]>
Commit | Line | Data |
---|---|---|
1 | /* | |
2 | * Tag allocation using scalable bitmaps. Uses active queue tracking to support | |
3 | * fairer distribution of tags between multiple submitters when a shared tag map | |
4 | * is used. | |
5 | * | |
6 | * Copyright (C) 2013-2014 Jens Axboe | |
7 | */ | |
8 | #include <linux/kernel.h> | |
9 | #include <linux/module.h> | |
10 | ||
11 | #include <linux/blk-mq.h> | |
12 | #include "blk.h" | |
13 | #include "blk-mq.h" | |
14 | #include "blk-mq-tag.h" | |
15 | ||
16 | bool blk_mq_has_free_tags(struct blk_mq_tags *tags) | |
17 | { | |
18 | if (!tags) | |
19 | return true; | |
20 | ||
21 | return sbitmap_any_bit_clear(&tags->bitmap_tags.sb); | |
22 | } | |
23 | ||
24 | /* | |
25 | * If a previously inactive queue goes active, bump the active user count. | |
26 | */ | |
27 | bool __blk_mq_tag_busy(struct blk_mq_hw_ctx *hctx) | |
28 | { | |
29 | if (!test_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state) && | |
30 | !test_and_set_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state)) | |
31 | atomic_inc(&hctx->tags->active_queues); | |
32 | ||
33 | return true; | |
34 | } | |
35 | ||
36 | /* | |
37 | * Wakeup all potentially sleeping on tags | |
38 | */ | |
39 | void blk_mq_tag_wakeup_all(struct blk_mq_tags *tags, bool include_reserve) | |
40 | { | |
41 | sbitmap_queue_wake_all(&tags->bitmap_tags); | |
42 | if (include_reserve) | |
43 | sbitmap_queue_wake_all(&tags->breserved_tags); | |
44 | } | |
45 | ||
46 | /* | |
47 | * If a previously busy queue goes inactive, potential waiters could now | |
48 | * be allowed to queue. Wake them up and check. | |
49 | */ | |
50 | void __blk_mq_tag_idle(struct blk_mq_hw_ctx *hctx) | |
51 | { | |
52 | struct blk_mq_tags *tags = hctx->tags; | |
53 | ||
54 | if (!test_and_clear_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state)) | |
55 | return; | |
56 | ||
57 | atomic_dec(&tags->active_queues); | |
58 | ||
59 | blk_mq_tag_wakeup_all(tags, false); | |
60 | } | |
61 | ||
62 | /* | |
63 | * For shared tag users, we track the number of currently active users | |
64 | * and attempt to provide a fair share of the tag depth for each of them. | |
65 | */ | |
66 | static inline bool hctx_may_queue(struct blk_mq_hw_ctx *hctx, | |
67 | struct sbitmap_queue *bt) | |
68 | { | |
69 | unsigned int depth, users; | |
70 | ||
71 | if (!hctx || !(hctx->flags & BLK_MQ_F_TAG_SHARED)) | |
72 | return true; | |
73 | if (!test_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state)) | |
74 | return true; | |
75 | ||
76 | /* | |
77 | * Don't try dividing an ant | |
78 | */ | |
79 | if (bt->sb.depth == 1) | |
80 | return true; | |
81 | ||
82 | users = atomic_read(&hctx->tags->active_queues); | |
83 | if (!users) | |
84 | return true; | |
85 | ||
86 | /* | |
87 | * Allow at least some tags | |
88 | */ | |
89 | depth = max((bt->sb.depth + users - 1) / users, 4U); | |
90 | return atomic_read(&hctx->nr_active) < depth; | |
91 | } | |
92 | ||
93 | static int __blk_mq_get_tag(struct blk_mq_alloc_data *data, | |
94 | struct sbitmap_queue *bt) | |
95 | { | |
96 | if (!(data->flags & BLK_MQ_REQ_INTERNAL) && | |
