]>
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); | |
144 | data->hctx = data->q->mq_ops->map_queue(data->q, | |
145 | data->ctx->cpu); | |
6f3b0e8b | 146 | if (data->flags & BLK_MQ_REQ_RESERVED) { |
cb96a42c ML |
147 | bt = &data->hctx->tags->breserved_tags; |
148 | } else { | |
cb96a42c ML |
149 | hctx = data->hctx; |
150 | bt = &hctx->tags->bitmap_tags; | |
151 | } | |
88459642 OS |
152 | finish_wait(&ws->wait, &wait); |
153 | ws = bt_wait_ptr(bt, hctx); | |
4bb659b1 JA |
154 | } while (1); |
155 | ||
88459642 | 156 | finish_wait(&ws->wait, &wait); |
4bb659b1 JA |
157 | return tag; |
158 | } | |
159 | ||
cb96a42c | 160 | static unsigned int __blk_mq_get_tag(struct blk_mq_alloc_data *data) |
4bb659b1 JA |
161 | { |
162 | int tag; | |
163 | ||
cb96a42c | 164 | tag = bt_get(data, &data->hctx->tags->bitmap_tags, data->hctx, |
40aabb67 | 165 | data->hctx->tags); |
4bb659b1 | 166 | if (tag >= 0) |
cb96a42c | 167 | return tag + data->hctx->tags->nr_reserved_tags; |
4bb659b1 JA |
168 | |
169 | return BLK_MQ_TAG_FAIL; | |
320ae51f JA |
170 | } |
171 | ||
cb96a42c | 172 | static unsigned int __blk_mq_get_reserved_tag(struct blk_mq_alloc_data *data) |
320ae51f | 173 | { |
40aabb67 | 174 | int tag; |
320ae51f | 175 | |
cb96a42c | 176 | if (unlikely(!data->hctx->tags->nr_reserved_tags)) { |
320ae51f JA |
177 | WARN_ON_ONCE(1); |
178 | return BLK_MQ_TAG_FAIL; | |
179 | } | |
180 | ||
40aabb67 OS |
181 | tag = bt_get(data, &data->hctx->tags->breserved_tags, NULL, |
182 | data->hctx->tags); | |
320ae51f JA |
183 | if (tag < 0) |
184 | return BLK_MQ_TAG_FAIL; | |
4bb659b1 | 185 | |
320ae51f JA |
186 | return tag; |
187 | } | |
188 | ||
cb96a42c | 189 | unsigned int blk_mq_get_tag(struct blk_mq_alloc_data *data) |
320ae51f | 190 | { |
6f3b0e8b CH |
191 | if (data->flags & BLK_MQ_REQ_RESERVED) |
192 | return __blk_mq_get_reserved_tag(data); | |
193 | return __blk_mq_get_tag(data); | |
320ae51f JA |
194 | } |
195 | ||
40aabb67 OS |
196 | void blk_mq_put_tag(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx, |
197 | unsigned int tag) | |
320ae51f | 198 | { |
0d2602ca JA |
199 | struct blk_mq_tags *tags = hctx->tags; |
200 | ||
4bb659b1 JA |
201 | if (tag >= tags->nr_reserved_tags) { |
202 | const int real_tag = tag - tags->nr_reserved_tags; | |
203 | ||
70114c39 | 204 | BUG_ON(real_tag >= tags->nr_tags); |
f4a644db | 205 | sbitmap_queue_clear(&tags->bitmap_tags, real_tag, ctx->cpu); |
70114c39 JA |
206 | } else { |
207 | BUG_ON(tag >= tags->nr_reserved_tags); | |
f4a644db | 208 | sbitmap_queue_clear(&tags->breserved_tags, tag, ctx->cpu); |
70114c39 | 209 | } |
320ae51f JA |
210 | } |
211 | ||
88459642 OS |
212 | struct bt_iter_data { |
213 | struct blk_mq_hw_ctx *hctx; | |
214 | busy_iter_fn *fn; | |
215 | void *data; | |
216 | bool reserved; | |
217 | }; | |
218 | ||
219 | static bool bt_iter(struct sbitmap *bitmap, unsigned int bitnr, void *data) | |
320ae51f | 220 | { |
88459642 OS |
221 | struct bt_iter_data *iter_data = data; |
222 | struct blk_mq_hw_ctx *hctx = iter_data->hctx; | |
223 | struct blk_mq_tags *tags = hctx->tags; | |
224 | bool reserved = iter_data->reserved; | |
