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 | rq->cmd_type = 0; |
203 | /* do not touch atomic flags, it needs atomic ops against the timer */ | |
204 | rq->cpu = -1; | |
205 | rq->__data_len = 0; | |
206 | rq->__sector = (sector_t) -1; | |
207 | rq->bio = NULL; | |
208 | rq->biotail = NULL; | |
209 | INIT_HLIST_NODE(&rq->hash); | |
210 | RB_CLEAR_NODE(&rq->rb_node); | |
211 | memset(&rq->flush, 0, max(sizeof(rq->flush), sizeof(rq->elv))); | |
212 | rq->rq_disk = NULL; | |
213 | rq->part = NULL; | |
0fec08b4 | 214 | rq->start_time = jiffies; |
af76e555 CH |
215 | #ifdef CONFIG_BLK_CGROUP |
216 | rq->rl = NULL; | |
0fec08b4 | 217 | set_start_time_ns(rq); |
af76e555 CH |
218 | rq->io_start_time_ns = 0; |
219 | #endif | |
220 | rq->nr_phys_segments = 0; | |
221 | #if defined(CONFIG_BLK_DEV_INTEGRITY) | |
222 | rq->nr_integrity_segments = 0; | |
223 | #endif | |
224 | rq->ioprio = 0; | |
225 | rq->special = NULL; | |
226 | /* tag was already set */ | |
227 | rq->errors = 0; | |
228 | memset(rq->__cmd, 0, sizeof(rq->__cmd)); | |
229 | rq->cmd = rq->__cmd; | |
230 | rq->cmd_len = BLK_MAX_CDB; | |
231 | ||
232 | rq->extra_len = 0; | |
233 | rq->sense_len = 0; | |
234 | rq->resid_len = 0; | |
235 | rq->sense = NULL; | |
236 | ||
237 | rq->deadline = 0; | |
238 | INIT_LIST_HEAD(&rq->timeout_list); | |
239 | rq->timeout = 0; | |
240 | rq->retries = 0; | |
241 | rq->end_io = NULL; | |
242 | rq->end_io_data = NULL; | |
243 | rq->next_rq = NULL; | |
244 | ||
320ae51f JA |
245 | ctx->rq_dispatched[rw_is_sync(rw_flags)]++; |
246 | } | |
247 | ||
5dee8577 CH |
248 | static struct request * |
249 | __blk_mq_alloc_request(struct request_queue *q, struct blk_mq_hw_ctx *hctx, | |
250 | struct blk_mq_ctx *ctx, int rw, gfp_t gfp, bool reserved) | |
251 | { | |
252 | struct request *rq; | |
253 | unsigned int tag; | |
254 | ||
255 | tag = blk_mq_get_tag(hctx, &ctx->last_tag, gfp, reserved); | |
256 | if (tag != BLK_MQ_TAG_FAIL) { | |
257 | rq = hctx->tags->rqs[tag]; | |
258 | ||
259 | rq->cmd_flags = 0; | |
260 | if (blk_mq_tag_busy(hctx)) { | |
261 | rq->cmd_flags = REQ_MQ_INFLIGHT; | |
262 | atomic_inc(&hctx->nr_active); | |
263 | } | |
264 | ||
265 | rq->tag = tag; | |
266 | blk_mq_rq_ctx_init(q, ctx, rq, rw); | |
267 | return rq; | |
268 | } | |
269 | ||
270 | return NULL; | |
271 | } | |
272 | ||
4ce01dd1 CH |
273 | struct request *blk_mq_alloc_request(struct request_queue *q, int rw, gfp_t gfp, |
274 | bool reserved) | |
320ae51f | 275 | { |
d852564f CH |
276 | struct blk_mq_ctx *ctx; |
277 | struct blk_mq_hw_ctx *hctx; | |
320ae51f JA |
278 | struct request *rq; |
279 | ||
280 | if (blk_mq_queue_enter(q)) | |
281 | return NULL; | |
282 | ||
d852564f CH |
283 | ctx = blk_mq_get_ctx(q); |
284 | hctx = q->mq_ops->map_queue(q, ctx->cpu); | |
285 | ||
286 | rq = __blk_mq_alloc_request(q, hctx, ctx, rw, gfp & ~__GFP_WAIT, | |
287 | reserved); | |
288 | if (!rq && (gfp & __GFP_WAIT)) { | |
289 | __blk_mq_run_hw_queue(hctx); | |
290 | blk_mq_put_ctx(ctx); | |
291 | ||
292 | ctx = blk_mq_get_ctx(q); | |
293 | hctx = q->mq_ops->map_queue(q, ctx->cpu); | |
294 | rq = __blk_mq_alloc_request(q, hctx, ctx, rw, gfp, reserved); | |
295 | } | |
296 | blk_mq_put_ctx(ctx); | |
320ae51f JA |
297 | return rq; |
298 | } | |
4bb659b1 | 299 | EXPORT_SYMBOL(blk_mq_alloc_request); |
320ae51f | 300 | |
320ae51f JA |
301 | static void __blk_mq_free_request(struct blk_mq_hw_ctx *hctx, |
302 | struct blk_mq_ctx *ctx, struct request *rq) | |
303 | { | |
304 | const int tag = rq->tag; | |
305 | struct request_queue *q = rq->q; | |
306 | ||
0d2602ca JA |
307 | if (rq->cmd_flags & REQ_MQ_INFLIGHT) |
308 | atomic_dec(&hctx->nr_active); | |
309 | ||
af76e555 | 310 | clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags); |
0d2602ca | 311 | blk_mq_put_tag(hctx, tag, &ctx->last_tag); |
320ae51f JA |
312 | blk_mq_queue_exit(q); |
313 | } | |
314 | ||
315 | void blk_mq_free_request(struct request *rq) | |
316 | { | |
317 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
318 | struct blk_mq_hw_ctx *hctx; | |
319 | struct request_queue *q = rq->q; | |
320 | ||
321 | ctx->rq_completed[rq_is_sync(rq)]++; | |
322 | ||
323 | hctx = q->mq_ops->map_queue(q, ctx->cpu); | |
324 | __blk_mq_free_request(hctx, ctx, rq); | |
325 | } | |
326 | ||
8727af4b CH |
327 | /* |
328 | * Clone all relevant state from a request that has been put on hold in | |
329 | * the flush state machine into the preallocated flush request that hangs | |
330 | * off the request queue. | |
331 | * | |
332 | * For a driver the flush request should be invisible, that's why we are | |
333 | * impersonating the original request here. | |
334 | */ | |
335 | void blk_mq_clone_flush_request(struct request *flush_rq, | |
336 | struct request *orig_rq) | |
337 | { | |
338 | struct blk_mq_hw_ctx *hctx = | |
339 | orig_rq->q->mq_ops->map_queue(orig_rq->q, orig_rq->mq_ctx->cpu); | |
340 | ||
341 | flush_rq->mq_ctx = orig_rq->mq_ctx; | |
342 | flush_rq->tag = orig_rq->tag; | |
343 | memcpy(blk_mq_rq_to_pdu(flush_rq), blk_mq_rq_to_pdu(orig_rq), | |
344 | hctx->cmd_size); | |
345 | } | |
346 | ||
63151a44 | 347 | inline void __blk_mq_end_io(struct request *rq, int error) |
320ae51f | 348 | { |
0d11e6ac ML |
349 | blk_account_io_done(rq); |
350 | ||
91b63639 | 351 | if (rq->end_io) { |
320ae51f | 352 | rq->end_io(rq, error); |
91b63639 CH |
353 | } else { |
354 | if (unlikely(blk_bidi_rq(rq))) | |
355 | blk_mq_free_request(rq->next_rq); | |
320ae51f | 356 | blk_mq_free_request(rq); |
91b63639 | 357 | } |
320ae51f | 358 | } |
63151a44 CH |
359 | EXPORT_SYMBOL(__blk_mq_end_io); |
360 | ||
361 | void blk_mq_end_io(struct request *rq, int error) | |
362 | { | |
363 | if (blk_update_request(rq, error, blk_rq_bytes(rq))) | |
364 | BUG(); | |
365 | __blk_mq_end_io(rq, error); | |
366 | } | |
367 | EXPORT_SYMBOL(blk_mq_end_io); | |
320ae51f | 368 | |
30a91cb4 | 369 | static void __blk_mq_complete_request_remote(void *data) |
320ae51f | 370 | { |
3d6efbf6 | 371 | struct request *rq = data; |
320ae51f | 372 | |
30a91cb4 | 373 | rq->q->softirq_done_fn(rq); |
320ae51f | 374 | } |
320ae51f | 375 | |
30a91cb4 | 376 | void __blk_mq_complete_request(struct request *rq) |
320ae51f JA |
377 | { |
378 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
38535201 | 379 | bool shared = false; |
320ae51f JA |
380 | int cpu; |
381 | ||
38535201 | 382 | if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) { |
30a91cb4 CH |
383 | rq->q->softirq_done_fn(rq); |
384 | return; | |
385 | } | |
320ae51f JA |
386 | |
387 | cpu = get_cpu(); | |
38535201 CH |
388 | if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags)) |
389 | shared = cpus_share_cache(cpu, ctx->cpu); | |
390 | ||
391 | if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) { | |
30a91cb4 | 392 | rq->csd.func = __blk_mq_complete_request_remote; |
3d6efbf6 CH |
393 | rq->csd.info = rq; |
394 | rq->csd.flags = 0; | |
c46fff2a | 395 | smp_call_function_single_async(ctx->cpu, &rq->csd); |
3d6efbf6 | 396 | } else { |
30a91cb4 | 397 | rq->q->softirq_done_fn(rq); |
3d6efbf6 | 398 | } |
320ae51f JA |
399 | put_cpu(); |
400 | } | |
30a91cb4 CH |
401 | |
402 | /** | |
403 | * blk_mq_complete_request - end I/O on a request | |
404 | * @rq: the request being processed | |
405 | * | |
406 | * Description: | |
407 | * Ends all I/O on a request. It does not handle partial completions. | |
408 | * The actual completion happens out-of-order, through a IPI handler. | |
409 | **/ | |
410 | void blk_mq_complete_request(struct request *rq) | |
411 | { | |
95f09684 JA |
412 | struct request_queue *q = rq->q; |
413 | ||
414 | if (unlikely(blk_should_fake_timeout(q))) | |
30a91cb4 | 415 | return; |
95f09684 JA |
416 | if (!blk_mark_rq_complete(rq)) { |
417 | if (q->softirq_done_fn) | |
418 | __blk_mq_complete_request(rq); | |
419 | else | |
420 | blk_mq_end_io(rq, rq->errors); | |
421 | } | |
30a91cb4 CH |
422 | } |
423 | EXPORT_SYMBOL(blk_mq_complete_request); | |
320ae51f | 424 | |
49f5baa5 | 425 | static void blk_mq_start_request(struct request *rq, bool last) |
320ae51f JA |
426 | { |
427 | struct request_queue *q = rq->q; | |
428 | ||
