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 | static struct blk_mq_ctx *__blk_mq_get_ctx(struct request_queue *q, |
37 | unsigned int cpu) | |
38 | { | |
39 | return per_cpu_ptr(q->queue_ctx, cpu); | |
40 | } | |
41 | ||
42 | /* | |
43 | * This assumes per-cpu software queueing queues. They could be per-node | |
44 | * as well, for instance. For now this is hardcoded as-is. Note that we don't | |
45 | * care about preemption, since we know the ctx's are persistent. This does | |
46 | * mean that we can't rely on ctx always matching the currently running CPU. | |
47 | */ | |
48 | static struct blk_mq_ctx *blk_mq_get_ctx(struct request_queue *q) | |
49 | { | |
50 | return __blk_mq_get_ctx(q, get_cpu()); | |
51 | } | |
52 | ||
53 | static void blk_mq_put_ctx(struct blk_mq_ctx *ctx) | |
54 | { | |
55 | put_cpu(); | |
56 | } | |
57 | ||
58 | /* | |
59 | * Check if any of the ctx's have pending work in this hardware queue | |
60 | */ | |
61 | static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx) | |
62 | { | |
63 | unsigned int i; | |
64 | ||
1429d7c9 JA |
65 | for (i = 0; i < hctx->ctx_map.map_size; i++) |
66 | if (hctx->ctx_map.map[i].word) | |
320ae51f JA |
67 | return true; |
68 | ||
69 | return false; | |
70 | } | |
71 | ||
1429d7c9 JA |
72 | static inline struct blk_align_bitmap *get_bm(struct blk_mq_hw_ctx *hctx, |
73 | struct blk_mq_ctx *ctx) | |
74 | { | |
75 | return &hctx->ctx_map.map[ctx->index_hw / hctx->ctx_map.bits_per_word]; | |
76 | } | |
77 | ||
78 | #define CTX_TO_BIT(hctx, ctx) \ | |
79 | ((ctx)->index_hw & ((hctx)->ctx_map.bits_per_word - 1)) | |
80 | ||
320ae51f JA |
81 | /* |
82 | * Mark this ctx as having pending work in this hardware queue | |
83 | */ | |
84 | static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx, | |
85 | struct blk_mq_ctx *ctx) | |
86 | { | |
1429d7c9 JA |
87 | struct blk_align_bitmap *bm = get_bm(hctx, ctx); |
88 | ||
89 | if (!test_bit(CTX_TO_BIT(hctx, ctx), &bm->word)) | |
90 | set_bit(CTX_TO_BIT(hctx, ctx), &bm->word); | |
91 | } | |
92 | ||
93 | static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx, | |
94 | struct blk_mq_ctx *ctx) | |
95 | { | |
96 | struct blk_align_bitmap *bm = get_bm(hctx, ctx); | |
97 | ||
98 | clear_bit(CTX_TO_BIT(hctx, ctx), &bm->word); | |
320ae51f JA |
99 | } |
100 | ||
320ae51f JA |
101 | static int blk_mq_queue_enter(struct request_queue *q) |
102 | { | |
103 | int ret; | |
104 | ||
105 | __percpu_counter_add(&q->mq_usage_counter, 1, 1000000); | |
106 | smp_wmb(); | |
107 | /* we have problems to freeze the queue if it's initializing */ | |
108 | if (!blk_queue_bypass(q) || !blk_queue_init_done(q)) | |
109 | return 0; | |
110 | ||
111 | __percpu_counter_add(&q->mq_usage_counter, -1, 1000000); | |
112 | ||
113 | spin_lock_irq(q->queue_lock); | |
114 | ret = wait_event_interruptible_lock_irq(q->mq_freeze_wq, | |
43a5e4e2 ML |
115 | !blk_queue_bypass(q) || blk_queue_dying(q), |
116 | *q->queue_lock); | |
320ae51f | 117 | /* inc usage with lock hold to avoid freeze_queue runs here */ |
43a5e4e2 | 118 | if (!ret && !blk_queue_dying(q)) |
320ae51f | 119 | __percpu_counter_add(&q->mq_usage_counter, 1, 1000000); |
43a5e4e2 ML |
120 | else if (blk_queue_dying(q)) |
121 | ret = -ENODEV; | |
320ae51f JA |
122 | spin_unlock_irq(q->queue_lock); |
123 | ||
124 | return ret; | |
125 | } | |
126 | ||
127 | static void blk_mq_queue_exit(struct request_queue *q) | |
128 | { | |
129 | __percpu_counter_add(&q->mq_usage_counter, -1, 1000000); | |
130 | } | |
131 | ||
43a5e4e2 ML |
132 | static void __blk_mq_drain_queue(struct request_queue *q) |
133 | { | |
134 | while (true) { | |
135 | s64 count; | |
136 | ||
137 | spin_lock_irq(q->queue_lock); | |
138 | count = percpu_counter_sum(&q->mq_usage_counter); | |
139 | spin_unlock_irq(q->queue_lock); | |
140 | ||
141 | if (count == 0) | |
142 | break; | |
143 | blk_mq_run_queues(q, false); | |
144 | msleep(10); | |
145 | } | |
146 | } | |
147 | ||
320ae51f JA |
148 | /* |
149 | * Guarantee no request is in use, so we can change any data structure of | |
150 | * the queue afterward. | |
151 | */ | |
152 | static void blk_mq_freeze_queue(struct request_queue *q) | |
153 | { | |
154 | bool drain; | |
155 | ||
156 | spin_lock_irq(q->queue_lock); | |
157 | drain = !q->bypass_depth++; | |
158 | queue_flag_set(QUEUE_FLAG_BYPASS, q); | |
159 | spin_unlock_irq(q->queue_lock); | |
160 | ||
43a5e4e2 ML |
161 | if (drain) |
162 | __blk_mq_drain_queue(q); | |
163 | } | |
320ae51f | 164 | |
43a5e4e2 ML |
165 | void blk_mq_drain_queue(struct request_queue *q) |
166 | { | |
167 | __blk_mq_drain_queue(q); | |
320ae51f JA |
168 | } |
169 | ||
170 | static void blk_mq_unfreeze_queue(struct request_queue *q) | |
171 | { | |
172 | bool wake = false; | |
173 | ||
174 | spin_lock_irq(q->queue_lock); | |
175 | if (!--q->bypass_depth) { | |
176 | queue_flag_clear(QUEUE_FLAG_BYPASS, q); | |
177 | wake = true; | |
178 | } | |
179 | WARN_ON_ONCE(q->bypass_depth < 0); | |
180 | spin_unlock_irq(q->queue_lock); | |
181 | if (wake) | |
182 | wake_up_all(&q->mq_freeze_wq); | |
183 | } | |
184 | ||
185 | bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx) | |
186 | { | |
187 | return blk_mq_has_free_tags(hctx->tags); | |
188 | } | |
189 | EXPORT_SYMBOL(blk_mq_can_queue); | |
190 | ||
94eddfbe JA |
191 | static void blk_mq_rq_ctx_init(struct request_queue *q, struct blk_mq_ctx *ctx, |
192 | struct request *rq, unsigned int rw_flags) | |
320ae51f | 193 | { |
94eddfbe JA |
194 | if (blk_queue_io_stat(q)) |
195 | rw_flags |= REQ_IO_STAT; | |
196 | ||
af76e555 CH |
197 | INIT_LIST_HEAD(&rq->queuelist); |
198 | /* csd/requeue_work/fifo_time is initialized before use */ | |
199 | rq->q = q; | |
320ae51f | 200 | rq->mq_ctx = ctx; |
0d2602ca | 201 | rq->cmd_flags |= rw_flags; |
af76e555 CH |
202 | /* do not touch atomic flags, it needs atomic ops against the timer */ |
203 | rq->cpu = -1; | |
af76e555 CH |
204 | INIT_HLIST_NODE(&rq->hash); |
205 | RB_CLEAR_NODE(&rq->rb_node); | |
af76e555 CH |
206 | rq->rq_disk = NULL; |
207 | rq->part = NULL; | |
af76e555 CH |
208 | #ifdef CONFIG_BLK_CGROUP |
209 | rq->rl = NULL; | |
0fec08b4 | 210 | set_start_time_ns(rq); |
af76e555 CH |
211 | rq->io_start_time_ns = 0; |
212 | #endif | |
213 | rq->nr_phys_segments = 0; | |
214 | #if defined(CONFIG_BLK_DEV_INTEGRITY) | |
215 | rq->nr_integrity_segments = 0; | |
216 | #endif | |
af76e555 CH |
217 | rq->special = NULL; |
218 | /* tag was already set */ | |
219 | rq->errors = 0; | |
af76e555 CH |
220 | |
221 | rq->extra_len = 0; | |
222 | rq->sense_len = 0; | |
223 | rq->resid_len = 0; | |
224 | rq->sense = NULL; | |
225 | ||
af76e555 | 226 | INIT_LIST_HEAD(&rq->timeout_list); |
af76e555 CH |
227 | rq->end_io = NULL; |
228 | rq->end_io_data = NULL; | |
229 | rq->next_rq = NULL; | |
230 | ||
320ae51f JA |
231 | ctx->rq_dispatched[rw_is_sync(rw_flags)]++; |
232 | } | |
233 | ||
5dee8577 CH |
234 | static struct request * |
235 | __blk_mq_alloc_request(struct request_queue *q, struct blk_mq_hw_ctx *hctx, | |
236 | struct blk_mq_ctx *ctx, int rw, gfp_t gfp, bool reserved) | |
237 | { | |
238 | struct request *rq; | |
239 | unsigned int tag; | |
240 | ||
241 | tag = blk_mq_get_tag(hctx, &ctx->last_tag, gfp, reserved); | |
242 | if (tag != BLK_MQ_TAG_FAIL) { | |
243 | rq = hctx->tags->rqs[tag]; | |
244 | ||
245 | rq->cmd_flags = 0; | |
246 | if (blk_mq_tag_busy(hctx)) { | |
247 | rq->cmd_flags = REQ_MQ_INFLIGHT; | |
248 | atomic_inc(&hctx->nr_active); | |
249 | } | |
250 | ||
251 | rq->tag = tag; | |
252 | blk_mq_rq_ctx_init(q, ctx, rq, rw); | |
253 | return rq; | |
254 | } | |
255 | ||
256 | return NULL; | |
257 | } | |
258 | ||
4ce01dd1 CH |
259 | struct request *blk_mq_alloc_request(struct request_queue *q, int rw, gfp_t gfp, |
260 | bool reserved) | |
320ae51f | 261 | { |
d852564f CH |
262 | struct blk_mq_ctx *ctx; |
263 | struct blk_mq_hw_ctx *hctx; | |
320ae51f JA |
264 | struct request *rq; |
265 | ||
266 | if (blk_mq_queue_enter(q)) | |
267 | return NULL; | |
268 | ||
d852564f CH |
269 | ctx = blk_mq_get_ctx(q); |
270 | hctx = q->mq_ops->map_queue(q, ctx->cpu); | |
271 | ||
272 | rq = __blk_mq_alloc_request(q, hctx, ctx, rw, gfp & ~__GFP_WAIT, | |
273 | reserved); | |
274 | if (!rq && (gfp & __GFP_WAIT)) { | |
275 | __blk_mq_run_hw_queue(hctx); | |
276 | blk_mq_put_ctx(ctx); | |
277 | ||
278 | ctx = blk_mq_get_ctx(q); | |
279 | hctx = q->mq_ops->map_queue(q, ctx->cpu); | |
280 | rq = __blk_mq_alloc_request(q, hctx, ctx, rw, gfp, reserved); | |
281 | } | |
282 | blk_mq_put_ctx(ctx); | |
320ae51f JA |
283 | return rq; |
284 | } | |
4bb659b1 | 285 | EXPORT_SYMBOL(blk_mq_alloc_request); |
320ae51f | 286 | |
320ae51f JA |
287 | static void __blk_mq_free_request(struct blk_mq_hw_ctx *hctx, |
288 | struct blk_mq_ctx *ctx, struct request *rq) | |
289 | { | |
