { | |
430 | if (rq->rq_flags & RQF_STATS) { | |
431 | /* | |
432 | * We could rq->mq_ctx here, but there's less of a risk | |
433 | * of races if we have the completion event add the stats | |
434 | * to the local software queue. | |
435 | */ | |
436 | struct blk_mq_ctx *ctx; | |
437 | ||
438 | ctx = __blk_mq_get_ctx(rq->q, raw_smp_processor_id()); | |
439 | blk_stat_add(&ctx->stat[rq_data_dir(rq)], rq); | |
440 | } | |
441 | } | |
442 | ||
1fa8cc52 | 443 | static void __blk_mq_complete_request(struct request *rq) |
ed851860 JA |
444 | { |
445 | struct request_queue *q = rq->q; | |
446 | ||
cf43e6be JA |
447 | blk_mq_stat_add(rq); |
448 | ||
ed851860 | 449 | if (!q->softirq_done_fn) |
c8a446ad | 450 | blk_mq_end_request(rq, rq->errors); |
ed851860 JA |
451 | else |
452 | blk_mq_ipi_complete_request(rq); | |
453 | } | |
454 | ||
30a91cb4 CH |
455 | /** |
456 | * blk_mq_complete_request - end I/O on a request | |
457 | * @rq: the request being processed | |
458 | * | |
459 | * Description: | |
460 | * Ends all I/O on a request. It does not handle partial completions. | |
461 | * The actual completion happens out-of-order, through a IPI handler. | |
462 | **/ | |
f4829a9b | 463 | void blk_mq_complete_request(struct request *rq, int error) |
30a91cb4 | 464 | { |
95f09684 JA |
465 | struct request_queue *q = rq->q; |
466 | ||
467 | if (unlikely(blk_should_fake_timeout(q))) | |
30a91cb4 | 468 | return; |
f4829a9b CH |
469 | if (!blk_mark_rq_complete(rq)) { |
470 | rq->errors = error; | |
ed851860 | 471 | __blk_mq_complete_request(rq); |
f4829a9b | 472 | } |
30a91cb4 CH |
473 | } |
474 | EXPORT_SYMBOL(blk_mq_complete_request); | |
320ae51f | 475 | |
973c0191 KB |
476 | int blk_mq_request_started(struct request *rq) |
477 | { | |
478 | return test_bit(REQ_ATOM_STARTED, &rq->atomic_flags); | |
479 | } | |
480 | EXPORT_SYMBOL_GPL(blk_mq_request_started); | |
481 | ||
e2490073 | 482 | void blk_mq_start_request(struct request *rq) |
320ae51f JA |
483 | { |
484 | struct request_queue *q = rq->q; | |
485 | ||
bd166ef1 JA |
486 | blk_mq_sched_started_request(rq); |
487 | ||
320ae51f JA |
488 | trace_block_rq_issue(q, rq); |
489 | ||
742ee69b | 490 | rq->resid_len = blk_rq_bytes(rq); |
91b63639 CH |
491 | if (unlikely(blk_bidi_rq(rq))) |
492 | rq->next_rq->resid_len = blk_rq_bytes(rq->next_rq); | |
742ee69b | 493 | |
cf43e6be JA |
494 | if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) { |
495 | blk_stat_set_issue_time(&rq->issue_stat); | |
496 | rq->rq_flags |= RQF_STATS; | |
87760e5e | 497 | wbt_issue(q->rq_wb, &rq->issue_stat); |
cf43e6be JA |
498 | } |
499 | ||
2b8393b4 | 500 | blk_add_timer(rq); |
87ee7b11 | 501 | |
538b7534 JA |
502 | /* |
503 | * Ensure that ->deadline is visible before set the started | |
504 | * flag and clear the completed flag. | |
505 | */ | |
506 | smp_mb__before_atomic(); | |
507 | ||
87ee7b11 JA |
508 | /* |
509 | * Mark us as started and clear complete. Complete might have been | |
510 | * set if requeue raced with timeout, which then marked it as | |
511 | * complete. So be sure to clear complete again when we start | |
512 | * the request, otherwise we'll ignore the completion event. | |
513 | */ | |
4b570521 JA |
514 | if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) |
515 | set_bit(REQ_ATOM_STARTED, &rq->atomic_flags); | |
516 | if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags)) | |
517 | clear_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags); | |
49f5baa5 CH |
518 | |
519 | if (q->dma_drain_size && blk_rq_bytes(rq)) { | |
520 | /* | |
521 | * Make sure space for the drain appears. We know we can do | |
522 | * this because max_hw_segments has been adjusted to be one | |
523 | * fewer than the device can handle. | |
524 | */ | |
525 | rq->nr_phys_segments++; | |
526 | } | |
320ae51f | 527 | } |
e2490073 | 528 | EXPORT_SYMBOL(blk_mq_start_request); |
320ae51f | 529 | |
ed0791b2 | 530 | static void __blk_mq_requeue_request(struct request *rq) |
320ae51f JA |
531 | { |
532 | struct request_queue *q = rq->q; | |
533 | ||
534 | trace_block_rq_requeue(q, rq); | |
87760e5e | 535 | wbt_requeue(q->rq_wb, &rq->issue_stat); |
bd166ef1 | 536 | blk_mq_sched_requeue_request(rq); |
49f5baa5 | 537 | |
e2490073 CH |
538 | if (test_and_clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) { |
539 | if (q->dma_drain_size && blk_rq_bytes(rq)) | |
540 | rq->nr_phys_segments--; | |
541 | } | |
320ae51f JA |
542 | } |
543 | ||
2b053aca | 544 | void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list) |
ed0791b2 | 545 | { |
ed0791b2 | 546 | __blk_mq_requeue_request(rq); |
ed0791b2 | 547 | |
ed0791b2 | 548 | BUG_ON(blk_queued_rq(rq)); |
2b053aca | 549 | blk_mq_add_to_requeue_list(rq, true, kick_requeue_list); |
ed0791b2 CH |
550 | } |
551 | EXPORT_SYMBOL(blk_mq_requeue_request); | |
552 | ||
6fca6a61 CH |
553 | static void blk_mq_requeue_work(struct work_struct *work) |
554 | { | |
555 | struct request_queue *q = | |
2849450a | 556 | container_of(work, struct request_queue, requeue_work.work); |
6fca6a61 CH |
557 | LIST_HEAD(rq_list); |
558 | struct request *rq, *next; | |
559 | unsigned long flags; | |
560 | ||
561 | spin_lock_irqsave(&q->requeue_lock, flags); | |
562 | list_splice_init(&q->requeue_list, &rq_list); | |
563 | spin_unlock_irqrestore(&q->requeue_lock, flags); | |
564 | ||
565 | list_for_each_entry_safe(rq, next, &rq_list, queuelist) { | |
e8064021 | 566 | if (!(rq->rq_flags & RQF_SOFTBARRIER)) |
6fca6a61 CH |
567 | continue; |
568 | ||
e8064021 | 569 | rq->rq_flags &= ~RQF_SOFTBARRIER; |
6fca6a61 | 570 | list_del_init(&rq->queuelist); |
bd166ef1 | 571 | blk_mq_sched_insert_request(rq, true, false, false); |
6fca6a61 CH |
572 | } |
573 | ||
574 | while (!list_empty(&rq_list)) { | |
575 | rq = list_entry(rq_list.next, struct request, queuelist); | |
576 | list_del_init(&rq->queuelist); | |
bd166ef1 | 577 | blk_mq_sched_insert_request(rq, false, false, false); |
6fca6a61 CH |
578 | } |
579 | ||
52d7f1b5 | 580 | blk_mq_run_hw_queues(q, false); |
6fca6a61 CH |
581 | } |
582 | ||
2b053aca BVA |
583 | void blk_mq_add_to_requeue_list(struct request *rq, bool at_head, |
584 | bool kick_requeue_list) | |
6fca6a61 CH |
585 | { |
586 | struct request_queue *q = rq->q; | |
587 | unsigned long flags; | |
588 | ||
589 | /* | |
590 | * We abuse this flag that is otherwise used by the I/O scheduler to | |
591 | * request head insertation from the workqueue. | |
592 | */ | |
e8064021 | 593 | BUG_ON(rq->rq_flags & RQF_SOFTBARRIER); |
6fca6a61 CH |
594 | |
595 | spin_lock_irqsave(&q->requeue_lock, flags); | |
596 | if (at_head) { | |
e8064021 | 597 | rq->rq_flags |= RQF_SOFTBARRIER; |
6fca6a61 CH |
598 | list_add(&rq->queuelist, &q->requeue_list); |
599 | } else { | |
600 | list_add_tail(&rq->queuelist, &q->requeue_list); | |
601 | } | |
602 | spin_unlock_irqrestore(&q->requeue_lock, flags); | |
2b053aca BVA |
603 | |
604 | if (kick_requeue_list) | |
605 | blk_mq_kick_requeue_list(q); | |
6fca6a61 CH |
606 | } |
607 | EXPORT_SYMBOL(blk_mq_add_to_requeue_list); | |
608 | ||
609 | void blk_mq_kick_requeue_list(struct request_queue *q) | |
610 | { | |
2849450a | 611 | kblockd_schedule_delayed_work(&q->requeue_work, 0); |
6fca6a61 CH |
612 | } |
613 | EXPORT_SYMBOL(blk_mq_kick_requeue_list); | |
614 | ||
2849450a MS |
615 | void blk_mq_delay_kick_requeue_list(struct request_queue *q, |
616 | unsigned long msecs) | |
617 | { | |
618 | kblockd_schedule_delayed_work(&q->requeue_work, | |
619 | msecs_to_jiffies(msecs)); | |
620 | } | |
621 | EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list); | |
622 | ||
1885b24d JA |
623 | void blk_mq_abort_requeue_list(struct request_queue *q) |
624 | { | |
625 | unsigned long flags; | |
626 | LIST_HEAD(rq_list); | |
627 | ||
628 | spin_lock_irqsave(&q->requeue_lock, flags); | |
629 | list_splice_init(&q->requeue_list, &rq_list); | |
630 | spin_unlock_irqrestore(&q->requeue_lock, flags); | |
631 | ||
632 | while (!list_empty(&rq_list)) { | |
633 | struct request *rq; | |
634 | ||
635 | rq = list_first_entry(&rq_list, struct request, queuelist); | |
636 | list_del_init(&rq->queuelist); | |
637 | rq->errors = -EIO; | |
638 | blk_mq_end_request(rq, rq->errors); | |
639 | } | |
640 | } | |
641 | EXPORT_SYMBOL(blk_mq_abort_requeue_list); | |
642 | ||
0e62f51f JA |
643 | struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag) |
644 | { | |
88c7b2b7 JA |
645 | if (tag < tags->nr_tags) { |
646 | prefetch(tags->rqs[tag]); | |
4ee86bab | 647 | return tags->rqs[tag]; |
88c7b2b7 | 648 | } |
4ee86bab HR |
649 | |
650 | return NULL; | |
24d2f903 CH |
651 | } |
652 | EXPORT_SYMBOL(blk_mq_tag_to_rq); | |
653 | ||
320ae51f | 654 | struct blk_mq_timeout_data { |
46f92d42 CH |
655 | unsigned long next; |
656 | unsigned int next_set; | |
320ae51f JA |
657 | }; |
658 | ||
90415837 | 659 | void blk_mq_rq_timed_out(struct request *req, bool reserved) |
320ae51f | 660 | { |
f8a5b122 | 661 | const struct blk_mq_ops *ops = req->q->mq_ops; |
46f92d42 | 662 | enum blk_eh_timer_return ret = BLK_EH_RESET_TIMER; |
87ee7b11 JA |
663 | |
664 | /* | |
665 | * We know that complete is set at this point. If STARTED isn't set | |
666 | * anymore, then the request isn't active and the "timeout" should | |
667 | * just be ignored. This can happen due to the bitflag ordering. | |
668 | * Timeout first checks if STARTED is set, and if it is, assumes | |
669 | * the request is active. But if we race with completion, then | |
670 | * we both flags will get cleared. So check here again, and ignore | |
671 | * a timeout event with a request that isn't active. | |
672 | */ | |
46f92d42 CH |
673 | if (!test_bit(REQ_ATOM_STARTED, &req->atomic_flags)) |
674 | return; | |
87ee7b11 | 675 | |
46f92d42 | 676 | if (ops->timeout) |
0152fb6b | 677 | ret = ops->timeout(req, reserved); |
46f92d42 CH |
678 | |
679 | switch (ret) { | |
680 | case BLK_EH_HANDLED: | |
681 | __blk_mq_complete_request(req); | |
682 | break; | |
683 | case BLK_EH_RESET_TIMER: | |
684 | blk_add_timer(req); | |
685 | blk_clear_rq_complete(req); | |
686 | break; | |
687 | case BLK_EH_NOT_HANDLED: | |
688 | break; | |
689 | default: | |
690 | printk(KERN_ERR "block: bad eh return: %d\n", ret); | |
691 | break; | |
692 | } | |
87ee7b11 | 693 | } |
5b3f25fc | 694 | |
81481eb4 CH |
695 | static void blk_mq_check_expired(struct blk_mq_hw_ctx *hctx, |
696 | struct request *rq, void *priv, bool reserved) | |
697 | { | |
698 | struct blk_mq_timeout_data *data = priv; | |
87ee7b11 | 699 | |
eb130dbf KB |
700 | if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) { |
701 | /* | |
702 | * If a request wasn't started before the queue was | |
703 | * marked dying, kill it here or it'll go unnoticed. | |
704 | */ | |
a59e0f57 KB |
705 | if (unlikely(blk_queue_dying(rq->q))) { |
706 | rq->errors = -EIO; | |
707 | blk_mq_end_request(rq, rq->errors); | |
708 | } | |
46f92d42 | 709 | return; |
eb130dbf | 710 | } |
87ee7b11 | 711 | |
46f92d42 CH |
712 | if (time_after_eq(jiffies, rq->deadline)) { |
