]>
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 | */ | |
50e1dab8 | 43 | bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx) |
320ae51f | 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 | rq->extra_len = 0; |
af76e555 | 203 | |
af76e555 | 204 | INIT_LIST_HEAD(&rq->timeout_list); |
f6be4fb4 JA |
205 | rq->timeout = 0; |
206 | ||
af76e555 CH |
207 | rq->end_io = NULL; |
208 | rq->end_io_data = NULL; | |
209 | rq->next_rq = NULL; | |
210 | ||
ef295ecf | 211 | ctx->rq_dispatched[op_is_sync(op)]++; |
320ae51f | 212 | } |
2c3ad667 | 213 | EXPORT_SYMBOL_GPL(blk_mq_rq_ctx_init); |
320ae51f | 214 | |
2c3ad667 JA |
215 | struct request *__blk_mq_alloc_request(struct blk_mq_alloc_data *data, |
216 | unsigned int op) | |
5dee8577 CH |
217 | { |
218 | struct request *rq; | |
219 | unsigned int tag; | |
220 | ||
cb96a42c | 221 | tag = blk_mq_get_tag(data); |
5dee8577 | 222 | if (tag != BLK_MQ_TAG_FAIL) { |
bd166ef1 JA |
223 | struct blk_mq_tags *tags = blk_mq_tags_from_data(data); |
224 | ||
225 | rq = tags->static_rqs[tag]; | |
5dee8577 | 226 | |
bd166ef1 JA |
227 | if (data->flags & BLK_MQ_REQ_INTERNAL) { |
228 | rq->tag = -1; | |
229 | rq->internal_tag = tag; | |
230 | } else { | |
200e86b3 JA |
231 | if (blk_mq_tag_busy(data->hctx)) { |
232 | rq->rq_flags = RQF_MQ_INFLIGHT; | |
233 | atomic_inc(&data->hctx->nr_active); | |
234 | } | |
bd166ef1 JA |
235 | rq->tag = tag; |
236 | rq->internal_tag = -1; | |
237 | } | |
238 | ||
ef295ecf | 239 | blk_mq_rq_ctx_init(data->q, data->ctx, rq, op); |
5dee8577 CH |
240 | return rq; |
241 | } | |
242 | ||
243 | return NULL; | |
244 | } | |
2c3ad667 | 245 | EXPORT_SYMBOL_GPL(__blk_mq_alloc_request); |
5dee8577 | 246 | |
6f3b0e8b CH |
247 | struct request *blk_mq_alloc_request(struct request_queue *q, int rw, |
248 | unsigned int flags) | |
320ae51f | 249 | { |
5a797e00 | 250 | struct blk_mq_alloc_data alloc_data = { .flags = flags }; |
bd166ef1 | 251 | struct request *rq; |
a492f075 | 252 | int ret; |
320ae51f | 253 | |
6f3b0e8b | 254 | ret = blk_queue_enter(q, flags & BLK_MQ_REQ_NOWAIT); |
a492f075 JL |
255 | if (ret) |
256 | return ERR_PTR(ret); | |
320ae51f | 257 | |
bd166ef1 | 258 | rq = blk_mq_sched_get_request(q, NULL, rw, &alloc_data); |
841bac2c | 259 | |
bd166ef1 JA |
260 | blk_mq_put_ctx(alloc_data.ctx); |
261 | blk_queue_exit(q); | |
262 | ||
263 | if (!rq) | |
a492f075 | 264 | return ERR_PTR(-EWOULDBLOCK); |
0c4de0f3 CH |
265 | |
266 | rq->__data_len = 0; | |
267 | rq->__sector = (sector_t) -1; | |
268 | rq->bio = rq->biotail = NULL; | |
320ae51f JA |
269 | return rq; |
270 | } | |
4bb659b1 | 271 | EXPORT_SYMBOL(blk_mq_alloc_request); |
320ae51f | 272 | |
1f5bd336 ML |
273 | struct request *blk_mq_alloc_request_hctx(struct request_queue *q, int rw, |
274 | unsigned int flags, unsigned int hctx_idx) | |
275 | { | |
276 | struct blk_mq_hw_ctx *hctx; | |
277 | struct blk_mq_ctx *ctx; | |
278 | struct request *rq; | |
279 | struct blk_mq_alloc_data alloc_data; | |
280 | int ret; | |
281 | ||
282 | /* | |
283 | * If the tag allocator sleeps we could get an allocation for a | |
284 | * different hardware context. No need to complicate the low level | |
285 | * allocator for this for the rare use case of a command tied to | |
286 | * a specific queue. | |
287 | */ | |
288 | if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT))) | |
289 | return ERR_PTR(-EINVAL); | |
290 | ||
291 | if (hctx_idx >= q->nr_hw_queues) | |
292 | return ERR_PTR(-EIO); | |
293 | ||
294 | ret = blk_queue_enter(q, true); | |
295 | if (ret) | |
296 | return ERR_PTR(ret); | |
297 | ||
c8712c6a CH |
298 | /* |
299 | * Check if the hardware context is actually mapped to anything. | |
300 | * If not tell the caller that it should skip this queue. | |
301 | */ | |
1f5bd336 | 302 | hctx = q->queue_hw_ctx[hctx_idx]; |
c8712c6a CH |
303 | if (!blk_mq_hw_queue_mapped(hctx)) { |
304 | ret = -EXDEV; | |
305 | goto out_queue_exit; | |
306 | } | |
1f5bd336 ML |
307 | ctx = __blk_mq_get_ctx(q, cpumask_first(hctx->cpumask)); |
308 | ||
309 | blk_mq_set_alloc_data(&alloc_data, q, flags, ctx, hctx); | |
ef295ecf | 310 | rq = __blk_mq_alloc_request(&alloc_data, rw); |
1f5bd336 | 311 | if (!rq) { |
c8712c6a CH |
312 | ret = -EWOULDBLOCK; |
313 | goto out_queue_exit; | |
1f5bd336 ML |
314 | } |
315 | ||
316 | return rq; | |
c8712c6a CH |
317 | |
318 | out_queue_exit: | |
319 | blk_queue_exit(q); | |
320 | return ERR_PTR(ret); | |
1f5bd336 ML |
321 | } |
322 | EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx); | |
323 | ||
bd166ef1 JA |
324 | void __blk_mq_finish_request(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx, |
325 | struct request *rq) | |
320ae51f | 326 | { |
bd166ef1 | 327 | const int sched_tag = rq->internal_tag; |
320ae51f JA |
328 | struct request_queue *q = rq->q; |
329 | ||
e8064021 | 330 | if (rq->rq_flags & RQF_MQ_INFLIGHT) |
0d2602ca | 331 | atomic_dec(&hctx->nr_active); |
87760e5e JA |
332 | |
333 | wbt_done(q->rq_wb, &rq->issue_stat); | |
e8064021 | 334 | rq->rq_flags = 0; |
0d2602ca | 335 | |
af76e555 | 336 | clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags); |
06426adf | 337 | clear_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags); |
bd166ef1 JA |
338 | if (rq->tag != -1) |
339 | blk_mq_put_tag(hctx, hctx->tags, ctx, rq->tag); | |
340 | if (sched_tag != -1) | |
341 | blk_mq_sched_completed_request(hctx, rq); | |
50e1dab8 | 342 | blk_mq_sched_restart_queues(hctx); |
3ef28e83 | 343 | blk_queue_exit(q); |
320ae51f JA |
344 | } |
345 | ||
bd166ef1 | 346 | static void blk_mq_finish_hctx_request(struct blk_mq_hw_ctx *hctx, |
16a3c2a7 | 347 | struct request *rq) |
320ae51f JA |
348 | { |
349 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
320ae51f JA |
350 | |
351 | ctx->rq_completed[rq_is_sync(rq)]++; | |
bd166ef1 JA |
352 | __blk_mq_finish_request(hctx, ctx, rq); |
353 | } | |
354 | ||
355 | void blk_mq_finish_request(struct request *rq) | |
356 | { | |
357 | blk_mq_finish_hctx_request(blk_mq_map_queue(rq->q, rq->mq_ctx->cpu), rq); | |
7c7f2f2b | 358 | } |
7c7f2f2b JA |
359 | |
360 | void blk_mq_free_request(struct request *rq) | |
361 | { | |
bd166ef1 | 362 | blk_mq_sched_put_request(rq); |
320ae51f | 363 | } |
1a3b595a | 364 | EXPORT_SYMBOL_GPL(blk_mq_free_request); |
320ae51f | 365 | |
c8a446ad | 366 | inline void __blk_mq_end_request(struct request *rq, int error) |
320ae51f | 367 | { |
0d11e6ac ML |
368 | blk_account_io_done(rq); |
369 | ||
91b63639 | 370 | if (rq->end_io) { |
87760e5e | 371 | wbt_done(rq->q->rq_wb, &rq->issue_stat); |
320ae51f | 372 | rq->end_io(rq, error); |
91b63639 CH |
373 | } else { |
374 | if (unlikely(blk_bidi_rq(rq))) | |
375 | blk_mq_free_request(rq->next_rq); | |
320ae51f | 376 | blk_mq_free_request(rq); |
91b63639 | 377 | } |
320ae51f | 378 | } |
c8a446ad | 379 | EXPORT_SYMBOL(__blk_mq_end_request); |
63151a44 | 380 | |
c8a446ad | 381 | void blk_mq_end_request(struct request *rq, int error) |
63151a44 CH |
382 | { |
383 | if (blk_update_request(rq, error, blk_rq_bytes(rq))) | |
384 | BUG(); | |
c8a446ad | 385 | __blk_mq_end_request(rq, error); |
63151a44 | 386 | } |
c8a446ad | 387 | EXPORT_SYMBOL(blk_mq_end_request); |
320ae51f | 388 | |
30a91cb4 | 389 | static void __blk_mq_complete_request_remote(void *data) |
320ae51f | 390 | { |
3d6efbf6 | 391 | struct request *rq = data; |
320ae51f | 392 | |
30a91cb4 | 393 | rq->q->softirq_done_fn(rq); |
320ae51f | 394 | } |
320ae51f | 395 | |
ed851860 | 396 | static void blk_mq_ipi_complete_request(struct request *rq) |
320ae51f JA |
397 | { |
398 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
38535201 | 399 | bool shared = false; |
320ae51f JA |
400 | int cpu; |
401 | ||
38535201 | 402 | if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) { |
30a91cb4 CH |
403 | rq->q->softirq_done_fn(rq); |
404 | return; | |
405 | } | |
320ae51f JA |
406 | |
407 | cpu = get_cpu(); | |
38535201 CH |
408 | if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags)) |
409 | shared = cpus_share_cache(cpu, ctx->cpu); | |
410 | ||
411 | if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) { | |
30a91cb4 | 412 | rq->csd.func = __blk_mq_complete_request_remote; |
3d6efbf6 CH |
413 | rq->csd.info = rq; |
414 | rq->csd.flags = 0; | |
c46fff2a | 415 | smp_call_function_single_async(ctx->cpu, &rq->csd); |
3d6efbf6 | 416 | } else { |
30a91cb4 | 417 | rq->q->softirq_done_fn(rq); |
3d6efbf6 | 418 | } |
320ae51f JA |
419 | put_cpu(); |
420 | } | |
30a91cb4 | 421 | |
cf43e6be JA |
422 | static void blk_mq_stat_add(struct request *rq) |
423 | { | |
424 | if (rq->rq_flags & RQF_STATS) { | |
425 | /* | |
426 | * We could rq->mq_ctx here, but there's less of a risk | |
427 | * of races if we have the completion event add the stats | |
428 | * to the local software queue. | |
429 | */ | |
430 | struct blk_mq_ctx *ctx; | |
431 | ||
432 | ctx = __blk_mq_get_ctx(rq->q, raw_smp_processor_id()); | |
433 | blk_stat_add(&ctx->stat[rq_data_dir(rq)], rq); | |
434 | } | |
435 | } | |
436 | ||
1fa8cc52 | 437 | static void __blk_mq_complete_request(struct request *rq) |
ed851860 JA |
438 | { |
439 | struct request_queue *q = rq->q; | |
440 | ||
cf43e6be JA |
441 | blk_mq_stat_add(rq); |
442 | ||
ed851860 | 443 | if (!q->softirq_done_fn) |
c8a446ad | 444 | blk_mq_end_request(rq, rq->errors); |
ed851860 JA |
445 | else |
446 | blk_mq_ipi_complete_request(rq); | |
447 | } | |
448 | ||
30a91cb4 CH |
449 | /** |
450 | * blk_mq_complete_request - end I/O on a request | |
451 | * @rq: the request being processed | |
452 | * | |
453 | * Description: | |
454 | * Ends all I/O on a request. It does not handle partial completions. | |
455 | * The actual completion happens out-of-order, through a IPI handler. | |
456 | **/ | |
f4829a9b | 457 | void blk_mq_complete_request(struct request *rq, int error) |
30a91cb4 | 458 | { |
95f09684 JA |
459 | struct request_queue *q = rq->q; |
460 | ||
461 | if (unlikely(blk_should_fake_timeout(q))) | |
30a91cb4 | 462 | return; |
f4829a9b CH |
463 | if (!blk_mark_rq_complete(rq)) { |
464 | rq->errors = error; | |
ed851860 | 465 | __blk_mq_complete_request(rq); |
f4829a9b | 466 | } |
30a91cb4 CH |
467 | } |
468 | EXPORT_SYMBOL(blk_mq_complete_request); | |
320ae51f | 469 | |
973c0191 KB |
470 | int blk_mq_request_started(struct request *rq) |
471 | { | |
472 | return test_bit(REQ_ATOM_STARTED, &rq->atomic_flags); | |
473 | } | |
474 | EXPORT_SYMBOL_GPL(blk_mq_request_started); | |
475 | ||
e2490073 | 476 | void blk_mq_start_request(struct request *rq) |
320ae51f JA |
477 | { |
478 | struct request_queue *q = rq->q; | |
479 | ||
bd166ef1 JA |
480 | blk_mq_sched_started_request(rq); |
481 | ||
320ae51f JA |
482 | trace_block_rq_issue(q, rq); |
483 | ||
cf43e6be JA |
484 | if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) { |
485 | blk_stat_set_issue_time(&rq->issue_stat); | |
486 | rq->rq_flags |= RQF_STATS; | |
87760e5e | 487 | wbt_issue(q->rq_wb, &rq->issue_stat); |
cf43e6be JA |
488 | } |
489 | ||
2b8393b4 | 490 | blk_add_timer(rq); |
87ee7b11 | 491 | |
538b7534 JA |
492 | /* |
493 | * Ensure that ->deadline is visible before set the started | |
494 | * flag and clear the completed flag. | |
495 | */ | |
496 | smp_mb__before_atomic(); | |
497 | ||
87ee7b11 JA |
498 | /* |
499 | * Mark us as started and clear complete. Complete might have been | |
500 | * set if requeue raced with timeout, which then marked it as | |
501 | * complete. So be sure to clear complete again when we start | |
502 | * the request, otherwise we'll ignore the completion event. | |
503 | */ | |
4b570521 JA |
504 | if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) |
505 | set_bit(REQ_ATOM_STARTED, &rq->atomic_flags); | |
506 | if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags)) | |
507 | clear_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags); | |
49f5baa5 CH |
508 | |
509 | if (q->dma_drain_size && blk_rq_bytes(rq)) { | |
510 | /* | |
511 | * Make sure space for the drain appears. We know we can do | |
512 | * this because max_hw_segments has been adjusted to be one | |
513 | * fewer than the device can handle. | |
514 | */ | |
515 | rq->nr_phys_segments++; | |
516 | } | |
320ae51f | 517 | } |
e2490073 | 518 | EXPORT_SYMBOL(blk_mq_start_request); |
320ae51f | 519 | |
ed0791b2 | 520 | static void __blk_mq_requeue_request(struct request *rq) |
