]>
Commit | Line | Data |
---|---|---|
75bb4625 JA |
1 | /* |
2 | * Block multiqueue core code | |
3 | * | |
4 | * Copyright (C) 2013-2014 Jens Axboe | |
5 | * Copyright (C) 2013-2014 Christoph Hellwig | |
6 | */ | |
320ae51f JA |
7 | #include <linux/kernel.h> |
8 | #include <linux/module.h> | |
9 | #include <linux/backing-dev.h> | |
10 | #include <linux/bio.h> | |
11 | #include <linux/blkdev.h> | |
f75782e4 | 12 | #include <linux/kmemleak.h> |
320ae51f JA |
13 | #include <linux/mm.h> |
14 | #include <linux/init.h> | |
15 | #include <linux/slab.h> | |
16 | #include <linux/workqueue.h> | |
17 | #include <linux/smp.h> | |
18 | #include <linux/llist.h> | |
19 | #include <linux/list_sort.h> | |
20 | #include <linux/cpu.h> | |
21 | #include <linux/cache.h> | |
22 | #include <linux/sched/sysctl.h> | |
23 | #include <linux/delay.h> | |
aedcd72f | 24 | #include <linux/crash_dump.h> |
88c7b2b7 | 25 | #include <linux/prefetch.h> |
320ae51f JA |
26 | |
27 | #include <trace/events/block.h> | |
28 | ||
29 | #include <linux/blk-mq.h> | |
30 | #include "blk.h" | |
31 | #include "blk-mq.h" | |
32 | #include "blk-mq-tag.h" | |
cf43e6be | 33 | #include "blk-stat.h" |
87760e5e | 34 | #include "blk-wbt.h" |
bd166ef1 | 35 | #include "blk-mq-sched.h" |
320ae51f JA |
36 | |
37 | static DEFINE_MUTEX(all_q_mutex); | |
38 | static LIST_HEAD(all_q_list); | |
39 | ||
320ae51f JA |
40 | /* |
41 | * Check if any of the ctx's have pending work in this hardware queue | |
42 | */ | |
43 | static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx) | |
44 | { | |
bd166ef1 JA |
45 | return sbitmap_any_bit_set(&hctx->ctx_map) || |
46 | !list_empty_careful(&hctx->dispatch) || | |
47 | blk_mq_sched_has_work(hctx); | |
1429d7c9 JA |
48 | } |
49 | ||
320ae51f JA |
50 | /* |
51 | * Mark this ctx as having pending work in this hardware queue | |
52 | */ | |
53 | static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx, | |
54 | struct blk_mq_ctx *ctx) | |
55 | { | |
88459642 OS |
56 | if (!sbitmap_test_bit(&hctx->ctx_map, ctx->index_hw)) |
57 | sbitmap_set_bit(&hctx->ctx_map, ctx->index_hw); | |
1429d7c9 JA |
58 | } |
59 | ||
60 | static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx, | |
61 | struct blk_mq_ctx *ctx) | |
62 | { | |
88459642 | 63 | sbitmap_clear_bit(&hctx->ctx_map, ctx->index_hw); |
320ae51f JA |
64 | } |
65 | ||
b4c6a028 | 66 | void blk_mq_freeze_queue_start(struct request_queue *q) |
43a5e4e2 | 67 | { |
4ecd4fef | 68 | int freeze_depth; |
cddd5d17 | 69 | |
4ecd4fef CH |
70 | freeze_depth = atomic_inc_return(&q->mq_freeze_depth); |
71 | if (freeze_depth == 1) { | |
3ef28e83 | 72 | percpu_ref_kill(&q->q_usage_counter); |
b94ec296 | 73 | blk_mq_run_hw_queues(q, false); |
cddd5d17 | 74 | } |
f3af020b | 75 | } |
b4c6a028 | 76 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_start); |
f3af020b TH |
77 | |
78 | static void blk_mq_freeze_queue_wait(struct request_queue *q) | |
79 | { | |
3ef28e83 | 80 | wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter)); |
43a5e4e2 ML |
81 | } |
82 | ||
f3af020b TH |
83 | /* |
84 | * Guarantee no request is in use, so we can change any data structure of | |
85 | * the queue afterward. | |
86 | */ | |
3ef28e83 | 87 | void blk_freeze_queue(struct request_queue *q) |
f3af020b | 88 | { |
3ef28e83 DW |
89 | /* |
90 | * In the !blk_mq case we are only calling this to kill the | |
91 | * q_usage_counter, otherwise this increases the freeze depth | |
92 | * and waits for it to return to zero. For this reason there is | |
93 | * no blk_unfreeze_queue(), and blk_freeze_queue() is not | |
94 | * exported to drivers as the only user for unfreeze is blk_mq. | |
95 | */ | |
f3af020b TH |
96 | blk_mq_freeze_queue_start(q); |
97 | blk_mq_freeze_queue_wait(q); | |
98 | } | |
3ef28e83 DW |
99 | |
100 | void blk_mq_freeze_queue(struct request_queue *q) | |
101 | { | |
102 | /* | |
103 | * ...just an alias to keep freeze and unfreeze actions balanced | |
104 | * in the blk_mq_* namespace | |
105 | */ | |
106 | blk_freeze_queue(q); | |
107 | } | |
c761d96b | 108 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue); |
f3af020b | 109 | |
b4c6a028 | 110 | void blk_mq_unfreeze_queue(struct request_queue *q) |
320ae51f | 111 | { |
4ecd4fef | 112 | int freeze_depth; |
320ae51f | 113 | |
4ecd4fef CH |
114 | freeze_depth = atomic_dec_return(&q->mq_freeze_depth); |
115 | WARN_ON_ONCE(freeze_depth < 0); | |
116 | if (!freeze_depth) { | |
3ef28e83 | 117 | percpu_ref_reinit(&q->q_usage_counter); |
320ae51f | 118 | wake_up_all(&q->mq_freeze_wq); |
add703fd | 119 | } |
320ae51f | 120 | } |
b4c6a028 | 121 | EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue); |
320ae51f | 122 | |
6a83e74d BVA |
123 | /** |
124 | * blk_mq_quiesce_queue() - wait until all ongoing queue_rq calls have finished | |
125 | * @q: request queue. | |
126 | * | |
127 | * Note: this function does not prevent that the struct request end_io() | |
128 | * callback function is invoked. Additionally, it is not prevented that | |
129 | * new queue_rq() calls occur unless the queue has been stopped first. | |
130 | */ | |
131 | void blk_mq_quiesce_queue(struct request_queue *q) | |
132 | { | |
133 | struct blk_mq_hw_ctx *hctx; | |
134 | unsigned int i; | |
135 | bool rcu = false; | |
136 | ||
137 | blk_mq_stop_hw_queues(q); | |
138 | ||
139 | queue_for_each_hw_ctx(q, hctx, i) { | |
140 | if (hctx->flags & BLK_MQ_F_BLOCKING) | |
141 | synchronize_srcu(&hctx->queue_rq_srcu); | |
142 | else | |
143 | rcu = true; | |
144 | } | |
145 | if (rcu) | |
146 | synchronize_rcu(); | |
147 | } | |
148 | EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue); | |
149 | ||
aed3ea94 JA |
150 | void blk_mq_wake_waiters(struct request_queue *q) |
151 | { | |
152 | struct blk_mq_hw_ctx *hctx; | |
153 | unsigned int i; | |
154 | ||
155 | queue_for_each_hw_ctx(q, hctx, i) | |
156 | if (blk_mq_hw_queue_mapped(hctx)) | |
157 | blk_mq_tag_wakeup_all(hctx->tags, true); | |
3fd5940c KB |
158 | |
159 | /* | |
160 | * If we are called because the queue has now been marked as | |
161 | * dying, we need to ensure that processes currently waiting on | |
162 | * the queue are notified as well. | |
163 | */ | |
164 | wake_up_all(&q->mq_freeze_wq); | |
aed3ea94 JA |
165 | } |
166 | ||
320ae51f JA |
167 | bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx) |
168 | { | |
169 | return blk_mq_has_free_tags(hctx->tags); | |
170 | } | |
171 | EXPORT_SYMBOL(blk_mq_can_queue); | |
172 | ||
2c3ad667 JA |
173 | void blk_mq_rq_ctx_init(struct request_queue *q, struct blk_mq_ctx *ctx, |
174 | struct request *rq, unsigned int op) | |
320ae51f | 175 | { |
af76e555 CH |
176 | INIT_LIST_HEAD(&rq->queuelist); |
177 | /* csd/requeue_work/fifo_time is initialized before use */ | |
178 | rq->q = q; | |
320ae51f | 179 | rq->mq_ctx = ctx; |
ef295ecf | 180 | rq->cmd_flags = op; |
e8064021 CH |
181 | if (blk_queue_io_stat(q)) |
182 | rq->rq_flags |= RQF_IO_STAT; | |
af76e555 CH |
183 | /* do not touch atomic flags, it needs atomic ops against the timer */ |
184 | rq->cpu = -1; | |
af76e555 CH |
185 | INIT_HLIST_NODE(&rq->hash); |
186 | RB_CLEAR_NODE(&rq->rb_node); | |
af76e555 CH |
187 | rq->rq_disk = NULL; |
188 | rq->part = NULL; | |
3ee32372 | 189 | rq->start_time = jiffies; |
af76e555 CH |
190 | #ifdef CONFIG_BLK_CGROUP |
191 | rq->rl = NULL; | |
0fec08b4 | 192 | set_start_time_ns(rq); |
af76e555 CH |
193 | rq->io_start_time_ns = 0; |
194 | #endif | |
195 | rq->nr_phys_segments = 0; | |
196 | #if defined(CONFIG_BLK_DEV_INTEGRITY) | |
197 | rq->nr_integrity_segments = 0; | |
198 | #endif | |
af76e555 CH |
199 | rq->special = NULL; |
200 | /* tag was already set */ | |
201 | rq->errors = 0; | |
af76e555 | 202 | |
6f4a1626 TB |
203 | rq->cmd = rq->__cmd; |
204 | ||
af76e555 CH |
205 | rq->extra_len = 0; |
206 | rq->sense_len = 0; | |
207 | rq->resid_len = 0; | |
208 | rq->sense = NULL; | |
209 | ||
af76e555 | 210 | INIT_LIST_HEAD(&rq->timeout_list); |
f6be4fb4 JA |
211 | rq->timeout = 0; |
212 | ||
af76e555 CH |
213 | rq->end_io = NULL; |
214 | rq->end_io_data = NULL; | |
215 | rq->next_rq = NULL; | |
216 | ||
ef295ecf | 217 | ctx->rq_dispatched[op_is_sync(op)]++; |
320ae51f | 218 | } |
2c3ad667 | 219 | EXPORT_SYMBOL_GPL(blk_mq_rq_ctx_init); |
320ae51f | 220 | |
2c3ad667 JA |
221 | struct request *__blk_mq_alloc_request(struct blk_mq_alloc_data *data, |
222 | unsigned int op) | |
5dee8577 CH |
223 | { |
224 | struct request *rq; | |
225 | unsigned int tag; | |
226 | ||
cb96a42c | 227 | tag = blk_mq_get_tag(data); |
5dee8577 | 228 | if (tag != BLK_MQ_TAG_FAIL) { |
bd166ef1 JA |
229 | struct blk_mq_tags *tags = blk_mq_tags_from_data(data); |
230 | ||
231 | rq = tags->static_rqs[tag]; | |
5dee8577 | 232 | |
bd166ef1 JA |
233 | if (data->flags & BLK_MQ_REQ_INTERNAL) { |
234 | rq->tag = -1; | |
235 | rq->internal_tag = tag; | |
236 | } else { | |
200e86b3 JA |
237 | if (blk_mq_tag_busy(data->hctx)) { |
238 | rq->rq_flags = RQF_MQ_INFLIGHT; | |
239 | atomic_inc(&data->hctx->nr_active); | |
240 | } | |
bd166ef1 JA |
241 | rq->tag = tag; |
242 | rq->internal_tag = -1; | |
243 | } | |
244 | ||
ef295ecf | 245 | blk_mq_rq_ctx_init(data->q, data->ctx, rq, op); |
5dee8577 CH |
246 | return rq; |
247 | } | |
248 | ||
249 | return NULL; | |
250 | } | |
2c3ad667 | 251 | EXPORT_SYMBOL_GPL(__blk_mq_alloc_request); |
5dee8577 | 252 | |
6f3b0e8b CH |
253 | struct request *blk_mq_alloc_request(struct request_queue *q, int rw, |
254 | unsigned int flags) | |
320ae51f | 255 | { |
5a797e00 | 256 | struct blk_mq_alloc_data alloc_data = { .flags = flags }; |
bd166ef1 | 257 | struct request *rq; |
a492f075 | 258 | int ret; |
320ae51f | 259 | |
6f3b0e8b | 260 | ret = blk_queue_enter(q, flags & BLK_MQ_REQ_NOWAIT); |
a492f075 JL |
261 | if (ret) |
262 | return ERR_PTR(ret); | |
320ae51f | 263 | |
bd166ef1 | 264 | rq = blk_mq_sched_get_request(q, NULL, rw, &alloc_data); |
841bac2c | 265 | |
bd166ef1 JA |
266 | blk_mq_put_ctx(alloc_data.ctx); |
267 | blk_queue_exit(q); | |
268 | ||
269 | if (!rq) | |
a492f075 | 270 | return ERR_PTR(-EWOULDBLOCK); |
0c4de0f3 CH |
271 | |
272 | rq->__data_len = 0; | |
273 | rq->__sector = (sector_t) -1; | |
274 | rq->bio = rq->biotail = NULL; | |
320ae51f JA |
275 | return rq; |
276 | } | |
4bb659b1 | 277 | EXPORT_SYMBOL(blk_mq_alloc_request); |
320ae51f | 278 | |
1f5bd336 ML |
279 | struct request *blk_mq_alloc_request_hctx(struct request_queue *q, int rw, |
280 | unsigned int flags, unsigned int hctx_idx) | |
281 | { | |
282 | struct blk_mq_hw_ctx *hctx; | |
283 | struct blk_mq_ctx *ctx; | |
284 | struct request *rq; | |
285 | struct blk_mq_alloc_data alloc_data; | |
286 | int ret; | |
287 | ||
288 | /* | |
289 | * If the tag allocator sleeps we could get an allocation for a | |
290 | * different hardware context. No need to complicate the low level | |
291 | * allocator for this for the rare use case of a command tied to | |
292 | * a specific queue. | |
293 | */ | |
294 | if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT))) | |
295 | return ERR_PTR(-EINVAL); | |
296 | ||
297 | if (hctx_idx >= q->nr_hw_queues) | |
298 | return ERR_PTR(-EIO); | |
299 | ||
300 | ret = blk_queue_enter(q, true); | |
301 | if (ret) | |
302 | return ERR_PTR(ret); | |
303 | ||
c8712c6a CH |
304 | /* |
305 | * Check if the hardware context is actually mapped to anything. | |
306 | * If not tell the caller that it should skip this queue. | |
307 | */ | |
1f5bd336 | 308 | hctx = q->queue_hw_ctx[hctx_idx]; |
c8712c6a CH |
309 | if (!blk_mq_hw_queue_mapped(hctx)) { |
310 | ret = -EXDEV; | |
311 | goto out_queue_exit; | |
312 | } | |
1f5bd336 ML |
313 | ctx = __blk_mq_get_ctx(q, cpumask_first(hctx->cpumask)); |
314 | ||
315 | blk_mq_set_alloc_data(&alloc_data, q, flags, ctx, hctx); | |
ef295ecf | 316 | rq = __blk_mq_alloc_request(&alloc_data, rw); |
1f5bd336 | 317 | if (!rq) { |
c8712c6a CH |
318 | ret = -EWOULDBLOCK; |
319 | goto out_queue_exit; | |
1f5bd336 ML |
320 | } |
321 | ||
322 | return rq; | |
c8712c6a CH |
323 | |
324 | out_queue_exit: | |
325 | blk_queue_exit(q); | |
326 | return ERR_PTR(ret); | |
1f5bd336 ML |
327 | } |
328 | EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx); | |
329 | ||
bd166ef1 JA |
330 | void __blk_mq_finish_request(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx, |
331 | struct request *rq) | |
320ae51f | 332 | { |
bd166ef1 | 333 | const int sched_tag = rq->internal_tag; |
320ae51f JA |
334 | struct request_queue *q = rq->q; |
335 | ||
e8064021 | 336 | if (rq->rq_flags & RQF_MQ_INFLIGHT) |
0d2602ca | 337 | atomic_dec(&hctx->nr_active); |
87760e5e JA |
338 | |
339 | wbt_done(q->rq_wb, &rq->issue_stat); | |
e8064021 | 340 | rq->rq_flags = 0; |
0d2602ca | 341 | |
af76e555 | 342 | clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags); |
06426adf | 343 | clear_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags); |
bd166ef1 JA |
344 | if (rq->tag != -1) |
345 | blk_mq_put_tag(hctx, hctx->tags, ctx, rq->tag); | |
346 | if (sched_tag != -1) | |
347 | blk_mq_sched_completed_request(hctx, rq); | |
3ef28e83 | 348 | blk_queue_exit(q); |
320ae51f JA |
349 | } |
350 | ||
bd166ef1 | 351 | static void blk_mq_finish_hctx_request(struct blk_mq_hw_ctx *hctx, |
