]>
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> |
320ae51f JA |
25 | |
26 | #include <trace/events/block.h> | |
27 | ||
28 | #include <linux/blk-mq.h> | |
29 | #include "blk.h" | |
30 | #include "blk-mq.h" | |
31 | #include "blk-mq-tag.h" | |
32 | ||
33 | static DEFINE_MUTEX(all_q_mutex); | |
34 | static LIST_HEAD(all_q_list); | |
35 | ||
36 | static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx); | |
37 | ||
320ae51f JA |
38 | /* |
39 | * Check if any of the ctx's have pending work in this hardware queue | |
40 | */ | |
41 | static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx) | |
42 | { | |
43 | unsigned int i; | |
44 | ||
569fd0ce | 45 | for (i = 0; i < hctx->ctx_map.size; i++) |
1429d7c9 | 46 | if (hctx->ctx_map.map[i].word) |
320ae51f JA |
47 | return true; |
48 | ||
49 | return false; | |
50 | } | |
51 | ||
1429d7c9 JA |
52 | static inline struct blk_align_bitmap *get_bm(struct blk_mq_hw_ctx *hctx, |
53 | struct blk_mq_ctx *ctx) | |
54 | { | |
55 | return &hctx->ctx_map.map[ctx->index_hw / hctx->ctx_map.bits_per_word]; | |
56 | } | |
57 | ||
58 | #define CTX_TO_BIT(hctx, ctx) \ | |
59 | ((ctx)->index_hw & ((hctx)->ctx_map.bits_per_word - 1)) | |
60 | ||
320ae51f JA |
61 | /* |
62 | * Mark this ctx as having pending work in this hardware queue | |
63 | */ | |
64 | static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx, | |
65 | struct blk_mq_ctx *ctx) | |
66 | { | |
1429d7c9 JA |
67 | struct blk_align_bitmap *bm = get_bm(hctx, ctx); |
68 | ||
69 | if (!test_bit(CTX_TO_BIT(hctx, ctx), &bm->word)) | |
70 | set_bit(CTX_TO_BIT(hctx, ctx), &bm->word); | |
71 | } | |
72 | ||
73 | static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx, | |
74 | struct blk_mq_ctx *ctx) | |
75 | { | |
76 | struct blk_align_bitmap *bm = get_bm(hctx, ctx); | |
77 | ||
78 | clear_bit(CTX_TO_BIT(hctx, ctx), &bm->word); | |
320ae51f JA |
79 | } |
80 | ||
b4c6a028 | 81 | void blk_mq_freeze_queue_start(struct request_queue *q) |
43a5e4e2 | 82 | { |
4ecd4fef | 83 | int freeze_depth; |
cddd5d17 | 84 | |
4ecd4fef CH |
85 | freeze_depth = atomic_inc_return(&q->mq_freeze_depth); |
86 | if (freeze_depth == 1) { | |
3ef28e83 | 87 | percpu_ref_kill(&q->q_usage_counter); |
b94ec296 | 88 | blk_mq_run_hw_queues(q, false); |
cddd5d17 | 89 | } |
f3af020b | 90 | } |
b4c6a028 | 91 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_start); |
f3af020b TH |
92 | |
93 | static void blk_mq_freeze_queue_wait(struct request_queue *q) | |
94 | { | |
3ef28e83 | 95 | wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter)); |
43a5e4e2 ML |
96 | } |
97 | ||
f3af020b TH |
98 | /* |
99 | * Guarantee no request is in use, so we can change any data structure of | |
100 | * the queue afterward. | |
101 | */ | |
3ef28e83 | 102 | void blk_freeze_queue(struct request_queue *q) |
f3af020b | 103 | { |
3ef28e83 DW |
104 | /* |
105 | * In the !blk_mq case we are only calling this to kill the | |
106 | * q_usage_counter, otherwise this increases the freeze depth | |
107 | * and waits for it to return to zero. For this reason there is | |
108 | * no blk_unfreeze_queue(), and blk_freeze_queue() is not | |
109 | * exported to drivers as the only user for unfreeze is blk_mq. | |
110 | */ | |
f3af020b TH |
111 | blk_mq_freeze_queue_start(q); |
112 | blk_mq_freeze_queue_wait(q); | |
113 | } | |
3ef28e83 DW |
114 | |
115 | void blk_mq_freeze_queue(struct request_queue *q) | |
116 | { | |
117 | /* | |
118 | * ...just an alias to keep freeze and unfreeze actions balanced | |
119 | * in the blk_mq_* namespace | |
120 | */ | |
121 | blk_freeze_queue(q); | |
122 | } | |
c761d96b | 123 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue); |
f3af020b | 124 | |
b4c6a028 | 125 | void blk_mq_unfreeze_queue(struct request_queue *q) |
320ae51f | 126 | { |
4ecd4fef | 127 | int freeze_depth; |
320ae51f | 128 | |
4ecd4fef CH |
129 | freeze_depth = atomic_dec_return(&q->mq_freeze_depth); |
130 | WARN_ON_ONCE(freeze_depth < 0); | |
131 | if (!freeze_depth) { | |
3ef28e83 | 132 | percpu_ref_reinit(&q->q_usage_counter); |
320ae51f | 133 | wake_up_all(&q->mq_freeze_wq); |
add703fd | 134 | } |
320ae51f | 135 | } |
b4c6a028 | 136 | EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue); |
320ae51f | 137 | |
aed3ea94 JA |
138 | void blk_mq_wake_waiters(struct request_queue *q) |
139 | { | |
140 | struct blk_mq_hw_ctx *hctx; | |
141 | unsigned int i; | |
142 | ||
143 | queue_for_each_hw_ctx(q, hctx, i) | |
144 | if (blk_mq_hw_queue_mapped(hctx)) | |
145 | blk_mq_tag_wakeup_all(hctx->tags, true); | |
3fd5940c KB |
146 | |
147 | /* | |
148 | * If we are called because the queue has now been marked as | |
149 | * dying, we need to ensure that processes currently waiting on | |
150 | * the queue are notified as well. | |
151 | */ | |
152 | wake_up_all(&q->mq_freeze_wq); | |
aed3ea94 JA |
153 | } |
154 | ||
320ae51f JA |
155 | bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx) |
156 | { | |
157 | return blk_mq_has_free_tags(hctx->tags); | |
158 | } | |
159 | EXPORT_SYMBOL(blk_mq_can_queue); | |
160 | ||
94eddfbe JA |
161 | static void blk_mq_rq_ctx_init(struct request_queue *q, struct blk_mq_ctx *ctx, |
162 | struct request *rq, unsigned int rw_flags) | |
320ae51f | 163 | { |
94eddfbe JA |
164 | if (blk_queue_io_stat(q)) |
165 | rw_flags |= REQ_IO_STAT; | |
166 | ||
af76e555 CH |
167 | INIT_LIST_HEAD(&rq->queuelist); |
168 | /* csd/requeue_work/fifo_time is initialized before use */ | |
169 | rq->q = q; | |
320ae51f | 170 | rq->mq_ctx = ctx; |
0d2602ca | 171 | rq->cmd_flags |= rw_flags; |
af76e555 CH |
172 | /* do not touch atomic flags, it needs atomic ops against the timer */ |
173 | rq->cpu = -1; | |
af76e555 CH |
174 | INIT_HLIST_NODE(&rq->hash); |
175 | RB_CLEAR_NODE(&rq->rb_node); | |
af76e555 CH |
176 | rq->rq_disk = NULL; |
177 | rq->part = NULL; | |
3ee32372 | 178 | rq->start_time = jiffies; |
af76e555 CH |
179 | #ifdef CONFIG_BLK_CGROUP |
180 | rq->rl = NULL; | |
0fec08b4 | 181 | set_start_time_ns(rq); |
af76e555 CH |
182 | rq->io_start_time_ns = 0; |
183 | #endif | |
184 | rq->nr_phys_segments = 0; | |
185 | #if defined(CONFIG_BLK_DEV_INTEGRITY) | |
186 | rq->nr_integrity_segments = 0; | |
187 | #endif | |
af76e555 CH |
188 | rq->special = NULL; |
189 | /* tag was already set */ | |
190 | rq->errors = 0; | |
af76e555 | 191 | |
6f4a1626 TB |
192 | rq->cmd = rq->__cmd; |
193 | ||
af76e555 CH |
194 | rq->extra_len = 0; |
195 | rq->sense_len = 0; | |
196 | rq->resid_len = 0; | |
197 | rq->sense = NULL; | |
198 | ||
af76e555 | 199 | INIT_LIST_HEAD(&rq->timeout_list); |
f6be4fb4 JA |
200 | rq->timeout = 0; |
201 | ||
af76e555 CH |
202 | rq->end_io = NULL; |
203 | rq->end_io_data = NULL; | |
204 | rq->next_rq = NULL; | |
205 | ||
320ae51f JA |
206 | ctx->rq_dispatched[rw_is_sync(rw_flags)]++; |
207 | } | |
208 | ||
5dee8577 | 209 | static struct request * |
cb96a42c | 210 | __blk_mq_alloc_request(struct blk_mq_alloc_data *data, int rw) |
5dee8577 CH |
211 | { |
212 | struct request *rq; | |
213 | unsigned int tag; | |
214 | ||
cb96a42c | 215 | tag = blk_mq_get_tag(data); |
5dee8577 | 216 | if (tag != BLK_MQ_TAG_FAIL) { |
cb96a42c | 217 | rq = data->hctx->tags->rqs[tag]; |
5dee8577 | 218 | |
cb96a42c | 219 | if (blk_mq_tag_busy(data->hctx)) { |
5dee8577 | 220 | rq->cmd_flags = REQ_MQ_INFLIGHT; |
cb96a42c | 221 | atomic_inc(&data->hctx->nr_active); |
5dee8577 CH |
222 | } |
223 | ||
224 | rq->tag = tag; | |
cb96a42c | 225 | blk_mq_rq_ctx_init(data->q, data->ctx, rq, rw); |
5dee8577 CH |
226 | return rq; |
227 | } | |
228 | ||
229 | return NULL; | |
230 | } | |
231 | ||
3c92db6a CH |
232 | struct request *blk_mq_alloc_request(struct request_queue *q, int rw, |
233 | unsigned int flags) | |
320ae51f | 234 | { |
d852564f CH |
235 | struct blk_mq_ctx *ctx; |
236 | struct blk_mq_hw_ctx *hctx; | |
320ae51f | 237 | struct request *rq; |
cb96a42c | 238 | struct blk_mq_alloc_data alloc_data; |
a492f075 | 239 | int ret; |
320ae51f | 240 | |
3c92db6a | 241 | ret = blk_queue_enter(q, flags & BLK_MQ_REQ_NOWAIT); |
a492f075 JL |
242 | if (ret) |
243 | return ERR_PTR(ret); | |
320ae51f | 244 | |
d852564f CH |
245 | ctx = blk_mq_get_ctx(q); |
246 | hctx = q->mq_ops->map_queue(q, ctx->cpu); | |
3c92db6a | 247 | blk_mq_set_alloc_data(&alloc_data, q, flags, ctx, hctx); |
d852564f | 248 | |
cb96a42c | 249 | rq = __blk_mq_alloc_request(&alloc_data, rw); |
3c92db6a | 250 | if (!rq && !(flags & BLK_MQ_REQ_NOWAIT)) { |
d852564f CH |
251 | __blk_mq_run_hw_queue(hctx); |
252 | blk_mq_put_ctx(ctx); | |
253 | ||
254 | ctx = blk_mq_get_ctx(q); | |
255 | hctx = q->mq_ops->map_queue(q, ctx->cpu); | |
3c92db6a | 256 | blk_mq_set_alloc_data(&alloc_data, q, flags, ctx, hctx); |
cb96a42c ML |
257 | rq = __blk_mq_alloc_request(&alloc_data, rw); |
258 | ctx = alloc_data.ctx; | |
d852564f CH |
259 | } |
260 | blk_mq_put_ctx(ctx); | |
c76541a9 | 261 | if (!rq) { |
3ef28e83 | 262 | blk_queue_exit(q); |
a492f075 | 263 | return ERR_PTR(-EWOULDBLOCK); |
c76541a9 | 264 | } |
320ae51f JA |
265 | return rq; |
266 | } | |
4bb659b1 | 267 | EXPORT_SYMBOL(blk_mq_alloc_request); |
320ae51f | 268 | |
2b67bc8d ML |
269 | struct request *blk_mq_alloc_request_hctx(struct request_queue *q, int rw, |
270 | unsigned int flags, unsigned int hctx_idx) | |
271 | { | |
272 | struct blk_mq_hw_ctx *hctx; | |
273 | struct blk_mq_ctx *ctx; | |
274 | struct request *rq; | |
275 | struct blk_mq_alloc_data alloc_data; | |
276 | int ret; | |
277 | ||
278 | /* | |
279 | * If the tag allocator sleeps we could get an allocation for a | |
280 | * different hardware context. No need to complicate the low level | |
281 | * allocator for this for the rare use case of a command tied to | |
282 | * a specific queue. | |
283 | */ | |
284 | if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT))) | |
285 | return ERR_PTR(-EINVAL); | |
286 | ||
287 | if (hctx_idx >= q->nr_hw_queues) | |
288 | return ERR_PTR(-EIO); | |
289 | ||
290 | ret = blk_queue_enter(q, true); | |
291 | if (ret) | |
292 | return ERR_PTR(ret); | |
293 | ||
294 | hctx = q->queue_hw_ctx[hctx_idx]; | |
295 | ctx = __blk_mq_get_ctx(q, cpumask_first(hctx->cpumask)); | |
296 | ||
297 | blk_mq_set_alloc_data(&alloc_data, q, flags, ctx, hctx); | |
298 | rq = __blk_mq_alloc_request(&alloc_data, rw); | |
299 | if (!rq) { | |
300 | blk_queue_exit(q); | |
301 | return ERR_PTR(-EWOULDBLOCK); | |
302 | } | |
303 | ||
304 | return rq; | |
305 | } | |
306 | EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx); | |
307 | ||
320ae51f JA |
308 | static void __blk_mq_free_request(struct blk_mq_hw_ctx *hctx, |
309 | struct blk_mq_ctx *ctx, struct request *rq) | |
310 | { | |
311 | const int tag = rq->tag; | |
312 | struct request_queue *q = rq->q; | |
313 | ||
0d2602ca JA |
314 | if (rq->cmd_flags & REQ_MQ_INFLIGHT) |
315 | atomic_dec(&hctx->nr_active); | |
683d0e12 | 316 | rq->cmd_flags = 0; |
0d2602ca | 317 | |
af76e555 | 318 | clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags); |
0d2602ca | 319 | blk_mq_put_tag(hctx, tag, &ctx->last_tag); |
3ef28e83 | 320 | blk_queue_exit(q); |
