]>
Commit | Line | Data |
---|---|---|
75bb4625 JA |
1 | /* |
2 | * Block multiqueue core code | |
3 | * | |
4 | * Copyright (C) 2013-2014 Jens Axboe | |
5 | * Copyright (C) 2013-2014 Christoph Hellwig | |
6 | */ | |
320ae51f JA |
7 | #include <linux/kernel.h> |
8 | #include <linux/module.h> | |
9 | #include <linux/backing-dev.h> | |
10 | #include <linux/bio.h> | |
11 | #include <linux/blkdev.h> | |
f75782e4 | 12 | #include <linux/kmemleak.h> |
320ae51f JA |
13 | #include <linux/mm.h> |
14 | #include <linux/init.h> | |
15 | #include <linux/slab.h> | |
16 | #include <linux/workqueue.h> | |
17 | #include <linux/smp.h> | |
18 | #include <linux/llist.h> | |
19 | #include <linux/list_sort.h> | |
20 | #include <linux/cpu.h> | |
21 | #include <linux/cache.h> | |
22 | #include <linux/sched/sysctl.h> | |
23 | #include <linux/delay.h> | |
aedcd72f | 24 | #include <linux/crash_dump.h> |
88c7b2b7 | 25 | #include <linux/prefetch.h> |
320ae51f JA |
26 | |
27 | #include <trace/events/block.h> | |
28 | ||
29 | #include <linux/blk-mq.h> | |
30 | #include "blk.h" | |
31 | #include "blk-mq.h" | |
32 | #include "blk-mq-tag.h" | |
cf43e6be | 33 | #include "blk-stat.h" |
87760e5e | 34 | #include "blk-wbt.h" |
320ae51f JA |
35 | |
36 | static DEFINE_MUTEX(all_q_mutex); | |
37 | static LIST_HEAD(all_q_list); | |
38 | ||
320ae51f JA |
39 | /* |
40 | * Check if any of the ctx's have pending work in this hardware queue | |
41 | */ | |
42 | static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx) | |
43 | { | |
88459642 | 44 | return sbitmap_any_bit_set(&hctx->ctx_map); |
1429d7c9 JA |
45 | } |
46 | ||
320ae51f JA |
47 | /* |
48 | * Mark this ctx as having pending work in this hardware queue | |
49 | */ | |
50 | static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx, | |
51 | struct blk_mq_ctx *ctx) | |
52 | { | |
88459642 OS |
53 | if (!sbitmap_test_bit(&hctx->ctx_map, ctx->index_hw)) |
54 | sbitmap_set_bit(&hctx->ctx_map, ctx->index_hw); | |
1429d7c9 JA |
55 | } |
56 | ||
57 | static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx, | |
58 | struct blk_mq_ctx *ctx) | |
59 | { | |
88459642 | 60 | sbitmap_clear_bit(&hctx->ctx_map, ctx->index_hw); |
320ae51f JA |
61 | } |
62 | ||
b4c6a028 | 63 | void blk_mq_freeze_queue_start(struct request_queue *q) |
43a5e4e2 | 64 | { |
4ecd4fef | 65 | int freeze_depth; |
cddd5d17 | 66 | |
4ecd4fef CH |
67 | freeze_depth = atomic_inc_return(&q->mq_freeze_depth); |
68 | if (freeze_depth == 1) { | |
3ef28e83 | 69 | percpu_ref_kill(&q->q_usage_counter); |
b94ec296 | 70 | blk_mq_run_hw_queues(q, false); |
cddd5d17 | 71 | } |
f3af020b | 72 | } |
b4c6a028 | 73 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_start); |
f3af020b TH |
74 | |
75 | static void blk_mq_freeze_queue_wait(struct request_queue *q) | |
76 | { | |
3ef28e83 | 77 | wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter)); |
43a5e4e2 ML |
78 | } |
79 | ||
f3af020b TH |
80 | /* |
81 | * Guarantee no request is in use, so we can change any data structure of | |
82 | * the queue afterward. | |
83 | */ | |
3ef28e83 | 84 | void blk_freeze_queue(struct request_queue *q) |
f3af020b | 85 | { |
3ef28e83 DW |
86 | /* |
87 | * In the !blk_mq case we are only calling this to kill the | |
88 | * q_usage_counter, otherwise this increases the freeze depth | |
89 | * and waits for it to return to zero. For this reason there is | |
90 | * no blk_unfreeze_queue(), and blk_freeze_queue() is not | |
91 | * exported to drivers as the only user for unfreeze is blk_mq. | |
92 | */ | |
f3af020b TH |
93 | blk_mq_freeze_queue_start(q); |
94 | blk_mq_freeze_queue_wait(q); | |
95 | } | |
3ef28e83 DW |
96 | |
97 | void blk_mq_freeze_queue(struct request_queue *q) | |
98 | { | |
99 | /* | |
100 | * ...just an alias to keep freeze and unfreeze actions balanced | |
101 | * in the blk_mq_* namespace | |
102 | */ | |
103 | blk_freeze_queue(q); | |
104 | } | |
c761d96b | 105 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue); |
f3af020b | 106 | |
b4c6a028 | 107 | void blk_mq_unfreeze_queue(struct request_queue *q) |
320ae51f | 108 | { |
4ecd4fef | 109 | int freeze_depth; |
320ae51f | 110 | |
4ecd4fef CH |
111 | freeze_depth = atomic_dec_return(&q->mq_freeze_depth); |
112 | WARN_ON_ONCE(freeze_depth < 0); | |
113 | if (!freeze_depth) { | |
3ef28e83 | 114 | percpu_ref_reinit(&q->q_usage_counter); |
320ae51f | 115 | wake_up_all(&q->mq_freeze_wq); |
add703fd | 116 | } |
320ae51f | 117 | } |
b4c6a028 | 118 | EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue); |
320ae51f | 119 | |
6a83e74d BVA |
120 | /** |
121 | * blk_mq_quiesce_queue() - wait until all ongoing queue_rq calls have finished | |
122 | * @q: request queue. | |
123 | * | |
124 | * Note: this function does not prevent that the struct request end_io() | |
125 | * callback function is invoked. Additionally, it is not prevented that | |
126 | * new queue_rq() calls occur unless the queue has been stopped first. | |
127 | */ | |
128 | void blk_mq_quiesce_queue(struct request_queue *q) | |
129 | { | |
130 | struct blk_mq_hw_ctx *hctx; | |
131 | unsigned int i; | |
132 | bool rcu = false; | |
133 | ||
134 | blk_mq_stop_hw_queues(q); | |
135 | ||
136 | queue_for_each_hw_ctx(q, hctx, i) { | |
137 | if (hctx->flags & BLK_MQ_F_BLOCKING) | |
138 | synchronize_srcu(&hctx->queue_rq_srcu); | |
139 | else | |
140 | rcu = true; | |
141 | } | |
142 | if (rcu) | |
143 | synchronize_rcu(); | |
144 | } | |
145 | EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue); | |
146 | ||
aed3ea94 JA |
147 | void blk_mq_wake_waiters(struct request_queue *q) |
148 | { | |
149 | struct blk_mq_hw_ctx *hctx; | |
150 | unsigned int i; | |
151 | ||
152 | queue_for_each_hw_ctx(q, hctx, i) | |
153 | if (blk_mq_hw_queue_mapped(hctx)) | |
154 | blk_mq_tag_wakeup_all(hctx->tags, true); | |
3fd5940c KB |
155 | |
156 | /* | |
157 | * If we are called because the queue has now been marked as | |
158 | * dying, we need to ensure that processes currently waiting on | |
159 | * the queue are notified as well. | |
160 | */ | |
161 | wake_up_all(&q->mq_freeze_wq); | |
aed3ea94 JA |
162 | } |
163 | ||
320ae51f JA |
164 | bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx) |
165 | { | |
166 | return blk_mq_has_free_tags(hctx->tags); | |
167 | } | |
168 | EXPORT_SYMBOL(blk_mq_can_queue); | |
169 | ||
94eddfbe | 170 | static void blk_mq_rq_ctx_init(struct request_queue *q, struct blk_mq_ctx *ctx, |
ef295ecf | 171 | struct request *rq, unsigned int op) |
320ae51f | 172 | { |
af76e555 CH |
173 | INIT_LIST_HEAD(&rq->queuelist); |
174 | /* csd/requeue_work/fifo_time is initialized before use */ | |
175 | rq->q = q; | |
320ae51f | 176 | rq->mq_ctx = ctx; |
ef295ecf | 177 | rq->cmd_flags = op; |
e8064021 CH |
178 | if (blk_queue_io_stat(q)) |
179 | rq->rq_flags |= RQF_IO_STAT; | |
af76e555 CH |
180 | /* do not touch atomic flags, it needs atomic ops against the timer */ |
181 | rq->cpu = -1; | |
af76e555 CH |
182 | INIT_HLIST_NODE(&rq->hash); |
183 | RB_CLEAR_NODE(&rq->rb_node); | |
af76e555 CH |
184 | rq->rq_disk = NULL; |
185 | rq->part = NULL; | |
3ee32372 | 186 | rq->start_time = jiffies; |
af76e555 CH |
187 | #ifdef CONFIG_BLK_CGROUP |
188 | rq->rl = NULL; | |
0fec08b4 | 189 | set_start_time_ns(rq); |
af76e555 CH |
190 | rq->io_start_time_ns = 0; |
191 | #endif | |
192 | rq->nr_phys_segments = 0; | |
193 | #if defined(CONFIG_BLK_DEV_INTEGRITY) | |
194 | rq->nr_integrity_segments = 0; | |
195 | #endif | |
af76e555 CH |
196 | rq->special = NULL; |
197 | /* tag was already set */ | |
198 | rq->errors = 0; | |
af76e555 | 199 | |
6f4a1626 TB |
200 | rq->cmd = rq->__cmd; |
201 | ||
af76e555 CH |
202 | rq->extra_len = 0; |
203 | rq->sense_len = 0; | |
204 | rq->resid_len = 0; | |
205 | rq->sense = NULL; | |
206 | ||
af76e555 | 207 | INIT_LIST_HEAD(&rq->timeout_list); |
f6be4fb4 JA |
208 | rq->timeout = 0; |
209 | ||
af76e555 CH |
210 | rq->end_io = NULL; |
211 | rq->end_io_data = NULL; | |
212 | rq->next_rq = NULL; | |
213 | ||
ef295ecf | 214 | ctx->rq_dispatched[op_is_sync(op)]++; |
320ae51f JA |
215 | } |
216 | ||
5dee8577 | 217 | static struct request * |
ef295ecf | 218 | __blk_mq_alloc_request(struct blk_mq_alloc_data *data, unsigned int op) |
5dee8577 CH |
219 | { |
220 | struct request *rq; | |
221 | unsigned int tag; | |
222 | ||
cb96a42c | 223 | tag = blk_mq_get_tag(data); |
5dee8577 | 224 | if (tag != BLK_MQ_TAG_FAIL) { |
cb96a42c | 225 | rq = data->hctx->tags->rqs[tag]; |
5dee8577 | 226 | |
cb96a42c | 227 | if (blk_mq_tag_busy(data->hctx)) { |
e8064021 | 228 | rq->rq_flags = RQF_MQ_INFLIGHT; |
cb96a42c | 229 | atomic_inc(&data->hctx->nr_active); |
5dee8577 CH |
230 | } |
231 | ||
232 | rq->tag = tag; | |
ef295ecf | 233 | blk_mq_rq_ctx_init(data->q, data->ctx, rq, op); |
5dee8577 CH |
234 | return rq; |
235 | } | |
236 | ||
237 | return NULL; | |
238 | } | |
239 | ||
6f3b0e8b CH |
240 | struct request *blk_mq_alloc_request(struct request_queue *q, int rw, |
241 | unsigned int flags) | |
320ae51f | 242 | { |
d852564f CH |
243 | struct blk_mq_ctx *ctx; |
244 | struct blk_mq_hw_ctx *hctx; | |
320ae51f | 245 | struct request *rq; |
cb96a42c | 246 | struct blk_mq_alloc_data alloc_data; |
a492f075 | 247 | int ret; |
320ae51f | 248 | |
6f3b0e8b | 249 | ret = blk_queue_enter(q, flags & BLK_MQ_REQ_NOWAIT); |
a492f075 JL |
250 | if (ret) |
251 | return ERR_PTR(ret); | |
320ae51f | 252 | |
d852564f | 253 | ctx = blk_mq_get_ctx(q); |
7d7e0f90 | 254 | hctx = blk_mq_map_queue(q, ctx->cpu); |
6f3b0e8b | 255 | blk_mq_set_alloc_data(&alloc_data, q, flags, ctx, hctx); |
ef295ecf | 256 | rq = __blk_mq_alloc_request(&alloc_data, rw); |
d852564f | 257 | blk_mq_put_ctx(ctx); |
841bac2c | 258 | |
c76541a9 | 259 | if (!rq) { |
3ef28e83 | 260 | blk_queue_exit(q); |
a492f075 | 261 | return ERR_PTR(-EWOULDBLOCK); |
c76541a9 | 262 | } |
0c4de0f3 CH |
263 | |
264 | rq->__data_len = 0; | |
265 | rq->__sector = (sector_t) -1; | |
266 | rq->bio = rq->biotail = NULL; | |
320ae51f JA |
267 | return rq; |
268 | } | |
4bb659b1 | 269 | EXPORT_SYMBOL(blk_mq_alloc_request); |
320ae51f | 270 | |
1f5bd336 ML |
271 | struct request *blk_mq_alloc_request_hctx(struct request_queue *q, int rw, |
272 | unsigned int flags, unsigned int hctx_idx) | |
273 | { | |
274 | struct blk_mq_hw_ctx *hctx; | |
275 | struct blk_mq_ctx *ctx; | |
276 | struct request *rq; | |
277 | struct blk_mq_alloc_data alloc_data; | |
278 | int ret; | |
279 | ||
280 | /* | |
281 | * If the tag allocator sleeps we could get an allocation for a | |
282 | * different hardware context. No need to complicate the low level | |
283 | * allocator for this for the rare use case of a command tied to | |
284 | * a specific queue. | |
285 | */ | |
286 | if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT))) | |
287 | return ERR_PTR(-EINVAL); | |
288 | ||
289 | if (hctx_idx >= q->nr_hw_queues) | |
290 | return ERR_PTR(-EIO); | |
291 | ||
292 | ret = blk_queue_enter(q, true); | |
293 | if (ret) | |
294 | return ERR_PTR(ret); | |
295 | ||
c8712c6a CH |
296 | /* |
297 | * Check if the hardware context is actually mapped to anything. | |
298 | * If not tell the caller that it should skip this queue. | |
299 | */ | |
1f5bd336 | 300 | hctx = q->queue_hw_ctx[hctx_idx]; |
c8712c6a CH |
301 | if (!blk_mq_hw_queue_mapped(hctx)) { |
302 | ret = -EXDEV; | |
303 | goto out_queue_exit; | |
304 | } | |
1f5bd336 ML |
305 | ctx = __blk_mq_get_ctx(q, cpumask_first(hctx->cpumask)); |
306 | ||
307 | blk_mq_set_alloc_data(&alloc_data, q, flags, ctx, hctx); | |
ef295ecf | 308 | rq = __blk_mq_alloc_request(&alloc_data, rw); |
1f5bd336 | 309 | if (!rq) { |
c8712c6a CH |
310 | ret = -EWOULDBLOCK; |
311 | goto out_queue_exit; | |
1f5bd336 ML |
312 | } |
313 | ||
314 | return rq; | |
c8712c6a CH |
315 | |
316 | out_queue_exit: | |
317 | blk_queue_exit(q); | |
318 | return ERR_PTR(ret); | |
1f5bd336 ML |
319 | } |
