Commit | Line | Data |
---|---|---|
75bb4625 | 1 | /* |
88459642 OS |
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. | |
75bb4625 JA |
5 | * |
6 | * Copyright (C) 2013-2014 Jens Axboe | |
7 | */ | |
320ae51f JA |
8 | #include <linux/kernel.h> |
9 | #include <linux/module.h> | |
320ae51f JA |
10 | |
11 | #include <linux/blk-mq.h> | |
12 | #include "blk.h" | |
13 | #include "blk-mq.h" | |
14 | #include "blk-mq-tag.h" | |
15 | ||
320ae51f JA |
16 | bool blk_mq_has_free_tags(struct blk_mq_tags *tags) |
17 | { | |
4bb659b1 JA |
18 | if (!tags) |
19 | return true; | |
20 | ||
88459642 | 21 | return sbitmap_any_bit_clear(&tags->bitmap_tags.sb); |
0d2602ca JA |
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 | /* | |
aed3ea94 | 37 | * Wakeup all potentially sleeping on tags |
0d2602ca | 38 | */ |
aed3ea94 | 39 | void blk_mq_tag_wakeup_all(struct blk_mq_tags *tags, bool include_reserve) |
0d2602ca | 40 | { |
88459642 OS |
41 | sbitmap_queue_wake_all(&tags->bitmap_tags); |
42 | if (include_reserve) | |
43 | sbitmap_queue_wake_all(&tags->breserved_tags); | |
0d2602ca JA |
44 | } |
45 | ||
e3a2b3f9 JA |
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 | ||
aed3ea94 | 59 | blk_mq_tag_wakeup_all(tags, false); |
e3a2b3f9 JA |
60 | } |
61 | ||
0d2602ca JA |
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, | |
88459642 | 67 | struct sbitmap_queue *bt) |
0d2602ca JA |
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 | */ | |
88459642 | 79 | if (bt->sb.depth == 1) |
0d2602ca JA |
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 | */ | |
88459642 | 89 | depth = max((bt->sb.depth + users - 1) / users, 4U); |
0d2602ca JA |
90 | return atomic_read(&hctx->nr_active) < depth; |
91 | } | |
92 | ||
f4a644db | 93 | static int __bt_get(struct blk_mq_hw_ctx *hctx, struct sbitmap_queue *bt) |
4bb659b1 | 94 | { |
0d2602ca JA |
95 | if (!hctx_may_queue(hctx, bt)) |
96 | return -1; | |
f4a644db | 97 | return __sbitmap_queue_get(bt); |
4bb659b1 JA |
98 | } |
99 | ||
40aabb67 OS |
100 | static int bt_get(struct blk_mq_alloc_data *data, struct sbitmap_queue *bt, |
101 | struct blk_mq_hw_ctx *hctx, struct blk_mq_tags *tags) | |
320ae51f | 102 | { |
88459642 | 103 | struct sbq_wait_state *ws; |
4bb659b1 | 104 | DEFINE_WAIT(wait); |
320ae51f JA |
105 | int tag; |
106 | ||
f4a644db | 107 | tag = __bt_get(hctx, bt); |
4bb659b1 JA |
108 | if (tag != -1) |
109 | return tag; | |
110 | ||
6f3b0e8b | 111 | if (data->flags & BLK_MQ_REQ_NOWAIT) |
4bb659b1 JA |
112 | return -1; |
113 | ||
88459642 | 114 | ws = bt_wait_ptr(bt, hctx); |
4bb659b1 | 115 | do { |
88459642 | 116 | prepare_to_wait(&ws->wait, &wait, TASK_UNINTERRUPTIBLE); |
4bb659b1 | 117 | |
f4a644db | 118 | tag = __bt_get(hctx, bt); |
4bb659b1 JA |
119 | if (tag != -1) |
120 | break; | |
121 | ||
b3223207 BVA |
122 | /* |
123 | * We're out of tags on this hardware queue, kick any | |
124 | * pending IO submits before going to sleep waiting for | |
bc188d81 SB |
125 | * some to complete. Note that hctx can be NULL here for |
126 | * reserved tag allocation. | |
b3223207 | 127 | */ |
bc188d81 SB |
128 | if (hctx) |
129 | blk_mq_run_hw_queue(hctx, false); | |
b3223207 | 130 | |
080ff351 JA |
131 | /* |
132 | * Retry tag allocation after running the hardware queue, | |