16a3c2a7 | 352 | struct request *rq) |
320ae51f JA |
353 | { |
354 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
320ae51f JA |
355 | |
356 | ctx->rq_completed[rq_is_sync(rq)]++; | |
bd166ef1 JA |
357 | __blk_mq_finish_request(hctx, ctx, rq); |
358 | } | |
359 | ||
360 | void blk_mq_finish_request(struct request *rq) | |
361 | { | |
362 | blk_mq_finish_hctx_request(blk_mq_map_queue(rq->q, rq->mq_ctx->cpu), rq); | |
7c7f2f2b | 363 | } |
7c7f2f2b JA |
364 | |
365 | void blk_mq_free_request(struct request *rq) | |
366 | { | |
bd166ef1 | 367 | blk_mq_sched_put_request(rq); |
320ae51f | 368 | } |
1a3b595a | 369 | EXPORT_SYMBOL_GPL(blk_mq_free_request); |
320ae51f | 370 | |
c8a446ad | 371 | inline void __blk_mq_end_request(struct request *rq, int error) |
320ae51f | 372 | { |
0d11e6ac ML |
373 | blk_account_io_done(rq); |
374 | ||
91b63639 | 375 | if (rq->end_io) { |
87760e5e | 376 | wbt_done(rq->q->rq_wb, &rq->issue_stat); |
320ae51f | 377 | rq->end_io(rq, error); |
91b63639 CH |
378 | } else { |
379 | if (unlikely(blk_bidi_rq(rq))) | |
380 | blk_mq_free_request(rq->next_rq); | |
320ae51f | 381 | blk_mq_free_request(rq); |
91b63639 | 382 | } |
320ae51f | 383 | } |
c8a446ad | 384 | EXPORT_SYMBOL(__blk_mq_end_request); |
63151a44 | 385 | |
c8a446ad | 386 | void blk_mq_end_request(struct request *rq, int error) |
63151a44 CH |
387 | { |
388 | if (blk_update_request(rq, error, blk_rq_bytes(rq))) | |
389 | BUG(); | |
c8a446ad | 390 | __blk_mq_end_request(rq, error); |
63151a44 | 391 | } |
c8a446ad | 392 | EXPORT_SYMBOL(blk_mq_end_request); |
320ae51f | 393 | |
30a91cb4 | 394 | static void __blk_mq_complete_request_remote(void *data) |
320ae51f | 395 | { |
3d6efbf6 | 396 | struct request *rq = data; |
320ae51f | 397 | |
30a91cb4 | 398 | rq->q->softirq_done_fn(rq); |
320ae51f | 399 | } |
320ae51f | 400 | |
ed851860 | 401 | static void blk_mq_ipi_complete_request(struct request *rq) |
320ae51f JA |
402 | { |
403 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
38535201 | 404 | bool shared = false; |
320ae51f JA |
405 | int cpu; |
406 | ||
38535201 | 407 | if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) { |
30a91cb4 CH |
408 | rq->q->softirq_done_fn(rq); |
409 | return; | |
410 | } | |
320ae51f JA |
411 | |
412 | cpu = get_cpu(); | |
38535201 CH |
413 | if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags)) |
414 | shared = cpus_share_cache(cpu, ctx->cpu); | |
415 | ||
416 | if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) { | |
30a91cb4 | 417 | rq->csd.func = __blk_mq_complete_request_remote; |
3d6efbf6 CH |
418 | rq->csd.info = rq; |
419 | rq->csd.flags = 0; | |
c46fff2a | 420 | smp_call_function_single_async(ctx->cpu, &rq->csd); |
3d6efbf6 | 421 | } else { |
30a91cb4 | 422 | rq->q->softirq_done_fn(rq); |
3d6efbf6 | 423 | } |
320ae51f JA |
424 | put_cpu(); |
425 | } | |
30a91cb4 | 426 | |
cf43e6be JA |
427 | static void blk_mq_stat_add(struct request *rq) |
428 | { | |
429 | if (rq->rq_flags & RQF_STATS) { | |
430 | /* | |
431 | * We could rq->mq_ctx here, but there's less of a risk | |
432 | * of races if we have the completion event add the stats | |
433 | * to the local software queue. | |
434 | */ | |
435 | struct blk_mq_ctx *ctx; | |
436 | ||
437 | ctx = __blk_mq_get_ctx(rq->q, raw_smp_processor_id()); | |
438 | blk_stat_add(&ctx->stat[rq_data_dir(rq)], rq); | |
439 | } | |
440 | } | |
441 | ||
1fa8cc52 | 442 | static void __blk_mq_complete_request(struct request *rq) |
ed851860 JA |
443 | { |
444 | struct request_queue *q = rq->q; | |
445 | ||
cf43e6be JA |
446 | blk_mq_stat_add(rq); |
447 | ||
ed851860 | 448 | if (!q->softirq_done_fn) |
c8a446ad | 449 | blk_mq_end_request(rq, rq->errors); |
ed851860 JA |
450 | else |
451 | blk_mq_ipi_complete_request(rq); | |
452 | } | |
453 | ||
30a91cb4 CH |
454 | /** |
455 | * blk_mq_complete_request - end I/O on a request | |
456 | * @rq: the request being processed | |
457 | * | |
458 | * Description: | |
459 | * Ends all I/O on a request. It does not handle partial completions. | |
460 | * The actual completion happens out-of-order, through a IPI handler. | |
461 | **/ | |
f4829a9b | 462 | void blk_mq_complete_request(struct request *rq, int error) |
30a91cb4 | 463 | { |
95f09684 JA |
464 | struct request_queue *q = rq->q; |
465 | ||
466 | if (unlikely(blk_should_fake_timeout(q))) | |
30a91cb4 | 467 | return; |
f4829a9b CH |
468 | if (!blk_mark_rq_complete(rq)) { |
469 | rq->errors = error; | |
ed851860 | 470 | __blk_mq_complete_request(rq); |
f4829a9b | 471 | } |
30a91cb4 CH |
472 | } |
473 | EXPORT_SYMBOL(blk_mq_complete_request); | |
320ae51f | 474 | |
973c0191 KB |
475 | int blk_mq_request_started(struct request *rq) |
476 | { | |
477 | return test_bit(REQ_ATOM_STARTED, &rq->atomic_flags); | |
478 | } | |
479 | EXPORT_SYMBOL_GPL(blk_mq_request_started); | |
480 | ||
e2490073 | 481 | void blk_mq_start_request(struct request *rq) |
320ae51f JA |
482 | { |
483 | struct request_queue *q = rq->q; | |
484 | ||
bd166ef1 JA |
485 | blk_mq_sched_started_request(rq); |
486 | ||
320ae51f JA |
487 | trace_block_rq_issue(q, rq); |
488 | ||
742ee69b | 489 | rq->resid_len = blk_rq_bytes(rq); |
91b63639 CH |
490 | if (unlikely(blk_bidi_rq(rq))) |
491 | rq->next_rq->resid_len = blk_rq_bytes(rq->next_rq); | |
742ee69b | 492 | |
cf43e6be JA |
493 | if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) { |
494 | blk_stat_set_issue_time(&rq->issue_stat); | |
495 | rq->rq_flags |= RQF_STATS; | |
87760e5e | 496 | wbt_issue(q->rq_wb, &rq->issue_stat); |
cf43e6be JA |
497 | } |
498 | ||
2b8393b4 | 499 | blk_add_timer(rq); |
87ee7b11 | 500 | |
538b7534 JA |
501 | /* |
502 | * Ensure that ->deadline is visible before set the started | |
503 | * flag and clear the completed flag. | |
504 | */ | |
505 | smp_mb__before_atomic(); | |
506 | ||
87ee7b11 JA |
507 | /* |
508 | * Mark us as started and clear complete. Complete might have been | |
509 | * set if requeue raced with timeout, which then marked it as | |
510 | * complete. So be sure to clear complete again when we start | |
511 | * the request, otherwise we'll ignore the completion event. | |
512 | */ | |
4b570521 JA |
513 | if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) |
514 | set_bit(REQ_ATOM_STARTED, &rq->atomic_flags); | |
515 | if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags)) | |
516 | clear_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags); | |
49f5baa5 CH |
517 | |
518 | if (q->dma_drain_size && blk_rq_bytes(rq)) { | |
519 | /* | |
520 | * Make sure space for the drain appears. We know we can do | |
521 | * this because max_hw_segments has been adjusted to be one | |
522 | * fewer than the device can handle. | |
523 | */ | |
524 | rq->nr_phys_segments++; | |
525 | } | |
320ae51f | 526 | } |
e2490073 | 527 | EXPORT_SYMBOL(blk_mq_start_request); |
320ae51f | 528 | |
ed0791b2 | 529 | static void __blk_mq_requeue_request(struct request *rq) |
320ae51f JA |
530 | { |
531 | struct request_queue *q = rq->q; | |
532 | ||
533 | trace_block_rq_requeue(q, rq); | |
87760e5e | 534 | wbt_requeue(q->rq_wb, &rq->issue_stat); |
bd166ef1 | 535 | blk_mq_sched_requeue_request(rq); |
49f5baa5 | 536 | |
e2490073 CH |
537 | if (test_and_clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) { |
538 | if (q->dma_drain_size && blk_rq_bytes(rq)) | |
539 | rq->nr_phys_segments--; | |
540 | } | |
320ae51f JA |
541 | } |
542 | ||
2b053aca | 543 | void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list) |
ed0791b2 | 544 | { |
ed0791b2 | 545 | __blk_mq_requeue_request(rq); |
ed0791b2 | 546 | |
ed0791b2 | 547 | BUG_ON(blk_queued_rq(rq)); |
2b053aca | 548 | blk_mq_add_to_requeue_list(rq, true, kick_requeue_list); |
ed0791b2 CH |
549 | } |
550 | EXPORT_SYMBOL(blk_mq_requeue_request); | |
551 | ||
6fca6a61 CH |
552 | static void blk_mq_requeue_work(struct work_struct *work) |
553 | { | |
554 | struct request_queue *q = | |
2849450a | 555 | container_of(work, struct request_queue, requeue_work.work); |
6fca6a61 CH |
556 | LIST_HEAD(rq_list); |
557 | struct request *rq, *next; | |
558 | unsigned long flags; | |
559 | ||
560 | spin_lock_irqsave(&q->requeue_lock, flags); | |
561 | list_splice_init(&q->requeue_list, &rq_list); | |
562 | spin_unlock_irqrestore(&q->requeue_lock, flags); | |
563 | ||
564 | list_for_each_entry_safe(rq, next, &rq_list, queuelist) { | |
e8064021 | 565 | if (!(rq->rq_flags & RQF_SOFTBARRIER)) |
6fca6a61 CH |
566 | continue; |
567 | ||
e8064021 | 568 | rq->rq_flags &= ~RQF_SOFTBARRIER; |
6fca6a61 | 569 | list_del_init(&rq->queuelist); |
bd166ef1 | 570 | blk_mq_sched_insert_request(rq, true, false, false); |
6fca6a61 CH |
571 | } |
572 | ||
573 | while (!list_empty(&rq_list)) { | |
574 | rq = list_entry(rq_list.next, struct request, queuelist); | |
575 | list_del_init(&rq->queuelist); | |
bd166ef1 | 576 | blk_mq_sched_insert_request(rq, false, false, false); |
6fca6a61 CH |
577 | } |
578 | ||
52d7f1b5 | 579 | blk_mq_run_hw_queues(q, false); |
6fca6a61 CH |
580 | } |
581 | ||
2b053aca BVA |
582 | void blk_mq_add_to_requeue_list(struct request *rq, bool at_head, |
583 | bool kick_requeue_list) | |
6fca6a61 CH |
584 | { |
585 | struct request_queue *q = rq->q; | |
586 | unsigned long flags; | |
587 | ||
588 | /* | |
589 | * We abuse this flag that is otherwise used by the I/O scheduler to | |
590 | * request head insertation from the workqueue. | |
591 | */ | |
e8064021 | 592 | BUG_ON(rq->rq_flags & RQF_SOFTBARRIER); |
6fca6a61 CH |
593 | |
594 | spin_lock_irqsave(&q->requeue_lock, flags); | |
595 | if (at_head) { | |
e8064021 | 596 | rq->rq_flags |= RQF_SOFTBARRIER; |
6fca6a61 CH |
597 | list_add(&rq->queuelist, &q->requeue_list); |
598 | } else { | |
599 | list_add_tail(&rq->queuelist, &q->requeue_list); | |
600 | } | |
601 | spin_unlock_irqrestore(&q->requeue_lock, flags); | |
2b053aca BVA |
602 | |
603 | if (kick_requeue_list) | |
604 | blk_mq_kick_requeue_list(q); | |
6fca6a61 CH |
605 | } |
606 | EXPORT_SYMBOL(blk_mq_add_to_requeue_list); | |
607 | ||
608 | void blk_mq_kick_requeue_list(struct request_queue *q) | |
609 | { | |
2849450a | 610 | kblockd_schedule_delayed_work(&q->requeue_work, 0); |
6fca6a61 CH |
611 | } |
612 | EXPORT_SYMBOL(blk_mq_kick_requeue_list); | |
613 | ||
2849450a MS |
614 | void blk_mq_delay_kick_requeue_list(struct request_queue *q, |
615 | unsigned long msecs) | |
616 | { | |
617 | kblockd_schedule_delayed_work(&q->requeue_work, | |
618 | msecs_to_jiffies(msecs)); | |
619 | } | |
620 | EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list); | |
621 | ||
1885b24d JA |
622 | void blk_mq_abort_requeue_list(struct request_queue *q) |
623 | { | |
624 | unsigned long flags; | |
625 | LIST_HEAD(rq_list); | |
626 | ||
627 | spin_lock_irqsave(&q->requeue_lock, flags); | |
628 | list_splice_init(&q->requeue_list, &rq_list); | |
629 | spin_unlock_irqrestore(&q->requeue_lock, flags); | |
630 | ||
631 | while (!list_empty(&rq_list)) { | |
632 | struct request *rq; | |
633 | ||
634 | rq = list_first_entry(&rq_list, struct request, queuelist); | |
635 | list_del_init(&rq->queuelist); | |
636 | rq->errors = -EIO; | |
637 | blk_mq_end_request(rq, rq->errors); | |
638 | } | |
639 | } | |
640 | EXPORT_SYMBOL(blk_mq_abort_requeue_list); | |
641 | ||
0e62f51f JA |
642 | struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag) |
643 | { | |
88c7b2b7 JA |
644 | if (tag < tags->nr_tags) { |
645 | prefetch(tags->rqs[tag]); | |
4ee86bab | 646 | return tags->rqs[tag]; |
88c7b2b7 | 647 | } |
4ee86bab HR |
648 | |
649 | return NULL; | |
24d2f903 CH |
650 | } |
651 | EXPORT_SYMBOL(blk_mq_tag_to_rq); | |
652 | ||
320ae51f | 653 | struct blk_mq_timeout_data { |
46f92d42 CH |
654 | unsigned long next; |
655 | unsigned int next_set; | |
320ae51f JA |
656 | }; |
657 | ||
90415837 | 658 | void blk_mq_rq_timed_out(struct request *req, bool reserved) |
320ae51f | 659 | { |
f8a5b122 | 660 | const struct blk_mq_ops *ops = req->q->mq_ops; |
46f92d42 | 661 | enum blk_eh_timer_return ret = BLK_EH_RESET_TIMER; |
87ee7b11 JA |
662 | |
663 | /* | |
664 | * We know that complete is set at this point. If STARTED isn't set | |
665 | * anymore, then the request isn't active and the "timeout" should | |
666 | * just be ignored. This can happen due to the bitflag ordering. | |
667 | * Timeout first checks if STARTED is set, and if it is, assumes | |
668 | * the request is active. But if we race with completion, then | |
669 | * we both flags will get cleared. So check here again, and ignore | |
670 | * a timeout event with a request that isn't active. | |
671 | */ | |
46f92d42 CH |
672 | if (!test_bit(REQ_ATOM_STARTED, &req->atomic_flags)) |
673 | return; | |
87ee7b11 | 674 | |
46f92d42 | 675 | if (ops->timeout) |
0152fb6b | 676 | ret = ops->timeout(req, reserved); |
46f92d42 CH |
677 | |
678 | switch (ret) { | |
679 | case BLK_EH_HANDLED: | |
680 | __blk_mq_complete_request(req); | |
681 | break; | |
682 | case BLK_EH_RESET_TIMER: | |
683 | blk_add_timer(req); | |
684 | blk_clear_rq_complete(req); | |
685 | break; | |
686 | case BLK_EH_NOT_HANDLED: | |
687 | break; | |
688 | default: | |
689 | printk(KERN_ERR "block: bad eh return: %d\n", ret); | |
690 | break; | |
691 | } | |
