Commit | Line | Data |
---|---|---|
75bb4625 JA |
1 | /* |
2 | * Fast and scalable bitmap tagging variant. Uses sparser bitmaps spread | |
3 | * over multiple cachelines to avoid ping-pong between multiple submitters | |
4 | * or submitter and completer. Uses rolling wakeups to avoid falling of | |
5 | * the scaling cliff when we run out of tags and have to start putting | |
6 | * submitters to sleep. | |
7 | * | |
8 | * Uses active queue tracking to support fairer distribution of tags | |
9 | * between multiple submitters when a shared tag map is used. | |
10 | * | |
11 | * Copyright (C) 2013-2014 Jens Axboe | |
12 | */ | |
320ae51f JA |
13 | #include <linux/kernel.h> |
14 | #include <linux/module.h> | |
4bb659b1 | 15 | #include <linux/random.h> |
320ae51f JA |
16 | |
17 | #include <linux/blk-mq.h> | |
18 | #include "blk.h" | |
19 | #include "blk-mq.h" | |
20 | #include "blk-mq-tag.h" | |
21 | ||
4bb659b1 JA |
22 | static bool bt_has_free_tags(struct blk_mq_bitmap_tags *bt) |
23 | { | |
24 | int i; | |
25 | ||
26 | for (i = 0; i < bt->map_nr; i++) { | |
e93ecf60 | 27 | struct blk_align_bitmap *bm = &bt->map[i]; |
4bb659b1 JA |
28 | int ret; |
29 | ||
30 | ret = find_first_zero_bit(&bm->word, bm->depth); | |
31 | if (ret < bm->depth) | |
32 | return true; | |
33 | } | |
34 | ||
35 | return false; | |
320ae51f JA |
36 | } |
37 | ||
38 | bool blk_mq_has_free_tags(struct blk_mq_tags *tags) | |
39 | { | |
4bb659b1 JA |
40 | if (!tags) |
41 | return true; | |
42 | ||
43 | return bt_has_free_tags(&tags->bitmap_tags); | |
44 | } | |
45 | ||
8537b120 | 46 | static inline int bt_index_inc(int index) |
0d2602ca | 47 | { |
8537b120 AG |
48 | return (index + 1) & (BT_WAIT_QUEUES - 1); |
49 | } | |
50 | ||
51 | static inline void bt_index_atomic_inc(atomic_t *index) | |
52 | { | |
53 | int old = atomic_read(index); | |
54 | int new = bt_index_inc(old); | |
55 | atomic_cmpxchg(index, old, new); | |
0d2602ca JA |
56 | } |
57 | ||
58 | /* | |
59 | * If a previously inactive queue goes active, bump the active user count. | |
60 | */ | |
61 | bool __blk_mq_tag_busy(struct blk_mq_hw_ctx *hctx) | |
62 | { | |
63 | if (!test_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state) && | |
64 | !test_and_set_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state)) | |
65 | atomic_inc(&hctx->tags->active_queues); | |
66 | ||
67 | return true; | |
68 | } | |
69 | ||
70 | /* | |
e3a2b3f9 | 71 | * Wakeup all potentially sleeping on normal (non-reserved) tags |
0d2602ca | 72 | */ |
e3a2b3f9 | 73 | static void blk_mq_tag_wakeup_all(struct blk_mq_tags *tags) |
0d2602ca | 74 | { |
0d2602ca JA |
75 | struct blk_mq_bitmap_tags *bt; |
76 | int i, wake_index; | |
77 | ||
0d2602ca | 78 | bt = &tags->bitmap_tags; |
8537b120 | 79 | wake_index = atomic_read(&bt->wake_index); |
0d2602ca JA |
80 | for (i = 0; i < BT_WAIT_QUEUES; i++) { |
81 | struct bt_wait_state *bs = &bt->bs[wake_index]; | |
82 | ||
83 | if (waitqueue_active(&bs->wait)) | |
84 | wake_up(&bs->wait); | |
