]>
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 | { | |
81 | int ret; | |
82 | ||
83 | __percpu_counter_add(&q->mq_usage_counter, 1, 1000000); | |
84 | smp_wmb(); | |
85 | /* we have problems to freeze the queue if it's initializing */ | |
86 | if (!blk_queue_bypass(q) || !blk_queue_init_done(q)) | |
87 | return 0; | |
88 | ||
89 | __percpu_counter_add(&q->mq_usage_counter, -1, 1000000); | |
90 | ||
91 | spin_lock_irq(q->queue_lock); | |
92 | ret = wait_event_interruptible_lock_irq(q->mq_freeze_wq, | |
43a5e4e2 ML |
93 | !blk_queue_bypass(q) || blk_queue_dying(q), |
94 | *q->queue_lock); | |
320ae51f | 95 | /* inc usage with lock hold to avoid freeze_queue runs here */ |
43a5e4e2 | 96 | if (!ret && !blk_queue_dying(q)) |
320ae51f | 97 | __percpu_counter_add(&q->mq_usage_counter, 1, 1000000); |
43a5e4e2 ML |
98 | else if (blk_queue_dying(q)) |
99 | ret = -ENODEV; | |
320ae51f JA |
100 | spin_unlock_irq(q->queue_lock); |
101 | ||
102 | return ret; | |
103 | } | |
104 | ||
105 | static void blk_mq_queue_exit(struct request_queue *q) | |
106 | { | |
107 | __percpu_counter_add(&q->mq_usage_counter, -1, 1000000); | |
108 | } | |
109 | ||
43a5e4e2 ML |
110 | static void __blk_mq_drain_queue(struct request_queue *q) |
111 | { | |
112 | while (true) { | |
113 | s64 count; | |
114 | ||
115 | spin_lock_irq(q->queue_lock); | |
116 | count = percpu_counter_sum(&q->mq_usage_counter); | |
117 | spin_unlock_irq(q->queue_lock); | |
118 | ||
119 | if (count == 0) | |
120 | break; | |
121 | blk_mq_run_queues(q, false); | |
122 | msleep(10); | |
123 | } | |
124 | } | |
125 | ||
320ae51f JA |
126 | /* |
127 | * Guarantee no request is in use, so we can change any data structure of | |
128 | * the queue afterward. | |
129 | */ | |
130 | static void blk_mq_freeze_queue(struct request_queue *q) | |
131 | { | |
132 | bool drain; | |
133 | ||
134 | spin_lock_irq(q->queue_lock); | |
135 | drain = !q->bypass_depth++; | |
136 | queue_flag_set(QUEUE_FLAG_BYPASS, q); | |
137 | spin_unlock_irq(q->queue_lock); | |
138 | ||
43a5e4e2 ML |
139 | if (drain) |
140 | __blk_mq_drain_queue(q); | |
141 | } | |
320ae51f | 142 | |
43a5e4e2 ML |
143 | void blk_mq_drain_queue(struct request_queue *q) |
144 | { | |
145 | __blk_mq_drain_queue(q); | |
320ae51f JA |
146 | } |
147 | ||
148 | static void blk_mq_unfreeze_queue(struct request_queue *q) | |
149 | { | |
150 | bool wake = false; | |
151 | ||
152 | spin_lock_irq(q->queue_lock); | |
153 | if (!--q->bypass_depth) { | |
154 | queue_flag_clear(QUEUE_FLAG_BYPASS, q); | |
155 | wake = true; | |
156 | } | |
157 | WARN_ON_ONCE(q->bypass_depth < 0); | |
158 | spin_unlock_irq(q->queue_lock); | |
159 | if (wake) | |
160 | wake_up_all(&q->mq_freeze_wq); | |
161 | } | |
162 | ||
163 | bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx) | |
164 | { | |
165 | return blk_mq_has_free_tags(hctx->tags); | |
166 | } | |
167 | EXPORT_SYMBOL(blk_mq_can_queue); | |
168 | ||
94eddfbe JA |
169 | static void blk_mq_rq_ctx_init(struct request_queue *q, struct blk_mq_ctx *ctx, |
170 | struct request *rq, unsigned int rw_flags) | |
320ae51f | 171 | { |
94eddfbe JA |
172 | if (blk_queue_io_stat(q)) |
173 | rw_flags |= REQ_IO_STAT; | |
174 | ||
af76e555 CH |
175 | INIT_LIST_HEAD(&rq->queuelist); |
176 | /* csd/requeue_work/fifo_time is initialized before use */ | |
177 | rq->q = q; | |
320ae51f | 178 | rq->mq_ctx = ctx; |
0d2602ca | 179 | rq->cmd_flags |= rw_flags; |
af76e555 CH |
180 | /* do not touch atomic flags, it needs atomic ops against the timer */ |
181 | rq->cpu = -1; | |
af76e555 CH |
182 | INIT_HLIST_NODE(&rq->hash); |
183 | RB_CLEAR_NODE(&rq->rb_node); | |
af76e555 CH |
184 | rq->rq_disk = NULL; |
185 | rq->part = NULL; | |
af76e555 CH |
186 | #ifdef CONFIG_BLK_CGROUP |
187 | rq->rl = NULL; | |
0fec08b4 | 188 | set_start_time_ns(rq); |
af76e555 CH |
189 | rq->io_start_time_ns = 0; |
190 | #endif | |
191 | rq->nr_phys_segments = 0; | |
192 | #if defined(CONFIG_BLK_DEV_INTEGRITY) | |
193 | rq->nr_integrity_segments = 0; | |
194 | #endif | |
af76e555 CH |
195 | rq->special = NULL; |
196 | /* tag was already set */ | |
197 | rq->errors = 0; | |
af76e555 CH |
198 | |
199 | rq->extra_len = 0; | |
200 | rq->sense_len = 0; | |
201 | rq->resid_len = 0; | |
202 | rq->sense = NULL; | |
203 | ||
af76e555 | 204 | INIT_LIST_HEAD(&rq->timeout_list); |
af76e555 CH |
205 | rq->end_io = NULL; |
206 | rq->end_io_data = NULL; | |
207 | rq->next_rq = NULL; | |
208 | ||
320ae51f JA |
209 | ctx->rq_dispatched[rw_is_sync(rw_flags)]++; |
210 | } | |
211 | ||
5dee8577 | 212 | static struct request * |
cb96a42c | 213 | __blk_mq_alloc_request(struct blk_mq_alloc_data *data, int rw) |
5dee8577 CH |
214 | { |
215 | struct request *rq; | |
216 | unsigned int tag; | |
217 | ||
cb96a42c | 218 | tag = blk_mq_get_tag(data); |
5dee8577 | 219 | if (tag != BLK_MQ_TAG_FAIL) { |
cb96a42c | 220 | rq = data->hctx->tags->rqs[tag]; |
5dee8577 CH |
221 | |
222 | rq->cmd_flags = 0; | |
cb96a42c | 223 | if (blk_mq_tag_busy(data->hctx)) { |
5dee8577 | 224 | rq->cmd_flags = REQ_MQ_INFLIGHT; |
cb96a42c | 225 | atomic_inc(&data->hctx->nr_active); |
5dee8577 CH |
226 | } |
227 | ||
228 | rq->tag = tag; | |
cb96a42c | 229 | blk_mq_rq_ctx_init(data->q, data->ctx, rq, rw); |
5dee8577 CH |
230 | return rq; |
231 | } | |
232 | ||
233 | return NULL; | |
234 | } | |
235 | ||
4ce01dd1 CH |
236 | struct request *blk_mq_alloc_request(struct request_queue *q, int rw, gfp_t gfp, |
237 | bool reserved) | |
320ae51f | 238 | { |
d852564f CH |
239 | struct blk_mq_ctx *ctx; |
240 | struct blk_mq_hw_ctx *hctx; | |
320ae51f | 241 | struct request *rq; |
cb96a42c | 242 | struct blk_mq_alloc_data alloc_data; |
320ae51f JA |
243 | |
244 | if (blk_mq_queue_enter(q)) | |
245 | return NULL; | |
246 | ||
d852564f CH |
247 | ctx = blk_mq_get_ctx(q); |
248 | hctx = q->mq_ops->map_queue(q, ctx->cpu); | |
cb96a42c ML |
249 | blk_mq_set_alloc_data(&alloc_data, q, gfp & ~__GFP_WAIT, |
250 | reserved, ctx, hctx); | |
d852564f | 251 | |
cb96a42c | 252 | rq = __blk_mq_alloc_request(&alloc_data, rw); |
d852564f CH |
253 | if (!rq && (gfp & __GFP_WAIT)) { |
254 | __blk_mq_run_hw_queue(hctx); | |
255 | blk_mq_put_ctx(ctx); | |
256 | ||
257 | ctx = blk_mq_get_ctx(q); | |
258 | hctx = q->mq_ops->map_queue(q, ctx->cpu); | |
cb96a42c ML |
259 | blk_mq_set_alloc_data(&alloc_data, q, gfp, reserved, ctx, |
260 | hctx); | |
261 | rq = __blk_mq_alloc_request(&alloc_data, rw); | |
262 | ctx = alloc_data.ctx; | |
d852564f CH |
263 | } |
264 | blk_mq_put_ctx(ctx); | |
320ae51f JA |
265 | return rq; |
266 | } | |
4bb659b1 | 267 | EXPORT_SYMBOL(blk_mq_alloc_request); |
320ae51f | 268 | |
320ae51f JA |
269 | static void __blk_mq_free_request(struct blk_mq_hw_ctx *hctx, |
270 | struct blk_mq_ctx *ctx, struct request *rq) | |
271 | { | |
272 | const int tag = rq->tag; | |
273 | struct request_queue *q = rq->q; | |
274 | ||
0d2602ca JA |
275 | if (rq->cmd_flags & REQ_MQ_INFLIGHT) |
276 | atomic_dec(&hctx->nr_active); | |
277 | ||
af76e555 | 278 | clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags); |
0d2602ca | 279 | blk_mq_put_tag(hctx, tag, &ctx->last_tag); |
320ae51f JA |
280 | blk_mq_queue_exit(q); |
281 | } | |
282 | ||
283 | void blk_mq_free_request(struct request *rq) | |
284 | { | |
285 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
286 | struct blk_mq_hw_ctx *hctx; | |
287 | struct request_queue *q = rq->q; | |
288 | ||
289 | ctx->rq_completed[rq_is_sync(rq)]++; | |
290 | ||
291 | hctx = q->mq_ops->map_queue(q, ctx->cpu); | |
292 | __blk_mq_free_request(hctx, ctx, rq); | |
293 | } | |
294 | ||
8727af4b CH |
295 | /* |
296 | * Clone all relevant state from a request that has been put on hold in | |
297 | * the flush state machine into the preallocated flush request that hangs | |
298 | * off the request queue. | |
299 | * | |
300 | * For a driver the flush request should be invisible, that's why we are | |
301 | * impersonating the original request here. | |
302 | */ | |
303 | void blk_mq_clone_flush_request(struct request *flush_rq, | |
304 | struct request *orig_rq) | |
305 | { | |
306 | struct blk_mq_hw_ctx *hctx = | |
307 | orig_rq->q->mq_ops->map_queue(orig_rq->q, orig_rq->mq_ctx->cpu); | |
308 | ||
309 | flush_rq->mq_ctx = orig_rq->mq_ctx; | |
310 | flush_rq->tag = orig_rq->tag; | |
311 | memcpy(blk_mq_rq_to_pdu(flush_rq), blk_mq_rq_to_pdu(orig_rq), | |
312 | hctx->cmd_size); | |
313 | } | |
314 | ||
63151a44 | 315 | inline void __blk_mq_end_io(struct request *rq, int error) |
320ae51f | 316 | { |
0d11e6ac ML |
317 | blk_account_io_done(rq); |
318 | ||
91b63639 | 319 | if (rq->end_io) { |
320ae51f | 320 | rq->end_io(rq, error); |
91b63639 CH |
