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 | ||
24d2f903 CH |
544 | struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag) |
545 | { | |
546 | return tags->rqs[tag]; | |
547 | } | |
548 | EXPORT_SYMBOL(blk_mq_tag_to_rq); | |
549 | ||
320ae51f JA |
550 | struct blk_mq_timeout_data { |
551 | struct blk_mq_hw_ctx *hctx; | |
552 | unsigned long *next; | |
553 | unsigned int *next_set; | |
554 | }; | |
555 | ||
556 | static void blk_mq_timeout_check(void *__data, unsigned long *free_tags) | |
557 | { | |
558 | struct blk_mq_timeout_data *data = __data; | |
559 | struct blk_mq_hw_ctx *hctx = data->hctx; | |
560 | unsigned int tag; | |
561 | ||
562 | /* It may not be in flight yet (this is where | |
563 | * the REQ_ATOMIC_STARTED flag comes in). The requests are | |
564 | * statically allocated, so we know it's always safe to access the | |
565 | * memory associated with a bit offset into ->rqs[]. | |
566 | */ | |
567 | tag = 0; | |
568 | do { | |
569 | struct request *rq; | |
570 | ||
24d2f903 CH |
571 | tag = find_next_zero_bit(free_tags, hctx->tags->nr_tags, tag); |
572 | if (tag >= hctx->tags->nr_tags) | |
320ae51f JA |
573 | break; |
574 | ||
24d2f903 CH |
575 | rq = blk_mq_tag_to_rq(hctx->tags, tag++); |
576 | if (rq->q != hctx->queue) | |
577 | continue; | |
320ae51f JA |
578 | if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) |
579 | continue; | |
580 | ||
581 | blk_rq_check_expired(rq, data->next, data->next_set); | |
582 | } while (1); | |
583 | } | |
584 | ||
585 | static void blk_mq_hw_ctx_check_timeout(struct blk_mq_hw_ctx *hctx, | |
586 | unsigned long *next, | |
587 | unsigned int *next_set) | |
588 | { | |
589 | struct blk_mq_timeout_data data = { | |
590 | .hctx = hctx, | |
591 | .next = next, | |
592 | .next_set = next_set, | |
593 | }; | |
594 | ||
595 | /* | |
596 | * Ask the tagging code to iterate busy requests, so we can | |
597 | * check them for timeout. | |
598 | */ | |
599 | blk_mq_tag_busy_iter(hctx->tags, blk_mq_timeout_check, &data); | |
600 | } | |
601 | ||
87ee7b11 JA |
602 | static enum blk_eh_timer_return blk_mq_rq_timed_out(struct request *rq) |
603 | { | |
604 | struct request_queue *q = rq->q; | |
605 | ||
606 | /* | |
607 | * We know that complete is set at this point. If STARTED isn't set | |
608 | * anymore, then the request isn't active and the "timeout" should | |
609 | * just be ignored. This can happen due to the bitflag ordering. | |
610 | * Timeout first checks if STARTED is set, and if it is, assumes | |
611 | * the request is active. But if we race with completion, then | |
612 | * we both flags will get cleared. So check here again, and ignore | |
613 | * a timeout event with a request that isn't active. | |
614 | */ | |
615 | if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) | |
616 | return BLK_EH_NOT_HANDLED; | |
617 | ||
618 | if (!q->mq_ops->timeout) | |
619 | return BLK_EH_RESET_TIMER; | |
620 | ||
621 | return q->mq_ops->timeout(rq); | |
622 | } | |
623 | ||
320ae51f JA |
624 | static void blk_mq_rq_timer(unsigned long data) |
625 | { | |
626 | struct request_queue *q = (struct request_queue *) data; | |
627 | struct blk_mq_hw_ctx *hctx; | |
628 | unsigned long next = 0; | |
629 | int i, next_set = 0; | |
630 | ||
484b4061 JA |
631 | queue_for_each_hw_ctx(q, hctx, i) { |
632 | /* | |
633 | * If not software queues are currently mapped to this | |
634 | * hardware queue, there's nothing to check | |
635 | */ | |
636 | if (!hctx->nr_ctx || !hctx->tags) | |
637 | continue; | |
638 | ||
320ae51f | 639 | blk_mq_hw_ctx_check_timeout(hctx, &next, &next_set); |
484b4061 | 640 | } |
320ae51f | 641 | |
0d2602ca JA |
642 | if (next_set) { |
643 | next = blk_rq_timeout(round_jiffies_up(next)); | |
644 | mod_timer(&q->timeout, next); | |
645 | } else { | |
646 | queue_for_each_hw_ctx(q, hctx, i) | |
647 | blk_mq_tag_idle(hctx); | |
648 | } | |
320ae51f JA |
649 | } |
650 | ||
651 | /* | |
652 | * Reverse check our software queue for entries that we could potentially | |
653 | * merge with. Currently includes a hand-wavy stop count of 8, to not spend | |
654 | * too much time checking for merges. | |
655 | */ | |
656 | static bool blk_mq_attempt_merge(struct request_queue *q, | |
657 | struct blk_mq_ctx *ctx, struct bio *bio) | |
658 | { | |
659 | struct request *rq; | |
660 | int checked = 8; | |
661 | ||
662 | list_for_each_entry_reverse(rq, &ctx->rq_list, queuelist) { | |
663 | int el_ret; | |
664 | ||
665 | if (!checked--) | |
666 | break; | |
667 | ||
668 | if (!blk_rq_merge_ok(rq, bio)) | |
669 | continue; | |
670 | ||
671 | el_ret = blk_try_merge(rq, bio); | |
672 | if (el_ret == ELEVATOR_BACK_MERGE) { | |
673 | if (bio_attempt_back_merge(q, rq, bio)) { | |
674 | ctx->rq_merged++; | |
675 | return true; | |
676 | } | |
677 | break; | |
678 | } else if (el_ret == ELEVATOR_FRONT_MERGE) { | |
679 | if (bio_attempt_front_merge(q, rq, bio)) { | |
680 | ctx->rq_merged++; | |
681 | return true; | |
682 | } | |
683 | break; | |
684 | } | |
685 | } | |
686 | ||
687 | return false; | |
688 | } | |
689 | ||
1429d7c9 JA |
690 | /* |
691 | * Process software queues that have been marked busy, splicing them | |
692 | * to the for-dispatch | |
693 | */ | |
694 | static void flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list) | |
695 | { | |
696 | struct blk_mq_ctx *ctx; | |
697 | int i; | |
698 | ||
699 | for (i = 0; i < hctx->ctx_map.map_size; i++) { | |
700 | struct blk_align_bitmap *bm = &hctx->ctx_map.map[i]; | |
701 | unsigned int off, bit; | |
702 | ||
703 | if (!bm->word) | |
704 | continue; | |
705 | ||
706 | bit = 0; | |
707 | off = i * hctx->ctx_map.bits_per_word; | |
708 | do { | |
709 | bit = find_next_bit(&bm->word, bm->depth, bit); | |
710 | if (bit >= bm->depth) | |
711 | break; | |
712 | ||
713 | ctx = hctx->ctxs[bit + off]; | |
714 | clear_bit(bit, &bm->word); | |
715 | spin_lock(&ctx->lock); | |
716 | list_splice_tail_init(&ctx->rq_list, list); | |
717 | spin_unlock(&ctx->lock); | |
718 | ||
719 | bit++; | |
720 | } while (1); | |
721 | } | |
722 | } | |
723 | ||
320ae51f JA |
724 | /* |
725 | * Run this hardware queue, pulling any software queues mapped to it in. | |
726 | * Note that this function currently has various problems around ordering | |
727 | * of IO. In particular, we'd like FIFO behaviour on handling existing | |