320ae51f JA |
321 | } |
322 | ||
7c7f2f2b | 323 | void blk_mq_free_hctx_request(struct blk_mq_hw_ctx *hctx, struct request *rq) |
320ae51f JA |
324 | { |
325 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
320ae51f JA |
326 | |
327 | ctx->rq_completed[rq_is_sync(rq)]++; | |
320ae51f | 328 | __blk_mq_free_request(hctx, ctx, rq); |
7c7f2f2b JA |
329 | |
330 | } | |
331 | EXPORT_SYMBOL_GPL(blk_mq_free_hctx_request); | |
332 | ||
333 | void blk_mq_free_request(struct request *rq) | |
334 | { | |
335 | struct blk_mq_hw_ctx *hctx; | |
336 | struct request_queue *q = rq->q; | |
337 | ||
338 | hctx = q->mq_ops->map_queue(q, rq->mq_ctx->cpu); | |
339 | blk_mq_free_hctx_request(hctx, rq); | |
320ae51f | 340 | } |
1a3b595a | 341 | EXPORT_SYMBOL_GPL(blk_mq_free_request); |
320ae51f | 342 | |
c8a446ad | 343 | inline void __blk_mq_end_request(struct request *rq, int error) |
320ae51f | 344 | { |
0d11e6ac ML |
345 | blk_account_io_done(rq); |
346 | ||
91b63639 | 347 | if (rq->end_io) { |
320ae51f | 348 | rq->end_io(rq, error); |
91b63639 CH |
349 | } else { |
350 | if (unlikely(blk_bidi_rq(rq))) | |
351 | blk_mq_free_request(rq->next_rq); | |
320ae51f | 352 | blk_mq_free_request(rq); |
91b63639 | 353 | } |
320ae51f | 354 | } |
c8a446ad | 355 | EXPORT_SYMBOL(__blk_mq_end_request); |
63151a44 | 356 | |
c8a446ad | 357 | void blk_mq_end_request(struct request *rq, int error) |
63151a44 CH |
358 | { |
359 | if (blk_update_request(rq, error, blk_rq_bytes(rq))) | |
360 | BUG(); | |
c8a446ad | 361 | __blk_mq_end_request(rq, error); |
63151a44 | 362 | } |
c8a446ad | 363 | EXPORT_SYMBOL(blk_mq_end_request); |
320ae51f | 364 | |
30a91cb4 | 365 | static void __blk_mq_complete_request_remote(void *data) |
320ae51f | 366 | { |
3d6efbf6 | 367 | struct request *rq = data; |
320ae51f | 368 | |
30a91cb4 | 369 | rq->q->softirq_done_fn(rq); |
320ae51f | 370 | } |
320ae51f | 371 | |
ed851860 | 372 | static void blk_mq_ipi_complete_request(struct request *rq) |
320ae51f JA |
373 | { |
374 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
38535201 | 375 | bool shared = false; |
320ae51f JA |
376 | int cpu; |
377 | ||
38535201 | 378 | if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) { |
30a91cb4 CH |
379 | rq->q->softirq_done_fn(rq); |
380 | return; | |
381 | } | |
320ae51f JA |
382 | |
383 | cpu = get_cpu(); | |
38535201 CH |
384 | if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags)) |
385 | shared = cpus_share_cache(cpu, ctx->cpu); | |
386 | ||
387 | if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) { | |
30a91cb4 | 388 | rq->csd.func = __blk_mq_complete_request_remote; |
3d6efbf6 CH |
389 | rq->csd.info = rq; |
390 | rq->csd.flags = 0; | |
c46fff2a | 391 | smp_call_function_single_async(ctx->cpu, &rq->csd); |
3d6efbf6 | 392 | } else { |
30a91cb4 | 393 | rq->q->softirq_done_fn(rq); |
3d6efbf6 | 394 | } |
320ae51f JA |
395 | put_cpu(); |
396 | } | |
30a91cb4 | 397 | |
1fa8cc52 | 398 | static void __blk_mq_complete_request(struct request *rq) |
ed851860 JA |
399 | { |
400 | struct request_queue *q = rq->q; | |
401 | ||
402 | if (!q->softirq_done_fn) | |
c8a446ad | 403 | blk_mq_end_request(rq, rq->errors); |
ed851860 JA |
404 | else |
405 | blk_mq_ipi_complete_request(rq); | |
406 | } | |
407 | ||
30a91cb4 CH |
408 | /** |
409 | * blk_mq_complete_request - end I/O on a request | |
410 | * @rq: the request being processed | |
411 | * | |
412 | * Description: | |
413 | * Ends all I/O on a request. It does not handle partial completions. | |
414 | * The actual completion happens out-of-order, through a IPI handler. | |
415 | **/ | |
f4829a9b | 416 | void blk_mq_complete_request(struct request *rq, int error) |
30a91cb4 | 417 | { |
95f09684 JA |
418 | struct request_queue *q = rq->q; |
419 | ||
420 | if (unlikely(blk_should_fake_timeout(q))) | |
30a91cb4 | 421 | return; |
f4829a9b CH |
422 | if (!blk_mark_rq_complete(rq)) { |
423 | rq->errors = error; | |
ed851860 | 424 | __blk_mq_complete_request(rq); |
f4829a9b | 425 | } |
30a91cb4 CH |
426 | } |
427 | EXPORT_SYMBOL(blk_mq_complete_request); | |
320ae51f | 428 | |
973c0191 KB |
429 | int blk_mq_request_started(struct request *rq) |
430 | { | |
431 | return test_bit(REQ_ATOM_STARTED, &rq->atomic_flags); | |
432 | } | |
433 | EXPORT_SYMBOL_GPL(blk_mq_request_started); | |
434 | ||
e2490073 | 435 | void blk_mq_start_request(struct request *rq) |
320ae51f JA |
436 | { |
437 | struct request_queue *q = rq->q; | |
438 | ||
439 | trace_block_rq_issue(q, rq); | |
440 | ||
742ee69b | 441 | rq->resid_len = blk_rq_bytes(rq); |
91b63639 CH |
442 | if (unlikely(blk_bidi_rq(rq))) |
443 | rq->next_rq->resid_len = blk_rq_bytes(rq->next_rq); | |
742ee69b | 444 | |
2b8393b4 | 445 | blk_add_timer(rq); |
87ee7b11 | 446 | |
538b7534 JA |
447 | /* |
448 | * Ensure that ->deadline is visible before set the started | |
449 | * flag and clear the completed flag. | |
450 | */ | |
451 | smp_mb__before_atomic(); | |
452 | ||
87ee7b11 JA |
453 | /* |
454 | * Mark us as started and clear complete. Complete might have been | |
455 | * set if requeue raced with timeout, which then marked it as | |
456 | * complete. So be sure to clear complete again when we start | |
457 | * the request, otherwise we'll ignore the completion event. | |
458 | */ | |
4b570521 JA |
459 | if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) |
460 | set_bit(REQ_ATOM_STARTED, &rq->atomic_flags); | |
461 | if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags)) | |
462 | clear_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags); | |
49f5baa5 CH |
463 | |
464 | if (q->dma_drain_size && blk_rq_bytes(rq)) { | |
465 | /* | |
466 | * Make sure space for the drain appears. We know we can do | |
467 | * this because max_hw_segments has been adjusted to be one | |
468 | * fewer than the device can handle. | |
469 | */ | |
470 | rq->nr_phys_segments++; | |
471 | } | |
320ae51f | 472 | } |
e2490073 | 473 | EXPORT_SYMBOL(blk_mq_start_request); |
320ae51f | 474 | |
ed0791b2 | 475 | static void __blk_mq_requeue_request(struct request *rq) |
320ae51f JA |
476 | { |
477 | struct request_queue *q = rq->q; | |
478 | ||
479 | trace_block_rq_requeue(q, rq); | |
49f5baa5 | 480 | |
e2490073 CH |
481 | if (test_and_clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) { |
482 | if (q->dma_drain_size && blk_rq_bytes(rq)) | |
483 | rq->nr_phys_segments--; | |
484 | } | |
320ae51f JA |
485 | } |
486 | ||
ed0791b2 CH |
487 | void blk_mq_requeue_request(struct request *rq) |
488 | { | |
ed0791b2 | 489 | __blk_mq_requeue_request(rq); |
ed0791b2 | 490 | |
ed0791b2 | 491 | BUG_ON(blk_queued_rq(rq)); |
6fca6a61 | 492 | blk_mq_add_to_requeue_list(rq, true); |
ed0791b2 CH |
493 | } |
494 | EXPORT_SYMBOL(blk_mq_requeue_request); | |
495 | ||
6fca6a61 CH |
496 | static void blk_mq_requeue_work(struct work_struct *work) |
497 | { | |
498 | struct request_queue *q = | |
499 | container_of(work, struct request_queue, requeue_work); | |
500 | LIST_HEAD(rq_list); | |
501 | struct request *rq, *next; | |
502 | unsigned long flags; | |
503 | ||
504 | spin_lock_irqsave(&q->requeue_lock, flags); | |
505 | list_splice_init(&q->requeue_list, &rq_list); | |
506 | spin_unlock_irqrestore(&q->requeue_lock, flags); | |
507 | ||
508 | list_for_each_entry_safe(rq, next, &rq_list, queuelist) { | |
509 | if (!(rq->cmd_flags & REQ_SOFTBARRIER)) | |
510 | continue; | |
511 | ||
512 | rq->cmd_flags &= ~REQ_SOFTBARRIER; | |
513 | list_del_init(&rq->queuelist); | |
514 | blk_mq_insert_request(rq, true, false, false); | |
515 | } | |
516 | ||
517 | while (!list_empty(&rq_list)) { | |
518 | rq = list_entry(rq_list.next, struct request, queuelist); | |
519 | list_del_init(&rq->queuelist); | |
520 | blk_mq_insert_request(rq, false, false, false); | |
521 | } | |
522 | ||
8b957415 JA |
523 | /* |
524 | * Use the start variant of queue running here, so that running | |
525 | * the requeue work will kick stopped queues. | |
526 | */ | |
527 | blk_mq_start_hw_queues(q); | |
6fca6a61 CH |
528 | } |
529 | ||
530 | void blk_mq_add_to_requeue_list(struct request *rq, bool at_head) | |
531 | { | |
532 | struct request_queue *q = rq->q; | |
533 | unsigned long flags; | |
534 | ||
535 | /* | |
536 | * We abuse this flag that is otherwise used by the I/O scheduler to | |
537 | * request head insertation from the workqueue. | |
538 | */ | |
539 | BUG_ON(rq->cmd_flags & REQ_SOFTBARRIER); | |
540 | ||
541 | spin_lock_irqsave(&q->requeue_lock, flags); | |
542 | if (at_head) { | |
543 | rq->cmd_flags |= REQ_SOFTBARRIER; | |
544 | list_add(&rq->queuelist, &q->requeue_list); | |
545 | } else { | |
546 | list_add_tail(&rq->queuelist, &q->requeue_list); | |
547 | } | |
548 | spin_unlock_irqrestore(&q->requeue_lock, flags); | |
549 | } | |
550 | EXPORT_SYMBOL(blk_mq_add_to_requeue_list); | |
551 | ||
c68ed59f KB |
552 | void blk_mq_cancel_requeue_work(struct request_queue *q) |
553 | { | |
554 | cancel_work_sync(&q->requeue_work); | |
555 | } | |
556 | EXPORT_SYMBOL_GPL(blk_mq_cancel_requeue_work); | |
557 | ||
6fca6a61 CH |
558 | void blk_mq_kick_requeue_list(struct request_queue *q) |
559 | { | |
560 | kblockd_schedule_work(&q->requeue_work); | |
561 | } | |
562 | EXPORT_SYMBOL(blk_mq_kick_requeue_list); | |
563 | ||
1885b24d JA |
564 | void blk_mq_abort_requeue_list(struct request_queue *q) |
565 | { | |
566 | unsigned long flags; | |
567 | LIST_HEAD(rq_list); | |
568 | ||
569 | spin_lock_irqsave(&q->requeue_lock, flags); | |
570 | list_splice_init(&q->requeue_list, &rq_list); | |
571 | spin_unlock_irqrestore(&q->requeue_lock, flags); | |
572 | ||
573 | while (!list_empty(&rq_list)) { | |
574 | struct request *rq; | |
575 | ||
576 | rq = list_first_entry(&rq_list, struct request, queuelist); | |
577 | list_del_init(&rq->queuelist); | |
578 | rq->errors = -EIO; | |
579 | blk_mq_end_request(rq, rq->errors); | |
580 | } | |
581 | } | |
582 | EXPORT_SYMBOL(blk_mq_abort_requeue_list); | |
583 | ||
0e62f51f JA |
584 | struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag) |
585 | { | |
0048b483 | 586 | return tags->rqs[tag]; |
24d2f903 CH |
587 | } |
588 | EXPORT_SYMBOL(blk_mq_tag_to_rq); | |
589 | ||
320ae51f | 590 | struct blk_mq_timeout_data { |
46f92d42 CH |
591 | unsigned long next; |
592 | unsigned int next_set; | |
320ae51f JA |
593 | }; |
594 | ||
90415837 | 595 | void blk_mq_rq_timed_out(struct request *req, bool reserved) |
320ae51f | 596 | { |
46f92d42 CH |
597 | struct blk_mq_ops *ops = req->q->mq_ops; |
598 | enum blk_eh_timer_return ret = BLK_EH_RESET_TIMER; | |
87ee7b11 JA |
599 | |
600 | /* | |
601 | * We know that complete is set at this point. If STARTED isn't set | |
602 | * anymore, then the request isn't active and the "timeout" should | |
603 | * just be ignored. This can happen due to the bitflag ordering. | |
604 | * Timeout first checks if STARTED is set, and if it is, assumes | |
605 | * the request is active. But if we race with completion, then | |
606 | * we both flags will get cleared. So check here again, and ignore | |