320ae51f JA |
521 | { |
522 | struct request_queue *q = rq->q; | |
523 | ||
524 | trace_block_rq_requeue(q, rq); | |
87760e5e | 525 | wbt_requeue(q->rq_wb, &rq->issue_stat); |
bd166ef1 | 526 | blk_mq_sched_requeue_request(rq); |
49f5baa5 | 527 | |
e2490073 CH |
528 | if (test_and_clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) { |
529 | if (q->dma_drain_size && blk_rq_bytes(rq)) | |
530 | rq->nr_phys_segments--; | |
531 | } | |
320ae51f JA |
532 | } |
533 | ||
2b053aca | 534 | void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list) |
ed0791b2 | 535 | { |
ed0791b2 | 536 | __blk_mq_requeue_request(rq); |
ed0791b2 | 537 | |
ed0791b2 | 538 | BUG_ON(blk_queued_rq(rq)); |
2b053aca | 539 | blk_mq_add_to_requeue_list(rq, true, kick_requeue_list); |
ed0791b2 CH |
540 | } |
541 | EXPORT_SYMBOL(blk_mq_requeue_request); | |
542 | ||
6fca6a61 CH |
543 | static void blk_mq_requeue_work(struct work_struct *work) |
544 | { | |
545 | struct request_queue *q = | |
2849450a | 546 | container_of(work, struct request_queue, requeue_work.work); |
6fca6a61 CH |
547 | LIST_HEAD(rq_list); |
548 | struct request *rq, *next; | |
549 | unsigned long flags; | |
550 | ||
551 | spin_lock_irqsave(&q->requeue_lock, flags); | |
552 | list_splice_init(&q->requeue_list, &rq_list); | |
553 | spin_unlock_irqrestore(&q->requeue_lock, flags); | |
554 | ||
555 | list_for_each_entry_safe(rq, next, &rq_list, queuelist) { | |
e8064021 | 556 | if (!(rq->rq_flags & RQF_SOFTBARRIER)) |
6fca6a61 CH |
557 | continue; |
558 | ||
e8064021 | 559 | rq->rq_flags &= ~RQF_SOFTBARRIER; |
6fca6a61 | 560 | list_del_init(&rq->queuelist); |
bd6737f1 | 561 | blk_mq_sched_insert_request(rq, true, false, false, true); |
6fca6a61 CH |
562 | } |
563 | ||
564 | while (!list_empty(&rq_list)) { | |
565 | rq = list_entry(rq_list.next, struct request, queuelist); | |
566 | list_del_init(&rq->queuelist); | |
bd6737f1 | 567 | blk_mq_sched_insert_request(rq, false, false, false, true); |
6fca6a61 CH |
568 | } |
569 | ||
52d7f1b5 | 570 | blk_mq_run_hw_queues(q, false); |
6fca6a61 CH |
571 | } |
572 | ||
2b053aca BVA |
573 | void blk_mq_add_to_requeue_list(struct request *rq, bool at_head, |
574 | bool kick_requeue_list) | |
6fca6a61 CH |
575 | { |
576 | struct request_queue *q = rq->q; | |
577 | unsigned long flags; | |
578 | ||
579 | /* | |
580 | * We abuse this flag that is otherwise used by the I/O scheduler to | |
581 | * request head insertation from the workqueue. | |
582 | */ | |
e8064021 | 583 | BUG_ON(rq->rq_flags & RQF_SOFTBARRIER); |
6fca6a61 CH |
584 | |
585 | spin_lock_irqsave(&q->requeue_lock, flags); | |
586 | if (at_head) { | |
e8064021 | 587 | rq->rq_flags |= RQF_SOFTBARRIER; |
6fca6a61 CH |
588 | list_add(&rq->queuelist, &q->requeue_list); |
589 | } else { | |
590 | list_add_tail(&rq->queuelist, &q->requeue_list); | |
591 | } | |
592 | spin_unlock_irqrestore(&q->requeue_lock, flags); | |
2b053aca BVA |
593 | |
594 | if (kick_requeue_list) | |
595 | blk_mq_kick_requeue_list(q); | |
6fca6a61 CH |
596 | } |
597 | EXPORT_SYMBOL(blk_mq_add_to_requeue_list); | |
598 | ||
599 | void blk_mq_kick_requeue_list(struct request_queue *q) | |
600 | { | |
2849450a | 601 | kblockd_schedule_delayed_work(&q->requeue_work, 0); |
6fca6a61 CH |
602 | } |
603 | EXPORT_SYMBOL(blk_mq_kick_requeue_list); | |
604 | ||
2849450a MS |
605 | void blk_mq_delay_kick_requeue_list(struct request_queue *q, |
606 | unsigned long msecs) | |
607 | { | |
608 | kblockd_schedule_delayed_work(&q->requeue_work, | |
609 | msecs_to_jiffies(msecs)); | |
610 | } | |
611 | EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list); | |
612 | ||
1885b24d JA |
613 | void blk_mq_abort_requeue_list(struct request_queue *q) |
614 | { | |
615 | unsigned long flags; | |
616 | LIST_HEAD(rq_list); | |
617 | ||
618 | spin_lock_irqsave(&q->requeue_lock, flags); | |
619 | list_splice_init(&q->requeue_list, &rq_list); | |
620 | spin_unlock_irqrestore(&q->requeue_lock, flags); | |
621 | ||
622 | while (!list_empty(&rq_list)) { | |
623 | struct request *rq; | |
624 | ||
625 | rq = list_first_entry(&rq_list, struct request, queuelist); | |
626 | list_del_init(&rq->queuelist); | |
627 | rq->errors = -EIO; | |
628 | blk_mq_end_request(rq, rq->errors); | |
629 | } | |
630 | } | |
631 | EXPORT_SYMBOL(blk_mq_abort_requeue_list); | |
632 | ||
0e62f51f JA |
633 | struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag) |
634 | { | |
88c7b2b7 JA |
635 | if (tag < tags->nr_tags) { |
636 | prefetch(tags->rqs[tag]); | |
4ee86bab | 637 | return tags->rqs[tag]; |
88c7b2b7 | 638 | } |
4ee86bab HR |
639 | |
640 | return NULL; | |
24d2f903 CH |
641 | } |
642 | EXPORT_SYMBOL(blk_mq_tag_to_rq); | |
643 | ||
320ae51f | 644 | struct blk_mq_timeout_data { |
46f92d42 CH |
645 | unsigned long next; |
646 | unsigned int next_set; | |
320ae51f JA |
647 | }; |
648 | ||
90415837 | 649 | void blk_mq_rq_timed_out(struct request *req, bool reserved) |
320ae51f | 650 | { |
f8a5b122 | 651 | const struct blk_mq_ops *ops = req->q->mq_ops; |
46f92d42 | 652 | enum blk_eh_timer_return ret = BLK_EH_RESET_TIMER; |
87ee7b11 JA |
653 | |
654 | /* | |
655 | * We know that complete is set at this point. If STARTED isn't set | |
656 | * anymore, then the request isn't active and the "timeout" should | |
657 | * just be ignored. This can happen due to the bitflag ordering. | |
658 | * Timeout first checks if STARTED is set, and if it is, assumes | |
659 | * the request is active. But if we race with completion, then | |
660 | * we both flags will get cleared. So check here again, and ignore | |
661 | * a timeout event with a request that isn't active. | |
662 | */ | |
46f92d42 CH |
663 | if (!test_bit(REQ_ATOM_STARTED, &req->atomic_flags)) |
664 | return; | |
87ee7b11 | 665 | |
46f92d42 | 666 | if (ops->timeout) |
0152fb6b | 667 | ret = ops->timeout(req, reserved); |
46f92d42 CH |
668 | |
669 | switch (ret) { | |
670 | case BLK_EH_HANDLED: | |
671 | __blk_mq_complete_request(req); | |
672 | break; | |
673 | case BLK_EH_RESET_TIMER: | |
674 | blk_add_timer(req); | |
675 | blk_clear_rq_complete(req); | |
676 | break; | |
677 | case BLK_EH_NOT_HANDLED: | |
678 | break; | |
679 | default: | |
680 | printk(KERN_ERR "block: bad eh return: %d\n", ret); | |
681 | break; | |
682 | } | |
87ee7b11 | 683 | } |
5b3f25fc | 684 | |
81481eb4 CH |
685 | static void blk_mq_check_expired(struct blk_mq_hw_ctx *hctx, |
686 | struct request *rq, void *priv, bool reserved) | |
687 | { | |
688 | struct blk_mq_timeout_data *data = priv; | |
87ee7b11 | 689 | |
eb130dbf KB |
690 | if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) { |
691 | /* | |
692 | * If a request wasn't started before the queue was | |
693 | * marked dying, kill it here or it'll go unnoticed. | |
694 | */ | |
a59e0f57 KB |
695 | if (unlikely(blk_queue_dying(rq->q))) { |
696 | rq->errors = -EIO; | |
697 | blk_mq_end_request(rq, rq->errors); | |
698 | } | |
46f92d42 | 699 | return; |
eb130dbf | 700 | } |
87ee7b11 | 701 | |
46f92d42 CH |
702 | if (time_after_eq(jiffies, rq->deadline)) { |
703 | if (!blk_mark_rq_complete(rq)) | |
0152fb6b | 704 | blk_mq_rq_timed_out(rq, reserved); |
46f92d42 CH |
705 | } else if (!data->next_set || time_after(data->next, rq->deadline)) { |
706 | data->next = rq->deadline; | |
707 | data->next_set = 1; | |
708 | } | |
87ee7b11 JA |
709 | } |
710 | ||
287922eb | 711 | static void blk_mq_timeout_work(struct work_struct *work) |
320ae51f | 712 | { |
287922eb CH |
713 | struct request_queue *q = |
714 | container_of(work, struct request_queue, timeout_work); | |
81481eb4 CH |
715 | struct blk_mq_timeout_data data = { |
716 | .next = 0, | |
717 | .next_set = 0, | |
718 | }; | |
81481eb4 | 719 | int i; |
320ae51f | 720 | |
71f79fb3 GKB |
721 | /* A deadlock might occur if a request is stuck requiring a |
722 | * timeout at the same time a queue freeze is waiting | |
723 | * completion, since the timeout code would not be able to | |
724 | * acquire the queue reference here. | |
725 | * | |
726 | * That's why we don't use blk_queue_enter here; instead, we use | |
727 | * percpu_ref_tryget directly, because we need to be able to | |
728 | * obtain a reference even in the short window between the queue | |
729 | * starting to freeze, by dropping the first reference in | |
730 | * blk_mq_freeze_queue_start, and the moment the last request is | |
731 | * consumed, marked by the instant q_usage_counter reaches | |
732 | * zero. | |
733 | */ | |
734 | if (!percpu_ref_tryget(&q->q_usage_counter)) | |
287922eb CH |
735 | return; |
736 | ||
0bf6cd5b | 737 | blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &data); |
320ae51f | 738 | |
81481eb4 CH |
739 | if (data.next_set) { |
740 | data.next = blk_rq_timeout(round_jiffies_up(data.next)); | |
741 | mod_timer(&q->timeout, data.next); | |
0d2602ca | 742 | } else { |
0bf6cd5b CH |
743 | struct blk_mq_hw_ctx *hctx; |
744 | ||
f054b56c ML |
745 | queue_for_each_hw_ctx(q, hctx, i) { |
746 | /* the hctx may be unmapped, so check it here */ | |
747 | if (blk_mq_hw_queue_mapped(hctx)) | |
748 | blk_mq_tag_idle(hctx); | |
749 | } | |
0d2602ca | 750 | } |
287922eb | 751 | blk_queue_exit(q); |
320ae51f JA |
752 | } |
753 | ||
754 | /* | |
755 | * Reverse check our software queue for entries that we could potentially | |
756 | * merge with. Currently includes a hand-wavy stop count of 8, to not spend | |
757 | * too much time checking for merges. | |
758 | */ | |
759 | static bool blk_mq_attempt_merge(struct request_queue *q, | |
760 | struct blk_mq_ctx *ctx, struct bio *bio) | |
761 | { | |
762 | struct request *rq; | |
763 | int checked = 8; | |
764 | ||
765 | list_for_each_entry_reverse(rq, &ctx->rq_list, queuelist) { | |
34fe7c05 | 766 | bool merged = false; |
320ae51f JA |
767 | |
768 | if (!checked--) | |
769 | break; | |
770 | ||
771 | if (!blk_rq_merge_ok(rq, bio)) | |
772 | continue; | |
773 | ||
34fe7c05 CH |
774 | switch (blk_try_merge(rq, bio)) { |
775 | case ELEVATOR_BACK_MERGE: | |
776 | if (blk_mq_sched_allow_merge(q, rq, bio)) | |
777 | merged = bio_attempt_back_merge(q, rq, bio); | |
bd166ef1 | 778 | break; |
34fe7c05 CH |
779 | case ELEVATOR_FRONT_MERGE: |
780 | if (blk_mq_sched_allow_merge(q, rq, bio)) | |
781 | merged = bio_attempt_front_merge(q, rq, bio); | |
320ae51f | 782 | break; |
1e739730 CH |
783 | case ELEVATOR_DISCARD_MERGE: |
784 | merged = bio_attempt_discard_merge(q, rq, bio); | |
320ae51f | 785 | break; |
34fe7c05 CH |
786 | default: |
787 | continue; | |
320ae51f | 788 | } |
34fe7c05 CH |
789 | |
790 | if (merged) | |
791 | ctx->rq_merged++; | |
792 | return merged; | |
320ae51f JA |
793 | } |
794 | ||
795 | return false; | |
796 | } | |
797 | ||
88459642 OS |
798 | struct flush_busy_ctx_data { |
799 | struct blk_mq_hw_ctx *hctx; | |
800 | struct list_head *list; | |
801 | }; | |
802 | ||
803 | static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data) | |
804 | { | |
805 | struct flush_busy_ctx_data *flush_data = data; | |
806 | struct blk_mq_hw_ctx *hctx = flush_data->hctx; | |
807 | struct blk_mq_ctx *ctx = hctx->ctxs[bitnr]; | |
808 | ||
809 | sbitmap_clear_bit(sb, bitnr); | |
810 | spin_lock(&ctx->lock); | |
811 | list_splice_tail_init(&ctx->rq_list, flush_data->list); | |
812 | spin_unlock(&ctx->lock); | |
813 | return true; | |
814 | } | |
815 | ||
1429d7c9 JA |
816 | /* |
817 | * Process software queues that have been marked busy, splicing them | |
818 | * to the for-dispatch | |
819 | */ | |
2c3ad667 | 820 | void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list) |
1429d7c9 | 821 | { |
88459642 OS |
822 | struct flush_busy_ctx_data data = { |
823 | .hctx = hctx, | |
824 | .list = list, | |
825 | }; | |
1429d7c9 | 826 | |
88459642 | 827 | sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data); |
1429d7c9 | 828 | } |
2c3ad667 | 829 | EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs); |
1429d7c9 | 830 | |
703fd1c0 JA |
831 | static inline unsigned int queued_to_index(unsigned int queued) |
832 | { | |
833 | if (!queued) | |
834 | return 0; | |
1429d7c9 | 835 | |
703fd1c0 | 836 | return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1); |
1429d7c9 JA |
837 | } |
838 | ||
bd6737f1 JA |
839 | bool blk_mq_get_driver_tag(struct request *rq, struct blk_mq_hw_ctx **hctx, |
840 | bool wait) | |
bd166ef1 JA |
841 | { |