290 | const int tag = rq->tag; | |
291 | struct request_queue *q = rq->q; | |
292 | ||
0d2602ca JA |
293 | if (rq->cmd_flags & REQ_MQ_INFLIGHT) |
294 | atomic_dec(&hctx->nr_active); | |
295 | ||
af76e555 | 296 | clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags); |
0d2602ca | 297 | blk_mq_put_tag(hctx, tag, &ctx->last_tag); |
320ae51f JA |
298 | blk_mq_queue_exit(q); |
299 | } | |
300 | ||
301 | void blk_mq_free_request(struct request *rq) | |
302 | { | |
303 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
304 | struct blk_mq_hw_ctx *hctx; | |
305 | struct request_queue *q = rq->q; | |
306 | ||
307 | ctx->rq_completed[rq_is_sync(rq)]++; | |
308 | ||
309 | hctx = q->mq_ops->map_queue(q, ctx->cpu); | |
310 | __blk_mq_free_request(hctx, ctx, rq); | |
311 | } | |
312 | ||
8727af4b CH |
313 | /* |
314 | * Clone all relevant state from a request that has been put on hold in | |
315 | * the flush state machine into the preallocated flush request that hangs | |
316 | * off the request queue. | |
317 | * | |
318 | * For a driver the flush request should be invisible, that's why we are | |
319 | * impersonating the original request here. | |
320 | */ | |
321 | void blk_mq_clone_flush_request(struct request *flush_rq, | |
322 | struct request *orig_rq) | |
323 | { | |
324 | struct blk_mq_hw_ctx *hctx = | |
325 | orig_rq->q->mq_ops->map_queue(orig_rq->q, orig_rq->mq_ctx->cpu); | |
326 | ||
327 | flush_rq->mq_ctx = orig_rq->mq_ctx; | |
328 | flush_rq->tag = orig_rq->tag; | |
329 | memcpy(blk_mq_rq_to_pdu(flush_rq), blk_mq_rq_to_pdu(orig_rq), | |
330 | hctx->cmd_size); | |
331 | } | |
332 | ||
63151a44 | 333 | inline void __blk_mq_end_io(struct request *rq, int error) |
320ae51f | 334 | { |
0d11e6ac ML |
335 | blk_account_io_done(rq); |
336 | ||
91b63639 | 337 | if (rq->end_io) { |
320ae51f | 338 | rq->end_io(rq, error); |
91b63639 CH |
339 | } else { |
340 | if (unlikely(blk_bidi_rq(rq))) | |
341 | blk_mq_free_request(rq->next_rq); | |
320ae51f | 342 | blk_mq_free_request(rq); |
91b63639 | 343 | } |
320ae51f | 344 | } |
63151a44 CH |
345 | EXPORT_SYMBOL(__blk_mq_end_io); |
346 | ||
347 | void blk_mq_end_io(struct request *rq, int error) | |
348 | { | |
349 | if (blk_update_request(rq, error, blk_rq_bytes(rq))) | |
350 | BUG(); | |
351 | __blk_mq_end_io(rq, error); | |
352 | } | |
353 | EXPORT_SYMBOL(blk_mq_end_io); | |
320ae51f | 354 | |
30a91cb4 | 355 | static void __blk_mq_complete_request_remote(void *data) |
320ae51f | 356 | { |
3d6efbf6 | 357 | struct request *rq = data; |
320ae51f | 358 | |
30a91cb4 | 359 | rq->q->softirq_done_fn(rq); |
320ae51f | 360 | } |
320ae51f | 361 | |
30a91cb4 | 362 | void __blk_mq_complete_request(struct request *rq) |
320ae51f JA |
363 | { |
364 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
38535201 | 365 | bool shared = false; |
320ae51f JA |
366 | int cpu; |
367 | ||
38535201 | 368 | if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) { |
30a91cb4 CH |
369 | rq->q->softirq_done_fn(rq); |
370 | return; | |
371 | } | |
320ae51f JA |
372 | |
373 | cpu = get_cpu(); | |
38535201 CH |
374 | if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags)) |
375 | shared = cpus_share_cache(cpu, ctx->cpu); | |
376 | ||
377 | if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) { | |
30a91cb4 | 378 | rq->csd.func = __blk_mq_complete_request_remote; |
3d6efbf6 CH |
379 | rq->csd.info = rq; |
380 | rq->csd.flags = 0; | |
c46fff2a | 381 | smp_call_function_single_async(ctx->cpu, &rq->csd); |
3d6efbf6 | 382 | } else { |
30a91cb4 | 383 | rq->q->softirq_done_fn(rq); |
3d6efbf6 | 384 | } |
320ae51f JA |
385 | put_cpu(); |
386 | } | |
30a91cb4 CH |
387 | |
388 | /** | |
389 | * blk_mq_complete_request - end I/O on a request | |
390 | * @rq: the request being processed | |
391 | * | |
392 | * Description: | |
393 | * Ends all I/O on a request. It does not handle partial completions. | |
394 | * The actual completion happens out-of-order, through a IPI handler. | |
395 | **/ | |
396 | void blk_mq_complete_request(struct request *rq) | |
397 | { | |
95f09684 JA |
398 | struct request_queue *q = rq->q; |
399 | ||
400 | if (unlikely(blk_should_fake_timeout(q))) | |
30a91cb4 | 401 | return; |
95f09684 JA |
402 | if (!blk_mark_rq_complete(rq)) { |
403 | if (q->softirq_done_fn) | |
404 | __blk_mq_complete_request(rq); | |
405 | else | |
406 | blk_mq_end_io(rq, rq->errors); | |
407 | } | |
30a91cb4 CH |
408 | } |
409 | EXPORT_SYMBOL(blk_mq_complete_request); | |
320ae51f | 410 | |
49f5baa5 | 411 | static void blk_mq_start_request(struct request *rq, bool last) |
320ae51f JA |
412 | { |
413 | struct request_queue *q = rq->q; | |
414 | ||
415 | trace_block_rq_issue(q, rq); | |
416 | ||
742ee69b | 417 | rq->resid_len = blk_rq_bytes(rq); |
91b63639 CH |
418 | if (unlikely(blk_bidi_rq(rq))) |
419 | rq->next_rq->resid_len = blk_rq_bytes(rq->next_rq); | |
742ee69b | 420 | |
320ae51f JA |
421 | /* |
422 | * Just mark start time and set the started bit. Due to memory | |
423 | * ordering, we know we'll see the correct deadline as long as | |
c22d9d8a JA |
424 | * REQ_ATOMIC_STARTED is seen. Use the default queue timeout, |
425 | * unless one has been set in the request. | |
320ae51f | 426 | */ |
c22d9d8a JA |
427 | if (!rq->timeout) |
428 | rq->deadline = jiffies + q->rq_timeout; | |
429 | else | |
430 | rq->deadline = jiffies + rq->timeout; | |
87ee7b11 JA |
431 | |
432 | /* | |
433 | * Mark us as started and clear complete. Complete might have been | |
434 | * set if requeue raced with timeout, which then marked it as | |
435 | * complete. So be sure to clear complete again when we start | |
436 | * the request, otherwise we'll ignore the completion event. | |
437 | */ | |
4b570521 JA |
438 | if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) |
439 | set_bit(REQ_ATOM_STARTED, &rq->atomic_flags); | |
440 | if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags)) | |
441 | clear_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags); | |
49f5baa5 CH |
442 | |
443 | if (q->dma_drain_size && blk_rq_bytes(rq)) { | |
444 | /* | |
445 | * Make sure space for the drain appears. We know we can do | |
446 | * this because max_hw_segments has been adjusted to be one | |
447 | * fewer than the device can handle. | |
448 | */ | |
449 | rq->nr_phys_segments++; | |
450 | } | |
451 | ||
452 | /* | |
453 | * Flag the last request in the series so that drivers know when IO | |
454 | * should be kicked off, if they don't do it on a per-request basis. | |
455 | * | |
456 | * Note: the flag isn't the only condition drivers should do kick off. | |
457 | * If drive is busy, the last request might not have the bit set. | |
458 | */ | |
459 | if (last) | |
460 | rq->cmd_flags |= REQ_END; | |
320ae51f JA |
461 | } |
462 | ||
ed0791b2 | 463 | static void __blk_mq_requeue_request(struct request *rq) |
320ae51f JA |
464 | { |
465 | struct request_queue *q = rq->q; | |
466 | ||
467 | trace_block_rq_requeue(q, rq); | |
468 | clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags); | |
49f5baa5 CH |
469 | |
470 | rq->cmd_flags &= ~REQ_END; | |
471 | ||
472 | if (q->dma_drain_size && blk_rq_bytes(rq)) | |
473 | rq->nr_phys_segments--; | |
320ae51f JA |
474 | } |
475 | ||
ed0791b2 CH |
476 | void blk_mq_requeue_request(struct request *rq) |
477 | { | |
ed0791b2 CH |
478 | __blk_mq_requeue_request(rq); |
479 | blk_clear_rq_complete(rq); | |
480 | ||
ed0791b2 | 481 | BUG_ON(blk_queued_rq(rq)); |
6fca6a61 | 482 | blk_mq_add_to_requeue_list(rq, true); |
ed0791b2 CH |
483 | } |
484 | EXPORT_SYMBOL(blk_mq_requeue_request); | |
485 | ||
6fca6a61 CH |
486 | static void blk_mq_requeue_work(struct work_struct *work) |
487 | { | |
488 | struct request_queue *q = | |
489 | container_of(work, struct request_queue, requeue_work); | |
490 | LIST_HEAD(rq_list); | |
491 | struct request *rq, *next; | |
492 | unsigned long flags; | |
493 | ||
494 | spin_lock_irqsave(&q->requeue_lock, flags); | |
495 | list_splice_init(&q->requeue_list, &rq_list); | |
496 | spin_unlock_irqrestore(&q->requeue_lock, flags); | |
497 | ||
498 | list_for_each_entry_safe(rq, next, &rq_list, queuelist) { | |
499 | if (!(rq->cmd_flags & REQ_SOFTBARRIER)) | |
500 | continue; | |
501 | ||
502 | rq->cmd_flags &= ~REQ_SOFTBARRIER; | |
503 | list_del_init(&rq->queuelist); | |
504 | blk_mq_insert_request(rq, true, false, false); | |
505 | } | |
506 | ||
507 | while (!list_empty(&rq_list)) { | |
508 | rq = list_entry(rq_list.next, struct request, queuelist); | |
509 | list_del_init(&rq->queuelist); | |
510 | blk_mq_insert_request(rq, false, false, false); | |
511 | } | |
512 | ||
513 | blk_mq_run_queues(q, false); | |
514 | } | |
515 | ||
516 | void blk_mq_add_to_requeue_list(struct request *rq, bool at_head) | |
517 | { | |
518 | struct request_queue *q = rq->q; | |
