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> | |
12 | #include <linux/mm.h> | |
13 | #include <linux/init.h> | |
14 | #include <linux/slab.h> | |
15 | #include <linux/workqueue.h> | |
16 | #include <linux/smp.h> | |
17 | #include <linux/llist.h> | |
18 | #include <linux/list_sort.h> | |
19 | #include <linux/cpu.h> | |
20 | #include <linux/cache.h> | |
21 | #include <linux/sched/sysctl.h> | |
22 | #include <linux/delay.h> | |
aedcd72f | 23 | #include <linux/crash_dump.h> |
320ae51f JA |
24 | |
25 | #include <trace/events/block.h> | |
26 | ||
27 | #include <linux/blk-mq.h> | |
28 | #include "blk.h" | |
29 | #include "blk-mq.h" | |
30 | #include "blk-mq-tag.h" | |
31 | ||
32 | static DEFINE_MUTEX(all_q_mutex); | |
33 | static LIST_HEAD(all_q_list); | |
34 | ||
35 | static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx); | |
36 | ||
320ae51f JA |
37 | /* |
38 | * Check if any of the ctx's have pending work in this hardware queue | |
39 | */ | |
40 | static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx) | |
41 | { | |
42 | unsigned int i; | |
43 | ||
1429d7c9 JA |
44 | for (i = 0; i < hctx->ctx_map.map_size; i++) |
45 | if (hctx->ctx_map.map[i].word) | |
320ae51f JA |
46 | return true; |
47 | ||
48 | return false; | |
49 | } | |
50 | ||
1429d7c9 JA |
51 | static inline struct blk_align_bitmap *get_bm(struct blk_mq_hw_ctx *hctx, |
52 | struct blk_mq_ctx *ctx) | |
53 | { | |
54 | return &hctx->ctx_map.map[ctx->index_hw / hctx->ctx_map.bits_per_word]; | |
55 | } | |
56 | ||
57 | #define CTX_TO_BIT(hctx, ctx) \ | |
58 | ((ctx)->index_hw & ((hctx)->ctx_map.bits_per_word - 1)) | |
59 | ||
320ae51f JA |
60 | /* |
61 | * Mark this ctx as having pending work in this hardware queue | |
62 | */ | |
63 | static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx, | |
64 | struct blk_mq_ctx *ctx) | |
65 | { | |
1429d7c9 JA |
66 | struct blk_align_bitmap *bm = get_bm(hctx, ctx); |
67 | ||
68 | if (!test_bit(CTX_TO_BIT(hctx, ctx), &bm->word)) | |
69 | set_bit(CTX_TO_BIT(hctx, ctx), &bm->word); | |
70 | } | |
71 | ||
72 | static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx, | |
73 | struct blk_mq_ctx *ctx) | |
74 | { | |
75 | struct blk_align_bitmap *bm = get_bm(hctx, ctx); | |
76 | ||
77 | clear_bit(CTX_TO_BIT(hctx, ctx), &bm->word); | |
320ae51f JA |
78 | } |
79 | ||
320ae51f JA |
80 | static int blk_mq_queue_enter(struct request_queue *q) |
81 | { | |
add703fd TH |
82 | while (true) { |
83 | int ret; | |
320ae51f | 84 | |
add703fd TH |
85 | if (percpu_ref_tryget_live(&q->mq_usage_counter)) |
86 | return 0; | |
320ae51f | 87 | |
add703fd TH |
88 | ret = wait_event_interruptible(q->mq_freeze_wq, |
89 | !q->mq_freeze_depth || blk_queue_dying(q)); | |
90 | if (blk_queue_dying(q)) | |
91 | return -ENODEV; | |
92 | if (ret) | |
93 | return ret; | |
94 | } | |
320ae51f JA |
95 | } |
96 | ||
97 | static void blk_mq_queue_exit(struct request_queue *q) | |
98 | { | |
add703fd TH |
99 | percpu_ref_put(&q->mq_usage_counter); |
100 | } | |
101 | ||
102 | static void blk_mq_usage_counter_release(struct percpu_ref *ref) | |
103 | { | |
104 | struct request_queue *q = | |
105 | container_of(ref, struct request_queue, mq_usage_counter); | |
106 | ||
107 | wake_up_all(&q->mq_freeze_wq); | |
320ae51f JA |
108 | } |
109 | ||
b4c6a028 | 110 | void blk_mq_freeze_queue_start(struct request_queue *q) |
43a5e4e2 | 111 | { |
cddd5d17 TH |
112 | bool freeze; |
113 | ||
72d6f02a | 114 | spin_lock_irq(q->queue_lock); |
cddd5d17 | 115 | freeze = !q->mq_freeze_depth++; |
72d6f02a TH |
116 | spin_unlock_irq(q->queue_lock); |
117 | ||
cddd5d17 | 118 | if (freeze) { |
9eca8046 | 119 | percpu_ref_kill(&q->mq_usage_counter); |
cddd5d17 TH |
120 | blk_mq_run_queues(q, false); |
121 | } | |
f3af020b | 122 | } |
b4c6a028 | 123 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_start); |
f3af020b TH |
124 | |
125 | static void blk_mq_freeze_queue_wait(struct request_queue *q) | |
126 | { | |
add703fd | 127 | wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->mq_usage_counter)); |
43a5e4e2 ML |
128 | } |
129 | ||
f3af020b TH |
130 | /* |
131 | * Guarantee no request is in use, so we can change any data structure of | |
132 | * the queue afterward. | |
133 | */ | |
134 | void blk_mq_freeze_queue(struct request_queue *q) | |
135 | { | |
136 | blk_mq_freeze_queue_start(q); | |
137 | blk_mq_freeze_queue_wait(q); | |
138 | } | |
139 | ||
b4c6a028 | 140 | void blk_mq_unfreeze_queue(struct request_queue *q) |
320ae51f | 141 | { |
cddd5d17 | 142 | bool wake; |
320ae51f JA |
143 | |
144 | spin_lock_irq(q->queue_lock); | |
780db207 TH |
145 | wake = !--q->mq_freeze_depth; |
146 | WARN_ON_ONCE(q->mq_freeze_depth < 0); | |
320ae51f | 147 | spin_unlock_irq(q->queue_lock); |
add703fd TH |
148 | if (wake) { |
149 | percpu_ref_reinit(&q->mq_usage_counter); | |
320ae51f | 150 | wake_up_all(&q->mq_freeze_wq); |
add703fd | 151 | } |
320ae51f | 152 | } |
b4c6a028 | 153 | EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue); |
320ae51f | 154 | |
aed3ea94 JA |
155 | void blk_mq_wake_waiters(struct request_queue *q) |
156 | { | |
157 | struct blk_mq_hw_ctx *hctx; | |
158 | unsigned int i; | |
159 | ||
160 | queue_for_each_hw_ctx(q, hctx, i) | |
161 | if (blk_mq_hw_queue_mapped(hctx)) | |
162 | blk_mq_tag_wakeup_all(hctx->tags, true); | |
3fd5940c KB |
163 | |
164 | /* | |
165 | * If we are called because the queue has now been marked as | |
166 | * dying, we need to ensure that processes currently waiting on | |
167 | * the queue are notified as well. | |
168 | */ | |
169 | wake_up_all(&q->mq_freeze_wq); | |
aed3ea94 JA |
170 | } |
171 | ||
320ae51f JA |
172 | bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx) |
173 | { | |
174 | return blk_mq_has_free_tags(hctx->tags); | |
175 | } | |
176 | EXPORT_SYMBOL(blk_mq_can_queue); | |
177 | ||
94eddfbe JA |
178 | static void blk_mq_rq_ctx_init(struct request_queue *q, struct blk_mq_ctx *ctx, |
179 | struct request *rq, unsigned int rw_flags) | |
320ae51f | 180 | { |
94eddfbe JA |
181 | if (blk_queue_io_stat(q)) |
182 | rw_flags |= REQ_IO_STAT; | |
183 | ||
af76e555 CH |
184 | INIT_LIST_HEAD(&rq->queuelist); |
185 | /* csd/requeue_work/fifo_time is initialized before use */ | |
186 | rq->q = q; | |
320ae51f | 187 | rq->mq_ctx = ctx; |
0d2602ca | 188 | rq->cmd_flags |= rw_flags; |
af76e555 CH |
189 | /* do not touch atomic flags, it needs atomic ops against the timer */ |
190 | rq->cpu = -1; | |
af76e555 CH |
191 | INIT_HLIST_NODE(&rq->hash); |
192 | RB_CLEAR_NODE(&rq->rb_node); | |
af76e555 CH |
193 | rq->rq_disk = NULL; |
194 | rq->part = NULL; | |
3ee32372 | 195 | rq->start_time = jiffies; |
af76e555 CH |
196 | #ifdef CONFIG_BLK_CGROUP |
197 | rq->rl = NULL; | |
0fec08b4 | 198 | set_start_time_ns(rq); |
af76e555 CH |
199 | rq->io_start_time_ns = 0; |
200 | #endif | |
201 | rq->nr_phys_segments = 0; | |
202 | #if defined(CONFIG_BLK_DEV_INTEGRITY) | |
203 | rq->nr_integrity_segments = 0; | |
204 | #endif | |
af76e555 CH |
205 | rq->special = NULL; |
206 | /* tag was already set */ | |
207 | rq->errors = 0; | |
af76e555 | 208 | |
6f4a1626 TB |
209 | rq->cmd = rq->__cmd; |
210 | ||
af76e555 CH |
211 | rq->extra_len = 0; |
212 | rq->sense_len = 0; | |
213 | rq->resid_len = 0; | |
214 | rq->sense = NULL; | |
215 | ||
af76e555 | 216 | INIT_LIST_HEAD(&rq->timeout_list); |
f6be4fb4 JA |
217 | rq->timeout = 0; |
218 | ||
af76e555 CH |
219 | rq->end_io = NULL; |
220 | rq->end_io_data = NULL; | |
221 | rq->next_rq = NULL; | |
222 | ||
320ae51f JA |
223 | ctx->rq_dispatched[rw_is_sync(rw_flags)]++; |
224 | } | |
225 | ||
5dee8577 | 226 | static struct request * |
cb96a42c | 227 | __blk_mq_alloc_request(struct blk_mq_alloc_data *data, int rw) |
5dee8577 CH |
228 | { |
229 | struct request *rq; | |
230 | unsigned int tag; | |
231 | ||
cb96a42c | 232 | tag = blk_mq_get_tag(data); |
5dee8577 | 233 | if (tag != BLK_MQ_TAG_FAIL) { |
cb96a42c | 234 | rq = data->hctx->tags->rqs[tag]; |
5dee8577 | 235 | |
cb96a42c | 236 | if (blk_mq_tag_busy(data->hctx)) { |
5dee8577 | 237 | rq->cmd_flags = REQ_MQ_INFLIGHT; |
cb96a42c | 238 | atomic_inc(&data->hctx->nr_active); |
5dee8577 CH |
239 | } |
240 | ||
241 | rq->tag = tag; | |
cb96a42c | 242 | blk_mq_rq_ctx_init(data->q, data->ctx, rq, rw); |
5dee8577 CH |
243 | return rq; |
244 | } | |
245 | ||
246 | return NULL; | |
247 | } | |
248 | ||
4ce01dd1 CH |
249 | struct request *blk_mq_alloc_request(struct request_queue *q, int rw, gfp_t gfp, |
250 | bool reserved) | |
320ae51f | 251 | { |
d852564f CH |
252 | struct blk_mq_ctx *ctx; |
253 | struct blk_mq_hw_ctx *hctx; | |
320ae51f | 254 | struct request *rq; |
cb96a42c | 255 | struct blk_mq_alloc_data alloc_data; |
a492f075 | 256 | int ret; |
320ae51f | 257 | |
a492f075 JL |
258 | ret = blk_mq_queue_enter(q); |
259 | if (ret) | |
260 | return ERR_PTR(ret); | |
320ae51f | 261 | |
d852564f CH |
262 | ctx = blk_mq_get_ctx(q); |
263 | hctx = q->mq_ops->map_queue(q, ctx->cpu); | |
cb96a42c ML |
264 | blk_mq_set_alloc_data(&alloc_data, q, gfp & ~__GFP_WAIT, |
265 | reserved, ctx, hctx); | |
d852564f | 266 | |
cb96a42c | 267 | rq = __blk_mq_alloc_request(&alloc_data, rw); |
d852564f CH |
268 | if (!rq && (gfp & __GFP_WAIT)) { |
269 | __blk_mq_run_hw_queue(hctx); | |
270 | blk_mq_put_ctx(ctx); | |
271 | ||
272 | ctx = blk_mq_get_ctx(q); | |
273 | hctx = q->mq_ops->map_queue(q, ctx->cpu); | |
cb96a42c ML |
274 | blk_mq_set_alloc_data(&alloc_data, q, gfp, reserved, ctx, |
275 | hctx); | |
276 | rq = __blk_mq_alloc_request(&alloc_data, rw); | |
277 | ctx = alloc_data.ctx; | |
d852564f CH |
278 | } |
279 | blk_mq_put_ctx(ctx); | |
c76541a9 KB |
280 | if (!rq) { |
281 | blk_mq_queue_exit(q); | |
a492f075 | 282 | return ERR_PTR(-EWOULDBLOCK); |
c76541a9 | 283 | } |
320ae51f JA |
284 | return rq; |
285 | } | |
4bb659b1 | 286 | EXPORT_SYMBOL(blk_mq_alloc_request); |
320ae51f | 287 | |
