]>
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 | |
7e79dadc | 1173 | blk_mq_stop_hw_queue(hctx); |
b657d7e6 CH |
1174 | kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx), |
1175 | &hctx->delay_work, msecs_to_jiffies(msecs)); | |
70f4db63 CH |
1176 | } |
1177 | EXPORT_SYMBOL(blk_mq_delay_queue); | |
1178 | ||
cfd0c552 | 1179 | static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx, |
cfd0c552 ML |
1180 | struct request *rq, |
1181 | bool at_head) | |
320ae51f | 1182 | { |
e57690fe JA |
1183 | struct blk_mq_ctx *ctx = rq->mq_ctx; |
1184 | ||
01b983c9 JA |
1185 | trace_block_rq_insert(hctx->queue, rq); |
1186 | ||
72a0a36e CH |
1187 | if (at_head) |
1188 | list_add(&rq->queuelist, &ctx->rq_list); | |
1189 | else | |
1190 | list_add_tail(&rq->queuelist, &ctx->rq_list); | |
cfd0c552 | 1191 | } |
4bb659b1 | 1192 | |
2c3ad667 JA |
1193 | void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq, |
1194 | bool at_head) | |
cfd0c552 ML |
1195 | { |
1196 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
1197 | ||
e57690fe | 1198 | __blk_mq_insert_req_list(hctx, rq, at_head); |
320ae51f | 1199 | blk_mq_hctx_mark_pending(hctx, ctx); |
320ae51f JA |
1200 | } |
1201 | ||
bd166ef1 JA |
1202 | void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx, |
1203 | struct list_head *list) | |
320ae51f JA |
1204 | |
1205 | { | |
320ae51f JA |
1206 | /* |
1207 | * preemption doesn't flush plug list, so it's possible ctx->cpu is | |
1208 | * offline now | |
1209 | */ | |
1210 | spin_lock(&ctx->lock); | |
1211 | while (!list_empty(list)) { | |
1212 | struct request *rq; | |
1213 | ||
1214 | rq = list_first_entry(list, struct request, queuelist); | |
e57690fe | 1215 | BUG_ON(rq->mq_ctx != ctx); |
320ae51f | 1216 | list_del_init(&rq->queuelist); |
e57690fe | 1217 | __blk_mq_insert_req_list(hctx, rq, false); |
320ae51f | 1218 | } |
cfd0c552 | 1219 | blk_mq_hctx_mark_pending(hctx, ctx); |
320ae51f | 1220 | spin_unlock(&ctx->lock); |
320ae51f JA |
1221 | } |
1222 | ||
1223 | static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b) | |
1224 | { | |
1225 | struct request *rqa = container_of(a, struct request, queuelist); | |
1226 | struct request *rqb = container_of(b, struct request, queuelist); | |
1227 | ||
1228 | return !(rqa->mq_ctx < rqb->mq_ctx || | |
1229 | (rqa->mq_ctx == rqb->mq_ctx && | |
1230 | blk_rq_pos(rqa) < blk_rq_pos(rqb))); | |
1231 | } | |
1232 | ||
1233 | void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule) | |
1234 | { | |
1235 | struct blk_mq_ctx *this_ctx; | |
1236 | struct request_queue *this_q; | |
1237 | struct request *rq; | |
1238 | LIST_HEAD(list); | |
1239 | LIST_HEAD(ctx_list); | |
1240 | unsigned int depth; | |
1241 | ||
1242 | list_splice_init(&plug->mq_list, &list); | |
1243 | ||
1244 | list_sort(NULL, &list, plug_ctx_cmp); | |
1245 | ||
1246 | this_q = NULL; | |
1247 | this_ctx = NULL; | |
1248 | depth = 0; | |
1249 | ||
1250 | while (!list_empty(&list)) { | |
1251 | rq = list_entry_rq(list.next); | |
1252 | list_del_init(&rq->queuelist); | |
1253 | BUG_ON(!rq->q); | |
1254 | if (rq->mq_ctx != this_ctx) { | |
1255 | if (this_ctx) { | |
bd166ef1 JA |
1256 | trace_block_unplug(this_q, depth, from_schedule); |
1257 | blk_mq_sched_insert_requests(this_q, this_ctx, | |
1258 | &ctx_list, | |
1259 | from_schedule); | |
320ae51f JA |
1260 | } |
1261 | ||
1262 | this_ctx = rq->mq_ctx; | |
1263 | this_q = rq->q; | |
1264 | depth = 0; | |
1265 | } | |
1266 | ||
1267 | depth++; | |
1268 | list_add_tail(&rq->queuelist, &ctx_list); | |
1269 | } | |
1270 | ||
1271 | /* | |
1272 | * If 'this_ctx' is set, we know we have entries to complete | |
1273 | * on 'ctx_list'. Do those. | |
1274 | */ | |
1275 | if (this_ctx) { | |
bd166ef1 JA |
1276 | trace_block_unplug(this_q, depth, from_schedule); |
1277 | blk_mq_sched_insert_requests(this_q, this_ctx, &ctx_list, | |
1278 | from_schedule); | |
320ae51f JA |
1279 | } |
1280 | } | |
1281 | ||
1282 | static void blk_mq_bio_to_request(struct request *rq, struct bio *bio) | |
1283 | { | |
1284 | init_request_from_bio(rq, bio); | |
4b570521 | 1285 | |
6e85eaf3 | 1286 | blk_account_io_start(rq, true); |
320ae51f JA |
1287 | } |
1288 | ||
274a5843 JA |
1289 | static inline bool hctx_allow_merges(struct blk_mq_hw_ctx *hctx) |
1290 | { | |
1291 | return (hctx->flags & BLK_MQ_F_SHOULD_MERGE) && | |
1292 | !blk_queue_nomerges(hctx->queue); | |
1293 | } | |
1294 | ||
07068d5b JA |
1295 | static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx *hctx, |
1296 | struct blk_mq_ctx *ctx, | |
1297 | struct request *rq, struct bio *bio) | |
320ae51f | 1298 | { |
e18378a6 | 1299 | if (!hctx_allow_merges(hctx) || !bio_mergeable(bio)) { |
07068d5b JA |
1300 | blk_mq_bio_to_request(rq, bio); |
1301 | spin_lock(&ctx->lock); | |
1302 | insert_rq: | |
1303 | __blk_mq_insert_request(hctx, rq, false); | |
1304 | spin_unlock(&ctx->lock); | |
1305 | return false; | |
1306 | } else { | |
274a5843 JA |
1307 | struct request_queue *q = hctx->queue; |
1308 | ||
07068d5b JA |
1309 | spin_lock(&ctx->lock); |
1310 | if (!blk_mq_attempt_merge(q, ctx, bio)) { | |
1311 | blk_mq_bio_to_request(rq, bio); | |
1312 | goto insert_rq; | |
1313 | } | |
320ae51f | 1314 | |
07068d5b | 1315 | spin_unlock(&ctx->lock); |
bd166ef1 | 1316 | __blk_mq_finish_request(hctx, ctx, rq); |
07068d5b | 1317 | return true; |
14ec77f3 | 1318 | } |
07068d5b | 1319 | } |
14ec77f3 | 1320 | |
fd2d3326 JA |
1321 | static blk_qc_t request_to_qc_t(struct blk_mq_hw_ctx *hctx, struct request *rq) |
1322 | { | |
bd166ef1 JA |
1323 | if (rq->tag != -1) |
1324 | return blk_tag_to_qc_t(rq->tag, hctx->queue_num, false); | |
1325 | ||
1326 | return blk_tag_to_qc_t(rq->internal_tag, hctx->queue_num, true); | |
fd2d3326 JA |
1327 | } |
1328 | ||
066a4a73 | 1329 | static void blk_mq_try_issue_directly(struct request *rq, blk_qc_t *cookie) |
f984df1f | 1330 | { |
f984df1f | 1331 | struct request_queue *q = rq->q; |
f984df1f SL |
1332 | struct blk_mq_queue_data bd = { |
1333 | .rq = rq, | |
1334 | .list = NULL, | |
1335 | .last = 1 | |
1336 | }; | |
bd166ef1 JA |
1337 | struct blk_mq_hw_ctx *hctx; |
1338 | blk_qc_t new_cookie; | |
1339 | int ret; | |
f984df1f | 1340 | |
bd166ef1 | 1341 | if (q->elevator) |
2253efc8 BVA |
1342 | goto insert; |
1343 | ||
bd166ef1 JA |
1344 | if (!blk_mq_get_driver_tag(rq, &hctx, false)) |
1345 | goto insert; | |
1346 | ||
1347 | new_cookie = request_to_qc_t(hctx, rq); | |
1348 | ||
f984df1f SL |
1349 | /* |
1350 | * For OK queue, we are done. For error, kill it. Any other | |
1351 | * error (busy), just add it to our list as we previously | |
1352 | * would have done | |
1353 | */ | |
1354 | ret = q->mq_ops->queue_rq(hctx, &bd); | |
7b371636 JA |
1355 | if (ret == BLK_MQ_RQ_QUEUE_OK) { |
1356 | *cookie = new_cookie; | |
2253efc8 | 1357 | return; |
7b371636 | 1358 | } |
f984df1f | 1359 | |
7b371636 JA |
1360 | __blk_mq_requeue_request(rq); |
1361 | ||
1362 | if (ret == BLK_MQ_RQ_QUEUE_ERROR) { | |
1363 | *cookie = BLK_QC_T_NONE; | |
1364 | rq->errors = -EIO; | |
1365 | blk_mq_end_request(rq, rq->errors); | |
2253efc8 | 1366 | return; |
f984df1f | 1367 | } |
7b371636 | 1368 | |
2253efc8 | 1369 | insert: |
bd166ef1 | 1370 | blk_mq_sched_insert_request(rq, false, true, true); |
f984df1f SL |
1371 | } |
1372 | ||
07068d5b JA |
1373 | /* |
1374 | * Multiple hardware queue variant. This will not use per-process plugs, | |
1375 | * but will attempt to bypass the hctx queueing if we can go straight to | |
1376 | * hardware for SYNC IO. | |
1377 | */ | |
dece1635 | 1378 | static blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio) |
