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 | /* | |
aed3ea94 | 71 | * Wakeup all potentially sleeping on tags |
0d2602ca | 72 | */ |
aed3ea94 | 73 | void blk_mq_tag_wakeup_all(struct blk_mq_tags *tags, bool include_reserve) |
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 | 87 | } |
aed3ea94 JA |
88 | |
89 | if (include_reserve) { | |
90 | bt = &tags->breserved_tags; | |
91 | if (waitqueue_active(&bt->bs[0].wait)) | |
92 | wake_up(&bt->bs[0].wait); | |
93 | } | |
0d2602ca JA |
94 | } |
95 | ||
e3a2b3f9 JA |
96 | /* |
97 | * If a previously busy queue goes inactive, potential waiters could now | |
98 | * be allowed to queue. Wake them up and check. | |
99 | */ | |
100 | void __blk_mq_tag_idle(struct blk_mq_hw_ctx *hctx) | |
101 | { | |
102 | struct blk_mq_tags *tags = hctx->tags; | |
103 | ||
104 | if (!test_and_clear_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state)) | |
105 | return; | |
106 | ||
107 | atomic_dec(&tags->active_queues); | |
108 | ||
aed3ea94 | 109 | blk_mq_tag_wakeup_all(tags, false); |
e3a2b3f9 JA |
110 | } |
111 | ||
0d2602ca JA |
112 | /* |
113 | * For shared tag users, we track the number of currently active users | |
114 | * and attempt to provide a fair share of the tag depth for each of them. | |
115 | */ | |
116 | static inline bool hctx_may_queue(struct blk_mq_hw_ctx *hctx, | |
117 | struct blk_mq_bitmap_tags *bt) | |
118 | { | |
119 | unsigned int depth, users; | |
120 | ||
121 | if (!hctx || !(hctx->flags & BLK_MQ_F_TAG_SHARED)) | |
122 | return true; | |
123 | if (!test_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state)) | |
124 | return true; | |
125 | ||
126 | /* | |
127 | * Don't try dividing an ant | |
128 | */ | |
129 | if (bt->depth == 1) | |
130 | return true; | |
131 | ||
132 | users = atomic_read(&hctx->tags->active_queues); | |
133 | if (!users) | |
134 | return true; | |
135 | ||
136 | /* | |
137 | * Allow at least some tags | |
138 | */ | |
139 | depth = max((bt->depth + users - 1) / users, 4U); | |
140 | return atomic_read(&hctx->nr_active) < depth; | |
141 | } | |
142 | ||
e93ecf60 | 143 | static int __bt_get_word(struct blk_align_bitmap *bm, unsigned int last_tag) |
4bb659b1 | 144 | { |
0bf36498 JA |
145 | int tag, org_last_tag = last_tag; |
146 | ||
147 | while (1) { | |
148 | tag = find_next_zero_bit(&bm->word, bm->depth, last_tag); | |
149 | if (unlikely(tag >= bm->depth)) { | |
4bb659b1 | 150 | /* |
0bf36498 JA |
151 | * We started with an offset, and we didn't reset the |
152 | * offset to 0 in a failure case, so start from 0 to | |
4bb659b1 JA |
153 | * exhaust the map. |
154 | */ | |
0bf36498 JA |
155 | if (org_last_tag && last_tag) { |
156 | last_tag = org_last_tag = 0; | |
157 | continue; | |
4bb659b1 JA |
158 | } |
159 | return -1; | |
160 | } | |
0bf36498 JA |
161 | |
162 | if (!test_and_set_bit(tag, &bm->word)) | |
163 | break; | |
164 | ||