519 | unsigned long flags; | |
520 | ||
521 | /* | |
522 | * We abuse this flag that is otherwise used by the I/O scheduler to | |
523 | * request head insertation from the workqueue. | |
524 | */ | |
525 | BUG_ON(rq->cmd_flags & REQ_SOFTBARRIER); | |
526 | ||
527 | spin_lock_irqsave(&q->requeue_lock, flags); | |
528 | if (at_head) { | |
529 | rq->cmd_flags |= REQ_SOFTBARRIER; | |
530 | list_add(&rq->queuelist, &q->requeue_list); | |
531 | } else { | |
532 | list_add_tail(&rq->queuelist, &q->requeue_list); | |
533 | } | |
534 | spin_unlock_irqrestore(&q->requeue_lock, flags); | |
535 | } | |
536 | EXPORT_SYMBOL(blk_mq_add_to_requeue_list); | |
537 | ||
538 | void blk_mq_kick_requeue_list(struct request_queue *q) | |
539 | { | |
540 | kblockd_schedule_work(&q->requeue_work); | |
541 | } | |
542 | EXPORT_SYMBOL(blk_mq_kick_requeue_list); | |
543 | ||
22302375 | 544 | struct request *blk_mq_tag_to_rq(struct blk_mq_hw_ctx *hctx, unsigned int tag) |
24d2f903 | 545 | { |
22302375 SL |
546 | struct request_queue *q = hctx->queue; |
547 | ||
548 | if ((q->flush_rq->cmd_flags & REQ_FLUSH_SEQ) && | |
549 | q->flush_rq->tag == tag) | |
550 | return q->flush_rq; | |
551 | ||
552 | return hctx->tags->rqs[tag]; | |
24d2f903 CH |
553 | } |
554 | EXPORT_SYMBOL(blk_mq_tag_to_rq); | |
555 | ||
320ae51f JA |
556 | struct blk_mq_timeout_data { |
557 | struct blk_mq_hw_ctx *hctx; | |
558 | unsigned long *next; | |
559 | unsigned int *next_set; | |
560 | }; | |
561 | ||
562 | static void blk_mq_timeout_check(void *__data, unsigned long *free_tags) | |
563 | { | |
564 | struct blk_mq_timeout_data *data = __data; | |
565 | struct blk_mq_hw_ctx *hctx = data->hctx; | |
566 | unsigned int tag; | |
567 | ||
568 | /* It may not be in flight yet (this is where | |
569 | * the REQ_ATOMIC_STARTED flag comes in). The requests are | |
570 | * statically allocated, so we know it's always safe to access the | |
571 | * memory associated with a bit offset into ->rqs[]. | |
572 | */ | |
573 | tag = 0; | |
574 | do { | |
575 | struct request *rq; | |
576 | ||
24d2f903 CH |
577 | tag = find_next_zero_bit(free_tags, hctx->tags->nr_tags, tag); |
578 | if (tag >= hctx->tags->nr_tags) | |
320ae51f JA |
579 | break; |
580 | ||
22302375 | 581 | rq = blk_mq_tag_to_rq(hctx, tag++); |
24d2f903 CH |
582 | if (rq->q != hctx->queue) |
583 | continue; | |
320ae51f JA |
584 | if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) |
585 | continue; | |
586 | ||
587 | blk_rq_check_expired(rq, data->next, data->next_set); | |
588 | } while (1); | |
589 | } | |
590 | ||
591 | static void blk_mq_hw_ctx_check_timeout(struct blk_mq_hw_ctx *hctx, | |
592 | unsigned long *next, | |
593 | unsigned int *next_set) | |
594 | { | |
595 | struct blk_mq_timeout_data data = { | |
596 | .hctx = hctx, | |
597 | .next = next, | |
598 | .next_set = next_set, | |
599 | }; | |
600 | ||
601 | /* | |
602 | * Ask the tagging code to iterate busy requests, so we can | |
603 | * check them for timeout. | |
604 | */ | |
605 | blk_mq_tag_busy_iter(hctx->tags, blk_mq_timeout_check, &data); | |
606 | } | |
607 | ||
87ee7b11 JA |
608 | static enum blk_eh_timer_return blk_mq_rq_timed_out(struct request *rq) |
609 | { | |
610 | struct request_queue *q = rq->q; | |
611 | ||
612 | /* | |
613 | * We know that complete is set at this point. If STARTED isn't set | |
614 | * anymore, then the request isn't active and the "timeout" should | |
615 | * just be ignored. This can happen due to the bitflag ordering. | |
616 | * Timeout first checks if STARTED is set, and if it is, assumes | |
617 | * the request is active. But if we race with completion, then | |
618 | * we both flags will get cleared. So check here again, and ignore | |
619 | * a timeout event with a request that isn't active. | |
620 | */ | |
621 | if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) | |
622 | return BLK_EH_NOT_HANDLED; | |
623 | ||
624 | if (!q->mq_ops->timeout) | |
625 | return BLK_EH_RESET_TIMER; | |
626 | ||
627 | return q->mq_ops->timeout(rq); | |
628 | } | |
629 | ||
320ae51f JA |
630 | static void blk_mq_rq_timer(unsigned long data) |
631 | { | |
632 | struct request_queue *q = (struct request_queue *) data; | |
633 | struct blk_mq_hw_ctx *hctx; | |
634 | unsigned long next = 0; | |
635 | int i, next_set = 0; | |
636 | ||
484b4061 JA |
637 | queue_for_each_hw_ctx(q, hctx, i) { |
638 | /* | |
639 | * If not software queues are currently mapped to this | |
640 | * hardware queue, there's nothing to check | |
641 | */ | |
642 | if (!hctx->nr_ctx || !hctx->tags) | |
643 | continue; | |
644 | ||
320ae51f | 645 | blk_mq_hw_ctx_check_timeout(hctx, &next, &next_set); |
484b4061 | 646 | } |
320ae51f | 647 | |
0d2602ca JA |
648 | if (next_set) { |
649 | next = blk_rq_timeout(round_jiffies_up(next)); | |
650 | mod_timer(&q->timeout, next); | |
651 | } else { | |
652 | queue_for_each_hw_ctx(q, hctx, i) | |
653 | blk_mq_tag_idle(hctx); | |
654 | } | |
320ae51f JA |
655 | } |
656 | ||
657 | /* | |
658 | * Reverse check our software queue for entries that we could potentially | |
659 | * merge with. Currently includes a hand-wavy stop count of 8, to not spend | |
660 | * too much time checking for merges. | |
661 | */ | |
662 | static bool blk_mq_attempt_merge(struct request_queue *q, | |
663 | struct blk_mq_ctx *ctx, struct bio *bio) | |
664 | { | |
665 | struct request *rq; | |
666 | int checked = 8; | |
667 | ||
668 | list_for_each_entry_reverse(rq, &ctx->rq_list, queuelist) { | |
669 | int el_ret; | |
670 | ||
671 | if (!checked--) | |
672 | break; | |
673 | ||
674 | if (!blk_rq_merge_ok(rq, bio)) | |
675 | continue; | |
676 | ||
677 | el_ret = blk_try_merge(rq, bio); | |
678 | if (el_ret == ELEVATOR_BACK_MERGE) { | |
679 | if (bio_attempt_back_merge(q, rq, bio)) { | |
680 | ctx->rq_merged++; | |
681 | return true; | |
682 | } | |
683 | break; | |
684 | } else if (el_ret == ELEVATOR_FRONT_MERGE) { | |
685 | if (bio_attempt_front_merge(q, rq, bio)) { | |
686 | ctx->rq_merged++; | |
687 | return true; | |
688 | } | |
689 | break; | |
690 | } | |
691 | } | |
692 | ||
693 | return false; | |
694 | } | |
695 | ||
1429d7c9 JA |
696 | /* |
697 | * Process software queues that have been marked busy, splicing them | |
698 | * to the for-dispatch | |
699 | */ | |
700 | static void flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list) | |
701 | { | |
702 | struct blk_mq_ctx *ctx; | |
703 | int i; | |
704 | ||
705 | for (i = 0; i < hctx->ctx_map.map_size; i++) { | |
706 | struct blk_align_bitmap *bm = &hctx->ctx_map.map[i]; | |
707 | unsigned int off, bit; | |
708 | ||
709 | if (!bm->word) | |
710 | continue; | |
711 | ||
712 | bit = 0; | |
713 | off = i * hctx->ctx_map.bits_per_word; | |
714 | do { | |
715 | bit = find_next_bit(&bm->word, bm->depth, bit); | |
716 | if (bit >= bm->depth) | |
717 | break; | |
718 | ||
719 | ctx = hctx->ctxs[bit + off]; | |
720 | clear_bit(bit, &bm->word); | |
721 | spin_lock(&ctx->lock); | |
722 | list_splice_tail_init(&ctx->rq_list, list); | |
723 | spin_unlock(&ctx->lock); | |
724 | ||
725 | bit++; | |
726 | } while (1); | |
727 | } | |
728 | } | |
729 | ||
320ae51f JA |
730 | /* |
731 | * Run this hardware queue, pulling any software queues mapped to it in. | |
732 | * Note that this function currently has various problems around ordering | |
733 | * of IO. In particular, we'd like FIFO behaviour on handling existing | |
734 | * items on the hctx->dispatch list. Ignore that for now. | |
735 | */ | |
736 | static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx) | |
737 | { | |
738 | struct request_queue *q = hctx->queue; | |
320ae51f JA |
739 | struct request *rq; |
740 | LIST_HEAD(rq_list); | |
1429d7c9 | 741 | int queued; |
320ae51f | 742 | |
fd1270d5 | 743 | WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask)); |
e4043dcf | 744 | |
5d12f905 | 745 | if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state))) |
320ae51f JA |
746 | return; |
747 | ||
748 | hctx->run++; | |
749 | ||
750 | /* | |
751 | * Touch any software queue that has pending entries. | |
752 | */ | |
1429d7c9 | 753 | flush_busy_ctxs(hctx, &rq_list); |
320ae51f JA |
754 | |
755 | /* | |
756 | * If we have previous entries on our dispatch list, grab them | |
757 | * and stuff them at the front for more fair dispatch. | |
758 | */ | |
759 | if (!list_empty_careful(&hctx->dispatch)) { | |
760 | spin_lock(&hctx->lock); | |
761 | if (!list_empty(&hctx->dispatch)) | |
762 | list_splice_init(&hctx->dispatch, &rq_list); | |
763 | spin_unlock(&hctx->lock); | |
764 | } | |
765 | ||
320ae51f JA |
766 | /* |
767 | * Now process all the entries, sending them to the driver. | |
768 | */ | |
1429d7c9 | 769 | queued = 0; |
320ae51f JA |
770 | while (!list_empty(&rq_list)) { |
771 | int ret; | |
772 | ||
773 | rq = list_first_entry(&rq_list, struct request, queuelist); | |