07068d5b | 1379 | { |
ef295ecf | 1380 | const int is_sync = op_is_sync(bio->bi_opf); |
1eff9d32 | 1381 | const int is_flush_fua = bio->bi_opf & (REQ_PREFLUSH | REQ_FUA); |
2552e3f8 | 1382 | struct blk_mq_alloc_data data; |
07068d5b | 1383 | struct request *rq; |
6a83e74d | 1384 | unsigned int request_count = 0, srcu_idx; |
f984df1f | 1385 | struct blk_plug *plug; |
5b3f341f | 1386 | struct request *same_queue_rq = NULL; |
7b371636 | 1387 | blk_qc_t cookie; |
87760e5e | 1388 | unsigned int wb_acct; |
07068d5b JA |
1389 | |
1390 | blk_queue_bounce(q, &bio); | |
1391 | ||
1392 | if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) { | |
4246a0b6 | 1393 | bio_io_error(bio); |
dece1635 | 1394 | return BLK_QC_T_NONE; |
07068d5b JA |
1395 | } |
1396 | ||
54efd50b KO |
1397 | blk_queue_split(q, &bio, q->bio_split); |
1398 | ||
87c279e6 OS |
1399 | if (!is_flush_fua && !blk_queue_nomerges(q) && |
1400 | blk_attempt_plug_merge(q, bio, &request_count, &same_queue_rq)) | |
1401 | return BLK_QC_T_NONE; | |
f984df1f | 1402 | |
bd166ef1 JA |
1403 | if (blk_mq_sched_bio_merge(q, bio)) |
1404 | return BLK_QC_T_NONE; | |
1405 | ||
87760e5e JA |
1406 | wb_acct = wbt_wait(q->rq_wb, bio, NULL); |
1407 | ||
bd166ef1 JA |
1408 | trace_block_getrq(q, bio, bio->bi_opf); |
1409 | ||
1410 | rq = blk_mq_sched_get_request(q, bio, bio->bi_opf, &data); | |
87760e5e JA |
1411 | if (unlikely(!rq)) { |
1412 | __wbt_done(q->rq_wb, wb_acct); | |
dece1635 | 1413 | return BLK_QC_T_NONE; |
87760e5e JA |
1414 | } |
1415 | ||
1416 | wbt_track(&rq->issue_stat, wb_acct); | |
07068d5b | 1417 | |
fd2d3326 | 1418 | cookie = request_to_qc_t(data.hctx, rq); |
07068d5b JA |
1419 | |
1420 | if (unlikely(is_flush_fua)) { | |
1421 | blk_mq_bio_to_request(rq, bio); | |
bd166ef1 | 1422 | blk_mq_get_driver_tag(rq, NULL, true); |
07068d5b JA |
1423 | blk_insert_flush(rq); |
1424 | goto run_queue; | |
1425 | } | |
1426 | ||
f984df1f | 1427 | plug = current->plug; |
e167dfb5 JA |
1428 | /* |
1429 | * If the driver supports defer issued based on 'last', then | |
1430 | * queue it up like normal since we can potentially save some | |
1431 | * CPU this way. | |
1432 | */ | |
f984df1f SL |
1433 | if (((plug && !blk_queue_nomerges(q)) || is_sync) && |
1434 | !(data.hctx->flags & BLK_MQ_F_DEFER_ISSUE)) { | |
1435 | struct request *old_rq = NULL; | |
07068d5b JA |
1436 | |
1437 | blk_mq_bio_to_request(rq, bio); | |
07068d5b JA |
1438 | |
1439 | /* | |
6a83e74d | 1440 | * We do limited plugging. If the bio can be merged, do that. |
f984df1f SL |
1441 | * Otherwise the existing request in the plug list will be |
1442 | * issued. So the plug list will have one request at most | |
07068d5b | 1443 | */ |
f984df1f | 1444 | if (plug) { |
5b3f341f SL |
1445 | /* |
1446 | * The plug list might get flushed before this. If that | |
b094f89c JA |
1447 | * happens, same_queue_rq is invalid and plug list is |
1448 | * empty | |
1449 | */ | |
5b3f341f SL |
1450 | if (same_queue_rq && !list_empty(&plug->mq_list)) { |
1451 | old_rq = same_queue_rq; | |
f984df1f | 1452 | list_del_init(&old_rq->queuelist); |
07068d5b | 1453 | } |
f984df1f SL |
1454 | list_add_tail(&rq->queuelist, &plug->mq_list); |
1455 | } else /* is_sync */ | |
1456 | old_rq = rq; | |
1457 | blk_mq_put_ctx(data.ctx); | |
1458 | if (!old_rq) | |
7b371636 | 1459 | goto done; |
6a83e74d BVA |
1460 | |
1461 | if (!(data.hctx->flags & BLK_MQ_F_BLOCKING)) { | |
1462 | rcu_read_lock(); | |
066a4a73 | 1463 | blk_mq_try_issue_directly(old_rq, &cookie); |
6a83e74d BVA |
1464 | rcu_read_unlock(); |
1465 | } else { | |
1466 | srcu_idx = srcu_read_lock(&data.hctx->queue_rq_srcu); | |
066a4a73 | 1467 | blk_mq_try_issue_directly(old_rq, &cookie); |
6a83e74d BVA |
1468 | srcu_read_unlock(&data.hctx->queue_rq_srcu, srcu_idx); |
1469 | } | |
7b371636 | 1470 | goto done; |
07068d5b JA |
1471 | } |
1472 | ||
bd166ef1 JA |
1473 | if (q->elevator) { |
1474 | blk_mq_put_ctx(data.ctx); | |
1475 | blk_mq_bio_to_request(rq, bio); | |
1476 | blk_mq_sched_insert_request(rq, false, true, true); | |
1477 | goto done; | |
1478 | } | |
07068d5b JA |
1479 | if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) { |
1480 | /* | |
1481 | * For a SYNC request, send it to the hardware immediately. For | |
1482 | * an ASYNC request, just ensure that we run it later on. The | |
1483 | * latter allows for merging opportunities and more efficient | |
1484 | * dispatching. | |
1485 | */ | |
1486 | run_queue: | |
1487 | blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua); | |
1488 | } | |
07068d5b | 1489 | blk_mq_put_ctx(data.ctx); |
7b371636 JA |
1490 | done: |
1491 | return cookie; | |
07068d5b JA |
1492 | } |
1493 | ||
1494 | /* | |
1495 | * Single hardware queue variant. This will attempt to use any per-process | |
1496 | * plug for merging and IO deferral. | |
1497 | */ | |
dece1635 | 1498 | static blk_qc_t blk_sq_make_request(struct request_queue *q, struct bio *bio) |
07068d5b | 1499 | { |
ef295ecf | 1500 | const int is_sync = op_is_sync(bio->bi_opf); |
1eff9d32 | 1501 | const int is_flush_fua = bio->bi_opf & (REQ_PREFLUSH | REQ_FUA); |
e6c4438b JM |
1502 | struct blk_plug *plug; |
1503 | unsigned int request_count = 0; | |
2552e3f8 | 1504 | struct blk_mq_alloc_data data; |
07068d5b | 1505 | struct request *rq; |
7b371636 | 1506 | blk_qc_t cookie; |
87760e5e | 1507 | unsigned int wb_acct; |
07068d5b | 1508 | |
07068d5b JA |
1509 | blk_queue_bounce(q, &bio); |
1510 | ||
1511 | if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) { | |
4246a0b6 | 1512 | bio_io_error(bio); |
dece1635 | 1513 | return BLK_QC_T_NONE; |
07068d5b JA |
1514 | } |
1515 | ||
54efd50b KO |
1516 | blk_queue_split(q, &bio, q->bio_split); |
1517 | ||
87c279e6 OS |
1518 | if (!is_flush_fua && !blk_queue_nomerges(q)) { |
1519 | if (blk_attempt_plug_merge(q, bio, &request_count, NULL)) | |
1520 | return BLK_QC_T_NONE; | |
1521 | } else | |
1522 | request_count = blk_plug_queued_count(q); | |
07068d5b | 1523 | |
bd166ef1 JA |
1524 | if (blk_mq_sched_bio_merge(q, bio)) |
1525 | return BLK_QC_T_NONE; | |
1526 | ||
87760e5e JA |
1527 | wb_acct = wbt_wait(q->rq_wb, bio, NULL); |
1528 | ||
bd166ef1 JA |
1529 | trace_block_getrq(q, bio, bio->bi_opf); |
1530 | ||
1531 | rq = blk_mq_sched_get_request(q, bio, bio->bi_opf, &data); | |
87760e5e JA |
1532 | if (unlikely(!rq)) { |
1533 | __wbt_done(q->rq_wb, wb_acct); | |
dece1635 | 1534 | return BLK_QC_T_NONE; |
87760e5e JA |
1535 | } |
1536 | ||
1537 | wbt_track(&rq->issue_stat, wb_acct); | |
320ae51f | 1538 | |
fd2d3326 | 1539 | cookie = request_to_qc_t(data.hctx, rq); |
320ae51f JA |
1540 | |
1541 | if (unlikely(is_flush_fua)) { | |
1542 | blk_mq_bio_to_request(rq, bio); | |
bd166ef1 | 1543 | blk_mq_get_driver_tag(rq, NULL, true); |
320ae51f JA |
1544 | blk_insert_flush(rq); |
1545 | goto run_queue; | |
1546 | } | |
1547 | ||
1548 | /* | |
1549 | * A task plug currently exists. Since this is completely lockless, | |
1550 | * utilize that to temporarily store requests until the task is | |
1551 | * either done or scheduled away. | |
1552 | */ | |
e6c4438b JM |
1553 | plug = current->plug; |
1554 | if (plug) { | |
600271d9 SL |
1555 | struct request *last = NULL; |
1556 | ||
e6c4438b | 1557 | blk_mq_bio_to_request(rq, bio); |
0a6219a9 ML |
1558 | |
1559 | /* | |
1560 | * @request_count may become stale because of schedule | |
1561 | * out, so check the list again. | |
1562 | */ | |
1563 | if (list_empty(&plug->mq_list)) | |
1564 | request_count = 0; | |
676d0607 | 1565 | if (!request_count) |
