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> | |
23 | ||
24 | #include <trace/events/block.h> | |
25 | ||
26 | #include <linux/blk-mq.h> | |
27 | #include "blk.h" | |
28 | #include "blk-mq.h" | |
29 | #include "blk-mq-tag.h" | |
30 | ||
31 | static DEFINE_MUTEX(all_q_mutex); | |
32 | static LIST_HEAD(all_q_list); | |
33 | ||
34 | static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx); | |
35 | ||
320ae51f JA |
36 | /* |
37 | * Check if any of the ctx's have pending work in this hardware queue | |
38 | */ | |
39 | static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx) | |
40 | { | |
41 | unsigned int i; | |
42 | ||
1429d7c9 JA |
43 | for (i = 0; i < hctx->ctx_map.map_size; i++) |
44 | if (hctx->ctx_map.map[i].word) | |
320ae51f JA |
45 | return true; |
46 | ||
47 | return false; | |
48 | } | |
49 | ||
1429d7c9 JA |
50 | static inline struct blk_align_bitmap *get_bm(struct blk_mq_hw_ctx *hctx, |
51 | struct blk_mq_ctx *ctx) | |
52 | { | |
53 | return &hctx->ctx_map.map[ctx->index_hw / hctx->ctx_map.bits_per_word]; | |
54 | } | |
55 | ||
56 | #define CTX_TO_BIT(hctx, ctx) \ | |
57 | ((ctx)->index_hw & ((hctx)->ctx_map.bits_per_word - 1)) | |
58 | ||
320ae51f JA |
59 | /* |
60 | * Mark this ctx as having pending work in this hardware queue | |
61 | */ | |
62 | static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx, | |
63 | struct blk_mq_ctx *ctx) | |
64 | { | |
1429d7c9 JA |
65 | struct blk_align_bitmap *bm = get_bm(hctx, ctx); |
66 | ||
67 | if (!test_bit(CTX_TO_BIT(hctx, ctx), &bm->word)) | |
68 | set_bit(CTX_TO_BIT(hctx, ctx), &bm->word); | |
69 | } | |
70 | ||
71 | static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx, | |
72 | struct blk_mq_ctx *ctx) | |
73 | { | |
74 | struct blk_align_bitmap *bm = get_bm(hctx, ctx); | |
75 | ||
76 | clear_bit(CTX_TO_BIT(hctx, ctx), &bm->word); | |
320ae51f JA |
77 | } |
78 | ||
320ae51f JA |
79 | static int blk_mq_queue_enter(struct request_queue *q) |
80 | { | |
add703fd TH |
81 | while (true) { |
82 | int ret; | |
320ae51f | 83 | |
add703fd TH |
84 | if (percpu_ref_tryget_live(&q->mq_usage_counter)) |
85 | return 0; | |
320ae51f | 86 | |
add703fd TH |
87 | ret = wait_event_interruptible(q->mq_freeze_wq, |
88 | !q->mq_freeze_depth || blk_queue_dying(q)); | |
89 | if (blk_queue_dying(q)) | |
90 | return -ENODEV; | |
91 | if (ret) | |
92 | return ret; | |
93 | } | |
320ae51f JA |
94 | } |
95 | ||
96 | static void blk_mq_queue_exit(struct request_queue *q) | |
97 | { | |
add703fd TH |
98 | percpu_ref_put(&q->mq_usage_counter); |
99 | } | |
100 | ||
101 | static void blk_mq_usage_counter_release(struct percpu_ref *ref) | |
102 | { | |
103 | struct request_queue *q = | |
104 | container_of(ref, struct request_queue, mq_usage_counter); | |
105 | ||
106 | wake_up_all(&q->mq_freeze_wq); | |
320ae51f JA |
107 | } |
108 | ||
72d6f02a TH |
109 | /* |
110 | * Guarantee no request is in use, so we can change any data structure of | |
111 | * the queue afterward. | |
112 | */ | |
113 | void blk_mq_freeze_queue(struct request_queue *q) | |
43a5e4e2 | 114 | { |
72d6f02a TH |
115 | spin_lock_irq(q->queue_lock); |
116 | q->mq_freeze_depth++; | |
117 | spin_unlock_irq(q->queue_lock); | |
118 | ||
add703fd TH |
119 | percpu_ref_kill(&q->mq_usage_counter); |
120 | blk_mq_run_queues(q, false); | |
121 | wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->mq_usage_counter)); | |
43a5e4e2 ML |
122 | } |
123 | ||
320ae51f JA |
124 | static void blk_mq_unfreeze_queue(struct request_queue *q) |
125 | { | |
126 | bool wake = false; | |
127 | ||
128 | spin_lock_irq(q->queue_lock); | |
780db207 TH |
129 | wake = !--q->mq_freeze_depth; |
130 | WARN_ON_ONCE(q->mq_freeze_depth < 0); | |
320ae51f | 131 | spin_unlock_irq(q->queue_lock); |
add703fd TH |
132 | if (wake) { |
133 | percpu_ref_reinit(&q->mq_usage_counter); | |
320ae51f | 134 | wake_up_all(&q->mq_freeze_wq); |
add703fd | 135 | } |
320ae51f JA |
136 | } |
137 | ||
138 | bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx) | |
139 | { | |
140 | return blk_mq_has_free_tags(hctx->tags); | |
141 | } | |
142 | EXPORT_SYMBOL(blk_mq_can_queue); | |
143 | ||
94eddfbe JA |
144 | static void blk_mq_rq_ctx_init(struct request_queue *q, struct blk_mq_ctx *ctx, |
145 | struct request *rq, unsigned int rw_flags) | |
320ae51f | 146 | { |
94eddfbe JA |
147 | if (blk_queue_io_stat(q)) |
148 | rw_flags |= REQ_IO_STAT; | |
149 | ||
af76e555 CH |
150 | INIT_LIST_HEAD(&rq->queuelist); |
151 | /* csd/requeue_work/fifo_time is initialized before use */ | |
152 | rq->q = q; | |
320ae51f | 153 | rq->mq_ctx = ctx; |
0d2602ca | 154 | rq->cmd_flags |= rw_flags; |
af76e555 CH |
155 | /* do not touch atomic flags, it needs atomic ops against the timer */ |
156 | rq->cpu = -1; | |
af76e555 CH |
157 | INIT_HLIST_NODE(&rq->hash); |
158 | RB_CLEAR_NODE(&rq->rb_node); | |
af76e555 CH |
159 | rq->rq_disk = NULL; |
160 | rq->part = NULL; | |
3ee32372 | 161 | rq->start_time = jiffies; |
af76e555 CH |
162 | #ifdef CONFIG_BLK_CGROUP |
163 | rq->rl = NULL; | |
0fec08b4 | 164 | set_start_time_ns(rq); |
af76e555 CH |
165 | rq->io_start_time_ns = 0; |
166 | #endif | |
167 | rq->nr_phys_segments = 0; | |
168 | #if defined(CONFIG_BLK_DEV_INTEGRITY) | |
169 | rq->nr_integrity_segments = 0; | |
170 | #endif | |
af76e555 CH |
171 | rq->special = NULL; |
172 | /* tag was already set */ | |
173 | rq->errors = 0; | |
af76e555 CH |
174 | |
175 | rq->extra_len = 0; | |
176 | rq->sense_len = 0; | |
177 | rq->resid_len = 0; | |
178 | rq->sense = NULL; | |
179 | ||
af76e555 | 180 | INIT_LIST_HEAD(&rq->timeout_list); |
f6be4fb4 JA |
181 | rq->timeout = 0; |
182 | ||
af76e555 CH |
183 | rq->end_io = NULL; |
184 | rq->end_io_data = NULL; | |
185 | rq->next_rq = NULL; | |
186 | ||
320ae51f JA |
187 | ctx->rq_dispatched[rw_is_sync(rw_flags)]++; |
188 | } | |
189 | ||
5dee8577 | 190 | static struct request * |
cb96a42c | 191 | __blk_mq_alloc_request(struct blk_mq_alloc_data *data, int rw) |
5dee8577 CH |
192 | { |
193 | struct request *rq; | |
194 | unsigned int tag; | |
195 | ||
cb96a42c | 196 | tag = blk_mq_get_tag(data); |
5dee8577 | 197 | if (tag != BLK_MQ_TAG_FAIL) { |
cb96a42c | 198 | rq = data->hctx->tags->rqs[tag]; |
5dee8577 CH |
199 | |
200 | rq->cmd_flags = 0; | |
cb96a42c | 201 | if (blk_mq_tag_busy(data->hctx)) { |
5dee8577 | 202 | rq->cmd_flags = REQ_MQ_INFLIGHT; |
cb96a42c | 203 | atomic_inc(&data->hctx->nr_active); |
5dee8577 CH |
204 | } |
205 | ||
206 | rq->tag = tag; | |
cb96a42c | 207 | blk_mq_rq_ctx_init(data->q, data->ctx, rq, rw); |
5dee8577 CH |
208 | return rq; |
209 | } | |
210 | ||
211 | return NULL; | |
212 | } | |
213 | ||
4ce01dd1 CH |
214 | struct request *blk_mq_alloc_request(struct request_queue *q, int rw, gfp_t gfp, |
215 | bool reserved) | |
320ae51f | 216 | { |
d852564f CH |
217 | struct blk_mq_ctx *ctx; |
218 | struct blk_mq_hw_ctx *hctx; | |
320ae51f | 219 | struct request *rq; |
cb96a42c | 220 | struct blk_mq_alloc_data alloc_data; |
a492f075 | 221 | int ret; |
320ae51f | 222 | |
a492f075 JL |
223 | ret = blk_mq_queue_enter(q); |
224 | if (ret) | |
225 | return ERR_PTR(ret); | |
320ae51f | 226 | |
d852564f CH |
227 | ctx = blk_mq_get_ctx(q); |
228 | hctx = q->mq_ops->map_queue(q, ctx->cpu); | |
cb96a42c ML |
229 | blk_mq_set_alloc_data(&alloc_data, q, gfp & ~__GFP_WAIT, |
230 | reserved, ctx, hctx); | |
d852564f | 231 | |
cb96a42c | 232 | rq = __blk_mq_alloc_request(&alloc_data, rw); |
d852564f CH |
233 | if (!rq && (gfp & __GFP_WAIT)) { |
234 | __blk_mq_run_hw_queue(hctx); | |
235 | blk_mq_put_ctx(ctx); | |
236 | ||
237 | ctx = blk_mq_get_ctx(q); | |
238 | hctx = q->mq_ops->map_queue(q, ctx->cpu); | |
cb96a42c ML |
239 | blk_mq_set_alloc_data(&alloc_data, q, gfp, reserved, ctx, |
240 | hctx); | |
241 | rq = __blk_mq_alloc_request(&alloc_data, rw); | |
242 | ctx = alloc_data.ctx; | |
d852564f CH |
243 | } |
244 | blk_mq_put_ctx(ctx); | |
a492f075 JL |
245 | if (!rq) |
246 | return ERR_PTR(-EWOULDBLOCK); | |
320ae51f JA |
247 | return rq; |
248 | } | |
4bb659b1 | 249 | EXPORT_SYMBOL(blk_mq_alloc_request); |
320ae51f | 250 | |
320ae51f JA |
251 | static void __blk_mq_free_request(struct blk_mq_hw_ctx *hctx, |
252 | struct blk_mq_ctx *ctx, struct request *rq) | |
253 | { | |
254 | const int tag = rq->tag; | |
255 | struct request_queue *q = rq->q; | |
256 | ||
0d2602ca JA |
257 | if (rq->cmd_flags & REQ_MQ_INFLIGHT) |
258 | atomic_dec(&hctx->nr_active); | |
259 | ||
af76e555 | 260 | clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags); |
0d2602ca | 261 | blk_mq_put_tag(hctx, tag, &ctx->last_tag); |
320ae51f JA |
262 | blk_mq_queue_exit(q); |
263 | } | |
264 | ||
265 | void blk_mq_free_request(struct request *rq) | |
266 | { | |
267 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
268 | struct blk_mq_hw_ctx *hctx; | |
269 | struct request_queue *q = rq->q; | |