320 | EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx); | |
321 | ||
320ae51f JA |
322 | static void __blk_mq_free_request(struct blk_mq_hw_ctx *hctx, |
323 | struct blk_mq_ctx *ctx, struct request *rq) | |
324 | { | |
325 | const int tag = rq->tag; | |
326 | struct request_queue *q = rq->q; | |
327 | ||
e8064021 | 328 | if (rq->rq_flags & RQF_MQ_INFLIGHT) |
0d2602ca | 329 | atomic_dec(&hctx->nr_active); |
87760e5e JA |
330 | |
331 | wbt_done(q->rq_wb, &rq->issue_stat); | |
e8064021 | 332 | rq->rq_flags = 0; |
0d2602ca | 333 | |
af76e555 | 334 | clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags); |
40aabb67 | 335 | blk_mq_put_tag(hctx, ctx, tag); |
3ef28e83 | 336 | blk_queue_exit(q); |
320ae51f JA |
337 | } |
338 | ||
7c7f2f2b | 339 | void blk_mq_free_hctx_request(struct blk_mq_hw_ctx *hctx, struct request *rq) |
320ae51f JA |
340 | { |
341 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
320ae51f JA |
342 | |
343 | ctx->rq_completed[rq_is_sync(rq)]++; | |
320ae51f | 344 | __blk_mq_free_request(hctx, ctx, rq); |
7c7f2f2b JA |
345 | |
346 | } | |
347 | EXPORT_SYMBOL_GPL(blk_mq_free_hctx_request); | |
348 | ||
349 | void blk_mq_free_request(struct request *rq) | |
350 | { | |
7d7e0f90 | 351 | blk_mq_free_hctx_request(blk_mq_map_queue(rq->q, rq->mq_ctx->cpu), rq); |
320ae51f | 352 | } |
1a3b595a | 353 | EXPORT_SYMBOL_GPL(blk_mq_free_request); |
320ae51f | 354 | |
c8a446ad | 355 | inline void __blk_mq_end_request(struct request *rq, int error) |
320ae51f | 356 | { |
0d11e6ac ML |
357 | blk_account_io_done(rq); |
358 | ||
91b63639 | 359 | if (rq->end_io) { |
87760e5e | 360 | wbt_done(rq->q->rq_wb, &rq->issue_stat); |
320ae51f | 361 | rq->end_io(rq, error); |
91b63639 CH |
362 | } else { |
363 | if (unlikely(blk_bidi_rq(rq))) | |
364 | blk_mq_free_request(rq->next_rq); | |
320ae51f | 365 | blk_mq_free_request(rq); |
91b63639 | 366 | } |
320ae51f | 367 | } |
c8a446ad | 368 | EXPORT_SYMBOL(__blk_mq_end_request); |
63151a44 | 369 | |
c8a446ad | 370 | void blk_mq_end_request(struct request *rq, int error) |
63151a44 CH |
371 | { |
372 | if (blk_update_request(rq, error, blk_rq_bytes(rq))) | |
373 | BUG(); | |
c8a446ad | 374 | __blk_mq_end_request(rq, error); |
63151a44 | 375 | } |
c8a446ad | 376 | EXPORT_SYMBOL(blk_mq_end_request); |
320ae51f | 377 | |
30a91cb4 | 378 | static void __blk_mq_complete_request_remote(void *data) |
320ae51f | 379 | { |
3d6efbf6 | 380 | struct request *rq = data; |
320ae51f | 381 | |
30a91cb4 | 382 | rq->q->softirq_done_fn(rq); |
320ae51f | 383 | } |
320ae51f | 384 | |
ed851860 | 385 | static void blk_mq_ipi_complete_request(struct request *rq) |
320ae51f JA |
386 | { |
387 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
38535201 | 388 | bool shared = false; |
320ae51f JA |
389 | int cpu; |
390 | ||
38535201 | 391 | if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) { |
30a91cb4 CH |
392 | rq->q->softirq_done_fn(rq); |
393 | return; | |
394 | } | |
320ae51f JA |
395 | |
396 | cpu = get_cpu(); | |
38535201 CH |
397 | if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags)) |
398 | shared = cpus_share_cache(cpu, ctx->cpu); | |
399 | ||
400 | if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) { | |
30a91cb4 | 401 | rq->csd.func = __blk_mq_complete_request_remote; |
3d6efbf6 CH |
402 | rq->csd.info = rq; |
403 | rq->csd.flags = 0; | |
c46fff2a | 404 | smp_call_function_single_async(ctx->cpu, &rq->csd); |
3d6efbf6 | 405 | } else { |
30a91cb4 | 406 | rq->q->softirq_done_fn(rq); |
3d6efbf6 | 407 | } |
320ae51f JA |
408 | put_cpu(); |
409 | } | |
30a91cb4 | 410 | |
cf43e6be JA |
411 | static void blk_mq_stat_add(struct request *rq) |
412 | { | |
413 | if (rq->rq_flags & RQF_STATS) { | |
414 | /* | |
415 | * We could rq->mq_ctx here, but there's less of a risk | |
416 | * of races if we have the completion event add the stats | |
417 | * to the local software queue. | |
418 | */ | |
419 | struct blk_mq_ctx *ctx; | |
420 | ||
421 | ctx = __blk_mq_get_ctx(rq->q, raw_smp_processor_id()); | |
422 | blk_stat_add(&ctx->stat[rq_data_dir(rq)], rq); | |
423 | } | |
424 | } | |
425 | ||
1fa8cc52 | 426 | static void __blk_mq_complete_request(struct request *rq) |
ed851860 JA |
427 | { |
428 | struct request_queue *q = rq->q; | |
429 | ||
cf43e6be JA |
430 | blk_mq_stat_add(rq); |
431 | ||
ed851860 | 432 | if (!q->softirq_done_fn) |
c8a446ad | 433 | blk_mq_end_request(rq, rq->errors); |
ed851860 JA |
434 | else |
435 | blk_mq_ipi_complete_request(rq); | |
436 | } | |
437 | ||
30a91cb4 CH |
438 | /** |
439 | * blk_mq_complete_request - end I/O on a request | |
440 | * @rq: the request being processed | |
441 | * | |
442 | * Description: | |
443 | * Ends all I/O on a request. It does not handle partial completions. | |
444 | * The actual completion happens out-of-order, through a IPI handler. | |
445 | **/ | |
f4829a9b | 446 | void blk_mq_complete_request(struct request *rq, int error) |
30a91cb4 | 447 | { |
95f09684 JA |
448 | struct request_queue *q = rq->q; |
449 | ||
450 | if (unlikely(blk_should_fake_timeout(q))) | |
30a91cb4 | 451 | return; |
f4829a9b CH |
452 | if (!blk_mark_rq_complete(rq)) { |
453 | rq->errors = error; | |
ed851860 | 454 | __blk_mq_complete_request(rq); |
f4829a9b | 455 | } |
30a91cb4 CH |
456 | } |
457 | EXPORT_SYMBOL(blk_mq_complete_request); | |
320ae51f | 458 | |
973c0191 KB |
459 | int blk_mq_request_started(struct request *rq) |
460 | { | |
461 | return test_bit(REQ_ATOM_STARTED, &rq->atomic_flags); | |
462 | } | |
463 | EXPORT_SYMBOL_GPL(blk_mq_request_started); | |
464 | ||
e2490073 | 465 | void blk_mq_start_request(struct request *rq) |
320ae51f JA |
466 | { |
467 | struct request_queue *q = rq->q; | |
468 | ||
469 | trace_block_rq_issue(q, rq); | |
470 | ||
742ee69b | 471 | rq->resid_len = blk_rq_bytes(rq); |
91b63639 CH |
472 | if (unlikely(blk_bidi_rq(rq))) |
473 | rq->next_rq->resid_len = blk_rq_bytes(rq->next_rq); | |
742ee69b | 474 | |
cf43e6be JA |
475 | if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) { |
476 | blk_stat_set_issue_time(&rq->issue_stat); | |
477 | rq->rq_flags |= RQF_STATS; | |
87760e5e | 478 | wbt_issue(q->rq_wb, &rq->issue_stat); |
cf43e6be JA |
479 | } |
480 | ||
2b8393b4 | 481 | blk_add_timer(rq); |
87ee7b11 | 482 | |
538b7534 JA |
483 | /* |
484 | * Ensure that ->deadline is visible before set the started | |
485 | * flag and clear the completed flag. | |
486 | */ | |
487 | smp_mb__before_atomic(); | |
488 | ||
87ee7b11 JA |
489 | /* |
490 | * Mark us as started and clear complete. Complete might have been | |
491 | * set if requeue raced with timeout, which then marked it as | |
492 | * complete. So be sure to clear complete again when we start | |
493 | * the request, otherwise we'll ignore the completion event. | |
494 | */ | |
4b570521 JA |
495 | if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) |
496 | set_bit(REQ_ATOM_STARTED, &rq->atomic_flags); | |
497 | if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags)) | |
498 | clear_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags); | |
49f5baa5 CH |
499 | |
500 | if (q->dma_drain_size && blk_rq_bytes(rq)) { | |
501 | /* | |
502 | * Make sure space for the drain appears. We know we can do | |
503 | * this because max_hw_segments has been adjusted to be one | |
504 | * fewer than the device can handle. | |
505 | */ | |
506 | rq->nr_phys_segments++; | |
507 | } | |
320ae51f | 508 | } |
e2490073 | 509 | EXPORT_SYMBOL(blk_mq_start_request); |
320ae51f | 510 | |
ed0791b2 | 511 | static void __blk_mq_requeue_request(struct request *rq) |
320ae51f JA |
512 | { |
513 | struct request_queue *q = rq->q; | |
514 | ||
515 | trace_block_rq_requeue(q, rq); | |
87760e5e | 516 | wbt_requeue(q->rq_wb, &rq->issue_stat); |
49f5baa5 | 517 | |
e2490073 CH |
518 | if (test_and_clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) { |
519 | if (q->dma_drain_size && blk_rq_bytes(rq)) | |
520 | rq->nr_phys_segments--; | |
521 | } | |
320ae51f JA |
522 | } |
523 | ||
2b053aca | 524 | void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list) |
ed0791b2 | 525 | { |
ed0791b2 | 526 | __blk_mq_requeue_request(rq); |
ed0791b2 | 527 | |
ed0791b2 | 528 | BUG_ON(blk_queued_rq(rq)); |
2b053aca | 529 | blk_mq_add_to_requeue_list(rq, true, kick_requeue_list); |
ed0791b2 CH |
530 | } |
531 | EXPORT_SYMBOL(blk_mq_requeue_request); | |
532 | ||
6fca6a61 CH |
533 | static void blk_mq_requeue_work(struct work_struct *work) |
534 | { | |
535 | struct request_queue *q = | |
2849450a | 536 | container_of(work, struct request_queue, requeue_work.work); |
6fca6a61 CH |
537 | LIST_HEAD(rq_list); |
538 | struct request *rq, *next; | |
539 | unsigned long flags; | |
540 | ||
541 | spin_lock_irqsave(&q->requeue_lock, flags); | |
542 | list_splice_init(&q->requeue_list, &rq_list); | |
543 | spin_unlock_irqrestore(&q->requeue_lock, flags); | |
544 | ||
545 | list_for_each_entry_safe(rq, next, &rq_list, queuelist) { | |
e8064021 | 546 | if (!(rq->rq_flags & RQF_SOFTBARRIER)) |
6fca6a61 CH |
547 | continue; |
548 | ||
e8064021 | 549 | rq->rq_flags &= ~RQF_SOFTBARRIER; |
6fca6a61 CH |
550 | list_del_init(&rq->queuelist); |
551 | blk_mq_insert_request(rq, true, false, false); | |
552 | } | |
553 | ||
554 | while (!list_empty(&rq_list)) { | |
555 | rq = list_entry(rq_list.next, struct request, queuelist); | |
556 | list_del_init(&rq->queuelist); | |
557 | blk_mq_insert_request(rq, false, false, false); | |
558 | } | |
559 | ||
52d7f1b5 | 560 | blk_mq_run_hw_queues(q, false); |
6fca6a61 CH |
561 | } |
562 | ||
2b053aca BVA |
563 | void blk_mq_add_to_requeue_list(struct request *rq, bool at_head, |
564 | bool kick_requeue_list) | |
6fca6a61 CH |
565 | { |
566 | struct request_queue *q = rq->q; | |
567 | unsigned long flags; | |
568 | ||
569 | /* | |
570 | * We abuse this flag that is otherwise used by the I/O scheduler to | |
571 | * request head insertation from the workqueue. | |
572 | */ | |
e8064021 | 573 | BUG_ON(rq->rq_flags & RQF_SOFTBARRIER); |
6fca6a61 CH |
574 | |
575 | spin_lock_irqsave(&q->requeue_lock, flags); | |
576 | if (at_head) { | |
e8064021 | 577 | rq->rq_flags |= RQF_SOFTBARRIER; |
6fca6a61 CH |
578 | list_add(&rq->queuelist, &q->requeue_list); |
579 | } else { | |
580 | list_add_tail(&rq->queuelist, &q->requeue_list); | |
581 | } | |
582 | spin_unlock_irqrestore(&q->requeue_lock, flags); | |
2b053aca BVA |
583 | |
584 | if (kick_requeue_list) | |
585 | blk_mq_kick_requeue_list(q); | |
6fca6a61 CH |
586 | } |
587 | EXPORT_SYMBOL(blk_mq_add_to_requeue_list); | |
588 | ||
589 | void blk_mq_kick_requeue_list(struct request_queue *q) | |
590 | { | |
2849450a | 591 | kblockd_schedule_delayed_work(&q->requeue_work, 0); |
6fca6a61 CH |
592 | } |
593 | EXPORT_SYMBOL(blk_mq_kick_requeue_list); | |
594 | ||
2849450a MS |
595 | void blk_mq_delay_kick_requeue_list(struct request_queue *q, |
596 | unsigned long msecs) | |
597 | { | |
598 | kblockd_schedule_delayed_work(&q->requeue_work, | |
599 | msecs_to_jiffies(msecs)); | |
600 | } | |
601 | EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list); | |
602 | ||
1885b24d JA |
603 | void blk_mq_abort_requeue_list(struct request_queue *q) |
604 | { | |
605 | unsigned long flags; | |
606 | LIST_HEAD(rq_list); | |
607 | ||
608 | spin_lock_irqsave(&q->requeue_lock, flags); | |
609 | list_splice_init(&q->requeue_list, &rq_list); | |
610 | spin_unlock_irqrestore(&q->requeue_lock, flags); | |
611 | ||
612 | while (!list_empty(&rq_list)) { | |
613 | struct request *rq; | |
614 | ||
615 | rq = list_first_entry(&rq_list, struct request, queuelist); | |
616 | list_del_init(&rq->queuelist); | |
617 | rq->errors = -EIO; | |
618 | blk_mq_end_request(rq, rq->errors); | |
619 | } | |
620 | } | |
621 | EXPORT_SYMBOL(blk_mq_abort_requeue_list); | |
622 | ||
0e62f51f JA |
623 | struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag) |
624 | { | |
88c7b2b7 JA |