429 | trace_block_rq_issue(q, rq); | |
430 | ||
742ee69b | 431 | rq->resid_len = blk_rq_bytes(rq); |
91b63639 CH |
432 | if (unlikely(blk_bidi_rq(rq))) |
433 | rq->next_rq->resid_len = blk_rq_bytes(rq->next_rq); | |
742ee69b | 434 | |
320ae51f JA |
435 | /* |
436 | * Just mark start time and set the started bit. Due to memory | |
437 | * ordering, we know we'll see the correct deadline as long as | |
c22d9d8a JA |
438 | * REQ_ATOMIC_STARTED is seen. Use the default queue timeout, |
439 | * unless one has been set in the request. | |
320ae51f | 440 | */ |
c22d9d8a JA |
441 | if (!rq->timeout) |
442 | rq->deadline = jiffies + q->rq_timeout; | |
443 | else | |
444 | rq->deadline = jiffies + rq->timeout; | |
87ee7b11 JA |
445 | |
446 | /* | |
447 | * Mark us as started and clear complete. Complete might have been | |
448 | * set if requeue raced with timeout, which then marked it as | |
449 | * complete. So be sure to clear complete again when we start | |
450 | * the request, otherwise we'll ignore the completion event. | |
451 | */ | |
320ae51f | 452 | set_bit(REQ_ATOM_STARTED, &rq->atomic_flags); |
87ee7b11 | 453 | clear_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags); |
49f5baa5 CH |
454 | |
455 | if (q->dma_drain_size && blk_rq_bytes(rq)) { | |
456 | /* | |
457 | * Make sure space for the drain appears. We know we can do | |
458 | * this because max_hw_segments has been adjusted to be one | |
459 | * fewer than the device can handle. | |
460 | */ | |
461 | rq->nr_phys_segments++; | |
462 | } | |
463 | ||
464 | /* | |
465 | * Flag the last request in the series so that drivers know when IO | |
466 | * should be kicked off, if they don't do it on a per-request basis. | |
467 | * | |
468 | * Note: the flag isn't the only condition drivers should do kick off. | |
469 | * If drive is busy, the last request might not have the bit set. | |
470 | */ | |
471 | if (last) | |
472 | rq->cmd_flags |= REQ_END; | |
320ae51f JA |
473 | } |
474 | ||
ed0791b2 | 475 | static void __blk_mq_requeue_request(struct request *rq) |
320ae51f JA |
476 | { |
477 | struct request_queue *q = rq->q; | |
478 | ||
479 | trace_block_rq_requeue(q, rq); | |
480 | clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags); | |
49f5baa5 CH |
481 | |
482 | rq->cmd_flags &= ~REQ_END; | |
483 | ||
484 | if (q->dma_drain_size && blk_rq_bytes(rq)) | |
485 | rq->nr_phys_segments--; | |
320ae51f JA |
486 | } |
487 | ||
ed0791b2 CH |
488 | void blk_mq_requeue_request(struct request *rq) |
489 | { | |
ed0791b2 CH |
490 | __blk_mq_requeue_request(rq); |
491 | blk_clear_rq_complete(rq); | |
492 | ||
ed0791b2 | 493 | BUG_ON(blk_queued_rq(rq)); |
6fca6a61 | 494 | blk_mq_add_to_requeue_list(rq, true); |
ed0791b2 CH |
495 | } |
496 | EXPORT_SYMBOL(blk_mq_requeue_request); | |
497 | ||
6fca6a61 CH |
498 | static void blk_mq_requeue_work(struct work_struct *work) |
499 | { | |
500 | struct request_queue *q = | |
501 | container_of(work, struct request_queue, requeue_work); | |
502 | LIST_HEAD(rq_list); | |
503 | struct request *rq, *next; | |
504 | unsigned long flags; | |
505 | ||
506 | spin_lock_irqsave(&q->requeue_lock, flags); | |
507 | list_splice_init(&q->requeue_list, &rq_list); | |
508 | spin_unlock_irqrestore(&q->requeue_lock, flags); | |
509 | ||
510 | list_for_each_entry_safe(rq, next, &rq_list, queuelist) { | |
511 | if (!(rq->cmd_flags & REQ_SOFTBARRIER)) | |
512 | continue; | |
513 | ||
514 | rq->cmd_flags &= ~REQ_SOFTBARRIER; | |
515 | list_del_init(&rq->queuelist); | |
516 | blk_mq_insert_request(rq, true, false, false); | |
517 | } | |
518 | ||
519 | while (!list_empty(&rq_list)) { | |
520 | rq = list_entry(rq_list.next, struct request, queuelist); | |
521 | list_del_init(&rq->queuelist); | |
522 | blk_mq_insert_request(rq, false, false, false); | |
523 | } | |
524 | ||
525 | blk_mq_run_queues(q, false); | |
526 | } | |
527 | ||
528 | void blk_mq_add_to_requeue_list(struct request *rq, bool at_head) | |
529 | { | |
530 | struct request_queue *q = rq->q; | |
531 | unsigned long flags; | |
532 | ||
533 | /* | |
534 | * We abuse this flag that is otherwise used by the I/O scheduler to | |
535 | * request head insertation from the workqueue. | |
536 | */ | |
537 | BUG_ON(rq->cmd_flags & REQ_SOFTBARRIER); | |
538 | ||
539 | spin_lock_irqsave(&q->requeue_lock, flags); | |
540 | if (at_head) { | |
541 | rq->cmd_flags |= REQ_SOFTBARRIER; | |
542 | list_add(&rq->queuelist, &q->requeue_list); | |
543 | } else { | |
544 | list_add_tail(&rq->queuelist, &q->requeue_list); | |
545 | } | |
546 | spin_unlock_irqrestore(&q->requeue_lock, flags); | |
547 | } | |
548 | EXPORT_SYMBOL(blk_mq_add_to_requeue_list); | |
549 | ||
550 | void blk_mq_kick_requeue_list(struct request_queue *q) | |
551 | { | |
552 | kblockd_schedule_work(&q->requeue_work); | |
553 | } | |
554 | EXPORT_SYMBOL(blk_mq_kick_requeue_list); | |
555 | ||
24d2f903 CH |
556 | struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag) |
557 | { | |
558 | return tags->rqs[tag]; | |
559 | } | |
560 | EXPORT_SYMBOL(blk_mq_tag_to_rq); | |
561 | ||
320ae51f JA |
562 | struct blk_mq_timeout_data { |
563 | struct blk_mq_hw_ctx *hctx; | |
564 | unsigned long *next; | |
565 | unsigned int *next_set; | |
566 | }; | |
567 | ||
568 | static void blk_mq_timeout_check(void *__data, unsigned long *free_tags) | |
569 | { | |
570 | struct blk_mq_timeout_data *data = __data; | |
571 | struct blk_mq_hw_ctx *hctx = data->hctx; | |
572 | unsigned int tag; | |
573 | ||
574 | /* It may not be in flight yet (this is where | |
575 | * the REQ_ATOMIC_STARTED flag comes in). The requests are | |
576 | * statically allocated, so we know it's always safe to access the | |
577 | * memory associated with a bit offset into ->rqs[]. | |
578 | */ | |
579 | tag = 0; | |
580 | do { | |
581 | struct request *rq; | |
582 | ||
24d2f903 CH |
583 | tag = find_next_zero_bit(free_tags, hctx->tags->nr_tags, tag); |
584 | if (tag >= hctx->tags->nr_tags) | |
320ae51f JA |
585 | break; |
586 | ||
24d2f903 CH |
587 | rq = blk_mq_tag_to_rq(hctx->tags, tag++); |
588 | if (rq->q != hctx->queue) | |
589 | continue; | |
320ae51f JA |
590 | if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) |
591 | continue; | |
592 | ||
593 | blk_rq_check_expired(rq, data->next, data->next_set); | |
594 | } while (1); | |
595 | } | |
596 | ||
597 | static void blk_mq_hw_ctx_check_timeout(struct blk_mq_hw_ctx *hctx, | |
598 | unsigned long *next, | |
599 | unsigned int *next_set) | |
600 | { | |
601 | struct blk_mq_timeout_data data = { | |
602 | .hctx = hctx, | |
603 | .next = next, | |
604 | .next_set = next_set, | |
605 | }; | |
606 | ||
607 | /* | |
608 | * Ask the tagging code to iterate busy requests, so we can | |
609 | * check them for timeout. | |
610 | */ | |
611 | blk_mq_tag_busy_iter(hctx->tags, blk_mq_timeout_check, &data); | |
612 | } | |
613 | ||
87ee7b11 JA |
614 | static enum blk_eh_timer_return blk_mq_rq_timed_out(struct request *rq) |
615 | { | |
616 | struct request_queue *q = rq->q; | |
617 | ||
618 | /* | |
619 | * We know that complete is set at this point. If STARTED isn't set | |
620 | * anymore, then the request isn't active and the "timeout" should | |
621 | * just be ignored. This can happen due to the bitflag ordering. | |
622 | * Timeout first checks if STARTED is set, and if it is, assumes | |
623 | * the request is active. But if we race with completion, then | |
624 | * we both flags will get cleared. So check here again, and ignore | |
625 | * a timeout event with a request that isn't active. | |
626 | */ | |
627 | if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) | |
628 | return BLK_EH_NOT_HANDLED; | |
629 | ||
630 | if (!q->mq_ops->timeout) | |
631 | return BLK_EH_RESET_TIMER; | |
632 | ||
633 | return q->mq_ops->timeout(rq); | |
634 | } | |
635 | ||
320ae51f JA |
636 | static void blk_mq_rq_timer(unsigned long data) |
637 | { | |
638 | struct request_queue *q = (struct request_queue *) data; | |
639 | struct blk_mq_hw_ctx *hctx; | |
640 | unsigned long next = 0; | |
641 | int i, next_set = 0; | |
642 | ||
484b4061 JA |
643 | queue_for_each_hw_ctx(q, hctx, i) { |
