]>
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 | ||
72d6f02a TH |
110 | /* |
111 | * Guarantee no request is in use, so we can change any data structure of | |
112 | * the queue afterward. | |
113 | */ | |
114 | void blk_mq_freeze_queue(struct request_queue *q) | |
43a5e4e2 | 115 | { |
cddd5d17 TH |
116 | bool freeze; |
117 | ||
72d6f02a | 118 | spin_lock_irq(q->queue_lock); |
cddd5d17 | 119 | freeze = !q->mq_freeze_depth++; |
72d6f02a TH |
120 | spin_unlock_irq(q->queue_lock); |
121 | ||
cddd5d17 | 122 | if (freeze) { |
9eca8046 | 123 | percpu_ref_kill(&q->mq_usage_counter); |
cddd5d17 TH |
124 | blk_mq_run_queues(q, false); |
125 | } | |
add703fd | 126 | wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->mq_usage_counter)); |
43a5e4e2 ML |
127 | } |
128 | ||
320ae51f JA |
129 | static void blk_mq_unfreeze_queue(struct request_queue *q) |
130 | { | |
cddd5d17 | 131 | bool wake; |
320ae51f JA |
132 | |
133 | spin_lock_irq(q->queue_lock); | |
780db207 TH |
134 | wake = !--q->mq_freeze_depth; |
135 | WARN_ON_ONCE(q->mq_freeze_depth < 0); | |
320ae51f | 136 | spin_unlock_irq(q->queue_lock); |
add703fd TH |
137 | if (wake) { |
138 | percpu_ref_reinit(&q->mq_usage_counter); | |
320ae51f | 139 | wake_up_all(&q->mq_freeze_wq); |
add703fd | 140 | } |
320ae51f JA |
141 | } |
142 | ||
143 | bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx) | |
144 | { | |
145 | return blk_mq_has_free_tags(hctx->tags); | |
146 | } | |
147 | EXPORT_SYMBOL(blk_mq_can_queue); | |
148 | ||
94eddfbe JA |
149 | static void blk_mq_rq_ctx_init(struct request_queue *q, struct blk_mq_ctx *ctx, |
150 | struct request *rq, unsigned int rw_flags) | |
320ae51f | 151 | { |
94eddfbe JA |
152 | if (blk_queue_io_stat(q)) |
153 | rw_flags |= REQ_IO_STAT; | |
154 | ||
af76e555 CH |
155 | INIT_LIST_HEAD(&rq->queuelist); |
156 | /* csd/requeue_work/fifo_time is initialized before use */ | |
157 | rq->q = q; | |
320ae51f | 158 | rq->mq_ctx = ctx; |
0d2602ca | 159 | rq->cmd_flags |= rw_flags; |
af76e555 CH |
160 | /* do not touch atomic flags, it needs atomic ops against the timer */ |
161 | rq->cpu = -1; | |
af76e555 CH |
162 | INIT_HLIST_NODE(&rq->hash); |
163 | RB_CLEAR_NODE(&rq->rb_node); | |
af76e555 CH |
164 | rq->rq_disk = NULL; |
165 | rq->part = NULL; | |
3ee32372 | 166 | rq->start_time = jiffies; |
af76e555 CH |
167 | #ifdef CONFIG_BLK_CGROUP |
168 | rq->rl = NULL; | |
0fec08b4 | 169 | set_start_time_ns(rq); |
af76e555 CH |
170 | rq->io_start_time_ns = 0; |
171 | #endif | |
172 | rq->nr_phys_segments = 0; | |
173 | #if defined(CONFIG_BLK_DEV_INTEGRITY) | |
174 | rq->nr_integrity_segments = 0; | |
175 | #endif | |
af76e555 CH |
176 | rq->special = NULL; |
177 | /* tag was already set */ | |
178 | rq->errors = 0; | |
af76e555 | 179 | |
6f4a1626 TB |
180 | rq->cmd = rq->__cmd; |
181 | ||
af76e555 CH |
182 | rq->extra_len = 0; |
183 | rq->sense_len = 0; | |
184 | rq->resid_len = 0; | |
185 | rq->sense = NULL; | |
186 | ||
af76e555 | 187 | INIT_LIST_HEAD(&rq->timeout_list); |
f6be4fb4 JA |
188 | rq->timeout = 0; |
189 | ||
af76e555 CH |
190 | rq->end_io = NULL; |
191 | rq->end_io_data = NULL; | |
192 | rq->next_rq = NULL; | |
193 | ||
320ae51f JA |
194 | ctx->rq_dispatched[rw_is_sync(rw_flags)]++; |
195 | } | |
196 | ||
5dee8577 | 197 | static struct request * |
cb96a42c | 198 | __blk_mq_alloc_request(struct blk_mq_alloc_data *data, int rw) |
5dee8577 CH |
199 | { |
200 | struct request *rq; | |
201 | unsigned int tag; | |
202 | ||
cb96a42c | 203 | tag = blk_mq_get_tag(data); |
5dee8577 | 204 | if (tag != BLK_MQ_TAG_FAIL) { |
cb96a42c | 205 | rq = data->hctx->tags->rqs[tag]; |
5dee8577 | 206 | |
cb96a42c | 207 | if (blk_mq_tag_busy(data->hctx)) { |
5dee8577 | 208 | rq->cmd_flags = REQ_MQ_INFLIGHT; |
cb96a42c | 209 | atomic_inc(&data->hctx->nr_active); |
5dee8577 CH |
210 | } |
211 | ||
212 | rq->tag = tag; | |
cb96a42c | 213 | blk_mq_rq_ctx_init(data->q, data->ctx, rq, rw); |
5dee8577 CH |
214 | return rq; |
215 | } | |
216 | ||
217 | return NULL; | |
218 | } | |
219 | ||
4ce01dd1 CH |
220 | struct request *blk_mq_alloc_request(struct request_queue *q, int rw, gfp_t gfp, |
221 | bool reserved) | |
320ae51f | 222 | { |
d852564f CH |
223 | struct blk_mq_ctx *ctx; |
224 | struct blk_mq_hw_ctx *hctx; | |
320ae51f | 225 | struct request *rq; |
cb96a42c | 226 | struct blk_mq_alloc_data alloc_data; |
a492f075 | 227 | int ret; |
320ae51f | 228 | |
a492f075 JL |
229 | ret = blk_mq_queue_enter(q); |
230 | if (ret) | |
231 | return ERR_PTR(ret); | |
320ae51f | 232 | |
d852564f CH |
233 | ctx = blk_mq_get_ctx(q); |
234 | hctx = q->mq_ops->map_queue(q, ctx->cpu); | |
cb96a42c ML |
235 | blk_mq_set_alloc_data(&alloc_data, q, gfp & ~__GFP_WAIT, |
236 | reserved, ctx, hctx); | |
d852564f | 237 | |
cb96a42c | 238 | rq = __blk_mq_alloc_request(&alloc_data, rw); |
d852564f CH |
239 | if (!rq && (gfp & __GFP_WAIT)) { |
240 | __blk_mq_run_hw_queue(hctx); | |
241 | blk_mq_put_ctx(ctx); | |
242 | ||
243 | ctx = blk_mq_get_ctx(q); | |
244 | hctx = q->mq_ops->map_queue(q, ctx->cpu); | |
cb96a42c ML |
245 | blk_mq_set_alloc_data(&alloc_data, q, gfp, reserved, ctx, |
246 | hctx); | |
247 | rq = __blk_mq_alloc_request(&alloc_data, rw); | |
248 | ctx = alloc_data.ctx; | |
d852564f CH |
249 | } |
250 | blk_mq_put_ctx(ctx); | |
a492f075 JL |
251 | if (!rq) |
252 | return ERR_PTR(-EWOULDBLOCK); | |
320ae51f JA |
253 | return rq; |
254 | } | |
4bb659b1 | 255 | EXPORT_SYMBOL(blk_mq_alloc_request); |
320ae51f | 256 | |
320ae51f JA |
257 | static void __blk_mq_free_request(struct blk_mq_hw_ctx *hctx, |
258 | struct blk_mq_ctx *ctx, struct request *rq) | |
259 | { | |
260 | const int tag = rq->tag; | |
261 | struct request_queue *q = rq->q; | |
262 | ||
0d2602ca JA |
263 | if (rq->cmd_flags & REQ_MQ_INFLIGHT) |
264 | atomic_dec(&hctx->nr_active); | |
683d0e12 | 265 | rq->cmd_flags = 0; |
0d2602ca | 266 | |
af76e555 | 267 | clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags); |
0d2602ca | 268 | blk_mq_put_tag(hctx, tag, &ctx->last_tag); |
320ae51f JA |
269 | blk_mq_queue_exit(q); |
270 | } | |
271 | ||
272 | void blk_mq_free_request(struct request *rq) | |
273 | { | |
274 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
275 | struct blk_mq_hw_ctx *hctx; | |
276 | struct request_queue *q = rq->q; | |
277 | ||
278 | ctx->rq_completed[rq_is_sync(rq)]++; | |
279 | ||
280 | hctx = q->mq_ops->map_queue(q, ctx->cpu); | |
281 | __blk_mq_free_request(hctx, ctx, rq); | |
282 | } | |
283 | ||
c8a446ad | 284 | inline void __blk_mq_end_request(struct request *rq, int error) |
320ae51f | 285 | { |
0d11e6ac ML |
286 | blk_account_io_done(rq); |
287 | ||
91b63639 | 288 | if (rq->end_io) { |
320ae51f | 289 | rq->end_io(rq, error); |
91b63639 CH |
290 | } else { |
291 | if (unlikely(blk_bidi_rq(rq))) | |
292 | blk_mq_free_request(rq->next_rq); | |
320ae51f | 293 | blk_mq_free_request(rq); |
91b63639 | 294 | } |
320ae51f | 295 | } |
c8a446ad | 296 | EXPORT_SYMBOL(__blk_mq_end_request); |
63151a44 | 297 | |
c8a446ad | 298 | void blk_mq_end_request(struct request *rq, int error) |
63151a44 CH |
299 | { |
300 | if (blk_update_request(rq, error, blk_rq_bytes(rq))) | |
301 | BUG(); | |
c8a446ad | 302 | __blk_mq_end_request(rq, error); |
63151a44 | 303 | } |
c8a446ad | 304 | EXPORT_SYMBOL(blk_mq_end_request); |
320ae51f | 305 | |
30a91cb4 | 306 | static void __blk_mq_complete_request_remote(void *data) |
320ae51f | 307 | { |
3d6efbf6 | 308 | struct request *rq = data; |
320ae51f | 309 | |
30a91cb4 | 310 | rq->q->softirq_done_fn(rq); |
320ae51f | 311 | } |
320ae51f | 312 | |
ed851860 | 313 | static void blk_mq_ipi_complete_request(struct request *rq) |
320ae51f JA |
314 | { |
315 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
38535201 | 316 | bool shared = false; |
320ae51f JA |
317 | int cpu; |
318 | ||
38535201 | 319 | if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) { |
30a91cb4 CH |
320 | rq->q->softirq_done_fn(rq); |
321 | return; | |
322 | } | |
320ae51f JA |
323 | |
324 | cpu = get_cpu(); | |
38535201 CH |
325 | if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags)) |
326 | shared = cpus_share_cache(cpu, ctx->cpu); | |
327 | ||
328 | if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) { | |
30a91cb4 | 329 | rq->csd.func = __blk_mq_complete_request_remote; |
3d6efbf6 CH |
330 | rq->csd.info = rq; |
331 | rq->csd.flags = 0; | |
c46fff2a | 332 | smp_call_function_single_async(ctx->cpu, &rq->csd); |
3d6efbf6 | 333 | } else { |
30a91cb4 | 334 | rq->q->softirq_done_fn(rq); |
3d6efbf6 | 335 | } |
320ae51f JA |
336 | put_cpu(); |
337 | } | |
30a91cb4 | 338 | |
ed851860 JA |
339 | void __blk_mq_complete_request(struct request *rq) |
340 | { | |
341 | struct request_queue *q = rq->q; | |
342 | ||
343 | if (!q->softirq_done_fn) | |
c8a446ad | 344 | blk_mq_end_request(rq, rq->errors); |
ed851860 JA |
345 | else |
346 | blk_mq_ipi_complete_request(rq); | |
347 | } | |
348 | ||
30a91cb4 CH |
349 | /** |
350 | * blk_mq_complete_request - end I/O on a request | |
351 | * @rq: the request being processed | |