728 | * items on the hctx->dispatch list. Ignore that for now. | |
729 | */ | |
730 | static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx) | |
731 | { | |
732 | struct request_queue *q = hctx->queue; | |
320ae51f JA |
733 | struct request *rq; |
734 | LIST_HEAD(rq_list); | |
1429d7c9 | 735 | int queued; |
320ae51f | 736 | |
fd1270d5 | 737 | WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask)); |
e4043dcf | 738 | |
5d12f905 | 739 | if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state))) |
320ae51f JA |
740 | return; |
741 | ||
742 | hctx->run++; | |
743 | ||
744 | /* | |
745 | * Touch any software queue that has pending entries. | |
746 | */ | |
1429d7c9 | 747 | flush_busy_ctxs(hctx, &rq_list); |
320ae51f JA |
748 | |
749 | /* | |
750 | * If we have previous entries on our dispatch list, grab them | |
751 | * and stuff them at the front for more fair dispatch. | |
752 | */ | |
753 | if (!list_empty_careful(&hctx->dispatch)) { | |
754 | spin_lock(&hctx->lock); | |
755 | if (!list_empty(&hctx->dispatch)) | |
756 | list_splice_init(&hctx->dispatch, &rq_list); | |
757 | spin_unlock(&hctx->lock); | |
758 | } | |
759 | ||
320ae51f JA |
760 | /* |
761 | * Now process all the entries, sending them to the driver. | |
762 | */ | |
1429d7c9 | 763 | queued = 0; |
320ae51f JA |
764 | while (!list_empty(&rq_list)) { |
765 | int ret; | |
766 | ||
767 | rq = list_first_entry(&rq_list, struct request, queuelist); | |
768 | list_del_init(&rq->queuelist); | |
320ae51f | 769 | |
49f5baa5 | 770 | blk_mq_start_request(rq, list_empty(&rq_list)); |
320ae51f JA |
771 | |
772 | ret = q->mq_ops->queue_rq(hctx, rq); | |
773 | switch (ret) { | |
774 | case BLK_MQ_RQ_QUEUE_OK: | |
775 | queued++; | |
776 | continue; | |
777 | case BLK_MQ_RQ_QUEUE_BUSY: | |
320ae51f | 778 | list_add(&rq->queuelist, &rq_list); |
ed0791b2 | 779 | __blk_mq_requeue_request(rq); |
320ae51f JA |
780 | break; |
781 | default: | |
782 | pr_err("blk-mq: bad return on queue: %d\n", ret); | |
320ae51f | 783 | case BLK_MQ_RQ_QUEUE_ERROR: |
1e93b8c2 | 784 | rq->errors = -EIO; |
320ae51f JA |
785 | blk_mq_end_io(rq, rq->errors); |
786 | break; | |
787 | } | |
788 | ||
789 | if (ret == BLK_MQ_RQ_QUEUE_BUSY) | |
790 | break; | |
791 | } | |
792 | ||
793 | if (!queued) | |
794 | hctx->dispatched[0]++; | |
795 | else if (queued < (1 << (BLK_MQ_MAX_DISPATCH_ORDER - 1))) | |
796 | hctx->dispatched[ilog2(queued) + 1]++; | |
797 | ||
798 | /* | |
799 | * Any items that need requeuing? Stuff them into hctx->dispatch, | |
800 | * that is where we will continue on next queue run. | |
801 | */ | |
802 | if (!list_empty(&rq_list)) { | |
803 | spin_lock(&hctx->lock); | |
804 | list_splice(&rq_list, &hctx->dispatch); | |
805 | spin_unlock(&hctx->lock); | |
806 | } | |
807 | } | |
808 | ||
506e931f JA |
809 | /* |
810 | * It'd be great if the workqueue API had a way to pass | |
811 | * in a mask and had some smarts for more clever placement. | |
812 | * For now we just round-robin here, switching for every | |
813 | * BLK_MQ_CPU_WORK_BATCH queued items. | |
814 | */ | |
815 | static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx) | |
816 | { | |
817 | int cpu = hctx->next_cpu; | |
818 | ||
819 | if (--hctx->next_cpu_batch <= 0) { | |
820 | int next_cpu; | |
821 | ||
822 | next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask); | |
823 | if (next_cpu >= nr_cpu_ids) | |
824 | next_cpu = cpumask_first(hctx->cpumask); | |
825 | ||
826 | hctx->next_cpu = next_cpu; | |
827 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; | |
828 | } | |
829 | ||
830 | return cpu; | |
831 | } | |
832 | ||
320ae51f JA |
833 | void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) |
834 | { | |
5d12f905 | 835 | if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state))) |
320ae51f JA |
836 | return; |
837 | ||
e4043dcf | 838 | if (!async && cpumask_test_cpu(smp_processor_id(), hctx->cpumask)) |
320ae51f | 839 | __blk_mq_run_hw_queue(hctx); |
e4043dcf | 840 | else if (hctx->queue->nr_hw_queues == 1) |
70f4db63 | 841 | kblockd_schedule_delayed_work(&hctx->run_work, 0); |
e4043dcf JA |
842 | else { |
843 | unsigned int cpu; | |
844 | ||
506e931f | 845 | cpu = blk_mq_hctx_next_cpu(hctx); |
70f4db63 | 846 | kblockd_schedule_delayed_work_on(cpu, &hctx->run_work, 0); |
e4043dcf | 847 | } |
320ae51f JA |
848 | } |
849 | ||
850 | void blk_mq_run_queues(struct request_queue *q, bool async) | |
851 | { | |
852 | struct blk_mq_hw_ctx *hctx; | |
853 | int i; | |
854 | ||
855 | queue_for_each_hw_ctx(q, hctx, i) { | |
856 | if ((!blk_mq_hctx_has_pending(hctx) && | |
857 | list_empty_careful(&hctx->dispatch)) || | |
5d12f905 | 858 | test_bit(BLK_MQ_S_STOPPED, &hctx->state)) |
320ae51f JA |
859 | continue; |
860 | ||
e4043dcf | 861 | preempt_disable(); |
320ae51f | 862 | blk_mq_run_hw_queue(hctx, async); |
e4043dcf | 863 | preempt_enable(); |
320ae51f JA |
864 | } |
865 | } | |
866 | EXPORT_SYMBOL(blk_mq_run_queues); | |
867 | ||
868 | void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx) | |
869 | { | |
70f4db63 CH |
870 | cancel_delayed_work(&hctx->run_work); |
871 | cancel_delayed_work(&hctx->delay_work); | |
320ae51f JA |
872 | set_bit(BLK_MQ_S_STOPPED, &hctx->state); |
873 | } | |
874 | EXPORT_SYMBOL(blk_mq_stop_hw_queue); | |
875 | ||
280d45f6 CH |
876 | void blk_mq_stop_hw_queues(struct request_queue *q) |
877 | { | |
878 | struct blk_mq_hw_ctx *hctx; | |
879 | int i; | |
880 | ||
881 | queue_for_each_hw_ctx(q, hctx, i) | |
882 | blk_mq_stop_hw_queue(hctx); | |
883 | } | |
884 | EXPORT_SYMBOL(blk_mq_stop_hw_queues); | |
885 | ||
320ae51f JA |
886 | void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx) |
887 | { | |
888 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
e4043dcf JA |
889 | |
890 | preempt_disable(); | |
320ae51f | 891 | __blk_mq_run_hw_queue(hctx); |
e4043dcf | 892 | preempt_enable(); |
320ae51f JA |
893 | } |
894 | EXPORT_SYMBOL(blk_mq_start_hw_queue); | |
895 | ||
2f268556 CH |
896 | void blk_mq_start_hw_queues(struct request_queue *q) |
897 | { | |
898 | struct blk_mq_hw_ctx *hctx; | |
899 | int i; | |
900 | ||
901 | queue_for_each_hw_ctx(q, hctx, i) | |
902 | blk_mq_start_hw_queue(hctx); | |
903 | } | |
904 | EXPORT_SYMBOL(blk_mq_start_hw_queues); | |
905 | ||
906 | ||