713 | if (!blk_mark_rq_complete(rq)) | |
0152fb6b | 714 | blk_mq_rq_timed_out(rq, reserved); |
46f92d42 CH |
715 | } else if (!data->next_set || time_after(data->next, rq->deadline)) { |
716 | data->next = rq->deadline; | |
717 | data->next_set = 1; | |
718 | } | |
87ee7b11 JA |
719 | } |
720 | ||
287922eb | 721 | static void blk_mq_timeout_work(struct work_struct *work) |
320ae51f | 722 | { |
287922eb CH |
723 | struct request_queue *q = |
724 | container_of(work, struct request_queue, timeout_work); | |
81481eb4 CH |
725 | struct blk_mq_timeout_data data = { |
726 | .next = 0, | |
727 | .next_set = 0, | |
728 | }; | |
81481eb4 | 729 | int i; |
320ae51f | 730 | |
71f79fb3 GKB |
731 | /* A deadlock might occur if a request is stuck requiring a |
732 | * timeout at the same time a queue freeze is waiting | |
733 | * completion, since the timeout code would not be able to | |
734 | * acquire the queue reference here. | |
735 | * | |
736 | * That's why we don't use blk_queue_enter here; instead, we use | |
737 | * percpu_ref_tryget directly, because we need to be able to | |
738 | * obtain a reference even in the short window between the queue | |
739 | * starting to freeze, by dropping the first reference in | |
740 | * blk_mq_freeze_queue_start, and the moment the last request is | |
741 | * consumed, marked by the instant q_usage_counter reaches | |
742 | * zero. | |
743 | */ | |
744 | if (!percpu_ref_tryget(&q->q_usage_counter)) | |
287922eb CH |
745 | return; |
746 | ||
0bf6cd5b | 747 | blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &data); |
320ae51f | 748 | |
81481eb4 CH |
749 | if (data.next_set) { |
750 | data.next = blk_rq_timeout(round_jiffies_up(data.next)); | |
751 | mod_timer(&q->timeout, data.next); | |
0d2602ca | 752 | } else { |
0bf6cd5b CH |
753 | struct blk_mq_hw_ctx *hctx; |
754 | ||
f054b56c ML |
755 | queue_for_each_hw_ctx(q, hctx, i) { |
756 | /* the hctx may be unmapped, so check it here */ | |
757 | if (blk_mq_hw_queue_mapped(hctx)) | |
758 | blk_mq_tag_idle(hctx); | |
759 | } | |
0d2602ca | 760 | } |
287922eb | 761 | blk_queue_exit(q); |
320ae51f JA |
762 | } |
763 | ||
764 | /* | |
765 | * Reverse check our software queue for entries that we could potentially | |
766 | * merge with. Currently includes a hand-wavy stop count of 8, to not spend | |
767 | * too much time checking for merges. | |
768 | */ | |
769 | static bool blk_mq_attempt_merge(struct request_queue *q, | |
770 | struct blk_mq_ctx *ctx, struct bio *bio) | |
771 | { | |
772 | struct request *rq; | |
773 | int checked = 8; | |
774 | ||
775 | list_for_each_entry_reverse(rq, &ctx->rq_list, queuelist) { | |
776 | int el_ret; | |
777 | ||
778 | if (!checked--) | |
779 | break; | |
780 | ||
781 | if (!blk_rq_merge_ok(rq, bio)) | |
782 | continue; | |
783 | ||
784 | el_ret = blk_try_merge(rq, bio); | |
bd166ef1 JA |
785 | if (el_ret == ELEVATOR_NO_MERGE) |
786 | continue; | |
787 | ||
788 | if (!blk_mq_sched_allow_merge(q, rq, bio)) | |
789 | break; | |
790 | ||
320ae51f JA |
791 | if (el_ret == ELEVATOR_BACK_MERGE) { |
792 | if (bio_attempt_back_merge(q, rq, bio)) { | |
793 | ctx->rq_merged++; | |
794 | return true; | |
795 | } | |
796 | break; | |
797 | } else if (el_ret == ELEVATOR_FRONT_MERGE) { | |
798 | if (bio_attempt_front_merge(q, rq, bio)) { | |
799 | ctx->rq_merged++; | |
800 | return true; | |
801 | } | |
802 | break; | |
803 | } | |
804 | } | |
805 | ||
806 | return false; | |
807 | } | |
808 | ||
88459642 OS |
809 | struct flush_busy_ctx_data { |
810 | struct blk_mq_hw_ctx *hctx; | |
811 | struct list_head *list; | |
812 | }; | |
813 | ||
814 | static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data) | |
815 | { | |
816 | struct flush_busy_ctx_data *flush_data = data; | |
817 | struct blk_mq_hw_ctx *hctx = flush_data->hctx; | |
818 | struct blk_mq_ctx *ctx = hctx->ctxs[bitnr]; | |
819 | ||
820 | sbitmap_clear_bit(sb, bitnr); | |
821 | spin_lock(&ctx->lock); | |
822 | list_splice_tail_init(&ctx->rq_list, flush_data->list); | |
823 | spin_unlock(&ctx->lock); | |
824 | return true; | |
825 | } | |
826 | ||
1429d7c9 JA |
827 | /* |
828 | * Process software queues that have been marked busy, splicing them | |
829 | * to the for-dispatch | |
830 | */ | |
2c3ad667 | 831 | void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list) |
1429d7c9 | 832 | { |
88459642 OS |
833 | struct flush_busy_ctx_data data = { |
834 | .hctx = hctx, | |
835 | .list = list, | |
836 | }; | |
1429d7c9 | 837 | |
88459642 | 838 | sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data); |
1429d7c9 | 839 | } |
2c3ad667 | 840 | EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs); |
1429d7c9 | 841 | |
703fd1c0 JA |
842 | static inline unsigned int queued_to_index(unsigned int queued) |
843 | { | |
844 | if (!queued) | |
845 | return 0; | |
1429d7c9 | 846 | |
703fd1c0 | 847 | return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1); |
1429d7c9 JA |
848 | } |
849 | ||
bd166ef1 JA |
850 | static bool blk_mq_get_driver_tag(struct request *rq, |
851 | struct blk_mq_hw_ctx **hctx, bool wait) | |
852 | { | |
853 | struct blk_mq_alloc_data data = { | |
854 | .q = rq->q, | |
855 | .ctx = rq->mq_ctx, | |
856 | .hctx = blk_mq_map_queue(rq->q, rq->mq_ctx->cpu), | |
857 | .flags = wait ? 0 : BLK_MQ_REQ_NOWAIT, | |
858 | }; | |
859 | ||
860 | if (blk_mq_hctx_stopped(data.hctx)) | |
861 | return false; | |
862 | ||
863 | if (rq->tag != -1) { | |
864 | done: | |
865 | if (hctx) | |
866 | *hctx = data.hctx; | |
867 | return true; | |
868 | } | |
869 | ||
870 | rq->tag = blk_mq_get_tag(&data); | |
871 | if (rq->tag >= 0) { | |
872 | data.hctx->tags->rqs[rq->tag] = rq; | |
873 | goto done; | |
874 | } | |
875 | ||
876 | return false; | |
877 | } | |
878 | ||
879 | /* | |
880 | * If we fail getting a driver tag because all the driver tags are already | |
881 | * assigned and on the dispatch list, BUT the first entry does not have a | |
882 | * tag, then we could deadlock. For that case, move entries with assigned | |
883 | * driver tags to the front, leaving the set of tagged requests in the | |
884 | * same order, and the untagged set in the same order. | |
885 | */ | |
886 | static bool reorder_tags_to_front(struct list_head *list) | |
887 | { | |
888 | struct request *rq, *tmp, *first = NULL; | |
889 | ||
890 | list_for_each_entry_safe_reverse(rq, tmp, list, queuelist) { | |
891 | if (rq == first) | |
892 | break; | |
893 | if (rq->tag != -1) { | |
894 | list_move(&rq->queuelist, list); | |
895 | if (!first) | |
896 | first = rq; | |
897 | } | |
898 | } | |
899 | ||
900 | return first != NULL; | |
901 | } | |
902 | ||
f04c3df3 | 903 | bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx *hctx, struct list_head *list) |
320ae51f JA |
904 | { |
905 | struct request_queue *q = hctx->queue; | |
320ae51f | 906 | struct request *rq; |
74c45052 JA |
907 | LIST_HEAD(driver_list); |
908 | struct list_head *dptr; | |
f04c3df3 | 909 | int queued, ret = BLK_MQ_RQ_QUEUE_OK; |
320ae51f | 910 | |
74c45052 JA |
911 | /* |
912 | * Start off with dptr being NULL, so we start the first request | |
913 | * immediately, even if we have more pending. | |
914 | */ | |
915 | dptr = NULL; | |
916 | ||
320ae51f JA |
917 | /* |
918 | * Now process all the entries, sending them to the driver. | |
919 | */ | |
1429d7c9 | 920 | queued = 0; |
f04c3df3 | 921 | while (!list_empty(list)) { |
74c45052 | 922 | struct blk_mq_queue_data bd; |
320ae51f | 923 | |
f04c3df3 | 924 | rq = list_first_entry(list, struct request, queuelist); |
bd166ef1 JA |
925 | if (!blk_mq_get_driver_tag(rq, &hctx, false)) { |
926 | if (!queued && reorder_tags_to_front(list)) | |
927 | continue; | |
928 | blk_mq_sched_mark_restart(hctx); | |
929 | break; | |
930 | } | |
320ae51f | 931 | list_del_init(&rq->queuelist); |
320ae51f | 932 | |
74c45052 JA |
933 | bd.rq = rq; |
934 | bd.list = dptr; | |
f04c3df3 | 935 | bd.last = list_empty(list); |
74c45052 JA |
936 | |
937 | ret = q->mq_ops->queue_rq(hctx, &bd); | |
320ae51f JA |
938 | switch (ret) { |
939 | case BLK_MQ_RQ_QUEUE_OK: | |
940 | queued++; | |
52b9c330 | 941 | break; |
320ae51f | 942 | case BLK_MQ_RQ_QUEUE_BUSY: |
f04c3df3 | 943 | list_add(&rq->queuelist, list); |
ed0791b2 | 944 | __blk_mq_requeue_request(rq); |
320ae51f JA |
945 | break; |
946 | default: | |
947 | pr_err("blk-mq: bad return on queue: %d\n", ret); | |
320ae51f | 948 | case BLK_MQ_RQ_QUEUE_ERROR: |
1e93b8c2 | 949 | rq->errors = -EIO; |
c8a446ad | 950 | blk_mq_end_request(rq, rq->errors); |
320ae51f JA |
951 | break; |
952 | } | |
953 | ||
954 | if (ret == BLK_MQ_RQ_QUEUE_BUSY) | |
955 | break; | |
74c45052 JA |
956 | |
957 | /* | |
958 | * We've done the first request. If we have more than 1 | |
959 | * left in the list, set dptr to defer issue. | |
960 | */ | |
f04c3df3 | 961 | if (!dptr && list->next != list->prev) |
74c45052 | 962 | dptr = &driver_list; |
320ae51f JA |
963 | } |
964 | ||
703fd1c0 | 965 | hctx->dispatched[queued_to_index(queued)]++; |
320ae51f JA |
966 | |
967 | /* | |
968 | * Any items that need requeuing? Stuff them into hctx->dispatch, | |
969 | * that is where we will continue on next queue run. | |
970 | */ | |
f04c3df3 | 971 | if (!list_empty(list)) { |
320ae51f | 972 | spin_lock(&hctx->lock); |
f04c3df3 | 973 | list_splice(list, &hctx->dispatch); |
320ae51f | 974 | spin_unlock(&hctx->lock); |
f04c3df3 | 975 | |
9ba52e58 SL |
976 | /* |
977 | * the queue is expected stopped with BLK_MQ_RQ_QUEUE_BUSY, but | |
978 | * it's possible the queue is stopped and restarted again | |
979 | * before this. Queue restart will dispatch requests. And since | |
980 | * requests in rq_list aren't added into hctx->dispatch yet, | |
981 | * the requests in rq_list might get lost. | |
982 | * | |
983 | * blk_mq_run_hw_queue() already checks the STOPPED bit | |
bd166ef1 JA |
984 | * |
985 | * If RESTART is set, then let completion restart the queue | |
986 | * instead of potentially looping here. | |
987 | */ | |
988 | if (!blk_mq_sched_needs_restart(hctx)) | |
989 | blk_mq_run_hw_queue(hctx, true); | |
320ae51f | 990 | } |
f04c3df3 JA |
991 | |
992 | return ret != BLK_MQ_RQ_QUEUE_BUSY; | |
993 | } | |
994 | ||
6a83e74d BVA |
995 | static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx) |
996 | { | |
997 | int srcu_idx; | |
998 | ||
999 | WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) && | |
1000 | cpu_online(hctx->next_cpu)); | |
1001 | ||
1002 | if (!(hctx->flags & BLK_MQ_F_BLOCKING)) { | |
1003 | rcu_read_lock(); | |
bd166ef1 | 1004 | blk_mq_sched_dispatch_requests(hctx); |
6a83e74d BVA |
1005 | rcu_read_unlock(); |
1006 | } else { | |
1007 | srcu_idx = srcu_read_lock(&hctx->queue_rq_srcu); | |
bd166ef1 | 1008 | blk_mq_sched_dispatch_requests(hctx); |
6a83e74d BVA |
1009 | srcu_read_unlock(&hctx->queue_rq_srcu, srcu_idx); |
1010 | } | |
1011 | } | |
1012 | ||
506e931f JA |
1013 | /* |
1014 | * It'd be great if the workqueue API had a way to pass | |
1015 | * in a mask and had some smarts for more clever placement. | |