97 | !hctx_may_queue(data->hctx, bt)) | |
98 | return -1; | |
99 | if (data->shallow_depth) | |
100 | return __sbitmap_queue_get_shallow(bt, data->shallow_depth); | |
101 | else | |
102 | return __sbitmap_queue_get(bt); | |
103 | } | |
104 | ||
105 | unsigned int blk_mq_get_tag(struct blk_mq_alloc_data *data) | |
106 | { | |
107 | struct blk_mq_tags *tags = blk_mq_tags_from_data(data); | |
108 | struct sbitmap_queue *bt; | |
109 | struct sbq_wait_state *ws; | |
110 | DEFINE_WAIT(wait); | |
111 | unsigned int tag_offset; | |
112 | bool drop_ctx; | |
113 | int tag; | |
114 | ||
115 | if (data->flags & BLK_MQ_REQ_RESERVED) { | |
116 | if (unlikely(!tags->nr_reserved_tags)) { | |
117 | WARN_ON_ONCE(1); | |
118 | return BLK_MQ_TAG_FAIL; | |
119 | } | |
120 | bt = &tags->breserved_tags; | |
121 | tag_offset = 0; | |
122 | } else { | |
123 | bt = &tags->bitmap_tags; | |
124 | tag_offset = tags->nr_reserved_tags; | |
125 | } | |
126 | ||
127 | tag = __blk_mq_get_tag(data, bt); | |
128 | if (tag != -1) | |
129 | goto found_tag; | |
130 | ||
131 | if (data->flags & BLK_MQ_REQ_NOWAIT) | |
132 | return BLK_MQ_TAG_FAIL; | |
133 | ||
134 | ws = bt_wait_ptr(bt, data->hctx); | |
135 | drop_ctx = data->ctx == NULL; | |
136 | do { | |
137 | prepare_to_wait(&ws->wait, &wait, TASK_UNINTERRUPTIBLE); | |
138 | ||
139 | tag = __blk_mq_get_tag(data, bt); | |
140 | if (tag != -1) | |
141 | break; | |
142 | ||
143 | /* | |
144 | * We're out of tags on this hardware queue, kick any | |
145 | * pending IO submits before going to sleep waiting for | |
146 | * some to complete. | |
147 | */ | |
148 | blk_mq_run_hw_queue(data->hctx, false); | |
149 | ||
150 | /* | |
151 | * Retry tag allocation after running the hardware queue, | |
152 | * as running the queue may also have found completions. | |
153 | */ | |
154 | tag = __blk_mq_get_tag(data, bt); | |
155 | if (tag != -1) | |
156 | break; | |
157 | ||
158 | if (data->ctx) | |
159 | blk_mq_put_ctx(data->ctx); | |
160 | ||
161 | io_schedule(); | |
162 | ||
163 | data->ctx = blk_mq_get_ctx(data->q); | |
164 | data->hctx = blk_mq_map_queue(data->q, data->ctx->cpu); | |
165 | tags = blk_mq_tags_from_data(data); | |
166 | if (data->flags & BLK_MQ_REQ_RESERVED) | |
167 | bt = &tags->breserved_tags; | |
168 | else | |
169 | bt = &tags->bitmap_tags; | |
170 | ||
171 | finish_wait(&ws->wait, &wait); | |
172 | ws = bt_wait_ptr(bt, data->hctx); | |
173 | } while (1); | |
174 | ||
175 | if (drop_ctx && data->ctx) | |
176 | blk_mq_put_ctx(data->ctx); | |
177 | ||
178 | finish_wait(&ws->wait, &wait); | |
179 | ||
180 | found_tag: | |
181 | return tag + tag_offset; | |
182 | } | |
183 | ||
184 | void blk_mq_put_tag(struct blk_mq_hw_ctx *hctx, struct blk_mq_tags *tags, | |
185 | struct blk_mq_ctx *ctx, unsigned int tag) | |
186 | { | |
187 | if (!blk_mq_tag_is_reserved(tags, tag)) { | |
188 | const int real_tag = tag - tags->nr_reserved_tags; | |
189 | ||
190 | BUG_ON(real_tag >= tags->nr_tags); | |
191 | sbitmap_queue_clear(&tags->bitmap_tags, real_tag, ctx->cpu); | |
192 | } else { | |
193 | BUG_ON(tag >= tags->nr_reserved_tags); | |
194 | sbitmap_queue_clear(&tags->breserved_tags, tag, ctx->cpu); | |