1b4a3258 | 907 | void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async) |
320ae51f JA |
908 | { |
909 | struct blk_mq_hw_ctx *hctx; | |
910 | int i; | |
911 | ||
912 | queue_for_each_hw_ctx(q, hctx, i) { | |
913 | if (!test_bit(BLK_MQ_S_STOPPED, &hctx->state)) | |
914 | continue; | |
915 | ||
916 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
e4043dcf | 917 | preempt_disable(); |
1b4a3258 | 918 | blk_mq_run_hw_queue(hctx, async); |
e4043dcf | 919 | preempt_enable(); |
320ae51f JA |
920 | } |
921 | } | |
922 | EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues); | |
923 | ||
70f4db63 | 924 | static void blk_mq_run_work_fn(struct work_struct *work) |
320ae51f JA |
925 | { |
926 | struct blk_mq_hw_ctx *hctx; | |
927 | ||
70f4db63 | 928 | hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work); |
e4043dcf | 929 | |
320ae51f JA |
930 | __blk_mq_run_hw_queue(hctx); |
931 | } | |
932 | ||
70f4db63 CH |
933 | static void blk_mq_delay_work_fn(struct work_struct *work) |
934 | { | |
935 | struct blk_mq_hw_ctx *hctx; | |
936 | ||
937 | hctx = container_of(work, struct blk_mq_hw_ctx, delay_work.work); | |
938 | ||
939 | if (test_and_clear_bit(BLK_MQ_S_STOPPED, &hctx->state)) | |
940 | __blk_mq_run_hw_queue(hctx); | |
941 | } | |
942 | ||
943 | void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs) | |
944 | { | |
945 | unsigned long tmo = msecs_to_jiffies(msecs); | |
946 | ||
947 | if (hctx->queue->nr_hw_queues == 1) | |
948 | kblockd_schedule_delayed_work(&hctx->delay_work, tmo); | |
949 | else { | |
950 | unsigned int cpu; | |
951 | ||
506e931f | 952 | cpu = blk_mq_hctx_next_cpu(hctx); |
70f4db63 CH |
953 | kblockd_schedule_delayed_work_on(cpu, &hctx->delay_work, tmo); |
954 | } | |
955 | } | |
956 | EXPORT_SYMBOL(blk_mq_delay_queue); | |
957 | ||
320ae51f | 958 | static void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, |
72a0a36e | 959 | struct request *rq, bool at_head) |
320ae51f JA |
960 | { |
961 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
962 | ||
01b983c9 JA |
963 | trace_block_rq_insert(hctx->queue, rq); |
964 | ||
72a0a36e CH |
965 | if (at_head) |
966 | list_add(&rq->queuelist, &ctx->rq_list); | |
967 | else | |
968 | list_add_tail(&rq->queuelist, &ctx->rq_list); | |
4bb659b1 | 969 | |
320ae51f JA |
970 | blk_mq_hctx_mark_pending(hctx, ctx); |
971 | ||
972 | /* | |
973 | * We do this early, to ensure we are on the right CPU. | |
974 | */ | |
87ee7b11 | 975 | blk_add_timer(rq); |
320ae51f JA |
976 | } |
977 | ||
eeabc850 CH |
978 | void blk_mq_insert_request(struct request *rq, bool at_head, bool run_queue, |
979 | bool async) | |
320ae51f | 980 | { |
eeabc850 | 981 | struct request_queue *q = rq->q; |
320ae51f | 982 | struct blk_mq_hw_ctx *hctx; |
eeabc850 CH |
983 | struct blk_mq_ctx *ctx = rq->mq_ctx, *current_ctx; |
984 | ||
985 | current_ctx = blk_mq_get_ctx(q); | |
986 | if (!cpu_online(ctx->cpu)) | |
987 | rq->mq_ctx = ctx = current_ctx; | |
320ae51f | 988 | |
320ae51f JA |
989 | hctx = q->mq_ops->map_queue(q, ctx->cpu); |
990 | ||
eeabc850 CH |
991 | if (rq->cmd_flags & (REQ_FLUSH | REQ_FUA) && |
992 | !(rq->cmd_flags & (REQ_FLUSH_SEQ))) { | |
320ae51f JA |
993 | blk_insert_flush(rq); |
994 | } else { | |
320ae51f | 995 | spin_lock(&ctx->lock); |
72a0a36e | 996 | __blk_mq_insert_request(hctx, rq, at_head); |
320ae51f | 997 | spin_unlock(&ctx->lock); |
320ae51f JA |
998 | } |
999 | ||
320ae51f JA |
1000 | if (run_queue) |
1001 | blk_mq_run_hw_queue(hctx, async); | |
e4043dcf JA |
1002 | |
1003 | blk_mq_put_ctx(current_ctx); | |
320ae51f JA |
1004 | } |
1005 | ||
1006 | static void blk_mq_insert_requests(struct request_queue *q, | |
1007 | struct blk_mq_ctx *ctx, | |
1008 | struct list_head *list, | |
1009 | int depth, | |
1010 | bool from_schedule) | |
1011 | ||
1012 | { | |
1013 | struct blk_mq_hw_ctx *hctx; | |
1014 | struct blk_mq_ctx *current_ctx; | |
1015 | ||
1016 | trace_block_unplug(q, depth, !from_schedule); | |
1017 | ||
1018 | current_ctx = blk_mq_get_ctx(q); | |
1019 | ||
1020 | if (!cpu_online(ctx->cpu)) | |
1021 | ctx = current_ctx; | |
1022 | hctx = q->mq_ops->map_queue(q, ctx->cpu); | |
1023 | ||
1024 | /* | |
1025 | * preemption doesn't flush plug list, so it's possible ctx->cpu is | |
1026 | * offline now | |
1027 | */ | |
1028 | spin_lock(&ctx->lock); | |
1029 | while (!list_empty(list)) { | |
1030 | struct request *rq; | |
1031 | ||
1032 | rq = list_first_entry(list, struct request, queuelist); | |
1033 | list_del_init(&rq->queuelist); | |
1034 | rq->mq_ctx = ctx; | |
72a0a36e | 1035 | __blk_mq_insert_request(hctx, rq, false); |
320ae51f JA |
1036 | } |
1037 | spin_unlock(&ctx->lock); | |
1038 | ||
320ae51f | 1039 | blk_mq_run_hw_queue(hctx, from_schedule); |
e4043dcf | 1040 | blk_mq_put_ctx(current_ctx); |
320ae51f JA |
1041 | } |
1042 | ||
1043 | static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b) | |
1044 | { | |
1045 | struct request *rqa = container_of(a, struct request, queuelist); | |
1046 | struct request *rqb = container_of(b, struct request, queuelist); | |
1047 | ||
1048 | return !(rqa->mq_ctx < rqb->mq_ctx || | |
1049 | (rqa->mq_ctx == rqb->mq_ctx && | |
1050 | blk_rq_pos(rqa) < blk_rq_pos(rqb))); | |
1051 | } | |
1052 | ||
1053 | void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule) | |
1054 | { | |
1055 | struct blk_mq_ctx *this_ctx; | |
1056 | struct request_queue *this_q; | |
1057 | struct request *rq; | |
1058 | LIST_HEAD(list); | |
1059 | LIST_HEAD(ctx_list); | |
1060 | unsigned int depth; | |
1061 | ||
1062 | list_splice_init(&plug->mq_list, &list); | |
1063 | ||
1064 | list_sort(NULL, &list, plug_ctx_cmp); | |
1065 | ||
1066 | this_q = NULL; | |
1067 | this_ctx = NULL; | |
1068 | depth = 0; | |
1069 | ||
1070 | while (!list_empty(&list)) { | |
1071 | rq = list_entry_rq(list.next); | |
1072 | list_del_init(&rq->queuelist); | |
1073 | BUG_ON(!rq->q); | |
1074 | if (rq->mq_ctx != this_ctx) { | |
1075 | if (this_ctx) { | |
1076 | blk_mq_insert_requests(this_q, this_ctx, | |
1077 | &ctx_list, depth, | |
1078 | from_schedule); | |
1079 | } | |
1080 | ||
1081 | this_ctx = rq->mq_ctx; | |
1082 | this_q = rq->q; | |
1083 | depth = 0; | |
1084 | } | |
1085 | ||
1086 | depth++; | |
1087 | list_add_tail(&rq->queuelist, &ctx_list); | |
1088 | } | |
1089 | ||
1090 | /* | |