774 | list_del_init(&rq->queuelist); | |
320ae51f | 775 | |
49f5baa5 | 776 | blk_mq_start_request(rq, list_empty(&rq_list)); |
320ae51f JA |
777 | |
778 | ret = q->mq_ops->queue_rq(hctx, rq); | |
779 | switch (ret) { | |
780 | case BLK_MQ_RQ_QUEUE_OK: | |
781 | queued++; | |
782 | continue; | |
783 | case BLK_MQ_RQ_QUEUE_BUSY: | |
320ae51f | 784 | list_add(&rq->queuelist, &rq_list); |
ed0791b2 | 785 | __blk_mq_requeue_request(rq); |
320ae51f JA |
786 | break; |
787 | default: | |
788 | pr_err("blk-mq: bad return on queue: %d\n", ret); | |
320ae51f | 789 | case BLK_MQ_RQ_QUEUE_ERROR: |
1e93b8c2 | 790 | rq->errors = -EIO; |
320ae51f JA |
791 | blk_mq_end_io(rq, rq->errors); |
792 | break; | |
793 | } | |
794 | ||
795 | if (ret == BLK_MQ_RQ_QUEUE_BUSY) | |
796 | break; | |
797 | } | |
798 | ||
799 | if (!queued) | |
800 | hctx->dispatched[0]++; | |
801 | else if (queued < (1 << (BLK_MQ_MAX_DISPATCH_ORDER - 1))) | |
802 | hctx->dispatched[ilog2(queued) + 1]++; | |
803 | ||
804 | /* | |
805 | * Any items that need requeuing? Stuff them into hctx->dispatch, | |
806 | * that is where we will continue on next queue run. | |
807 | */ | |
808 | if (!list_empty(&rq_list)) { | |
809 | spin_lock(&hctx->lock); | |
810 | list_splice(&rq_list, &hctx->dispatch); | |
811 | spin_unlock(&hctx->lock); | |
812 | } | |
813 | } | |
814 | ||
506e931f JA |
815 | /* |
816 | * It'd be great if the workqueue API had a way to pass | |
817 | * in a mask and had some smarts for more clever placement. | |
818 | * For now we just round-robin here, switching for every | |
819 | * BLK_MQ_CPU_WORK_BATCH queued items. | |
820 | */ | |
821 | static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx) | |
822 | { | |
823 | int cpu = hctx->next_cpu; | |
824 | ||
825 | if (--hctx->next_cpu_batch <= 0) { | |
826 | int next_cpu; | |
827 | ||
828 | next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask); | |
829 | if (next_cpu >= nr_cpu_ids) | |
830 | next_cpu = cpumask_first(hctx->cpumask); | |
831 | ||
832 | hctx->next_cpu = next_cpu; | |
833 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; | |
834 | } | |
835 | ||
836 | return cpu; | |
837 | } | |
838 | ||
320ae51f JA |
839 | void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) |
840 | { | |
5d12f905 | 841 | if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state))) |
320ae51f JA |
842 | return; |
843 | ||
e4043dcf | 844 | if (!async && cpumask_test_cpu(smp_processor_id(), hctx->cpumask)) |
320ae51f | 845 | __blk_mq_run_hw_queue(hctx); |
e4043dcf | 846 | else if (hctx->queue->nr_hw_queues == 1) |
70f4db63 | 847 | kblockd_schedule_delayed_work(&hctx->run_work, 0); |
e4043dcf JA |
848 | else { |
849 | unsigned int cpu; | |
850 | ||
506e931f | 851 | cpu = blk_mq_hctx_next_cpu(hctx); |
70f4db63 | 852 | kblockd_schedule_delayed_work_on(cpu, &hctx->run_work, 0); |
e4043dcf | 853 | } |
320ae51f JA |
854 | } |
855 | ||
856 | void blk_mq_run_queues(struct request_queue *q, bool async) | |
857 | { | |
858 | struct blk_mq_hw_ctx *hctx; | |
859 | int i; | |
860 | ||
861 | queue_for_each_hw_ctx(q, hctx, i) { | |
862 | if ((!blk_mq_hctx_has_pending(hctx) && | |
863 | list_empty_careful(&hctx->dispatch)) || | |
5d12f905 | 864 | test_bit(BLK_MQ_S_STOPPED, &hctx->state)) |
320ae51f JA |
865 | continue; |
866 | ||
e4043dcf | 867 | preempt_disable(); |
320ae51f | 868 | blk_mq_run_hw_queue(hctx, async); |
e4043dcf | 869 | preempt_enable(); |
320ae51f JA |
870 | } |
871 | } | |
872 | EXPORT_SYMBOL(blk_mq_run_queues); | |
873 | ||
874 | void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx) | |
875 | { | |
70f4db63 CH |
876 | cancel_delayed_work(&hctx->run_work); |
877 | cancel_delayed_work(&hctx->delay_work); | |
320ae51f JA |
878 | set_bit(BLK_MQ_S_STOPPED, &hctx->state); |
879 | } | |
880 | EXPORT_SYMBOL(blk_mq_stop_hw_queue); | |
881 | ||
280d45f6 CH |
882 | void blk_mq_stop_hw_queues(struct request_queue *q) |
883 | { | |
884 | struct blk_mq_hw_ctx *hctx; | |
885 | int i; | |
886 | ||
887 | queue_for_each_hw_ctx(q, hctx, i) | |
888 | blk_mq_stop_hw_queue(hctx); | |
889 | } | |
890 | EXPORT_SYMBOL(blk_mq_stop_hw_queues); | |
891 | ||
320ae51f JA |
892 | void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx) |
893 | { | |
894 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
e4043dcf JA |
895 | |
896 | preempt_disable(); | |
320ae51f | 897 | __blk_mq_run_hw_queue(hctx); |
e4043dcf | 898 | preempt_enable(); |
320ae51f JA |
899 | } |
900 | EXPORT_SYMBOL(blk_mq_start_hw_queue); | |
901 | ||
2f268556 CH |
902 | void blk_mq_start_hw_queues(struct request_queue *q) |
903 | { | |
904 | struct blk_mq_hw_ctx *hctx; | |
905 | int i; | |
906 | ||
907 | queue_for_each_hw_ctx(q, hctx, i) | |
908 | blk_mq_start_hw_queue(hctx); | |
909 | } | |
910 | EXPORT_SYMBOL(blk_mq_start_hw_queues); | |
911 | ||
912 | ||
1b4a3258 | 913 | void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async) |
320ae51f JA |
914 | { |
915 | struct blk_mq_hw_ctx *hctx; | |
916 | int i; | |
917 | ||
918 | queue_for_each_hw_ctx(q, hctx, i) { | |
919 | if (!test_bit(BLK_MQ_S_STOPPED, &hctx->state)) | |
920 | continue; | |
921 | ||
922 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
e4043dcf | 923 | preempt_disable(); |
1b4a3258 | 924 | blk_mq_run_hw_queue(hctx, async); |
e4043dcf | 925 | preempt_enable(); |
320ae51f JA |
926 | } |
927 | } | |
928 | EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues); | |
929 | ||
70f4db63 | 930 | static void blk_mq_run_work_fn(struct work_struct *work) |
320ae51f JA |
931 | { |
932 | struct blk_mq_hw_ctx *hctx; | |
933 | ||
70f4db63 | 934 | hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work); |
e4043dcf | 935 | |
320ae51f JA |
936 | __blk_mq_run_hw_queue(hctx); |
937 | } | |
938 | ||
70f4db63 CH |
939 | static void blk_mq_delay_work_fn(struct work_struct *work) |
940 | { | |
941 | struct blk_mq_hw_ctx *hctx; | |
942 | ||
943 | hctx = container_of(work, struct blk_mq_hw_ctx, delay_work.work); | |
944 | ||
945 | if (test_and_clear_bit(BLK_MQ_S_STOPPED, &hctx->state)) | |
946 | __blk_mq_run_hw_queue(hctx); | |
947 | } | |
948 | ||
949 | void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs) | |
950 | { | |
951 | unsigned long tmo = msecs_to_jiffies(msecs); | |
952 | ||
953 | if (hctx->queue->nr_hw_queues == 1) | |
954 | kblockd_schedule_delayed_work(&hctx->delay_work, tmo); | |
955 | else { | |
956 | unsigned int cpu; | |
957 | ||
506e931f | 958 | cpu = blk_mq_hctx_next_cpu(hctx); |
70f4db63 CH |
959 | kblockd_schedule_delayed_work_on(cpu, &hctx->delay_work, tmo); |
960 | } | |
961 | } | |
962 | EXPORT_SYMBOL(blk_mq_delay_queue); | |
963 | ||
320ae51f | 964 | static void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, |
72a0a36e | 965 | struct request *rq, bool at_head) |
320ae51f JA |
966 | { |
967 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
968 | ||
01b983c9 JA |
969 | trace_block_rq_insert(hctx->queue, rq); |
970 | ||
72a0a36e CH |
971 | if (at_head) |
972 | list_add(&rq->queuelist, &ctx->rq_list); | |
973 | else | |
974 | list_add_tail(&rq->queuelist, &ctx->rq_list); | |
4bb659b1 | 975 | |
320ae51f JA |
976 | blk_mq_hctx_mark_pending(hctx, ctx); |
977 | ||
978 | /* | |
979 | * We do this early, to ensure we are on the right CPU. | |
980 | */ | |
87ee7b11 | 981 | blk_add_timer(rq); |
320ae51f JA |
982 | } |
983 | ||
eeabc850 CH |
984 | void blk_mq_insert_request(struct request *rq, bool at_head, bool run_queue, |
985 | bool async) | |
320ae51f | 986 | { |
eeabc850 | 987 | struct request_queue *q = rq->q; |
320ae51f | 988 | struct blk_mq_hw_ctx *hctx; |
eeabc850 CH |
989 | struct blk_mq_ctx *ctx = rq->mq_ctx, *current_ctx; |
990 | ||
991 | current_ctx = blk_mq_get_ctx(q); | |
992 | if (!cpu_online(ctx->cpu)) | |
993 | rq->mq_ctx = ctx = current_ctx; | |
320ae51f | 994 | |
320ae51f JA |
995 | hctx = q->mq_ops->map_queue(q, ctx->cpu); |
996 | ||
eeabc850 CH |
997 | if (rq->cmd_flags & (REQ_FLUSH | REQ_FUA) && |
998 | !(rq->cmd_flags & (REQ_FLUSH_SEQ))) { | |
320ae51f JA |
999 | blk_insert_flush(rq); |
1000 | } else { | |
320ae51f | 1001 | spin_lock(&ctx->lock); |
72a0a36e | 1002 | __blk_mq_insert_request(hctx, rq, at_head); |
320ae51f | 1003 | spin_unlock(&ctx->lock); |
320ae51f JA |
1004 | } |
1005 | ||
320ae51f JA |
1006 | if (run_queue) |
1007 | blk_mq_run_hw_queue(hctx, async); | |
e4043dcf JA |
1008 | |
1009 | blk_mq_put_ctx(current_ctx); | |
320ae51f JA |
1010 | } |
1011 | ||
1012 | static void blk_mq_insert_requests(struct request_queue *q, | |
1013 | struct blk_mq_ctx *ctx, | |
1014 | struct list_head *list, | |
1015 | int depth, | |
1016 | bool from_schedule) | |
1017 | ||
1018 | { | |
1019 | struct blk_mq_hw_ctx *hctx; | |
1020 | struct blk_mq_ctx *current_ctx; | |
1021 | ||
1022 | trace_block_unplug(q, depth, !from_schedule); | |