320ae51f JA |
288 | static void __blk_mq_free_request(struct blk_mq_hw_ctx *hctx, |
289 | struct blk_mq_ctx *ctx, struct request *rq) | |
290 | { | |
291 | const int tag = rq->tag; | |
292 | struct request_queue *q = rq->q; | |
293 | ||
0d2602ca JA |
294 | if (rq->cmd_flags & REQ_MQ_INFLIGHT) |
295 | atomic_dec(&hctx->nr_active); | |
683d0e12 | 296 | rq->cmd_flags = 0; |
0d2602ca | 297 | |
af76e555 | 298 | clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags); |
0d2602ca | 299 | blk_mq_put_tag(hctx, tag, &ctx->last_tag); |
320ae51f JA |
300 | blk_mq_queue_exit(q); |
301 | } | |
302 | ||
7c7f2f2b | 303 | void blk_mq_free_hctx_request(struct blk_mq_hw_ctx *hctx, struct request *rq) |
320ae51f JA |
304 | { |
305 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
320ae51f JA |
306 | |
307 | ctx->rq_completed[rq_is_sync(rq)]++; | |
320ae51f | 308 | __blk_mq_free_request(hctx, ctx, rq); |
7c7f2f2b JA |
309 | |
310 | } | |
311 | EXPORT_SYMBOL_GPL(blk_mq_free_hctx_request); | |
312 | ||
313 | void blk_mq_free_request(struct request *rq) | |
314 | { | |
315 | struct blk_mq_hw_ctx *hctx; | |
316 | struct request_queue *q = rq->q; | |
317 | ||
318 | hctx = q->mq_ops->map_queue(q, rq->mq_ctx->cpu); | |
319 | blk_mq_free_hctx_request(hctx, rq); | |
320ae51f | 320 | } |
1a3b595a | 321 | EXPORT_SYMBOL_GPL(blk_mq_free_request); |
320ae51f | 322 | |
c8a446ad | 323 | inline void __blk_mq_end_request(struct request *rq, int error) |
320ae51f | 324 | { |
0d11e6ac ML |
325 | blk_account_io_done(rq); |
326 | ||
91b63639 | 327 | if (rq->end_io) { |
320ae51f | 328 | rq->end_io(rq, error); |
91b63639 CH |
329 | } else { |
330 | if (unlikely(blk_bidi_rq(rq))) | |
331 | blk_mq_free_request(rq->next_rq); | |
320ae51f | 332 | blk_mq_free_request(rq); |
91b63639 | 333 | } |
320ae51f | 334 | } |
c8a446ad | 335 | EXPORT_SYMBOL(__blk_mq_end_request); |
63151a44 | 336 | |
c8a446ad | 337 | void blk_mq_end_request(struct request *rq, int error) |
63151a44 CH |
338 | { |
339 | if (blk_update_request(rq, error, blk_rq_bytes(rq))) | |
340 | BUG(); | |
c8a446ad | 341 | __blk_mq_end_request(rq, error); |
63151a44 | 342 | } |
c8a446ad | 343 | EXPORT_SYMBOL(blk_mq_end_request); |
320ae51f | 344 | |
30a91cb4 | 345 | static void __blk_mq_complete_request_remote(void *data) |
320ae51f | 346 | { |
3d6efbf6 | 347 | struct request *rq = data; |
320ae51f | 348 | |
30a91cb4 | 349 | rq->q->softirq_done_fn(rq); |
320ae51f | 350 | } |
320ae51f | 351 | |
ed851860 | 352 | static void blk_mq_ipi_complete_request(struct request *rq) |
320ae51f JA |
353 | { |
354 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
38535201 | 355 | bool shared = false; |
320ae51f JA |
356 | int cpu; |
357 | ||
38535201 | 358 | if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) { |
30a91cb4 CH |
359 | rq->q->softirq_done_fn(rq); |
360 | return; | |
361 | } | |
320ae51f JA |
362 | |
363 | cpu = get_cpu(); | |
38535201 CH |
364 | if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags)) |
365 | shared = cpus_share_cache(cpu, ctx->cpu); | |
366 | ||
367 | if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) { | |
30a91cb4 | 368 | rq->csd.func = __blk_mq_complete_request_remote; |
3d6efbf6 CH |
369 | rq->csd.info = rq; |
370 | rq->csd.flags = 0; | |
c46fff2a | 371 | smp_call_function_single_async(ctx->cpu, &rq->csd); |
3d6efbf6 | 372 | } else { |
30a91cb4 | 373 | rq->q->softirq_done_fn(rq); |
3d6efbf6 | 374 | } |
320ae51f JA |
375 | put_cpu(); |
376 | } | |
30a91cb4 | 377 | |
ed851860 JA |
378 | void __blk_mq_complete_request(struct request *rq) |
379 | { | |
380 | struct request_queue *q = rq->q; | |
381 | ||
382 | if (!q->softirq_done_fn) | |
c8a446ad | 383 | blk_mq_end_request(rq, rq->errors); |
ed851860 JA |
384 | else |
385 | blk_mq_ipi_complete_request(rq); | |
386 | } | |
387 | ||
30a91cb4 CH |
388 | /** |
389 | * blk_mq_complete_request - end I/O on a request | |
390 | * @rq: the request being processed | |
391 | * | |
392 | * Description: | |
393 | * Ends all I/O on a request. It does not handle partial completions. | |
394 | * The actual completion happens out-of-order, through a IPI handler. | |
395 | **/ | |
396 | void blk_mq_complete_request(struct request *rq) | |
397 | { | |
95f09684 JA |
398 | struct request_queue *q = rq->q; |
399 | ||
400 | if (unlikely(blk_should_fake_timeout(q))) | |
30a91cb4 | 401 | return; |
ed851860 JA |
402 | if (!blk_mark_rq_complete(rq)) |
403 | __blk_mq_complete_request(rq); | |
30a91cb4 CH |
404 | } |
405 | EXPORT_SYMBOL(blk_mq_complete_request); | |
320ae51f | 406 | |
973c0191 KB |
407 | int blk_mq_request_started(struct request *rq) |
408 | { | |
409 | return test_bit(REQ_ATOM_STARTED, &rq->atomic_flags); | |
410 | } | |
411 | EXPORT_SYMBOL_GPL(blk_mq_request_started); | |
412 | ||
e2490073 | 413 | void blk_mq_start_request(struct request *rq) |
320ae51f JA |
414 | { |
415 | struct request_queue *q = rq->q; | |
416 | ||
417 | trace_block_rq_issue(q, rq); | |
418 | ||
742ee69b | 419 | rq->resid_len = blk_rq_bytes(rq); |
91b63639 CH |
420 | if (unlikely(blk_bidi_rq(rq))) |
421 | rq->next_rq->resid_len = blk_rq_bytes(rq->next_rq); | |
742ee69b | 422 | |
2b8393b4 | 423 | blk_add_timer(rq); |
87ee7b11 | 424 | |
538b7534 JA |
425 | /* |
426 | * Ensure that ->deadline is visible before set the started | |
427 | * flag and clear the completed flag. | |
428 | */ | |
429 | smp_mb__before_atomic(); | |
430 | ||
87ee7b11 JA |
431 | /* |
432 | * Mark us as started and clear complete. Complete might have been | |
433 | * set if requeue raced with timeout, which then marked it as | |
434 | * complete. So be sure to clear complete again when we start | |
435 | * the request, otherwise we'll ignore the completion event. | |
436 | */ | |
4b570521 JA |
437 | if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) |
438 | set_bit(REQ_ATOM_STARTED, &rq->atomic_flags); | |
439 | if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags)) | |
440 | clear_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags); | |
49f5baa5 CH |
441 | |
442 | if (q->dma_drain_size && blk_rq_bytes(rq)) { | |
443 | /* | |
444 | * Make sure space for the drain appears. We know we can do | |
445 | * this because max_hw_segments has been adjusted to be one | |
446 | * fewer than the device can handle. | |
447 | */ | |
448 | rq->nr_phys_segments++; | |
449 | } | |
320ae51f | 450 | } |
e2490073 | 451 | EXPORT_SYMBOL(blk_mq_start_request); |
320ae51f | 452 | |
ed0791b2 | 453 | static void __blk_mq_requeue_request(struct request *rq) |
320ae51f JA |
454 | { |
455 | struct request_queue *q = rq->q; | |
456 | ||
457 | trace_block_rq_requeue(q, rq); | |
49f5baa5 | 458 | |
e2490073 CH |
459 | if (test_and_clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) { |
460 | if (q->dma_drain_size && blk_rq_bytes(rq)) | |
461 | rq->nr_phys_segments--; | |
462 | } | |
320ae51f JA |
463 | } |
464 | ||
ed0791b2 CH |
465 | void blk_mq_requeue_request(struct request *rq) |
466 | { | |
ed0791b2 | 467 | __blk_mq_requeue_request(rq); |
ed0791b2 | 468 | |
ed0791b2 | 469 | BUG_ON(blk_queued_rq(rq)); |
6fca6a61 | 470 | blk_mq_add_to_requeue_list(rq, true); |
ed0791b2 CH |
471 | } |
472 | EXPORT_SYMBOL(blk_mq_requeue_request); | |
473 | ||
6fca6a61 CH |
474 | static void blk_mq_requeue_work(struct work_struct *work) |
475 | { | |
476 | struct request_queue *q = | |
477 | container_of(work, struct request_queue, requeue_work); | |
478 | LIST_HEAD(rq_list); | |
479 | struct request *rq, *next; | |
480 | unsigned long flags; | |
481 | ||
482 | spin_lock_irqsave(&q->requeue_lock, flags); | |
483 | list_splice_init(&q->requeue_list, &rq_list); | |
484 | spin_unlock_irqrestore(&q->requeue_lock, flags); | |
485 | ||
486 | list_for_each_entry_safe(rq, next, &rq_list, queuelist) { | |
487 | if (!(rq->cmd_flags & REQ_SOFTBARRIER)) | |
488 | continue; | |
489 | ||
490 | rq->cmd_flags &= ~REQ_SOFTBARRIER; | |
491 | list_del_init(&rq->queuelist); | |
492 | blk_mq_insert_request(rq, true, false, false); | |
493 | } | |
494 | ||
495 | while (!list_empty(&rq_list)) { | |
496 | rq = list_entry(rq_list.next, struct request, queuelist); | |
497 | list_del_init(&rq->queuelist); | |
498 | blk_mq_insert_request(rq, false, false, false); | |
499 | } | |
500 | ||
8b957415 JA |
501 | /* |
502 | * Use the start variant of queue running here, so that running | |
503 | * the requeue work will kick stopped queues. | |
504 | */ | |
505 | blk_mq_start_hw_queues(q); | |
6fca6a61 CH |
506 | } |
507 | ||
508 | void blk_mq_add_to_requeue_list(struct request *rq, bool at_head) | |
509 | { | |
510 | struct request_queue *q = rq->q; | |
511 | unsigned long flags; | |
512 | ||
513 | /* | |
514 | * We abuse this flag that is otherwise used by the I/O scheduler to | |
515 | * request head insertation from the workqueue. | |
516 | */ | |
517 | BUG_ON(rq->cmd_flags & REQ_SOFTBARRIER); | |
518 | ||
519 | spin_lock_irqsave(&q->requeue_lock, flags); | |
520 | if (at_head) { | |
521 | rq->cmd_flags |= REQ_SOFTBARRIER; | |
522 | list_add(&rq->queuelist, &q->requeue_list); | |
523 | } else { | |
524 | list_add_tail(&rq->queuelist, &q->requeue_list); | |
525 | } | |
526 | spin_unlock_irqrestore(&q->requeue_lock, flags); | |
527 | } | |
528 | EXPORT_SYMBOL(blk_mq_add_to_requeue_list); | |
529 | ||
c68ed59f KB |
530 | void blk_mq_cancel_requeue_work(struct request_queue *q) |
531 | { | |
532 | cancel_work_sync(&q->requeue_work); | |
533 | } | |
534 | EXPORT_SYMBOL_GPL(blk_mq_cancel_requeue_work); | |
535 | ||
6fca6a61 CH |
536 | void blk_mq_kick_requeue_list(struct request_queue *q) |
537 | { | |
538 | kblockd_schedule_work(&q->requeue_work); | |
539 | } | |
540 | EXPORT_SYMBOL(blk_mq_kick_requeue_list); | |
541 | ||
1885b24d JA |
542 | void blk_mq_abort_requeue_list(struct request_queue *q) |
543 | { | |
544 | unsigned long flags; | |
545 | LIST_HEAD(rq_list); | |
546 | ||