352 | * | |
353 | * Description: | |
354 | * Ends all I/O on a request. It does not handle partial completions. | |
355 | * The actual completion happens out-of-order, through a IPI handler. | |
356 | **/ | |
357 | void blk_mq_complete_request(struct request *rq) | |
358 | { | |
95f09684 JA |
359 | struct request_queue *q = rq->q; |
360 | ||
361 | if (unlikely(blk_should_fake_timeout(q))) | |
30a91cb4 | 362 | return; |
ed851860 JA |
363 | if (!blk_mark_rq_complete(rq)) |
364 | __blk_mq_complete_request(rq); | |
30a91cb4 CH |
365 | } |
366 | EXPORT_SYMBOL(blk_mq_complete_request); | |
320ae51f | 367 | |
e2490073 | 368 | void blk_mq_start_request(struct request *rq) |
320ae51f JA |
369 | { |
370 | struct request_queue *q = rq->q; | |
371 | ||
372 | trace_block_rq_issue(q, rq); | |
373 | ||
742ee69b | 374 | rq->resid_len = blk_rq_bytes(rq); |
91b63639 CH |
375 | if (unlikely(blk_bidi_rq(rq))) |
376 | rq->next_rq->resid_len = blk_rq_bytes(rq->next_rq); | |
742ee69b | 377 | |
2b8393b4 | 378 | blk_add_timer(rq); |
87ee7b11 | 379 | |
538b7534 JA |
380 | /* |
381 | * Ensure that ->deadline is visible before set the started | |
382 | * flag and clear the completed flag. | |
383 | */ | |
384 | smp_mb__before_atomic(); | |
385 | ||
87ee7b11 JA |
386 | /* |
387 | * Mark us as started and clear complete. Complete might have been | |
388 | * set if requeue raced with timeout, which then marked it as | |
389 | * complete. So be sure to clear complete again when we start | |
390 | * the request, otherwise we'll ignore the completion event. | |
391 | */ | |
4b570521 JA |
392 | if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) |
393 | set_bit(REQ_ATOM_STARTED, &rq->atomic_flags); | |
394 | if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags)) | |
395 | clear_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags); | |
49f5baa5 CH |
396 | |
397 | if (q->dma_drain_size && blk_rq_bytes(rq)) { | |
398 | /* | |
399 | * Make sure space for the drain appears. We know we can do | |
400 | * this because max_hw_segments has been adjusted to be one | |
401 | * fewer than the device can handle. | |
402 | */ | |
403 | rq->nr_phys_segments++; | |
404 | } | |
320ae51f | 405 | } |
e2490073 | 406 | EXPORT_SYMBOL(blk_mq_start_request); |
320ae51f | 407 | |
ed0791b2 | 408 | static void __blk_mq_requeue_request(struct request *rq) |
320ae51f JA |
409 | { |
410 | struct request_queue *q = rq->q; | |
411 | ||
412 | trace_block_rq_requeue(q, rq); | |
49f5baa5 | 413 | |
e2490073 CH |
414 | if (test_and_clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) { |
415 | if (q->dma_drain_size && blk_rq_bytes(rq)) | |
416 | rq->nr_phys_segments--; | |
417 | } | |
320ae51f JA |
418 | } |
419 | ||
ed0791b2 CH |
420 | void blk_mq_requeue_request(struct request *rq) |
421 | { | |
ed0791b2 | 422 | __blk_mq_requeue_request(rq); |
ed0791b2 | 423 | |
ed0791b2 | 424 | BUG_ON(blk_queued_rq(rq)); |
6fca6a61 | 425 | blk_mq_add_to_requeue_list(rq, true); |
ed0791b2 CH |
426 | } |
427 | EXPORT_SYMBOL(blk_mq_requeue_request); | |
428 | ||
6fca6a61 CH |
429 | static void blk_mq_requeue_work(struct work_struct *work) |
430 | { | |
431 | struct request_queue *q = | |
432 | container_of(work, struct request_queue, requeue_work); | |
433 | LIST_HEAD(rq_list); | |
434 | struct request *rq, *next; | |
435 | unsigned long flags; | |
436 | ||
437 | spin_lock_irqsave(&q->requeue_lock, flags); | |
438 | list_splice_init(&q->requeue_list, &rq_list); | |
439 | spin_unlock_irqrestore(&q->requeue_lock, flags); | |
440 | ||
441 | list_for_each_entry_safe(rq, next, &rq_list, queuelist) { | |
442 | if (!(rq->cmd_flags & REQ_SOFTBARRIER)) | |
443 | continue; | |
444 | ||
445 | rq->cmd_flags &= ~REQ_SOFTBARRIER; | |
446 | list_del_init(&rq->queuelist); | |
447 | blk_mq_insert_request(rq, true, false, false); | |
448 | } | |
449 | ||
450 | while (!list_empty(&rq_list)) { | |
451 | rq = list_entry(rq_list.next, struct request, queuelist); | |
452 | list_del_init(&rq->queuelist); | |
453 | blk_mq_insert_request(rq, false, false, false); | |
454 | } | |
455 | ||
8b957415 JA |
456 | /* |
457 | * Use the start variant of queue running here, so that running | |
458 | * the requeue work will kick stopped queues. | |
459 | */ | |
460 | blk_mq_start_hw_queues(q); | |
6fca6a61 CH |
461 | } |
462 | ||
463 | void blk_mq_add_to_requeue_list(struct request *rq, bool at_head) | |
464 | { | |
465 | struct request_queue *q = rq->q; | |
466 | unsigned long flags; | |
467 | ||
468 | /* | |
469 | * We abuse this flag that is otherwise used by the I/O scheduler to | |
470 | * request head insertation from the workqueue. | |
471 | */ | |
472 | BUG_ON(rq->cmd_flags & REQ_SOFTBARRIER); | |
473 | ||
474 | spin_lock_irqsave(&q->requeue_lock, flags); | |
475 | if (at_head) { | |
476 | rq->cmd_flags |= REQ_SOFTBARRIER; | |
477 | list_add(&rq->queuelist, &q->requeue_list); | |
478 | } else { | |
479 | list_add_tail(&rq->queuelist, &q->requeue_list); | |
480 | } | |
481 | spin_unlock_irqrestore(&q->requeue_lock, flags); | |
482 | } | |
483 | EXPORT_SYMBOL(blk_mq_add_to_requeue_list); | |
484 | ||
485 | void blk_mq_kick_requeue_list(struct request_queue *q) | |
486 | { | |
487 | kblockd_schedule_work(&q->requeue_work); | |
488 | } | |
489 | EXPORT_SYMBOL(blk_mq_kick_requeue_list); | |
490 | ||
7c94e1c1 ML |
491 | static inline bool is_flush_request(struct request *rq, |
492 | struct blk_flush_queue *fq, unsigned int tag) | |
24d2f903 | 493 | { |
0e62f51f | 494 | return ((rq->cmd_flags & REQ_FLUSH_SEQ) && |
7c94e1c1 | 495 | fq->flush_rq->tag == tag); |
0e62f51f JA |
496 | } |
497 | ||
498 | struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag) | |
499 | { | |
500 | struct request *rq = tags->rqs[tag]; | |
e97c293c ML |
501 | /* mq_ctx of flush rq is always cloned from the corresponding req */ |
502 | struct blk_flush_queue *fq = blk_get_flush_queue(rq->q, rq->mq_ctx); | |
22302375 | 503 | |
7c94e1c1 | 504 | if (!is_flush_request(rq, fq, tag)) |
0e62f51f | 505 | return rq; |
22302375 | 506 | |
7c94e1c1 | 507 | return fq->flush_rq; |
24d2f903 CH |
508 | } |
509 | EXPORT_SYMBOL(blk_mq_tag_to_rq); | |
510 | ||
320ae51f | 511 | struct blk_mq_timeout_data { |
46f92d42 CH |
512 | unsigned long next; |
513 | unsigned int next_set; | |
320ae51f JA |
514 | }; |
515 | ||
90415837 | 516 | void blk_mq_rq_timed_out(struct request *req, bool reserved) |
320ae51f | 517 | { |
46f92d42 CH |
518 | struct blk_mq_ops *ops = req->q->mq_ops; |
519 | enum blk_eh_timer_return ret = BLK_EH_RESET_TIMER; | |
87ee7b11 JA |
520 | |
521 | /* | |
522 | * We know that complete is set at this point. If STARTED isn't set | |
523 | * anymore, then the request isn't active and the "timeout" should | |
524 | * just be ignored. This can happen due to the bitflag ordering. | |
525 | * Timeout first checks if STARTED is set, and if it is, assumes | |
526 | * the request is active. But if we race with completion, then | |
527 | * we both flags will get cleared. So check here again, and ignore | |
528 | * a timeout event with a request that isn't active. | |
529 | */ | |
46f92d42 CH |
530 | if (!test_bit(REQ_ATOM_STARTED, &req->atomic_flags)) |
531 | return; | |
87ee7b11 | 532 | |
46f92d42 | 533 | if (ops->timeout) |
0152fb6b | 534 | ret = ops->timeout(req, reserved); |
46f92d42 CH |
535 | |
536 | switch (ret) { | |
537 | case BLK_EH_HANDLED: | |
538 | __blk_mq_complete_request(req); | |
539 | break; | |
540 | case BLK_EH_RESET_TIMER: | |
541 | blk_add_timer(req); | |
542 | blk_clear_rq_complete(req); | |
543 | break; | |
544 | case BLK_EH_NOT_HANDLED: | |
545 | break; | |
546 | default: | |
547 | printk(KERN_ERR "block: bad eh return: %d\n", ret); | |
548 | break; | |
549 | } | |
87ee7b11 | 550 | } |
81481eb4 | 551 | |
81481eb4 CH |
552 | static void blk_mq_check_expired(struct blk_mq_hw_ctx *hctx, |
553 | struct request *rq, void *priv, bool reserved) | |
554 | { | |
555 | struct blk_mq_timeout_data *data = priv; | |
87ee7b11 | 556 | |
46f92d42 CH |
557 | if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) |
558 | return; | |
87ee7b11 | 559 | |
46f92d42 CH |
560 | if (time_after_eq(jiffies, rq->deadline)) { |
561 | if (!blk_mark_rq_complete(rq)) | |
0152fb6b | 562 | blk_mq_rq_timed_out(rq, reserved); |
46f92d42 CH |
563 | } else if (!data->next_set || time_after(data->next, rq->deadline)) { |
564 | data->next = rq->deadline; | |
565 | data->next_set = 1; | |
566 | } | |
87ee7b11 JA |
567 | } |
568 | ||
81481eb4 | 569 | static void blk_mq_rq_timer(unsigned long priv) |
320ae51f | 570 | { |
81481eb4 CH |
571 | struct request_queue *q = (struct request_queue *)priv; |
572 | struct blk_mq_timeout_data data = { | |
573 | .next = 0, | |
574 | .next_set = 0, | |
575 | }; | |
320ae51f | 576 | struct blk_mq_hw_ctx *hctx; |
81481eb4 | 577 | int i; |
320ae51f | 578 | |
484b4061 JA |
579 | queue_for_each_hw_ctx(q, hctx, i) { |
580 | /* | |
581 | * If not software queues are currently mapped to this | |
582 | * hardware queue, there's nothing to check | |
583 | */ | |
584 | if (!hctx->nr_ctx || !hctx->tags) | |