1091 | * If 'this_ctx' is set, we know we have entries to complete | |
1092 | * on 'ctx_list'. Do those. | |
1093 | */ | |
1094 | if (this_ctx) { | |
1095 | blk_mq_insert_requests(this_q, this_ctx, &ctx_list, depth, | |
1096 | from_schedule); | |
1097 | } | |
1098 | } | |
1099 | ||
1100 | static void blk_mq_bio_to_request(struct request *rq, struct bio *bio) | |
1101 | { | |
1102 | init_request_from_bio(rq, bio); | |
4b570521 JA |
1103 | |
1104 | if (blk_do_io_stat(rq)) { | |
1105 | rq->start_time = jiffies; | |
1106 | blk_account_io_start(rq, 1); | |
1107 | } | |
320ae51f JA |
1108 | } |
1109 | ||
07068d5b JA |
1110 | static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx *hctx, |
1111 | struct blk_mq_ctx *ctx, | |
1112 | struct request *rq, struct bio *bio) | |
320ae51f | 1113 | { |
07068d5b | 1114 | struct request_queue *q = hctx->queue; |
320ae51f | 1115 | |
07068d5b JA |
1116 | if (!(hctx->flags & BLK_MQ_F_SHOULD_MERGE)) { |
1117 | blk_mq_bio_to_request(rq, bio); | |
1118 | spin_lock(&ctx->lock); | |
1119 | insert_rq: | |
1120 | __blk_mq_insert_request(hctx, rq, false); | |
1121 | spin_unlock(&ctx->lock); | |
1122 | return false; | |
1123 | } else { | |
1124 | spin_lock(&ctx->lock); | |
1125 | if (!blk_mq_attempt_merge(q, ctx, bio)) { | |
1126 | blk_mq_bio_to_request(rq, bio); | |
1127 | goto insert_rq; | |
1128 | } | |
320ae51f | 1129 | |
07068d5b JA |
1130 | spin_unlock(&ctx->lock); |
1131 | __blk_mq_free_request(hctx, ctx, rq); | |
1132 | return true; | |
14ec77f3 | 1133 | } |
07068d5b | 1134 | } |
14ec77f3 | 1135 | |
07068d5b JA |
1136 | struct blk_map_ctx { |
1137 | struct blk_mq_hw_ctx *hctx; | |
1138 | struct blk_mq_ctx *ctx; | |
1139 | }; | |
1140 | ||
1141 | static struct request *blk_mq_map_request(struct request_queue *q, | |
1142 | struct bio *bio, | |
1143 | struct blk_map_ctx *data) | |
1144 | { | |
1145 | struct blk_mq_hw_ctx *hctx; | |
1146 | struct blk_mq_ctx *ctx; | |
1147 | struct request *rq; | |
1148 | int rw = bio_data_dir(bio); | |
320ae51f | 1149 | |
07068d5b | 1150 | if (unlikely(blk_mq_queue_enter(q))) { |
320ae51f | 1151 | bio_endio(bio, -EIO); |
07068d5b | 1152 | return NULL; |
320ae51f JA |
1153 | } |
1154 | ||
1155 | ctx = blk_mq_get_ctx(q); | |
1156 | hctx = q->mq_ops->map_queue(q, ctx->cpu); | |
1157 | ||
07068d5b | 1158 | if (rw_is_sync(bio->bi_rw)) |
27fbf4e8 | 1159 | rw |= REQ_SYNC; |
07068d5b | 1160 | |
320ae51f | 1161 | trace_block_getrq(q, bio, rw); |
5dee8577 CH |
1162 | rq = __blk_mq_alloc_request(q, hctx, ctx, rw, GFP_ATOMIC, false); |
1163 | if (unlikely(!rq)) { | |
793597a6 | 1164 | __blk_mq_run_hw_queue(hctx); |
320ae51f JA |
1165 | blk_mq_put_ctx(ctx); |
1166 | trace_block_sleeprq(q, bio, rw); | |
793597a6 CH |
1167 | |
1168 | ctx = blk_mq_get_ctx(q); | |
320ae51f | 1169 | hctx = q->mq_ops->map_queue(q, ctx->cpu); |
793597a6 CH |
1170 | rq = __blk_mq_alloc_request(q, hctx, ctx, rw, |
1171 | __GFP_WAIT|GFP_ATOMIC, false); | |
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); | |
1214 | ||
1215 | /* | |
1216 | * For OK queue, we are done. For error, kill it. Any other | |
1217 | * error (busy), just add it to our list as we previously | |
1218 | * would have done | |
1219 | */ | |
1220 | ret = q->mq_ops->queue_rq(data.hctx, rq); | |
1221 | if (ret == BLK_MQ_RQ_QUEUE_OK) | |
1222 | goto done; | |
1223 | else { | |
1224 | __blk_mq_requeue_request(rq); | |
1225 | ||
1226 | if (ret == BLK_MQ_RQ_QUEUE_ERROR) { | |
1227 | rq->errors = -EIO; | |
1228 | blk_mq_end_io(rq, rq->errors); | |
1229 | goto done; | |
1230 | } | |
1231 | } | |
1232 | } | |
1233 | ||
1234 | if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) { | |
1235 | /* | |
1236 | * For a SYNC request, send it to the hardware immediately. For | |
1237 | * an ASYNC request, just ensure that we run it later on. The | |
1238 | * latter allows for merging opportunities and more efficient | |
1239 | * dispatching. | |
1240 | */ | |
1241 | run_queue: | |
1242 | blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua); | |
1243 | } | |
1244 | done: | |
1245 | blk_mq_put_ctx(data.ctx); | |
1246 | } | |
1247 | ||
1248 | /* | |
1249 | * Single hardware queue variant. This will attempt to use any per-process | |
1250 | * plug for merging and IO deferral. | |
1251 | */ | |
1252 | static void blk_sq_make_request(struct request_queue *q, struct bio *bio) | |
1253 | { | |
1254 | const int is_sync = rw_is_sync(bio->bi_rw); | |
1255 | const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA); | |
1256 | unsigned int use_plug, request_count = 0; | |
1257 | struct blk_map_ctx data; | |
1258 | struct request *rq; | |
1259 | ||
1260 | /* | |
1261 | * If we have multiple hardware queues, just go directly to | |
1262 | * one of those for sync IO. | |
1263 | */ | |
1264 | use_plug = !is_flush_fua && !is_sync; | |
1265 | ||
1266 | blk_queue_bounce(q, &bio); | |
1267 | ||
1268 | if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) { | |
1269 | bio_endio(bio, -EIO); | |
1270 | return; | |
1271 | } | |
1272 | ||
1273 | if (use_plug && !blk_queue_nomerges(q) && | |
1274 | blk_attempt_plug_merge(q, bio, &request_count)) | |
1275 | return; | |
1276 | ||
1277 | rq = blk_mq_map_request(q, bio, &data); | |
320ae51f JA |
1278 | |
1279 | if (unlikely(is_flush_fua)) { | |
1280 | blk_mq_bio_to_request(rq, bio); | |
320ae51f JA |
1281 | blk_insert_flush(rq); |
1282 | goto run_queue; | |
1283 | } | |
1284 | ||
1285 | /* | |
1286 | * A task plug currently exists. Since this is completely lockless, | |
1287 | * utilize that to temporarily store requests until the task is | |
1288 | * either done or scheduled away. | |
1289 | */ | |
1290 | if (use_plug) { | |
1291 | struct blk_plug *plug = current->plug; | |
1292 | ||
1293 | if (plug) { | |
1294 | blk_mq_bio_to_request(rq, bio); | |
92f399c7 | 1295 | if (list_empty(&plug->mq_list)) |
320ae51f JA |
1296 | trace_block_plug(q); |
1297 | else if (request_count >= BLK_MAX_REQUEST_COUNT) { | |
1298 | blk_flush_plug_list(plug, false); | |
1299 | trace_block_plug(q); | |
1300 | } | |
1301 | list_add_tail(&rq->queuelist, &plug->mq_list); | |
07068d5b | 1302 | blk_mq_put_ctx(data.ctx); |
320ae51f JA |
1303 | return; |
1304 | } | |
1305 | } | |
1306 | ||
07068d5b JA |
1307 | if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) { |
1308 | /* | |
1309 | * For a SYNC request, send it to the hardware immediately. For | |