644 | /* | |
645 | * If not software queues are currently mapped to this | |
646 | * hardware queue, there's nothing to check | |
647 | */ | |
648 | if (!hctx->nr_ctx || !hctx->tags) | |
649 | continue; | |
650 | ||
320ae51f | 651 | blk_mq_hw_ctx_check_timeout(hctx, &next, &next_set); |
484b4061 | 652 | } |
320ae51f | 653 | |
0d2602ca JA |
654 | if (next_set) { |
655 | next = blk_rq_timeout(round_jiffies_up(next)); | |
656 | mod_timer(&q->timeout, next); | |
657 | } else { | |
658 | queue_for_each_hw_ctx(q, hctx, i) | |
659 | blk_mq_tag_idle(hctx); | |
660 | } | |
320ae51f JA |
661 | } |
662 | ||
663 | /* | |
664 | * Reverse check our software queue for entries that we could potentially | |
665 | * merge with. Currently includes a hand-wavy stop count of 8, to not spend | |
666 | * too much time checking for merges. | |
667 | */ | |
668 | static bool blk_mq_attempt_merge(struct request_queue *q, | |
669 | struct blk_mq_ctx *ctx, struct bio *bio) | |
670 | { | |
671 | struct request *rq; | |
672 | int checked = 8; | |
673 | ||
674 | list_for_each_entry_reverse(rq, &ctx->rq_list, queuelist) { | |
675 | int el_ret; | |
676 | ||
677 | if (!checked--) | |
678 | break; | |
679 | ||
680 | if (!blk_rq_merge_ok(rq, bio)) | |
681 | continue; | |
682 | ||
683 | el_ret = blk_try_merge(rq, bio); | |
684 | if (el_ret == ELEVATOR_BACK_MERGE) { | |
685 | if (bio_attempt_back_merge(q, rq, bio)) { | |
686 | ctx->rq_merged++; | |
687 | return true; | |
688 | } | |
689 | break; | |
690 | } else if (el_ret == ELEVATOR_FRONT_MERGE) { | |
691 | if (bio_attempt_front_merge(q, rq, bio)) { | |
692 | ctx->rq_merged++; | |
693 | return true; | |
694 | } | |
695 | break; | |
696 | } | |
697 | } | |
698 | ||
699 | return false; | |
700 | } | |
701 | ||
1429d7c9 JA |
702 | /* |
703 | * Process software queues that have been marked busy, splicing them | |
704 | * to the for-dispatch | |
705 | */ | |
706 | static void flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list) | |
707 | { | |
708 | struct blk_mq_ctx *ctx; | |
709 | int i; | |
710 | ||
711 | for (i = 0; i < hctx->ctx_map.map_size; i++) { | |
712 | struct blk_align_bitmap *bm = &hctx->ctx_map.map[i]; | |
713 | unsigned int off, bit; | |
714 | ||
715 | if (!bm->word) | |
716 | continue; | |
717 | ||
718 | bit = 0; | |
719 | off = i * hctx->ctx_map.bits_per_word; | |
720 | do { | |
721 | bit = find_next_bit(&bm->word, bm->depth, bit); | |
722 | if (bit >= bm->depth) | |
723 | break; | |
724 | ||
725 | ctx = hctx->ctxs[bit + off]; | |
726 | clear_bit(bit, &bm->word); | |
727 | spin_lock(&ctx->lock); | |
728 | list_splice_tail_init(&ctx->rq_list, list); | |
729 | spin_unlock(&ctx->lock); | |
730 | ||
731 | bit++; | |
732 | } while (1); | |
733 | } | |
734 | } | |
735 | ||
320ae51f JA |
736 | /* |
737 | * Run this hardware queue, pulling any software queues mapped to it in. | |
738 | * Note that this function currently has various problems around ordering | |
739 | * of IO. In particular, we'd like FIFO behaviour on handling existing | |
740 | * items on the hctx->dispatch list. Ignore that for now. | |
741 | */ | |
742 | static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx) | |
743 | { | |
744 | struct request_queue *q = hctx->queue; | |
320ae51f JA |
745 | struct request *rq; |
746 | LIST_HEAD(rq_list); | |
1429d7c9 | 747 | int queued; |
320ae51f | 748 | |
fd1270d5 | 749 | WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask)); |
e4043dcf | 750 | |
5d12f905 | 751 | if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state))) |
320ae51f JA |
752 | return; |
753 | ||
754 | hctx->run++; | |
755 | ||
756 | /* | |
757 | * Touch any software queue that has pending entries. | |
758 | */ | |
1429d7c9 | 759 | flush_busy_ctxs(hctx, &rq_list); |
320ae51f JA |
760 | |
761 | /* | |
762 | * If we have previous entries on our dispatch list, grab them | |
763 | * and stuff them at the front for more fair dispatch. | |
764 | */ | |
765 | if (!list_empty_careful(&hctx->dispatch)) { | |
766 | spin_lock(&hctx->lock); | |
767 | if (!list_empty(&hctx->dispatch)) | |
768 | list_splice_init(&hctx->dispatch, &rq_list); | |
769 | spin_unlock(&hctx->lock); | |
770 | } | |
771 | ||
320ae51f JA |
772 | /* |
773 | * Now process all the entries, sending them to the driver. | |
774 | */ | |
1429d7c9 | 775 | queued = 0; |
320ae51f JA |
776 | while (!list_empty(&rq_list)) { |
777 | int ret; | |
778 | ||
779 | rq = list_first_entry(&rq_list, struct request, queuelist); | |
780 | list_del_init(&rq->queuelist); | |
320ae51f | 781 | |
49f5baa5 | 782 | blk_mq_start_request(rq, list_empty(&rq_list)); |
320ae51f JA |
783 | |
784 | ret = q->mq_ops->queue_rq(hctx, rq); | |
785 | switch (ret) { | |
786 | case BLK_MQ_RQ_QUEUE_OK: | |
787 | queued++; | |
788 | continue; | |
789 | case BLK_MQ_RQ_QUEUE_BUSY: | |
320ae51f | 790 | list_add(&rq->queuelist, &rq_list); |
ed0791b2 | 791 | __blk_mq_requeue_request(rq); |
320ae51f JA |
792 | break; |
793 | default: | |
794 | pr_err("blk-mq: bad return on queue: %d\n", ret); | |
320ae51f | 795 | case BLK_MQ_RQ_QUEUE_ERROR: |
1e93b8c2 | 796 | rq->errors = -EIO; |
320ae51f JA |
797 | blk_mq_end_io(rq, rq->errors); |
798 | break; | |
799 | } | |
800 | ||
801 | if (ret == BLK_MQ_RQ_QUEUE_BUSY) | |
802 | break; | |
803 | } | |
804 | ||
805 | if (!queued) | |
806 | hctx->dispatched[0]++; | |
807 | else if (queued < (1 << (BLK_MQ_MAX_DISPATCH_ORDER - 1))) | |
808 | hctx->dispatched[ilog2(queued) + 1]++; | |
809 | ||
810 | /* | |
811 | * Any items that need requeuing? Stuff them into hctx->dispatch, | |
812 | * that is where we will continue on next queue run. | |
813 | */ | |
814 | if (!list_empty(&rq_list)) { | |
815 | spin_lock(&hctx->lock); | |
816 | list_splice(&rq_list, &hctx->dispatch); | |
817 | spin_unlock(&hctx->lock); | |
818 | } | |
819 | } | |
820 | ||
506e931f JA |
821 | /* |
822 | * It'd be great if the workqueue API had a way to pass | |
823 | * in a mask and had some smarts for more clever placement. | |
824 | * For now we just round-robin here, switching for every | |
825 | * BLK_MQ_CPU_WORK_BATCH queued items. | |
826 | */ | |
827 | static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx) | |
828 | { | |
829 | int cpu = hctx->next_cpu; | |
830 | ||
831 | if (--hctx->next_cpu_batch <= 0) { | |
832 | int next_cpu; | |
833 | ||
834 | next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask); | |
835 | if (next_cpu >= nr_cpu_ids) | |
836 | next_cpu = cpumask_first(hctx->cpumask); | |
837 | ||
838 | hctx->next_cpu = next_cpu; | |
839 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; | |
840 | } | |
841 | ||
842 | return cpu; | |
843 | } | |
844 | ||
320ae51f JA |
845 | void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) |
846 | { | |
5d12f905 | 847 | if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state))) |
320ae51f JA |
848 | return; |
849 | ||
e4043dcf | 850 | if (!async && cpumask_test_cpu(smp_processor_id(), hctx->cpumask)) |
320ae51f | 851 | __blk_mq_run_hw_queue(hctx); |
e4043dcf | 852 | else if (hctx->queue->nr_hw_queues == 1) |
70f4db63 | 853 | kblockd_schedule_delayed_work(&hctx->run_work, 0); |
e4043dcf JA |
854 | else { |
855 | unsigned int cpu; | |
856 | ||
506e931f | 857 | cpu = blk_mq_hctx_next_cpu(hctx); |
70f4db63 | 858 | kblockd_schedule_delayed_work_on(cpu, &hctx->run_work, 0); |
e4043dcf | 859 | } |
320ae51f JA |
860 | } |
861 | ||
862 | void blk_mq_run_queues(struct request_queue *q, bool async) | |
863 | { | |
864 | struct blk_mq_hw_ctx *hctx; | |
865 | int i; | |
866 | ||
867 | queue_for_each_hw_ctx(q, hctx, i) { | |
868 | if ((!blk_mq_hctx_has_pending(hctx) && | |
869 | list_empty_careful(&hctx->dispatch)) || | |
5d12f905 | 870 | test_bit(BLK_MQ_S_STOPPED, &hctx->state)) |
320ae51f JA |
871 | continue; |
872 | ||
e4043dcf | 873 | preempt_disable(); |
320ae51f | 874 | blk_mq_run_hw_queue(hctx, async); |
e4043dcf | 875 | preempt_enable(); |
320ae51f JA |
876 | } |
877 | } | |
878 | EXPORT_SYMBOL(blk_mq_run_queues); | |
879 | ||
880 | void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx) | |
881 | { | |
70f4db63 CH |
882 | cancel_delayed_work(&hctx->run_work); |