585 | continue; | |
586 | ||
81481eb4 | 587 | blk_mq_tag_busy_iter(hctx, blk_mq_check_expired, &data); |
484b4061 | 588 | } |
320ae51f | 589 | |
81481eb4 CH |
590 | if (data.next_set) { |
591 | data.next = blk_rq_timeout(round_jiffies_up(data.next)); | |
592 | mod_timer(&q->timeout, data.next); | |
0d2602ca JA |
593 | } else { |
594 | queue_for_each_hw_ctx(q, hctx, i) | |
595 | blk_mq_tag_idle(hctx); | |
596 | } | |
320ae51f JA |
597 | } |
598 | ||
599 | /* | |
600 | * Reverse check our software queue for entries that we could potentially | |
601 | * merge with. Currently includes a hand-wavy stop count of 8, to not spend | |
602 | * too much time checking for merges. | |
603 | */ | |
604 | static bool blk_mq_attempt_merge(struct request_queue *q, | |
605 | struct blk_mq_ctx *ctx, struct bio *bio) | |
606 | { | |
607 | struct request *rq; | |
608 | int checked = 8; | |
609 | ||
610 | list_for_each_entry_reverse(rq, &ctx->rq_list, queuelist) { | |
611 | int el_ret; | |
612 | ||
613 | if (!checked--) | |
614 | break; | |
615 | ||
616 | if (!blk_rq_merge_ok(rq, bio)) | |
617 | continue; | |
618 | ||
619 | el_ret = blk_try_merge(rq, bio); | |
620 | if (el_ret == ELEVATOR_BACK_MERGE) { | |
621 | if (bio_attempt_back_merge(q, rq, bio)) { | |
622 | ctx->rq_merged++; | |
623 | return true; | |
624 | } | |
625 | break; | |
626 | } else if (el_ret == ELEVATOR_FRONT_MERGE) { | |
627 | if (bio_attempt_front_merge(q, rq, bio)) { | |
628 | ctx->rq_merged++; | |
629 | return true; | |
630 | } | |
631 | break; | |
632 | } | |
633 | } | |
634 | ||
635 | return false; | |
636 | } | |
637 | ||
1429d7c9 JA |
638 | /* |
639 | * Process software queues that have been marked busy, splicing them | |
640 | * to the for-dispatch | |
641 | */ | |
642 | static void flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list) | |
643 | { | |
644 | struct blk_mq_ctx *ctx; | |
645 | int i; | |
646 | ||
647 | for (i = 0; i < hctx->ctx_map.map_size; i++) { | |
648 | struct blk_align_bitmap *bm = &hctx->ctx_map.map[i]; | |
649 | unsigned int off, bit; | |
650 | ||
651 | if (!bm->word) | |
652 | continue; | |
653 | ||
654 | bit = 0; | |
655 | off = i * hctx->ctx_map.bits_per_word; | |
656 | do { | |
657 | bit = find_next_bit(&bm->word, bm->depth, bit); | |
658 | if (bit >= bm->depth) | |
659 | break; | |
660 | ||
661 | ctx = hctx->ctxs[bit + off]; | |
662 | clear_bit(bit, &bm->word); | |
663 | spin_lock(&ctx->lock); | |
664 | list_splice_tail_init(&ctx->rq_list, list); | |
665 | spin_unlock(&ctx->lock); | |
666 | ||
667 | bit++; | |
668 | } while (1); | |
669 | } | |
670 | } | |
671 | ||
320ae51f JA |
672 | /* |
673 | * Run this hardware queue, pulling any software queues mapped to it in. | |
674 | * Note that this function currently has various problems around ordering | |
675 | * of IO. In particular, we'd like FIFO behaviour on handling existing | |
676 | * items on the hctx->dispatch list. Ignore that for now. | |
677 | */ | |
678 | static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx) | |
679 | { | |
680 | struct request_queue *q = hctx->queue; | |
320ae51f JA |
681 | struct request *rq; |
682 | LIST_HEAD(rq_list); | |
74c45052 JA |
683 | LIST_HEAD(driver_list); |
684 | struct list_head *dptr; | |
1429d7c9 | 685 | int queued; |
320ae51f | 686 | |
fd1270d5 | 687 | WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask)); |
e4043dcf | 688 | |
5d12f905 | 689 | if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state))) |
320ae51f JA |
690 | return; |
691 | ||
692 | hctx->run++; | |
693 | ||
694 | /* | |
695 | * Touch any software queue that has pending entries. | |
696 | */ | |
1429d7c9 | 697 | flush_busy_ctxs(hctx, &rq_list); |
320ae51f JA |
698 | |
699 | /* | |
700 | * If we have previous entries on our dispatch list, grab them | |
701 | * and stuff them at the front for more fair dispatch. | |
702 | */ | |
703 | if (!list_empty_careful(&hctx->dispatch)) { | |
704 | spin_lock(&hctx->lock); | |
705 | if (!list_empty(&hctx->dispatch)) | |
706 | list_splice_init(&hctx->dispatch, &rq_list); | |
707 | spin_unlock(&hctx->lock); | |
708 | } | |
709 | ||
74c45052 JA |
710 | /* |
711 | * Start off with dptr being NULL, so we start the first request | |
712 | * immediately, even if we have more pending. | |
713 | */ | |
714 | dptr = NULL; | |
715 | ||
320ae51f JA |
716 | /* |
717 | * Now process all the entries, sending them to the driver. | |
718 | */ | |
1429d7c9 | 719 | queued = 0; |
320ae51f | 720 | while (!list_empty(&rq_list)) { |
74c45052 | 721 | struct blk_mq_queue_data bd; |
320ae51f JA |
722 | int ret; |
723 | ||
724 | rq = list_first_entry(&rq_list, struct request, queuelist); | |
725 | list_del_init(&rq->queuelist); | |
320ae51f | 726 | |
74c45052 JA |
727 | bd.rq = rq; |
728 | bd.list = dptr; | |
729 | bd.last = list_empty(&rq_list); | |
730 | ||
731 | ret = q->mq_ops->queue_rq(hctx, &bd); | |
320ae51f JA |
732 | switch (ret) { |
733 | case BLK_MQ_RQ_QUEUE_OK: | |
734 | queued++; | |
735 | continue; | |
736 | case BLK_MQ_RQ_QUEUE_BUSY: | |
320ae51f | 737 | list_add(&rq->queuelist, &rq_list); |
ed0791b2 | 738 | __blk_mq_requeue_request(rq); |
320ae51f JA |
739 | break; |
740 | default: | |
741 | pr_err("blk-mq: bad return on queue: %d\n", ret); | |
320ae51f | 742 | case BLK_MQ_RQ_QUEUE_ERROR: |
1e93b8c2 | 743 | rq->errors = -EIO; |
c8a446ad | 744 | blk_mq_end_request(rq, rq->errors); |
320ae51f JA |
745 | break; |
746 | } | |
747 | ||
748 | if (ret == BLK_MQ_RQ_QUEUE_BUSY) | |
749 | break; | |
74c45052 JA |
750 | |
751 | /* | |
752 | * We've done the first request. If we have more than 1 | |
753 | * left in the list, set dptr to defer issue. | |
754 | */ | |
755 | if (!dptr && rq_list.next != rq_list.prev) | |
756 | dptr = &driver_list; | |
320ae51f JA |
757 | } |
758 | ||
759 | if (!queued) | |
760 | hctx->dispatched[0]++; | |
761 | else if (queued < (1 << (BLK_MQ_MAX_DISPATCH_ORDER - 1))) | |
762 | hctx->dispatched[ilog2(queued) + 1]++; | |
763 | ||
764 | /* | |
765 | * Any items that need requeuing? Stuff them into hctx->dispatch, | |
766 | * that is where we will continue on next queue run. | |
767 | */ | |
768 | if (!list_empty(&rq_list)) { | |
769 | spin_lock(&hctx->lock); | |
770 | list_splice(&rq_list, &hctx->dispatch); | |
771 | spin_unlock(&hctx->lock); | |
772 | } | |
773 | } | |
774 | ||
506e931f JA |
775 | /* |
776 | * It'd be great if the workqueue API had a way to pass | |
777 | * in a mask and had some smarts for more clever placement. | |
778 | * For now we just round-robin here, switching for every | |
779 | * BLK_MQ_CPU_WORK_BATCH queued items. | |
780 | */ | |
781 | static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx) | |
782 | { | |
783 | int cpu = hctx->next_cpu; | |
784 | ||
785 | if (--hctx->next_cpu_batch <= 0) { | |
786 | int next_cpu; | |
787 | ||
788 | next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask); | |
789 | if (next_cpu >= nr_cpu_ids) | |
790 | next_cpu = cpumask_first(hctx->cpumask); | |
791 | ||
792 | hctx->next_cpu = next_cpu; | |
793 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; | |
794 | } | |
795 | ||
796 | return cpu; | |
797 | } | |
798 | ||
320ae51f JA |
799 | void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) |
800 | { | |
5d12f905 | 801 | if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state))) |
320ae51f JA |
802 | return; |
803 | ||
398205b8 | 804 | if (!async) { |
2a90d4aa PB |
805 | int cpu = get_cpu(); |
806 | if (cpumask_test_cpu(cpu, hctx->cpumask)) { | |
398205b8 | 807 | __blk_mq_run_hw_queue(hctx); |
2a90d4aa | 808 | put_cpu(); |
398205b8 PB |
809 | return; |
810 | } | |
811 | ||
2a90d4aa | 812 | put_cpu(); |
398205b8 PB |
813 | } |
814 | ||
815 | if (hctx->queue->nr_hw_queues == 1) | |
70f4db63 | 816 | kblockd_schedule_delayed_work(&hctx->run_work, 0); |
e4043dcf JA |
817 | else { |
818 | unsigned int cpu; | |
819 | ||
506e931f | 820 | cpu = blk_mq_hctx_next_cpu(hctx); |
70f4db63 | 821 | kblockd_schedule_delayed_work_on(cpu, &hctx->run_work, 0); |
e4043dcf | 822 | } |
320ae51f JA |
823 | } |
824 | ||
825 | void blk_mq_run_queues(struct request_queue *q, bool async) | |
826 | { | |
827 | struct blk_mq_hw_ctx *hctx; | |
828 | int i; | |
829 | ||
830 | queue_for_each_hw_ctx(q, hctx, i) { | |
831 | if ((!blk_mq_hctx_has_pending(hctx) && | |
832 | list_empty_careful(&hctx->dispatch)) || | |
5d12f905 | 833 | test_bit(BLK_MQ_S_STOPPED, &hctx->state)) |
320ae51f JA |
834 | continue; |
835 | ||
836 | blk_mq_run_hw_queue(hctx, async); | |
837 | } | |
838 | } | |
839 | EXPORT_SYMBOL(blk_mq_run_queues); | |
840 | ||
841 | void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx) | |
842 | { | |
70f4db63 CH |
843 | cancel_delayed_work(&hctx->run_work); |
844 | cancel_delayed_work(&hctx->delay_work); | |
320ae51f JA |
845 | set_bit(BLK_MQ_S_STOPPED, &hctx->state); |
846 | } | |
847 | EXPORT_SYMBOL(blk_mq_stop_hw_queue); | |
848 | ||
280d45f6 CH |
849 | void blk_mq_stop_hw_queues(struct request_queue *q) |
850 | { | |
851 | struct blk_mq_hw_ctx *hctx; | |
852 | int i; | |