842 | struct blk_mq_alloc_data data = { | |
843 | .q = rq->q, | |
bd166ef1 JA |
844 | .hctx = blk_mq_map_queue(rq->q, rq->mq_ctx->cpu), |
845 | .flags = wait ? 0 : BLK_MQ_REQ_NOWAIT, | |
846 | }; | |
847 | ||
bd166ef1 JA |
848 | if (rq->tag != -1) { |
849 | done: | |
850 | if (hctx) | |
851 | *hctx = data.hctx; | |
852 | return true; | |
853 | } | |
854 | ||
855 | rq->tag = blk_mq_get_tag(&data); | |
856 | if (rq->tag >= 0) { | |
200e86b3 JA |
857 | if (blk_mq_tag_busy(data.hctx)) { |
858 | rq->rq_flags |= RQF_MQ_INFLIGHT; | |
859 | atomic_inc(&data.hctx->nr_active); | |
860 | } | |
bd166ef1 JA |
861 | data.hctx->tags->rqs[rq->tag] = rq; |
862 | goto done; | |
863 | } | |
864 | ||
865 | return false; | |
866 | } | |
867 | ||
99cf1dc5 JA |
868 | static void blk_mq_put_driver_tag(struct blk_mq_hw_ctx *hctx, |
869 | struct request *rq) | |
870 | { | |
871 | if (rq->tag == -1 || rq->internal_tag == -1) | |
872 | return; | |
873 | ||
874 | blk_mq_put_tag(hctx, hctx->tags, rq->mq_ctx, rq->tag); | |
875 | rq->tag = -1; | |
876 | ||
877 | if (rq->rq_flags & RQF_MQ_INFLIGHT) { | |
878 | rq->rq_flags &= ~RQF_MQ_INFLIGHT; | |
879 | atomic_dec(&hctx->nr_active); | |
880 | } | |
881 | } | |
882 | ||
bd166ef1 JA |
883 | /* |
884 | * If we fail getting a driver tag because all the driver tags are already | |
885 | * assigned and on the dispatch list, BUT the first entry does not have a | |
886 | * tag, then we could deadlock. For that case, move entries with assigned | |
887 | * driver tags to the front, leaving the set of tagged requests in the | |
888 | * same order, and the untagged set in the same order. | |
889 | */ | |
890 | static bool reorder_tags_to_front(struct list_head *list) | |
891 | { | |
892 | struct request *rq, *tmp, *first = NULL; | |
893 | ||
894 | list_for_each_entry_safe_reverse(rq, tmp, list, queuelist) { | |
895 | if (rq == first) | |
896 | break; | |
897 | if (rq->tag != -1) { | |
898 | list_move(&rq->queuelist, list); | |
899 | if (!first) | |
900 | first = rq; | |
901 | } | |
902 | } | |
903 | ||
904 | return first != NULL; | |
905 | } | |
906 | ||
f04c3df3 | 907 | bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx *hctx, struct list_head *list) |
320ae51f JA |
908 | { |
909 | struct request_queue *q = hctx->queue; | |
320ae51f | 910 | struct request *rq; |
74c45052 JA |
911 | LIST_HEAD(driver_list); |
912 | struct list_head *dptr; | |
f04c3df3 | 913 | int queued, ret = BLK_MQ_RQ_QUEUE_OK; |
320ae51f | 914 | |
74c45052 JA |
915 | /* |
916 | * Start off with dptr being NULL, so we start the first request | |
917 | * immediately, even if we have more pending. | |
918 | */ | |
919 | dptr = NULL; | |
920 | ||
320ae51f JA |
921 | /* |
922 | * Now process all the entries, sending them to the driver. | |
923 | */ | |
1429d7c9 | 924 | queued = 0; |
f04c3df3 | 925 | while (!list_empty(list)) { |
74c45052 | 926 | struct blk_mq_queue_data bd; |
320ae51f | 927 | |
f04c3df3 | 928 | rq = list_first_entry(list, struct request, queuelist); |
bd166ef1 JA |
929 | if (!blk_mq_get_driver_tag(rq, &hctx, false)) { |
930 | if (!queued && reorder_tags_to_front(list)) | |
931 | continue; | |
3c782d67 JA |
932 | |
933 | /* | |
934 | * We failed getting a driver tag. Mark the queue(s) | |
935 | * as needing a restart. Retry getting a tag again, | |
936 | * in case the needed IO completed right before we | |
937 | * marked the queue as needing a restart. | |
938 | */ | |
bd166ef1 | 939 | blk_mq_sched_mark_restart(hctx); |
3c782d67 JA |
940 | if (!blk_mq_get_driver_tag(rq, &hctx, false)) |
941 | break; | |
bd166ef1 | 942 | } |
320ae51f | 943 | list_del_init(&rq->queuelist); |
320ae51f | 944 | |
74c45052 JA |
945 | bd.rq = rq; |
946 | bd.list = dptr; | |
f04c3df3 | 947 | bd.last = list_empty(list); |
74c45052 JA |
948 | |
949 | ret = q->mq_ops->queue_rq(hctx, &bd); | |
320ae51f JA |
950 | switch (ret) { |
951 | case BLK_MQ_RQ_QUEUE_OK: | |
952 | queued++; | |
52b9c330 | 953 | break; |
320ae51f | 954 | case BLK_MQ_RQ_QUEUE_BUSY: |
99cf1dc5 | 955 | blk_mq_put_driver_tag(hctx, rq); |
f04c3df3 | 956 | list_add(&rq->queuelist, list); |
ed0791b2 | 957 | __blk_mq_requeue_request(rq); |
320ae51f JA |
958 | break; |
959 | default: | |
960 | pr_err("blk-mq: bad return on queue: %d\n", ret); | |
320ae51f | 961 | case BLK_MQ_RQ_QUEUE_ERROR: |
1e93b8c2 | 962 | rq->errors = -EIO; |
c8a446ad | 963 | blk_mq_end_request(rq, rq->errors); |
320ae51f JA |
964 | break; |
965 | } | |
966 | ||
967 | if (ret == BLK_MQ_RQ_QUEUE_BUSY) | |
968 | break; | |
74c45052 JA |
969 | |
970 | /* | |
971 | * We've done the first request. If we have more than 1 | |
972 | * left in the list, set dptr to defer issue. | |
973 | */ | |
f04c3df3 | 974 | if (!dptr && list->next != list->prev) |
74c45052 | 975 | dptr = &driver_list; |
320ae51f JA |
976 | } |
977 | ||
703fd1c0 | 978 | hctx->dispatched[queued_to_index(queued)]++; |
320ae51f JA |
979 | |
980 | /* | |
981 | * Any items that need requeuing? Stuff them into hctx->dispatch, | |
982 | * that is where we will continue on next queue run. | |
983 | */ | |
f04c3df3 | 984 | if (!list_empty(list)) { |
320ae51f | 985 | spin_lock(&hctx->lock); |
c13660a0 | 986 | list_splice_init(list, &hctx->dispatch); |
320ae51f | 987 | spin_unlock(&hctx->lock); |
f04c3df3 | 988 | |
9ba52e58 SL |
989 | /* |
990 | * the queue is expected stopped with BLK_MQ_RQ_QUEUE_BUSY, but | |
991 | * it's possible the queue is stopped and restarted again | |
992 | * before this. Queue restart will dispatch requests. And since | |
993 | * requests in rq_list aren't added into hctx->dispatch yet, | |
994 | * the requests in rq_list might get lost. | |
995 | * | |
996 | * blk_mq_run_hw_queue() already checks the STOPPED bit | |
bd166ef1 JA |
997 | * |
998 | * If RESTART is set, then let completion restart the queue | |
999 | * instead of potentially looping here. | |
1000 | */ | |
1001 | if (!blk_mq_sched_needs_restart(hctx)) | |
1002 | blk_mq_run_hw_queue(hctx, true); | |
320ae51f | 1003 | } |
f04c3df3 | 1004 | |
2aa0f21d | 1005 | return queued != 0; |
f04c3df3 JA |
1006 | } |
1007 | ||
6a83e74d BVA |
1008 | static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx) |
1009 | { | |
1010 | int srcu_idx; | |
1011 | ||
1012 | WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) && | |
1013 | cpu_online(hctx->next_cpu)); | |
1014 | ||
1015 | if (!(hctx->flags & BLK_MQ_F_BLOCKING)) { | |
1016 | rcu_read_lock(); | |
bd166ef1 | 1017 | blk_mq_sched_dispatch_requests(hctx); |
6a83e74d BVA |
1018 | rcu_read_unlock(); |
1019 | } else { | |
1020 | srcu_idx = srcu_read_lock(&hctx->queue_rq_srcu); | |
bd166ef1 | 1021 | blk_mq_sched_dispatch_requests(hctx); |
6a83e74d BVA |
1022 | srcu_read_unlock(&hctx->queue_rq_srcu, srcu_idx); |
1023 | } | |
1024 | } | |
1025 | ||
506e931f JA |
1026 | /* |
1027 | * It'd be great if the workqueue API had a way to pass | |
1028 | * in a mask and had some smarts for more clever placement. | |
1029 | * For now we just round-robin here, switching for every | |
1030 | * BLK_MQ_CPU_WORK_BATCH queued items. | |
1031 | */ | |
1032 | static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx) | |
1033 | { | |
b657d7e6 CH |
1034 | if (hctx->queue->nr_hw_queues == 1) |
1035 | return WORK_CPU_UNBOUND; | |
506e931f JA |
1036 | |
1037 | if (--hctx->next_cpu_batch <= 0) { | |
c02ebfdd | 1038 | int next_cpu; |
506e931f JA |
1039 | |
1040 | next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask); | |
1041 | if (next_cpu >= nr_cpu_ids) | |
1042 | next_cpu = cpumask_first(hctx->cpumask); | |
1043 | ||
1044 | hctx->next_cpu = next_cpu; | |
1045 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; | |
1046 | } | |
1047 | ||
b657d7e6 | 1048 | return hctx->next_cpu; |
506e931f JA |
1049 | } |
1050 | ||
320ae51f JA |
1051 | void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) |
1052 | { | |
5d1b25c1 BVA |
1053 | if (unlikely(blk_mq_hctx_stopped(hctx) || |
1054 | !blk_mq_hw_queue_mapped(hctx))) | |
320ae51f JA |
1055 | return; |
1056 | ||
1b792f2f | 1057 | if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) { |
2a90d4aa PB |
1058 | int cpu = get_cpu(); |
1059 | if (cpumask_test_cpu(cpu, hctx->cpumask)) { | |
398205b8 | 1060 | __blk_mq_run_hw_queue(hctx); |
2a90d4aa | 1061 | put_cpu(); |
398205b8 PB |
1062 | return; |
1063 | } | |
e4043dcf | 1064 | |
2a90d4aa | 1065 | put_cpu(); |
e4043dcf | 1066 | } |
398205b8 | 1067 | |
27489a3c | 1068 | kblockd_schedule_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work); |
320ae51f JA |
1069 | } |
1070 | ||
b94ec296 | 1071 | void blk_mq_run_hw_queues(struct request_queue *q, bool async) |
320ae51f JA |
1072 | { |
1073 | struct blk_mq_hw_ctx *hctx; | |
1074 | int i; | |
1075 | ||
1076 | queue_for_each_hw_ctx(q, hctx, i) { | |
bd166ef1 | 1077 | if (!blk_mq_hctx_has_pending(hctx) || |
5d1b25c1 | 1078 | blk_mq_hctx_stopped(hctx)) |
320ae51f JA |
1079 | continue; |
1080 | ||
b94ec296 | 1081 | blk_mq_run_hw_queue(hctx, async); |
320ae51f JA |
1082 | } |
1083 | } | |
b94ec296 | 1084 | EXPORT_SYMBOL(blk_mq_run_hw_queues); |
320ae51f | 1085 | |
fd001443 BVA |
1086 | /** |
1087 | * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped | |
1088 | * @q: request queue. | |
1089 | * | |
1090 | * The caller is responsible for serializing this function against | |
1091 | * blk_mq_{start,stop}_hw_queue(). | |
1092 | */ | |
1093 | bool blk_mq_queue_stopped(struct request_queue *q) | |
1094 | { | |
1095 | struct blk_mq_hw_ctx *hctx; | |
1096 | int i; | |
1097 | ||
1098 | queue_for_each_hw_ctx(q, hctx, i) | |
1099 | if (blk_mq_hctx_stopped(hctx)) | |
1100 | return true; | |
1101 | ||
1102 | return false; | |
1103 | } | |
1104 | EXPORT_SYMBOL(blk_mq_queue_stopped); | |
1105 | ||
320ae51f JA |
1106 | void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx) |
1107 | { | |
27489a3c | 1108 | cancel_work(&hctx->run_work); |
70f4db63 | 1109 | cancel_delayed_work(&hctx->delay_work); |
320ae51f JA |
1110 | set_bit(BLK_MQ_S_STOPPED, &hctx->state); |
1111 | } | |
1112 | EXPORT_SYMBOL(blk_mq_stop_hw_queue); | |
1113 | ||
280d45f6 CH |
1114 | void blk_mq_stop_hw_queues(struct request_queue *q) |
1115 | { | |
1116 | struct blk_mq_hw_ctx *hctx; | |
1117 | int i; | |
1118 | ||
1119 | queue_for_each_hw_ctx(q, hctx, i) | |
1120 | blk_mq_stop_hw_queue(hctx); | |
1121 | } | |
1122 | EXPORT_SYMBOL(blk_mq_stop_hw_queues); | |
1123 | ||
320ae51f JA |
1124 | void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx) |
1125 | { | |
1126 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
e4043dcf | 1127 | |
0ffbce80 | 1128 | blk_mq_run_hw_queue(hctx, false); |
320ae51f JA |
1129 | } |
1130 | EXPORT_SYMBOL(blk_mq_start_hw_queue); | |
1131 | ||
2f268556 CH |
1132 | void blk_mq_start_hw_queues(struct request_queue *q) |
1133 | { | |
1134 | struct blk_mq_hw_ctx *hctx; | |
1135 | int i; | |
1136 | ||
1137 | queue_for_each_hw_ctx(q, hctx, i) | |
1138 | blk_mq_start_hw_queue(hctx); | |
1139 | } | |
1140 | EXPORT_SYMBOL(blk_mq_start_hw_queues); | |
1141 | ||
ae911c5e JA |
1142 | void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) |
1143 | { | |
1144 | if (!blk_mq_hctx_stopped(hctx)) | |
1145 | return; | |
1146 | ||
1147 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
1148 | blk_mq_run_hw_queue(hctx, async); | |
1149 | } | |
1150 | EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue); | |
1151 | ||
1b4a3258 | 1152 | void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async) |
320ae51f JA |
1153 | { |
1154 | struct blk_mq_hw_ctx *hctx; | |
1155 | int i; | |
1156 | ||
ae911c5e JA |
1157 | queue_for_each_hw_ctx(q, hctx, i) |
1158 | blk_mq_start_stopped_hw_queue(hctx, async); | |
320ae51f JA |
1159 | } |
1160 | EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues); | |
1161 | ||
70f4db63 | 1162 | static void blk_mq_run_work_fn(struct work_struct *work) |
320ae51f JA |
1163 | { |
1164 | struct blk_mq_hw_ctx *hctx; | |
1165 | ||
27489a3c | 1166 | hctx = container_of(work, struct blk_mq_hw_ctx, run_work); |
e4043dcf | 1167 | |
320ae51f JA |
1168 | __blk_mq_run_hw_queue(hctx); |
1169 | } | |
1170 | ||
70f4db63 CH |
1171 | static void blk_mq_delay_work_fn(struct work_struct *work) |
1172 | { | |
1173 | struct blk_mq_hw_ctx *hctx; | |
1174 | ||
1175 | hctx = container_of(work, struct blk_mq_hw_ctx, delay_work.work); | |
1176 | ||
1177 | if (test_and_clear_bit(BLK_MQ_S_STOPPED, &hctx->state)) | |