625 | if (tag < tags->nr_tags) { |
626 | prefetch(tags->rqs[tag]); | |
4ee86bab | 627 | return tags->rqs[tag]; |
88c7b2b7 | 628 | } |
4ee86bab HR |
629 | |
630 | return NULL; | |
24d2f903 CH |
631 | } |
632 | EXPORT_SYMBOL(blk_mq_tag_to_rq); | |
633 | ||
320ae51f | 634 | struct blk_mq_timeout_data { |
46f92d42 CH |
635 | unsigned long next; |
636 | unsigned int next_set; | |
320ae51f JA |
637 | }; |
638 | ||
90415837 | 639 | void blk_mq_rq_timed_out(struct request *req, bool reserved) |
320ae51f | 640 | { |
46f92d42 CH |
641 | struct blk_mq_ops *ops = req->q->mq_ops; |
642 | enum blk_eh_timer_return ret = BLK_EH_RESET_TIMER; | |
87ee7b11 JA |
643 | |
644 | /* | |
645 | * We know that complete is set at this point. If STARTED isn't set | |
646 | * anymore, then the request isn't active and the "timeout" should | |
647 | * just be ignored. This can happen due to the bitflag ordering. | |
648 | * Timeout first checks if STARTED is set, and if it is, assumes | |
649 | * the request is active. But if we race with completion, then | |
650 | * we both flags will get cleared. So check here again, and ignore | |
651 | * a timeout event with a request that isn't active. | |
652 | */ | |
46f92d42 CH |
653 | if (!test_bit(REQ_ATOM_STARTED, &req->atomic_flags)) |
654 | return; | |
87ee7b11 | 655 | |
46f92d42 | 656 | if (ops->timeout) |
0152fb6b | 657 | ret = ops->timeout(req, reserved); |
46f92d42 CH |
658 | |
659 | switch (ret) { | |
660 | case BLK_EH_HANDLED: | |
661 | __blk_mq_complete_request(req); | |
662 | break; | |
663 | case BLK_EH_RESET_TIMER: | |
664 | blk_add_timer(req); | |
665 | blk_clear_rq_complete(req); | |
666 | break; | |
667 | case BLK_EH_NOT_HANDLED: | |
668 | break; | |
669 | default: | |
670 | printk(KERN_ERR "block: bad eh return: %d\n", ret); | |
671 | break; | |
672 | } | |
87ee7b11 | 673 | } |
5b3f25fc | 674 | |
81481eb4 CH |
675 | static void blk_mq_check_expired(struct blk_mq_hw_ctx *hctx, |
676 | struct request *rq, void *priv, bool reserved) | |
677 | { | |
678 | struct blk_mq_timeout_data *data = priv; | |
87ee7b11 | 679 | |
eb130dbf KB |
680 | if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) { |
681 | /* | |
682 | * If a request wasn't started before the queue was | |
683 | * marked dying, kill it here or it'll go unnoticed. | |
684 | */ | |
a59e0f57 KB |
685 | if (unlikely(blk_queue_dying(rq->q))) { |
686 | rq->errors = -EIO; | |
687 | blk_mq_end_request(rq, rq->errors); | |
688 | } | |
46f92d42 | 689 | return; |
eb130dbf | 690 | } |
87ee7b11 | 691 | |
46f92d42 CH |
692 | if (time_after_eq(jiffies, rq->deadline)) { |
693 | if (!blk_mark_rq_complete(rq)) | |
0152fb6b | 694 | blk_mq_rq_timed_out(rq, reserved); |
46f92d42 CH |
695 | } else if (!data->next_set || time_after(data->next, rq->deadline)) { |
696 | data->next = rq->deadline; | |
697 | data->next_set = 1; | |
698 | } | |
87ee7b11 JA |
699 | } |
700 | ||
287922eb | 701 | static void blk_mq_timeout_work(struct work_struct *work) |
320ae51f | 702 | { |
287922eb CH |
703 | struct request_queue *q = |
704 | container_of(work, struct request_queue, timeout_work); | |
81481eb4 CH |
705 | struct blk_mq_timeout_data data = { |
706 | .next = 0, | |
707 | .next_set = 0, | |
708 | }; | |
81481eb4 | 709 | int i; |
320ae51f | 710 | |
71f79fb3 GKB |
711 | /* A deadlock might occur if a request is stuck requiring a |
712 | * timeout at the same time a queue freeze is waiting | |
713 | * completion, since the timeout code would not be able to | |
714 | * acquire the queue reference here. | |
715 | * | |
716 | * That's why we don't use blk_queue_enter here; instead, we use | |
717 | * percpu_ref_tryget directly, because we need to be able to | |
718 | * obtain a reference even in the short window between the queue | |
719 | * starting to freeze, by dropping the first reference in | |
720 | * blk_mq_freeze_queue_start, and the moment the last request is | |
721 | * consumed, marked by the instant q_usage_counter reaches | |
722 | * zero. | |
723 | */ | |
724 | if (!percpu_ref_tryget(&q->q_usage_counter)) | |
287922eb CH |
725 | return; |
726 | ||
0bf6cd5b | 727 | blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &data); |
320ae51f | 728 | |
81481eb4 CH |
729 | if (data.next_set) { |
730 | data.next = blk_rq_timeout(round_jiffies_up(data.next)); | |
731 | mod_timer(&q->timeout, data.next); | |
0d2602ca | 732 | } else { |
0bf6cd5b CH |
733 | struct blk_mq_hw_ctx *hctx; |
734 | ||
f054b56c ML |
735 | queue_for_each_hw_ctx(q, hctx, i) { |
736 | /* the hctx may be unmapped, so check it here */ | |
737 | if (blk_mq_hw_queue_mapped(hctx)) | |
738 | blk_mq_tag_idle(hctx); | |
739 | } | |
0d2602ca | 740 | } |
287922eb | 741 | blk_queue_exit(q); |
320ae51f JA |
742 | } |
743 | ||
744 | /* | |
745 | * Reverse check our software queue for entries that we could potentially | |
746 | * merge with. Currently includes a hand-wavy stop count of 8, to not spend | |
747 | * too much time checking for merges. | |
748 | */ | |
749 | static bool blk_mq_attempt_merge(struct request_queue *q, | |
750 | struct blk_mq_ctx *ctx, struct bio *bio) | |
751 | { | |
752 | struct request *rq; | |
753 | int checked = 8; | |
754 | ||
755 | list_for_each_entry_reverse(rq, &ctx->rq_list, queuelist) { | |
756 | int el_ret; | |
757 | ||
758 | if (!checked--) | |
759 | break; | |
760 | ||
761 | if (!blk_rq_merge_ok(rq, bio)) | |
762 | continue; | |
763 | ||
764 | el_ret = blk_try_merge(rq, bio); | |
765 | if (el_ret == ELEVATOR_BACK_MERGE) { | |
766 | if (bio_attempt_back_merge(q, rq, bio)) { | |
767 | ctx->rq_merged++; | |
768 | return true; | |
769 | } | |
770 | break; | |
771 | } else if (el_ret == ELEVATOR_FRONT_MERGE) { | |
772 | if (bio_attempt_front_merge(q, rq, bio)) { | |
773 | ctx->rq_merged++; | |
774 | return true; | |
775 | } | |
776 | break; | |
777 | } | |
778 | } | |
779 | ||
780 | return false; | |
781 | } | |
782 | ||
88459642 OS |
783 | struct flush_busy_ctx_data { |
784 | struct blk_mq_hw_ctx *hctx; | |
785 | struct list_head *list; | |
786 | }; | |
787 | ||
788 | static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data) | |
789 | { | |
790 | struct flush_busy_ctx_data *flush_data = data; | |
791 | struct blk_mq_hw_ctx *hctx = flush_data->hctx; | |
792 | struct blk_mq_ctx *ctx = hctx->ctxs[bitnr]; | |
793 | ||
794 | sbitmap_clear_bit(sb, bitnr); | |
795 | spin_lock(&ctx->lock); | |
796 | list_splice_tail_init(&ctx->rq_list, flush_data->list); | |
797 | spin_unlock(&ctx->lock); | |
798 | return true; | |
799 | } | |
800 | ||
1429d7c9 JA |
801 | /* |
802 | * Process software queues that have been marked busy, splicing them | |
803 | * to the for-dispatch | |
804 | */ | |
805 | static void flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list) | |
806 | { | |
88459642 OS |
807 | struct flush_busy_ctx_data data = { |
808 | .hctx = hctx, | |
809 | .list = list, | |
810 | }; | |
1429d7c9 | 811 | |
88459642 | 812 | sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data); |
1429d7c9 | 813 | } |
1429d7c9 | 814 | |
703fd1c0 JA |
815 | static inline unsigned int queued_to_index(unsigned int queued) |
816 | { | |
817 | if (!queued) | |
818 | return 0; | |
1429d7c9 | 819 | |
703fd1c0 | 820 | return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1); |
1429d7c9 JA |
821 | } |
822 | ||
320ae51f JA |
823 | /* |
824 | * Run this hardware queue, pulling any software queues mapped to it in. | |
825 | * Note that this function currently has various problems around ordering | |
826 | * of IO. In particular, we'd like FIFO behaviour on handling existing | |
827 | * items on the hctx->dispatch list. Ignore that for now. | |
828 | */ | |
6a83e74d | 829 | static void blk_mq_process_rq_list(struct blk_mq_hw_ctx *hctx) |
320ae51f JA |
830 | { |
831 | struct request_queue *q = hctx->queue; | |
320ae51f JA |
832 | struct request *rq; |
833 | LIST_HEAD(rq_list); | |
74c45052 JA |
834 | LIST_HEAD(driver_list); |
835 | struct list_head *dptr; | |
1429d7c9 | 836 | int queued; |
320ae51f | 837 | |
5d1b25c1 | 838 | if (unlikely(blk_mq_hctx_stopped(hctx))) |
320ae51f JA |
839 | return; |
840 | ||
841 | hctx->run++; | |
842 | ||
843 | /* | |
844 | * Touch any software queue that has pending entries. | |
845 | */ | |
1429d7c9 | 846 | flush_busy_ctxs(hctx, &rq_list); |
320ae51f JA |
847 | |
848 | /* | |
849 | * If we have previous entries on our dispatch list, grab them | |
850 | * and stuff them at the front for more fair dispatch. | |
851 | */ | |
852 | if (!list_empty_careful(&hctx->dispatch)) { | |
853 | spin_lock(&hctx->lock); | |
854 | if (!list_empty(&hctx->dispatch)) | |
855 | list_splice_init(&hctx->dispatch, &rq_list); | |
856 | spin_unlock(&hctx->lock); | |
857 | } | |
858 | ||
74c45052 JA |
859 | /* |
860 | * Start off with dptr being NULL, so we start the first request | |
861 | * immediately, even if we have more pending. | |
862 | */ | |
863 | dptr = NULL; | |
864 | ||
320ae51f JA |
865 | /* |
866 | * Now process all the entries, sending them to the driver. | |
867 | */ | |
1429d7c9 | 868 | queued = 0; |
320ae51f | 869 | while (!list_empty(&rq_list)) { |
74c45052 | 870 | struct blk_mq_queue_data bd; |
320ae51f JA |
871 | int ret; |
872 | ||
873 | rq = list_first_entry(&rq_list, struct request, queuelist); | |
874 | list_del_init(&rq->queuelist); | |
320ae51f | 875 | |
74c45052 JA |
876 | bd.rq = rq; |
877 | bd.list = dptr; | |
878 | bd.last = list_empty(&rq_list); | |
879 | ||
880 | ret = q->mq_ops->queue_rq(hctx, &bd); | |
320ae51f JA |
881 | switch (ret) { |
882 | case BLK_MQ_RQ_QUEUE_OK: | |
883 | queued++; | |
52b9c330 | 884 | break; |
320ae51f | 885 | case BLK_MQ_RQ_QUEUE_BUSY: |
320ae51f | 886 | list_add(&rq->queuelist, &rq_list); |
ed0791b2 | 887 | __blk_mq_requeue_request(rq); |
320ae51f JA |
888 | break; |
889 | default: | |
890 | pr_err("blk-mq: bad return on queue: %d\n", ret); | |
320ae51f | 891 | case BLK_MQ_RQ_QUEUE_ERROR: |
1e93b8c2 | 892 | rq->errors = -EIO; |
c8a446ad | 893 | blk_mq_end_request(rq, rq->errors); |
320ae51f JA |
894 | break; |
895 | } | |
896 | ||
897 | if (ret == BLK_MQ_RQ_QUEUE_BUSY) | |
898 | break; | |
74c45052 JA |
899 | |
900 | /* | |
901 | * We've done the first request. If we have more than 1 | |
902 | * left in the list, set dptr to defer issue. | |
903 | */ | |
904 | if (!dptr && rq_list.next != rq_list.prev) | |
905 | dptr = &driver_list; | |
320ae51f JA |
906 | } |
907 | ||
703fd1c0 | 908 | hctx->dispatched[queued_to_index(queued)]++; |
320ae51f JA |
909 | |
910 | /* | |
911 | * Any items that need requeuing? Stuff them into hctx->dispatch, | |
912 | * that is where we will continue on next queue run. | |
913 | */ | |
914 | if (!list_empty(&rq_list)) { | |
915 | spin_lock(&hctx->lock); | |
916 | list_splice(&rq_list, &hctx->dispatch); | |
917 | spin_unlock(&hctx->lock); | |
9ba52e58 SL |
918 | /* |
919 | * the queue is expected stopped with BLK_MQ_RQ_QUEUE_BUSY, but | |
920 | * it's possible the queue is stopped and restarted again | |
921 | * before this. Queue restart will dispatch requests. And since | |
922 | * requests in rq_list aren't added into hctx->dispatch yet, | |
923 | * the requests in rq_list might get lost. | |
924 | * | |
925 | * blk_mq_run_hw_queue() already checks the STOPPED bit | |
926 | **/ | |
927 | blk_mq_run_hw_queue(hctx, true); | |
320ae51f JA |
928 | } |
929 | } | |
930 | ||
6a83e74d BVA |
931 | static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx) |
932 | { | |
933 | int srcu_idx; | |
934 | ||
935 | WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) && | |
936 | cpu_online(hctx->next_cpu)); | |
937 | ||
938 | if (!(hctx->flags & BLK_MQ_F_BLOCKING)) { | |
939 | rcu_read_lock(); | |
940 | blk_mq_process_rq_list(hctx); | |
941 | rcu_read_unlock(); | |
942 | } else { | |
943 | srcu_idx = srcu_read_lock(&hctx->queue_rq_srcu); | |
944 | blk_mq_process_rq_list(hctx); | |
945 | srcu_read_unlock(&hctx->queue_rq_srcu, srcu_idx); | |