85 | ||
8537b120 | 86 | wake_index = bt_index_inc(wake_index); |
0d2602ca JA |
87 | } |
88 | } | |
89 | ||
e3a2b3f9 JA |
90 | /* |
91 | * If a previously busy queue goes inactive, potential waiters could now | |
92 | * be allowed to queue. Wake them up and check. | |
93 | */ | |
94 | void __blk_mq_tag_idle(struct blk_mq_hw_ctx *hctx) | |
95 | { | |
96 | struct blk_mq_tags *tags = hctx->tags; | |
97 | ||
98 | if (!test_and_clear_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state)) | |
99 | return; | |
100 | ||
101 | atomic_dec(&tags->active_queues); | |
102 | ||
103 | blk_mq_tag_wakeup_all(tags); | |
104 | } | |
105 | ||
0d2602ca JA |
106 | /* |
107 | * For shared tag users, we track the number of currently active users | |
108 | * and attempt to provide a fair share of the tag depth for each of them. | |
109 | */ | |
110 | static inline bool hctx_may_queue(struct blk_mq_hw_ctx *hctx, | |
111 | struct blk_mq_bitmap_tags *bt) | |
112 | { | |
113 | unsigned int depth, users; | |
114 | ||
115 | if (!hctx || !(hctx->flags & BLK_MQ_F_TAG_SHARED)) | |
116 | return true; | |
117 | if (!test_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state)) | |
118 | return true; | |
119 | ||
120 | /* | |
121 | * Don't try dividing an ant | |
122 | */ | |
123 | if (bt->depth == 1) | |
124 | return true; | |
125 | ||
126 | users = atomic_read(&hctx->tags->active_queues); | |
127 | if (!users) | |
128 | return true; | |
129 | ||
130 | /* | |
131 | * Allow at least some tags | |
132 | */ | |
133 | depth = max((bt->depth + users - 1) / users, 4U); | |
134 | return atomic_read(&hctx->nr_active) < depth; | |
135 | } | |
136 | ||
e93ecf60 | 137 | static int __bt_get_word(struct blk_align_bitmap *bm, unsigned int last_tag) |
4bb659b1 JA |
138 | { |
139 | int tag, org_last_tag, end; | |
140 | ||
59d13bf5 | 141 | org_last_tag = last_tag; |
4bb659b1 JA |
142 | end = bm->depth; |
143 | do { | |
144 | restart: | |
145 | tag = find_next_zero_bit(&bm->word, end, last_tag); | |
146 | if (unlikely(tag >= end)) { | |
147 | /* | |
148 | * We started with an offset, start from 0 to | |
149 | * exhaust the map. | |
150 | */ | |
151 | if (org_last_tag && last_tag) { | |
152 | end = last_tag; | |
153 | last_tag = 0; | |
154 | goto restart; | |
155 | } | |
156 | return -1; | |
157 | } | |
158 | last_tag = tag + 1; | |
159 | } while (test_and_set_bit_lock(tag, &bm->word)); | |
160 | ||
161 | return tag; | |
162 | } | |
163 | ||
164 | /* | |
165 | * Straight forward bitmap tag implementation, where each bit is a tag | |
166 | * (cleared == free, and set == busy). The small twist is using per-cpu | |
167 | * last_tag caches, which blk-mq stores in the blk_mq_ctx software queue | |
168 | * contexts. This enables us to drastically limit the space searched, | |
169 | * without dirtying an extra shared cacheline like we would if we stored | |
170 | * the cache value inside the shared blk_mq_bitmap_tags structure. On top | |
171 | * of that, each word of tags is in a separate cacheline. This means that | |