547 | spin_lock_irqsave(&q->requeue_lock, flags); | |
548 | list_splice_init(&q->requeue_list, &rq_list); | |
549 | spin_unlock_irqrestore(&q->requeue_lock, flags); | |
550 | ||
551 | while (!list_empty(&rq_list)) { | |
552 | struct request *rq; | |
553 | ||
554 | rq = list_first_entry(&rq_list, struct request, queuelist); | |
555 | list_del_init(&rq->queuelist); | |
556 | rq->errors = -EIO; | |
557 | blk_mq_end_request(rq, rq->errors); | |
558 | } | |
559 | } | |
560 | EXPORT_SYMBOL(blk_mq_abort_requeue_list); | |
561 | ||
7c94e1c1 ML |
562 | static inline bool is_flush_request(struct request *rq, |
563 | struct blk_flush_queue *fq, unsigned int tag) | |
24d2f903 | 564 | { |
0e62f51f | 565 | return ((rq->cmd_flags & REQ_FLUSH_SEQ) && |
7c94e1c1 | 566 | fq->flush_rq->tag == tag); |
0e62f51f JA |
567 | } |
568 | ||
569 | struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag) | |
570 | { | |
571 | struct request *rq = tags->rqs[tag]; | |
e97c293c ML |
572 | /* mq_ctx of flush rq is always cloned from the corresponding req */ |
573 | struct blk_flush_queue *fq = blk_get_flush_queue(rq->q, rq->mq_ctx); | |
22302375 | 574 | |
7c94e1c1 | 575 | if (!is_flush_request(rq, fq, tag)) |
0e62f51f | 576 | return rq; |
22302375 | 577 | |
7c94e1c1 | 578 | return fq->flush_rq; |
24d2f903 CH |
579 | } |
580 | EXPORT_SYMBOL(blk_mq_tag_to_rq); | |
581 | ||
320ae51f | 582 | struct blk_mq_timeout_data { |
46f92d42 CH |
583 | unsigned long next; |
584 | unsigned int next_set; | |
320ae51f JA |
585 | }; |
586 | ||
90415837 | 587 | void blk_mq_rq_timed_out(struct request *req, bool reserved) |
320ae51f | 588 | { |
46f92d42 CH |
589 | struct blk_mq_ops *ops = req->q->mq_ops; |
590 | enum blk_eh_timer_return ret = BLK_EH_RESET_TIMER; | |
87ee7b11 JA |
591 | |
592 | /* | |
593 | * We know that complete is set at this point. If STARTED isn't set | |
594 | * anymore, then the request isn't active and the "timeout" should | |
595 | * just be ignored. This can happen due to the bitflag ordering. | |
596 | * Timeout first checks if STARTED is set, and if it is, assumes | |
597 | * the request is active. But if we race with completion, then | |
598 | * we both flags will get cleared. So check here again, and ignore | |
599 | * a timeout event with a request that isn't active. | |
600 | */ | |
46f92d42 CH |
601 | if (!test_bit(REQ_ATOM_STARTED, &req->atomic_flags)) |
602 | return; | |
87ee7b11 | 603 | |
46f92d42 | 604 | if (ops->timeout) |
0152fb6b | 605 | ret = ops->timeout(req, reserved); |
46f92d42 CH |
606 | |
607 | switch (ret) { | |
608 | case BLK_EH_HANDLED: | |
609 | __blk_mq_complete_request(req); | |
610 | break; | |
611 | case BLK_EH_RESET_TIMER: | |
612 | blk_add_timer(req); | |
613 | blk_clear_rq_complete(req); | |
614 | break; | |
615 | case BLK_EH_NOT_HANDLED: | |
616 | break; | |
617 | default: | |
618 | printk(KERN_ERR "block: bad eh return: %d\n", ret); | |
619 | break; | |
620 | } | |
87ee7b11 | 621 | } |
81481eb4 | 622 | |
81481eb4 CH |
623 | static void blk_mq_check_expired(struct blk_mq_hw_ctx *hctx, |
624 | struct request *rq, void *priv, bool reserved) | |
625 | { | |
626 | struct blk_mq_timeout_data *data = priv; | |
87ee7b11 | 627 | |
46f92d42 CH |
628 | if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) |
629 | return; | |
87ee7b11 | 630 | |
46f92d42 CH |
631 | if (time_after_eq(jiffies, rq->deadline)) { |
632 | if (!blk_mark_rq_complete(rq)) | |
0152fb6b | 633 | blk_mq_rq_timed_out(rq, reserved); |
46f92d42 CH |
634 | } else if (!data->next_set || time_after(data->next, rq->deadline)) { |
635 | data->next = rq->deadline; | |
636 | data->next_set = 1; | |
637 | } | |
87ee7b11 JA |
638 | } |
639 | ||
81481eb4 | 640 | static void blk_mq_rq_timer(unsigned long priv) |
320ae51f | 641 | { |
81481eb4 CH |
642 | struct request_queue *q = (struct request_queue *)priv; |
643 | struct blk_mq_timeout_data data = { | |
644 | .next = 0, | |
645 | .next_set = 0, | |
646 | }; | |
320ae51f | 647 | struct blk_mq_hw_ctx *hctx; |
81481eb4 | 648 | int i; |
320ae51f | 649 | |
484b4061 JA |
650 | queue_for_each_hw_ctx(q, hctx, i) { |
651 | /* | |
652 | * If not software queues are currently mapped to this | |
653 | * hardware queue, there's nothing to check | |
654 | */ | |
19c66e59 | 655 | if (!blk_mq_hw_queue_mapped(hctx)) |
484b4061 JA |
656 | continue; |
657 | ||
81481eb4 | 658 | blk_mq_tag_busy_iter(hctx, blk_mq_check_expired, &data); |
484b4061 | 659 | } |
320ae51f | 660 | |
81481eb4 CH |
661 | if (data.next_set) { |
662 | data.next = blk_rq_timeout(round_jiffies_up(data.next)); | |
663 | mod_timer(&q->timeout, data.next); | |
0d2602ca JA |
664 | } else { |
665 | queue_for_each_hw_ctx(q, hctx, i) | |
666 | blk_mq_tag_idle(hctx); | |
667 | } | |
320ae51f JA |
668 | } |
669 | ||
670 | /* | |
671 | * Reverse check our software queue for entries that we could potentially | |
672 | * merge with. Currently includes a hand-wavy stop count of 8, to not spend | |
673 | * too much time checking for merges. | |
674 | */ | |
675 | static bool blk_mq_attempt_merge(struct request_queue *q, | |
676 | struct blk_mq_ctx *ctx, struct bio *bio) | |
677 | { | |
678 | struct request *rq; | |
679 | int checked = 8; | |
680 | ||
681 | list_for_each_entry_reverse(rq, &ctx->rq_list, queuelist) { | |
682 | int el_ret; | |
683 | ||
684 | if (!checked--) | |
685 | break; | |
686 | ||
687 | if (!blk_rq_merge_ok(rq, bio)) | |
688 | continue; | |
689 | ||
690 | el_ret = blk_try_merge(rq, bio); | |
691 | if (el_ret == ELEVATOR_BACK_MERGE) { | |
692 | if (bio_attempt_back_merge(q, rq, bio)) { | |
693 | ctx->rq_merged++; | |
694 | return true; | |
695 | } | |
696 | break; | |
697 | } else if (el_ret == ELEVATOR_FRONT_MERGE) { | |
698 | if (bio_attempt_front_merge(q, rq, bio)) { | |
699 | ctx->rq_merged++; | |
700 | return true; | |
701 | } | |
702 | break; | |
703 | } | |
704 | } | |
705 | ||
706 | return false; | |
707 | } | |
708 | ||
1429d7c9 JA |
709 | /* |
710 | * Process software queues that have been marked busy, splicing them | |
711 | * to the for-dispatch | |
712 | */ | |
713 | static void flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list) | |
714 | { | |
715 | struct blk_mq_ctx *ctx; | |
716 | int i; | |
717 | ||
718 | for (i = 0; i < hctx->ctx_map.map_size; i++) { | |
719 | struct blk_align_bitmap *bm = &hctx->ctx_map.map[i]; | |
720 | unsigned int off, bit; | |
721 | ||
722 | if (!bm->word) | |
723 | continue; | |
724 | ||
725 | bit = 0; | |
726 | off = i * hctx->ctx_map.bits_per_word; | |
727 | do { | |
728 | bit = find_next_bit(&bm->word, bm->depth, bit); | |
729 | if (bit >= bm->depth) | |
730 | break; | |
731 | ||
732 | ctx = hctx->ctxs[bit + off]; | |
733 | clear_bit(bit, &bm->word); | |
734 | spin_lock(&ctx->lock); | |
735 | list_splice_tail_init(&ctx->rq_list, list); | |
736 | spin_unlock(&ctx->lock); | |
737 | ||
738 | bit++; | |
739 | } while (1); | |
740 | } | |
741 | } | |
742 | ||
320ae51f JA |
743 | /* |
744 | * Run this hardware queue, pulling any software queues mapped to it in. | |
745 | * Note that this function currently has various problems around ordering | |
746 | * of IO. In particular, we'd like FIFO behaviour on handling existing | |
747 | * items on the hctx->dispatch list. Ignore that for now. | |
748 | */ | |
749 | static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx) | |
750 | { | |
751 | struct request_queue *q = hctx->queue; | |
320ae51f JA |
752 | struct request *rq; |
753 | LIST_HEAD(rq_list); | |
74c45052 JA |
754 | LIST_HEAD(driver_list); |
755 | struct list_head *dptr; | |
1429d7c9 | 756 | int queued; |
320ae51f | 757 | |
fd1270d5 | 758 | WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask)); |
e4043dcf | 759 | |
5d12f905 | 760 | if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state))) |
320ae51f JA |
761 | return; |
762 | ||
763 | hctx->run++; | |
764 | ||
765 | /* | |
766 | * Touch any software queue that has pending entries. | |
767 | */ | |
1429d7c9 | 768 | flush_busy_ctxs(hctx, &rq_list); |
320ae51f JA |
769 | |
770 | /* | |
771 | * If we have previous entries on our dispatch list, grab them | |
772 | * and stuff them at the front for more fair dispatch. | |
773 | */ | |
774 | if (!list_empty_careful(&hctx->dispatch)) { | |
775 | spin_lock(&hctx->lock); | |
776 | if (!list_empty(&hctx->dispatch)) | |
777 | list_splice_init(&hctx->dispatch, &rq_list); | |
778 | spin_unlock(&hctx->lock); | |
779 | } | |
780 | ||
74c45052 JA |
781 | /* |
782 | * Start off with dptr being NULL, so we start the first request | |
783 | * immediately, even if we have more pending. | |
784 | */ | |
785 | dptr = NULL; | |
786 | ||
320ae51f JA |
787 | /* |
788 | * Now process all the entries, sending them to the driver. | |
789 | */ | |
1429d7c9 | 790 | queued = 0; |
320ae51f | 791 | while (!list_empty(&rq_list)) { |
74c45052 | 792 | struct blk_mq_queue_data bd; |
320ae51f JA |
793 | int ret; |
794 | ||
795 | rq = list_first_entry(&rq_list, struct request, queuelist); | |
796 | list_del_init(&rq->queuelist); | |
320ae51f | 797 | |
74c45052 JA |
798 | bd.rq = rq; |
799 | bd.list = dptr; | |
800 | bd.last = list_empty(&rq_list); | |
801 | ||
802 | ret = q->mq_ops->queue_rq(hctx, &bd); | |
320ae51f JA |
803 | switch (ret) { |
804 | case BLK_MQ_RQ_QUEUE_OK: | |
805 | queued++; | |
806 | continue; | |
807 | case BLK_MQ_RQ_QUEUE_BUSY: | |
320ae51f | 808 | list_add(&rq->queuelist, &rq_list); |
ed0791b2 | 809 | __blk_mq_requeue_request(rq); |
320ae51f JA |
810 | break; |
811 | default: | |
812 | pr_err("blk-mq: bad return on queue: %d\n", ret); | |
320ae51f | 813 | case BLK_MQ_RQ_QUEUE_ERROR: |
1e93b8c2 | 814 | rq->errors = -EIO; |
c8a446ad | 815 | blk_mq_end_request(rq, rq->errors); |
320ae51f JA |
816 | break; |
817 | } | |
818 | ||
819 | if (ret == BLK_MQ_RQ_QUEUE_BUSY) | |
820 | break; | |
74c45052 JA |
821 | |
822 | /* | |
823 | * We've done the first request. If we have more than 1 | |