946 | } | |
947 | } | |
948 | ||
506e931f JA |
949 | /* |
950 | * It'd be great if the workqueue API had a way to pass | |
951 | * in a mask and had some smarts for more clever placement. | |
952 | * For now we just round-robin here, switching for every | |
953 | * BLK_MQ_CPU_WORK_BATCH queued items. | |
954 | */ | |
955 | static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx) | |
956 | { | |
b657d7e6 CH |
957 | if (hctx->queue->nr_hw_queues == 1) |
958 | return WORK_CPU_UNBOUND; | |
506e931f JA |
959 | |
960 | if (--hctx->next_cpu_batch <= 0) { | |
c02ebfdd | 961 | int next_cpu; |
506e931f JA |
962 | |
963 | next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask); | |
964 | if (next_cpu >= nr_cpu_ids) | |
965 | next_cpu = cpumask_first(hctx->cpumask); | |
966 | ||
967 | hctx->next_cpu = next_cpu; | |
968 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; | |
969 | } | |
970 | ||
b657d7e6 | 971 | return hctx->next_cpu; |
506e931f JA |
972 | } |
973 | ||
320ae51f JA |
974 | void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) |
975 | { | |
5d1b25c1 BVA |
976 | if (unlikely(blk_mq_hctx_stopped(hctx) || |
977 | !blk_mq_hw_queue_mapped(hctx))) | |
320ae51f JA |
978 | return; |
979 | ||
1b792f2f | 980 | if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) { |
2a90d4aa PB |
981 | int cpu = get_cpu(); |
982 | if (cpumask_test_cpu(cpu, hctx->cpumask)) { | |
398205b8 | 983 | __blk_mq_run_hw_queue(hctx); |
2a90d4aa | 984 | put_cpu(); |
398205b8 PB |
985 | return; |
986 | } | |
e4043dcf | 987 | |
2a90d4aa | 988 | put_cpu(); |
e4043dcf | 989 | } |
398205b8 | 990 | |
27489a3c | 991 | kblockd_schedule_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work); |
320ae51f JA |
992 | } |
993 | ||
b94ec296 | 994 | void blk_mq_run_hw_queues(struct request_queue *q, bool async) |
320ae51f JA |
995 | { |
996 | struct blk_mq_hw_ctx *hctx; | |
997 | int i; | |
998 | ||
999 | queue_for_each_hw_ctx(q, hctx, i) { | |
1000 | if ((!blk_mq_hctx_has_pending(hctx) && | |
1001 | list_empty_careful(&hctx->dispatch)) || | |
5d1b25c1 | 1002 | blk_mq_hctx_stopped(hctx)) |
320ae51f JA |
1003 | continue; |
1004 | ||
b94ec296 | 1005 | blk_mq_run_hw_queue(hctx, async); |
320ae51f JA |
1006 | } |
1007 | } | |
b94ec296 | 1008 | EXPORT_SYMBOL(blk_mq_run_hw_queues); |
320ae51f | 1009 | |
fd001443 BVA |
1010 | /** |
1011 | * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped | |
1012 | * @q: request queue. | |
1013 | * | |
1014 | * The caller is responsible for serializing this function against | |
1015 | * blk_mq_{start,stop}_hw_queue(). | |
1016 | */ | |
1017 | bool blk_mq_queue_stopped(struct request_queue *q) | |
1018 | { | |
1019 | struct blk_mq_hw_ctx *hctx; | |
1020 | int i; | |
1021 | ||
1022 | queue_for_each_hw_ctx(q, hctx, i) | |
1023 | if (blk_mq_hctx_stopped(hctx)) | |
1024 | return true; | |
1025 | ||
1026 | return false; | |
1027 | } | |
1028 | EXPORT_SYMBOL(blk_mq_queue_stopped); | |
1029 | ||
320ae51f JA |
1030 | void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx) |
1031 | { | |
27489a3c | 1032 | cancel_work(&hctx->run_work); |
70f4db63 | 1033 | cancel_delayed_work(&hctx->delay_work); |
320ae51f JA |
1034 | set_bit(BLK_MQ_S_STOPPED, &hctx->state); |
1035 | } | |
1036 | EXPORT_SYMBOL(blk_mq_stop_hw_queue); | |
1037 | ||
280d45f6 CH |
1038 | void blk_mq_stop_hw_queues(struct request_queue *q) |
1039 | { | |
1040 | struct blk_mq_hw_ctx *hctx; | |
1041 | int i; | |
1042 | ||
1043 | queue_for_each_hw_ctx(q, hctx, i) | |
1044 | blk_mq_stop_hw_queue(hctx); | |
1045 | } | |
1046 | EXPORT_SYMBOL(blk_mq_stop_hw_queues); | |
1047 | ||
320ae51f JA |
1048 | void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx) |
1049 | { | |
1050 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
e4043dcf | 1051 | |
0ffbce80 | 1052 | blk_mq_run_hw_queue(hctx, false); |
320ae51f JA |
1053 | } |
1054 | EXPORT_SYMBOL(blk_mq_start_hw_queue); | |
1055 | ||
2f268556 CH |
1056 | void blk_mq_start_hw_queues(struct request_queue *q) |
1057 | { | |
1058 | struct blk_mq_hw_ctx *hctx; | |
1059 | int i; | |
1060 | ||
1061 | queue_for_each_hw_ctx(q, hctx, i) | |
1062 | blk_mq_start_hw_queue(hctx); | |
1063 | } | |
1064 | EXPORT_SYMBOL(blk_mq_start_hw_queues); | |
1065 | ||
1b4a3258 | 1066 | void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async) |
320ae51f JA |
1067 | { |
1068 | struct blk_mq_hw_ctx *hctx; | |
1069 | int i; | |
1070 | ||
1071 | queue_for_each_hw_ctx(q, hctx, i) { | |
5d1b25c1 | 1072 | if (!blk_mq_hctx_stopped(hctx)) |
320ae51f JA |
1073 | continue; |
1074 | ||
1075 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
1b4a3258 | 1076 | blk_mq_run_hw_queue(hctx, async); |
320ae51f JA |
1077 | } |
1078 | } | |
1079 | EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues); | |
1080 | ||
70f4db63 | 1081 | static void blk_mq_run_work_fn(struct work_struct *work) |
320ae51f JA |
1082 | { |
1083 | struct blk_mq_hw_ctx *hctx; | |
1084 | ||
27489a3c | 1085 | hctx = container_of(work, struct blk_mq_hw_ctx, run_work); |
e4043dcf | 1086 | |
320ae51f JA |
1087 | __blk_mq_run_hw_queue(hctx); |
1088 | } | |
1089 | ||
70f4db63 CH |
1090 | static void blk_mq_delay_work_fn(struct work_struct *work) |
1091 | { | |
1092 | struct blk_mq_hw_ctx *hctx; | |
1093 | ||
1094 | hctx = container_of(work, struct blk_mq_hw_ctx, delay_work.work); | |
1095 | ||
1096 | if (test_and_clear_bit(BLK_MQ_S_STOPPED, &hctx->state)) | |
1097 | __blk_mq_run_hw_queue(hctx); | |
1098 | } | |
1099 | ||
1100 | void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs) | |
1101 | { | |
19c66e59 ML |
1102 | if (unlikely(!blk_mq_hw_queue_mapped(hctx))) |
1103 | return; | |
70f4db63 | 1104 | |
b657d7e6 CH |
1105 | kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx), |
1106 | &hctx->delay_work, msecs_to_jiffies(msecs)); | |
70f4db63 CH |
1107 | } |
1108 | EXPORT_SYMBOL(blk_mq_delay_queue); | |
1109 | ||
cfd0c552 | 1110 | static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx, |
cfd0c552 ML |
1111 | struct request *rq, |
1112 | bool at_head) | |
320ae51f | 1113 | { |
e57690fe JA |
1114 | struct blk_mq_ctx *ctx = rq->mq_ctx; |
1115 | ||
01b983c9 JA |
1116 | trace_block_rq_insert(hctx->queue, rq); |
1117 | ||
72a0a36e CH |
1118 | if (at_head) |
1119 | list_add(&rq->queuelist, &ctx->rq_list); | |
1120 | else | |
1121 | list_add_tail(&rq->queuelist, &ctx->rq_list); | |
cfd0c552 | 1122 | } |
4bb659b1 | 1123 | |
cfd0c552 ML |
1124 | static void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, |
1125 | struct request *rq, bool at_head) | |
1126 | { | |
1127 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
1128 | ||
e57690fe | 1129 | __blk_mq_insert_req_list(hctx, rq, at_head); |
320ae51f | 1130 | blk_mq_hctx_mark_pending(hctx, ctx); |
320ae51f JA |
1131 | } |
1132 | ||
eeabc850 | 1133 | void blk_mq_insert_request(struct request *rq, bool at_head, bool run_queue, |
e57690fe | 1134 | bool async) |
320ae51f | 1135 | { |
e57690fe | 1136 | struct blk_mq_ctx *ctx = rq->mq_ctx; |
eeabc850 | 1137 | struct request_queue *q = rq->q; |
7d7e0f90 | 1138 | struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, ctx->cpu); |
320ae51f | 1139 | |
a57a178a CH |
1140 | spin_lock(&ctx->lock); |
1141 | __blk_mq_insert_request(hctx, rq, at_head); | |
1142 | spin_unlock(&ctx->lock); | |
320ae51f | 1143 | |
320ae51f JA |
1144 | if (run_queue) |
1145 | blk_mq_run_hw_queue(hctx, async); | |
1146 | } | |
1147 | ||
1148 | static void blk_mq_insert_requests(struct request_queue *q, | |
1149 | struct blk_mq_ctx *ctx, | |
1150 | struct list_head *list, | |
1151 | int depth, | |
1152 | bool from_schedule) | |
1153 | ||
1154 | { | |
7d7e0f90 | 1155 | struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, ctx->cpu); |
320ae51f JA |
1156 | |
1157 | trace_block_unplug(q, depth, !from_schedule); | |
1158 | ||
320ae51f JA |
1159 | /* |
1160 | * preemption doesn't flush plug list, so it's possible ctx->cpu is | |
1161 | * offline now | |
1162 | */ | |
1163 | spin_lock(&ctx->lock); | |
1164 | while (!list_empty(list)) { | |
1165 | struct request *rq; | |
1166 | ||
1167 | rq = list_first_entry(list, struct request, queuelist); | |
e57690fe | 1168 | BUG_ON(rq->mq_ctx != ctx); |
320ae51f | 1169 | list_del_init(&rq->queuelist); |
e57690fe | 1170 | __blk_mq_insert_req_list(hctx, rq, false); |
320ae51f | 1171 | } |
cfd0c552 | 1172 | blk_mq_hctx_mark_pending(hctx, ctx); |
320ae51f JA |
1173 | spin_unlock(&ctx->lock); |
1174 | ||
320ae51f JA |
1175 | blk_mq_run_hw_queue(hctx, from_schedule); |
1176 | } | |
1177 | ||
1178 | static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b) | |
1179 | { | |
1180 | struct request *rqa = container_of(a, struct request, queuelist); | |
1181 | struct request *rqb = container_of(b, struct request, queuelist); | |
1182 | ||
1183 | return !(rqa->mq_ctx < rqb->mq_ctx || | |
1184 | (rqa->mq_ctx == rqb->mq_ctx && | |
1185 | blk_rq_pos(rqa) < blk_rq_pos(rqb))); | |
1186 | } | |
1187 | ||
1188 | void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule) | |
1189 | { | |
1190 | struct blk_mq_ctx *this_ctx; | |
1191 | struct request_queue *this_q; | |
1192 | struct request *rq; | |
1193 | LIST_HEAD(list); | |
1194 | LIST_HEAD(ctx_list); | |
1195 | unsigned int depth; | |
1196 | ||
1197 | list_splice_init(&plug->mq_list, &list); | |
1198 | ||
1199 | list_sort(NULL, &list, plug_ctx_cmp); | |
1200 | ||
1201 | this_q = NULL; | |
1202 | this_ctx = NULL; | |
1203 | depth = 0; | |
1204 | ||
1205 | while (!list_empty(&list)) { | |
1206 | rq = list_entry_rq(list.next); | |
1207 | list_del_init(&rq->queuelist); | |
1208 | BUG_ON(!rq->q); | |
1209 | if (rq->mq_ctx != this_ctx) { | |
1210 | if (this_ctx) { | |
1211 | blk_mq_insert_requests(this_q, this_ctx, | |
1212 | &ctx_list, depth, | |
1213 | from_schedule); | |
1214 | } | |
1215 | ||
1216 | this_ctx = rq->mq_ctx; | |
1217 | this_q = rq->q; | |
1218 | depth = 0; | |
1219 | } | |
1220 | ||
1221 | depth++; | |
1222 | list_add_tail(&rq->queuelist, &ctx_list); | |
1223 | } | |
1224 | ||
1225 | /* | |
1226 | * If 'this_ctx' is set, we know we have entries to complete | |
1227 | * on 'ctx_list'. Do those. | |
1228 | */ | |
1229 | if (this_ctx) { | |
1230 | blk_mq_insert_requests(this_q, this_ctx, &ctx_list, depth, | |
1231 | from_schedule); | |
1232 | } | |
1233 | } | |
1234 | ||
1235 | static void blk_mq_bio_to_request(struct request *rq, struct bio *bio) | |
1236 | { | |
1237 | init_request_from_bio(rq, bio); | |
4b570521 | 1238 | |
a21f2a3e | 1239 | blk_account_io_start(rq, 1); |
320ae51f JA |
1240 | } |
1241 | ||
274a5843 JA |
1242 | static inline bool hctx_allow_merges(struct blk_mq_hw_ctx *hctx) |
1243 | { | |
1244 | return (hctx->flags & BLK_MQ_F_SHOULD_MERGE) && | |
1245 | !blk_queue_nomerges(hctx->queue); | |
1246 | } | |
1247 | ||
07068d5b JA |
1248 | static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx *hctx, |
1249 | struct blk_mq_ctx *ctx, | |
1250 | struct request *rq, struct bio *bio) | |
320ae51f | 1251 | { |
e18378a6 | 1252 | if (!hctx_allow_merges(hctx) || !bio_mergeable(bio)) { |
07068d5b JA |
1253 | blk_mq_bio_to_request(rq, bio); |
1254 | spin_lock(&ctx->lock); | |
1255 | insert_rq: | |
1256 | __blk_mq_insert_request(hctx, rq, false); | |
1257 | spin_unlock(&ctx->lock); | |
1258 | return false; | |
1259 | } else { | |
274a5843 JA |
1260 | struct request_queue *q = hctx->queue; |
1261 | ||
07068d5b JA |
1262 | spin_lock(&ctx->lock); |
1263 | if (!blk_mq_attempt_merge(q, ctx, bio)) { | |
1264 | blk_mq_bio_to_request(rq, bio); | |
1265 | goto insert_rq; | |
1266 | } | |
320ae51f | 1267 | |
07068d5b JA |
1268 | spin_unlock(&ctx->lock); |
1269 | __blk_mq_free_request(hctx, ctx, rq); | |
1270 | return true; | |
14ec77f3 | 1271 | } |
07068d5b | 1272 | } |
14ec77f3 | 1273 | |
07068d5b JA |
1274 | static struct request *blk_mq_map_request(struct request_queue *q, |