87ee7b11 | 692 | } |
5b3f25fc | 693 | |
81481eb4 CH |
694 | static void blk_mq_check_expired(struct blk_mq_hw_ctx *hctx, |
695 | struct request *rq, void *priv, bool reserved) | |
696 | { | |
697 | struct blk_mq_timeout_data *data = priv; | |
87ee7b11 | 698 | |
eb130dbf KB |
699 | if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) { |
700 | /* | |
701 | * If a request wasn't started before the queue was | |
702 | * marked dying, kill it here or it'll go unnoticed. | |
703 | */ | |
a59e0f57 KB |
704 | if (unlikely(blk_queue_dying(rq->q))) { |
705 | rq->errors = -EIO; | |
706 | blk_mq_end_request(rq, rq->errors); | |
707 | } | |
46f92d42 | 708 | return; |
eb130dbf | 709 | } |
87ee7b11 | 710 | |
46f92d42 CH |
711 | if (time_after_eq(jiffies, rq->deadline)) { |
712 | if (!blk_mark_rq_complete(rq)) | |
0152fb6b | 713 | blk_mq_rq_timed_out(rq, reserved); |
46f92d42 CH |
714 | } else if (!data->next_set || time_after(data->next, rq->deadline)) { |
715 | data->next = rq->deadline; | |
716 | data->next_set = 1; | |
717 | } | |
87ee7b11 JA |
718 | } |
719 | ||
287922eb | 720 | static void blk_mq_timeout_work(struct work_struct *work) |
320ae51f | 721 | { |
287922eb CH |
722 | struct request_queue *q = |
723 | container_of(work, struct request_queue, timeout_work); | |
81481eb4 CH |
724 | struct blk_mq_timeout_data data = { |
725 | .next = 0, | |
726 | .next_set = 0, | |
727 | }; | |
81481eb4 | 728 | int i; |
320ae51f | 729 | |
71f79fb3 GKB |
730 | /* A deadlock might occur if a request is stuck requiring a |
731 | * timeout at the same time a queue freeze is waiting | |
732 | * completion, since the timeout code would not be able to | |
733 | * acquire the queue reference here. | |
734 | * | |
735 | * That's why we don't use blk_queue_enter here; instead, we use | |
736 | * percpu_ref_tryget directly, because we need to be able to | |
737 | * obtain a reference even in the short window between the queue | |
738 | * starting to freeze, by dropping the first reference in | |
739 | * blk_mq_freeze_queue_start, and the moment the last request is | |
740 | * consumed, marked by the instant q_usage_counter reaches | |
741 | * zero. | |
742 | */ | |
743 | if (!percpu_ref_tryget(&q->q_usage_counter)) | |
287922eb CH |
744 | return; |
745 | ||
0bf6cd5b | 746 | blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &data); |
320ae51f | 747 | |
81481eb4 CH |
748 | if (data.next_set) { |
749 | data.next = blk_rq_timeout(round_jiffies_up(data.next)); | |
750 | mod_timer(&q->timeout, data.next); | |
0d2602ca | 751 | } else { |
0bf6cd5b CH |
752 | struct blk_mq_hw_ctx *hctx; |
753 | ||
f054b56c ML |
754 | queue_for_each_hw_ctx(q, hctx, i) { |
755 | /* the hctx may be unmapped, so check it here */ | |
756 | if (blk_mq_hw_queue_mapped(hctx)) | |
757 | blk_mq_tag_idle(hctx); | |
758 | } | |
0d2602ca | 759 | } |
287922eb | 760 | blk_queue_exit(q); |
320ae51f JA |
761 | } |
762 | ||
763 | /* | |
764 | * Reverse check our software queue for entries that we could potentially | |
765 | * merge with. Currently includes a hand-wavy stop count of 8, to not spend | |
766 | * too much time checking for merges. | |
767 | */ | |
768 | static bool blk_mq_attempt_merge(struct request_queue *q, | |
769 | struct blk_mq_ctx *ctx, struct bio *bio) | |
770 | { | |
771 | struct request *rq; | |
772 | int checked = 8; | |
773 | ||
774 | list_for_each_entry_reverse(rq, &ctx->rq_list, queuelist) { | |
775 | int el_ret; | |
776 | ||
777 | if (!checked--) | |
778 | break; | |
779 | ||
780 | if (!blk_rq_merge_ok(rq, bio)) | |
781 | continue; | |
782 | ||
783 | el_ret = blk_try_merge(rq, bio); | |
bd166ef1 JA |
784 | if (el_ret == ELEVATOR_NO_MERGE) |
785 | continue; | |
786 | ||
787 | if (!blk_mq_sched_allow_merge(q, rq, bio)) | |
788 | break; | |
789 | ||
320ae51f JA |
790 | if (el_ret == ELEVATOR_BACK_MERGE) { |
791 | if (bio_attempt_back_merge(q, rq, bio)) { | |
792 | ctx->rq_merged++; | |
793 | return true; | |
794 | } | |
795 | break; | |
796 | } else if (el_ret == ELEVATOR_FRONT_MERGE) { | |
797 | if (bio_attempt_front_merge(q, rq, bio)) { | |
798 | ctx->rq_merged++; | |
799 | return true; | |
800 | } | |
801 | break; | |
802 | } | |
803 | } | |
804 | ||
805 | return false; | |
806 | } | |
807 | ||
88459642 OS |
808 | struct flush_busy_ctx_data { |
809 | struct blk_mq_hw_ctx *hctx; | |
810 | struct list_head *list; | |
811 | }; | |
812 | ||
813 | static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data) | |
814 | { | |
815 | struct flush_busy_ctx_data *flush_data = data; | |
816 | struct blk_mq_hw_ctx *hctx = flush_data->hctx; | |
817 | struct blk_mq_ctx *ctx = hctx->ctxs[bitnr]; | |
818 | ||
819 | sbitmap_clear_bit(sb, bitnr); | |
820 | spin_lock(&ctx->lock); | |
821 | list_splice_tail_init(&ctx->rq_list, flush_data->list); | |
822 | spin_unlock(&ctx->lock); | |
823 | return true; | |
824 | } | |
825 | ||
1429d7c9 JA |
826 | /* |
827 | * Process software queues that have been marked busy, splicing them | |
828 | * to the for-dispatch | |
829 | */ | |
2c3ad667 | 830 | void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list) |
1429d7c9 | 831 | { |
88459642 OS |
832 | struct flush_busy_ctx_data data = { |
833 | .hctx = hctx, | |
834 | .list = list, | |
835 | }; | |
1429d7c9 | 836 | |
88459642 | 837 | sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data); |
1429d7c9 | 838 | } |
2c3ad667 | 839 | EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs); |
1429d7c9 | 840 | |
703fd1c0 JA |
841 | static inline unsigned int queued_to_index(unsigned int queued) |
842 | { | |
843 | if (!queued) | |
844 | return 0; | |
1429d7c9 | 845 | |
703fd1c0 | 846 | return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1); |
1429d7c9 JA |
847 | } |
848 | ||
bd166ef1 JA |
849 | static bool blk_mq_get_driver_tag(struct request *rq, |
850 | struct blk_mq_hw_ctx **hctx, bool wait) | |
851 | { | |
852 | struct blk_mq_alloc_data data = { | |
853 | .q = rq->q, | |
854 | .ctx = rq->mq_ctx, | |
855 | .hctx = blk_mq_map_queue(rq->q, rq->mq_ctx->cpu), | |
856 | .flags = wait ? 0 : BLK_MQ_REQ_NOWAIT, | |
857 | }; | |
858 | ||
859 | if (blk_mq_hctx_stopped(data.hctx)) | |
860 | return false; | |
861 | ||
862 | if (rq->tag != -1) { | |
863 | done: | |
864 | if (hctx) | |
865 | *hctx = data.hctx; | |
866 | return true; | |
867 | } | |
868 | ||
869 | rq->tag = blk_mq_get_tag(&data); | |
870 | if (rq->tag >= 0) { | |
200e86b3 JA |
871 | if (blk_mq_tag_busy(data.hctx)) { |
872 | rq->rq_flags |= RQF_MQ_INFLIGHT; | |
873 | atomic_inc(&data.hctx->nr_active); | |
874 | } | |
bd166ef1 JA |
875 | data.hctx->tags->rqs[rq->tag] = rq; |
876 | goto done; | |
877 | } | |
878 | ||
879 | return false; | |
880 | } | |
881 | ||
882 | /* | |
883 | * If we fail getting a driver tag because all the driver tags are already | |
884 | * assigned and on the dispatch list, BUT the first entry does not have a | |
885 | * tag, then we could deadlock. For that case, move entries with assigned | |
886 | * driver tags to the front, leaving the set of tagged requests in the | |
887 | * same order, and the untagged set in the same order. | |
888 | */ | |
889 | static bool reorder_tags_to_front(struct list_head *list) | |
890 | { | |
891 | struct request *rq, *tmp, *first = NULL; | |
892 | ||
893 | list_for_each_entry_safe_reverse(rq, tmp, list, queuelist) { | |
894 | if (rq == first) | |
895 | break; | |
896 | if (rq->tag != -1) { | |
897 | list_move(&rq->queuelist, list); | |
898 | if (!first) | |
899 | first = rq; | |
900 | } | |
901 | } | |
902 | ||
903 | return first != NULL; | |
904 | } | |
905 | ||
f04c3df3 | 906 | bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx *hctx, struct list_head *list) |
320ae51f JA |
907 | { |
908 | struct request_queue *q = hctx->queue; | |
320ae51f | 909 | struct request *rq; |
74c45052 JA |
910 | LIST_HEAD(driver_list); |
911 | struct list_head *dptr; | |
f04c3df3 | 912 | int queued, ret = BLK_MQ_RQ_QUEUE_OK; |
320ae51f | 913 | |
74c45052 JA |
914 | /* |
915 | * Start off with dptr being NULL, so we start the first request | |
916 | * immediately, even if we have more pending. | |
917 | */ | |
918 | dptr = NULL; | |
919 | ||
320ae51f JA |
920 | /* |
921 | * Now process all the entries, sending them to the driver. | |
922 | */ | |
1429d7c9 | 923 | queued = 0; |
f04c3df3 | 924 | while (!list_empty(list)) { |
74c45052 | 925 | struct blk_mq_queue_data bd; |
320ae51f | 926 | |
f04c3df3 | 927 | rq = list_first_entry(list, struct request, queuelist); |
bd166ef1 JA |
928 | if (!blk_mq_get_driver_tag(rq, &hctx, false)) { |
929 | if (!queued && reorder_tags_to_front(list)) | |
930 | continue; | |
3c782d67 JA |
931 | |
932 | /* | |
933 | * We failed getting a driver tag. Mark the queue(s) | |
934 | * as needing a restart. Retry getting a tag again, | |
935 | * in case the needed IO completed right before we | |
936 | * marked the queue as needing a restart. | |
937 | */ | |
bd166ef1 | 938 | blk_mq_sched_mark_restart(hctx); |
3c782d67 JA |
939 | if (!blk_mq_get_driver_tag(rq, &hctx, false)) |
940 | break; | |
bd166ef1 | 941 | } |
320ae51f | 942 | list_del_init(&rq->queuelist); |
320ae51f | 943 | |
74c45052 JA |
944 | bd.rq = rq; |
945 | bd.list = dptr; | |
f04c3df3 | 946 | bd.last = list_empty(list); |
74c45052 JA |
947 | |
948 | ret = q->mq_ops->queue_rq(hctx, &bd); | |
320ae51f JA |
949 | switch (ret) { |
950 | case BLK_MQ_RQ_QUEUE_OK: | |
951 | queued++; | |
52b9c330 | 952 | break; |
320ae51f | 953 | case BLK_MQ_RQ_QUEUE_BUSY: |
f04c3df3 | 954 | list_add(&rq->queuelist, list); |
ed0791b2 | 955 | __blk_mq_requeue_request(rq); |
320ae51f JA |
956 | break; |
957 | default: | |
958 | pr_err("blk-mq: bad return on queue: %d\n", ret); | |
320ae51f | 959 | case BLK_MQ_RQ_QUEUE_ERROR: |
1e93b8c2 | 960 | rq->errors = -EIO; |
c8a446ad | 961 | blk_mq_end_request(rq, rq->errors); |
320ae51f JA |
962 | break; |
963 | } | |
964 | ||
965 | if (ret == BLK_MQ_RQ_QUEUE_BUSY) | |
966 | break; | |
74c45052 JA |
967 | |
968 | /* | |
969 | * We've done the first request. If we have more than 1 | |
970 | * left in the list, set dptr to defer issue. | |
971 | */ | |
f04c3df3 | 972 | if (!dptr && list->next != list->prev) |
74c45052 | 973 | dptr = &driver_list; |
320ae51f JA |
974 | } |
975 | ||
703fd1c0 | 976 | hctx->dispatched[queued_to_index(queued)]++; |
320ae51f JA |
977 | |
978 | /* | |
979 | * Any items that need requeuing? Stuff them into hctx->dispatch, | |
980 | * that is where we will continue on next queue run. | |
981 | */ | |
f04c3df3 | 982 | if (!list_empty(list)) { |
320ae51f | 983 | spin_lock(&hctx->lock); |
f04c3df3 | 984 | list_splice(list, &hctx->dispatch); |
320ae51f | 985 | spin_unlock(&hctx->lock); |
f04c3df3 | 986 | |
9ba52e58 SL |
987 | /* |
988 | * the queue is expected stopped with BLK_MQ_RQ_QUEUE_BUSY, but | |
989 | * it's possible the queue is stopped and restarted again | |
990 | * before this. Queue restart will dispatch requests. And since | |
991 | * requests in rq_list aren't added into hctx->dispatch yet, | |
992 | * the requests in rq_list might get lost. | |
993 | * | |
994 | * blk_mq_run_hw_queue() already checks the STOPPED bit | |
bd166ef1 JA |
995 | * |
996 | * If RESTART is set, then let completion restart the queue | |
997 | * instead of potentially looping here. | |
998 | */ | |
999 | if (!blk_mq_sched_needs_restart(hctx)) | |
1000 | blk_mq_run_hw_queue(hctx, true); | |
320ae51f | 1001 | } |
f04c3df3 JA |
1002 | |
1003 | return ret != BLK_MQ_RQ_QUEUE_BUSY; | |
1004 | } | |
1005 | ||
6a83e74d BVA |
1006 | static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx) |
1007 | { | |
1008 | int srcu_idx; | |
1009 | ||
1010 | WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) && | |
1011 | cpu_online(hctx->next_cpu)); | |
1012 | ||
1013 | if (!(hctx->flags & BLK_MQ_F_BLOCKING)) { | |
1014 | rcu_read_lock(); | |
bd166ef1 | 1015 | blk_mq_sched_dispatch_requests(hctx); |
6a83e74d BVA |
1016 | rcu_read_unlock(); |
1017 | } else { | |
1018 | srcu_idx = srcu_read_lock(&hctx->queue_rq_srcu); | |
bd166ef1 | 1019 | blk_mq_sched_dispatch_requests(hctx); |
6a83e74d BVA |
1020 | srcu_read_unlock(&hctx->queue_rq_srcu, srcu_idx); |
1021 | } | |
1022 | } | |
1023 | ||
506e931f JA |
1024 | /* |
1025 | * It'd be great if the workqueue API had a way to pass | |
1026 | * in a mask and had some smarts for more clever placement. | |
1027 | * For now we just round-robin here, switching for every | |
1028 | * BLK_MQ_CPU_WORK_BATCH queued items. | |
1029 | */ | |
1030 | static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx) | |
1031 | { | |
b657d7e6 CH |
1032 | if (hctx->queue->nr_hw_queues == 1) |
1033 | return WORK_CPU_UNBOUND; | |
506e931f JA |
1034 | |
1035 | if (--hctx->next_cpu_batch <= 0) { | |
c02ebfdd | 1036 | int next_cpu; |
506e931f JA |
1037 | |
1038 | next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask); | |
1039 | if (next_cpu >= nr_cpu_ids) | |
1040 | next_cpu = cpumask_first(hctx->cpumask); | |
1041 | ||
1042 | hctx->next_cpu = next_cpu; | |
1043 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; | |
1044 | } | |
1045 | ||
b657d7e6 | 1046 | return hctx->next_cpu; |
506e931f JA |
1047 | } |
1048 | ||
320ae51f JA |