1310 | * an ASYNC request, just ensure that we run it later on. The | |
1311 | * latter allows for merging opportunities and more efficient | |
1312 | * dispatching. | |
1313 | */ | |
1314 | run_queue: | |
1315 | blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua); | |
320ae51f JA |
1316 | } |
1317 | ||
07068d5b | 1318 | blk_mq_put_ctx(data.ctx); |
320ae51f JA |
1319 | } |
1320 | ||
1321 | /* | |
1322 | * Default mapping to a software queue, since we use one per CPU. | |
1323 | */ | |
1324 | struct blk_mq_hw_ctx *blk_mq_map_queue(struct request_queue *q, const int cpu) | |
1325 | { | |
1326 | return q->queue_hw_ctx[q->mq_map[cpu]]; | |
1327 | } | |
1328 | EXPORT_SYMBOL(blk_mq_map_queue); | |
1329 | ||
24d2f903 CH |
1330 | static void blk_mq_free_rq_map(struct blk_mq_tag_set *set, |
1331 | struct blk_mq_tags *tags, unsigned int hctx_idx) | |
95363efd | 1332 | { |
e9b267d9 | 1333 | struct page *page; |
320ae51f | 1334 | |
24d2f903 | 1335 | if (tags->rqs && set->ops->exit_request) { |
e9b267d9 | 1336 | int i; |
320ae51f | 1337 | |
24d2f903 CH |
1338 | for (i = 0; i < tags->nr_tags; i++) { |
1339 | if (!tags->rqs[i]) | |
e9b267d9 | 1340 | continue; |
24d2f903 CH |
1341 | set->ops->exit_request(set->driver_data, tags->rqs[i], |
1342 | hctx_idx, i); | |
e9b267d9 | 1343 | } |
320ae51f | 1344 | } |
320ae51f | 1345 | |
24d2f903 CH |
1346 | while (!list_empty(&tags->page_list)) { |
1347 | page = list_first_entry(&tags->page_list, struct page, lru); | |
6753471c | 1348 | list_del_init(&page->lru); |
320ae51f JA |
1349 | __free_pages(page, page->private); |
1350 | } | |
1351 | ||
24d2f903 | 1352 | kfree(tags->rqs); |
320ae51f | 1353 | |
24d2f903 | 1354 | blk_mq_free_tags(tags); |
320ae51f JA |
1355 | } |
1356 | ||
1357 | static size_t order_to_size(unsigned int order) | |
1358 | { | |
4ca08500 | 1359 | return (size_t)PAGE_SIZE << order; |
320ae51f JA |
1360 | } |
1361 | ||
24d2f903 CH |
1362 | static struct blk_mq_tags *blk_mq_init_rq_map(struct blk_mq_tag_set *set, |
1363 | unsigned int hctx_idx) | |
320ae51f | 1364 | { |
24d2f903 | 1365 | struct blk_mq_tags *tags; |
320ae51f JA |
1366 | unsigned int i, j, entries_per_page, max_order = 4; |
1367 | size_t rq_size, left; | |
1368 | ||
24d2f903 CH |
1369 | tags = blk_mq_init_tags(set->queue_depth, set->reserved_tags, |
1370 | set->numa_node); | |
1371 | if (!tags) | |
1372 | return NULL; | |
320ae51f | 1373 | |
24d2f903 CH |
1374 | INIT_LIST_HEAD(&tags->page_list); |
1375 | ||
1376 | tags->rqs = kmalloc_node(set->queue_depth * sizeof(struct request *), | |
1377 | GFP_KERNEL, set->numa_node); | |
1378 | if (!tags->rqs) { | |
1379 | blk_mq_free_tags(tags); | |
1380 | return NULL; | |
1381 | } | |
320ae51f JA |
1382 | |
1383 | /* | |
1384 | * rq_size is the size of the request plus driver payload, rounded | |
1385 | * to the cacheline size | |
1386 | */ | |
24d2f903 | 1387 | rq_size = round_up(sizeof(struct request) + set->cmd_size, |
320ae51f | 1388 | cache_line_size()); |
24d2f903 | 1389 | left = rq_size * set->queue_depth; |
320ae51f | 1390 | |
24d2f903 | 1391 | for (i = 0; i < set->queue_depth; ) { |
320ae51f JA |
1392 | int this_order = max_order; |
1393 | struct page *page; | |
1394 | int to_do; | |
1395 | void *p; | |
1396 | ||
1397 | while (left < order_to_size(this_order - 1) && this_order) | |
1398 | this_order--; | |
1399 | ||
1400 | do { | |
24d2f903 CH |
1401 | page = alloc_pages_node(set->numa_node, GFP_KERNEL, |
1402 | this_order); | |
320ae51f JA |
1403 | if (page) |
1404 | break; | |
1405 | if (!this_order--) | |
1406 | break; | |
1407 | if (order_to_size(this_order) < rq_size) | |
1408 | break; | |
1409 | } while (1); | |
1410 | ||
1411 | if (!page) | |
24d2f903 | 1412 | goto fail; |
320ae51f JA |
1413 | |
1414 | page->private = this_order; | |
24d2f903 | 1415 | list_add_tail(&page->lru, &tags->page_list); |
320ae51f JA |
1416 | |
1417 | p = page_address(page); | |
1418 | entries_per_page = order_to_size(this_order) / rq_size; | |
24d2f903 | 1419 | to_do = min(entries_per_page, set->queue_depth - i); |
320ae51f JA |
1420 | left -= to_do * rq_size; |
1421 | for (j = 0; j < to_do; j++) { | |
24d2f903 CH |
1422 | tags->rqs[i] = p; |
1423 | if (set->ops->init_request) { | |
1424 | if (set->ops->init_request(set->driver_data, | |
1425 | tags->rqs[i], hctx_idx, i, | |
1426 | set->numa_node)) | |
1427 | goto fail; | |
e9b267d9 CH |
1428 | } |
1429 | ||
320ae51f JA |
1430 | p += rq_size; |
1431 | i++; | |
1432 | } | |
1433 | } | |
1434 | ||
24d2f903 | 1435 | return tags; |
320ae51f | 1436 | |
24d2f903 CH |
1437 | fail: |
1438 | pr_warn("%s: failed to allocate requests\n", __func__); | |
1439 | blk_mq_free_rq_map(set, tags, hctx_idx); | |
1440 | return NULL; | |
320ae51f JA |
1441 | } |
1442 | ||
1429d7c9 JA |
1443 | static void blk_mq_free_bitmap(struct blk_mq_ctxmap *bitmap) |
1444 | { | |
1445 | kfree(bitmap->map); | |
1446 | } | |
1447 | ||
1448 | static int blk_mq_alloc_bitmap(struct blk_mq_ctxmap *bitmap, int node) | |
1449 | { | |
1450 | unsigned int bpw = 8, total, num_maps, i; | |
1451 | ||
1452 | bitmap->bits_per_word = bpw; | |
1453 | ||
1454 | num_maps = ALIGN(nr_cpu_ids, bpw) / bpw; | |
1455 | bitmap->map = kzalloc_node(num_maps * sizeof(struct blk_align_bitmap), | |
1456 | GFP_KERNEL, node); | |
1457 | if (!bitmap->map) | |
1458 | return -ENOMEM; | |
1459 | ||
1460 | bitmap->map_size = num_maps; | |
1461 | ||
1462 | total = nr_cpu_ids; | |
1463 | for (i = 0; i < num_maps; i++) { | |
1464 | bitmap->map[i].depth = min(total, bitmap->bits_per_word); | |
1465 | total -= bitmap->map[i].depth; | |
1466 | } | |
1467 | ||
1468 | return 0; | |
1469 | } | |
1470 | ||
484b4061 JA |
1471 | static int blk_mq_hctx_cpu_offline(struct blk_mq_hw_ctx *hctx, int cpu) |
1472 | { | |
1473 | struct request_queue *q = hctx->queue; | |
1474 | struct blk_mq_ctx *ctx; | |
1475 | LIST_HEAD(tmp); | |
1476 | ||
1477 | /* | |
1478 | * Move ctx entries to new CPU, if this one is going away. | |
1479 | */ | |
1480 | ctx = __blk_mq_get_ctx(q, cpu); | |
1481 | ||
1482 | spin_lock(&ctx->lock); | |
1483 | if (!list_empty(&ctx->rq_list)) { | |
1484 | list_splice_init(&ctx->rq_list, &tmp); | |
1485 | blk_mq_hctx_clear_pending(hctx, ctx); | |
1486 | } | |
1487 | spin_unlock(&ctx->lock); | |
1488 | ||
1489 | if (list_empty(&tmp)) | |