81481eb4 | 225 | struct request *rq; |
4bb659b1 | 226 | |
88459642 OS |
227 | if (!reserved) |
228 | bitnr += tags->nr_reserved_tags; | |
229 | rq = tags->rqs[bitnr]; | |
4bb659b1 | 230 | |
88459642 OS |
231 | if (rq->q == hctx->queue) |
232 | iter_data->fn(hctx, rq, iter_data->data, reserved); | |
233 | return true; | |
234 | } | |
4bb659b1 | 235 | |
88459642 OS |
236 | static void bt_for_each(struct blk_mq_hw_ctx *hctx, struct sbitmap_queue *bt, |
237 | busy_iter_fn *fn, void *data, bool reserved) | |
238 | { | |
239 | struct bt_iter_data iter_data = { | |
240 | .hctx = hctx, | |
241 | .fn = fn, | |
242 | .data = data, | |
243 | .reserved = reserved, | |
244 | }; | |
245 | ||
246 | sbitmap_for_each_set(&bt->sb, bt_iter, &iter_data); | |
320ae51f JA |
247 | } |
248 | ||
88459642 OS |
249 | struct bt_tags_iter_data { |
250 | struct blk_mq_tags *tags; | |
251 | busy_tag_iter_fn *fn; | |
252 | void *data; | |
253 | bool reserved; | |
254 | }; | |
255 | ||
256 | static bool bt_tags_iter(struct sbitmap *bitmap, unsigned int bitnr, void *data) | |
f26cdc85 | 257 | { |
88459642 OS |
258 | struct bt_tags_iter_data *iter_data = data; |
259 | struct blk_mq_tags *tags = iter_data->tags; | |
260 | bool reserved = iter_data->reserved; | |
f26cdc85 | 261 | struct request *rq; |
f26cdc85 | 262 | |
88459642 OS |
263 | if (!reserved) |
264 | bitnr += tags->nr_reserved_tags; | |
265 | rq = tags->rqs[bitnr]; | |
f26cdc85 | 266 | |
88459642 OS |
267 | iter_data->fn(rq, iter_data->data, reserved); |
268 | return true; | |
269 | } | |
270 | ||
271 | static void bt_tags_for_each(struct blk_mq_tags *tags, struct sbitmap_queue *bt, | |
272 | busy_tag_iter_fn *fn, void *data, bool reserved) | |
273 | { | |
274 | struct bt_tags_iter_data iter_data = { | |
275 | .tags = tags, | |
276 | .fn = fn, | |
277 | .data = data, | |
278 | .reserved = reserved, | |
279 | }; | |
280 | ||
281 | if (tags->rqs) | |
282 | sbitmap_for_each_set(&bt->sb, bt_tags_iter, &iter_data); | |
f26cdc85 KB |
283 | } |
284 | ||
e8f1e163 SG |
285 | static void blk_mq_all_tag_busy_iter(struct blk_mq_tags *tags, |
286 | busy_tag_iter_fn *fn, void *priv) | |
f26cdc85 KB |
287 | { |
288 | if (tags->nr_reserved_tags) | |
88459642 OS |
289 | bt_tags_for_each(tags, &tags->breserved_tags, fn, priv, true); |
290 | bt_tags_for_each(tags, &tags->bitmap_tags, fn, priv, false); | |
f26cdc85 | 291 | } |
f26cdc85 | 292 | |
e0489487 SG |
293 | void blk_mq_tagset_busy_iter(struct blk_mq_tag_set *tagset, |
294 | busy_tag_iter_fn *fn, void *priv) | |
295 | { | |
296 | int i; | |
297 | ||
298 | for (i = 0; i < tagset->nr_hw_queues; i++) { | |
299 | if (tagset->tags && tagset->tags[i]) | |
300 | blk_mq_all_tag_busy_iter(tagset->tags[i], fn, priv); | |
301 | } | |
302 | } | |
303 | EXPORT_SYMBOL(blk_mq_tagset_busy_iter); | |
304 | ||
486cf989 SG |
305 | int blk_mq_reinit_tagset(struct blk_mq_tag_set *set) |
306 | { | |
307 | int i, j, ret = 0; | |
308 | ||
309 | if (!set->ops->reinit_request) | |
310 | goto out; | |
311 | ||
312 | for (i = 0; i < set->nr_hw_queues; i++) { | |
313 | struct blk_mq_tags *tags = set->tags[i]; | |