270 | ||
271 | ctx->rq_completed[rq_is_sync(rq)]++; | |
272 | ||
273 | hctx = q->mq_ops->map_queue(q, ctx->cpu); | |
274 | __blk_mq_free_request(hctx, ctx, rq); | |
275 | } | |
276 | ||
8727af4b CH |
277 | /* |
278 | * Clone all relevant state from a request that has been put on hold in | |
279 | * the flush state machine into the preallocated flush request that hangs | |
280 | * off the request queue. | |
281 | * | |
282 | * For a driver the flush request should be invisible, that's why we are | |
283 | * impersonating the original request here. | |
284 | */ | |
285 | void blk_mq_clone_flush_request(struct request *flush_rq, | |
286 | struct request *orig_rq) | |
287 | { | |
288 | struct blk_mq_hw_ctx *hctx = | |
289 | orig_rq->q->mq_ops->map_queue(orig_rq->q, orig_rq->mq_ctx->cpu); | |
290 | ||
291 | flush_rq->mq_ctx = orig_rq->mq_ctx; | |
292 | flush_rq->tag = orig_rq->tag; | |
293 | memcpy(blk_mq_rq_to_pdu(flush_rq), blk_mq_rq_to_pdu(orig_rq), | |
294 | hctx->cmd_size); | |
295 | } | |
296 | ||
63151a44 | 297 | inline void __blk_mq_end_io(struct request *rq, int error) |
320ae51f | 298 | { |
0d11e6ac ML |
299 | blk_account_io_done(rq); |
300 | ||
91b63639 | 301 | if (rq->end_io) { |
320ae51f | 302 | rq->end_io(rq, error); |
91b63639 CH |
303 | } else { |
304 | if (unlikely(blk_bidi_rq(rq))) | |
305 | blk_mq_free_request(rq->next_rq); | |
320ae51f | 306 | blk_mq_free_request(rq); |
91b63639 | 307 | } |
320ae51f | 308 | } |
63151a44 CH |
309 | EXPORT_SYMBOL(__blk_mq_end_io); |
310 | ||
311 | void blk_mq_end_io(struct request *rq, int error) | |
312 | { | |
313 | if (blk_update_request(rq, error, blk_rq_bytes(rq))) | |
314 | BUG(); | |
315 | __blk_mq_end_io(rq, error); | |
316 | } | |
317 | EXPORT_SYMBOL(blk_mq_end_io); | |
320ae51f | 318 | |
30a91cb4 | 319 | static void __blk_mq_complete_request_remote(void *data) |
320ae51f | 320 | { |
3d6efbf6 | 321 | struct request *rq = data; |
320ae51f | 322 | |
30a91cb4 | 323 | rq->q->softirq_done_fn(rq); |
320ae51f | 324 | } |
320ae51f | 325 | |
ed851860 | 326 | static void blk_mq_ipi_complete_request(struct request *rq) |
320ae51f JA |
327 | { |
328 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
38535201 | 329 | bool shared = false; |
320ae51f JA |
330 | int cpu; |
331 | ||
38535201 | 332 | if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) { |
30a91cb4 CH |
333 | rq->q->softirq_done_fn(rq); |
334 | return; | |
335 | } | |
320ae51f JA |
336 | |
337 | cpu = get_cpu(); | |
38535201 CH |
338 | if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags)) |
339 | shared = cpus_share_cache(cpu, ctx->cpu); | |
340 | ||
341 | if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) { | |
30a91cb4 | 342 | rq->csd.func = __blk_mq_complete_request_remote; |
3d6efbf6 CH |
343 | rq->csd.info = rq; |
344 | rq->csd.flags = 0; | |
c46fff2a | 345 | smp_call_function_single_async(ctx->cpu, &rq->csd); |
3d6efbf6 | 346 | } else { |
30a91cb4 | 347 | rq->q->softirq_done_fn(rq); |
3d6efbf6 | 348 | } |
320ae51f JA |
349 | put_cpu(); |
350 | } | |
30a91cb4 | 351 | |
ed851860 JA |
352 | void __blk_mq_complete_request(struct request *rq) |
353 | { | |
354 | struct request_queue *q = rq->q; | |
355 | ||
356 | if (!q->softirq_done_fn) | |
357 | blk_mq_end_io(rq, rq->errors); | |
358 | else | |
359 | blk_mq_ipi_complete_request(rq); | |
360 | } | |
361 | ||
30a91cb4 CH |
362 | /** |
363 | * blk_mq_complete_request - end I/O on a request | |
364 | * @rq: the request being processed | |
365 | * | |
366 | * Description: | |
367 | * Ends all I/O on a request. It does not handle partial completions. | |
368 | * The actual completion happens out-of-order, through a IPI handler. | |
369 | **/ | |
370 | void blk_mq_complete_request(struct request *rq) | |
371 | { | |
95f09684 JA |
372 | struct request_queue *q = rq->q; |
373 | ||
374 | if (unlikely(blk_should_fake_timeout(q))) | |
30a91cb4 | 375 | return; |
ed851860 JA |
376 | if (!blk_mark_rq_complete(rq)) |
377 | __blk_mq_complete_request(rq); | |
30a91cb4 CH |
378 | } |
379 | EXPORT_SYMBOL(blk_mq_complete_request); | |
320ae51f | 380 | |
49f5baa5 | 381 | static void blk_mq_start_request(struct request *rq, bool last) |
320ae51f JA |
382 | { |
383 | struct request_queue *q = rq->q; | |
384 | ||
385 | trace_block_rq_issue(q, rq); | |
386 | ||
742ee69b | 387 | rq->resid_len = blk_rq_bytes(rq); |
91b63639 CH |
388 | if (unlikely(blk_bidi_rq(rq))) |
389 | rq->next_rq->resid_len = blk_rq_bytes(rq->next_rq); | |
742ee69b | 390 | |
2b8393b4 | 391 | blk_add_timer(rq); |
87ee7b11 JA |
392 | |
393 | /* | |
394 | * Mark us as started and clear complete. Complete might have been | |
395 | * set if requeue raced with timeout, which then marked it as | |
396 | * complete. So be sure to clear complete again when we start | |
397 | * the request, otherwise we'll ignore the completion event. | |
398 | */ | |
4b570521 JA |
399 | if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) |
400 | set_bit(REQ_ATOM_STARTED, &rq->atomic_flags); | |
401 | if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags)) | |
402 | clear_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags); | |
49f5baa5 CH |
403 | |
404 | if (q->dma_drain_size && blk_rq_bytes(rq)) { | |
405 | /* | |
406 | * Make sure space for the drain appears. We know we can do | |
407 | * this because max_hw_segments has been adjusted to be one | |
408 | * fewer than the device can handle. | |
409 | */ | |
410 | rq->nr_phys_segments++; | |
411 | } | |
412 | ||
413 | /* | |
414 | * Flag the last request in the series so that drivers know when IO | |
415 | * should be kicked off, if they don't do it on a per-request basis. | |
416 | * | |
417 | * Note: the flag isn't the only condition drivers should do kick off. | |
418 | * If drive is busy, the last request might not have the bit set. | |
419 | */ | |
420 | if (last) | |
421 | rq->cmd_flags |= REQ_END; | |
320ae51f JA |
422 | } |
423 | ||
ed0791b2 | 424 | static void __blk_mq_requeue_request(struct request *rq) |
320ae51f JA |
425 | { |
426 | struct request_queue *q = rq->q; | |
427 | ||
428 | trace_block_rq_requeue(q, rq); | |
429 | clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags); | |
49f5baa5 CH |
430 | |
431 | rq->cmd_flags &= ~REQ_END; | |
432 | ||
433 | if (q->dma_drain_size && blk_rq_bytes(rq)) | |
434 | rq->nr_phys_segments--; | |
320ae51f JA |
435 | } |
436 | ||
ed0791b2 CH |
437 | void blk_mq_requeue_request(struct request *rq) |
438 | { | |
ed0791b2 CH |
439 | __blk_mq_requeue_request(rq); |
440 | blk_clear_rq_complete(rq); | |
441 | ||
ed0791b2 | 442 | BUG_ON(blk_queued_rq(rq)); |
6fca6a61 | 443 | blk_mq_add_to_requeue_list(rq, true); |
ed0791b2 CH |
444 | } |
445 | EXPORT_SYMBOL(blk_mq_requeue_request); | |
446 | ||
6fca6a61 CH |
447 | static void blk_mq_requeue_work(struct work_struct *work) |
448 | { | |
449 | struct request_queue *q = | |
450 | container_of(work, struct request_queue, requeue_work); | |
451 | LIST_HEAD(rq_list); | |
452 | struct request *rq, *next; | |
453 | unsigned long flags; | |
454 | ||
455 | spin_lock_irqsave(&q->requeue_lock, flags); | |
456 | list_splice_init(&q->requeue_list, &rq_list); | |
457 | spin_unlock_irqrestore(&q->requeue_lock, flags); | |
458 | ||
459 | list_for_each_entry_safe(rq, next, &rq_list, queuelist) { | |
460 | if (!(rq->cmd_flags & REQ_SOFTBARRIER)) | |
461 | continue; | |
462 | ||
463 | rq->cmd_flags &= ~REQ_SOFTBARRIER; | |
464 | list_del_init(&rq->queuelist); | |
465 | blk_mq_insert_request(rq, true, false, false); | |
466 | } | |
467 | ||
468 | while (!list_empty(&rq_list)) { | |
469 | rq = list_entry(rq_list.next, struct request, queuelist); | |
470 | list_del_init(&rq->queuelist); | |
471 | blk_mq_insert_request(rq, false, false, false); | |
472 | } | |
473 | ||
474 | blk_mq_run_queues(q, false); | |
475 | } | |
476 | ||
477 | void blk_mq_add_to_requeue_list(struct request *rq, bool at_head) | |
478 | { | |
479 | struct request_queue *q = rq->q; | |
480 | unsigned long flags; | |
481 | ||
482 | /* | |
483 | * We abuse this flag that is otherwise used by the I/O scheduler to | |
484 | * request head insertation from the workqueue. | |
485 | */ | |
486 | BUG_ON(rq->cmd_flags & REQ_SOFTBARRIER); | |
487 | ||
488 | spin_lock_irqsave(&q->requeue_lock, flags); | |
489 | if (at_head) { | |
490 | rq->cmd_flags |= REQ_SOFTBARRIER; | |
491 | list_add(&rq->queuelist, &q->requeue_list); | |
492 | } else { | |
493 | list_add_tail(&rq->queuelist, &q->requeue_list); | |
494 | } | |
495 | spin_unlock_irqrestore(&q->requeue_lock, flags); | |
496 | } | |
497 | EXPORT_SYMBOL(blk_mq_add_to_requeue_list); | |
498 | ||
499 | void blk_mq_kick_requeue_list(struct request_queue *q) | |
500 | { | |
501 | kblockd_schedule_work(&q->requeue_work); | |
502 | } | |
503 | EXPORT_SYMBOL(blk_mq_kick_requeue_list); | |
504 | ||