e6c4438b | 1566 | trace_block_plug(q); |
600271d9 SL |
1567 | else |
1568 | last = list_entry_rq(plug->mq_list.prev); | |
b094f89c JA |
1569 | |
1570 | blk_mq_put_ctx(data.ctx); | |
1571 | ||
600271d9 SL |
1572 | if (request_count >= BLK_MAX_REQUEST_COUNT || (last && |
1573 | blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) { | |
e6c4438b JM |
1574 | blk_flush_plug_list(plug, false); |
1575 | trace_block_plug(q); | |
320ae51f | 1576 | } |
b094f89c | 1577 | |
e6c4438b | 1578 | list_add_tail(&rq->queuelist, &plug->mq_list); |
7b371636 | 1579 | return cookie; |
320ae51f JA |
1580 | } |
1581 | ||
bd166ef1 JA |
1582 | if (q->elevator) { |
1583 | blk_mq_put_ctx(data.ctx); | |
1584 | blk_mq_bio_to_request(rq, bio); | |
1585 | blk_mq_sched_insert_request(rq, false, true, true); | |
1586 | goto done; | |
1587 | } | |
07068d5b JA |
1588 | if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) { |
1589 | /* | |
1590 | * For a SYNC request, send it to the hardware immediately. For | |
1591 | * an ASYNC request, just ensure that we run it later on. The | |
1592 | * latter allows for merging opportunities and more efficient | |
1593 | * dispatching. | |
1594 | */ | |
1595 | run_queue: | |
1596 | blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua); | |
320ae51f JA |
1597 | } |
1598 | ||
07068d5b | 1599 | blk_mq_put_ctx(data.ctx); |
bd166ef1 | 1600 | done: |
7b371636 | 1601 | return cookie; |
320ae51f JA |
1602 | } |
1603 | ||
cc71a6f4 JA |
1604 | void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags, |
1605 | unsigned int hctx_idx) | |
95363efd | 1606 | { |
e9b267d9 | 1607 | struct page *page; |
320ae51f | 1608 | |
24d2f903 | 1609 | if (tags->rqs && set->ops->exit_request) { |
e9b267d9 | 1610 | int i; |
320ae51f | 1611 | |
24d2f903 | 1612 | for (i = 0; i < tags->nr_tags; i++) { |
2af8cbe3 JA |
1613 | struct request *rq = tags->static_rqs[i]; |
1614 | ||
1615 | if (!rq) | |
e9b267d9 | 1616 | continue; |
2af8cbe3 | 1617 | set->ops->exit_request(set->driver_data, rq, |
24d2f903 | 1618 | hctx_idx, i); |
2af8cbe3 | 1619 | tags->static_rqs[i] = NULL; |
e9b267d9 | 1620 | } |
320ae51f | 1621 | } |
320ae51f | 1622 | |
24d2f903 CH |
1623 | while (!list_empty(&tags->page_list)) { |
1624 | page = list_first_entry(&tags->page_list, struct page, lru); | |
6753471c | 1625 | list_del_init(&page->lru); |
f75782e4 CM |
1626 | /* |
1627 | * Remove kmemleak object previously allocated in | |
1628 | * blk_mq_init_rq_map(). | |
1629 | */ | |
1630 | kmemleak_free(page_address(page)); | |
320ae51f JA |
1631 | __free_pages(page, page->private); |
1632 | } | |
cc71a6f4 | 1633 | } |
320ae51f | 1634 | |
cc71a6f4 JA |
1635 | void blk_mq_free_rq_map(struct blk_mq_tags *tags) |
1636 | { | |
24d2f903 | 1637 | kfree(tags->rqs); |
cc71a6f4 | 1638 | tags->rqs = NULL; |
2af8cbe3 JA |
1639 | kfree(tags->static_rqs); |
1640 | tags->static_rqs = NULL; | |
320ae51f | 1641 | |
24d2f903 | 1642 | blk_mq_free_tags(tags); |
320ae51f JA |
1643 | } |
1644 | ||
cc71a6f4 JA |
1645 | struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, |
1646 | unsigned int hctx_idx, | |
1647 | unsigned int nr_tags, | |
1648 | unsigned int reserved_tags) | |
320ae51f | 1649 | { |
24d2f903 | 1650 | struct blk_mq_tags *tags; |
320ae51f | 1651 | |
cc71a6f4 | 1652 | tags = blk_mq_init_tags(nr_tags, reserved_tags, |
24391c0d SL |
1653 | set->numa_node, |
1654 | BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags)); | |
24d2f903 CH |
1655 | if (!tags) |
1656 | return NULL; | |
320ae51f | 1657 | |
cc71a6f4 | 1658 | tags->rqs = kzalloc_node(nr_tags * sizeof(struct request *), |
36e1f3d1 | 1659 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, |
a5164405 | 1660 | set->numa_node); |
24d2f903 CH |
1661 | if (!tags->rqs) { |
1662 | blk_mq_free_tags(tags); | |
1663 | return NULL; | |
1664 | } | |
320ae51f | 1665 | |
2af8cbe3 JA |
1666 | tags->static_rqs = kzalloc_node(nr_tags * sizeof(struct request *), |
1667 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, | |
1668 | set->numa_node); | |
1669 | if (!tags->static_rqs) { | |
1670 | kfree(tags->rqs); | |
1671 | blk_mq_free_tags(tags); | |
1672 | return NULL; | |
1673 | } | |
1674 | ||
cc71a6f4 JA |
1675 | return tags; |
1676 | } | |
1677 | ||
1678 | static size_t order_to_size(unsigned int order) | |
1679 | { | |
1680 | return (size_t)PAGE_SIZE << order; | |
1681 | } | |
1682 | ||
1683 | int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags, | |
1684 | unsigned int hctx_idx, unsigned int depth) | |
1685 | { | |
1686 | unsigned int i, j, entries_per_page, max_order = 4; | |
1687 | size_t rq_size, left; | |
1688 | ||
1689 | INIT_LIST_HEAD(&tags->page_list); | |
1690 | ||
320ae51f JA |
1691 | /* |
1692 | * rq_size is the size of the request plus driver payload, rounded | |
1693 | * to the cacheline size | |
1694 | */ | |
24d2f903 | 1695 | rq_size = round_up(sizeof(struct request) + set->cmd_size, |
320ae51f | 1696 | cache_line_size()); |
cc71a6f4 | 1697 | left = rq_size * depth; |
320ae51f | 1698 | |
cc71a6f4 | 1699 | for (i = 0; i < depth; ) { |
320ae51f JA |
1700 | int this_order = max_order; |
1701 | struct page *page; | |
1702 | int to_do; | |
1703 | void *p; | |
1704 | ||
b3a834b1 | 1705 | while (this_order && left < order_to_size(this_order - 1)) |
320ae51f JA |
1706 | this_order--; |
1707 | ||
1708 | do { | |
a5164405 | 1709 | page = alloc_pages_node(set->numa_node, |
36e1f3d1 | 1710 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO, |
a5164405 | 1711 | this_order); |
320ae51f JA |
1712 | if (page) |
1713 | break; | |
1714 | if (!this_order--) | |
1715 | break; | |
1716 | if (order_to_size(this_order) < rq_size) | |
1717 | break; | |
1718 | } while (1); | |
1719 | ||
1720 | if (!page) | |
24d2f903 | 1721 | goto fail; |
320ae51f JA |
1722 | |
1723 | page->private = this_order; | |
24d2f903 | 1724 | list_add_tail(&page->lru, &tags->page_list); |
320ae51f JA |
1725 | |
1726 | p = page_address(page); | |
f75782e4 CM |
1727 | /* |
1728 | * Allow kmemleak to scan these pages as they contain pointers | |
1729 | * to additional allocations like via ops->init_request(). | |
1730 | */ | |
36e1f3d1 | 1731 | kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO); |
320ae51f | 1732 | entries_per_page = order_to_size(this_order) / rq_size; |
cc71a6f4 | 1733 | to_do = min(entries_per_page, depth - i); |
320ae51f JA |
1734 | left -= to_do * rq_size; |
1735 | for (j = 0; j < to_do; j++) { | |
2af8cbe3 JA |
1736 | struct request *rq = p; |
1737 | ||
1738 | tags->static_rqs[i] = rq; | |
24d2f903 CH |
1739 | if (set->ops->init_request) { |
1740 | if (set->ops->init_request(set->driver_data, | |
2af8cbe3 | 1741 | rq, hctx_idx, i, |
a5164405 | 1742 | set->numa_node)) { |
2af8cbe3 | 1743 | tags->static_rqs[i] = NULL; |
24d2f903 | 1744 | goto fail; |
a5164405 | 1745 | } |
e9b267d9 CH |
1746 | } |
1747 | ||
320ae51f JA |
1748 | p += rq_size; |
1749 | i++; | |
1750 | } | |
1751 | } | |
cc71a6f4 | 1752 | return 0; |
320ae51f | 1753 | |
24d2f903 | 1754 | fail: |
cc71a6f4 JA |
1755 | blk_mq_free_rqs(set, tags, hctx_idx); |
1756 | return -ENOMEM; | |
320ae51f JA |
1757 | } |
1758 | ||
e57690fe JA |
1759 | /* |
1760 | * 'cpu' is going away. splice any existing rq_list entries from this | |
1761 | * software queue to the hw queue dispatch list, and ensure that it | |
1762 | * gets run. | |
1763 | */ | |
9467f859 | 1764 | static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node) |
484b4061 | 1765 | { |
9467f859 | 1766 | struct blk_mq_hw_ctx *hctx; |
484b4061 JA |
1767 | struct blk_mq_ctx *ctx; |
1768 | LIST_HEAD(tmp); | |
1769 | ||