607 | * a timeout event with a request that isn't active. | |
608 | */ | |
46f92d42 CH |
609 | if (!test_bit(REQ_ATOM_STARTED, &req->atomic_flags)) |
610 | return; | |
87ee7b11 | 611 | |
46f92d42 | 612 | if (ops->timeout) |
0152fb6b | 613 | ret = ops->timeout(req, reserved); |
46f92d42 CH |
614 | |
615 | switch (ret) { | |
616 | case BLK_EH_HANDLED: | |
617 | __blk_mq_complete_request(req); | |
618 | break; | |
619 | case BLK_EH_RESET_TIMER: | |
620 | blk_add_timer(req); | |
621 | blk_clear_rq_complete(req); | |
622 | break; | |
623 | case BLK_EH_NOT_HANDLED: | |
624 | break; | |
625 | default: | |
626 | printk(KERN_ERR "block: bad eh return: %d\n", ret); | |
627 | break; | |
628 | } | |
87ee7b11 | 629 | } |
5b3f25fc | 630 | |
81481eb4 CH |
631 | static void blk_mq_check_expired(struct blk_mq_hw_ctx *hctx, |
632 | struct request *rq, void *priv, bool reserved) | |
633 | { | |
634 | struct blk_mq_timeout_data *data = priv; | |
87ee7b11 | 635 | |
eb130dbf KB |
636 | if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) { |
637 | /* | |
638 | * If a request wasn't started before the queue was | |
639 | * marked dying, kill it here or it'll go unnoticed. | |
640 | */ | |
c0772d51 KB |
641 | if (unlikely(blk_queue_dying(rq->q))) { |
642 | rq->errors = -EIO; | |
643 | blk_mq_end_request(rq, rq->errors); | |
644 | } | |
46f92d42 | 645 | return; |
eb130dbf | 646 | } |
5b3f25fc KB |
647 | if (rq->cmd_flags & REQ_NO_TIMEOUT) |
648 | return; | |
87ee7b11 | 649 | |
46f92d42 CH |
650 | if (time_after_eq(jiffies, rq->deadline)) { |
651 | if (!blk_mark_rq_complete(rq)) | |
0152fb6b | 652 | blk_mq_rq_timed_out(rq, reserved); |
46f92d42 CH |
653 | } else if (!data->next_set || time_after(data->next, rq->deadline)) { |
654 | data->next = rq->deadline; | |
655 | data->next_set = 1; | |
656 | } | |
87ee7b11 JA |
657 | } |
658 | ||
fc3bda55 | 659 | static void blk_mq_timeout_work(struct work_struct *work) |
320ae51f | 660 | { |
fc3bda55 CH |
661 | struct request_queue *q = |
662 | container_of(work, struct request_queue, timeout_work); | |
81481eb4 CH |
663 | struct blk_mq_timeout_data data = { |
664 | .next = 0, | |
665 | .next_set = 0, | |
666 | }; | |
81481eb4 | 667 | int i; |
320ae51f | 668 | |
702b11d1 GKB |
669 | /* A deadlock might occur if a request is stuck requiring a |
670 | * timeout at the same time a queue freeze is waiting | |
671 | * completion, since the timeout code would not be able to | |
672 | * acquire the queue reference here. | |
673 | * | |
674 | * That's why we don't use blk_queue_enter here; instead, we use | |
675 | * percpu_ref_tryget directly, because we need to be able to | |
676 | * obtain a reference even in the short window between the queue | |
677 | * starting to freeze, by dropping the first reference in | |
678 | * blk_mq_freeze_queue_start, and the moment the last request is | |
679 | * consumed, marked by the instant q_usage_counter reaches | |
680 | * zero. | |
681 | */ | |
682 | if (!percpu_ref_tryget(&q->q_usage_counter)) | |
fc3bda55 CH |
683 | return; |
684 | ||
0bf6cd5b | 685 | blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &data); |
320ae51f | 686 | |
81481eb4 CH |
687 | if (data.next_set) { |
688 | data.next = blk_rq_timeout(round_jiffies_up(data.next)); | |
689 | mod_timer(&q->timeout, data.next); | |
0d2602ca | 690 | } else { |
0bf6cd5b CH |
691 | struct blk_mq_hw_ctx *hctx; |
692 | ||
f054b56c ML |
693 | queue_for_each_hw_ctx(q, hctx, i) { |
694 | /* the hctx may be unmapped, so check it here */ | |
695 | if (blk_mq_hw_queue_mapped(hctx)) | |
696 | blk_mq_tag_idle(hctx); | |
697 | } | |
0d2602ca | 698 | } |
fc3bda55 | 699 | blk_queue_exit(q); |
320ae51f JA |
700 | } |
701 | ||
702 | /* | |
703 | * Reverse check our software queue for entries that we could potentially | |
704 | * merge with. Currently includes a hand-wavy stop count of 8, to not spend | |
705 | * too much time checking for merges. | |
706 | */ | |
707 | static bool blk_mq_attempt_merge(struct request_queue *q, | |
708 | struct blk_mq_ctx *ctx, struct bio *bio) | |
709 | { | |
710 | struct request *rq; | |
711 | int checked = 8; | |
712 | ||
713 | list_for_each_entry_reverse(rq, &ctx->rq_list, queuelist) { | |
714 | int el_ret; | |
715 | ||
716 | if (!checked--) | |
717 | break; | |
718 | ||
719 | if (!blk_rq_merge_ok(rq, bio)) | |
720 | continue; | |
721 | ||
722 | el_ret = blk_try_merge(rq, bio); | |
723 | if (el_ret == ELEVATOR_BACK_MERGE) { | |
724 | if (bio_attempt_back_merge(q, rq, bio)) { | |
725 | ctx->rq_merged++; | |
726 | return true; | |
727 | } | |
728 | break; | |
729 | } else if (el_ret == ELEVATOR_FRONT_MERGE) { | |
730 | if (bio_attempt_front_merge(q, rq, bio)) { | |
731 | ctx->rq_merged++; | |
732 | return true; | |
733 | } | |
734 | break; | |
735 | } | |
736 | } | |
737 | ||
738 | return false; | |
739 | } | |
740 | ||
1429d7c9 JA |
741 | /* |
742 | * Process software queues that have been marked busy, splicing them | |
743 | * to the for-dispatch | |
744 | */ | |
745 | static void flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list) | |
746 | { | |
747 | struct blk_mq_ctx *ctx; | |
748 | int i; | |
749 | ||
569fd0ce | 750 | for (i = 0; i < hctx->ctx_map.size; i++) { |
1429d7c9 JA |
751 | struct blk_align_bitmap *bm = &hctx->ctx_map.map[i]; |
752 | unsigned int off, bit; | |
753 | ||
754 | if (!bm->word) | |
755 | continue; | |
756 | ||
757 | bit = 0; | |
758 | off = i * hctx->ctx_map.bits_per_word; | |
759 | do { | |
760 | bit = find_next_bit(&bm->word, bm->depth, bit); | |
761 | if (bit >= bm->depth) | |
762 | break; | |
763 | ||
764 | ctx = hctx->ctxs[bit + off]; | |
765 | clear_bit(bit, &bm->word); | |
766 | spin_lock(&ctx->lock); | |
767 | list_splice_tail_init(&ctx->rq_list, list); | |
768 | spin_unlock(&ctx->lock); | |
769 | ||
770 | bit++; | |
771 | } while (1); | |
772 | } | |
773 | } | |
774 | ||
320ae51f JA |
775 | /* |
776 | * Run this hardware queue, pulling any software queues mapped to it in. | |
777 | * Note that this function currently has various problems around ordering | |
778 | * of IO. In particular, we'd like FIFO behaviour on handling existing | |
779 | * items on the hctx->dispatch list. Ignore that for now. | |
780 | */ | |
781 | static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx) | |
782 | { | |
783 | struct request_queue *q = hctx->queue; | |
320ae51f JA |
784 | struct request *rq; |
785 | LIST_HEAD(rq_list); | |
74c45052 JA |
786 | LIST_HEAD(driver_list); |
787 | struct list_head *dptr; | |
1429d7c9 | 788 | int queued; |
320ae51f | 789 | |
5d12f905 | 790 | if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state))) |
320ae51f JA |
791 | return; |
792 | ||
8b6f6135 JA |
793 | WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) && |
794 | cpu_online(hctx->next_cpu)); | |
795 | ||
320ae51f JA |
796 | hctx->run++; |
797 | ||
798 | /* | |
799 | * Touch any software queue that has pending entries. | |
800 | */ | |
1429d7c9 | 801 | flush_busy_ctxs(hctx, &rq_list); |
320ae51f JA |
802 | |
803 | /* | |
804 | * If we have previous entries on our dispatch list, grab them | |
805 | * and stuff them at the front for more fair dispatch. | |
806 | */ | |
807 | if (!list_empty_careful(&hctx->dispatch)) { | |
808 | spin_lock(&hctx->lock); | |
809 | if (!list_empty(&hctx->dispatch)) | |
810 | list_splice_init(&hctx->dispatch, &rq_list); | |
811 | spin_unlock(&hctx->lock); | |
812 | } | |
813 | ||
74c45052 JA |
814 | /* |
815 | * Start off with dptr being NULL, so we start the first request | |
816 | * immediately, even if we have more pending. | |
817 | */ | |
818 | dptr = NULL; | |
819 | ||
320ae51f JA |
820 | /* |
821 | * Now process all the entries, sending them to the driver. | |
822 | */ | |
1429d7c9 | 823 | queued = 0; |
320ae51f | 824 | while (!list_empty(&rq_list)) { |
74c45052 | 825 | struct blk_mq_queue_data bd; |
320ae51f JA |
826 | int ret; |
827 | ||
828 | rq = list_first_entry(&rq_list, struct request, queuelist); | |
829 | list_del_init(&rq->queuelist); | |
320ae51f | 830 | |
74c45052 JA |
831 | bd.rq = rq; |
832 | bd.list = dptr; | |
833 | bd.last = list_empty(&rq_list); | |
834 | ||
835 | ret = q->mq_ops->queue_rq(hctx, &bd); | |
320ae51f JA |
836 | switch (ret) { |
837 | case BLK_MQ_RQ_QUEUE_OK: | |
838 | queued++; | |
bbde8de2 | 839 | break; |
320ae51f | 840 | case BLK_MQ_RQ_QUEUE_BUSY: |
320ae51f | 841 | list_add(&rq->queuelist, &rq_list); |
ed0791b2 | 842 | __blk_mq_requeue_request(rq); |
320ae51f JA |
843 | break; |
844 | default: | |
845 | pr_err("blk-mq: bad return on queue: %d\n", ret); | |
320ae51f | 846 | case BLK_MQ_RQ_QUEUE_ERROR: |
1e93b8c2 | 847 | rq->errors = -EIO; |
c8a446ad | 848 | blk_mq_end_request(rq, rq->errors); |
320ae51f JA |
849 | break; |
850 | } | |
851 | ||
852 | if (ret == BLK_MQ_RQ_QUEUE_BUSY) | |
853 | break; | |
74c45052 JA |
854 | |
855 | /* | |
856 | * We've done the first request. If we have more than 1 | |
857 | * left in the list, set dptr to defer issue. | |
858 | */ | |
859 | if (!dptr && rq_list.next != rq_list.prev) | |
860 | dptr = &driver_list; | |
320ae51f JA |
861 | } |
862 | ||
863 | if (!queued) | |
864 | hctx->dispatched[0]++; | |
865 | else if (queued < (1 << (BLK_MQ_MAX_DISPATCH_ORDER - 1))) | |
866 | hctx->dispatched[ilog2(queued) + 1]++; | |
867 | ||
868 | /* | |
869 | * Any items that need requeuing? Stuff them into hctx->dispatch, | |
870 | * that is where we will continue on next queue run. | |
871 | */ | |
872 | if (!list_empty(&rq_list)) { | |
873 | spin_lock(&hctx->lock); | |
874 | list_splice(&rq_list, &hctx->dispatch); | |
875 | spin_unlock(&hctx->lock); | |
9ba52e58 SL |
876 | /* |
877 | * the queue is expected stopped with BLK_MQ_RQ_QUEUE_BUSY, but | |
878 | * it's possible the queue is stopped and restarted again | |
879 | * before this. Queue restart will dispatch requests. And since | |
880 | * requests in rq_list aren't added into hctx->dispatch yet, | |
881 | * the requests in rq_list might get lost. | |
882 | * | |
883 | * blk_mq_run_hw_queue() already checks the STOPPED bit | |
884 | **/ | |
885 | blk_mq_run_hw_queue(hctx, true); | |
320ae51f JA |
886 | } |
887 | } | |
888 | ||
506e931f JA |
889 | /* |
890 | * It'd be great if the workqueue API had a way to pass | |
891 | * in a mask and had some smarts for more clever placement. | |
892 | * For now we just round-robin here, switching for every | |
893 | * BLK_MQ_CPU_WORK_BATCH queued items. | |
894 | */ | |
895 | static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx) | |
896 | { | |
b657d7e6 CH |
897 | if (hctx->queue->nr_hw_queues == 1) |
898 | return WORK_CPU_UNBOUND; | |
506e931f JA |
899 | |
900 | if (--hctx->next_cpu_batch <= 0) { | |
5273bc28 | 901 | int next_cpu; |
506e931f JA |
902 | |
903 | next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask); | |
904 | if (next_cpu >= nr_cpu_ids) | |
905 | next_cpu = cpumask_first(hctx->cpumask); | |
906 | ||
907 | hctx->next_cpu = next_cpu; | |
908 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; | |
909 | } | |
910 | ||
b657d7e6 | 911 | return hctx->next_cpu; |
506e931f JA |