0e62f51f | 505 | static inline bool is_flush_request(struct request *rq, unsigned int tag) |
24d2f903 | 506 | { |
0e62f51f JA |
507 | return ((rq->cmd_flags & REQ_FLUSH_SEQ) && |
508 | rq->q->flush_rq->tag == tag); | |
509 | } | |
510 | ||
511 | struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag) | |
512 | { | |
513 | struct request *rq = tags->rqs[tag]; | |
22302375 | 514 | |
0e62f51f JA |
515 | if (!is_flush_request(rq, tag)) |
516 | return rq; | |
22302375 | 517 | |
0e62f51f | 518 | return rq->q->flush_rq; |
24d2f903 CH |
519 | } |
520 | EXPORT_SYMBOL(blk_mq_tag_to_rq); | |
521 | ||
320ae51f JA |
522 | struct blk_mq_timeout_data { |
523 | struct blk_mq_hw_ctx *hctx; | |
524 | unsigned long *next; | |
525 | unsigned int *next_set; | |
526 | }; | |
527 | ||
528 | static void blk_mq_timeout_check(void *__data, unsigned long *free_tags) | |
529 | { | |
530 | struct blk_mq_timeout_data *data = __data; | |
531 | struct blk_mq_hw_ctx *hctx = data->hctx; | |
532 | unsigned int tag; | |
533 | ||
534 | /* It may not be in flight yet (this is where | |
535 | * the REQ_ATOMIC_STARTED flag comes in). The requests are | |
536 | * statically allocated, so we know it's always safe to access the | |
537 | * memory associated with a bit offset into ->rqs[]. | |
538 | */ | |
539 | tag = 0; | |
540 | do { | |
541 | struct request *rq; | |
542 | ||
24d2f903 CH |
543 | tag = find_next_zero_bit(free_tags, hctx->tags->nr_tags, tag); |
544 | if (tag >= hctx->tags->nr_tags) | |
320ae51f JA |
545 | break; |
546 | ||
0e62f51f | 547 | rq = blk_mq_tag_to_rq(hctx->tags, tag++); |
24d2f903 CH |
548 | if (rq->q != hctx->queue) |
549 | continue; | |
320ae51f JA |
550 | if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) |
551 | continue; | |
552 | ||
553 | blk_rq_check_expired(rq, data->next, data->next_set); | |
554 | } while (1); | |
555 | } | |
556 | ||
557 | static void blk_mq_hw_ctx_check_timeout(struct blk_mq_hw_ctx *hctx, | |
558 | unsigned long *next, | |
559 | unsigned int *next_set) | |
560 | { | |
561 | struct blk_mq_timeout_data data = { | |
562 | .hctx = hctx, | |
563 | .next = next, | |
564 | .next_set = next_set, | |
565 | }; | |
566 | ||
567 | /* | |
568 | * Ask the tagging code to iterate busy requests, so we can | |
569 | * check them for timeout. | |
570 | */ | |
571 | blk_mq_tag_busy_iter(hctx->tags, blk_mq_timeout_check, &data); | |
572 | } | |
573 | ||
87ee7b11 JA |
574 | static enum blk_eh_timer_return blk_mq_rq_timed_out(struct request *rq) |
575 | { | |
576 | struct request_queue *q = rq->q; | |
577 | ||
578 | /* | |
579 | * We know that complete is set at this point. If STARTED isn't set | |
580 | * anymore, then the request isn't active and the "timeout" should | |
581 | * just be ignored. This can happen due to the bitflag ordering. | |
582 | * Timeout first checks if STARTED is set, and if it is, assumes | |
583 | * the request is active. But if we race with completion, then | |
584 | * we both flags will get cleared. So check here again, and ignore | |
585 | * a timeout event with a request that isn't active. | |
586 | */ | |
587 | if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) | |
588 | return BLK_EH_NOT_HANDLED; | |
589 | ||
590 | if (!q->mq_ops->timeout) | |
591 | return BLK_EH_RESET_TIMER; | |
592 | ||
593 | return q->mq_ops->timeout(rq); | |
594 | } | |
595 | ||
320ae51f JA |
596 | static void blk_mq_rq_timer(unsigned long data) |
597 | { | |
598 | struct request_queue *q = (struct request_queue *) data; | |
599 | struct blk_mq_hw_ctx *hctx; | |
600 | unsigned long next = 0; | |
601 | int i, next_set = 0; | |
602 | ||
484b4061 JA |
603 | queue_for_each_hw_ctx(q, hctx, i) { |
604 | /* | |
605 | * If not software queues are currently mapped to this | |
606 | * hardware queue, there's nothing to check | |
607 | */ | |
608 | if (!hctx->nr_ctx || !hctx->tags) | |
609 | continue; | |
610 | ||
320ae51f | 611 | blk_mq_hw_ctx_check_timeout(hctx, &next, &next_set); |
484b4061 | 612 | } |
320ae51f | 613 | |
0d2602ca JA |
614 | if (next_set) { |
615 | next = blk_rq_timeout(round_jiffies_up(next)); | |
616 | mod_timer(&q->timeout, next); | |
617 | } else { | |
618 | queue_for_each_hw_ctx(q, hctx, i) | |
619 | blk_mq_tag_idle(hctx); | |
620 | } | |
320ae51f JA |
621 | } |
622 | ||
623 | /* | |
624 | * Reverse check our software queue for entries that we could potentially | |
625 | * merge with. Currently includes a hand-wavy stop count of 8, to not spend | |
626 | * too much time checking for merges. | |
627 | */ | |
628 | static bool blk_mq_attempt_merge(struct request_queue *q, | |
629 | struct blk_mq_ctx *ctx, struct bio *bio) | |
630 | { | |
631 | struct request *rq; | |
632 | int checked = 8; | |
633 | ||
634 | list_for_each_entry_reverse(rq, &ctx->rq_list, queuelist) { | |
635 | int el_ret; | |
636 | ||
637 | if (!checked--) | |
638 | break; | |
639 | ||
640 | if (!blk_rq_merge_ok(rq, bio)) | |
641 | continue; | |
642 | ||
643 | el_ret = blk_try_merge(rq, bio); | |
644 | if (el_ret == ELEVATOR_BACK_MERGE) { | |
645 | if (bio_attempt_back_merge(q, rq, bio)) { | |
646 | ctx->rq_merged++; | |
647 | return true; | |
648 | } | |
649 | break; | |
650 | } else if (el_ret == ELEVATOR_FRONT_MERGE) { | |
651 | if (bio_attempt_front_merge(q, rq, bio)) { | |
652 | ctx->rq_merged++; | |
653 | return true; | |
654 | } | |
655 | break; | |
656 | } | |
657 | } | |
658 | ||
659 | return false; | |
660 | } | |
661 | ||
1429d7c9 JA |
662 | /* |
663 | * Process software queues that have been marked busy, splicing them | |
664 | * to the for-dispatch | |
665 | */ | |
666 | static void flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list) | |
667 | { | |
668 | struct blk_mq_ctx *ctx; | |
669 | int i; | |
670 | ||
671 | for (i = 0; i < hctx->ctx_map.map_size; i++) { | |
672 | struct blk_align_bitmap *bm = &hctx->ctx_map.map[i]; | |
673 | unsigned int off, bit; | |
674 | ||
675 | if (!bm->word) | |
676 | continue; | |
677 | ||
678 | bit = 0; | |
679 | off = i * hctx->ctx_map.bits_per_word; | |
680 | do { | |
681 | bit = find_next_bit(&bm->word, bm->depth, bit); | |
682 | if (bit >= bm->depth) | |
683 | break; | |
684 | ||
685 | ctx = hctx->ctxs[bit + off]; | |
686 | clear_bit(bit, &bm->word); | |
687 | spin_lock(&ctx->lock); | |
688 | list_splice_tail_init(&ctx->rq_list, list); | |
689 | spin_unlock(&ctx->lock); | |
690 | ||
691 | bit++; | |
692 | } while (1); | |
693 | } | |
694 | } | |
695 | ||
320ae51f JA |
696 | /* |
697 | * Run this hardware queue, pulling any software queues mapped to it in. | |
698 | * Note that this function currently has various problems around ordering | |
699 | * of IO. In particular, we'd like FIFO behaviour on handling existing | |
700 | * items on the hctx->dispatch list. Ignore that for now. | |
701 | */ | |
702 | static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx) | |
703 | { | |
704 | struct request_queue *q = hctx->queue; | |
320ae51f JA |
705 | struct request *rq; |
706 | LIST_HEAD(rq_list); | |
1429d7c9 | 707 | int queued; |
320ae51f | 708 | |
fd1270d5 | 709 | WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask)); |
e4043dcf | 710 | |
5d12f905 | 711 | if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state))) |
320ae51f JA |
712 | return; |
713 | ||
714 | hctx->run++; | |
715 | ||
716 | /* | |
717 | * Touch any software queue that has pending entries. | |
718 | */ | |
1429d7c9 | 719 | flush_busy_ctxs(hctx, &rq_list); |
320ae51f JA |
720 | |
721 | /* | |
722 | * If we have previous entries on our dispatch list, grab them | |
723 | * and stuff them at the front for more fair dispatch. | |
724 | */ | |
725 | if (!list_empty_careful(&hctx->dispatch)) { | |
726 | spin_lock(&hctx->lock); | |
727 | if (!list_empty(&hctx->dispatch)) | |
728 | list_splice_init(&hctx->dispatch, &rq_list); | |
729 | spin_unlock(&hctx->lock); | |
730 | } | |
731 | ||
320ae51f JA |
732 | /* |
733 | * Now process all the entries, sending them to the driver. | |
734 | */ | |
1429d7c9 | 735 | queued = 0; |
320ae51f JA |
736 | while (!list_empty(&rq_list)) { |
737 | int ret; | |
738 | ||
739 | rq = list_first_entry(&rq_list, struct request, queuelist); | |
740 | list_del_init(&rq->queuelist); | |
320ae51f | 741 | |
49f5baa5 | 742 | blk_mq_start_request(rq, list_empty(&rq_list)); |
320ae51f JA |
743 | |
744 | ret = q->mq_ops->queue_rq(hctx, rq); | |
745 | switch (ret) { | |
746 | case BLK_MQ_RQ_QUEUE_OK: | |
747 | queued++; | |
748 | continue; | |
749 | case BLK_MQ_RQ_QUEUE_BUSY: | |
320ae51f | 750 | list_add(&rq->queuelist, &rq_list); |
ed0791b2 | 751 | __blk_mq_requeue_request(rq); |
320ae51f JA |
752 | break; |
753 | default: | |
754 | pr_err("blk-mq: bad return on queue: %d\n", ret); | |
320ae51f | 755 | case BLK_MQ_RQ_QUEUE_ERROR: |
1e93b8c2 | 756 | rq->errors = -EIO; |
320ae51f JA |
757 | blk_mq_end_io(rq, rq->errors); |
758 | break; | |