195 | } | |
196 | } | |
197 | ||
198 | struct bt_iter_data { | |
199 | struct blk_mq_hw_ctx *hctx; | |
200 | busy_iter_fn *fn; | |
201 | void *data; | |
202 | bool reserved; | |
203 | }; | |
204 | ||
205 | static bool bt_iter(struct sbitmap *bitmap, unsigned int bitnr, void *data) | |
206 | { | |
207 | struct bt_iter_data *iter_data = data; | |
208 | struct blk_mq_hw_ctx *hctx = iter_data->hctx; | |
209 | struct blk_mq_tags *tags = hctx->tags; | |
210 | bool reserved = iter_data->reserved; | |
211 | struct request *rq; | |
212 | ||
213 | if (!reserved) | |
214 | bitnr += tags->nr_reserved_tags; | |
215 | rq = tags->rqs[bitnr]; | |
216 | ||
217 | /* | |
218 | * We can hit rq == NULL here, because the tagging functions | |
219 | * test and set the bit before assining ->rqs[]. | |
220 | */ | |
221 | if (rq && rq->q == hctx->queue) | |
222 | iter_data->fn(hctx, rq, iter_data->data, reserved); | |
223 | return true; | |
224 | } | |
225 | ||
226 | static void bt_for_each(struct blk_mq_hw_ctx *hctx, struct sbitmap_queue *bt, | |
227 | busy_iter_fn *fn, void *data, bool reserved) | |
228 | { | |
229 | struct bt_iter_data iter_data = { | |
230 | .hctx = hctx, | |
231 | .fn = fn, | |
232 | .data = data, | |
233 | .reserved = reserved, | |
234 | }; | |
235 | ||
236 | sbitmap_for_each_set(&bt->sb, bt_iter, &iter_data); | |
237 | } | |
238 | ||
239 | struct bt_tags_iter_data { | |
240 | struct blk_mq_tags *tags; | |
241 | busy_tag_iter_fn *fn; | |
242 | void *data; | |
243 | bool reserved; | |
244 | }; | |
245 | ||
246 | static bool bt_tags_iter(struct sbitmap *bitmap, unsigned int bitnr, void *data) | |
247 | { | |
248 | struct bt_tags_iter_data *iter_data = data; | |
249 | struct blk_mq_tags *tags = iter_data->tags; | |
250 | bool reserved = iter_data->reserved; | |
251 | struct request *rq; | |
252 | ||
253 | if (!reserved) | |
254 | bitnr += tags->nr_reserved_tags; | |
255 | ||
256 | /* | |
257 | * We can hit rq == NULL here, because the tagging functions | |
258 | * test and set the bit before assining ->rqs[]. | |
259 | */ | |
260 | rq = tags->rqs[bitnr]; | |
261 | if (rq) | |
262 | iter_data->fn(rq, iter_data->data, reserved); | |
263 | ||
264 | return true; | |
265 | } | |
266 | ||
267 | static void bt_tags_for_each(struct blk_mq_tags *tags, struct sbitmap_queue *bt, | |
268 | busy_tag_iter_fn *fn, void *data, bool reserved) | |
269 | { | |
270 | struct bt_tags_iter_data iter_data = { | |
271 | .tags = tags, | |
272 | .fn = fn, | |
273 | .data = data, | |
274 | .reserved = reserved, | |
275 | }; | |
276 | ||
277 | if (tags->rqs) | |
278 | sbitmap_for_each_set(&bt->sb, bt_tags_iter, &iter_data); | |
279 | } | |
280 | ||
281 | static void blk_mq_all_tag_busy_iter(struct blk_mq_tags *tags, | |
282 | busy_tag_iter_fn *fn, void *priv) | |
283 | { | |
284 | if (tags->nr_reserved_tags) | |
285 | bt_tags_for_each(tags, &tags->breserved_tags, fn, priv, true); | |
286 | bt_tags_for_each(tags, &tags->bitmap_tags, fn, priv, false); | |
287 | } | |
288 | ||
289 | void blk_mq_tagset_busy_iter(struct blk_mq_tag_set *tagset, | |
290 | busy_tag_iter_fn *fn, void *priv) | |
291 | { | |