824 | * left in the list, set dptr to defer issue. | |
825 | */ | |
826 | if (!dptr && rq_list.next != rq_list.prev) | |
827 | dptr = &driver_list; | |
320ae51f JA |
828 | } |
829 | ||
830 | if (!queued) | |
831 | hctx->dispatched[0]++; | |
832 | else if (queued < (1 << (BLK_MQ_MAX_DISPATCH_ORDER - 1))) | |
833 | hctx->dispatched[ilog2(queued) + 1]++; | |
834 | ||
835 | /* | |
836 | * Any items that need requeuing? Stuff them into hctx->dispatch, | |
837 | * that is where we will continue on next queue run. | |
838 | */ | |
839 | if (!list_empty(&rq_list)) { | |
840 | spin_lock(&hctx->lock); | |
841 | list_splice(&rq_list, &hctx->dispatch); | |
842 | spin_unlock(&hctx->lock); | |
843 | } | |
844 | } | |
845 | ||
506e931f JA |
846 | /* |
847 | * It'd be great if the workqueue API had a way to pass | |
848 | * in a mask and had some smarts for more clever placement. | |
849 | * For now we just round-robin here, switching for every | |
850 | * BLK_MQ_CPU_WORK_BATCH queued items. | |
851 | */ | |
852 | static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx) | |
853 | { | |
b657d7e6 CH |
854 | if (hctx->queue->nr_hw_queues == 1) |
855 | return WORK_CPU_UNBOUND; | |
506e931f JA |
856 | |
857 | if (--hctx->next_cpu_batch <= 0) { | |
b657d7e6 | 858 | int cpu = hctx->next_cpu, next_cpu; |
506e931f JA |
859 | |
860 | next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask); | |
861 | if (next_cpu >= nr_cpu_ids) | |
862 | next_cpu = cpumask_first(hctx->cpumask); | |
863 | ||
864 | hctx->next_cpu = next_cpu; | |
865 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; | |
b657d7e6 CH |
866 | |
867 | return cpu; | |
506e931f JA |
868 | } |
869 | ||
b657d7e6 | 870 | return hctx->next_cpu; |
506e931f JA |
871 | } |
872 | ||
320ae51f JA |
873 | void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) |
874 | { | |
19c66e59 ML |
875 | if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state) || |
876 | !blk_mq_hw_queue_mapped(hctx))) | |
320ae51f JA |
877 | return; |
878 | ||
398205b8 | 879 | if (!async) { |
2a90d4aa PB |
880 | int cpu = get_cpu(); |
881 | if (cpumask_test_cpu(cpu, hctx->cpumask)) { | |
398205b8 | 882 | __blk_mq_run_hw_queue(hctx); |
2a90d4aa | 883 | put_cpu(); |
398205b8 PB |
884 | return; |
885 | } | |
e4043dcf | 886 | |
2a90d4aa | 887 | put_cpu(); |
e4043dcf | 888 | } |
398205b8 | 889 | |
b657d7e6 CH |
890 | kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx), |
891 | &hctx->run_work, 0); | |
320ae51f JA |
892 | } |
893 | ||
894 | void blk_mq_run_queues(struct request_queue *q, bool async) | |
895 | { | |
896 | struct blk_mq_hw_ctx *hctx; | |
897 | int i; | |
898 | ||
899 | queue_for_each_hw_ctx(q, hctx, i) { | |
900 | if ((!blk_mq_hctx_has_pending(hctx) && | |
901 | list_empty_careful(&hctx->dispatch)) || | |
5d12f905 | 902 | test_bit(BLK_MQ_S_STOPPED, &hctx->state)) |
320ae51f JA |
903 | continue; |
904 | ||
905 | blk_mq_run_hw_queue(hctx, async); | |
906 | } | |
907 | } | |
908 | EXPORT_SYMBOL(blk_mq_run_queues); | |
909 | ||
910 | void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx) | |
911 | { | |
70f4db63 CH |
912 | cancel_delayed_work(&hctx->run_work); |
913 | cancel_delayed_work(&hctx->delay_work); | |
320ae51f JA |
914 | set_bit(BLK_MQ_S_STOPPED, &hctx->state); |
915 | } | |
916 | EXPORT_SYMBOL(blk_mq_stop_hw_queue); | |
917 | ||
280d45f6 CH |
918 | void blk_mq_stop_hw_queues(struct request_queue *q) |
919 | { | |
920 | struct blk_mq_hw_ctx *hctx; | |
921 | int i; | |
922 | ||
923 | queue_for_each_hw_ctx(q, hctx, i) | |
924 | blk_mq_stop_hw_queue(hctx); | |
925 | } | |
926 | EXPORT_SYMBOL(blk_mq_stop_hw_queues); | |
927 | ||
320ae51f JA |
928 | void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx) |
929 | { | |
930 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
e4043dcf | 931 | |
0ffbce80 | 932 | blk_mq_run_hw_queue(hctx, false); |
320ae51f JA |
933 | } |
934 | EXPORT_SYMBOL(blk_mq_start_hw_queue); | |
935 | ||
2f268556 CH |
936 | void blk_mq_start_hw_queues(struct request_queue *q) |
937 | { | |
938 | struct blk_mq_hw_ctx *hctx; | |
939 | int i; | |
940 | ||
941 | queue_for_each_hw_ctx(q, hctx, i) | |
942 | blk_mq_start_hw_queue(hctx); | |
943 | } | |
944 | EXPORT_SYMBOL(blk_mq_start_hw_queues); | |
945 | ||
946 | ||
1b4a3258 | 947 | void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async) |
320ae51f JA |
948 | { |
949 | struct blk_mq_hw_ctx *hctx; | |
950 | int i; | |
951 | ||
952 | queue_for_each_hw_ctx(q, hctx, i) { | |
953 | if (!test_bit(BLK_MQ_S_STOPPED, &hctx->state)) | |
954 | continue; | |
955 | ||
956 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
1b4a3258 | 957 | blk_mq_run_hw_queue(hctx, async); |
320ae51f JA |
958 | } |
959 | } | |
960 | EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues); | |
961 | ||
70f4db63 | 962 | static void blk_mq_run_work_fn(struct work_struct *work) |
320ae51f JA |
963 | { |
964 | struct blk_mq_hw_ctx *hctx; | |
965 | ||
70f4db63 | 966 | hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work); |
e4043dcf | 967 | |
320ae51f JA |
968 | __blk_mq_run_hw_queue(hctx); |
969 | } | |
970 | ||
70f4db63 CH |
971 | static void blk_mq_delay_work_fn(struct work_struct *work) |
972 | { | |
973 | struct blk_mq_hw_ctx *hctx; | |
974 | ||
975 | hctx = container_of(work, struct blk_mq_hw_ctx, delay_work.work); | |
976 | ||
977 | if (test_and_clear_bit(BLK_MQ_S_STOPPED, &hctx->state)) | |
978 | __blk_mq_run_hw_queue(hctx); | |
979 | } | |
980 | ||
981 | void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs) | |
982 | { | |
19c66e59 ML |
983 | if (unlikely(!blk_mq_hw_queue_mapped(hctx))) |
984 | return; | |
70f4db63 | 985 | |
b657d7e6 CH |
986 | kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx), |
987 | &hctx->delay_work, msecs_to_jiffies(msecs)); | |
70f4db63 CH |
988 | } |
989 | EXPORT_SYMBOL(blk_mq_delay_queue); | |
990 | ||
320ae51f | 991 | static void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, |
72a0a36e | 992 | struct request *rq, bool at_head) |
320ae51f JA |
993 | { |
994 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
995 | ||
01b983c9 JA |
996 | trace_block_rq_insert(hctx->queue, rq); |
997 | ||
72a0a36e CH |
998 | if (at_head) |
999 | list_add(&rq->queuelist, &ctx->rq_list); | |
1000 | else | |
1001 | list_add_tail(&rq->queuelist, &ctx->rq_list); | |
4bb659b1 | 1002 | |
320ae51f | 1003 | blk_mq_hctx_mark_pending(hctx, ctx); |
320ae51f JA |
1004 | } |
1005 | ||
eeabc850 CH |
1006 | void blk_mq_insert_request(struct request *rq, bool at_head, bool run_queue, |
1007 | bool async) | |
320ae51f | 1008 | { |
eeabc850 | 1009 | struct request_queue *q = rq->q; |
320ae51f | 1010 | struct blk_mq_hw_ctx *hctx; |
eeabc850 CH |
1011 | struct blk_mq_ctx *ctx = rq->mq_ctx, *current_ctx; |
1012 | ||
1013 | current_ctx = blk_mq_get_ctx(q); | |
1014 | if (!cpu_online(ctx->cpu)) | |
1015 | rq->mq_ctx = ctx = current_ctx; | |
320ae51f | 1016 | |
320ae51f JA |
1017 | hctx = q->mq_ops->map_queue(q, ctx->cpu); |
1018 | ||
a57a178a CH |
1019 | spin_lock(&ctx->lock); |
1020 | __blk_mq_insert_request(hctx, rq, at_head); | |
1021 | spin_unlock(&ctx->lock); | |
320ae51f | 1022 | |
320ae51f JA |
1023 | if (run_queue) |
1024 | blk_mq_run_hw_queue(hctx, async); | |
e4043dcf JA |
1025 | |
1026 | blk_mq_put_ctx(current_ctx); | |
320ae51f JA |
1027 | } |
1028 | ||
1029 | static void blk_mq_insert_requests(struct request_queue *q, | |
1030 | struct blk_mq_ctx *ctx, | |
1031 | struct list_head *list, | |
1032 | int depth, | |
1033 | bool from_schedule) | |
1034 | ||
1035 | { | |
1036 | struct blk_mq_hw_ctx *hctx; | |
1037 | struct blk_mq_ctx *current_ctx; | |
1038 | ||
1039 | trace_block_unplug(q, depth, !from_schedule); | |
1040 | ||
1041 | current_ctx = blk_mq_get_ctx(q); | |
1042 | ||
1043 | if (!cpu_online(ctx->cpu)) | |
1044 | ctx = current_ctx; | |
1045 | hctx = q->mq_ops->map_queue(q, ctx->cpu); | |
1046 | ||
1047 | /* | |
1048 | * preemption doesn't flush plug list, so it's possible ctx->cpu is | |
1049 | * offline now | |
1050 | */ | |
1051 | spin_lock(&ctx->lock); | |
1052 | while (!list_empty(list)) { | |
1053 | struct request *rq; | |
1054 | ||
1055 | rq = list_first_entry(list, struct request, queuelist); | |
1056 | list_del_init(&rq->queuelist); | |
1057 | rq->mq_ctx = ctx; | |
72a0a36e | 1058 | __blk_mq_insert_request(hctx, rq, false); |
320ae51f JA |
1059 | } |
1060 | spin_unlock(&ctx->lock); | |
1061 | ||
320ae51f | 1062 | blk_mq_run_hw_queue(hctx, from_schedule); |
e4043dcf | 1063 | blk_mq_put_ctx(current_ctx); |
320ae51f JA |
1064 | } |
1065 | ||
1066 | static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b) | |
1067 | { | |
1068 | struct request *rqa = container_of(a, struct request, queuelist); | |
1069 | struct request *rqb = container_of(b, struct request, queuelist); | |
1070 | ||
1071 | return !(rqa->mq_ctx < rqb->mq_ctx || | |
1072 | (rqa->mq_ctx == rqb->mq_ctx && | |
1073 | blk_rq_pos(rqa) < blk_rq_pos(rqb))); | |
1074 | } | |
1075 | ||
1076 | void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule) | |
1077 | { | |
1078 | struct blk_mq_ctx *this_ctx; | |
1079 | struct request_queue *this_q; | |
1080 | struct request *rq; | |
1081 | LIST_HEAD(list); | |
1082 | LIST_HEAD(ctx_list); | |
1083 | unsigned int depth; | |
1084 | ||
1085 | list_splice_init(&plug->mq_list, &list); | |
1086 | ||
1087 | list_sort(NULL, &list, plug_ctx_cmp); | |
1088 | ||
1089 | this_q = NULL; | |
1090 | this_ctx = NULL; | |
1091 | depth = 0; | |
1092 | ||
1093 | while (!list_empty(&list)) { | |
1094 | rq = list_entry_rq(list.next); | |
1095 | list_del_init(&rq->queuelist); | |
1096 | BUG_ON(!rq->q); | |
1097 | if (rq->mq_ctx != this_ctx) { | |
1098 | if (this_ctx) { | |
1099 | blk_mq_insert_requests(this_q, this_ctx, | |
1100 | &ctx_list, depth, | |
1101 | from_schedule); | |
1102 | } | |
1103 | ||