321 | } else { |
322 | if (unlikely(blk_bidi_rq(rq))) | |
323 | blk_mq_free_request(rq->next_rq); | |
320ae51f | 324 | blk_mq_free_request(rq); |
91b63639 | 325 | } |
320ae51f | 326 | } |
63151a44 CH |
327 | EXPORT_SYMBOL(__blk_mq_end_io); |
328 | ||
329 | void blk_mq_end_io(struct request *rq, int error) | |
330 | { | |
331 | if (blk_update_request(rq, error, blk_rq_bytes(rq))) | |
332 | BUG(); | |
333 | __blk_mq_end_io(rq, error); | |
334 | } | |
335 | EXPORT_SYMBOL(blk_mq_end_io); | |
320ae51f | 336 | |
30a91cb4 | 337 | static void __blk_mq_complete_request_remote(void *data) |
320ae51f | 338 | { |
3d6efbf6 | 339 | struct request *rq = data; |
320ae51f | 340 | |
30a91cb4 | 341 | rq->q->softirq_done_fn(rq); |
320ae51f | 342 | } |
320ae51f | 343 | |
ed851860 | 344 | static void blk_mq_ipi_complete_request(struct request *rq) |
320ae51f JA |
345 | { |
346 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
38535201 | 347 | bool shared = false; |
320ae51f JA |
348 | int cpu; |
349 | ||
38535201 | 350 | if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) { |
30a91cb4 CH |
351 | rq->q->softirq_done_fn(rq); |
352 | return; | |
353 | } | |
320ae51f JA |
354 | |
355 | cpu = get_cpu(); | |
38535201 CH |
356 | if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags)) |
357 | shared = cpus_share_cache(cpu, ctx->cpu); | |
358 | ||
359 | if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) { | |
30a91cb4 | 360 | rq->csd.func = __blk_mq_complete_request_remote; |
3d6efbf6 CH |
361 | rq->csd.info = rq; |
362 | rq->csd.flags = 0; | |
c46fff2a | 363 | smp_call_function_single_async(ctx->cpu, &rq->csd); |
3d6efbf6 | 364 | } else { |
30a91cb4 | 365 | rq->q->softirq_done_fn(rq); |
3d6efbf6 | 366 | } |
320ae51f JA |
367 | put_cpu(); |
368 | } | |
30a91cb4 | 369 | |
ed851860 JA |
370 | void __blk_mq_complete_request(struct request *rq) |
371 | { | |
372 | struct request_queue *q = rq->q; | |
373 | ||
374 | if (!q->softirq_done_fn) | |
375 | blk_mq_end_io(rq, rq->errors); | |
376 | else | |
377 | blk_mq_ipi_complete_request(rq); | |
378 | } | |
379 | ||
30a91cb4 CH |
380 | /** |
381 | * blk_mq_complete_request - end I/O on a request | |
382 | * @rq: the request being processed | |
383 | * | |
384 | * Description: | |
385 | * Ends all I/O on a request. It does not handle partial completions. | |
386 | * The actual completion happens out-of-order, through a IPI handler. | |
387 | **/ | |
388 | void blk_mq_complete_request(struct request *rq) | |
389 | { | |
95f09684 JA |
390 | struct request_queue *q = rq->q; |
391 | ||
392 | if (unlikely(blk_should_fake_timeout(q))) | |
30a91cb4 | 393 | return; |
ed851860 JA |
394 | if (!blk_mark_rq_complete(rq)) |
395 | __blk_mq_complete_request(rq); | |
30a91cb4 CH |
396 | } |
397 | EXPORT_SYMBOL(blk_mq_complete_request); | |
320ae51f | 398 | |
49f5baa5 | 399 | static void blk_mq_start_request(struct request *rq, bool last) |
320ae51f JA |
400 | { |
401 | struct request_queue *q = rq->q; | |
402 | ||
403 | trace_block_rq_issue(q, rq); | |
404 | ||
742ee69b | 405 | rq->resid_len = blk_rq_bytes(rq); |
91b63639 CH |
406 | if (unlikely(blk_bidi_rq(rq))) |
407 | rq->next_rq->resid_len = blk_rq_bytes(rq->next_rq); | |
742ee69b | 408 | |
320ae51f JA |
409 | /* |
410 | * Just mark start time and set the started bit. Due to memory | |
411 | * ordering, we know we'll see the correct deadline as long as | |
c22d9d8a JA |
412 | * REQ_ATOMIC_STARTED is seen. Use the default queue timeout, |
413 | * unless one has been set in the request. | |
320ae51f | 414 | */ |
c22d9d8a JA |
415 | if (!rq->timeout) |
416 | rq->deadline = jiffies + q->rq_timeout; | |
417 | else | |
418 | rq->deadline = jiffies + rq->timeout; | |
87ee7b11 JA |
419 | |
420 | /* | |
421 | * Mark us as started and clear complete. Complete might have been | |
422 | * set if requeue raced with timeout, which then marked it as | |
423 | * complete. So be sure to clear complete again when we start | |
424 | * the request, otherwise we'll ignore the completion event. | |
425 | */ | |
4b570521 JA |
426 | if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) |
427 | set_bit(REQ_ATOM_STARTED, &rq->atomic_flags); | |
428 | if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags)) | |
429 | clear_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags); | |
49f5baa5 CH |
430 | |
431 | if (q->dma_drain_size && blk_rq_bytes(rq)) { | |
432 | /* | |
433 | * Make sure space for the drain appears. We know we can do | |
434 | * this because max_hw_segments has been adjusted to be one | |
435 | * fewer than the device can handle. | |
436 | */ | |
437 | rq->nr_phys_segments++; | |
438 | } | |
439 | ||
440 | /* | |
441 | * Flag the last request in the series so that drivers know when IO | |
442 | * should be kicked off, if they don't do it on a per-request basis. | |
443 | * | |
444 | * Note: the flag isn't the only condition drivers should do kick off. | |
445 | * If drive is busy, the last request might not have the bit set. | |
446 | */ | |
447 | if (last) | |
448 | rq->cmd_flags |= REQ_END; | |
320ae51f JA |
449 | } |
450 | ||
ed0791b2 | 451 | static void __blk_mq_requeue_request(struct request *rq) |
320ae51f JA |
452 | { |
453 | struct request_queue *q = rq->q; | |
454 | ||
455 | trace_block_rq_requeue(q, rq); | |
456 | clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags); | |
49f5baa5 CH |
457 | |
458 | rq->cmd_flags &= ~REQ_END; | |
459 | ||
460 | if (q->dma_drain_size && blk_rq_bytes(rq)) | |
461 | rq->nr_phys_segments--; | |
320ae51f JA |
462 | } |
463 | ||
ed0791b2 CH |
464 | void blk_mq_requeue_request(struct request *rq) |
465 | { | |
ed0791b2 CH |
466 | __blk_mq_requeue_request(rq); |
467 | blk_clear_rq_complete(rq); | |
468 | ||
ed0791b2 | 469 | BUG_ON(blk_queued_rq(rq)); |
6fca6a61 | 470 | blk_mq_add_to_requeue_list(rq, true); |
ed0791b2 CH |
471 | } |
472 | EXPORT_SYMBOL(blk_mq_requeue_request); | |
473 | ||
6fca6a61 CH |
474 | static void blk_mq_requeue_work(struct work_struct *work) |
475 | { | |
476 | struct request_queue *q = | |
477 | container_of(work, struct request_queue, requeue_work); | |
478 | LIST_HEAD(rq_list); | |
479 | struct request *rq, *next; | |
480 | unsigned long flags; | |
481 | ||
482 | spin_lock_irqsave(&q->requeue_lock, flags); | |
483 | list_splice_init(&q->requeue_list, &rq_list); | |
484 | spin_unlock_irqrestore(&q->requeue_lock, flags); | |
485 | ||
486 | list_for_each_entry_safe(rq, next, &rq_list, queuelist) { | |
487 | if (!(rq->cmd_flags & REQ_SOFTBARRIER)) | |
488 | continue; | |
489 | ||
490 | rq->cmd_flags &= ~REQ_SOFTBARRIER; | |
491 | list_del_init(&rq->queuelist); | |
492 | blk_mq_insert_request(rq, true, false, false); | |
493 | } | |
494 | ||
495 | while (!list_empty(&rq_list)) { | |
496 | rq = list_entry(rq_list.next, struct request, queuelist); | |
497 | list_del_init(&rq->queuelist); | |
498 | blk_mq_insert_request(rq, false, false, false); | |
499 | } | |
500 | ||
501 | blk_mq_run_queues(q, false); | |
502 | } | |
503 | ||
504 | void blk_mq_add_to_requeue_list(struct request *rq, bool at_head) | |
505 | { | |
506 | struct request_queue *q = rq->q; | |
507 | unsigned long flags; | |
508 | ||
509 | /* | |
510 | * We abuse this flag that is otherwise used by the I/O scheduler to | |
511 | * request head insertation from the workqueue. | |
512 | */ | |
513 | BUG_ON(rq->cmd_flags & REQ_SOFTBARRIER); | |
514 | ||
515 | spin_lock_irqsave(&q->requeue_lock, flags); | |
516 | if (at_head) { | |
517 | rq->cmd_flags |= REQ_SOFTBARRIER; | |
518 | list_add(&rq->queuelist, &q->requeue_list); | |
519 | } else { | |
520 | list_add_tail(&rq->queuelist, &q->requeue_list); | |
521 | } | |
522 | spin_unlock_irqrestore(&q->requeue_lock, flags); | |
523 | } | |
524 | EXPORT_SYMBOL(blk_mq_add_to_requeue_list); | |
525 | ||
526 | void blk_mq_kick_requeue_list(struct request_queue *q) | |
527 | { | |
528 | kblockd_schedule_work(&q->requeue_work); | |
529 | } | |
530 | EXPORT_SYMBOL(blk_mq_kick_requeue_list); | |
531 | ||
22302375 | 532 | struct request *blk_mq_tag_to_rq(struct blk_mq_hw_ctx *hctx, unsigned int tag) |
24d2f903 | 533 | { |
22302375 SL |
534 | struct request_queue *q = hctx->queue; |
535 | ||
536 | if ((q->flush_rq->cmd_flags & REQ_FLUSH_SEQ) && | |
537 | q->flush_rq->tag == tag) | |
538 | return q->flush_rq; | |
539 | ||
540 | return hctx->tags->rqs[tag]; | |
24d2f903 CH |
541 | } |
542 | EXPORT_SYMBOL(blk_mq_tag_to_rq); | |
543 | ||
320ae51f JA |
544 | struct blk_mq_timeout_data { |
545 | struct blk_mq_hw_ctx *hctx; | |
546 | unsigned long *next; | |
547 | unsigned int *next_set; | |
548 | }; | |
549 | ||
550 | static void blk_mq_timeout_check(void *__data, unsigned long *free_tags) | |