172 | * multiple users will tend to stick to different cachelines, at least | |
173 | * until the map is exhausted. | |
174 | */ | |
0d2602ca JA |
175 | static int __bt_get(struct blk_mq_hw_ctx *hctx, struct blk_mq_bitmap_tags *bt, |
176 | unsigned int *tag_cache) | |
4bb659b1 JA |
177 | { |
178 | unsigned int last_tag, org_last_tag; | |
179 | int index, i, tag; | |
180 | ||
0d2602ca JA |
181 | if (!hctx_may_queue(hctx, bt)) |
182 | return -1; | |
183 | ||
4bb659b1 | 184 | last_tag = org_last_tag = *tag_cache; |
59d13bf5 | 185 | index = TAG_TO_INDEX(bt, last_tag); |
4bb659b1 JA |
186 | |
187 | for (i = 0; i < bt->map_nr; i++) { | |
59d13bf5 | 188 | tag = __bt_get_word(&bt->map[index], TAG_TO_BIT(bt, last_tag)); |
4bb659b1 | 189 | if (tag != -1) { |
59d13bf5 | 190 | tag += (index << bt->bits_per_word); |
4bb659b1 JA |
191 | goto done; |
192 | } | |
193 | ||
194 | last_tag = 0; | |
195 | if (++index >= bt->map_nr) | |
196 | index = 0; | |
197 | } | |
198 | ||
199 | *tag_cache = 0; | |
200 | return -1; | |
201 | ||
202 | /* | |
203 | * Only update the cache from the allocation path, if we ended | |
204 | * up using the specific cached tag. | |
205 | */ | |
206 | done: | |
207 | if (tag == org_last_tag) { | |
208 | last_tag = tag + 1; | |
209 | if (last_tag >= bt->depth - 1) | |
210 | last_tag = 0; | |
211 | ||
212 | *tag_cache = last_tag; | |
213 | } | |
214 | ||
215 | return tag; | |
216 | } | |
217 | ||
4bb659b1 JA |
218 | static struct bt_wait_state *bt_wait_ptr(struct blk_mq_bitmap_tags *bt, |
219 | struct blk_mq_hw_ctx *hctx) | |
220 | { | |
221 | struct bt_wait_state *bs; | |
8537b120 | 222 | int wait_index; |
4bb659b1 JA |
223 | |
224 | if (!hctx) | |
225 | return &bt->bs[0]; | |
226 | ||
8537b120 AG |
227 | wait_index = atomic_read(&hctx->wait_index); |
228 | bs = &bt->bs[wait_index]; | |
229 | bt_index_atomic_inc(&hctx->wait_index); | |
4bb659b1 | 230 | return bs; |
320ae51f JA |
231 | } |
232 | ||
cb96a42c ML |
233 | static int bt_get(struct blk_mq_alloc_data *data, |
234 | struct blk_mq_bitmap_tags *bt, | |
235 | struct blk_mq_hw_ctx *hctx, | |
236 | unsigned int *last_tag) | |
320ae51f | 237 | { |
4bb659b1 JA |
238 | struct bt_wait_state *bs; |
239 | DEFINE_WAIT(wait); | |
320ae51f JA |
240 | int tag; |
241 | ||
0d2602ca | 242 | tag = __bt_get(hctx, bt, last_tag); |
4bb659b1 JA |
243 | if (tag != -1) |
244 | return tag; | |
245 | ||
cb96a42c | 246 | if (!(data->gfp & __GFP_WAIT)) |
4bb659b1 JA |
247 | return -1; |
248 | ||
249 | bs = bt_wait_ptr(bt, hctx); | |
250 | do { | |
4bb659b1 JA |
251 | prepare_to_wait(&bs->wait, &wait, TASK_UNINTERRUPTIBLE); |
252 | ||
0d2602ca | 253 | tag = __bt_get(hctx, bt, last_tag); |
4bb659b1 JA |
254 | if (tag != -1) |
255 | break; | |
256 | ||
b3223207 BVA |
257 | /* |
258 | * We're out of tags on this hardware queue, kick any | |
259 | * pending IO submits before going to sleep waiting for | |
260 | * some to complete. | |
261 | */ | |