4bb659b1 | 165 | last_tag = tag + 1; |
0bf36498 JA |
166 | if (last_tag >= bm->depth - 1) |
167 | last_tag = 0; | |
168 | } | |
4bb659b1 JA |
169 | |
170 | return tag; | |
171 | } | |
172 | ||
173 | /* | |
174 | * Straight forward bitmap tag implementation, where each bit is a tag | |
175 | * (cleared == free, and set == busy). The small twist is using per-cpu | |
176 | * last_tag caches, which blk-mq stores in the blk_mq_ctx software queue | |
177 | * contexts. This enables us to drastically limit the space searched, | |
178 | * without dirtying an extra shared cacheline like we would if we stored | |
179 | * the cache value inside the shared blk_mq_bitmap_tags structure. On top | |
180 | * of that, each word of tags is in a separate cacheline. This means that | |
181 | * multiple users will tend to stick to different cachelines, at least | |
182 | * until the map is exhausted. | |
183 | */ | |
0d2602ca JA |
184 | static int __bt_get(struct blk_mq_hw_ctx *hctx, struct blk_mq_bitmap_tags *bt, |
185 | unsigned int *tag_cache) | |
4bb659b1 JA |
186 | { |
187 | unsigned int last_tag, org_last_tag; | |
188 | int index, i, tag; | |
189 | ||
0d2602ca JA |
190 | if (!hctx_may_queue(hctx, bt)) |
191 | return -1; | |
192 | ||
4bb659b1 | 193 | last_tag = org_last_tag = *tag_cache; |
59d13bf5 | 194 | index = TAG_TO_INDEX(bt, last_tag); |
4bb659b1 JA |
195 | |
196 | for (i = 0; i < bt->map_nr; i++) { | |
59d13bf5 | 197 | tag = __bt_get_word(&bt->map[index], TAG_TO_BIT(bt, last_tag)); |
4bb659b1 | 198 | if (tag != -1) { |
59d13bf5 | 199 | tag += (index << bt->bits_per_word); |
4bb659b1 JA |
200 | goto done; |
201 | } | |
202 | ||
0bf36498 JA |
203 | /* |
204 | * Jump to next index, and reset the last tag to be the | |
205 | * first tag of that index | |
206 | */ | |
207 | index++; | |
208 | last_tag = (index << bt->bits_per_word); | |
209 | ||
210 | if (index >= bt->map_nr) { | |
4bb659b1 | 211 | index = 0; |
0bf36498 JA |
212 | last_tag = 0; |
213 | } | |
4bb659b1 JA |
214 | } |
215 | ||
216 | *tag_cache = 0; | |
217 | return -1; | |
218 | ||
219 | /* | |
220 | * Only update the cache from the allocation path, if we ended | |
221 | * up using the specific cached tag. | |
222 | */ | |
223 | done: | |
224 | if (tag == org_last_tag) { | |
225 | last_tag = tag + 1; | |
226 | if (last_tag >= bt->depth - 1) | |
227 | last_tag = 0; | |
228 | ||
229 | *tag_cache = last_tag; | |
230 | } | |
231 | ||
232 | return tag; | |
233 | } | |
234 | ||
4bb659b1 JA |
235 | static struct bt_wait_state *bt_wait_ptr(struct blk_mq_bitmap_tags *bt, |
236 | struct blk_mq_hw_ctx *hctx) | |
237 | { | |
238 | struct bt_wait_state *bs; | |
8537b120 | 239 | int wait_index; |
4bb659b1 JA |
240 | |
241 | if (!hctx) | |
242 | return &bt->bs[0]; | |
243 | ||
8537b120 AG |
244 | wait_index = atomic_read(&hctx->wait_index); |
245 | bs = &bt->bs[wait_index]; | |
246 | bt_index_atomic_inc(&hctx->wait_index); | |