551 | { | |
552 | struct blk_mq_timeout_data *data = __data; | |
553 | struct blk_mq_hw_ctx *hctx = data->hctx; | |
554 | unsigned int tag; | |
555 | ||
556 | /* It may not be in flight yet (this is where | |
557 | * the REQ_ATOMIC_STARTED flag comes in). The requests are | |
558 | * statically allocated, so we know it's always safe to access the | |
559 | * memory associated with a bit offset into ->rqs[]. | |
560 | */ | |
561 | tag = 0; | |
562 | do { | |
563 | struct request *rq; | |
564 | ||
24d2f903 CH |
565 | tag = find_next_zero_bit(free_tags, hctx->tags->nr_tags, tag); |
566 | if (tag >= hctx->tags->nr_tags) | |
320ae51f JA |
567 | break; |
568 | ||
22302375 | 569 | rq = blk_mq_tag_to_rq(hctx, tag++); |
24d2f903 CH |
570 | if (rq->q != hctx->queue) |
571 | continue; | |
320ae51f JA |
572 | if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) |
573 | continue; | |
574 | ||
575 | blk_rq_check_expired(rq, data->next, data->next_set); | |
576 | } while (1); | |
577 | } | |
578 | ||
579 | static void blk_mq_hw_ctx_check_timeout(struct blk_mq_hw_ctx *hctx, | |
580 | unsigned long *next, | |
581 | unsigned int *next_set) | |
582 | { | |
583 | struct blk_mq_timeout_data data = { | |
584 | .hctx = hctx, | |
585 | .next = next, | |
586 | .next_set = next_set, | |
587 | }; | |
588 | ||
589 | /* | |
590 | * Ask the tagging code to iterate busy requests, so we can | |
591 | * check them for timeout. | |
592 | */ | |
593 | blk_mq_tag_busy_iter(hctx->tags, blk_mq_timeout_check, &data); | |
594 | } | |
595 | ||
87ee7b11 JA |
596 | static enum blk_eh_timer_return blk_mq_rq_timed_out(struct request *rq) |
597 | { | |
598 | struct request_queue *q = rq->q; | |
599 | ||
600 | /* | |
601 | * We know that complete is set at this point. If STARTED isn't set | |
602 | * anymore, then the request isn't active and the "timeout" should | |
603 | * just be ignored. This can happen due to the bitflag ordering. | |
604 | * Timeout first checks if STARTED is set, and if it is, assumes | |
605 | * the request is active. But if we race with completion, then | |
606 | * we both flags will get cleared. So check here again, and ignore | |
607 | * a timeout event with a request that isn't active. | |
608 | */ | |
609 | if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) | |
610 | return BLK_EH_NOT_HANDLED; | |
611 | ||
612 | if (!q->mq_ops->timeout) | |
613 | return BLK_EH_RESET_TIMER; | |
614 | ||
615 | return q->mq_ops->timeout(rq); | |
616 | } | |
617 | ||
320ae51f JA |
618 | static void blk_mq_rq_timer(unsigned long data) |
619 | { | |
620 | struct request_queue *q = (struct request_queue *) data; | |
621 | struct blk_mq_hw_ctx *hctx; | |
622 | unsigned long next = 0; | |
623 | int i, next_set = 0; | |
624 | ||
484b4061 JA |
625 | queue_for_each_hw_ctx(q, hctx, i) { |
626 | /* | |
627 | * If not software queues are currently mapped to this | |
628 | * hardware queue, there's nothing to check | |
629 | */ | |
630 | if (!hctx->nr_ctx || !hctx->tags) | |
631 | continue; | |
632 | ||
320ae51f | 633 | blk_mq_hw_ctx_check_timeout(hctx, &next, &next_set); |
484b4061 | 634 | } |
320ae51f | 635 | |
0d2602ca JA |
636 | if (next_set) { |
637 | next = blk_rq_timeout(round_jiffies_up(next)); | |
638 | mod_timer(&q->timeout, next); | |
639 | } else { | |
640 | queue_for_each_hw_ctx(q, hctx, i) | |
641 | blk_mq_tag_idle(hctx); | |
642 | } | |
320ae51f JA |
643 | } |
644 | ||
645 | /* | |
646 | * Reverse check our software queue for entries that we could potentially | |
647 | * merge with. Currently includes a hand-wavy stop count of 8, to not spend | |
648 | * too much time checking for merges. | |
649 | */ | |
650 | static bool blk_mq_attempt_merge(struct request_queue *q, | |
651 | struct blk_mq_ctx *ctx, struct bio *bio) | |
652 | { | |
653 | struct request *rq; | |
654 | int checked = 8; | |
655 | ||
656 | list_for_each_entry_reverse(rq, &ctx->rq_list, queuelist) { | |
657 | int el_ret; | |
658 | ||
659 | if (!checked--) | |
660 | break; | |
661 | ||
662 | if (!blk_rq_merge_ok(rq, bio)) | |
663 | continue; | |
664 | ||
665 | el_ret = blk_try_merge(rq, bio); | |
666 | if (el_ret == ELEVATOR_BACK_MERGE) { | |
667 | if (bio_attempt_back_merge(q, rq, bio)) { | |
668 | ctx->rq_merged++; | |
669 | return true; | |
670 | } | |
671 | break; | |
672 | } else if (el_ret == ELEVATOR_FRONT_MERGE) { | |
673 | if (bio_attempt_front_merge(q, rq, bio)) { | |
674 | ctx->rq_merged++; | |
675 | return true; | |
676 | } | |
677 | break; | |
678 | } | |
679 | } | |
680 | ||
681 | return false; | |
682 | } | |
683 | ||
1429d7c9 JA |
684 | /* |
685 | * Process software queues that have been marked busy, splicing them | |
686 | * to the for-dispatch | |
687 | */ | |
688 | static void flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list) | |
689 | { | |
690 | struct blk_mq_ctx *ctx; | |
691 | int i; | |
692 | ||
693 | for (i = 0; i < hctx->ctx_map.map_size; i++) { | |
694 | struct blk_align_bitmap *bm = &hctx->ctx_map.map[i]; | |
695 | unsigned int off, bit; | |
696 | ||
697 | if (!bm->word) | |
698 | continue; | |
699 | ||
700 | bit = 0; | |
701 | off = i * hctx->ctx_map.bits_per_word; | |
702 | do { | |
703 | bit = find_next_bit(&bm->word, bm->depth, bit); | |
704 | if (bit >= bm->depth) | |
705 | break; | |
706 | ||
707 | ctx = hctx->ctxs[bit + off]; | |
708 | clear_bit(bit, &bm->word); | |
709 | spin_lock(&ctx->lock); | |
710 | list_splice_tail_init(&ctx->rq_list, list); | |
711 | spin_unlock(&ctx->lock); | |
712 | ||
713 | bit++; | |
714 | } while (1); | |
715 | } | |
716 | } | |
717 | ||
320ae51f JA |
718 | /* |
719 | * Run this hardware queue, pulling any software queues mapped to it in. | |
720 | * Note that this function currently has various problems around ordering | |
721 | * of IO. In particular, we'd like FIFO behaviour on handling existing | |
722 | * items on the hctx->dispatch list. Ignore that for now. | |
723 | */ | |
724 | static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx) | |
725 | { | |
726 | struct request_queue *q = hctx->queue; | |
320ae51f JA |
727 | struct request *rq; |
728 | LIST_HEAD(rq_list); | |
1429d7c9 | 729 | int queued; |
320ae51f | 730 | |
fd1270d5 | 731 | WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask)); |
e4043dcf | 732 | |
5d12f905 | 733 | if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state))) |
320ae51f JA |
734 | return; |
735 | ||
736 | hctx->run++; | |
737 | ||
738 | /* | |
739 | * Touch any software queue that has pending entries. | |
740 | */ | |
1429d7c9 | 741 | flush_busy_ctxs(hctx, &rq_list); |
320ae51f JA |
742 | |
743 | /* | |
744 | * If we have previous entries on our dispatch list, grab them | |
745 | * and stuff them at the front for more fair dispatch. | |
746 | */ | |
747 | if (!list_empty_careful(&hctx->dispatch)) { | |
748 | spin_lock(&hctx->lock); | |
749 | if (!list_empty(&hctx->dispatch)) | |
750 | list_splice_init(&hctx->dispatch, &rq_list); | |
751 | spin_unlock(&hctx->lock); | |
752 | } | |
753 | ||
320ae51f JA |
754 | /* |
755 | * Now process all the entries, sending them to the driver. | |
756 | */ | |
1429d7c9 | 757 | queued = 0; |
320ae51f JA |
758 | while (!list_empty(&rq_list)) { |
759 | int ret; | |
760 | ||
761 | rq = list_first_entry(&rq_list, struct request, queuelist); | |
762 | list_del_init(&rq->queuelist); | |
320ae51f | 763 | |
49f5baa5 | 764 | blk_mq_start_request(rq, list_empty(&rq_list)); |
320ae51f JA |
765 | |
766 | ret = q->mq_ops->queue_rq(hctx, rq); | |
767 | switch (ret) { | |
768 | case BLK_MQ_RQ_QUEUE_OK: | |
769 | queued++; | |
770 | continue; | |
771 | case BLK_MQ_RQ_QUEUE_BUSY: | |
320ae51f | 772 | list_add(&rq->queuelist, &rq_list); |
ed0791b2 | 773 | __blk_mq_requeue_request(rq); |
320ae51f JA |
774 | break; |
775 | default: | |
776 | pr_err("blk-mq: bad return on queue: %d\n", ret); | |
320ae51f | 777 | case BLK_MQ_RQ_QUEUE_ERROR: |
1e93b8c2 | 778 | rq->errors = -EIO; |
320ae51f JA |
779 | blk_mq_end_io(rq, rq->errors); |
780 | break; | |
781 | } | |
782 | ||
783 | if (ret == BLK_MQ_RQ_QUEUE_BUSY) | |
784 | break; | |
785 | } | |
786 | ||
787 | if (!queued) | |
788 | hctx->dispatched[0]++; | |
789 | else if (queued < (1 << (BLK_MQ_MAX_DISPATCH_ORDER - 1))) | |
790 | hctx->dispatched[ilog2(queued) + 1]++; | |
791 | ||
792 | /* | |
793 | * Any items that need requeuing? Stuff them into hctx->dispatch, | |
794 | * that is where we will continue on next queue run. | |
795 | */ | |
796 | if (!list_empty(&rq_list)) { | |
797 | spin_lock(&hctx->lock); | |
798 | list_splice(&rq_list, &hctx->dispatch); | |
799 | spin_unlock(&hctx->lock); | |
800 | } | |
801 | } | |
802 | ||