1049 | void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) |
1050 | { | |
5d1b25c1 BVA |
1051 | if (unlikely(blk_mq_hctx_stopped(hctx) || |
1052 | !blk_mq_hw_queue_mapped(hctx))) | |
320ae51f JA |
1053 | return; |
1054 | ||
1b792f2f | 1055 | if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) { |
2a90d4aa PB |
1056 | int cpu = get_cpu(); |
1057 | if (cpumask_test_cpu(cpu, hctx->cpumask)) { | |
398205b8 | 1058 | __blk_mq_run_hw_queue(hctx); |
2a90d4aa | 1059 | put_cpu(); |
398205b8 PB |
1060 | return; |
1061 | } | |
e4043dcf | 1062 | |
2a90d4aa | 1063 | put_cpu(); |
e4043dcf | 1064 | } |
398205b8 | 1065 | |
27489a3c | 1066 | kblockd_schedule_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work); |
320ae51f JA |
1067 | } |
1068 | ||
b94ec296 | 1069 | void blk_mq_run_hw_queues(struct request_queue *q, bool async) |
320ae51f JA |
1070 | { |
1071 | struct blk_mq_hw_ctx *hctx; | |
1072 | int i; | |
1073 | ||
1074 | queue_for_each_hw_ctx(q, hctx, i) { | |
bd166ef1 | 1075 | if (!blk_mq_hctx_has_pending(hctx) || |
5d1b25c1 | 1076 | blk_mq_hctx_stopped(hctx)) |
320ae51f JA |
1077 | continue; |
1078 | ||
b94ec296 | 1079 | blk_mq_run_hw_queue(hctx, async); |
320ae51f JA |
1080 | } |
1081 | } | |
b94ec296 | 1082 | EXPORT_SYMBOL(blk_mq_run_hw_queues); |
320ae51f | 1083 | |
fd001443 BVA |
1084 | /** |
1085 | * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped | |
1086 | * @q: request queue. | |
1087 | * | |
1088 | * The caller is responsible for serializing this function against | |
1089 | * blk_mq_{start,stop}_hw_queue(). | |
1090 | */ | |
1091 | bool blk_mq_queue_stopped(struct request_queue *q) | |
1092 | { | |
1093 | struct blk_mq_hw_ctx *hctx; | |
1094 | int i; | |
1095 | ||
1096 | queue_for_each_hw_ctx(q, hctx, i) | |
1097 | if (blk_mq_hctx_stopped(hctx)) | |
1098 | return true; | |
1099 | ||
1100 | return false; | |
1101 | } | |
1102 | EXPORT_SYMBOL(blk_mq_queue_stopped); | |
1103 | ||
320ae51f JA |
1104 | void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx) |
1105 | { | |
27489a3c | 1106 | cancel_work(&hctx->run_work); |
70f4db63 | 1107 | cancel_delayed_work(&hctx->delay_work); |
320ae51f JA |
1108 | set_bit(BLK_MQ_S_STOPPED, &hctx->state); |
1109 | } | |
1110 | EXPORT_SYMBOL(blk_mq_stop_hw_queue); | |
1111 | ||
280d45f6 CH |
1112 | void blk_mq_stop_hw_queues(struct request_queue *q) |
1113 | { | |
1114 | struct blk_mq_hw_ctx *hctx; | |
1115 | int i; | |
1116 | ||
1117 | queue_for_each_hw_ctx(q, hctx, i) | |
1118 | blk_mq_stop_hw_queue(hctx); | |
1119 | } | |
1120 | EXPORT_SYMBOL(blk_mq_stop_hw_queues); | |
1121 | ||
320ae51f JA |
1122 | void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx) |
1123 | { | |
1124 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
e4043dcf | 1125 | |
0ffbce80 | 1126 | blk_mq_run_hw_queue(hctx, false); |
320ae51f JA |
1127 | } |
1128 | EXPORT_SYMBOL(blk_mq_start_hw_queue); | |
1129 | ||
2f268556 CH |
1130 | void blk_mq_start_hw_queues(struct request_queue *q) |
1131 | { | |
1132 | struct blk_mq_hw_ctx *hctx; | |
1133 | int i; | |
1134 | ||
1135 | queue_for_each_hw_ctx(q, hctx, i) | |
1136 | blk_mq_start_hw_queue(hctx); | |
1137 | } | |
1138 | EXPORT_SYMBOL(blk_mq_start_hw_queues); | |
1139 | ||
ae911c5e JA |
1140 | void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) |
1141 | { | |
1142 | if (!blk_mq_hctx_stopped(hctx)) | |
1143 | return; | |
1144 | ||
1145 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
1146 | blk_mq_run_hw_queue(hctx, async); | |
1147 | } | |
1148 | EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue); | |
1149 | ||
1b4a3258 | 1150 | void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async) |
320ae51f JA |
1151 | { |
1152 | struct blk_mq_hw_ctx *hctx; | |
1153 | int i; | |
1154 | ||
ae911c5e JA |
1155 | queue_for_each_hw_ctx(q, hctx, i) |
1156 | blk_mq_start_stopped_hw_queue(hctx, async); | |
320ae51f JA |
1157 | } |
1158 | EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues); | |
1159 | ||
70f4db63 | 1160 | static void blk_mq_run_work_fn(struct work_struct *work) |
320ae51f JA |
1161 | { |
1162 | struct blk_mq_hw_ctx *hctx; | |
1163 | ||
27489a3c | 1164 | hctx = container_of(work, struct blk_mq_hw_ctx, run_work); |
e4043dcf | 1165 | |
320ae51f JA |
1166 | __blk_mq_run_hw_queue(hctx); |
1167 | } | |
1168 | ||
70f4db63 CH |
1169 | static void blk_mq_delay_work_fn(struct work_struct *work) |
1170 | { | |
1171 | struct blk_mq_hw_ctx *hctx; | |
1172 | ||
1173 | hctx = container_of(work, struct blk_mq_hw_ctx, delay_work.work); | |
1174 | ||
1175 | if (test_and_clear_bit(BLK_MQ_S_STOPPED, &hctx->state)) | |
1176 | __blk_mq_run_hw_queue(hctx); | |
1177 | } | |
1178 | ||
1179 | void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs) | |
1180 | { | |
19c66e59 ML |
1181 | if (unlikely(!blk_mq_hw_queue_mapped(hctx))) |
1182 | return; | |
70f4db63 | 1183 | |
7e79dadc | 1184 | blk_mq_stop_hw_queue(hctx); |
b657d7e6 CH |
1185 | kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx), |
1186 | &hctx->delay_work, msecs_to_jiffies(msecs)); | |
70f4db63 CH |
1187 | } |
1188 | EXPORT_SYMBOL(blk_mq_delay_queue); | |
1189 | ||
cfd0c552 | 1190 | static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx, |
cfd0c552 ML |
1191 | struct request *rq, |
1192 | bool at_head) | |
320ae51f | 1193 | { |
e57690fe JA |
1194 | struct blk_mq_ctx *ctx = rq->mq_ctx; |
1195 | ||
01b983c9 JA |
1196 | trace_block_rq_insert(hctx->queue, rq); |
1197 | ||
72a0a36e CH |
1198 | if (at_head) |
1199 | list_add(&rq->queuelist, &ctx->rq_list); | |
1200 | else | |
1201 | list_add_tail(&rq->queuelist, &ctx->rq_list); | |
cfd0c552 | 1202 | } |
4bb659b1 | 1203 | |
2c3ad667 JA |
1204 | void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq, |
1205 | bool at_head) | |
cfd0c552 ML |
1206 | { |
1207 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
1208 | ||
e57690fe | 1209 | __blk_mq_insert_req_list(hctx, rq, at_head); |
320ae51f | 1210 | blk_mq_hctx_mark_pending(hctx, ctx); |
320ae51f JA |
1211 | } |
1212 | ||
bd166ef1 JA |
1213 | void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx, |
1214 | struct list_head *list) | |
320ae51f JA |
1215 | |
1216 | { | |
320ae51f JA |
1217 | /* |
1218 | * preemption doesn't flush plug list, so it's possible ctx->cpu is | |
1219 | * offline now | |
1220 | */ | |
1221 | spin_lock(&ctx->lock); | |
1222 | while (!list_empty(list)) { | |
1223 | struct request *rq; | |
1224 | ||
1225 | rq = list_first_entry(list, struct request, queuelist); | |
e57690fe | 1226 | BUG_ON(rq->mq_ctx != ctx); |
320ae51f | 1227 | list_del_init(&rq->queuelist); |
e57690fe | 1228 | __blk_mq_insert_req_list(hctx, rq, false); |
320ae51f | 1229 | } |
cfd0c552 | 1230 | blk_mq_hctx_mark_pending(hctx, ctx); |
320ae51f | 1231 | spin_unlock(&ctx->lock); |
320ae51f JA |
1232 | } |
1233 | ||
1234 | static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b) | |
1235 | { | |
1236 | struct request *rqa = container_of(a, struct request, queuelist); | |
1237 | struct request *rqb = container_of(b, struct request, queuelist); | |
1238 | ||
1239 | return !(rqa->mq_ctx < rqb->mq_ctx || | |
1240 | (rqa->mq_ctx == rqb->mq_ctx && | |
1241 | blk_rq_pos(rqa) < blk_rq_pos(rqb))); | |
1242 | } | |
1243 | ||
1244 | void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule) | |
1245 | { | |
1246 | struct blk_mq_ctx *this_ctx; | |
1247 | struct request_queue *this_q; | |
1248 | struct request *rq; | |
1249 | LIST_HEAD(list); | |
1250 | LIST_HEAD(ctx_list); | |
1251 | unsigned int depth; | |
1252 | ||
1253 | list_splice_init(&plug->mq_list, &list); | |
1254 | ||
1255 | list_sort(NULL, &list, plug_ctx_cmp); | |
1256 | ||
1257 | this_q = NULL; | |
1258 | this_ctx = NULL; | |
1259 | depth = 0; | |
1260 | ||
1261 | while (!list_empty(&list)) { | |
1262 | rq = list_entry_rq(list.next); | |
1263 | list_del_init(&rq->queuelist); | |
1264 | BUG_ON(!rq->q); | |
1265 | if (rq->mq_ctx != this_ctx) { | |
1266 | if (this_ctx) { | |
bd166ef1 JA |
1267 | trace_block_unplug(this_q, depth, from_schedule); |
1268 | blk_mq_sched_insert_requests(this_q, this_ctx, | |
1269 | &ctx_list, | |
1270 | from_schedule); | |
320ae51f JA |
1271 | } |
1272 | ||
1273 | this_ctx = rq->mq_ctx; | |
1274 | this_q = rq->q; | |
1275 | depth = 0; | |
1276 | } | |
1277 | ||
1278 | depth++; | |
1279 | list_add_tail(&rq->queuelist, &ctx_list); | |
1280 | } | |
1281 | ||
1282 | /* | |
1283 | * If 'this_ctx' is set, we know we have entries to complete | |
1284 | * on 'ctx_list'. Do those. | |
1285 | */ | |
1286 | if (this_ctx) { | |
bd166ef1 JA |
1287 | trace_block_unplug(this_q, depth, from_schedule); |
1288 | blk_mq_sched_insert_requests(this_q, this_ctx, &ctx_list, | |
1289 | from_schedule); | |
320ae51f JA |
1290 | } |
1291 | } | |
1292 | ||
1293 | static void blk_mq_bio_to_request(struct request *rq, struct bio *bio) | |
1294 | { | |
1295 | init_request_from_bio(rq, bio); | |
4b570521 | 1296 | |
6e85eaf3 | 1297 | blk_account_io_start(rq, true); |
320ae51f JA |
1298 | } |
1299 | ||
274a5843 JA |
1300 | static inline bool hctx_allow_merges(struct blk_mq_hw_ctx *hctx) |
1301 | { | |
1302 | return (hctx->flags & BLK_MQ_F_SHOULD_MERGE) && | |
1303 | !blk_queue_nomerges(hctx->queue); | |
1304 | } | |
1305 | ||
07068d5b JA |
1306 | static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx *hctx, |
1307 | struct blk_mq_ctx *ctx, | |
1308 | struct request *rq, struct bio *bio) | |
320ae51f | 1309 | { |
e18378a6 | 1310 | if (!hctx_allow_merges(hctx) || !bio_mergeable(bio)) { |
07068d5b JA |
1311 | blk_mq_bio_to_request(rq, bio); |
1312 | spin_lock(&ctx->lock); | |
1313 | insert_rq: | |
1314 | __blk_mq_insert_request(hctx, rq, false); | |
1315 | spin_unlock(&ctx->lock); | |
1316 | return false; | |
1317 | } else { | |
274a5843 JA |
1318 | struct request_queue *q = hctx->queue; |
1319 | ||
07068d5b JA |
1320 | spin_lock(&ctx->lock); |
1321 | if (!blk_mq_attempt_merge(q, ctx, bio)) { | |
1322 | blk_mq_bio_to_request(rq, bio); | |
1323 | goto insert_rq; | |
1324 | } | |
320ae51f | 1325 | |
07068d5b | 1326 | spin_unlock(&ctx->lock); |
bd166ef1 | 1327 | __blk_mq_finish_request(hctx, ctx, rq); |
07068d5b | 1328 | return true; |
14ec77f3 | 1329 | } |
07068d5b | 1330 | } |
14ec77f3 | 1331 | |
fd2d3326 JA |
1332 | static blk_qc_t request_to_qc_t(struct blk_mq_hw_ctx *hctx, struct request *rq) |
1333 | { | |
bd166ef1 JA |
1334 | if (rq->tag != -1) |
1335 | return blk_tag_to_qc_t(rq->tag, hctx->queue_num, false); | |
1336 | ||
1337 | return blk_tag_to_qc_t(rq->internal_tag, hctx->queue_num, true); | |
fd2d3326 JA |
1338 | } |
1339 | ||
066a4a73 | 1340 | static void blk_mq_try_issue_directly(struct request *rq, blk_qc_t *cookie) |
f984df1f | 1341 | { |
f984df1f | 1342 | struct request_queue *q = rq->q; |
f984df1f SL |
1343 | struct blk_mq_queue_data bd = { |
1344 | .rq = rq, | |
1345 | .list = NULL, | |
1346 | .last = 1 | |
1347 | }; | |
bd166ef1 JA |
1348 | struct blk_mq_hw_ctx *hctx; |
1349 | blk_qc_t new_cookie; | |
1350 | int ret; | |
f984df1f | 1351 | |
bd166ef1 | 1352 | if (q->elevator) |
2253efc8 BVA |
1353 | goto insert; |
1354 | ||
bd166ef1 JA |
1355 | if (!blk_mq_get_driver_tag(rq, &hctx, false)) |
1356 | goto insert; | |
1357 | ||
1358 | new_cookie = request_to_qc_t(hctx, rq); | |
1359 | ||
f984df1f SL |
1360 | /* |
1361 | * For OK queue, we are done. For error, kill it. Any other | |
1362 | * error (busy), just add it to our list as we previously | |
1363 | * would have done | |
1364 | */ | |
1365 | ret = q->mq_ops->queue_rq(hctx, &bd); | |
7b371636 JA |
1366 | if (ret == BLK_MQ_RQ_QUEUE_OK) { |
1367 | *cookie = new_cookie; | |
2253efc8 | 1368 | return; |
7b371636 | 1369 | } |
f984df1f | 1370 | |
7b371636 JA |
1371 | __blk_mq_requeue_request(rq); |
1372 | ||
1373 | if (ret == BLK_MQ_RQ_QUEUE_ERROR) { | |
1374 | *cookie = BLK_QC_T_NONE; | |
1375 | rq->errors = -EIO; | |
1376 | blk_mq_end_request(rq, rq->errors); | |
2253efc8 | 1377 | return; |
f984df1f | 1378 | } |
7b371636 | 1379 | |
2253efc8 | 1380 | insert: |
bd166ef1 | 1381 | blk_mq_sched_insert_request(rq, false, true, true); |
f984df1f SL |
1382 | } |
1383 | ||
07068d5b JA |
1384 | /* |
1385 | * Multiple hardware queue variant. This will not use per-process plugs, | |
1386 | * but will attempt to bypass the hctx queueing if we can go straight to | |
1387 | * hardware for SYNC IO. | |
1388 | */ | |
dece1635 | 1389 | static blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio) |
07068d5b | 1390 | { |
ef295ecf | 1391 | const int is_sync = op_is_sync(bio->bi_opf); |
1eff9d32 | 1392 | const int is_flush_fua = bio->bi_opf & (REQ_PREFLUSH | REQ_FUA); |
5a797e00 | 1393 | struct blk_mq_alloc_data data = { .flags = 0 }; |
07068d5b | 1394 | struct request *rq; |
6a83e74d | 1395 | unsigned int request_count = 0, srcu_idx; |
f984df1f | 1396 | struct blk_plug *plug; |
5b3f341f | 1397 | struct request *same_queue_rq = NULL; |
7b371636 | 1398 | blk_qc_t cookie; |
87760e5e | 1399 | unsigned int wb_acct; |
07068d5b JA |
1400 | |
1401 | blk_queue_bounce(q, &bio); | |
1402 | ||
1403 | if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) { | |
4246a0b6 | 1404 | bio_io_error(bio); |
dece1635 | 1405 | return BLK_QC_T_NONE; |
07068d5b JA |
1406 | } |
1407 | ||