1023 | ||
1024 | current_ctx = blk_mq_get_ctx(q); | |
1025 | ||
1026 | if (!cpu_online(ctx->cpu)) | |
1027 | ctx = current_ctx; | |
1028 | hctx = q->mq_ops->map_queue(q, ctx->cpu); | |
1029 | ||
1030 | /* | |
1031 | * preemption doesn't flush plug list, so it's possible ctx->cpu is | |
1032 | * offline now | |
1033 | */ | |
1034 | spin_lock(&ctx->lock); | |
1035 | while (!list_empty(list)) { | |
1036 | struct request *rq; | |
1037 | ||
1038 | rq = list_first_entry(list, struct request, queuelist); | |
1039 | list_del_init(&rq->queuelist); | |
1040 | rq->mq_ctx = ctx; | |
72a0a36e | 1041 | __blk_mq_insert_request(hctx, rq, false); |
320ae51f JA |
1042 | } |
1043 | spin_unlock(&ctx->lock); | |
1044 | ||
320ae51f | 1045 | blk_mq_run_hw_queue(hctx, from_schedule); |
e4043dcf | 1046 | blk_mq_put_ctx(current_ctx); |
320ae51f JA |
1047 | } |
1048 | ||
1049 | static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b) | |
1050 | { | |
1051 | struct request *rqa = container_of(a, struct request, queuelist); | |
1052 | struct request *rqb = container_of(b, struct request, queuelist); | |
1053 | ||
1054 | return !(rqa->mq_ctx < rqb->mq_ctx || | |
1055 | (rqa->mq_ctx == rqb->mq_ctx && | |
1056 | blk_rq_pos(rqa) < blk_rq_pos(rqb))); | |
1057 | } | |
1058 | ||
1059 | void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule) | |
1060 | { | |
1061 | struct blk_mq_ctx *this_ctx; | |
1062 | struct request_queue *this_q; | |
1063 | struct request *rq; | |
1064 | LIST_HEAD(list); | |
1065 | LIST_HEAD(ctx_list); | |
1066 | unsigned int depth; | |
1067 | ||
1068 | list_splice_init(&plug->mq_list, &list); | |
1069 | ||
1070 | list_sort(NULL, &list, plug_ctx_cmp); | |
1071 | ||
1072 | this_q = NULL; | |
1073 | this_ctx = NULL; | |
1074 | depth = 0; | |
1075 | ||
1076 | while (!list_empty(&list)) { | |
1077 | rq = list_entry_rq(list.next); | |
1078 | list_del_init(&rq->queuelist); | |
1079 | BUG_ON(!rq->q); | |
1080 | if (rq->mq_ctx != this_ctx) { | |
1081 | if (this_ctx) { | |
1082 | blk_mq_insert_requests(this_q, this_ctx, | |
1083 | &ctx_list, depth, | |
1084 | from_schedule); | |
1085 | } | |
1086 | ||
1087 | this_ctx = rq->mq_ctx; | |
1088 | this_q = rq->q; | |
1089 | depth = 0; | |
1090 | } | |
1091 | ||
1092 | depth++; | |
1093 | list_add_tail(&rq->queuelist, &ctx_list); | |
1094 | } | |
1095 | ||
1096 | /* | |
1097 | * If 'this_ctx' is set, we know we have entries to complete | |
1098 | * on 'ctx_list'. Do those. | |
1099 | */ | |
1100 | if (this_ctx) { | |
1101 | blk_mq_insert_requests(this_q, this_ctx, &ctx_list, depth, | |
1102 | from_schedule); | |
1103 | } | |
1104 | } | |
1105 | ||
1106 | static void blk_mq_bio_to_request(struct request *rq, struct bio *bio) | |
1107 | { | |
1108 | init_request_from_bio(rq, bio); | |
4b570521 JA |
1109 | |
1110 | if (blk_do_io_stat(rq)) { | |
1111 | rq->start_time = jiffies; | |
1112 | blk_account_io_start(rq, 1); | |
1113 | } | |
320ae51f JA |
1114 | } |
1115 | ||
07068d5b JA |
1116 | static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx *hctx, |
1117 | struct blk_mq_ctx *ctx, | |
1118 | struct request *rq, struct bio *bio) | |
320ae51f | 1119 | { |
07068d5b | 1120 | struct request_queue *q = hctx->queue; |
320ae51f | 1121 | |
07068d5b JA |
1122 | if (!(hctx->flags & BLK_MQ_F_SHOULD_MERGE)) { |
1123 | blk_mq_bio_to_request(rq, bio); | |
1124 | spin_lock(&ctx->lock); | |
1125 | insert_rq: | |
1126 | __blk_mq_insert_request(hctx, rq, false); | |
1127 | spin_unlock(&ctx->lock); | |
1128 | return false; | |
1129 | } else { | |
1130 | spin_lock(&ctx->lock); | |
1131 | if (!blk_mq_attempt_merge(q, ctx, bio)) { | |
1132 | blk_mq_bio_to_request(rq, bio); | |
1133 | goto insert_rq; | |
1134 | } | |
320ae51f | 1135 | |
07068d5b JA |
1136 | spin_unlock(&ctx->lock); |
1137 | __blk_mq_free_request(hctx, ctx, rq); | |
1138 | return true; | |
14ec77f3 | 1139 | } |
07068d5b | 1140 | } |
14ec77f3 | 1141 | |
07068d5b JA |
1142 | struct blk_map_ctx { |
1143 | struct blk_mq_hw_ctx *hctx; | |
1144 | struct blk_mq_ctx *ctx; | |
1145 | }; | |
1146 | ||
1147 | static struct request *blk_mq_map_request(struct request_queue *q, | |
1148 | struct bio *bio, | |
1149 | struct blk_map_ctx *data) | |
1150 | { | |
1151 | struct blk_mq_hw_ctx *hctx; | |
1152 | struct blk_mq_ctx *ctx; | |
1153 | struct request *rq; | |
1154 | int rw = bio_data_dir(bio); | |
320ae51f | 1155 | |
07068d5b | 1156 | if (unlikely(blk_mq_queue_enter(q))) { |
320ae51f | 1157 | bio_endio(bio, -EIO); |
07068d5b | 1158 | return NULL; |
320ae51f JA |
1159 | } |
1160 | ||
1161 | ctx = blk_mq_get_ctx(q); | |
1162 | hctx = q->mq_ops->map_queue(q, ctx->cpu); | |
1163 | ||
07068d5b | 1164 | if (rw_is_sync(bio->bi_rw)) |
27fbf4e8 | 1165 | rw |= REQ_SYNC; |
07068d5b | 1166 | |
320ae51f | 1167 | trace_block_getrq(q, bio, rw); |
5dee8577 CH |
1168 | rq = __blk_mq_alloc_request(q, hctx, ctx, rw, GFP_ATOMIC, false); |
1169 | if (unlikely(!rq)) { | |
793597a6 | 1170 | __blk_mq_run_hw_queue(hctx); |
320ae51f JA |
1171 | blk_mq_put_ctx(ctx); |
1172 | trace_block_sleeprq(q, bio, rw); | |
793597a6 CH |
1173 | |
1174 | ctx = blk_mq_get_ctx(q); | |
320ae51f | 1175 | hctx = q->mq_ops->map_queue(q, ctx->cpu); |
793597a6 CH |
1176 | rq = __blk_mq_alloc_request(q, hctx, ctx, rw, |
1177 | __GFP_WAIT|GFP_ATOMIC, false); | |
320ae51f JA |
1178 | } |
1179 | ||
1180 | hctx->queued++; | |
07068d5b JA |
1181 | data->hctx = hctx; |
1182 | data->ctx = ctx; | |
1183 | return rq; | |
1184 | } | |
1185 | ||
1186 | /* | |
1187 | * Multiple hardware queue variant. This will not use per-process plugs, | |
1188 | * but will attempt to bypass the hctx queueing if we can go straight to | |
1189 | * hardware for SYNC IO. | |
1190 | */ | |
1191 | static void blk_mq_make_request(struct request_queue *q, struct bio *bio) | |
1192 | { | |
1193 | const int is_sync = rw_is_sync(bio->bi_rw); | |
1194 | const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA); | |
1195 | struct blk_map_ctx data; | |
1196 | struct request *rq; | |
1197 | ||
1198 | blk_queue_bounce(q, &bio); | |
1199 | ||
1200 | if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) { | |
1201 | bio_endio(bio, -EIO); | |
1202 | return; | |
1203 | } | |
1204 | ||
1205 | rq = blk_mq_map_request(q, bio, &data); | |
1206 | if (unlikely(!rq)) | |
1207 | return; | |
1208 | ||
1209 | if (unlikely(is_flush_fua)) { | |
1210 | blk_mq_bio_to_request(rq, bio); | |
1211 | blk_insert_flush(rq); | |
1212 | goto run_queue; | |
1213 | } | |
1214 | ||
1215 | if (is_sync) { | |
1216 | int ret; | |
1217 | ||
1218 | blk_mq_bio_to_request(rq, bio); | |
1219 | blk_mq_start_request(rq, true); | |
1220 | ||
1221 | /* | |
1222 | * For OK queue, we are done. For error, kill it. Any other | |
1223 | * error (busy), just add it to our list as we previously | |
1224 | * would have done | |
1225 | */ | |
1226 | ret = q->mq_ops->queue_rq(data.hctx, rq); | |
1227 | if (ret == BLK_MQ_RQ_QUEUE_OK) | |
1228 | goto done; | |
1229 | else { | |
1230 | __blk_mq_requeue_request(rq); | |
1231 | ||
1232 | if (ret == BLK_MQ_RQ_QUEUE_ERROR) { | |
1233 | rq->errors = -EIO; | |
1234 | blk_mq_end_io(rq, rq->errors); | |
1235 | goto done; | |
1236 | } | |
1237 | } | |
1238 | } | |
1239 | ||
1240 | if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) { | |
1241 | /* | |
1242 | * For a SYNC request, send it to the hardware immediately. For | |
1243 | * an ASYNC request, just ensure that we run it later on. The | |
1244 | * latter allows for merging opportunities and more efficient | |
1245 | * dispatching. | |
1246 | */ | |
1247 | run_queue: | |
1248 | blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua); | |
1249 | } | |
1250 | done: | |
1251 | blk_mq_put_ctx(data.ctx); | |
1252 | } | |
1253 | ||
1254 | /* | |
1255 | * Single hardware queue variant. This will attempt to use any per-process | |
1256 | * plug for merging and IO deferral. | |
1257 | */ | |
1258 | static void blk_sq_make_request(struct request_queue *q, struct bio *bio) | |
1259 | { | |
1260 | const int is_sync = rw_is_sync(bio->bi_rw); | |
1261 | const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA); | |
1262 | unsigned int use_plug, request_count = 0; | |
1263 | struct blk_map_ctx data; | |
1264 | struct request *rq; | |
1265 | ||
1266 | /* | |
1267 | * If we have multiple hardware queues, just go directly to | |
1268 | * one of those for sync IO. | |
1269 | */ | |
1270 | use_plug = !is_flush_fua && !is_sync; | |
1271 | ||
1272 | blk_queue_bounce(q, &bio); | |
1273 | ||
1274 | if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) { | |
1275 | bio_endio(bio, -EIO); | |
1276 | return; | |
1277 | } | |
1278 | ||
1279 | if (use_plug && !blk_queue_nomerges(q) && | |
1280 | blk_attempt_plug_merge(q, bio, &request_count)) | |
1281 | return; | |
1282 | ||
1283 | rq = blk_mq_map_request(q, bio, &data); | |