506e931f JA |
803 | /* |
804 | * It'd be great if the workqueue API had a way to pass | |
805 | * in a mask and had some smarts for more clever placement. | |
806 | * For now we just round-robin here, switching for every | |
807 | * BLK_MQ_CPU_WORK_BATCH queued items. | |
808 | */ | |
809 | static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx) | |
810 | { | |
811 | int cpu = hctx->next_cpu; | |
812 | ||
813 | if (--hctx->next_cpu_batch <= 0) { | |
814 | int next_cpu; | |
815 | ||
816 | next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask); | |
817 | if (next_cpu >= nr_cpu_ids) | |
818 | next_cpu = cpumask_first(hctx->cpumask); | |
819 | ||
820 | hctx->next_cpu = next_cpu; | |
821 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; | |
822 | } | |
823 | ||
824 | return cpu; | |
825 | } | |
826 | ||
320ae51f JA |
827 | void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) |
828 | { | |
5d12f905 | 829 | if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state))) |
320ae51f JA |
830 | return; |
831 | ||
e4043dcf | 832 | if (!async && cpumask_test_cpu(smp_processor_id(), hctx->cpumask)) |
320ae51f | 833 | __blk_mq_run_hw_queue(hctx); |
e4043dcf | 834 | else if (hctx->queue->nr_hw_queues == 1) |
70f4db63 | 835 | kblockd_schedule_delayed_work(&hctx->run_work, 0); |
e4043dcf JA |
836 | else { |
837 | unsigned int cpu; | |
838 | ||
506e931f | 839 | cpu = blk_mq_hctx_next_cpu(hctx); |
70f4db63 | 840 | kblockd_schedule_delayed_work_on(cpu, &hctx->run_work, 0); |
e4043dcf | 841 | } |
320ae51f JA |
842 | } |
843 | ||
844 | void blk_mq_run_queues(struct request_queue *q, bool async) | |
845 | { | |
846 | struct blk_mq_hw_ctx *hctx; | |
847 | int i; | |
848 | ||
849 | queue_for_each_hw_ctx(q, hctx, i) { | |
850 | if ((!blk_mq_hctx_has_pending(hctx) && | |
851 | list_empty_careful(&hctx->dispatch)) || | |
5d12f905 | 852 | test_bit(BLK_MQ_S_STOPPED, &hctx->state)) |
320ae51f JA |
853 | continue; |
854 | ||
e4043dcf | 855 | preempt_disable(); |
320ae51f | 856 | blk_mq_run_hw_queue(hctx, async); |
e4043dcf | 857 | preempt_enable(); |
320ae51f JA |
858 | } |
859 | } | |
860 | EXPORT_SYMBOL(blk_mq_run_queues); | |
861 | ||
862 | void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx) | |
863 | { | |
70f4db63 CH |
864 | cancel_delayed_work(&hctx->run_work); |
865 | cancel_delayed_work(&hctx->delay_work); | |
320ae51f JA |
866 | set_bit(BLK_MQ_S_STOPPED, &hctx->state); |
867 | } | |
868 | EXPORT_SYMBOL(blk_mq_stop_hw_queue); | |
869 | ||
280d45f6 CH |
870 | void blk_mq_stop_hw_queues(struct request_queue *q) |
871 | { | |
872 | struct blk_mq_hw_ctx *hctx; | |
873 | int i; | |
874 | ||
875 | queue_for_each_hw_ctx(q, hctx, i) | |
876 | blk_mq_stop_hw_queue(hctx); | |
877 | } | |
878 | EXPORT_SYMBOL(blk_mq_stop_hw_queues); | |
879 | ||
320ae51f JA |
880 | void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx) |
881 | { | |
882 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
e4043dcf JA |
883 | |
884 | preempt_disable(); | |
320ae51f | 885 | __blk_mq_run_hw_queue(hctx); |
e4043dcf | 886 | preempt_enable(); |
320ae51f JA |
887 | } |
888 | EXPORT_SYMBOL(blk_mq_start_hw_queue); | |
889 | ||
2f268556 CH |
890 | void blk_mq_start_hw_queues(struct request_queue *q) |
891 | { | |
892 | struct blk_mq_hw_ctx *hctx; | |
893 | int i; | |
894 | ||
895 | queue_for_each_hw_ctx(q, hctx, i) | |
896 | blk_mq_start_hw_queue(hctx); | |
897 | } | |
898 | EXPORT_SYMBOL(blk_mq_start_hw_queues); | |
899 | ||
900 | ||
1b4a3258 | 901 | void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async) |
320ae51f JA |
902 | { |
903 | struct blk_mq_hw_ctx *hctx; | |
904 | int i; | |
905 | ||
906 | queue_for_each_hw_ctx(q, hctx, i) { | |
907 | if (!test_bit(BLK_MQ_S_STOPPED, &hctx->state)) | |
908 | continue; | |
909 | ||
910 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
e4043dcf | 911 | preempt_disable(); |
1b4a3258 | 912 | blk_mq_run_hw_queue(hctx, async); |
e4043dcf | 913 | preempt_enable(); |
320ae51f JA |
914 | } |
915 | } | |
916 | EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues); | |
917 | ||
70f4db63 | 918 | static void blk_mq_run_work_fn(struct work_struct *work) |
320ae51f JA |
919 | { |
920 | struct blk_mq_hw_ctx *hctx; | |
921 | ||
70f4db63 | 922 | hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work); |
e4043dcf | 923 | |
320ae51f JA |
924 | __blk_mq_run_hw_queue(hctx); |
925 | } | |
926 | ||
70f4db63 CH |
927 | static void blk_mq_delay_work_fn(struct work_struct *work) |
928 | { | |
929 | struct blk_mq_hw_ctx *hctx; | |
930 | ||
931 | hctx = container_of(work, struct blk_mq_hw_ctx, delay_work.work); | |
932 | ||
933 | if (test_and_clear_bit(BLK_MQ_S_STOPPED, &hctx->state)) | |
934 | __blk_mq_run_hw_queue(hctx); | |
935 | } | |
936 | ||
937 | void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs) | |
938 | { | |
939 | unsigned long tmo = msecs_to_jiffies(msecs); | |
940 | ||
941 | if (hctx->queue->nr_hw_queues == 1) | |
942 | kblockd_schedule_delayed_work(&hctx->delay_work, tmo); | |
943 | else { | |
944 | unsigned int cpu; | |
945 | ||
506e931f | 946 | cpu = blk_mq_hctx_next_cpu(hctx); |
70f4db63 CH |
947 | kblockd_schedule_delayed_work_on(cpu, &hctx->delay_work, tmo); |
948 | } | |
949 | } | |
950 | EXPORT_SYMBOL(blk_mq_delay_queue); | |
951 | ||
320ae51f | 952 | static void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, |
72a0a36e | 953 | struct request *rq, bool at_head) |
320ae51f JA |
954 | { |
955 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
956 | ||
01b983c9 JA |
957 | trace_block_rq_insert(hctx->queue, rq); |
958 | ||
72a0a36e CH |
959 | if (at_head) |
960 | list_add(&rq->queuelist, &ctx->rq_list); | |
961 | else | |
962 | list_add_tail(&rq->queuelist, &ctx->rq_list); | |
4bb659b1 | 963 | |
320ae51f JA |
964 | blk_mq_hctx_mark_pending(hctx, ctx); |
965 | ||
966 | /* | |
967 | * We do this early, to ensure we are on the right CPU. | |
968 | */ | |
87ee7b11 | 969 | blk_add_timer(rq); |
320ae51f JA |
970 | } |
971 | ||
eeabc850 CH |
972 | void blk_mq_insert_request(struct request *rq, bool at_head, bool run_queue, |
973 | bool async) | |
320ae51f | 974 | { |
eeabc850 | 975 | struct request_queue *q = rq->q; |
320ae51f | 976 | struct blk_mq_hw_ctx *hctx; |
eeabc850 CH |
977 | struct blk_mq_ctx *ctx = rq->mq_ctx, *current_ctx; |
978 | ||
979 | current_ctx = blk_mq_get_ctx(q); | |
980 | if (!cpu_online(ctx->cpu)) | |
981 | rq->mq_ctx = ctx = current_ctx; | |
320ae51f | 982 | |
320ae51f JA |
983 | hctx = q->mq_ops->map_queue(q, ctx->cpu); |
984 | ||
eeabc850 CH |
985 | if (rq->cmd_flags & (REQ_FLUSH | REQ_FUA) && |
986 | !(rq->cmd_flags & (REQ_FLUSH_SEQ))) { | |
320ae51f JA |
987 | blk_insert_flush(rq); |
988 | } else { | |
320ae51f | 989 | spin_lock(&ctx->lock); |
72a0a36e | 990 | __blk_mq_insert_request(hctx, rq, at_head); |
320ae51f | 991 | spin_unlock(&ctx->lock); |
320ae51f JA |
992 | } |
993 | ||
320ae51f JA |
994 | if (run_queue) |
995 | blk_mq_run_hw_queue(hctx, async); | |
e4043dcf JA |
996 | |
997 | blk_mq_put_ctx(current_ctx); | |
320ae51f JA |
998 | } |
999 | ||
1000 | static void blk_mq_insert_requests(struct request_queue *q, | |
1001 | struct blk_mq_ctx *ctx, | |
1002 | struct list_head *list, | |
1003 | int depth, | |
1004 | bool from_schedule) | |
1005 | ||
1006 | { | |
1007 | struct blk_mq_hw_ctx *hctx; | |
1008 | struct blk_mq_ctx *current_ctx; | |
1009 | ||
1010 | trace_block_unplug(q, depth, !from_schedule); | |
1011 | ||
1012 | current_ctx = blk_mq_get_ctx(q); | |
1013 | ||
1014 | if (!cpu_online(ctx->cpu)) | |
1015 | ctx = current_ctx; | |
1016 | hctx = q->mq_ops->map_queue(q, ctx->cpu); | |
1017 | ||
1018 | /* | |
1019 | * preemption doesn't flush plug list, so it's possible ctx->cpu is | |
1020 | * offline now | |
1021 | */ | |
1022 | spin_lock(&ctx->lock); | |
1023 | while (!list_empty(list)) { | |
1024 | struct request *rq; | |
1025 | ||
1026 | rq = list_first_entry(list, struct request, queuelist); | |
1027 | list_del_init(&rq->queuelist); | |
1028 | rq->mq_ctx = ctx; | |
72a0a36e | 1029 | __blk_mq_insert_request(hctx, rq, false); |
320ae51f JA |
1030 | } |
1031 | spin_unlock(&ctx->lock); | |
1032 | ||
320ae51f | 1033 | blk_mq_run_hw_queue(hctx, from_schedule); |
e4043dcf | 1034 | blk_mq_put_ctx(current_ctx); |
320ae51f JA |
1035 | } |
1036 | ||
1037 | static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b) | |
1038 | { | |
1039 | struct request *rqa = container_of(a, struct request, queuelist); | |
1040 | struct request *rqb = container_of(b, struct request, queuelist); | |
1041 | ||
1042 | return !(rqa->mq_ctx < rqb->mq_ctx || | |
1043 | (rqa->mq_ctx == rqb->mq_ctx && | |