883 | cancel_delayed_work(&hctx->delay_work); | |
320ae51f JA |
884 | set_bit(BLK_MQ_S_STOPPED, &hctx->state); |
885 | } | |
886 | EXPORT_SYMBOL(blk_mq_stop_hw_queue); | |
887 | ||
280d45f6 CH |
888 | void blk_mq_stop_hw_queues(struct request_queue *q) |
889 | { | |
890 | struct blk_mq_hw_ctx *hctx; | |
891 | int i; | |
892 | ||
893 | queue_for_each_hw_ctx(q, hctx, i) | |
894 | blk_mq_stop_hw_queue(hctx); | |
895 | } | |
896 | EXPORT_SYMBOL(blk_mq_stop_hw_queues); | |
897 | ||
320ae51f JA |
898 | void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx) |
899 | { | |
900 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
e4043dcf JA |
901 | |
902 | preempt_disable(); | |
320ae51f | 903 | __blk_mq_run_hw_queue(hctx); |
e4043dcf | 904 | preempt_enable(); |
320ae51f JA |
905 | } |
906 | EXPORT_SYMBOL(blk_mq_start_hw_queue); | |
907 | ||
2f268556 CH |
908 | void blk_mq_start_hw_queues(struct request_queue *q) |
909 | { | |
910 | struct blk_mq_hw_ctx *hctx; | |
911 | int i; | |
912 | ||
913 | queue_for_each_hw_ctx(q, hctx, i) | |
914 | blk_mq_start_hw_queue(hctx); | |
915 | } | |
916 | EXPORT_SYMBOL(blk_mq_start_hw_queues); | |
917 | ||
918 | ||
1b4a3258 | 919 | void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async) |
320ae51f JA |
920 | { |
921 | struct blk_mq_hw_ctx *hctx; | |
922 | int i; | |
923 | ||
924 | queue_for_each_hw_ctx(q, hctx, i) { | |
925 | if (!test_bit(BLK_MQ_S_STOPPED, &hctx->state)) | |
926 | continue; | |
927 | ||
928 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
e4043dcf | 929 | preempt_disable(); |
1b4a3258 | 930 | blk_mq_run_hw_queue(hctx, async); |
e4043dcf | 931 | preempt_enable(); |
320ae51f JA |
932 | } |
933 | } | |
934 | EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues); | |
935 | ||
70f4db63 | 936 | static void blk_mq_run_work_fn(struct work_struct *work) |
320ae51f JA |
937 | { |
938 | struct blk_mq_hw_ctx *hctx; | |
939 | ||
70f4db63 | 940 | hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work); |
e4043dcf | 941 | |
320ae51f JA |
942 | __blk_mq_run_hw_queue(hctx); |
943 | } | |
944 | ||
70f4db63 CH |
945 | static void blk_mq_delay_work_fn(struct work_struct *work) |
946 | { | |
947 | struct blk_mq_hw_ctx *hctx; | |
948 | ||
949 | hctx = container_of(work, struct blk_mq_hw_ctx, delay_work.work); | |
950 | ||
951 | if (test_and_clear_bit(BLK_MQ_S_STOPPED, &hctx->state)) | |
952 | __blk_mq_run_hw_queue(hctx); | |
953 | } | |
954 | ||
955 | void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs) | |
956 | { | |
957 | unsigned long tmo = msecs_to_jiffies(msecs); | |
958 | ||
959 | if (hctx->queue->nr_hw_queues == 1) | |
960 | kblockd_schedule_delayed_work(&hctx->delay_work, tmo); | |
961 | else { | |
962 | unsigned int cpu; | |
963 | ||
506e931f | 964 | cpu = blk_mq_hctx_next_cpu(hctx); |
70f4db63 CH |
965 | kblockd_schedule_delayed_work_on(cpu, &hctx->delay_work, tmo); |
966 | } | |
967 | } | |
968 | EXPORT_SYMBOL(blk_mq_delay_queue); | |
969 | ||
320ae51f | 970 | static void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, |
72a0a36e | 971 | struct request *rq, bool at_head) |
320ae51f JA |
972 | { |
973 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
974 | ||
01b983c9 JA |
975 | trace_block_rq_insert(hctx->queue, rq); |
976 | ||
72a0a36e CH |
977 | if (at_head) |
978 | list_add(&rq->queuelist, &ctx->rq_list); | |
979 | else | |
980 | list_add_tail(&rq->queuelist, &ctx->rq_list); | |
4bb659b1 | 981 | |
320ae51f JA |
982 | blk_mq_hctx_mark_pending(hctx, ctx); |
983 | ||
984 | /* | |
985 | * We do this early, to ensure we are on the right CPU. | |
986 | */ | |
87ee7b11 | 987 | blk_add_timer(rq); |
320ae51f JA |
988 | } |
989 | ||
eeabc850 CH |
990 | void blk_mq_insert_request(struct request *rq, bool at_head, bool run_queue, |
991 | bool async) | |
320ae51f | 992 | { |
eeabc850 | 993 | struct request_queue *q = rq->q; |
320ae51f | 994 | struct blk_mq_hw_ctx *hctx; |
eeabc850 CH |
995 | struct blk_mq_ctx *ctx = rq->mq_ctx, *current_ctx; |
996 | ||
997 | current_ctx = blk_mq_get_ctx(q); | |
998 | if (!cpu_online(ctx->cpu)) | |
999 | rq->mq_ctx = ctx = current_ctx; | |
320ae51f | 1000 | |
320ae51f JA |
1001 | hctx = q->mq_ops->map_queue(q, ctx->cpu); |
1002 | ||
eeabc850 CH |
1003 | if (rq->cmd_flags & (REQ_FLUSH | REQ_FUA) && |
1004 | !(rq->cmd_flags & (REQ_FLUSH_SEQ))) { | |
320ae51f JA |
1005 | blk_insert_flush(rq); |
1006 | } else { | |
320ae51f | 1007 | spin_lock(&ctx->lock); |
72a0a36e | 1008 | __blk_mq_insert_request(hctx, rq, at_head); |
320ae51f | 1009 | spin_unlock(&ctx->lock); |
320ae51f JA |
1010 | } |
1011 | ||
320ae51f JA |
1012 | if (run_queue) |
1013 | blk_mq_run_hw_queue(hctx, async); | |
e4043dcf JA |
1014 | |
1015 | blk_mq_put_ctx(current_ctx); | |
320ae51f JA |
1016 | } |
1017 | ||
1018 | static void blk_mq_insert_requests(struct request_queue *q, | |
1019 | struct blk_mq_ctx *ctx, | |
1020 | struct list_head *list, | |
1021 | int depth, | |
1022 | bool from_schedule) | |
1023 | ||
1024 | { | |
1025 | struct blk_mq_hw_ctx *hctx; | |
1026 | struct blk_mq_ctx *current_ctx; | |
1027 | ||
1028 | trace_block_unplug(q, depth, !from_schedule); | |
1029 | ||
1030 | current_ctx = blk_mq_get_ctx(q); | |
1031 | ||
1032 | if (!cpu_online(ctx->cpu)) | |
1033 | ctx = current_ctx; | |
1034 | hctx = q->mq_ops->map_queue(q, ctx->cpu); | |
1035 | ||
1036 | /* | |
1037 | * preemption doesn't flush plug list, so it's possible ctx->cpu is | |
1038 | * offline now | |
1039 | */ | |
1040 | spin_lock(&ctx->lock); | |
1041 | while (!list_empty(list)) { | |
1042 | struct request *rq; | |
1043 | ||
1044 | rq = list_first_entry(list, struct request, queuelist); | |
1045 | list_del_init(&rq->queuelist); | |
1046 | rq->mq_ctx = ctx; | |
72a0a36e | 1047 | __blk_mq_insert_request(hctx, rq, false); |
320ae51f JA |
1048 | } |
1049 | spin_unlock(&ctx->lock); | |
1050 | ||
320ae51f | 1051 | blk_mq_run_hw_queue(hctx, from_schedule); |
e4043dcf | 1052 | blk_mq_put_ctx(current_ctx); |
320ae51f JA |
1053 | } |
1054 | ||
1055 | static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b) | |
1056 | { | |
1057 | struct request *rqa = container_of(a, struct request, queuelist); | |
1058 | struct request *rqb = container_of(b, struct request, queuelist); | |
1059 | ||
1060 | return !(rqa->mq_ctx < rqb->mq_ctx || | |
1061 | (rqa->mq_ctx == rqb->mq_ctx && | |
1062 | blk_rq_pos(rqa) < blk_rq_pos(rqb))); | |
1063 | } | |
1064 | ||
1065 | void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule) | |
1066 | { | |
1067 | struct blk_mq_ctx *this_ctx; | |
1068 | struct request_queue *this_q; | |
1069 | struct request *rq; | |
1070 | LIST_HEAD(list); | |
1071 | LIST_HEAD(ctx_list); | |
1072 | unsigned int depth; | |
1073 | ||
1074 | list_splice_init(&plug->mq_list, &list); | |
1075 | ||
1076 | list_sort(NULL, &list, plug_ctx_cmp); | |
1077 | ||
1078 | this_q = NULL; | |
1079 | this_ctx = NULL; | |
1080 | depth = 0; | |
1081 | ||
1082 | while (!list_empty(&list)) { | |
1083 | rq = list_entry_rq(list.next); | |
1084 | list_del_init(&rq->queuelist); | |
1085 | BUG_ON(!rq->q); | |
1086 | if (rq->mq_ctx != this_ctx) { | |
1087 | if (this_ctx) { | |
1088 | blk_mq_insert_requests(this_q, this_ctx, | |
1089 | &ctx_list, depth, | |
1090 | from_schedule); | |
1091 | } | |
1092 | ||
1093 | this_ctx = rq->mq_ctx; | |
1094 | this_q = rq->q; | |
1095 | depth = 0; | |
1096 | } | |
1097 | ||
1098 | depth++; | |
1099 | list_add_tail(&rq->queuelist, &ctx_list); | |
1100 | } | |
1101 | ||
1102 | /* | |
1103 | * If 'this_ctx' is set, we know we have entries to complete | |
1104 | * on 'ctx_list'. Do those. | |
1105 | */ | |
1106 | if (this_ctx) { | |
1107 | blk_mq_insert_requests(this_q, this_ctx, &ctx_list, depth, | |
1108 | from_schedule); | |
1109 | } | |
1110 | } | |
1111 | ||
1112 | static void blk_mq_bio_to_request(struct request *rq, struct bio *bio) | |
1113 | { | |
1114 | init_request_from_bio(rq, bio); | |
1115 | blk_account_io_start(rq, 1); | |
1116 | } | |
1117 | ||
07068d5b JA |
1118 | static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx *hctx, |
1119 | struct blk_mq_ctx *ctx, | |
1120 | struct request *rq, struct bio *bio) | |