292 | int i; | |
293 | ||
294 | for (i = 0; i < tagset->nr_hw_queues; i++) { | |
295 | if (tagset->tags && tagset->tags[i]) | |
296 | blk_mq_all_tag_busy_iter(tagset->tags[i], fn, priv); | |
297 | } | |
298 | } | |
299 | EXPORT_SYMBOL(blk_mq_tagset_busy_iter); | |
300 | ||
301 | int blk_mq_tagset_iter(struct blk_mq_tag_set *set, void *data, | |
302 | int (fn)(void *, struct request *)) | |
303 | { | |
304 | int i, j, ret = 0; | |
305 | ||
306 | if (WARN_ON_ONCE(!fn)) | |
307 | goto out; | |
308 | ||
309 | for (i = 0; i < set->nr_hw_queues; i++) { | |
310 | struct blk_mq_tags *tags = set->tags[i]; | |
311 | ||
312 | if (!tags) | |
313 | continue; | |
314 | ||
315 | for (j = 0; j < tags->nr_tags; j++) { | |
316 | if (!tags->static_rqs[j]) | |
317 | continue; | |
318 | ||
319 | ret = fn(data, tags->static_rqs[j]); | |
320 | if (ret) | |
321 | goto out; | |
322 | } | |
323 | } | |
324 | ||
325 | out: | |
326 | return ret; | |
327 | } | |
328 | EXPORT_SYMBOL_GPL(blk_mq_tagset_iter); | |
329 | ||
330 | void blk_mq_queue_tag_busy_iter(struct request_queue *q, busy_iter_fn *fn, | |
331 | void *priv) | |
332 | { | |
333 | struct blk_mq_hw_ctx *hctx; | |
334 | int i; | |
335 | ||
336 | ||
337 | queue_for_each_hw_ctx(q, hctx, i) { | |
338 | struct blk_mq_tags *tags = hctx->tags; | |
339 | ||
340 | /* | |
341 | * If not software queues are currently mapped to this | |
342 | * hardware queue, there's nothing to check | |
343 | */ | |
344 | if (!blk_mq_hw_queue_mapped(hctx)) | |
345 | continue; | |
346 | ||
347 | if (tags->nr_reserved_tags) | |
348 | bt_for_each(hctx, &tags->breserved_tags, fn, priv, true); | |
349 | bt_for_each(hctx, &tags->bitmap_tags, fn, priv, false); | |
350 | } | |
351 | ||
352 | } | |
353 | ||
354 | static int bt_alloc(struct sbitmap_queue *bt, unsigned int depth, | |
355 | bool round_robin, int node) | |
356 | { | |
357 | return sbitmap_queue_init_node(bt, depth, -1, round_robin, GFP_KERNEL, | |
358 | node); | |
359 | } | |
360 | ||
361 | static struct blk_mq_tags *blk_mq_init_bitmap_tags(struct blk_mq_tags *tags, | |
362 | int node, int alloc_policy) | |
363 | { | |
364 | unsigned int depth = tags->nr_tags - tags->nr_reserved_tags; | |
365 | bool round_robin = alloc_policy == BLK_TAG_ALLOC_RR; | |
366 | ||
367 | if (bt_alloc(&tags->bitmap_tags, depth, round_robin, node)) | |
368 | goto free_tags; | |
369 | if (bt_alloc(&tags->breserved_tags, tags->nr_reserved_tags, round_robin, | |
370 | node)) | |
371 | goto free_bitmap_tags; | |
372 | ||
373 | return tags; | |
374 | free_bitmap_tags: | |
375 | sbitmap_queue_free(&tags->bitmap_tags); | |
376 | free_tags: | |
377 | kfree(tags); | |
378 | return NULL; | |
379 | } | |
380 | ||
381 | struct blk_mq_tags *blk_mq_init_tags(unsigned int total_tags, | |
382 | unsigned int reserved_tags, | |
383 | int node, int alloc_policy) | |
384 | { | |
385 | struct blk_mq_tags *tags; | |
386 | ||
387 | if (total_tags > BLK_MQ_TAG_MAX) { | |
388 | pr_err("blk-mq: tag depth too large\n"); | |
389 | return NULL; | |
390 | } | |
391 | ||
392 | tags = kzalloc_node(sizeof(*tags), GFP_KERNEL, node); | |
393 | if (!tags) | |
394 | return NULL; | |
395 | ||