853 | ||
854 | queue_for_each_hw_ctx(q, hctx, i) | |
855 | blk_mq_stop_hw_queue(hctx); | |
856 | } | |
857 | EXPORT_SYMBOL(blk_mq_stop_hw_queues); | |
858 | ||
320ae51f JA |
859 | void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx) |
860 | { | |
861 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
e4043dcf | 862 | |
0ffbce80 | 863 | blk_mq_run_hw_queue(hctx, false); |
320ae51f JA |
864 | } |
865 | EXPORT_SYMBOL(blk_mq_start_hw_queue); | |
866 | ||
2f268556 CH |
867 | void blk_mq_start_hw_queues(struct request_queue *q) |
868 | { | |
869 | struct blk_mq_hw_ctx *hctx; | |
870 | int i; | |
871 | ||
872 | queue_for_each_hw_ctx(q, hctx, i) | |
873 | blk_mq_start_hw_queue(hctx); | |
874 | } | |
875 | EXPORT_SYMBOL(blk_mq_start_hw_queues); | |
876 | ||
877 | ||
1b4a3258 | 878 | void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async) |
320ae51f JA |
879 | { |
880 | struct blk_mq_hw_ctx *hctx; | |
881 | int i; | |
882 | ||
883 | queue_for_each_hw_ctx(q, hctx, i) { | |
884 | if (!test_bit(BLK_MQ_S_STOPPED, &hctx->state)) | |
885 | continue; | |
886 | ||
887 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
1b4a3258 | 888 | blk_mq_run_hw_queue(hctx, async); |
320ae51f JA |
889 | } |
890 | } | |
891 | EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues); | |
892 | ||
70f4db63 | 893 | static void blk_mq_run_work_fn(struct work_struct *work) |
320ae51f JA |
894 | { |
895 | struct blk_mq_hw_ctx *hctx; | |
896 | ||
70f4db63 | 897 | hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work); |
e4043dcf | 898 | |
320ae51f JA |
899 | __blk_mq_run_hw_queue(hctx); |
900 | } | |
901 | ||
70f4db63 CH |
902 | static void blk_mq_delay_work_fn(struct work_struct *work) |
903 | { | |
904 | struct blk_mq_hw_ctx *hctx; | |
905 | ||
906 | hctx = container_of(work, struct blk_mq_hw_ctx, delay_work.work); | |
907 | ||
908 | if (test_and_clear_bit(BLK_MQ_S_STOPPED, &hctx->state)) | |
909 | __blk_mq_run_hw_queue(hctx); | |
910 | } | |
911 | ||
912 | void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs) | |
913 | { | |
914 | unsigned long tmo = msecs_to_jiffies(msecs); | |
915 | ||
916 | if (hctx->queue->nr_hw_queues == 1) | |
917 | kblockd_schedule_delayed_work(&hctx->delay_work, tmo); | |
918 | else { | |
919 | unsigned int cpu; | |
920 | ||
506e931f | 921 | cpu = blk_mq_hctx_next_cpu(hctx); |
70f4db63 CH |
922 | kblockd_schedule_delayed_work_on(cpu, &hctx->delay_work, tmo); |
923 | } | |
924 | } | |
925 | EXPORT_SYMBOL(blk_mq_delay_queue); | |
926 | ||
320ae51f | 927 | static void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, |
72a0a36e | 928 | struct request *rq, bool at_head) |
320ae51f JA |
929 | { |
930 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
931 | ||
01b983c9 JA |
932 | trace_block_rq_insert(hctx->queue, rq); |
933 | ||
72a0a36e CH |
934 | if (at_head) |
935 | list_add(&rq->queuelist, &ctx->rq_list); | |
936 | else | |
937 | list_add_tail(&rq->queuelist, &ctx->rq_list); | |
4bb659b1 | 938 | |
320ae51f | 939 | blk_mq_hctx_mark_pending(hctx, ctx); |
320ae51f JA |
940 | } |
941 | ||
eeabc850 CH |
942 | void blk_mq_insert_request(struct request *rq, bool at_head, bool run_queue, |
943 | bool async) | |
320ae51f | 944 | { |
eeabc850 | 945 | struct request_queue *q = rq->q; |
320ae51f | 946 | struct blk_mq_hw_ctx *hctx; |
eeabc850 CH |
947 | struct blk_mq_ctx *ctx = rq->mq_ctx, *current_ctx; |
948 | ||
949 | current_ctx = blk_mq_get_ctx(q); | |
950 | if (!cpu_online(ctx->cpu)) | |
951 | rq->mq_ctx = ctx = current_ctx; | |
320ae51f | 952 | |
320ae51f JA |
953 | hctx = q->mq_ops->map_queue(q, ctx->cpu); |
954 | ||
a57a178a CH |
955 | spin_lock(&ctx->lock); |
956 | __blk_mq_insert_request(hctx, rq, at_head); | |
957 | spin_unlock(&ctx->lock); | |
320ae51f | 958 | |
320ae51f JA |
959 | if (run_queue) |
960 | blk_mq_run_hw_queue(hctx, async); | |
e4043dcf JA |
961 | |
962 | blk_mq_put_ctx(current_ctx); | |
320ae51f JA |
963 | } |
964 | ||
965 | static void blk_mq_insert_requests(struct request_queue *q, | |
966 | struct blk_mq_ctx *ctx, | |
967 | struct list_head *list, | |
968 | int depth, | |
969 | bool from_schedule) | |
970 | ||
971 | { | |
972 | struct blk_mq_hw_ctx *hctx; | |
973 | struct blk_mq_ctx *current_ctx; | |
974 | ||
975 | trace_block_unplug(q, depth, !from_schedule); | |
976 | ||
977 | current_ctx = blk_mq_get_ctx(q); | |
978 | ||
979 | if (!cpu_online(ctx->cpu)) | |
980 | ctx = current_ctx; | |
981 | hctx = q->mq_ops->map_queue(q, ctx->cpu); | |
982 | ||
983 | /* | |
984 | * preemption doesn't flush plug list, so it's possible ctx->cpu is | |
985 | * offline now | |
986 | */ | |
987 | spin_lock(&ctx->lock); | |
988 | while (!list_empty(list)) { | |
989 | struct request *rq; | |
990 | ||
991 | rq = list_first_entry(list, struct request, queuelist); | |
992 | list_del_init(&rq->queuelist); | |
993 | rq->mq_ctx = ctx; | |
72a0a36e | 994 | __blk_mq_insert_request(hctx, rq, false); |
320ae51f JA |
995 | } |
996 | spin_unlock(&ctx->lock); | |
997 | ||
320ae51f | 998 | blk_mq_run_hw_queue(hctx, from_schedule); |
e4043dcf | 999 | blk_mq_put_ctx(current_ctx); |
320ae51f JA |
1000 | } |
1001 | ||
1002 | static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b) | |
1003 | { | |
1004 | struct request *rqa = container_of(a, struct request, queuelist); | |
1005 | struct request *rqb = container_of(b, struct request, queuelist); | |
1006 | ||
1007 | return !(rqa->mq_ctx < rqb->mq_ctx || | |
1008 | (rqa->mq_ctx == rqb->mq_ctx && | |
1009 | blk_rq_pos(rqa) < blk_rq_pos(rqb))); | |
1010 | } | |
1011 | ||
1012 | void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule) | |
1013 | { | |
1014 | struct blk_mq_ctx *this_ctx; | |
1015 | struct request_queue *this_q; | |
1016 | struct request *rq; | |
1017 | LIST_HEAD(list); | |
1018 | LIST_HEAD(ctx_list); | |
1019 | unsigned int depth; | |
1020 | ||
1021 | list_splice_init(&plug->mq_list, &list); | |
1022 | ||
1023 | list_sort(NULL, &list, plug_ctx_cmp); | |
1024 | ||
1025 | this_q = NULL; | |
1026 | this_ctx = NULL; | |
1027 | depth = 0; | |
1028 | ||
1029 | while (!list_empty(&list)) { | |
1030 | rq = list_entry_rq(list.next); | |
1031 | list_del_init(&rq->queuelist); | |
1032 | BUG_ON(!rq->q); | |
1033 | if (rq->mq_ctx != this_ctx) { | |
1034 | if (this_ctx) { | |
1035 | blk_mq_insert_requests(this_q, this_ctx, | |
1036 | &ctx_list, depth, | |
1037 | from_schedule); | |
1038 | } | |
1039 | ||
1040 | this_ctx = rq->mq_ctx; | |
1041 | this_q = rq->q; | |
1042 | depth = 0; | |
1043 | } | |
1044 | ||
1045 | depth++; | |
1046 | list_add_tail(&rq->queuelist, &ctx_list); | |
1047 | } | |
1048 | ||
1049 | /* | |
1050 | * If 'this_ctx' is set, we know we have entries to complete | |
1051 | * on 'ctx_list'. Do those. | |
1052 | */ | |
1053 | if (this_ctx) { | |
1054 | blk_mq_insert_requests(this_q, this_ctx, &ctx_list, depth, | |
1055 | from_schedule); | |
1056 | } | |
1057 | } | |
1058 | ||
1059 | static void blk_mq_bio_to_request(struct request *rq, struct bio *bio) | |
1060 | { | |
1061 | init_request_from_bio(rq, bio); | |
4b570521 | 1062 | |
3ee32372 | 1063 | if (blk_do_io_stat(rq)) |
4b570521 | 1064 | blk_account_io_start(rq, 1); |
320ae51f JA |
1065 | } |
1066 | ||
274a5843 JA |
1067 | static inline bool hctx_allow_merges(struct blk_mq_hw_ctx *hctx) |
1068 | { | |
1069 | return (hctx->flags & BLK_MQ_F_SHOULD_MERGE) && | |
1070 | !blk_queue_nomerges(hctx->queue); | |
1071 | } | |
1072 | ||
07068d5b JA |
1073 | static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx *hctx, |
1074 | struct blk_mq_ctx *ctx, | |
1075 | struct request *rq, struct bio *bio) | |
320ae51f | 1076 | { |
274a5843 | 1077 | if (!hctx_allow_merges(hctx)) { |
07068d5b JA |
1078 | blk_mq_bio_to_request(rq, bio); |
1079 | spin_lock(&ctx->lock); | |
1080 | insert_rq: | |
1081 | __blk_mq_insert_request(hctx, rq, false); | |
1082 | spin_unlock(&ctx->lock); | |
1083 | return false; | |
1084 | } else { | |
274a5843 JA |
1085 | struct request_queue *q = hctx->queue; |
1086 | ||
07068d5b JA |
1087 | spin_lock(&ctx->lock); |
1088 | if (!blk_mq_attempt_merge(q, ctx, bio)) { | |
1089 | blk_mq_bio_to_request(rq, bio); | |
1090 | goto insert_rq; | |
1091 | } | |
320ae51f | 1092 | |
07068d5b JA |
1093 | spin_unlock(&ctx->lock); |
1094 | __blk_mq_free_request(hctx, ctx, rq); | |
1095 | return true; | |
14ec77f3 | 1096 | } |
07068d5b | 1097 | } |
14ec77f3 | 1098 | |
07068d5b JA |
1099 | struct blk_map_ctx { |
1100 | struct blk_mq_hw_ctx *hctx; | |
1101 | struct blk_mq_ctx *ctx; | |
1102 | }; | |
1103 | ||
1104 | static struct request *blk_mq_map_request(struct request_queue *q, | |
1105 | struct bio *bio, | |
1106 | struct blk_map_ctx *data) | |
1107 | { | |
1108 | struct blk_mq_hw_ctx *hctx; | |
1109 | struct blk_mq_ctx *ctx; | |
1110 | struct request *rq; | |
1111 | int rw = bio_data_dir(bio); | |
cb96a42c | 1112 | struct blk_mq_alloc_data alloc_data; |
320ae51f | 1113 | |
07068d5b | 1114 | if (unlikely(blk_mq_queue_enter(q))) { |
320ae51f | 1115 | bio_endio(bio, -EIO); |
07068d5b | 1116 | return NULL; |
320ae51f JA |