320ae51f | 1121 | { |
07068d5b | 1122 | struct request_queue *q = hctx->queue; |
320ae51f | 1123 | |
07068d5b JA |
1124 | if (!(hctx->flags & BLK_MQ_F_SHOULD_MERGE)) { |
1125 | blk_mq_bio_to_request(rq, bio); | |
1126 | spin_lock(&ctx->lock); | |
1127 | insert_rq: | |
1128 | __blk_mq_insert_request(hctx, rq, false); | |
1129 | spin_unlock(&ctx->lock); | |
1130 | return false; | |
1131 | } else { | |
1132 | spin_lock(&ctx->lock); | |
1133 | if (!blk_mq_attempt_merge(q, ctx, bio)) { | |
1134 | blk_mq_bio_to_request(rq, bio); | |
1135 | goto insert_rq; | |
1136 | } | |
320ae51f | 1137 | |
07068d5b JA |
1138 | spin_unlock(&ctx->lock); |
1139 | __blk_mq_free_request(hctx, ctx, rq); | |
1140 | return true; | |
14ec77f3 | 1141 | } |
07068d5b | 1142 | } |
14ec77f3 | 1143 | |
07068d5b JA |
1144 | struct blk_map_ctx { |
1145 | struct blk_mq_hw_ctx *hctx; | |
1146 | struct blk_mq_ctx *ctx; | |
1147 | }; | |
1148 | ||
1149 | static struct request *blk_mq_map_request(struct request_queue *q, | |
1150 | struct bio *bio, | |
1151 | struct blk_map_ctx *data) | |
1152 | { | |
1153 | struct blk_mq_hw_ctx *hctx; | |
1154 | struct blk_mq_ctx *ctx; | |
1155 | struct request *rq; | |
1156 | int rw = bio_data_dir(bio); | |
320ae51f | 1157 | |
07068d5b | 1158 | if (unlikely(blk_mq_queue_enter(q))) { |
320ae51f | 1159 | bio_endio(bio, -EIO); |
07068d5b | 1160 | return NULL; |
320ae51f JA |
1161 | } |
1162 | ||
1163 | ctx = blk_mq_get_ctx(q); | |
1164 | hctx = q->mq_ops->map_queue(q, ctx->cpu); | |
1165 | ||
07068d5b | 1166 | if (rw_is_sync(bio->bi_rw)) |
27fbf4e8 | 1167 | rw |= REQ_SYNC; |
07068d5b | 1168 | |
320ae51f | 1169 | trace_block_getrq(q, bio, rw); |
5dee8577 CH |
1170 | rq = __blk_mq_alloc_request(q, hctx, ctx, rw, GFP_ATOMIC, false); |
1171 | if (unlikely(!rq)) { | |
793597a6 | 1172 | __blk_mq_run_hw_queue(hctx); |
320ae51f JA |
1173 | blk_mq_put_ctx(ctx); |
1174 | trace_block_sleeprq(q, bio, rw); | |
793597a6 CH |
1175 | |
1176 | ctx = blk_mq_get_ctx(q); | |
320ae51f | 1177 | hctx = q->mq_ops->map_queue(q, ctx->cpu); |
793597a6 CH |
1178 | rq = __blk_mq_alloc_request(q, hctx, ctx, rw, |
1179 | __GFP_WAIT|GFP_ATOMIC, false); | |
320ae51f JA |
1180 | } |
1181 | ||
1182 | hctx->queued++; | |
07068d5b JA |
1183 | data->hctx = hctx; |
1184 | data->ctx = ctx; | |
1185 | return rq; | |
1186 | } | |
1187 | ||
1188 | /* | |
1189 | * Multiple hardware queue variant. This will not use per-process plugs, | |
1190 | * but will attempt to bypass the hctx queueing if we can go straight to | |
1191 | * hardware for SYNC IO. | |
1192 | */ | |
1193 | static void blk_mq_make_request(struct request_queue *q, struct bio *bio) | |
1194 | { | |
1195 | const int is_sync = rw_is_sync(bio->bi_rw); | |
1196 | const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA); | |
1197 | struct blk_map_ctx data; | |
1198 | struct request *rq; | |
1199 | ||
1200 | blk_queue_bounce(q, &bio); | |
1201 | ||
1202 | if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) { | |
1203 | bio_endio(bio, -EIO); | |
1204 | return; | |
1205 | } | |
1206 | ||
1207 | rq = blk_mq_map_request(q, bio, &data); | |
1208 | if (unlikely(!rq)) | |
1209 | return; | |
1210 | ||
1211 | if (unlikely(is_flush_fua)) { | |
1212 | blk_mq_bio_to_request(rq, bio); | |
1213 | blk_insert_flush(rq); | |
1214 | goto run_queue; | |
1215 | } | |
1216 | ||
1217 | if (is_sync) { | |
1218 | int ret; | |
1219 | ||
1220 | blk_mq_bio_to_request(rq, bio); | |
1221 | blk_mq_start_request(rq, true); | |
1222 | ||
1223 | /* | |
1224 | * For OK queue, we are done. For error, kill it. Any other | |
1225 | * error (busy), just add it to our list as we previously | |
1226 | * would have done | |
1227 | */ | |
1228 | ret = q->mq_ops->queue_rq(data.hctx, rq); | |
1229 | if (ret == BLK_MQ_RQ_QUEUE_OK) | |
1230 | goto done; | |
1231 | else { | |
1232 | __blk_mq_requeue_request(rq); | |
1233 | ||
1234 | if (ret == BLK_MQ_RQ_QUEUE_ERROR) { | |
1235 | rq->errors = -EIO; | |
1236 | blk_mq_end_io(rq, rq->errors); | |
1237 | goto done; | |
1238 | } | |
1239 | } | |
1240 | } | |
1241 | ||
1242 | if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) { | |
1243 | /* | |
1244 | * For a SYNC request, send it to the hardware immediately. For | |
1245 | * an ASYNC request, just ensure that we run it later on. The | |
1246 | * latter allows for merging opportunities and more efficient | |
1247 | * dispatching. | |
1248 | */ | |
1249 | run_queue: | |
1250 | blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua); | |
1251 | } | |
1252 | done: | |
1253 | blk_mq_put_ctx(data.ctx); | |
1254 | } | |
1255 | ||
1256 | /* | |
1257 | * Single hardware queue variant. This will attempt to use any per-process | |
1258 | * plug for merging and IO deferral. | |
1259 | */ | |
1260 | static void blk_sq_make_request(struct request_queue *q, struct bio *bio) | |
1261 | { | |
1262 | const int is_sync = rw_is_sync(bio->bi_rw); | |
1263 | const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA); | |
1264 | unsigned int use_plug, request_count = 0; | |
1265 | struct blk_map_ctx data; | |
1266 | struct request *rq; | |
1267 | ||
1268 | /* | |
1269 | * If we have multiple hardware queues, just go directly to | |
1270 | * one of those for sync IO. | |
1271 | */ | |
1272 | use_plug = !is_flush_fua && !is_sync; | |
1273 | ||
1274 | blk_queue_bounce(q, &bio); | |
1275 | ||
1276 | if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) { | |
1277 | bio_endio(bio, -EIO); | |
1278 | return; | |
1279 | } | |
1280 | ||
1281 | if (use_plug && !blk_queue_nomerges(q) && | |
1282 | blk_attempt_plug_merge(q, bio, &request_count)) | |
1283 | return; | |
1284 | ||
1285 | rq = blk_mq_map_request(q, bio, &data); | |
320ae51f JA |
1286 | |
1287 | if (unlikely(is_flush_fua)) { | |
1288 | blk_mq_bio_to_request(rq, bio); | |
320ae51f JA |
1289 | blk_insert_flush(rq); |
1290 | goto run_queue; | |
1291 | } | |
1292 | ||
1293 | /* | |
1294 | * A task plug currently exists. Since this is completely lockless, | |
1295 | * utilize that to temporarily store requests until the task is | |
1296 | * either done or scheduled away. | |
1297 | */ | |
1298 | if (use_plug) { | |
1299 | struct blk_plug *plug = current->plug; | |
1300 | ||
1301 | if (plug) { | |
1302 | blk_mq_bio_to_request(rq, bio); | |
92f399c7 | 1303 | if (list_empty(&plug->mq_list)) |
320ae51f JA |
1304 | trace_block_plug(q); |
1305 | else if (request_count >= BLK_MAX_REQUEST_COUNT) { | |
1306 | blk_flush_plug_list(plug, false); | |
1307 | trace_block_plug(q); | |
1308 | } | |
1309 | list_add_tail(&rq->queuelist, &plug->mq_list); | |
07068d5b | 1310 | blk_mq_put_ctx(data.ctx); |
320ae51f JA |
1311 | return; |
1312 | } | |
1313 | } | |
1314 | ||
07068d5b JA |
1315 | if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) { |
1316 | /* | |
1317 | * For a SYNC request, send it to the hardware immediately. For | |
1318 | * an ASYNC request, just ensure that we run it later on. The | |
1319 | * latter allows for merging opportunities and more efficient | |
1320 | * dispatching. | |
1321 | */ | |
1322 | run_queue: | |
1323 | blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua); | |
320ae51f JA |
1324 | } |
1325 | ||
07068d5b | 1326 | blk_mq_put_ctx(data.ctx); |
320ae51f JA |
1327 | } |
1328 | ||
1329 | /* | |
1330 | * Default mapping to a software queue, since we use one per CPU. | |
1331 | */ | |
1332 | struct blk_mq_hw_ctx *blk_mq_map_queue(struct request_queue *q, const int cpu) | |
1333 | { | |
1334 | return q->queue_hw_ctx[q->mq_map[cpu]]; | |
1335 | } | |
1336 | EXPORT_SYMBOL(blk_mq_map_queue); | |
1337 | ||
24d2f903 | 1338 | struct blk_mq_hw_ctx *blk_mq_alloc_single_hw_queue(struct blk_mq_tag_set *set, |
f14bbe77 JA |
1339 | unsigned int hctx_index, |
1340 | int node) | |
320ae51f | 1341 | { |
f14bbe77 | 1342 | return kzalloc_node(sizeof(struct blk_mq_hw_ctx), GFP_KERNEL, node); |
320ae51f JA |
1343 | } |
1344 | EXPORT_SYMBOL(blk_mq_alloc_single_hw_queue); | |
1345 | ||
1346 | void blk_mq_free_single_hw_queue(struct blk_mq_hw_ctx *hctx, | |
1347 | unsigned int hctx_index) | |
1348 | { | |
1349 | kfree(hctx); | |
1350 | } | |
1351 | EXPORT_SYMBOL(blk_mq_free_single_hw_queue); | |