133 | * as running the queue may also have found completions. | |
134 | */ | |
f4a644db | 135 | tag = __bt_get(hctx, bt); |
080ff351 JA |
136 | if (tag != -1) |
137 | break; | |
138 | ||
cb96a42c ML |
139 | blk_mq_put_ctx(data->ctx); |
140 | ||
4bb659b1 | 141 | io_schedule(); |
cb96a42c ML |
142 | |
143 | data->ctx = blk_mq_get_ctx(data->q); | |
7d7e0f90 | 144 | data->hctx = blk_mq_map_queue(data->q, data->ctx->cpu); |
6f3b0e8b | 145 | if (data->flags & BLK_MQ_REQ_RESERVED) { |
cb96a42c ML |
146 | bt = &data->hctx->tags->breserved_tags; |
147 | } else { | |
cb96a42c ML |
148 | hctx = data->hctx; |
149 | bt = &hctx->tags->bitmap_tags; | |
150 | } | |
88459642 OS |
151 | finish_wait(&ws->wait, &wait); |
152 | ws = bt_wait_ptr(bt, hctx); | |
4bb659b1 JA |
153 | } while (1); |
154 | ||
88459642 | 155 | finish_wait(&ws->wait, &wait); |
4bb659b1 JA |
156 | return tag; |
157 | } | |
158 | ||
cb96a42c | 159 | static unsigned int __blk_mq_get_tag(struct blk_mq_alloc_data *data) |
4bb659b1 JA |
160 | { |
161 | int tag; | |
162 | ||
cb96a42c | 163 | tag = bt_get(data, &data->hctx->tags->bitmap_tags, data->hctx, |
40aabb67 | 164 | data->hctx->tags); |
4bb659b1 | 165 | if (tag >= 0) |
cb96a42c | 166 | return tag + data->hctx->tags->nr_reserved_tags; |
4bb659b1 JA |
167 | |
168 | return BLK_MQ_TAG_FAIL; | |
320ae51f JA |
169 | } |
170 | ||
cb96a42c | 171 | static unsigned int __blk_mq_get_reserved_tag(struct blk_mq_alloc_data *data) |
320ae51f | 172 | { |
40aabb67 | 173 | int tag; |
320ae51f | 174 | |
cb96a42c | 175 | if (unlikely(!data->hctx->tags->nr_reserved_tags)) { |
320ae51f JA |
176 | WARN_ON_ONCE(1); |
177 | return BLK_MQ_TAG_FAIL; | |
178 | } | |
179 | ||
40aabb67 OS |
180 | tag = bt_get(data, &data->hctx->tags->breserved_tags, NULL, |
181 | data->hctx->tags); | |
320ae51f JA |
182 | if (tag < 0) |
183 | return BLK_MQ_TAG_FAIL; | |
4bb659b1 | 184 | |
320ae51f JA |
185 | return tag; |
186 | } | |
187 | ||
cb96a42c | 188 | unsigned int blk_mq_get_tag(struct blk_mq_alloc_data *data) |
320ae51f | 189 | { |
6f3b0e8b CH |
190 | if (data->flags & BLK_MQ_REQ_RESERVED) |
191 | return __blk_mq_get_reserved_tag(data); | |
192 | return __blk_mq_get_tag(data); | |
320ae51f JA |
193 | } |
194 | ||
40aabb67 OS |
195 | void blk_mq_put_tag(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx, |
196 | unsigned int tag) | |
320ae51f | 197 | { |
0d2602ca JA |
198 | struct blk_mq_tags *tags = hctx->tags; |
199 | ||
4bb659b1 JA |
200 | if (tag >= tags->nr_reserved_tags) { |
201 | const int real_tag = tag - tags->nr_reserved_tags; | |
202 | ||
70114c39 | 203 | BUG_ON(real_tag >= tags->nr_tags); |
f4a644db | 204 | sbitmap_queue_clear(&tags->bitmap_tags, real_tag, ctx->cpu); |
70114c39 JA |
205 | } else { |
206 | BUG_ON(tag >= tags->nr_reserved_tags); | |
f4a644db | 207 | sbitmap_queue_clear(&tags->breserved_tags, tag, ctx->cpu); |
70114c39 | 208 | } |
320ae51f JA |
209 | } |
210 | ||
88459642 OS |
211 | struct bt_iter_data { |
212 | struct blk_mq_hw_ctx *hctx; | |
213 | busy_iter_fn *fn; | |
214 | void *data; | |
215 | bool reserved; | |
216 | }; | |
217 | ||
218 | static bool bt_iter(struct sbitmap *bitmap, unsigned int bitnr, void *data) | |
320ae51f | 219 | { |
88459642 OS |
220 | struct bt_iter_data *iter_data = data; |