1178 | __blk_mq_run_hw_queue(hctx); | |
1179 | } | |
1180 | ||
1181 | void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs) | |
1182 | { | |
19c66e59 ML |
1183 | if (unlikely(!blk_mq_hw_queue_mapped(hctx))) |
1184 | return; | |
70f4db63 | 1185 | |
7e79dadc | 1186 | blk_mq_stop_hw_queue(hctx); |
b657d7e6 CH |
1187 | kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx), |
1188 | &hctx->delay_work, msecs_to_jiffies(msecs)); | |
70f4db63 CH |
1189 | } |
1190 | EXPORT_SYMBOL(blk_mq_delay_queue); | |
1191 | ||
cfd0c552 | 1192 | static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx, |
cfd0c552 ML |
1193 | struct request *rq, |
1194 | bool at_head) | |
320ae51f | 1195 | { |
e57690fe JA |
1196 | struct blk_mq_ctx *ctx = rq->mq_ctx; |
1197 | ||
01b983c9 JA |
1198 | trace_block_rq_insert(hctx->queue, rq); |
1199 | ||
72a0a36e CH |
1200 | if (at_head) |
1201 | list_add(&rq->queuelist, &ctx->rq_list); | |
1202 | else | |
1203 | list_add_tail(&rq->queuelist, &ctx->rq_list); | |
cfd0c552 | 1204 | } |
4bb659b1 | 1205 | |
2c3ad667 JA |
1206 | void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq, |
1207 | bool at_head) | |
cfd0c552 ML |
1208 | { |
1209 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
1210 | ||
e57690fe | 1211 | __blk_mq_insert_req_list(hctx, rq, at_head); |
320ae51f | 1212 | blk_mq_hctx_mark_pending(hctx, ctx); |
320ae51f JA |
1213 | } |
1214 | ||
bd166ef1 JA |
1215 | void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx, |
1216 | struct list_head *list) | |
320ae51f JA |
1217 | |
1218 | { | |
320ae51f JA |
1219 | /* |
1220 | * preemption doesn't flush plug list, so it's possible ctx->cpu is | |
1221 | * offline now | |
1222 | */ | |
1223 | spin_lock(&ctx->lock); | |
1224 | while (!list_empty(list)) { | |
1225 | struct request *rq; | |
1226 | ||
1227 | rq = list_first_entry(list, struct request, queuelist); | |
e57690fe | 1228 | BUG_ON(rq->mq_ctx != ctx); |
320ae51f | 1229 | list_del_init(&rq->queuelist); |
e57690fe | 1230 | __blk_mq_insert_req_list(hctx, rq, false); |
320ae51f | 1231 | } |
cfd0c552 | 1232 | blk_mq_hctx_mark_pending(hctx, ctx); |
320ae51f | 1233 | spin_unlock(&ctx->lock); |
320ae51f JA |
1234 | } |
1235 | ||
1236 | static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b) | |
1237 | { | |
1238 | struct request *rqa = container_of(a, struct request, queuelist); | |
1239 | struct request *rqb = container_of(b, struct request, queuelist); | |
1240 | ||
1241 | return !(rqa->mq_ctx < rqb->mq_ctx || | |
1242 | (rqa->mq_ctx == rqb->mq_ctx && | |
1243 | blk_rq_pos(rqa) < blk_rq_pos(rqb))); | |
1244 | } | |
1245 | ||
1246 | void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule) | |
1247 | { | |
1248 | struct blk_mq_ctx *this_ctx; | |
1249 | struct request_queue *this_q; | |
1250 | struct request *rq; | |
1251 | LIST_HEAD(list); | |
1252 | LIST_HEAD(ctx_list); | |
1253 | unsigned int depth; | |
1254 | ||
1255 | list_splice_init(&plug->mq_list, &list); | |
1256 | ||
1257 | list_sort(NULL, &list, plug_ctx_cmp); | |
1258 | ||
1259 | this_q = NULL; | |
1260 | this_ctx = NULL; | |
1261 | depth = 0; | |
1262 | ||
1263 | while (!list_empty(&list)) { | |
1264 | rq = list_entry_rq(list.next); | |
1265 | list_del_init(&rq->queuelist); | |
1266 | BUG_ON(!rq->q); | |
1267 | if (rq->mq_ctx != this_ctx) { | |
1268 | if (this_ctx) { | |
bd166ef1 JA |
1269 | trace_block_unplug(this_q, depth, from_schedule); |
1270 | blk_mq_sched_insert_requests(this_q, this_ctx, | |
1271 | &ctx_list, | |
1272 | from_schedule); | |
320ae51f JA |
1273 | } |
1274 | ||
1275 | this_ctx = rq->mq_ctx; | |
1276 | this_q = rq->q; | |
1277 | depth = 0; | |
1278 | } | |
1279 | ||
1280 | depth++; | |
1281 | list_add_tail(&rq->queuelist, &ctx_list); | |
1282 | } | |
1283 | ||
1284 | /* | |
1285 | * If 'this_ctx' is set, we know we have entries to complete | |
1286 | * on 'ctx_list'. Do those. | |
1287 | */ | |
1288 | if (this_ctx) { | |
bd166ef1 JA |
1289 | trace_block_unplug(this_q, depth, from_schedule); |
1290 | blk_mq_sched_insert_requests(this_q, this_ctx, &ctx_list, | |
1291 | from_schedule); | |
320ae51f JA |
1292 | } |
1293 | } | |
1294 | ||
1295 | static void blk_mq_bio_to_request(struct request *rq, struct bio *bio) | |
1296 | { | |
1297 | init_request_from_bio(rq, bio); | |
4b570521 | 1298 | |
6e85eaf3 | 1299 | blk_account_io_start(rq, true); |
320ae51f JA |
1300 | } |
1301 | ||
274a5843 JA |
1302 | static inline bool hctx_allow_merges(struct blk_mq_hw_ctx *hctx) |
1303 | { | |
1304 | return (hctx->flags & BLK_MQ_F_SHOULD_MERGE) && | |
1305 | !blk_queue_nomerges(hctx->queue); | |
1306 | } | |
1307 | ||
07068d5b JA |
1308 | static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx *hctx, |
1309 | struct blk_mq_ctx *ctx, | |
1310 | struct request *rq, struct bio *bio) | |
320ae51f | 1311 | { |
e18378a6 | 1312 | if (!hctx_allow_merges(hctx) || !bio_mergeable(bio)) { |
07068d5b JA |
1313 | blk_mq_bio_to_request(rq, bio); |
1314 | spin_lock(&ctx->lock); | |
1315 | insert_rq: | |
1316 | __blk_mq_insert_request(hctx, rq, false); | |
1317 | spin_unlock(&ctx->lock); | |
1318 | return false; | |
1319 | } else { | |
274a5843 JA |
1320 | struct request_queue *q = hctx->queue; |
1321 | ||
07068d5b JA |
1322 | spin_lock(&ctx->lock); |
1323 | if (!blk_mq_attempt_merge(q, ctx, bio)) { | |
1324 | blk_mq_bio_to_request(rq, bio); | |
1325 | goto insert_rq; | |
1326 | } | |
320ae51f | 1327 | |
07068d5b | 1328 | spin_unlock(&ctx->lock); |
bd166ef1 | 1329 | __blk_mq_finish_request(hctx, ctx, rq); |
07068d5b | 1330 | return true; |
14ec77f3 | 1331 | } |
07068d5b | 1332 | } |
14ec77f3 | 1333 | |
fd2d3326 JA |
1334 | static blk_qc_t request_to_qc_t(struct blk_mq_hw_ctx *hctx, struct request *rq) |
1335 | { | |
bd166ef1 JA |
1336 | if (rq->tag != -1) |
1337 | return blk_tag_to_qc_t(rq->tag, hctx->queue_num, false); | |
1338 | ||
1339 | return blk_tag_to_qc_t(rq->internal_tag, hctx->queue_num, true); | |
fd2d3326 JA |
1340 | } |
1341 | ||
066a4a73 | 1342 | static void blk_mq_try_issue_directly(struct request *rq, blk_qc_t *cookie) |
f984df1f | 1343 | { |
f984df1f | 1344 | struct request_queue *q = rq->q; |
f984df1f SL |
1345 | struct blk_mq_queue_data bd = { |
1346 | .rq = rq, | |
1347 | .list = NULL, | |
1348 | .last = 1 | |
1349 | }; | |
bd166ef1 JA |
1350 | struct blk_mq_hw_ctx *hctx; |
1351 | blk_qc_t new_cookie; | |
1352 | int ret; | |
f984df1f | 1353 | |
bd166ef1 | 1354 | if (q->elevator) |
2253efc8 BVA |
1355 | goto insert; |
1356 | ||
bd166ef1 JA |
1357 | if (!blk_mq_get_driver_tag(rq, &hctx, false)) |
1358 | goto insert; | |
1359 | ||
1360 | new_cookie = request_to_qc_t(hctx, rq); | |
1361 | ||
f984df1f SL |
1362 | /* |
1363 | * For OK queue, we are done. For error, kill it. Any other | |
1364 | * error (busy), just add it to our list as we previously | |
1365 | * would have done | |
1366 | */ | |
1367 | ret = q->mq_ops->queue_rq(hctx, &bd); | |
7b371636 JA |
1368 | if (ret == BLK_MQ_RQ_QUEUE_OK) { |
1369 | *cookie = new_cookie; | |
2253efc8 | 1370 | return; |
7b371636 | 1371 | } |
f984df1f | 1372 | |
7b371636 JA |
1373 | __blk_mq_requeue_request(rq); |
1374 | ||
1375 | if (ret == BLK_MQ_RQ_QUEUE_ERROR) { | |
1376 | *cookie = BLK_QC_T_NONE; | |
1377 | rq->errors = -EIO; | |
1378 | blk_mq_end_request(rq, rq->errors); | |
2253efc8 | 1379 | return; |
f984df1f | 1380 | } |
7b371636 | 1381 | |
2253efc8 | 1382 | insert: |
bd6737f1 | 1383 | blk_mq_sched_insert_request(rq, false, true, true, false); |
f984df1f SL |
1384 | } |
1385 | ||
07068d5b JA |
1386 | /* |
1387 | * Multiple hardware queue variant. This will not use per-process plugs, | |
1388 | * but will attempt to bypass the hctx queueing if we can go straight to | |
1389 | * hardware for SYNC IO. | |
1390 | */ | |
dece1635 | 1391 | static blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio) |
07068d5b | 1392 | { |
ef295ecf | 1393 | const int is_sync = op_is_sync(bio->bi_opf); |
f73f44eb | 1394 | const int is_flush_fua = op_is_flush(bio->bi_opf); |
5a797e00 | 1395 | struct blk_mq_alloc_data data = { .flags = 0 }; |
07068d5b | 1396 | struct request *rq; |
6a83e74d | 1397 | unsigned int request_count = 0, srcu_idx; |
f984df1f | 1398 | struct blk_plug *plug; |
5b3f341f | 1399 | struct request *same_queue_rq = NULL; |
7b371636 | 1400 | blk_qc_t cookie; |
87760e5e | 1401 | unsigned int wb_acct; |
07068d5b JA |
1402 | |
1403 | blk_queue_bounce(q, &bio); | |
1404 | ||
1405 | if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) { | |
4246a0b6 | 1406 | bio_io_error(bio); |
dece1635 | 1407 | return BLK_QC_T_NONE; |
07068d5b JA |
1408 | } |
1409 | ||
54efd50b KO |
1410 | blk_queue_split(q, &bio, q->bio_split); |
1411 | ||
87c279e6 OS |
1412 | if (!is_flush_fua && !blk_queue_nomerges(q) && |
1413 | blk_attempt_plug_merge(q, bio, &request_count, &same_queue_rq)) | |
1414 | return BLK_QC_T_NONE; | |
f984df1f | 1415 | |
bd166ef1 JA |
1416 | if (blk_mq_sched_bio_merge(q, bio)) |
1417 | return BLK_QC_T_NONE; | |
1418 | ||
87760e5e JA |
1419 | wb_acct = wbt_wait(q->rq_wb, bio, NULL); |
1420 | ||
bd166ef1 JA |
1421 | trace_block_getrq(q, bio, bio->bi_opf); |
1422 | ||
1423 | rq = blk_mq_sched_get_request(q, bio, bio->bi_opf, &data); | |
87760e5e JA |
1424 | if (unlikely(!rq)) { |
1425 | __wbt_done(q->rq_wb, wb_acct); | |
dece1635 | 1426 | return BLK_QC_T_NONE; |
87760e5e JA |
1427 | } |
1428 | ||
1429 | wbt_track(&rq->issue_stat, wb_acct); | |
07068d5b | 1430 | |
fd2d3326 | 1431 | cookie = request_to_qc_t(data.hctx, rq); |
07068d5b JA |
1432 | |
1433 | if (unlikely(is_flush_fua)) { | |
0c2a6fe4 JA |
1434 | if (q->elevator) |
1435 | goto elv_insert; | |
07068d5b JA |
1436 | blk_mq_bio_to_request(rq, bio); |
1437 | blk_insert_flush(rq); | |
0c2a6fe4 | 1438 | goto run_queue; |
07068d5b JA |
1439 | } |
1440 | ||
f984df1f | 1441 | plug = current->plug; |
e167dfb5 JA |
1442 | /* |
1443 | * If the driver supports defer issued based on 'last', then | |
1444 | * queue it up like normal since we can potentially save some | |
1445 | * CPU this way. | |
1446 | */ | |
f984df1f SL |
1447 | if (((plug && !blk_queue_nomerges(q)) || is_sync) && |
1448 | !(data.hctx->flags & BLK_MQ_F_DEFER_ISSUE)) { | |
1449 | struct request *old_rq = NULL; | |
07068d5b JA |
1450 | |
1451 | blk_mq_bio_to_request(rq, bio); | |
07068d5b JA |
1452 | |
1453 | /* | |
6a83e74d | 1454 | * We do limited plugging. If the bio can be merged, do that. |
f984df1f SL |
1455 | * Otherwise the existing request in the plug list will be |
1456 | * issued. So the plug list will have one request at most | |
07068d5b | 1457 | */ |
f984df1f | 1458 | if (plug) { |
5b3f341f SL |
1459 | /* |
1460 | * The plug list might get flushed before this. If that | |
b094f89c JA |
1461 | * happens, same_queue_rq is invalid and plug list is |
1462 | * empty | |
1463 | */ | |
5b3f341f SL |
1464 | if (same_queue_rq && !list_empty(&plug->mq_list)) { |
1465 | old_rq = same_queue_rq; | |
f984df1f | 1466 | list_del_init(&old_rq->queuelist); |
07068d5b | 1467 | } |
f984df1f SL |
1468 | list_add_tail(&rq->queuelist, &plug->mq_list); |
1469 | } else /* is_sync */ | |
1470 | old_rq = rq; | |
1471 | blk_mq_put_ctx(data.ctx); | |
1472 | if (!old_rq) | |
7b371636 | 1473 | goto done; |
6a83e74d BVA |
1474 | |
1475 | if (!(data.hctx->flags & BLK_MQ_F_BLOCKING)) { | |
1476 | rcu_read_lock(); | |
066a4a73 | 1477 | blk_mq_try_issue_directly(old_rq, &cookie); |
6a83e74d BVA |
1478 | rcu_read_unlock(); |
1479 | } else { | |
1480 | srcu_idx = srcu_read_lock(&data.hctx->queue_rq_srcu); | |
066a4a73 | 1481 | blk_mq_try_issue_directly(old_rq, &cookie); |
6a83e74d BVA |
1482 | srcu_read_unlock(&data.hctx->queue_rq_srcu, srcu_idx); |
1483 | } | |
7b371636 | 1484 | goto done; |
07068d5b JA |
1485 | } |
1486 | ||
bd166ef1 | 1487 | if (q->elevator) { |
0c2a6fe4 | 1488 | elv_insert: |
bd166ef1 JA |
1489 | blk_mq_put_ctx(data.ctx); |
1490 | blk_mq_bio_to_request(rq, bio); | |
0abad774 | 1491 | blk_mq_sched_insert_request(rq, false, true, |
bd6737f1 | 1492 | !is_sync || is_flush_fua, true); |
bd166ef1 JA |
1493 | goto done; |
1494 | } | |
07068d5b JA |
1495 | if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) { |
1496 | /* | |
1497 | * For a SYNC request, send it to the hardware immediately. For | |
1498 | * an ASYNC request, just ensure that we run it later on. The | |
1499 | * latter allows for merging opportunities and more efficient | |