54efd50b KO |
1408 | blk_queue_split(q, &bio, q->bio_split); |
1409 | ||
87c279e6 OS |
1410 | if (!is_flush_fua && !blk_queue_nomerges(q) && |
1411 | blk_attempt_plug_merge(q, bio, &request_count, &same_queue_rq)) | |
1412 | return BLK_QC_T_NONE; | |
f984df1f | 1413 | |
bd166ef1 JA |
1414 | if (blk_mq_sched_bio_merge(q, bio)) |
1415 | return BLK_QC_T_NONE; | |
1416 | ||
87760e5e JA |
1417 | wb_acct = wbt_wait(q->rq_wb, bio, NULL); |
1418 | ||
bd166ef1 JA |
1419 | trace_block_getrq(q, bio, bio->bi_opf); |
1420 | ||
1421 | rq = blk_mq_sched_get_request(q, bio, bio->bi_opf, &data); | |
87760e5e JA |
1422 | if (unlikely(!rq)) { |
1423 | __wbt_done(q->rq_wb, wb_acct); | |
dece1635 | 1424 | return BLK_QC_T_NONE; |
87760e5e JA |
1425 | } |
1426 | ||
1427 | wbt_track(&rq->issue_stat, wb_acct); | |
07068d5b | 1428 | |
fd2d3326 | 1429 | cookie = request_to_qc_t(data.hctx, rq); |
07068d5b JA |
1430 | |
1431 | if (unlikely(is_flush_fua)) { | |
1432 | blk_mq_bio_to_request(rq, bio); | |
bd166ef1 | 1433 | blk_mq_get_driver_tag(rq, NULL, true); |
07068d5b JA |
1434 | blk_insert_flush(rq); |
1435 | goto run_queue; | |
1436 | } | |
1437 | ||
f984df1f | 1438 | plug = current->plug; |
e167dfb5 JA |
1439 | /* |
1440 | * If the driver supports defer issued based on 'last', then | |
1441 | * queue it up like normal since we can potentially save some | |
1442 | * CPU this way. | |
1443 | */ | |
f984df1f SL |
1444 | if (((plug && !blk_queue_nomerges(q)) || is_sync) && |
1445 | !(data.hctx->flags & BLK_MQ_F_DEFER_ISSUE)) { | |
1446 | struct request *old_rq = NULL; | |
07068d5b JA |
1447 | |
1448 | blk_mq_bio_to_request(rq, bio); | |
07068d5b JA |
1449 | |
1450 | /* | |
6a83e74d | 1451 | * We do limited plugging. If the bio can be merged, do that. |
f984df1f SL |
1452 | * Otherwise the existing request in the plug list will be |
1453 | * issued. So the plug list will have one request at most | |
07068d5b | 1454 | */ |
f984df1f | 1455 | if (plug) { |
5b3f341f SL |
1456 | /* |
1457 | * The plug list might get flushed before this. If that | |
b094f89c JA |
1458 | * happens, same_queue_rq is invalid and plug list is |
1459 | * empty | |
1460 | */ | |
5b3f341f SL |
1461 | if (same_queue_rq && !list_empty(&plug->mq_list)) { |
1462 | old_rq = same_queue_rq; | |
f984df1f | 1463 | list_del_init(&old_rq->queuelist); |
07068d5b | 1464 | } |
f984df1f SL |
1465 | list_add_tail(&rq->queuelist, &plug->mq_list); |
1466 | } else /* is_sync */ | |
1467 | old_rq = rq; | |
1468 | blk_mq_put_ctx(data.ctx); | |
1469 | if (!old_rq) | |
7b371636 | 1470 | goto done; |
6a83e74d BVA |
1471 | |
1472 | if (!(data.hctx->flags & BLK_MQ_F_BLOCKING)) { | |
1473 | rcu_read_lock(); | |
066a4a73 | 1474 | blk_mq_try_issue_directly(old_rq, &cookie); |
6a83e74d BVA |
1475 | rcu_read_unlock(); |
1476 | } else { | |
1477 | srcu_idx = srcu_read_lock(&data.hctx->queue_rq_srcu); | |
066a4a73 | 1478 | blk_mq_try_issue_directly(old_rq, &cookie); |
6a83e74d BVA |
1479 | srcu_read_unlock(&data.hctx->queue_rq_srcu, srcu_idx); |
1480 | } | |
7b371636 | 1481 | goto done; |
07068d5b JA |
1482 | } |
1483 | ||
bd166ef1 JA |
1484 | if (q->elevator) { |
1485 | blk_mq_put_ctx(data.ctx); | |
1486 | blk_mq_bio_to_request(rq, bio); | |
0abad774 JA |
1487 | blk_mq_sched_insert_request(rq, false, true, |
1488 | !is_sync || is_flush_fua); | |
bd166ef1 JA |
1489 | goto done; |
1490 | } | |
07068d5b JA |
1491 | if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) { |
1492 | /* | |
1493 | * For a SYNC request, send it to the hardware immediately. For | |
1494 | * an ASYNC request, just ensure that we run it later on. The | |
1495 | * latter allows for merging opportunities and more efficient | |
1496 | * dispatching. | |
1497 | */ | |
1498 | run_queue: | |
1499 | blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua); | |
1500 | } | |
07068d5b | 1501 | blk_mq_put_ctx(data.ctx); |
7b371636 JA |
1502 | done: |
1503 | return cookie; | |
07068d5b JA |
1504 | } |
1505 | ||
1506 | /* | |
1507 | * Single hardware queue variant. This will attempt to use any per-process | |
1508 | * plug for merging and IO deferral. | |
1509 | */ | |
dece1635 | 1510 | static blk_qc_t blk_sq_make_request(struct request_queue *q, struct bio *bio) |
07068d5b | 1511 | { |
ef295ecf | 1512 | const int is_sync = op_is_sync(bio->bi_opf); |
1eff9d32 | 1513 | const int is_flush_fua = bio->bi_opf & (REQ_PREFLUSH | REQ_FUA); |
e6c4438b JM |
1514 | struct blk_plug *plug; |
1515 | unsigned int request_count = 0; | |
5a797e00 | 1516 | struct blk_mq_alloc_data data = { .flags = 0 }; |
07068d5b | 1517 | struct request *rq; |
7b371636 | 1518 | blk_qc_t cookie; |
87760e5e | 1519 | unsigned int wb_acct; |
07068d5b | 1520 | |
07068d5b JA |
1521 | blk_queue_bounce(q, &bio); |
1522 | ||
1523 | if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) { | |
4246a0b6 | 1524 | bio_io_error(bio); |
dece1635 | 1525 | return BLK_QC_T_NONE; |
07068d5b JA |
1526 | } |
1527 | ||
54efd50b KO |
1528 | blk_queue_split(q, &bio, q->bio_split); |
1529 | ||
87c279e6 OS |
1530 | if (!is_flush_fua && !blk_queue_nomerges(q)) { |
1531 | if (blk_attempt_plug_merge(q, bio, &request_count, NULL)) | |
1532 | return BLK_QC_T_NONE; | |
1533 | } else | |
1534 | request_count = blk_plug_queued_count(q); | |
07068d5b | 1535 | |
bd166ef1 JA |
1536 | if (blk_mq_sched_bio_merge(q, bio)) |
1537 | return BLK_QC_T_NONE; | |
1538 | ||
87760e5e JA |
1539 | wb_acct = wbt_wait(q->rq_wb, bio, NULL); |
1540 | ||
bd166ef1 JA |
1541 | trace_block_getrq(q, bio, bio->bi_opf); |
1542 | ||
1543 | rq = blk_mq_sched_get_request(q, bio, bio->bi_opf, &data); | |
87760e5e JA |
1544 | if (unlikely(!rq)) { |
1545 | __wbt_done(q->rq_wb, wb_acct); | |
dece1635 | 1546 | return BLK_QC_T_NONE; |
87760e5e JA |
1547 | } |
1548 | ||
1549 | wbt_track(&rq->issue_stat, wb_acct); | |
320ae51f | 1550 | |
fd2d3326 | 1551 | cookie = request_to_qc_t(data.hctx, rq); |
320ae51f JA |
1552 | |
1553 | if (unlikely(is_flush_fua)) { | |
1554 | blk_mq_bio_to_request(rq, bio); | |
bd166ef1 | 1555 | blk_mq_get_driver_tag(rq, NULL, true); |
320ae51f JA |
1556 | blk_insert_flush(rq); |
1557 | goto run_queue; | |
1558 | } | |
1559 | ||
1560 | /* | |
1561 | * A task plug currently exists. Since this is completely lockless, | |
1562 | * utilize that to temporarily store requests until the task is | |
1563 | * either done or scheduled away. | |
1564 | */ | |
e6c4438b JM |
1565 | plug = current->plug; |
1566 | if (plug) { | |
600271d9 SL |
1567 | struct request *last = NULL; |
1568 | ||
e6c4438b | 1569 | blk_mq_bio_to_request(rq, bio); |
0a6219a9 ML |
1570 | |
1571 | /* | |
1572 | * @request_count may become stale because of schedule | |
1573 | * out, so check the list again. | |
1574 | */ | |
1575 | if (list_empty(&plug->mq_list)) | |
1576 | request_count = 0; | |
676d0607 | 1577 | if (!request_count) |
e6c4438b | 1578 | trace_block_plug(q); |
600271d9 SL |
1579 | else |
1580 | last = list_entry_rq(plug->mq_list.prev); | |
b094f89c JA |
1581 | |
1582 | blk_mq_put_ctx(data.ctx); | |
1583 | ||
600271d9 SL |
1584 | if (request_count >= BLK_MAX_REQUEST_COUNT || (last && |
1585 | blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) { | |
e6c4438b JM |
1586 | blk_flush_plug_list(plug, false); |
1587 | trace_block_plug(q); | |
320ae51f | 1588 | } |
b094f89c | 1589 | |
e6c4438b | 1590 | list_add_tail(&rq->queuelist, &plug->mq_list); |
7b371636 | 1591 | return cookie; |
320ae51f JA |
1592 | } |
1593 | ||
bd166ef1 JA |
1594 | if (q->elevator) { |
1595 | blk_mq_put_ctx(data.ctx); | |
1596 | blk_mq_bio_to_request(rq, bio); | |
0abad774 JA |
1597 | blk_mq_sched_insert_request(rq, false, true, |
1598 | !is_sync || is_flush_fua); | |
bd166ef1 JA |
1599 | goto done; |
1600 | } | |
07068d5b JA |
1601 | if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) { |
1602 | /* | |
1603 | * For a SYNC request, send it to the hardware immediately. For | |
1604 | * an ASYNC request, just ensure that we run it later on. The | |
1605 | * latter allows for merging opportunities and more efficient | |
1606 | * dispatching. | |
1607 | */ | |
1608 | run_queue: | |
1609 | blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua); | |
320ae51f JA |
1610 | } |
1611 | ||
07068d5b | 1612 | blk_mq_put_ctx(data.ctx); |
bd166ef1 | 1613 | done: |
7b371636 | 1614 | return cookie; |
320ae51f JA |
1615 | } |
1616 | ||
cc71a6f4 JA |
1617 | void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags, |
1618 | unsigned int hctx_idx) | |
95363efd | 1619 | { |
e9b267d9 | 1620 | struct page *page; |
320ae51f | 1621 | |
24d2f903 | 1622 | if (tags->rqs && set->ops->exit_request) { |
e9b267d9 | 1623 | int i; |
320ae51f | 1624 | |
24d2f903 | 1625 | for (i = 0; i < tags->nr_tags; i++) { |
2af8cbe3 JA |
1626 | struct request *rq = tags->static_rqs[i]; |
1627 | ||
1628 | if (!rq) | |
e9b267d9 | 1629 | continue; |
2af8cbe3 | 1630 | set->ops->exit_request(set->driver_data, rq, |
24d2f903 | 1631 | hctx_idx, i); |
2af8cbe3 | 1632 | tags->static_rqs[i] = NULL; |
e9b267d9 | 1633 | } |
320ae51f | 1634 | } |
320ae51f | 1635 | |
24d2f903 CH |
1636 | while (!list_empty(&tags->page_list)) { |
1637 | page = list_first_entry(&tags->page_list, struct page, lru); | |
6753471c | 1638 | list_del_init(&page->lru); |
f75782e4 CM |
1639 | /* |
1640 | * Remove kmemleak object previously allocated in | |
1641 | * blk_mq_init_rq_map(). | |
1642 | */ | |
1643 | kmemleak_free(page_address(page)); | |
320ae51f JA |
1644 | __free_pages(page, page->private); |
1645 | } | |
cc71a6f4 | 1646 | } |
320ae51f | 1647 | |
cc71a6f4 JA |
1648 | void blk_mq_free_rq_map(struct blk_mq_tags *tags) |
1649 | { | |
24d2f903 | 1650 | kfree(tags->rqs); |
cc71a6f4 | 1651 | tags->rqs = NULL; |
2af8cbe3 JA |
1652 | kfree(tags->static_rqs); |
1653 | tags->static_rqs = NULL; | |
320ae51f | 1654 | |
24d2f903 | 1655 | blk_mq_free_tags(tags); |
320ae51f JA |
1656 | } |
1657 | ||
cc71a6f4 JA |
1658 | struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, |
1659 | unsigned int hctx_idx, | |
1660 | unsigned int nr_tags, | |
1661 | unsigned int reserved_tags) | |
320ae51f | 1662 | { |
24d2f903 | 1663 | struct blk_mq_tags *tags; |
320ae51f | 1664 | |
cc71a6f4 | 1665 | tags = blk_mq_init_tags(nr_tags, reserved_tags, |
24391c0d SL |
1666 | set->numa_node, |
1667 | BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags)); | |
24d2f903 CH |
1668 | if (!tags) |
1669 | return NULL; | |
320ae51f | 1670 | |
cc71a6f4 | 1671 | tags->rqs = kzalloc_node(nr_tags * sizeof(struct request *), |
36e1f3d1 | 1672 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, |
a5164405 | 1673 | set->numa_node); |
24d2f903 CH |
1674 | if (!tags->rqs) { |
1675 | blk_mq_free_tags(tags); | |
1676 | return NULL; | |
1677 | } | |
320ae51f | 1678 | |
2af8cbe3 JA |
1679 | tags->static_rqs = kzalloc_node(nr_tags * sizeof(struct request *), |
1680 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, | |
1681 | set->numa_node); | |
1682 | if (!tags->static_rqs) { | |
1683 | kfree(tags->rqs); | |
1684 | blk_mq_free_tags(tags); | |
1685 | return NULL; | |
1686 | } | |
1687 | ||
cc71a6f4 JA |
1688 | return tags; |
1689 | } | |
1690 | ||
1691 | static size_t order_to_size(unsigned int order) | |
1692 | { | |
1693 | return (size_t)PAGE_SIZE << order; | |
1694 | } | |
1695 | ||
1696 | int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags, | |
1697 | unsigned int hctx_idx, unsigned int depth) | |
1698 | { | |
1699 | unsigned int i, j, entries_per_page, max_order = 4; | |
1700 | size_t rq_size, left; | |
1701 | ||
1702 | INIT_LIST_HEAD(&tags->page_list); | |
1703 | ||
320ae51f JA |
1704 | /* |
1705 | * rq_size is the size of the request plus driver payload, rounded | |
1706 | * to the cacheline size | |
1707 | */ | |
24d2f903 | 1708 | rq_size = round_up(sizeof(struct request) + set->cmd_size, |
320ae51f | 1709 | cache_line_size()); |
cc71a6f4 | 1710 | left = rq_size * depth; |
320ae51f | 1711 | |
cc71a6f4 | 1712 | for (i = 0; i < depth; ) { |
320ae51f JA |
1713 | int this_order = max_order; |
1714 | struct page *page; | |
1715 | int to_do; | |
1716 | void *p; | |
1717 | ||
b3a834b1 | 1718 | while (this_order && left < order_to_size(this_order - 1)) |
320ae51f JA |
1719 | this_order--; |
1720 | ||
1721 | do { | |
a5164405 | 1722 | page = alloc_pages_node(set->numa_node, |
36e1f3d1 | 1723 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO, |
a5164405 | 1724 | this_order); |
320ae51f JA |
1725 | if (page) |
1726 | break; | |
1727 | if (!this_order--) | |
1728 | break; | |
1729 | if (order_to_size(this_order) < rq_size) | |
1730 | break; | |
1731 | } while (1); | |
1732 | ||
1733 | if (!page) | |
24d2f903 | 1734 | goto fail; |
320ae51f JA |
1735 | |
1736 | page->private = this_order; | |
24d2f903 | 1737 | list_add_tail(&page->lru, &tags->page_list); |
320ae51f JA |
1738 | |
1739 | p = page_address(page); | |
f75782e4 CM |
1740 | /* |
1741 | * Allow kmemleak to scan these pages as they contain pointers | |
1742 | * to additional allocations like via ops->init_request(). | |
1743 | */ | |
36e1f3d1 | 1744 | kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO); |
320ae51f | 1745 | entries_per_page = order_to_size(this_order) / rq_size; |