1016 | * For now we just round-robin here, switching for every | |
1017 | * BLK_MQ_CPU_WORK_BATCH queued items. | |
1018 | */ | |
1019 | static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx) | |
1020 | { | |
b657d7e6 CH |
1021 | if (hctx->queue->nr_hw_queues == 1) |
1022 | return WORK_CPU_UNBOUND; | |
506e931f JA |
1023 | |
1024 | if (--hctx->next_cpu_batch <= 0) { | |
c02ebfdd | 1025 | int next_cpu; |
506e931f JA |
1026 | |
1027 | next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask); | |
1028 | if (next_cpu >= nr_cpu_ids) | |
1029 | next_cpu = cpumask_first(hctx->cpumask); | |
1030 | ||
1031 | hctx->next_cpu = next_cpu; | |
1032 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; | |
1033 | } | |
1034 | ||
b657d7e6 | 1035 | return hctx->next_cpu; |
506e931f JA |
1036 | } |
1037 | ||
320ae51f JA |
1038 | void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) |
1039 | { | |
5d1b25c1 BVA |
1040 | if (unlikely(blk_mq_hctx_stopped(hctx) || |
1041 | !blk_mq_hw_queue_mapped(hctx))) | |
320ae51f JA |
1042 | return; |
1043 | ||
1b792f2f | 1044 | if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) { |
2a90d4aa PB |
1045 | int cpu = get_cpu(); |
1046 | if (cpumask_test_cpu(cpu, hctx->cpumask)) { | |
398205b8 | 1047 | __blk_mq_run_hw_queue(hctx); |
2a90d4aa | 1048 | put_cpu(); |
398205b8 PB |
1049 | return; |
1050 | } | |
e4043dcf | 1051 | |
2a90d4aa | 1052 | put_cpu(); |
e4043dcf | 1053 | } |
398205b8 | 1054 | |
27489a3c | 1055 | kblockd_schedule_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work); |
320ae51f JA |
1056 | } |
1057 | ||
b94ec296 | 1058 | void blk_mq_run_hw_queues(struct request_queue *q, bool async) |
320ae51f JA |
1059 | { |
1060 | struct blk_mq_hw_ctx *hctx; | |
1061 | int i; | |
1062 | ||
1063 | queue_for_each_hw_ctx(q, hctx, i) { | |
bd166ef1 | 1064 | if (!blk_mq_hctx_has_pending(hctx) || |
5d1b25c1 | 1065 | blk_mq_hctx_stopped(hctx)) |
320ae51f JA |
1066 | continue; |
1067 | ||
b94ec296 | 1068 | blk_mq_run_hw_queue(hctx, async); |
320ae51f JA |
1069 | } |
1070 | } | |
b94ec296 | 1071 | EXPORT_SYMBOL(blk_mq_run_hw_queues); |
320ae51f | 1072 | |
fd001443 BVA |
1073 | /** |
1074 | * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped | |
1075 | * @q: request queue. | |
1076 | * | |
1077 | * The caller is responsible for serializing this function against | |
1078 | * blk_mq_{start,stop}_hw_queue(). | |
1079 | */ | |
1080 | bool blk_mq_queue_stopped(struct request_queue *q) | |
1081 | { | |
1082 | struct blk_mq_hw_ctx *hctx; | |
1083 | int i; | |
1084 | ||
1085 | queue_for_each_hw_ctx(q, hctx, i) | |
1086 | if (blk_mq_hctx_stopped(hctx)) | |
1087 | return true; | |
1088 | ||
1089 | return false; | |
1090 | } | |
1091 | EXPORT_SYMBOL(blk_mq_queue_stopped); | |
1092 | ||
320ae51f JA |
1093 | void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx) |
1094 | { | |
27489a3c | 1095 | cancel_work(&hctx->run_work); |
70f4db63 | 1096 | cancel_delayed_work(&hctx->delay_work); |
320ae51f JA |
1097 | set_bit(BLK_MQ_S_STOPPED, &hctx->state); |
1098 | } | |
1099 | EXPORT_SYMBOL(blk_mq_stop_hw_queue); | |
1100 | ||
280d45f6 CH |
1101 | void blk_mq_stop_hw_queues(struct request_queue *q) |
1102 | { | |
1103 | struct blk_mq_hw_ctx *hctx; | |
1104 | int i; | |
1105 | ||
1106 | queue_for_each_hw_ctx(q, hctx, i) | |
1107 | blk_mq_stop_hw_queue(hctx); | |
1108 | } | |
1109 | EXPORT_SYMBOL(blk_mq_stop_hw_queues); | |
1110 | ||
320ae51f JA |
1111 | void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx) |
1112 | { | |
1113 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
e4043dcf | 1114 | |
0ffbce80 | 1115 | blk_mq_run_hw_queue(hctx, false); |
320ae51f JA |
1116 | } |
1117 | EXPORT_SYMBOL(blk_mq_start_hw_queue); | |
1118 | ||
2f268556 CH |
1119 | void blk_mq_start_hw_queues(struct request_queue *q) |
1120 | { | |
1121 | struct blk_mq_hw_ctx *hctx; | |
1122 | int i; | |
1123 | ||
1124 | queue_for_each_hw_ctx(q, hctx, i) | |
1125 | blk_mq_start_hw_queue(hctx); | |
1126 | } | |
1127 | EXPORT_SYMBOL(blk_mq_start_hw_queues); | |
1128 | ||
ae911c5e JA |
1129 | void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) |
1130 | { | |
1131 | if (!blk_mq_hctx_stopped(hctx)) | |
1132 | return; | |
1133 | ||
1134 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
1135 | blk_mq_run_hw_queue(hctx, async); | |
1136 | } | |
1137 | EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue); | |
1138 | ||
1b4a3258 | 1139 | void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async) |
320ae51f JA |
1140 | { |
1141 | struct blk_mq_hw_ctx *hctx; | |
1142 | int i; | |
1143 | ||
ae911c5e JA |
1144 | queue_for_each_hw_ctx(q, hctx, i) |
1145 | blk_mq_start_stopped_hw_queue(hctx, async); | |
320ae51f JA |
1146 | } |
1147 | EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues); | |
1148 | ||
70f4db63 | 1149 | static void blk_mq_run_work_fn(struct work_struct *work) |
320ae51f JA |
1150 | { |
1151 | struct blk_mq_hw_ctx *hctx; | |
1152 | ||
27489a3c | 1153 | hctx = container_of(work, struct blk_mq_hw_ctx, run_work); |
e4043dcf | 1154 | |
320ae51f JA |
1155 | __blk_mq_run_hw_queue(hctx); |
1156 | } | |
1157 | ||
70f4db63 CH |
1158 | static void blk_mq_delay_work_fn(struct work_struct *work) |
1159 | { | |
1160 | struct blk_mq_hw_ctx *hctx; | |
1161 | ||
1162 | hctx = container_of(work, struct blk_mq_hw_ctx, delay_work.work); | |
1163 | ||
1164 | if (test_and_clear_bit(BLK_MQ_S_STOPPED, &hctx->state)) | |
1165 | __blk_mq_run_hw_queue(hctx); | |
1166 | } | |
1167 | ||
1168 | void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs) | |
1169 | { | |
19c66e59 ML |
1170 | if (unlikely(!blk_mq_hw_queue_mapped(hctx))) |
1171 | return; | |
70f4db63 | 1172 | |
b657d7e6 CH |
1173 | kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx), |
1174 | &hctx->delay_work, msecs_to_jiffies(msecs)); | |
70f4db63 CH |
1175 | } |
1176 | EXPORT_SYMBOL(blk_mq_delay_queue); | |
1177 | ||
cfd0c552 | 1178 | static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx, |
cfd0c552 ML |
1179 | struct request *rq, |
1180 | bool at_head) | |
320ae51f | 1181 | { |
e57690fe JA |
1182 | struct blk_mq_ctx *ctx = rq->mq_ctx; |
1183 | ||
01b983c9 JA |
1184 | trace_block_rq_insert(hctx->queue, rq); |
1185 | ||
72a0a36e CH |
1186 | if (at_head) |
1187 | list_add(&rq->queuelist, &ctx->rq_list); | |
1188 | else | |
1189 | list_add_tail(&rq->queuelist, &ctx->rq_list); | |
cfd0c552 | 1190 | } |
4bb659b1 | 1191 | |
2c3ad667 JA |
1192 | void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq, |
1193 | bool at_head) | |
cfd0c552 ML |
1194 | { |
1195 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
1196 | ||
e57690fe | 1197 | __blk_mq_insert_req_list(hctx, rq, at_head); |
320ae51f | 1198 | blk_mq_hctx_mark_pending(hctx, ctx); |
320ae51f JA |
1199 | } |
1200 | ||
bd166ef1 JA |
1201 | void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx, |
1202 | struct list_head *list) | |
320ae51f JA |
1203 | |
1204 | { | |
320ae51f JA |
1205 | /* |
1206 | * preemption doesn't flush plug list, so it's possible ctx->cpu is | |
1207 | * offline now | |
1208 | */ | |
1209 | spin_lock(&ctx->lock); | |
1210 | while (!list_empty(list)) { | |
1211 | struct request *rq; | |
1212 | ||
1213 | rq = list_first_entry(list, struct request, queuelist); | |
e57690fe | 1214 | BUG_ON(rq->mq_ctx != ctx); |
320ae51f | 1215 | list_del_init(&rq->queuelist); |
e57690fe | 1216 | __blk_mq_insert_req_list(hctx, rq, false); |
320ae51f | 1217 | } |
cfd0c552 | 1218 | blk_mq_hctx_mark_pending(hctx, ctx); |
320ae51f | 1219 | spin_unlock(&ctx->lock); |
320ae51f JA |
1220 | } |
1221 | ||
1222 | static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b) | |
1223 | { | |
1224 | struct request *rqa = container_of(a, struct request, queuelist); | |
1225 | struct request *rqb = container_of(b, struct request, queuelist); | |
1226 | ||
1227 | return !(rqa->mq_ctx < rqb->mq_ctx || | |
1228 | (rqa->mq_ctx == rqb->mq_ctx && | |
1229 | blk_rq_pos(rqa) < blk_rq_pos(rqb))); | |
1230 | } | |
1231 | ||
1232 | void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule) | |
1233 | { | |
1234 | struct blk_mq_ctx *this_ctx; | |
1235 | struct request_queue *this_q; | |
1236 | struct request *rq; | |
1237 | LIST_HEAD(list); | |
1238 | LIST_HEAD(ctx_list); | |
1239 | unsigned int depth; | |
1240 | ||
1241 | list_splice_init(&plug->mq_list, &list); | |
1242 | ||
1243 | list_sort(NULL, &list, plug_ctx_cmp); | |
1244 | ||
1245 | this_q = NULL; | |
1246 | this_ctx = NULL; | |
1247 | depth = 0; | |
1248 | ||
1249 | while (!list_empty(&list)) { | |
1250 | rq = list_entry_rq(list.next); | |
1251 | list_del_init(&rq->queuelist); | |
1252 | BUG_ON(!rq->q); | |
1253 | if (rq->mq_ctx != this_ctx) { | |
1254 | if (this_ctx) { | |
bd166ef1 JA |
1255 | trace_block_unplug(this_q, depth, from_schedule); |
1256 | blk_mq_sched_insert_requests(this_q, this_ctx, | |
1257 | &ctx_list, | |
1258 | from_schedule); | |
320ae51f JA |
1259 | } |
1260 | ||
1261 | this_ctx = rq->mq_ctx; | |
1262 | this_q = rq->q; | |
1263 | depth = 0; | |
1264 | } | |
1265 | ||
1266 | depth++; | |
1267 | list_add_tail(&rq->queuelist, &ctx_list); | |
1268 | } | |
1269 | ||
1270 | /* | |
1271 | * If 'this_ctx' is set, we know we have entries to complete | |
1272 | * on 'ctx_list'. Do those. | |
1273 | */ | |
1274 | if (this_ctx) { | |
bd166ef1 JA |
1275 | trace_block_unplug(this_q, depth, from_schedule); |
1276 | blk_mq_sched_insert_requests(this_q, this_ctx, &ctx_list, | |
1277 | from_schedule); | |
320ae51f JA |
1278 | } |
1279 | } | |
1280 | ||
1281 | static void blk_mq_bio_to_request(struct request *rq, struct bio *bio) | |
1282 | { | |
1283 | init_request_from_bio(rq, bio); | |
4b570521 | 1284 | |
6e85eaf3 | 1285 | blk_account_io_start(rq, true); |
320ae51f JA |
1286 | } |
1287 | ||
274a5843 JA |
1288 | static inline bool hctx_allow_merges(struct blk_mq_hw_ctx *hctx) |
1289 | { | |
1290 | return (hctx->flags & BLK_MQ_F_SHOULD_MERGE) && | |
1291 | !blk_queue_nomerges(hctx->queue); | |
1292 | } | |
1293 | ||
07068d5b JA |
1294 | static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx *hctx, |
1295 | struct blk_mq_ctx *ctx, | |
1296 | struct request *rq, struct bio *bio) | |
320ae51f | 1297 | { |
e18378a6 | 1298 | if (!hctx_allow_merges(hctx) || !bio_mergeable(bio)) { |
07068d5b JA |
1299 | blk_mq_bio_to_request(rq, bio); |
1300 | spin_lock(&ctx->lock); | |
1301 | insert_rq: | |
1302 | __blk_mq_insert_request(hctx, rq, false); | |
1303 | spin_unlock(&ctx->lock); | |
1304 | return false; | |
1305 | } else { | |
274a5843 JA |
1306 | struct request_queue *q = hctx->queue; |
1307 | ||
07068d5b JA |
1308 | spin_lock(&ctx->lock); |
1309 | if (!blk_mq_attempt_merge(q, ctx, bio)) { | |
1310 | blk_mq_bio_to_request(rq, bio); | |
1311 | goto insert_rq; | |
1312 | } | |
320ae51f | 1313 | |
07068d5b | 1314 | spin_unlock(&ctx->lock); |
bd166ef1 | 1315 | __blk_mq_finish_request(hctx, ctx, rq); |
07068d5b | 1316 | return true; |
14ec77f3 | 1317 | } |
07068d5b | 1318 | } |