1117 | } |
1118 | ||
1119 | ctx = blk_mq_get_ctx(q); | |
1120 | hctx = q->mq_ops->map_queue(q, ctx->cpu); | |
1121 | ||
07068d5b | 1122 | if (rw_is_sync(bio->bi_rw)) |
27fbf4e8 | 1123 | rw |= REQ_SYNC; |
07068d5b | 1124 | |
320ae51f | 1125 | trace_block_getrq(q, bio, rw); |
cb96a42c ML |
1126 | blk_mq_set_alloc_data(&alloc_data, q, GFP_ATOMIC, false, ctx, |
1127 | hctx); | |
1128 | rq = __blk_mq_alloc_request(&alloc_data, rw); | |
5dee8577 | 1129 | if (unlikely(!rq)) { |
793597a6 | 1130 | __blk_mq_run_hw_queue(hctx); |
320ae51f JA |
1131 | blk_mq_put_ctx(ctx); |
1132 | trace_block_sleeprq(q, bio, rw); | |
793597a6 CH |
1133 | |
1134 | ctx = blk_mq_get_ctx(q); | |
320ae51f | 1135 | hctx = q->mq_ops->map_queue(q, ctx->cpu); |
cb96a42c ML |
1136 | blk_mq_set_alloc_data(&alloc_data, q, |
1137 | __GFP_WAIT|GFP_ATOMIC, false, ctx, hctx); | |
1138 | rq = __blk_mq_alloc_request(&alloc_data, rw); | |
1139 | ctx = alloc_data.ctx; | |
1140 | hctx = alloc_data.hctx; | |
320ae51f JA |
1141 | } |
1142 | ||
1143 | hctx->queued++; | |
07068d5b JA |
1144 | data->hctx = hctx; |
1145 | data->ctx = ctx; | |
1146 | return rq; | |
1147 | } | |
1148 | ||
1149 | /* | |
1150 | * Multiple hardware queue variant. This will not use per-process plugs, | |
1151 | * but will attempt to bypass the hctx queueing if we can go straight to | |
1152 | * hardware for SYNC IO. | |
1153 | */ | |
1154 | static void blk_mq_make_request(struct request_queue *q, struct bio *bio) | |
1155 | { | |
1156 | const int is_sync = rw_is_sync(bio->bi_rw); | |
1157 | const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA); | |
1158 | struct blk_map_ctx data; | |
1159 | struct request *rq; | |
1160 | ||
1161 | blk_queue_bounce(q, &bio); | |
1162 | ||
1163 | if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) { | |
1164 | bio_endio(bio, -EIO); | |
1165 | return; | |
1166 | } | |
1167 | ||
1168 | rq = blk_mq_map_request(q, bio, &data); | |
1169 | if (unlikely(!rq)) | |
1170 | return; | |
1171 | ||
1172 | if (unlikely(is_flush_fua)) { | |
1173 | blk_mq_bio_to_request(rq, bio); | |
1174 | blk_insert_flush(rq); | |
1175 | goto run_queue; | |
1176 | } | |
1177 | ||
e167dfb5 JA |
1178 | /* |
1179 | * If the driver supports defer issued based on 'last', then | |
1180 | * queue it up like normal since we can potentially save some | |
1181 | * CPU this way. | |
1182 | */ | |
1183 | if (is_sync && !(data.hctx->flags & BLK_MQ_F_DEFER_ISSUE)) { | |
74c45052 JA |
1184 | struct blk_mq_queue_data bd = { |
1185 | .rq = rq, | |
1186 | .list = NULL, | |
1187 | .last = 1 | |
1188 | }; | |
07068d5b JA |
1189 | int ret; |
1190 | ||
1191 | blk_mq_bio_to_request(rq, bio); | |
07068d5b JA |
1192 | |
1193 | /* | |
1194 | * For OK queue, we are done. For error, kill it. Any other | |
1195 | * error (busy), just add it to our list as we previously | |
1196 | * would have done | |
1197 | */ | |
74c45052 | 1198 | ret = q->mq_ops->queue_rq(data.hctx, &bd); |
07068d5b JA |
1199 | if (ret == BLK_MQ_RQ_QUEUE_OK) |
1200 | goto done; | |
1201 | else { | |
1202 | __blk_mq_requeue_request(rq); | |
1203 | ||
1204 | if (ret == BLK_MQ_RQ_QUEUE_ERROR) { | |
1205 | rq->errors = -EIO; | |
c8a446ad | 1206 | blk_mq_end_request(rq, rq->errors); |
07068d5b JA |
1207 | goto done; |
1208 | } | |
1209 | } | |
1210 | } | |
1211 | ||
1212 | if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) { | |
1213 | /* | |
1214 | * For a SYNC request, send it to the hardware immediately. For | |
1215 | * an ASYNC request, just ensure that we run it later on. The | |
1216 | * latter allows for merging opportunities and more efficient | |
1217 | * dispatching. | |
1218 | */ | |
1219 | run_queue: | |
1220 | blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua); | |
1221 | } | |
1222 | done: | |
1223 | blk_mq_put_ctx(data.ctx); | |
1224 | } | |
1225 | ||
1226 | /* | |
1227 | * Single hardware queue variant. This will attempt to use any per-process | |
1228 | * plug for merging and IO deferral. | |
1229 | */ | |
1230 | static void blk_sq_make_request(struct request_queue *q, struct bio *bio) | |
1231 | { | |
1232 | const int is_sync = rw_is_sync(bio->bi_rw); | |
1233 | const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA); | |
1234 | unsigned int use_plug, request_count = 0; | |
1235 | struct blk_map_ctx data; | |
1236 | struct request *rq; | |
1237 | ||
1238 | /* | |
1239 | * If we have multiple hardware queues, just go directly to | |
1240 | * one of those for sync IO. | |
1241 | */ | |
1242 | use_plug = !is_flush_fua && !is_sync; | |
1243 | ||
1244 | blk_queue_bounce(q, &bio); | |
1245 | ||
1246 | if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) { | |
1247 | bio_endio(bio, -EIO); | |
1248 | return; | |
1249 | } | |
1250 | ||
1251 | if (use_plug && !blk_queue_nomerges(q) && | |
1252 | blk_attempt_plug_merge(q, bio, &request_count)) | |
1253 | return; | |
1254 | ||
1255 | rq = blk_mq_map_request(q, bio, &data); | |
ff87bcec JA |
1256 | if (unlikely(!rq)) |
1257 | return; | |
320ae51f JA |
1258 | |
1259 | if (unlikely(is_flush_fua)) { | |
1260 | blk_mq_bio_to_request(rq, bio); | |
320ae51f JA |
1261 | blk_insert_flush(rq); |
1262 | goto run_queue; | |
1263 | } | |
1264 | ||
1265 | /* | |
1266 | * A task plug currently exists. Since this is completely lockless, | |
1267 | * utilize that to temporarily store requests until the task is | |
1268 | * either done or scheduled away. | |
1269 | */ | |
1270 | if (use_plug) { | |
1271 | struct blk_plug *plug = current->plug; | |
1272 | ||
1273 | if (plug) { | |
1274 | blk_mq_bio_to_request(rq, bio); | |
92f399c7 | 1275 | if (list_empty(&plug->mq_list)) |
320ae51f JA |
1276 | trace_block_plug(q); |
1277 | else if (request_count >= BLK_MAX_REQUEST_COUNT) { | |
1278 | blk_flush_plug_list(plug, false); | |
1279 | trace_block_plug(q); | |
1280 | } | |
1281 | list_add_tail(&rq->queuelist, &plug->mq_list); | |
07068d5b | 1282 | blk_mq_put_ctx(data.ctx); |
320ae51f JA |
1283 | return; |
1284 | } | |
1285 | } | |
1286 | ||
07068d5b JA |
1287 | if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) { |
1288 | /* | |
1289 | * For a SYNC request, send it to the hardware immediately. For | |
1290 | * an ASYNC request, just ensure that we run it later on. The | |
1291 | * latter allows for merging opportunities and more efficient | |
1292 | * dispatching. | |
1293 | */ | |
1294 | run_queue: | |
1295 | blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua); | |
320ae51f JA |
1296 | } |
1297 | ||
07068d5b | 1298 | blk_mq_put_ctx(data.ctx); |
320ae51f JA |
1299 | } |
1300 | ||
1301 | /* | |
1302 | * Default mapping to a software queue, since we use one per CPU. | |
1303 | */ | |
1304 | struct blk_mq_hw_ctx *blk_mq_map_queue(struct request_queue *q, const int cpu) | |
1305 | { | |
1306 | return q->queue_hw_ctx[q->mq_map[cpu]]; | |
1307 | } | |
1308 | EXPORT_SYMBOL(blk_mq_map_queue); | |
1309 | ||
24d2f903 CH |
1310 | static void blk_mq_free_rq_map(struct blk_mq_tag_set *set, |
1311 | struct blk_mq_tags *tags, unsigned int hctx_idx) | |
95363efd | 1312 | { |
e9b267d9 | 1313 | struct page *page; |
320ae51f | 1314 | |
24d2f903 | 1315 | if (tags->rqs && set->ops->exit_request) { |
e9b267d9 | 1316 | int i; |
320ae51f | 1317 | |
24d2f903 CH |
1318 | for (i = 0; i < tags->nr_tags; i++) { |
1319 | if (!tags->rqs[i]) | |
e9b267d9 | 1320 | continue; |
24d2f903 CH |
1321 | set->ops->exit_request(set->driver_data, tags->rqs[i], |
1322 | hctx_idx, i); | |
a5164405 | 1323 | tags->rqs[i] = NULL; |
e9b267d9 | 1324 | } |
320ae51f | 1325 | } |
320ae51f | 1326 | |
24d2f903 CH |
1327 | while (!list_empty(&tags->page_list)) { |
1328 | page = list_first_entry(&tags->page_list, struct page, lru); | |
6753471c | 1329 | list_del_init(&page->lru); |
320ae51f JA |
1330 | __free_pages(page, page->private); |
1331 | } | |
1332 | ||
24d2f903 | 1333 | kfree(tags->rqs); |
320ae51f | 1334 | |
24d2f903 | 1335 | blk_mq_free_tags(tags); |
320ae51f JA |
1336 | } |
1337 | ||
1338 | static size_t order_to_size(unsigned int order) | |
1339 | { | |
4ca08500 | 1340 | return (size_t)PAGE_SIZE << order; |
320ae51f JA |
1341 | } |
1342 | ||
24d2f903 CH |
1343 | static struct blk_mq_tags *blk_mq_init_rq_map(struct blk_mq_tag_set *set, |
1344 | unsigned int hctx_idx) | |
320ae51f | 1345 | { |
24d2f903 | 1346 | struct blk_mq_tags *tags; |
320ae51f JA |
1347 | unsigned int i, j, entries_per_page, max_order = 4; |
1348 | size_t rq_size, left; | |
1349 | ||
24d2f903 CH |
1350 | tags = blk_mq_init_tags(set->queue_depth, set->reserved_tags, |
1351 | set->numa_node); | |
1352 | if (!tags) | |
1353 | return NULL; | |
320ae51f | 1354 | |
24d2f903 CH |
1355 | INIT_LIST_HEAD(&tags->page_list); |
1356 | ||
a5164405 JA |
1357 | tags->rqs = kzalloc_node(set->queue_depth * sizeof(struct request *), |
1358 | GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY, | |
1359 | set->numa_node); | |
24d2f903 CH |
1360 | if (!tags->rqs) { |
1361 | blk_mq_free_tags(tags); | |
1362 | return NULL; | |
1363 | } | |
320ae51f JA |
1364 | |
1365 | /* | |