221 | struct blk_mq_hw_ctx *hctx = iter_data->hctx; | |
222 | struct blk_mq_tags *tags = hctx->tags; | |
223 | bool reserved = iter_data->reserved; | |
81481eb4 | 224 | struct request *rq; |
4bb659b1 | 225 | |
88459642 OS |
226 | if (!reserved) |
227 | bitnr += tags->nr_reserved_tags; | |
228 | rq = tags->rqs[bitnr]; | |
4bb659b1 | 229 | |
88459642 OS |
230 | if (rq->q == hctx->queue) |
231 | iter_data->fn(hctx, rq, iter_data->data, reserved); | |
232 | return true; | |
233 | } | |
4bb659b1 | 234 | |
88459642 OS |
235 | static void bt_for_each(struct blk_mq_hw_ctx *hctx, struct sbitmap_queue *bt, |
236 | busy_iter_fn *fn, void *data, bool reserved) | |
237 | { | |
238 | struct bt_iter_data iter_data = { | |
239 | .hctx = hctx, | |
240 | .fn = fn, | |
241 | .data = data, | |
242 | .reserved = reserved, | |
243 | }; | |
244 | ||
245 | sbitmap_for_each_set(&bt->sb, bt_iter, &iter_data); | |
320ae51f JA |
246 | } |
247 | ||
88459642 OS |
248 | struct bt_tags_iter_data { |
249 | struct blk_mq_tags *tags; | |
250 | busy_tag_iter_fn *fn; | |
251 | void *data; | |
252 | bool reserved; | |
253 | }; | |
254 | ||
255 | static bool bt_tags_iter(struct sbitmap *bitmap, unsigned int bitnr, void *data) | |
f26cdc85 | 256 | { |
88459642 OS |
257 | struct bt_tags_iter_data *iter_data = data; |
258 | struct blk_mq_tags *tags = iter_data->tags; | |
259 | bool reserved = iter_data->reserved; | |
f26cdc85 | 260 | struct request *rq; |
f26cdc85 | 261 | |
88459642 OS |
262 | if (!reserved) |
263 | bitnr += tags->nr_reserved_tags; | |
264 | rq = tags->rqs[bitnr]; | |
f26cdc85 | 265 | |
88459642 OS |
266 | iter_data->fn(rq, iter_data->data, reserved); |
267 | return true; | |
268 | } | |
269 | ||
270 | static void bt_tags_for_each(struct blk_mq_tags *tags, struct sbitmap_queue *bt, | |
271 | busy_tag_iter_fn *fn, void *data, bool reserved) | |
272 | { | |
273 | struct bt_tags_iter_data iter_data = { | |
274 | .tags = tags, | |
275 | .fn = fn, | |
276 | .data = data, | |
277 | .reserved = reserved, | |
278 | }; | |
279 | ||
280 | if (tags->rqs) | |
281 | sbitmap_for_each_set(&bt->sb, bt_tags_iter, &iter_data); | |
f26cdc85 KB |
282 | } |
283 | ||
e8f1e163 SG |
284 | static void blk_mq_all_tag_busy_iter(struct blk_mq_tags *tags, |
285 | busy_tag_iter_fn *fn, void *priv) | |
f26cdc85 KB |
286 | { |
287 | if (tags->nr_reserved_tags) | |
88459642 OS |
288 | bt_tags_for_each(tags, &tags->breserved_tags, fn, priv, true); |
289 | bt_tags_for_each(tags, &tags->bitmap_tags, fn, priv, false); | |
f26cdc85 | 290 | } |
f26cdc85 | 291 | |
e0489487 SG |
292 | void blk_mq_tagset_busy_iter(struct blk_mq_tag_set *tagset, |
293 | busy_tag_iter_fn *fn, void *priv) | |
294 | { | |
295 | int i; | |
296 | ||
297 | for (i = 0; i < tagset->nr_hw_queues; i++) { | |
298 | if (tagset->tags && tagset->tags[i]) | |
299 | blk_mq_all_tag_busy_iter(tagset->tags[i], fn, priv); | |
300 | } | |
301 | } | |
302 | EXPORT_SYMBOL(blk_mq_tagset_busy_iter); | |
303 | ||
486cf989 SG |
304 | int blk_mq_reinit_tagset(struct blk_mq_tag_set *set) |
305 | { | |
306 | int i, j, ret = 0; | |
307 | ||
308 | if (!set->ops->reinit_request) | |
309 | goto out; | |
310 | ||
311 | for (i = 0; i < set->nr_hw_queues; i++) { | |
312 | struct blk_mq_tags *tags = set->tags[i]; | |