314 | ||
315 | for (j = 0; j < tags->nr_tags; j++) { | |
316 | if (!tags->rqs[j]) | |
317 | continue; | |
318 | ||
319 | ret = set->ops->reinit_request(set->driver_data, | |
320 | tags->rqs[j]); | |
321 | if (ret) | |
322 | goto out; | |
323 | } | |
324 | } | |
325 | ||
326 | out: | |
327 | return ret; | |
328 | } | |
329 | EXPORT_SYMBOL_GPL(blk_mq_reinit_tagset); | |
330 | ||
0bf6cd5b | 331 | void blk_mq_queue_tag_busy_iter(struct request_queue *q, busy_iter_fn *fn, |
81481eb4 | 332 | void *priv) |
320ae51f | 333 | { |
0bf6cd5b CH |
334 | struct blk_mq_hw_ctx *hctx; |
335 | int i; | |
336 | ||
337 | ||
338 | queue_for_each_hw_ctx(q, hctx, i) { | |
339 | struct blk_mq_tags *tags = hctx->tags; | |
340 | ||
341 | /* | |
342 | * If not software queues are currently mapped to this | |
343 | * hardware queue, there's nothing to check | |
344 | */ | |
345 | if (!blk_mq_hw_queue_mapped(hctx)) | |
346 | continue; | |
347 | ||
348 | if (tags->nr_reserved_tags) | |
88459642 OS |
349 | bt_for_each(hctx, &tags->breserved_tags, fn, priv, true); |
350 | bt_for_each(hctx, &tags->bitmap_tags, fn, priv, false); | |
4bb659b1 JA |
351 | } |
352 | ||
4bb659b1 JA |
353 | } |
354 | ||
88459642 | 355 | static unsigned int bt_unused_tags(const struct sbitmap_queue *bt) |
e3a2b3f9 | 356 | { |
88459642 | 357 | return bt->sb.depth - sbitmap_weight(&bt->sb); |
e3a2b3f9 JA |
358 | } |
359 | ||
f4a644db OS |
360 | static int bt_alloc(struct sbitmap_queue *bt, unsigned int depth, |
361 | bool round_robin, int node) | |
4bb659b1 | 362 | { |
f4a644db OS |
363 | return sbitmap_queue_init_node(bt, depth, -1, round_robin, GFP_KERNEL, |
364 | node); | |
4bb659b1 JA |
365 | } |
366 | ||
367 | static struct blk_mq_tags *blk_mq_init_bitmap_tags(struct blk_mq_tags *tags, | |
24391c0d | 368 | int node, int alloc_policy) |
4bb659b1 JA |
369 | { |
370 | unsigned int depth = tags->nr_tags - tags->nr_reserved_tags; | |
f4a644db | 371 | bool round_robin = alloc_policy == BLK_TAG_ALLOC_RR; |
4bb659b1 | 372 | |
f4a644db | 373 | if (bt_alloc(&tags->bitmap_tags, depth, round_robin, node)) |
88459642 | 374 | goto free_tags; |
f4a644db OS |
375 | if (bt_alloc(&tags->breserved_tags, tags->nr_reserved_tags, round_robin, |
376 | node)) | |
88459642 | 377 | goto free_bitmap_tags; |
4bb659b1 JA |
378 | |
379 | return tags; | |
88459642 OS |
380 | free_bitmap_tags: |
381 | sbitmap_queue_free(&tags->bitmap_tags); | |
382 | free_tags: | |
4bb659b1 JA |
383 | kfree(tags); |
384 | return NULL; | |
385 | } | |
386 | ||
320ae51f | 387 | struct blk_mq_tags *blk_mq_init_tags(unsigned int total_tags, |
24391c0d SL |
388 | unsigned int reserved_tags, |
389 | int node, int alloc_policy) | |
320ae51f | 390 | { |
320ae51f | 391 | struct blk_mq_tags *tags; |
320ae51f JA |
392 | |
393 | if (total_tags > BLK_MQ_TAG_MAX) { | |
394 | pr_err("blk-mq: tag depth too large\n"); | |
395 | return NULL; | |
396 | } | |
397 | ||
398 | tags = kzalloc_node(sizeof(*tags), GFP_KERNEL, node); | |
399 | if (!tags) | |
400 | return NULL; | |
401 | ||
f26cdc85 KB |
402 | if (!zalloc_cpumask_var(&tags->cpumask, GFP_KERNEL)) { |
403 | kfree(tags); | |