1104 | this_ctx = rq->mq_ctx; | |
1105 | this_q = rq->q; | |
1106 | depth = 0; | |
1107 | } | |
1108 | ||
1109 | depth++; | |
1110 | list_add_tail(&rq->queuelist, &ctx_list); | |
1111 | } | |
1112 | ||
1113 | /* | |
1114 | * If 'this_ctx' is set, we know we have entries to complete | |
1115 | * on 'ctx_list'. Do those. | |
1116 | */ | |
1117 | if (this_ctx) { | |
1118 | blk_mq_insert_requests(this_q, this_ctx, &ctx_list, depth, | |
1119 | from_schedule); | |
1120 | } | |
1121 | } | |
1122 | ||
1123 | static void blk_mq_bio_to_request(struct request *rq, struct bio *bio) | |
1124 | { | |
1125 | init_request_from_bio(rq, bio); | |
4b570521 | 1126 | |
3ee32372 | 1127 | if (blk_do_io_stat(rq)) |
4b570521 | 1128 | blk_account_io_start(rq, 1); |
320ae51f JA |
1129 | } |
1130 | ||
274a5843 JA |
1131 | static inline bool hctx_allow_merges(struct blk_mq_hw_ctx *hctx) |
1132 | { | |
1133 | return (hctx->flags & BLK_MQ_F_SHOULD_MERGE) && | |
1134 | !blk_queue_nomerges(hctx->queue); | |
1135 | } | |
1136 | ||
07068d5b JA |
1137 | static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx *hctx, |
1138 | struct blk_mq_ctx *ctx, | |
1139 | struct request *rq, struct bio *bio) | |
320ae51f | 1140 | { |
274a5843 | 1141 | if (!hctx_allow_merges(hctx)) { |
07068d5b JA |
1142 | blk_mq_bio_to_request(rq, bio); |
1143 | spin_lock(&ctx->lock); | |
1144 | insert_rq: | |
1145 | __blk_mq_insert_request(hctx, rq, false); | |
1146 | spin_unlock(&ctx->lock); | |
1147 | return false; | |
1148 | } else { | |
274a5843 JA |
1149 | struct request_queue *q = hctx->queue; |
1150 | ||
07068d5b JA |
1151 | spin_lock(&ctx->lock); |
1152 | if (!blk_mq_attempt_merge(q, ctx, bio)) { | |
1153 | blk_mq_bio_to_request(rq, bio); | |
1154 | goto insert_rq; | |
1155 | } | |
320ae51f | 1156 | |
07068d5b JA |
1157 | spin_unlock(&ctx->lock); |
1158 | __blk_mq_free_request(hctx, ctx, rq); | |
1159 | return true; | |
14ec77f3 | 1160 | } |
07068d5b | 1161 | } |
14ec77f3 | 1162 | |
07068d5b JA |
1163 | struct blk_map_ctx { |
1164 | struct blk_mq_hw_ctx *hctx; | |
1165 | struct blk_mq_ctx *ctx; | |
1166 | }; | |
1167 | ||
1168 | static struct request *blk_mq_map_request(struct request_queue *q, | |
1169 | struct bio *bio, | |
1170 | struct blk_map_ctx *data) | |
1171 | { | |
1172 | struct blk_mq_hw_ctx *hctx; | |
1173 | struct blk_mq_ctx *ctx; | |
1174 | struct request *rq; | |
1175 | int rw = bio_data_dir(bio); | |
cb96a42c | 1176 | struct blk_mq_alloc_data alloc_data; |
320ae51f | 1177 | |
07068d5b | 1178 | if (unlikely(blk_mq_queue_enter(q))) { |
320ae51f | 1179 | bio_endio(bio, -EIO); |
07068d5b | 1180 | return NULL; |
320ae51f JA |
1181 | } |
1182 | ||
1183 | ctx = blk_mq_get_ctx(q); | |
1184 | hctx = q->mq_ops->map_queue(q, ctx->cpu); | |
1185 | ||
07068d5b | 1186 | if (rw_is_sync(bio->bi_rw)) |
27fbf4e8 | 1187 | rw |= REQ_SYNC; |
07068d5b | 1188 | |
320ae51f | 1189 | trace_block_getrq(q, bio, rw); |
cb96a42c ML |
1190 | blk_mq_set_alloc_data(&alloc_data, q, GFP_ATOMIC, false, ctx, |
1191 | hctx); | |
1192 | rq = __blk_mq_alloc_request(&alloc_data, rw); | |
5dee8577 | 1193 | if (unlikely(!rq)) { |
793597a6 | 1194 | __blk_mq_run_hw_queue(hctx); |
320ae51f JA |
1195 | blk_mq_put_ctx(ctx); |
1196 | trace_block_sleeprq(q, bio, rw); | |
793597a6 CH |
1197 | |
1198 | ctx = blk_mq_get_ctx(q); | |
320ae51f | 1199 | hctx = q->mq_ops->map_queue(q, ctx->cpu); |
cb96a42c ML |
1200 | blk_mq_set_alloc_data(&alloc_data, q, |
1201 | __GFP_WAIT|GFP_ATOMIC, false, ctx, hctx); | |
1202 | rq = __blk_mq_alloc_request(&alloc_data, rw); | |
1203 | ctx = alloc_data.ctx; | |
1204 | hctx = alloc_data.hctx; | |
320ae51f JA |
1205 | } |
1206 | ||
1207 | hctx->queued++; | |
07068d5b JA |
1208 | data->hctx = hctx; |
1209 | data->ctx = ctx; | |
1210 | return rq; | |
1211 | } | |
1212 | ||
1213 | /* | |
1214 | * Multiple hardware queue variant. This will not use per-process plugs, | |
1215 | * but will attempt to bypass the hctx queueing if we can go straight to | |
1216 | * hardware for SYNC IO. | |
1217 | */ | |
1218 | static void blk_mq_make_request(struct request_queue *q, struct bio *bio) | |
1219 | { | |
1220 | const int is_sync = rw_is_sync(bio->bi_rw); | |
1221 | const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA); | |
1222 | struct blk_map_ctx data; | |
1223 | struct request *rq; | |
1224 | ||
1225 | blk_queue_bounce(q, &bio); | |
1226 | ||
1227 | if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) { | |
1228 | bio_endio(bio, -EIO); | |
1229 | return; | |
1230 | } | |
1231 | ||
1232 | rq = blk_mq_map_request(q, bio, &data); | |
1233 | if (unlikely(!rq)) | |
1234 | return; | |
1235 | ||
1236 | if (unlikely(is_flush_fua)) { | |
1237 | blk_mq_bio_to_request(rq, bio); | |
1238 | blk_insert_flush(rq); | |
1239 | goto run_queue; | |
1240 | } | |
1241 | ||
e167dfb5 JA |
1242 | /* |
1243 | * If the driver supports defer issued based on 'last', then | |
1244 | * queue it up like normal since we can potentially save some | |
1245 | * CPU this way. | |
1246 | */ | |
1247 | if (is_sync && !(data.hctx->flags & BLK_MQ_F_DEFER_ISSUE)) { | |
74c45052 JA |
1248 | struct blk_mq_queue_data bd = { |
1249 | .rq = rq, | |
1250 | .list = NULL, | |
1251 | .last = 1 | |
1252 | }; | |
07068d5b JA |
1253 | int ret; |
1254 | ||
1255 | blk_mq_bio_to_request(rq, bio); | |
07068d5b JA |
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 | */ | |
74c45052 | 1262 | ret = q->mq_ops->queue_rq(data.hctx, &bd); |
07068d5b JA |
1263 | if (ret == BLK_MQ_RQ_QUEUE_OK) |
1264 | goto done; | |
1265 | else { | |
1266 | __blk_mq_requeue_request(rq); | |
1267 | ||
1268 | if (ret == BLK_MQ_RQ_QUEUE_ERROR) { | |
1269 | rq->errors = -EIO; | |
c8a446ad | 1270 | blk_mq_end_request(rq, rq->errors); |
07068d5b JA |
1271 | goto done; |
1272 | } | |
1273 | } | |
1274 | } | |
1275 | ||
1276 | if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) { | |
1277 | /* | |
1278 | * For a SYNC request, send it to the hardware immediately. For | |
1279 | * an ASYNC request, just ensure that we run it later on. The | |
1280 | * latter allows for merging opportunities and more efficient | |
1281 | * dispatching. | |
1282 | */ | |
1283 | run_queue: | |
1284 | blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua); | |
1285 | } | |
1286 | done: | |
1287 | blk_mq_put_ctx(data.ctx); | |
1288 | } | |
1289 | ||
1290 | /* | |
1291 | * Single hardware queue variant. This will attempt to use any per-process | |
1292 | * plug for merging and IO deferral. | |
1293 | */ | |
1294 | static void blk_sq_make_request(struct request_queue *q, struct bio *bio) | |
1295 | { | |
1296 | const int is_sync = rw_is_sync(bio->bi_rw); | |
1297 | const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA); | |
1298 | unsigned int use_plug, request_count = 0; | |
1299 | struct blk_map_ctx data; | |
1300 | struct request *rq; | |
1301 | ||
1302 | /* | |
1303 | * If we have multiple hardware queues, just go directly to | |
1304 | * one of those for sync IO. | |
1305 | */ | |
1306 | use_plug = !is_flush_fua && !is_sync; | |
1307 | ||
1308 | blk_queue_bounce(q, &bio); | |
1309 | ||
1310 | if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) { | |
1311 | bio_endio(bio, -EIO); | |
1312 | return; | |
1313 | } | |
1314 | ||
1315 | if (use_plug && !blk_queue_nomerges(q) && | |
1316 | blk_attempt_plug_merge(q, bio, &request_count)) | |
1317 | return; | |
1318 | ||
1319 | rq = blk_mq_map_request(q, bio, &data); | |
ff87bcec JA |
1320 | if (unlikely(!rq)) |
1321 | return; | |
320ae51f JA |
1322 | |
1323 | if (unlikely(is_flush_fua)) { | |
1324 | blk_mq_bio_to_request(rq, bio); | |
320ae51f JA |
1325 | blk_insert_flush(rq); |
1326 | goto run_queue; | |
1327 | } | |
1328 | ||
1329 | /* | |
1330 | * A task plug currently exists. Since this is completely lockless, | |
1331 | * utilize that to temporarily store requests until the task is | |
1332 | * either done or scheduled away. | |
1333 | */ | |
1334 | if (use_plug) { | |
1335 | struct blk_plug *plug = current->plug; | |
1336 | ||
1337 | if (plug) { | |
1338 | blk_mq_bio_to_request(rq, bio); | |
92f399c7 | 1339 | if (list_empty(&plug->mq_list)) |
320ae51f JA |
1340 | trace_block_plug(q); |
1341 | else if (request_count >= BLK_MAX_REQUEST_COUNT) { | |
1342 | blk_flush_plug_list(plug, false); | |
1343 | trace_block_plug(q); | |
1344 | } | |
1345 | list_add_tail(&rq->queuelist, &plug->mq_list); | |
07068d5b | 1346 | blk_mq_put_ctx(data.ctx); |
320ae51f JA |
1347 | return; |
1348 | } | |
1349 | } | |
1350 | ||
07068d5b JA |
1351 | if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) { |
1352 | /* | |
1353 | * For a SYNC request, send it to the hardware immediately. For | |
1354 | * an ASYNC request, just ensure that we run it later on. The | |
1355 | * latter allows for merging opportunities and more efficient | |
1356 | * dispatching. | |
1357 | */ | |
1358 | run_queue: | |
1359 | blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua); | |
320ae51f JA |
1360 | } |
1361 | ||
07068d5b | 1362 | blk_mq_put_ctx(data.ctx); |
320ae51f JA |
1363 | } |
1364 | ||
1365 | /* | |
1366 | * Default mapping to a software queue, since we use one per CPU. | |
1367 | */ | |
1368 | struct blk_mq_hw_ctx *blk_mq_map_queue(struct request_queue *q, const int cpu) | |
1369 | { | |
1370 | return q->queue_hw_ctx[q->mq_map[cpu]]; | |
1371 | } | |
1372 | EXPORT_SYMBOL(blk_mq_map_queue); | |
1373 | ||
24d2f903 CH |
1374 | static void blk_mq_free_rq_map(struct blk_mq_tag_set *set, |
1375 | struct blk_mq_tags *tags, unsigned int hctx_idx) | |
95363efd | 1376 | { |
e9b267d9 | 1377 | struct page *page; |
320ae51f | 1378 | |
24d2f903 | 1379 | if (tags->rqs && set->ops->exit_request) { |
e9b267d9 | 1380 | int i; |
320ae51f | 1381 | |
24d2f903 CH |
1382 | for (i = 0; i < tags->nr_tags; i++) { |
1383 | if (!tags->rqs[i]) | |
e9b267d9 | 1384 | continue; |
24d2f903 CH |