4bb659b1 | 247 | return bs; |
320ae51f JA |
248 | } |
249 | ||
cb96a42c ML |
250 | static int bt_get(struct blk_mq_alloc_data *data, |
251 | struct blk_mq_bitmap_tags *bt, | |
252 | struct blk_mq_hw_ctx *hctx, | |
253 | unsigned int *last_tag) | |
320ae51f | 254 | { |
4bb659b1 JA |
255 | struct bt_wait_state *bs; |
256 | DEFINE_WAIT(wait); | |
320ae51f JA |
257 | int tag; |
258 | ||
0d2602ca | 259 | tag = __bt_get(hctx, bt, last_tag); |
4bb659b1 JA |
260 | if (tag != -1) |
261 | return tag; | |
262 | ||
cb96a42c | 263 | if (!(data->gfp & __GFP_WAIT)) |
4bb659b1 JA |
264 | return -1; |
265 | ||
35d37c66 | 266 | bs = bt_wait_ptr(bt, hctx); |
4bb659b1 | 267 | do { |
4bb659b1 JA |
268 | prepare_to_wait(&bs->wait, &wait, TASK_UNINTERRUPTIBLE); |
269 | ||
0d2602ca | 270 | tag = __bt_get(hctx, bt, last_tag); |
4bb659b1 JA |
271 | if (tag != -1) |
272 | break; | |
273 | ||
b3223207 BVA |
274 | /* |
275 | * We're out of tags on this hardware queue, kick any | |
276 | * pending IO submits before going to sleep waiting for | |
277 | * some to complete. | |
278 | */ | |
279 | blk_mq_run_hw_queue(hctx, false); | |
280 | ||
080ff351 JA |
281 | /* |
282 | * Retry tag allocation after running the hardware queue, | |
283 | * as running the queue may also have found completions. | |
284 | */ | |
285 | tag = __bt_get(hctx, bt, last_tag); | |
286 | if (tag != -1) | |
287 | break; | |
288 | ||
cb96a42c ML |
289 | blk_mq_put_ctx(data->ctx); |
290 | ||
4bb659b1 | 291 | io_schedule(); |
cb96a42c ML |
292 | |
293 | data->ctx = blk_mq_get_ctx(data->q); | |
294 | data->hctx = data->q->mq_ops->map_queue(data->q, | |
295 | data->ctx->cpu); | |
296 | if (data->reserved) { | |
297 | bt = &data->hctx->tags->breserved_tags; | |
298 | } else { | |
299 | last_tag = &data->ctx->last_tag; | |
300 | hctx = data->hctx; | |
301 | bt = &hctx->tags->bitmap_tags; | |
302 | } | |
35d37c66 JA |
303 | finish_wait(&bs->wait, &wait); |
304 | bs = bt_wait_ptr(bt, hctx); | |
4bb659b1 JA |
305 | } while (1); |
306 | ||
307 | finish_wait(&bs->wait, &wait); | |
308 | return tag; | |
309 | } | |
310 | ||
cb96a42c | 311 | static unsigned int __blk_mq_get_tag(struct blk_mq_alloc_data *data) |
4bb659b1 JA |
312 | { |
313 | int tag; | |
314 | ||
cb96a42c ML |
315 | tag = bt_get(data, &data->hctx->tags->bitmap_tags, data->hctx, |
316 | &data->ctx->last_tag); | |
4bb659b1 | 317 | if (tag >= 0) |
cb96a42c | 318 | return tag + data->hctx->tags->nr_reserved_tags; |
4bb659b1 JA |
319 | |
320 | return BLK_MQ_TAG_FAIL; | |
320ae51f JA |
321 | } |
322 | ||
cb96a42c | 323 | static unsigned int __blk_mq_get_reserved_tag(struct blk_mq_alloc_data *data) |
320ae51f | 324 | { |
4bb659b1 | 325 | int tag, zero = 0; |
320ae51f | 326 | |
cb96a42c | 327 | if (unlikely(!data->hctx->tags->nr_reserved_tags)) { |
320ae51f JA |
328 | WARN_ON_ONCE(1); |
329 | return BLK_MQ_TAG_FAIL; | |
330 | } | |