759 | } | |
760 | ||
761 | if (ret == BLK_MQ_RQ_QUEUE_BUSY) | |
762 | break; | |
763 | } | |
764 | ||
765 | if (!queued) | |
766 | hctx->dispatched[0]++; | |
767 | else if (queued < (1 << (BLK_MQ_MAX_DISPATCH_ORDER - 1))) | |
768 | hctx->dispatched[ilog2(queued) + 1]++; | |
769 | ||
770 | /* | |
771 | * Any items that need requeuing? Stuff them into hctx->dispatch, | |
772 | * that is where we will continue on next queue run. | |
773 | */ | |
774 | if (!list_empty(&rq_list)) { | |
775 | spin_lock(&hctx->lock); | |
776 | list_splice(&rq_list, &hctx->dispatch); | |
777 | spin_unlock(&hctx->lock); | |
778 | } | |
779 | } | |
780 | ||
506e931f JA |
781 | /* |
782 | * It'd be great if the workqueue API had a way to pass | |
783 | * in a mask and had some smarts for more clever placement. | |
784 | * For now we just round-robin here, switching for every | |
785 | * BLK_MQ_CPU_WORK_BATCH queued items. | |
786 | */ | |
787 | static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx) | |
788 | { | |
789 | int cpu = hctx->next_cpu; | |
790 | ||
791 | if (--hctx->next_cpu_batch <= 0) { | |
792 | int next_cpu; | |
793 | ||
794 | next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask); | |
795 | if (next_cpu >= nr_cpu_ids) | |
796 | next_cpu = cpumask_first(hctx->cpumask); | |
797 | ||
798 | hctx->next_cpu = next_cpu; | |
799 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; | |
800 | } | |
801 | ||
802 | return cpu; | |
803 | } | |
804 | ||
320ae51f JA |
805 | void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) |
806 | { | |
5d12f905 | 807 | if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state))) |
320ae51f JA |
808 | return; |
809 | ||
e4043dcf | 810 | if (!async && cpumask_test_cpu(smp_processor_id(), hctx->cpumask)) |
320ae51f | 811 | __blk_mq_run_hw_queue(hctx); |
e4043dcf | 812 | else if (hctx->queue->nr_hw_queues == 1) |
70f4db63 | 813 | kblockd_schedule_delayed_work(&hctx->run_work, 0); |
e4043dcf JA |
814 | else { |
815 | unsigned int cpu; | |
816 | ||
506e931f | 817 | cpu = blk_mq_hctx_next_cpu(hctx); |
70f4db63 | 818 | kblockd_schedule_delayed_work_on(cpu, &hctx->run_work, 0); |
e4043dcf | 819 | } |
320ae51f JA |
820 | } |
821 | ||
822 | void blk_mq_run_queues(struct request_queue *q, bool async) | |
823 | { | |
824 | struct blk_mq_hw_ctx *hctx; | |
825 | int i; | |
826 | ||
827 | queue_for_each_hw_ctx(q, hctx, i) { | |
828 | if ((!blk_mq_hctx_has_pending(hctx) && | |
829 | list_empty_careful(&hctx->dispatch)) || | |
5d12f905 | 830 | test_bit(BLK_MQ_S_STOPPED, &hctx->state)) |
320ae51f JA |
831 | continue; |
832 | ||
e4043dcf | 833 | preempt_disable(); |
320ae51f | 834 | blk_mq_run_hw_queue(hctx, async); |
e4043dcf | 835 | preempt_enable(); |
320ae51f JA |
836 | } |
837 | } | |
838 | EXPORT_SYMBOL(blk_mq_run_queues); | |
839 | ||
840 | void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx) | |
841 | { | |
70f4db63 CH |
842 | cancel_delayed_work(&hctx->run_work); |
843 | cancel_delayed_work(&hctx->delay_work); | |
320ae51f JA |
844 | set_bit(BLK_MQ_S_STOPPED, &hctx->state); |
845 | } | |
846 | EXPORT_SYMBOL(blk_mq_stop_hw_queue); | |
847 | ||
280d45f6 CH |
848 | void blk_mq_stop_hw_queues(struct request_queue *q) |
849 | { | |
850 | struct blk_mq_hw_ctx *hctx; | |
851 | int i; | |
852 | ||
853 | queue_for_each_hw_ctx(q, hctx, i) | |
854 | blk_mq_stop_hw_queue(hctx); | |
855 | } | |
856 | EXPORT_SYMBOL(blk_mq_stop_hw_queues); | |
857 | ||
320ae51f JA |
858 | void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx) |
859 | { | |
860 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
e4043dcf JA |
861 | |
862 | preempt_disable(); | |
0ffbce80 | 863 | blk_mq_run_hw_queue(hctx, false); |
e4043dcf | 864 | preempt_enable(); |
320ae51f JA |
865 | } |
866 | EXPORT_SYMBOL(blk_mq_start_hw_queue); | |
867 | ||
2f268556 CH |
868 | void blk_mq_start_hw_queues(struct request_queue *q) |
869 | { | |
870 | struct blk_mq_hw_ctx *hctx; | |
871 | int i; | |
872 | ||
873 | queue_for_each_hw_ctx(q, hctx, i) | |
874 | blk_mq_start_hw_queue(hctx); | |
875 | } | |
876 | EXPORT_SYMBOL(blk_mq_start_hw_queues); | |
877 | ||
878 | ||
1b4a3258 | 879 | void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async) |
320ae51f JA |
880 | { |
881 | struct blk_mq_hw_ctx *hctx; | |
882 | int i; | |
883 | ||
884 | queue_for_each_hw_ctx(q, hctx, i) { | |
885 | if (!test_bit(BLK_MQ_S_STOPPED, &hctx->state)) | |
886 | continue; | |
887 | ||
888 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
e4043dcf | 889 | preempt_disable(); |
1b4a3258 | 890 | blk_mq_run_hw_queue(hctx, async); |
e4043dcf | 891 | preempt_enable(); |
320ae51f JA |
892 | } |
893 | } | |
894 | EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues); | |
895 | ||
70f4db63 | 896 | static void blk_mq_run_work_fn(struct work_struct *work) |
320ae51f JA |
897 | { |
898 | struct blk_mq_hw_ctx *hctx; | |
899 | ||
70f4db63 | 900 | hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work); |
e4043dcf | 901 | |
320ae51f JA |
902 | __blk_mq_run_hw_queue(hctx); |
903 | } | |
904 | ||
70f4db63 CH |
905 | static void blk_mq_delay_work_fn(struct work_struct *work) |
906 | { | |
907 | struct blk_mq_hw_ctx *hctx; | |
908 | ||
909 | hctx = container_of(work, struct blk_mq_hw_ctx, delay_work.work); | |
910 | ||
911 | if (test_and_clear_bit(BLK_MQ_S_STOPPED, &hctx->state)) | |
912 | __blk_mq_run_hw_queue(hctx); | |
913 | } | |
914 | ||
915 | void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs) | |
916 | { | |
917 | unsigned long tmo = msecs_to_jiffies(msecs); | |
918 | ||
919 | if (hctx->queue->nr_hw_queues == 1) | |
920 | kblockd_schedule_delayed_work(&hctx->delay_work, tmo); | |
921 | else { | |
922 | unsigned int cpu; | |
923 | ||
506e931f | 924 | cpu = blk_mq_hctx_next_cpu(hctx); |
70f4db63 CH |
925 | kblockd_schedule_delayed_work_on(cpu, &hctx->delay_work, tmo); |
926 | } | |
927 | } | |
928 | EXPORT_SYMBOL(blk_mq_delay_queue); | |
929 | ||
320ae51f | 930 | static void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, |
72a0a36e | 931 | struct request *rq, bool at_head) |
320ae51f JA |
932 | { |
933 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
934 | ||
01b983c9 JA |
935 | trace_block_rq_insert(hctx->queue, rq); |
936 | ||
72a0a36e CH |
937 | if (at_head) |
938 | list_add(&rq->queuelist, &ctx->rq_list); | |
939 | else | |
940 | list_add_tail(&rq->queuelist, &ctx->rq_list); | |
4bb659b1 | 941 | |
320ae51f | 942 | blk_mq_hctx_mark_pending(hctx, ctx); |
320ae51f JA |
943 | } |
944 | ||
eeabc850 CH |
945 | void blk_mq_insert_request(struct request *rq, bool at_head, bool run_queue, |
946 | bool async) | |
320ae51f | 947 | { |
eeabc850 | 948 | struct request_queue *q = rq->q; |
320ae51f | 949 | struct blk_mq_hw_ctx *hctx; |
eeabc850 CH |
950 | struct blk_mq_ctx *ctx = rq->mq_ctx, *current_ctx; |
951 | ||
952 | current_ctx = blk_mq_get_ctx(q); | |
953 | if (!cpu_online(ctx->cpu)) | |
954 | rq->mq_ctx = ctx = current_ctx; | |
320ae51f | 955 | |
320ae51f JA |
956 | hctx = q->mq_ops->map_queue(q, ctx->cpu); |
957 | ||
eeabc850 CH |
958 | if (rq->cmd_flags & (REQ_FLUSH | REQ_FUA) && |
959 | !(rq->cmd_flags & (REQ_FLUSH_SEQ))) { | |
320ae51f JA |
960 | blk_insert_flush(rq); |
961 | } else { | |
320ae51f | 962 | spin_lock(&ctx->lock); |
72a0a36e | 963 | __blk_mq_insert_request(hctx, rq, at_head); |
320ae51f | 964 | spin_unlock(&ctx->lock); |
320ae51f JA |
965 | } |
966 | ||
320ae51f JA |
967 | if (run_queue) |
968 | blk_mq_run_hw_queue(hctx, async); | |
e4043dcf JA |
969 | |
970 | blk_mq_put_ctx(current_ctx); | |
320ae51f JA |
971 | } |
972 | ||
973 | static void blk_mq_insert_requests(struct request_queue *q, | |
974 | struct blk_mq_ctx *ctx, | |
975 | struct list_head *list, | |
976 | int depth, | |
977 | bool from_schedule) | |
978 | ||
979 | { | |
980 | struct blk_mq_hw_ctx *hctx; | |
981 | struct blk_mq_ctx *current_ctx; | |
982 | ||
983 | trace_block_unplug(q, depth, !from_schedule); | |
984 | ||
985 | current_ctx = blk_mq_get_ctx(q); | |
986 | ||
987 | if (!cpu_online(ctx->cpu)) | |
988 | ctx = current_ctx; | |
989 | hctx = q->mq_ops->map_queue(q, ctx->cpu); | |
990 | ||
991 | /* | |
992 | * preemption doesn't flush plug list, so it's possible ctx->cpu is | |
993 | * offline now | |
994 | */ | |
995 | spin_lock(&ctx->lock); | |
996 | while (!list_empty(list)) { | |
997 | struct request *rq; | |
998 | ||
999 | rq = list_first_entry(list, struct request, queuelist); | |
1000 | list_del_init(&rq->queuelist); | |
1001 | rq->mq_ctx = ctx; | |
72a0a36e | 1002 | __blk_mq_insert_request(hctx, rq, false); |
320ae51f JA |
1003 | } |
1004 | spin_unlock(&ctx->lock); | |
1005 | ||
320ae51f | 1006 | blk_mq_run_hw_queue(hctx, from_schedule); |
e4043dcf | 1007 | blk_mq_put_ctx(current_ctx); |
320ae51f JA |
1008 | } |
1009 | ||
1010 | static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b) | |