1385 | set->ops->exit_request(set->driver_data, tags->rqs[i], |
1386 | hctx_idx, i); | |
a5164405 | 1387 | tags->rqs[i] = NULL; |
e9b267d9 | 1388 | } |
320ae51f | 1389 | } |
320ae51f | 1390 | |
24d2f903 CH |
1391 | while (!list_empty(&tags->page_list)) { |
1392 | page = list_first_entry(&tags->page_list, struct page, lru); | |
6753471c | 1393 | list_del_init(&page->lru); |
320ae51f JA |
1394 | __free_pages(page, page->private); |
1395 | } | |
1396 | ||
24d2f903 | 1397 | kfree(tags->rqs); |
320ae51f | 1398 | |
24d2f903 | 1399 | blk_mq_free_tags(tags); |
320ae51f JA |
1400 | } |
1401 | ||
1402 | static size_t order_to_size(unsigned int order) | |
1403 | { | |
4ca08500 | 1404 | return (size_t)PAGE_SIZE << order; |
320ae51f JA |
1405 | } |
1406 | ||
24d2f903 CH |
1407 | static struct blk_mq_tags *blk_mq_init_rq_map(struct blk_mq_tag_set *set, |
1408 | unsigned int hctx_idx) | |
320ae51f | 1409 | { |
24d2f903 | 1410 | struct blk_mq_tags *tags; |
320ae51f JA |
1411 | unsigned int i, j, entries_per_page, max_order = 4; |
1412 | size_t rq_size, left; | |
1413 | ||
24d2f903 CH |
1414 | tags = blk_mq_init_tags(set->queue_depth, set->reserved_tags, |
1415 | set->numa_node); | |
1416 | if (!tags) | |
1417 | return NULL; | |
320ae51f | 1418 | |
24d2f903 CH |
1419 | INIT_LIST_HEAD(&tags->page_list); |
1420 | ||
a5164405 JA |
1421 | tags->rqs = kzalloc_node(set->queue_depth * sizeof(struct request *), |
1422 | GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY, | |
1423 | set->numa_node); | |
24d2f903 CH |
1424 | if (!tags->rqs) { |
1425 | blk_mq_free_tags(tags); | |
1426 | return NULL; | |
1427 | } | |
320ae51f JA |
1428 | |
1429 | /* | |
1430 | * rq_size is the size of the request plus driver payload, rounded | |
1431 | * to the cacheline size | |
1432 | */ | |
24d2f903 | 1433 | rq_size = round_up(sizeof(struct request) + set->cmd_size, |
320ae51f | 1434 | cache_line_size()); |
24d2f903 | 1435 | left = rq_size * set->queue_depth; |
320ae51f | 1436 | |
24d2f903 | 1437 | for (i = 0; i < set->queue_depth; ) { |
320ae51f JA |
1438 | int this_order = max_order; |
1439 | struct page *page; | |
1440 | int to_do; | |
1441 | void *p; | |
1442 | ||
1443 | while (left < order_to_size(this_order - 1) && this_order) | |
1444 | this_order--; | |
1445 | ||
1446 | do { | |
a5164405 JA |
1447 | page = alloc_pages_node(set->numa_node, |
1448 | GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY, | |
1449 | this_order); | |
320ae51f JA |
1450 | if (page) |
1451 | break; | |
1452 | if (!this_order--) | |
1453 | break; | |
1454 | if (order_to_size(this_order) < rq_size) | |
1455 | break; | |
1456 | } while (1); | |
1457 | ||
1458 | if (!page) | |
24d2f903 | 1459 | goto fail; |
320ae51f JA |
1460 | |
1461 | page->private = this_order; | |
24d2f903 | 1462 | list_add_tail(&page->lru, &tags->page_list); |
320ae51f JA |
1463 | |
1464 | p = page_address(page); | |
1465 | entries_per_page = order_to_size(this_order) / rq_size; | |
24d2f903 | 1466 | to_do = min(entries_per_page, set->queue_depth - i); |
320ae51f JA |
1467 | left -= to_do * rq_size; |
1468 | for (j = 0; j < to_do; j++) { | |
24d2f903 | 1469 | tags->rqs[i] = p; |
683d0e12 DH |
1470 | tags->rqs[i]->atomic_flags = 0; |
1471 | tags->rqs[i]->cmd_flags = 0; | |
24d2f903 CH |
1472 | if (set->ops->init_request) { |
1473 | if (set->ops->init_request(set->driver_data, | |
1474 | tags->rqs[i], hctx_idx, i, | |
a5164405 JA |
1475 | set->numa_node)) { |
1476 | tags->rqs[i] = NULL; | |
24d2f903 | 1477 | goto fail; |
a5164405 | 1478 | } |
e9b267d9 CH |
1479 | } |
1480 | ||
320ae51f JA |
1481 | p += rq_size; |
1482 | i++; | |
1483 | } | |
1484 | } | |
1485 | ||
24d2f903 | 1486 | return tags; |
320ae51f | 1487 | |
24d2f903 | 1488 | fail: |
24d2f903 CH |
1489 | blk_mq_free_rq_map(set, tags, hctx_idx); |
1490 | return NULL; | |
320ae51f JA |
1491 | } |
1492 | ||
1429d7c9 JA |
1493 | static void blk_mq_free_bitmap(struct blk_mq_ctxmap *bitmap) |
1494 | { | |
1495 | kfree(bitmap->map); | |
1496 | } | |
1497 | ||
1498 | static int blk_mq_alloc_bitmap(struct blk_mq_ctxmap *bitmap, int node) | |
1499 | { | |
1500 | unsigned int bpw = 8, total, num_maps, i; | |
1501 | ||
1502 | bitmap->bits_per_word = bpw; | |
1503 | ||
1504 | num_maps = ALIGN(nr_cpu_ids, bpw) / bpw; | |
1505 | bitmap->map = kzalloc_node(num_maps * sizeof(struct blk_align_bitmap), | |
1506 | GFP_KERNEL, node); | |
1507 | if (!bitmap->map) | |
1508 | return -ENOMEM; | |
1509 | ||
1510 | bitmap->map_size = num_maps; | |
1511 | ||
1512 | total = nr_cpu_ids; | |
1513 | for (i = 0; i < num_maps; i++) { | |
1514 | bitmap->map[i].depth = min(total, bitmap->bits_per_word); | |
1515 | total -= bitmap->map[i].depth; | |
1516 | } | |
1517 | ||
1518 | return 0; | |
1519 | } | |
1520 | ||
484b4061 JA |
1521 | static int blk_mq_hctx_cpu_offline(struct blk_mq_hw_ctx *hctx, int cpu) |
1522 | { | |
1523 | struct request_queue *q = hctx->queue; | |
1524 | struct blk_mq_ctx *ctx; | |
1525 | LIST_HEAD(tmp); | |
1526 | ||
1527 | /* | |
1528 | * Move ctx entries to new CPU, if this one is going away. | |
1529 | */ | |
1530 | ctx = __blk_mq_get_ctx(q, cpu); | |
1531 | ||
1532 | spin_lock(&ctx->lock); | |
1533 | if (!list_empty(&ctx->rq_list)) { | |
1534 | list_splice_init(&ctx->rq_list, &tmp); | |
1535 | blk_mq_hctx_clear_pending(hctx, ctx); | |
1536 | } | |
1537 | spin_unlock(&ctx->lock); | |
1538 | ||
1539 | if (list_empty(&tmp)) | |
1540 | return NOTIFY_OK; | |
1541 | ||
1542 | ctx = blk_mq_get_ctx(q); | |
1543 | spin_lock(&ctx->lock); | |
1544 | ||
1545 | while (!list_empty(&tmp)) { | |
1546 | struct request *rq; | |
1547 | ||
1548 | rq = list_first_entry(&tmp, struct request, queuelist); | |
1549 | rq->mq_ctx = ctx; | |
1550 | list_move_tail(&rq->queuelist, &ctx->rq_list); | |
1551 | } | |
1552 | ||
1553 | hctx = q->mq_ops->map_queue(q, ctx->cpu); | |
1554 | blk_mq_hctx_mark_pending(hctx, ctx); | |
1555 | ||
1556 | spin_unlock(&ctx->lock); | |
1557 | ||
1558 | blk_mq_run_hw_queue(hctx, true); | |
1559 | blk_mq_put_ctx(ctx); | |
1560 | return NOTIFY_OK; | |
1561 | } | |
1562 | ||
1563 | static int blk_mq_hctx_cpu_online(struct blk_mq_hw_ctx *hctx, int cpu) | |
1564 | { | |
1565 | struct request_queue *q = hctx->queue; | |
1566 | struct blk_mq_tag_set *set = q->tag_set; | |
1567 | ||
1568 | if (set->tags[hctx->queue_num]) | |
1569 | return NOTIFY_OK; | |
1570 | ||
1571 | set->tags[hctx->queue_num] = blk_mq_init_rq_map(set, hctx->queue_num); | |
1572 | if (!set->tags[hctx->queue_num]) | |
1573 | return NOTIFY_STOP; | |
1574 | ||
1575 | hctx->tags = set->tags[hctx->queue_num]; | |
1576 | return NOTIFY_OK; | |
1577 | } | |
1578 | ||
1579 | static int blk_mq_hctx_notify(void *data, unsigned long action, | |
1580 | unsigned int cpu) | |
1581 | { | |
1582 | struct blk_mq_hw_ctx *hctx = data; | |
1583 | ||
1584 | if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) | |
1585 | return blk_mq_hctx_cpu_offline(hctx, cpu); | |
1586 | else if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) | |
1587 | return blk_mq_hctx_cpu_online(hctx, cpu); | |
1588 | ||
1589 | return NOTIFY_OK; | |
1590 | } | |
1591 | ||
08e98fc6 ML |
1592 | static void blk_mq_exit_hctx(struct request_queue *q, |
1593 | struct blk_mq_tag_set *set, | |
1594 | struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx) | |
1595 | { | |
f70ced09 ML |
1596 | unsigned flush_start_tag = set->queue_depth; |
1597 | ||
08e98fc6 ML |
1598 | blk_mq_tag_idle(hctx); |
1599 | ||
f70ced09 ML |
1600 | if (set->ops->exit_request) |
1601 | set->ops->exit_request(set->driver_data, | |
1602 | hctx->fq->flush_rq, hctx_idx, | |
1603 | flush_start_tag + hctx_idx); | |
1604 | ||
08e98fc6 ML |
1605 | if (set->ops->exit_hctx) |
1606 | set->ops->exit_hctx(hctx, hctx_idx); | |
1607 | ||
1608 | blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier); | |
f70ced09 | 1609 | blk_free_flush_queue(hctx->fq); |
08e98fc6 ML |
1610 | kfree(hctx->ctxs); |
1611 | blk_mq_free_bitmap(&hctx->ctx_map); | |
1612 | } | |
1613 | ||
624dbe47 ML |
1614 | static void blk_mq_exit_hw_queues(struct request_queue *q, |
1615 | struct blk_mq_tag_set *set, int nr_queue) | |
1616 | { | |
1617 | struct blk_mq_hw_ctx *hctx; | |
1618 | unsigned int i; | |
1619 | ||
1620 | queue_for_each_hw_ctx(q, hctx, i) { | |
1621 | if (i == nr_queue) | |
1622 | break; | |
08e98fc6 | 1623 | blk_mq_exit_hctx(q, set, hctx, i); |
624dbe47 | 1624 | } |
624dbe47 ML |
1625 | } |
1626 | ||
1627 | static void blk_mq_free_hw_queues(struct request_queue *q, | |
1628 | struct blk_mq_tag_set *set) | |
1629 | { | |
1630 | struct blk_mq_hw_ctx *hctx; | |
1631 | unsigned int i; | |
1632 | ||
1633 | queue_for_each_hw_ctx(q, hctx, i) { | |
1634 | free_cpumask_var(hctx->cpumask); | |
cdef54dd | 1635 | kfree(hctx); |
624dbe47 ML |
1636 | } |
1637 | } | |
1638 | ||
08e98fc6 ML |
1639 | static int blk_mq_init_hctx(struct request_queue *q, |
1640 | struct blk_mq_tag_set *set, | |
1641 | struct blk_mq_hw_ctx *hctx, unsigned hctx_idx) | |
320ae51f | 1642 | { |
08e98fc6 | 1643 | int node; |
f70ced09 | 1644 | unsigned flush_start_tag = set->queue_depth; |
08e98fc6 ML |
1645 | |
1646 | node = hctx->numa_node; | |
1647 | if (node == NUMA_NO_NODE) | |
1648 | node = hctx->numa_node = set->numa_node; | |
1649 | ||
1650 | INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn); | |
1651 | INIT_DELAYED_WORK(&hctx->delay_work, blk_mq_delay_work_fn); | |
1652 | spin_lock_init(&hctx->lock); | |
1653 | INIT_LIST_HEAD(&hctx->dispatch); | |
1654 | hctx->queue = q; | |
1655 | hctx->queue_num = hctx_idx; | |
1656 | hctx->flags = set->flags; | |
08e98fc6 ML |
1657 | |
1658 | blk_mq_init_cpu_notifier(&hctx->cpu_notifier, | |
1659 | blk_mq_hctx_notify, hctx); | |
1660 | blk_mq_register_cpu_notifier(&hctx->cpu_notifier); | |
1661 | ||
1662 | hctx->tags = set->tags[hctx_idx]; | |
320ae51f JA |
1663 | |
1664 | /* | |
08e98fc6 ML |
1665 | * Allocate space for all possible cpus to avoid allocation at |