1366 | * rq_size is the size of the request plus driver payload, rounded | |
1367 | * to the cacheline size | |
1368 | */ | |
24d2f903 | 1369 | rq_size = round_up(sizeof(struct request) + set->cmd_size, |
320ae51f | 1370 | cache_line_size()); |
24d2f903 | 1371 | left = rq_size * set->queue_depth; |
320ae51f | 1372 | |
24d2f903 | 1373 | for (i = 0; i < set->queue_depth; ) { |
320ae51f JA |
1374 | int this_order = max_order; |
1375 | struct page *page; | |
1376 | int to_do; | |
1377 | void *p; | |
1378 | ||
1379 | while (left < order_to_size(this_order - 1) && this_order) | |
1380 | this_order--; | |
1381 | ||
1382 | do { | |
a5164405 JA |
1383 | page = alloc_pages_node(set->numa_node, |
1384 | GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY, | |
1385 | this_order); | |
320ae51f JA |
1386 | if (page) |
1387 | break; | |
1388 | if (!this_order--) | |
1389 | break; | |
1390 | if (order_to_size(this_order) < rq_size) | |
1391 | break; | |
1392 | } while (1); | |
1393 | ||
1394 | if (!page) | |
24d2f903 | 1395 | goto fail; |
320ae51f JA |
1396 | |
1397 | page->private = this_order; | |
24d2f903 | 1398 | list_add_tail(&page->lru, &tags->page_list); |
320ae51f JA |
1399 | |
1400 | p = page_address(page); | |
1401 | entries_per_page = order_to_size(this_order) / rq_size; | |
24d2f903 | 1402 | to_do = min(entries_per_page, set->queue_depth - i); |
320ae51f JA |
1403 | left -= to_do * rq_size; |
1404 | for (j = 0; j < to_do; j++) { | |
24d2f903 | 1405 | tags->rqs[i] = p; |
683d0e12 DH |
1406 | tags->rqs[i]->atomic_flags = 0; |
1407 | tags->rqs[i]->cmd_flags = 0; | |
24d2f903 CH |
1408 | if (set->ops->init_request) { |
1409 | if (set->ops->init_request(set->driver_data, | |
1410 | tags->rqs[i], hctx_idx, i, | |
a5164405 JA |
1411 | set->numa_node)) { |
1412 | tags->rqs[i] = NULL; | |
24d2f903 | 1413 | goto fail; |
a5164405 | 1414 | } |
e9b267d9 CH |
1415 | } |
1416 | ||
320ae51f JA |
1417 | p += rq_size; |
1418 | i++; | |
1419 | } | |
1420 | } | |
1421 | ||
24d2f903 | 1422 | return tags; |
320ae51f | 1423 | |
24d2f903 | 1424 | fail: |
24d2f903 CH |
1425 | blk_mq_free_rq_map(set, tags, hctx_idx); |
1426 | return NULL; | |
320ae51f JA |
1427 | } |
1428 | ||
1429d7c9 JA |
1429 | static void blk_mq_free_bitmap(struct blk_mq_ctxmap *bitmap) |
1430 | { | |
1431 | kfree(bitmap->map); | |
1432 | } | |
1433 | ||
1434 | static int blk_mq_alloc_bitmap(struct blk_mq_ctxmap *bitmap, int node) | |
1435 | { | |
1436 | unsigned int bpw = 8, total, num_maps, i; | |
1437 | ||
1438 | bitmap->bits_per_word = bpw; | |
1439 | ||
1440 | num_maps = ALIGN(nr_cpu_ids, bpw) / bpw; | |
1441 | bitmap->map = kzalloc_node(num_maps * sizeof(struct blk_align_bitmap), | |
1442 | GFP_KERNEL, node); | |
1443 | if (!bitmap->map) | |
1444 | return -ENOMEM; | |
1445 | ||
1446 | bitmap->map_size = num_maps; | |
1447 | ||
1448 | total = nr_cpu_ids; | |
1449 | for (i = 0; i < num_maps; i++) { | |
1450 | bitmap->map[i].depth = min(total, bitmap->bits_per_word); | |
1451 | total -= bitmap->map[i].depth; | |
1452 | } | |
1453 | ||
1454 | return 0; | |
1455 | } | |
1456 | ||
484b4061 JA |
1457 | static int blk_mq_hctx_cpu_offline(struct blk_mq_hw_ctx *hctx, int cpu) |
1458 | { | |
1459 | struct request_queue *q = hctx->queue; | |
1460 | struct blk_mq_ctx *ctx; | |
1461 | LIST_HEAD(tmp); | |
1462 | ||
1463 | /* | |
1464 | * Move ctx entries to new CPU, if this one is going away. | |
1465 | */ | |
1466 | ctx = __blk_mq_get_ctx(q, cpu); | |
1467 | ||
1468 | spin_lock(&ctx->lock); | |
1469 | if (!list_empty(&ctx->rq_list)) { | |
1470 | list_splice_init(&ctx->rq_list, &tmp); | |
1471 | blk_mq_hctx_clear_pending(hctx, ctx); | |
1472 | } | |
1473 | spin_unlock(&ctx->lock); | |
1474 | ||
1475 | if (list_empty(&tmp)) | |
1476 | return NOTIFY_OK; | |
1477 | ||
1478 | ctx = blk_mq_get_ctx(q); | |
1479 | spin_lock(&ctx->lock); | |
1480 | ||
1481 | while (!list_empty(&tmp)) { | |
1482 | struct request *rq; | |
1483 | ||
1484 | rq = list_first_entry(&tmp, struct request, queuelist); | |
1485 | rq->mq_ctx = ctx; | |
1486 | list_move_tail(&rq->queuelist, &ctx->rq_list); | |
1487 | } | |
1488 | ||
1489 | hctx = q->mq_ops->map_queue(q, ctx->cpu); | |
1490 | blk_mq_hctx_mark_pending(hctx, ctx); | |
1491 | ||
1492 | spin_unlock(&ctx->lock); | |
1493 | ||
1494 | blk_mq_run_hw_queue(hctx, true); | |
1495 | blk_mq_put_ctx(ctx); | |
1496 | return NOTIFY_OK; | |
1497 | } | |
1498 | ||
1499 | static int blk_mq_hctx_cpu_online(struct blk_mq_hw_ctx *hctx, int cpu) | |
1500 | { | |
1501 | struct request_queue *q = hctx->queue; | |
1502 | struct blk_mq_tag_set *set = q->tag_set; | |
1503 | ||
1504 | if (set->tags[hctx->queue_num]) | |
1505 | return NOTIFY_OK; | |
1506 | ||
1507 | set->tags[hctx->queue_num] = blk_mq_init_rq_map(set, hctx->queue_num); | |
1508 | if (!set->tags[hctx->queue_num]) | |
1509 | return NOTIFY_STOP; | |
1510 | ||
1511 | hctx->tags = set->tags[hctx->queue_num]; | |
1512 | return NOTIFY_OK; | |
1513 | } | |
1514 | ||
1515 | static int blk_mq_hctx_notify(void *data, unsigned long action, | |
1516 | unsigned int cpu) | |
1517 | { | |
1518 | struct blk_mq_hw_ctx *hctx = data; | |
1519 | ||
1520 | if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) | |
1521 | return blk_mq_hctx_cpu_offline(hctx, cpu); | |
1522 | else if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) | |
1523 | return blk_mq_hctx_cpu_online(hctx, cpu); | |
1524 | ||
1525 | return NOTIFY_OK; | |
1526 | } | |
1527 | ||
08e98fc6 ML |
1528 | static void blk_mq_exit_hctx(struct request_queue *q, |
1529 | struct blk_mq_tag_set *set, | |
1530 | struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx) | |
1531 | { | |
f70ced09 ML |
1532 | unsigned flush_start_tag = set->queue_depth; |
1533 | ||
08e98fc6 ML |
1534 | blk_mq_tag_idle(hctx); |
1535 | ||
f70ced09 ML |
1536 | if (set->ops->exit_request) |
1537 | set->ops->exit_request(set->driver_data, | |
1538 | hctx->fq->flush_rq, hctx_idx, | |
1539 | flush_start_tag + hctx_idx); | |
1540 | ||
08e98fc6 ML |
1541 | if (set->ops->exit_hctx) |
1542 | set->ops->exit_hctx(hctx, hctx_idx); | |
1543 | ||
1544 | blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier); | |
f70ced09 | 1545 | blk_free_flush_queue(hctx->fq); |
08e98fc6 ML |
1546 | kfree(hctx->ctxs); |
1547 | blk_mq_free_bitmap(&hctx->ctx_map); | |
1548 | } | |
1549 | ||
624dbe47 ML |
1550 | static void blk_mq_exit_hw_queues(struct request_queue *q, |
1551 | struct blk_mq_tag_set *set, int nr_queue) | |
1552 | { | |
1553 | struct blk_mq_hw_ctx *hctx; | |
1554 | unsigned int i; | |
1555 | ||
1556 | queue_for_each_hw_ctx(q, hctx, i) { | |
1557 | if (i == nr_queue) | |
1558 | break; | |
08e98fc6 | 1559 | blk_mq_exit_hctx(q, set, hctx, i); |
624dbe47 | 1560 | } |
624dbe47 ML |
1561 | } |
1562 | ||
1563 | static void blk_mq_free_hw_queues(struct request_queue *q, | |
1564 | struct blk_mq_tag_set *set) | |
1565 | { | |
1566 | struct blk_mq_hw_ctx *hctx; | |
1567 | unsigned int i; | |
1568 | ||
1569 | queue_for_each_hw_ctx(q, hctx, i) { | |
1570 | free_cpumask_var(hctx->cpumask); | |
cdef54dd | 1571 | kfree(hctx); |
624dbe47 ML |
1572 | } |
1573 | } | |
1574 | ||
08e98fc6 ML |
1575 | static int blk_mq_init_hctx(struct request_queue *q, |
1576 | struct blk_mq_tag_set *set, | |
1577 | struct blk_mq_hw_ctx *hctx, unsigned hctx_idx) | |
320ae51f | 1578 | { |
08e98fc6 | 1579 | int node; |
f70ced09 | 1580 | unsigned flush_start_tag = set->queue_depth; |
08e98fc6 ML |
1581 | |
1582 | node = hctx->numa_node; | |
1583 | if (node == NUMA_NO_NODE) | |
1584 | node = hctx->numa_node = set->numa_node; | |
1585 | ||
1586 | INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn); | |
1587 | INIT_DELAYED_WORK(&hctx->delay_work, blk_mq_delay_work_fn); | |
1588 | spin_lock_init(&hctx->lock); | |
1589 | INIT_LIST_HEAD(&hctx->dispatch); | |
1590 | hctx->queue = q; | |
1591 | hctx->queue_num = hctx_idx; | |
1592 | hctx->flags = set->flags; | |
1593 | hctx->cmd_size = set->cmd_size; | |
1594 | ||
1595 | blk_mq_init_cpu_notifier(&hctx->cpu_notifier, | |
1596 | blk_mq_hctx_notify, hctx); | |
1597 | blk_mq_register_cpu_notifier(&hctx->cpu_notifier); | |
1598 | ||
1599 | hctx->tags = set->tags[hctx_idx]; | |
320ae51f JA |
1600 | |
1601 | /* | |
08e98fc6 ML |
1602 | * Allocate space for all possible cpus to avoid allocation at |
1603 | * runtime | |
320ae51f | 1604 | */ |
08e98fc6 ML |
1605 | hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *), |
1606 | GFP_KERNEL, node); | |
1607 | if (!hctx->ctxs) | |
1608 | goto unregister_cpu_notifier; | |
320ae51f | 1609 | |
08e98fc6 ML |
1610 | if (blk_mq_alloc_bitmap(&hctx->ctx_map, node)) |
1611 | goto free_ctxs; | |
320ae51f | 1612 | |
08e98fc6 | 1613 | hctx->nr_ctx = 0; |
320ae51f | 1614 | |
08e98fc6 ML |
1615 | if (set->ops->init_hctx && |
1616 | set->ops->init_hctx(hctx, set->driver_data, hctx_idx)) | |
1617 | goto free_bitmap; | |
320ae51f | 1618 | |
f70ced09 ML |
1619 | hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size); |
1620 | if (!hctx->fq) | |
1621 | goto exit_hctx; | |
320ae51f | 1622 | |
f70ced09 ML |
1623 | if (set->ops->init_request && |