1011 | { | |
1012 | struct request *rqa = container_of(a, struct request, queuelist); | |
1013 | struct request *rqb = container_of(b, struct request, queuelist); | |
1014 | ||
1015 | return !(rqa->mq_ctx < rqb->mq_ctx || | |
1016 | (rqa->mq_ctx == rqb->mq_ctx && | |
1017 | blk_rq_pos(rqa) < blk_rq_pos(rqb))); | |
1018 | } | |
1019 | ||
1020 | void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule) | |
1021 | { | |
1022 | struct blk_mq_ctx *this_ctx; | |
1023 | struct request_queue *this_q; | |
1024 | struct request *rq; | |
1025 | LIST_HEAD(list); | |
1026 | LIST_HEAD(ctx_list); | |
1027 | unsigned int depth; | |
1028 | ||
1029 | list_splice_init(&plug->mq_list, &list); | |
1030 | ||
1031 | list_sort(NULL, &list, plug_ctx_cmp); | |
1032 | ||
1033 | this_q = NULL; | |
1034 | this_ctx = NULL; | |
1035 | depth = 0; | |
1036 | ||
1037 | while (!list_empty(&list)) { | |
1038 | rq = list_entry_rq(list.next); | |
1039 | list_del_init(&rq->queuelist); | |
1040 | BUG_ON(!rq->q); | |
1041 | if (rq->mq_ctx != this_ctx) { | |
1042 | if (this_ctx) { | |
1043 | blk_mq_insert_requests(this_q, this_ctx, | |
1044 | &ctx_list, depth, | |
1045 | from_schedule); | |
1046 | } | |
1047 | ||
1048 | this_ctx = rq->mq_ctx; | |
1049 | this_q = rq->q; | |
1050 | depth = 0; | |
1051 | } | |
1052 | ||
1053 | depth++; | |
1054 | list_add_tail(&rq->queuelist, &ctx_list); | |
1055 | } | |
1056 | ||
1057 | /* | |
1058 | * If 'this_ctx' is set, we know we have entries to complete | |
1059 | * on 'ctx_list'. Do those. | |
1060 | */ | |
1061 | if (this_ctx) { | |
1062 | blk_mq_insert_requests(this_q, this_ctx, &ctx_list, depth, | |
1063 | from_schedule); | |
1064 | } | |
1065 | } | |
1066 | ||
1067 | static void blk_mq_bio_to_request(struct request *rq, struct bio *bio) | |
1068 | { | |
1069 | init_request_from_bio(rq, bio); | |
4b570521 | 1070 | |
3ee32372 | 1071 | if (blk_do_io_stat(rq)) |
4b570521 | 1072 | blk_account_io_start(rq, 1); |
320ae51f JA |
1073 | } |
1074 | ||
07068d5b JA |
1075 | static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx *hctx, |
1076 | struct blk_mq_ctx *ctx, | |
1077 | struct request *rq, struct bio *bio) | |
320ae51f | 1078 | { |
07068d5b | 1079 | struct request_queue *q = hctx->queue; |
320ae51f | 1080 | |
07068d5b JA |
1081 | if (!(hctx->flags & BLK_MQ_F_SHOULD_MERGE)) { |
1082 | blk_mq_bio_to_request(rq, bio); | |
1083 | spin_lock(&ctx->lock); | |
1084 | insert_rq: | |
1085 | __blk_mq_insert_request(hctx, rq, false); | |
1086 | spin_unlock(&ctx->lock); | |
1087 | return false; | |
1088 | } else { | |
1089 | spin_lock(&ctx->lock); | |
1090 | if (!blk_mq_attempt_merge(q, ctx, bio)) { | |
1091 | blk_mq_bio_to_request(rq, bio); | |
1092 | goto insert_rq; | |
1093 | } | |
320ae51f | 1094 | |
07068d5b JA |
1095 | spin_unlock(&ctx->lock); |
1096 | __blk_mq_free_request(hctx, ctx, rq); | |
1097 | return true; | |
14ec77f3 | 1098 | } |
07068d5b | 1099 | } |
14ec77f3 | 1100 | |
07068d5b JA |
1101 | struct blk_map_ctx { |
1102 | struct blk_mq_hw_ctx *hctx; | |
1103 | struct blk_mq_ctx *ctx; | |
1104 | }; | |
1105 | ||
1106 | static struct request *blk_mq_map_request(struct request_queue *q, | |
1107 | struct bio *bio, | |
1108 | struct blk_map_ctx *data) | |
1109 | { | |
1110 | struct blk_mq_hw_ctx *hctx; | |
1111 | struct blk_mq_ctx *ctx; | |
1112 | struct request *rq; | |
1113 | int rw = bio_data_dir(bio); | |
cb96a42c | 1114 | struct blk_mq_alloc_data alloc_data; |
320ae51f | 1115 | |
07068d5b | 1116 | if (unlikely(blk_mq_queue_enter(q))) { |
320ae51f | 1117 | bio_endio(bio, -EIO); |
07068d5b | 1118 | return NULL; |
320ae51f JA |
1119 | } |
1120 | ||
1121 | ctx = blk_mq_get_ctx(q); | |
1122 | hctx = q->mq_ops->map_queue(q, ctx->cpu); | |
1123 | ||
07068d5b | 1124 | if (rw_is_sync(bio->bi_rw)) |
27fbf4e8 | 1125 | rw |= REQ_SYNC; |
07068d5b | 1126 | |
320ae51f | 1127 | trace_block_getrq(q, bio, rw); |
cb96a42c ML |
1128 | blk_mq_set_alloc_data(&alloc_data, q, GFP_ATOMIC, false, ctx, |
1129 | hctx); | |
1130 | rq = __blk_mq_alloc_request(&alloc_data, rw); | |
5dee8577 | 1131 | if (unlikely(!rq)) { |
793597a6 | 1132 | __blk_mq_run_hw_queue(hctx); |
320ae51f JA |
1133 | blk_mq_put_ctx(ctx); |
1134 | trace_block_sleeprq(q, bio, rw); | |
793597a6 CH |
1135 | |
1136 | ctx = blk_mq_get_ctx(q); | |
320ae51f | 1137 | hctx = q->mq_ops->map_queue(q, ctx->cpu); |
cb96a42c ML |
1138 | blk_mq_set_alloc_data(&alloc_data, q, |
1139 | __GFP_WAIT|GFP_ATOMIC, false, ctx, hctx); | |
1140 | rq = __blk_mq_alloc_request(&alloc_data, rw); | |
1141 | ctx = alloc_data.ctx; | |
1142 | hctx = alloc_data.hctx; | |
320ae51f JA |
1143 | } |
1144 | ||
1145 | hctx->queued++; | |
07068d5b JA |
1146 | data->hctx = hctx; |
1147 | data->ctx = ctx; | |
1148 | return rq; | |
1149 | } | |
1150 | ||
1151 | /* | |
1152 | * Multiple hardware queue variant. This will not use per-process plugs, | |
1153 | * but will attempt to bypass the hctx queueing if we can go straight to | |
1154 | * hardware for SYNC IO. | |
1155 | */ | |
1156 | static void blk_mq_make_request(struct request_queue *q, struct bio *bio) | |
1157 | { | |
1158 | const int is_sync = rw_is_sync(bio->bi_rw); | |
1159 | const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA); | |
1160 | struct blk_map_ctx data; | |
1161 | struct request *rq; | |
1162 | ||
1163 | blk_queue_bounce(q, &bio); | |
1164 | ||
1165 | if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) { | |
1166 | bio_endio(bio, -EIO); | |
1167 | return; | |
1168 | } | |
1169 | ||
1170 | rq = blk_mq_map_request(q, bio, &data); | |
1171 | if (unlikely(!rq)) | |
1172 | return; | |
1173 | ||
1174 | if (unlikely(is_flush_fua)) { | |
1175 | blk_mq_bio_to_request(rq, bio); | |
1176 | blk_insert_flush(rq); | |
1177 | goto run_queue; | |
1178 | } | |
1179 | ||
1180 | if (is_sync) { | |
1181 | int ret; | |
1182 | ||
1183 | blk_mq_bio_to_request(rq, bio); | |
1184 | blk_mq_start_request(rq, true); | |
1185 | ||
1186 | /* | |
1187 | * For OK queue, we are done. For error, kill it. Any other | |
1188 | * error (busy), just add it to our list as we previously | |
1189 | * would have done | |
1190 | */ | |
1191 | ret = q->mq_ops->queue_rq(data.hctx, rq); | |
1192 | if (ret == BLK_MQ_RQ_QUEUE_OK) | |
1193 | goto done; | |
1194 | else { | |
1195 | __blk_mq_requeue_request(rq); | |
1196 | ||
1197 | if (ret == BLK_MQ_RQ_QUEUE_ERROR) { | |
1198 | rq->errors = -EIO; | |
1199 | blk_mq_end_io(rq, rq->errors); | |
1200 | goto done; | |
1201 | } | |
1202 | } | |
1203 | } | |
1204 | ||
1205 | if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) { | |
1206 | /* | |
1207 | * For a SYNC request, send it to the hardware immediately. For | |
1208 | * an ASYNC request, just ensure that we run it later on. The | |
1209 | * latter allows for merging opportunities and more efficient | |
1210 | * dispatching. | |
1211 | */ | |
1212 | run_queue: | |
1213 | blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua); | |
1214 | } | |
1215 | done: | |
1216 | blk_mq_put_ctx(data.ctx); | |
1217 | } | |
1218 | ||
1219 | /* | |
1220 | * Single hardware queue variant. This will attempt to use any per-process | |
1221 | * plug for merging and IO deferral. | |
1222 | */ | |
1223 | static void blk_sq_make_request(struct request_queue *q, struct bio *bio) | |
1224 | { | |
1225 | const int is_sync = rw_is_sync(bio->bi_rw); | |
1226 | const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA); | |
1227 | unsigned int use_plug, request_count = 0; | |
1228 | struct blk_map_ctx data; | |
1229 | struct request *rq; | |
1230 | ||
1231 | /* | |
1232 | * If we have multiple hardware queues, just go directly to | |
1233 | * one of those for sync IO. | |
1234 | */ | |
1235 | use_plug = !is_flush_fua && !is_sync; | |
1236 | ||
1237 | blk_queue_bounce(q, &bio); | |
1238 | ||
1239 | if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) { | |
1240 | bio_endio(bio, -EIO); | |
1241 | return; | |
1242 | } | |
1243 | ||
1244 | if (use_plug && !blk_queue_nomerges(q) && | |
1245 | blk_attempt_plug_merge(q, bio, &request_count)) | |
1246 | return; | |
1247 | ||
1248 | rq = blk_mq_map_request(q, bio, &data); | |
ff87bcec JA |
1249 | if (unlikely(!rq)) |
1250 | return; | |
320ae51f JA |
1251 | |
1252 | if (unlikely(is_flush_fua)) { | |
1253 | blk_mq_bio_to_request(rq, bio); | |
320ae51f JA |
1254 | blk_insert_flush(rq); |
1255 | goto run_queue; | |
1256 | } | |
1257 | ||
1258 | /* | |
1259 | * A task plug currently exists. Since this is completely lockless, | |
1260 | * utilize that to temporarily store requests until the task is | |
1261 | * either done or scheduled away. | |
1262 | */ | |
1263 | if (use_plug) { | |
1264 | struct blk_plug *plug = current->plug; | |
1265 | ||
1266 | if (plug) { | |
1267 | blk_mq_bio_to_request(rq, bio); | |
92f399c7 | 1268 | if (list_empty(&plug->mq_list)) |
320ae51f JA |