262 | blk_mq_run_hw_queue(hctx, false); | |
263 | ||
cb96a42c ML |
264 | blk_mq_put_ctx(data->ctx); |
265 | ||
4bb659b1 | 266 | io_schedule(); |
cb96a42c ML |
267 | |
268 | data->ctx = blk_mq_get_ctx(data->q); | |
269 | data->hctx = data->q->mq_ops->map_queue(data->q, | |
270 | data->ctx->cpu); | |
271 | if (data->reserved) { | |
272 | bt = &data->hctx->tags->breserved_tags; | |
273 | } else { | |
274 | last_tag = &data->ctx->last_tag; | |
275 | hctx = data->hctx; | |
276 | bt = &hctx->tags->bitmap_tags; | |
277 | } | |
278 | finish_wait(&bs->wait, &wait); | |
279 | bs = bt_wait_ptr(bt, hctx); | |
4bb659b1 JA |
280 | } while (1); |
281 | ||
282 | finish_wait(&bs->wait, &wait); | |
283 | return tag; | |
284 | } | |
285 | ||
cb96a42c | 286 | static unsigned int __blk_mq_get_tag(struct blk_mq_alloc_data *data) |
4bb659b1 JA |
287 | { |
288 | int tag; | |
289 | ||
cb96a42c ML |
290 | tag = bt_get(data, &data->hctx->tags->bitmap_tags, data->hctx, |
291 | &data->ctx->last_tag); | |
4bb659b1 | 292 | if (tag >= 0) |
cb96a42c | 293 | return tag + data->hctx->tags->nr_reserved_tags; |
4bb659b1 JA |
294 | |
295 | return BLK_MQ_TAG_FAIL; | |
320ae51f JA |
296 | } |
297 | ||
cb96a42c | 298 | static unsigned int __blk_mq_get_reserved_tag(struct blk_mq_alloc_data *data) |
320ae51f | 299 | { |
4bb659b1 | 300 | int tag, zero = 0; |
320ae51f | 301 | |
cb96a42c | 302 | if (unlikely(!data->hctx->tags->nr_reserved_tags)) { |
320ae51f JA |
303 | WARN_ON_ONCE(1); |
304 | return BLK_MQ_TAG_FAIL; | |
305 | } | |
306 | ||
cb96a42c | 307 | tag = bt_get(data, &data->hctx->tags->breserved_tags, NULL, &zero); |
320ae51f JA |
308 | if (tag < 0) |
309 | return BLK_MQ_TAG_FAIL; | |
4bb659b1 | 310 | |
320ae51f JA |
311 | return tag; |
312 | } | |
313 | ||
cb96a42c | 314 | unsigned int blk_mq_get_tag(struct blk_mq_alloc_data *data) |
320ae51f | 315 | { |
cb96a42c ML |
316 | if (!data->reserved) |
317 | return __blk_mq_get_tag(data); | |
320ae51f | 318 | |
cb96a42c | 319 | return __blk_mq_get_reserved_tag(data); |
320ae51f JA |
320 | } |
321 | ||
4bb659b1 JA |
322 | static struct bt_wait_state *bt_wake_ptr(struct blk_mq_bitmap_tags *bt) |
323 | { | |
324 | int i, wake_index; | |
325 | ||
8537b120 | 326 | wake_index = atomic_read(&bt->wake_index); |
4bb659b1 JA |
327 | for (i = 0; i < BT_WAIT_QUEUES; i++) { |
328 | struct bt_wait_state *bs = &bt->bs[wake_index]; | |
329 | ||
330 | if (waitqueue_active(&bs->wait)) { | |
8537b120 AG |
331 | int o = atomic_read(&bt->wake_index); |
332 | if (wake_index != o) | |
333 | atomic_cmpxchg(&bt->wake_index, o, wake_index); | |
4bb659b1 JA |
334 | |
335 | return bs; | |
336 | } | |
337 | ||
8537b120 | 338 | wake_index = bt_index_inc(wake_index); |
4bb659b1 JA |
339 | } |
340 | ||
341 | return NULL; | |
342 | } | |
343 | ||
344 | static void bt_clear_tag(struct blk_mq_bitmap_tags *bt, unsigned int tag) | |
345 | { | |
59d13bf5 | 346 | const int index = TAG_TO_INDEX(bt, tag); |
4bb659b1 | 347 | struct bt_wait_state *bs; |
2971c35f | 348 | int wait_cnt; |