1666 | * runtime | |
320ae51f | 1667 | */ |
08e98fc6 ML |
1668 | hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *), |
1669 | GFP_KERNEL, node); | |
1670 | if (!hctx->ctxs) | |
1671 | goto unregister_cpu_notifier; | |
320ae51f | 1672 | |
08e98fc6 ML |
1673 | if (blk_mq_alloc_bitmap(&hctx->ctx_map, node)) |
1674 | goto free_ctxs; | |
320ae51f | 1675 | |
08e98fc6 | 1676 | hctx->nr_ctx = 0; |
320ae51f | 1677 | |
08e98fc6 ML |
1678 | if (set->ops->init_hctx && |
1679 | set->ops->init_hctx(hctx, set->driver_data, hctx_idx)) | |
1680 | goto free_bitmap; | |
320ae51f | 1681 | |
f70ced09 ML |
1682 | hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size); |
1683 | if (!hctx->fq) | |
1684 | goto exit_hctx; | |
320ae51f | 1685 | |
f70ced09 ML |
1686 | if (set->ops->init_request && |
1687 | set->ops->init_request(set->driver_data, | |
1688 | hctx->fq->flush_rq, hctx_idx, | |
1689 | flush_start_tag + hctx_idx, node)) | |
1690 | goto free_fq; | |
320ae51f | 1691 | |
08e98fc6 | 1692 | return 0; |
320ae51f | 1693 | |
f70ced09 ML |
1694 | free_fq: |
1695 | kfree(hctx->fq); | |
1696 | exit_hctx: | |
1697 | if (set->ops->exit_hctx) | |
1698 | set->ops->exit_hctx(hctx, hctx_idx); | |
08e98fc6 ML |
1699 | free_bitmap: |
1700 | blk_mq_free_bitmap(&hctx->ctx_map); | |
1701 | free_ctxs: | |
1702 | kfree(hctx->ctxs); | |
1703 | unregister_cpu_notifier: | |
1704 | blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier); | |
320ae51f | 1705 | |
08e98fc6 ML |
1706 | return -1; |
1707 | } | |
320ae51f | 1708 | |
08e98fc6 ML |
1709 | static int blk_mq_init_hw_queues(struct request_queue *q, |
1710 | struct blk_mq_tag_set *set) | |
1711 | { | |
1712 | struct blk_mq_hw_ctx *hctx; | |
1713 | unsigned int i; | |
320ae51f | 1714 | |
08e98fc6 ML |
1715 | /* |
1716 | * Initialize hardware queues | |
1717 | */ | |
1718 | queue_for_each_hw_ctx(q, hctx, i) { | |
1719 | if (blk_mq_init_hctx(q, set, hctx, i)) | |
320ae51f JA |
1720 | break; |
1721 | } | |
1722 | ||
1723 | if (i == q->nr_hw_queues) | |
1724 | return 0; | |
1725 | ||
1726 | /* | |
1727 | * Init failed | |
1728 | */ | |
624dbe47 | 1729 | blk_mq_exit_hw_queues(q, set, i); |
320ae51f JA |
1730 | |
1731 | return 1; | |
1732 | } | |
1733 | ||
1734 | static void blk_mq_init_cpu_queues(struct request_queue *q, | |
1735 | unsigned int nr_hw_queues) | |
1736 | { | |
1737 | unsigned int i; | |
1738 | ||
1739 | for_each_possible_cpu(i) { | |
1740 | struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i); | |
1741 | struct blk_mq_hw_ctx *hctx; | |
1742 | ||
1743 | memset(__ctx, 0, sizeof(*__ctx)); | |
1744 | __ctx->cpu = i; | |
1745 | spin_lock_init(&__ctx->lock); | |
1746 | INIT_LIST_HEAD(&__ctx->rq_list); | |
1747 | __ctx->queue = q; | |
1748 | ||
1749 | /* If the cpu isn't online, the cpu is mapped to first hctx */ | |
320ae51f JA |
1750 | if (!cpu_online(i)) |
1751 | continue; | |
1752 | ||
e4043dcf JA |
1753 | hctx = q->mq_ops->map_queue(q, i); |
1754 | cpumask_set_cpu(i, hctx->cpumask); | |
1755 | hctx->nr_ctx++; | |
1756 | ||
320ae51f JA |
1757 | /* |
1758 | * Set local node, IFF we have more than one hw queue. If | |
1759 | * not, we remain on the home node of the device | |
1760 | */ | |
1761 | if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE) | |
1762 | hctx->numa_node = cpu_to_node(i); | |
1763 | } | |
1764 | } | |
1765 | ||
1766 | static void blk_mq_map_swqueue(struct request_queue *q) | |
1767 | { | |
1768 | unsigned int i; | |
1769 | struct blk_mq_hw_ctx *hctx; | |
1770 | struct blk_mq_ctx *ctx; | |
1771 | ||
1772 | queue_for_each_hw_ctx(q, hctx, i) { | |
e4043dcf | 1773 | cpumask_clear(hctx->cpumask); |
320ae51f JA |
1774 | hctx->nr_ctx = 0; |
1775 | } | |
1776 | ||
1777 | /* | |
1778 | * Map software to hardware queues | |
1779 | */ | |
1780 | queue_for_each_ctx(q, ctx, i) { | |
1781 | /* If the cpu isn't online, the cpu is mapped to first hctx */ | |
e4043dcf JA |
1782 | if (!cpu_online(i)) |
1783 | continue; | |
1784 | ||
320ae51f | 1785 | hctx = q->mq_ops->map_queue(q, i); |
e4043dcf | 1786 | cpumask_set_cpu(i, hctx->cpumask); |
320ae51f JA |
1787 | ctx->index_hw = hctx->nr_ctx; |
1788 | hctx->ctxs[hctx->nr_ctx++] = ctx; | |
1789 | } | |
506e931f JA |
1790 | |
1791 | queue_for_each_hw_ctx(q, hctx, i) { | |
484b4061 | 1792 | /* |
a68aafa5 JA |
1793 | * If no software queues are mapped to this hardware queue, |
1794 | * disable it and free the request entries. | |
484b4061 JA |
1795 | */ |
1796 | if (!hctx->nr_ctx) { | |
1797 | struct blk_mq_tag_set *set = q->tag_set; | |
1798 | ||
1799 | if (set->tags[i]) { | |
1800 | blk_mq_free_rq_map(set, set->tags[i], i); | |
1801 | set->tags[i] = NULL; | |
1802 | hctx->tags = NULL; | |
1803 | } | |
1804 | continue; | |
1805 | } | |
1806 | ||
1807 | /* | |
1808 | * Initialize batch roundrobin counts | |
1809 | */ | |
506e931f JA |
1810 | hctx->next_cpu = cpumask_first(hctx->cpumask); |
1811 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; | |
1812 | } | |
320ae51f JA |
1813 | } |
1814 | ||
0d2602ca JA |
1815 | static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set) |
1816 | { | |
1817 | struct blk_mq_hw_ctx *hctx; | |
1818 | struct request_queue *q; | |
1819 | bool shared; | |
1820 | int i; | |
1821 | ||
1822 | if (set->tag_list.next == set->tag_list.prev) | |
1823 | shared = false; | |
1824 | else | |
1825 | shared = true; | |
1826 | ||
1827 | list_for_each_entry(q, &set->tag_list, tag_set_list) { | |
1828 | blk_mq_freeze_queue(q); | |
1829 | ||
1830 | queue_for_each_hw_ctx(q, hctx, i) { | |
1831 | if (shared) | |
1832 | hctx->flags |= BLK_MQ_F_TAG_SHARED; | |
1833 | else | |
1834 | hctx->flags &= ~BLK_MQ_F_TAG_SHARED; | |
1835 | } | |
1836 | blk_mq_unfreeze_queue(q); | |
1837 | } | |
1838 | } | |
1839 | ||
1840 | static void blk_mq_del_queue_tag_set(struct request_queue *q) | |
1841 | { | |
1842 | struct blk_mq_tag_set *set = q->tag_set; | |
1843 | ||
0d2602ca JA |
1844 | mutex_lock(&set->tag_list_lock); |
1845 | list_del_init(&q->tag_set_list); | |
1846 | blk_mq_update_tag_set_depth(set); | |
1847 | mutex_unlock(&set->tag_list_lock); | |
0d2602ca JA |
1848 | } |
1849 | ||
1850 | static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set, | |
1851 | struct request_queue *q) | |
1852 | { | |
1853 | q->tag_set = set; | |
1854 | ||
1855 | mutex_lock(&set->tag_list_lock); | |
1856 | list_add_tail(&q->tag_set_list, &set->tag_list); | |
1857 | blk_mq_update_tag_set_depth(set); | |
1858 | mutex_unlock(&set->tag_list_lock); | |
1859 | } | |
1860 | ||
24d2f903 | 1861 | struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set) |
320ae51f JA |
1862 | { |
1863 | struct blk_mq_hw_ctx **hctxs; | |
e6cdb092 | 1864 | struct blk_mq_ctx __percpu *ctx; |
320ae51f | 1865 | struct request_queue *q; |
f14bbe77 | 1866 | unsigned int *map; |
320ae51f JA |
1867 | int i; |
1868 | ||
320ae51f JA |
1869 | ctx = alloc_percpu(struct blk_mq_ctx); |
1870 | if (!ctx) | |
1871 | return ERR_PTR(-ENOMEM); | |
1872 | ||
24d2f903 CH |
1873 | hctxs = kmalloc_node(set->nr_hw_queues * sizeof(*hctxs), GFP_KERNEL, |
1874 | set->numa_node); | |
320ae51f JA |
1875 | |
1876 | if (!hctxs) | |
1877 | goto err_percpu; | |
1878 | ||
f14bbe77 JA |
1879 | map = blk_mq_make_queue_map(set); |
1880 | if (!map) | |
1881 | goto err_map; | |
1882 | ||
24d2f903 | 1883 | for (i = 0; i < set->nr_hw_queues; i++) { |
f14bbe77 JA |
1884 | int node = blk_mq_hw_queue_to_node(map, i); |
1885 | ||
cdef54dd CH |
1886 | hctxs[i] = kzalloc_node(sizeof(struct blk_mq_hw_ctx), |
1887 | GFP_KERNEL, node); | |
320ae51f JA |
1888 | if (!hctxs[i]) |
1889 | goto err_hctxs; | |
1890 | ||
a86073e4 JA |
1891 | if (!zalloc_cpumask_var_node(&hctxs[i]->cpumask, GFP_KERNEL, |
1892 | node)) | |
e4043dcf JA |
1893 | goto err_hctxs; |
1894 | ||
0d2602ca | 1895 | atomic_set(&hctxs[i]->nr_active, 0); |
f14bbe77 | 1896 | hctxs[i]->numa_node = node; |
320ae51f JA |
1897 | hctxs[i]->queue_num = i; |
1898 | } | |
1899 | ||
24d2f903 | 1900 | q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node); |
320ae51f JA |
1901 | if (!q) |
1902 | goto err_hctxs; | |
1903 | ||
17497acb TH |
1904 | /* |
1905 | * Init percpu_ref in atomic mode so that it's faster to shutdown. | |
1906 | * See blk_register_queue() for details. | |
1907 | */ | |
a34375ef | 1908 | if (percpu_ref_init(&q->mq_usage_counter, blk_mq_usage_counter_release, |
17497acb | 1909 | PERCPU_REF_INIT_ATOMIC, GFP_KERNEL)) |
3d2936f4 ML |
1910 | goto err_map; |
1911 | ||
320ae51f JA |
1912 | setup_timer(&q->timeout, blk_mq_rq_timer, (unsigned long) q); |
1913 | blk_queue_rq_timeout(q, 30000); | |
1914 | ||
1915 | q->nr_queues = nr_cpu_ids; | |
24d2f903 | 1916 | q->nr_hw_queues = set->nr_hw_queues; |
f14bbe77 | 1917 | q->mq_map = map; |
320ae51f JA |
1918 | |
1919 | q->queue_ctx = ctx; | |
1920 | q->queue_hw_ctx = hctxs; | |
1921 | ||
24d2f903 | 1922 | q->mq_ops = set->ops; |
94eddfbe | 1923 | q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT; |
320ae51f | 1924 | |
05f1dd53 JA |
1925 | if (!(set->flags & BLK_MQ_F_SG_MERGE)) |
1926 | q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE; | |
1927 | ||
1be036e9 CH |
1928 | q->sg_reserved_size = INT_MAX; |
1929 | ||
6fca6a61 CH |
1930 | INIT_WORK(&q->requeue_work, blk_mq_requeue_work); |
1931 | INIT_LIST_HEAD(&q->requeue_list); | |
1932 | spin_lock_init(&q->requeue_lock); | |
1933 | ||
07068d5b JA |
1934 | if (q->nr_hw_queues > 1) |
1935 | blk_queue_make_request(q, blk_mq_make_request); | |
1936 | else | |
1937 | blk_queue_make_request(q, blk_sq_make_request); | |
1938 | ||
24d2f903 CH |
1939 | if (set->timeout) |
1940 | blk_queue_rq_timeout(q, set->timeout); | |
320ae51f | 1941 | |
eba71768 JA |
1942 | /* |
1943 | * Do this after blk_queue_make_request() overrides it... | |
1944 | */ | |
1945 | q->nr_requests = set->queue_depth; | |
1946 | ||
24d2f903 CH |
1947 | if (set->ops->complete) |