1044 | blk_rq_pos(rqa) < blk_rq_pos(rqb))); | |
1045 | } | |
1046 | ||
1047 | void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule) | |
1048 | { | |
1049 | struct blk_mq_ctx *this_ctx; | |
1050 | struct request_queue *this_q; | |
1051 | struct request *rq; | |
1052 | LIST_HEAD(list); | |
1053 | LIST_HEAD(ctx_list); | |
1054 | unsigned int depth; | |
1055 | ||
1056 | list_splice_init(&plug->mq_list, &list); | |
1057 | ||
1058 | list_sort(NULL, &list, plug_ctx_cmp); | |
1059 | ||
1060 | this_q = NULL; | |
1061 | this_ctx = NULL; | |
1062 | depth = 0; | |
1063 | ||
1064 | while (!list_empty(&list)) { | |
1065 | rq = list_entry_rq(list.next); | |
1066 | list_del_init(&rq->queuelist); | |
1067 | BUG_ON(!rq->q); | |
1068 | if (rq->mq_ctx != this_ctx) { | |
1069 | if (this_ctx) { | |
1070 | blk_mq_insert_requests(this_q, this_ctx, | |
1071 | &ctx_list, depth, | |
1072 | from_schedule); | |
1073 | } | |
1074 | ||
1075 | this_ctx = rq->mq_ctx; | |
1076 | this_q = rq->q; | |
1077 | depth = 0; | |
1078 | } | |
1079 | ||
1080 | depth++; | |
1081 | list_add_tail(&rq->queuelist, &ctx_list); | |
1082 | } | |
1083 | ||
1084 | /* | |
1085 | * If 'this_ctx' is set, we know we have entries to complete | |
1086 | * on 'ctx_list'. Do those. | |
1087 | */ | |
1088 | if (this_ctx) { | |
1089 | blk_mq_insert_requests(this_q, this_ctx, &ctx_list, depth, | |
1090 | from_schedule); | |
1091 | } | |
1092 | } | |
1093 | ||
1094 | static void blk_mq_bio_to_request(struct request *rq, struct bio *bio) | |
1095 | { | |
1096 | init_request_from_bio(rq, bio); | |
4b570521 JA |
1097 | |
1098 | if (blk_do_io_stat(rq)) { | |
1099 | rq->start_time = jiffies; | |
1100 | blk_account_io_start(rq, 1); | |
1101 | } | |
320ae51f JA |
1102 | } |
1103 | ||
07068d5b JA |
1104 | static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx *hctx, |
1105 | struct blk_mq_ctx *ctx, | |
1106 | struct request *rq, struct bio *bio) | |
320ae51f | 1107 | { |
07068d5b | 1108 | struct request_queue *q = hctx->queue; |
320ae51f | 1109 | |
07068d5b JA |
1110 | if (!(hctx->flags & BLK_MQ_F_SHOULD_MERGE)) { |
1111 | blk_mq_bio_to_request(rq, bio); | |
1112 | spin_lock(&ctx->lock); | |
1113 | insert_rq: | |
1114 | __blk_mq_insert_request(hctx, rq, false); | |
1115 | spin_unlock(&ctx->lock); | |
1116 | return false; | |
1117 | } else { | |
1118 | spin_lock(&ctx->lock); | |
1119 | if (!blk_mq_attempt_merge(q, ctx, bio)) { | |
1120 | blk_mq_bio_to_request(rq, bio); | |
1121 | goto insert_rq; | |
1122 | } | |
320ae51f | 1123 | |
07068d5b JA |
1124 | spin_unlock(&ctx->lock); |
1125 | __blk_mq_free_request(hctx, ctx, rq); | |
1126 | return true; | |
14ec77f3 | 1127 | } |
07068d5b | 1128 | } |
14ec77f3 | 1129 | |
07068d5b JA |
1130 | struct blk_map_ctx { |
1131 | struct blk_mq_hw_ctx *hctx; | |
1132 | struct blk_mq_ctx *ctx; | |
1133 | }; | |
1134 | ||
1135 | static struct request *blk_mq_map_request(struct request_queue *q, | |
1136 | struct bio *bio, | |
1137 | struct blk_map_ctx *data) | |
1138 | { | |
1139 | struct blk_mq_hw_ctx *hctx; | |
1140 | struct blk_mq_ctx *ctx; | |
1141 | struct request *rq; | |
1142 | int rw = bio_data_dir(bio); | |
cb96a42c | 1143 | struct blk_mq_alloc_data alloc_data; |
320ae51f | 1144 | |
07068d5b | 1145 | if (unlikely(blk_mq_queue_enter(q))) { |
320ae51f | 1146 | bio_endio(bio, -EIO); |
07068d5b | 1147 | return NULL; |
320ae51f JA |
1148 | } |
1149 | ||
1150 | ctx = blk_mq_get_ctx(q); | |
1151 | hctx = q->mq_ops->map_queue(q, ctx->cpu); | |
1152 | ||
07068d5b | 1153 | if (rw_is_sync(bio->bi_rw)) |
27fbf4e8 | 1154 | rw |= REQ_SYNC; |
07068d5b | 1155 | |
320ae51f | 1156 | trace_block_getrq(q, bio, rw); |
cb96a42c ML |
1157 | blk_mq_set_alloc_data(&alloc_data, q, GFP_ATOMIC, false, ctx, |
1158 | hctx); | |
1159 | rq = __blk_mq_alloc_request(&alloc_data, rw); | |
5dee8577 | 1160 | if (unlikely(!rq)) { |
793597a6 | 1161 | __blk_mq_run_hw_queue(hctx); |
320ae51f JA |
1162 | blk_mq_put_ctx(ctx); |
1163 | trace_block_sleeprq(q, bio, rw); | |
793597a6 CH |
1164 | |
1165 | ctx = blk_mq_get_ctx(q); | |
320ae51f | 1166 | hctx = q->mq_ops->map_queue(q, ctx->cpu); |
cb96a42c ML |
1167 | blk_mq_set_alloc_data(&alloc_data, q, |
1168 | __GFP_WAIT|GFP_ATOMIC, false, ctx, hctx); | |
1169 | rq = __blk_mq_alloc_request(&alloc_data, rw); | |
1170 | ctx = alloc_data.ctx; | |
1171 | hctx = alloc_data.hctx; | |
320ae51f JA |
1172 | } |
1173 | ||
1174 | hctx->queued++; | |
07068d5b JA |
1175 | data->hctx = hctx; |
1176 | data->ctx = ctx; | |
1177 | return rq; | |
1178 | } | |
1179 | ||
1180 | /* | |
1181 | * Multiple hardware queue variant. This will not use per-process plugs, | |
1182 | * but will attempt to bypass the hctx queueing if we can go straight to | |
1183 | * hardware for SYNC IO. | |
1184 | */ | |
1185 | static void blk_mq_make_request(struct request_queue *q, struct bio *bio) | |
1186 | { | |
1187 | const int is_sync = rw_is_sync(bio->bi_rw); | |
1188 | const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA); | |
1189 | struct blk_map_ctx data; | |
1190 | struct request *rq; | |
1191 | ||
1192 | blk_queue_bounce(q, &bio); | |
1193 | ||
1194 | if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) { | |
1195 | bio_endio(bio, -EIO); | |
1196 | return; | |
1197 | } | |
1198 | ||
1199 | rq = blk_mq_map_request(q, bio, &data); | |
1200 | if (unlikely(!rq)) | |
1201 | return; | |
1202 | ||
1203 | if (unlikely(is_flush_fua)) { | |
1204 | blk_mq_bio_to_request(rq, bio); | |
1205 | blk_insert_flush(rq); | |
1206 | goto run_queue; | |
1207 | } | |
1208 | ||
1209 | if (is_sync) { | |
1210 | int ret; | |
1211 | ||
1212 | blk_mq_bio_to_request(rq, bio); | |
1213 | blk_mq_start_request(rq, true); | |
feff6894 | 1214 | blk_add_timer(rq); |
07068d5b JA |
1215 | |
1216 | /* | |
1217 | * For OK queue, we are done. For error, kill it. Any other | |
1218 | * error (busy), just add it to our list as we previously | |
1219 | * would have done | |
1220 | */ | |
1221 | ret = q->mq_ops->queue_rq(data.hctx, rq); | |
1222 | if (ret == BLK_MQ_RQ_QUEUE_OK) | |
1223 | goto done; | |
1224 | else { | |
1225 | __blk_mq_requeue_request(rq); | |
1226 | ||
1227 | if (ret == BLK_MQ_RQ_QUEUE_ERROR) { | |
1228 | rq->errors = -EIO; | |
1229 | blk_mq_end_io(rq, rq->errors); | |
1230 | goto done; | |
1231 | } | |
1232 | } | |
1233 | } | |
1234 | ||
1235 | if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) { | |
1236 | /* | |
1237 | * For a SYNC request, send it to the hardware immediately. For | |
1238 | * an ASYNC request, just ensure that we run it later on. The | |
1239 | * latter allows for merging opportunities and more efficient | |
1240 | * dispatching. | |
1241 | */ | |
1242 | run_queue: | |
1243 | blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua); | |
1244 | } | |
1245 | done: | |
1246 | blk_mq_put_ctx(data.ctx); | |
1247 | } | |
1248 | ||
1249 | /* | |
1250 | * Single hardware queue variant. This will attempt to use any per-process | |
1251 | * plug for merging and IO deferral. | |
1252 | */ | |
1253 | static void blk_sq_make_request(struct request_queue *q, struct bio *bio) | |
1254 | { | |
1255 | const int is_sync = rw_is_sync(bio->bi_rw); | |
1256 | const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA); | |
1257 | unsigned int use_plug, request_count = 0; | |
1258 | struct blk_map_ctx data; | |
1259 | struct request *rq; | |
1260 | ||
1261 | /* | |
1262 | * If we have multiple hardware queues, just go directly to | |
1263 | * one of those for sync IO. | |
1264 | */ | |
1265 | use_plug = !is_flush_fua && !is_sync; | |
1266 | ||
1267 | blk_queue_bounce(q, &bio); | |
1268 | ||
1269 | if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) { | |
1270 | bio_endio(bio, -EIO); | |
1271 | return; | |
1272 | } | |
1273 | ||
1274 | if (use_plug && !blk_queue_nomerges(q) && | |
1275 | blk_attempt_plug_merge(q, bio, &request_count)) | |
1276 | return; | |
1277 | ||
1278 | rq = blk_mq_map_request(q, bio, &data); | |
ff87bcec JA |
1279 | if (unlikely(!rq)) |
1280 | return; | |
320ae51f JA |
1281 | |
1282 | if (unlikely(is_flush_fua)) { | |
1283 | blk_mq_bio_to_request(rq, bio); | |
320ae51f JA |
1284 | blk_insert_flush(rq); |
1285 | goto run_queue; | |
1286 | } | |
1287 | ||
1288 | /* | |
1289 | * A task plug currently exists. Since this is completely lockless, | |
1290 | * utilize that to temporarily store requests until the task is | |
1291 | * either done or scheduled away. | |
1292 | */ | |
1293 | if (use_plug) { | |
1294 | struct blk_plug *plug = current->plug; | |
1295 | ||
1296 | if (plug) { | |
1297 | blk_mq_bio_to_request(rq, bio); | |
92f399c7 | 1298 | if (list_empty(&plug->mq_list)) |
320ae51f JA |
1299 | trace_block_plug(q); |