1275 | struct bio *bio, | |
2552e3f8 | 1276 | struct blk_mq_alloc_data *data) |
07068d5b JA |
1277 | { |
1278 | struct blk_mq_hw_ctx *hctx; | |
1279 | struct blk_mq_ctx *ctx; | |
1280 | struct request *rq; | |
320ae51f | 1281 | |
3ef28e83 | 1282 | blk_queue_enter_live(q); |
320ae51f | 1283 | ctx = blk_mq_get_ctx(q); |
7d7e0f90 | 1284 | hctx = blk_mq_map_queue(q, ctx->cpu); |
320ae51f | 1285 | |
ef295ecf | 1286 | trace_block_getrq(q, bio, bio->bi_opf); |
2552e3f8 | 1287 | blk_mq_set_alloc_data(data, q, 0, ctx, hctx); |
ef295ecf | 1288 | rq = __blk_mq_alloc_request(data, bio->bi_opf); |
320ae51f | 1289 | |
7dd2fb68 | 1290 | data->hctx->queued++; |
07068d5b JA |
1291 | return rq; |
1292 | } | |
1293 | ||
066a4a73 | 1294 | static void blk_mq_try_issue_directly(struct request *rq, blk_qc_t *cookie) |
f984df1f SL |
1295 | { |
1296 | int ret; | |
1297 | struct request_queue *q = rq->q; | |
066a4a73 | 1298 | struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, rq->mq_ctx->cpu); |
f984df1f SL |
1299 | struct blk_mq_queue_data bd = { |
1300 | .rq = rq, | |
1301 | .list = NULL, | |
1302 | .last = 1 | |
1303 | }; | |
7b371636 | 1304 | blk_qc_t new_cookie = blk_tag_to_qc_t(rq->tag, hctx->queue_num); |
f984df1f | 1305 | |
2253efc8 BVA |
1306 | if (blk_mq_hctx_stopped(hctx)) |
1307 | goto insert; | |
1308 | ||
f984df1f SL |
1309 | /* |
1310 | * For OK queue, we are done. For error, kill it. Any other | |
1311 | * error (busy), just add it to our list as we previously | |
1312 | * would have done | |
1313 | */ | |
1314 | ret = q->mq_ops->queue_rq(hctx, &bd); | |
7b371636 JA |
1315 | if (ret == BLK_MQ_RQ_QUEUE_OK) { |
1316 | *cookie = new_cookie; | |
2253efc8 | 1317 | return; |
7b371636 | 1318 | } |
f984df1f | 1319 | |
7b371636 JA |
1320 | __blk_mq_requeue_request(rq); |
1321 | ||
1322 | if (ret == BLK_MQ_RQ_QUEUE_ERROR) { | |
1323 | *cookie = BLK_QC_T_NONE; | |
1324 | rq->errors = -EIO; | |
1325 | blk_mq_end_request(rq, rq->errors); | |
2253efc8 | 1326 | return; |
f984df1f | 1327 | } |
7b371636 | 1328 | |
2253efc8 BVA |
1329 | insert: |
1330 | blk_mq_insert_request(rq, false, true, true); | |
f984df1f SL |
1331 | } |
1332 | ||
07068d5b JA |
1333 | /* |
1334 | * Multiple hardware queue variant. This will not use per-process plugs, | |
1335 | * but will attempt to bypass the hctx queueing if we can go straight to | |
1336 | * hardware for SYNC IO. | |
1337 | */ | |
dece1635 | 1338 | static blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio) |
07068d5b | 1339 | { |
ef295ecf | 1340 | const int is_sync = op_is_sync(bio->bi_opf); |
1eff9d32 | 1341 | const int is_flush_fua = bio->bi_opf & (REQ_PREFLUSH | REQ_FUA); |
2552e3f8 | 1342 | struct blk_mq_alloc_data data; |
07068d5b | 1343 | struct request *rq; |
6a83e74d | 1344 | unsigned int request_count = 0, srcu_idx; |
f984df1f | 1345 | struct blk_plug *plug; |
5b3f341f | 1346 | struct request *same_queue_rq = NULL; |
7b371636 | 1347 | blk_qc_t cookie; |
87760e5e | 1348 | unsigned int wb_acct; |
07068d5b JA |
1349 | |
1350 | blk_queue_bounce(q, &bio); | |
1351 | ||
1352 | if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) { | |
4246a0b6 | 1353 | bio_io_error(bio); |
dece1635 | 1354 | return BLK_QC_T_NONE; |
07068d5b JA |
1355 | } |
1356 | ||
54efd50b KO |
1357 | blk_queue_split(q, &bio, q->bio_split); |
1358 | ||
87c279e6 OS |
1359 | if (!is_flush_fua && !blk_queue_nomerges(q) && |
1360 | blk_attempt_plug_merge(q, bio, &request_count, &same_queue_rq)) | |
1361 | return BLK_QC_T_NONE; | |
f984df1f | 1362 | |
87760e5e JA |
1363 | wb_acct = wbt_wait(q->rq_wb, bio, NULL); |
1364 | ||
07068d5b | 1365 | rq = blk_mq_map_request(q, bio, &data); |
87760e5e JA |
1366 | if (unlikely(!rq)) { |
1367 | __wbt_done(q->rq_wb, wb_acct); | |
dece1635 | 1368 | return BLK_QC_T_NONE; |
87760e5e JA |
1369 | } |
1370 | ||
1371 | wbt_track(&rq->issue_stat, wb_acct); | |
07068d5b | 1372 | |
7b371636 | 1373 | cookie = blk_tag_to_qc_t(rq->tag, data.hctx->queue_num); |
07068d5b JA |
1374 | |
1375 | if (unlikely(is_flush_fua)) { | |
1376 | blk_mq_bio_to_request(rq, bio); | |
1377 | blk_insert_flush(rq); | |
1378 | goto run_queue; | |
1379 | } | |
1380 | ||
f984df1f | 1381 | plug = current->plug; |
e167dfb5 JA |
1382 | /* |
1383 | * If the driver supports defer issued based on 'last', then | |
1384 | * queue it up like normal since we can potentially save some | |
1385 | * CPU this way. | |
1386 | */ | |
f984df1f SL |
1387 | if (((plug && !blk_queue_nomerges(q)) || is_sync) && |
1388 | !(data.hctx->flags & BLK_MQ_F_DEFER_ISSUE)) { | |
1389 | struct request *old_rq = NULL; | |
07068d5b JA |
1390 | |
1391 | blk_mq_bio_to_request(rq, bio); | |
07068d5b JA |
1392 | |
1393 | /* | |
6a83e74d | 1394 | * We do limited plugging. If the bio can be merged, do that. |
f984df1f SL |
1395 | * Otherwise the existing request in the plug list will be |
1396 | * issued. So the plug list will have one request at most | |
07068d5b | 1397 | */ |
f984df1f | 1398 | if (plug) { |
5b3f341f SL |
1399 | /* |
1400 | * The plug list might get flushed before this. If that | |
b094f89c JA |
1401 | * happens, same_queue_rq is invalid and plug list is |
1402 | * empty | |
1403 | */ | |
5b3f341f SL |
1404 | if (same_queue_rq && !list_empty(&plug->mq_list)) { |
1405 | old_rq = same_queue_rq; | |
f984df1f | 1406 | list_del_init(&old_rq->queuelist); |
07068d5b | 1407 | } |
f984df1f SL |
1408 | list_add_tail(&rq->queuelist, &plug->mq_list); |
1409 | } else /* is_sync */ | |
1410 | old_rq = rq; | |
1411 | blk_mq_put_ctx(data.ctx); | |
1412 | if (!old_rq) | |
7b371636 | 1413 | goto done; |
6a83e74d BVA |
1414 | |
1415 | if (!(data.hctx->flags & BLK_MQ_F_BLOCKING)) { | |
1416 | rcu_read_lock(); | |
066a4a73 | 1417 | blk_mq_try_issue_directly(old_rq, &cookie); |
6a83e74d BVA |
1418 | rcu_read_unlock(); |
1419 | } else { | |
1420 | srcu_idx = srcu_read_lock(&data.hctx->queue_rq_srcu); | |
066a4a73 | 1421 | blk_mq_try_issue_directly(old_rq, &cookie); |
6a83e74d BVA |
1422 | srcu_read_unlock(&data.hctx->queue_rq_srcu, srcu_idx); |
1423 | } | |
7b371636 | 1424 | goto done; |
07068d5b JA |
1425 | } |
1426 | ||
1427 | if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) { | |
1428 | /* | |
1429 | * For a SYNC request, send it to the hardware immediately. For | |
1430 | * an ASYNC request, just ensure that we run it later on. The | |
1431 | * latter allows for merging opportunities and more efficient | |
1432 | * dispatching. | |
1433 | */ | |
1434 | run_queue: | |
1435 | blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua); | |
1436 | } | |
07068d5b | 1437 | blk_mq_put_ctx(data.ctx); |
7b371636 JA |
1438 | done: |
1439 | return cookie; | |
07068d5b JA |
1440 | } |
1441 | ||
1442 | /* | |
1443 | * Single hardware queue variant. This will attempt to use any per-process | |
1444 | * plug for merging and IO deferral. | |
1445 | */ | |
dece1635 | 1446 | static blk_qc_t blk_sq_make_request(struct request_queue *q, struct bio *bio) |
07068d5b | 1447 | { |
ef295ecf | 1448 | const int is_sync = op_is_sync(bio->bi_opf); |
1eff9d32 | 1449 | const int is_flush_fua = bio->bi_opf & (REQ_PREFLUSH | REQ_FUA); |
e6c4438b JM |
1450 | struct blk_plug *plug; |
1451 | unsigned int request_count = 0; | |
2552e3f8 | 1452 | struct blk_mq_alloc_data data; |
07068d5b | 1453 | struct request *rq; |
7b371636 | 1454 | blk_qc_t cookie; |
87760e5e | 1455 | unsigned int wb_acct; |
07068d5b | 1456 | |
07068d5b JA |
1457 | blk_queue_bounce(q, &bio); |
1458 | ||
1459 | if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) { | |
4246a0b6 | 1460 | bio_io_error(bio); |
dece1635 | 1461 | return BLK_QC_T_NONE; |
07068d5b JA |
1462 | } |
1463 | ||
54efd50b KO |
1464 | blk_queue_split(q, &bio, q->bio_split); |
1465 | ||
87c279e6 OS |
1466 | if (!is_flush_fua && !blk_queue_nomerges(q)) { |
1467 | if (blk_attempt_plug_merge(q, bio, &request_count, NULL)) | |
1468 | return BLK_QC_T_NONE; | |
1469 | } else | |
1470 | request_count = blk_plug_queued_count(q); | |
07068d5b | 1471 | |
87760e5e JA |
1472 | wb_acct = wbt_wait(q->rq_wb, bio, NULL); |
1473 | ||
07068d5b | 1474 | rq = blk_mq_map_request(q, bio, &data); |
87760e5e JA |
1475 | if (unlikely(!rq)) { |
1476 | __wbt_done(q->rq_wb, wb_acct); | |
dece1635 | 1477 | return BLK_QC_T_NONE; |
87760e5e JA |
1478 | } |
1479 | ||
1480 | wbt_track(&rq->issue_stat, wb_acct); | |
320ae51f | 1481 | |
7b371636 | 1482 | cookie = blk_tag_to_qc_t(rq->tag, data.hctx->queue_num); |
320ae51f JA |
1483 | |
1484 | if (unlikely(is_flush_fua)) { | |
1485 | blk_mq_bio_to_request(rq, bio); | |
320ae51f JA |
1486 | blk_insert_flush(rq); |
1487 | goto run_queue; | |
1488 | } | |
1489 | ||
1490 | /* | |
1491 | * A task plug currently exists. Since this is completely lockless, | |
1492 | * utilize that to temporarily store requests until the task is | |
1493 | * either done or scheduled away. | |
1494 | */ | |
e6c4438b JM |
1495 | plug = current->plug; |
1496 | if (plug) { | |
600271d9 SL |
1497 | struct request *last = NULL; |
1498 | ||
e6c4438b | 1499 | blk_mq_bio_to_request(rq, bio); |
0a6219a9 ML |
1500 | |
1501 | /* | |
1502 | * @request_count may become stale because of schedule | |
1503 | * out, so check the list again. | |
1504 | */ | |
1505 | if (list_empty(&plug->mq_list)) | |
1506 | request_count = 0; | |
676d0607 | 1507 | if (!request_count) |
e6c4438b | 1508 | trace_block_plug(q); |
600271d9 SL |
1509 | else |
1510 | last = list_entry_rq(plug->mq_list.prev); | |
b094f89c JA |
1511 | |
1512 | blk_mq_put_ctx(data.ctx); | |
1513 | ||
600271d9 SL |
1514 | if (request_count >= BLK_MAX_REQUEST_COUNT || (last && |
1515 | blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) { | |
e6c4438b JM |
1516 | blk_flush_plug_list(plug, false); |
1517 | trace_block_plug(q); | |
320ae51f | 1518 | } |
b094f89c | 1519 | |
e6c4438b | 1520 | list_add_tail(&rq->queuelist, &plug->mq_list); |
7b371636 | 1521 | return cookie; |
320ae51f JA |
1522 | } |
1523 | ||
07068d5b JA |
1524 | if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) { |
1525 | /* | |
1526 | * For a SYNC request, send it to the hardware immediately. For | |
1527 | * an ASYNC request, just ensure that we run it later on. The | |
1528 | * latter allows for merging opportunities and more efficient | |
1529 | * dispatching. | |
1530 | */ | |
1531 | run_queue: | |
1532 | blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua); | |
320ae51f JA |
1533 | } |
1534 | ||
07068d5b | 1535 | blk_mq_put_ctx(data.ctx); |
7b371636 | 1536 | return cookie; |
320ae51f JA |
1537 | } |
1538 | ||
24d2f903 CH |
1539 | static void blk_mq_free_rq_map(struct blk_mq_tag_set *set, |
1540 | struct blk_mq_tags *tags, unsigned int hctx_idx) | |
95363efd | 1541 | { |
e9b267d9 | 1542 | struct page *page; |
320ae51f | 1543 | |
24d2f903 | 1544 | if (tags->rqs && set->ops->exit_request) { |
e9b267d9 | 1545 | int i; |
320ae51f | 1546 | |
24d2f903 CH |
1547 | for (i = 0; i < tags->nr_tags; i++) { |
1548 | if (!tags->rqs[i]) | |
e9b267d9 | 1549 | continue; |
24d2f903 CH |
1550 | set->ops->exit_request(set->driver_data, tags->rqs[i], |
1551 | hctx_idx, i); | |
a5164405 | 1552 | tags->rqs[i] = NULL; |
e9b267d9 | 1553 | } |
320ae51f | 1554 | } |
320ae51f | 1555 | |
24d2f903 CH |
1556 | while (!list_empty(&tags->page_list)) { |
1557 | page = list_first_entry(&tags->page_list, struct page, lru); | |
6753471c | 1558 | list_del_init(&page->lru); |
f75782e4 CM |
1559 | /* |
1560 | * Remove kmemleak object previously allocated in | |
1561 | * blk_mq_init_rq_map(). | |
1562 | */ | |
1563 | kmemleak_free(page_address(page)); | |
320ae51f JA |
1564 | __free_pages(page, page->private); |
1565 | } | |
1566 | ||
24d2f903 | 1567 | kfree(tags->rqs); |
320ae51f | 1568 | |
24d2f903 | 1569 | blk_mq_free_tags(tags); |
320ae51f JA |
1570 | } |
1571 | ||
1572 | static size_t order_to_size(unsigned int order) | |
1573 | { | |
4ca08500 | 1574 | return (size_t)PAGE_SIZE << order; |