912 | } |
913 | ||
320ae51f JA |
914 | void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) |
915 | { | |
19c66e59 ML |
916 | if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state) || |
917 | !blk_mq_hw_queue_mapped(hctx))) | |
320ae51f JA |
918 | return; |
919 | ||
398205b8 | 920 | if (!async) { |
2a90d4aa PB |
921 | int cpu = get_cpu(); |
922 | if (cpumask_test_cpu(cpu, hctx->cpumask)) { | |
398205b8 | 923 | __blk_mq_run_hw_queue(hctx); |
2a90d4aa | 924 | put_cpu(); |
398205b8 PB |
925 | return; |
926 | } | |
e4043dcf | 927 | |
2a90d4aa | 928 | put_cpu(); |
e4043dcf | 929 | } |
398205b8 | 930 | |
b657d7e6 CH |
931 | kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx), |
932 | &hctx->run_work, 0); | |
320ae51f JA |
933 | } |
934 | ||
b94ec296 | 935 | void blk_mq_run_hw_queues(struct request_queue *q, bool async) |
320ae51f JA |
936 | { |
937 | struct blk_mq_hw_ctx *hctx; | |
938 | int i; | |
939 | ||
940 | queue_for_each_hw_ctx(q, hctx, i) { | |
941 | if ((!blk_mq_hctx_has_pending(hctx) && | |
942 | list_empty_careful(&hctx->dispatch)) || | |
5d12f905 | 943 | test_bit(BLK_MQ_S_STOPPED, &hctx->state)) |
320ae51f JA |
944 | continue; |
945 | ||
b94ec296 | 946 | blk_mq_run_hw_queue(hctx, async); |
320ae51f JA |
947 | } |
948 | } | |
b94ec296 | 949 | EXPORT_SYMBOL(blk_mq_run_hw_queues); |
320ae51f JA |
950 | |
951 | void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx) | |
952 | { | |
70f4db63 CH |
953 | cancel_delayed_work(&hctx->run_work); |
954 | cancel_delayed_work(&hctx->delay_work); | |
320ae51f JA |
955 | set_bit(BLK_MQ_S_STOPPED, &hctx->state); |
956 | } | |
957 | EXPORT_SYMBOL(blk_mq_stop_hw_queue); | |
958 | ||
280d45f6 CH |
959 | void blk_mq_stop_hw_queues(struct request_queue *q) |
960 | { | |
961 | struct blk_mq_hw_ctx *hctx; | |
962 | int i; | |
963 | ||
964 | queue_for_each_hw_ctx(q, hctx, i) | |
965 | blk_mq_stop_hw_queue(hctx); | |
966 | } | |
967 | EXPORT_SYMBOL(blk_mq_stop_hw_queues); | |
968 | ||
320ae51f JA |
969 | void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx) |
970 | { | |
971 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
e4043dcf | 972 | |
0ffbce80 | 973 | blk_mq_run_hw_queue(hctx, false); |
320ae51f JA |
974 | } |
975 | EXPORT_SYMBOL(blk_mq_start_hw_queue); | |
976 | ||
2f268556 CH |
977 | void blk_mq_start_hw_queues(struct request_queue *q) |
978 | { | |
979 | struct blk_mq_hw_ctx *hctx; | |
980 | int i; | |
981 | ||
982 | queue_for_each_hw_ctx(q, hctx, i) | |
983 | blk_mq_start_hw_queue(hctx); | |
984 | } | |
985 | EXPORT_SYMBOL(blk_mq_start_hw_queues); | |
986 | ||
1b4a3258 | 987 | void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async) |
320ae51f JA |
988 | { |
989 | struct blk_mq_hw_ctx *hctx; | |
990 | int i; | |
991 | ||
992 | queue_for_each_hw_ctx(q, hctx, i) { | |
993 | if (!test_bit(BLK_MQ_S_STOPPED, &hctx->state)) | |
994 | continue; | |
995 | ||
996 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
1b4a3258 | 997 | blk_mq_run_hw_queue(hctx, async); |
320ae51f JA |
998 | } |
999 | } | |
1000 | EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues); | |
1001 | ||
70f4db63 | 1002 | static void blk_mq_run_work_fn(struct work_struct *work) |
320ae51f JA |
1003 | { |
1004 | struct blk_mq_hw_ctx *hctx; | |
1005 | ||
70f4db63 | 1006 | hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work); |
e4043dcf | 1007 | |
320ae51f JA |
1008 | __blk_mq_run_hw_queue(hctx); |
1009 | } | |
1010 | ||
70f4db63 CH |
1011 | static void blk_mq_delay_work_fn(struct work_struct *work) |
1012 | { | |
1013 | struct blk_mq_hw_ctx *hctx; | |
1014 | ||
1015 | hctx = container_of(work, struct blk_mq_hw_ctx, delay_work.work); | |
1016 | ||
1017 | if (test_and_clear_bit(BLK_MQ_S_STOPPED, &hctx->state)) | |
1018 | __blk_mq_run_hw_queue(hctx); | |
1019 | } | |
1020 | ||
1021 | void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs) | |
1022 | { | |
19c66e59 ML |
1023 | if (unlikely(!blk_mq_hw_queue_mapped(hctx))) |
1024 | return; | |
70f4db63 | 1025 | |
b657d7e6 CH |
1026 | kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx), |
1027 | &hctx->delay_work, msecs_to_jiffies(msecs)); | |
70f4db63 CH |
1028 | } |
1029 | EXPORT_SYMBOL(blk_mq_delay_queue); | |
1030 | ||
cfd0c552 | 1031 | static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx, |
cfd0c552 ML |
1032 | struct request *rq, |
1033 | bool at_head) | |
320ae51f | 1034 | { |
f463371e JA |
1035 | struct blk_mq_ctx *ctx = rq->mq_ctx; |
1036 | ||
01b983c9 JA |
1037 | trace_block_rq_insert(hctx->queue, rq); |
1038 | ||
72a0a36e CH |
1039 | if (at_head) |
1040 | list_add(&rq->queuelist, &ctx->rq_list); | |
1041 | else | |
1042 | list_add_tail(&rq->queuelist, &ctx->rq_list); | |
cfd0c552 | 1043 | } |
4bb659b1 | 1044 | |
cfd0c552 ML |
1045 | static void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, |
1046 | struct request *rq, bool at_head) | |
1047 | { | |
1048 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
1049 | ||
f463371e | 1050 | __blk_mq_insert_req_list(hctx, rq, at_head); |
320ae51f | 1051 | blk_mq_hctx_mark_pending(hctx, ctx); |
320ae51f JA |
1052 | } |
1053 | ||
eeabc850 | 1054 | void blk_mq_insert_request(struct request *rq, bool at_head, bool run_queue, |
f463371e | 1055 | bool async) |
320ae51f | 1056 | { |
f463371e | 1057 | struct blk_mq_ctx *ctx = rq->mq_ctx; |
eeabc850 | 1058 | struct request_queue *q = rq->q; |
320ae51f | 1059 | struct blk_mq_hw_ctx *hctx; |
320ae51f | 1060 | |
320ae51f JA |
1061 | hctx = q->mq_ops->map_queue(q, ctx->cpu); |
1062 | ||
a57a178a CH |
1063 | spin_lock(&ctx->lock); |
1064 | __blk_mq_insert_request(hctx, rq, at_head); | |
1065 | spin_unlock(&ctx->lock); | |
320ae51f | 1066 | |
320ae51f JA |
1067 | if (run_queue) |
1068 | blk_mq_run_hw_queue(hctx, async); | |
1069 | } | |
1070 | ||
1071 | static void blk_mq_insert_requests(struct request_queue *q, | |
1072 | struct blk_mq_ctx *ctx, | |
1073 | struct list_head *list, | |
1074 | int depth, | |
1075 | bool from_schedule) | |
1076 | ||
1077 | { | |
1078 | struct blk_mq_hw_ctx *hctx; | |
320ae51f JA |
1079 | |
1080 | trace_block_unplug(q, depth, !from_schedule); | |
1081 | ||
320ae51f JA |
1082 | hctx = q->mq_ops->map_queue(q, ctx->cpu); |
1083 | ||
1084 | /* | |
1085 | * preemption doesn't flush plug list, so it's possible ctx->cpu is | |
1086 | * offline now | |
1087 | */ | |
1088 | spin_lock(&ctx->lock); | |
1089 | while (!list_empty(list)) { | |
1090 | struct request *rq; | |
1091 | ||
1092 | rq = list_first_entry(list, struct request, queuelist); | |
f463371e | 1093 | BUG_ON(rq->mq_ctx != ctx); |
320ae51f | 1094 | list_del_init(&rq->queuelist); |
f463371e | 1095 | __blk_mq_insert_req_list(hctx, rq, false); |
320ae51f | 1096 | } |
cfd0c552 | 1097 | blk_mq_hctx_mark_pending(hctx, ctx); |
320ae51f JA |
1098 | spin_unlock(&ctx->lock); |
1099 | ||
320ae51f JA |
1100 | blk_mq_run_hw_queue(hctx, from_schedule); |
1101 | } | |
1102 | ||
1103 | static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b) | |
1104 | { | |
1105 | struct request *rqa = container_of(a, struct request, queuelist); | |
1106 | struct request *rqb = container_of(b, struct request, queuelist); | |
1107 | ||
1108 | return !(rqa->mq_ctx < rqb->mq_ctx || | |
1109 | (rqa->mq_ctx == rqb->mq_ctx && | |
1110 | blk_rq_pos(rqa) < blk_rq_pos(rqb))); | |
1111 | } | |
1112 | ||
1113 | void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule) | |
1114 | { | |
1115 | struct blk_mq_ctx *this_ctx; | |
1116 | struct request_queue *this_q; | |
1117 | struct request *rq; | |
1118 | LIST_HEAD(list); | |
1119 | LIST_HEAD(ctx_list); | |
1120 | unsigned int depth; | |
1121 | ||
1122 | list_splice_init(&plug->mq_list, &list); | |
1123 | ||
1124 | list_sort(NULL, &list, plug_ctx_cmp); | |
1125 | ||
1126 | this_q = NULL; | |
1127 | this_ctx = NULL; | |
1128 | depth = 0; | |
1129 | ||
1130 | while (!list_empty(&list)) { | |
1131 | rq = list_entry_rq(list.next); | |
1132 | list_del_init(&rq->queuelist); | |
1133 | BUG_ON(!rq->q); | |
1134 | if (rq->mq_ctx != this_ctx) { | |
1135 | if (this_ctx) { | |
1136 | blk_mq_insert_requests(this_q, this_ctx, | |
1137 | &ctx_list, depth, | |
1138 | from_schedule); | |
1139 | } | |
1140 | ||
1141 | this_ctx = rq->mq_ctx; | |
1142 | this_q = rq->q; | |
1143 | depth = 0; | |
1144 | } | |
1145 | ||
1146 | depth++; | |
1147 | list_add_tail(&rq->queuelist, &ctx_list); | |
1148 | } | |
1149 | ||
1150 | /* | |
1151 | * If 'this_ctx' is set, we know we have entries to complete | |
1152 | * on 'ctx_list'. Do those. | |
1153 | */ | |
1154 | if (this_ctx) { | |
1155 | blk_mq_insert_requests(this_q, this_ctx, &ctx_list, depth, | |
1156 | from_schedule); | |
1157 | } | |
1158 | } | |
1159 | ||
1160 | static void blk_mq_bio_to_request(struct request *rq, struct bio *bio) | |
1161 | { | |
1162 | init_request_from_bio(rq, bio); | |
4b570521 | 1163 | |
3ee32372 | 1164 | if (blk_do_io_stat(rq)) |
4b570521 | 1165 | blk_account_io_start(rq, 1); |
320ae51f JA |
1166 | } |
1167 | ||
274a5843 JA |
1168 | static inline bool hctx_allow_merges(struct blk_mq_hw_ctx *hctx) |
1169 | { | |
1170 | return (hctx->flags & BLK_MQ_F_SHOULD_MERGE) && | |
1171 | !blk_queue_nomerges(hctx->queue); | |
1172 | } | |
1173 | ||
07068d5b JA |
1174 | static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx *hctx, |
1175 | struct blk_mq_ctx *ctx, | |
1176 | struct request *rq, struct bio *bio) | |
320ae51f | 1177 | { |
e18378a6 | 1178 | if (!hctx_allow_merges(hctx) || !bio_mergeable(bio)) { |
07068d5b JA |
1179 | blk_mq_bio_to_request(rq, bio); |
1180 | spin_lock(&ctx->lock); | |
1181 | insert_rq: | |
1182 | __blk_mq_insert_request(hctx, rq, false); | |
1183 | spin_unlock(&ctx->lock); | |
1184 | return false; | |
1185 | } else { | |
274a5843 JA |
1186 | struct request_queue *q = hctx->queue; |
1187 | ||
07068d5b JA |
1188 | spin_lock(&ctx->lock); |
1189 | if (!blk_mq_attempt_merge(q, ctx, bio)) { | |
1190 | blk_mq_bio_to_request(rq, bio); | |
1191 | goto insert_rq; | |
1192 | } | |
320ae51f | 1193 | |
07068d5b JA |
1194 | spin_unlock(&ctx->lock); |
1195 | __blk_mq_free_request(hctx, ctx, rq); | |
1196 | return true; | |
14ec77f3 | 1197 | } |
07068d5b | 1198 | } |
14ec77f3 | 1199 | |
07068d5b JA |
1200 | struct blk_map_ctx { |
1201 | struct blk_mq_hw_ctx *hctx; | |
1202 | struct blk_mq_ctx *ctx; | |
1203 | }; | |
1204 | ||
1205 | static struct request *blk_mq_map_request(struct request_queue *q, | |
1206 | struct bio *bio, | |
1207 | struct blk_map_ctx *data) | |
1208 | { | |
1209 | struct blk_mq_hw_ctx *hctx; | |
1210 | struct blk_mq_ctx *ctx; | |
1211 | struct request *rq; | |
1212 | int rw = bio_data_dir(bio); | |
cb96a42c | 1213 | struct blk_mq_alloc_data alloc_data; |
320ae51f | 1214 | |
3ef28e83 | 1215 | blk_queue_enter_live(q); |
320ae51f JA |
1216 | ctx = blk_mq_get_ctx(q); |
1217 | hctx = q->mq_ops->map_queue(q, ctx->cpu); | |
1218 | ||
07068d5b | 1219 | if (rw_is_sync(bio->bi_rw)) |
27fbf4e8 | 1220 | rw |= REQ_SYNC; |
07068d5b | 1221 | |
320ae51f | 1222 | trace_block_getrq(q, bio, rw); |
3c92db6a | 1223 | blk_mq_set_alloc_data(&alloc_data, q, BLK_MQ_REQ_NOWAIT, ctx, hctx); |