320ae51f JA |
1284 | |
1285 | if (unlikely(is_flush_fua)) { | |
1286 | blk_mq_bio_to_request(rq, bio); | |
320ae51f JA |
1287 | blk_insert_flush(rq); |
1288 | goto run_queue; | |
1289 | } | |
1290 | ||
1291 | /* | |
1292 | * A task plug currently exists. Since this is completely lockless, | |
1293 | * utilize that to temporarily store requests until the task is | |
1294 | * either done or scheduled away. | |
1295 | */ | |
1296 | if (use_plug) { | |
1297 | struct blk_plug *plug = current->plug; | |
1298 | ||
1299 | if (plug) { | |
1300 | blk_mq_bio_to_request(rq, bio); | |
92f399c7 | 1301 | if (list_empty(&plug->mq_list)) |
320ae51f JA |
1302 | trace_block_plug(q); |
1303 | else if (request_count >= BLK_MAX_REQUEST_COUNT) { | |
1304 | blk_flush_plug_list(plug, false); | |
1305 | trace_block_plug(q); | |
1306 | } | |
1307 | list_add_tail(&rq->queuelist, &plug->mq_list); | |
07068d5b | 1308 | blk_mq_put_ctx(data.ctx); |
320ae51f JA |
1309 | return; |
1310 | } | |
1311 | } | |
1312 | ||
07068d5b JA |
1313 | if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) { |
1314 | /* | |
1315 | * For a SYNC request, send it to the hardware immediately. For | |
1316 | * an ASYNC request, just ensure that we run it later on. The | |
1317 | * latter allows for merging opportunities and more efficient | |
1318 | * dispatching. | |
1319 | */ | |
1320 | run_queue: | |
1321 | blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua); | |
320ae51f JA |
1322 | } |
1323 | ||
07068d5b | 1324 | blk_mq_put_ctx(data.ctx); |
320ae51f JA |
1325 | } |
1326 | ||
1327 | /* | |
1328 | * Default mapping to a software queue, since we use one per CPU. | |
1329 | */ | |
1330 | struct blk_mq_hw_ctx *blk_mq_map_queue(struct request_queue *q, const int cpu) | |
1331 | { | |
1332 | return q->queue_hw_ctx[q->mq_map[cpu]]; | |
1333 | } | |
1334 | EXPORT_SYMBOL(blk_mq_map_queue); | |
1335 | ||
24d2f903 CH |
1336 | static void blk_mq_free_rq_map(struct blk_mq_tag_set *set, |
1337 | struct blk_mq_tags *tags, unsigned int hctx_idx) | |
95363efd | 1338 | { |
e9b267d9 | 1339 | struct page *page; |
320ae51f | 1340 | |
24d2f903 | 1341 | if (tags->rqs && set->ops->exit_request) { |
e9b267d9 | 1342 | int i; |
320ae51f | 1343 | |
24d2f903 CH |
1344 | for (i = 0; i < tags->nr_tags; i++) { |
1345 | if (!tags->rqs[i]) | |
e9b267d9 | 1346 | continue; |
24d2f903 CH |
1347 | set->ops->exit_request(set->driver_data, tags->rqs[i], |
1348 | hctx_idx, i); | |
e9b267d9 | 1349 | } |
320ae51f | 1350 | } |
320ae51f | 1351 | |
24d2f903 CH |
1352 | while (!list_empty(&tags->page_list)) { |
1353 | page = list_first_entry(&tags->page_list, struct page, lru); | |
6753471c | 1354 | list_del_init(&page->lru); |
320ae51f JA |
1355 | __free_pages(page, page->private); |
1356 | } | |
1357 | ||
24d2f903 | 1358 | kfree(tags->rqs); |
320ae51f | 1359 | |
24d2f903 | 1360 | blk_mq_free_tags(tags); |
320ae51f JA |
1361 | } |
1362 | ||
1363 | static size_t order_to_size(unsigned int order) | |
1364 | { | |
4ca08500 | 1365 | return (size_t)PAGE_SIZE << order; |
320ae51f JA |
1366 | } |
1367 | ||
24d2f903 CH |
1368 | static struct blk_mq_tags *blk_mq_init_rq_map(struct blk_mq_tag_set *set, |
1369 | unsigned int hctx_idx) | |
320ae51f | 1370 | { |
24d2f903 | 1371 | struct blk_mq_tags *tags; |
320ae51f JA |
1372 | unsigned int i, j, entries_per_page, max_order = 4; |
1373 | size_t rq_size, left; | |
1374 | ||
24d2f903 CH |
1375 | tags = blk_mq_init_tags(set->queue_depth, set->reserved_tags, |
1376 | set->numa_node); | |
1377 | if (!tags) | |
1378 | return NULL; | |
320ae51f | 1379 | |
24d2f903 CH |
1380 | INIT_LIST_HEAD(&tags->page_list); |
1381 | ||
1382 | tags->rqs = kmalloc_node(set->queue_depth * sizeof(struct request *), | |
1383 | GFP_KERNEL, set->numa_node); | |
1384 | if (!tags->rqs) { | |
1385 | blk_mq_free_tags(tags); | |
1386 | return NULL; | |
1387 | } | |
320ae51f JA |
1388 | |
1389 | /* | |
1390 | * rq_size is the size of the request plus driver payload, rounded | |
1391 | * to the cacheline size | |
1392 | */ | |
24d2f903 | 1393 | rq_size = round_up(sizeof(struct request) + set->cmd_size, |
320ae51f | 1394 | cache_line_size()); |
24d2f903 | 1395 | left = rq_size * set->queue_depth; |
320ae51f | 1396 | |
24d2f903 | 1397 | for (i = 0; i < set->queue_depth; ) { |
320ae51f JA |
1398 | int this_order = max_order; |
1399 | struct page *page; | |
1400 | int to_do; | |
1401 | void *p; | |
1402 | ||
1403 | while (left < order_to_size(this_order - 1) && this_order) | |
1404 | this_order--; | |
1405 | ||
1406 | do { | |
24d2f903 CH |
1407 | page = alloc_pages_node(set->numa_node, GFP_KERNEL, |
1408 | this_order); | |
320ae51f JA |
1409 | if (page) |
1410 | break; | |
1411 | if (!this_order--) | |
1412 | break; | |
1413 | if (order_to_size(this_order) < rq_size) | |
1414 | break; | |
1415 | } while (1); | |
1416 | ||
1417 | if (!page) | |
24d2f903 | 1418 | goto fail; |
320ae51f JA |
1419 | |
1420 | page->private = this_order; | |
24d2f903 | 1421 | list_add_tail(&page->lru, &tags->page_list); |
320ae51f JA |
1422 | |
1423 | p = page_address(page); | |
1424 | entries_per_page = order_to_size(this_order) / rq_size; | |
24d2f903 | 1425 | to_do = min(entries_per_page, set->queue_depth - i); |
320ae51f JA |
1426 | left -= to_do * rq_size; |
1427 | for (j = 0; j < to_do; j++) { | |
24d2f903 CH |
1428 | tags->rqs[i] = p; |
1429 | if (set->ops->init_request) { | |
1430 | if (set->ops->init_request(set->driver_data, | |
1431 | tags->rqs[i], hctx_idx, i, | |
1432 | set->numa_node)) | |
1433 | goto fail; | |
e9b267d9 CH |
1434 | } |
1435 | ||
320ae51f JA |
1436 | p += rq_size; |
1437 | i++; | |
1438 | } | |
1439 | } | |
1440 | ||
24d2f903 | 1441 | return tags; |
320ae51f | 1442 | |
24d2f903 CH |
1443 | fail: |
1444 | pr_warn("%s: failed to allocate requests\n", __func__); | |
1445 | blk_mq_free_rq_map(set, tags, hctx_idx); | |
1446 | return NULL; | |
320ae51f JA |
1447 | } |
1448 | ||
1429d7c9 JA |
1449 | static void blk_mq_free_bitmap(struct blk_mq_ctxmap *bitmap) |
1450 | { | |
1451 | kfree(bitmap->map); | |
1452 | } | |
1453 | ||
1454 | static int blk_mq_alloc_bitmap(struct blk_mq_ctxmap *bitmap, int node) | |
1455 | { | |
1456 | unsigned int bpw = 8, total, num_maps, i; | |
1457 | ||
1458 | bitmap->bits_per_word = bpw; | |
1459 | ||
1460 | num_maps = ALIGN(nr_cpu_ids, bpw) / bpw; | |
1461 | bitmap->map = kzalloc_node(num_maps * sizeof(struct blk_align_bitmap), | |
1462 | GFP_KERNEL, node); | |
1463 | if (!bitmap->map) | |
1464 | return -ENOMEM; | |
1465 | ||
1466 | bitmap->map_size = num_maps; | |
1467 | ||
1468 | total = nr_cpu_ids; | |
1469 | for (i = 0; i < num_maps; i++) { | |
1470 | bitmap->map[i].depth = min(total, bitmap->bits_per_word); | |
1471 | total -= bitmap->map[i].depth; | |
1472 | } | |
1473 | ||
1474 | return 0; | |
1475 | } | |
1476 | ||
484b4061 JA |
1477 | static int blk_mq_hctx_cpu_offline(struct blk_mq_hw_ctx *hctx, int cpu) |
1478 | { | |
1479 | struct request_queue *q = hctx->queue; | |
1480 | struct blk_mq_ctx *ctx; | |
1481 | LIST_HEAD(tmp); | |
1482 | ||
1483 | /* | |
1484 | * Move ctx entries to new CPU, if this one is going away. | |
1485 | */ | |
1486 | ctx = __blk_mq_get_ctx(q, cpu); | |
1487 | ||
1488 | spin_lock(&ctx->lock); | |
1489 | if (!list_empty(&ctx->rq_list)) { | |
1490 | list_splice_init(&ctx->rq_list, &tmp); | |
1491 | blk_mq_hctx_clear_pending(hctx, ctx); | |
1492 | } | |
1493 | spin_unlock(&ctx->lock); | |
1494 | ||
1495 | if (list_empty(&tmp)) | |
1496 | return NOTIFY_OK; | |
1497 | ||
1498 | ctx = blk_mq_get_ctx(q); | |
1499 | spin_lock(&ctx->lock); | |
1500 | ||
1501 | while (!list_empty(&tmp)) { | |
1502 | struct request *rq; | |
1503 | ||
1504 | rq = list_first_entry(&tmp, struct request, queuelist); | |
1505 | rq->mq_ctx = ctx; | |
1506 | list_move_tail(&rq->queuelist, &ctx->rq_list); | |
1507 | } | |
1508 | ||
1509 | hctx = q->mq_ops->map_queue(q, ctx->cpu); | |
1510 | blk_mq_hctx_mark_pending(hctx, ctx); | |
1511 | ||
1512 | spin_unlock(&ctx->lock); | |
1513 | ||
1514 | blk_mq_run_hw_queue(hctx, true); | |
1515 | blk_mq_put_ctx(ctx); | |
1516 | return NOTIFY_OK; | |
1517 | } | |
1518 | ||
1519 | static int blk_mq_hctx_cpu_online(struct blk_mq_hw_ctx *hctx, int cpu) | |
1520 | { | |
1521 | struct request_queue *q = hctx->queue; | |
1522 | struct blk_mq_tag_set *set = q->tag_set; | |
1523 | ||
1524 | if (set->tags[hctx->queue_num]) | |
1525 | return NOTIFY_OK; | |
1526 | ||
1527 | set->tags[hctx->queue_num] = blk_mq_init_rq_map(set, hctx->queue_num); | |
1528 | if (!set->tags[hctx->queue_num]) | |
1529 | return NOTIFY_STOP; | |
1530 | ||
1531 | hctx->tags = set->tags[hctx->queue_num]; | |
1532 | return NOTIFY_OK; | |
1533 | } | |
1534 | ||
1535 | static int blk_mq_hctx_notify(void *data, unsigned long action, | |
1536 | unsigned int cpu) | |
1537 | { | |