320ae51f JA |
1575 | } |
1576 | ||
24d2f903 CH |
1577 | static struct blk_mq_tags *blk_mq_init_rq_map(struct blk_mq_tag_set *set, |
1578 | unsigned int hctx_idx) | |
320ae51f | 1579 | { |
24d2f903 | 1580 | struct blk_mq_tags *tags; |
320ae51f JA |
1581 | unsigned int i, j, entries_per_page, max_order = 4; |
1582 | size_t rq_size, left; | |
1583 | ||
24d2f903 | 1584 | tags = blk_mq_init_tags(set->queue_depth, set->reserved_tags, |
24391c0d SL |
1585 | set->numa_node, |
1586 | BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags)); | |
24d2f903 CH |
1587 | if (!tags) |
1588 | return NULL; | |
320ae51f | 1589 | |
24d2f903 CH |
1590 | INIT_LIST_HEAD(&tags->page_list); |
1591 | ||
a5164405 JA |
1592 | tags->rqs = kzalloc_node(set->queue_depth * sizeof(struct request *), |
1593 | GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY, | |
1594 | set->numa_node); | |
24d2f903 CH |
1595 | if (!tags->rqs) { |
1596 | blk_mq_free_tags(tags); | |
1597 | return NULL; | |
1598 | } | |
320ae51f JA |
1599 | |
1600 | /* | |
1601 | * rq_size is the size of the request plus driver payload, rounded | |
1602 | * to the cacheline size | |
1603 | */ | |
24d2f903 | 1604 | rq_size = round_up(sizeof(struct request) + set->cmd_size, |
320ae51f | 1605 | cache_line_size()); |
24d2f903 | 1606 | left = rq_size * set->queue_depth; |
320ae51f | 1607 | |
24d2f903 | 1608 | for (i = 0; i < set->queue_depth; ) { |
320ae51f JA |
1609 | int this_order = max_order; |
1610 | struct page *page; | |
1611 | int to_do; | |
1612 | void *p; | |
1613 | ||
b3a834b1 | 1614 | while (this_order && left < order_to_size(this_order - 1)) |
320ae51f JA |
1615 | this_order--; |
1616 | ||
1617 | do { | |
a5164405 | 1618 | page = alloc_pages_node(set->numa_node, |
ac211175 | 1619 | GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO, |
a5164405 | 1620 | this_order); |
320ae51f JA |
1621 | if (page) |
1622 | break; | |
1623 | if (!this_order--) | |
1624 | break; | |
1625 | if (order_to_size(this_order) < rq_size) | |
1626 | break; | |
1627 | } while (1); | |
1628 | ||
1629 | if (!page) | |
24d2f903 | 1630 | goto fail; |
320ae51f JA |
1631 | |
1632 | page->private = this_order; | |
24d2f903 | 1633 | list_add_tail(&page->lru, &tags->page_list); |
320ae51f JA |
1634 | |
1635 | p = page_address(page); | |
f75782e4 CM |
1636 | /* |
1637 | * Allow kmemleak to scan these pages as they contain pointers | |
1638 | * to additional allocations like via ops->init_request(). | |
1639 | */ | |
1640 | kmemleak_alloc(p, order_to_size(this_order), 1, GFP_KERNEL); | |
320ae51f | 1641 | entries_per_page = order_to_size(this_order) / rq_size; |
24d2f903 | 1642 | to_do = min(entries_per_page, set->queue_depth - i); |
320ae51f JA |
1643 | left -= to_do * rq_size; |
1644 | for (j = 0; j < to_do; j++) { | |
24d2f903 CH |
1645 | tags->rqs[i] = p; |
1646 | if (set->ops->init_request) { | |
1647 | if (set->ops->init_request(set->driver_data, | |
1648 | tags->rqs[i], hctx_idx, i, | |
a5164405 JA |
1649 | set->numa_node)) { |
1650 | tags->rqs[i] = NULL; | |
24d2f903 | 1651 | goto fail; |
a5164405 | 1652 | } |
e9b267d9 CH |
1653 | } |
1654 | ||
320ae51f JA |
1655 | p += rq_size; |
1656 | i++; | |
1657 | } | |
1658 | } | |
24d2f903 | 1659 | return tags; |
320ae51f | 1660 | |
24d2f903 | 1661 | fail: |
24d2f903 CH |
1662 | blk_mq_free_rq_map(set, tags, hctx_idx); |
1663 | return NULL; | |
320ae51f JA |
1664 | } |
1665 | ||
e57690fe JA |
1666 | /* |
1667 | * 'cpu' is going away. splice any existing rq_list entries from this | |
1668 | * software queue to the hw queue dispatch list, and ensure that it | |
1669 | * gets run. | |
1670 | */ | |
9467f859 | 1671 | static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node) |
484b4061 | 1672 | { |
9467f859 | 1673 | struct blk_mq_hw_ctx *hctx; |
484b4061 JA |
1674 | struct blk_mq_ctx *ctx; |
1675 | LIST_HEAD(tmp); | |
1676 | ||
9467f859 | 1677 | hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead); |
e57690fe | 1678 | ctx = __blk_mq_get_ctx(hctx->queue, cpu); |
484b4061 JA |
1679 | |
1680 | spin_lock(&ctx->lock); | |
1681 | if (!list_empty(&ctx->rq_list)) { | |
1682 | list_splice_init(&ctx->rq_list, &tmp); | |
1683 | blk_mq_hctx_clear_pending(hctx, ctx); | |
1684 | } | |
1685 | spin_unlock(&ctx->lock); | |
1686 | ||
1687 | if (list_empty(&tmp)) | |
9467f859 | 1688 | return 0; |
484b4061 | 1689 | |
e57690fe JA |
1690 | spin_lock(&hctx->lock); |
1691 | list_splice_tail_init(&tmp, &hctx->dispatch); | |
1692 | spin_unlock(&hctx->lock); | |
484b4061 JA |
1693 | |
1694 | blk_mq_run_hw_queue(hctx, true); | |
9467f859 | 1695 | return 0; |
484b4061 JA |
1696 | } |
1697 | ||
9467f859 | 1698 | static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx) |
484b4061 | 1699 | { |
9467f859 TG |
1700 | cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD, |
1701 | &hctx->cpuhp_dead); | |
484b4061 JA |
1702 | } |
1703 | ||
c3b4afca | 1704 | /* hctx->ctxs will be freed in queue's release handler */ |
08e98fc6 ML |
1705 | static void blk_mq_exit_hctx(struct request_queue *q, |
1706 | struct blk_mq_tag_set *set, | |
1707 | struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx) | |
1708 | { | |
f70ced09 ML |
1709 | unsigned flush_start_tag = set->queue_depth; |
1710 | ||
08e98fc6 ML |
1711 | blk_mq_tag_idle(hctx); |
1712 | ||
f70ced09 ML |
1713 | if (set->ops->exit_request) |
1714 | set->ops->exit_request(set->driver_data, | |
1715 | hctx->fq->flush_rq, hctx_idx, | |
1716 | flush_start_tag + hctx_idx); | |
1717 | ||
08e98fc6 ML |
1718 | if (set->ops->exit_hctx) |
1719 | set->ops->exit_hctx(hctx, hctx_idx); | |
1720 | ||
6a83e74d BVA |
1721 | if (hctx->flags & BLK_MQ_F_BLOCKING) |
1722 | cleanup_srcu_struct(&hctx->queue_rq_srcu); | |
1723 | ||
9467f859 | 1724 | blk_mq_remove_cpuhp(hctx); |
f70ced09 | 1725 | blk_free_flush_queue(hctx->fq); |
88459642 | 1726 | sbitmap_free(&hctx->ctx_map); |
08e98fc6 ML |
1727 | } |
1728 | ||
624dbe47 ML |
1729 | static void blk_mq_exit_hw_queues(struct request_queue *q, |
1730 | struct blk_mq_tag_set *set, int nr_queue) | |
1731 | { | |
1732 | struct blk_mq_hw_ctx *hctx; | |
1733 | unsigned int i; | |
1734 | ||
1735 | queue_for_each_hw_ctx(q, hctx, i) { | |
1736 | if (i == nr_queue) | |
1737 | break; | |
08e98fc6 | 1738 | blk_mq_exit_hctx(q, set, hctx, i); |
624dbe47 | 1739 | } |
624dbe47 ML |
1740 | } |
1741 | ||
1742 | static void blk_mq_free_hw_queues(struct request_queue *q, | |
1743 | struct blk_mq_tag_set *set) | |
1744 | { | |
1745 | struct blk_mq_hw_ctx *hctx; | |
1746 | unsigned int i; | |
1747 | ||
e09aae7e | 1748 | queue_for_each_hw_ctx(q, hctx, i) |
624dbe47 | 1749 | free_cpumask_var(hctx->cpumask); |
624dbe47 ML |
1750 | } |
1751 | ||
08e98fc6 ML |
1752 | static int blk_mq_init_hctx(struct request_queue *q, |
1753 | struct blk_mq_tag_set *set, | |
1754 | struct blk_mq_hw_ctx *hctx, unsigned hctx_idx) | |
320ae51f | 1755 | { |
08e98fc6 | 1756 | int node; |
f70ced09 | 1757 | unsigned flush_start_tag = set->queue_depth; |
08e98fc6 ML |
1758 | |
1759 | node = hctx->numa_node; | |
1760 | if (node == NUMA_NO_NODE) | |
1761 | node = hctx->numa_node = set->numa_node; | |
1762 | ||
27489a3c | 1763 | INIT_WORK(&hctx->run_work, blk_mq_run_work_fn); |
08e98fc6 ML |
1764 | INIT_DELAYED_WORK(&hctx->delay_work, blk_mq_delay_work_fn); |
1765 | spin_lock_init(&hctx->lock); | |
1766 | INIT_LIST_HEAD(&hctx->dispatch); | |
1767 | hctx->queue = q; | |
1768 | hctx->queue_num = hctx_idx; | |
2404e607 | 1769 | hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED; |
08e98fc6 | 1770 | |
9467f859 | 1771 | cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead); |
08e98fc6 ML |
1772 | |
1773 | hctx->tags = set->tags[hctx_idx]; | |
320ae51f JA |
1774 | |
1775 | /* | |
08e98fc6 ML |
1776 | * Allocate space for all possible cpus to avoid allocation at |
1777 | * runtime | |
320ae51f | 1778 | */ |
08e98fc6 ML |
1779 | hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *), |
1780 | GFP_KERNEL, node); | |
1781 | if (!hctx->ctxs) | |
1782 | goto unregister_cpu_notifier; | |
320ae51f | 1783 | |
88459642 OS |
1784 | if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8), GFP_KERNEL, |
1785 | node)) | |
08e98fc6 | 1786 | goto free_ctxs; |
320ae51f | 1787 | |
08e98fc6 | 1788 | hctx->nr_ctx = 0; |
320ae51f | 1789 | |
08e98fc6 ML |
1790 | if (set->ops->init_hctx && |
1791 | set->ops->init_hctx(hctx, set->driver_data, hctx_idx)) | |
1792 | goto free_bitmap; | |
320ae51f | 1793 | |
f70ced09 ML |
1794 | hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size); |
1795 | if (!hctx->fq) | |
1796 | goto exit_hctx; | |
320ae51f | 1797 | |
f70ced09 ML |
1798 | if (set->ops->init_request && |
1799 | set->ops->init_request(set->driver_data, | |
1800 | hctx->fq->flush_rq, hctx_idx, | |
1801 | flush_start_tag + hctx_idx, node)) | |
1802 | goto free_fq; | |
320ae51f | 1803 | |
6a83e74d BVA |
1804 | if (hctx->flags & BLK_MQ_F_BLOCKING) |
1805 | init_srcu_struct(&hctx->queue_rq_srcu); | |
1806 | ||
08e98fc6 | 1807 | return 0; |
320ae51f | 1808 | |
f70ced09 ML |
1809 | free_fq: |
1810 | kfree(hctx->fq); | |
1811 | exit_hctx: | |
1812 | if (set->ops->exit_hctx) | |
1813 | set->ops->exit_hctx(hctx, hctx_idx); | |
08e98fc6 | 1814 | free_bitmap: |
88459642 | 1815 | sbitmap_free(&hctx->ctx_map); |
08e98fc6 ML |
1816 | free_ctxs: |
1817 | kfree(hctx->ctxs); | |
1818 | unregister_cpu_notifier: | |
9467f859 | 1819 | blk_mq_remove_cpuhp(hctx); |
08e98fc6 ML |
1820 | return -1; |
1821 | } | |
320ae51f | 1822 | |
320ae51f JA |
1823 | static void blk_mq_init_cpu_queues(struct request_queue *q, |
1824 | unsigned int nr_hw_queues) | |
1825 | { | |
1826 | unsigned int i; | |
1827 | ||
1828 | for_each_possible_cpu(i) { | |
1829 | struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i); | |
1830 | struct blk_mq_hw_ctx *hctx; | |
1831 | ||
1832 | memset(__ctx, 0, sizeof(*__ctx)); | |
1833 | __ctx->cpu = i; | |
1834 | spin_lock_init(&__ctx->lock); | |
1835 | INIT_LIST_HEAD(&__ctx->rq_list); | |
1836 | __ctx->queue = q; | |
cf43e6be JA |
1837 | blk_stat_init(&__ctx->stat[BLK_STAT_READ]); |
1838 | blk_stat_init(&__ctx->stat[BLK_STAT_WRITE]); | |
320ae51f JA |
1839 | |
1840 | /* If the cpu isn't online, the cpu is mapped to first hctx */ | |
320ae51f JA |
1841 | if (!cpu_online(i)) |
1842 | continue; | |
1843 | ||
7d7e0f90 | 1844 | hctx = blk_mq_map_queue(q, i); |
e4043dcf | 1845 | |
320ae51f JA |
1846 | /* |
1847 | * Set local node, IFF we have more than one hw queue. If | |
1848 | * not, we remain on the home node of the device | |
1849 | */ | |
1850 | if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE) | |
bffed457 | 1851 | hctx->numa_node = local_memory_node(cpu_to_node(i)); |
320ae51f JA |
1852 | } |
1853 | } | |
1854 | ||
5778322e AM |
1855 | static void blk_mq_map_swqueue(struct request_queue *q, |
1856 | const struct cpumask *online_mask) | |
320ae51f JA |
1857 | { |
1858 | unsigned int i; | |
1859 | struct blk_mq_hw_ctx *hctx; | |
1860 | struct blk_mq_ctx *ctx; | |
2a34c087 | 1861 | struct blk_mq_tag_set *set = q->tag_set; |
320ae51f | 1862 | |
60de074b AM |
1863 | /* |
1864 | * Avoid others reading imcomplete hctx->cpumask through sysfs | |
1865 | */ | |
1866 | mutex_lock(&q->sysfs_lock); | |
1867 | ||
320ae51f | 1868 | queue_for_each_hw_ctx(q, hctx, i) { |
e4043dcf | 1869 | cpumask_clear(hctx->cpumask); |
320ae51f JA |
1870 | hctx->nr_ctx = 0; |
1871 | } | |
1872 | ||
1873 | /* | |
1874 | * Map software to hardware queues | |
1875 | */ | |
897bb0c7 | 1876 | for_each_possible_cpu(i) { |
320ae51f | 1877 | /* If the cpu isn't online, the cpu is mapped to first hctx */ |
5778322e | 1878 | if (!cpumask_test_cpu(i, online_mask)) |
e4043dcf JA |
1879 | continue; |
1880 | ||
897bb0c7 | 1881 | ctx = per_cpu_ptr(q->queue_ctx, i); |
7d7e0f90 | 1882 | hctx = blk_mq_map_queue(q, i); |
868f2f0b | 1883 | |
e4043dcf | 1884 | cpumask_set_cpu(i, hctx->cpumask); |
320ae51f JA |