9467f859 | 1770 | hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead); |
e57690fe | 1771 | ctx = __blk_mq_get_ctx(hctx->queue, cpu); |
484b4061 JA |
1772 | |
1773 | spin_lock(&ctx->lock); | |
1774 | if (!list_empty(&ctx->rq_list)) { | |
1775 | list_splice_init(&ctx->rq_list, &tmp); | |
1776 | blk_mq_hctx_clear_pending(hctx, ctx); | |
1777 | } | |
1778 | spin_unlock(&ctx->lock); | |
1779 | ||
1780 | if (list_empty(&tmp)) | |
9467f859 | 1781 | return 0; |
484b4061 | 1782 | |
e57690fe JA |
1783 | spin_lock(&hctx->lock); |
1784 | list_splice_tail_init(&tmp, &hctx->dispatch); | |
1785 | spin_unlock(&hctx->lock); | |
484b4061 JA |
1786 | |
1787 | blk_mq_run_hw_queue(hctx, true); | |
9467f859 | 1788 | return 0; |
484b4061 JA |
1789 | } |
1790 | ||
9467f859 | 1791 | static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx) |
484b4061 | 1792 | { |
9467f859 TG |
1793 | cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD, |
1794 | &hctx->cpuhp_dead); | |
484b4061 JA |
1795 | } |
1796 | ||
c3b4afca | 1797 | /* hctx->ctxs will be freed in queue's release handler */ |
08e98fc6 ML |
1798 | static void blk_mq_exit_hctx(struct request_queue *q, |
1799 | struct blk_mq_tag_set *set, | |
1800 | struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx) | |
1801 | { | |
f70ced09 ML |
1802 | unsigned flush_start_tag = set->queue_depth; |
1803 | ||
08e98fc6 ML |
1804 | blk_mq_tag_idle(hctx); |
1805 | ||
f70ced09 ML |
1806 | if (set->ops->exit_request) |
1807 | set->ops->exit_request(set->driver_data, | |
1808 | hctx->fq->flush_rq, hctx_idx, | |
1809 | flush_start_tag + hctx_idx); | |
1810 | ||
08e98fc6 ML |
1811 | if (set->ops->exit_hctx) |
1812 | set->ops->exit_hctx(hctx, hctx_idx); | |
1813 | ||
6a83e74d BVA |
1814 | if (hctx->flags & BLK_MQ_F_BLOCKING) |
1815 | cleanup_srcu_struct(&hctx->queue_rq_srcu); | |
1816 | ||
9467f859 | 1817 | blk_mq_remove_cpuhp(hctx); |
f70ced09 | 1818 | blk_free_flush_queue(hctx->fq); |
88459642 | 1819 | sbitmap_free(&hctx->ctx_map); |
08e98fc6 ML |
1820 | } |
1821 | ||
624dbe47 ML |
1822 | static void blk_mq_exit_hw_queues(struct request_queue *q, |
1823 | struct blk_mq_tag_set *set, int nr_queue) | |
1824 | { | |
1825 | struct blk_mq_hw_ctx *hctx; | |
1826 | unsigned int i; | |
1827 | ||
1828 | queue_for_each_hw_ctx(q, hctx, i) { | |
1829 | if (i == nr_queue) | |
1830 | break; | |
08e98fc6 | 1831 | blk_mq_exit_hctx(q, set, hctx, i); |
624dbe47 | 1832 | } |
624dbe47 ML |
1833 | } |
1834 | ||
1835 | static void blk_mq_free_hw_queues(struct request_queue *q, | |
1836 | struct blk_mq_tag_set *set) | |
1837 | { | |
1838 | struct blk_mq_hw_ctx *hctx; | |
1839 | unsigned int i; | |
1840 | ||
e09aae7e | 1841 | queue_for_each_hw_ctx(q, hctx, i) |
624dbe47 | 1842 | free_cpumask_var(hctx->cpumask); |
624dbe47 ML |
1843 | } |
1844 | ||
08e98fc6 ML |
1845 | static int blk_mq_init_hctx(struct request_queue *q, |
1846 | struct blk_mq_tag_set *set, | |
1847 | struct blk_mq_hw_ctx *hctx, unsigned hctx_idx) | |
320ae51f | 1848 | { |
08e98fc6 | 1849 | int node; |
f70ced09 | 1850 | unsigned flush_start_tag = set->queue_depth; |
08e98fc6 ML |
1851 | |
1852 | node = hctx->numa_node; | |
1853 | if (node == NUMA_NO_NODE) | |
1854 | node = hctx->numa_node = set->numa_node; | |
1855 | ||
27489a3c | 1856 | INIT_WORK(&hctx->run_work, blk_mq_run_work_fn); |
08e98fc6 ML |
1857 | INIT_DELAYED_WORK(&hctx->delay_work, blk_mq_delay_work_fn); |
1858 | spin_lock_init(&hctx->lock); | |
1859 | INIT_LIST_HEAD(&hctx->dispatch); | |
1860 | hctx->queue = q; | |
1861 | hctx->queue_num = hctx_idx; | |
2404e607 | 1862 | hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED; |
08e98fc6 | 1863 | |
9467f859 | 1864 | cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead); |
08e98fc6 ML |
1865 | |
1866 | hctx->tags = set->tags[hctx_idx]; | |
320ae51f JA |
1867 | |
1868 | /* | |
08e98fc6 ML |
1869 | * Allocate space for all possible cpus to avoid allocation at |
1870 | * runtime | |
320ae51f | 1871 | */ |
08e98fc6 ML |
1872 | hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *), |
1873 | GFP_KERNEL, node); | |
1874 | if (!hctx->ctxs) | |
1875 | goto unregister_cpu_notifier; | |
320ae51f | 1876 | |
88459642 OS |
1877 | if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8), GFP_KERNEL, |
1878 | node)) | |
08e98fc6 | 1879 | goto free_ctxs; |
320ae51f | 1880 | |
08e98fc6 | 1881 | hctx->nr_ctx = 0; |
320ae51f | 1882 | |
08e98fc6 ML |
1883 | if (set->ops->init_hctx && |
1884 | set->ops->init_hctx(hctx, set->driver_data, hctx_idx)) | |
1885 | goto free_bitmap; | |
320ae51f | 1886 | |
f70ced09 ML |
1887 | hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size); |
1888 | if (!hctx->fq) | |
1889 | goto exit_hctx; | |
320ae51f | 1890 | |
f70ced09 ML |
1891 | if (set->ops->init_request && |
1892 | set->ops->init_request(set->driver_data, | |
1893 | hctx->fq->flush_rq, hctx_idx, | |
1894 | flush_start_tag + hctx_idx, node)) | |
1895 | goto free_fq; | |
320ae51f | 1896 | |
6a83e74d BVA |
1897 | if (hctx->flags & BLK_MQ_F_BLOCKING) |
1898 | init_srcu_struct(&hctx->queue_rq_srcu); | |
1899 | ||
08e98fc6 | 1900 | return 0; |
320ae51f | 1901 | |
f70ced09 ML |
1902 | free_fq: |
1903 | kfree(hctx->fq); | |
1904 | exit_hctx: | |
1905 | if (set->ops->exit_hctx) | |
1906 | set->ops->exit_hctx(hctx, hctx_idx); | |
08e98fc6 | 1907 | free_bitmap: |
88459642 | 1908 | sbitmap_free(&hctx->ctx_map); |
08e98fc6 ML |
1909 | free_ctxs: |
1910 | kfree(hctx->ctxs); | |
1911 | unregister_cpu_notifier: | |
9467f859 | 1912 | blk_mq_remove_cpuhp(hctx); |
08e98fc6 ML |
1913 | return -1; |
1914 | } | |
320ae51f | 1915 | |
320ae51f JA |
1916 | static void blk_mq_init_cpu_queues(struct request_queue *q, |
1917 | unsigned int nr_hw_queues) | |
1918 | { | |
1919 | unsigned int i; | |
1920 | ||
1921 | for_each_possible_cpu(i) { | |
1922 | struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i); | |
1923 | struct blk_mq_hw_ctx *hctx; | |
1924 | ||
1925 | memset(__ctx, 0, sizeof(*__ctx)); | |
1926 | __ctx->cpu = i; | |
1927 | spin_lock_init(&__ctx->lock); | |
1928 | INIT_LIST_HEAD(&__ctx->rq_list); | |
1929 | __ctx->queue = q; | |
cf43e6be JA |
1930 | blk_stat_init(&__ctx->stat[BLK_STAT_READ]); |
1931 | blk_stat_init(&__ctx->stat[BLK_STAT_WRITE]); | |
320ae51f JA |
1932 | |
1933 | /* If the cpu isn't online, the cpu is mapped to first hctx */ | |
320ae51f JA |
1934 | if (!cpu_online(i)) |
1935 | continue; | |
1936 | ||
7d7e0f90 | 1937 | hctx = blk_mq_map_queue(q, i); |
e4043dcf | 1938 | |
320ae51f JA |
1939 | /* |
1940 | * Set local node, IFF we have more than one hw queue. If | |
1941 | * not, we remain on the home node of the device | |
1942 | */ | |
1943 | if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE) | |
bffed457 | 1944 | hctx->numa_node = local_memory_node(cpu_to_node(i)); |
320ae51f JA |
1945 | } |
1946 | } | |
1947 | ||
cc71a6f4 JA |
1948 | static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, int hctx_idx) |
1949 | { | |
1950 | int ret = 0; | |
1951 | ||
1952 | set->tags[hctx_idx] = blk_mq_alloc_rq_map(set, hctx_idx, | |
1953 | set->queue_depth, set->reserved_tags); | |
1954 | if (!set->tags[hctx_idx]) | |
1955 | return false; | |
1956 | ||
1957 | ret = blk_mq_alloc_rqs(set, set->tags[hctx_idx], hctx_idx, | |
1958 | set->queue_depth); | |
1959 | if (!ret) | |
1960 | return true; | |
1961 | ||
1962 | blk_mq_free_rq_map(set->tags[hctx_idx]); | |
1963 | set->tags[hctx_idx] = NULL; | |
1964 | return false; | |
1965 | } | |
1966 | ||
1967 | static void blk_mq_free_map_and_requests(struct blk_mq_tag_set *set, | |
1968 | unsigned int hctx_idx) | |