1269 | trace_block_plug(q); |
1270 | else if (request_count >= BLK_MAX_REQUEST_COUNT) { | |
1271 | blk_flush_plug_list(plug, false); | |
1272 | trace_block_plug(q); | |
1273 | } | |
1274 | list_add_tail(&rq->queuelist, &plug->mq_list); | |
07068d5b | 1275 | blk_mq_put_ctx(data.ctx); |
320ae51f JA |
1276 | return; |
1277 | } | |
1278 | } | |
1279 | ||
07068d5b JA |
1280 | if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) { |
1281 | /* | |
1282 | * For a SYNC request, send it to the hardware immediately. For | |
1283 | * an ASYNC request, just ensure that we run it later on. The | |
1284 | * latter allows for merging opportunities and more efficient | |
1285 | * dispatching. | |
1286 | */ | |
1287 | run_queue: | |
1288 | blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua); | |
320ae51f JA |
1289 | } |
1290 | ||
07068d5b | 1291 | blk_mq_put_ctx(data.ctx); |
320ae51f JA |
1292 | } |
1293 | ||
1294 | /* | |
1295 | * Default mapping to a software queue, since we use one per CPU. | |
1296 | */ | |
1297 | struct blk_mq_hw_ctx *blk_mq_map_queue(struct request_queue *q, const int cpu) | |
1298 | { | |
1299 | return q->queue_hw_ctx[q->mq_map[cpu]]; | |
1300 | } | |
1301 | EXPORT_SYMBOL(blk_mq_map_queue); | |
1302 | ||
24d2f903 CH |
1303 | static void blk_mq_free_rq_map(struct blk_mq_tag_set *set, |
1304 | struct blk_mq_tags *tags, unsigned int hctx_idx) | |
95363efd | 1305 | { |
e9b267d9 | 1306 | struct page *page; |
320ae51f | 1307 | |
24d2f903 | 1308 | if (tags->rqs && set->ops->exit_request) { |
e9b267d9 | 1309 | int i; |
320ae51f | 1310 | |
24d2f903 CH |
1311 | for (i = 0; i < tags->nr_tags; i++) { |
1312 | if (!tags->rqs[i]) | |
e9b267d9 | 1313 | continue; |
24d2f903 CH |
1314 | set->ops->exit_request(set->driver_data, tags->rqs[i], |
1315 | hctx_idx, i); | |
e9b267d9 | 1316 | } |
320ae51f | 1317 | } |
320ae51f | 1318 | |
24d2f903 CH |
1319 | while (!list_empty(&tags->page_list)) { |
1320 | page = list_first_entry(&tags->page_list, struct page, lru); | |
6753471c | 1321 | list_del_init(&page->lru); |
320ae51f JA |
1322 | __free_pages(page, page->private); |
1323 | } | |
1324 | ||
24d2f903 | 1325 | kfree(tags->rqs); |
320ae51f | 1326 | |
24d2f903 | 1327 | blk_mq_free_tags(tags); |
320ae51f JA |
1328 | } |
1329 | ||
1330 | static size_t order_to_size(unsigned int order) | |
1331 | { | |
4ca08500 | 1332 | return (size_t)PAGE_SIZE << order; |
320ae51f JA |
1333 | } |
1334 | ||
24d2f903 CH |
1335 | static struct blk_mq_tags *blk_mq_init_rq_map(struct blk_mq_tag_set *set, |
1336 | unsigned int hctx_idx) | |
320ae51f | 1337 | { |
24d2f903 | 1338 | struct blk_mq_tags *tags; |
320ae51f JA |
1339 | unsigned int i, j, entries_per_page, max_order = 4; |
1340 | size_t rq_size, left; | |
1341 | ||
24d2f903 CH |
1342 | tags = blk_mq_init_tags(set->queue_depth, set->reserved_tags, |
1343 | set->numa_node); | |
1344 | if (!tags) | |
1345 | return NULL; | |
320ae51f | 1346 | |
24d2f903 CH |
1347 | INIT_LIST_HEAD(&tags->page_list); |
1348 | ||
1349 | tags->rqs = kmalloc_node(set->queue_depth * sizeof(struct request *), | |
1350 | GFP_KERNEL, set->numa_node); | |
1351 | if (!tags->rqs) { | |
1352 | blk_mq_free_tags(tags); | |
1353 | return NULL; | |
1354 | } | |
320ae51f JA |
1355 | |
1356 | /* | |
1357 | * rq_size is the size of the request plus driver payload, rounded | |
1358 | * to the cacheline size | |
1359 | */ | |
24d2f903 | 1360 | rq_size = round_up(sizeof(struct request) + set->cmd_size, |
320ae51f | 1361 | cache_line_size()); |
24d2f903 | 1362 | left = rq_size * set->queue_depth; |
320ae51f | 1363 | |
24d2f903 | 1364 | for (i = 0; i < set->queue_depth; ) { |
320ae51f JA |
1365 | int this_order = max_order; |
1366 | struct page *page; | |
1367 | int to_do; | |
1368 | void *p; | |
1369 | ||
1370 | while (left < order_to_size(this_order - 1) && this_order) | |
1371 | this_order--; | |
1372 | ||
1373 | do { | |
24d2f903 CH |
1374 | page = alloc_pages_node(set->numa_node, GFP_KERNEL, |
1375 | this_order); | |
320ae51f JA |
1376 | if (page) |
1377 | break; | |
1378 | if (!this_order--) | |
1379 | break; | |
1380 | if (order_to_size(this_order) < rq_size) | |
1381 | break; | |
1382 | } while (1); | |
1383 | ||
1384 | if (!page) | |
24d2f903 | 1385 | goto fail; |
320ae51f JA |
1386 | |
1387 | page->private = this_order; | |
24d2f903 | 1388 | list_add_tail(&page->lru, &tags->page_list); |
320ae51f JA |
1389 | |
1390 | p = page_address(page); | |
1391 | entries_per_page = order_to_size(this_order) / rq_size; | |
24d2f903 | 1392 | to_do = min(entries_per_page, set->queue_depth - i); |
320ae51f JA |
1393 | left -= to_do * rq_size; |
1394 | for (j = 0; j < to_do; j++) { | |
24d2f903 CH |
1395 | tags->rqs[i] = p; |
1396 | if (set->ops->init_request) { | |
1397 | if (set->ops->init_request(set->driver_data, | |
1398 | tags->rqs[i], hctx_idx, i, | |
1399 | set->numa_node)) | |
1400 | goto fail; | |
e9b267d9 CH |
1401 | } |
1402 | ||
320ae51f JA |
1403 | p += rq_size; |
1404 | i++; | |
1405 | } | |
1406 | } | |
1407 | ||
24d2f903 | 1408 | return tags; |
320ae51f | 1409 | |
24d2f903 CH |
1410 | fail: |
1411 | pr_warn("%s: failed to allocate requests\n", __func__); | |
1412 | blk_mq_free_rq_map(set, tags, hctx_idx); | |
1413 | return NULL; | |
320ae51f JA |
1414 | } |
1415 | ||
1429d7c9 JA |
1416 | static void blk_mq_free_bitmap(struct blk_mq_ctxmap *bitmap) |
1417 | { | |
1418 | kfree(bitmap->map); | |
1419 | } | |
1420 | ||
1421 | static int blk_mq_alloc_bitmap(struct blk_mq_ctxmap *bitmap, int node) | |
1422 | { | |
1423 | unsigned int bpw = 8, total, num_maps, i; | |
1424 | ||
1425 | bitmap->bits_per_word = bpw; | |
1426 | ||
1427 | num_maps = ALIGN(nr_cpu_ids, bpw) / bpw; | |
1428 | bitmap->map = kzalloc_node(num_maps * sizeof(struct blk_align_bitmap), | |
1429 | GFP_KERNEL, node); | |
1430 | if (!bitmap->map) | |
1431 | return -ENOMEM; | |
1432 | ||
1433 | bitmap->map_size = num_maps; | |
1434 | ||
1435 | total = nr_cpu_ids; | |
1436 | for (i = 0; i < num_maps; i++) { | |
1437 | bitmap->map[i].depth = min(total, bitmap->bits_per_word); | |
1438 | total -= bitmap->map[i].depth; | |
1439 | } | |
1440 | ||
1441 | return 0; | |
1442 | } | |
1443 | ||
484b4061 JA |
1444 | static int blk_mq_hctx_cpu_offline(struct blk_mq_hw_ctx *hctx, int cpu) |
1445 | { | |
1446 | struct request_queue *q = hctx->queue; | |
1447 | struct blk_mq_ctx *ctx; | |
1448 | LIST_HEAD(tmp); | |
1449 | ||
1450 | /* | |
1451 | * Move ctx entries to new CPU, if this one is going away. | |
1452 | */ | |
1453 | ctx = __blk_mq_get_ctx(q, cpu); | |
1454 | ||
1455 | spin_lock(&ctx->lock); | |
1456 | if (!list_empty(&ctx->rq_list)) { | |
1457 | list_splice_init(&ctx->rq_list, &tmp); | |
1458 | blk_mq_hctx_clear_pending(hctx, ctx); | |
1459 | } | |
1460 | spin_unlock(&ctx->lock); | |
1461 | ||
1462 | if (list_empty(&tmp)) | |
1463 | return NOTIFY_OK; | |
1464 | ||
1465 | ctx = blk_mq_get_ctx(q); | |
1466 | spin_lock(&ctx->lock); | |
1467 | ||
1468 | while (!list_empty(&tmp)) { | |
1469 | struct request *rq; | |
1470 | ||
1471 | rq = list_first_entry(&tmp, struct request, queuelist); | |
1472 | rq->mq_ctx = ctx; | |
1473 | list_move_tail(&rq->queuelist, &ctx->rq_list); | |
1474 | } | |
1475 | ||
1476 | hctx = q->mq_ops->map_queue(q, ctx->cpu); | |
1477 | blk_mq_hctx_mark_pending(hctx, ctx); | |
1478 | ||
1479 | spin_unlock(&ctx->lock); | |
1480 | ||
1481 | blk_mq_run_hw_queue(hctx, true); | |
1482 | blk_mq_put_ctx(ctx); | |
1483 | return NOTIFY_OK; | |
1484 | } | |
1485 | ||
1486 | static int blk_mq_hctx_cpu_online(struct blk_mq_hw_ctx *hctx, int cpu) | |
1487 | { | |
1488 | struct request_queue *q = hctx->queue; | |
1489 | struct blk_mq_tag_set *set = q->tag_set; | |
1490 | ||
1491 | if (set->tags[hctx->queue_num]) | |
1492 | return NOTIFY_OK; | |
1493 | ||
1494 | set->tags[hctx->queue_num] = blk_mq_init_rq_map(set, hctx->queue_num); | |
1495 | if (!set->tags[hctx->queue_num]) | |
1496 | return NOTIFY_STOP; | |
1497 | ||
1498 | hctx->tags = set->tags[hctx->queue_num]; | |
1499 | return NOTIFY_OK; | |
1500 | } | |
1501 | ||
1502 | static int blk_mq_hctx_notify(void *data, unsigned long action, | |
1503 | unsigned int cpu) | |
1504 | { | |
1505 | struct blk_mq_hw_ctx *hctx = data; | |
1506 | ||
1507 | if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) | |
1508 | return blk_mq_hctx_cpu_offline(hctx, cpu); | |
1509 | else if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) | |
1510 | return blk_mq_hctx_cpu_online(hctx, cpu); | |
1511 | ||
1512 | return NOTIFY_OK; | |
1513 | } | |
1514 | ||
624dbe47 ML |
1515 | static void blk_mq_exit_hw_queues(struct request_queue *q, |
1516 | struct blk_mq_tag_set *set, int nr_queue) | |
1517 | { | |
1518 | struct blk_mq_hw_ctx *hctx; | |
1519 | unsigned int i; | |
1520 | ||
1521 | queue_for_each_hw_ctx(q, hctx, i) { | |
1522 | if (i == nr_queue) | |
1523 | break; | |
1524 | ||