1352 | ||
24d2f903 CH |
1353 | static void blk_mq_free_rq_map(struct blk_mq_tag_set *set, |
1354 | struct blk_mq_tags *tags, unsigned int hctx_idx) | |
95363efd | 1355 | { |
e9b267d9 | 1356 | struct page *page; |
320ae51f | 1357 | |
24d2f903 | 1358 | if (tags->rqs && set->ops->exit_request) { |
e9b267d9 | 1359 | int i; |
320ae51f | 1360 | |
24d2f903 CH |
1361 | for (i = 0; i < tags->nr_tags; i++) { |
1362 | if (!tags->rqs[i]) | |
e9b267d9 | 1363 | continue; |
24d2f903 CH |
1364 | set->ops->exit_request(set->driver_data, tags->rqs[i], |
1365 | hctx_idx, i); | |
e9b267d9 | 1366 | } |
320ae51f | 1367 | } |
320ae51f | 1368 | |
24d2f903 CH |
1369 | while (!list_empty(&tags->page_list)) { |
1370 | page = list_first_entry(&tags->page_list, struct page, lru); | |
6753471c | 1371 | list_del_init(&page->lru); |
320ae51f JA |
1372 | __free_pages(page, page->private); |
1373 | } | |
1374 | ||
24d2f903 | 1375 | kfree(tags->rqs); |
320ae51f | 1376 | |
24d2f903 | 1377 | blk_mq_free_tags(tags); |
320ae51f JA |
1378 | } |
1379 | ||
1380 | static size_t order_to_size(unsigned int order) | |
1381 | { | |
4ca08500 | 1382 | return (size_t)PAGE_SIZE << order; |
320ae51f JA |
1383 | } |
1384 | ||
24d2f903 CH |
1385 | static struct blk_mq_tags *blk_mq_init_rq_map(struct blk_mq_tag_set *set, |
1386 | unsigned int hctx_idx) | |
320ae51f | 1387 | { |
24d2f903 | 1388 | struct blk_mq_tags *tags; |
320ae51f JA |
1389 | unsigned int i, j, entries_per_page, max_order = 4; |
1390 | size_t rq_size, left; | |
1391 | ||
24d2f903 CH |
1392 | tags = blk_mq_init_tags(set->queue_depth, set->reserved_tags, |
1393 | set->numa_node); | |
1394 | if (!tags) | |
1395 | return NULL; | |
320ae51f | 1396 | |
24d2f903 CH |
1397 | INIT_LIST_HEAD(&tags->page_list); |
1398 | ||
1399 | tags->rqs = kmalloc_node(set->queue_depth * sizeof(struct request *), | |
1400 | GFP_KERNEL, set->numa_node); | |
1401 | if (!tags->rqs) { | |
1402 | blk_mq_free_tags(tags); | |
1403 | return NULL; | |
1404 | } | |
320ae51f JA |
1405 | |
1406 | /* | |
1407 | * rq_size is the size of the request plus driver payload, rounded | |
1408 | * to the cacheline size | |
1409 | */ | |
24d2f903 | 1410 | rq_size = round_up(sizeof(struct request) + set->cmd_size, |
320ae51f | 1411 | cache_line_size()); |
24d2f903 | 1412 | left = rq_size * set->queue_depth; |
320ae51f | 1413 | |
24d2f903 | 1414 | for (i = 0; i < set->queue_depth; ) { |
320ae51f JA |
1415 | int this_order = max_order; |
1416 | struct page *page; | |
1417 | int to_do; | |
1418 | void *p; | |
1419 | ||
1420 | while (left < order_to_size(this_order - 1) && this_order) | |
1421 | this_order--; | |
1422 | ||
1423 | do { | |
24d2f903 CH |
1424 | page = alloc_pages_node(set->numa_node, GFP_KERNEL, |
1425 | this_order); | |
320ae51f JA |
1426 | if (page) |
1427 | break; | |
1428 | if (!this_order--) | |
1429 | break; | |
1430 | if (order_to_size(this_order) < rq_size) | |
1431 | break; | |
1432 | } while (1); | |
1433 | ||
1434 | if (!page) | |
24d2f903 | 1435 | goto fail; |
320ae51f JA |
1436 | |
1437 | page->private = this_order; | |
24d2f903 | 1438 | list_add_tail(&page->lru, &tags->page_list); |
320ae51f JA |
1439 | |
1440 | p = page_address(page); | |
1441 | entries_per_page = order_to_size(this_order) / rq_size; | |
24d2f903 | 1442 | to_do = min(entries_per_page, set->queue_depth - i); |
320ae51f JA |
1443 | left -= to_do * rq_size; |
1444 | for (j = 0; j < to_do; j++) { | |
24d2f903 CH |
1445 | tags->rqs[i] = p; |
1446 | if (set->ops->init_request) { | |
1447 | if (set->ops->init_request(set->driver_data, | |
1448 | tags->rqs[i], hctx_idx, i, | |
1449 | set->numa_node)) | |
1450 | goto fail; | |
e9b267d9 CH |
1451 | } |
1452 | ||
320ae51f JA |
1453 | p += rq_size; |
1454 | i++; | |
1455 | } | |
1456 | } | |
1457 | ||
24d2f903 | 1458 | return tags; |
320ae51f | 1459 | |
24d2f903 CH |
1460 | fail: |
1461 | pr_warn("%s: failed to allocate requests\n", __func__); | |
1462 | blk_mq_free_rq_map(set, tags, hctx_idx); | |
1463 | return NULL; | |
320ae51f JA |
1464 | } |
1465 | ||
1429d7c9 JA |
1466 | static void blk_mq_free_bitmap(struct blk_mq_ctxmap *bitmap) |
1467 | { | |
1468 | kfree(bitmap->map); | |
1469 | } | |
1470 | ||
1471 | static int blk_mq_alloc_bitmap(struct blk_mq_ctxmap *bitmap, int node) | |
1472 | { | |
1473 | unsigned int bpw = 8, total, num_maps, i; | |
1474 | ||
1475 | bitmap->bits_per_word = bpw; | |
1476 | ||
1477 | num_maps = ALIGN(nr_cpu_ids, bpw) / bpw; | |
1478 | bitmap->map = kzalloc_node(num_maps * sizeof(struct blk_align_bitmap), | |
1479 | GFP_KERNEL, node); | |
1480 | if (!bitmap->map) | |
1481 | return -ENOMEM; | |
1482 | ||
1483 | bitmap->map_size = num_maps; | |
1484 | ||
1485 | total = nr_cpu_ids; | |
1486 | for (i = 0; i < num_maps; i++) { | |
1487 | bitmap->map[i].depth = min(total, bitmap->bits_per_word); | |
1488 | total -= bitmap->map[i].depth; | |
1489 | } | |
1490 | ||
1491 | return 0; | |
1492 | } | |
1493 | ||
484b4061 JA |
1494 | static int blk_mq_hctx_cpu_offline(struct blk_mq_hw_ctx *hctx, int cpu) |
1495 | { | |
1496 | struct request_queue *q = hctx->queue; | |
1497 | struct blk_mq_ctx *ctx; | |
1498 | LIST_HEAD(tmp); | |
1499 | ||
1500 | /* | |
1501 | * Move ctx entries to new CPU, if this one is going away. | |
1502 | */ | |
1503 | ctx = __blk_mq_get_ctx(q, cpu); | |
1504 | ||
1505 | spin_lock(&ctx->lock); | |
1506 | if (!list_empty(&ctx->rq_list)) { | |
1507 | list_splice_init(&ctx->rq_list, &tmp); | |
1508 | blk_mq_hctx_clear_pending(hctx, ctx); | |
1509 | } | |
1510 | spin_unlock(&ctx->lock); | |
1511 | ||
1512 | if (list_empty(&tmp)) | |
1513 | return NOTIFY_OK; | |
1514 | ||
1515 | ctx = blk_mq_get_ctx(q); | |
1516 | spin_lock(&ctx->lock); | |
1517 | ||
1518 | while (!list_empty(&tmp)) { | |
1519 | struct request *rq; | |
1520 | ||
1521 | rq = list_first_entry(&tmp, struct request, queuelist); | |
1522 | rq->mq_ctx = ctx; | |
1523 | list_move_tail(&rq->queuelist, &ctx->rq_list); | |
1524 | } | |
1525 | ||
1526 | hctx = q->mq_ops->map_queue(q, ctx->cpu); | |
1527 | blk_mq_hctx_mark_pending(hctx, ctx); | |
1528 | ||
1529 | spin_unlock(&ctx->lock); | |
1530 | ||
1531 | blk_mq_run_hw_queue(hctx, true); | |
1532 | blk_mq_put_ctx(ctx); | |
1533 | return NOTIFY_OK; | |
1534 | } | |
1535 | ||
1536 | static int blk_mq_hctx_cpu_online(struct blk_mq_hw_ctx *hctx, int cpu) | |
1537 | { | |
1538 | struct request_queue *q = hctx->queue; | |
1539 | struct blk_mq_tag_set *set = q->tag_set; | |
1540 | ||
1541 | if (set->tags[hctx->queue_num]) | |
1542 | return NOTIFY_OK; | |
1543 | ||
1544 | set->tags[hctx->queue_num] = blk_mq_init_rq_map(set, hctx->queue_num); | |
1545 | if (!set->tags[hctx->queue_num]) | |
1546 | return NOTIFY_STOP; | |
1547 | ||
1548 | hctx->tags = set->tags[hctx->queue_num]; | |
1549 | return NOTIFY_OK; | |
1550 | } | |
1551 | ||
1552 | static int blk_mq_hctx_notify(void *data, unsigned long action, | |
1553 | unsigned int cpu) | |
1554 | { | |
1555 | struct blk_mq_hw_ctx *hctx = data; | |
1556 | ||
1557 | if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) | |
1558 | return blk_mq_hctx_cpu_offline(hctx, cpu); | |
1559 | else if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) | |
1560 | return blk_mq_hctx_cpu_online(hctx, cpu); | |
1561 | ||
1562 | return NOTIFY_OK; | |
1563 | } | |
1564 | ||
624dbe47 ML |
1565 | static void blk_mq_exit_hw_queues(struct request_queue *q, |
1566 | struct blk_mq_tag_set *set, int nr_queue) | |
1567 | { | |
1568 | struct blk_mq_hw_ctx *hctx; | |
1569 | unsigned int i; | |
1570 | ||
1571 | queue_for_each_hw_ctx(q, hctx, i) { | |
1572 | if (i == nr_queue) | |
1573 | break; | |
1574 | ||
1575 | if (set->ops->exit_hctx) | |
1576 | set->ops->exit_hctx(hctx, i); | |
1577 | ||
1578 | blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier); | |
1579 | kfree(hctx->ctxs); | |
1580 | blk_mq_free_bitmap(&hctx->ctx_map); | |
1581 | } | |
1582 | ||
1583 | } | |
1584 | ||
1585 | static void blk_mq_free_hw_queues(struct request_queue *q, | |
1586 | struct blk_mq_tag_set *set) | |
1587 | { | |
1588 | struct blk_mq_hw_ctx *hctx; | |
1589 | unsigned int i; | |
1590 | ||
1591 | queue_for_each_hw_ctx(q, hctx, i) { | |