4bb659b1 | 349 | |
0289b2e1 ML |
350 | /* |
351 | * The unlock memory barrier need to order access to req in free | |
352 | * path and clearing tag bit | |
353 | */ | |
354 | clear_bit_unlock(TAG_TO_BIT(bt, tag), &bt->map[index].word); | |
4bb659b1 JA |
355 | |
356 | bs = bt_wake_ptr(bt); | |
2971c35f AG |
357 | if (!bs) |
358 | return; | |
359 | ||
360 | wait_cnt = atomic_dec_return(&bs->wait_cnt); | |
9d8f0bcc BVA |
361 | if (unlikely(wait_cnt < 0)) |
362 | wait_cnt = atomic_inc_return(&bs->wait_cnt); | |
2971c35f | 363 | if (wait_cnt == 0) { |
2971c35f | 364 | atomic_add(bt->wake_cnt, &bs->wait_cnt); |
8537b120 | 365 | bt_index_atomic_inc(&bt->wake_index); |
4bb659b1 JA |
366 | wake_up(&bs->wait); |
367 | } | |
368 | } | |
369 | ||
0d2602ca | 370 | void blk_mq_put_tag(struct blk_mq_hw_ctx *hctx, unsigned int tag, |
4bb659b1 | 371 | unsigned int *last_tag) |
320ae51f | 372 | { |
0d2602ca JA |
373 | struct blk_mq_tags *tags = hctx->tags; |
374 | ||
4bb659b1 JA |
375 | if (tag >= tags->nr_reserved_tags) { |
376 | const int real_tag = tag - tags->nr_reserved_tags; | |
377 | ||
70114c39 JA |
378 | BUG_ON(real_tag >= tags->nr_tags); |
379 | bt_clear_tag(&tags->bitmap_tags, real_tag); | |
4bb659b1 | 380 | *last_tag = real_tag; |
70114c39 JA |
381 | } else { |
382 | BUG_ON(tag >= tags->nr_reserved_tags); | |
383 | bt_clear_tag(&tags->breserved_tags, tag); | |
384 | } | |
320ae51f JA |
385 | } |
386 | ||
81481eb4 CH |
387 | static void bt_for_each(struct blk_mq_hw_ctx *hctx, |
388 | struct blk_mq_bitmap_tags *bt, unsigned int off, | |
389 | busy_iter_fn *fn, void *data, bool reserved) | |
320ae51f | 390 | { |
81481eb4 CH |
391 | struct request *rq; |
392 | int bit, i; | |
4bb659b1 JA |
393 | |
394 | for (i = 0; i < bt->map_nr; i++) { | |
e93ecf60 | 395 | struct blk_align_bitmap *bm = &bt->map[i]; |
4bb659b1 | 396 | |
81481eb4 CH |
397 | for (bit = find_first_bit(&bm->word, bm->depth); |
398 | bit < bm->depth; | |
399 | bit = find_next_bit(&bm->word, bm->depth, bit + 1)) { | |
400 | rq = blk_mq_tag_to_rq(hctx->tags, off + bit); | |
401 | if (rq->q == hctx->queue) | |
402 | fn(hctx, rq, data, reserved); | |
403 | } | |
4bb659b1 | 404 | |
59d13bf5 | 405 | off += (1 << bt->bits_per_word); |
4bb659b1 | 406 | } |
320ae51f JA |
407 | } |
408 | ||
81481eb4 CH |
409 | void blk_mq_tag_busy_iter(struct blk_mq_hw_ctx *hctx, busy_iter_fn *fn, |
410 | void *priv) | |
320ae51f | 411 | { |
81481eb4 | 412 | struct blk_mq_tags *tags = hctx->tags; |
320ae51f | 413 | |
320ae51f | 414 | if (tags->nr_reserved_tags) |
81481eb4 CH |
415 | bt_for_each(hctx, &tags->breserved_tags, 0, fn, priv, true); |
416 | bt_for_each(hctx, &tags->bitmap_tags, tags->nr_reserved_tags, fn, priv, | |
417 | false); | |
320ae51f | 418 | } |
edf866b3 | 419 | EXPORT_SYMBOL(blk_mq_tag_busy_iter); |
320ae51f | 420 | |
4bb659b1 JA |
421 | static unsigned int bt_unused_tags(struct blk_mq_bitmap_tags *bt) |
422 | { | |
423 | unsigned int i, used; | |
424 | ||