1490 | return NOTIFY_OK; | |
1491 | ||
1492 | ctx = blk_mq_get_ctx(q); | |
1493 | spin_lock(&ctx->lock); | |
1494 | ||
1495 | while (!list_empty(&tmp)) { | |
1496 | struct request *rq; | |
1497 | ||
1498 | rq = list_first_entry(&tmp, struct request, queuelist); | |
1499 | rq->mq_ctx = ctx; | |
1500 | list_move_tail(&rq->queuelist, &ctx->rq_list); | |
1501 | } | |
1502 | ||
1503 | hctx = q->mq_ops->map_queue(q, ctx->cpu); | |
1504 | blk_mq_hctx_mark_pending(hctx, ctx); | |
1505 | ||
1506 | spin_unlock(&ctx->lock); | |
1507 | ||
1508 | blk_mq_run_hw_queue(hctx, true); | |
1509 | blk_mq_put_ctx(ctx); | |
1510 | return NOTIFY_OK; | |
1511 | } | |
1512 | ||
1513 | static int blk_mq_hctx_cpu_online(struct blk_mq_hw_ctx *hctx, int cpu) | |
1514 | { | |
1515 | struct request_queue *q = hctx->queue; | |
1516 | struct blk_mq_tag_set *set = q->tag_set; | |
1517 | ||
1518 | if (set->tags[hctx->queue_num]) | |
1519 | return NOTIFY_OK; | |
1520 | ||
1521 | set->tags[hctx->queue_num] = blk_mq_init_rq_map(set, hctx->queue_num); | |
1522 | if (!set->tags[hctx->queue_num]) | |
1523 | return NOTIFY_STOP; | |
1524 | ||
1525 | hctx->tags = set->tags[hctx->queue_num]; | |
1526 | return NOTIFY_OK; | |
1527 | } | |
1528 | ||
1529 | static int blk_mq_hctx_notify(void *data, unsigned long action, | |
1530 | unsigned int cpu) | |
1531 | { | |
1532 | struct blk_mq_hw_ctx *hctx = data; | |
1533 | ||
1534 | if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) | |
1535 | return blk_mq_hctx_cpu_offline(hctx, cpu); | |
1536 | else if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) | |
1537 | return blk_mq_hctx_cpu_online(hctx, cpu); | |
1538 | ||
1539 | return NOTIFY_OK; | |
1540 | } | |
1541 | ||
624dbe47 ML |
1542 | static void blk_mq_exit_hw_queues(struct request_queue *q, |
1543 | struct blk_mq_tag_set *set, int nr_queue) | |
1544 | { | |
1545 | struct blk_mq_hw_ctx *hctx; | |
1546 | unsigned int i; | |
1547 | ||
1548 | queue_for_each_hw_ctx(q, hctx, i) { | |
1549 | if (i == nr_queue) | |
1550 | break; | |
1551 | ||
1552 | if (set->ops->exit_hctx) | |
1553 | set->ops->exit_hctx(hctx, i); | |
1554 | ||
1555 | blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier); | |
1556 | kfree(hctx->ctxs); | |
1557 | blk_mq_free_bitmap(&hctx->ctx_map); | |
1558 | } | |
1559 | ||
1560 | } | |
1561 | ||
1562 | static void blk_mq_free_hw_queues(struct request_queue *q, | |
1563 | struct blk_mq_tag_set *set) | |
1564 | { | |
1565 | struct blk_mq_hw_ctx *hctx; | |
1566 | unsigned int i; | |
1567 | ||
1568 | queue_for_each_hw_ctx(q, hctx, i) { | |
1569 | free_cpumask_var(hctx->cpumask); | |
cdef54dd | 1570 | kfree(hctx); |
624dbe47 ML |
1571 | } |
1572 | } | |
1573 | ||
320ae51f | 1574 | static int blk_mq_init_hw_queues(struct request_queue *q, |
24d2f903 | 1575 | struct blk_mq_tag_set *set) |
320ae51f JA |
1576 | { |
1577 | struct blk_mq_hw_ctx *hctx; | |
624dbe47 | 1578 | unsigned int i; |
320ae51f JA |
1579 | |
1580 | /* | |
1581 | * Initialize hardware queues | |
1582 | */ | |
1583 | queue_for_each_hw_ctx(q, hctx, i) { | |
320ae51f JA |
1584 | int node; |
1585 | ||
1586 | node = hctx->numa_node; | |
1587 | if (node == NUMA_NO_NODE) | |
24d2f903 | 1588 | node = hctx->numa_node = set->numa_node; |
320ae51f | 1589 | |
70f4db63 CH |
1590 | INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn); |
1591 | INIT_DELAYED_WORK(&hctx->delay_work, blk_mq_delay_work_fn); | |
320ae51f JA |
1592 | spin_lock_init(&hctx->lock); |
1593 | INIT_LIST_HEAD(&hctx->dispatch); | |
1594 | hctx->queue = q; | |
1595 | hctx->queue_num = i; | |
24d2f903 CH |
1596 | hctx->flags = set->flags; |
1597 | hctx->cmd_size = set->cmd_size; | |
320ae51f JA |
1598 | |
1599 | blk_mq_init_cpu_notifier(&hctx->cpu_notifier, | |
1600 | blk_mq_hctx_notify, hctx); | |
1601 | blk_mq_register_cpu_notifier(&hctx->cpu_notifier); | |
1602 | ||
24d2f903 | 1603 | hctx->tags = set->tags[i]; |
320ae51f JA |
1604 | |
1605 | /* | |
1606 | * Allocate space for all possible cpus to avoid allocation in | |
1607 | * runtime | |
1608 | */ | |
1609 | hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *), | |
1610 | GFP_KERNEL, node); | |
1611 | if (!hctx->ctxs) | |
1612 | break; | |
1613 | ||
1429d7c9 | 1614 | if (blk_mq_alloc_bitmap(&hctx->ctx_map, node)) |
320ae51f JA |
1615 | break; |
1616 | ||
320ae51f JA |
1617 | hctx->nr_ctx = 0; |
1618 | ||
24d2f903 CH |
1619 | if (set->ops->init_hctx && |
1620 | set->ops->init_hctx(hctx, set->driver_data, i)) | |
320ae51f JA |
1621 | break; |
1622 | } | |
1623 | ||
1624 | if (i == q->nr_hw_queues) | |
1625 | return 0; | |
1626 | ||
1627 | /* | |
1628 | * Init failed | |
1629 | */ | |
624dbe47 | 1630 | blk_mq_exit_hw_queues(q, set, i); |
320ae51f JA |
1631 | |
1632 | return 1; | |
1633 | } | |
1634 | ||
1635 | static void blk_mq_init_cpu_queues(struct request_queue *q, | |
1636 | unsigned int nr_hw_queues) | |
1637 | { | |
1638 | unsigned int i; | |
1639 | ||
1640 | for_each_possible_cpu(i) { | |
1641 | struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i); | |
1642 | struct blk_mq_hw_ctx *hctx; | |
1643 | ||
1644 | memset(__ctx, 0, sizeof(*__ctx)); | |
1645 | __ctx->cpu = i; | |
1646 | spin_lock_init(&__ctx->lock); | |
1647 | INIT_LIST_HEAD(&__ctx->rq_list); | |
1648 | __ctx->queue = q; | |
1649 | ||
1650 | /* If the cpu isn't online, the cpu is mapped to first hctx */ | |
320ae51f JA |
1651 | if (!cpu_online(i)) |
1652 | continue; | |
1653 | ||
e4043dcf JA |
1654 | hctx = q->mq_ops->map_queue(q, i); |
1655 | cpumask_set_cpu(i, hctx->cpumask); | |
1656 | hctx->nr_ctx++; | |
1657 | ||
320ae51f JA |
1658 | /* |
1659 | * Set local node, IFF we have more than one hw queue. If | |
1660 | * not, we remain on the home node of the device | |
1661 | */ | |
1662 | if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE) | |
1663 | hctx->numa_node = cpu_to_node(i); | |
1664 | } | |
1665 | } | |
1666 | ||
1667 | static void blk_mq_map_swqueue(struct request_queue *q) | |
1668 | { | |
1669 | unsigned int i; | |
1670 | struct blk_mq_hw_ctx *hctx; | |
1671 | struct blk_mq_ctx *ctx; | |
1672 | ||
1673 | queue_for_each_hw_ctx(q, hctx, i) { | |
e4043dcf | 1674 | cpumask_clear(hctx->cpumask); |
320ae51f JA |
1675 | hctx->nr_ctx = 0; |
1676 | } | |
1677 | ||
1678 | /* | |
1679 | * Map software to hardware queues | |
1680 | */ | |