14ec77f3 | 1319 | |
fd2d3326 JA |
1320 | static blk_qc_t request_to_qc_t(struct blk_mq_hw_ctx *hctx, struct request *rq) |
1321 | { | |
bd166ef1 JA |
1322 | if (rq->tag != -1) |
1323 | return blk_tag_to_qc_t(rq->tag, hctx->queue_num, false); | |
1324 | ||
1325 | return blk_tag_to_qc_t(rq->internal_tag, hctx->queue_num, true); | |
fd2d3326 JA |
1326 | } |
1327 | ||
066a4a73 | 1328 | static void blk_mq_try_issue_directly(struct request *rq, blk_qc_t *cookie) |
f984df1f | 1329 | { |
f984df1f | 1330 | struct request_queue *q = rq->q; |
f984df1f SL |
1331 | struct blk_mq_queue_data bd = { |
1332 | .rq = rq, | |
1333 | .list = NULL, | |
1334 | .last = 1 | |
1335 | }; | |
bd166ef1 JA |
1336 | struct blk_mq_hw_ctx *hctx; |
1337 | blk_qc_t new_cookie; | |
1338 | int ret; | |
f984df1f | 1339 | |
bd166ef1 | 1340 | if (q->elevator) |
2253efc8 BVA |
1341 | goto insert; |
1342 | ||
bd166ef1 JA |
1343 | if (!blk_mq_get_driver_tag(rq, &hctx, false)) |
1344 | goto insert; | |
1345 | ||
1346 | new_cookie = request_to_qc_t(hctx, rq); | |
1347 | ||
f984df1f SL |
1348 | /* |
1349 | * For OK queue, we are done. For error, kill it. Any other | |
1350 | * error (busy), just add it to our list as we previously | |
1351 | * would have done | |
1352 | */ | |
1353 | ret = q->mq_ops->queue_rq(hctx, &bd); | |
7b371636 JA |
1354 | if (ret == BLK_MQ_RQ_QUEUE_OK) { |
1355 | *cookie = new_cookie; | |
2253efc8 | 1356 | return; |
7b371636 | 1357 | } |
f984df1f | 1358 | |
7b371636 JA |
1359 | __blk_mq_requeue_request(rq); |
1360 | ||
1361 | if (ret == BLK_MQ_RQ_QUEUE_ERROR) { | |
1362 | *cookie = BLK_QC_T_NONE; | |
1363 | rq->errors = -EIO; | |
1364 | blk_mq_end_request(rq, rq->errors); | |
2253efc8 | 1365 | return; |
f984df1f | 1366 | } |
7b371636 | 1367 | |
2253efc8 | 1368 | insert: |
bd166ef1 | 1369 | blk_mq_sched_insert_request(rq, false, true, true); |
f984df1f SL |
1370 | } |
1371 | ||
07068d5b JA |
1372 | /* |
1373 | * Multiple hardware queue variant. This will not use per-process plugs, | |
1374 | * but will attempt to bypass the hctx queueing if we can go straight to | |
1375 | * hardware for SYNC IO. | |
1376 | */ | |
dece1635 | 1377 | static blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio) |
07068d5b | 1378 | { |
ef295ecf | 1379 | const int is_sync = op_is_sync(bio->bi_opf); |
1eff9d32 | 1380 | const int is_flush_fua = bio->bi_opf & (REQ_PREFLUSH | REQ_FUA); |
2552e3f8 | 1381 | struct blk_mq_alloc_data data; |
07068d5b | 1382 | struct request *rq; |
6a83e74d | 1383 | unsigned int request_count = 0, srcu_idx; |
f984df1f | 1384 | struct blk_plug *plug; |
5b3f341f | 1385 | struct request *same_queue_rq = NULL; |
7b371636 | 1386 | blk_qc_t cookie; |
87760e5e | 1387 | unsigned int wb_acct; |
07068d5b JA |
1388 | |
1389 | blk_queue_bounce(q, &bio); | |
1390 | ||
1391 | if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) { | |
4246a0b6 | 1392 | bio_io_error(bio); |
dece1635 | 1393 | return BLK_QC_T_NONE; |
07068d5b JA |
1394 | } |
1395 | ||
54efd50b KO |
1396 | blk_queue_split(q, &bio, q->bio_split); |
1397 | ||
87c279e6 OS |
1398 | if (!is_flush_fua && !blk_queue_nomerges(q) && |
1399 | blk_attempt_plug_merge(q, bio, &request_count, &same_queue_rq)) | |
1400 | return BLK_QC_T_NONE; | |
f984df1f | 1401 | |
bd166ef1 JA |
1402 | if (blk_mq_sched_bio_merge(q, bio)) |
1403 | return BLK_QC_T_NONE; | |
1404 | ||
87760e5e JA |
1405 | wb_acct = wbt_wait(q->rq_wb, bio, NULL); |
1406 | ||
bd166ef1 JA |
1407 | trace_block_getrq(q, bio, bio->bi_opf); |
1408 | ||
1409 | rq = blk_mq_sched_get_request(q, bio, bio->bi_opf, &data); | |
87760e5e JA |
1410 | if (unlikely(!rq)) { |
1411 | __wbt_done(q->rq_wb, wb_acct); | |
dece1635 | 1412 | return BLK_QC_T_NONE; |
87760e5e JA |
1413 | } |
1414 | ||
1415 | wbt_track(&rq->issue_stat, wb_acct); | |
07068d5b | 1416 | |
fd2d3326 | 1417 | cookie = request_to_qc_t(data.hctx, rq); |
07068d5b JA |
1418 | |
1419 | if (unlikely(is_flush_fua)) { | |
1420 | blk_mq_bio_to_request(rq, bio); | |
bd166ef1 | 1421 | blk_mq_get_driver_tag(rq, NULL, true); |
07068d5b JA |
1422 | blk_insert_flush(rq); |
1423 | goto run_queue; | |
1424 | } | |
1425 | ||
f984df1f | 1426 | plug = current->plug; |
e167dfb5 JA |
1427 | /* |
1428 | * If the driver supports defer issued based on 'last', then | |
1429 | * queue it up like normal since we can potentially save some | |
1430 | * CPU this way. | |
1431 | */ | |
f984df1f SL |
1432 | if (((plug && !blk_queue_nomerges(q)) || is_sync) && |
1433 | !(data.hctx->flags & BLK_MQ_F_DEFER_ISSUE)) { | |
1434 | struct request *old_rq = NULL; | |
07068d5b JA |
1435 | |
1436 | blk_mq_bio_to_request(rq, bio); | |
07068d5b JA |
1437 | |
1438 | /* | |
6a83e74d | 1439 | * We do limited plugging. If the bio can be merged, do that. |
f984df1f SL |
1440 | * Otherwise the existing request in the plug list will be |
1441 | * issued. So the plug list will have one request at most | |
07068d5b | 1442 | */ |
f984df1f | 1443 | if (plug) { |
5b3f341f SL |
1444 | /* |
1445 | * The plug list might get flushed before this. If that | |
b094f89c JA |
1446 | * happens, same_queue_rq is invalid and plug list is |
1447 | * empty | |
1448 | */ | |
5b3f341f SL |
1449 | if (same_queue_rq && !list_empty(&plug->mq_list)) { |
1450 | old_rq = same_queue_rq; | |
f984df1f | 1451 | list_del_init(&old_rq->queuelist); |
07068d5b | 1452 | } |
f984df1f SL |
1453 | list_add_tail(&rq->queuelist, &plug->mq_list); |
1454 | } else /* is_sync */ | |
1455 | old_rq = rq; | |
1456 | blk_mq_put_ctx(data.ctx); | |
1457 | if (!old_rq) | |
7b371636 | 1458 | goto done; |
6a83e74d BVA |
1459 | |
1460 | if (!(data.hctx->flags & BLK_MQ_F_BLOCKING)) { | |
1461 | rcu_read_lock(); | |
066a4a73 | 1462 | blk_mq_try_issue_directly(old_rq, &cookie); |
6a83e74d BVA |
1463 | rcu_read_unlock(); |
1464 | } else { | |
1465 | srcu_idx = srcu_read_lock(&data.hctx->queue_rq_srcu); | |
066a4a73 | 1466 | blk_mq_try_issue_directly(old_rq, &cookie); |
6a83e74d BVA |
1467 | srcu_read_unlock(&data.hctx->queue_rq_srcu, srcu_idx); |
1468 | } | |
7b371636 | 1469 | goto done; |
07068d5b JA |
1470 | } |
1471 | ||
bd166ef1 JA |
1472 | if (q->elevator) { |
1473 | blk_mq_put_ctx(data.ctx); | |
1474 | blk_mq_bio_to_request(rq, bio); | |
1475 | blk_mq_sched_insert_request(rq, false, true, true); | |
1476 | goto done; | |
1477 | } | |
07068d5b JA |
1478 | if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) { |
1479 | /* | |
1480 | * For a SYNC request, send it to the hardware immediately. For | |
1481 | * an ASYNC request, just ensure that we run it later on. The | |
1482 | * latter allows for merging opportunities and more efficient | |
1483 | * dispatching. | |
1484 | */ | |
1485 | run_queue: | |
1486 | blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua); | |
1487 | } | |
07068d5b | 1488 | blk_mq_put_ctx(data.ctx); |
7b371636 JA |
1489 | done: |
1490 | return cookie; | |
07068d5b JA |
1491 | } |
1492 | ||
1493 | /* | |
1494 | * Single hardware queue variant. This will attempt to use any per-process | |
1495 | * plug for merging and IO deferral. | |
1496 | */ | |
dece1635 | 1497 | static blk_qc_t blk_sq_make_request(struct request_queue *q, struct bio *bio) |
07068d5b | 1498 | { |
ef295ecf | 1499 | const int is_sync = op_is_sync(bio->bi_opf); |
1eff9d32 | 1500 | const int is_flush_fua = bio->bi_opf & (REQ_PREFLUSH | REQ_FUA); |
e6c4438b JM |
1501 | struct blk_plug *plug; |
1502 | unsigned int request_count = 0; | |
2552e3f8 | 1503 | struct blk_mq_alloc_data data; |
07068d5b | 1504 | struct request *rq; |
7b371636 | 1505 | blk_qc_t cookie; |
87760e5e | 1506 | unsigned int wb_acct; |
07068d5b | 1507 | |
07068d5b JA |
1508 | blk_queue_bounce(q, &bio); |
1509 | ||
1510 | if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) { | |
4246a0b6 | 1511 | bio_io_error(bio); |
dece1635 | 1512 | return BLK_QC_T_NONE; |
07068d5b JA |
1513 | } |
1514 | ||
54efd50b KO |
1515 | blk_queue_split(q, &bio, q->bio_split); |
1516 | ||
87c279e6 OS |
1517 | if (!is_flush_fua && !blk_queue_nomerges(q)) { |
1518 | if (blk_attempt_plug_merge(q, bio, &request_count, NULL)) | |
1519 | return BLK_QC_T_NONE; | |
1520 | } else | |
1521 | request_count = blk_plug_queued_count(q); | |
07068d5b | 1522 | |
bd166ef1 JA |
1523 | if (blk_mq_sched_bio_merge(q, bio)) |
1524 | return BLK_QC_T_NONE; | |
1525 | ||
87760e5e JA |
1526 | wb_acct = wbt_wait(q->rq_wb, bio, NULL); |
1527 | ||
bd166ef1 JA |
1528 | trace_block_getrq(q, bio, bio->bi_opf); |
1529 | ||
1530 | rq = blk_mq_sched_get_request(q, bio, bio->bi_opf, &data); | |
87760e5e JA |
1531 | if (unlikely(!rq)) { |
1532 | __wbt_done(q->rq_wb, wb_acct); | |
dece1635 | 1533 | return BLK_QC_T_NONE; |
87760e5e JA |
1534 | } |
1535 | ||
1536 | wbt_track(&rq->issue_stat, wb_acct); | |
320ae51f | 1537 | |
fd2d3326 | 1538 | cookie = request_to_qc_t(data.hctx, rq); |
320ae51f JA |
1539 | |
1540 | if (unlikely(is_flush_fua)) { | |
1541 | blk_mq_bio_to_request(rq, bio); | |
bd166ef1 | 1542 | blk_mq_get_driver_tag(rq, NULL, true); |
320ae51f JA |
1543 | blk_insert_flush(rq); |
1544 | goto run_queue; | |
1545 | } | |
1546 | ||
1547 | /* | |
1548 | * A task plug currently exists. Since this is completely lockless, | |
1549 | * utilize that to temporarily store requests until the task is | |
1550 | * either done or scheduled away. | |
1551 | */ | |
e6c4438b JM |
1552 | plug = current->plug; |
1553 | if (plug) { | |
600271d9 SL |
1554 | struct request *last = NULL; |
1555 | ||
e6c4438b | 1556 | blk_mq_bio_to_request(rq, bio); |
0a6219a9 ML |
1557 | |
1558 | /* | |
1559 | * @request_count may become stale because of schedule | |
1560 | * out, so check the list again. | |
1561 | */ | |
1562 | if (list_empty(&plug->mq_list)) | |
1563 | request_count = 0; | |
676d0607 | 1564 | if (!request_count) |
e6c4438b | 1565 | trace_block_plug(q); |
600271d9 SL |
1566 | else |
1567 | last = list_entry_rq(plug->mq_list.prev); | |
b094f89c JA |
1568 | |
1569 | blk_mq_put_ctx(data.ctx); | |
1570 | ||
600271d9 SL |
1571 | if (request_count >= BLK_MAX_REQUEST_COUNT || (last && |
1572 | blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) { | |
e6c4438b JM |
1573 | blk_flush_plug_list(plug, false); |
1574 | trace_block_plug(q); | |
320ae51f | 1575 | } |
b094f89c | 1576 | |
e6c4438b | 1577 | list_add_tail(&rq->queuelist, &plug->mq_list); |
7b371636 | 1578 | return cookie; |
320ae51f JA |
1579 | } |
1580 | ||
bd166ef1 JA |
1581 | if (q->elevator) { |
1582 | blk_mq_put_ctx(data.ctx); | |
1583 | blk_mq_bio_to_request(rq, bio); | |
1584 | blk_mq_sched_insert_request(rq, false, true, true); | |
1585 | goto done; | |
1586 | } | |
07068d5b JA |
1587 | if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) { |
1588 | /* | |
1589 | * For a SYNC request, send it to the hardware immediately. For | |
1590 | * an ASYNC request, just ensure that we run it later on. The | |
1591 | * latter allows for merging opportunities and more efficient | |
1592 | * dispatching. | |
1593 | */ | |
1594 | run_queue: | |
1595 | blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua); | |