425 | for (i = 0, used = 0; i < bt->map_nr; i++) { | |
e93ecf60 | 426 | struct blk_align_bitmap *bm = &bt->map[i]; |
4bb659b1 JA |
427 | |
428 | used += bitmap_weight(&bm->word, bm->depth); | |
429 | } | |
430 | ||
431 | return bt->depth - used; | |
432 | } | |
433 | ||
e3a2b3f9 JA |
434 | static void bt_update_count(struct blk_mq_bitmap_tags *bt, |
435 | unsigned int depth) | |
436 | { | |
437 | unsigned int tags_per_word = 1U << bt->bits_per_word; | |
438 | unsigned int map_depth = depth; | |
439 | ||
440 | if (depth) { | |
441 | int i; | |
442 | ||
443 | for (i = 0; i < bt->map_nr; i++) { | |
444 | bt->map[i].depth = min(map_depth, tags_per_word); | |
445 | map_depth -= bt->map[i].depth; | |
446 | } | |
447 | } | |
448 | ||
449 | bt->wake_cnt = BT_WAIT_BATCH; | |
abab13b5 JA |
450 | if (bt->wake_cnt > depth / BT_WAIT_QUEUES) |
451 | bt->wake_cnt = max(1U, depth / BT_WAIT_QUEUES); | |
e3a2b3f9 JA |
452 | |
453 | bt->depth = depth; | |
454 | } | |
455 | ||
4bb659b1 JA |
456 | static int bt_alloc(struct blk_mq_bitmap_tags *bt, unsigned int depth, |
457 | int node, bool reserved) | |
458 | { | |
459 | int i; | |
460 | ||
59d13bf5 JA |
461 | bt->bits_per_word = ilog2(BITS_PER_LONG); |
462 | ||
4bb659b1 JA |
463 | /* |
464 | * Depth can be zero for reserved tags, that's not a failure | |
465 | * condition. | |
466 | */ | |
467 | if (depth) { | |
e3a2b3f9 | 468 | unsigned int nr, tags_per_word; |
59d13bf5 JA |
469 | |
470 | tags_per_word = (1 << bt->bits_per_word); | |
471 | ||
472 | /* | |
473 | * If the tag space is small, shrink the number of tags | |
474 | * per word so we spread over a few cachelines, at least. | |
475 | * If less than 4 tags, just forget about it, it's not | |
476 | * going to work optimally anyway. | |
477 | */ | |
478 | if (depth >= 4) { | |
479 | while (tags_per_word * 4 > depth) { | |
480 | bt->bits_per_word--; | |
481 | tags_per_word = (1 << bt->bits_per_word); | |
482 | } | |
483 | } | |
4bb659b1 | 484 | |
59d13bf5 | 485 | nr = ALIGN(depth, tags_per_word) / tags_per_word; |
e93ecf60 | 486 | bt->map = kzalloc_node(nr * sizeof(struct blk_align_bitmap), |
4bb659b1 JA |
487 | GFP_KERNEL, node); |
488 | if (!bt->map) | |
489 | return -ENOMEM; | |
490 | ||
491 | bt->map_nr = nr; | |
4bb659b1 JA |
492 | } |
493 | ||
494 | bt->bs = kzalloc(BT_WAIT_QUEUES * sizeof(*bt->bs), GFP_KERNEL); | |
495 | if (!bt->bs) { | |
496 | kfree(bt->map); | |
497 | return -ENOMEM; | |
498 | } | |
499 | ||
86fb5c56 AG |
500 | bt_update_count(bt, depth); |
501 | ||
502 | for (i = 0; i < BT_WAIT_QUEUES; i++) { | |
4bb659b1 | 503 | init_waitqueue_head(&bt->bs[i].wait); |
86fb5c56 AG |
504 | atomic_set(&bt->bs[i].wait_cnt, bt->wake_cnt); |
505 | } | |
4bb659b1 | 506 | |
4bb659b1 JA |
507 | return 0; |
508 | } | |
509 | ||
510 | static void bt_free(struct blk_mq_bitmap_tags *bt) | |
511 | { | |
512 | kfree(bt->map); | |
513 | kfree(bt->bs); | |
514 | } | |
515 | ||
516 | static struct blk_mq_tags *blk_mq_init_bitmap_tags(struct blk_mq_tags *tags, | |
517 | int node) | |