cb96a42c | 1224 | rq = __blk_mq_alloc_request(&alloc_data, rw); |
5dee8577 | 1225 | if (unlikely(!rq)) { |
793597a6 | 1226 | __blk_mq_run_hw_queue(hctx); |
320ae51f JA |
1227 | blk_mq_put_ctx(ctx); |
1228 | trace_block_sleeprq(q, bio, rw); | |
793597a6 CH |
1229 | |
1230 | ctx = blk_mq_get_ctx(q); | |
320ae51f | 1231 | hctx = q->mq_ops->map_queue(q, ctx->cpu); |
3c92db6a | 1232 | blk_mq_set_alloc_data(&alloc_data, q, 0, ctx, hctx); |
cb96a42c ML |
1233 | rq = __blk_mq_alloc_request(&alloc_data, rw); |
1234 | ctx = alloc_data.ctx; | |
1235 | hctx = alloc_data.hctx; | |
320ae51f JA |
1236 | } |
1237 | ||
1238 | hctx->queued++; | |
07068d5b JA |
1239 | data->hctx = hctx; |
1240 | data->ctx = ctx; | |
1241 | return rq; | |
1242 | } | |
1243 | ||
7b371636 | 1244 | static int blk_mq_direct_issue_request(struct request *rq, blk_qc_t *cookie) |
f984df1f SL |
1245 | { |
1246 | int ret; | |
1247 | struct request_queue *q = rq->q; | |
1248 | struct blk_mq_hw_ctx *hctx = q->mq_ops->map_queue(q, | |
1249 | rq->mq_ctx->cpu); | |
1250 | struct blk_mq_queue_data bd = { | |
1251 | .rq = rq, | |
1252 | .list = NULL, | |
1253 | .last = 1 | |
1254 | }; | |
7b371636 | 1255 | blk_qc_t new_cookie = blk_tag_to_qc_t(rq->tag, hctx->queue_num); |
f984df1f SL |
1256 | |
1257 | /* | |
1258 | * For OK queue, we are done. For error, kill it. Any other | |
1259 | * error (busy), just add it to our list as we previously | |
1260 | * would have done | |
1261 | */ | |
1262 | ret = q->mq_ops->queue_rq(hctx, &bd); | |
7b371636 JA |
1263 | if (ret == BLK_MQ_RQ_QUEUE_OK) { |
1264 | *cookie = new_cookie; | |
f984df1f | 1265 | return 0; |
7b371636 | 1266 | } |
f984df1f | 1267 | |
7b371636 JA |
1268 | __blk_mq_requeue_request(rq); |
1269 | ||
1270 | if (ret == BLK_MQ_RQ_QUEUE_ERROR) { | |
1271 | *cookie = BLK_QC_T_NONE; | |
1272 | rq->errors = -EIO; | |
1273 | blk_mq_end_request(rq, rq->errors); | |
1274 | return 0; | |
f984df1f | 1275 | } |
7b371636 JA |
1276 | |
1277 | return -1; | |
f984df1f SL |
1278 | } |
1279 | ||
07068d5b JA |
1280 | /* |
1281 | * Multiple hardware queue variant. This will not use per-process plugs, | |
1282 | * but will attempt to bypass the hctx queueing if we can go straight to | |
1283 | * hardware for SYNC IO. | |
1284 | */ | |
dece1635 | 1285 | static blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio) |
07068d5b JA |
1286 | { |
1287 | const int is_sync = rw_is_sync(bio->bi_rw); | |
1288 | const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA); | |
1289 | struct blk_map_ctx data; | |
1290 | struct request *rq; | |
f984df1f SL |
1291 | unsigned int request_count = 0; |
1292 | struct blk_plug *plug; | |
5b3f341f | 1293 | struct request *same_queue_rq = NULL; |
7b371636 | 1294 | blk_qc_t cookie; |
07068d5b JA |
1295 | |
1296 | blk_queue_bounce(q, &bio); | |
1297 | ||
8ffac713 WX |
1298 | blk_queue_split(q, &bio, q->bio_split); |
1299 | ||
07068d5b | 1300 | if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) { |
4246a0b6 | 1301 | bio_io_error(bio); |
dece1635 | 1302 | return BLK_QC_T_NONE; |
07068d5b JA |
1303 | } |
1304 | ||
7392f29e OS |
1305 | if (!is_flush_fua && !blk_queue_nomerges(q) && |
1306 | blk_attempt_plug_merge(q, bio, &request_count, &same_queue_rq)) | |
1307 | return BLK_QC_T_NONE; | |
f984df1f | 1308 | |
07068d5b JA |
1309 | rq = blk_mq_map_request(q, bio, &data); |
1310 | if (unlikely(!rq)) | |
dece1635 | 1311 | return BLK_QC_T_NONE; |
07068d5b | 1312 | |
7b371636 | 1313 | cookie = blk_tag_to_qc_t(rq->tag, data.hctx->queue_num); |
07068d5b JA |
1314 | |
1315 | if (unlikely(is_flush_fua)) { | |
1316 | blk_mq_bio_to_request(rq, bio); | |
1317 | blk_insert_flush(rq); | |
1318 | goto run_queue; | |
1319 | } | |
1320 | ||
f984df1f | 1321 | plug = current->plug; |
e167dfb5 JA |
1322 | /* |
1323 | * If the driver supports defer issued based on 'last', then | |
1324 | * queue it up like normal since we can potentially save some | |
1325 | * CPU this way. | |
1326 | */ | |
f984df1f SL |
1327 | if (((plug && !blk_queue_nomerges(q)) || is_sync) && |
1328 | !(data.hctx->flags & BLK_MQ_F_DEFER_ISSUE)) { | |
1329 | struct request *old_rq = NULL; | |
07068d5b JA |
1330 | |
1331 | blk_mq_bio_to_request(rq, bio); | |
07068d5b JA |
1332 | |
1333 | /* | |
b094f89c | 1334 | * We do limited pluging. If the bio can be merged, do that. |
f984df1f SL |
1335 | * Otherwise the existing request in the plug list will be |
1336 | * issued. So the plug list will have one request at most | |
07068d5b | 1337 | */ |
f984df1f | 1338 | if (plug) { |
5b3f341f SL |
1339 | /* |
1340 | * The plug list might get flushed before this. If that | |
b094f89c JA |
1341 | * happens, same_queue_rq is invalid and plug list is |
1342 | * empty | |
1343 | */ | |
5b3f341f SL |
1344 | if (same_queue_rq && !list_empty(&plug->mq_list)) { |
1345 | old_rq = same_queue_rq; | |
f984df1f | 1346 | list_del_init(&old_rq->queuelist); |
07068d5b | 1347 | } |
f984df1f SL |
1348 | list_add_tail(&rq->queuelist, &plug->mq_list); |
1349 | } else /* is_sync */ | |
1350 | old_rq = rq; | |
1351 | blk_mq_put_ctx(data.ctx); | |
1352 | if (!old_rq) | |
7b371636 | 1353 | goto done; |
b455d7c1 BVA |
1354 | if (test_bit(BLK_MQ_S_STOPPED, &data.hctx->state) || |
1355 | blk_mq_direct_issue_request(old_rq, &cookie) != 0) | |
1356 | blk_mq_insert_request(old_rq, false, true, true); | |
7b371636 | 1357 | goto done; |
07068d5b JA |
1358 | } |
1359 | ||
1360 | if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) { | |
1361 | /* | |
1362 | * For a SYNC request, send it to the hardware immediately. For | |
1363 | * an ASYNC request, just ensure that we run it later on. The | |
1364 | * latter allows for merging opportunities and more efficient | |
1365 | * dispatching. | |
1366 | */ | |
1367 | run_queue: | |
1368 | blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua); | |
1369 | } | |
07068d5b | 1370 | blk_mq_put_ctx(data.ctx); |
7b371636 JA |
1371 | done: |
1372 | return cookie; | |
07068d5b JA |
1373 | } |
1374 | ||
1375 | /* | |
1376 | * Single hardware queue variant. This will attempt to use any per-process | |
1377 | * plug for merging and IO deferral. | |
1378 | */ | |
dece1635 | 1379 | static blk_qc_t blk_sq_make_request(struct request_queue *q, struct bio *bio) |
07068d5b JA |
1380 | { |
1381 | const int is_sync = rw_is_sync(bio->bi_rw); | |
1382 | const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA); | |
e6c4438b JM |
1383 | struct blk_plug *plug; |
1384 | unsigned int request_count = 0; | |
07068d5b JA |
1385 | struct blk_map_ctx data; |
1386 | struct request *rq; | |
7b371636 | 1387 | blk_qc_t cookie; |
07068d5b | 1388 | |
07068d5b JA |
1389 | blk_queue_bounce(q, &bio); |
1390 | ||
1391 | if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) { | |
4246a0b6 | 1392 | bio_io_error(bio); |
dece1635 | 1393 | return BLK_QC_T_NONE; |
07068d5b JA |
1394 | } |
1395 | ||
54efd50b KO |
1396 | blk_queue_split(q, &bio, q->bio_split); |
1397 | ||
7392f29e OS |
1398 | if (!is_flush_fua && !blk_queue_nomerges(q)) { |
1399 | if (blk_attempt_plug_merge(q, bio, &request_count, NULL)) | |
1400 | return BLK_QC_T_NONE; | |
1401 | } else | |
1402 | request_count = blk_plug_queued_count(q); | |
07068d5b JA |
1403 | |
1404 | rq = blk_mq_map_request(q, bio, &data); | |
ff87bcec | 1405 | if (unlikely(!rq)) |
dece1635 | 1406 | return BLK_QC_T_NONE; |
320ae51f | 1407 | |
7b371636 | 1408 | cookie = blk_tag_to_qc_t(rq->tag, data.hctx->queue_num); |
320ae51f JA |
1409 | |
1410 | if (unlikely(is_flush_fua)) { | |
1411 | blk_mq_bio_to_request(rq, bio); | |
320ae51f JA |
1412 | blk_insert_flush(rq); |
1413 | goto run_queue; | |
1414 | } | |
1415 | ||
1416 | /* | |
1417 | * A task plug currently exists. Since this is completely lockless, | |
1418 | * utilize that to temporarily store requests until the task is | |
1419 | * either done or scheduled away. | |
1420 | */ | |
e6c4438b JM |
1421 | plug = current->plug; |
1422 | if (plug) { | |
1423 | blk_mq_bio_to_request(rq, bio); | |
676d0607 | 1424 | if (!request_count) |
e6c4438b | 1425 | trace_block_plug(q); |
b094f89c JA |
1426 | |
1427 | blk_mq_put_ctx(data.ctx); | |
1428 | ||
1429 | if (request_count >= BLK_MAX_REQUEST_COUNT) { | |
e6c4438b JM |
1430 | blk_flush_plug_list(plug, false); |
1431 | trace_block_plug(q); | |
320ae51f | 1432 | } |
b094f89c | 1433 | |
e6c4438b | 1434 | list_add_tail(&rq->queuelist, &plug->mq_list); |
7b371636 | 1435 | return cookie; |
320ae51f JA |
1436 | } |
1437 | ||
07068d5b JA |
1438 | if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) { |
1439 | /* | |
1440 | * For a SYNC request, send it to the hardware immediately. For | |
1441 | * an ASYNC request, just ensure that we run it later on. The | |
1442 | * latter allows for merging opportunities and more efficient | |
1443 | * dispatching. | |
1444 | */ | |
1445 | run_queue: | |
1446 | blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua); | |
320ae51f JA |
1447 | } |
1448 | ||
07068d5b | 1449 | blk_mq_put_ctx(data.ctx); |
7b371636 | 1450 | return cookie; |
320ae51f JA |
1451 | } |
1452 | ||
1453 | /* | |
1454 | * Default mapping to a software queue, since we use one per CPU. | |
1455 | */ | |
1456 | struct blk_mq_hw_ctx *blk_mq_map_queue(struct request_queue *q, const int cpu) | |
1457 | { | |
1458 | return q->queue_hw_ctx[q->mq_map[cpu]]; | |
1459 | } | |
1460 | EXPORT_SYMBOL(blk_mq_map_queue); | |
1461 | ||
24d2f903 CH |
1462 | static void blk_mq_free_rq_map(struct blk_mq_tag_set *set, |
1463 | struct blk_mq_tags *tags, unsigned int hctx_idx) | |
95363efd | 1464 | { |
e9b267d9 | 1465 | struct page *page; |
320ae51f | 1466 | |
24d2f903 | 1467 | if (tags->rqs && set->ops->exit_request) { |
e9b267d9 | 1468 | int i; |
320ae51f | 1469 | |
24d2f903 CH |
1470 | for (i = 0; i < tags->nr_tags; i++) { |
1471 | if (!tags->rqs[i]) | |
e9b267d9 | 1472 | continue; |
24d2f903 CH |
1473 | set->ops->exit_request(set->driver_data, tags->rqs[i], |
1474 | hctx_idx, i); | |
a5164405 | 1475 | tags->rqs[i] = NULL; |
e9b267d9 | 1476 | } |
320ae51f | 1477 | } |
320ae51f | 1478 | |
24d2f903 CH |
1479 | while (!list_empty(&tags->page_list)) { |
1480 | page = list_first_entry(&tags->page_list, struct page, lru); | |
6753471c | 1481 | list_del_init(&page->lru); |
f75782e4 CM |
1482 | /* |
1483 | * Remove kmemleak object previously allocated in | |
1484 | * blk_mq_init_rq_map(). | |
1485 | */ | |
1486 | kmemleak_free(page_address(page)); | |
320ae51f JA |
1487 | __free_pages(page, page->private); |
1488 | } | |
1489 | ||
24d2f903 | 1490 | kfree(tags->rqs); |
320ae51f | 1491 | |
24d2f903 | 1492 | blk_mq_free_tags(tags); |
320ae51f JA |
1493 | } |
1494 | ||
1495 | static size_t order_to_size(unsigned int order) | |
1496 | { | |
4ca08500 | 1497 | return (size_t)PAGE_SIZE << order; |
320ae51f JA |
1498 | } |
1499 | ||
24d2f903 CH |
1500 | static struct blk_mq_tags *blk_mq_init_rq_map(struct blk_mq_tag_set *set, |
1501 | unsigned int hctx_idx) | |
320ae51f | 1502 | { |
24d2f903 | 1503 | struct blk_mq_tags *tags; |
320ae51f JA |