1969 | { | |
bd166ef1 JA |
1970 | if (set->tags[hctx_idx]) { |
1971 | blk_mq_free_rqs(set, set->tags[hctx_idx], hctx_idx); | |
1972 | blk_mq_free_rq_map(set->tags[hctx_idx]); | |
1973 | set->tags[hctx_idx] = NULL; | |
1974 | } | |
cc71a6f4 JA |
1975 | } |
1976 | ||
5778322e AM |
1977 | static void blk_mq_map_swqueue(struct request_queue *q, |
1978 | const struct cpumask *online_mask) | |
320ae51f | 1979 | { |
d1b1cea1 | 1980 | unsigned int i, hctx_idx; |
320ae51f JA |
1981 | struct blk_mq_hw_ctx *hctx; |
1982 | struct blk_mq_ctx *ctx; | |
2a34c087 | 1983 | struct blk_mq_tag_set *set = q->tag_set; |
320ae51f | 1984 | |
60de074b AM |
1985 | /* |
1986 | * Avoid others reading imcomplete hctx->cpumask through sysfs | |
1987 | */ | |
1988 | mutex_lock(&q->sysfs_lock); | |
1989 | ||
320ae51f | 1990 | queue_for_each_hw_ctx(q, hctx, i) { |
e4043dcf | 1991 | cpumask_clear(hctx->cpumask); |
320ae51f JA |
1992 | hctx->nr_ctx = 0; |
1993 | } | |
1994 | ||
1995 | /* | |
1996 | * Map software to hardware queues | |
1997 | */ | |
897bb0c7 | 1998 | for_each_possible_cpu(i) { |
320ae51f | 1999 | /* If the cpu isn't online, the cpu is mapped to first hctx */ |
5778322e | 2000 | if (!cpumask_test_cpu(i, online_mask)) |
e4043dcf JA |
2001 | continue; |
2002 | ||
d1b1cea1 GKB |
2003 | hctx_idx = q->mq_map[i]; |
2004 | /* unmapped hw queue can be remapped after CPU topo changed */ | |
cc71a6f4 JA |
2005 | if (!set->tags[hctx_idx] && |
2006 | !__blk_mq_alloc_rq_map(set, hctx_idx)) { | |
d1b1cea1 GKB |
2007 | /* |
2008 | * If tags initialization fail for some hctx, | |
2009 | * that hctx won't be brought online. In this | |
2010 | * case, remap the current ctx to hctx[0] which | |
2011 | * is guaranteed to always have tags allocated | |
2012 | */ | |
cc71a6f4 | 2013 | q->mq_map[i] = 0; |
d1b1cea1 GKB |
2014 | } |
2015 | ||
897bb0c7 | 2016 | ctx = per_cpu_ptr(q->queue_ctx, i); |
7d7e0f90 | 2017 | hctx = blk_mq_map_queue(q, i); |
868f2f0b | 2018 | |
e4043dcf | 2019 | cpumask_set_cpu(i, hctx->cpumask); |
320ae51f JA |
2020 | ctx->index_hw = hctx->nr_ctx; |
2021 | hctx->ctxs[hctx->nr_ctx++] = ctx; | |
2022 | } | |
506e931f | 2023 | |
60de074b AM |
2024 | mutex_unlock(&q->sysfs_lock); |
2025 | ||
506e931f | 2026 | queue_for_each_hw_ctx(q, hctx, i) { |
484b4061 | 2027 | /* |
a68aafa5 JA |
2028 | * If no software queues are mapped to this hardware queue, |
2029 | * disable it and free the request entries. | |
484b4061 JA |
2030 | */ |
2031 | if (!hctx->nr_ctx) { | |
d1b1cea1 GKB |
2032 | /* Never unmap queue 0. We need it as a |
2033 | * fallback in case of a new remap fails | |
2034 | * allocation | |
2035 | */ | |
cc71a6f4 JA |
2036 | if (i && set->tags[i]) |
2037 | blk_mq_free_map_and_requests(set, i); | |
2038 | ||
2a34c087 | 2039 | hctx->tags = NULL; |
484b4061 JA |
2040 | continue; |
2041 | } | |
2042 | ||
2a34c087 ML |
2043 | hctx->tags = set->tags[i]; |
2044 | WARN_ON(!hctx->tags); | |
2045 | ||
889fa31f CY |
2046 | /* |
2047 | * Set the map size to the number of mapped software queues. | |
2048 | * This is more accurate and more efficient than looping | |
2049 | * over all possibly mapped software queues. | |
2050 | */ | |
88459642 | 2051 | sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx); |
889fa31f | 2052 | |
484b4061 JA |
2053 | /* |
2054 | * Initialize batch roundrobin counts | |
2055 | */ | |
506e931f JA |
2056 | hctx->next_cpu = cpumask_first(hctx->cpumask); |
2057 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; | |
2058 | } | |
320ae51f JA |
2059 | } |
2060 | ||
2404e607 | 2061 | static void queue_set_hctx_shared(struct request_queue *q, bool shared) |
0d2602ca JA |
2062 | { |
2063 | struct blk_mq_hw_ctx *hctx; | |
0d2602ca JA |
2064 | int i; |
2065 | ||
2404e607 JM |
2066 | queue_for_each_hw_ctx(q, hctx, i) { |
2067 | if (shared) | |
2068 | hctx->flags |= BLK_MQ_F_TAG_SHARED; | |
2069 | else | |
2070 | hctx->flags &= ~BLK_MQ_F_TAG_SHARED; | |
2071 | } | |
2072 | } | |
2073 | ||
2074 | static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set, bool shared) | |
2075 | { | |
2076 | struct request_queue *q; | |
0d2602ca JA |
2077 | |
2078 | list_for_each_entry(q, &set->tag_list, tag_set_list) { | |
2079 | blk_mq_freeze_queue(q); | |
2404e607 | 2080 | queue_set_hctx_shared(q, shared); |
0d2602ca JA |
2081 | blk_mq_unfreeze_queue(q); |
2082 | } | |
2083 | } | |
2084 | ||
2085 | static void blk_mq_del_queue_tag_set(struct request_queue *q) | |
2086 | { | |
2087 | struct blk_mq_tag_set *set = q->tag_set; | |
2088 | ||
0d2602ca JA |
2089 | mutex_lock(&set->tag_list_lock); |
2090 | list_del_init(&q->tag_set_list); | |
2404e607 JM |
2091 | if (list_is_singular(&set->tag_list)) { |
2092 | /* just transitioned to unshared */ | |
2093 | set->flags &= ~BLK_MQ_F_TAG_SHARED; | |
2094 | /* update existing queue */ | |
2095 | blk_mq_update_tag_set_depth(set, false); | |
2096 | } | |
0d2602ca | 2097 | mutex_unlock(&set->tag_list_lock); |
0d2602ca JA |
2098 | } |
2099 | ||
2100 | static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set, | |
2101 | struct request_queue *q) | |
2102 | { | |
2103 | q->tag_set = set; | |
2104 | ||
2105 | mutex_lock(&set->tag_list_lock); | |
2404e607 JM |
2106 | |
2107 | /* Check to see if we're transitioning to shared (from 1 to 2 queues). */ | |
2108 | if (!list_empty(&set->tag_list) && !(set->flags & BLK_MQ_F_TAG_SHARED)) { | |
2109 | set->flags |= BLK_MQ_F_TAG_SHARED; | |
2110 | /* update existing queue */ | |
2111 | blk_mq_update_tag_set_depth(set, true); | |
2112 | } | |
2113 | if (set->flags & BLK_MQ_F_TAG_SHARED) | |
2114 | queue_set_hctx_shared(q, true); | |
0d2602ca | 2115 | list_add_tail(&q->tag_set_list, &set->tag_list); |
2404e607 | 2116 | |
0d2602ca JA |
2117 | mutex_unlock(&set->tag_list_lock); |
2118 | } | |
2119 | ||
e09aae7e ML |
2120 | /* |
2121 | * It is the actual release handler for mq, but we do it from | |
2122 | * request queue's release handler for avoiding use-after-free | |
2123 | * and headache because q->mq_kobj shouldn't have been introduced, | |
2124 | * but we can't group ctx/kctx kobj without it. | |
2125 | */ | |
2126 | void blk_mq_release(struct request_queue *q) | |
2127 | { | |
2128 | struct blk_mq_hw_ctx *hctx; | |
2129 | unsigned int i; | |
2130 | ||
bd166ef1 JA |
2131 | blk_mq_sched_teardown(q); |
2132 | ||
e09aae7e | 2133 | /* hctx kobj stays in hctx */ |
c3b4afca ML |
2134 | queue_for_each_hw_ctx(q, hctx, i) { |
2135 | if (!hctx) | |
2136 | continue; | |
2137 | kfree(hctx->ctxs); | |
e09aae7e | 2138 | kfree(hctx); |
c3b4afca | 2139 | } |
e09aae7e | 2140 | |
a723bab3 AM |
2141 | q->mq_map = NULL; |
2142 | ||
e09aae7e ML |
2143 | kfree(q->queue_hw_ctx); |
2144 | ||
2145 | /* ctx kobj stays in queue_ctx */ | |
2146 | free_percpu(q->queue_ctx); | |
2147 | } | |
2148 | ||
24d2f903 | 2149 | struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set) |
b62c21b7 MS |
2150 | { |
2151 | struct request_queue *uninit_q, *q; | |
2152 | ||
2153 | uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node); | |
2154 | if (!uninit_q) | |
2155 | return ERR_PTR(-ENOMEM); | |
2156 | ||
2157 | q = blk_mq_init_allocated_queue(set, uninit_q); | |
2158 | if (IS_ERR(q)) | |
2159 | blk_cleanup_queue(uninit_q); | |
2160 | ||
2161 | return q; | |
2162 | } | |
2163 | EXPORT_SYMBOL(blk_mq_init_queue); | |
2164 | ||
868f2f0b KB |
2165 | static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set, |
2166 | struct request_queue *q) | |
320ae51f | 2167 | { |
868f2f0b KB |
2168 | int i, j; |
2169 | struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx; | |
f14bbe77 | 2170 | |