1500 | * dispatching. | |
1501 | */ | |
0c2a6fe4 | 1502 | run_queue: |
07068d5b JA |
1503 | blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua); |
1504 | } | |
07068d5b | 1505 | blk_mq_put_ctx(data.ctx); |
7b371636 JA |
1506 | done: |
1507 | return cookie; | |
07068d5b JA |
1508 | } |
1509 | ||
1510 | /* | |
1511 | * Single hardware queue variant. This will attempt to use any per-process | |
1512 | * plug for merging and IO deferral. | |
1513 | */ | |
dece1635 | 1514 | static blk_qc_t blk_sq_make_request(struct request_queue *q, struct bio *bio) |
07068d5b | 1515 | { |
ef295ecf | 1516 | const int is_sync = op_is_sync(bio->bi_opf); |
f73f44eb | 1517 | const int is_flush_fua = op_is_flush(bio->bi_opf); |
e6c4438b JM |
1518 | struct blk_plug *plug; |
1519 | unsigned int request_count = 0; | |
5a797e00 | 1520 | struct blk_mq_alloc_data data = { .flags = 0 }; |
07068d5b | 1521 | struct request *rq; |
7b371636 | 1522 | blk_qc_t cookie; |
87760e5e | 1523 | unsigned int wb_acct; |
07068d5b | 1524 | |
07068d5b JA |
1525 | blk_queue_bounce(q, &bio); |
1526 | ||
1527 | if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) { | |
4246a0b6 | 1528 | bio_io_error(bio); |
dece1635 | 1529 | return BLK_QC_T_NONE; |
07068d5b JA |
1530 | } |
1531 | ||
54efd50b KO |
1532 | blk_queue_split(q, &bio, q->bio_split); |
1533 | ||
87c279e6 OS |
1534 | if (!is_flush_fua && !blk_queue_nomerges(q)) { |
1535 | if (blk_attempt_plug_merge(q, bio, &request_count, NULL)) | |
1536 | return BLK_QC_T_NONE; | |
1537 | } else | |
1538 | request_count = blk_plug_queued_count(q); | |
07068d5b | 1539 | |
bd166ef1 JA |
1540 | if (blk_mq_sched_bio_merge(q, bio)) |
1541 | return BLK_QC_T_NONE; | |
1542 | ||
87760e5e JA |
1543 | wb_acct = wbt_wait(q->rq_wb, bio, NULL); |
1544 | ||
bd166ef1 JA |
1545 | trace_block_getrq(q, bio, bio->bi_opf); |
1546 | ||
1547 | rq = blk_mq_sched_get_request(q, bio, bio->bi_opf, &data); | |
87760e5e JA |
1548 | if (unlikely(!rq)) { |
1549 | __wbt_done(q->rq_wb, wb_acct); | |
dece1635 | 1550 | return BLK_QC_T_NONE; |
87760e5e JA |
1551 | } |
1552 | ||
1553 | wbt_track(&rq->issue_stat, wb_acct); | |
320ae51f | 1554 | |
fd2d3326 | 1555 | cookie = request_to_qc_t(data.hctx, rq); |
320ae51f JA |
1556 | |
1557 | if (unlikely(is_flush_fua)) { | |
0c2a6fe4 JA |
1558 | if (q->elevator) |
1559 | goto elv_insert; | |
320ae51f | 1560 | blk_mq_bio_to_request(rq, bio); |
320ae51f | 1561 | blk_insert_flush(rq); |
0c2a6fe4 | 1562 | goto run_queue; |
320ae51f JA |
1563 | } |
1564 | ||
1565 | /* | |
1566 | * A task plug currently exists. Since this is completely lockless, | |
1567 | * utilize that to temporarily store requests until the task is | |
1568 | * either done or scheduled away. | |
1569 | */ | |
e6c4438b JM |
1570 | plug = current->plug; |
1571 | if (plug) { | |
600271d9 SL |
1572 | struct request *last = NULL; |
1573 | ||
e6c4438b | 1574 | blk_mq_bio_to_request(rq, bio); |
0a6219a9 ML |
1575 | |
1576 | /* | |
1577 | * @request_count may become stale because of schedule | |
1578 | * out, so check the list again. | |
1579 | */ | |
1580 | if (list_empty(&plug->mq_list)) | |
1581 | request_count = 0; | |
676d0607 | 1582 | if (!request_count) |
e6c4438b | 1583 | trace_block_plug(q); |
600271d9 SL |
1584 | else |
1585 | last = list_entry_rq(plug->mq_list.prev); | |
b094f89c JA |
1586 | |
1587 | blk_mq_put_ctx(data.ctx); | |
1588 | ||
600271d9 SL |
1589 | if (request_count >= BLK_MAX_REQUEST_COUNT || (last && |
1590 | blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) { | |
e6c4438b JM |
1591 | blk_flush_plug_list(plug, false); |
1592 | trace_block_plug(q); | |
320ae51f | 1593 | } |
b094f89c | 1594 | |
e6c4438b | 1595 | list_add_tail(&rq->queuelist, &plug->mq_list); |
7b371636 | 1596 | return cookie; |
320ae51f JA |
1597 | } |
1598 | ||
bd166ef1 | 1599 | if (q->elevator) { |
0c2a6fe4 | 1600 | elv_insert: |
bd166ef1 JA |
1601 | blk_mq_put_ctx(data.ctx); |
1602 | blk_mq_bio_to_request(rq, bio); | |
0abad774 | 1603 | blk_mq_sched_insert_request(rq, false, true, |
bd6737f1 | 1604 | !is_sync || is_flush_fua, true); |
bd166ef1 JA |
1605 | goto done; |
1606 | } | |
07068d5b JA |
1607 | if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) { |
1608 | /* | |
1609 | * For a SYNC request, send it to the hardware immediately. For | |
1610 | * an ASYNC request, just ensure that we run it later on. The | |
1611 | * latter allows for merging opportunities and more efficient | |
1612 | * dispatching. | |
1613 | */ | |
0c2a6fe4 | 1614 | run_queue: |
07068d5b | 1615 | blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua); |
320ae51f JA |
1616 | } |
1617 | ||
07068d5b | 1618 | blk_mq_put_ctx(data.ctx); |
bd166ef1 | 1619 | done: |
7b371636 | 1620 | return cookie; |
320ae51f JA |
1621 | } |
1622 | ||
cc71a6f4 JA |
1623 | void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags, |
1624 | unsigned int hctx_idx) | |
95363efd | 1625 | { |
e9b267d9 | 1626 | struct page *page; |
320ae51f | 1627 | |
24d2f903 | 1628 | if (tags->rqs && set->ops->exit_request) { |
e9b267d9 | 1629 | int i; |
320ae51f | 1630 | |
24d2f903 | 1631 | for (i = 0; i < tags->nr_tags; i++) { |
2af8cbe3 JA |
1632 | struct request *rq = tags->static_rqs[i]; |
1633 | ||
1634 | if (!rq) | |
e9b267d9 | 1635 | continue; |
2af8cbe3 | 1636 | set->ops->exit_request(set->driver_data, rq, |
24d2f903 | 1637 | hctx_idx, i); |
2af8cbe3 | 1638 | tags->static_rqs[i] = NULL; |
e9b267d9 | 1639 | } |
320ae51f | 1640 | } |
320ae51f | 1641 | |
24d2f903 CH |
1642 | while (!list_empty(&tags->page_list)) { |
1643 | page = list_first_entry(&tags->page_list, struct page, lru); | |
6753471c | 1644 | list_del_init(&page->lru); |
f75782e4 CM |
1645 | /* |
1646 | * Remove kmemleak object previously allocated in | |
1647 | * blk_mq_init_rq_map(). | |
1648 | */ | |
1649 | kmemleak_free(page_address(page)); | |
320ae51f JA |
1650 | __free_pages(page, page->private); |
1651 | } | |
cc71a6f4 | 1652 | } |
320ae51f | 1653 | |
cc71a6f4 JA |
1654 | void blk_mq_free_rq_map(struct blk_mq_tags *tags) |
1655 | { | |
24d2f903 | 1656 | kfree(tags->rqs); |
cc71a6f4 | 1657 | tags->rqs = NULL; |
2af8cbe3 JA |
1658 | kfree(tags->static_rqs); |
1659 | tags->static_rqs = NULL; | |
320ae51f | 1660 | |
24d2f903 | 1661 | blk_mq_free_tags(tags); |
320ae51f JA |
1662 | } |
1663 | ||
cc71a6f4 JA |
1664 | struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, |
1665 | unsigned int hctx_idx, | |
1666 | unsigned int nr_tags, | |
1667 | unsigned int reserved_tags) | |
320ae51f | 1668 | { |
24d2f903 | 1669 | struct blk_mq_tags *tags; |
320ae51f | 1670 | |
cc71a6f4 | 1671 | tags = blk_mq_init_tags(nr_tags, reserved_tags, |
24391c0d SL |
1672 | set->numa_node, |
1673 | BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags)); | |
24d2f903 CH |
1674 | if (!tags) |
1675 | return NULL; | |
320ae51f | 1676 | |
cc71a6f4 | 1677 | tags->rqs = kzalloc_node(nr_tags * sizeof(struct request *), |
36e1f3d1 | 1678 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, |
a5164405 | 1679 | set->numa_node); |
24d2f903 CH |
1680 | if (!tags->rqs) { |
1681 | blk_mq_free_tags(tags); | |
1682 | return NULL; | |
1683 | } | |
320ae51f | 1684 | |
2af8cbe3 JA |
1685 | tags->static_rqs = kzalloc_node(nr_tags * sizeof(struct request *), |
1686 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, | |
1687 | set->numa_node); | |
1688 | if (!tags->static_rqs) { | |
1689 | kfree(tags->rqs); | |
1690 | blk_mq_free_tags(tags); | |
1691 | return NULL; | |
1692 | } | |
1693 | ||
cc71a6f4 JA |
1694 | return tags; |
1695 | } | |
1696 | ||
1697 | static size_t order_to_size(unsigned int order) | |
1698 | { | |
1699 | return (size_t)PAGE_SIZE << order; | |
1700 | } | |
1701 | ||
1702 | int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags, | |
1703 | unsigned int hctx_idx, unsigned int depth) | |
1704 | { | |
1705 | unsigned int i, j, entries_per_page, max_order = 4; | |
1706 | size_t rq_size, left; | |
1707 | ||
1708 | INIT_LIST_HEAD(&tags->page_list); | |
1709 | ||
320ae51f JA |
1710 | /* |
1711 | * rq_size is the size of the request plus driver payload, rounded | |
1712 | * to the cacheline size | |
1713 | */ | |
24d2f903 | 1714 | rq_size = round_up(sizeof(struct request) + set->cmd_size, |
320ae51f | 1715 | cache_line_size()); |
cc71a6f4 | 1716 | left = rq_size * depth; |
320ae51f | 1717 | |
cc71a6f4 | 1718 | for (i = 0; i < depth; ) { |
320ae51f JA |
1719 | int this_order = max_order; |
1720 | struct page *page; | |
1721 | int to_do; | |
1722 | void *p; | |
1723 | ||
b3a834b1 | 1724 | while (this_order && left < order_to_size(this_order - 1)) |
320ae51f JA |
1725 | this_order--; |
1726 | ||
1727 | do { | |
a5164405 | 1728 | page = alloc_pages_node(set->numa_node, |
36e1f3d1 | 1729 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO, |
a5164405 | 1730 | this_order); |
320ae51f JA |
1731 | if (page) |
1732 | break; | |
1733 | if (!this_order--) | |
1734 | break; | |
1735 | if (order_to_size(this_order) < rq_size) | |
1736 | break; | |
1737 | } while (1); | |
1738 | ||
1739 | if (!page) | |
24d2f903 | 1740 | goto fail; |
320ae51f JA |
1741 | |
1742 | page->private = this_order; | |
24d2f903 | 1743 | list_add_tail(&page->lru, &tags->page_list); |
320ae51f JA |
1744 | |
1745 | p = page_address(page); | |
f75782e4 CM |
1746 | /* |
1747 | * Allow kmemleak to scan these pages as they contain pointers | |
1748 | * to additional allocations like via ops->init_request(). | |
1749 | */ | |
36e1f3d1 | 1750 | kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO); |
320ae51f | 1751 | entries_per_page = order_to_size(this_order) / rq_size; |
cc71a6f4 | 1752 | to_do = min(entries_per_page, depth - i); |
320ae51f JA |
1753 | left -= to_do * rq_size; |
1754 | for (j = 0; j < to_do; j++) { | |
2af8cbe3 JA |
1755 | struct request *rq = p; |
1756 | ||
1757 | tags->static_rqs[i] = rq; | |
24d2f903 CH |
1758 | if (set->ops->init_request) { |
1759 | if (set->ops->init_request(set->driver_data, | |
2af8cbe3 | 1760 | rq, hctx_idx, i, |
a5164405 | 1761 | set->numa_node)) { |
2af8cbe3 | 1762 | tags->static_rqs[i] = NULL; |
24d2f903 | 1763 | goto fail; |
a5164405 | 1764 | } |
e9b267d9 CH |
1765 | } |
1766 | ||
320ae51f JA |
1767 | p += rq_size; |
1768 | i++; | |
1769 | } | |
1770 | } | |
cc71a6f4 | 1771 | return 0; |
320ae51f | 1772 | |
24d2f903 | 1773 | fail: |
cc71a6f4 JA |
1774 | blk_mq_free_rqs(set, tags, hctx_idx); |
1775 | return -ENOMEM; | |
320ae51f JA |
1776 | } |
1777 | ||
e57690fe JA |
1778 | /* |
1779 | * 'cpu' is going away. splice any existing rq_list entries from this | |
1780 | * software queue to the hw queue dispatch list, and ensure that it | |
1781 | * gets run. | |
1782 | */ | |
9467f859 | 1783 | static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node) |
484b4061 | 1784 | { |
9467f859 | 1785 | struct blk_mq_hw_ctx *hctx; |
484b4061 JA |
1786 | struct blk_mq_ctx *ctx; |
1787 | LIST_HEAD(tmp); | |
1788 | ||
9467f859 | 1789 | hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead); |
e57690fe | 1790 | ctx = __blk_mq_get_ctx(hctx->queue, cpu); |
484b4061 JA |
1791 | |
1792 | spin_lock(&ctx->lock); | |
1793 | if (!list_empty(&ctx->rq_list)) { | |
1794 | list_splice_init(&ctx->rq_list, &tmp); | |
1795 | blk_mq_hctx_clear_pending(hctx, ctx); | |
1796 | } | |
1797 | spin_unlock(&ctx->lock); | |
1798 | ||
1799 | if (list_empty(&tmp)) | |
9467f859 | 1800 | return 0; |
484b4061 | 1801 | |
e57690fe JA |
1802 | spin_lock(&hctx->lock); |
1803 | list_splice_tail_init(&tmp, &hctx->dispatch); | |
1804 | spin_unlock(&hctx->lock); | |
484b4061 JA |
1805 | |
1806 | blk_mq_run_hw_queue(hctx, true); | |
9467f859 | 1807 | return 0; |
484b4061 JA |
1808 | } |
1809 | ||
9467f859 | 1810 | static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx) |
484b4061 | 1811 | { |
9467f859 TG |
1812 | cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD, |
1813 | &hctx->cpuhp_dead); | |
484b4061 JA |
1814 | } |
1815 | ||
c3b4afca | 1816 | /* hctx->ctxs will be freed in queue's release handler */ |
08e98fc6 ML |
1817 | static void blk_mq_exit_hctx(struct request_queue *q, |
1818 | struct blk_mq_tag_set *set, | |
1819 | struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx) | |
1820 | { | |
f70ced09 ML |
1821 | unsigned flush_start_tag = set->queue_depth; |
1822 | ||
08e98fc6 ML |
1823 | blk_mq_tag_idle(hctx); |