331 | ||
cb96a42c | 332 | tag = bt_get(data, &data->hctx->tags->breserved_tags, NULL, &zero); |
320ae51f JA |
333 | if (tag < 0) |
334 | return BLK_MQ_TAG_FAIL; | |
4bb659b1 | 335 | |
320ae51f JA |
336 | return tag; |
337 | } | |
338 | ||
cb96a42c | 339 | unsigned int blk_mq_get_tag(struct blk_mq_alloc_data *data) |
320ae51f | 340 | { |
cb96a42c ML |
341 | if (!data->reserved) |
342 | return __blk_mq_get_tag(data); | |
320ae51f | 343 | |
cb96a42c | 344 | return __blk_mq_get_reserved_tag(data); |
320ae51f JA |
345 | } |
346 | ||
4bb659b1 JA |
347 | static struct bt_wait_state *bt_wake_ptr(struct blk_mq_bitmap_tags *bt) |
348 | { | |
349 | int i, wake_index; | |
350 | ||
8537b120 | 351 | wake_index = atomic_read(&bt->wake_index); |
4bb659b1 JA |
352 | for (i = 0; i < BT_WAIT_QUEUES; i++) { |
353 | struct bt_wait_state *bs = &bt->bs[wake_index]; | |
354 | ||
355 | if (waitqueue_active(&bs->wait)) { | |
8537b120 AG |
356 | int o = atomic_read(&bt->wake_index); |
357 | if (wake_index != o) | |
358 | atomic_cmpxchg(&bt->wake_index, o, wake_index); | |
4bb659b1 JA |
359 | |
360 | return bs; | |
361 | } | |
362 | ||
8537b120 | 363 | wake_index = bt_index_inc(wake_index); |
4bb659b1 JA |
364 | } |
365 | ||
366 | return NULL; | |
367 | } | |
368 | ||
369 | static void bt_clear_tag(struct blk_mq_bitmap_tags *bt, unsigned int tag) | |
370 | { | |
59d13bf5 | 371 | const int index = TAG_TO_INDEX(bt, tag); |
4bb659b1 | 372 | struct bt_wait_state *bs; |
2971c35f | 373 | int wait_cnt; |
4bb659b1 | 374 | |
c38d185d BVA |
375 | clear_bit(TAG_TO_BIT(bt, tag), &bt->map[index].word); |
376 | ||
377 | /* Ensure that the wait list checks occur after clear_bit(). */ | |
378 | smp_mb(); | |
4bb659b1 JA |
379 | |
380 | bs = bt_wake_ptr(bt); | |
2971c35f AG |
381 | if (!bs) |
382 | return; | |
383 | ||
384 | wait_cnt = atomic_dec_return(&bs->wait_cnt); | |
9d8f0bcc BVA |
385 | if (unlikely(wait_cnt < 0)) |
386 | wait_cnt = atomic_inc_return(&bs->wait_cnt); | |
2971c35f | 387 | if (wait_cnt == 0) { |
2971c35f | 388 | atomic_add(bt->wake_cnt, &bs->wait_cnt); |
8537b120 | 389 | bt_index_atomic_inc(&bt->wake_index); |
4bb659b1 JA |
390 | wake_up(&bs->wait); |
391 | } | |
392 | } | |
393 | ||
0d2602ca | 394 | void blk_mq_put_tag(struct blk_mq_hw_ctx *hctx, unsigned int tag, |
4bb659b1 | 395 | unsigned int *last_tag) |
320ae51f | 396 | { |
0d2602ca JA |
397 | struct blk_mq_tags *tags = hctx->tags; |
398 | ||
4bb659b1 JA |
399 | if (tag >= tags->nr_reserved_tags) { |
400 | const int real_tag = tag - tags->nr_reserved_tags; | |
401 | ||
70114c39 JA |
402 | BUG_ON(real_tag >= tags->nr_tags); |
403 | bt_clear_tag(&tags->bitmap_tags, real_tag); | |
4bb659b1 | 404 | *last_tag = real_tag; |
70114c39 JA |
405 | } else { |
406 | BUG_ON(tag >= tags->nr_reserved_tags); | |
407 | bt_clear_tag(&tags->breserved_tags, tag); | |
408 | } | |
320ae51f JA |
409 | } |
410 | ||
81481eb4 CH |