cc71a6f4 | 1746 | to_do = min(entries_per_page, depth - i); |
320ae51f JA |
1747 | left -= to_do * rq_size; |
1748 | for (j = 0; j < to_do; j++) { | |
2af8cbe3 JA |
1749 | struct request *rq = p; |
1750 | ||
1751 | tags->static_rqs[i] = rq; | |
24d2f903 CH |
1752 | if (set->ops->init_request) { |
1753 | if (set->ops->init_request(set->driver_data, | |
2af8cbe3 | 1754 | rq, hctx_idx, i, |
a5164405 | 1755 | set->numa_node)) { |
2af8cbe3 | 1756 | tags->static_rqs[i] = NULL; |
24d2f903 | 1757 | goto fail; |
a5164405 | 1758 | } |
e9b267d9 CH |
1759 | } |
1760 | ||
320ae51f JA |
1761 | p += rq_size; |
1762 | i++; | |
1763 | } | |
1764 | } | |
cc71a6f4 | 1765 | return 0; |
320ae51f | 1766 | |
24d2f903 | 1767 | fail: |
cc71a6f4 JA |
1768 | blk_mq_free_rqs(set, tags, hctx_idx); |
1769 | return -ENOMEM; | |
320ae51f JA |
1770 | } |
1771 | ||
e57690fe JA |
1772 | /* |
1773 | * 'cpu' is going away. splice any existing rq_list entries from this | |
1774 | * software queue to the hw queue dispatch list, and ensure that it | |
1775 | * gets run. | |
1776 | */ | |
9467f859 | 1777 | static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node) |
484b4061 | 1778 | { |
9467f859 | 1779 | struct blk_mq_hw_ctx *hctx; |
484b4061 JA |
1780 | struct blk_mq_ctx *ctx; |
1781 | LIST_HEAD(tmp); | |
1782 | ||
9467f859 | 1783 | hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead); |
e57690fe | 1784 | ctx = __blk_mq_get_ctx(hctx->queue, cpu); |
484b4061 JA |
1785 | |
1786 | spin_lock(&ctx->lock); | |
1787 | if (!list_empty(&ctx->rq_list)) { | |
1788 | list_splice_init(&ctx->rq_list, &tmp); | |
1789 | blk_mq_hctx_clear_pending(hctx, ctx); | |
1790 | } | |
1791 | spin_unlock(&ctx->lock); | |
1792 | ||
1793 | if (list_empty(&tmp)) | |
9467f859 | 1794 | return 0; |
484b4061 | 1795 | |
e57690fe JA |
1796 | spin_lock(&hctx->lock); |
1797 | list_splice_tail_init(&tmp, &hctx->dispatch); | |
1798 | spin_unlock(&hctx->lock); | |
484b4061 JA |
1799 | |
1800 | blk_mq_run_hw_queue(hctx, true); | |
9467f859 | 1801 | return 0; |
484b4061 JA |
1802 | } |
1803 | ||
9467f859 | 1804 | static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx) |
484b4061 | 1805 | { |
9467f859 TG |
1806 | cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD, |
1807 | &hctx->cpuhp_dead); | |
484b4061 JA |
1808 | } |
1809 | ||
c3b4afca | 1810 | /* hctx->ctxs will be freed in queue's release handler */ |
08e98fc6 ML |
1811 | static void blk_mq_exit_hctx(struct request_queue *q, |
1812 | struct blk_mq_tag_set *set, | |
1813 | struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx) | |
1814 | { | |
f70ced09 ML |
1815 | unsigned flush_start_tag = set->queue_depth; |
1816 | ||
08e98fc6 ML |
1817 | blk_mq_tag_idle(hctx); |
1818 | ||
f70ced09 ML |
1819 | if (set->ops->exit_request) |
1820 | set->ops->exit_request(set->driver_data, | |
1821 | hctx->fq->flush_rq, hctx_idx, | |
1822 | flush_start_tag + hctx_idx); | |
1823 | ||
08e98fc6 ML |
1824 | if (set->ops->exit_hctx) |
1825 | set->ops->exit_hctx(hctx, hctx_idx); | |
1826 | ||
6a83e74d BVA |
1827 | if (hctx->flags & BLK_MQ_F_BLOCKING) |
1828 | cleanup_srcu_struct(&hctx->queue_rq_srcu); | |
1829 | ||
9467f859 | 1830 | blk_mq_remove_cpuhp(hctx); |
f70ced09 | 1831 | blk_free_flush_queue(hctx->fq); |
88459642 | 1832 | sbitmap_free(&hctx->ctx_map); |
08e98fc6 ML |
1833 | } |
1834 | ||
624dbe47 ML |
1835 | static void blk_mq_exit_hw_queues(struct request_queue *q, |
1836 | struct blk_mq_tag_set *set, int nr_queue) | |
1837 | { | |
1838 | struct blk_mq_hw_ctx *hctx; | |
1839 | unsigned int i; | |
1840 | ||
1841 | queue_for_each_hw_ctx(q, hctx, i) { | |
1842 | if (i == nr_queue) | |
1843 | break; | |
08e98fc6 | 1844 | blk_mq_exit_hctx(q, set, hctx, i); |
624dbe47 | 1845 | } |
624dbe47 ML |
1846 | } |
1847 | ||
1848 | static void blk_mq_free_hw_queues(struct request_queue *q, | |
1849 | struct blk_mq_tag_set *set) | |
1850 | { | |
1851 | struct blk_mq_hw_ctx *hctx; | |
1852 | unsigned int i; | |
1853 | ||
e09aae7e | 1854 | queue_for_each_hw_ctx(q, hctx, i) |
624dbe47 | 1855 | free_cpumask_var(hctx->cpumask); |
624dbe47 ML |
1856 | } |
1857 | ||
08e98fc6 ML |
1858 | static int blk_mq_init_hctx(struct request_queue *q, |
1859 | struct blk_mq_tag_set *set, | |
1860 | struct blk_mq_hw_ctx *hctx, unsigned hctx_idx) | |
320ae51f | 1861 | { |
08e98fc6 | 1862 | int node; |
f70ced09 | 1863 | unsigned flush_start_tag = set->queue_depth; |
08e98fc6 ML |
1864 | |
1865 | node = hctx->numa_node; | |
1866 | if (node == NUMA_NO_NODE) | |
1867 | node = hctx->numa_node = set->numa_node; | |
1868 | ||
27489a3c | 1869 | INIT_WORK(&hctx->run_work, blk_mq_run_work_fn); |
08e98fc6 ML |
1870 | INIT_DELAYED_WORK(&hctx->delay_work, blk_mq_delay_work_fn); |
1871 | spin_lock_init(&hctx->lock); | |
1872 | INIT_LIST_HEAD(&hctx->dispatch); | |
1873 | hctx->queue = q; | |
1874 | hctx->queue_num = hctx_idx; | |
2404e607 | 1875 | hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED; |
08e98fc6 | 1876 | |
9467f859 | 1877 | cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead); |
08e98fc6 ML |
1878 | |
1879 | hctx->tags = set->tags[hctx_idx]; | |
320ae51f JA |
1880 | |
1881 | /* | |
08e98fc6 ML |
1882 | * Allocate space for all possible cpus to avoid allocation at |
1883 | * runtime | |
320ae51f | 1884 | */ |
08e98fc6 ML |
1885 | hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *), |
1886 | GFP_KERNEL, node); | |
1887 | if (!hctx->ctxs) | |
1888 | goto unregister_cpu_notifier; | |
320ae51f | 1889 | |
88459642 OS |
1890 | if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8), GFP_KERNEL, |
1891 | node)) | |
08e98fc6 | 1892 | goto free_ctxs; |
320ae51f | 1893 | |
08e98fc6 | 1894 | hctx->nr_ctx = 0; |
320ae51f | 1895 | |
08e98fc6 ML |
1896 | if (set->ops->init_hctx && |
1897 | set->ops->init_hctx(hctx, set->driver_data, hctx_idx)) | |
1898 | goto free_bitmap; | |
320ae51f | 1899 | |
f70ced09 ML |
1900 | hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size); |
1901 | if (!hctx->fq) | |
1902 | goto exit_hctx; | |
320ae51f | 1903 | |
f70ced09 ML |
1904 | if (set->ops->init_request && |
1905 | set->ops->init_request(set->driver_data, | |
1906 | hctx->fq->flush_rq, hctx_idx, | |
1907 | flush_start_tag + hctx_idx, node)) | |
1908 | goto free_fq; | |
320ae51f | 1909 | |
6a83e74d BVA |
1910 | if (hctx->flags & BLK_MQ_F_BLOCKING) |
1911 | init_srcu_struct(&hctx->queue_rq_srcu); | |
1912 | ||
08e98fc6 | 1913 | return 0; |
320ae51f | 1914 | |
f70ced09 ML |
1915 | free_fq: |
1916 | kfree(hctx->fq); | |
1917 | exit_hctx: | |
1918 | if (set->ops->exit_hctx) | |
1919 | set->ops->exit_hctx(hctx, hctx_idx); | |
08e98fc6 | 1920 | free_bitmap: |
88459642 | 1921 | sbitmap_free(&hctx->ctx_map); |
08e98fc6 ML |
1922 | free_ctxs: |
1923 | kfree(hctx->ctxs); | |
1924 | unregister_cpu_notifier: | |
9467f859 | 1925 | blk_mq_remove_cpuhp(hctx); |
08e98fc6 ML |
1926 | return -1; |
1927 | } | |
320ae51f | 1928 | |
320ae51f JA |
1929 | static void blk_mq_init_cpu_queues(struct request_queue *q, |
1930 | unsigned int nr_hw_queues) | |
1931 | { | |
1932 | unsigned int i; | |
1933 | ||
1934 | for_each_possible_cpu(i) { | |
1935 | struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i); | |
1936 | struct blk_mq_hw_ctx *hctx; | |
1937 | ||
1938 | memset(__ctx, 0, sizeof(*__ctx)); | |
1939 | __ctx->cpu = i; | |
1940 | spin_lock_init(&__ctx->lock); | |
1941 | INIT_LIST_HEAD(&__ctx->rq_list); | |
1942 | __ctx->queue = q; | |
cf43e6be JA |
1943 | blk_stat_init(&__ctx->stat[BLK_STAT_READ]); |
1944 | blk_stat_init(&__ctx->stat[BLK_STAT_WRITE]); | |
320ae51f JA |
1945 | |
1946 | /* If the cpu isn't online, the cpu is mapped to first hctx */ | |
320ae51f JA |
1947 | if (!cpu_online(i)) |
1948 | continue; | |
1949 | ||
7d7e0f90 | 1950 | hctx = blk_mq_map_queue(q, i); |
e4043dcf | 1951 | |
320ae51f JA |
1952 | /* |
1953 | * Set local node, IFF we have more than one hw queue. If | |
1954 | * not, we remain on the home node of the device | |
1955 | */ | |
1956 | if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE) | |
bffed457 | 1957 | hctx->numa_node = local_memory_node(cpu_to_node(i)); |
320ae51f JA |
1958 | } |
1959 | } | |
1960 | ||
cc71a6f4 JA |
1961 | static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, int hctx_idx) |
1962 | { | |
1963 | int ret = 0; | |
1964 | ||
1965 | set->tags[hctx_idx] = blk_mq_alloc_rq_map(set, hctx_idx, | |
1966 | set->queue_depth, set->reserved_tags); | |
1967 | if (!set->tags[hctx_idx]) | |
1968 | return false; | |
1969 | ||
1970 | ret = blk_mq_alloc_rqs(set, set->tags[hctx_idx], hctx_idx, | |
1971 | set->queue_depth); | |
1972 | if (!ret) | |
1973 | return true; | |
1974 | ||
1975 | blk_mq_free_rq_map(set->tags[hctx_idx]); | |
1976 | set->tags[hctx_idx] = NULL; | |
1977 | return false; | |
1978 | } | |
1979 | ||
1980 | static void blk_mq_free_map_and_requests(struct blk_mq_tag_set *set, | |
1981 | unsigned int hctx_idx) | |
1982 | { | |
bd166ef1 JA |
1983 | if (set->tags[hctx_idx]) { |
1984 | blk_mq_free_rqs(set, set->tags[hctx_idx], hctx_idx); | |
1985 | blk_mq_free_rq_map(set->tags[hctx_idx]); | |
1986 | set->tags[hctx_idx] = NULL; | |
1987 | } | |
cc71a6f4 JA |
1988 | } |
1989 | ||
5778322e AM |
1990 | static void blk_mq_map_swqueue(struct request_queue *q, |
1991 | const struct cpumask *online_mask) | |
320ae51f | 1992 | { |
d1b1cea1 | 1993 | unsigned int i, hctx_idx; |
320ae51f JA |
1994 | struct blk_mq_hw_ctx *hctx; |
1995 | struct blk_mq_ctx *ctx; | |
2a34c087 | 1996 | struct blk_mq_tag_set *set = q->tag_set; |
320ae51f | 1997 | |
60de074b AM |
1998 | /* |
1999 | * Avoid others reading imcomplete hctx->cpumask through sysfs | |
2000 | */ | |
2001 | mutex_lock(&q->sysfs_lock); | |
2002 | ||
320ae51f | 2003 | queue_for_each_hw_ctx(q, hctx, i) { |
e4043dcf | 2004 | cpumask_clear(hctx->cpumask); |
320ae51f JA |
2005 | hctx->nr_ctx = 0; |
2006 | } | |
2007 | ||
2008 | /* | |
2009 | * Map software to hardware queues | |
2010 | */ | |
897bb0c7 | 2011 | for_each_possible_cpu(i) { |
320ae51f | 2012 | /* If the cpu isn't online, the cpu is mapped to first hctx */ |
5778322e | 2013 | if (!cpumask_test_cpu(i, online_mask)) |
e4043dcf JA |
2014 | continue; |
2015 | ||
d1b1cea1 GKB |
2016 | hctx_idx = q->mq_map[i]; |
2017 | /* unmapped hw queue can be remapped after CPU topo changed */ | |
cc71a6f4 JA |
2018 | if (!set->tags[hctx_idx] && |
2019 | !__blk_mq_alloc_rq_map(set, hctx_idx)) { | |
d1b1cea1 GKB |
2020 | /* |
2021 | * If tags initialization fail for some hctx, | |
2022 | * that hctx won't be brought online. In this | |
2023 | * case, remap the current ctx to hctx[0] which | |
2024 | * is guaranteed to always have tags allocated | |
2025 | */ | |
cc71a6f4 | 2026 | q->mq_map[i] = 0; |
d1b1cea1 GKB |
2027 | } |
2028 | ||
897bb0c7 | 2029 | ctx = per_cpu_ptr(q->queue_ctx, i); |
7d7e0f90 | 2030 | hctx = blk_mq_map_queue(q, i); |
868f2f0b | 2031 | |
e4043dcf | 2032 | cpumask_set_cpu(i, hctx->cpumask); |
320ae51f JA |
2033 | ctx->index_hw = hctx->nr_ctx; |
2034 | hctx->ctxs[hctx->nr_ctx++] = ctx; | |
2035 | } | |
506e931f | 2036 | |
60de074b AM |
2037 | mutex_unlock(&q->sysfs_lock); |
2038 | ||
506e931f | 2039 | queue_for_each_hw_ctx(q, hctx, i) { |
484b4061 | 2040 | /* |
a68aafa5 JA |
2041 | * If no software queues are mapped to this hardware queue, |
2042 | * disable it and free the request entries. | |
484b4061 JA |
2043 | */ |
2044 | if (!hctx->nr_ctx) { | |
d1b1cea1 GKB |
2045 | /* Never unmap queue 0. We need it as a |
2046 | * fallback in case of a new remap fails | |
2047 | * allocation | |
2048 | */ | |
cc71a6f4 JA |
2049 | if (i && set->tags[i]) |
2050 | blk_mq_free_map_and_requests(set, i); | |
2051 | ||
2a34c087 | 2052 | hctx->tags = NULL; |
484b4061 JA |
2053 | continue; |
2054 | } | |
2055 | ||
2a34c087 ML |
2056 | hctx->tags = set->tags[i]; |
2057 | WARN_ON(!hctx->tags); | |
2058 | ||
889fa31f CY |
2059 | /* |
2060 | * Set the map size to the number of mapped software queues. | |
2061 | * This is more accurate and more efficient than looping | |
2062 | * over all possibly mapped software queues. | |
2063 | */ | |
88459642 | 2064 | sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx); |
889fa31f | 2065 | |
484b4061 JA |
2066 | /* |
2067 | * Initialize batch roundrobin counts | |
2068 | */ | |
506e931f JA |
2069 | hctx->next_cpu = cpumask_first(hctx->cpumask); |
2070 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; | |
2071 | } | |
320ae51f JA |
2072 | } |
2073 | ||
2404e607 | 2074 | static void queue_set_hctx_shared(struct request_queue *q, bool shared) |
0d2602ca JA |
2075 | { |
2076 | struct blk_mq_hw_ctx *hctx; | |
0d2602ca JA |
2077 | int i; |
2078 | ||
2404e607 JM |
2079 | queue_for_each_hw_ctx(q, hctx, i) { |
2080 | if (shared) | |
2081 | hctx->flags |= BLK_MQ_F_TAG_SHARED; | |
2082 | else | |
2083 | hctx->flags &= ~BLK_MQ_F_TAG_SHARED; | |
2084 | } | |
2085 | } | |
2086 | ||
2087 | static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set, bool shared) | |
2088 | { | |
2089 | struct request_queue *q; | |
0d2602ca JA |
2090 | |
2091 | list_for_each_entry(q, &set->tag_list, tag_set_list) { | |
2092 | blk_mq_freeze_queue(q); | |
2404e607 | 2093 | queue_set_hctx_shared(q, shared); |
0d2602ca JA |
2094 | blk_mq_unfreeze_queue(q); |