1538 | struct blk_mq_hw_ctx *hctx = data; | |
1539 | ||
1540 | if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) | |
1541 | return blk_mq_hctx_cpu_offline(hctx, cpu); | |
1542 | else if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) | |
1543 | return blk_mq_hctx_cpu_online(hctx, cpu); | |
1544 | ||
1545 | return NOTIFY_OK; | |
1546 | } | |
1547 | ||
624dbe47 ML |
1548 | static void blk_mq_exit_hw_queues(struct request_queue *q, |
1549 | struct blk_mq_tag_set *set, int nr_queue) | |
1550 | { | |
1551 | struct blk_mq_hw_ctx *hctx; | |
1552 | unsigned int i; | |
1553 | ||
1554 | queue_for_each_hw_ctx(q, hctx, i) { | |
1555 | if (i == nr_queue) | |
1556 | break; | |
1557 | ||
1558 | if (set->ops->exit_hctx) | |
1559 | set->ops->exit_hctx(hctx, i); | |
1560 | ||
1561 | blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier); | |
1562 | kfree(hctx->ctxs); | |
1563 | blk_mq_free_bitmap(&hctx->ctx_map); | |
1564 | } | |
1565 | ||
1566 | } | |
1567 | ||
1568 | static void blk_mq_free_hw_queues(struct request_queue *q, | |
1569 | struct blk_mq_tag_set *set) | |
1570 | { | |
1571 | struct blk_mq_hw_ctx *hctx; | |
1572 | unsigned int i; | |
1573 | ||
1574 | queue_for_each_hw_ctx(q, hctx, i) { | |
1575 | free_cpumask_var(hctx->cpumask); | |
cdef54dd | 1576 | kfree(hctx); |
624dbe47 ML |
1577 | } |
1578 | } | |
1579 | ||
320ae51f | 1580 | static int blk_mq_init_hw_queues(struct request_queue *q, |
24d2f903 | 1581 | struct blk_mq_tag_set *set) |
320ae51f JA |
1582 | { |
1583 | struct blk_mq_hw_ctx *hctx; | |
624dbe47 | 1584 | unsigned int i; |
320ae51f JA |
1585 | |
1586 | /* | |
1587 | * Initialize hardware queues | |
1588 | */ | |
1589 | queue_for_each_hw_ctx(q, hctx, i) { | |
320ae51f JA |
1590 | int node; |
1591 | ||
1592 | node = hctx->numa_node; | |
1593 | if (node == NUMA_NO_NODE) | |
24d2f903 | 1594 | node = hctx->numa_node = set->numa_node; |
320ae51f | 1595 | |
70f4db63 CH |
1596 | INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn); |
1597 | INIT_DELAYED_WORK(&hctx->delay_work, blk_mq_delay_work_fn); | |
320ae51f JA |
1598 | spin_lock_init(&hctx->lock); |
1599 | INIT_LIST_HEAD(&hctx->dispatch); | |
1600 | hctx->queue = q; | |
1601 | hctx->queue_num = i; | |
24d2f903 CH |
1602 | hctx->flags = set->flags; |
1603 | hctx->cmd_size = set->cmd_size; | |
320ae51f JA |
1604 | |
1605 | blk_mq_init_cpu_notifier(&hctx->cpu_notifier, | |
1606 | blk_mq_hctx_notify, hctx); | |
1607 | blk_mq_register_cpu_notifier(&hctx->cpu_notifier); | |
1608 | ||
24d2f903 | 1609 | hctx->tags = set->tags[i]; |
320ae51f JA |
1610 | |
1611 | /* | |
1612 | * Allocate space for all possible cpus to avoid allocation in | |
1613 | * runtime | |
1614 | */ | |
1615 | hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *), | |
1616 | GFP_KERNEL, node); | |
1617 | if (!hctx->ctxs) | |
1618 | break; | |
1619 | ||
1429d7c9 | 1620 | if (blk_mq_alloc_bitmap(&hctx->ctx_map, node)) |
320ae51f JA |
1621 | break; |
1622 | ||
320ae51f JA |
1623 | hctx->nr_ctx = 0; |
1624 | ||
24d2f903 CH |
1625 | if (set->ops->init_hctx && |
1626 | set->ops->init_hctx(hctx, set->driver_data, i)) | |
320ae51f JA |
1627 | break; |
1628 | } | |
1629 | ||
1630 | if (i == q->nr_hw_queues) | |
1631 | return 0; | |
1632 | ||
1633 | /* | |
1634 | * Init failed | |
1635 | */ | |
624dbe47 | 1636 | blk_mq_exit_hw_queues(q, set, i); |
320ae51f JA |
1637 | |
1638 | return 1; | |
1639 | } | |
1640 | ||
1641 | static void blk_mq_init_cpu_queues(struct request_queue *q, | |
1642 | unsigned int nr_hw_queues) | |
1643 | { | |
1644 | unsigned int i; | |
1645 | ||
1646 | for_each_possible_cpu(i) { | |
1647 | struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i); | |
1648 | struct blk_mq_hw_ctx *hctx; | |
1649 | ||
1650 | memset(__ctx, 0, sizeof(*__ctx)); | |
1651 | __ctx->cpu = i; | |
1652 | spin_lock_init(&__ctx->lock); | |
1653 | INIT_LIST_HEAD(&__ctx->rq_list); | |
1654 | __ctx->queue = q; | |
1655 | ||
1656 | /* If the cpu isn't online, the cpu is mapped to first hctx */ | |
320ae51f JA |
1657 | if (!cpu_online(i)) |
1658 | continue; | |
1659 | ||
e4043dcf JA |
1660 | hctx = q->mq_ops->map_queue(q, i); |
1661 | cpumask_set_cpu(i, hctx->cpumask); | |
1662 | hctx->nr_ctx++; | |
1663 | ||
320ae51f JA |
1664 | /* |
1665 | * Set local node, IFF we have more than one hw queue. If | |
1666 | * not, we remain on the home node of the device | |
1667 | */ | |
1668 | if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE) | |
1669 | hctx->numa_node = cpu_to_node(i); | |
1670 | } | |
1671 | } | |
1672 | ||
1673 | static void blk_mq_map_swqueue(struct request_queue *q) | |
1674 | { | |
1675 | unsigned int i; | |
1676 | struct blk_mq_hw_ctx *hctx; | |
1677 | struct blk_mq_ctx *ctx; | |
1678 | ||
1679 | queue_for_each_hw_ctx(q, hctx, i) { | |
e4043dcf | 1680 | cpumask_clear(hctx->cpumask); |
320ae51f JA |
1681 | hctx->nr_ctx = 0; |
1682 | } | |
1683 | ||
1684 | /* | |
1685 | * Map software to hardware queues | |
1686 | */ | |
1687 | queue_for_each_ctx(q, ctx, i) { | |
1688 | /* If the cpu isn't online, the cpu is mapped to first hctx */ | |
e4043dcf JA |
1689 | if (!cpu_online(i)) |
1690 | continue; | |
1691 | ||
320ae51f | 1692 | hctx = q->mq_ops->map_queue(q, i); |
e4043dcf | 1693 | cpumask_set_cpu(i, hctx->cpumask); |
320ae51f JA |
1694 | ctx->index_hw = hctx->nr_ctx; |
1695 | hctx->ctxs[hctx->nr_ctx++] = ctx; | |
1696 | } | |
506e931f JA |
1697 | |
1698 | queue_for_each_hw_ctx(q, hctx, i) { | |
484b4061 JA |
1699 | /* |
1700 | * If not software queues are mapped to this hardware queue, | |
1701 | * disable it and free the request entries | |
1702 | */ | |
1703 | if (!hctx->nr_ctx) { | |
1704 | struct blk_mq_tag_set *set = q->tag_set; | |
1705 | ||
1706 | if (set->tags[i]) { | |
1707 | blk_mq_free_rq_map(set, set->tags[i], i); | |
1708 | set->tags[i] = NULL; | |
1709 | hctx->tags = NULL; | |
1710 | } | |
1711 | continue; | |
1712 | } | |
1713 | ||
1714 | /* | |
1715 | * Initialize batch roundrobin counts | |
1716 | */ | |
506e931f JA |
1717 | hctx->next_cpu = cpumask_first(hctx->cpumask); |
1718 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; | |
1719 | } | |
320ae51f JA |
1720 | } |
1721 | ||
0d2602ca JA |
1722 | static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set) |
1723 | { | |
1724 | struct blk_mq_hw_ctx *hctx; | |
1725 | struct request_queue *q; | |
1726 | bool shared; | |
1727 | int i; | |
1728 | ||
1729 | if (set->tag_list.next == set->tag_list.prev) | |
1730 | shared = false; | |
1731 | else | |
1732 | shared = true; | |
1733 | ||
1734 | list_for_each_entry(q, &set->tag_list, tag_set_list) { | |
1735 | blk_mq_freeze_queue(q); | |
1736 | ||
1737 | queue_for_each_hw_ctx(q, hctx, i) { | |
1738 | if (shared) | |
1739 | hctx->flags |= BLK_MQ_F_TAG_SHARED; | |
1740 | else | |
1741 | hctx->flags &= ~BLK_MQ_F_TAG_SHARED; | |
1742 | } | |
1743 | blk_mq_unfreeze_queue(q); | |
1744 | } | |
1745 | } | |
1746 | ||
1747 | static void blk_mq_del_queue_tag_set(struct request_queue *q) | |
1748 | { | |
1749 | struct blk_mq_tag_set *set = q->tag_set; | |
1750 | ||
1751 | blk_mq_freeze_queue(q); | |
1752 | ||
1753 | mutex_lock(&set->tag_list_lock); | |
1754 | list_del_init(&q->tag_set_list); | |
1755 | blk_mq_update_tag_set_depth(set); | |
1756 | mutex_unlock(&set->tag_list_lock); | |
1757 | ||
1758 | blk_mq_unfreeze_queue(q); | |
1759 | } | |
1760 | ||
1761 | static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set, | |
1762 | struct request_queue *q) | |
1763 | { | |
1764 | q->tag_set = set; | |
1765 | ||
1766 | mutex_lock(&set->tag_list_lock); | |
1767 | list_add_tail(&q->tag_set_list, &set->tag_list); | |
1768 | blk_mq_update_tag_set_depth(set); | |
1769 | mutex_unlock(&set->tag_list_lock); | |
1770 | } | |
1771 | ||
24d2f903 | 1772 | struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set) |
320ae51f JA |
1773 | { |
1774 | struct blk_mq_hw_ctx **hctxs; | |
1775 | struct blk_mq_ctx *ctx; | |
1776 | struct request_queue *q; | |
f14bbe77 | 1777 | unsigned int *map; |
320ae51f JA |
1778 | int i; |
1779 | ||
320ae51f JA |
1780 | ctx = alloc_percpu(struct blk_mq_ctx); |
1781 | if (!ctx) | |
1782 | return ERR_PTR(-ENOMEM); | |
1783 | ||
24d2f903 CH |
1784 | hctxs = kmalloc_node(set->nr_hw_queues * sizeof(*hctxs), GFP_KERNEL, |
1785 | set->numa_node); | |
320ae51f JA |
1786 | |
1787 | if (!hctxs) | |
1788 | goto err_percpu; | |
1789 | ||
f14bbe77 JA |
1790 | map = blk_mq_make_queue_map(set); |
1791 | if (!map) | |
1792 | goto err_map; | |
1793 | ||
24d2f903 | 1794 | for (i = 0; i < set->nr_hw_queues; i++) { |
f14bbe77 JA |
1795 | int node = blk_mq_hw_queue_to_node(map, i); |
1796 | ||
cdef54dd CH |
1797 | hctxs[i] = kzalloc_node(sizeof(struct blk_mq_hw_ctx), |
1798 | GFP_KERNEL, node); | |
320ae51f JA |
1799 | if (!hctxs[i]) |
1800 | goto err_hctxs; | |
1801 | ||
e4043dcf JA |