411 | static void bt_for_each(struct blk_mq_hw_ctx *hctx, |
412 | struct blk_mq_bitmap_tags *bt, unsigned int off, | |
413 | busy_iter_fn *fn, void *data, bool reserved) | |
320ae51f | 414 | { |
81481eb4 CH |
415 | struct request *rq; |
416 | int bit, i; | |
4bb659b1 JA |
417 | |
418 | for (i = 0; i < bt->map_nr; i++) { | |
e93ecf60 | 419 | struct blk_align_bitmap *bm = &bt->map[i]; |
4bb659b1 | 420 | |
81481eb4 CH |
421 | for (bit = find_first_bit(&bm->word, bm->depth); |
422 | bit < bm->depth; | |
423 | bit = find_next_bit(&bm->word, bm->depth, bit + 1)) { | |
424 | rq = blk_mq_tag_to_rq(hctx->tags, off + bit); | |
425 | if (rq->q == hctx->queue) | |
426 | fn(hctx, rq, data, reserved); | |
427 | } | |
4bb659b1 | 428 | |
59d13bf5 | 429 | off += (1 << bt->bits_per_word); |
4bb659b1 | 430 | } |
320ae51f JA |
431 | } |
432 | ||
81481eb4 CH |
433 | void blk_mq_tag_busy_iter(struct blk_mq_hw_ctx *hctx, busy_iter_fn *fn, |
434 | void *priv) | |
320ae51f | 435 | { |
81481eb4 | 436 | struct blk_mq_tags *tags = hctx->tags; |
320ae51f | 437 | |
320ae51f | 438 | if (tags->nr_reserved_tags) |
81481eb4 CH |
439 | bt_for_each(hctx, &tags->breserved_tags, 0, fn, priv, true); |
440 | bt_for_each(hctx, &tags->bitmap_tags, tags->nr_reserved_tags, fn, priv, | |
441 | false); | |
320ae51f | 442 | } |
edf866b3 | 443 | EXPORT_SYMBOL(blk_mq_tag_busy_iter); |
320ae51f | 444 | |
4bb659b1 JA |
445 | static unsigned int bt_unused_tags(struct blk_mq_bitmap_tags *bt) |
446 | { | |
447 | unsigned int i, used; | |
448 | ||
449 | for (i = 0, used = 0; i < bt->map_nr; i++) { | |
e93ecf60 | 450 | struct blk_align_bitmap *bm = &bt->map[i]; |
4bb659b1 JA |
451 | |
452 | used += bitmap_weight(&bm->word, bm->depth); | |
453 | } | |
454 | ||
455 | return bt->depth - used; | |
456 | } | |
457 | ||
e3a2b3f9 JA |
458 | static void bt_update_count(struct blk_mq_bitmap_tags *bt, |
459 | unsigned int depth) | |
460 | { | |
461 | unsigned int tags_per_word = 1U << bt->bits_per_word; | |
462 | unsigned int map_depth = depth; | |
463 | ||
464 | if (depth) { | |
465 | int i; | |
466 | ||
467 | for (i = 0; i < bt->map_nr; i++) { | |
468 | bt->map[i].depth = min(map_depth, tags_per_word); | |
469 | map_depth -= bt->map[i].depth; | |
470 | } | |
471 | } | |
472 | ||
473 | bt->wake_cnt = BT_WAIT_BATCH; | |
abab13b5 JA |
474 | if (bt->wake_cnt > depth / BT_WAIT_QUEUES) |
475 | bt->wake_cnt = max(1U, depth / BT_WAIT_QUEUES); | |
e3a2b3f9 JA |
476 | |
477 | bt->depth = depth; | |
478 | } | |
479 | ||
4bb659b1 JA |
480 | static int bt_alloc(struct blk_mq_bitmap_tags *bt, unsigned int depth, |
481 | int node, bool reserved) | |
482 | { | |
483 | int i; | |
484 | ||
59d13bf5 JA |
485 | bt->bits_per_word = ilog2(BITS_PER_LONG); |
486 | ||
4bb659b1 JA |
487 | /* |
488 | * Depth can be zero for reserved tags, that's not a failure | |
489 | * condition. | |
490 | */ | |
491 | if (depth) { | |