1592 | free_cpumask_var(hctx->cpumask); | |
1593 | set->ops->free_hctx(hctx, i); | |
1594 | } | |
1595 | } | |
1596 | ||
320ae51f | 1597 | static int blk_mq_init_hw_queues(struct request_queue *q, |
24d2f903 | 1598 | struct blk_mq_tag_set *set) |
320ae51f JA |
1599 | { |
1600 | struct blk_mq_hw_ctx *hctx; | |
624dbe47 | 1601 | unsigned int i; |
320ae51f JA |
1602 | |
1603 | /* | |
1604 | * Initialize hardware queues | |
1605 | */ | |
1606 | queue_for_each_hw_ctx(q, hctx, i) { | |
320ae51f JA |
1607 | int node; |
1608 | ||
1609 | node = hctx->numa_node; | |
1610 | if (node == NUMA_NO_NODE) | |
24d2f903 | 1611 | node = hctx->numa_node = set->numa_node; |
320ae51f | 1612 | |
70f4db63 CH |
1613 | INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn); |
1614 | INIT_DELAYED_WORK(&hctx->delay_work, blk_mq_delay_work_fn); | |
320ae51f JA |
1615 | spin_lock_init(&hctx->lock); |
1616 | INIT_LIST_HEAD(&hctx->dispatch); | |
1617 | hctx->queue = q; | |
1618 | hctx->queue_num = i; | |
24d2f903 CH |
1619 | hctx->flags = set->flags; |
1620 | hctx->cmd_size = set->cmd_size; | |
320ae51f JA |
1621 | |
1622 | blk_mq_init_cpu_notifier(&hctx->cpu_notifier, | |
1623 | blk_mq_hctx_notify, hctx); | |
1624 | blk_mq_register_cpu_notifier(&hctx->cpu_notifier); | |
1625 | ||
24d2f903 | 1626 | hctx->tags = set->tags[i]; |
320ae51f JA |
1627 | |
1628 | /* | |
1629 | * Allocate space for all possible cpus to avoid allocation in | |
1630 | * runtime | |
1631 | */ | |
1632 | hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *), | |
1633 | GFP_KERNEL, node); | |
1634 | if (!hctx->ctxs) | |
1635 | break; | |
1636 | ||
1429d7c9 | 1637 | if (blk_mq_alloc_bitmap(&hctx->ctx_map, node)) |
320ae51f JA |
1638 | break; |
1639 | ||
320ae51f JA |
1640 | hctx->nr_ctx = 0; |
1641 | ||
24d2f903 CH |
1642 | if (set->ops->init_hctx && |
1643 | set->ops->init_hctx(hctx, set->driver_data, i)) | |
320ae51f JA |
1644 | break; |
1645 | } | |
1646 | ||
1647 | if (i == q->nr_hw_queues) | |
1648 | return 0; | |
1649 | ||
1650 | /* | |
1651 | * Init failed | |
1652 | */ | |
624dbe47 | 1653 | blk_mq_exit_hw_queues(q, set, i); |
320ae51f JA |
1654 | |
1655 | return 1; | |
1656 | } | |
1657 | ||
1658 | static void blk_mq_init_cpu_queues(struct request_queue *q, | |
1659 | unsigned int nr_hw_queues) | |
1660 | { | |
1661 | unsigned int i; | |
1662 | ||
1663 | for_each_possible_cpu(i) { | |
1664 | struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i); | |
1665 | struct blk_mq_hw_ctx *hctx; | |
1666 | ||
1667 | memset(__ctx, 0, sizeof(*__ctx)); | |
1668 | __ctx->cpu = i; | |
1669 | spin_lock_init(&__ctx->lock); | |
1670 | INIT_LIST_HEAD(&__ctx->rq_list); | |
1671 | __ctx->queue = q; | |
1672 | ||
1673 | /* If the cpu isn't online, the cpu is mapped to first hctx */ | |
320ae51f JA |
1674 | if (!cpu_online(i)) |
1675 | continue; | |
1676 | ||
e4043dcf JA |
1677 | hctx = q->mq_ops->map_queue(q, i); |
1678 | cpumask_set_cpu(i, hctx->cpumask); | |
1679 | hctx->nr_ctx++; | |
1680 | ||
320ae51f JA |
1681 | /* |
1682 | * Set local node, IFF we have more than one hw queue. If | |
1683 | * not, we remain on the home node of the device | |
1684 | */ | |
1685 | if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE) | |
1686 | hctx->numa_node = cpu_to_node(i); | |
1687 | } | |
1688 | } | |
1689 | ||
1690 | static void blk_mq_map_swqueue(struct request_queue *q) | |
1691 | { | |
1692 | unsigned int i; | |
1693 | struct blk_mq_hw_ctx *hctx; | |
1694 | struct blk_mq_ctx *ctx; | |
1695 | ||
1696 | queue_for_each_hw_ctx(q, hctx, i) { | |
e4043dcf | 1697 | cpumask_clear(hctx->cpumask); |
320ae51f JA |
1698 | hctx->nr_ctx = 0; |
1699 | } | |
1700 | ||
1701 | /* | |
1702 | * Map software to hardware queues | |
1703 | */ | |
1704 | queue_for_each_ctx(q, ctx, i) { | |
1705 | /* If the cpu isn't online, the cpu is mapped to first hctx */ | |
e4043dcf JA |
1706 | if (!cpu_online(i)) |
1707 | continue; | |
1708 | ||
320ae51f | 1709 | hctx = q->mq_ops->map_queue(q, i); |
e4043dcf | 1710 | cpumask_set_cpu(i, hctx->cpumask); |
320ae51f JA |
1711 | ctx->index_hw = hctx->nr_ctx; |
1712 | hctx->ctxs[hctx->nr_ctx++] = ctx; | |
1713 | } | |
506e931f JA |
1714 | |
1715 | queue_for_each_hw_ctx(q, hctx, i) { | |
484b4061 JA |
1716 | /* |
1717 | * If not software queues are mapped to this hardware queue, | |
1718 | * disable it and free the request entries | |
1719 | */ | |
1720 | if (!hctx->nr_ctx) { | |
1721 | struct blk_mq_tag_set *set = q->tag_set; | |
1722 | ||
1723 | if (set->tags[i]) { | |
1724 | blk_mq_free_rq_map(set, set->tags[i], i); | |
1725 | set->tags[i] = NULL; | |
1726 | hctx->tags = NULL; | |
1727 | } | |
1728 | continue; | |
1729 | } | |
1730 | ||
1731 | /* | |
1732 | * Initialize batch roundrobin counts | |
1733 | */ | |
506e931f JA |
1734 | hctx->next_cpu = cpumask_first(hctx->cpumask); |
1735 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; | |
1736 | } | |
320ae51f JA |
1737 | } |
1738 | ||
0d2602ca JA |
1739 | static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set) |
1740 | { | |
1741 | struct blk_mq_hw_ctx *hctx; | |
1742 | struct request_queue *q; | |
1743 | bool shared; | |
1744 | int i; | |
1745 | ||
1746 | if (set->tag_list.next == set->tag_list.prev) | |
1747 | shared = false; | |
1748 | else | |
1749 | shared = true; | |
1750 | ||
1751 | list_for_each_entry(q, &set->tag_list, tag_set_list) { | |
1752 | blk_mq_freeze_queue(q); | |
1753 | ||
1754 | queue_for_each_hw_ctx(q, hctx, i) { | |
1755 | if (shared) | |
1756 | hctx->flags |= BLK_MQ_F_TAG_SHARED; | |
1757 | else | |
1758 | hctx->flags &= ~BLK_MQ_F_TAG_SHARED; | |
1759 | } | |
1760 | blk_mq_unfreeze_queue(q); | |
1761 | } | |
1762 | } | |
1763 | ||
1764 | static void blk_mq_del_queue_tag_set(struct request_queue *q) | |
1765 | { | |
1766 | struct blk_mq_tag_set *set = q->tag_set; | |
1767 | ||
1768 | blk_mq_freeze_queue(q); | |
1769 | ||
1770 | mutex_lock(&set->tag_list_lock); | |
1771 | list_del_init(&q->tag_set_list); | |
1772 | blk_mq_update_tag_set_depth(set); | |
1773 | mutex_unlock(&set->tag_list_lock); | |
1774 | ||
1775 | blk_mq_unfreeze_queue(q); | |
1776 | } | |
1777 | ||
1778 | static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set, | |
1779 | struct request_queue *q) | |
1780 | { | |
1781 | q->tag_set = set; | |
1782 | ||
1783 | mutex_lock(&set->tag_list_lock); | |
1784 | list_add_tail(&q->tag_set_list, &set->tag_list); | |
1785 | blk_mq_update_tag_set_depth(set); | |
1786 | mutex_unlock(&set->tag_list_lock); | |
1787 | } | |
1788 | ||
24d2f903 | 1789 | struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set) |
320ae51f JA |
1790 | { |
1791 | struct blk_mq_hw_ctx **hctxs; | |
1792 | struct blk_mq_ctx *ctx; | |
1793 | struct request_queue *q; | |
f14bbe77 | 1794 | unsigned int *map; |
320ae51f JA |
1795 | int i; |
1796 | ||
320ae51f JA |
1797 | ctx = alloc_percpu(struct blk_mq_ctx); |
1798 | if (!ctx) | |
1799 | return ERR_PTR(-ENOMEM); | |
1800 | ||
24d2f903 CH |
1801 | hctxs = kmalloc_node(set->nr_hw_queues * sizeof(*hctxs), GFP_KERNEL, |
1802 | set->numa_node); | |
320ae51f JA |
1803 | |
1804 | if (!hctxs) | |
1805 | goto err_percpu; | |
1806 | ||
f14bbe77 JA |
1807 | map = blk_mq_make_queue_map(set); |
1808 | if (!map) | |
1809 | goto err_map; | |
1810 | ||
24d2f903 | 1811 | for (i = 0; i < set->nr_hw_queues; i++) { |
f14bbe77 JA |
1812 | int node = blk_mq_hw_queue_to_node(map, i); |
1813 | ||
1814 | hctxs[i] = set->ops->alloc_hctx(set, i, node); | |
320ae51f JA |
1815 | if (!hctxs[i]) |
1816 | goto err_hctxs; | |
1817 | ||
e4043dcf JA |
1818 | if (!zalloc_cpumask_var(&hctxs[i]->cpumask, GFP_KERNEL)) |
1819 | goto err_hctxs; | |
1820 | ||
0d2602ca | 1821 | atomic_set(&hctxs[i]->nr_active, 0); |
f14bbe77 | 1822 | hctxs[i]->numa_node = node; |
320ae51f JA |
1823 | hctxs[i]->queue_num = i; |
1824 | } | |
1825 | ||
24d2f903 | 1826 | q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node); |
320ae51f JA |
1827 | if (!q) |
1828 | goto err_hctxs; | |
1829 | ||
3d2936f4 ML |
1830 | if (percpu_counter_init(&q->mq_usage_counter, 0)) |
1831 | goto err_map; | |
1832 | ||
320ae51f JA |
1833 | setup_timer(&q->timeout, blk_mq_rq_timer, (unsigned long) q); |