404 | return NULL; | |
405 | } | |
406 | ||
320ae51f JA |
407 | tags->nr_tags = total_tags; |
408 | tags->nr_reserved_tags = reserved_tags; | |
320ae51f | 409 | |
24391c0d | 410 | return blk_mq_init_bitmap_tags(tags, node, alloc_policy); |
320ae51f JA |
411 | } |
412 | ||
413 | void blk_mq_free_tags(struct blk_mq_tags *tags) | |
414 | { | |
88459642 OS |
415 | sbitmap_queue_free(&tags->bitmap_tags); |
416 | sbitmap_queue_free(&tags->breserved_tags); | |
f42d79ab | 417 | free_cpumask_var(tags->cpumask); |
320ae51f JA |
418 | kfree(tags); |
419 | } | |
420 | ||
e3a2b3f9 JA |
421 | int blk_mq_tag_update_depth(struct blk_mq_tags *tags, unsigned int tdepth) |
422 | { | |
423 | tdepth -= tags->nr_reserved_tags; | |
424 | if (tdepth > tags->nr_tags) | |
425 | return -EINVAL; | |
426 | ||
427 | /* | |
428 | * Don't need (or can't) update reserved tags here, they remain | |
429 | * static and should never need resizing. | |
430 | */ | |
88459642 OS |
431 | sbitmap_queue_resize(&tags->bitmap_tags, tdepth); |
432 | ||
aed3ea94 | 433 | blk_mq_tag_wakeup_all(tags, false); |
e3a2b3f9 JA |
434 | return 0; |
435 | } | |
436 | ||
205fb5f5 BVA |
437 | /** |
438 | * blk_mq_unique_tag() - return a tag that is unique queue-wide | |
439 | * @rq: request for which to compute a unique tag | |
440 | * | |
441 | * The tag field in struct request is unique per hardware queue but not over | |
442 | * all hardware queues. Hence this function that returns a tag with the | |
443 | * hardware context index in the upper bits and the per hardware queue tag in | |
444 | * the lower bits. | |
445 | * | |
446 | * Note: When called for a request that is queued on a non-multiqueue request | |
447 | * queue, the hardware context index is set to zero. | |
448 | */ | |
449 | u32 blk_mq_unique_tag(struct request *rq) | |
450 | { | |
451 | struct request_queue *q = rq->q; | |
452 | struct blk_mq_hw_ctx *hctx; | |
453 | int hwq = 0; | |
454 | ||
455 | if (q->mq_ops) { | |
456 | hctx = q->mq_ops->map_queue(q, rq->mq_ctx->cpu); | |
457 | hwq = hctx->queue_num; | |
458 | } | |
459 | ||
460 | return (hwq << BLK_MQ_UNIQUE_TAG_BITS) | | |
461 | (rq->tag & BLK_MQ_UNIQUE_TAG_MASK); | |
462 | } | |
463 | EXPORT_SYMBOL(blk_mq_unique_tag); | |
464 | ||
320ae51f JA |
465 | ssize_t blk_mq_tag_sysfs_show(struct blk_mq_tags *tags, char *page) |
466 | { | |
467 | char *orig_page = page; | |
4bb659b1 | 468 | unsigned int free, res; |
320ae51f JA |
469 | |
470 | if (!tags) | |
471 | return 0; | |
472 | ||
59d13bf5 JA |
473 | page += sprintf(page, "nr_tags=%u, reserved_tags=%u, " |
474 | "bits_per_word=%u\n", | |
475 | tags->nr_tags, tags->nr_reserved_tags, | |
88459642 | 476 | 1U << tags->bitmap_tags.sb.shift); |
320ae51f | 477 | |
4bb659b1 JA |
478 | free = bt_unused_tags(&tags->bitmap_tags); |
479 | res = bt_unused_tags(&tags->breserved_tags); | |
320ae51f | 480 | |
4bb659b1 | 481 | page += sprintf(page, "nr_free=%u, nr_reserved=%u\n", free, res); |
0d2602ca | 482 | page += sprintf(page, "active_queues=%u\n", atomic_read(&tags->active_queues)); |
320ae51f JA |
483 | |
484 | return page - orig_page; | |
485 | } |