2095 | } | |
2096 | } | |
2097 | ||
2098 | static void blk_mq_del_queue_tag_set(struct request_queue *q) | |
2099 | { | |
2100 | struct blk_mq_tag_set *set = q->tag_set; | |
2101 | ||
0d2602ca JA |
2102 | mutex_lock(&set->tag_list_lock); |
2103 | list_del_init(&q->tag_set_list); | |
2404e607 JM |
2104 | if (list_is_singular(&set->tag_list)) { |
2105 | /* just transitioned to unshared */ | |
2106 | set->flags &= ~BLK_MQ_F_TAG_SHARED; | |
2107 | /* update existing queue */ | |
2108 | blk_mq_update_tag_set_depth(set, false); | |
2109 | } | |
0d2602ca | 2110 | mutex_unlock(&set->tag_list_lock); |
0d2602ca JA |
2111 | } |
2112 | ||
2113 | static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set, | |
2114 | struct request_queue *q) | |
2115 | { | |
2116 | q->tag_set = set; | |
2117 | ||
2118 | mutex_lock(&set->tag_list_lock); | |
2404e607 JM |
2119 | |
2120 | /* Check to see if we're transitioning to shared (from 1 to 2 queues). */ | |
2121 | if (!list_empty(&set->tag_list) && !(set->flags & BLK_MQ_F_TAG_SHARED)) { | |
2122 | set->flags |= BLK_MQ_F_TAG_SHARED; | |
2123 | /* update existing queue */ | |
2124 | blk_mq_update_tag_set_depth(set, true); | |
2125 | } | |
2126 | if (set->flags & BLK_MQ_F_TAG_SHARED) | |
2127 | queue_set_hctx_shared(q, true); | |
0d2602ca | 2128 | list_add_tail(&q->tag_set_list, &set->tag_list); |
2404e607 | 2129 | |
0d2602ca JA |
2130 | mutex_unlock(&set->tag_list_lock); |
2131 | } | |
2132 | ||
e09aae7e ML |
2133 | /* |
2134 | * It is the actual release handler for mq, but we do it from | |
2135 | * request queue's release handler for avoiding use-after-free | |
2136 | * and headache because q->mq_kobj shouldn't have been introduced, | |
2137 | * but we can't group ctx/kctx kobj without it. | |
2138 | */ | |
2139 | void blk_mq_release(struct request_queue *q) | |
2140 | { | |
2141 | struct blk_mq_hw_ctx *hctx; | |
2142 | unsigned int i; | |
2143 | ||
bd166ef1 JA |
2144 | blk_mq_sched_teardown(q); |
2145 | ||
e09aae7e | 2146 | /* hctx kobj stays in hctx */ |
c3b4afca ML |
2147 | queue_for_each_hw_ctx(q, hctx, i) { |
2148 | if (!hctx) | |
2149 | continue; | |
2150 | kfree(hctx->ctxs); | |
e09aae7e | 2151 | kfree(hctx); |
c3b4afca | 2152 | } |
e09aae7e | 2153 | |
a723bab3 AM |
2154 | q->mq_map = NULL; |
2155 | ||
e09aae7e ML |
2156 | kfree(q->queue_hw_ctx); |
2157 | ||
2158 | /* ctx kobj stays in queue_ctx */ | |
2159 | free_percpu(q->queue_ctx); | |
2160 | } | |
2161 | ||
24d2f903 | 2162 | struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set) |
b62c21b7 MS |
2163 | { |
2164 | struct request_queue *uninit_q, *q; | |
2165 | ||
2166 | uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node); | |
2167 | if (!uninit_q) | |
2168 | return ERR_PTR(-ENOMEM); | |
2169 | ||
2170 | q = blk_mq_init_allocated_queue(set, uninit_q); | |
2171 | if (IS_ERR(q)) | |
2172 | blk_cleanup_queue(uninit_q); | |
2173 | ||
2174 | return q; | |
2175 | } | |
2176 | EXPORT_SYMBOL(blk_mq_init_queue); | |
2177 | ||
868f2f0b KB |
2178 | static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set, |
2179 | struct request_queue *q) | |
320ae51f | 2180 | { |
868f2f0b KB |
2181 | int i, j; |
2182 | struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx; | |
f14bbe77 | 2183 | |
868f2f0b | 2184 | blk_mq_sysfs_unregister(q); |
24d2f903 | 2185 | for (i = 0; i < set->nr_hw_queues; i++) { |
868f2f0b | 2186 | int node; |
f14bbe77 | 2187 | |
868f2f0b KB |
2188 | if (hctxs[i]) |
2189 | continue; | |
2190 | ||
2191 | node = blk_mq_hw_queue_to_node(q->mq_map, i); | |
cdef54dd CH |
2192 | hctxs[i] = kzalloc_node(sizeof(struct blk_mq_hw_ctx), |
2193 | GFP_KERNEL, node); | |
320ae51f | 2194 | if (!hctxs[i]) |
868f2f0b | 2195 | break; |
320ae51f | 2196 | |
a86073e4 | 2197 | if (!zalloc_cpumask_var_node(&hctxs[i]->cpumask, GFP_KERNEL, |
868f2f0b KB |
2198 | node)) { |
2199 | kfree(hctxs[i]); | |
2200 | hctxs[i] = NULL; | |
2201 | break; | |
2202 | } | |
e4043dcf | 2203 | |
0d2602ca | 2204 | atomic_set(&hctxs[i]->nr_active, 0); |
f14bbe77 | 2205 | hctxs[i]->numa_node = node; |
320ae51f | 2206 | hctxs[i]->queue_num = i; |
868f2f0b KB |
2207 | |
2208 | if (blk_mq_init_hctx(q, set, hctxs[i], i)) { | |
2209 | free_cpumask_var(hctxs[i]->cpumask); | |
2210 | kfree(hctxs[i]); | |
2211 | hctxs[i] = NULL; | |
2212 | break; | |
2213 | } | |
2214 | blk_mq_hctx_kobj_init(hctxs[i]); | |
320ae51f | 2215 | } |
868f2f0b KB |
2216 | for (j = i; j < q->nr_hw_queues; j++) { |
2217 | struct blk_mq_hw_ctx *hctx = hctxs[j]; | |
2218 | ||
2219 | if (hctx) { | |
cc71a6f4 JA |
2220 | if (hctx->tags) |
2221 | blk_mq_free_map_and_requests(set, j); | |
868f2f0b KB |
2222 | blk_mq_exit_hctx(q, set, hctx, j); |
2223 | free_cpumask_var(hctx->cpumask); | |
2224 | kobject_put(&hctx->kobj); | |
2225 | kfree(hctx->ctxs); | |
2226 | kfree(hctx); | |
2227 | hctxs[j] = NULL; | |
2228 | ||
2229 | } | |
2230 | } | |
2231 | q->nr_hw_queues = i; | |
2232 | blk_mq_sysfs_register(q); | |
2233 | } | |
2234 | ||
2235 | struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set, | |
2236 | struct request_queue *q) | |
2237 | { | |
66841672 ML |
2238 | /* mark the queue as mq asap */ |
2239 | q->mq_ops = set->ops; | |
2240 | ||
868f2f0b KB |
2241 | q->queue_ctx = alloc_percpu(struct blk_mq_ctx); |
2242 | if (!q->queue_ctx) | |
c7de5726 | 2243 | goto err_exit; |
868f2f0b KB |
2244 | |
2245 | q->queue_hw_ctx = kzalloc_node(nr_cpu_ids * sizeof(*(q->queue_hw_ctx)), | |
2246 | GFP_KERNEL, set->numa_node); | |
2247 | if (!q->queue_hw_ctx) | |
2248 | goto err_percpu; | |
2249 | ||
bdd17e75 | 2250 | q->mq_map = set->mq_map; |
868f2f0b KB |
2251 | |
2252 | blk_mq_realloc_hw_ctxs(set, q); | |
2253 | if (!q->nr_hw_queues) | |
2254 | goto err_hctxs; | |
320ae51f | 2255 | |
287922eb | 2256 | INIT_WORK(&q->timeout_work, blk_mq_timeout_work); |
e56f698b | 2257 | blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ); |
320ae51f JA |
2258 | |
2259 | q->nr_queues = nr_cpu_ids; | |
320ae51f | 2260 | |
94eddfbe | 2261 | q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT; |
320ae51f | 2262 | |
05f1dd53 JA |
2263 | if (!(set->flags & BLK_MQ_F_SG_MERGE)) |
2264 | q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE; | |
2265 | ||
1be036e9 CH |
2266 | q->sg_reserved_size = INT_MAX; |
2267 | ||
2849450a | 2268 | INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work); |
6fca6a61 CH |
2269 | INIT_LIST_HEAD(&q->requeue_list); |
2270 | spin_lock_init(&q->requeue_lock); | |
2271 | ||
07068d5b JA |
2272 | if (q->nr_hw_queues > 1) |
2273 | blk_queue_make_request(q, blk_mq_make_request); | |
2274 | else | |
2275 | blk_queue_make_request(q, blk_sq_make_request); | |
2276 | ||
eba71768 JA |
2277 | /* |
2278 | * Do this after blk_queue_make_request() overrides it... | |
2279 | */ | |
2280 | q->nr_requests = set->queue_depth; | |
2281 | ||
64f1c21e JA |
2282 | /* |
2283 | * Default to classic polling | |
2284 | */ | |
2285 | q->poll_nsec = -1; | |
2286 | ||
24d2f903 CH |
2287 | if (set->ops->complete) |
2288 | blk_queue_softirq_done(q, set->ops->complete); | |
30a91cb4 | 2289 | |
24d2f903 | 2290 | blk_mq_init_cpu_queues(q, set->nr_hw_queues); |
320ae51f | 2291 | |
5778322e | 2292 | get_online_cpus(); |
320ae51f | 2293 | mutex_lock(&all_q_mutex); |
320ae51f | 2294 | |
4593fdbe | 2295 | list_add_tail(&q->all_q_node, &all_q_list); |
0d2602ca | 2296 | blk_mq_add_queue_tag_set(set, q); |
5778322e | 2297 | blk_mq_map_swqueue(q, cpu_online_mask); |
484b4061 | 2298 | |
4593fdbe | 2299 | mutex_unlock(&all_q_mutex); |
5778322e | 2300 | put_online_cpus(); |
4593fdbe | 2301 | |
d3484991 JA |
2302 | if (!(set->flags & BLK_MQ_F_NO_SCHED)) { |
2303 | int ret; | |
2304 | ||
2305 | ret = blk_mq_sched_init(q); | |
2306 | if (ret) | |
2307 | return ERR_PTR(ret); | |
2308 | } | |
2309 | ||
320ae51f | 2310 | return q; |
18741986 | 2311 | |
320ae51f | 2312 | err_hctxs: |
868f2f0b | 2313 | kfree(q->queue_hw_ctx); |
320ae51f | 2314 | err_percpu: |
868f2f0b | 2315 | free_percpu(q->queue_ctx); |
c7de5726 ML |
2316 | err_exit: |
2317 | q->mq_ops = NULL; | |
320ae51f JA |
2318 | return ERR_PTR(-ENOMEM); |
2319 | } | |
b62c21b7 | 2320 | EXPORT_SYMBOL(blk_mq_init_allocated_queue); |
320ae51f JA |
2321 | |
2322 | void blk_mq_free_queue(struct request_queue *q) | |
2323 | { | |
624dbe47 | 2324 | struct blk_mq_tag_set *set = q->tag_set; |
320ae51f | 2325 | |
0e626368 AM |
2326 | mutex_lock(&all_q_mutex); |
2327 | list_del_init(&q->all_q_node); | |
2328 | mutex_unlock(&all_q_mutex); | |
2329 | ||
87760e5e JA |
2330 | wbt_exit(q); |
2331 | ||
0d2602ca JA |
2332 | blk_mq_del_queue_tag_set(q); |
2333 | ||
624dbe47 ML |
2334 | blk_mq_exit_hw_queues(q, set, set->nr_hw_queues); |
2335 | blk_mq_free_hw_queues(q, set); | |
320ae51f | 2336 | } |
320ae51f JA |
2337 | |
2338 | /* Basically redo blk_mq_init_queue with queue frozen */ | |
5778322e AM |
2339 | static void blk_mq_queue_reinit(struct request_queue *q, |
2340 | const struct cpumask *online_mask) | |
320ae51f | 2341 | { |
4ecd4fef | 2342 | WARN_ON_ONCE(!atomic_read(&q->mq_freeze_depth)); |
320ae51f | 2343 | |
67aec14c JA |
2344 | blk_mq_sysfs_unregister(q); |
2345 | ||
320ae51f JA |
2346 | /* |
2347 | * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe | |
2348 | * we should change hctx numa_node according to new topology (this | |
2349 | * involves free and re-allocate memory, worthy doing?) | |
2350 | */ | |
2351 | ||
5778322e | 2352 | blk_mq_map_swqueue(q, online_mask); |
320ae51f | 2353 | |
67aec14c | 2354 | blk_mq_sysfs_register(q); |
320ae51f JA |
2355 | } |
2356 | ||
65d5291e SAS |
2357 | /* |
2358 | * New online cpumask which is going to be set in this hotplug event. | |
2359 | * Declare this cpumasks as global as cpu-hotplug operation is invoked | |
2360 | * one-by-one and dynamically allocating this could result in a failure. | |
2361 | */ | |
2362 | static struct cpumask cpuhp_online_new; | |
2363 | ||
2364 | static void blk_mq_queue_reinit_work(void) | |
320ae51f JA |
2365 | { |
2366 | struct request_queue *q; | |
320ae51f JA |
2367 | |
2368 | mutex_lock(&all_q_mutex); | |
f3af020b TH |
2369 | /* |
2370 | * We need to freeze and reinit all existing queues. Freezing | |
2371 | * involves synchronous wait for an RCU grace period and doing it | |
2372 | * one by one may take a long time. Start freezing all queues in | |
2373 | * one swoop and then wait for the completions so that freezing can | |
2374 | * take place in parallel. | |
2375 | */ | |
2376 | list_for_each_entry(q, &all_q_list, all_q_node) | |
2377 | blk_mq_freeze_queue_start(q); | |
415d3dab | 2378 | list_for_each_entry(q, &all_q_list, all_q_node) |
f3af020b TH |
2379 | blk_mq_freeze_queue_wait(q); |
2380 | ||
320ae51f | 2381 | list_for_each_entry(q, &all_q_list, all_q_node) |
65d5291e | 2382 | blk_mq_queue_reinit(q, &cpuhp_online_new); |
f3af020b TH |
2383 | |
2384 | list_for_each_entry(q, &all_q_list, all_q_node) | |
2385 | blk_mq_unfreeze_queue(q); | |
2386 | ||
320ae51f | 2387 | mutex_unlock(&all_q_mutex); |
65d5291e SAS |
2388 | } |
2389 | ||
2390 | static int blk_mq_queue_reinit_dead(unsigned int cpu) | |
2391 | { | |
97a32864 | 2392 | cpumask_copy(&cpuhp_online_new, cpu_online_mask); |
65d5291e SAS |
2393 | blk_mq_queue_reinit_work(); |
2394 | return 0; | |
2395 | } | |
2396 | ||
2397 | /* | |
2398 | * Before hotadded cpu starts handling requests, new mappings must be | |
2399 | * established. Otherwise, these requests in hw queue might never be | |
2400 | * dispatched. | |
2401 | * | |
2402 | * For example, there is a single hw queue (hctx) and two CPU queues (ctx0 | |
2403 | * for CPU0, and ctx1 for CPU1). | |
2404 | * | |
2405 | * Now CPU1 is just onlined and a request is inserted into ctx1->rq_list | |
2406 | * and set bit0 in pending bitmap as ctx1->index_hw is still zero. | |
2407 | * | |
2c3ad667 JA |
2408 | * And then while running hw queue, blk_mq_flush_busy_ctxs() finds bit0 is set |
2409 | * in pending bitmap and tries to retrieve requests in hctx->ctxs[0]->rq_list. | |
2410 | * But htx->ctxs[0] is a pointer to ctx0, so the request in ctx1->rq_list is | |
2411 | * ignored. | |
65d5291e SAS |
2412 | */ |
2413 | static int blk_mq_queue_reinit_prepare(unsigned int cpu) | |
2414 | { | |
2415 | cpumask_copy(&cpuhp_online_new, cpu_online_mask); | |
2416 | cpumask_set_cpu(cpu, &cpuhp_online_new); | |
2417 | blk_mq_queue_reinit_work(); | |
2418 | return 0; | |
320ae51f JA |
2419 | } |
2420 | ||
a5164405 JA |
2421 | static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set) |
2422 | { | |
2423 | int i; | |
2424 | ||
cc71a6f4 JA |
2425 | for (i = 0; i < set->nr_hw_queues; i++) |
2426 | if (!__blk_mq_alloc_rq_map(set, i)) | |
a5164405 | 2427 | goto out_unwind; |
a5164405 JA |
2428 | |
2429 | return 0; | |
2430 | ||
2431 | out_unwind: | |
2432 | while (--i >= 0) | |
cc71a6f4 | 2433 | blk_mq_free_rq_map(set->tags[i]); |
a5164405 | 2434 | |
a5164405 JA |
2435 | return -ENOMEM; |
2436 | } | |
2437 | ||
2438 | /* | |
2439 | * Allocate the request maps associated with this tag_set. Note that this | |
2440 | * may reduce the depth asked for, if memory is tight. set->queue_depth | |
2441 | * will be updated to reflect the allocated depth. | |
2442 | */ | |