f899fed4 JA |
1525 | blk_mq_tag_idle(hctx); |
1526 | ||
624dbe47 ML |
1527 | if (set->ops->exit_hctx) |
1528 | set->ops->exit_hctx(hctx, i); | |
1529 | ||
1530 | blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier); | |
1531 | kfree(hctx->ctxs); | |
1532 | blk_mq_free_bitmap(&hctx->ctx_map); | |
1533 | } | |
1534 | ||
1535 | } | |
1536 | ||
1537 | static void blk_mq_free_hw_queues(struct request_queue *q, | |
1538 | struct blk_mq_tag_set *set) | |
1539 | { | |
1540 | struct blk_mq_hw_ctx *hctx; | |
1541 | unsigned int i; | |
1542 | ||
1543 | queue_for_each_hw_ctx(q, hctx, i) { | |
1544 | free_cpumask_var(hctx->cpumask); | |
cdef54dd | 1545 | kfree(hctx); |
624dbe47 ML |
1546 | } |
1547 | } | |
1548 | ||
320ae51f | 1549 | static int blk_mq_init_hw_queues(struct request_queue *q, |
24d2f903 | 1550 | struct blk_mq_tag_set *set) |
320ae51f JA |
1551 | { |
1552 | struct blk_mq_hw_ctx *hctx; | |
624dbe47 | 1553 | unsigned int i; |
320ae51f JA |
1554 | |
1555 | /* | |
1556 | * Initialize hardware queues | |
1557 | */ | |
1558 | queue_for_each_hw_ctx(q, hctx, i) { | |
320ae51f JA |
1559 | int node; |
1560 | ||
1561 | node = hctx->numa_node; | |
1562 | if (node == NUMA_NO_NODE) | |
24d2f903 | 1563 | node = hctx->numa_node = set->numa_node; |
320ae51f | 1564 | |
70f4db63 CH |
1565 | INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn); |
1566 | INIT_DELAYED_WORK(&hctx->delay_work, blk_mq_delay_work_fn); | |
320ae51f JA |
1567 | spin_lock_init(&hctx->lock); |
1568 | INIT_LIST_HEAD(&hctx->dispatch); | |
1569 | hctx->queue = q; | |
1570 | hctx->queue_num = i; | |
24d2f903 CH |
1571 | hctx->flags = set->flags; |
1572 | hctx->cmd_size = set->cmd_size; | |
320ae51f JA |
1573 | |
1574 | blk_mq_init_cpu_notifier(&hctx->cpu_notifier, | |
1575 | blk_mq_hctx_notify, hctx); | |
1576 | blk_mq_register_cpu_notifier(&hctx->cpu_notifier); | |
1577 | ||
24d2f903 | 1578 | hctx->tags = set->tags[i]; |
320ae51f JA |
1579 | |
1580 | /* | |
1581 | * Allocate space for all possible cpus to avoid allocation in | |
1582 | * runtime | |
1583 | */ | |
1584 | hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *), | |
1585 | GFP_KERNEL, node); | |
1586 | if (!hctx->ctxs) | |
1587 | break; | |
1588 | ||
1429d7c9 | 1589 | if (blk_mq_alloc_bitmap(&hctx->ctx_map, node)) |
320ae51f JA |
1590 | break; |
1591 | ||
320ae51f JA |
1592 | hctx->nr_ctx = 0; |
1593 | ||
24d2f903 CH |
1594 | if (set->ops->init_hctx && |
1595 | set->ops->init_hctx(hctx, set->driver_data, i)) | |
320ae51f JA |
1596 | break; |
1597 | } | |
1598 | ||
1599 | if (i == q->nr_hw_queues) | |
1600 | return 0; | |
1601 | ||
1602 | /* | |
1603 | * Init failed | |
1604 | */ | |
624dbe47 | 1605 | blk_mq_exit_hw_queues(q, set, i); |
320ae51f JA |
1606 | |
1607 | return 1; | |
1608 | } | |
1609 | ||
1610 | static void blk_mq_init_cpu_queues(struct request_queue *q, | |
1611 | unsigned int nr_hw_queues) | |
1612 | { | |
1613 | unsigned int i; | |
1614 | ||
1615 | for_each_possible_cpu(i) { | |
1616 | struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i); | |
1617 | struct blk_mq_hw_ctx *hctx; | |
1618 | ||
1619 | memset(__ctx, 0, sizeof(*__ctx)); | |
1620 | __ctx->cpu = i; | |
1621 | spin_lock_init(&__ctx->lock); | |
1622 | INIT_LIST_HEAD(&__ctx->rq_list); | |
1623 | __ctx->queue = q; | |
1624 | ||
1625 | /* If the cpu isn't online, the cpu is mapped to first hctx */ | |
320ae51f JA |
1626 | if (!cpu_online(i)) |
1627 | continue; | |
1628 | ||
e4043dcf JA |
1629 | hctx = q->mq_ops->map_queue(q, i); |
1630 | cpumask_set_cpu(i, hctx->cpumask); | |
1631 | hctx->nr_ctx++; | |
1632 | ||
320ae51f JA |
1633 | /* |
1634 | * Set local node, IFF we have more than one hw queue. If | |
1635 | * not, we remain on the home node of the device | |
1636 | */ | |
1637 | if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE) | |
1638 | hctx->numa_node = cpu_to_node(i); | |
1639 | } | |
1640 | } | |
1641 | ||
1642 | static void blk_mq_map_swqueue(struct request_queue *q) | |
1643 | { | |
1644 | unsigned int i; | |
1645 | struct blk_mq_hw_ctx *hctx; | |
1646 | struct blk_mq_ctx *ctx; | |
1647 | ||
1648 | queue_for_each_hw_ctx(q, hctx, i) { | |
e4043dcf | 1649 | cpumask_clear(hctx->cpumask); |
320ae51f JA |
1650 | hctx->nr_ctx = 0; |
1651 | } | |
1652 | ||
1653 | /* | |
1654 | * Map software to hardware queues | |
1655 | */ | |
1656 | queue_for_each_ctx(q, ctx, i) { | |
1657 | /* If the cpu isn't online, the cpu is mapped to first hctx */ | |
e4043dcf JA |
1658 | if (!cpu_online(i)) |
1659 | continue; | |
1660 | ||
320ae51f | 1661 | hctx = q->mq_ops->map_queue(q, i); |
e4043dcf | 1662 | cpumask_set_cpu(i, hctx->cpumask); |
320ae51f JA |
1663 | ctx->index_hw = hctx->nr_ctx; |
1664 | hctx->ctxs[hctx->nr_ctx++] = ctx; | |
1665 | } | |
506e931f JA |
1666 | |
1667 | queue_for_each_hw_ctx(q, hctx, i) { | |
484b4061 JA |
1668 | /* |
1669 | * If not software queues are mapped to this hardware queue, | |
1670 | * disable it and free the request entries | |
1671 | */ | |
1672 | if (!hctx->nr_ctx) { | |
1673 | struct blk_mq_tag_set *set = q->tag_set; | |
1674 | ||
1675 | if (set->tags[i]) { | |
1676 | blk_mq_free_rq_map(set, set->tags[i], i); | |
1677 | set->tags[i] = NULL; | |
1678 | hctx->tags = NULL; | |
1679 | } | |
1680 | continue; | |
1681 | } | |
1682 | ||
1683 | /* | |
1684 | * Initialize batch roundrobin counts | |
1685 | */ | |
506e931f JA |
1686 | hctx->next_cpu = cpumask_first(hctx->cpumask); |
1687 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; | |
1688 | } | |
320ae51f JA |
1689 | } |
1690 | ||
0d2602ca JA |
1691 | static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set) |
1692 | { | |
1693 | struct blk_mq_hw_ctx *hctx; | |
1694 | struct request_queue *q; | |
1695 | bool shared; | |
1696 | int i; | |
1697 | ||
1698 | if (set->tag_list.next == set->tag_list.prev) | |
1699 | shared = false; | |
1700 | else | |
1701 | shared = true; | |
1702 | ||
1703 | list_for_each_entry(q, &set->tag_list, tag_set_list) { | |
1704 | blk_mq_freeze_queue(q); | |
1705 | ||
1706 | queue_for_each_hw_ctx(q, hctx, i) { | |
1707 | if (shared) | |
1708 | hctx->flags |= BLK_MQ_F_TAG_SHARED; | |
1709 | else | |
1710 | hctx->flags &= ~BLK_MQ_F_TAG_SHARED; | |
1711 | } | |
1712 | blk_mq_unfreeze_queue(q); | |
1713 | } | |
1714 | } | |
1715 | ||
1716 | static void blk_mq_del_queue_tag_set(struct request_queue *q) | |
1717 | { | |
1718 | struct blk_mq_tag_set *set = q->tag_set; | |
1719 | ||
1720 | blk_mq_freeze_queue(q); | |
1721 | ||
1722 | mutex_lock(&set->tag_list_lock); | |
1723 | list_del_init(&q->tag_set_list); | |
1724 | blk_mq_update_tag_set_depth(set); | |
1725 | mutex_unlock(&set->tag_list_lock); | |
1726 | ||
1727 | blk_mq_unfreeze_queue(q); | |
1728 | } | |
1729 | ||
1730 | static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set, | |
1731 | struct request_queue *q) | |
1732 | { | |
1733 | q->tag_set = set; | |
1734 | ||
1735 | mutex_lock(&set->tag_list_lock); | |
1736 | list_add_tail(&q->tag_set_list, &set->tag_list); | |
1737 | blk_mq_update_tag_set_depth(set); | |
1738 | mutex_unlock(&set->tag_list_lock); | |
1739 | } | |
1740 | ||
24d2f903 | 1741 | struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set) |
320ae51f JA |
1742 | { |
1743 | struct blk_mq_hw_ctx **hctxs; | |
e6cdb092 | 1744 | struct blk_mq_ctx __percpu *ctx; |
320ae51f | 1745 | struct request_queue *q; |
f14bbe77 | 1746 | unsigned int *map; |
320ae51f JA |
1747 | int i; |
1748 | ||
320ae51f JA |
1749 | ctx = alloc_percpu(struct blk_mq_ctx); |
1750 | if (!ctx) | |
1751 | return ERR_PTR(-ENOMEM); | |
1752 | ||
24d2f903 CH |
1753 | hctxs = kmalloc_node(set->nr_hw_queues * sizeof(*hctxs), GFP_KERNEL, |
1754 | set->numa_node); | |
320ae51f JA |
1755 | |
1756 | if (!hctxs) | |
1757 | goto err_percpu; | |
1758 | ||
f14bbe77 JA |
1759 | map = blk_mq_make_queue_map(set); |
1760 | if (!map) | |
1761 | goto err_map; | |
1762 | ||
24d2f903 | 1763 | for (i = 0; i < set->nr_hw_queues; i++) { |
f14bbe77 JA |
1764 | int node = blk_mq_hw_queue_to_node(map, i); |
1765 | ||
cdef54dd CH |
1766 | hctxs[i] = kzalloc_node(sizeof(struct blk_mq_hw_ctx), |
1767 | GFP_KERNEL, node); | |
320ae51f JA |
1768 | if (!hctxs[i]) |
1769 | goto err_hctxs; | |
1770 | ||
e4043dcf JA |
1771 | if (!zalloc_cpumask_var(&hctxs[i]->cpumask, GFP_KERNEL)) |
1772 | goto err_hctxs; | |
1773 | ||
0d2602ca | 1774 | atomic_set(&hctxs[i]->nr_active, 0); |
f14bbe77 | 1775 | hctxs[i]->numa_node = node; |
320ae51f JA |
1776 | hctxs[i]->queue_num = i; |
1777 | } | |
1778 | ||
24d2f903 | 1779 | q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node); |
320ae51f JA |
1780 | if (!q) |
1781 | goto err_hctxs; | |
1782 | ||
add703fd | 1783 | if (percpu_ref_init(&q->mq_usage_counter, blk_mq_usage_counter_release)) |
3d2936f4 ML |
1784 | goto err_map; |
1785 | ||
320ae51f JA |
1786 | setup_timer(&q->timeout, blk_mq_rq_timer, (unsigned long) q); |