1681 | queue_for_each_ctx(q, ctx, i) { | |
1682 | /* If the cpu isn't online, the cpu is mapped to first hctx */ | |
e4043dcf JA |
1683 | if (!cpu_online(i)) |
1684 | continue; | |
1685 | ||
320ae51f | 1686 | hctx = q->mq_ops->map_queue(q, i); |
e4043dcf | 1687 | cpumask_set_cpu(i, hctx->cpumask); |
320ae51f JA |
1688 | ctx->index_hw = hctx->nr_ctx; |
1689 | hctx->ctxs[hctx->nr_ctx++] = ctx; | |
1690 | } | |
506e931f JA |
1691 | |
1692 | queue_for_each_hw_ctx(q, hctx, i) { | |
484b4061 JA |
1693 | /* |
1694 | * If not software queues are mapped to this hardware queue, | |
1695 | * disable it and free the request entries | |
1696 | */ | |
1697 | if (!hctx->nr_ctx) { | |
1698 | struct blk_mq_tag_set *set = q->tag_set; | |
1699 | ||
1700 | if (set->tags[i]) { | |
1701 | blk_mq_free_rq_map(set, set->tags[i], i); | |
1702 | set->tags[i] = NULL; | |
1703 | hctx->tags = NULL; | |
1704 | } | |
1705 | continue; | |
1706 | } | |
1707 | ||
1708 | /* | |
1709 | * Initialize batch roundrobin counts | |
1710 | */ | |
506e931f JA |
1711 | hctx->next_cpu = cpumask_first(hctx->cpumask); |
1712 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; | |
1713 | } | |
320ae51f JA |
1714 | } |
1715 | ||
0d2602ca JA |
1716 | static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set) |
1717 | { | |
1718 | struct blk_mq_hw_ctx *hctx; | |
1719 | struct request_queue *q; | |
1720 | bool shared; | |
1721 | int i; | |
1722 | ||
1723 | if (set->tag_list.next == set->tag_list.prev) | |
1724 | shared = false; | |
1725 | else | |
1726 | shared = true; | |
1727 | ||
1728 | list_for_each_entry(q, &set->tag_list, tag_set_list) { | |
1729 | blk_mq_freeze_queue(q); | |
1730 | ||
1731 | queue_for_each_hw_ctx(q, hctx, i) { | |
1732 | if (shared) | |
1733 | hctx->flags |= BLK_MQ_F_TAG_SHARED; | |
1734 | else | |
1735 | hctx->flags &= ~BLK_MQ_F_TAG_SHARED; | |
1736 | } | |
1737 | blk_mq_unfreeze_queue(q); | |
1738 | } | |
1739 | } | |
1740 | ||
1741 | static void blk_mq_del_queue_tag_set(struct request_queue *q) | |
1742 | { | |
1743 | struct blk_mq_tag_set *set = q->tag_set; | |
1744 | ||
1745 | blk_mq_freeze_queue(q); | |
1746 | ||
1747 | mutex_lock(&set->tag_list_lock); | |
1748 | list_del_init(&q->tag_set_list); | |
1749 | blk_mq_update_tag_set_depth(set); | |
1750 | mutex_unlock(&set->tag_list_lock); | |
1751 | ||
1752 | blk_mq_unfreeze_queue(q); | |
1753 | } | |
1754 | ||
1755 | static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set, | |
1756 | struct request_queue *q) | |
1757 | { | |
1758 | q->tag_set = set; | |
1759 | ||
1760 | mutex_lock(&set->tag_list_lock); | |
1761 | list_add_tail(&q->tag_set_list, &set->tag_list); | |
1762 | blk_mq_update_tag_set_depth(set); | |
1763 | mutex_unlock(&set->tag_list_lock); | |
1764 | } | |
1765 | ||
24d2f903 | 1766 | struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set) |
320ae51f JA |
1767 | { |
1768 | struct blk_mq_hw_ctx **hctxs; | |
1769 | struct blk_mq_ctx *ctx; | |
1770 | struct request_queue *q; | |
f14bbe77 | 1771 | unsigned int *map; |
320ae51f JA |
1772 | int i; |
1773 | ||
320ae51f JA |
1774 | ctx = alloc_percpu(struct blk_mq_ctx); |
1775 | if (!ctx) | |
1776 | return ERR_PTR(-ENOMEM); | |
1777 | ||
24d2f903 CH |
1778 | hctxs = kmalloc_node(set->nr_hw_queues * sizeof(*hctxs), GFP_KERNEL, |
1779 | set->numa_node); | |
320ae51f JA |
1780 | |
1781 | if (!hctxs) | |
1782 | goto err_percpu; | |
1783 | ||
f14bbe77 JA |
1784 | map = blk_mq_make_queue_map(set); |
1785 | if (!map) | |
1786 | goto err_map; | |
1787 | ||
24d2f903 | 1788 | for (i = 0; i < set->nr_hw_queues; i++) { |
f14bbe77 JA |
1789 | int node = blk_mq_hw_queue_to_node(map, i); |
1790 | ||
cdef54dd CH |
1791 | hctxs[i] = kzalloc_node(sizeof(struct blk_mq_hw_ctx), |
1792 | GFP_KERNEL, node); | |
320ae51f JA |
1793 | if (!hctxs[i]) |
1794 | goto err_hctxs; | |
1795 | ||
e4043dcf JA |
1796 | if (!zalloc_cpumask_var(&hctxs[i]->cpumask, GFP_KERNEL)) |
1797 | goto err_hctxs; | |
1798 | ||
0d2602ca | 1799 | atomic_set(&hctxs[i]->nr_active, 0); |
f14bbe77 | 1800 | hctxs[i]->numa_node = node; |
320ae51f JA |
1801 | hctxs[i]->queue_num = i; |
1802 | } | |
1803 | ||
24d2f903 | 1804 | q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node); |
320ae51f JA |
1805 | if (!q) |
1806 | goto err_hctxs; | |
1807 | ||
3d2936f4 ML |
1808 | if (percpu_counter_init(&q->mq_usage_counter, 0)) |
1809 | goto err_map; | |
1810 | ||
320ae51f JA |
1811 | setup_timer(&q->timeout, blk_mq_rq_timer, (unsigned long) q); |
1812 | blk_queue_rq_timeout(q, 30000); | |
1813 | ||
1814 | q->nr_queues = nr_cpu_ids; | |
24d2f903 | 1815 | q->nr_hw_queues = set->nr_hw_queues; |
f14bbe77 | 1816 | q->mq_map = map; |
320ae51f JA |
1817 | |
1818 | q->queue_ctx = ctx; | |
1819 | q->queue_hw_ctx = hctxs; | |
1820 | ||
24d2f903 | 1821 | q->mq_ops = set->ops; |
94eddfbe | 1822 | q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT; |
320ae51f | 1823 | |
05f1dd53 JA |
1824 | if (!(set->flags & BLK_MQ_F_SG_MERGE)) |
1825 | q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE; | |
1826 | ||
1be036e9 CH |
1827 | q->sg_reserved_size = INT_MAX; |
1828 | ||
6fca6a61 CH |
1829 | INIT_WORK(&q->requeue_work, blk_mq_requeue_work); |
1830 | INIT_LIST_HEAD(&q->requeue_list); | |
1831 | spin_lock_init(&q->requeue_lock); | |
1832 | ||
07068d5b JA |
1833 | if (q->nr_hw_queues > 1) |
1834 | blk_queue_make_request(q, blk_mq_make_request); | |
1835 | else | |
1836 | blk_queue_make_request(q, blk_sq_make_request); | |
1837 | ||
87ee7b11 | 1838 | blk_queue_rq_timed_out(q, blk_mq_rq_timed_out); |
24d2f903 CH |
1839 | if (set->timeout) |
1840 | blk_queue_rq_timeout(q, set->timeout); | |
320ae51f | 1841 | |
eba71768 JA |
1842 | /* |
1843 | * Do this after blk_queue_make_request() overrides it... | |
1844 | */ | |
1845 | q->nr_requests = set->queue_depth; | |
1846 | ||
24d2f903 CH |
1847 | if (set->ops->complete) |
1848 | blk_queue_softirq_done(q, set->ops->complete); | |
30a91cb4 | 1849 | |
320ae51f | 1850 | blk_mq_init_flush(q); |
24d2f903 | 1851 | blk_mq_init_cpu_queues(q, set->nr_hw_queues); |
320ae51f | 1852 | |
24d2f903 CH |
1853 | q->flush_rq = kzalloc(round_up(sizeof(struct request) + |
1854 | set->cmd_size, cache_line_size()), | |
1855 | GFP_KERNEL); | |
18741986 | 1856 | if (!q->flush_rq) |
320ae51f JA |
1857 | goto err_hw; |
1858 | ||
24d2f903 | 1859 | if (blk_mq_init_hw_queues(q, set)) |