1300 | else if (request_count >= BLK_MAX_REQUEST_COUNT) { | |
1301 | blk_flush_plug_list(plug, false); | |
1302 | trace_block_plug(q); | |
1303 | } | |
1304 | list_add_tail(&rq->queuelist, &plug->mq_list); | |
07068d5b | 1305 | blk_mq_put_ctx(data.ctx); |
320ae51f JA |
1306 | return; |
1307 | } | |
1308 | } | |
1309 | ||
07068d5b JA |
1310 | if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) { |
1311 | /* | |
1312 | * For a SYNC request, send it to the hardware immediately. For | |
1313 | * an ASYNC request, just ensure that we run it later on. The | |
1314 | * latter allows for merging opportunities and more efficient | |
1315 | * dispatching. | |
1316 | */ | |
1317 | run_queue: | |
1318 | blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua); | |
320ae51f JA |
1319 | } |
1320 | ||
07068d5b | 1321 | blk_mq_put_ctx(data.ctx); |
320ae51f JA |
1322 | } |
1323 | ||
1324 | /* | |
1325 | * Default mapping to a software queue, since we use one per CPU. | |
1326 | */ | |
1327 | struct blk_mq_hw_ctx *blk_mq_map_queue(struct request_queue *q, const int cpu) | |
1328 | { | |
1329 | return q->queue_hw_ctx[q->mq_map[cpu]]; | |
1330 | } | |
1331 | EXPORT_SYMBOL(blk_mq_map_queue); | |
1332 | ||
24d2f903 CH |
1333 | static void blk_mq_free_rq_map(struct blk_mq_tag_set *set, |
1334 | struct blk_mq_tags *tags, unsigned int hctx_idx) | |
95363efd | 1335 | { |
e9b267d9 | 1336 | struct page *page; |
320ae51f | 1337 | |
24d2f903 | 1338 | if (tags->rqs && set->ops->exit_request) { |
e9b267d9 | 1339 | int i; |
320ae51f | 1340 | |
24d2f903 CH |
1341 | for (i = 0; i < tags->nr_tags; i++) { |
1342 | if (!tags->rqs[i]) | |
e9b267d9 | 1343 | continue; |
24d2f903 CH |
1344 | set->ops->exit_request(set->driver_data, tags->rqs[i], |
1345 | hctx_idx, i); | |
e9b267d9 | 1346 | } |
320ae51f | 1347 | } |
320ae51f | 1348 | |
24d2f903 CH |
1349 | while (!list_empty(&tags->page_list)) { |
1350 | page = list_first_entry(&tags->page_list, struct page, lru); | |
6753471c | 1351 | list_del_init(&page->lru); |
320ae51f JA |
1352 | __free_pages(page, page->private); |
1353 | } | |
1354 | ||
24d2f903 | 1355 | kfree(tags->rqs); |
320ae51f | 1356 | |
24d2f903 | 1357 | blk_mq_free_tags(tags); |
320ae51f JA |
1358 | } |
1359 | ||
1360 | static size_t order_to_size(unsigned int order) | |
1361 | { | |
4ca08500 | 1362 | return (size_t)PAGE_SIZE << order; |
320ae51f JA |
1363 | } |
1364 | ||
24d2f903 CH |
1365 | static struct blk_mq_tags *blk_mq_init_rq_map(struct blk_mq_tag_set *set, |
1366 | unsigned int hctx_idx) | |
320ae51f | 1367 | { |
24d2f903 | 1368 | struct blk_mq_tags *tags; |
320ae51f JA |
1369 | unsigned int i, j, entries_per_page, max_order = 4; |
1370 | size_t rq_size, left; | |
1371 | ||
24d2f903 CH |
1372 | tags = blk_mq_init_tags(set->queue_depth, set->reserved_tags, |
1373 | set->numa_node); | |
1374 | if (!tags) | |
1375 | return NULL; | |
320ae51f | 1376 | |
24d2f903 CH |
1377 | INIT_LIST_HEAD(&tags->page_list); |
1378 | ||
1379 | tags->rqs = kmalloc_node(set->queue_depth * sizeof(struct request *), | |
1380 | GFP_KERNEL, set->numa_node); | |
1381 | if (!tags->rqs) { | |
1382 | blk_mq_free_tags(tags); | |
1383 | return NULL; | |
1384 | } | |
320ae51f JA |
1385 | |
1386 | /* | |
1387 | * rq_size is the size of the request plus driver payload, rounded | |
1388 | * to the cacheline size | |
1389 | */ | |
24d2f903 | 1390 | rq_size = round_up(sizeof(struct request) + set->cmd_size, |
320ae51f | 1391 | cache_line_size()); |
24d2f903 | 1392 | left = rq_size * set->queue_depth; |
320ae51f | 1393 | |
24d2f903 | 1394 | for (i = 0; i < set->queue_depth; ) { |
320ae51f JA |
1395 | int this_order = max_order; |
1396 | struct page *page; | |
1397 | int to_do; | |
1398 | void *p; | |
1399 | ||
1400 | while (left < order_to_size(this_order - 1) && this_order) | |
1401 | this_order--; | |
1402 | ||
1403 | do { | |
24d2f903 CH |
1404 | page = alloc_pages_node(set->numa_node, GFP_KERNEL, |
1405 | this_order); | |
320ae51f JA |
1406 | if (page) |
1407 | break; | |
1408 | if (!this_order--) | |
1409 | break; | |
1410 | if (order_to_size(this_order) < rq_size) | |
1411 | break; | |
1412 | } while (1); | |
1413 | ||
1414 | if (!page) | |
24d2f903 | 1415 | goto fail; |
320ae51f JA |
1416 | |
1417 | page->private = this_order; | |
24d2f903 | 1418 | list_add_tail(&page->lru, &tags->page_list); |
320ae51f JA |
1419 | |
1420 | p = page_address(page); | |
1421 | entries_per_page = order_to_size(this_order) / rq_size; | |
24d2f903 | 1422 | to_do = min(entries_per_page, set->queue_depth - i); |
320ae51f JA |
1423 | left -= to_do * rq_size; |
1424 | for (j = 0; j < to_do; j++) { | |
24d2f903 CH |
1425 | tags->rqs[i] = p; |
1426 | if (set->ops->init_request) { | |
1427 | if (set->ops->init_request(set->driver_data, | |
1428 | tags->rqs[i], hctx_idx, i, | |
1429 | set->numa_node)) | |
1430 | goto fail; | |
e9b267d9 CH |
1431 | } |
1432 | ||
320ae51f JA |
1433 | p += rq_size; |
1434 | i++; | |
1435 | } | |
1436 | } | |
1437 | ||
24d2f903 | 1438 | return tags; |
320ae51f | 1439 | |
24d2f903 CH |
1440 | fail: |
1441 | pr_warn("%s: failed to allocate requests\n", __func__); | |
1442 | blk_mq_free_rq_map(set, tags, hctx_idx); | |
1443 | return NULL; | |
320ae51f JA |
1444 | } |
1445 | ||
1429d7c9 JA |
1446 | static void blk_mq_free_bitmap(struct blk_mq_ctxmap *bitmap) |
1447 | { | |
1448 | kfree(bitmap->map); | |
1449 | } | |
1450 | ||
1451 | static int blk_mq_alloc_bitmap(struct blk_mq_ctxmap *bitmap, int node) | |
1452 | { | |
1453 | unsigned int bpw = 8, total, num_maps, i; | |
1454 | ||
1455 | bitmap->bits_per_word = bpw; | |
1456 | ||
1457 | num_maps = ALIGN(nr_cpu_ids, bpw) / bpw; | |
1458 | bitmap->map = kzalloc_node(num_maps * sizeof(struct blk_align_bitmap), | |
1459 | GFP_KERNEL, node); | |
1460 | if (!bitmap->map) | |
1461 | return -ENOMEM; | |
1462 | ||
1463 | bitmap->map_size = num_maps; | |
1464 | ||
1465 | total = nr_cpu_ids; | |
1466 | for (i = 0; i < num_maps; i++) { | |
1467 | bitmap->map[i].depth = min(total, bitmap->bits_per_word); | |
1468 | total -= bitmap->map[i].depth; | |
1469 | } | |
1470 | ||
1471 | return 0; | |
1472 | } | |
1473 | ||
484b4061 JA |
1474 | static int blk_mq_hctx_cpu_offline(struct blk_mq_hw_ctx *hctx, int cpu) |
1475 | { | |
1476 | struct request_queue *q = hctx->queue; | |
1477 | struct blk_mq_ctx *ctx; | |
1478 | LIST_HEAD(tmp); | |
1479 | ||
1480 | /* | |
1481 | * Move ctx entries to new CPU, if this one is going away. | |
1482 | */ | |
1483 | ctx = __blk_mq_get_ctx(q, cpu); | |
1484 | ||
1485 | spin_lock(&ctx->lock); | |
1486 | if (!list_empty(&ctx->rq_list)) { | |
1487 | list_splice_init(&ctx->rq_list, &tmp); | |
1488 | blk_mq_hctx_clear_pending(hctx, ctx); | |
1489 | } | |
1490 | spin_unlock(&ctx->lock); | |
1491 | ||
1492 | if (list_empty(&tmp)) | |
1493 | return NOTIFY_OK; | |
1494 | ||
1495 | ctx = blk_mq_get_ctx(q); | |
1496 | spin_lock(&ctx->lock); | |
1497 | ||
1498 | while (!list_empty(&tmp)) { | |
1499 | struct request *rq; | |
1500 | ||
1501 | rq = list_first_entry(&tmp, struct request, queuelist); | |
1502 | rq->mq_ctx = ctx; | |
1503 | list_move_tail(&rq->queuelist, &ctx->rq_list); | |
1504 | } | |
1505 | ||
1506 | hctx = q->mq_ops->map_queue(q, ctx->cpu); | |
1507 | blk_mq_hctx_mark_pending(hctx, ctx); | |
1508 | ||
1509 | spin_unlock(&ctx->lock); | |
1510 | ||
1511 | blk_mq_run_hw_queue(hctx, true); | |
1512 | blk_mq_put_ctx(ctx); | |
1513 | return NOTIFY_OK; | |
1514 | } | |
1515 | ||
1516 | static int blk_mq_hctx_cpu_online(struct blk_mq_hw_ctx *hctx, int cpu) | |
1517 | { | |
1518 | struct request_queue *q = hctx->queue; | |
1519 | struct blk_mq_tag_set *set = q->tag_set; | |
1520 | ||
1521 | if (set->tags[hctx->queue_num]) | |
1522 | return NOTIFY_OK; | |
1523 | ||
1524 | set->tags[hctx->queue_num] = blk_mq_init_rq_map(set, hctx->queue_num); | |
1525 | if (!set->tags[hctx->queue_num]) | |
1526 | return NOTIFY_STOP; | |
1527 | ||
1528 | hctx->tags = set->tags[hctx->queue_num]; | |
1529 | return NOTIFY_OK; | |
1530 | } | |
1531 | ||
1532 | static int blk_mq_hctx_notify(void *data, unsigned long action, | |
1533 | unsigned int cpu) | |
1534 | { | |
1535 | struct blk_mq_hw_ctx *hctx = data; | |
1536 | ||
1537 | if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) | |
1538 | return blk_mq_hctx_cpu_offline(hctx, cpu); | |
1539 | else if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) | |
1540 | return blk_mq_hctx_cpu_online(hctx, cpu); | |
1541 | ||
1542 | return NOTIFY_OK; | |
1543 | } | |
1544 | ||
624dbe47 ML |
1545 | static void blk_mq_exit_hw_queues(struct request_queue *q, |
1546 | struct blk_mq_tag_set *set, int nr_queue) | |
1547 | { | |
1548 | struct blk_mq_hw_ctx *hctx; | |