e3a2b3f9 | 492 | unsigned int nr, tags_per_word; |
59d13bf5 JA |
493 | |
494 | tags_per_word = (1 << bt->bits_per_word); | |
495 | ||
496 | /* | |
497 | * If the tag space is small, shrink the number of tags | |
498 | * per word so we spread over a few cachelines, at least. | |
499 | * If less than 4 tags, just forget about it, it's not | |
500 | * going to work optimally anyway. | |
501 | */ | |
502 | if (depth >= 4) { | |
503 | while (tags_per_word * 4 > depth) { | |
504 | bt->bits_per_word--; | |
505 | tags_per_word = (1 << bt->bits_per_word); | |
506 | } | |
507 | } | |
4bb659b1 | 508 | |
59d13bf5 | 509 | nr = ALIGN(depth, tags_per_word) / tags_per_word; |
e93ecf60 | 510 | bt->map = kzalloc_node(nr * sizeof(struct blk_align_bitmap), |
4bb659b1 JA |
511 | GFP_KERNEL, node); |
512 | if (!bt->map) | |
513 | return -ENOMEM; | |
514 | ||
515 | bt->map_nr = nr; | |
4bb659b1 JA |
516 | } |
517 | ||
518 | bt->bs = kzalloc(BT_WAIT_QUEUES * sizeof(*bt->bs), GFP_KERNEL); | |
519 | if (!bt->bs) { | |
520 | kfree(bt->map); | |
521 | return -ENOMEM; | |
522 | } | |
523 | ||
86fb5c56 AG |
524 | bt_update_count(bt, depth); |
525 | ||
526 | for (i = 0; i < BT_WAIT_QUEUES; i++) { | |
4bb659b1 | 527 | init_waitqueue_head(&bt->bs[i].wait); |
86fb5c56 AG |
528 | atomic_set(&bt->bs[i].wait_cnt, bt->wake_cnt); |
529 | } | |
4bb659b1 | 530 | |
4bb659b1 JA |
531 | return 0; |
532 | } | |
533 | ||
534 | static void bt_free(struct blk_mq_bitmap_tags *bt) | |
535 | { | |
536 | kfree(bt->map); | |
537 | kfree(bt->bs); | |
538 | } | |
539 | ||
540 | static struct blk_mq_tags *blk_mq_init_bitmap_tags(struct blk_mq_tags *tags, | |
541 | int node) | |
542 | { | |
543 | unsigned int depth = tags->nr_tags - tags->nr_reserved_tags; | |
544 | ||
545 | if (bt_alloc(&tags->bitmap_tags, depth, node, false)) | |
546 | goto enomem; | |
547 | if (bt_alloc(&tags->breserved_tags, tags->nr_reserved_tags, node, true)) | |
548 | goto enomem; | |
549 | ||
550 | return tags; | |
551 | enomem: | |
552 | bt_free(&tags->bitmap_tags); | |
553 | kfree(tags); | |
554 | return NULL; | |
555 | } | |
556 | ||
320ae51f JA |
557 | struct blk_mq_tags *blk_mq_init_tags(unsigned int total_tags, |
558 | unsigned int reserved_tags, int node) | |
559 | { | |
320ae51f | 560 | struct blk_mq_tags *tags; |
320ae51f JA |
561 | |
562 | if (total_tags > BLK_MQ_TAG_MAX) { | |
563 | pr_err("blk-mq: tag depth too large\n"); | |
564 | return NULL; | |
565 | } | |
566 | ||
567 | tags = kzalloc_node(sizeof(*tags), GFP_KERNEL, node); | |
568 | if (!tags) | |
569 | return NULL; | |
570 | ||
320ae51f JA |
571 | tags->nr_tags = total_tags; |
572 | tags->nr_reserved_tags = reserved_tags; | |
320ae51f | 573 | |
4bb659b1 | 574 | return blk_mq_init_bitmap_tags(tags, node); |
320ae51f JA |
575 | } |
576 | ||
577 | void blk_mq_free_tags(struct blk_mq_tags *tags) | |
578 | { | |
4bb659b1 JA |
579 | bt_free(&tags->bitmap_tags); |