1824 | ||
f70ced09 ML |
1825 | if (set->ops->exit_request) |
1826 | set->ops->exit_request(set->driver_data, | |
1827 | hctx->fq->flush_rq, hctx_idx, | |
1828 | flush_start_tag + hctx_idx); | |
1829 | ||
08e98fc6 ML |
1830 | if (set->ops->exit_hctx) |
1831 | set->ops->exit_hctx(hctx, hctx_idx); | |
1832 | ||
6a83e74d BVA |
1833 | if (hctx->flags & BLK_MQ_F_BLOCKING) |
1834 | cleanup_srcu_struct(&hctx->queue_rq_srcu); | |
1835 | ||
9467f859 | 1836 | blk_mq_remove_cpuhp(hctx); |
f70ced09 | 1837 | blk_free_flush_queue(hctx->fq); |
88459642 | 1838 | sbitmap_free(&hctx->ctx_map); |
08e98fc6 ML |
1839 | } |
1840 | ||
624dbe47 ML |
1841 | static void blk_mq_exit_hw_queues(struct request_queue *q, |
1842 | struct blk_mq_tag_set *set, int nr_queue) | |
1843 | { | |
1844 | struct blk_mq_hw_ctx *hctx; | |
1845 | unsigned int i; | |
1846 | ||
1847 | queue_for_each_hw_ctx(q, hctx, i) { | |
1848 | if (i == nr_queue) | |
1849 | break; | |
08e98fc6 | 1850 | blk_mq_exit_hctx(q, set, hctx, i); |
624dbe47 | 1851 | } |
624dbe47 ML |
1852 | } |
1853 | ||
1854 | static void blk_mq_free_hw_queues(struct request_queue *q, | |
1855 | struct blk_mq_tag_set *set) | |
1856 | { | |
1857 | struct blk_mq_hw_ctx *hctx; | |
1858 | unsigned int i; | |
1859 | ||
e09aae7e | 1860 | queue_for_each_hw_ctx(q, hctx, i) |
624dbe47 | 1861 | free_cpumask_var(hctx->cpumask); |
624dbe47 ML |
1862 | } |
1863 | ||
08e98fc6 ML |
1864 | static int blk_mq_init_hctx(struct request_queue *q, |
1865 | struct blk_mq_tag_set *set, | |
1866 | struct blk_mq_hw_ctx *hctx, unsigned hctx_idx) | |
320ae51f | 1867 | { |
08e98fc6 | 1868 | int node; |
f70ced09 | 1869 | unsigned flush_start_tag = set->queue_depth; |
08e98fc6 ML |
1870 | |
1871 | node = hctx->numa_node; | |
1872 | if (node == NUMA_NO_NODE) | |
1873 | node = hctx->numa_node = set->numa_node; | |
1874 | ||
27489a3c | 1875 | INIT_WORK(&hctx->run_work, blk_mq_run_work_fn); |
08e98fc6 ML |
1876 | INIT_DELAYED_WORK(&hctx->delay_work, blk_mq_delay_work_fn); |
1877 | spin_lock_init(&hctx->lock); | |
1878 | INIT_LIST_HEAD(&hctx->dispatch); | |
1879 | hctx->queue = q; | |
1880 | hctx->queue_num = hctx_idx; | |
2404e607 | 1881 | hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED; |
08e98fc6 | 1882 | |
9467f859 | 1883 | cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead); |
08e98fc6 ML |
1884 | |
1885 | hctx->tags = set->tags[hctx_idx]; | |
320ae51f JA |
1886 | |
1887 | /* | |
08e98fc6 ML |
1888 | * Allocate space for all possible cpus to avoid allocation at |
1889 | * runtime | |
320ae51f | 1890 | */ |
08e98fc6 ML |
1891 | hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *), |
1892 | GFP_KERNEL, node); | |
1893 | if (!hctx->ctxs) | |
1894 | goto unregister_cpu_notifier; | |
320ae51f | 1895 | |
88459642 OS |
1896 | if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8), GFP_KERNEL, |
1897 | node)) | |
08e98fc6 | 1898 | goto free_ctxs; |
320ae51f | 1899 | |
08e98fc6 | 1900 | hctx->nr_ctx = 0; |
320ae51f | 1901 | |
08e98fc6 ML |
1902 | if (set->ops->init_hctx && |
1903 | set->ops->init_hctx(hctx, set->driver_data, hctx_idx)) | |
1904 | goto free_bitmap; | |
320ae51f | 1905 | |
f70ced09 ML |
1906 | hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size); |
1907 | if (!hctx->fq) | |
1908 | goto exit_hctx; | |
320ae51f | 1909 | |
f70ced09 ML |
1910 | if (set->ops->init_request && |
1911 | set->ops->init_request(set->driver_data, | |
1912 | hctx->fq->flush_rq, hctx_idx, | |
1913 | flush_start_tag + hctx_idx, node)) | |
1914 | goto free_fq; | |
320ae51f | 1915 | |
6a83e74d BVA |
1916 | if (hctx->flags & BLK_MQ_F_BLOCKING) |
1917 | init_srcu_struct(&hctx->queue_rq_srcu); | |
1918 | ||
08e98fc6 | 1919 | return 0; |
320ae51f | 1920 | |
f70ced09 ML |
1921 | free_fq: |
1922 | kfree(hctx->fq); | |
1923 | exit_hctx: | |
1924 | if (set->ops->exit_hctx) | |
1925 | set->ops->exit_hctx(hctx, hctx_idx); | |
08e98fc6 | 1926 | free_bitmap: |
88459642 | 1927 | sbitmap_free(&hctx->ctx_map); |
08e98fc6 ML |
1928 | free_ctxs: |
1929 | kfree(hctx->ctxs); | |
1930 | unregister_cpu_notifier: | |
9467f859 | 1931 | blk_mq_remove_cpuhp(hctx); |
08e98fc6 ML |
1932 | return -1; |
1933 | } | |
320ae51f | 1934 | |
320ae51f JA |
1935 | static void blk_mq_init_cpu_queues(struct request_queue *q, |
1936 | unsigned int nr_hw_queues) | |
1937 | { | |
1938 | unsigned int i; | |
1939 | ||
1940 | for_each_possible_cpu(i) { | |
1941 | struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i); | |
1942 | struct blk_mq_hw_ctx *hctx; | |
1943 | ||
1944 | memset(__ctx, 0, sizeof(*__ctx)); | |
1945 | __ctx->cpu = i; | |
1946 | spin_lock_init(&__ctx->lock); | |
1947 | INIT_LIST_HEAD(&__ctx->rq_list); | |
1948 | __ctx->queue = q; | |
cf43e6be JA |
1949 | blk_stat_init(&__ctx->stat[BLK_STAT_READ]); |
1950 | blk_stat_init(&__ctx->stat[BLK_STAT_WRITE]); | |
320ae51f JA |
1951 | |
1952 | /* If the cpu isn't online, the cpu is mapped to first hctx */ | |
320ae51f JA |
1953 | if (!cpu_online(i)) |
1954 | continue; | |
1955 | ||
7d7e0f90 | 1956 | hctx = blk_mq_map_queue(q, i); |
e4043dcf | 1957 | |
320ae51f JA |
1958 | /* |
1959 | * Set local node, IFF we have more than one hw queue. If | |
1960 | * not, we remain on the home node of the device | |
1961 | */ | |
1962 | if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE) | |
bffed457 | 1963 | hctx->numa_node = local_memory_node(cpu_to_node(i)); |
320ae51f JA |
1964 | } |
1965 | } | |
1966 | ||
cc71a6f4 JA |
1967 | static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, int hctx_idx) |
1968 | { | |
1969 | int ret = 0; | |
1970 | ||
1971 | set->tags[hctx_idx] = blk_mq_alloc_rq_map(set, hctx_idx, | |
1972 | set->queue_depth, set->reserved_tags); | |
1973 | if (!set->tags[hctx_idx]) | |
1974 | return false; | |
1975 | ||
1976 | ret = blk_mq_alloc_rqs(set, set->tags[hctx_idx], hctx_idx, | |
1977 | set->queue_depth); | |
1978 | if (!ret) | |
1979 | return true; | |
1980 | ||
1981 | blk_mq_free_rq_map(set->tags[hctx_idx]); | |
1982 | set->tags[hctx_idx] = NULL; | |
1983 | return false; | |
1984 | } | |
1985 | ||
1986 | static void blk_mq_free_map_and_requests(struct blk_mq_tag_set *set, | |
1987 | unsigned int hctx_idx) | |
1988 | { | |
bd166ef1 JA |
1989 | if (set->tags[hctx_idx]) { |
1990 | blk_mq_free_rqs(set, set->tags[hctx_idx], hctx_idx); | |
1991 | blk_mq_free_rq_map(set->tags[hctx_idx]); | |
1992 | set->tags[hctx_idx] = NULL; | |
1993 | } | |
cc71a6f4 JA |
1994 | } |
1995 | ||
5778322e AM |
1996 | static void blk_mq_map_swqueue(struct request_queue *q, |
1997 | const struct cpumask *online_mask) | |
320ae51f | 1998 | { |
d1b1cea1 | 1999 | unsigned int i, hctx_idx; |
320ae51f JA |
2000 | struct blk_mq_hw_ctx *hctx; |
2001 | struct blk_mq_ctx *ctx; | |
2a34c087 | 2002 | struct blk_mq_tag_set *set = q->tag_set; |
320ae51f | 2003 | |
60de074b AM |
2004 | /* |
2005 | * Avoid others reading imcomplete hctx->cpumask through sysfs | |
2006 | */ | |
2007 | mutex_lock(&q->sysfs_lock); | |
2008 | ||
320ae51f | 2009 | queue_for_each_hw_ctx(q, hctx, i) { |
e4043dcf | 2010 | cpumask_clear(hctx->cpumask); |
320ae51f JA |
2011 | hctx->nr_ctx = 0; |
2012 | } | |
2013 | ||
2014 | /* | |
2015 | * Map software to hardware queues | |
2016 | */ | |
897bb0c7 | 2017 | for_each_possible_cpu(i) { |
320ae51f | 2018 | /* If the cpu isn't online, the cpu is mapped to first hctx */ |
5778322e | 2019 | if (!cpumask_test_cpu(i, online_mask)) |
e4043dcf JA |
2020 | continue; |
2021 | ||
d1b1cea1 GKB |
2022 | hctx_idx = q->mq_map[i]; |
2023 | /* unmapped hw queue can be remapped after CPU topo changed */ | |
cc71a6f4 JA |
2024 | if (!set->tags[hctx_idx] && |
2025 | !__blk_mq_alloc_rq_map(set, hctx_idx)) { | |
d1b1cea1 GKB |
2026 | /* |
2027 | * If tags initialization fail for some hctx, | |
2028 | * that hctx won't be brought online. In this | |
2029 | * case, remap the current ctx to hctx[0] which | |
2030 | * is guaranteed to always have tags allocated | |
2031 | */ | |
cc71a6f4 | 2032 | q->mq_map[i] = 0; |
d1b1cea1 GKB |
2033 | } |
2034 | ||
897bb0c7 | 2035 | ctx = per_cpu_ptr(q->queue_ctx, i); |
7d7e0f90 | 2036 | hctx = blk_mq_map_queue(q, i); |
868f2f0b | 2037 | |
e4043dcf | 2038 | cpumask_set_cpu(i, hctx->cpumask); |
320ae51f JA |
2039 | ctx->index_hw = hctx->nr_ctx; |
2040 | hctx->ctxs[hctx->nr_ctx++] = ctx; | |
2041 | } | |
506e931f | 2042 | |
60de074b AM |
2043 | mutex_unlock(&q->sysfs_lock); |
2044 | ||
506e931f | 2045 | queue_for_each_hw_ctx(q, hctx, i) { |
484b4061 | 2046 | /* |
a68aafa5 JA |
2047 | * If no software queues are mapped to this hardware queue, |
2048 | * disable it and free the request entries. | |
484b4061 JA |
2049 | */ |
2050 | if (!hctx->nr_ctx) { | |
d1b1cea1 GKB |
2051 | /* Never unmap queue 0. We need it as a |
2052 | * fallback in case of a new remap fails | |
2053 | * allocation | |
2054 | */ | |
cc71a6f4 JA |
2055 | if (i && set->tags[i]) |
2056 | blk_mq_free_map_and_requests(set, i); | |
2057 | ||
2a34c087 | 2058 | hctx->tags = NULL; |
484b4061 JA |
2059 | continue; |
2060 | } | |
2061 | ||
2a34c087 ML |
2062 | hctx->tags = set->tags[i]; |
2063 | WARN_ON(!hctx->tags); | |
2064 | ||
889fa31f CY |
2065 | /* |
2066 | * Set the map size to the number of mapped software queues. | |
2067 | * This is more accurate and more efficient than looping | |
2068 | * over all possibly mapped software queues. | |
2069 | */ | |
88459642 | 2070 | sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx); |
889fa31f | 2071 | |
484b4061 JA |
2072 | /* |
2073 | * Initialize batch roundrobin counts | |
2074 | */ | |
506e931f JA |
2075 | hctx->next_cpu = cpumask_first(hctx->cpumask); |
2076 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; | |
2077 | } | |
320ae51f JA |
2078 | } |
2079 | ||
2404e607 | 2080 | static void queue_set_hctx_shared(struct request_queue *q, bool shared) |
0d2602ca JA |
2081 | { |
2082 | struct blk_mq_hw_ctx *hctx; | |
0d2602ca JA |
2083 | int i; |
2084 | ||
2404e607 JM |
2085 | queue_for_each_hw_ctx(q, hctx, i) { |
2086 | if (shared) | |
2087 | hctx->flags |= BLK_MQ_F_TAG_SHARED; | |
2088 | else | |
2089 | hctx->flags &= ~BLK_MQ_F_TAG_SHARED; | |
2090 | } | |
2091 | } | |
2092 | ||
2093 | static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set, bool shared) | |
2094 | { | |
2095 | struct request_queue *q; | |
0d2602ca JA |
2096 | |
2097 | list_for_each_entry(q, &set->tag_list, tag_set_list) { | |
2098 | blk_mq_freeze_queue(q); | |
2404e607 | 2099 | queue_set_hctx_shared(q, shared); |
0d2602ca JA |
2100 | blk_mq_unfreeze_queue(q); |
2101 | } | |
2102 | } | |
2103 | ||
2104 | static void blk_mq_del_queue_tag_set(struct request_queue *q) | |
2105 | { | |
2106 | struct blk_mq_tag_set *set = q->tag_set; | |
2107 | ||
0d2602ca JA |
2108 | mutex_lock(&set->tag_list_lock); |
2109 | list_del_init(&q->tag_set_list); | |
2404e607 JM |
2110 | if (list_is_singular(&set->tag_list)) { |
2111 | /* just transitioned to unshared */ | |
2112 | set->flags &= ~BLK_MQ_F_TAG_SHARED; | |
2113 | /* update existing queue */ | |
2114 | blk_mq_update_tag_set_depth(set, false); | |
2115 | } | |
0d2602ca | 2116 | mutex_unlock(&set->tag_list_lock); |
0d2602ca JA |
2117 | } |
2118 | ||
2119 | static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set, | |
2120 | struct request_queue *q) | |
2121 | { | |
2122 | q->tag_set = set; | |
2123 | ||
2124 | mutex_lock(&set->tag_list_lock); | |
2404e607 JM |
2125 | |
2126 | /* Check to see if we're transitioning to shared (from 1 to 2 queues). */ | |
2127 | if (!list_empty(&set->tag_list) && !(set->flags & BLK_MQ_F_TAG_SHARED)) { | |
2128 | set->flags |= BLK_MQ_F_TAG_SHARED; | |
2129 | /* update existing queue */ | |
2130 | blk_mq_update_tag_set_depth(set, true); | |
2131 | } | |
2132 | if (set->flags & BLK_MQ_F_TAG_SHARED) | |
2133 | queue_set_hctx_shared(q, true); | |
0d2602ca | 2134 | list_add_tail(&q->tag_set_list, &set->tag_list); |
2404e607 | 2135 | |
0d2602ca JA |
2136 | mutex_unlock(&set->tag_list_lock); |
2137 | } | |
2138 | ||
e09aae7e ML |
2139 | /* |
2140 | * It is the actual release handler for mq, but we do it from | |
2141 | * request queue's release handler for avoiding use-after-free | |
2142 | * and headache because q->mq_kobj shouldn't have been introduced, | |
2143 | * but we can't group ctx/kctx kobj without it. | |
2144 | */ | |
2145 | void blk_mq_release(struct request_queue *q) | |