1504 | unsigned int i, j, entries_per_page, max_order = 4; |
1505 | size_t rq_size, left; | |
1506 | ||
24d2f903 | 1507 | tags = blk_mq_init_tags(set->queue_depth, set->reserved_tags, |
24391c0d SL |
1508 | set->numa_node, |
1509 | BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags)); | |
24d2f903 CH |
1510 | if (!tags) |
1511 | return NULL; | |
320ae51f | 1512 | |
24d2f903 CH |
1513 | INIT_LIST_HEAD(&tags->page_list); |
1514 | ||
a5164405 | 1515 | tags->rqs = kzalloc_node(set->queue_depth * sizeof(struct request *), |
2fd8868a | 1516 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, |
a5164405 | 1517 | set->numa_node); |
24d2f903 CH |
1518 | if (!tags->rqs) { |
1519 | blk_mq_free_tags(tags); | |
1520 | return NULL; | |
1521 | } | |
320ae51f JA |
1522 | |
1523 | /* | |
1524 | * rq_size is the size of the request plus driver payload, rounded | |
1525 | * to the cacheline size | |
1526 | */ | |
24d2f903 | 1527 | rq_size = round_up(sizeof(struct request) + set->cmd_size, |
320ae51f | 1528 | cache_line_size()); |
24d2f903 | 1529 | left = rq_size * set->queue_depth; |
320ae51f | 1530 | |
24d2f903 | 1531 | for (i = 0; i < set->queue_depth; ) { |
320ae51f JA |
1532 | int this_order = max_order; |
1533 | struct page *page; | |
1534 | int to_do; | |
1535 | void *p; | |
1536 | ||
1537 | while (left < order_to_size(this_order - 1) && this_order) | |
1538 | this_order--; | |
1539 | ||
1540 | do { | |
a5164405 | 1541 | page = alloc_pages_node(set->numa_node, |
2fd8868a | 1542 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO, |
a5164405 | 1543 | this_order); |
320ae51f JA |
1544 | if (page) |
1545 | break; | |
1546 | if (!this_order--) | |
1547 | break; | |
1548 | if (order_to_size(this_order) < rq_size) | |
1549 | break; | |
1550 | } while (1); | |
1551 | ||
1552 | if (!page) | |
24d2f903 | 1553 | goto fail; |
320ae51f JA |
1554 | |
1555 | page->private = this_order; | |
24d2f903 | 1556 | list_add_tail(&page->lru, &tags->page_list); |
320ae51f JA |
1557 | |
1558 | p = page_address(page); | |
f75782e4 CM |
1559 | /* |
1560 | * Allow kmemleak to scan these pages as they contain pointers | |
1561 | * to additional allocations like via ops->init_request(). | |
1562 | */ | |
2fd8868a | 1563 | kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO); |
320ae51f | 1564 | entries_per_page = order_to_size(this_order) / rq_size; |
24d2f903 | 1565 | to_do = min(entries_per_page, set->queue_depth - i); |
320ae51f JA |
1566 | left -= to_do * rq_size; |
1567 | for (j = 0; j < to_do; j++) { | |
24d2f903 CH |
1568 | tags->rqs[i] = p; |
1569 | if (set->ops->init_request) { | |
1570 | if (set->ops->init_request(set->driver_data, | |
1571 | tags->rqs[i], hctx_idx, i, | |
a5164405 JA |
1572 | set->numa_node)) { |
1573 | tags->rqs[i] = NULL; | |
24d2f903 | 1574 | goto fail; |
a5164405 | 1575 | } |
e9b267d9 CH |
1576 | } |
1577 | ||
320ae51f JA |
1578 | p += rq_size; |
1579 | i++; | |
1580 | } | |
1581 | } | |
24d2f903 | 1582 | return tags; |
320ae51f | 1583 | |
24d2f903 | 1584 | fail: |
24d2f903 CH |
1585 | blk_mq_free_rq_map(set, tags, hctx_idx); |
1586 | return NULL; | |
320ae51f JA |
1587 | } |
1588 | ||
1429d7c9 JA |
1589 | static void blk_mq_free_bitmap(struct blk_mq_ctxmap *bitmap) |
1590 | { | |
1591 | kfree(bitmap->map); | |
1592 | } | |
1593 | ||
1594 | static int blk_mq_alloc_bitmap(struct blk_mq_ctxmap *bitmap, int node) | |
1595 | { | |
1596 | unsigned int bpw = 8, total, num_maps, i; | |
1597 | ||
1598 | bitmap->bits_per_word = bpw; | |
1599 | ||
1600 | num_maps = ALIGN(nr_cpu_ids, bpw) / bpw; | |
1601 | bitmap->map = kzalloc_node(num_maps * sizeof(struct blk_align_bitmap), | |
1602 | GFP_KERNEL, node); | |
1603 | if (!bitmap->map) | |
1604 | return -ENOMEM; | |
1605 | ||
1429d7c9 JA |
1606 | total = nr_cpu_ids; |
1607 | for (i = 0; i < num_maps; i++) { | |
1608 | bitmap->map[i].depth = min(total, bitmap->bits_per_word); | |
1609 | total -= bitmap->map[i].depth; | |
1610 | } | |
1611 | ||
1612 | return 0; | |
1613 | } | |
1614 | ||
f463371e JA |
1615 | /* |
1616 | * 'cpu' is going away. splice any existing rq_list entries from this | |
1617 | * software queue to the hw queue dispatch list, and ensure that it | |
1618 | * gets run. | |
1619 | */ | |
484b4061 JA |
1620 | static int blk_mq_hctx_cpu_offline(struct blk_mq_hw_ctx *hctx, int cpu) |
1621 | { | |
484b4061 JA |
1622 | struct blk_mq_ctx *ctx; |
1623 | LIST_HEAD(tmp); | |
1624 | ||
f463371e | 1625 | ctx = __blk_mq_get_ctx(hctx->queue, cpu); |
484b4061 JA |
1626 | |
1627 | spin_lock(&ctx->lock); | |
1628 | if (!list_empty(&ctx->rq_list)) { | |
1629 | list_splice_init(&ctx->rq_list, &tmp); | |
1630 | blk_mq_hctx_clear_pending(hctx, ctx); | |
1631 | } | |
1632 | spin_unlock(&ctx->lock); | |
1633 | ||
1634 | if (list_empty(&tmp)) | |
1635 | return NOTIFY_OK; | |
1636 | ||
f463371e JA |
1637 | spin_lock(&hctx->lock); |
1638 | list_splice_tail_init(&tmp, &hctx->dispatch); | |
1639 | spin_unlock(&hctx->lock); | |
484b4061 JA |
1640 | |
1641 | blk_mq_run_hw_queue(hctx, true); | |
484b4061 JA |
1642 | return NOTIFY_OK; |
1643 | } | |
1644 | ||
484b4061 JA |
1645 | static int blk_mq_hctx_notify(void *data, unsigned long action, |
1646 | unsigned int cpu) | |
1647 | { | |
1648 | struct blk_mq_hw_ctx *hctx = data; | |
1649 | ||
1650 | if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) | |
1651 | return blk_mq_hctx_cpu_offline(hctx, cpu); | |
2a34c087 ML |
1652 | |
1653 | /* | |
1654 | * In case of CPU online, tags may be reallocated | |
1655 | * in blk_mq_map_swqueue() after mapping is updated. | |
1656 | */ | |
484b4061 JA |
1657 | |
1658 | return NOTIFY_OK; | |
1659 | } | |
1660 | ||
c3b4afca | 1661 | /* hctx->ctxs will be freed in queue's release handler */ |
08e98fc6 ML |
1662 | static void blk_mq_exit_hctx(struct request_queue *q, |
1663 | struct blk_mq_tag_set *set, | |
1664 | struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx) | |
1665 | { | |
f70ced09 ML |
1666 | unsigned flush_start_tag = set->queue_depth; |
1667 | ||
08e98fc6 ML |
1668 | blk_mq_tag_idle(hctx); |
1669 | ||
f70ced09 ML |
1670 | if (set->ops->exit_request) |
1671 | set->ops->exit_request(set->driver_data, | |
1672 | hctx->fq->flush_rq, hctx_idx, | |
1673 | flush_start_tag + hctx_idx); | |
1674 | ||
08e98fc6 ML |
1675 | if (set->ops->exit_hctx) |
1676 | set->ops->exit_hctx(hctx, hctx_idx); | |
1677 | ||
1678 | blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier); | |
f70ced09 | 1679 | blk_free_flush_queue(hctx->fq); |
08e98fc6 ML |
1680 | blk_mq_free_bitmap(&hctx->ctx_map); |
1681 | } | |
1682 | ||
624dbe47 ML |
1683 | static void blk_mq_exit_hw_queues(struct request_queue *q, |
1684 | struct blk_mq_tag_set *set, int nr_queue) | |
1685 | { | |
1686 | struct blk_mq_hw_ctx *hctx; | |
1687 | unsigned int i; | |
1688 | ||
1689 | queue_for_each_hw_ctx(q, hctx, i) { | |
1690 | if (i == nr_queue) | |
1691 | break; | |
08e98fc6 | 1692 | blk_mq_exit_hctx(q, set, hctx, i); |
624dbe47 | 1693 | } |
624dbe47 ML |
1694 | } |
1695 | ||
1696 | static void blk_mq_free_hw_queues(struct request_queue *q, | |
1697 | struct blk_mq_tag_set *set) | |
1698 | { | |
1699 | struct blk_mq_hw_ctx *hctx; | |
1700 | unsigned int i; | |
1701 | ||
e09aae7e | 1702 | queue_for_each_hw_ctx(q, hctx, i) |
624dbe47 | 1703 | free_cpumask_var(hctx->cpumask); |
624dbe47 ML |
1704 | } |
1705 | ||
08e98fc6 ML |
1706 | static int blk_mq_init_hctx(struct request_queue *q, |
1707 | struct blk_mq_tag_set *set, | |
1708 | struct blk_mq_hw_ctx *hctx, unsigned hctx_idx) | |
320ae51f | 1709 | { |
08e98fc6 | 1710 | int node; |
f70ced09 | 1711 | unsigned flush_start_tag = set->queue_depth; |
08e98fc6 ML |
1712 | |
1713 | node = hctx->numa_node; | |
1714 | if (node == NUMA_NO_NODE) | |
1715 | node = hctx->numa_node = set->numa_node; | |
1716 | ||
1717 | INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn); | |
1718 | INIT_DELAYED_WORK(&hctx->delay_work, blk_mq_delay_work_fn); | |
1719 | spin_lock_init(&hctx->lock); | |
1720 | INIT_LIST_HEAD(&hctx->dispatch); | |
1721 | hctx->queue = q; | |
1722 | hctx->queue_num = hctx_idx; | |
2404e607 | 1723 | hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED; |
08e98fc6 ML |
1724 | |
1725 | blk_mq_init_cpu_notifier(&hctx->cpu_notifier, | |
1726 | blk_mq_hctx_notify, hctx); | |
1727 | blk_mq_register_cpu_notifier(&hctx->cpu_notifier); | |
1728 | ||
1729 | hctx->tags = set->tags[hctx_idx]; | |
320ae51f JA |
1730 | |
1731 | /* | |
08e98fc6 ML |
1732 | * Allocate space for all possible cpus to avoid allocation at |
1733 | * runtime | |
320ae51f | 1734 | */ |
08e98fc6 ML |
1735 | hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *), |
1736 | GFP_KERNEL, node); | |
1737 | if (!hctx->ctxs) | |
1738 | goto unregister_cpu_notifier; | |
320ae51f | 1739 | |
08e98fc6 ML |
1740 | if (blk_mq_alloc_bitmap(&hctx->ctx_map, node)) |
1741 | goto free_ctxs; | |
320ae51f | 1742 | |
08e98fc6 | 1743 | hctx->nr_ctx = 0; |
320ae51f | 1744 | |
08e98fc6 ML |
1745 | if (set->ops->init_hctx && |
1746 | set->ops->init_hctx(hctx, set->driver_data, hctx_idx)) | |
1747 | goto free_bitmap; | |
320ae51f | 1748 | |
f70ced09 ML |
1749 | hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size); |
1750 | if (!hctx->fq) | |
1751 | goto exit_hctx; | |
320ae51f | 1752 | |
f70ced09 ML |
1753 | if (set->ops->init_request && |
1754 | set->ops->init_request(set->driver_data, | |
1755 | hctx->fq->flush_rq, hctx_idx, | |
1756 | flush_start_tag + hctx_idx, node)) | |
1757 | goto free_fq; | |
320ae51f | 1758 | |
08e98fc6 | 1759 | return 0; |
320ae51f | 1760 | |
f70ced09 ML |
1761 | free_fq: |
1762 | kfree(hctx->fq); | |
1763 | exit_hctx: | |
1764 | if (set->ops->exit_hctx) | |
1765 | set->ops->exit_hctx(hctx, hctx_idx); | |
08e98fc6 ML |
1766 | free_bitmap: |
1767 | blk_mq_free_bitmap(&hctx->ctx_map); | |
1768 | free_ctxs: | |
1769 | kfree(hctx->ctxs); | |
1770 | unregister_cpu_notifier: | |
1771 | blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier); | |
320ae51f | 1772 | |
08e98fc6 ML |
1773 | return -1; |
1774 | } | |
320ae51f | 1775 | |
08e98fc6 ML |
1776 | static int blk_mq_init_hw_queues(struct request_queue *q, |
1777 | struct blk_mq_tag_set *set) | |
1778 | { | |
1779 | struct blk_mq_hw_ctx *hctx; | |
1780 | unsigned int i; | |
320ae51f | 1781 | |
08e98fc6 ML |
1782 | /* |
1783 | * Initialize hardware queues | |
1784 | */ | |
1785 | queue_for_each_hw_ctx(q, hctx, i) { | |
1786 | if (blk_mq_init_hctx(q, set, hctx, i)) | |
320ae51f JA |
1787 | break; |
1788 | } | |
1789 | ||
1790 | if (i == q->nr_hw_queues) | |
1791 | return 0; | |
1792 | ||
1793 | /* | |
1794 | * Init failed | |
1795 | */ | |
624dbe47 | 1796 | blk_mq_exit_hw_queues(q, set, i); |
320ae51f JA |
1797 | |
1798 | return 1; | |
1799 | } | |
1800 | ||
1801 | static void blk_mq_init_cpu_queues(struct request_queue *q, | |
1802 | unsigned int nr_hw_queues) | |
1803 | { | |
1804 | unsigned int i; | |
1805 | ||
1806 | for_each_possible_cpu(i) { | |
1807 | struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i); | |
1808 | struct blk_mq_hw_ctx *hctx; | |
1809 | ||