1549 | unsigned int i; | |
1550 | ||
1551 | queue_for_each_hw_ctx(q, hctx, i) { | |
1552 | if (i == nr_queue) | |
1553 | break; | |
1554 | ||
f899fed4 JA |
1555 | blk_mq_tag_idle(hctx); |
1556 | ||
624dbe47 ML |
1557 | if (set->ops->exit_hctx) |
1558 | set->ops->exit_hctx(hctx, i); | |
1559 | ||
1560 | blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier); | |
1561 | kfree(hctx->ctxs); | |
1562 | blk_mq_free_bitmap(&hctx->ctx_map); | |
1563 | } | |
1564 | ||
1565 | } | |
1566 | ||
1567 | static void blk_mq_free_hw_queues(struct request_queue *q, | |
1568 | struct blk_mq_tag_set *set) | |
1569 | { | |
1570 | struct blk_mq_hw_ctx *hctx; | |
1571 | unsigned int i; | |
1572 | ||
1573 | queue_for_each_hw_ctx(q, hctx, i) { | |
1574 | free_cpumask_var(hctx->cpumask); | |
cdef54dd | 1575 | kfree(hctx); |
624dbe47 ML |
1576 | } |
1577 | } | |
1578 | ||
320ae51f | 1579 | static int blk_mq_init_hw_queues(struct request_queue *q, |
24d2f903 | 1580 | struct blk_mq_tag_set *set) |
320ae51f JA |
1581 | { |
1582 | struct blk_mq_hw_ctx *hctx; | |
624dbe47 | 1583 | unsigned int i; |
320ae51f JA |
1584 | |
1585 | /* | |
1586 | * Initialize hardware queues | |
1587 | */ | |
1588 | queue_for_each_hw_ctx(q, hctx, i) { | |
320ae51f JA |
1589 | int node; |
1590 | ||
1591 | node = hctx->numa_node; | |
1592 | if (node == NUMA_NO_NODE) | |
24d2f903 | 1593 | node = hctx->numa_node = set->numa_node; |
320ae51f | 1594 | |
70f4db63 CH |
1595 | INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn); |
1596 | INIT_DELAYED_WORK(&hctx->delay_work, blk_mq_delay_work_fn); | |
320ae51f JA |
1597 | spin_lock_init(&hctx->lock); |
1598 | INIT_LIST_HEAD(&hctx->dispatch); | |
1599 | hctx->queue = q; | |
1600 | hctx->queue_num = i; | |
24d2f903 CH |
1601 | hctx->flags = set->flags; |
1602 | hctx->cmd_size = set->cmd_size; | |
320ae51f JA |
1603 | |
1604 | blk_mq_init_cpu_notifier(&hctx->cpu_notifier, | |
1605 | blk_mq_hctx_notify, hctx); | |
1606 | blk_mq_register_cpu_notifier(&hctx->cpu_notifier); | |
1607 | ||
24d2f903 | 1608 | hctx->tags = set->tags[i]; |
320ae51f JA |
1609 | |
1610 | /* | |
1611 | * Allocate space for all possible cpus to avoid allocation in | |
1612 | * runtime | |
1613 | */ | |
1614 | hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *), | |
1615 | GFP_KERNEL, node); | |
1616 | if (!hctx->ctxs) | |
1617 | break; | |
1618 | ||
1429d7c9 | 1619 | if (blk_mq_alloc_bitmap(&hctx->ctx_map, node)) |
320ae51f JA |
1620 | break; |
1621 | ||
320ae51f JA |
1622 | hctx->nr_ctx = 0; |
1623 | ||
24d2f903 CH |
1624 | if (set->ops->init_hctx && |
1625 | set->ops->init_hctx(hctx, set->driver_data, i)) | |
320ae51f JA |
1626 | break; |
1627 | } | |
1628 | ||
1629 | if (i == q->nr_hw_queues) | |
1630 | return 0; | |
1631 | ||
1632 | /* | |
1633 | * Init failed | |
1634 | */ | |
624dbe47 | 1635 | blk_mq_exit_hw_queues(q, set, i); |
320ae51f JA |
1636 | |
1637 | return 1; | |
1638 | } | |
1639 | ||
1640 | static void blk_mq_init_cpu_queues(struct request_queue *q, | |
1641 | unsigned int nr_hw_queues) | |
1642 | { | |
1643 | unsigned int i; | |
1644 | ||
1645 | for_each_possible_cpu(i) { | |
1646 | struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i); | |
1647 | struct blk_mq_hw_ctx *hctx; | |
1648 | ||
1649 | memset(__ctx, 0, sizeof(*__ctx)); | |
1650 | __ctx->cpu = i; | |
1651 | spin_lock_init(&__ctx->lock); | |
1652 | INIT_LIST_HEAD(&__ctx->rq_list); | |
1653 | __ctx->queue = q; | |
1654 | ||
1655 | /* If the cpu isn't online, the cpu is mapped to first hctx */ | |
320ae51f JA |
1656 | if (!cpu_online(i)) |
1657 | continue; | |
1658 | ||
e4043dcf JA |
1659 | hctx = q->mq_ops->map_queue(q, i); |
1660 | cpumask_set_cpu(i, hctx->cpumask); | |
1661 | hctx->nr_ctx++; | |
1662 | ||
320ae51f JA |
1663 | /* |
1664 | * Set local node, IFF we have more than one hw queue. If | |
1665 | * not, we remain on the home node of the device | |
1666 | */ | |
1667 | if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE) | |
1668 | hctx->numa_node = cpu_to_node(i); | |
1669 | } | |
1670 | } | |
1671 | ||
1672 | static void blk_mq_map_swqueue(struct request_queue *q) | |
1673 | { | |
1674 | unsigned int i; | |
1675 | struct blk_mq_hw_ctx *hctx; | |
1676 | struct blk_mq_ctx *ctx; | |
1677 | ||
1678 | queue_for_each_hw_ctx(q, hctx, i) { | |
e4043dcf | 1679 | cpumask_clear(hctx->cpumask); |
320ae51f JA |
1680 | hctx->nr_ctx = 0; |
1681 | } | |
1682 | ||
1683 | /* | |
1684 | * Map software to hardware queues | |
1685 | */ | |
1686 | queue_for_each_ctx(q, ctx, i) { | |
1687 | /* If the cpu isn't online, the cpu is mapped to first hctx */ | |
e4043dcf JA |
1688 | if (!cpu_online(i)) |
1689 | continue; | |
1690 | ||
320ae51f | 1691 | hctx = q->mq_ops->map_queue(q, i); |
e4043dcf | 1692 | cpumask_set_cpu(i, hctx->cpumask); |
320ae51f JA |
1693 | ctx->index_hw = hctx->nr_ctx; |
1694 | hctx->ctxs[hctx->nr_ctx++] = ctx; | |
1695 | } | |
506e931f JA |
1696 | |
1697 | queue_for_each_hw_ctx(q, hctx, i) { | |
484b4061 JA |
1698 | /* |
1699 | * If not software queues are mapped to this hardware queue, | |
1700 | * disable it and free the request entries | |
1701 | */ | |
1702 | if (!hctx->nr_ctx) { | |
1703 | struct blk_mq_tag_set *set = q->tag_set; | |
1704 | ||
1705 | if (set->tags[i]) { | |
1706 | blk_mq_free_rq_map(set, set->tags[i], i); | |
1707 | set->tags[i] = NULL; | |
1708 | hctx->tags = NULL; | |
1709 | } | |
1710 | continue; | |
1711 | } | |
1712 | ||
1713 | /* | |
1714 | * Initialize batch roundrobin counts | |
1715 | */ | |
506e931f JA |
1716 | hctx->next_cpu = cpumask_first(hctx->cpumask); |
1717 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; | |
1718 | } | |
320ae51f JA |
1719 | } |
1720 | ||
0d2602ca JA |
1721 | static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set) |
1722 | { | |
1723 | struct blk_mq_hw_ctx *hctx; | |
1724 | struct request_queue *q; | |
1725 | bool shared; | |
1726 | int i; | |
1727 | ||
1728 | if (set->tag_list.next == set->tag_list.prev) | |
1729 | shared = false; | |
1730 | else | |
1731 | shared = true; | |
1732 | ||
1733 | list_for_each_entry(q, &set->tag_list, tag_set_list) { | |
1734 | blk_mq_freeze_queue(q); | |
1735 | ||
1736 | queue_for_each_hw_ctx(q, hctx, i) { | |
1737 | if (shared) | |
1738 | hctx->flags |= BLK_MQ_F_TAG_SHARED; | |
1739 | else | |
1740 | hctx->flags &= ~BLK_MQ_F_TAG_SHARED; | |
1741 | } | |
1742 | blk_mq_unfreeze_queue(q); | |
1743 | } | |
1744 | } | |
1745 | ||
1746 | static void blk_mq_del_queue_tag_set(struct request_queue *q) | |
1747 | { | |
1748 | struct blk_mq_tag_set *set = q->tag_set; | |
1749 | ||
1750 | blk_mq_freeze_queue(q); | |
1751 | ||
1752 | mutex_lock(&set->tag_list_lock); | |
1753 | list_del_init(&q->tag_set_list); | |
1754 | blk_mq_update_tag_set_depth(set); | |
1755 | mutex_unlock(&set->tag_list_lock); | |
1756 | ||
1757 | blk_mq_unfreeze_queue(q); | |
1758 | } | |
1759 | ||
1760 | static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set, | |
1761 | struct request_queue *q) | |
1762 | { | |
1763 | q->tag_set = set; | |
1764 | ||
1765 | mutex_lock(&set->tag_list_lock); | |
1766 | list_add_tail(&q->tag_set_list, &set->tag_list); | |
1767 | blk_mq_update_tag_set_depth(set); | |
1768 | mutex_unlock(&set->tag_list_lock); | |
1769 | } | |
1770 | ||
24d2f903 | 1771 | struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set) |
320ae51f JA |
1772 | { |
1773 | struct blk_mq_hw_ctx **hctxs; | |
e6cdb092 | 1774 | struct blk_mq_ctx __percpu *ctx; |
320ae51f | 1775 | struct request_queue *q; |
f14bbe77 | 1776 | unsigned int *map; |
320ae51f JA |
1777 | int i; |
1778 | ||
320ae51f JA |
1779 | ctx = alloc_percpu(struct blk_mq_ctx); |
1780 | if (!ctx) | |
1781 | return ERR_PTR(-ENOMEM); | |
1782 | ||
24d2f903 CH |
1783 | hctxs = kmalloc_node(set->nr_hw_queues * sizeof(*hctxs), GFP_KERNEL, |
1784 | set->numa_node); | |
320ae51f JA |
1785 | |
1786 | if (!hctxs) | |
1787 | goto err_percpu; | |
1788 | ||
f14bbe77 JA |
1789 | map = blk_mq_make_queue_map(set); |
1790 | if (!map) | |
1791 | goto err_map; | |
1792 | ||
24d2f903 | 1793 | for (i = 0; i < set->nr_hw_queues; i++) { |
f14bbe77 JA |
1794 | int node = blk_mq_hw_queue_to_node(map, i); |
1795 | ||
cdef54dd CH |
1796 | hctxs[i] = kzalloc_node(sizeof(struct blk_mq_hw_ctx), |
1797 | GFP_KERNEL, node); | |
320ae51f JA |
1798 | if (!hctxs[i]) |
1799 | goto err_hctxs; | |
1800 | ||
e4043dcf JA |
1801 | if (!zalloc_cpumask_var(&hctxs[i]->cpumask, GFP_KERNEL)) |
1802 | goto err_hctxs; | |
1803 | ||
0d2602ca | 1804 | atomic_set(&hctxs[i]->nr_active, 0); |
f14bbe77 | 1805 | hctxs[i]->numa_node = node; |
320ae51f JA |
1806 | hctxs[i]->queue_num = i; |
1807 | } | |
1808 | ||