518 | { | |
519 | unsigned int depth = tags->nr_tags - tags->nr_reserved_tags; | |
520 | ||
521 | if (bt_alloc(&tags->bitmap_tags, depth, node, false)) | |
522 | goto enomem; | |
523 | if (bt_alloc(&tags->breserved_tags, tags->nr_reserved_tags, node, true)) | |
524 | goto enomem; | |
525 | ||
526 | return tags; | |
527 | enomem: | |
528 | bt_free(&tags->bitmap_tags); | |
529 | kfree(tags); | |
530 | return NULL; | |
531 | } | |
532 | ||
320ae51f JA |
533 | struct blk_mq_tags *blk_mq_init_tags(unsigned int total_tags, |
534 | unsigned int reserved_tags, int node) | |
535 | { | |
320ae51f | 536 | struct blk_mq_tags *tags; |
320ae51f JA |
537 | |
538 | if (total_tags > BLK_MQ_TAG_MAX) { | |
539 | pr_err("blk-mq: tag depth too large\n"); | |
540 | return NULL; | |
541 | } | |
542 | ||
543 | tags = kzalloc_node(sizeof(*tags), GFP_KERNEL, node); | |
544 | if (!tags) | |
545 | return NULL; | |
546 | ||
320ae51f JA |
547 | tags->nr_tags = total_tags; |
548 | tags->nr_reserved_tags = reserved_tags; | |
320ae51f | 549 | |
4bb659b1 | 550 | return blk_mq_init_bitmap_tags(tags, node); |
320ae51f JA |
551 | } |
552 | ||
553 | void blk_mq_free_tags(struct blk_mq_tags *tags) | |
554 | { | |
4bb659b1 JA |
555 | bt_free(&tags->bitmap_tags); |
556 | bt_free(&tags->breserved_tags); | |
320ae51f JA |
557 | kfree(tags); |
558 | } | |
559 | ||
4bb659b1 JA |
560 | void blk_mq_tag_init_last_tag(struct blk_mq_tags *tags, unsigned int *tag) |
561 | { | |
562 | unsigned int depth = tags->nr_tags - tags->nr_reserved_tags; | |
563 | ||
9d3d21ae | 564 | *tag = prandom_u32() % depth; |
4bb659b1 JA |
565 | } |
566 | ||
e3a2b3f9 JA |
567 | int blk_mq_tag_update_depth(struct blk_mq_tags *tags, unsigned int tdepth) |
568 | { | |
569 | tdepth -= tags->nr_reserved_tags; | |
570 | if (tdepth > tags->nr_tags) | |
571 | return -EINVAL; | |
572 | ||
573 | /* | |
574 | * Don't need (or can't) update reserved tags here, they remain | |
575 | * static and should never need resizing. | |
576 | */ | |
577 | bt_update_count(&tags->bitmap_tags, tdepth); | |
578 | blk_mq_tag_wakeup_all(tags); | |
579 | return 0; | |
580 | } | |
581 | ||
320ae51f JA |
582 | ssize_t blk_mq_tag_sysfs_show(struct blk_mq_tags *tags, char *page) |
583 | { | |
584 | char *orig_page = page; | |
4bb659b1 | 585 | unsigned int free, res; |
320ae51f JA |
586 | |
587 | if (!tags) | |
588 | return 0; | |
589 | ||
59d13bf5 JA |
590 | page += sprintf(page, "nr_tags=%u, reserved_tags=%u, " |
591 | "bits_per_word=%u\n", | |
592 | tags->nr_tags, tags->nr_reserved_tags, | |
593 | tags->bitmap_tags.bits_per_word); | |
320ae51f | 594 | |
4bb659b1 JA |
595 | free = bt_unused_tags(&tags->bitmap_tags); |
596 | res = bt_unused_tags(&tags->breserved_tags); | |
320ae51f | 597 | |
4bb659b1 | 598 | page += sprintf(page, "nr_free=%u, nr_reserved=%u\n", free, res); |
0d2602ca | 599 | page += sprintf(page, "active_queues=%u\n", atomic_read(&tags->active_queues)); |
320ae51f JA |
600 | |
601 | return page - orig_page; | |
602 | } |