313 | ||
d28046fb SG |
314 | if (!tags) |
315 | continue; | |
316 | ||
486cf989 SG |
317 | for (j = 0; j < tags->nr_tags; j++) { |
318 | if (!tags->rqs[j]) | |
319 | continue; | |
320 | ||
321 | ret = set->ops->reinit_request(set->driver_data, | |
322 | tags->rqs[j]); | |
323 | if (ret) | |
324 | goto out; | |
325 | } | |
326 | } | |
327 | ||
328 | out: | |
329 | return ret; | |
330 | } | |
331 | EXPORT_SYMBOL_GPL(blk_mq_reinit_tagset); | |
332 | ||
0bf6cd5b | 333 | void blk_mq_queue_tag_busy_iter(struct request_queue *q, busy_iter_fn *fn, |
81481eb4 | 334 | void *priv) |
320ae51f | 335 | { |
0bf6cd5b CH |
336 | struct blk_mq_hw_ctx *hctx; |
337 | int i; | |
338 | ||
f530afb9 JW |
339 | /* |
340 | * __blk_mq_update_nr_hw_queues will update the nr_hw_queues and | |
d4d74449 KB |
341 | * queue_hw_ctx after freeze the queue, so we use q_usage_counter |
342 | * to avoid race with it. | |
f530afb9 | 343 | */ |
d4d74449 | 344 | if (!percpu_ref_tryget(&q->q_usage_counter)) |
f530afb9 | 345 | return; |
0bf6cd5b CH |
346 | |
347 | queue_for_each_hw_ctx(q, hctx, i) { | |
348 | struct blk_mq_tags *tags = hctx->tags; | |
349 | ||
350 | /* | |
351 | * If not software queues are currently mapped to this | |
352 | * hardware queue, there's nothing to check | |
353 | */ | |
354 | if (!blk_mq_hw_queue_mapped(hctx)) | |
355 | continue; | |
356 | ||
357 | if (tags->nr_reserved_tags) | |
88459642 OS |
358 | bt_for_each(hctx, &tags->breserved_tags, fn, priv, true); |
359 | bt_for_each(hctx, &tags->bitmap_tags, fn, priv, false); | |
4bb659b1 | 360 | } |
d4d74449 | 361 | blk_queue_exit(q); |
4bb659b1 JA |
362 | } |
363 | ||
88459642 | 364 | static unsigned int bt_unused_tags(const struct sbitmap_queue *bt) |
e3a2b3f9 | 365 | { |
88459642 | 366 | return bt->sb.depth - sbitmap_weight(&bt->sb); |
e3a2b3f9 JA |
367 | } |
368 | ||
f4a644db OS |
369 | static int bt_alloc(struct sbitmap_queue *bt, unsigned int depth, |
370 | bool round_robin, int node) | |
4bb659b1 | 371 | { |
f4a644db OS |
372 | return sbitmap_queue_init_node(bt, depth, -1, round_robin, GFP_KERNEL, |
373 | node); | |
4bb659b1 JA |
374 | } |
375 | ||
376 | static struct blk_mq_tags *blk_mq_init_bitmap_tags(struct blk_mq_tags *tags, | |
24391c0d | 377 | int node, int alloc_policy) |
4bb659b1 JA |
378 | { |
379 | unsigned int depth = tags->nr_tags - tags->nr_reserved_tags; | |
f4a644db | 380 | bool round_robin = alloc_policy == BLK_TAG_ALLOC_RR; |
4bb659b1 | 381 | |
f4a644db | 382 | if (bt_alloc(&tags->bitmap_tags, depth, round_robin, node)) |
88459642 | 383 | goto free_tags; |
f4a644db OS |
384 | if (bt_alloc(&tags->breserved_tags, tags->nr_reserved_tags, round_robin, |
385 | node)) | |
88459642 | 386 | goto free_bitmap_tags; |
4bb659b1 JA |
387 | |
388 | return tags; | |
88459642 OS |
389 | free_bitmap_tags: |
390 | sbitmap_queue_free(&tags->bitmap_tags); | |
391 | free_tags: | |
4bb659b1 JA |
392 | kfree(tags); |
393 | return NULL; | |
394 | } | |
395 | ||
320ae51f | 396 | struct blk_mq_tags *blk_mq_init_tags(unsigned int total_tags, |
24391c0d SL |
397 | unsigned int reserved_tags, |
398 | int node, int alloc_policy) | |
320ae51f | 399 | { |
320ae51f | 400 | struct blk_mq_tags *tags; |
320ae51f JA |
401 | |
402 | if (total_tags > BLK_MQ_TAG_MAX) { | |