320ae51f JA |
1596 | } |
1597 | ||
07068d5b | 1598 | blk_mq_put_ctx(data.ctx); |
bd166ef1 | 1599 | done: |
7b371636 | 1600 | return cookie; |
320ae51f JA |
1601 | } |
1602 | ||
cc71a6f4 JA |
1603 | void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags, |
1604 | unsigned int hctx_idx) | |
95363efd | 1605 | { |
e9b267d9 | 1606 | struct page *page; |
320ae51f | 1607 | |
24d2f903 | 1608 | if (tags->rqs && set->ops->exit_request) { |
e9b267d9 | 1609 | int i; |
320ae51f | 1610 | |
24d2f903 | 1611 | for (i = 0; i < tags->nr_tags; i++) { |
2af8cbe3 JA |
1612 | struct request *rq = tags->static_rqs[i]; |
1613 | ||
1614 | if (!rq) | |
e9b267d9 | 1615 | continue; |
2af8cbe3 | 1616 | set->ops->exit_request(set->driver_data, rq, |
24d2f903 | 1617 | hctx_idx, i); |
2af8cbe3 | 1618 | tags->static_rqs[i] = NULL; |
e9b267d9 | 1619 | } |
320ae51f | 1620 | } |
320ae51f | 1621 | |
24d2f903 CH |
1622 | while (!list_empty(&tags->page_list)) { |
1623 | page = list_first_entry(&tags->page_list, struct page, lru); | |
6753471c | 1624 | list_del_init(&page->lru); |
f75782e4 CM |
1625 | /* |
1626 | * Remove kmemleak object previously allocated in | |
1627 | * blk_mq_init_rq_map(). | |
1628 | */ | |
1629 | kmemleak_free(page_address(page)); | |
320ae51f JA |
1630 | __free_pages(page, page->private); |
1631 | } | |
cc71a6f4 | 1632 | } |
320ae51f | 1633 | |
cc71a6f4 JA |
1634 | void blk_mq_free_rq_map(struct blk_mq_tags *tags) |
1635 | { | |
24d2f903 | 1636 | kfree(tags->rqs); |
cc71a6f4 | 1637 | tags->rqs = NULL; |
2af8cbe3 JA |
1638 | kfree(tags->static_rqs); |
1639 | tags->static_rqs = NULL; | |
320ae51f | 1640 | |
24d2f903 | 1641 | blk_mq_free_tags(tags); |
320ae51f JA |
1642 | } |
1643 | ||
cc71a6f4 JA |
1644 | struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, |
1645 | unsigned int hctx_idx, | |
1646 | unsigned int nr_tags, | |
1647 | unsigned int reserved_tags) | |
320ae51f | 1648 | { |
24d2f903 | 1649 | struct blk_mq_tags *tags; |
320ae51f | 1650 | |
cc71a6f4 | 1651 | tags = blk_mq_init_tags(nr_tags, reserved_tags, |
24391c0d SL |
1652 | set->numa_node, |
1653 | BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags)); | |
24d2f903 CH |
1654 | if (!tags) |
1655 | return NULL; | |
320ae51f | 1656 | |
cc71a6f4 | 1657 | tags->rqs = kzalloc_node(nr_tags * sizeof(struct request *), |
36e1f3d1 | 1658 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, |
a5164405 | 1659 | set->numa_node); |
24d2f903 CH |
1660 | if (!tags->rqs) { |
1661 | blk_mq_free_tags(tags); | |
1662 | return NULL; | |
1663 | } | |
320ae51f | 1664 | |
2af8cbe3 JA |
1665 | tags->static_rqs = kzalloc_node(nr_tags * sizeof(struct request *), |
1666 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, | |
1667 | set->numa_node); | |
1668 | if (!tags->static_rqs) { | |
1669 | kfree(tags->rqs); | |
1670 | blk_mq_free_tags(tags); | |
1671 | return NULL; | |
1672 | } | |
1673 | ||
cc71a6f4 JA |
1674 | return tags; |
1675 | } | |
1676 | ||
1677 | static size_t order_to_size(unsigned int order) | |
1678 | { | |
1679 | return (size_t)PAGE_SIZE << order; | |
1680 | } | |
1681 | ||
1682 | int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags, | |
1683 | unsigned int hctx_idx, unsigned int depth) | |
1684 | { | |
1685 | unsigned int i, j, entries_per_page, max_order = 4; | |
1686 | size_t rq_size, left; | |
1687 | ||
1688 | INIT_LIST_HEAD(&tags->page_list); | |
1689 | ||
320ae51f JA |
1690 | /* |
1691 | * rq_size is the size of the request plus driver payload, rounded | |
1692 | * to the cacheline size | |
1693 | */ | |
24d2f903 | 1694 | rq_size = round_up(sizeof(struct request) + set->cmd_size, |
320ae51f | 1695 | cache_line_size()); |
cc71a6f4 | 1696 | left = rq_size * depth; |
320ae51f | 1697 | |
cc71a6f4 | 1698 | for (i = 0; i < depth; ) { |
320ae51f JA |
1699 | int this_order = max_order; |
1700 | struct page *page; | |
1701 | int to_do; | |
1702 | void *p; | |
1703 | ||
b3a834b1 | 1704 | while (this_order && left < order_to_size(this_order - 1)) |
320ae51f JA |
1705 | this_order--; |
1706 | ||
1707 | do { | |
a5164405 | 1708 | page = alloc_pages_node(set->numa_node, |
36e1f3d1 | 1709 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO, |
a5164405 | 1710 | this_order); |
320ae51f JA |
1711 | if (page) |
1712 | break; | |
1713 | if (!this_order--) | |
1714 | break; | |
1715 | if (order_to_size(this_order) < rq_size) | |
1716 | break; | |
1717 | } while (1); | |
1718 | ||
1719 | if (!page) | |
24d2f903 | 1720 | goto fail; |
320ae51f JA |
1721 | |
1722 | page->private = this_order; | |
24d2f903 | 1723 | list_add_tail(&page->lru, &tags->page_list); |
320ae51f JA |
1724 | |
1725 | p = page_address(page); | |
f75782e4 CM |
1726 | /* |
1727 | * Allow kmemleak to scan these pages as they contain pointers | |
1728 | * to additional allocations like via ops->init_request(). | |
1729 | */ | |
36e1f3d1 | 1730 | kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO); |
320ae51f | 1731 | entries_per_page = order_to_size(this_order) / rq_size; |
cc71a6f4 | 1732 | to_do = min(entries_per_page, depth - i); |
320ae51f JA |
1733 | left -= to_do * rq_size; |
1734 | for (j = 0; j < to_do; j++) { | |
2af8cbe3 JA |
1735 | struct request *rq = p; |
1736 | ||
1737 | tags->static_rqs[i] = rq; | |
24d2f903 CH |
1738 | if (set->ops->init_request) { |
1739 | if (set->ops->init_request(set->driver_data, | |
2af8cbe3 | 1740 | rq, hctx_idx, i, |
a5164405 | 1741 | set->numa_node)) { |
2af8cbe3 | 1742 | tags->static_rqs[i] = NULL; |
24d2f903 | 1743 | goto fail; |
a5164405 | 1744 | } |
e9b267d9 CH |
1745 | } |
1746 | ||
320ae51f JA |
1747 | p += rq_size; |
1748 | i++; | |
1749 | } | |
1750 | } | |
cc71a6f4 | 1751 | return 0; |
320ae51f | 1752 | |
24d2f903 | 1753 | fail: |
cc71a6f4 JA |
1754 | blk_mq_free_rqs(set, tags, hctx_idx); |
1755 | return -ENOMEM; | |
320ae51f JA |
1756 | } |
1757 | ||
e57690fe JA |
1758 | /* |
1759 | * 'cpu' is going away. splice any existing rq_list entries from this | |
1760 | * software queue to the hw queue dispatch list, and ensure that it | |
1761 | * gets run. | |
1762 | */ | |
9467f859 | 1763 | static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node) |
484b4061 | 1764 | { |
9467f859 | 1765 | struct blk_mq_hw_ctx *hctx; |
484b4061 JA |
1766 | struct blk_mq_ctx *ctx; |
1767 | LIST_HEAD(tmp); | |
1768 | ||
9467f859 | 1769 | hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead); |
e57690fe | 1770 | ctx = __blk_mq_get_ctx(hctx->queue, cpu); |
484b4061 JA |
1771 | |
1772 | spin_lock(&ctx->lock); | |
1773 | if (!list_empty(&ctx->rq_list)) { | |
1774 | list_splice_init(&ctx->rq_list, &tmp); | |
1775 | blk_mq_hctx_clear_pending(hctx, ctx); | |
1776 | } | |
1777 | spin_unlock(&ctx->lock); | |
1778 | ||
1779 | if (list_empty(&tmp)) | |
9467f859 | 1780 | return 0; |
484b4061 | 1781 | |
e57690fe JA |
1782 | spin_lock(&hctx->lock); |
1783 | list_splice_tail_init(&tmp, &hctx->dispatch); | |
1784 | spin_unlock(&hctx->lock); | |
484b4061 JA |
1785 | |
1786 | blk_mq_run_hw_queue(hctx, true); | |
9467f859 | 1787 | return 0; |
484b4061 JA |
1788 | } |
1789 | ||
9467f859 | 1790 | static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx) |
484b4061 | 1791 | { |
9467f859 TG |
1792 | cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD, |
1793 | &hctx->cpuhp_dead); | |
484b4061 JA |
1794 | } |
1795 | ||
c3b4afca | 1796 | /* hctx->ctxs will be freed in queue's release handler */ |
08e98fc6 ML |
1797 | static void blk_mq_exit_hctx(struct request_queue *q, |
1798 | struct blk_mq_tag_set *set, | |
1799 | struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx) | |
1800 | { | |
f70ced09 ML |
1801 | unsigned flush_start_tag = set->queue_depth; |
1802 | ||
08e98fc6 ML |
1803 | blk_mq_tag_idle(hctx); |
1804 | ||
f70ced09 ML |
1805 | if (set->ops->exit_request) |
1806 | set->ops->exit_request(set->driver_data, | |
1807 | hctx->fq->flush_rq, hctx_idx, | |
1808 | flush_start_tag + hctx_idx); | |
1809 | ||
08e98fc6 ML |
1810 | if (set->ops->exit_hctx) |
1811 | set->ops->exit_hctx(hctx, hctx_idx); | |
1812 | ||
6a83e74d BVA |
1813 | if (hctx->flags & BLK_MQ_F_BLOCKING) |
1814 | cleanup_srcu_struct(&hctx->queue_rq_srcu); | |
1815 | ||
9467f859 | 1816 | blk_mq_remove_cpuhp(hctx); |
f70ced09 | 1817 | blk_free_flush_queue(hctx->fq); |
88459642 | 1818 | sbitmap_free(&hctx->ctx_map); |
08e98fc6 ML |
1819 | } |
1820 | ||
624dbe47 ML |
1821 | static void blk_mq_exit_hw_queues(struct request_queue *q, |
1822 | struct blk_mq_tag_set *set, int nr_queue) | |
1823 | { | |
1824 | struct blk_mq_hw_ctx *hctx; | |
1825 | unsigned int i; | |
1826 | ||
1827 | queue_for_each_hw_ctx(q, hctx, i) { | |
1828 | if (i == nr_queue) | |
1829 | break; | |
08e98fc6 | 1830 | blk_mq_exit_hctx(q, set, hctx, i); |
624dbe47 | 1831 | } |
624dbe47 ML |
1832 | } |
1833 | ||
1834 | static void blk_mq_free_hw_queues(struct request_queue *q, | |
1835 | struct blk_mq_tag_set *set) | |
1836 | { | |
1837 | struct blk_mq_hw_ctx *hctx; | |
1838 | unsigned int i; | |
1839 | ||
e09aae7e | 1840 | queue_for_each_hw_ctx(q, hctx, i) |
624dbe47 | 1841 | free_cpumask_var(hctx->cpumask); |
624dbe47 ML |
1842 | } |
1843 | ||
08e98fc6 ML |
1844 | static int blk_mq_init_hctx(struct request_queue *q, |
1845 | struct blk_mq_tag_set *set, | |
1846 | struct blk_mq_hw_ctx *hctx, unsigned hctx_idx) | |
320ae51f | 1847 | { |
08e98fc6 | 1848 | int node; |
f70ced09 | 1849 | unsigned flush_start_tag = set->queue_depth; |
08e98fc6 ML |
1850 | |
1851 | node = hctx->numa_node; | |
1852 | if (node == NUMA_NO_NODE) | |
1853 | node = hctx->numa_node = set->numa_node; | |
1854 | ||
27489a3c | 1855 | INIT_WORK(&hctx->run_work, blk_mq_run_work_fn); |
08e98fc6 ML |
1856 | INIT_DELAYED_WORK(&hctx->delay_work, blk_mq_delay_work_fn); |
1857 | spin_lock_init(&hctx->lock); | |
1858 | INIT_LIST_HEAD(&hctx->dispatch); | |
1859 | hctx->queue = q; | |
1860 | hctx->queue_num = hctx_idx; | |
2404e607 | 1861 | hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED; |
08e98fc6 | 1862 | |
9467f859 | 1863 | cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead); |
08e98fc6 ML |
1864 | |
1865 | hctx->tags = set->tags[hctx_idx]; | |
320ae51f JA |
1866 | |
1867 | /* | |
08e98fc6 ML |
1868 | * Allocate space for all possible cpus to avoid allocation at |
1869 | * runtime | |
320ae51f | 1870 | */ |
08e98fc6 ML |
1871 | hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *), |
1872 | GFP_KERNEL, node); | |
1873 | if (!hctx->ctxs) | |
1874 | goto unregister_cpu_notifier; | |
320ae51f | 1875 | |
88459642 OS |
1876 | if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8), GFP_KERNEL, |
1877 | node)) | |
08e98fc6 | 1878 | goto free_ctxs; |
320ae51f | 1879 | |
08e98fc6 | 1880 | hctx->nr_ctx = 0; |
320ae51f | 1881 | |
08e98fc6 ML |
1882 | if (set->ops->init_hctx && |
1883 | set->ops->init_hctx(hctx, set->driver_data, hctx_idx)) | |
1884 | goto free_bitmap; | |
320ae51f | 1885 | |
f70ced09 ML |
1886 | hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size); |
1887 | if (!hctx->fq) | |