2146 | { | |
2147 | struct blk_mq_hw_ctx *hctx; | |
2148 | unsigned int i; | |
2149 | ||
bd166ef1 JA |
2150 | blk_mq_sched_teardown(q); |
2151 | ||
e09aae7e | 2152 | /* hctx kobj stays in hctx */ |
c3b4afca ML |
2153 | queue_for_each_hw_ctx(q, hctx, i) { |
2154 | if (!hctx) | |
2155 | continue; | |
2156 | kfree(hctx->ctxs); | |
e09aae7e | 2157 | kfree(hctx); |
c3b4afca | 2158 | } |
e09aae7e | 2159 | |
a723bab3 AM |
2160 | q->mq_map = NULL; |
2161 | ||
e09aae7e ML |
2162 | kfree(q->queue_hw_ctx); |
2163 | ||
2164 | /* ctx kobj stays in queue_ctx */ | |
2165 | free_percpu(q->queue_ctx); | |
2166 | } | |
2167 | ||
24d2f903 | 2168 | struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set) |
b62c21b7 MS |
2169 | { |
2170 | struct request_queue *uninit_q, *q; | |
2171 | ||
2172 | uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node); | |
2173 | if (!uninit_q) | |
2174 | return ERR_PTR(-ENOMEM); | |
2175 | ||
2176 | q = blk_mq_init_allocated_queue(set, uninit_q); | |
2177 | if (IS_ERR(q)) | |
2178 | blk_cleanup_queue(uninit_q); | |
2179 | ||
2180 | return q; | |
2181 | } | |
2182 | EXPORT_SYMBOL(blk_mq_init_queue); | |
2183 | ||
868f2f0b KB |
2184 | static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set, |
2185 | struct request_queue *q) | |
320ae51f | 2186 | { |
868f2f0b KB |
2187 | int i, j; |
2188 | struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx; | |
f14bbe77 | 2189 | |
868f2f0b | 2190 | blk_mq_sysfs_unregister(q); |
24d2f903 | 2191 | for (i = 0; i < set->nr_hw_queues; i++) { |
868f2f0b | 2192 | int node; |
f14bbe77 | 2193 | |
868f2f0b KB |
2194 | if (hctxs[i]) |
2195 | continue; | |
2196 | ||
2197 | node = blk_mq_hw_queue_to_node(q->mq_map, i); | |
cdef54dd CH |
2198 | hctxs[i] = kzalloc_node(sizeof(struct blk_mq_hw_ctx), |
2199 | GFP_KERNEL, node); | |
320ae51f | 2200 | if (!hctxs[i]) |
868f2f0b | 2201 | break; |
320ae51f | 2202 | |
a86073e4 | 2203 | if (!zalloc_cpumask_var_node(&hctxs[i]->cpumask, GFP_KERNEL, |
868f2f0b KB |
2204 | node)) { |
2205 | kfree(hctxs[i]); | |
2206 | hctxs[i] = NULL; | |
2207 | break; | |
2208 | } | |
e4043dcf | 2209 | |
0d2602ca | 2210 | atomic_set(&hctxs[i]->nr_active, 0); |
f14bbe77 | 2211 | hctxs[i]->numa_node = node; |
320ae51f | 2212 | hctxs[i]->queue_num = i; |
868f2f0b KB |
2213 | |
2214 | if (blk_mq_init_hctx(q, set, hctxs[i], i)) { | |
2215 | free_cpumask_var(hctxs[i]->cpumask); | |
2216 | kfree(hctxs[i]); | |
2217 | hctxs[i] = NULL; | |
2218 | break; | |
2219 | } | |
2220 | blk_mq_hctx_kobj_init(hctxs[i]); | |
320ae51f | 2221 | } |
868f2f0b KB |
2222 | for (j = i; j < q->nr_hw_queues; j++) { |
2223 | struct blk_mq_hw_ctx *hctx = hctxs[j]; | |
2224 | ||
2225 | if (hctx) { | |
cc71a6f4 JA |
2226 | if (hctx->tags) |
2227 | blk_mq_free_map_and_requests(set, j); | |
868f2f0b KB |
2228 | blk_mq_exit_hctx(q, set, hctx, j); |
2229 | free_cpumask_var(hctx->cpumask); | |
2230 | kobject_put(&hctx->kobj); | |
2231 | kfree(hctx->ctxs); | |
2232 | kfree(hctx); | |
2233 | hctxs[j] = NULL; | |
2234 | ||
2235 | } | |
2236 | } | |
2237 | q->nr_hw_queues = i; | |
2238 | blk_mq_sysfs_register(q); | |
2239 | } | |
2240 | ||
2241 | struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set, | |
2242 | struct request_queue *q) | |
2243 | { | |
66841672 ML |
2244 | /* mark the queue as mq asap */ |
2245 | q->mq_ops = set->ops; | |
2246 | ||
868f2f0b KB |
2247 | q->queue_ctx = alloc_percpu(struct blk_mq_ctx); |
2248 | if (!q->queue_ctx) | |
c7de5726 | 2249 | goto err_exit; |
868f2f0b KB |
2250 | |
2251 | q->queue_hw_ctx = kzalloc_node(nr_cpu_ids * sizeof(*(q->queue_hw_ctx)), | |
2252 | GFP_KERNEL, set->numa_node); | |
2253 | if (!q->queue_hw_ctx) | |
2254 | goto err_percpu; | |
2255 | ||
bdd17e75 | 2256 | q->mq_map = set->mq_map; |
868f2f0b KB |
2257 | |
2258 | blk_mq_realloc_hw_ctxs(set, q); | |
2259 | if (!q->nr_hw_queues) | |
2260 | goto err_hctxs; | |
320ae51f | 2261 | |
287922eb | 2262 | INIT_WORK(&q->timeout_work, blk_mq_timeout_work); |
e56f698b | 2263 | blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ); |
320ae51f JA |
2264 | |
2265 | q->nr_queues = nr_cpu_ids; | |
320ae51f | 2266 | |
94eddfbe | 2267 | q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT; |
320ae51f | 2268 | |
05f1dd53 JA |
2269 | if (!(set->flags & BLK_MQ_F_SG_MERGE)) |
2270 | q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE; | |
2271 | ||
1be036e9 CH |
2272 | q->sg_reserved_size = INT_MAX; |
2273 | ||
2849450a | 2274 | INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work); |
6fca6a61 CH |
2275 | INIT_LIST_HEAD(&q->requeue_list); |
2276 | spin_lock_init(&q->requeue_lock); | |
2277 | ||
07068d5b JA |
2278 | if (q->nr_hw_queues > 1) |
2279 | blk_queue_make_request(q, blk_mq_make_request); | |
2280 | else | |
2281 | blk_queue_make_request(q, blk_sq_make_request); | |
2282 | ||
eba71768 JA |
2283 | /* |
2284 | * Do this after blk_queue_make_request() overrides it... | |
2285 | */ | |
2286 | q->nr_requests = set->queue_depth; | |
2287 | ||
64f1c21e JA |
2288 | /* |
2289 | * Default to classic polling | |
2290 | */ | |
2291 | q->poll_nsec = -1; | |
2292 | ||
24d2f903 CH |
2293 | if (set->ops->complete) |
2294 | blk_queue_softirq_done(q, set->ops->complete); | |
30a91cb4 | 2295 | |
24d2f903 | 2296 | blk_mq_init_cpu_queues(q, set->nr_hw_queues); |
320ae51f | 2297 | |
5778322e | 2298 | get_online_cpus(); |
320ae51f | 2299 | mutex_lock(&all_q_mutex); |
320ae51f | 2300 | |
4593fdbe | 2301 | list_add_tail(&q->all_q_node, &all_q_list); |
0d2602ca | 2302 | blk_mq_add_queue_tag_set(set, q); |
5778322e | 2303 | blk_mq_map_swqueue(q, cpu_online_mask); |
484b4061 | 2304 | |
4593fdbe | 2305 | mutex_unlock(&all_q_mutex); |
5778322e | 2306 | put_online_cpus(); |
4593fdbe | 2307 | |
d3484991 JA |
2308 | if (!(set->flags & BLK_MQ_F_NO_SCHED)) { |
2309 | int ret; | |
2310 | ||
2311 | ret = blk_mq_sched_init(q); | |
2312 | if (ret) | |
2313 | return ERR_PTR(ret); | |
2314 | } | |
2315 | ||
320ae51f | 2316 | return q; |
18741986 | 2317 | |
320ae51f | 2318 | err_hctxs: |
868f2f0b | 2319 | kfree(q->queue_hw_ctx); |
320ae51f | 2320 | err_percpu: |
868f2f0b | 2321 | free_percpu(q->queue_ctx); |
c7de5726 ML |
2322 | err_exit: |
2323 | q->mq_ops = NULL; | |
320ae51f JA |
2324 | return ERR_PTR(-ENOMEM); |
2325 | } | |
b62c21b7 | 2326 | EXPORT_SYMBOL(blk_mq_init_allocated_queue); |
320ae51f JA |
2327 | |
2328 | void blk_mq_free_queue(struct request_queue *q) | |
2329 | { | |
624dbe47 | 2330 | struct blk_mq_tag_set *set = q->tag_set; |
320ae51f | 2331 | |
0e626368 AM |
2332 | mutex_lock(&all_q_mutex); |
2333 | list_del_init(&q->all_q_node); | |
2334 | mutex_unlock(&all_q_mutex); | |
2335 | ||
87760e5e JA |
2336 | wbt_exit(q); |
2337 | ||
0d2602ca JA |
2338 | blk_mq_del_queue_tag_set(q); |
2339 | ||
624dbe47 ML |
2340 | blk_mq_exit_hw_queues(q, set, set->nr_hw_queues); |
2341 | blk_mq_free_hw_queues(q, set); | |
320ae51f | 2342 | } |
320ae51f JA |
2343 | |
2344 | /* Basically redo blk_mq_init_queue with queue frozen */ | |
5778322e AM |
2345 | static void blk_mq_queue_reinit(struct request_queue *q, |
2346 | const struct cpumask *online_mask) | |
320ae51f | 2347 | { |
4ecd4fef | 2348 | WARN_ON_ONCE(!atomic_read(&q->mq_freeze_depth)); |
320ae51f | 2349 | |
67aec14c JA |
2350 | blk_mq_sysfs_unregister(q); |
2351 | ||
320ae51f JA |
2352 | /* |
2353 | * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe | |
2354 | * we should change hctx numa_node according to new topology (this | |
2355 | * involves free and re-allocate memory, worthy doing?) | |
2356 | */ | |
2357 | ||
5778322e | 2358 | blk_mq_map_swqueue(q, online_mask); |
320ae51f | 2359 | |
67aec14c | 2360 | blk_mq_sysfs_register(q); |
320ae51f JA |
2361 | } |
2362 | ||
65d5291e SAS |
2363 | /* |
2364 | * New online cpumask which is going to be set in this hotplug event. | |
2365 | * Declare this cpumasks as global as cpu-hotplug operation is invoked | |
2366 | * one-by-one and dynamically allocating this could result in a failure. | |
2367 | */ | |
2368 | static struct cpumask cpuhp_online_new; | |
2369 | ||
2370 | static void blk_mq_queue_reinit_work(void) | |
320ae51f JA |
2371 | { |
2372 | struct request_queue *q; | |
320ae51f JA |
2373 | |
2374 | mutex_lock(&all_q_mutex); | |
f3af020b TH |
2375 | /* |
2376 | * We need to freeze and reinit all existing queues. Freezing | |
2377 | * involves synchronous wait for an RCU grace period and doing it | |
2378 | * one by one may take a long time. Start freezing all queues in | |
2379 | * one swoop and then wait for the completions so that freezing can | |
2380 | * take place in parallel. | |
2381 | */ | |
2382 | list_for_each_entry(q, &all_q_list, all_q_node) | |
2383 | blk_mq_freeze_queue_start(q); | |
415d3dab | 2384 | list_for_each_entry(q, &all_q_list, all_q_node) |
f3af020b TH |
2385 | blk_mq_freeze_queue_wait(q); |
2386 | ||
320ae51f | 2387 | list_for_each_entry(q, &all_q_list, all_q_node) |
65d5291e | 2388 | blk_mq_queue_reinit(q, &cpuhp_online_new); |
f3af020b TH |
2389 | |
2390 | list_for_each_entry(q, &all_q_list, all_q_node) | |
2391 | blk_mq_unfreeze_queue(q); | |
2392 | ||
320ae51f | 2393 | mutex_unlock(&all_q_mutex); |
65d5291e SAS |
2394 | } |
2395 | ||
2396 | static int blk_mq_queue_reinit_dead(unsigned int cpu) | |
2397 | { | |
97a32864 | 2398 | cpumask_copy(&cpuhp_online_new, cpu_online_mask); |
65d5291e SAS |
2399 | blk_mq_queue_reinit_work(); |
2400 | return 0; | |
2401 | } | |
2402 | ||
2403 | /* | |
2404 | * Before hotadded cpu starts handling requests, new mappings must be | |
2405 | * established. Otherwise, these requests in hw queue might never be | |
2406 | * dispatched. | |
2407 | * | |
2408 | * For example, there is a single hw queue (hctx) and two CPU queues (ctx0 | |
2409 | * for CPU0, and ctx1 for CPU1). | |
2410 | * | |
2411 | * Now CPU1 is just onlined and a request is inserted into ctx1->rq_list | |
2412 | * and set bit0 in pending bitmap as ctx1->index_hw is still zero. | |
2413 | * | |
2c3ad667 JA |
2414 | * And then while running hw queue, blk_mq_flush_busy_ctxs() finds bit0 is set |
2415 | * in pending bitmap and tries to retrieve requests in hctx->ctxs[0]->rq_list. | |
2416 | * But htx->ctxs[0] is a pointer to ctx0, so the request in ctx1->rq_list is | |
2417 | * ignored. | |
65d5291e SAS |
2418 | */ |
2419 | static int blk_mq_queue_reinit_prepare(unsigned int cpu) | |
2420 | { | |
2421 | cpumask_copy(&cpuhp_online_new, cpu_online_mask); | |
2422 | cpumask_set_cpu(cpu, &cpuhp_online_new); | |
2423 | blk_mq_queue_reinit_work(); | |
2424 | return 0; | |
320ae51f JA |
2425 | } |
2426 | ||
a5164405 JA |
2427 | static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set) |
2428 | { | |
2429 | int i; | |
2430 | ||
cc71a6f4 JA |
2431 | for (i = 0; i < set->nr_hw_queues; i++) |
2432 | if (!__blk_mq_alloc_rq_map(set, i)) | |
a5164405 | 2433 | goto out_unwind; |
a5164405 JA |
2434 | |
2435 | return 0; | |
2436 | ||
2437 | out_unwind: | |
2438 | while (--i >= 0) | |
cc71a6f4 | 2439 | blk_mq_free_rq_map(set->tags[i]); |
a5164405 | 2440 | |
a5164405 JA |
2441 | return -ENOMEM; |
2442 | } | |
2443 | ||
2444 | /* | |
2445 | * Allocate the request maps associated with this tag_set. Note that this | |
2446 | * may reduce the depth asked for, if memory is tight. set->queue_depth | |
2447 | * will be updated to reflect the allocated depth. | |
2448 | */ | |
2449 | static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set) | |
2450 | { | |
2451 | unsigned int depth; | |
2452 | int err; | |
2453 | ||
2454 | depth = set->queue_depth; | |
2455 | do { | |
2456 | err = __blk_mq_alloc_rq_maps(set); | |
2457 | if (!err) | |
2458 | break; | |
2459 | ||
2460 | set->queue_depth >>= 1; | |
2461 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) { | |
2462 | err = -ENOMEM; | |
2463 | break; | |
2464 | } | |
2465 | } while (set->queue_depth); | |
2466 | ||
2467 | if (!set->queue_depth || err) { | |
2468 | pr_err("blk-mq: failed to allocate request map\n"); | |
2469 | return -ENOMEM; | |
2470 | } | |
2471 | ||
2472 | if (depth != set->queue_depth) | |
2473 | pr_info("blk-mq: reduced tag depth (%u -> %u)\n", | |
2474 | depth, set->queue_depth); | |
2475 | ||
2476 | return 0; | |
2477 | } | |
2478 | ||
a4391c64 JA |
2479 | /* |
2480 | * Alloc a tag set to be associated with one or more request queues. | |
2481 | * May fail with EINVAL for various error conditions. May adjust the | |