24d2f903 | 1809 | q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node); |
320ae51f JA |
1810 | if (!q) |
1811 | goto err_hctxs; | |
1812 | ||
3d2936f4 ML |
1813 | if (percpu_counter_init(&q->mq_usage_counter, 0)) |
1814 | goto err_map; | |
1815 | ||
320ae51f JA |
1816 | setup_timer(&q->timeout, blk_mq_rq_timer, (unsigned long) q); |
1817 | blk_queue_rq_timeout(q, 30000); | |
1818 | ||
1819 | q->nr_queues = nr_cpu_ids; | |
24d2f903 | 1820 | q->nr_hw_queues = set->nr_hw_queues; |
f14bbe77 | 1821 | q->mq_map = map; |
320ae51f JA |
1822 | |
1823 | q->queue_ctx = ctx; | |
1824 | q->queue_hw_ctx = hctxs; | |
1825 | ||
24d2f903 | 1826 | q->mq_ops = set->ops; |
94eddfbe | 1827 | q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT; |
320ae51f | 1828 | |
05f1dd53 JA |
1829 | if (!(set->flags & BLK_MQ_F_SG_MERGE)) |
1830 | q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE; | |
1831 | ||
1be036e9 CH |
1832 | q->sg_reserved_size = INT_MAX; |
1833 | ||
6fca6a61 CH |
1834 | INIT_WORK(&q->requeue_work, blk_mq_requeue_work); |
1835 | INIT_LIST_HEAD(&q->requeue_list); | |
1836 | spin_lock_init(&q->requeue_lock); | |
1837 | ||
07068d5b JA |
1838 | if (q->nr_hw_queues > 1) |
1839 | blk_queue_make_request(q, blk_mq_make_request); | |
1840 | else | |
1841 | blk_queue_make_request(q, blk_sq_make_request); | |
1842 | ||
87ee7b11 | 1843 | blk_queue_rq_timed_out(q, blk_mq_rq_timed_out); |
24d2f903 CH |
1844 | if (set->timeout) |
1845 | blk_queue_rq_timeout(q, set->timeout); | |
320ae51f | 1846 | |
eba71768 JA |
1847 | /* |
1848 | * Do this after blk_queue_make_request() overrides it... | |
1849 | */ | |
1850 | q->nr_requests = set->queue_depth; | |
1851 | ||
24d2f903 CH |
1852 | if (set->ops->complete) |
1853 | blk_queue_softirq_done(q, set->ops->complete); | |
30a91cb4 | 1854 | |
320ae51f | 1855 | blk_mq_init_flush(q); |
24d2f903 | 1856 | blk_mq_init_cpu_queues(q, set->nr_hw_queues); |
320ae51f | 1857 | |
24d2f903 CH |
1858 | q->flush_rq = kzalloc(round_up(sizeof(struct request) + |
1859 | set->cmd_size, cache_line_size()), | |
1860 | GFP_KERNEL); | |
18741986 | 1861 | if (!q->flush_rq) |
320ae51f JA |
1862 | goto err_hw; |
1863 | ||
24d2f903 | 1864 | if (blk_mq_init_hw_queues(q, set)) |
18741986 CH |
1865 | goto err_flush_rq; |
1866 | ||
320ae51f JA |
1867 | mutex_lock(&all_q_mutex); |
1868 | list_add_tail(&q->all_q_node, &all_q_list); | |
1869 | mutex_unlock(&all_q_mutex); | |
1870 | ||
0d2602ca JA |
1871 | blk_mq_add_queue_tag_set(set, q); |
1872 | ||
484b4061 JA |
1873 | blk_mq_map_swqueue(q); |
1874 | ||
320ae51f | 1875 | return q; |
18741986 CH |
1876 | |
1877 | err_flush_rq: | |
1878 | kfree(q->flush_rq); | |
320ae51f | 1879 | err_hw: |
320ae51f JA |
1880 | blk_cleanup_queue(q); |
1881 | err_hctxs: | |
f14bbe77 | 1882 | kfree(map); |
24d2f903 | 1883 | for (i = 0; i < set->nr_hw_queues; i++) { |
320ae51f JA |
1884 | if (!hctxs[i]) |
1885 | break; | |
e4043dcf | 1886 | free_cpumask_var(hctxs[i]->cpumask); |
cdef54dd | 1887 | kfree(hctxs[i]); |
320ae51f | 1888 | } |
f14bbe77 | 1889 | err_map: |
320ae51f JA |
1890 | kfree(hctxs); |
1891 | err_percpu: | |
1892 | free_percpu(ctx); | |
1893 | return ERR_PTR(-ENOMEM); | |
1894 | } | |
1895 | EXPORT_SYMBOL(blk_mq_init_queue); | |
1896 | ||
1897 | void blk_mq_free_queue(struct request_queue *q) | |
1898 | { | |
624dbe47 | 1899 | struct blk_mq_tag_set *set = q->tag_set; |
320ae51f | 1900 | |
0d2602ca JA |
1901 | blk_mq_del_queue_tag_set(q); |
1902 | ||
624dbe47 ML |
1903 | blk_mq_exit_hw_queues(q, set, set->nr_hw_queues); |
1904 | blk_mq_free_hw_queues(q, set); | |
320ae51f | 1905 | |
3d2936f4 ML |
1906 | percpu_counter_destroy(&q->mq_usage_counter); |
1907 | ||
320ae51f JA |
1908 | free_percpu(q->queue_ctx); |
1909 | kfree(q->queue_hw_ctx); | |
1910 | kfree(q->mq_map); | |
1911 | ||
1912 | q->queue_ctx = NULL; | |
1913 | q->queue_hw_ctx = NULL; | |
1914 | q->mq_map = NULL; | |
1915 | ||
1916 | mutex_lock(&all_q_mutex); | |
1917 | list_del_init(&q->all_q_node); | |
1918 | mutex_unlock(&all_q_mutex); | |
1919 | } | |
320ae51f JA |
1920 | |
1921 | /* Basically redo blk_mq_init_queue with queue frozen */ | |
f618ef7c | 1922 | static void blk_mq_queue_reinit(struct request_queue *q) |
320ae51f JA |
1923 | { |
1924 | blk_mq_freeze_queue(q); | |
1925 | ||
67aec14c JA |
1926 | blk_mq_sysfs_unregister(q); |
1927 | ||
320ae51f JA |
1928 | blk_mq_update_queue_map(q->mq_map, q->nr_hw_queues); |
1929 | ||
1930 | /* | |
1931 | * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe | |
1932 | * we should change hctx numa_node according to new topology (this | |
1933 | * involves free and re-allocate memory, worthy doing?) | |
1934 | */ | |
1935 | ||
1936 | blk_mq_map_swqueue(q); | |
1937 | ||
67aec14c JA |
1938 | blk_mq_sysfs_register(q); |
1939 | ||
320ae51f JA |
1940 | blk_mq_unfreeze_queue(q); |
1941 | } | |
1942 | ||
f618ef7c PG |
1943 | static int blk_mq_queue_reinit_notify(struct notifier_block *nb, |
1944 | unsigned long action, void *hcpu) | |
320ae51f JA |
1945 | { |
1946 | struct request_queue *q; | |
1947 | ||
1948 | /* | |
9fccfed8 JA |
1949 | * Before new mappings are established, hotadded cpu might already |
1950 | * start handling requests. This doesn't break anything as we map | |
1951 | * offline CPUs to first hardware queue. We will re-init the queue | |
1952 | * below to get optimal settings. | |
320ae51f JA |
1953 | */ |
1954 | if (action != CPU_DEAD && action != CPU_DEAD_FROZEN && | |
1955 | action != CPU_ONLINE && action != CPU_ONLINE_FROZEN) | |
1956 | return NOTIFY_OK; | |
1957 | ||
1958 | mutex_lock(&all_q_mutex); | |
1959 | list_for_each_entry(q, &all_q_list, all_q_node) | |
1960 | blk_mq_queue_reinit(q); | |
1961 | mutex_unlock(&all_q_mutex); | |
1962 | return NOTIFY_OK; | |
1963 | } | |
1964 | ||
24d2f903 CH |
1965 | int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set) |
1966 | { | |
1967 | int i; | |
1968 | ||
1969 | if (!set->nr_hw_queues) | |
1970 | return -EINVAL; | |
1971 | if (!set->queue_depth || set->queue_depth > BLK_MQ_MAX_DEPTH) | |
1972 | return -EINVAL; | |
1973 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) | |
1974 | return -EINVAL; | |
1975 | ||
cdef54dd | 1976 | if (!set->nr_hw_queues || !set->ops->queue_rq || !set->ops->map_queue) |
24d2f903 CH |
1977 | return -EINVAL; |
1978 | ||
1979 | ||
48479005 ML |
1980 | set->tags = kmalloc_node(set->nr_hw_queues * |
1981 | sizeof(struct blk_mq_tags *), | |
24d2f903 CH |
1982 | GFP_KERNEL, set->numa_node); |
1983 | if (!set->tags) | |
1984 | goto out; | |
1985 | ||
1986 | for (i = 0; i < set->nr_hw_queues; i++) { | |
1987 | set->tags[i] = blk_mq_init_rq_map(set, i); | |
1988 | if (!set->tags[i]) | |
1989 | goto out_unwind; | |
1990 | } | |
1991 | ||
0d2602ca JA |
1992 | mutex_init(&set->tag_list_lock); |
1993 | INIT_LIST_HEAD(&set->tag_list); | |
1994 | ||
24d2f903 CH |
1995 | return 0; |
1996 | ||
1997 | out_unwind: | |
1998 | while (--i >= 0) | |
1999 | blk_mq_free_rq_map(set, set->tags[i], i); | |
2000 | out: | |
2001 | return -ENOMEM; | |
2002 | } | |
2003 | EXPORT_SYMBOL(blk_mq_alloc_tag_set); | |
2004 | ||
2005 | void blk_mq_free_tag_set(struct blk_mq_tag_set *set) | |
2006 | { | |
2007 | int i; | |
2008 | ||
484b4061 JA |
2009 | for (i = 0; i < set->nr_hw_queues; i++) { |
2010 | if (set->tags[i]) | |
2011 | blk_mq_free_rq_map(set, set->tags[i], i); | |
2012 | } | |
2013 | ||
981bd189 | 2014 | kfree(set->tags); |
24d2f903 CH |
2015 | } |
2016 | EXPORT_SYMBOL(blk_mq_free_tag_set); | |
2017 | ||
e3a2b3f9 JA |
2018 | int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr) |
2019 | { | |
2020 | struct blk_mq_tag_set *set = q->tag_set; | |
2021 | struct blk_mq_hw_ctx *hctx; | |
2022 | int i, ret; | |
2023 | ||
2024 | if (!set || nr > set->queue_depth) | |
2025 | return -EINVAL; | |
2026 | ||
2027 | ret = 0; | |
2028 | queue_for_each_hw_ctx(q, hctx, i) { | |
2029 | ret = blk_mq_tag_update_depth(hctx->tags, nr); | |
2030 | if (ret) | |
2031 | break; | |
2032 | } | |
2033 | ||
2034 | if (!ret) | |
2035 | q->nr_requests = nr; | |
2036 | ||
2037 | return ret; | |
2038 | } | |
2039 | ||
676141e4 JA |
2040 | void blk_mq_disable_hotplug(void) |
2041 | { | |
2042 | mutex_lock(&all_q_mutex); | |
2043 | } | |
2044 | ||
2045 | void blk_mq_enable_hotplug(void) | |
2046 | { | |
2047 | mutex_unlock(&all_q_mutex); | |
2048 | } | |
2049 | ||
320ae51f JA |
2050 | static int __init blk_mq_init(void) |
2051 | { | |
320ae51f JA |
2052 | blk_mq_cpu_init(); |
2053 | ||
2054 | /* Must be called after percpu_counter_hotcpu_callback() */ | |
2055 | hotcpu_notifier(blk_mq_queue_reinit_notify, -10); | |
2056 | ||
2057 | return 0; | |
2058 | } | |
2059 | subsys_initcall(blk_mq_init); |