1885 | ctx->index_hw = hctx->nr_ctx; |
1886 | hctx->ctxs[hctx->nr_ctx++] = ctx; | |
1887 | } | |
506e931f | 1888 | |
60de074b AM |
1889 | mutex_unlock(&q->sysfs_lock); |
1890 | ||
506e931f | 1891 | queue_for_each_hw_ctx(q, hctx, i) { |
484b4061 | 1892 | /* |
a68aafa5 JA |
1893 | * If no software queues are mapped to this hardware queue, |
1894 | * disable it and free the request entries. | |
484b4061 JA |
1895 | */ |
1896 | if (!hctx->nr_ctx) { | |
484b4061 JA |
1897 | if (set->tags[i]) { |
1898 | blk_mq_free_rq_map(set, set->tags[i], i); | |
1899 | set->tags[i] = NULL; | |
484b4061 | 1900 | } |
2a34c087 | 1901 | hctx->tags = NULL; |
484b4061 JA |
1902 | continue; |
1903 | } | |
1904 | ||
2a34c087 ML |
1905 | /* unmapped hw queue can be remapped after CPU topo changed */ |
1906 | if (!set->tags[i]) | |
1907 | set->tags[i] = blk_mq_init_rq_map(set, i); | |
1908 | hctx->tags = set->tags[i]; | |
1909 | WARN_ON(!hctx->tags); | |
1910 | ||
889fa31f CY |
1911 | /* |
1912 | * Set the map size to the number of mapped software queues. | |
1913 | * This is more accurate and more efficient than looping | |
1914 | * over all possibly mapped software queues. | |
1915 | */ | |
88459642 | 1916 | sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx); |
889fa31f | 1917 | |
484b4061 JA |
1918 | /* |
1919 | * Initialize batch roundrobin counts | |
1920 | */ | |
506e931f JA |
1921 | hctx->next_cpu = cpumask_first(hctx->cpumask); |
1922 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; | |
1923 | } | |
320ae51f JA |
1924 | } |
1925 | ||
2404e607 | 1926 | static void queue_set_hctx_shared(struct request_queue *q, bool shared) |
0d2602ca JA |
1927 | { |
1928 | struct blk_mq_hw_ctx *hctx; | |
0d2602ca JA |
1929 | int i; |
1930 | ||
2404e607 JM |
1931 | queue_for_each_hw_ctx(q, hctx, i) { |
1932 | if (shared) | |
1933 | hctx->flags |= BLK_MQ_F_TAG_SHARED; | |
1934 | else | |
1935 | hctx->flags &= ~BLK_MQ_F_TAG_SHARED; | |
1936 | } | |
1937 | } | |
1938 | ||
1939 | static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set, bool shared) | |
1940 | { | |
1941 | struct request_queue *q; | |
0d2602ca JA |
1942 | |
1943 | list_for_each_entry(q, &set->tag_list, tag_set_list) { | |
1944 | blk_mq_freeze_queue(q); | |
2404e607 | 1945 | queue_set_hctx_shared(q, shared); |
0d2602ca JA |
1946 | blk_mq_unfreeze_queue(q); |
1947 | } | |
1948 | } | |
1949 | ||
1950 | static void blk_mq_del_queue_tag_set(struct request_queue *q) | |
1951 | { | |
1952 | struct blk_mq_tag_set *set = q->tag_set; | |
1953 | ||
0d2602ca JA |
1954 | mutex_lock(&set->tag_list_lock); |
1955 | list_del_init(&q->tag_set_list); | |
2404e607 JM |
1956 | if (list_is_singular(&set->tag_list)) { |
1957 | /* just transitioned to unshared */ | |
1958 | set->flags &= ~BLK_MQ_F_TAG_SHARED; | |
1959 | /* update existing queue */ | |
1960 | blk_mq_update_tag_set_depth(set, false); | |
1961 | } | |
0d2602ca | 1962 | mutex_unlock(&set->tag_list_lock); |
0d2602ca JA |
1963 | } |
1964 | ||
1965 | static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set, | |
1966 | struct request_queue *q) | |
1967 | { | |
1968 | q->tag_set = set; | |
1969 | ||
1970 | mutex_lock(&set->tag_list_lock); | |
2404e607 JM |
1971 | |
1972 | /* Check to see if we're transitioning to shared (from 1 to 2 queues). */ | |
1973 | if (!list_empty(&set->tag_list) && !(set->flags & BLK_MQ_F_TAG_SHARED)) { | |
1974 | set->flags |= BLK_MQ_F_TAG_SHARED; | |
1975 | /* update existing queue */ | |
1976 | blk_mq_update_tag_set_depth(set, true); | |
1977 | } | |
1978 | if (set->flags & BLK_MQ_F_TAG_SHARED) | |
1979 | queue_set_hctx_shared(q, true); | |
0d2602ca | 1980 | list_add_tail(&q->tag_set_list, &set->tag_list); |
2404e607 | 1981 | |
0d2602ca JA |
1982 | mutex_unlock(&set->tag_list_lock); |
1983 | } | |
1984 | ||
e09aae7e ML |
1985 | /* |
1986 | * It is the actual release handler for mq, but we do it from | |
1987 | * request queue's release handler for avoiding use-after-free | |
1988 | * and headache because q->mq_kobj shouldn't have been introduced, | |
1989 | * but we can't group ctx/kctx kobj without it. | |
1990 | */ | |
1991 | void blk_mq_release(struct request_queue *q) | |
1992 | { | |
1993 | struct blk_mq_hw_ctx *hctx; | |
1994 | unsigned int i; | |
1995 | ||
1996 | /* hctx kobj stays in hctx */ | |
c3b4afca ML |
1997 | queue_for_each_hw_ctx(q, hctx, i) { |
1998 | if (!hctx) | |
1999 | continue; | |
2000 | kfree(hctx->ctxs); | |
e09aae7e | 2001 | kfree(hctx); |
c3b4afca | 2002 | } |
e09aae7e | 2003 | |
a723bab3 AM |
2004 | q->mq_map = NULL; |
2005 | ||
e09aae7e ML |
2006 | kfree(q->queue_hw_ctx); |
2007 | ||
2008 | /* ctx kobj stays in queue_ctx */ | |
2009 | free_percpu(q->queue_ctx); | |
2010 | } | |
2011 | ||
24d2f903 | 2012 | struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set) |
b62c21b7 MS |
2013 | { |
2014 | struct request_queue *uninit_q, *q; | |
2015 | ||
2016 | uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node); | |
2017 | if (!uninit_q) | |
2018 | return ERR_PTR(-ENOMEM); | |
2019 | ||
2020 | q = blk_mq_init_allocated_queue(set, uninit_q); | |
2021 | if (IS_ERR(q)) | |
2022 | blk_cleanup_queue(uninit_q); | |
2023 | ||
2024 | return q; | |
2025 | } | |
2026 | EXPORT_SYMBOL(blk_mq_init_queue); | |
2027 | ||
868f2f0b KB |
2028 | static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set, |
2029 | struct request_queue *q) | |
320ae51f | 2030 | { |
868f2f0b KB |
2031 | int i, j; |
2032 | struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx; | |
f14bbe77 | 2033 | |
868f2f0b | 2034 | blk_mq_sysfs_unregister(q); |
24d2f903 | 2035 | for (i = 0; i < set->nr_hw_queues; i++) { |
868f2f0b | 2036 | int node; |
f14bbe77 | 2037 | |
868f2f0b KB |
2038 | if (hctxs[i]) |
2039 | continue; | |
2040 | ||
2041 | node = blk_mq_hw_queue_to_node(q->mq_map, i); | |
cdef54dd CH |
2042 | hctxs[i] = kzalloc_node(sizeof(struct blk_mq_hw_ctx), |
2043 | GFP_KERNEL, node); | |
320ae51f | 2044 | if (!hctxs[i]) |
868f2f0b | 2045 | break; |
320ae51f | 2046 | |
a86073e4 | 2047 | if (!zalloc_cpumask_var_node(&hctxs[i]->cpumask, GFP_KERNEL, |
868f2f0b KB |
2048 | node)) { |
2049 | kfree(hctxs[i]); | |
2050 | hctxs[i] = NULL; | |
2051 | break; | |
2052 | } | |
e4043dcf | 2053 | |
0d2602ca | 2054 | atomic_set(&hctxs[i]->nr_active, 0); |
f14bbe77 | 2055 | hctxs[i]->numa_node = node; |
320ae51f | 2056 | hctxs[i]->queue_num = i; |
868f2f0b KB |
2057 | |
2058 | if (blk_mq_init_hctx(q, set, hctxs[i], i)) { | |
2059 | free_cpumask_var(hctxs[i]->cpumask); | |
2060 | kfree(hctxs[i]); | |
2061 | hctxs[i] = NULL; | |
2062 | break; | |
2063 | } | |
2064 | blk_mq_hctx_kobj_init(hctxs[i]); | |
320ae51f | 2065 | } |
868f2f0b KB |
2066 | for (j = i; j < q->nr_hw_queues; j++) { |
2067 | struct blk_mq_hw_ctx *hctx = hctxs[j]; | |
2068 | ||
2069 | if (hctx) { | |
2070 | if (hctx->tags) { | |
2071 | blk_mq_free_rq_map(set, hctx->tags, j); | |
2072 | set->tags[j] = NULL; | |
2073 | } | |
2074 | blk_mq_exit_hctx(q, set, hctx, j); | |
2075 | free_cpumask_var(hctx->cpumask); | |
2076 | kobject_put(&hctx->kobj); | |
2077 | kfree(hctx->ctxs); | |
2078 | kfree(hctx); | |
2079 | hctxs[j] = NULL; | |
2080 | ||
2081 | } | |
2082 | } | |
2083 | q->nr_hw_queues = i; | |
2084 | blk_mq_sysfs_register(q); | |
2085 | } | |
2086 | ||
2087 | struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set, | |
2088 | struct request_queue *q) | |
2089 | { | |
66841672 ML |
2090 | /* mark the queue as mq asap */ |
2091 | q->mq_ops = set->ops; | |
2092 | ||
868f2f0b KB |
2093 | q->queue_ctx = alloc_percpu(struct blk_mq_ctx); |
2094 | if (!q->queue_ctx) | |
c7de5726 | 2095 | goto err_exit; |
868f2f0b KB |
2096 | |
2097 | q->queue_hw_ctx = kzalloc_node(nr_cpu_ids * sizeof(*(q->queue_hw_ctx)), | |
2098 | GFP_KERNEL, set->numa_node); | |
2099 | if (!q->queue_hw_ctx) | |
2100 | goto err_percpu; | |
2101 | ||
bdd17e75 | 2102 | q->mq_map = set->mq_map; |
868f2f0b KB |
2103 | |
2104 | blk_mq_realloc_hw_ctxs(set, q); | |
2105 | if (!q->nr_hw_queues) | |
2106 | goto err_hctxs; | |
320ae51f | 2107 | |
287922eb | 2108 | INIT_WORK(&q->timeout_work, blk_mq_timeout_work); |
e56f698b | 2109 | blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ); |
320ae51f JA |
2110 | |
2111 | q->nr_queues = nr_cpu_ids; | |
320ae51f | 2112 | |
94eddfbe | 2113 | q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT; |
320ae51f | 2114 | |
05f1dd53 JA |
2115 | if (!(set->flags & BLK_MQ_F_SG_MERGE)) |
2116 | q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE; | |
2117 | ||
1be036e9 CH |
2118 | q->sg_reserved_size = INT_MAX; |
2119 | ||
2849450a | 2120 | INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work); |
6fca6a61 CH |
2121 | INIT_LIST_HEAD(&q->requeue_list); |
2122 | spin_lock_init(&q->requeue_lock); | |
2123 | ||
07068d5b JA |
2124 | if (q->nr_hw_queues > 1) |
2125 | blk_queue_make_request(q, blk_mq_make_request); | |
2126 | else | |
2127 | blk_queue_make_request(q, blk_sq_make_request); | |
2128 | ||
eba71768 JA |
2129 | /* |
2130 | * Do this after blk_queue_make_request() overrides it... | |
2131 | */ | |
2132 | q->nr_requests = set->queue_depth; | |
2133 | ||
24d2f903 CH |
2134 | if (set->ops->complete) |
2135 | blk_queue_softirq_done(q, set->ops->complete); | |
30a91cb4 | 2136 | |
24d2f903 | 2137 | blk_mq_init_cpu_queues(q, set->nr_hw_queues); |
320ae51f | 2138 | |
5778322e | 2139 | get_online_cpus(); |
320ae51f | 2140 | mutex_lock(&all_q_mutex); |
320ae51f | 2141 | |
4593fdbe | 2142 | list_add_tail(&q->all_q_node, &all_q_list); |
0d2602ca | 2143 | blk_mq_add_queue_tag_set(set, q); |
5778322e | 2144 | blk_mq_map_swqueue(q, cpu_online_mask); |
484b4061 | 2145 | |
4593fdbe | 2146 | mutex_unlock(&all_q_mutex); |
5778322e | 2147 | put_online_cpus(); |
4593fdbe | 2148 | |
320ae51f | 2149 | return q; |
18741986 | 2150 | |
320ae51f | 2151 | err_hctxs: |
868f2f0b | 2152 | kfree(q->queue_hw_ctx); |
320ae51f | 2153 | err_percpu: |
868f2f0b | 2154 | free_percpu(q->queue_ctx); |
c7de5726 ML |
2155 | err_exit: |
2156 | q->mq_ops = NULL; | |
320ae51f JA |
2157 | return ERR_PTR(-ENOMEM); |
2158 | } | |
b62c21b7 | 2159 | EXPORT_SYMBOL(blk_mq_init_allocated_queue); |
320ae51f JA |
2160 | |
2161 | void blk_mq_free_queue(struct request_queue *q) | |
2162 | { | |
624dbe47 | 2163 | struct blk_mq_tag_set *set = q->tag_set; |
320ae51f | 2164 | |
0e626368 AM |
2165 | mutex_lock(&all_q_mutex); |
2166 | list_del_init(&q->all_q_node); | |
2167 | mutex_unlock(&all_q_mutex); | |
2168 | ||
87760e5e JA |
2169 | wbt_exit(q); |
2170 | ||
0d2602ca JA |
2171 | blk_mq_del_queue_tag_set(q); |
2172 | ||
624dbe47 ML |
2173 | blk_mq_exit_hw_queues(q, set, set->nr_hw_queues); |
2174 | blk_mq_free_hw_queues(q, set); | |
320ae51f | 2175 | } |
320ae51f JA |
2176 | |
2177 | /* Basically redo blk_mq_init_queue with queue frozen */ | |
5778322e AM |
2178 | static void blk_mq_queue_reinit(struct request_queue *q, |
2179 | const struct cpumask *online_mask) | |
320ae51f | 2180 | { |
4ecd4fef | 2181 | WARN_ON_ONCE(!atomic_read(&q->mq_freeze_depth)); |
320ae51f | 2182 | |
67aec14c JA |
2183 | blk_mq_sysfs_unregister(q); |
2184 | ||
320ae51f JA |
2185 | /* |
2186 | * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe | |
2187 | * we should change hctx numa_node according to new topology (this | |
2188 | * involves free and re-allocate memory, worthy doing?) | |
2189 | */ | |
2190 | ||
5778322e | 2191 | blk_mq_map_swqueue(q, online_mask); |
320ae51f | 2192 | |
67aec14c | 2193 | blk_mq_sysfs_register(q); |
320ae51f JA |
2194 | } |
2195 | ||
65d5291e SAS |
2196 | /* |
2197 | * New online cpumask which is going to be set in this hotplug event. | |
2198 | * Declare this cpumasks as global as cpu-hotplug operation is invoked | |
2199 | * one-by-one and dynamically allocating this could result in a failure. | |
2200 | */ | |
2201 | static struct cpumask cpuhp_online_new; | |
2202 | ||
2203 | static void blk_mq_queue_reinit_work(void) | |