2443 | static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set) | |
2444 | { | |
2445 | unsigned int depth; | |
2446 | int err; | |
2447 | ||
2448 | depth = set->queue_depth; | |
2449 | do { | |
2450 | err = __blk_mq_alloc_rq_maps(set); | |
2451 | if (!err) | |
2452 | break; | |
2453 | ||
2454 | set->queue_depth >>= 1; | |
2455 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) { | |
2456 | err = -ENOMEM; | |
2457 | break; | |
2458 | } | |
2459 | } while (set->queue_depth); | |
2460 | ||
2461 | if (!set->queue_depth || err) { | |
2462 | pr_err("blk-mq: failed to allocate request map\n"); | |
2463 | return -ENOMEM; | |
2464 | } | |
2465 | ||
2466 | if (depth != set->queue_depth) | |
2467 | pr_info("blk-mq: reduced tag depth (%u -> %u)\n", | |
2468 | depth, set->queue_depth); | |
2469 | ||
2470 | return 0; | |
2471 | } | |
2472 | ||
a4391c64 JA |
2473 | /* |
2474 | * Alloc a tag set to be associated with one or more request queues. | |
2475 | * May fail with EINVAL for various error conditions. May adjust the | |
2476 | * requested depth down, if if it too large. In that case, the set | |
2477 | * value will be stored in set->queue_depth. | |
2478 | */ | |
24d2f903 CH |
2479 | int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set) |
2480 | { | |
da695ba2 CH |
2481 | int ret; |
2482 | ||
205fb5f5 BVA |
2483 | BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS); |
2484 | ||
24d2f903 CH |
2485 | if (!set->nr_hw_queues) |
2486 | return -EINVAL; | |
a4391c64 | 2487 | if (!set->queue_depth) |
24d2f903 CH |
2488 | return -EINVAL; |
2489 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) | |
2490 | return -EINVAL; | |
2491 | ||
7d7e0f90 | 2492 | if (!set->ops->queue_rq) |
24d2f903 CH |
2493 | return -EINVAL; |
2494 | ||
a4391c64 JA |
2495 | if (set->queue_depth > BLK_MQ_MAX_DEPTH) { |
2496 | pr_info("blk-mq: reduced tag depth to %u\n", | |
2497 | BLK_MQ_MAX_DEPTH); | |
2498 | set->queue_depth = BLK_MQ_MAX_DEPTH; | |
2499 | } | |
24d2f903 | 2500 | |
6637fadf SL |
2501 | /* |
2502 | * If a crashdump is active, then we are potentially in a very | |
2503 | * memory constrained environment. Limit us to 1 queue and | |
2504 | * 64 tags to prevent using too much memory. | |
2505 | */ | |
2506 | if (is_kdump_kernel()) { | |
2507 | set->nr_hw_queues = 1; | |
2508 | set->queue_depth = min(64U, set->queue_depth); | |
2509 | } | |
868f2f0b KB |
2510 | /* |
2511 | * There is no use for more h/w queues than cpus. | |
2512 | */ | |
2513 | if (set->nr_hw_queues > nr_cpu_ids) | |
2514 | set->nr_hw_queues = nr_cpu_ids; | |
6637fadf | 2515 | |
868f2f0b | 2516 | set->tags = kzalloc_node(nr_cpu_ids * sizeof(struct blk_mq_tags *), |
24d2f903 CH |
2517 | GFP_KERNEL, set->numa_node); |
2518 | if (!set->tags) | |
a5164405 | 2519 | return -ENOMEM; |
24d2f903 | 2520 | |
da695ba2 CH |
2521 | ret = -ENOMEM; |
2522 | set->mq_map = kzalloc_node(sizeof(*set->mq_map) * nr_cpu_ids, | |
2523 | GFP_KERNEL, set->numa_node); | |
bdd17e75 CH |
2524 | if (!set->mq_map) |
2525 | goto out_free_tags; | |
2526 | ||
da695ba2 CH |
2527 | if (set->ops->map_queues) |
2528 | ret = set->ops->map_queues(set); | |
2529 | else | |
2530 | ret = blk_mq_map_queues(set); | |
2531 | if (ret) | |
2532 | goto out_free_mq_map; | |
2533 | ||
2534 | ret = blk_mq_alloc_rq_maps(set); | |
2535 | if (ret) | |
bdd17e75 | 2536 | goto out_free_mq_map; |
24d2f903 | 2537 | |
0d2602ca JA |
2538 | mutex_init(&set->tag_list_lock); |
2539 | INIT_LIST_HEAD(&set->tag_list); | |
2540 | ||
24d2f903 | 2541 | return 0; |
bdd17e75 CH |
2542 | |
2543 | out_free_mq_map: | |
2544 | kfree(set->mq_map); | |
2545 | set->mq_map = NULL; | |
2546 | out_free_tags: | |
5676e7b6 RE |
2547 | kfree(set->tags); |
2548 | set->tags = NULL; | |
da695ba2 | 2549 | return ret; |
24d2f903 CH |
2550 | } |
2551 | EXPORT_SYMBOL(blk_mq_alloc_tag_set); | |
2552 | ||
2553 | void blk_mq_free_tag_set(struct blk_mq_tag_set *set) | |
2554 | { | |
2555 | int i; | |
2556 | ||
cc71a6f4 JA |
2557 | for (i = 0; i < nr_cpu_ids; i++) |
2558 | blk_mq_free_map_and_requests(set, i); | |
484b4061 | 2559 | |
bdd17e75 CH |
2560 | kfree(set->mq_map); |
2561 | set->mq_map = NULL; | |
2562 | ||
981bd189 | 2563 | kfree(set->tags); |
5676e7b6 | 2564 | set->tags = NULL; |
24d2f903 CH |
2565 | } |
2566 | EXPORT_SYMBOL(blk_mq_free_tag_set); | |
2567 | ||
e3a2b3f9 JA |
2568 | int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr) |
2569 | { | |
2570 | struct blk_mq_tag_set *set = q->tag_set; | |
2571 | struct blk_mq_hw_ctx *hctx; | |
2572 | int i, ret; | |
2573 | ||
bd166ef1 | 2574 | if (!set) |
e3a2b3f9 JA |
2575 | return -EINVAL; |
2576 | ||
70f36b60 JA |
2577 | blk_mq_freeze_queue(q); |
2578 | blk_mq_quiesce_queue(q); | |
2579 | ||
e3a2b3f9 JA |
2580 | ret = 0; |
2581 | queue_for_each_hw_ctx(q, hctx, i) { | |
e9137d4b KB |
2582 | if (!hctx->tags) |
2583 | continue; | |
bd166ef1 JA |
2584 | /* |
2585 | * If we're using an MQ scheduler, just update the scheduler | |
2586 | * queue depth. This is similar to what the old code would do. | |
2587 | */ | |
70f36b60 JA |
2588 | if (!hctx->sched_tags) { |
2589 | ret = blk_mq_tag_update_depth(hctx, &hctx->tags, | |
2590 | min(nr, set->queue_depth), | |
2591 | false); | |
2592 | } else { | |
2593 | ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags, | |
2594 | nr, true); | |
2595 | } | |
e3a2b3f9 JA |
2596 | if (ret) |
2597 | break; | |
2598 | } | |
2599 | ||
2600 | if (!ret) | |
2601 | q->nr_requests = nr; | |
2602 | ||
70f36b60 JA |
2603 | blk_mq_unfreeze_queue(q); |
2604 | blk_mq_start_stopped_hw_queues(q, true); | |
2605 | ||
e3a2b3f9 JA |
2606 | return ret; |
2607 | } | |
2608 | ||
868f2f0b KB |
2609 | void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues) |
2610 | { | |
2611 | struct request_queue *q; | |
2612 | ||
2613 | if (nr_hw_queues > nr_cpu_ids) | |
2614 | nr_hw_queues = nr_cpu_ids; | |
2615 | if (nr_hw_queues < 1 || nr_hw_queues == set->nr_hw_queues) | |
2616 | return; | |
2617 | ||
2618 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
2619 | blk_mq_freeze_queue(q); | |
2620 | ||
2621 | set->nr_hw_queues = nr_hw_queues; | |
2622 | list_for_each_entry(q, &set->tag_list, tag_set_list) { | |
2623 | blk_mq_realloc_hw_ctxs(set, q); | |
2624 | ||
2625 | if (q->nr_hw_queues > 1) | |
2626 | blk_queue_make_request(q, blk_mq_make_request); | |
2627 | else | |
2628 | blk_queue_make_request(q, blk_sq_make_request); | |
2629 | ||
2630 | blk_mq_queue_reinit(q, cpu_online_mask); | |
2631 | } | |
2632 | ||
2633 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
2634 | blk_mq_unfreeze_queue(q); | |
2635 | } | |
2636 | EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues); | |
2637 | ||
64f1c21e JA |
2638 | static unsigned long blk_mq_poll_nsecs(struct request_queue *q, |
2639 | struct blk_mq_hw_ctx *hctx, | |
2640 | struct request *rq) | |
2641 | { | |
2642 | struct blk_rq_stat stat[2]; | |
2643 | unsigned long ret = 0; | |
2644 | ||
2645 | /* | |
2646 | * If stats collection isn't on, don't sleep but turn it on for | |
2647 | * future users | |
2648 | */ | |
2649 | if (!blk_stat_enable(q)) | |
2650 | return 0; | |
2651 | ||
2652 | /* | |
2653 | * We don't have to do this once per IO, should optimize this | |
2654 | * to just use the current window of stats until it changes | |
2655 | */ | |
2656 | memset(&stat, 0, sizeof(stat)); | |
2657 | blk_hctx_stat_get(hctx, stat); | |
2658 | ||
2659 | /* | |
2660 | * As an optimistic guess, use half of the mean service time | |
2661 | * for this type of request. We can (and should) make this smarter. | |
2662 | * For instance, if the completion latencies are tight, we can | |
2663 | * get closer than just half the mean. This is especially | |
2664 | * important on devices where the completion latencies are longer | |
2665 | * than ~10 usec. | |
2666 | */ | |
2667 | if (req_op(rq) == REQ_OP_READ && stat[BLK_STAT_READ].nr_samples) | |
2668 | ret = (stat[BLK_STAT_READ].mean + 1) / 2; | |
2669 | else if (req_op(rq) == REQ_OP_WRITE && stat[BLK_STAT_WRITE].nr_samples) | |
2670 | ret = (stat[BLK_STAT_WRITE].mean + 1) / 2; | |
2671 | ||
2672 | return ret; | |
2673 | } | |
2674 | ||
06426adf | 2675 | static bool blk_mq_poll_hybrid_sleep(struct request_queue *q, |
64f1c21e | 2676 | struct blk_mq_hw_ctx *hctx, |
06426adf JA |
2677 | struct request *rq) |
2678 | { | |
2679 | struct hrtimer_sleeper hs; | |
2680 | enum hrtimer_mode mode; | |
64f1c21e | 2681 | unsigned int nsecs; |
06426adf JA |
2682 | ktime_t kt; |
2683 | ||
64f1c21e JA |
2684 | if (test_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags)) |
2685 | return false; | |
2686 | ||
2687 | /* | |
2688 | * poll_nsec can be: | |
2689 | * | |
2690 | * -1: don't ever hybrid sleep | |
2691 | * 0: use half of prev avg | |
2692 | * >0: use this specific value | |
2693 | */ | |
2694 | if (q->poll_nsec == -1) | |
2695 | return false; | |
2696 | else if (q->poll_nsec > 0) | |
2697 | nsecs = q->poll_nsec; | |
2698 | else | |
2699 | nsecs = blk_mq_poll_nsecs(q, hctx, rq); | |
2700 | ||
2701 | if (!nsecs) | |
06426adf JA |
2702 | return false; |
2703 | ||
2704 | set_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags); | |
2705 | ||
2706 | /* | |
2707 | * This will be replaced with the stats tracking code, using | |
2708 | * 'avg_completion_time / 2' as the pre-sleep target. | |
2709 | */ | |
8b0e1953 | 2710 | kt = nsecs; |
06426adf JA |
2711 | |
2712 | mode = HRTIMER_MODE_REL; | |
2713 | hrtimer_init_on_stack(&hs.timer, CLOCK_MONOTONIC, mode); | |
2714 | hrtimer_set_expires(&hs.timer, kt); | |
2715 | ||
2716 | hrtimer_init_sleeper(&hs, current); | |
2717 | do { | |
2718 | if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags)) | |
2719 | break; | |
2720 | set_current_state(TASK_UNINTERRUPTIBLE); | |
2721 | hrtimer_start_expires(&hs.timer, mode); | |
2722 | if (hs.task) | |
2723 | io_schedule(); | |
2724 | hrtimer_cancel(&hs.timer); | |
2725 | mode = HRTIMER_MODE_ABS; | |
2726 | } while (hs.task && !signal_pending(current)); | |
2727 | ||
2728 | __set_current_state(TASK_RUNNING); | |
2729 | destroy_hrtimer_on_stack(&hs.timer); | |
2730 | return true; | |
2731 | } | |
2732 | ||
bbd7bb70 JA |
2733 | static bool __blk_mq_poll(struct blk_mq_hw_ctx *hctx, struct request *rq) |
2734 | { | |
2735 | struct request_queue *q = hctx->queue; | |
2736 | long state; | |
2737 | ||
06426adf JA |
2738 | /* |
2739 | * If we sleep, have the caller restart the poll loop to reset | |
2740 | * the state. Like for the other success return cases, the | |
2741 | * caller is responsible for checking if the IO completed. If | |
2742 | * the IO isn't complete, we'll get called again and will go | |
2743 | * straight to the busy poll loop. | |
2744 | */ | |
64f1c21e | 2745 | if (blk_mq_poll_hybrid_sleep(q, hctx, rq)) |
06426adf JA |
2746 | return true; |
2747 | ||
bbd7bb70 JA |
2748 | hctx->poll_considered++; |
2749 | ||
2750 | state = current->state; | |
2751 | while (!need_resched()) { | |
2752 | int ret; | |
2753 | ||
2754 | hctx->poll_invoked++; | |
2755 | ||
2756 | ret = q->mq_ops->poll(hctx, rq->tag); | |
2757 | if (ret > 0) { | |
2758 | hctx->poll_success++; | |
2759 | set_current_state(TASK_RUNNING); | |
2760 | return true; | |
2761 | } | |
2762 | ||
2763 | if (signal_pending_state(state, current)) | |
2764 | set_current_state(TASK_RUNNING); | |
2765 | ||
2766 | if (current->state == TASK_RUNNING) | |
2767 | return true; | |
2768 | if (ret < 0) | |
2769 | break; | |
2770 | cpu_relax(); | |
2771 | } | |
2772 | ||
2773 | return false; | |
2774 | } | |
2775 | ||
2776 | bool blk_mq_poll(struct request_queue *q, blk_qc_t cookie) | |
2777 | { | |
2778 | struct blk_mq_hw_ctx *hctx; | |
2779 | struct blk_plug *plug; | |
2780 | struct request *rq; | |
2781 | ||
2782 | if (!q->mq_ops || !q->mq_ops->poll || !blk_qc_t_valid(cookie) || | |
2783 | !test_bit(QUEUE_FLAG_POLL, &q->queue_flags)) | |
2784 | return false; | |
2785 | ||
2786 | plug = current->plug; | |
2787 | if (plug) | |
2788 | blk_flush_plug_list(plug, false); | |
2789 | ||
2790 | hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)]; | |
bd166ef1 JA |
2791 | if (!blk_qc_t_is_internal(cookie)) |
2792 | rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie)); | |
2793 | else | |
2794 | rq = blk_mq_tag_to_rq(hctx->sched_tags, blk_qc_t_to_tag(cookie)); | |
bbd7bb70 JA |
2795 | |
2796 | return __blk_mq_poll(hctx, rq); | |
2797 | } | |
2798 | EXPORT_SYMBOL_GPL(blk_mq_poll); | |
2799 | ||
676141e4 JA |
2800 | void blk_mq_disable_hotplug(void) |
2801 | { | |
2802 | mutex_lock(&all_q_mutex); | |
2803 | } | |
2804 | ||
2805 | void blk_mq_enable_hotplug(void) | |
2806 | { | |
2807 | mutex_unlock(&all_q_mutex); | |
2808 | } | |
2809 | ||
320ae51f JA |
2810 | static int __init blk_mq_init(void) |
2811 | { | |
07e4fead OS |
2812 | blk_mq_debugfs_init(); |
2813 | ||
9467f859 TG |
2814 | cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL, |
2815 | blk_mq_hctx_notify_dead); | |
320ae51f | 2816 | |
65d5291e SAS |
2817 | cpuhp_setup_state_nocalls(CPUHP_BLK_MQ_PREPARE, "block/mq:prepare", |
2818 | blk_mq_queue_reinit_prepare, | |
2819 | blk_mq_queue_reinit_dead); | |
320ae51f JA |
2820 | return 0; |
2821 | } | |
2822 | subsys_initcall(blk_mq_init); |