2482 | * requested depth down, if if it too large. In that case, the set | |
2483 | * value will be stored in set->queue_depth. | |
2484 | */ | |
24d2f903 CH |
2485 | int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set) |
2486 | { | |
da695ba2 CH |
2487 | int ret; |
2488 | ||
205fb5f5 BVA |
2489 | BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS); |
2490 | ||
24d2f903 CH |
2491 | if (!set->nr_hw_queues) |
2492 | return -EINVAL; | |
a4391c64 | 2493 | if (!set->queue_depth) |
24d2f903 CH |
2494 | return -EINVAL; |
2495 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) | |
2496 | return -EINVAL; | |
2497 | ||
7d7e0f90 | 2498 | if (!set->ops->queue_rq) |
24d2f903 CH |
2499 | return -EINVAL; |
2500 | ||
a4391c64 JA |
2501 | if (set->queue_depth > BLK_MQ_MAX_DEPTH) { |
2502 | pr_info("blk-mq: reduced tag depth to %u\n", | |
2503 | BLK_MQ_MAX_DEPTH); | |
2504 | set->queue_depth = BLK_MQ_MAX_DEPTH; | |
2505 | } | |
24d2f903 | 2506 | |
6637fadf SL |
2507 | /* |
2508 | * If a crashdump is active, then we are potentially in a very | |
2509 | * memory constrained environment. Limit us to 1 queue and | |
2510 | * 64 tags to prevent using too much memory. | |
2511 | */ | |
2512 | if (is_kdump_kernel()) { | |
2513 | set->nr_hw_queues = 1; | |
2514 | set->queue_depth = min(64U, set->queue_depth); | |
2515 | } | |
868f2f0b KB |
2516 | /* |
2517 | * There is no use for more h/w queues than cpus. | |
2518 | */ | |
2519 | if (set->nr_hw_queues > nr_cpu_ids) | |
2520 | set->nr_hw_queues = nr_cpu_ids; | |
6637fadf | 2521 | |
868f2f0b | 2522 | set->tags = kzalloc_node(nr_cpu_ids * sizeof(struct blk_mq_tags *), |
24d2f903 CH |
2523 | GFP_KERNEL, set->numa_node); |
2524 | if (!set->tags) | |
a5164405 | 2525 | return -ENOMEM; |
24d2f903 | 2526 | |
da695ba2 CH |
2527 | ret = -ENOMEM; |
2528 | set->mq_map = kzalloc_node(sizeof(*set->mq_map) * nr_cpu_ids, | |
2529 | GFP_KERNEL, set->numa_node); | |
bdd17e75 CH |
2530 | if (!set->mq_map) |
2531 | goto out_free_tags; | |
2532 | ||
da695ba2 CH |
2533 | if (set->ops->map_queues) |
2534 | ret = set->ops->map_queues(set); | |
2535 | else | |
2536 | ret = blk_mq_map_queues(set); | |
2537 | if (ret) | |
2538 | goto out_free_mq_map; | |
2539 | ||
2540 | ret = blk_mq_alloc_rq_maps(set); | |
2541 | if (ret) | |
bdd17e75 | 2542 | goto out_free_mq_map; |
24d2f903 | 2543 | |
0d2602ca JA |
2544 | mutex_init(&set->tag_list_lock); |
2545 | INIT_LIST_HEAD(&set->tag_list); | |
2546 | ||
24d2f903 | 2547 | return 0; |
bdd17e75 CH |
2548 | |
2549 | out_free_mq_map: | |
2550 | kfree(set->mq_map); | |
2551 | set->mq_map = NULL; | |
2552 | out_free_tags: | |
5676e7b6 RE |
2553 | kfree(set->tags); |
2554 | set->tags = NULL; | |
da695ba2 | 2555 | return ret; |
24d2f903 CH |
2556 | } |
2557 | EXPORT_SYMBOL(blk_mq_alloc_tag_set); | |
2558 | ||
2559 | void blk_mq_free_tag_set(struct blk_mq_tag_set *set) | |
2560 | { | |
2561 | int i; | |
2562 | ||
cc71a6f4 JA |
2563 | for (i = 0; i < nr_cpu_ids; i++) |
2564 | blk_mq_free_map_and_requests(set, i); | |
484b4061 | 2565 | |
bdd17e75 CH |
2566 | kfree(set->mq_map); |
2567 | set->mq_map = NULL; | |
2568 | ||
981bd189 | 2569 | kfree(set->tags); |
5676e7b6 | 2570 | set->tags = NULL; |
24d2f903 CH |
2571 | } |
2572 | EXPORT_SYMBOL(blk_mq_free_tag_set); | |
2573 | ||
e3a2b3f9 JA |
2574 | int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr) |
2575 | { | |
2576 | struct blk_mq_tag_set *set = q->tag_set; | |
2577 | struct blk_mq_hw_ctx *hctx; | |
2578 | int i, ret; | |
2579 | ||
bd166ef1 | 2580 | if (!set) |
e3a2b3f9 JA |
2581 | return -EINVAL; |
2582 | ||
70f36b60 JA |
2583 | blk_mq_freeze_queue(q); |
2584 | blk_mq_quiesce_queue(q); | |
2585 | ||
e3a2b3f9 JA |
2586 | ret = 0; |
2587 | queue_for_each_hw_ctx(q, hctx, i) { | |
e9137d4b KB |
2588 | if (!hctx->tags) |
2589 | continue; | |
bd166ef1 JA |
2590 | /* |
2591 | * If we're using an MQ scheduler, just update the scheduler | |
2592 | * queue depth. This is similar to what the old code would do. | |
2593 | */ | |
70f36b60 JA |
2594 | if (!hctx->sched_tags) { |
2595 | ret = blk_mq_tag_update_depth(hctx, &hctx->tags, | |
2596 | min(nr, set->queue_depth), | |
2597 | false); | |
2598 | } else { | |
2599 | ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags, | |
2600 | nr, true); | |
2601 | } | |
e3a2b3f9 JA |
2602 | if (ret) |
2603 | break; | |
2604 | } | |
2605 | ||
2606 | if (!ret) | |
2607 | q->nr_requests = nr; | |
2608 | ||
70f36b60 JA |
2609 | blk_mq_unfreeze_queue(q); |
2610 | blk_mq_start_stopped_hw_queues(q, true); | |
2611 | ||
e3a2b3f9 JA |
2612 | return ret; |
2613 | } | |
2614 | ||
868f2f0b KB |
2615 | void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues) |
2616 | { | |
2617 | struct request_queue *q; | |
2618 | ||
2619 | if (nr_hw_queues > nr_cpu_ids) | |
2620 | nr_hw_queues = nr_cpu_ids; | |
2621 | if (nr_hw_queues < 1 || nr_hw_queues == set->nr_hw_queues) | |
2622 | return; | |
2623 | ||
2624 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
2625 | blk_mq_freeze_queue(q); | |
2626 | ||
2627 | set->nr_hw_queues = nr_hw_queues; | |
2628 | list_for_each_entry(q, &set->tag_list, tag_set_list) { | |
2629 | blk_mq_realloc_hw_ctxs(set, q); | |
2630 | ||
f6f94300 JB |
2631 | /* |
2632 | * Manually set the make_request_fn as blk_queue_make_request | |
2633 | * resets a lot of the queue settings. | |
2634 | */ | |
868f2f0b | 2635 | if (q->nr_hw_queues > 1) |
f6f94300 | 2636 | q->make_request_fn = blk_mq_make_request; |
868f2f0b | 2637 | else |
f6f94300 | 2638 | q->make_request_fn = blk_sq_make_request; |
868f2f0b KB |
2639 | |
2640 | blk_mq_queue_reinit(q, cpu_online_mask); | |
2641 | } | |
2642 | ||
2643 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
2644 | blk_mq_unfreeze_queue(q); | |
2645 | } | |
2646 | EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues); | |
2647 | ||
64f1c21e JA |
2648 | static unsigned long blk_mq_poll_nsecs(struct request_queue *q, |
2649 | struct blk_mq_hw_ctx *hctx, | |
2650 | struct request *rq) | |
2651 | { | |
2652 | struct blk_rq_stat stat[2]; | |
2653 | unsigned long ret = 0; | |
2654 | ||
2655 | /* | |
2656 | * If stats collection isn't on, don't sleep but turn it on for | |
2657 | * future users | |
2658 | */ | |
2659 | if (!blk_stat_enable(q)) | |
2660 | return 0; | |
2661 | ||
2662 | /* | |
2663 | * We don't have to do this once per IO, should optimize this | |
2664 | * to just use the current window of stats until it changes | |
2665 | */ | |
2666 | memset(&stat, 0, sizeof(stat)); | |
2667 | blk_hctx_stat_get(hctx, stat); | |
2668 | ||
2669 | /* | |
2670 | * As an optimistic guess, use half of the mean service time | |
2671 | * for this type of request. We can (and should) make this smarter. | |
2672 | * For instance, if the completion latencies are tight, we can | |
2673 | * get closer than just half the mean. This is especially | |
2674 | * important on devices where the completion latencies are longer | |
2675 | * than ~10 usec. | |
2676 | */ | |
2677 | if (req_op(rq) == REQ_OP_READ && stat[BLK_STAT_READ].nr_samples) | |
2678 | ret = (stat[BLK_STAT_READ].mean + 1) / 2; | |
2679 | else if (req_op(rq) == REQ_OP_WRITE && stat[BLK_STAT_WRITE].nr_samples) | |
2680 | ret = (stat[BLK_STAT_WRITE].mean + 1) / 2; | |
2681 | ||
2682 | return ret; | |
2683 | } | |
2684 | ||
06426adf | 2685 | static bool blk_mq_poll_hybrid_sleep(struct request_queue *q, |
64f1c21e | 2686 | struct blk_mq_hw_ctx *hctx, |
06426adf JA |
2687 | struct request *rq) |
2688 | { | |
2689 | struct hrtimer_sleeper hs; | |
2690 | enum hrtimer_mode mode; | |
64f1c21e | 2691 | unsigned int nsecs; |
06426adf JA |
2692 | ktime_t kt; |
2693 | ||
64f1c21e JA |
2694 | if (test_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags)) |
2695 | return false; | |
2696 | ||
2697 | /* | |
2698 | * poll_nsec can be: | |
2699 | * | |
2700 | * -1: don't ever hybrid sleep | |
2701 | * 0: use half of prev avg | |
2702 | * >0: use this specific value | |
2703 | */ | |
2704 | if (q->poll_nsec == -1) | |
2705 | return false; | |
2706 | else if (q->poll_nsec > 0) | |
2707 | nsecs = q->poll_nsec; | |
2708 | else | |
2709 | nsecs = blk_mq_poll_nsecs(q, hctx, rq); | |
2710 | ||
2711 | if (!nsecs) | |
06426adf JA |
2712 | return false; |
2713 | ||
2714 | set_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags); | |
2715 | ||
2716 | /* | |
2717 | * This will be replaced with the stats tracking code, using | |
2718 | * 'avg_completion_time / 2' as the pre-sleep target. | |
2719 | */ | |
8b0e1953 | 2720 | kt = nsecs; |
06426adf JA |
2721 | |
2722 | mode = HRTIMER_MODE_REL; | |
2723 | hrtimer_init_on_stack(&hs.timer, CLOCK_MONOTONIC, mode); | |
2724 | hrtimer_set_expires(&hs.timer, kt); | |
2725 | ||
2726 | hrtimer_init_sleeper(&hs, current); | |
2727 | do { | |
2728 | if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags)) | |
2729 | break; | |
2730 | set_current_state(TASK_UNINTERRUPTIBLE); | |
2731 | hrtimer_start_expires(&hs.timer, mode); | |
2732 | if (hs.task) | |
2733 | io_schedule(); | |
2734 | hrtimer_cancel(&hs.timer); | |
2735 | mode = HRTIMER_MODE_ABS; | |
2736 | } while (hs.task && !signal_pending(current)); | |
2737 | ||
2738 | __set_current_state(TASK_RUNNING); | |
2739 | destroy_hrtimer_on_stack(&hs.timer); | |
2740 | return true; | |
2741 | } | |
2742 | ||
bbd7bb70 JA |
2743 | static bool __blk_mq_poll(struct blk_mq_hw_ctx *hctx, struct request *rq) |
2744 | { | |
2745 | struct request_queue *q = hctx->queue; | |
2746 | long state; | |
2747 | ||
06426adf JA |
2748 | /* |
2749 | * If we sleep, have the caller restart the poll loop to reset | |
2750 | * the state. Like for the other success return cases, the | |
2751 | * caller is responsible for checking if the IO completed. If | |
2752 | * the IO isn't complete, we'll get called again and will go | |
2753 | * straight to the busy poll loop. | |
2754 | */ | |
64f1c21e | 2755 | if (blk_mq_poll_hybrid_sleep(q, hctx, rq)) |
06426adf JA |
2756 | return true; |
2757 | ||
bbd7bb70 JA |
2758 | hctx->poll_considered++; |
2759 | ||
2760 | state = current->state; | |
2761 | while (!need_resched()) { | |
2762 | int ret; | |
2763 | ||
2764 | hctx->poll_invoked++; | |
2765 | ||
2766 | ret = q->mq_ops->poll(hctx, rq->tag); | |
2767 | if (ret > 0) { | |
2768 | hctx->poll_success++; | |
2769 | set_current_state(TASK_RUNNING); | |
2770 | return true; | |
2771 | } | |
2772 | ||
2773 | if (signal_pending_state(state, current)) | |
2774 | set_current_state(TASK_RUNNING); | |
2775 | ||
2776 | if (current->state == TASK_RUNNING) | |
2777 | return true; | |
2778 | if (ret < 0) | |
2779 | break; | |
2780 | cpu_relax(); | |
2781 | } | |
2782 | ||
2783 | return false; | |
2784 | } | |
2785 | ||
2786 | bool blk_mq_poll(struct request_queue *q, blk_qc_t cookie) | |
2787 | { | |
2788 | struct blk_mq_hw_ctx *hctx; | |
2789 | struct blk_plug *plug; | |
2790 | struct request *rq; | |
2791 | ||
2792 | if (!q->mq_ops || !q->mq_ops->poll || !blk_qc_t_valid(cookie) || | |
2793 | !test_bit(QUEUE_FLAG_POLL, &q->queue_flags)) | |
2794 | return false; | |
2795 | ||
2796 | plug = current->plug; | |
2797 | if (plug) | |
2798 | blk_flush_plug_list(plug, false); | |
2799 | ||
2800 | hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)]; | |
bd166ef1 JA |
2801 | if (!blk_qc_t_is_internal(cookie)) |
2802 | rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie)); | |
2803 | else | |
2804 | rq = blk_mq_tag_to_rq(hctx->sched_tags, blk_qc_t_to_tag(cookie)); | |
bbd7bb70 JA |
2805 | |
2806 | return __blk_mq_poll(hctx, rq); | |
2807 | } | |
2808 | EXPORT_SYMBOL_GPL(blk_mq_poll); | |
2809 | ||
676141e4 JA |
2810 | void blk_mq_disable_hotplug(void) |
2811 | { | |
2812 | mutex_lock(&all_q_mutex); | |
2813 | } | |
2814 | ||
2815 | void blk_mq_enable_hotplug(void) | |
2816 | { | |
2817 | mutex_unlock(&all_q_mutex); | |
2818 | } | |
2819 | ||
320ae51f JA |
2820 | static int __init blk_mq_init(void) |
2821 | { | |
9467f859 TG |
2822 | cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL, |
2823 | blk_mq_hctx_notify_dead); | |
320ae51f | 2824 | |
65d5291e SAS |
2825 | cpuhp_setup_state_nocalls(CPUHP_BLK_MQ_PREPARE, "block/mq:prepare", |
2826 | blk_mq_queue_reinit_prepare, | |
2827 | blk_mq_queue_reinit_dead); | |
320ae51f JA |
2828 | return 0; |
2829 | } | |
2830 | subsys_initcall(blk_mq_init); |