868f2f0b | 2171 | blk_mq_sysfs_unregister(q); |
24d2f903 | 2172 | for (i = 0; i < set->nr_hw_queues; i++) { |
868f2f0b | 2173 | int node; |
f14bbe77 | 2174 | |
868f2f0b KB |
2175 | if (hctxs[i]) |
2176 | continue; | |
2177 | ||
2178 | node = blk_mq_hw_queue_to_node(q->mq_map, i); | |
cdef54dd CH |
2179 | hctxs[i] = kzalloc_node(sizeof(struct blk_mq_hw_ctx), |
2180 | GFP_KERNEL, node); | |
320ae51f | 2181 | if (!hctxs[i]) |
868f2f0b | 2182 | break; |
320ae51f | 2183 | |
a86073e4 | 2184 | if (!zalloc_cpumask_var_node(&hctxs[i]->cpumask, GFP_KERNEL, |
868f2f0b KB |
2185 | node)) { |
2186 | kfree(hctxs[i]); | |
2187 | hctxs[i] = NULL; | |
2188 | break; | |
2189 | } | |
e4043dcf | 2190 | |
0d2602ca | 2191 | atomic_set(&hctxs[i]->nr_active, 0); |
f14bbe77 | 2192 | hctxs[i]->numa_node = node; |
320ae51f | 2193 | hctxs[i]->queue_num = i; |
868f2f0b KB |
2194 | |
2195 | if (blk_mq_init_hctx(q, set, hctxs[i], i)) { | |
2196 | free_cpumask_var(hctxs[i]->cpumask); | |
2197 | kfree(hctxs[i]); | |
2198 | hctxs[i] = NULL; | |
2199 | break; | |
2200 | } | |
2201 | blk_mq_hctx_kobj_init(hctxs[i]); | |
320ae51f | 2202 | } |
868f2f0b KB |
2203 | for (j = i; j < q->nr_hw_queues; j++) { |
2204 | struct blk_mq_hw_ctx *hctx = hctxs[j]; | |
2205 | ||
2206 | if (hctx) { | |
cc71a6f4 JA |
2207 | if (hctx->tags) |
2208 | blk_mq_free_map_and_requests(set, j); | |
868f2f0b KB |
2209 | blk_mq_exit_hctx(q, set, hctx, j); |
2210 | free_cpumask_var(hctx->cpumask); | |
2211 | kobject_put(&hctx->kobj); | |
2212 | kfree(hctx->ctxs); | |
2213 | kfree(hctx); | |
2214 | hctxs[j] = NULL; | |
2215 | ||
2216 | } | |
2217 | } | |
2218 | q->nr_hw_queues = i; | |
2219 | blk_mq_sysfs_register(q); | |
2220 | } | |
2221 | ||
2222 | struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set, | |
2223 | struct request_queue *q) | |
2224 | { | |
66841672 ML |
2225 | /* mark the queue as mq asap */ |
2226 | q->mq_ops = set->ops; | |
2227 | ||
868f2f0b KB |
2228 | q->queue_ctx = alloc_percpu(struct blk_mq_ctx); |
2229 | if (!q->queue_ctx) | |
c7de5726 | 2230 | goto err_exit; |
868f2f0b KB |
2231 | |
2232 | q->queue_hw_ctx = kzalloc_node(nr_cpu_ids * sizeof(*(q->queue_hw_ctx)), | |
2233 | GFP_KERNEL, set->numa_node); | |
2234 | if (!q->queue_hw_ctx) | |
2235 | goto err_percpu; | |
2236 | ||
bdd17e75 | 2237 | q->mq_map = set->mq_map; |
868f2f0b KB |
2238 | |
2239 | blk_mq_realloc_hw_ctxs(set, q); | |
2240 | if (!q->nr_hw_queues) | |
2241 | goto err_hctxs; | |
320ae51f | 2242 | |
287922eb | 2243 | INIT_WORK(&q->timeout_work, blk_mq_timeout_work); |
e56f698b | 2244 | blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ); |
320ae51f JA |
2245 | |
2246 | q->nr_queues = nr_cpu_ids; | |
320ae51f | 2247 | |
94eddfbe | 2248 | q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT; |
320ae51f | 2249 | |
05f1dd53 JA |
2250 | if (!(set->flags & BLK_MQ_F_SG_MERGE)) |
2251 | q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE; | |
2252 | ||
1be036e9 CH |
2253 | q->sg_reserved_size = INT_MAX; |
2254 | ||
2849450a | 2255 | INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work); |
6fca6a61 CH |
2256 | INIT_LIST_HEAD(&q->requeue_list); |
2257 | spin_lock_init(&q->requeue_lock); | |
2258 | ||
07068d5b JA |
2259 | if (q->nr_hw_queues > 1) |
2260 | blk_queue_make_request(q, blk_mq_make_request); | |
2261 | else | |
2262 | blk_queue_make_request(q, blk_sq_make_request); | |
2263 | ||
eba71768 JA |
2264 | /* |
2265 | * Do this after blk_queue_make_request() overrides it... | |
2266 | */ | |
2267 | q->nr_requests = set->queue_depth; | |
2268 | ||
64f1c21e JA |
2269 | /* |
2270 | * Default to classic polling | |
2271 | */ | |
2272 | q->poll_nsec = -1; | |
2273 | ||
24d2f903 CH |
2274 | if (set->ops->complete) |
2275 | blk_queue_softirq_done(q, set->ops->complete); | |
30a91cb4 | 2276 | |
24d2f903 | 2277 | blk_mq_init_cpu_queues(q, set->nr_hw_queues); |
320ae51f | 2278 | |
5778322e | 2279 | get_online_cpus(); |
320ae51f | 2280 | mutex_lock(&all_q_mutex); |
320ae51f | 2281 | |
4593fdbe | 2282 | list_add_tail(&q->all_q_node, &all_q_list); |
0d2602ca | 2283 | blk_mq_add_queue_tag_set(set, q); |
5778322e | 2284 | blk_mq_map_swqueue(q, cpu_online_mask); |
484b4061 | 2285 | |
4593fdbe | 2286 | mutex_unlock(&all_q_mutex); |
5778322e | 2287 | put_online_cpus(); |
4593fdbe | 2288 | |
d3484991 JA |
2289 | if (!(set->flags & BLK_MQ_F_NO_SCHED)) { |
2290 | int ret; | |
2291 | ||
2292 | ret = blk_mq_sched_init(q); | |
2293 | if (ret) | |
2294 | return ERR_PTR(ret); | |
2295 | } | |
2296 | ||
320ae51f | 2297 | return q; |
18741986 | 2298 | |
320ae51f | 2299 | err_hctxs: |
868f2f0b | 2300 | kfree(q->queue_hw_ctx); |
320ae51f | 2301 | err_percpu: |
868f2f0b | 2302 | free_percpu(q->queue_ctx); |
c7de5726 ML |
2303 | err_exit: |
2304 | q->mq_ops = NULL; | |
320ae51f JA |
2305 | return ERR_PTR(-ENOMEM); |
2306 | } | |
b62c21b7 | 2307 | EXPORT_SYMBOL(blk_mq_init_allocated_queue); |
320ae51f JA |
2308 | |
2309 | void blk_mq_free_queue(struct request_queue *q) | |
2310 | { | |
624dbe47 | 2311 | struct blk_mq_tag_set *set = q->tag_set; |
320ae51f | 2312 | |
0e626368 AM |
2313 | mutex_lock(&all_q_mutex); |
2314 | list_del_init(&q->all_q_node); | |
2315 | mutex_unlock(&all_q_mutex); | |
2316 | ||
87760e5e JA |
2317 | wbt_exit(q); |
2318 | ||
0d2602ca JA |
2319 | blk_mq_del_queue_tag_set(q); |
2320 | ||
624dbe47 ML |
2321 | blk_mq_exit_hw_queues(q, set, set->nr_hw_queues); |
2322 | blk_mq_free_hw_queues(q, set); | |
320ae51f | 2323 | } |
320ae51f JA |
2324 | |
2325 | /* Basically redo blk_mq_init_queue with queue frozen */ | |
5778322e AM |
2326 | static void blk_mq_queue_reinit(struct request_queue *q, |
2327 | const struct cpumask *online_mask) | |
320ae51f | 2328 | { |
4ecd4fef | 2329 | WARN_ON_ONCE(!atomic_read(&q->mq_freeze_depth)); |
320ae51f | 2330 | |
67aec14c JA |
2331 | blk_mq_sysfs_unregister(q); |
2332 | ||
320ae51f JA |
2333 | /* |
2334 | * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe | |
2335 | * we should change hctx numa_node according to new topology (this | |
2336 | * involves free and re-allocate memory, worthy doing?) | |
2337 | */ | |
2338 | ||
5778322e | 2339 | blk_mq_map_swqueue(q, online_mask); |
320ae51f | 2340 | |
67aec14c | 2341 | blk_mq_sysfs_register(q); |
320ae51f JA |
2342 | } |
2343 | ||
65d5291e SAS |
2344 | /* |
2345 | * New online cpumask which is going to be set in this hotplug event. | |
2346 | * Declare this cpumasks as global as cpu-hotplug operation is invoked | |
2347 | * one-by-one and dynamically allocating this could result in a failure. | |
2348 | */ | |
2349 | static struct cpumask cpuhp_online_new; | |
2350 | ||
2351 | static void blk_mq_queue_reinit_work(void) | |
320ae51f JA |
2352 | { |
2353 | struct request_queue *q; | |
320ae51f JA |
2354 | |
2355 | mutex_lock(&all_q_mutex); | |
f3af020b TH |
2356 | /* |
2357 | * We need to freeze and reinit all existing queues. Freezing | |
2358 | * involves synchronous wait for an RCU grace period and doing it | |
2359 | * one by one may take a long time. Start freezing all queues in | |
2360 | * one swoop and then wait for the completions so that freezing can | |
2361 | * take place in parallel. | |
2362 | */ | |
2363 | list_for_each_entry(q, &all_q_list, all_q_node) | |
2364 | blk_mq_freeze_queue_start(q); | |
415d3dab | 2365 | list_for_each_entry(q, &all_q_list, all_q_node) |
f3af020b TH |
2366 | blk_mq_freeze_queue_wait(q); |
2367 | ||
320ae51f | 2368 | list_for_each_entry(q, &all_q_list, all_q_node) |
65d5291e | 2369 | blk_mq_queue_reinit(q, &cpuhp_online_new); |
f3af020b TH |
2370 | |