1802 | if (!zalloc_cpumask_var(&hctxs[i]->cpumask, GFP_KERNEL)) |
1803 | goto err_hctxs; | |
1804 | ||
0d2602ca | 1805 | atomic_set(&hctxs[i]->nr_active, 0); |
f14bbe77 | 1806 | hctxs[i]->numa_node = node; |
320ae51f JA |
1807 | hctxs[i]->queue_num = i; |
1808 | } | |
1809 | ||
24d2f903 | 1810 | q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node); |
320ae51f JA |
1811 | if (!q) |
1812 | goto err_hctxs; | |
1813 | ||
3d2936f4 ML |
1814 | if (percpu_counter_init(&q->mq_usage_counter, 0)) |
1815 | goto err_map; | |
1816 | ||
320ae51f JA |
1817 | setup_timer(&q->timeout, blk_mq_rq_timer, (unsigned long) q); |
1818 | blk_queue_rq_timeout(q, 30000); | |
1819 | ||
1820 | q->nr_queues = nr_cpu_ids; | |
24d2f903 | 1821 | q->nr_hw_queues = set->nr_hw_queues; |
f14bbe77 | 1822 | q->mq_map = map; |
320ae51f JA |
1823 | |
1824 | q->queue_ctx = ctx; | |
1825 | q->queue_hw_ctx = hctxs; | |
1826 | ||
24d2f903 | 1827 | q->mq_ops = set->ops; |
94eddfbe | 1828 | q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT; |
320ae51f | 1829 | |
05f1dd53 JA |
1830 | if (!(set->flags & BLK_MQ_F_SG_MERGE)) |
1831 | q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE; | |
1832 | ||
1be036e9 CH |
1833 | q->sg_reserved_size = INT_MAX; |
1834 | ||
6fca6a61 CH |
1835 | INIT_WORK(&q->requeue_work, blk_mq_requeue_work); |
1836 | INIT_LIST_HEAD(&q->requeue_list); | |
1837 | spin_lock_init(&q->requeue_lock); | |
1838 | ||
07068d5b JA |
1839 | if (q->nr_hw_queues > 1) |
1840 | blk_queue_make_request(q, blk_mq_make_request); | |
1841 | else | |
1842 | blk_queue_make_request(q, blk_sq_make_request); | |
1843 | ||
87ee7b11 | 1844 | blk_queue_rq_timed_out(q, blk_mq_rq_timed_out); |
24d2f903 CH |
1845 | if (set->timeout) |
1846 | blk_queue_rq_timeout(q, set->timeout); | |
320ae51f | 1847 | |
eba71768 JA |
1848 | /* |
1849 | * Do this after blk_queue_make_request() overrides it... | |
1850 | */ | |
1851 | q->nr_requests = set->queue_depth; | |
1852 | ||
24d2f903 CH |
1853 | if (set->ops->complete) |
1854 | blk_queue_softirq_done(q, set->ops->complete); | |
30a91cb4 | 1855 | |
320ae51f | 1856 | blk_mq_init_flush(q); |
24d2f903 | 1857 | blk_mq_init_cpu_queues(q, set->nr_hw_queues); |
320ae51f | 1858 | |
24d2f903 CH |
1859 | q->flush_rq = kzalloc(round_up(sizeof(struct request) + |
1860 | set->cmd_size, cache_line_size()), | |
1861 | GFP_KERNEL); | |
18741986 | 1862 | if (!q->flush_rq) |
320ae51f JA |
1863 | goto err_hw; |
1864 | ||
24d2f903 | 1865 | if (blk_mq_init_hw_queues(q, set)) |
18741986 CH |
1866 | goto err_flush_rq; |
1867 | ||
320ae51f JA |
1868 | mutex_lock(&all_q_mutex); |
1869 | list_add_tail(&q->all_q_node, &all_q_list); | |
1870 | mutex_unlock(&all_q_mutex); | |
1871 | ||
0d2602ca JA |
1872 | blk_mq_add_queue_tag_set(set, q); |
1873 | ||
484b4061 JA |
1874 | blk_mq_map_swqueue(q); |
1875 | ||
320ae51f | 1876 | return q; |
18741986 CH |
1877 | |
1878 | err_flush_rq: | |
1879 | kfree(q->flush_rq); | |
320ae51f | 1880 | err_hw: |
320ae51f JA |
1881 | blk_cleanup_queue(q); |
1882 | err_hctxs: | |
f14bbe77 | 1883 | kfree(map); |
24d2f903 | 1884 | for (i = 0; i < set->nr_hw_queues; i++) { |
320ae51f JA |
1885 | if (!hctxs[i]) |
1886 | break; | |
e4043dcf | 1887 | free_cpumask_var(hctxs[i]->cpumask); |
cdef54dd | 1888 | kfree(hctxs[i]); |
320ae51f | 1889 | } |
f14bbe77 | 1890 | err_map: |
320ae51f JA |
1891 | kfree(hctxs); |
1892 | err_percpu: | |
1893 | free_percpu(ctx); | |
1894 | return ERR_PTR(-ENOMEM); | |
1895 | } | |
1896 | EXPORT_SYMBOL(blk_mq_init_queue); | |
1897 | ||
1898 | void blk_mq_free_queue(struct request_queue *q) | |
1899 | { | |
624dbe47 | 1900 | struct blk_mq_tag_set *set = q->tag_set; |
320ae51f | 1901 | |
0d2602ca JA |
1902 | blk_mq_del_queue_tag_set(q); |
1903 | ||
624dbe47 ML |
1904 | blk_mq_exit_hw_queues(q, set, set->nr_hw_queues); |
1905 | blk_mq_free_hw_queues(q, set); | |
320ae51f | 1906 | |
3d2936f4 ML |
1907 | percpu_counter_destroy(&q->mq_usage_counter); |
1908 | ||
320ae51f JA |
1909 | free_percpu(q->queue_ctx); |
1910 | kfree(q->queue_hw_ctx); | |
1911 | kfree(q->mq_map); | |
1912 | ||
1913 | q->queue_ctx = NULL; | |
1914 | q->queue_hw_ctx = NULL; | |
1915 | q->mq_map = NULL; | |
1916 | ||
1917 | mutex_lock(&all_q_mutex); | |
1918 | list_del_init(&q->all_q_node); | |
1919 | mutex_unlock(&all_q_mutex); | |
1920 | } | |
320ae51f JA |
1921 | |
1922 | /* Basically redo blk_mq_init_queue with queue frozen */ | |
f618ef7c | 1923 | static void blk_mq_queue_reinit(struct request_queue *q) |
320ae51f JA |
1924 | { |
1925 | blk_mq_freeze_queue(q); | |
1926 | ||
1927 | blk_mq_update_queue_map(q->mq_map, q->nr_hw_queues); | |
1928 | ||
1929 | /* | |
1930 | * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe | |
1931 | * we should change hctx numa_node according to new topology (this | |
1932 | * involves free and re-allocate memory, worthy doing?) | |
1933 | */ | |
1934 | ||
1935 | blk_mq_map_swqueue(q); | |
1936 | ||
1937 | blk_mq_unfreeze_queue(q); | |
1938 | } | |
1939 | ||
f618ef7c PG |
1940 | static int blk_mq_queue_reinit_notify(struct notifier_block *nb, |
1941 | unsigned long action, void *hcpu) | |
320ae51f JA |
1942 | { |
1943 | struct request_queue *q; | |
1944 | ||
1945 | /* | |
9fccfed8 JA |
1946 | * Before new mappings are established, hotadded cpu might already |
1947 | * start handling requests. This doesn't break anything as we map | |
1948 | * offline CPUs to first hardware queue. We will re-init the queue | |
1949 | * below to get optimal settings. | |
320ae51f JA |
1950 | */ |
1951 | if (action != CPU_DEAD && action != CPU_DEAD_FROZEN && | |
1952 | action != CPU_ONLINE && action != CPU_ONLINE_FROZEN) | |
1953 | return NOTIFY_OK; | |
1954 | ||
1955 | mutex_lock(&all_q_mutex); | |
1956 | list_for_each_entry(q, &all_q_list, all_q_node) | |
1957 | blk_mq_queue_reinit(q); | |
1958 | mutex_unlock(&all_q_mutex); | |
1959 | return NOTIFY_OK; | |
1960 | } | |
1961 | ||
24d2f903 CH |
1962 | int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set) |
1963 | { | |
1964 | int i; | |
1965 | ||
1966 | if (!set->nr_hw_queues) | |
1967 | return -EINVAL; | |
1968 | if (!set->queue_depth || set->queue_depth > BLK_MQ_MAX_DEPTH) | |
1969 | return -EINVAL; | |
1970 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) | |
1971 | return -EINVAL; | |
1972 | ||
cdef54dd | 1973 | if (!set->nr_hw_queues || !set->ops->queue_rq || !set->ops->map_queue) |
24d2f903 CH |
1974 | return -EINVAL; |
1975 | ||
1976 | ||
48479005 ML |
1977 | set->tags = kmalloc_node(set->nr_hw_queues * |
1978 | sizeof(struct blk_mq_tags *), | |
24d2f903 CH |
1979 | GFP_KERNEL, set->numa_node); |
1980 | if (!set->tags) | |
1981 | goto out; | |
1982 | ||
1983 | for (i = 0; i < set->nr_hw_queues; i++) { | |
1984 | set->tags[i] = blk_mq_init_rq_map(set, i); | |
1985 | if (!set->tags[i]) | |
1986 | goto out_unwind; | |
1987 | } | |
1988 | ||
0d2602ca JA |
1989 | mutex_init(&set->tag_list_lock); |
1990 | INIT_LIST_HEAD(&set->tag_list); | |
1991 | ||
24d2f903 CH |
1992 | return 0; |
1993 | ||
1994 | out_unwind: | |
1995 | while (--i >= 0) | |
1996 | blk_mq_free_rq_map(set, set->tags[i], i); | |
1997 | out: | |
1998 | return -ENOMEM; | |
1999 | } | |
2000 | EXPORT_SYMBOL(blk_mq_alloc_tag_set); | |
2001 | ||
2002 | void blk_mq_free_tag_set(struct blk_mq_tag_set *set) | |
2003 | { | |
2004 | int i; | |
2005 | ||
484b4061 JA |
2006 | for (i = 0; i < set->nr_hw_queues; i++) { |
2007 | if (set->tags[i]) | |
2008 | blk_mq_free_rq_map(set, set->tags[i], i); | |
2009 | } | |
2010 | ||
981bd189 | 2011 | kfree(set->tags); |
24d2f903 CH |
2012 | } |
2013 | EXPORT_SYMBOL(blk_mq_free_tag_set); | |
2014 | ||
e3a2b3f9 JA |
2015 | int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr) |
2016 | { | |
2017 | struct blk_mq_tag_set *set = q->tag_set; | |
2018 | struct blk_mq_hw_ctx *hctx; | |
2019 | int i, ret; | |
2020 | ||
2021 | if (!set || nr > set->queue_depth) | |
2022 | return -EINVAL; | |
2023 | ||
2024 | ret = 0; | |
2025 | queue_for_each_hw_ctx(q, hctx, i) { | |
2026 | ret = blk_mq_tag_update_depth(hctx->tags, nr); | |
2027 | if (ret) | |
2028 | break; | |
2029 | } | |
2030 | ||
2031 | if (!ret) | |
2032 | q->nr_requests = nr; | |
2033 | ||
2034 | return ret; | |
2035 | } | |
2036 | ||
676141e4 JA |
2037 | void blk_mq_disable_hotplug(void) |
2038 | { | |
2039 | mutex_lock(&all_q_mutex); | |
2040 | } | |
2041 | ||
2042 | void blk_mq_enable_hotplug(void) | |
2043 | { | |
2044 | mutex_unlock(&all_q_mutex); | |
2045 | } | |
2046 | ||
320ae51f JA |
2047 | static int __init blk_mq_init(void) |
2048 | { | |
320ae51f JA |
2049 | blk_mq_cpu_init(); |
2050 | ||
2051 | /* Must be called after percpu_counter_hotcpu_callback() */ | |
2052 | hotcpu_notifier(blk_mq_queue_reinit_notify, -10); | |
2053 | ||
2054 | return 0; | |
2055 | } | |
2056 | subsys_initcall(blk_mq_init); |