1888 | goto exit_hctx; | |
320ae51f | 1889 | |
f70ced09 ML |
1890 | if (set->ops->init_request && |
1891 | set->ops->init_request(set->driver_data, | |
1892 | hctx->fq->flush_rq, hctx_idx, | |
1893 | flush_start_tag + hctx_idx, node)) | |
1894 | goto free_fq; | |
320ae51f | 1895 | |
6a83e74d BVA |
1896 | if (hctx->flags & BLK_MQ_F_BLOCKING) |
1897 | init_srcu_struct(&hctx->queue_rq_srcu); | |
1898 | ||
08e98fc6 | 1899 | return 0; |
320ae51f | 1900 | |
f70ced09 ML |
1901 | free_fq: |
1902 | kfree(hctx->fq); | |
1903 | exit_hctx: | |
1904 | if (set->ops->exit_hctx) | |
1905 | set->ops->exit_hctx(hctx, hctx_idx); | |
08e98fc6 | 1906 | free_bitmap: |
88459642 | 1907 | sbitmap_free(&hctx->ctx_map); |
08e98fc6 ML |
1908 | free_ctxs: |
1909 | kfree(hctx->ctxs); | |
1910 | unregister_cpu_notifier: | |
9467f859 | 1911 | blk_mq_remove_cpuhp(hctx); |
08e98fc6 ML |
1912 | return -1; |
1913 | } | |
320ae51f | 1914 | |
320ae51f JA |
1915 | static void blk_mq_init_cpu_queues(struct request_queue *q, |
1916 | unsigned int nr_hw_queues) | |
1917 | { | |
1918 | unsigned int i; | |
1919 | ||
1920 | for_each_possible_cpu(i) { | |
1921 | struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i); | |
1922 | struct blk_mq_hw_ctx *hctx; | |
1923 | ||
1924 | memset(__ctx, 0, sizeof(*__ctx)); | |
1925 | __ctx->cpu = i; | |
1926 | spin_lock_init(&__ctx->lock); | |
1927 | INIT_LIST_HEAD(&__ctx->rq_list); | |
1928 | __ctx->queue = q; | |
cf43e6be JA |
1929 | blk_stat_init(&__ctx->stat[BLK_STAT_READ]); |
1930 | blk_stat_init(&__ctx->stat[BLK_STAT_WRITE]); | |
320ae51f JA |
1931 | |
1932 | /* If the cpu isn't online, the cpu is mapped to first hctx */ | |
320ae51f JA |
1933 | if (!cpu_online(i)) |
1934 | continue; | |
1935 | ||
7d7e0f90 | 1936 | hctx = blk_mq_map_queue(q, i); |
e4043dcf | 1937 | |
320ae51f JA |
1938 | /* |
1939 | * Set local node, IFF we have more than one hw queue. If | |
1940 | * not, we remain on the home node of the device | |
1941 | */ | |
1942 | if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE) | |
bffed457 | 1943 | hctx->numa_node = local_memory_node(cpu_to_node(i)); |
320ae51f JA |
1944 | } |
1945 | } | |
1946 | ||
cc71a6f4 JA |
1947 | static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, int hctx_idx) |
1948 | { | |
1949 | int ret = 0; | |
1950 | ||
1951 | set->tags[hctx_idx] = blk_mq_alloc_rq_map(set, hctx_idx, | |
1952 | set->queue_depth, set->reserved_tags); | |
1953 | if (!set->tags[hctx_idx]) | |
1954 | return false; | |
1955 | ||
1956 | ret = blk_mq_alloc_rqs(set, set->tags[hctx_idx], hctx_idx, | |
1957 | set->queue_depth); | |
1958 | if (!ret) | |
1959 | return true; | |
1960 | ||
1961 | blk_mq_free_rq_map(set->tags[hctx_idx]); | |
1962 | set->tags[hctx_idx] = NULL; | |
1963 | return false; | |
1964 | } | |
1965 | ||
1966 | static void blk_mq_free_map_and_requests(struct blk_mq_tag_set *set, | |
1967 | unsigned int hctx_idx) | |
1968 | { | |
bd166ef1 JA |
1969 | if (set->tags[hctx_idx]) { |
1970 | blk_mq_free_rqs(set, set->tags[hctx_idx], hctx_idx); | |
1971 | blk_mq_free_rq_map(set->tags[hctx_idx]); | |
1972 | set->tags[hctx_idx] = NULL; | |
1973 | } | |
cc71a6f4 JA |
1974 | } |
1975 | ||
5778322e AM |
1976 | static void blk_mq_map_swqueue(struct request_queue *q, |
1977 | const struct cpumask *online_mask) | |
320ae51f | 1978 | { |
d1b1cea1 | 1979 | unsigned int i, hctx_idx; |
320ae51f JA |
1980 | struct blk_mq_hw_ctx *hctx; |
1981 | struct blk_mq_ctx *ctx; | |
2a34c087 | 1982 | struct blk_mq_tag_set *set = q->tag_set; |
320ae51f | 1983 | |
60de074b AM |
1984 | /* |
1985 | * Avoid others reading imcomplete hctx->cpumask through sysfs | |
1986 | */ | |
1987 | mutex_lock(&q->sysfs_lock); | |
1988 | ||
320ae51f | 1989 | queue_for_each_hw_ctx(q, hctx, i) { |
e4043dcf | 1990 | cpumask_clear(hctx->cpumask); |
320ae51f JA |
1991 | hctx->nr_ctx = 0; |
1992 | } | |
1993 | ||
1994 | /* | |
1995 | * Map software to hardware queues | |
1996 | */ | |
897bb0c7 | 1997 | for_each_possible_cpu(i) { |
320ae51f | 1998 | /* If the cpu isn't online, the cpu is mapped to first hctx */ |
5778322e | 1999 | if (!cpumask_test_cpu(i, online_mask)) |
e4043dcf JA |
2000 | continue; |
2001 | ||
d1b1cea1 GKB |
2002 | hctx_idx = q->mq_map[i]; |
2003 | /* unmapped hw queue can be remapped after CPU topo changed */ | |
cc71a6f4 JA |
2004 | if (!set->tags[hctx_idx] && |
2005 | !__blk_mq_alloc_rq_map(set, hctx_idx)) { | |
d1b1cea1 GKB |
2006 | /* |
2007 | * If tags initialization fail for some hctx, | |
2008 | * that hctx won't be brought online. In this | |
2009 | * case, remap the current ctx to hctx[0] which | |
2010 | * is guaranteed to always have tags allocated | |
2011 | */ | |
cc71a6f4 | 2012 | q->mq_map[i] = 0; |
d1b1cea1 GKB |
2013 | } |
2014 | ||
897bb0c7 | 2015 | ctx = per_cpu_ptr(q->queue_ctx, i); |
7d7e0f90 | 2016 | hctx = blk_mq_map_queue(q, i); |
868f2f0b | 2017 | |
e4043dcf | 2018 | cpumask_set_cpu(i, hctx->cpumask); |
320ae51f JA |
2019 | ctx->index_hw = hctx->nr_ctx; |
2020 | hctx->ctxs[hctx->nr_ctx++] = ctx; | |
2021 | } | |
506e931f | 2022 | |
60de074b AM |
2023 | mutex_unlock(&q->sysfs_lock); |
2024 | ||
506e931f | 2025 | queue_for_each_hw_ctx(q, hctx, i) { |
484b4061 | 2026 | /* |
a68aafa5 JA |
2027 | * If no software queues are mapped to this hardware queue, |
2028 | * disable it and free the request entries. | |
484b4061 JA |
2029 | */ |
2030 | if (!hctx->nr_ctx) { | |
d1b1cea1 GKB |
2031 | /* Never unmap queue 0. We need it as a |
2032 | * fallback in case of a new remap fails | |
2033 | * allocation | |
2034 | */ | |
cc71a6f4 JA |
2035 | if (i && set->tags[i]) |
2036 | blk_mq_free_map_and_requests(set, i); | |
2037 | ||
2a34c087 | 2038 | hctx->tags = NULL; |
484b4061 JA |
2039 | continue; |
2040 | } | |
2041 | ||
2a34c087 ML |
2042 | hctx->tags = set->tags[i]; |
2043 | WARN_ON(!hctx->tags); | |
2044 | ||
889fa31f CY |
2045 | /* |
2046 | * Set the map size to the number of mapped software queues. | |
2047 | * This is more accurate and more efficient than looping | |
2048 | * over all possibly mapped software queues. | |
2049 | */ | |
88459642 | 2050 | sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx); |
889fa31f | 2051 | |
484b4061 JA |
2052 | /* |
2053 | * Initialize batch roundrobin counts | |
2054 | */ | |
506e931f JA |
2055 | hctx->next_cpu = cpumask_first(hctx->cpumask); |
2056 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; | |
2057 | } | |
320ae51f JA |
2058 | } |
2059 | ||
2404e607 | 2060 | static void queue_set_hctx_shared(struct request_queue *q, bool shared) |
0d2602ca JA |
2061 | { |
2062 | struct blk_mq_hw_ctx *hctx; | |
0d2602ca JA |
2063 | int i; |
2064 | ||
2404e607 JM |
2065 | queue_for_each_hw_ctx(q, hctx, i) { |
2066 | if (shared) | |
2067 | hctx->flags |= BLK_MQ_F_TAG_SHARED; | |
2068 | else | |
2069 | hctx->flags &= ~BLK_MQ_F_TAG_SHARED; | |
2070 | } | |
2071 | } | |
2072 | ||
2073 | static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set, bool shared) | |
2074 | { | |
2075 | struct request_queue *q; | |
0d2602ca JA |
2076 | |
2077 | list_for_each_entry(q, &set->tag_list, tag_set_list) { | |
2078 | blk_mq_freeze_queue(q); | |
2404e607 | 2079 | queue_set_hctx_shared(q, shared); |
0d2602ca JA |
2080 | blk_mq_unfreeze_queue(q); |
2081 | } | |
2082 | } | |
2083 | ||
2084 | static void blk_mq_del_queue_tag_set(struct request_queue *q) | |
2085 | { | |
2086 | struct blk_mq_tag_set *set = q->tag_set; | |
2087 | ||
0d2602ca JA |
2088 | mutex_lock(&set->tag_list_lock); |
2089 | list_del_init(&q->tag_set_list); | |
2404e607 JM |
2090 | if (list_is_singular(&set->tag_list)) { |
2091 | /* just transitioned to unshared */ | |
2092 | set->flags &= ~BLK_MQ_F_TAG_SHARED; | |
2093 | /* update existing queue */ | |
2094 | blk_mq_update_tag_set_depth(set, false); | |
2095 | } | |
0d2602ca | 2096 | mutex_unlock(&set->tag_list_lock); |
0d2602ca JA |
2097 | } |
2098 | ||
2099 | static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set, | |
2100 | struct request_queue *q) | |
2101 | { | |
2102 | q->tag_set = set; | |
2103 | ||
2104 | mutex_lock(&set->tag_list_lock); | |
2404e607 JM |
2105 | |
2106 | /* Check to see if we're transitioning to shared (from 1 to 2 queues). */ | |
2107 | if (!list_empty(&set->tag_list) && !(set->flags & BLK_MQ_F_TAG_SHARED)) { | |
2108 | set->flags |= BLK_MQ_F_TAG_SHARED; | |
2109 | /* update existing queue */ | |
2110 | blk_mq_update_tag_set_depth(set, true); | |
2111 | } | |
2112 | if (set->flags & BLK_MQ_F_TAG_SHARED) | |
2113 | queue_set_hctx_shared(q, true); | |
0d2602ca | 2114 | list_add_tail(&q->tag_set_list, &set->tag_list); |
2404e607 | 2115 | |
0d2602ca JA |
2116 | mutex_unlock(&set->tag_list_lock); |
2117 | } | |
2118 | ||
e09aae7e ML |
2119 | /* |
2120 | * It is the actual release handler for mq, but we do it from | |
2121 | * request queue's release handler for avoiding use-after-free | |
2122 | * and headache because q->mq_kobj shouldn't have been introduced, | |
2123 | * but we can't group ctx/kctx kobj without it. | |
2124 | */ | |
2125 | void blk_mq_release(struct request_queue *q) | |
2126 | { | |
2127 | struct blk_mq_hw_ctx *hctx; | |
2128 | unsigned int i; | |
2129 | ||
bd166ef1 JA |
2130 | blk_mq_sched_teardown(q); |
2131 | ||
e09aae7e | 2132 | /* hctx kobj stays in hctx */ |
c3b4afca ML |
2133 | queue_for_each_hw_ctx(q, hctx, i) { |
2134 | if (!hctx) | |
2135 | continue; | |
2136 | kfree(hctx->ctxs); | |
e09aae7e | 2137 | kfree(hctx); |
c3b4afca | 2138 | } |
e09aae7e | 2139 | |
a723bab3 AM |
2140 | q->mq_map = NULL; |
2141 | ||
e09aae7e ML |
2142 | kfree(q->queue_hw_ctx); |
2143 | ||
2144 | /* ctx kobj stays in queue_ctx */ | |
2145 | free_percpu(q->queue_ctx); | |
2146 | } | |
2147 | ||
24d2f903 | 2148 | struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set) |
b62c21b7 MS |
2149 | { |
2150 | struct request_queue *uninit_q, *q; | |
2151 | ||
2152 | uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node); | |
2153 | if (!uninit_q) | |
2154 | return ERR_PTR(-ENOMEM); | |
2155 | ||
2156 | q = blk_mq_init_allocated_queue(set, uninit_q); | |
2157 | if (IS_ERR(q)) | |
2158 | blk_cleanup_queue(uninit_q); | |
2159 | ||
2160 | return q; | |
2161 | } | |
2162 | EXPORT_SYMBOL(blk_mq_init_queue); | |
2163 | ||
868f2f0b KB |
2164 | static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set, |
2165 | struct request_queue *q) | |
320ae51f | 2166 | { |
868f2f0b KB |
2167 | int i, j; |
2168 | struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx; | |
f14bbe77 | 2169 | |
868f2f0b | 2170 | blk_mq_sysfs_unregister(q); |
24d2f903 | 2171 | for (i = 0; i < set->nr_hw_queues; i++) { |
868f2f0b | 2172 | int node; |
f14bbe77 | 2173 | |
868f2f0b KB |
2174 | if (hctxs[i]) |
2175 | continue; | |
2176 | ||
2177 | node = blk_mq_hw_queue_to_node(q->mq_map, i); | |
cdef54dd CH |
2178 | hctxs[i] = kzalloc_node(sizeof(struct blk_mq_hw_ctx), |
2179 | GFP_KERNEL, node); | |
320ae51f | 2180 | if (!hctxs[i]) |
868f2f0b | 2181 | break; |
320ae51f | 2182 | |
a86073e4 | 2183 | if (!zalloc_cpumask_var_node(&hctxs[i]->cpumask, GFP_KERNEL, |