1624 | set->ops->init_request(set->driver_data, | |
1625 | hctx->fq->flush_rq, hctx_idx, | |
1626 | flush_start_tag + hctx_idx, node)) | |
1627 | goto free_fq; | |
320ae51f | 1628 | |
08e98fc6 | 1629 | return 0; |
320ae51f | 1630 | |
f70ced09 ML |
1631 | free_fq: |
1632 | kfree(hctx->fq); | |
1633 | exit_hctx: | |
1634 | if (set->ops->exit_hctx) | |
1635 | set->ops->exit_hctx(hctx, hctx_idx); | |
08e98fc6 ML |
1636 | free_bitmap: |
1637 | blk_mq_free_bitmap(&hctx->ctx_map); | |
1638 | free_ctxs: | |
1639 | kfree(hctx->ctxs); | |
1640 | unregister_cpu_notifier: | |
1641 | blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier); | |
320ae51f | 1642 | |
08e98fc6 ML |
1643 | return -1; |
1644 | } | |
320ae51f | 1645 | |
08e98fc6 ML |
1646 | static int blk_mq_init_hw_queues(struct request_queue *q, |
1647 | struct blk_mq_tag_set *set) | |
1648 | { | |
1649 | struct blk_mq_hw_ctx *hctx; | |
1650 | unsigned int i; | |
320ae51f | 1651 | |
08e98fc6 ML |
1652 | /* |
1653 | * Initialize hardware queues | |
1654 | */ | |
1655 | queue_for_each_hw_ctx(q, hctx, i) { | |
1656 | if (blk_mq_init_hctx(q, set, hctx, i)) | |
320ae51f JA |
1657 | break; |
1658 | } | |
1659 | ||
1660 | if (i == q->nr_hw_queues) | |
1661 | return 0; | |
1662 | ||
1663 | /* | |
1664 | * Init failed | |
1665 | */ | |
624dbe47 | 1666 | blk_mq_exit_hw_queues(q, set, i); |
320ae51f JA |
1667 | |
1668 | return 1; | |
1669 | } | |
1670 | ||
1671 | static void blk_mq_init_cpu_queues(struct request_queue *q, | |
1672 | unsigned int nr_hw_queues) | |
1673 | { | |
1674 | unsigned int i; | |
1675 | ||
1676 | for_each_possible_cpu(i) { | |
1677 | struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i); | |
1678 | struct blk_mq_hw_ctx *hctx; | |
1679 | ||
1680 | memset(__ctx, 0, sizeof(*__ctx)); | |
1681 | __ctx->cpu = i; | |
1682 | spin_lock_init(&__ctx->lock); | |
1683 | INIT_LIST_HEAD(&__ctx->rq_list); | |
1684 | __ctx->queue = q; | |
1685 | ||
1686 | /* If the cpu isn't online, the cpu is mapped to first hctx */ | |
320ae51f JA |
1687 | if (!cpu_online(i)) |
1688 | continue; | |
1689 | ||
e4043dcf JA |
1690 | hctx = q->mq_ops->map_queue(q, i); |
1691 | cpumask_set_cpu(i, hctx->cpumask); | |
1692 | hctx->nr_ctx++; | |
1693 | ||
320ae51f JA |
1694 | /* |
1695 | * Set local node, IFF we have more than one hw queue. If | |
1696 | * not, we remain on the home node of the device | |
1697 | */ | |
1698 | if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE) | |
1699 | hctx->numa_node = cpu_to_node(i); | |
1700 | } | |
1701 | } | |
1702 | ||
1703 | static void blk_mq_map_swqueue(struct request_queue *q) | |
1704 | { | |
1705 | unsigned int i; | |
1706 | struct blk_mq_hw_ctx *hctx; | |
1707 | struct blk_mq_ctx *ctx; | |
1708 | ||
1709 | queue_for_each_hw_ctx(q, hctx, i) { | |
e4043dcf | 1710 | cpumask_clear(hctx->cpumask); |
320ae51f JA |
1711 | hctx->nr_ctx = 0; |
1712 | } | |
1713 | ||
1714 | /* | |
1715 | * Map software to hardware queues | |
1716 | */ | |
1717 | queue_for_each_ctx(q, ctx, i) { | |
1718 | /* If the cpu isn't online, the cpu is mapped to first hctx */ | |
e4043dcf JA |
1719 | if (!cpu_online(i)) |
1720 | continue; | |
1721 | ||
320ae51f | 1722 | hctx = q->mq_ops->map_queue(q, i); |
e4043dcf | 1723 | cpumask_set_cpu(i, hctx->cpumask); |
320ae51f JA |
1724 | ctx->index_hw = hctx->nr_ctx; |
1725 | hctx->ctxs[hctx->nr_ctx++] = ctx; | |
1726 | } | |
506e931f JA |
1727 | |
1728 | queue_for_each_hw_ctx(q, hctx, i) { | |
484b4061 | 1729 | /* |
a68aafa5 JA |
1730 | * If no software queues are mapped to this hardware queue, |
1731 | * disable it and free the request entries. | |
484b4061 JA |
1732 | */ |
1733 | if (!hctx->nr_ctx) { | |
1734 | struct blk_mq_tag_set *set = q->tag_set; | |
1735 | ||
1736 | if (set->tags[i]) { | |
1737 | blk_mq_free_rq_map(set, set->tags[i], i); | |
1738 | set->tags[i] = NULL; | |
1739 | hctx->tags = NULL; | |
1740 | } | |
1741 | continue; | |
1742 | } | |
1743 | ||
1744 | /* | |
1745 | * Initialize batch roundrobin counts | |
1746 | */ | |
506e931f JA |
1747 | hctx->next_cpu = cpumask_first(hctx->cpumask); |
1748 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; | |
1749 | } | |
320ae51f JA |
1750 | } |
1751 | ||
0d2602ca JA |
1752 | static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set) |
1753 | { | |
1754 | struct blk_mq_hw_ctx *hctx; | |
1755 | struct request_queue *q; | |
1756 | bool shared; | |
1757 | int i; | |
1758 | ||
1759 | if (set->tag_list.next == set->tag_list.prev) | |
1760 | shared = false; | |
1761 | else | |
1762 | shared = true; | |
1763 | ||
1764 | list_for_each_entry(q, &set->tag_list, tag_set_list) { | |
1765 | blk_mq_freeze_queue(q); | |
1766 | ||
1767 | queue_for_each_hw_ctx(q, hctx, i) { | |
1768 | if (shared) | |
1769 | hctx->flags |= BLK_MQ_F_TAG_SHARED; | |
1770 | else | |
1771 | hctx->flags &= ~BLK_MQ_F_TAG_SHARED; | |
1772 | } | |
1773 | blk_mq_unfreeze_queue(q); | |
1774 | } | |
1775 | } | |
1776 | ||
1777 | static void blk_mq_del_queue_tag_set(struct request_queue *q) | |
1778 | { | |
1779 | struct blk_mq_tag_set *set = q->tag_set; | |
1780 | ||
0d2602ca JA |
1781 | mutex_lock(&set->tag_list_lock); |
1782 | list_del_init(&q->tag_set_list); | |
1783 | blk_mq_update_tag_set_depth(set); | |
1784 | mutex_unlock(&set->tag_list_lock); | |
0d2602ca JA |
1785 | } |
1786 | ||
1787 | static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set, | |
1788 | struct request_queue *q) | |
1789 | { | |
1790 | q->tag_set = set; | |
1791 | ||
1792 | mutex_lock(&set->tag_list_lock); | |
1793 | list_add_tail(&q->tag_set_list, &set->tag_list); | |
1794 | blk_mq_update_tag_set_depth(set); | |
1795 | mutex_unlock(&set->tag_list_lock); | |
1796 | } | |
1797 | ||
24d2f903 | 1798 | struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set) |
320ae51f JA |
1799 | { |
1800 | struct blk_mq_hw_ctx **hctxs; | |
e6cdb092 | 1801 | struct blk_mq_ctx __percpu *ctx; |
320ae51f | 1802 | struct request_queue *q; |
f14bbe77 | 1803 | unsigned int *map; |
320ae51f JA |
1804 | int i; |
1805 | ||
320ae51f JA |
1806 | ctx = alloc_percpu(struct blk_mq_ctx); |
1807 | if (!ctx) | |
1808 | return ERR_PTR(-ENOMEM); | |
1809 | ||
aedcd72f JA |
1810 | /* |
1811 | * If a crashdump is active, then we are potentially in a very | |
1812 | * memory constrained environment. Limit us to 1 queue and | |
1813 | * 64 tags to prevent using too much memory. | |
1814 | */ | |
1815 | if (is_kdump_kernel()) { | |
1816 | set->nr_hw_queues = 1; | |
1817 | set->queue_depth = min(64U, set->queue_depth); | |
1818 | } | |
1819 | ||
24d2f903 CH |
1820 | hctxs = kmalloc_node(set->nr_hw_queues * sizeof(*hctxs), GFP_KERNEL, |
1821 | set->numa_node); | |
320ae51f JA |
1822 | |
1823 | if (!hctxs) | |
1824 | goto err_percpu; | |
1825 | ||
f14bbe77 JA |
1826 | map = blk_mq_make_queue_map(set); |
1827 | if (!map) | |
1828 | goto err_map; | |
1829 | ||
24d2f903 | 1830 | for (i = 0; i < set->nr_hw_queues; i++) { |
f14bbe77 JA |
1831 | int node = blk_mq_hw_queue_to_node(map, i); |
1832 | ||
cdef54dd CH |
1833 | hctxs[i] = kzalloc_node(sizeof(struct blk_mq_hw_ctx), |
1834 | GFP_KERNEL, node); | |
320ae51f JA |
1835 | if (!hctxs[i]) |
1836 | goto err_hctxs; | |
1837 | ||
a86073e4 JA |
1838 | if (!zalloc_cpumask_var_node(&hctxs[i]->cpumask, GFP_KERNEL, |
1839 | node)) | |
e4043dcf JA |
1840 | goto err_hctxs; |
1841 | ||
0d2602ca | 1842 | atomic_set(&hctxs[i]->nr_active, 0); |
f14bbe77 | 1843 | hctxs[i]->numa_node = node; |
320ae51f JA |
1844 | hctxs[i]->queue_num = i; |
1845 | } | |
1846 | ||
24d2f903 | 1847 | q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node); |
320ae51f JA |
1848 | if (!q) |
1849 | goto err_hctxs; | |
1850 | ||
17497acb TH |
1851 | /* |
1852 | * Init percpu_ref in atomic mode so that it's faster to shutdown. | |
1853 | * See blk_register_queue() for details. | |
1854 | */ | |
a34375ef | 1855 | if (percpu_ref_init(&q->mq_usage_counter, blk_mq_usage_counter_release, |
17497acb | 1856 | PERCPU_REF_INIT_ATOMIC, GFP_KERNEL)) |
3d2936f4 ML |
1857 | goto err_map; |
1858 | ||
320ae51f JA |
1859 | setup_timer(&q->timeout, blk_mq_rq_timer, (unsigned long) q); |
1860 | blk_queue_rq_timeout(q, 30000); | |
1861 | ||
1862 | q->nr_queues = nr_cpu_ids; | |
24d2f903 | 1863 | q->nr_hw_queues = set->nr_hw_queues; |
f14bbe77 | 1864 | q->mq_map = map; |
320ae51f JA |
1865 | |
1866 | q->queue_ctx = ctx; | |
1867 | q->queue_hw_ctx = hctxs; | |
1868 | ||
24d2f903 | 1869 | q->mq_ops = set->ops; |
94eddfbe | 1870 | q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT; |
320ae51f | 1871 | |
05f1dd53 JA |
1872 | if (!(set->flags & BLK_MQ_F_SG_MERGE)) |
1873 | q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE; | |
1874 | ||
1be036e9 CH |
1875 | q->sg_reserved_size = INT_MAX; |
1876 | ||
6fca6a61 CH |
1877 | INIT_WORK(&q->requeue_work, blk_mq_requeue_work); |
1878 | INIT_LIST_HEAD(&q->requeue_list); | |
1879 | spin_lock_init(&q->requeue_lock); | |
1880 | ||
07068d5b JA |
1881 | if (q->nr_hw_queues > 1) |
1882 | blk_queue_make_request(q, blk_mq_make_request); | |
1883 | else | |