1810 | memset(__ctx, 0, sizeof(*__ctx)); | |
1811 | __ctx->cpu = i; | |
1812 | spin_lock_init(&__ctx->lock); | |
1813 | INIT_LIST_HEAD(&__ctx->rq_list); | |
1814 | __ctx->queue = q; | |
1815 | ||
1816 | /* If the cpu isn't online, the cpu is mapped to first hctx */ | |
320ae51f JA |
1817 | if (!cpu_online(i)) |
1818 | continue; | |
1819 | ||
e4043dcf | 1820 | hctx = q->mq_ops->map_queue(q, i); |
e4043dcf | 1821 | |
320ae51f JA |
1822 | /* |
1823 | * Set local node, IFF we have more than one hw queue. If | |
1824 | * not, we remain on the home node of the device | |
1825 | */ | |
1826 | if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE) | |
1827 | hctx->numa_node = cpu_to_node(i); | |
1828 | } | |
1829 | } | |
1830 | ||
5778322e AM |
1831 | static void blk_mq_map_swqueue(struct request_queue *q, |
1832 | const struct cpumask *online_mask) | |
320ae51f | 1833 | { |
ce7a95c5 | 1834 | unsigned int i, hctx_idx; |
320ae51f JA |
1835 | struct blk_mq_hw_ctx *hctx; |
1836 | struct blk_mq_ctx *ctx; | |
2a34c087 | 1837 | struct blk_mq_tag_set *set = q->tag_set; |
320ae51f | 1838 | |
60de074b AM |
1839 | /* |
1840 | * Avoid others reading imcomplete hctx->cpumask through sysfs | |
1841 | */ | |
1842 | mutex_lock(&q->sysfs_lock); | |
1843 | ||
320ae51f | 1844 | queue_for_each_hw_ctx(q, hctx, i) { |
e4043dcf | 1845 | cpumask_clear(hctx->cpumask); |
320ae51f JA |
1846 | hctx->nr_ctx = 0; |
1847 | } | |
1848 | ||
1849 | /* | |
1850 | * Map software to hardware queues | |
1851 | */ | |
c4cf5fcd | 1852 | for_each_possible_cpu(i) { |
320ae51f | 1853 | /* If the cpu isn't online, the cpu is mapped to first hctx */ |
5778322e | 1854 | if (!cpumask_test_cpu(i, online_mask)) |
e4043dcf JA |
1855 | continue; |
1856 | ||
ce7a95c5 GKB |
1857 | hctx_idx = q->mq_map[i]; |
1858 | /* unmapped hw queue can be remapped after CPU topo changed */ | |
1859 | if (!set->tags[hctx_idx]) { | |
1860 | set->tags[hctx_idx] = blk_mq_init_rq_map(set, hctx_idx); | |
1861 | ||
1862 | /* | |
1863 | * If tags initialization fail for some hctx, | |
1864 | * that hctx won't be brought online. In this | |
1865 | * case, remap the current ctx to hctx[0] which | |
1866 | * is guaranteed to always have tags allocated | |
1867 | */ | |
1868 | if (!set->tags[hctx_idx]) | |
1869 | q->mq_map[i] = 0; | |
1870 | } | |
1871 | ||
c4cf5fcd | 1872 | ctx = per_cpu_ptr(q->queue_ctx, i); |
320ae51f | 1873 | hctx = q->mq_ops->map_queue(q, i); |
e4043dcf | 1874 | cpumask_set_cpu(i, hctx->cpumask); |
320ae51f JA |
1875 | ctx->index_hw = hctx->nr_ctx; |
1876 | hctx->ctxs[hctx->nr_ctx++] = ctx; | |
1877 | } | |
506e931f | 1878 | |
60de074b AM |
1879 | mutex_unlock(&q->sysfs_lock); |
1880 | ||
506e931f | 1881 | queue_for_each_hw_ctx(q, hctx, i) { |
889fa31f CY |
1882 | struct blk_mq_ctxmap *map = &hctx->ctx_map; |
1883 | ||
484b4061 | 1884 | /* |
a68aafa5 JA |
1885 | * If no software queues are mapped to this hardware queue, |
1886 | * disable it and free the request entries. | |
484b4061 JA |
1887 | */ |
1888 | if (!hctx->nr_ctx) { | |
ce7a95c5 GKB |
1889 | /* Never unmap queue 0. We need it as a |
1890 | * fallback in case of a new remap fails | |
1891 | * allocation | |
1892 | */ | |
1893 | if (i && set->tags[i]) { | |
484b4061 JA |
1894 | blk_mq_free_rq_map(set, set->tags[i], i); |
1895 | set->tags[i] = NULL; | |
484b4061 | 1896 | } |
2a34c087 | 1897 | hctx->tags = NULL; |
484b4061 JA |
1898 | continue; |
1899 | } | |
1900 | ||
2a34c087 ML |
1901 | hctx->tags = set->tags[i]; |
1902 | WARN_ON(!hctx->tags); | |
1903 | ||
4ab4f927 | 1904 | cpumask_copy(hctx->tags->cpumask, hctx->cpumask); |
889fa31f CY |
1905 | /* |
1906 | * Set the map size to the number of mapped software queues. | |
1907 | * This is more accurate and more efficient than looping | |
1908 | * over all possibly mapped software queues. | |
1909 | */ | |
569fd0ce | 1910 | map->size = DIV_ROUND_UP(hctx->nr_ctx, map->bits_per_word); |
889fa31f | 1911 | |
484b4061 JA |
1912 | /* |
1913 | * Initialize batch roundrobin counts | |
1914 | */ | |
506e931f JA |
1915 | hctx->next_cpu = cpumask_first(hctx->cpumask); |
1916 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; | |
1917 | } | |
320ae51f JA |
1918 | } |
1919 | ||
2404e607 | 1920 | static void queue_set_hctx_shared(struct request_queue *q, bool shared) |
0d2602ca JA |
1921 | { |
1922 | struct blk_mq_hw_ctx *hctx; | |
0d2602ca JA |
1923 | int i; |
1924 | ||
2404e607 JM |
1925 | queue_for_each_hw_ctx(q, hctx, i) { |
1926 | if (shared) | |
1927 | hctx->flags |= BLK_MQ_F_TAG_SHARED; | |
1928 | else | |
1929 | hctx->flags &= ~BLK_MQ_F_TAG_SHARED; | |
1930 | } | |
1931 | } | |
1932 | ||
1933 | static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set, bool shared) | |
1934 | { | |
1935 | struct request_queue *q; | |
0d2602ca JA |
1936 | |
1937 | list_for_each_entry(q, &set->tag_list, tag_set_list) { | |
1938 | blk_mq_freeze_queue(q); | |
2404e607 | 1939 | queue_set_hctx_shared(q, shared); |
0d2602ca JA |
1940 | blk_mq_unfreeze_queue(q); |
1941 | } | |
1942 | } | |
1943 | ||
1944 | static void blk_mq_del_queue_tag_set(struct request_queue *q) | |
1945 | { | |
1946 | struct blk_mq_tag_set *set = q->tag_set; | |
1947 | ||
0d2602ca JA |
1948 | mutex_lock(&set->tag_list_lock); |
1949 | list_del_init(&q->tag_set_list); | |
2404e607 JM |
1950 | if (list_is_singular(&set->tag_list)) { |
1951 | /* just transitioned to unshared */ | |
1952 | set->flags &= ~BLK_MQ_F_TAG_SHARED; | |
1953 | /* update existing queue */ | |
1954 | blk_mq_update_tag_set_depth(set, false); | |
1955 | } | |
0d2602ca | 1956 | mutex_unlock(&set->tag_list_lock); |
0d2602ca JA |
1957 | } |
1958 | ||
1959 | static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set, | |
1960 | struct request_queue *q) | |
1961 | { | |
1962 | q->tag_set = set; | |
1963 | ||
1964 | mutex_lock(&set->tag_list_lock); | |
2404e607 JM |
1965 | |
1966 | /* Check to see if we're transitioning to shared (from 1 to 2 queues). */ | |
1967 | if (!list_empty(&set->tag_list) && !(set->flags & BLK_MQ_F_TAG_SHARED)) { | |
1968 | set->flags |= BLK_MQ_F_TAG_SHARED; | |
1969 | /* update existing queue */ | |
1970 | blk_mq_update_tag_set_depth(set, true); | |
1971 | } | |
1972 | if (set->flags & BLK_MQ_F_TAG_SHARED) | |
1973 | queue_set_hctx_shared(q, true); | |
0d2602ca | 1974 | list_add_tail(&q->tag_set_list, &set->tag_list); |
2404e607 | 1975 | |
0d2602ca JA |
1976 | mutex_unlock(&set->tag_list_lock); |
1977 | } | |
1978 | ||
e09aae7e ML |
1979 | /* |
1980 | * It is the actual release handler for mq, but we do it from | |
1981 | * request queue's release handler for avoiding use-after-free | |
1982 | * and headache because q->mq_kobj shouldn't have been introduced, | |
1983 | * but we can't group ctx/kctx kobj without it. | |
1984 | */ | |
1985 | void blk_mq_release(struct request_queue *q) | |
1986 | { | |
1987 | struct blk_mq_hw_ctx *hctx; | |
1988 | unsigned int i; | |
1989 | ||
1990 | /* hctx kobj stays in hctx */ | |
c3b4afca ML |
1991 | queue_for_each_hw_ctx(q, hctx, i) { |
1992 | if (!hctx) | |
1993 | continue; | |
1994 | kfree(hctx->ctxs); | |
e09aae7e | 1995 | kfree(hctx); |
c3b4afca | 1996 | } |
e09aae7e | 1997 | |
a723bab3 AM |
1998 | kfree(q->mq_map); |
1999 | q->mq_map = NULL; | |
2000 | ||
e09aae7e ML |
2001 | kfree(q->queue_hw_ctx); |
2002 | ||
2003 | /* ctx kobj stays in queue_ctx */ | |
2004 | free_percpu(q->queue_ctx); | |
2005 | } | |
2006 | ||
24d2f903 | 2007 | struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set) |
b62c21b7 MS |
2008 | { |
2009 | struct request_queue *uninit_q, *q; | |
2010 | ||
2011 | uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node); | |
2012 | if (!uninit_q) | |
2013 | return ERR_PTR(-ENOMEM); | |
2014 | ||
2015 | q = blk_mq_init_allocated_queue(set, uninit_q); | |
2016 | if (IS_ERR(q)) | |
2017 | blk_cleanup_queue(uninit_q); | |
2018 | ||
2019 | return q; | |
2020 | } | |
2021 | EXPORT_SYMBOL(blk_mq_init_queue); | |
2022 | ||
2023 | struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set, | |
2024 | struct request_queue *q) | |
320ae51f JA |
2025 | { |
2026 | struct blk_mq_hw_ctx **hctxs; | |
e6cdb092 | 2027 | struct blk_mq_ctx __percpu *ctx; |
f14bbe77 | 2028 | unsigned int *map; |
320ae51f JA |
2029 | int i; |
2030 | ||
320ae51f JA |
2031 | ctx = alloc_percpu(struct blk_mq_ctx); |
2032 | if (!ctx) | |
2033 | return ERR_PTR(-ENOMEM); | |
2034 | ||
24d2f903 CH |
2035 | hctxs = kmalloc_node(set->nr_hw_queues * sizeof(*hctxs), GFP_KERNEL, |
2036 | set->numa_node); | |
320ae51f JA |
2037 | |
2038 | if (!hctxs) | |
2039 | goto err_percpu; | |
2040 | ||
f14bbe77 JA |
2041 | map = blk_mq_make_queue_map(set); |
2042 | if (!map) | |
2043 | goto err_map; | |
2044 | ||
24d2f903 | 2045 | for (i = 0; i < set->nr_hw_queues; i++) { |
f14bbe77 JA |
2046 | int node = blk_mq_hw_queue_to_node(map, i); |
2047 | ||
cdef54dd CH |
2048 | hctxs[i] = kzalloc_node(sizeof(struct blk_mq_hw_ctx), |
2049 | GFP_KERNEL, node); | |
320ae51f JA |
2050 | if (!hctxs[i]) |
2051 | goto err_hctxs; | |
2052 | ||
a86073e4 JA |
2053 | if (!zalloc_cpumask_var_node(&hctxs[i]->cpumask, GFP_KERNEL, |
2054 | node)) | |
e4043dcf JA |
2055 | goto err_hctxs; |
2056 | ||
0d2602ca | 2057 | atomic_set(&hctxs[i]->nr_active, 0); |
f14bbe77 | 2058 | hctxs[i]->numa_node = node; |
320ae51f JA |
2059 | hctxs[i]->queue_num = i; |
2060 | } | |
2061 | ||
fc3bda55 | 2062 | INIT_WORK(&q->timeout_work, blk_mq_timeout_work); |
e56f698b | 2063 | blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ); |
320ae51f JA |
2064 | |
2065 | q->nr_queues = nr_cpu_ids; | |
24d2f903 | 2066 | q->nr_hw_queues = set->nr_hw_queues; |
f14bbe77 | 2067 | q->mq_map = map; |
320ae51f JA |
2068 | |
2069 | q->queue_ctx = ctx; | |
2070 | q->queue_hw_ctx = hctxs; | |
2071 | ||
24d2f903 | 2072 | q->mq_ops = set->ops; |
94eddfbe | 2073 | q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT; |
320ae51f | 2074 | |
05f1dd53 JA |
2075 | if (!(set->flags & BLK_MQ_F_SG_MERGE)) |
2076 | q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE; | |
2077 | ||
1be036e9 CH |
2078 | q->sg_reserved_size = INT_MAX; |
2079 | ||
6fca6a61 CH |
2080 | INIT_WORK(&q->requeue_work, blk_mq_requeue_work); |
2081 | INIT_LIST_HEAD(&q->requeue_list); | |
2082 | spin_lock_init(&q->requeue_lock); | |
2083 | ||
07068d5b JA |
2084 | if (q->nr_hw_queues > 1) |
2085 | blk_queue_make_request(q, blk_mq_make_request); | |
2086 | else | |
2087 | blk_queue_make_request(q, blk_sq_make_request); | |
2088 | ||
eba71768 JA |
2089 | /* |
2090 | * Do this after blk_queue_make_request() overrides it... | |
2091 | */ | |
2092 | q->nr_requests = set->queue_depth; | |
2093 | ||
24d2f903 CH |
2094 | if (set->ops->complete) |
2095 | blk_queue_softirq_done(q, set->ops->complete); | |
30a91cb4 | 2096 | |
24d2f903 | 2097 | blk_mq_init_cpu_queues(q, set->nr_hw_queues); |
320ae51f | 2098 | |
24d2f903 | 2099 | if (blk_mq_init_hw_queues(q, set)) |
b62c21b7 | 2100 | goto err_hctxs; |
18741986 | 2101 | |
5778322e | 2102 | get_online_cpus(); |
320ae51f | 2103 | mutex_lock(&all_q_mutex); |
320ae51f | 2104 | |