2371 | list_for_each_entry(q, &all_q_list, all_q_node) | |
2372 | blk_mq_unfreeze_queue(q); | |
2373 | ||
320ae51f | 2374 | mutex_unlock(&all_q_mutex); |
65d5291e SAS |
2375 | } |
2376 | ||
2377 | static int blk_mq_queue_reinit_dead(unsigned int cpu) | |
2378 | { | |
97a32864 | 2379 | cpumask_copy(&cpuhp_online_new, cpu_online_mask); |
65d5291e SAS |
2380 | blk_mq_queue_reinit_work(); |
2381 | return 0; | |
2382 | } | |
2383 | ||
2384 | /* | |
2385 | * Before hotadded cpu starts handling requests, new mappings must be | |
2386 | * established. Otherwise, these requests in hw queue might never be | |
2387 | * dispatched. | |
2388 | * | |
2389 | * For example, there is a single hw queue (hctx) and two CPU queues (ctx0 | |
2390 | * for CPU0, and ctx1 for CPU1). | |
2391 | * | |
2392 | * Now CPU1 is just onlined and a request is inserted into ctx1->rq_list | |
2393 | * and set bit0 in pending bitmap as ctx1->index_hw is still zero. | |
2394 | * | |
2c3ad667 JA |
2395 | * And then while running hw queue, blk_mq_flush_busy_ctxs() finds bit0 is set |
2396 | * in pending bitmap and tries to retrieve requests in hctx->ctxs[0]->rq_list. | |
2397 | * But htx->ctxs[0] is a pointer to ctx0, so the request in ctx1->rq_list is | |
2398 | * ignored. | |
65d5291e SAS |
2399 | */ |
2400 | static int blk_mq_queue_reinit_prepare(unsigned int cpu) | |
2401 | { | |
2402 | cpumask_copy(&cpuhp_online_new, cpu_online_mask); | |
2403 | cpumask_set_cpu(cpu, &cpuhp_online_new); | |
2404 | blk_mq_queue_reinit_work(); | |
2405 | return 0; | |
320ae51f JA |
2406 | } |
2407 | ||
a5164405 JA |
2408 | static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set) |
2409 | { | |
2410 | int i; | |
2411 | ||
cc71a6f4 JA |
2412 | for (i = 0; i < set->nr_hw_queues; i++) |
2413 | if (!__blk_mq_alloc_rq_map(set, i)) | |
a5164405 | 2414 | goto out_unwind; |
a5164405 JA |
2415 | |
2416 | return 0; | |
2417 | ||
2418 | out_unwind: | |
2419 | while (--i >= 0) | |
cc71a6f4 | 2420 | blk_mq_free_rq_map(set->tags[i]); |
a5164405 | 2421 | |
a5164405 JA |
2422 | return -ENOMEM; |
2423 | } | |
2424 | ||
2425 | /* | |
2426 | * Allocate the request maps associated with this tag_set. Note that this | |
2427 | * may reduce the depth asked for, if memory is tight. set->queue_depth | |
2428 | * will be updated to reflect the allocated depth. | |
2429 | */ | |
2430 | static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set) | |
2431 | { | |
2432 | unsigned int depth; | |
2433 | int err; | |
2434 | ||
2435 | depth = set->queue_depth; | |
2436 | do { | |
2437 | err = __blk_mq_alloc_rq_maps(set); | |
2438 | if (!err) | |
2439 | break; | |
2440 | ||
2441 | set->queue_depth >>= 1; | |
2442 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) { | |
2443 | err = -ENOMEM; | |
2444 | break; | |
2445 | } | |
2446 | } while (set->queue_depth); | |
2447 | ||
2448 | if (!set->queue_depth || err) { | |
2449 | pr_err("blk-mq: failed to allocate request map\n"); | |
2450 | return -ENOMEM; | |
2451 | } | |
2452 | ||
2453 | if (depth != set->queue_depth) | |
2454 | pr_info("blk-mq: reduced tag depth (%u -> %u)\n", | |
2455 | depth, set->queue_depth); | |
2456 | ||
2457 | return 0; | |
2458 | } | |
2459 | ||
a4391c64 JA |
2460 | /* |
2461 | * Alloc a tag set to be associated with one or more request queues. | |
2462 | * May fail with EINVAL for various error conditions. May adjust the | |
2463 | * requested depth down, if if it too large. In that case, the set | |
2464 | * value will be stored in set->queue_depth. | |
2465 | */ | |
24d2f903 CH |
2466 | int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set) |
2467 | { | |
da695ba2 CH |
2468 | int ret; |
2469 | ||
205fb5f5 BVA |
2470 | BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS); |
2471 | ||
24d2f903 CH |
2472 | if (!set->nr_hw_queues) |
2473 | return -EINVAL; | |
a4391c64 | 2474 | if (!set->queue_depth) |
24d2f903 CH |
2475 | return -EINVAL; |
2476 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) | |
2477 | return -EINVAL; | |
2478 | ||
7d7e0f90 | 2479 | if (!set->ops->queue_rq) |
24d2f903 CH |
2480 | return -EINVAL; |
2481 | ||
a4391c64 JA |
2482 | if (set->queue_depth > BLK_MQ_MAX_DEPTH) { |
2483 | pr_info("blk-mq: reduced tag depth to %u\n", | |
2484 | BLK_MQ_MAX_DEPTH); | |
2485 | set->queue_depth = BLK_MQ_MAX_DEPTH; | |
2486 | } | |
24d2f903 | 2487 | |
6637fadf SL |
2488 | /* |
2489 | * If a crashdump is active, then we are potentially in a very | |
2490 | * memory constrained environment. Limit us to 1 queue and | |
2491 | * 64 tags to prevent using too much memory. | |
2492 | */ | |
2493 | if (is_kdump_kernel()) { | |
2494 | set->nr_hw_queues = 1; | |
2495 | set->queue_depth = min(64U, set->queue_depth); | |
2496 | } | |
868f2f0b KB |
2497 | /* |
2498 | * There is no use for more h/w queues than cpus. | |
2499 | */ | |
2500 | if (set->nr_hw_queues > nr_cpu_ids) | |
2501 | set->nr_hw_queues = nr_cpu_ids; | |
6637fadf | 2502 | |
868f2f0b | 2503 | set->tags = kzalloc_node(nr_cpu_ids * sizeof(struct blk_mq_tags *), |
24d2f903 CH |
2504 | GFP_KERNEL, set->numa_node); |
2505 | if (!set->tags) | |
a5164405 | 2506 | return -ENOMEM; |
24d2f903 | 2507 | |
da695ba2 CH |
2508 | ret = -ENOMEM; |
2509 | set->mq_map = kzalloc_node(sizeof(*set->mq_map) * nr_cpu_ids, | |
2510 | GFP_KERNEL, set->numa_node); | |
bdd17e75 CH |
2511 | if (!set->mq_map) |
2512 | goto out_free_tags; | |
2513 | ||
da695ba2 CH |
2514 | if (set->ops->map_queues) |
2515 | ret = set->ops->map_queues(set); | |
2516 | else | |
2517 | ret = blk_mq_map_queues(set); | |
2518 | if (ret) | |
2519 | goto out_free_mq_map; | |
2520 | ||
2521 | ret = blk_mq_alloc_rq_maps(set); | |
2522 | if (ret) | |
bdd17e75 | 2523 | goto out_free_mq_map; |
24d2f903 | 2524 | |
0d2602ca JA |
2525 | mutex_init(&set->tag_list_lock); |
2526 | INIT_LIST_HEAD(&set->tag_list); | |
2527 | ||
24d2f903 | 2528 | return 0; |
bdd17e75 CH |
2529 | |
2530 | out_free_mq_map: | |
2531 | kfree(set->mq_map); | |
2532 | set->mq_map = NULL; | |
2533 | out_free_tags: | |
5676e7b6 RE |
2534 | kfree(set->tags); |
2535 | set->tags = NULL; | |
da695ba2 | 2536 | return ret; |
24d2f903 CH |
2537 | } |
2538 | EXPORT_SYMBOL(blk_mq_alloc_tag_set); | |
2539 | ||
2540 | void blk_mq_free_tag_set(struct blk_mq_tag_set *set) | |
2541 | { | |
2542 | int i; | |
2543 | ||
cc71a6f4 JA |
2544 | for (i = 0; i < nr_cpu_ids; i++) |
2545 | blk_mq_free_map_and_requests(set, i); | |
484b4061 | 2546 | |
bdd17e75 CH |
2547 | kfree(set->mq_map); |
2548 | set->mq_map = NULL; | |
2549 | ||
981bd189 | 2550 | kfree(set->tags); |
5676e7b6 | 2551 | set->tags = NULL; |
24d2f903 CH |
2552 | } |
2553 | EXPORT_SYMBOL(blk_mq_free_tag_set); | |
2554 | ||
e3a2b3f9 JA |
2555 | int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr) |
2556 | { | |
2557 | struct blk_mq_tag_set *set = q->tag_set; | |
2558 | struct blk_mq_hw_ctx *hctx; | |
2559 | int i, ret; | |
2560 | ||
bd166ef1 | 2561 | if (!set) |
e3a2b3f9 JA |
2562 | return -EINVAL; |
2563 | ||
2564 | ret = 0; | |
2565 | queue_for_each_hw_ctx(q, hctx, i) { | |
e9137d4b KB |
2566 | if (!hctx->tags) |
2567 | continue; | |
bd166ef1 JA |
2568 | /* |
2569 | * If we're using an MQ scheduler, just update the scheduler | |
2570 | * queue depth. This is similar to what the old code would do. | |
2571 | */ | |
2572 | if (!hctx->sched_tags) | |
2573 | ret = blk_mq_tag_update_depth(hctx->tags, | |
2574 | min(nr, set->queue_depth)); | |
2575 | else | |
2576 | ret = blk_mq_tag_update_depth(hctx->sched_tags, nr); | |
e3a2b3f9 JA |
2577 | if (ret) |
2578 | break; | |
2579 | } | |
2580 | ||
2581 | if (!ret) | |