1948 | blk_queue_softirq_done(q, set->ops->complete); | |
30a91cb4 | 1949 | |
24d2f903 | 1950 | blk_mq_init_cpu_queues(q, set->nr_hw_queues); |
320ae51f | 1951 | |
24d2f903 | 1952 | if (blk_mq_init_hw_queues(q, set)) |
1bcb1ead | 1953 | goto err_hw; |
18741986 | 1954 | |
320ae51f JA |
1955 | mutex_lock(&all_q_mutex); |
1956 | list_add_tail(&q->all_q_node, &all_q_list); | |
1957 | mutex_unlock(&all_q_mutex); | |
1958 | ||
0d2602ca JA |
1959 | blk_mq_add_queue_tag_set(set, q); |
1960 | ||
484b4061 JA |
1961 | blk_mq_map_swqueue(q); |
1962 | ||
320ae51f | 1963 | return q; |
18741986 | 1964 | |
320ae51f | 1965 | err_hw: |
320ae51f JA |
1966 | blk_cleanup_queue(q); |
1967 | err_hctxs: | |
f14bbe77 | 1968 | kfree(map); |
24d2f903 | 1969 | for (i = 0; i < set->nr_hw_queues; i++) { |
320ae51f JA |
1970 | if (!hctxs[i]) |
1971 | break; | |
e4043dcf | 1972 | free_cpumask_var(hctxs[i]->cpumask); |
cdef54dd | 1973 | kfree(hctxs[i]); |
320ae51f | 1974 | } |
f14bbe77 | 1975 | err_map: |
320ae51f JA |
1976 | kfree(hctxs); |
1977 | err_percpu: | |
1978 | free_percpu(ctx); | |
1979 | return ERR_PTR(-ENOMEM); | |
1980 | } | |
1981 | EXPORT_SYMBOL(blk_mq_init_queue); | |
1982 | ||
1983 | void blk_mq_free_queue(struct request_queue *q) | |
1984 | { | |
624dbe47 | 1985 | struct blk_mq_tag_set *set = q->tag_set; |
320ae51f | 1986 | |
0d2602ca JA |
1987 | blk_mq_del_queue_tag_set(q); |
1988 | ||
624dbe47 ML |
1989 | blk_mq_exit_hw_queues(q, set, set->nr_hw_queues); |
1990 | blk_mq_free_hw_queues(q, set); | |
320ae51f | 1991 | |
add703fd | 1992 | percpu_ref_exit(&q->mq_usage_counter); |
3d2936f4 | 1993 | |
320ae51f JA |
1994 | free_percpu(q->queue_ctx); |
1995 | kfree(q->queue_hw_ctx); | |
1996 | kfree(q->mq_map); | |
1997 | ||
1998 | q->queue_ctx = NULL; | |
1999 | q->queue_hw_ctx = NULL; | |
2000 | q->mq_map = NULL; | |
2001 | ||
2002 | mutex_lock(&all_q_mutex); | |
2003 | list_del_init(&q->all_q_node); | |
2004 | mutex_unlock(&all_q_mutex); | |
2005 | } | |
320ae51f JA |
2006 | |
2007 | /* Basically redo blk_mq_init_queue with queue frozen */ | |
f618ef7c | 2008 | static void blk_mq_queue_reinit(struct request_queue *q) |
320ae51f | 2009 | { |
f3af020b | 2010 | WARN_ON_ONCE(!q->mq_freeze_depth); |
320ae51f | 2011 | |
67aec14c JA |
2012 | blk_mq_sysfs_unregister(q); |
2013 | ||
320ae51f JA |
2014 | blk_mq_update_queue_map(q->mq_map, q->nr_hw_queues); |
2015 | ||
2016 | /* | |
2017 | * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe | |
2018 | * we should change hctx numa_node according to new topology (this | |
2019 | * involves free and re-allocate memory, worthy doing?) | |
2020 | */ | |
2021 | ||
2022 | blk_mq_map_swqueue(q); | |
2023 | ||
67aec14c | 2024 | blk_mq_sysfs_register(q); |
320ae51f JA |
2025 | } |
2026 | ||
f618ef7c PG |
2027 | static int blk_mq_queue_reinit_notify(struct notifier_block *nb, |
2028 | unsigned long action, void *hcpu) | |
320ae51f JA |
2029 | { |
2030 | struct request_queue *q; | |
2031 | ||
2032 | /* | |
9fccfed8 JA |
2033 | * Before new mappings are established, hotadded cpu might already |
2034 | * start handling requests. This doesn't break anything as we map | |
2035 | * offline CPUs to first hardware queue. We will re-init the queue | |
2036 | * below to get optimal settings. | |
320ae51f JA |
2037 | */ |
2038 | if (action != CPU_DEAD && action != CPU_DEAD_FROZEN && | |
2039 | action != CPU_ONLINE && action != CPU_ONLINE_FROZEN) | |
2040 | return NOTIFY_OK; | |
2041 | ||
2042 | mutex_lock(&all_q_mutex); | |
f3af020b TH |
2043 | |
2044 | /* | |
2045 | * We need to freeze and reinit all existing queues. Freezing | |
2046 | * involves synchronous wait for an RCU grace period and doing it | |
2047 | * one by one may take a long time. Start freezing all queues in | |
2048 | * one swoop and then wait for the completions so that freezing can | |
2049 | * take place in parallel. | |
2050 | */ | |
2051 | list_for_each_entry(q, &all_q_list, all_q_node) | |
2052 | blk_mq_freeze_queue_start(q); | |
2053 | list_for_each_entry(q, &all_q_list, all_q_node) | |
2054 | blk_mq_freeze_queue_wait(q); | |
2055 | ||
320ae51f JA |
2056 | list_for_each_entry(q, &all_q_list, all_q_node) |
2057 | blk_mq_queue_reinit(q); | |
f3af020b TH |
2058 | |
2059 | list_for_each_entry(q, &all_q_list, all_q_node) | |
2060 | blk_mq_unfreeze_queue(q); | |
2061 | ||
320ae51f JA |
2062 | mutex_unlock(&all_q_mutex); |
2063 | return NOTIFY_OK; | |
2064 | } | |
2065 | ||
a5164405 JA |
2066 | static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set) |
2067 | { | |
2068 | int i; | |
2069 | ||
2070 | for (i = 0; i < set->nr_hw_queues; i++) { | |
2071 | set->tags[i] = blk_mq_init_rq_map(set, i); | |
2072 | if (!set->tags[i]) | |
2073 | goto out_unwind; | |
2074 | } | |
2075 | ||
2076 | return 0; | |
2077 | ||
2078 | out_unwind: | |
2079 | while (--i >= 0) | |
2080 | blk_mq_free_rq_map(set, set->tags[i], i); | |
2081 | ||
a5164405 JA |
2082 | return -ENOMEM; |
2083 | } | |
2084 | ||
2085 | /* | |
2086 | * Allocate the request maps associated with this tag_set. Note that this | |
2087 | * may reduce the depth asked for, if memory is tight. set->queue_depth | |
2088 | * will be updated to reflect the allocated depth. | |
2089 | */ | |
2090 | static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set) | |
2091 | { | |
2092 | unsigned int depth; | |
2093 | int err; | |
2094 | ||
2095 | depth = set->queue_depth; | |
2096 | do { | |
2097 | err = __blk_mq_alloc_rq_maps(set); | |
2098 | if (!err) | |
2099 | break; | |
2100 | ||
2101 | set->queue_depth >>= 1; | |
2102 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) { | |
2103 | err = -ENOMEM; | |
2104 | break; | |
2105 | } | |
2106 | } while (set->queue_depth); | |
2107 | ||
2108 | if (!set->queue_depth || err) { | |
2109 | pr_err("blk-mq: failed to allocate request map\n"); | |
2110 | return -ENOMEM; | |
2111 | } | |
2112 | ||
2113 | if (depth != set->queue_depth) | |
2114 | pr_info("blk-mq: reduced tag depth (%u -> %u)\n", | |
2115 | depth, set->queue_depth); | |
2116 | ||
2117 | return 0; | |
2118 | } | |
2119 | ||
a4391c64 JA |
2120 | /* |
2121 | * Alloc a tag set to be associated with one or more request queues. | |
2122 | * May fail with EINVAL for various error conditions. May adjust the | |
2123 | * requested depth down, if if it too large. In that case, the set | |
2124 | * value will be stored in set->queue_depth. | |
2125 | */ | |
24d2f903 CH |
2126 | int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set) |
2127 | { | |
205fb5f5 BVA |
2128 | BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS); |
2129 | ||
24d2f903 CH |
2130 | if (!set->nr_hw_queues) |
2131 | return -EINVAL; | |
a4391c64 | 2132 | if (!set->queue_depth) |
24d2f903 CH |
2133 | return -EINVAL; |
2134 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) | |
2135 | return -EINVAL; | |
2136 | ||
cdef54dd | 2137 | if (!set->nr_hw_queues || !set->ops->queue_rq || !set->ops->map_queue) |
24d2f903 CH |
2138 | return -EINVAL; |
2139 | ||
a4391c64 JA |
2140 | if (set->queue_depth > BLK_MQ_MAX_DEPTH) { |
2141 | pr_info("blk-mq: reduced tag depth to %u\n", | |
2142 | BLK_MQ_MAX_DEPTH); | |
2143 | set->queue_depth = BLK_MQ_MAX_DEPTH; | |
2144 | } | |
24d2f903 | 2145 | |
6637fadf SL |
2146 | /* |
2147 | * If a crashdump is active, then we are potentially in a very | |
2148 | * memory constrained environment. Limit us to 1 queue and | |
2149 | * 64 tags to prevent using too much memory. | |
2150 | */ | |
2151 | if (is_kdump_kernel()) { | |
2152 | set->nr_hw_queues = 1; | |
2153 | set->queue_depth = min(64U, set->queue_depth); | |
2154 | } | |
2155 | ||
48479005 ML |
2156 | set->tags = kmalloc_node(set->nr_hw_queues * |
2157 | sizeof(struct blk_mq_tags *), | |
24d2f903 CH |
2158 | GFP_KERNEL, set->numa_node); |
2159 | if (!set->tags) | |
a5164405 | 2160 | return -ENOMEM; |
24d2f903 | 2161 | |
a5164405 JA |
2162 | if (blk_mq_alloc_rq_maps(set)) |
2163 | goto enomem; | |
24d2f903 | 2164 | |
0d2602ca JA |
2165 | mutex_init(&set->tag_list_lock); |
2166 | INIT_LIST_HEAD(&set->tag_list); | |
2167 | ||
24d2f903 | 2168 | return 0; |
a5164405 | 2169 | enomem: |
5676e7b6 RE |
2170 | kfree(set->tags); |
2171 | set->tags = NULL; | |
24d2f903 CH |
2172 | return -ENOMEM; |
2173 | } | |
2174 | EXPORT_SYMBOL(blk_mq_alloc_tag_set); | |
2175 | ||
2176 | void blk_mq_free_tag_set(struct blk_mq_tag_set *set) | |
2177 | { | |
2178 | int i; | |
2179 | ||
484b4061 JA |
2180 | for (i = 0; i < set->nr_hw_queues; i++) { |
2181 | if (set->tags[i]) | |
2182 | blk_mq_free_rq_map(set, set->tags[i], i); | |
2183 | } | |
2184 | ||
981bd189 | 2185 | kfree(set->tags); |
5676e7b6 | 2186 | set->tags = NULL; |
24d2f903 CH |
2187 | } |
2188 | EXPORT_SYMBOL(blk_mq_free_tag_set); | |
2189 | ||
e3a2b3f9 JA |
2190 | int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr) |
2191 | { | |
2192 | struct blk_mq_tag_set *set = q->tag_set; | |
2193 | struct blk_mq_hw_ctx *hctx; | |
2194 | int i, ret; | |
2195 | ||
2196 | if (!set || nr > set->queue_depth) | |
2197 | return -EINVAL; | |
2198 | ||
2199 | ret = 0; | |
2200 | queue_for_each_hw_ctx(q, hctx, i) { | |
2201 | ret = blk_mq_tag_update_depth(hctx->tags, nr); | |
2202 | if (ret) | |
2203 | break; | |
2204 | } | |
2205 | ||
2206 | if (!ret) | |
2207 | q->nr_requests = nr; | |
2208 | ||
2209 | return ret; | |
2210 | } | |
2211 | ||
676141e4 JA |
2212 | void blk_mq_disable_hotplug(void) |
2213 | { | |
2214 | mutex_lock(&all_q_mutex); | |
2215 | } | |
2216 | ||
2217 | void blk_mq_enable_hotplug(void) | |
2218 | { | |
2219 | mutex_unlock(&all_q_mutex); | |
2220 | } | |
2221 | ||
320ae51f JA |
2222 | static int __init blk_mq_init(void) |
2223 | { | |
320ae51f JA |
2224 | blk_mq_cpu_init(); |
2225 | ||
add703fd | 2226 | hotcpu_notifier(blk_mq_queue_reinit_notify, 0); |
320ae51f JA |
2227 | |
2228 | return 0; | |
2229 | } | |
2230 | subsys_initcall(blk_mq_init); |