18741986 CH |
1860 | goto err_flush_rq; |
1861 | ||
320ae51f JA |
1862 | mutex_lock(&all_q_mutex); |
1863 | list_add_tail(&q->all_q_node, &all_q_list); | |
1864 | mutex_unlock(&all_q_mutex); | |
1865 | ||
0d2602ca JA |
1866 | blk_mq_add_queue_tag_set(set, q); |
1867 | ||
484b4061 JA |
1868 | blk_mq_map_swqueue(q); |
1869 | ||
320ae51f | 1870 | return q; |
18741986 CH |
1871 | |
1872 | err_flush_rq: | |
1873 | kfree(q->flush_rq); | |
320ae51f | 1874 | err_hw: |
320ae51f JA |
1875 | blk_cleanup_queue(q); |
1876 | err_hctxs: | |
f14bbe77 | 1877 | kfree(map); |
24d2f903 | 1878 | for (i = 0; i < set->nr_hw_queues; i++) { |
320ae51f JA |
1879 | if (!hctxs[i]) |
1880 | break; | |
e4043dcf | 1881 | free_cpumask_var(hctxs[i]->cpumask); |
cdef54dd | 1882 | kfree(hctxs[i]); |
320ae51f | 1883 | } |
f14bbe77 | 1884 | err_map: |
320ae51f JA |
1885 | kfree(hctxs); |
1886 | err_percpu: | |
1887 | free_percpu(ctx); | |
1888 | return ERR_PTR(-ENOMEM); | |
1889 | } | |
1890 | EXPORT_SYMBOL(blk_mq_init_queue); | |
1891 | ||
1892 | void blk_mq_free_queue(struct request_queue *q) | |
1893 | { | |
624dbe47 | 1894 | struct blk_mq_tag_set *set = q->tag_set; |
320ae51f | 1895 | |
0d2602ca JA |
1896 | blk_mq_del_queue_tag_set(q); |
1897 | ||
624dbe47 ML |
1898 | blk_mq_exit_hw_queues(q, set, set->nr_hw_queues); |
1899 | blk_mq_free_hw_queues(q, set); | |
320ae51f | 1900 | |
3d2936f4 ML |
1901 | percpu_counter_destroy(&q->mq_usage_counter); |
1902 | ||
320ae51f JA |
1903 | free_percpu(q->queue_ctx); |
1904 | kfree(q->queue_hw_ctx); | |
1905 | kfree(q->mq_map); | |
1906 | ||
1907 | q->queue_ctx = NULL; | |
1908 | q->queue_hw_ctx = NULL; | |
1909 | q->mq_map = NULL; | |
1910 | ||
1911 | mutex_lock(&all_q_mutex); | |
1912 | list_del_init(&q->all_q_node); | |
1913 | mutex_unlock(&all_q_mutex); | |
1914 | } | |
320ae51f JA |
1915 | |
1916 | /* Basically redo blk_mq_init_queue with queue frozen */ | |
f618ef7c | 1917 | static void blk_mq_queue_reinit(struct request_queue *q) |
320ae51f JA |
1918 | { |
1919 | blk_mq_freeze_queue(q); | |
1920 | ||
1921 | blk_mq_update_queue_map(q->mq_map, q->nr_hw_queues); | |
1922 | ||
1923 | /* | |
1924 | * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe | |
1925 | * we should change hctx numa_node according to new topology (this | |
1926 | * involves free and re-allocate memory, worthy doing?) | |
1927 | */ | |
1928 | ||
1929 | blk_mq_map_swqueue(q); | |
1930 | ||
1931 | blk_mq_unfreeze_queue(q); | |
1932 | } | |
1933 | ||
f618ef7c PG |
1934 | static int blk_mq_queue_reinit_notify(struct notifier_block *nb, |
1935 | unsigned long action, void *hcpu) | |
320ae51f JA |
1936 | { |
1937 | struct request_queue *q; | |
1938 | ||
1939 | /* | |
9fccfed8 JA |
1940 | * Before new mappings are established, hotadded cpu might already |
1941 | * start handling requests. This doesn't break anything as we map | |
1942 | * offline CPUs to first hardware queue. We will re-init the queue | |
1943 | * below to get optimal settings. | |
320ae51f JA |
1944 | */ |
1945 | if (action != CPU_DEAD && action != CPU_DEAD_FROZEN && | |
1946 | action != CPU_ONLINE && action != CPU_ONLINE_FROZEN) | |
1947 | return NOTIFY_OK; | |
1948 | ||
1949 | mutex_lock(&all_q_mutex); | |
1950 | list_for_each_entry(q, &all_q_list, all_q_node) | |
1951 | blk_mq_queue_reinit(q); | |
1952 | mutex_unlock(&all_q_mutex); | |
1953 | return NOTIFY_OK; | |
1954 | } | |
1955 | ||
24d2f903 CH |
1956 | int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set) |
1957 | { | |
1958 | int i; | |
1959 | ||
1960 | if (!set->nr_hw_queues) | |
1961 | return -EINVAL; | |
1962 | if (!set->queue_depth || set->queue_depth > BLK_MQ_MAX_DEPTH) | |
1963 | return -EINVAL; | |
1964 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) | |
1965 | return -EINVAL; | |
1966 | ||
cdef54dd | 1967 | if (!set->nr_hw_queues || !set->ops->queue_rq || !set->ops->map_queue) |
24d2f903 CH |
1968 | return -EINVAL; |
1969 | ||
1970 | ||
48479005 ML |
1971 | set->tags = kmalloc_node(set->nr_hw_queues * |
1972 | sizeof(struct blk_mq_tags *), | |
24d2f903 CH |
1973 | GFP_KERNEL, set->numa_node); |
1974 | if (!set->tags) | |
1975 | goto out; | |
1976 | ||
1977 | for (i = 0; i < set->nr_hw_queues; i++) { | |
1978 | set->tags[i] = blk_mq_init_rq_map(set, i); | |
1979 | if (!set->tags[i]) | |
1980 | goto out_unwind; | |
1981 | } | |
1982 | ||
0d2602ca JA |
1983 | mutex_init(&set->tag_list_lock); |
1984 | INIT_LIST_HEAD(&set->tag_list); | |
1985 | ||
24d2f903 CH |
1986 | return 0; |
1987 | ||
1988 | out_unwind: | |
1989 | while (--i >= 0) | |
1990 | blk_mq_free_rq_map(set, set->tags[i], i); | |
1991 | out: | |
1992 | return -ENOMEM; | |
1993 | } | |
1994 | EXPORT_SYMBOL(blk_mq_alloc_tag_set); | |
1995 | ||
1996 | void blk_mq_free_tag_set(struct blk_mq_tag_set *set) | |
1997 | { | |
1998 | int i; | |
1999 | ||
484b4061 JA |
2000 | for (i = 0; i < set->nr_hw_queues; i++) { |
2001 | if (set->tags[i]) | |
2002 | blk_mq_free_rq_map(set, set->tags[i], i); | |
2003 | } | |
2004 | ||
981bd189 | 2005 | kfree(set->tags); |
24d2f903 CH |
2006 | } |
2007 | EXPORT_SYMBOL(blk_mq_free_tag_set); | |
2008 | ||
e3a2b3f9 JA |
2009 | int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr) |
2010 | { | |
2011 | struct blk_mq_tag_set *set = q->tag_set; | |
2012 | struct blk_mq_hw_ctx *hctx; | |
2013 | int i, ret; | |
2014 | ||
2015 | if (!set || nr > set->queue_depth) | |
2016 | return -EINVAL; | |
2017 | ||
2018 | ret = 0; | |
2019 | queue_for_each_hw_ctx(q, hctx, i) { | |
2020 | ret = blk_mq_tag_update_depth(hctx->tags, nr); | |
2021 | if (ret) | |
2022 | break; | |
2023 | } | |
2024 | ||
2025 | if (!ret) | |
2026 | q->nr_requests = nr; | |
2027 | ||
2028 | return ret; | |
2029 | } | |
2030 | ||
676141e4 JA |
2031 | void blk_mq_disable_hotplug(void) |
2032 | { | |
2033 | mutex_lock(&all_q_mutex); | |
2034 | } | |
2035 | ||
2036 | void blk_mq_enable_hotplug(void) | |
2037 | { | |
2038 | mutex_unlock(&all_q_mutex); | |
2039 | } | |
2040 | ||
320ae51f JA |
2041 | static int __init blk_mq_init(void) |
2042 | { | |
320ae51f JA |
2043 | blk_mq_cpu_init(); |
2044 | ||
2045 | /* Must be called after percpu_counter_hotcpu_callback() */ | |
2046 | hotcpu_notifier(blk_mq_queue_reinit_notify, -10); | |
2047 | ||
2048 | return 0; | |
2049 | } | |
2050 | subsys_initcall(blk_mq_init); |