2582 | q->nr_requests = nr; | |
2583 | ||
2584 | return ret; | |
2585 | } | |
2586 | ||
868f2f0b KB |
2587 | void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues) |
2588 | { | |
2589 | struct request_queue *q; | |
2590 | ||
2591 | if (nr_hw_queues > nr_cpu_ids) | |
2592 | nr_hw_queues = nr_cpu_ids; | |
2593 | if (nr_hw_queues < 1 || nr_hw_queues == set->nr_hw_queues) | |
2594 | return; | |
2595 | ||
2596 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
2597 | blk_mq_freeze_queue(q); | |
2598 | ||
2599 | set->nr_hw_queues = nr_hw_queues; | |
2600 | list_for_each_entry(q, &set->tag_list, tag_set_list) { | |
2601 | blk_mq_realloc_hw_ctxs(set, q); | |
2602 | ||
2603 | if (q->nr_hw_queues > 1) | |
2604 | blk_queue_make_request(q, blk_mq_make_request); | |
2605 | else | |
2606 | blk_queue_make_request(q, blk_sq_make_request); | |
2607 | ||
2608 | blk_mq_queue_reinit(q, cpu_online_mask); | |
2609 | } | |
2610 | ||
2611 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
2612 | blk_mq_unfreeze_queue(q); | |
2613 | } | |
2614 | EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues); | |
2615 | ||
64f1c21e JA |
2616 | static unsigned long blk_mq_poll_nsecs(struct request_queue *q, |
2617 | struct blk_mq_hw_ctx *hctx, | |
2618 | struct request *rq) | |
2619 | { | |
2620 | struct blk_rq_stat stat[2]; | |
2621 | unsigned long ret = 0; | |
2622 | ||
2623 | /* | |
2624 | * If stats collection isn't on, don't sleep but turn it on for | |
2625 | * future users | |
2626 | */ | |
2627 | if (!blk_stat_enable(q)) | |
2628 | return 0; | |
2629 | ||
2630 | /* | |
2631 | * We don't have to do this once per IO, should optimize this | |
2632 | * to just use the current window of stats until it changes | |
2633 | */ | |
2634 | memset(&stat, 0, sizeof(stat)); | |
2635 | blk_hctx_stat_get(hctx, stat); | |
2636 | ||
2637 | /* | |
2638 | * As an optimistic guess, use half of the mean service time | |
2639 | * for this type of request. We can (and should) make this smarter. | |
2640 | * For instance, if the completion latencies are tight, we can | |
2641 | * get closer than just half the mean. This is especially | |
2642 | * important on devices where the completion latencies are longer | |
2643 | * than ~10 usec. | |
2644 | */ | |
2645 | if (req_op(rq) == REQ_OP_READ && stat[BLK_STAT_READ].nr_samples) | |
2646 | ret = (stat[BLK_STAT_READ].mean + 1) / 2; | |
2647 | else if (req_op(rq) == REQ_OP_WRITE && stat[BLK_STAT_WRITE].nr_samples) | |
2648 | ret = (stat[BLK_STAT_WRITE].mean + 1) / 2; | |
2649 | ||
2650 | return ret; | |
2651 | } | |
2652 | ||
06426adf | 2653 | static bool blk_mq_poll_hybrid_sleep(struct request_queue *q, |
64f1c21e | 2654 | struct blk_mq_hw_ctx *hctx, |
06426adf JA |
2655 | struct request *rq) |
2656 | { | |
2657 | struct hrtimer_sleeper hs; | |
2658 | enum hrtimer_mode mode; | |
64f1c21e | 2659 | unsigned int nsecs; |
06426adf JA |
2660 | ktime_t kt; |
2661 | ||
64f1c21e JA |
2662 | if (test_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags)) |
2663 | return false; | |
2664 | ||
2665 | /* | |
2666 | * poll_nsec can be: | |
2667 | * | |
2668 | * -1: don't ever hybrid sleep | |
2669 | * 0: use half of prev avg | |
2670 | * >0: use this specific value | |
2671 | */ | |
2672 | if (q->poll_nsec == -1) | |
2673 | return false; | |
2674 | else if (q->poll_nsec > 0) | |
2675 | nsecs = q->poll_nsec; | |
2676 | else | |
2677 | nsecs = blk_mq_poll_nsecs(q, hctx, rq); | |
2678 | ||
2679 | if (!nsecs) | |
06426adf JA |
2680 | return false; |
2681 | ||
2682 | set_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags); | |
2683 | ||
2684 | /* | |
2685 | * This will be replaced with the stats tracking code, using | |
2686 | * 'avg_completion_time / 2' as the pre-sleep target. | |
2687 | */ | |
8b0e1953 | 2688 | kt = nsecs; |
06426adf JA |
2689 | |
2690 | mode = HRTIMER_MODE_REL; | |
2691 | hrtimer_init_on_stack(&hs.timer, CLOCK_MONOTONIC, mode); | |
2692 | hrtimer_set_expires(&hs.timer, kt); | |
2693 | ||
2694 | hrtimer_init_sleeper(&hs, current); | |
2695 | do { | |
2696 | if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags)) | |
2697 | break; | |
2698 | set_current_state(TASK_UNINTERRUPTIBLE); | |
2699 | hrtimer_start_expires(&hs.timer, mode); | |
2700 | if (hs.task) | |
2701 | io_schedule(); | |
2702 | hrtimer_cancel(&hs.timer); | |
2703 | mode = HRTIMER_MODE_ABS; | |
2704 | } while (hs.task && !signal_pending(current)); | |
2705 | ||
2706 | __set_current_state(TASK_RUNNING); | |
2707 | destroy_hrtimer_on_stack(&hs.timer); | |
2708 | return true; | |
2709 | } | |
2710 | ||
bbd7bb70 JA |
2711 | static bool __blk_mq_poll(struct blk_mq_hw_ctx *hctx, struct request *rq) |
2712 | { | |
2713 | struct request_queue *q = hctx->queue; | |
2714 | long state; | |
2715 | ||
06426adf JA |
2716 | /* |
2717 | * If we sleep, have the caller restart the poll loop to reset | |
2718 | * the state. Like for the other success return cases, the | |
2719 | * caller is responsible for checking if the IO completed. If | |
2720 | * the IO isn't complete, we'll get called again and will go | |
2721 | * straight to the busy poll loop. | |
2722 | */ | |
64f1c21e | 2723 | if (blk_mq_poll_hybrid_sleep(q, hctx, rq)) |
06426adf JA |
2724 | return true; |
2725 | ||
bbd7bb70 JA |
2726 | hctx->poll_considered++; |
2727 | ||
2728 | state = current->state; | |
2729 | while (!need_resched()) { | |
2730 | int ret; | |
2731 | ||
2732 | hctx->poll_invoked++; | |
2733 | ||
2734 | ret = q->mq_ops->poll(hctx, rq->tag); | |
2735 | if (ret > 0) { | |
2736 | hctx->poll_success++; | |
2737 | set_current_state(TASK_RUNNING); | |
2738 | return true; | |
2739 | } | |
2740 | ||
2741 | if (signal_pending_state(state, current)) | |
2742 | set_current_state(TASK_RUNNING); | |
2743 | ||
2744 | if (current->state == TASK_RUNNING) | |
2745 | return true; | |
2746 | if (ret < 0) | |
2747 | break; | |
2748 | cpu_relax(); | |
2749 | } | |
2750 | ||
2751 | return false; | |
2752 | } | |
2753 | ||
2754 | bool blk_mq_poll(struct request_queue *q, blk_qc_t cookie) | |
2755 | { | |
2756 | struct blk_mq_hw_ctx *hctx; | |
2757 | struct blk_plug *plug; | |
2758 | struct request *rq; | |
2759 | ||
2760 | if (!q->mq_ops || !q->mq_ops->poll || !blk_qc_t_valid(cookie) || | |
2761 | !test_bit(QUEUE_FLAG_POLL, &q->queue_flags)) | |
2762 | return false; | |
2763 | ||
2764 | plug = current->plug; | |
2765 | if (plug) | |
2766 | blk_flush_plug_list(plug, false); | |
2767 | ||
2768 | hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)]; | |
bd166ef1 JA |
2769 | if (!blk_qc_t_is_internal(cookie)) |
2770 | rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie)); | |
2771 | else | |
2772 | rq = blk_mq_tag_to_rq(hctx->sched_tags, blk_qc_t_to_tag(cookie)); | |
bbd7bb70 JA |
2773 | |
2774 | return __blk_mq_poll(hctx, rq); | |
2775 | } | |
2776 | EXPORT_SYMBOL_GPL(blk_mq_poll); | |
2777 | ||
676141e4 JA |
2778 | void blk_mq_disable_hotplug(void) |
2779 | { | |
2780 | mutex_lock(&all_q_mutex); | |
2781 | } | |
2782 | ||
2783 | void blk_mq_enable_hotplug(void) | |
2784 | { | |
2785 | mutex_unlock(&all_q_mutex); | |
2786 | } | |
2787 | ||
320ae51f JA |
2788 | static int __init blk_mq_init(void) |
2789 | { | |
9467f859 TG |
2790 | cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL, |
2791 | blk_mq_hctx_notify_dead); | |
320ae51f | 2792 | |
65d5291e SAS |
2793 | cpuhp_setup_state_nocalls(CPUHP_BLK_MQ_PREPARE, "block/mq:prepare", |
2794 | blk_mq_queue_reinit_prepare, | |
2795 | blk_mq_queue_reinit_dead); | |
320ae51f JA |
2796 | return 0; |
2797 | } | |
2798 | subsys_initcall(blk_mq_init); |