580 | bt_free(&tags->breserved_tags); | |
320ae51f JA |
581 | kfree(tags); |
582 | } | |
583 | ||
4bb659b1 JA |
584 | void blk_mq_tag_init_last_tag(struct blk_mq_tags *tags, unsigned int *tag) |
585 | { | |
586 | unsigned int depth = tags->nr_tags - tags->nr_reserved_tags; | |
587 | ||
9d3d21ae | 588 | *tag = prandom_u32() % depth; |
4bb659b1 JA |
589 | } |
590 | ||
e3a2b3f9 JA |
591 | int blk_mq_tag_update_depth(struct blk_mq_tags *tags, unsigned int tdepth) |
592 | { | |
593 | tdepth -= tags->nr_reserved_tags; | |
594 | if (tdepth > tags->nr_tags) | |
595 | return -EINVAL; | |
596 | ||
597 | /* | |
598 | * Don't need (or can't) update reserved tags here, they remain | |
599 | * static and should never need resizing. | |
600 | */ | |
601 | bt_update_count(&tags->bitmap_tags, tdepth); | |
aed3ea94 | 602 | blk_mq_tag_wakeup_all(tags, false); |
e3a2b3f9 JA |
603 | return 0; |
604 | } | |
605 | ||
205fb5f5 BVA |
606 | /** |
607 | * blk_mq_unique_tag() - return a tag that is unique queue-wide | |
608 | * @rq: request for which to compute a unique tag | |
609 | * | |
610 | * The tag field in struct request is unique per hardware queue but not over | |
611 | * all hardware queues. Hence this function that returns a tag with the | |
612 | * hardware context index in the upper bits and the per hardware queue tag in | |
613 | * the lower bits. | |
614 | * | |
615 | * Note: When called for a request that is queued on a non-multiqueue request | |
616 | * queue, the hardware context index is set to zero. | |
617 | */ | |
618 | u32 blk_mq_unique_tag(struct request *rq) | |
619 | { | |
620 | struct request_queue *q = rq->q; | |
621 | struct blk_mq_hw_ctx *hctx; | |
622 | int hwq = 0; | |
623 | ||
624 | if (q->mq_ops) { | |
625 | hctx = q->mq_ops->map_queue(q, rq->mq_ctx->cpu); | |
626 | hwq = hctx->queue_num; | |
627 | } | |
628 | ||
629 | return (hwq << BLK_MQ_UNIQUE_TAG_BITS) | | |
630 | (rq->tag & BLK_MQ_UNIQUE_TAG_MASK); | |
631 | } | |
632 | EXPORT_SYMBOL(blk_mq_unique_tag); | |
633 | ||
320ae51f JA |
634 | ssize_t blk_mq_tag_sysfs_show(struct blk_mq_tags *tags, char *page) |
635 | { | |
636 | char *orig_page = page; | |
4bb659b1 | 637 | unsigned int free, res; |
320ae51f JA |
638 | |
639 | if (!tags) | |
640 | return 0; | |
641 | ||
59d13bf5 JA |
642 | page += sprintf(page, "nr_tags=%u, reserved_tags=%u, " |
643 | "bits_per_word=%u\n", | |
644 | tags->nr_tags, tags->nr_reserved_tags, | |
645 | tags->bitmap_tags.bits_per_word); | |
320ae51f | 646 | |
4bb659b1 JA |
647 | free = bt_unused_tags(&tags->bitmap_tags); |
648 | res = bt_unused_tags(&tags->breserved_tags); | |
320ae51f | 649 | |
4bb659b1 | 650 | page += sprintf(page, "nr_free=%u, nr_reserved=%u\n", free, res); |
0d2602ca | 651 | page += sprintf(page, "active_queues=%u\n", atomic_read(&tags->active_queues)); |
320ae51f JA |
652 | |
653 | return page - orig_page; | |
654 | } |