403 | pr_err("blk-mq: tag depth too large\n"); | |
404 | return NULL; | |
405 | } | |
406 | ||
407 | tags = kzalloc_node(sizeof(*tags), GFP_KERNEL, node); | |
408 | if (!tags) | |
409 | return NULL; | |
410 | ||
320ae51f JA |
411 | tags->nr_tags = total_tags; |
412 | tags->nr_reserved_tags = reserved_tags; | |
320ae51f | 413 | |
24391c0d | 414 | return blk_mq_init_bitmap_tags(tags, node, alloc_policy); |
320ae51f JA |
415 | } |
416 | ||
417 | void blk_mq_free_tags(struct blk_mq_tags *tags) | |
418 | { | |
88459642 OS |
419 | sbitmap_queue_free(&tags->bitmap_tags); |
420 | sbitmap_queue_free(&tags->breserved_tags); | |
320ae51f JA |
421 | kfree(tags); |
422 | } | |
423 | ||
e3a2b3f9 JA |
424 | int blk_mq_tag_update_depth(struct blk_mq_tags *tags, unsigned int tdepth) |
425 | { | |
426 | tdepth -= tags->nr_reserved_tags; | |
427 | if (tdepth > tags->nr_tags) | |
428 | return -EINVAL; | |
429 | ||
430 | /* | |
431 | * Don't need (or can't) update reserved tags here, they remain | |
432 | * static and should never need resizing. | |
433 | */ | |
88459642 OS |
434 | sbitmap_queue_resize(&tags->bitmap_tags, tdepth); |
435 | ||
aed3ea94 | 436 | blk_mq_tag_wakeup_all(tags, false); |
e3a2b3f9 JA |
437 | return 0; |
438 | } | |
439 | ||
205fb5f5 BVA |
440 | /** |
441 | * blk_mq_unique_tag() - return a tag that is unique queue-wide | |
442 | * @rq: request for which to compute a unique tag | |
443 | * | |
444 | * The tag field in struct request is unique per hardware queue but not over | |
445 | * all hardware queues. Hence this function that returns a tag with the | |
446 | * hardware context index in the upper bits and the per hardware queue tag in | |
447 | * the lower bits. | |
448 | * | |
449 | * Note: When called for a request that is queued on a non-multiqueue request | |
450 | * queue, the hardware context index is set to zero. | |
451 | */ | |
452 | u32 blk_mq_unique_tag(struct request *rq) | |
453 | { | |
454 | struct request_queue *q = rq->q; | |
455 | struct blk_mq_hw_ctx *hctx; | |
456 | int hwq = 0; | |
457 | ||
458 | if (q->mq_ops) { | |
7d7e0f90 | 459 | hctx = blk_mq_map_queue(q, rq->mq_ctx->cpu); |
205fb5f5 BVA |
460 | hwq = hctx->queue_num; |
461 | } | |
462 | ||
463 | return (hwq << BLK_MQ_UNIQUE_TAG_BITS) | | |
464 | (rq->tag & BLK_MQ_UNIQUE_TAG_MASK); | |
465 | } | |
466 | EXPORT_SYMBOL(blk_mq_unique_tag); | |
467 | ||
320ae51f JA |
468 | ssize_t blk_mq_tag_sysfs_show(struct blk_mq_tags *tags, char *page) |
469 | { | |
470 | char *orig_page = page; | |
4bb659b1 | 471 | unsigned int free, res; |
320ae51f JA |
472 | |
473 | if (!tags) | |
474 | return 0; | |
475 | ||
59d13bf5 JA |
476 | page += sprintf(page, "nr_tags=%u, reserved_tags=%u, " |
477 | "bits_per_word=%u\n", | |
478 | tags->nr_tags, tags->nr_reserved_tags, | |
88459642 | 479 | 1U << tags->bitmap_tags.sb.shift); |
320ae51f | 480 | |
4bb659b1 JA |
481 | free = bt_unused_tags(&tags->bitmap_tags); |
482 | res = bt_unused_tags(&tags->breserved_tags); | |
320ae51f | 483 | |
4bb659b1 | 484 | page += sprintf(page, "nr_free=%u, nr_reserved=%u\n", free, res); |
0d2602ca | 485 | page += sprintf(page, "active_queues=%u\n", atomic_read(&tags->active_queues)); |
320ae51f JA |
486 | |
487 | return page - orig_page; | |
488 | } |