1834 | blk_queue_rq_timeout(q, 30000); | |
1835 | ||
1836 | q->nr_queues = nr_cpu_ids; | |
24d2f903 | 1837 | q->nr_hw_queues = set->nr_hw_queues; |
f14bbe77 | 1838 | q->mq_map = map; |
320ae51f JA |
1839 | |
1840 | q->queue_ctx = ctx; | |
1841 | q->queue_hw_ctx = hctxs; | |
1842 | ||
24d2f903 | 1843 | q->mq_ops = set->ops; |
94eddfbe | 1844 | q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT; |
320ae51f | 1845 | |
1be036e9 CH |
1846 | q->sg_reserved_size = INT_MAX; |
1847 | ||
6fca6a61 CH |
1848 | INIT_WORK(&q->requeue_work, blk_mq_requeue_work); |
1849 | INIT_LIST_HEAD(&q->requeue_list); | |
1850 | spin_lock_init(&q->requeue_lock); | |
1851 | ||
07068d5b JA |
1852 | if (q->nr_hw_queues > 1) |
1853 | blk_queue_make_request(q, blk_mq_make_request); | |
1854 | else | |
1855 | blk_queue_make_request(q, blk_sq_make_request); | |
1856 | ||
87ee7b11 | 1857 | blk_queue_rq_timed_out(q, blk_mq_rq_timed_out); |
24d2f903 CH |
1858 | if (set->timeout) |
1859 | blk_queue_rq_timeout(q, set->timeout); | |
320ae51f | 1860 | |
eba71768 JA |
1861 | /* |
1862 | * Do this after blk_queue_make_request() overrides it... | |
1863 | */ | |
1864 | q->nr_requests = set->queue_depth; | |
1865 | ||
24d2f903 CH |
1866 | if (set->ops->complete) |
1867 | blk_queue_softirq_done(q, set->ops->complete); | |
30a91cb4 | 1868 | |
320ae51f | 1869 | blk_mq_init_flush(q); |
24d2f903 | 1870 | blk_mq_init_cpu_queues(q, set->nr_hw_queues); |
320ae51f | 1871 | |
24d2f903 CH |
1872 | q->flush_rq = kzalloc(round_up(sizeof(struct request) + |
1873 | set->cmd_size, cache_line_size()), | |
1874 | GFP_KERNEL); | |
18741986 | 1875 | if (!q->flush_rq) |
320ae51f JA |
1876 | goto err_hw; |
1877 | ||
24d2f903 | 1878 | if (blk_mq_init_hw_queues(q, set)) |
18741986 CH |
1879 | goto err_flush_rq; |
1880 | ||
320ae51f JA |
1881 | mutex_lock(&all_q_mutex); |
1882 | list_add_tail(&q->all_q_node, &all_q_list); | |
1883 | mutex_unlock(&all_q_mutex); | |
1884 | ||
0d2602ca JA |
1885 | blk_mq_add_queue_tag_set(set, q); |
1886 | ||
484b4061 JA |
1887 | blk_mq_map_swqueue(q); |
1888 | ||
320ae51f | 1889 | return q; |
18741986 CH |
1890 | |
1891 | err_flush_rq: | |
1892 | kfree(q->flush_rq); | |
320ae51f | 1893 | err_hw: |
320ae51f JA |
1894 | blk_cleanup_queue(q); |
1895 | err_hctxs: | |
f14bbe77 | 1896 | kfree(map); |
24d2f903 | 1897 | for (i = 0; i < set->nr_hw_queues; i++) { |
320ae51f JA |
1898 | if (!hctxs[i]) |
1899 | break; | |
e4043dcf | 1900 | free_cpumask_var(hctxs[i]->cpumask); |
24d2f903 | 1901 | set->ops->free_hctx(hctxs[i], i); |
320ae51f | 1902 | } |
f14bbe77 | 1903 | err_map: |
320ae51f JA |
1904 | kfree(hctxs); |
1905 | err_percpu: | |
1906 | free_percpu(ctx); | |
1907 | return ERR_PTR(-ENOMEM); | |
1908 | } | |
1909 | EXPORT_SYMBOL(blk_mq_init_queue); | |
1910 | ||
1911 | void blk_mq_free_queue(struct request_queue *q) | |
1912 | { | |
624dbe47 | 1913 | struct blk_mq_tag_set *set = q->tag_set; |
320ae51f | 1914 | |
0d2602ca JA |
1915 | blk_mq_del_queue_tag_set(q); |
1916 | ||
624dbe47 ML |
1917 | blk_mq_exit_hw_queues(q, set, set->nr_hw_queues); |
1918 | blk_mq_free_hw_queues(q, set); | |
320ae51f | 1919 | |
3d2936f4 ML |
1920 | percpu_counter_destroy(&q->mq_usage_counter); |
1921 | ||
320ae51f JA |
1922 | free_percpu(q->queue_ctx); |
1923 | kfree(q->queue_hw_ctx); | |
1924 | kfree(q->mq_map); | |
1925 | ||
1926 | q->queue_ctx = NULL; | |
1927 | q->queue_hw_ctx = NULL; | |
1928 | q->mq_map = NULL; | |
1929 | ||
1930 | mutex_lock(&all_q_mutex); | |
1931 | list_del_init(&q->all_q_node); | |
1932 | mutex_unlock(&all_q_mutex); | |
1933 | } | |
320ae51f JA |
1934 | |
1935 | /* Basically redo blk_mq_init_queue with queue frozen */ | |
f618ef7c | 1936 | static void blk_mq_queue_reinit(struct request_queue *q) |
320ae51f JA |
1937 | { |
1938 | blk_mq_freeze_queue(q); | |
1939 | ||
1940 | blk_mq_update_queue_map(q->mq_map, q->nr_hw_queues); | |
1941 | ||
1942 | /* | |
1943 | * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe | |
1944 | * we should change hctx numa_node according to new topology (this | |
1945 | * involves free and re-allocate memory, worthy doing?) | |
1946 | */ | |
1947 | ||
1948 | blk_mq_map_swqueue(q); | |
1949 | ||
1950 | blk_mq_unfreeze_queue(q); | |
1951 | } | |
1952 | ||
f618ef7c PG |
1953 | static int blk_mq_queue_reinit_notify(struct notifier_block *nb, |
1954 | unsigned long action, void *hcpu) | |
320ae51f JA |
1955 | { |
1956 | struct request_queue *q; | |
1957 | ||
1958 | /* | |
9fccfed8 JA |
1959 | * Before new mappings are established, hotadded cpu might already |
1960 | * start handling requests. This doesn't break anything as we map | |
1961 | * offline CPUs to first hardware queue. We will re-init the queue | |
1962 | * below to get optimal settings. | |
320ae51f JA |
1963 | */ |
1964 | if (action != CPU_DEAD && action != CPU_DEAD_FROZEN && | |
1965 | action != CPU_ONLINE && action != CPU_ONLINE_FROZEN) | |
1966 | return NOTIFY_OK; | |
1967 | ||
1968 | mutex_lock(&all_q_mutex); | |
1969 | list_for_each_entry(q, &all_q_list, all_q_node) | |
1970 | blk_mq_queue_reinit(q); | |
1971 | mutex_unlock(&all_q_mutex); | |
1972 | return NOTIFY_OK; | |
1973 | } | |
1974 | ||
24d2f903 CH |
1975 | int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set) |
1976 | { | |
1977 | int i; | |
1978 | ||
1979 | if (!set->nr_hw_queues) | |
1980 | return -EINVAL; | |
1981 | if (!set->queue_depth || set->queue_depth > BLK_MQ_MAX_DEPTH) | |
1982 | return -EINVAL; | |
1983 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) | |
1984 | return -EINVAL; | |
1985 | ||
1986 | if (!set->nr_hw_queues || | |
1987 | !set->ops->queue_rq || !set->ops->map_queue || | |
1988 | !set->ops->alloc_hctx || !set->ops->free_hctx) | |
1989 | return -EINVAL; | |
1990 | ||
1991 | ||
48479005 ML |
1992 | set->tags = kmalloc_node(set->nr_hw_queues * |
1993 | sizeof(struct blk_mq_tags *), | |
24d2f903 CH |
1994 | GFP_KERNEL, set->numa_node); |
1995 | if (!set->tags) | |
1996 | goto out; | |
1997 | ||
1998 | for (i = 0; i < set->nr_hw_queues; i++) { | |
1999 | set->tags[i] = blk_mq_init_rq_map(set, i); | |
2000 | if (!set->tags[i]) | |
2001 | goto out_unwind; | |
2002 | } | |
2003 | ||
0d2602ca JA |
2004 | mutex_init(&set->tag_list_lock); |
2005 | INIT_LIST_HEAD(&set->tag_list); | |
2006 | ||
24d2f903 CH |
2007 | return 0; |
2008 | ||
2009 | out_unwind: | |
2010 | while (--i >= 0) | |
2011 | blk_mq_free_rq_map(set, set->tags[i], i); | |
2012 | out: | |
2013 | return -ENOMEM; | |
2014 | } | |
2015 | EXPORT_SYMBOL(blk_mq_alloc_tag_set); | |
2016 | ||
2017 | void blk_mq_free_tag_set(struct blk_mq_tag_set *set) | |
2018 | { | |
2019 | int i; | |
2020 | ||
484b4061 JA |
2021 | for (i = 0; i < set->nr_hw_queues; i++) { |
2022 | if (set->tags[i]) | |
2023 | blk_mq_free_rq_map(set, set->tags[i], i); | |
2024 | } | |
2025 | ||
981bd189 | 2026 | kfree(set->tags); |
24d2f903 CH |
2027 | } |
2028 | EXPORT_SYMBOL(blk_mq_free_tag_set); | |
2029 | ||
e3a2b3f9 JA |
2030 | int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr) |
2031 | { | |
2032 | struct blk_mq_tag_set *set = q->tag_set; | |
2033 | struct blk_mq_hw_ctx *hctx; | |
2034 | int i, ret; | |
2035 | ||
2036 | if (!set || nr > set->queue_depth) | |
2037 | return -EINVAL; | |
2038 | ||
2039 | ret = 0; | |
2040 | queue_for_each_hw_ctx(q, hctx, i) { | |
2041 | ret = blk_mq_tag_update_depth(hctx->tags, nr); | |
2042 | if (ret) | |
2043 | break; | |
2044 | } | |
2045 | ||
2046 | if (!ret) | |
2047 | q->nr_requests = nr; | |
2048 | ||
2049 | return ret; | |
2050 | } | |
2051 | ||
676141e4 JA |
2052 | void blk_mq_disable_hotplug(void) |
2053 | { | |
2054 | mutex_lock(&all_q_mutex); | |
2055 | } | |
2056 | ||
2057 | void blk_mq_enable_hotplug(void) | |
2058 | { | |
2059 | mutex_unlock(&all_q_mutex); | |
2060 | } | |
2061 | ||
320ae51f JA |
2062 | static int __init blk_mq_init(void) |
2063 | { | |
320ae51f JA |
2064 | blk_mq_cpu_init(); |
2065 | ||
2066 | /* Must be called after percpu_counter_hotcpu_callback() */ | |
2067 | hotcpu_notifier(blk_mq_queue_reinit_notify, -10); | |
2068 | ||
2069 | return 0; | |
2070 | } | |
2071 | subsys_initcall(blk_mq_init); |