320ae51f JA |
2204 | { |
2205 | struct request_queue *q; | |
320ae51f JA |
2206 | |
2207 | mutex_lock(&all_q_mutex); | |
f3af020b TH |
2208 | /* |
2209 | * We need to freeze and reinit all existing queues. Freezing | |
2210 | * involves synchronous wait for an RCU grace period and doing it | |
2211 | * one by one may take a long time. Start freezing all queues in | |
2212 | * one swoop and then wait for the completions so that freezing can | |
2213 | * take place in parallel. | |
2214 | */ | |
2215 | list_for_each_entry(q, &all_q_list, all_q_node) | |
2216 | blk_mq_freeze_queue_start(q); | |
f054b56c | 2217 | list_for_each_entry(q, &all_q_list, all_q_node) { |
f3af020b TH |
2218 | blk_mq_freeze_queue_wait(q); |
2219 | ||
f054b56c ML |
2220 | /* |
2221 | * timeout handler can't touch hw queue during the | |
2222 | * reinitialization | |
2223 | */ | |
2224 | del_timer_sync(&q->timeout); | |
2225 | } | |
2226 | ||
320ae51f | 2227 | list_for_each_entry(q, &all_q_list, all_q_node) |
65d5291e | 2228 | blk_mq_queue_reinit(q, &cpuhp_online_new); |
f3af020b TH |
2229 | |
2230 | list_for_each_entry(q, &all_q_list, all_q_node) | |
2231 | blk_mq_unfreeze_queue(q); | |
2232 | ||
320ae51f | 2233 | mutex_unlock(&all_q_mutex); |
65d5291e SAS |
2234 | } |
2235 | ||
2236 | static int blk_mq_queue_reinit_dead(unsigned int cpu) | |
2237 | { | |
97a32864 | 2238 | cpumask_copy(&cpuhp_online_new, cpu_online_mask); |
65d5291e SAS |
2239 | blk_mq_queue_reinit_work(); |
2240 | return 0; | |
2241 | } | |
2242 | ||
2243 | /* | |
2244 | * Before hotadded cpu starts handling requests, new mappings must be | |
2245 | * established. Otherwise, these requests in hw queue might never be | |
2246 | * dispatched. | |
2247 | * | |
2248 | * For example, there is a single hw queue (hctx) and two CPU queues (ctx0 | |
2249 | * for CPU0, and ctx1 for CPU1). | |
2250 | * | |
2251 | * Now CPU1 is just onlined and a request is inserted into ctx1->rq_list | |
2252 | * and set bit0 in pending bitmap as ctx1->index_hw is still zero. | |
2253 | * | |
2254 | * And then while running hw queue, flush_busy_ctxs() finds bit0 is set in | |
2255 | * pending bitmap and tries to retrieve requests in hctx->ctxs[0]->rq_list. | |
2256 | * But htx->ctxs[0] is a pointer to ctx0, so the request in ctx1->rq_list | |
2257 | * is ignored. | |
2258 | */ | |
2259 | static int blk_mq_queue_reinit_prepare(unsigned int cpu) | |
2260 | { | |
2261 | cpumask_copy(&cpuhp_online_new, cpu_online_mask); | |
2262 | cpumask_set_cpu(cpu, &cpuhp_online_new); | |
2263 | blk_mq_queue_reinit_work(); | |
2264 | return 0; | |
320ae51f JA |
2265 | } |
2266 | ||
a5164405 JA |
2267 | static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set) |
2268 | { | |
2269 | int i; | |
2270 | ||
2271 | for (i = 0; i < set->nr_hw_queues; i++) { | |
2272 | set->tags[i] = blk_mq_init_rq_map(set, i); | |
2273 | if (!set->tags[i]) | |
2274 | goto out_unwind; | |
2275 | } | |
2276 | ||
2277 | return 0; | |
2278 | ||
2279 | out_unwind: | |
2280 | while (--i >= 0) | |
2281 | blk_mq_free_rq_map(set, set->tags[i], i); | |
2282 | ||
a5164405 JA |
2283 | return -ENOMEM; |
2284 | } | |
2285 | ||
2286 | /* | |
2287 | * Allocate the request maps associated with this tag_set. Note that this | |
2288 | * may reduce the depth asked for, if memory is tight. set->queue_depth | |
2289 | * will be updated to reflect the allocated depth. | |
2290 | */ | |
2291 | static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set) | |
2292 | { | |
2293 | unsigned int depth; | |
2294 | int err; | |
2295 | ||
2296 | depth = set->queue_depth; | |
2297 | do { | |
2298 | err = __blk_mq_alloc_rq_maps(set); | |
2299 | if (!err) | |
2300 | break; | |
2301 | ||
2302 | set->queue_depth >>= 1; | |
2303 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) { | |
2304 | err = -ENOMEM; | |
2305 | break; | |
2306 | } | |
2307 | } while (set->queue_depth); | |
2308 | ||
2309 | if (!set->queue_depth || err) { | |
2310 | pr_err("blk-mq: failed to allocate request map\n"); | |
2311 | return -ENOMEM; | |
2312 | } | |
2313 | ||
2314 | if (depth != set->queue_depth) | |
2315 | pr_info("blk-mq: reduced tag depth (%u -> %u)\n", | |
2316 | depth, set->queue_depth); | |
2317 | ||
2318 | return 0; | |
2319 | } | |
2320 | ||
a4391c64 JA |
2321 | /* |
2322 | * Alloc a tag set to be associated with one or more request queues. | |
2323 | * May fail with EINVAL for various error conditions. May adjust the | |
2324 | * requested depth down, if if it too large. In that case, the set | |
2325 | * value will be stored in set->queue_depth. | |
2326 | */ | |
24d2f903 CH |
2327 | int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set) |
2328 | { | |
da695ba2 CH |
2329 | int ret; |
2330 | ||
205fb5f5 BVA |
2331 | BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS); |
2332 | ||
24d2f903 CH |
2333 | if (!set->nr_hw_queues) |
2334 | return -EINVAL; | |
a4391c64 | 2335 | if (!set->queue_depth) |
24d2f903 CH |
2336 | return -EINVAL; |
2337 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) | |
2338 | return -EINVAL; | |
2339 | ||
7d7e0f90 | 2340 | if (!set->ops->queue_rq) |
24d2f903 CH |
2341 | return -EINVAL; |
2342 | ||
a4391c64 JA |
2343 | if (set->queue_depth > BLK_MQ_MAX_DEPTH) { |
2344 | pr_info("blk-mq: reduced tag depth to %u\n", | |
2345 | BLK_MQ_MAX_DEPTH); | |
2346 | set->queue_depth = BLK_MQ_MAX_DEPTH; | |
2347 | } | |
24d2f903 | 2348 | |
6637fadf SL |
2349 | /* |
2350 | * If a crashdump is active, then we are potentially in a very | |
2351 | * memory constrained environment. Limit us to 1 queue and | |
2352 | * 64 tags to prevent using too much memory. | |
2353 | */ | |
2354 | if (is_kdump_kernel()) { | |
2355 | set->nr_hw_queues = 1; | |
2356 | set->queue_depth = min(64U, set->queue_depth); | |
2357 | } | |
868f2f0b KB |
2358 | /* |
2359 | * There is no use for more h/w queues than cpus. | |
2360 | */ | |
2361 | if (set->nr_hw_queues > nr_cpu_ids) | |
2362 | set->nr_hw_queues = nr_cpu_ids; | |
6637fadf | 2363 | |
868f2f0b | 2364 | set->tags = kzalloc_node(nr_cpu_ids * sizeof(struct blk_mq_tags *), |
24d2f903 CH |
2365 | GFP_KERNEL, set->numa_node); |
2366 | if (!set->tags) | |
a5164405 | 2367 | return -ENOMEM; |
24d2f903 | 2368 | |
da695ba2 CH |
2369 | ret = -ENOMEM; |
2370 | set->mq_map = kzalloc_node(sizeof(*set->mq_map) * nr_cpu_ids, | |
2371 | GFP_KERNEL, set->numa_node); | |
bdd17e75 CH |
2372 | if (!set->mq_map) |
2373 | goto out_free_tags; | |
2374 | ||
da695ba2 CH |
2375 | if (set->ops->map_queues) |
2376 | ret = set->ops->map_queues(set); | |
2377 | else | |
2378 | ret = blk_mq_map_queues(set); | |
2379 | if (ret) | |
2380 | goto out_free_mq_map; | |
2381 | ||
2382 | ret = blk_mq_alloc_rq_maps(set); | |
2383 | if (ret) | |
bdd17e75 | 2384 | goto out_free_mq_map; |
24d2f903 | 2385 | |
0d2602ca JA |
2386 | mutex_init(&set->tag_list_lock); |
2387 | INIT_LIST_HEAD(&set->tag_list); | |
2388 | ||
24d2f903 | 2389 | return 0; |
bdd17e75 CH |
2390 | |
2391 | out_free_mq_map: | |
2392 | kfree(set->mq_map); | |
2393 | set->mq_map = NULL; | |
2394 | out_free_tags: | |
5676e7b6 RE |
2395 | kfree(set->tags); |
2396 | set->tags = NULL; | |
da695ba2 | 2397 | return ret; |
24d2f903 CH |
2398 | } |
2399 | EXPORT_SYMBOL(blk_mq_alloc_tag_set); | |
2400 | ||
2401 | void blk_mq_free_tag_set(struct blk_mq_tag_set *set) | |
2402 | { | |
2403 | int i; | |
2404 | ||
868f2f0b | 2405 | for (i = 0; i < nr_cpu_ids; i++) { |
f42d79ab | 2406 | if (set->tags[i]) |
484b4061 JA |
2407 | blk_mq_free_rq_map(set, set->tags[i], i); |
2408 | } | |
2409 | ||
bdd17e75 CH |
2410 | kfree(set->mq_map); |
2411 | set->mq_map = NULL; | |
2412 | ||
981bd189 | 2413 | kfree(set->tags); |
5676e7b6 | 2414 | set->tags = NULL; |
24d2f903 CH |
2415 | } |
2416 | EXPORT_SYMBOL(blk_mq_free_tag_set); | |
2417 | ||
e3a2b3f9 JA |
2418 | int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr) |
2419 | { | |
2420 | struct blk_mq_tag_set *set = q->tag_set; | |
2421 | struct blk_mq_hw_ctx *hctx; | |
2422 | int i, ret; | |
2423 | ||
2424 | if (!set || nr > set->queue_depth) | |
2425 | return -EINVAL; | |
2426 | ||
2427 | ret = 0; | |
2428 | queue_for_each_hw_ctx(q, hctx, i) { | |
e9137d4b KB |
2429 | if (!hctx->tags) |
2430 | continue; | |
e3a2b3f9 JA |
2431 | ret = blk_mq_tag_update_depth(hctx->tags, nr); |
2432 | if (ret) | |
2433 | break; | |
2434 | } | |
2435 | ||
2436 | if (!ret) | |
2437 | q->nr_requests = nr; | |
2438 | ||
2439 | return ret; | |
2440 | } | |
2441 | ||
868f2f0b KB |
2442 | void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues) |
2443 | { | |
2444 | struct request_queue *q; | |
2445 | ||
2446 | if (nr_hw_queues > nr_cpu_ids) | |
2447 | nr_hw_queues = nr_cpu_ids; | |
2448 | if (nr_hw_queues < 1 || nr_hw_queues == set->nr_hw_queues) | |
2449 | return; | |
2450 | ||
2451 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
2452 | blk_mq_freeze_queue(q); | |
2453 | ||
2454 | set->nr_hw_queues = nr_hw_queues; | |
2455 | list_for_each_entry(q, &set->tag_list, tag_set_list) { | |
2456 | blk_mq_realloc_hw_ctxs(set, q); | |
2457 | ||
2458 | if (q->nr_hw_queues > 1) | |
2459 | blk_queue_make_request(q, blk_mq_make_request); | |
2460 | else | |
2461 | blk_queue_make_request(q, blk_sq_make_request); | |
2462 | ||
2463 | blk_mq_queue_reinit(q, cpu_online_mask); | |
2464 | } | |
2465 | ||
2466 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
2467 | blk_mq_unfreeze_queue(q); | |
2468 | } | |
2469 | EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues); | |
2470 | ||
bbd7bb70 JA |
2471 | static bool __blk_mq_poll(struct blk_mq_hw_ctx *hctx, struct request *rq) |
2472 | { | |
2473 | struct request_queue *q = hctx->queue; | |
2474 | long state; | |
2475 | ||
2476 | hctx->poll_considered++; | |
2477 | ||
2478 | state = current->state; | |
2479 | while (!need_resched()) { | |
2480 | int ret; | |
2481 | ||
2482 | hctx->poll_invoked++; | |
2483 | ||
2484 | ret = q->mq_ops->poll(hctx, rq->tag); | |
2485 | if (ret > 0) { | |
2486 | hctx->poll_success++; | |
2487 | set_current_state(TASK_RUNNING); | |
2488 | return true; | |
2489 | } | |
2490 | ||
2491 | if (signal_pending_state(state, current)) | |
2492 | set_current_state(TASK_RUNNING); | |
2493 | ||
2494 | if (current->state == TASK_RUNNING) | |
2495 | return true; | |
2496 | if (ret < 0) | |
2497 | break; | |
2498 | cpu_relax(); | |
2499 | } | |
2500 | ||
2501 | return false; | |
2502 | } | |
2503 | ||
2504 | bool blk_mq_poll(struct request_queue *q, blk_qc_t cookie) | |
2505 | { | |
2506 | struct blk_mq_hw_ctx *hctx; | |
2507 | struct blk_plug *plug; | |
2508 | struct request *rq; | |
2509 | ||
2510 | if (!q->mq_ops || !q->mq_ops->poll || !blk_qc_t_valid(cookie) || | |
2511 | !test_bit(QUEUE_FLAG_POLL, &q->queue_flags)) | |
2512 | return false; | |
2513 | ||
2514 | plug = current->plug; | |
2515 | if (plug) | |
2516 | blk_flush_plug_list(plug, false); | |
2517 | ||
2518 | hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)]; | |
2519 | rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie)); | |
2520 | ||
2521 | return __blk_mq_poll(hctx, rq); | |
2522 | } | |
2523 | EXPORT_SYMBOL_GPL(blk_mq_poll); | |
2524 | ||
676141e4 JA |
2525 | void blk_mq_disable_hotplug(void) |
2526 | { | |
2527 | mutex_lock(&all_q_mutex); | |
2528 | } | |
2529 | ||
2530 | void blk_mq_enable_hotplug(void) | |
2531 | { | |
2532 | mutex_unlock(&all_q_mutex); | |
2533 | } | |
2534 | ||
320ae51f JA |
2535 | static int __init blk_mq_init(void) |
2536 | { | |
9467f859 TG |
2537 | cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL, |
2538 | blk_mq_hctx_notify_dead); | |
320ae51f | 2539 | |
65d5291e SAS |
2540 | cpuhp_setup_state_nocalls(CPUHP_BLK_MQ_PREPARE, "block/mq:prepare", |
2541 | blk_mq_queue_reinit_prepare, | |
2542 | blk_mq_queue_reinit_dead); | |
320ae51f JA |
2543 | return 0; |
2544 | } | |
2545 | subsys_initcall(blk_mq_init); |