4593fdbe | 2105 | list_add_tail(&q->all_q_node, &all_q_list); |
0d2602ca | 2106 | blk_mq_add_queue_tag_set(set, q); |
5778322e | 2107 | blk_mq_map_swqueue(q, cpu_online_mask); |
484b4061 | 2108 | |
4593fdbe | 2109 | mutex_unlock(&all_q_mutex); |
5778322e | 2110 | put_online_cpus(); |
4593fdbe | 2111 | |
320ae51f | 2112 | return q; |
18741986 | 2113 | |
320ae51f | 2114 | err_hctxs: |
f14bbe77 | 2115 | kfree(map); |
24d2f903 | 2116 | for (i = 0; i < set->nr_hw_queues; i++) { |
320ae51f JA |
2117 | if (!hctxs[i]) |
2118 | break; | |
e4043dcf | 2119 | free_cpumask_var(hctxs[i]->cpumask); |
cdef54dd | 2120 | kfree(hctxs[i]); |
320ae51f | 2121 | } |
f14bbe77 | 2122 | err_map: |
320ae51f JA |
2123 | kfree(hctxs); |
2124 | err_percpu: | |
2125 | free_percpu(ctx); | |
2126 | return ERR_PTR(-ENOMEM); | |
2127 | } | |
b62c21b7 | 2128 | EXPORT_SYMBOL(blk_mq_init_allocated_queue); |
320ae51f JA |
2129 | |
2130 | void blk_mq_free_queue(struct request_queue *q) | |
2131 | { | |
624dbe47 | 2132 | struct blk_mq_tag_set *set = q->tag_set; |
320ae51f | 2133 | |
0e626368 AM |
2134 | mutex_lock(&all_q_mutex); |
2135 | list_del_init(&q->all_q_node); | |
2136 | mutex_unlock(&all_q_mutex); | |
2137 | ||
0d2602ca JA |
2138 | blk_mq_del_queue_tag_set(q); |
2139 | ||
624dbe47 ML |
2140 | blk_mq_exit_hw_queues(q, set, set->nr_hw_queues); |
2141 | blk_mq_free_hw_queues(q, set); | |
320ae51f | 2142 | } |
320ae51f JA |
2143 | |
2144 | /* Basically redo blk_mq_init_queue with queue frozen */ | |
5778322e AM |
2145 | static void blk_mq_queue_reinit(struct request_queue *q, |
2146 | const struct cpumask *online_mask) | |
320ae51f | 2147 | { |
4ecd4fef | 2148 | WARN_ON_ONCE(!atomic_read(&q->mq_freeze_depth)); |
320ae51f | 2149 | |
67aec14c JA |
2150 | blk_mq_sysfs_unregister(q); |
2151 | ||
5778322e | 2152 | blk_mq_update_queue_map(q->mq_map, q->nr_hw_queues, online_mask); |
320ae51f JA |
2153 | |
2154 | /* | |
2155 | * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe | |
2156 | * we should change hctx numa_node according to new topology (this | |
2157 | * involves free and re-allocate memory, worthy doing?) | |
2158 | */ | |
2159 | ||
5778322e | 2160 | blk_mq_map_swqueue(q, online_mask); |
320ae51f | 2161 | |
67aec14c | 2162 | blk_mq_sysfs_register(q); |
320ae51f JA |
2163 | } |
2164 | ||
f618ef7c PG |
2165 | static int blk_mq_queue_reinit_notify(struct notifier_block *nb, |
2166 | unsigned long action, void *hcpu) | |
320ae51f JA |
2167 | { |
2168 | struct request_queue *q; | |
5778322e AM |
2169 | int cpu = (unsigned long)hcpu; |
2170 | /* | |
2171 | * New online cpumask which is going to be set in this hotplug event. | |
2172 | * Declare this cpumasks as global as cpu-hotplug operation is invoked | |
2173 | * one-by-one and dynamically allocating this could result in a failure. | |
2174 | */ | |
2175 | static struct cpumask online_new; | |
320ae51f JA |
2176 | |
2177 | /* | |
5778322e AM |
2178 | * Before hotadded cpu starts handling requests, new mappings must |
2179 | * be established. Otherwise, these requests in hw queue might | |
2180 | * never be dispatched. | |
2181 | * | |
2182 | * For example, there is a single hw queue (hctx) and two CPU queues | |
2183 | * (ctx0 for CPU0, and ctx1 for CPU1). | |
2184 | * | |
2185 | * Now CPU1 is just onlined and a request is inserted into | |
2186 | * ctx1->rq_list and set bit0 in pending bitmap as ctx1->index_hw is | |
2187 | * still zero. | |
2188 | * | |
2189 | * And then while running hw queue, flush_busy_ctxs() finds bit0 is | |
2190 | * set in pending bitmap and tries to retrieve requests in | |
2191 | * hctx->ctxs[0]->rq_list. But htx->ctxs[0] is a pointer to ctx0, | |
2192 | * so the request in ctx1->rq_list is ignored. | |
320ae51f | 2193 | */ |
5778322e AM |
2194 | switch (action & ~CPU_TASKS_FROZEN) { |
2195 | case CPU_DEAD: | |
2196 | case CPU_UP_CANCELED: | |
2197 | cpumask_copy(&online_new, cpu_online_mask); | |
2198 | break; | |
2199 | case CPU_UP_PREPARE: | |
2200 | cpumask_copy(&online_new, cpu_online_mask); | |
2201 | cpumask_set_cpu(cpu, &online_new); | |
2202 | break; | |
2203 | default: | |
320ae51f | 2204 | return NOTIFY_OK; |
5778322e | 2205 | } |
320ae51f JA |
2206 | |
2207 | mutex_lock(&all_q_mutex); | |
f3af020b TH |
2208 | |
2209 | /* | |
2210 | * We need to freeze and reinit all existing queues. Freezing | |
2211 | * involves synchronous wait for an RCU grace period and doing it | |
2212 | * one by one may take a long time. Start freezing all queues in | |
2213 | * one swoop and then wait for the completions so that freezing can | |
2214 | * take place in parallel. | |
2215 | */ | |
2216 | list_for_each_entry(q, &all_q_list, all_q_node) | |
2217 | blk_mq_freeze_queue_start(q); | |
f054b56c | 2218 | list_for_each_entry(q, &all_q_list, all_q_node) { |
f3af020b TH |
2219 | blk_mq_freeze_queue_wait(q); |
2220 | ||
f054b56c ML |
2221 | /* |
2222 | * timeout handler can't touch hw queue during the | |
2223 | * reinitialization | |
2224 | */ | |
2225 | del_timer_sync(&q->timeout); | |
2226 | } | |
2227 | ||
320ae51f | 2228 | list_for_each_entry(q, &all_q_list, all_q_node) |
5778322e | 2229 | blk_mq_queue_reinit(q, &online_new); |
f3af020b TH |
2230 | |
2231 | list_for_each_entry(q, &all_q_list, all_q_node) | |
2232 | blk_mq_unfreeze_queue(q); | |
2233 | ||
320ae51f JA |
2234 | mutex_unlock(&all_q_mutex); |
2235 | return NOTIFY_OK; | |
2236 | } | |
2237 | ||
a5164405 JA |
2238 | static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set) |
2239 | { | |
2240 | int i; | |
2241 | ||
2242 | for (i = 0; i < set->nr_hw_queues; i++) { | |
2243 | set->tags[i] = blk_mq_init_rq_map(set, i); | |
2244 | if (!set->tags[i]) | |
2245 | goto out_unwind; | |
2246 | } | |
2247 | ||
2248 | return 0; | |
2249 | ||
2250 | out_unwind: | |
2251 | while (--i >= 0) | |
2252 | blk_mq_free_rq_map(set, set->tags[i], i); | |
2253 | ||
a5164405 JA |
2254 | return -ENOMEM; |
2255 | } | |
2256 | ||
2257 | /* | |
2258 | * Allocate the request maps associated with this tag_set. Note that this | |
2259 | * may reduce the depth asked for, if memory is tight. set->queue_depth | |
2260 | * will be updated to reflect the allocated depth. | |
2261 | */ | |
2262 | static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set) | |
2263 | { | |
2264 | unsigned int depth; | |
2265 | int err; | |
2266 | ||
2267 | depth = set->queue_depth; | |
2268 | do { | |
2269 | err = __blk_mq_alloc_rq_maps(set); | |
2270 | if (!err) | |
2271 | break; | |
2272 | ||
2273 | set->queue_depth >>= 1; | |
2274 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) { | |
2275 | err = -ENOMEM; | |
2276 | break; | |
2277 | } | |
2278 | } while (set->queue_depth); | |
2279 | ||
2280 | if (!set->queue_depth || err) { | |
2281 | pr_err("blk-mq: failed to allocate request map\n"); | |
2282 | return -ENOMEM; | |
2283 | } | |
2284 | ||
2285 | if (depth != set->queue_depth) | |
2286 | pr_info("blk-mq: reduced tag depth (%u -> %u)\n", | |
2287 | depth, set->queue_depth); | |
2288 | ||
2289 | return 0; | |
2290 | } | |
2291 | ||
f26cdc85 KB |
2292 | struct cpumask *blk_mq_tags_cpumask(struct blk_mq_tags *tags) |
2293 | { | |
2294 | return tags->cpumask; | |
2295 | } | |
2296 | EXPORT_SYMBOL_GPL(blk_mq_tags_cpumask); | |
2297 | ||
a4391c64 JA |
2298 | /* |
2299 | * Alloc a tag set to be associated with one or more request queues. | |
2300 | * May fail with EINVAL for various error conditions. May adjust the | |
2301 | * requested depth down, if if it too large. In that case, the set | |
2302 | * value will be stored in set->queue_depth. | |
2303 | */ | |
24d2f903 CH |
2304 | int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set) |
2305 | { | |
205fb5f5 BVA |
2306 | BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS); |
2307 | ||
24d2f903 CH |
2308 | if (!set->nr_hw_queues) |
2309 | return -EINVAL; | |
a4391c64 | 2310 | if (!set->queue_depth) |
24d2f903 CH |
2311 | return -EINVAL; |
2312 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) | |
2313 | return -EINVAL; | |
2314 | ||
f9018ac9 | 2315 | if (!set->ops->queue_rq || !set->ops->map_queue) |
24d2f903 CH |
2316 | return -EINVAL; |
2317 | ||
a4391c64 JA |
2318 | if (set->queue_depth > BLK_MQ_MAX_DEPTH) { |
2319 | pr_info("blk-mq: reduced tag depth to %u\n", | |
2320 | BLK_MQ_MAX_DEPTH); | |
2321 | set->queue_depth = BLK_MQ_MAX_DEPTH; | |
2322 | } | |
24d2f903 | 2323 | |
6637fadf SL |
2324 | /* |
2325 | * If a crashdump is active, then we are potentially in a very | |
2326 | * memory constrained environment. Limit us to 1 queue and | |
2327 | * 64 tags to prevent using too much memory. | |
2328 | */ | |
2329 | if (is_kdump_kernel()) { | |
2330 | set->nr_hw_queues = 1; | |
2331 | set->queue_depth = min(64U, set->queue_depth); | |
2332 | } | |
2333 | ||
48479005 ML |
2334 | set->tags = kmalloc_node(set->nr_hw_queues * |
2335 | sizeof(struct blk_mq_tags *), | |
24d2f903 CH |
2336 | GFP_KERNEL, set->numa_node); |
2337 | if (!set->tags) | |
a5164405 | 2338 | return -ENOMEM; |
24d2f903 | 2339 | |
a5164405 JA |
2340 | if (blk_mq_alloc_rq_maps(set)) |
2341 | goto enomem; | |
24d2f903 | 2342 | |
0d2602ca JA |
2343 | mutex_init(&set->tag_list_lock); |
2344 | INIT_LIST_HEAD(&set->tag_list); | |
2345 | ||
24d2f903 | 2346 | return 0; |
a5164405 | 2347 | enomem: |
5676e7b6 RE |
2348 | kfree(set->tags); |
2349 | set->tags = NULL; | |
24d2f903 CH |
2350 | return -ENOMEM; |
2351 | } | |
2352 | EXPORT_SYMBOL(blk_mq_alloc_tag_set); | |
2353 | ||
2354 | void blk_mq_free_tag_set(struct blk_mq_tag_set *set) | |
2355 | { | |
2356 | int i; | |
2357 | ||
484b4061 | 2358 | for (i = 0; i < set->nr_hw_queues; i++) { |
f42d79ab | 2359 | if (set->tags[i]) |
484b4061 JA |
2360 | blk_mq_free_rq_map(set, set->tags[i], i); |
2361 | } | |
2362 | ||
981bd189 | 2363 | kfree(set->tags); |
5676e7b6 | 2364 | set->tags = NULL; |
24d2f903 CH |
2365 | } |
2366 | EXPORT_SYMBOL(blk_mq_free_tag_set); | |
2367 | ||
e3a2b3f9 JA |
2368 | int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr) |
2369 | { | |
2370 | struct blk_mq_tag_set *set = q->tag_set; | |
2371 | struct blk_mq_hw_ctx *hctx; | |
2372 | int i, ret; | |
2373 | ||
2374 | if (!set || nr > set->queue_depth) | |
2375 | return -EINVAL; | |
2376 | ||
2377 | ret = 0; | |
2378 | queue_for_each_hw_ctx(q, hctx, i) { | |
2379 | ret = blk_mq_tag_update_depth(hctx->tags, nr); | |
2380 | if (ret) | |
2381 | break; | |
2382 | } | |
2383 | ||
2384 | if (!ret) | |
2385 | q->nr_requests = nr; | |
2386 | ||
2387 | return ret; | |
2388 | } | |
2389 | ||
676141e4 JA |
2390 | void blk_mq_disable_hotplug(void) |
2391 | { | |
2392 | mutex_lock(&all_q_mutex); | |
2393 | } | |
2394 | ||
2395 | void blk_mq_enable_hotplug(void) | |
2396 | { | |
2397 | mutex_unlock(&all_q_mutex); | |
2398 | } | |
2399 | ||
320ae51f JA |
2400 | static int __init blk_mq_init(void) |
2401 | { | |
320ae51f JA |
2402 | blk_mq_cpu_init(); |
2403 | ||
add703fd | 2404 | hotcpu_notifier(blk_mq_queue_reinit_notify, 0); |
320ae51f JA |
2405 | |
2406 | return 0; | |
2407 | } | |
2408 | subsys_initcall(blk_mq_init); |