868f2f0b KB |
2184 | node)) { |
2185 | kfree(hctxs[i]); | |
2186 | hctxs[i] = NULL; | |
2187 | break; | |
2188 | } | |
e4043dcf | 2189 | |
0d2602ca | 2190 | atomic_set(&hctxs[i]->nr_active, 0); |
f14bbe77 | 2191 | hctxs[i]->numa_node = node; |
320ae51f | 2192 | hctxs[i]->queue_num = i; |
868f2f0b KB |
2193 | |
2194 | if (blk_mq_init_hctx(q, set, hctxs[i], i)) { | |
2195 | free_cpumask_var(hctxs[i]->cpumask); | |
2196 | kfree(hctxs[i]); | |
2197 | hctxs[i] = NULL; | |
2198 | break; | |
2199 | } | |
2200 | blk_mq_hctx_kobj_init(hctxs[i]); | |
320ae51f | 2201 | } |
868f2f0b KB |
2202 | for (j = i; j < q->nr_hw_queues; j++) { |
2203 | struct blk_mq_hw_ctx *hctx = hctxs[j]; | |
2204 | ||
2205 | if (hctx) { | |
cc71a6f4 JA |
2206 | if (hctx->tags) |
2207 | blk_mq_free_map_and_requests(set, j); | |
868f2f0b KB |
2208 | blk_mq_exit_hctx(q, set, hctx, j); |
2209 | free_cpumask_var(hctx->cpumask); | |
2210 | kobject_put(&hctx->kobj); | |
2211 | kfree(hctx->ctxs); | |
2212 | kfree(hctx); | |
2213 | hctxs[j] = NULL; | |
2214 | ||
2215 | } | |
2216 | } | |
2217 | q->nr_hw_queues = i; | |
2218 | blk_mq_sysfs_register(q); | |
2219 | } | |
2220 | ||
2221 | struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set, | |
2222 | struct request_queue *q) | |
2223 | { | |
66841672 ML |
2224 | /* mark the queue as mq asap */ |
2225 | q->mq_ops = set->ops; | |
2226 | ||
868f2f0b KB |
2227 | q->queue_ctx = alloc_percpu(struct blk_mq_ctx); |
2228 | if (!q->queue_ctx) | |
c7de5726 | 2229 | goto err_exit; |
868f2f0b KB |
2230 | |
2231 | q->queue_hw_ctx = kzalloc_node(nr_cpu_ids * sizeof(*(q->queue_hw_ctx)), | |
2232 | GFP_KERNEL, set->numa_node); | |
2233 | if (!q->queue_hw_ctx) | |
2234 | goto err_percpu; | |
2235 | ||
bdd17e75 | 2236 | q->mq_map = set->mq_map; |
868f2f0b KB |
2237 | |
2238 | blk_mq_realloc_hw_ctxs(set, q); | |
2239 | if (!q->nr_hw_queues) | |
2240 | goto err_hctxs; | |
320ae51f | 2241 | |
287922eb | 2242 | INIT_WORK(&q->timeout_work, blk_mq_timeout_work); |
e56f698b | 2243 | blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ); |
320ae51f JA |
2244 | |
2245 | q->nr_queues = nr_cpu_ids; | |
320ae51f | 2246 | |
94eddfbe | 2247 | q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT; |
320ae51f | 2248 | |
05f1dd53 JA |
2249 | if (!(set->flags & BLK_MQ_F_SG_MERGE)) |
2250 | q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE; | |
2251 | ||
1be036e9 CH |
2252 | q->sg_reserved_size = INT_MAX; |
2253 | ||
2849450a | 2254 | INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work); |
6fca6a61 CH |
2255 | INIT_LIST_HEAD(&q->requeue_list); |
2256 | spin_lock_init(&q->requeue_lock); | |
2257 | ||
07068d5b JA |
2258 | if (q->nr_hw_queues > 1) |
2259 | blk_queue_make_request(q, blk_mq_make_request); | |
2260 | else | |
2261 | blk_queue_make_request(q, blk_sq_make_request); | |
2262 | ||
eba71768 JA |
2263 | /* |
2264 | * Do this after blk_queue_make_request() overrides it... | |
2265 | */ | |
2266 | q->nr_requests = set->queue_depth; | |
2267 | ||
64f1c21e JA |
2268 | /* |
2269 | * Default to classic polling | |
2270 | */ | |
2271 | q->poll_nsec = -1; | |
2272 | ||
24d2f903 CH |
2273 | if (set->ops->complete) |
2274 | blk_queue_softirq_done(q, set->ops->complete); | |
30a91cb4 | 2275 | |
24d2f903 | 2276 | blk_mq_init_cpu_queues(q, set->nr_hw_queues); |
320ae51f | 2277 | |
5778322e | 2278 | get_online_cpus(); |
320ae51f | 2279 | mutex_lock(&all_q_mutex); |
320ae51f | 2280 | |
4593fdbe | 2281 | list_add_tail(&q->all_q_node, &all_q_list); |
0d2602ca | 2282 | blk_mq_add_queue_tag_set(set, q); |
5778322e | 2283 | blk_mq_map_swqueue(q, cpu_online_mask); |
484b4061 | 2284 | |
4593fdbe | 2285 | mutex_unlock(&all_q_mutex); |
5778322e | 2286 | put_online_cpus(); |
4593fdbe | 2287 | |
320ae51f | 2288 | return q; |
18741986 | 2289 | |
320ae51f | 2290 | err_hctxs: |
868f2f0b | 2291 | kfree(q->queue_hw_ctx); |
320ae51f | 2292 | err_percpu: |
868f2f0b | 2293 | free_percpu(q->queue_ctx); |
c7de5726 ML |
2294 | err_exit: |
2295 | q->mq_ops = NULL; | |
320ae51f JA |
2296 | return ERR_PTR(-ENOMEM); |
2297 | } | |
b62c21b7 | 2298 | EXPORT_SYMBOL(blk_mq_init_allocated_queue); |
320ae51f JA |
2299 | |
2300 | void blk_mq_free_queue(struct request_queue *q) | |
2301 | { | |
624dbe47 | 2302 | struct blk_mq_tag_set *set = q->tag_set; |
320ae51f | 2303 | |
0e626368 AM |
2304 | mutex_lock(&all_q_mutex); |
2305 | list_del_init(&q->all_q_node); | |
2306 | mutex_unlock(&all_q_mutex); | |
2307 | ||
87760e5e JA |
2308 | wbt_exit(q); |
2309 | ||
0d2602ca JA |
2310 | blk_mq_del_queue_tag_set(q); |
2311 | ||
624dbe47 ML |
2312 | blk_mq_exit_hw_queues(q, set, set->nr_hw_queues); |
2313 | blk_mq_free_hw_queues(q, set); | |
320ae51f | 2314 | } |
320ae51f JA |
2315 | |
2316 | /* Basically redo blk_mq_init_queue with queue frozen */ | |
5778322e AM |
2317 | static void blk_mq_queue_reinit(struct request_queue *q, |
2318 | const struct cpumask *online_mask) | |
320ae51f | 2319 | { |
4ecd4fef | 2320 | WARN_ON_ONCE(!atomic_read(&q->mq_freeze_depth)); |
320ae51f | 2321 | |
67aec14c JA |
2322 | blk_mq_sysfs_unregister(q); |
2323 | ||
320ae51f JA |
2324 | /* |
2325 | * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe | |
2326 | * we should change hctx numa_node according to new topology (this | |
2327 | * involves free and re-allocate memory, worthy doing?) | |
2328 | */ | |
2329 | ||
5778322e | 2330 | blk_mq_map_swqueue(q, online_mask); |
320ae51f | 2331 | |
67aec14c | 2332 | blk_mq_sysfs_register(q); |
320ae51f JA |
2333 | } |
2334 | ||
65d5291e SAS |
2335 | /* |
2336 | * New online cpumask which is going to be set in this hotplug event. | |
2337 | * Declare this cpumasks as global as cpu-hotplug operation is invoked | |
2338 | * one-by-one and dynamically allocating this could result in a failure. | |
2339 | */ | |
2340 | static struct cpumask cpuhp_online_new; | |
2341 | ||
2342 | static void blk_mq_queue_reinit_work(void) | |
320ae51f JA |
2343 | { |
2344 | struct request_queue *q; | |
320ae51f JA |
2345 | |
2346 | mutex_lock(&all_q_mutex); | |
f3af020b TH |
2347 | /* |
2348 | * We need to freeze and reinit all existing queues. Freezing | |
2349 | * involves synchronous wait for an RCU grace period and doing it | |
2350 | * one by one may take a long time. Start freezing all queues in | |
2351 | * one swoop and then wait for the completions so that freezing can | |
2352 | * take place in parallel. | |
2353 | */ | |
2354 | list_for_each_entry(q, &all_q_list, all_q_node) | |
2355 | blk_mq_freeze_queue_start(q); | |
415d3dab | 2356 | list_for_each_entry(q, &all_q_list, all_q_node) |
f3af020b TH |
2357 | blk_mq_freeze_queue_wait(q); |
2358 | ||
320ae51f | 2359 | list_for_each_entry(q, &all_q_list, all_q_node) |
65d5291e | 2360 | blk_mq_queue_reinit(q, &cpuhp_online_new); |
f3af020b TH |
2361 | |
2362 | list_for_each_entry(q, &all_q_list, all_q_node) | |
2363 | blk_mq_unfreeze_queue(q); | |
2364 | ||
320ae51f | 2365 | mutex_unlock(&all_q_mutex); |
65d5291e SAS |
2366 | } |
2367 | ||
2368 | static int blk_mq_queue_reinit_dead(unsigned int cpu) | |
2369 | { | |
97a32864 | 2370 | cpumask_copy(&cpuhp_online_new, cpu_online_mask); |
65d5291e SAS |
2371 | blk_mq_queue_reinit_work(); |
2372 | return 0; | |
2373 | } | |
2374 | ||
2375 | /* | |
2376 | * Before hotadded cpu starts handling requests, new mappings must be | |
2377 | * established. Otherwise, these requests in hw queue might never be | |
2378 | * dispatched. | |
2379 | * | |
2380 | * For example, there is a single hw queue (hctx) and two CPU queues (ctx0 | |
2381 | * for CPU0, and ctx1 for CPU1). | |
2382 | * | |
2383 | * Now CPU1 is just onlined and a request is inserted into ctx1->rq_list | |
2384 | * and set bit0 in pending bitmap as ctx1->index_hw is still zero. | |
2385 | * | |
2c3ad667 JA |
2386 | * And then while running hw queue, blk_mq_flush_busy_ctxs() finds bit0 is set |
2387 | * in pending bitmap and tries to retrieve requests in hctx->ctxs[0]->rq_list. | |
2388 | * But htx->ctxs[0] is a pointer to ctx0, so the request in ctx1->rq_list is | |
2389 | * ignored. | |
65d5291e SAS |
2390 | */ |
2391 | static int blk_mq_queue_reinit_prepare(unsigned int cpu) | |
2392 | { | |
2393 | cpumask_copy(&cpuhp_online_new, cpu_online_mask); | |
2394 | cpumask_set_cpu(cpu, &cpuhp_online_new); | |
2395 | blk_mq_queue_reinit_work(); | |
2396 | return 0; | |
320ae51f JA |
2397 | } |
2398 | ||
a5164405 JA |
2399 | static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set) |
2400 | { | |
2401 | int i; | |
2402 | ||
cc71a6f4 JA |
2403 | for (i = 0; i < set->nr_hw_queues; i++) |
2404 | if (!__blk_mq_alloc_rq_map(set, i)) | |
a5164405 | 2405 | goto out_unwind; |
a5164405 JA |
2406 | |
2407 | return 0; | |
2408 | ||
2409 | out_unwind: | |
2410 | while (--i >= 0) | |
cc71a6f4 | 2411 | blk_mq_free_rq_map(set->tags[i]); |
a5164405 | 2412 | |
a5164405 JA |
2413 | return -ENOMEM; |
2414 | } | |
2415 | ||
2416 | /* | |
2417 | * Allocate the request maps associated with this tag_set. Note that this | |
2418 | * may reduce the depth asked for, if memory is tight. set->queue_depth | |
2419 | * will be updated to reflect the allocated depth. | |
2420 | */ | |
2421 | static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set) | |
2422 | { | |
2423 | unsigned int depth; | |
2424 | int err; | |
2425 | ||
2426 | depth = set->queue_depth; | |
2427 | do { | |
2428 | err = __blk_mq_alloc_rq_maps(set); | |
2429 | if (!err) | |
2430 | break; | |
2431 | ||
2432 | set->queue_depth >>= 1; | |
2433 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) { | |
2434 | err = -ENOMEM; | |
2435 | break; | |
2436 | } | |
2437 | } while (set->queue_depth); | |
2438 | ||
2439 | if (!set->queue_depth || err) { | |
2440 | pr_err("blk-mq: failed to allocate request map\n"); | |
2441 | return -ENOMEM; | |
2442 | } | |
2443 | ||
2444 | if (depth != set->queue_depth) | |
2445 | pr_info("blk-mq: reduced tag depth (%u -> %u)\n", | |
2446 | depth, set->queue_depth); | |
2447 | ||
2448 | return 0; | |
2449 | } | |
2450 | ||
a4391c64 JA |
2451 | /* |
2452 | * Alloc a tag set to be associated with one or more request queues. | |
2453 | * May fail with EINVAL for various error conditions. May adjust the | |
2454 | * requested depth down, if if it too large. In that case, the set | |
2455 | * value will be stored in set->queue_depth. | |
2456 | */ | |
24d2f903 CH |
2457 | int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set) |
2458 | { | |
da695ba2 CH |
2459 | int ret; |
2460 | ||
205fb5f5 BVA |
2461 | BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS); |
2462 | ||
24d2f903 CH |
2463 | if (!set->nr_hw_queues) |
2464 | return -EINVAL; | |
a4391c64 | 2465 | if (!set->queue_depth) |
24d2f903 CH |
2466 | return -EINVAL; |
2467 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) | |
2468 | return -EINVAL; | |
2469 | ||
7d7e0f90 | 2470 | if (!set->ops->queue_rq) |
24d2f903 CH |
2471 | return -EINVAL; |
2472 | ||
a4391c64 JA |
2473 | if (set->queue_depth > BLK_MQ_MAX_DEPTH) { |