1787 | blk_queue_rq_timeout(q, 30000); | |
1788 | ||
1789 | q->nr_queues = nr_cpu_ids; | |
24d2f903 | 1790 | q->nr_hw_queues = set->nr_hw_queues; |
f14bbe77 | 1791 | q->mq_map = map; |
320ae51f JA |
1792 | |
1793 | q->queue_ctx = ctx; | |
1794 | q->queue_hw_ctx = hctxs; | |
1795 | ||
24d2f903 | 1796 | q->mq_ops = set->ops; |
94eddfbe | 1797 | q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT; |
320ae51f | 1798 | |
05f1dd53 JA |
1799 | if (!(set->flags & BLK_MQ_F_SG_MERGE)) |
1800 | q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE; | |
1801 | ||
1be036e9 CH |
1802 | q->sg_reserved_size = INT_MAX; |
1803 | ||
6fca6a61 CH |
1804 | INIT_WORK(&q->requeue_work, blk_mq_requeue_work); |
1805 | INIT_LIST_HEAD(&q->requeue_list); | |
1806 | spin_lock_init(&q->requeue_lock); | |
1807 | ||
07068d5b JA |
1808 | if (q->nr_hw_queues > 1) |
1809 | blk_queue_make_request(q, blk_mq_make_request); | |
1810 | else | |
1811 | blk_queue_make_request(q, blk_sq_make_request); | |
1812 | ||
87ee7b11 | 1813 | blk_queue_rq_timed_out(q, blk_mq_rq_timed_out); |
24d2f903 CH |
1814 | if (set->timeout) |
1815 | blk_queue_rq_timeout(q, set->timeout); | |
320ae51f | 1816 | |
eba71768 JA |
1817 | /* |
1818 | * Do this after blk_queue_make_request() overrides it... | |
1819 | */ | |
1820 | q->nr_requests = set->queue_depth; | |
1821 | ||
24d2f903 CH |
1822 | if (set->ops->complete) |
1823 | blk_queue_softirq_done(q, set->ops->complete); | |
30a91cb4 | 1824 | |
320ae51f | 1825 | blk_mq_init_flush(q); |
24d2f903 | 1826 | blk_mq_init_cpu_queues(q, set->nr_hw_queues); |
320ae51f | 1827 | |
24d2f903 CH |
1828 | q->flush_rq = kzalloc(round_up(sizeof(struct request) + |
1829 | set->cmd_size, cache_line_size()), | |
1830 | GFP_KERNEL); | |
18741986 | 1831 | if (!q->flush_rq) |
320ae51f JA |
1832 | goto err_hw; |
1833 | ||
24d2f903 | 1834 | if (blk_mq_init_hw_queues(q, set)) |
18741986 CH |
1835 | goto err_flush_rq; |
1836 | ||
320ae51f JA |
1837 | mutex_lock(&all_q_mutex); |
1838 | list_add_tail(&q->all_q_node, &all_q_list); | |
1839 | mutex_unlock(&all_q_mutex); | |
1840 | ||
0d2602ca JA |
1841 | blk_mq_add_queue_tag_set(set, q); |
1842 | ||
484b4061 JA |
1843 | blk_mq_map_swqueue(q); |
1844 | ||
320ae51f | 1845 | return q; |
18741986 CH |
1846 | |
1847 | err_flush_rq: | |
1848 | kfree(q->flush_rq); | |
320ae51f | 1849 | err_hw: |
320ae51f JA |
1850 | blk_cleanup_queue(q); |
1851 | err_hctxs: | |
f14bbe77 | 1852 | kfree(map); |
24d2f903 | 1853 | for (i = 0; i < set->nr_hw_queues; i++) { |
320ae51f JA |
1854 | if (!hctxs[i]) |
1855 | break; | |
e4043dcf | 1856 | free_cpumask_var(hctxs[i]->cpumask); |
cdef54dd | 1857 | kfree(hctxs[i]); |
320ae51f | 1858 | } |
f14bbe77 | 1859 | err_map: |
320ae51f JA |
1860 | kfree(hctxs); |
1861 | err_percpu: | |
1862 | free_percpu(ctx); | |
1863 | return ERR_PTR(-ENOMEM); | |
1864 | } | |
1865 | EXPORT_SYMBOL(blk_mq_init_queue); | |
1866 | ||
1867 | void blk_mq_free_queue(struct request_queue *q) | |
1868 | { | |
624dbe47 | 1869 | struct blk_mq_tag_set *set = q->tag_set; |
320ae51f | 1870 | |
0d2602ca JA |
1871 | blk_mq_del_queue_tag_set(q); |
1872 | ||
624dbe47 ML |
1873 | blk_mq_exit_hw_queues(q, set, set->nr_hw_queues); |
1874 | blk_mq_free_hw_queues(q, set); | |
320ae51f | 1875 | |
add703fd | 1876 | percpu_ref_exit(&q->mq_usage_counter); |
3d2936f4 | 1877 | |
320ae51f JA |
1878 | free_percpu(q->queue_ctx); |
1879 | kfree(q->queue_hw_ctx); | |
1880 | kfree(q->mq_map); | |
1881 | ||
1882 | q->queue_ctx = NULL; | |
1883 | q->queue_hw_ctx = NULL; | |
1884 | q->mq_map = NULL; | |
1885 | ||
1886 | mutex_lock(&all_q_mutex); | |
1887 | list_del_init(&q->all_q_node); | |
1888 | mutex_unlock(&all_q_mutex); | |
1889 | } | |
320ae51f JA |
1890 | |
1891 | /* Basically redo blk_mq_init_queue with queue frozen */ | |
f618ef7c | 1892 | static void blk_mq_queue_reinit(struct request_queue *q) |
320ae51f JA |
1893 | { |
1894 | blk_mq_freeze_queue(q); | |
1895 | ||
67aec14c JA |
1896 | blk_mq_sysfs_unregister(q); |
1897 | ||
320ae51f JA |
1898 | blk_mq_update_queue_map(q->mq_map, q->nr_hw_queues); |
1899 | ||
1900 | /* | |
1901 | * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe | |
1902 | * we should change hctx numa_node according to new topology (this | |
1903 | * involves free and re-allocate memory, worthy doing?) | |
1904 | */ | |
1905 | ||
1906 | blk_mq_map_swqueue(q); | |
1907 | ||
67aec14c JA |
1908 | blk_mq_sysfs_register(q); |
1909 | ||
320ae51f JA |
1910 | blk_mq_unfreeze_queue(q); |
1911 | } | |
1912 | ||
f618ef7c PG |
1913 | static int blk_mq_queue_reinit_notify(struct notifier_block *nb, |
1914 | unsigned long action, void *hcpu) | |
320ae51f JA |
1915 | { |
1916 | struct request_queue *q; | |
1917 | ||
1918 | /* | |
9fccfed8 JA |
1919 | * Before new mappings are established, hotadded cpu might already |
1920 | * start handling requests. This doesn't break anything as we map | |
1921 | * offline CPUs to first hardware queue. We will re-init the queue | |
1922 | * below to get optimal settings. | |
320ae51f JA |
1923 | */ |
1924 | if (action != CPU_DEAD && action != CPU_DEAD_FROZEN && | |
1925 | action != CPU_ONLINE && action != CPU_ONLINE_FROZEN) | |
1926 | return NOTIFY_OK; | |
1927 | ||
1928 | mutex_lock(&all_q_mutex); | |
1929 | list_for_each_entry(q, &all_q_list, all_q_node) | |
1930 | blk_mq_queue_reinit(q); | |
1931 | mutex_unlock(&all_q_mutex); | |
1932 | return NOTIFY_OK; | |
1933 | } | |
1934 | ||
a4391c64 JA |
1935 | /* |
1936 | * Alloc a tag set to be associated with one or more request queues. | |
1937 | * May fail with EINVAL for various error conditions. May adjust the | |
1938 | * requested depth down, if if it too large. In that case, the set | |
1939 | * value will be stored in set->queue_depth. | |
1940 | */ | |
24d2f903 CH |
1941 | int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set) |
1942 | { | |
1943 | int i; | |
1944 | ||
1945 | if (!set->nr_hw_queues) | |
1946 | return -EINVAL; | |
a4391c64 | 1947 | if (!set->queue_depth) |
24d2f903 CH |
1948 | return -EINVAL; |
1949 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) | |
1950 | return -EINVAL; | |
1951 | ||
cdef54dd | 1952 | if (!set->nr_hw_queues || !set->ops->queue_rq || !set->ops->map_queue) |
24d2f903 CH |
1953 | return -EINVAL; |
1954 | ||
a4391c64 JA |
1955 | if (set->queue_depth > BLK_MQ_MAX_DEPTH) { |
1956 | pr_info("blk-mq: reduced tag depth to %u\n", | |
1957 | BLK_MQ_MAX_DEPTH); | |
1958 | set->queue_depth = BLK_MQ_MAX_DEPTH; | |
1959 | } | |
24d2f903 | 1960 | |
48479005 ML |
1961 | set->tags = kmalloc_node(set->nr_hw_queues * |
1962 | sizeof(struct blk_mq_tags *), | |
24d2f903 CH |
1963 | GFP_KERNEL, set->numa_node); |
1964 | if (!set->tags) | |
1965 | goto out; | |
1966 | ||
1967 | for (i = 0; i < set->nr_hw_queues; i++) { | |
1968 | set->tags[i] = blk_mq_init_rq_map(set, i); | |
1969 | if (!set->tags[i]) | |
1970 | goto out_unwind; | |
1971 | } | |
1972 | ||
0d2602ca JA |
1973 | mutex_init(&set->tag_list_lock); |
1974 | INIT_LIST_HEAD(&set->tag_list); | |
1975 | ||
24d2f903 CH |
1976 | return 0; |
1977 | ||
1978 | out_unwind: | |
1979 | while (--i >= 0) | |
1980 | blk_mq_free_rq_map(set, set->tags[i], i); | |
1981 | out: | |
1982 | return -ENOMEM; | |
1983 | } | |
1984 | EXPORT_SYMBOL(blk_mq_alloc_tag_set); | |
1985 | ||
1986 | void blk_mq_free_tag_set(struct blk_mq_tag_set *set) | |
1987 | { | |
1988 | int i; | |
1989 | ||
484b4061 JA |
1990 | for (i = 0; i < set->nr_hw_queues; i++) { |
1991 | if (set->tags[i]) | |
1992 | blk_mq_free_rq_map(set, set->tags[i], i); | |
1993 | } | |
1994 | ||
981bd189 | 1995 | kfree(set->tags); |
24d2f903 CH |
1996 | } |
1997 | EXPORT_SYMBOL(blk_mq_free_tag_set); | |
1998 | ||
e3a2b3f9 JA |
1999 | int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr) |
2000 | { | |
2001 | struct blk_mq_tag_set *set = q->tag_set; | |
2002 | struct blk_mq_hw_ctx *hctx; | |
2003 | int i, ret; | |
2004 | ||
2005 | if (!set || nr > set->queue_depth) | |
2006 | return -EINVAL; | |
2007 | ||
2008 | ret = 0; | |
2009 | queue_for_each_hw_ctx(q, hctx, i) { | |
2010 | ret = blk_mq_tag_update_depth(hctx->tags, nr); | |
2011 | if (ret) | |
2012 | break; | |
2013 | } | |
2014 | ||
2015 | if (!ret) | |
2016 | q->nr_requests = nr; | |
2017 | ||
2018 | return ret; | |
2019 | } | |
2020 | ||
676141e4 JA |
2021 | void blk_mq_disable_hotplug(void) |
2022 | { | |
2023 | mutex_lock(&all_q_mutex); | |
2024 | } | |
2025 | ||
2026 | void blk_mq_enable_hotplug(void) | |
2027 | { | |
2028 | mutex_unlock(&all_q_mutex); | |
2029 | } | |
2030 | ||
320ae51f JA |
2031 | static int __init blk_mq_init(void) |
2032 | { | |
320ae51f JA |
2033 | blk_mq_cpu_init(); |
2034 | ||
add703fd | 2035 | hotcpu_notifier(blk_mq_queue_reinit_notify, 0); |
320ae51f JA |
2036 | |
2037 | return 0; | |
2038 | } | |
2039 | subsys_initcall(blk_mq_init); |