1884 | blk_queue_make_request(q, blk_sq_make_request); | |
1885 | ||
24d2f903 CH |
1886 | if (set->timeout) |
1887 | blk_queue_rq_timeout(q, set->timeout); | |
320ae51f | 1888 | |
eba71768 JA |
1889 | /* |
1890 | * Do this after blk_queue_make_request() overrides it... | |
1891 | */ | |
1892 | q->nr_requests = set->queue_depth; | |
1893 | ||
24d2f903 CH |
1894 | if (set->ops->complete) |
1895 | blk_queue_softirq_done(q, set->ops->complete); | |
30a91cb4 | 1896 | |
24d2f903 | 1897 | blk_mq_init_cpu_queues(q, set->nr_hw_queues); |
320ae51f | 1898 | |
24d2f903 | 1899 | if (blk_mq_init_hw_queues(q, set)) |
1bcb1ead | 1900 | goto err_hw; |
18741986 | 1901 | |
320ae51f JA |
1902 | mutex_lock(&all_q_mutex); |
1903 | list_add_tail(&q->all_q_node, &all_q_list); | |
1904 | mutex_unlock(&all_q_mutex); | |
1905 | ||
0d2602ca JA |
1906 | blk_mq_add_queue_tag_set(set, q); |
1907 | ||
484b4061 JA |
1908 | blk_mq_map_swqueue(q); |
1909 | ||
320ae51f | 1910 | return q; |
18741986 | 1911 | |
320ae51f | 1912 | err_hw: |
320ae51f JA |
1913 | blk_cleanup_queue(q); |
1914 | err_hctxs: | |
f14bbe77 | 1915 | kfree(map); |
24d2f903 | 1916 | for (i = 0; i < set->nr_hw_queues; i++) { |
320ae51f JA |
1917 | if (!hctxs[i]) |
1918 | break; | |
e4043dcf | 1919 | free_cpumask_var(hctxs[i]->cpumask); |
cdef54dd | 1920 | kfree(hctxs[i]); |
320ae51f | 1921 | } |
f14bbe77 | 1922 | err_map: |
320ae51f JA |
1923 | kfree(hctxs); |
1924 | err_percpu: | |
1925 | free_percpu(ctx); | |
1926 | return ERR_PTR(-ENOMEM); | |
1927 | } | |
1928 | EXPORT_SYMBOL(blk_mq_init_queue); | |
1929 | ||
1930 | void blk_mq_free_queue(struct request_queue *q) | |
1931 | { | |
624dbe47 | 1932 | struct blk_mq_tag_set *set = q->tag_set; |
320ae51f | 1933 | |
0d2602ca JA |
1934 | blk_mq_del_queue_tag_set(q); |
1935 | ||
624dbe47 ML |
1936 | blk_mq_exit_hw_queues(q, set, set->nr_hw_queues); |
1937 | blk_mq_free_hw_queues(q, set); | |
320ae51f | 1938 | |
add703fd | 1939 | percpu_ref_exit(&q->mq_usage_counter); |
3d2936f4 | 1940 | |
320ae51f JA |
1941 | free_percpu(q->queue_ctx); |
1942 | kfree(q->queue_hw_ctx); | |
1943 | kfree(q->mq_map); | |
1944 | ||
1945 | q->queue_ctx = NULL; | |
1946 | q->queue_hw_ctx = NULL; | |
1947 | q->mq_map = NULL; | |
1948 | ||
1949 | mutex_lock(&all_q_mutex); | |
1950 | list_del_init(&q->all_q_node); | |
1951 | mutex_unlock(&all_q_mutex); | |
1952 | } | |
320ae51f JA |
1953 | |
1954 | /* Basically redo blk_mq_init_queue with queue frozen */ | |
f618ef7c | 1955 | static void blk_mq_queue_reinit(struct request_queue *q) |
320ae51f JA |
1956 | { |
1957 | blk_mq_freeze_queue(q); | |
1958 | ||
67aec14c JA |
1959 | blk_mq_sysfs_unregister(q); |
1960 | ||
320ae51f JA |
1961 | blk_mq_update_queue_map(q->mq_map, q->nr_hw_queues); |
1962 | ||
1963 | /* | |
1964 | * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe | |
1965 | * we should change hctx numa_node according to new topology (this | |
1966 | * involves free and re-allocate memory, worthy doing?) | |
1967 | */ | |
1968 | ||
1969 | blk_mq_map_swqueue(q); | |
1970 | ||
67aec14c JA |
1971 | blk_mq_sysfs_register(q); |
1972 | ||
320ae51f JA |
1973 | blk_mq_unfreeze_queue(q); |
1974 | } | |
1975 | ||
f618ef7c PG |
1976 | static int blk_mq_queue_reinit_notify(struct notifier_block *nb, |
1977 | unsigned long action, void *hcpu) | |
320ae51f JA |
1978 | { |
1979 | struct request_queue *q; | |
1980 | ||
1981 | /* | |
9fccfed8 JA |
1982 | * Before new mappings are established, hotadded cpu might already |
1983 | * start handling requests. This doesn't break anything as we map | |
1984 | * offline CPUs to first hardware queue. We will re-init the queue | |
1985 | * below to get optimal settings. | |
320ae51f JA |
1986 | */ |
1987 | if (action != CPU_DEAD && action != CPU_DEAD_FROZEN && | |
1988 | action != CPU_ONLINE && action != CPU_ONLINE_FROZEN) | |
1989 | return NOTIFY_OK; | |
1990 | ||
1991 | mutex_lock(&all_q_mutex); | |
1992 | list_for_each_entry(q, &all_q_list, all_q_node) | |
1993 | blk_mq_queue_reinit(q); | |
1994 | mutex_unlock(&all_q_mutex); | |
1995 | return NOTIFY_OK; | |
1996 | } | |
1997 | ||
a5164405 JA |
1998 | static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set) |
1999 | { | |
2000 | int i; | |
2001 | ||
2002 | for (i = 0; i < set->nr_hw_queues; i++) { | |
2003 | set->tags[i] = blk_mq_init_rq_map(set, i); | |
2004 | if (!set->tags[i]) | |
2005 | goto out_unwind; | |
2006 | } | |
2007 | ||
2008 | return 0; | |
2009 | ||
2010 | out_unwind: | |
2011 | while (--i >= 0) | |
2012 | blk_mq_free_rq_map(set, set->tags[i], i); | |
2013 | ||
a5164405 JA |
2014 | return -ENOMEM; |
2015 | } | |
2016 | ||
2017 | /* | |
2018 | * Allocate the request maps associated with this tag_set. Note that this | |
2019 | * may reduce the depth asked for, if memory is tight. set->queue_depth | |
2020 | * will be updated to reflect the allocated depth. | |
2021 | */ | |
2022 | static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set) | |
2023 | { | |
2024 | unsigned int depth; | |
2025 | int err; | |
2026 | ||
2027 | depth = set->queue_depth; | |
2028 | do { | |
2029 | err = __blk_mq_alloc_rq_maps(set); | |
2030 | if (!err) | |
2031 | break; | |
2032 | ||
2033 | set->queue_depth >>= 1; | |
2034 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) { | |
2035 | err = -ENOMEM; | |
2036 | break; | |
2037 | } | |
2038 | } while (set->queue_depth); | |
2039 | ||
2040 | if (!set->queue_depth || err) { | |
2041 | pr_err("blk-mq: failed to allocate request map\n"); | |
2042 | return -ENOMEM; | |
2043 | } | |
2044 | ||
2045 | if (depth != set->queue_depth) | |
2046 | pr_info("blk-mq: reduced tag depth (%u -> %u)\n", | |
2047 | depth, set->queue_depth); | |
2048 | ||
2049 | return 0; | |
2050 | } | |
2051 | ||
a4391c64 JA |
2052 | /* |
2053 | * Alloc a tag set to be associated with one or more request queues. | |
2054 | * May fail with EINVAL for various error conditions. May adjust the | |
2055 | * requested depth down, if if it too large. In that case, the set | |
2056 | * value will be stored in set->queue_depth. | |
2057 | */ | |
24d2f903 CH |
2058 | int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set) |
2059 | { | |
24d2f903 CH |
2060 | if (!set->nr_hw_queues) |
2061 | return -EINVAL; | |
a4391c64 | 2062 | if (!set->queue_depth) |
24d2f903 CH |
2063 | return -EINVAL; |
2064 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) | |
2065 | return -EINVAL; | |
2066 | ||
cdef54dd | 2067 | if (!set->nr_hw_queues || !set->ops->queue_rq || !set->ops->map_queue) |
24d2f903 CH |
2068 | return -EINVAL; |
2069 | ||
a4391c64 JA |
2070 | if (set->queue_depth > BLK_MQ_MAX_DEPTH) { |
2071 | pr_info("blk-mq: reduced tag depth to %u\n", | |
2072 | BLK_MQ_MAX_DEPTH); | |
2073 | set->queue_depth = BLK_MQ_MAX_DEPTH; | |
2074 | } | |
24d2f903 | 2075 | |
48479005 ML |
2076 | set->tags = kmalloc_node(set->nr_hw_queues * |
2077 | sizeof(struct blk_mq_tags *), | |
24d2f903 CH |
2078 | GFP_KERNEL, set->numa_node); |
2079 | if (!set->tags) | |
a5164405 | 2080 | return -ENOMEM; |
24d2f903 | 2081 | |
a5164405 JA |
2082 | if (blk_mq_alloc_rq_maps(set)) |
2083 | goto enomem; | |
24d2f903 | 2084 | |
0d2602ca JA |
2085 | mutex_init(&set->tag_list_lock); |
2086 | INIT_LIST_HEAD(&set->tag_list); | |
2087 | ||
24d2f903 | 2088 | return 0; |
a5164405 | 2089 | enomem: |
5676e7b6 RE |
2090 | kfree(set->tags); |
2091 | set->tags = NULL; | |
24d2f903 CH |
2092 | return -ENOMEM; |
2093 | } | |
2094 | EXPORT_SYMBOL(blk_mq_alloc_tag_set); | |
2095 | ||
2096 | void blk_mq_free_tag_set(struct blk_mq_tag_set *set) | |
2097 | { | |
2098 | int i; | |
2099 | ||
484b4061 JA |
2100 | for (i = 0; i < set->nr_hw_queues; i++) { |
2101 | if (set->tags[i]) | |
2102 | blk_mq_free_rq_map(set, set->tags[i], i); | |
2103 | } | |
2104 | ||
981bd189 | 2105 | kfree(set->tags); |
5676e7b6 | 2106 | set->tags = NULL; |
24d2f903 CH |
2107 | } |
2108 | EXPORT_SYMBOL(blk_mq_free_tag_set); | |
2109 | ||
e3a2b3f9 JA |
2110 | int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr) |
2111 | { | |
2112 | struct blk_mq_tag_set *set = q->tag_set; | |
2113 | struct blk_mq_hw_ctx *hctx; | |
2114 | int i, ret; | |
2115 | ||
2116 | if (!set || nr > set->queue_depth) | |
2117 | return -EINVAL; | |
2118 | ||
2119 | ret = 0; | |
2120 | queue_for_each_hw_ctx(q, hctx, i) { | |
2121 | ret = blk_mq_tag_update_depth(hctx->tags, nr); | |
2122 | if (ret) | |
2123 | break; | |
2124 | } | |
2125 | ||
2126 | if (!ret) | |
2127 | q->nr_requests = nr; | |
2128 | ||
2129 | return ret; | |
2130 | } | |
2131 | ||
676141e4 JA |
2132 | void blk_mq_disable_hotplug(void) |
2133 | { | |
2134 | mutex_lock(&all_q_mutex); | |
2135 | } | |
2136 | ||
2137 | void blk_mq_enable_hotplug(void) | |
2138 | { | |
2139 | mutex_unlock(&all_q_mutex); | |
2140 | } | |
2141 | ||
320ae51f JA |
2142 | static int __init blk_mq_init(void) |
2143 | { | |
320ae51f JA |
2144 | blk_mq_cpu_init(); |
2145 | ||
add703fd | 2146 | hotcpu_notifier(blk_mq_queue_reinit_notify, 0); |
320ae51f JA |
2147 | |
2148 | return 0; | |
2149 | } | |
2150 | subsys_initcall(blk_mq_init); |