2474 | pr_info("blk-mq: reduced tag depth to %u\n", | |
2475 | BLK_MQ_MAX_DEPTH); | |
2476 | set->queue_depth = BLK_MQ_MAX_DEPTH; | |
2477 | } | |
24d2f903 | 2478 | |
6637fadf SL |
2479 | /* |
2480 | * If a crashdump is active, then we are potentially in a very | |
2481 | * memory constrained environment. Limit us to 1 queue and | |
2482 | * 64 tags to prevent using too much memory. | |
2483 | */ | |
2484 | if (is_kdump_kernel()) { | |
2485 | set->nr_hw_queues = 1; | |
2486 | set->queue_depth = min(64U, set->queue_depth); | |
2487 | } | |
868f2f0b KB |
2488 | /* |
2489 | * There is no use for more h/w queues than cpus. | |
2490 | */ | |
2491 | if (set->nr_hw_queues > nr_cpu_ids) | |
2492 | set->nr_hw_queues = nr_cpu_ids; | |
6637fadf | 2493 | |
868f2f0b | 2494 | set->tags = kzalloc_node(nr_cpu_ids * sizeof(struct blk_mq_tags *), |
24d2f903 CH |
2495 | GFP_KERNEL, set->numa_node); |
2496 | if (!set->tags) | |
a5164405 | 2497 | return -ENOMEM; |
24d2f903 | 2498 | |
da695ba2 CH |
2499 | ret = -ENOMEM; |
2500 | set->mq_map = kzalloc_node(sizeof(*set->mq_map) * nr_cpu_ids, | |
2501 | GFP_KERNEL, set->numa_node); | |
bdd17e75 CH |
2502 | if (!set->mq_map) |
2503 | goto out_free_tags; | |
2504 | ||
da695ba2 CH |
2505 | if (set->ops->map_queues) |
2506 | ret = set->ops->map_queues(set); | |
2507 | else | |
2508 | ret = blk_mq_map_queues(set); | |
2509 | if (ret) | |
2510 | goto out_free_mq_map; | |
2511 | ||
2512 | ret = blk_mq_alloc_rq_maps(set); | |
2513 | if (ret) | |
bdd17e75 | 2514 | goto out_free_mq_map; |
24d2f903 | 2515 | |
0d2602ca JA |
2516 | mutex_init(&set->tag_list_lock); |
2517 | INIT_LIST_HEAD(&set->tag_list); | |
2518 | ||
24d2f903 | 2519 | return 0; |
bdd17e75 CH |
2520 | |
2521 | out_free_mq_map: | |
2522 | kfree(set->mq_map); | |
2523 | set->mq_map = NULL; | |
2524 | out_free_tags: | |
5676e7b6 RE |
2525 | kfree(set->tags); |
2526 | set->tags = NULL; | |
da695ba2 | 2527 | return ret; |
24d2f903 CH |
2528 | } |
2529 | EXPORT_SYMBOL(blk_mq_alloc_tag_set); | |
2530 | ||
2531 | void blk_mq_free_tag_set(struct blk_mq_tag_set *set) | |
2532 | { | |
2533 | int i; | |
2534 | ||
cc71a6f4 JA |
2535 | for (i = 0; i < nr_cpu_ids; i++) |
2536 | blk_mq_free_map_and_requests(set, i); | |
484b4061 | 2537 | |
bdd17e75 CH |
2538 | kfree(set->mq_map); |
2539 | set->mq_map = NULL; | |
2540 | ||
981bd189 | 2541 | kfree(set->tags); |
5676e7b6 | 2542 | set->tags = NULL; |
24d2f903 CH |
2543 | } |
2544 | EXPORT_SYMBOL(blk_mq_free_tag_set); | |
2545 | ||
e3a2b3f9 JA |
2546 | int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr) |
2547 | { | |
2548 | struct blk_mq_tag_set *set = q->tag_set; | |
2549 | struct blk_mq_hw_ctx *hctx; | |
2550 | int i, ret; | |
2551 | ||
bd166ef1 | 2552 | if (!set) |
e3a2b3f9 JA |
2553 | return -EINVAL; |
2554 | ||
2555 | ret = 0; | |
2556 | queue_for_each_hw_ctx(q, hctx, i) { | |
e9137d4b KB |
2557 | if (!hctx->tags) |
2558 | continue; | |
bd166ef1 JA |
2559 | /* |
2560 | * If we're using an MQ scheduler, just update the scheduler | |
2561 | * queue depth. This is similar to what the old code would do. | |
2562 | */ | |
2563 | if (!hctx->sched_tags) | |
2564 | ret = blk_mq_tag_update_depth(hctx->tags, | |
2565 | min(nr, set->queue_depth)); | |
2566 | else | |
2567 | ret = blk_mq_tag_update_depth(hctx->sched_tags, nr); | |
e3a2b3f9 JA |
2568 | if (ret) |
2569 | break; | |
2570 | } | |
2571 | ||
2572 | if (!ret) | |
2573 | q->nr_requests = nr; | |
2574 | ||
2575 | return ret; | |
2576 | } | |
2577 | ||
868f2f0b KB |
2578 | void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues) |
2579 | { | |
2580 | struct request_queue *q; | |
2581 | ||
2582 | if (nr_hw_queues > nr_cpu_ids) | |
2583 | nr_hw_queues = nr_cpu_ids; | |
2584 | if (nr_hw_queues < 1 || nr_hw_queues == set->nr_hw_queues) | |
2585 | return; | |
2586 | ||
2587 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
2588 | blk_mq_freeze_queue(q); | |
2589 | ||
2590 | set->nr_hw_queues = nr_hw_queues; | |
2591 | list_for_each_entry(q, &set->tag_list, tag_set_list) { | |
2592 | blk_mq_realloc_hw_ctxs(set, q); | |
2593 | ||
2594 | if (q->nr_hw_queues > 1) | |
2595 | blk_queue_make_request(q, blk_mq_make_request); | |
2596 | else | |
2597 | blk_queue_make_request(q, blk_sq_make_request); | |
2598 | ||
2599 | blk_mq_queue_reinit(q, cpu_online_mask); | |
2600 | } | |
2601 | ||
2602 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
2603 | blk_mq_unfreeze_queue(q); | |
2604 | } | |
2605 | EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues); | |
2606 | ||
64f1c21e JA |
2607 | static unsigned long blk_mq_poll_nsecs(struct request_queue *q, |
2608 | struct blk_mq_hw_ctx *hctx, | |
2609 | struct request *rq) | |
2610 | { | |
2611 | struct blk_rq_stat stat[2]; | |
2612 | unsigned long ret = 0; | |
2613 | ||
2614 | /* | |
2615 | * If stats collection isn't on, don't sleep but turn it on for | |
2616 | * future users | |
2617 | */ | |
2618 | if (!blk_stat_enable(q)) | |
2619 | return 0; | |
2620 | ||
2621 | /* | |
2622 | * We don't have to do this once per IO, should optimize this | |
2623 | * to just use the current window of stats until it changes | |
2624 | */ | |
2625 | memset(&stat, 0, sizeof(stat)); | |
2626 | blk_hctx_stat_get(hctx, stat); | |
2627 | ||
2628 | /* | |
2629 | * As an optimistic guess, use half of the mean service time | |
2630 | * for this type of request. We can (and should) make this smarter. | |
2631 | * For instance, if the completion latencies are tight, we can | |
2632 | * get closer than just half the mean. This is especially | |
2633 | * important on devices where the completion latencies are longer | |
2634 | * than ~10 usec. | |
2635 | */ | |
2636 | if (req_op(rq) == REQ_OP_READ && stat[BLK_STAT_READ].nr_samples) | |
2637 | ret = (stat[BLK_STAT_READ].mean + 1) / 2; | |
2638 | else if (req_op(rq) == REQ_OP_WRITE && stat[BLK_STAT_WRITE].nr_samples) | |
2639 | ret = (stat[BLK_STAT_WRITE].mean + 1) / 2; | |
2640 | ||
2641 | return ret; | |
2642 | } | |
2643 | ||
06426adf | 2644 | static bool blk_mq_poll_hybrid_sleep(struct request_queue *q, |
64f1c21e | 2645 | struct blk_mq_hw_ctx *hctx, |
06426adf JA |
2646 | struct request *rq) |
2647 | { | |
2648 | struct hrtimer_sleeper hs; | |
2649 | enum hrtimer_mode mode; | |
64f1c21e | 2650 | unsigned int nsecs; |
06426adf JA |
2651 | ktime_t kt; |
2652 | ||
64f1c21e JA |
2653 | if (test_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags)) |
2654 | return false; | |
2655 | ||
2656 | /* | |
2657 | * poll_nsec can be: | |
2658 | * | |
2659 | * -1: don't ever hybrid sleep | |
2660 | * 0: use half of prev avg | |
2661 | * >0: use this specific value | |
2662 | */ | |
2663 | if (q->poll_nsec == -1) | |
2664 | return false; | |
2665 | else if (q->poll_nsec > 0) | |
2666 | nsecs = q->poll_nsec; | |
2667 | else | |
2668 | nsecs = blk_mq_poll_nsecs(q, hctx, rq); | |
2669 | ||
2670 | if (!nsecs) | |
06426adf JA |
2671 | return false; |
2672 | ||
2673 | set_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags); | |
2674 | ||
2675 | /* | |
2676 | * This will be replaced with the stats tracking code, using | |
2677 | * 'avg_completion_time / 2' as the pre-sleep target. | |
2678 | */ | |
8b0e1953 | 2679 | kt = nsecs; |
06426adf JA |
2680 | |
2681 | mode = HRTIMER_MODE_REL; | |
2682 | hrtimer_init_on_stack(&hs.timer, CLOCK_MONOTONIC, mode); | |
2683 | hrtimer_set_expires(&hs.timer, kt); | |
2684 | ||
2685 | hrtimer_init_sleeper(&hs, current); | |
2686 | do { | |
2687 | if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags)) | |
2688 | break; | |
2689 | set_current_state(TASK_UNINTERRUPTIBLE); | |
2690 | hrtimer_start_expires(&hs.timer, mode); | |
2691 | if (hs.task) | |
2692 | io_schedule(); | |
2693 | hrtimer_cancel(&hs.timer); | |
2694 | mode = HRTIMER_MODE_ABS; | |
2695 | } while (hs.task && !signal_pending(current)); | |
2696 | ||
2697 | __set_current_state(TASK_RUNNING); | |
2698 | destroy_hrtimer_on_stack(&hs.timer); | |
2699 | return true; | |
2700 | } | |
2701 | ||
bbd7bb70 JA |
2702 | static bool __blk_mq_poll(struct blk_mq_hw_ctx *hctx, struct request *rq) |
2703 | { | |
2704 | struct request_queue *q = hctx->queue; | |
2705 | long state; | |
2706 | ||
06426adf JA |
2707 | /* |
2708 | * If we sleep, have the caller restart the poll loop to reset | |
2709 | * the state. Like for the other success return cases, the | |
2710 | * caller is responsible for checking if the IO completed. If | |
2711 | * the IO isn't complete, we'll get called again and will go | |
2712 | * straight to the busy poll loop. | |
2713 | */ | |
64f1c21e | 2714 | if (blk_mq_poll_hybrid_sleep(q, hctx, rq)) |
06426adf JA |
2715 | return true; |
2716 | ||
bbd7bb70 JA |
2717 | hctx->poll_considered++; |
2718 | ||
2719 | state = current->state; | |
2720 | while (!need_resched()) { | |
2721 | int ret; | |
2722 | ||
2723 | hctx->poll_invoked++; | |
2724 | ||
2725 | ret = q->mq_ops->poll(hctx, rq->tag); | |
2726 | if (ret > 0) { | |
2727 | hctx->poll_success++; | |
2728 | set_current_state(TASK_RUNNING); | |
2729 | return true; | |
2730 | } | |
2731 | ||
2732 | if (signal_pending_state(state, current)) | |
2733 | set_current_state(TASK_RUNNING); | |
2734 | ||
2735 | if (current->state == TASK_RUNNING) | |
2736 | return true; | |
2737 | if (ret < 0) | |
2738 | break; | |
2739 | cpu_relax(); | |
2740 | } | |
2741 | ||
2742 | return false; | |
2743 | } | |
2744 | ||
2745 | bool blk_mq_poll(struct request_queue *q, blk_qc_t cookie) | |
2746 | { | |
2747 | struct blk_mq_hw_ctx *hctx; | |
2748 | struct blk_plug *plug; | |
2749 | struct request *rq; | |
2750 | ||
2751 | if (!q->mq_ops || !q->mq_ops->poll || !blk_qc_t_valid(cookie) || | |
2752 | !test_bit(QUEUE_FLAG_POLL, &q->queue_flags)) | |
2753 | return false; | |
2754 | ||
2755 | plug = current->plug; | |
2756 | if (plug) | |
2757 | blk_flush_plug_list(plug, false); | |
2758 | ||
2759 | hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)]; | |
bd166ef1 JA |
2760 | if (!blk_qc_t_is_internal(cookie)) |
2761 | rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie)); | |
2762 | else | |
2763 | rq = blk_mq_tag_to_rq(hctx->sched_tags, blk_qc_t_to_tag(cookie)); | |
bbd7bb70 JA |
2764 | |
2765 | return __blk_mq_poll(hctx, rq); | |
2766 | } | |
2767 | EXPORT_SYMBOL_GPL(blk_mq_poll); | |
2768 | ||
676141e4 JA |
2769 | void blk_mq_disable_hotplug(void) |
2770 | { | |
2771 | mutex_lock(&all_q_mutex); | |
2772 | } | |
2773 | ||
2774 | void blk_mq_enable_hotplug(void) | |
2775 | { | |
2776 | mutex_unlock(&all_q_mutex); | |
2777 | } | |
2778 | ||
320ae51f JA |
2779 | static int __init blk_mq_init(void) |
2780 | { | |
9467f859 TG |
2781 | cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL, |
2782 | blk_mq_hctx_notify_dead); | |
320ae51f | 2783 | |
65d5291e SAS |
2784 | cpuhp_setup_state_nocalls(CPUHP_BLK_MQ_PREPARE, "block/mq:prepare", |
2785 | blk_mq_queue_reinit_prepare, | |
2786 | blk_mq_queue_reinit_dead); | |
320ae51f JA |
2787 | return 0; |
2788 | } | |
2789 | subsys_initcall(blk_mq_init); |