396 | tags->nr_tags = total_tags; | |
397 | tags->nr_reserved_tags = reserved_tags; | |
398 | ||
399 | return blk_mq_init_bitmap_tags(tags, node, alloc_policy); | |
400 | } | |
401 | ||
402 | void blk_mq_free_tags(struct blk_mq_tags *tags) | |
403 | { | |
404 | sbitmap_queue_free(&tags->bitmap_tags); | |
405 | sbitmap_queue_free(&tags->breserved_tags); | |
406 | kfree(tags); | |
407 | } | |
408 | ||
409 | int blk_mq_tag_update_depth(struct blk_mq_hw_ctx *hctx, | |
410 | struct blk_mq_tags **tagsptr, unsigned int tdepth, | |
411 | bool can_grow) | |
412 | { | |
413 | struct blk_mq_tags *tags = *tagsptr; | |
414 | ||
415 | if (tdepth <= tags->nr_reserved_tags) | |
416 | return -EINVAL; | |
417 | ||
418 | tdepth -= tags->nr_reserved_tags; | |
419 | ||
420 | /* | |
421 | * If we are allowed to grow beyond the original size, allocate | |
422 | * a new set of tags before freeing the old one. | |
423 | */ | |
424 | if (tdepth > tags->nr_tags) { | |
425 | struct blk_mq_tag_set *set = hctx->queue->tag_set; | |
426 | struct blk_mq_tags *new; | |
427 | bool ret; | |
428 | ||
429 | if (!can_grow) | |
430 | return -EINVAL; | |
431 | ||
432 | /* | |
433 | * We need some sort of upper limit, set it high enough that | |
434 | * no valid use cases should require more. | |
435 | */ | |
436 | if (tdepth > 16 * BLKDEV_MAX_RQ) | |
437 | return -EINVAL; | |
438 | ||
439 | new = blk_mq_alloc_rq_map(set, hctx->queue_num, tdepth, 0); | |
440 | if (!new) | |
441 | return -ENOMEM; | |
442 | ret = blk_mq_alloc_rqs(set, new, hctx->queue_num, tdepth); | |
443 | if (ret) { | |
444 | blk_mq_free_rq_map(new); | |
445 | return -ENOMEM; | |
446 | } | |
447 | ||
448 | blk_mq_free_rqs(set, *tagsptr, hctx->queue_num); | |
449 | blk_mq_free_rq_map(*tagsptr); | |
450 | *tagsptr = new; | |
451 | } else { | |
452 | /* | |
453 | * Don't need (or can't) update reserved tags here, they | |
454 | * remain static and should never need resizing. | |
455 | */ | |
456 | sbitmap_queue_resize(&tags->bitmap_tags, tdepth); | |
457 | } | |
458 | ||
459 | return 0; | |
460 | } | |
461 | ||
462 | /** | |
463 | * blk_mq_unique_tag() - return a tag that is unique queue-wide | |
464 | * @rq: request for which to compute a unique tag | |
465 | * | |
466 | * The tag field in struct request is unique per hardware queue but not over | |
467 | * all hardware queues. Hence this function that returns a tag with the | |
468 | * hardware context index in the upper bits and the per hardware queue tag in | |
469 | * the lower bits. | |
470 | * | |
471 | * Note: When called for a request that is queued on a non-multiqueue request | |
472 | * queue, the hardware context index is set to zero. | |
473 | */ | |
474 | u32 blk_mq_unique_tag(struct request *rq) | |
475 | { | |
476 | struct request_queue *q = rq->q; | |
477 | struct blk_mq_hw_ctx *hctx; | |
478 | int hwq = 0; | |
479 | ||
480 | if (q->mq_ops) { | |
481 | hctx = blk_mq_map_queue(q, rq->mq_ctx->cpu); | |
482 | hwq = hctx->queue_num; | |
483 | } | |
484 | ||
485 | return (hwq << BLK_MQ_UNIQUE_TAG_BITS) | | |
486 | (rq->tag & BLK_MQ_UNIQUE_TAG_MASK); | |
487 | } | |
488 | EXPORT_SYMBOL(blk_mq_unique_tag); |