Merge tag 'scsi-fixes' of git://git.kernel.org/pub/scm/linux/kernel/git/jejb/scsi
[cris-mirror.git] / block / blk-mq-tag.c
blob336dde07b230636b82ef81581a63620734298c5d
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.
6 * Copyright (C) 2013-2014 Jens Axboe
7 */
8 #include <linux/kernel.h>
9 #include <linux/module.h>
11 #include <linux/blk-mq.h>
12 #include "blk.h"
13 #include "blk-mq.h"
14 #include "blk-mq-tag.h"
16 bool blk_mq_has_free_tags(struct blk_mq_tags *tags)
18 if (!tags)
19 return true;
21 return sbitmap_any_bit_clear(&tags->bitmap_tags.sb);
25 * If a previously inactive queue goes active, bump the active user count.
27 bool __blk_mq_tag_busy(struct blk_mq_hw_ctx *hctx)
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);
33 return true;
37 * Wakeup all potentially sleeping on tags
39 void blk_mq_tag_wakeup_all(struct blk_mq_tags *tags, bool include_reserve)
41 sbitmap_queue_wake_all(&tags->bitmap_tags);
42 if (include_reserve)
43 sbitmap_queue_wake_all(&tags->breserved_tags);
47 * If a previously busy queue goes inactive, potential waiters could now
48 * be allowed to queue. Wake them up and check.
50 void __blk_mq_tag_idle(struct blk_mq_hw_ctx *hctx)
52 struct blk_mq_tags *tags = hctx->tags;
54 if (!test_and_clear_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state))
55 return;
57 atomic_dec(&tags->active_queues);
59 blk_mq_tag_wakeup_all(tags, false);
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.
66 static inline bool hctx_may_queue(struct blk_mq_hw_ctx *hctx,
67 struct sbitmap_queue *bt)
69 unsigned int depth, users;
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;
77 * Don't try dividing an ant
79 if (bt->sb.depth == 1)
80 return true;
82 users = atomic_read(&hctx->tags->active_queues);
83 if (!users)
84 return true;
87 * Allow at least some tags
89 depth = max((bt->sb.depth + users - 1) / users, 4U);
90 return atomic_read(&hctx->nr_active) < depth;
93 static int __blk_mq_get_tag(struct blk_mq_alloc_data *data,
94 struct sbitmap_queue *bt)
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);
105 unsigned int blk_mq_get_tag(struct blk_mq_alloc_data *data)
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;
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;
120 bt = &tags->breserved_tags;
121 tag_offset = 0;
122 } else {
123 bt = &tags->bitmap_tags;
124 tag_offset = tags->nr_reserved_tags;
127 tag = __blk_mq_get_tag(data, bt);
128 if (tag != -1)
129 goto found_tag;
131 if (data->flags & BLK_MQ_REQ_NOWAIT)
132 return BLK_MQ_TAG_FAIL;
134 ws = bt_wait_ptr(bt, data->hctx);
135 drop_ctx = data->ctx == NULL;
136 do {
138 * We're out of tags on this hardware queue, kick any
139 * pending IO submits before going to sleep waiting for
140 * some to complete.
142 blk_mq_run_hw_queue(data->hctx, false);
145 * Retry tag allocation after running the hardware queue,
146 * as running the queue may also have found completions.
148 tag = __blk_mq_get_tag(data, bt);
149 if (tag != -1)
150 break;
152 prepare_to_wait_exclusive(&ws->wait, &wait,
153 TASK_UNINTERRUPTIBLE);
155 tag = __blk_mq_get_tag(data, bt);
156 if (tag != -1)
157 break;
159 if (data->ctx)
160 blk_mq_put_ctx(data->ctx);
162 io_schedule();
164 data->ctx = blk_mq_get_ctx(data->q);
165 data->hctx = blk_mq_map_queue(data->q, data->ctx->cpu);
166 tags = blk_mq_tags_from_data(data);
167 if (data->flags & BLK_MQ_REQ_RESERVED)
168 bt = &tags->breserved_tags;
169 else
170 bt = &tags->bitmap_tags;
172 finish_wait(&ws->wait, &wait);
173 ws = bt_wait_ptr(bt, data->hctx);
174 } while (1);
176 if (drop_ctx && data->ctx)
177 blk_mq_put_ctx(data->ctx);
179 finish_wait(&ws->wait, &wait);
181 found_tag:
182 return tag + tag_offset;
185 void blk_mq_put_tag(struct blk_mq_hw_ctx *hctx, struct blk_mq_tags *tags,
186 struct blk_mq_ctx *ctx, unsigned int tag)
188 if (!blk_mq_tag_is_reserved(tags, tag)) {
189 const int real_tag = tag - tags->nr_reserved_tags;
191 BUG_ON(real_tag >= tags->nr_tags);
192 sbitmap_queue_clear(&tags->bitmap_tags, real_tag, ctx->cpu);
193 } else {
194 BUG_ON(tag >= tags->nr_reserved_tags);
195 sbitmap_queue_clear(&tags->breserved_tags, tag, ctx->cpu);
199 struct bt_iter_data {
200 struct blk_mq_hw_ctx *hctx;
201 busy_iter_fn *fn;
202 void *data;
203 bool reserved;
206 static bool bt_iter(struct sbitmap *bitmap, unsigned int bitnr, void *data)
208 struct bt_iter_data *iter_data = data;
209 struct blk_mq_hw_ctx *hctx = iter_data->hctx;
210 struct blk_mq_tags *tags = hctx->tags;
211 bool reserved = iter_data->reserved;
212 struct request *rq;
214 if (!reserved)
215 bitnr += tags->nr_reserved_tags;
216 rq = tags->rqs[bitnr];
219 * We can hit rq == NULL here, because the tagging functions
220 * test and set the bit before assining ->rqs[].
222 if (rq && rq->q == hctx->queue)
223 iter_data->fn(hctx, rq, iter_data->data, reserved);
224 return true;
227 static void bt_for_each(struct blk_mq_hw_ctx *hctx, struct sbitmap_queue *bt,
228 busy_iter_fn *fn, void *data, bool reserved)
230 struct bt_iter_data iter_data = {
231 .hctx = hctx,
232 .fn = fn,
233 .data = data,
234 .reserved = reserved,
237 sbitmap_for_each_set(&bt->sb, bt_iter, &iter_data);
240 struct bt_tags_iter_data {
241 struct blk_mq_tags *tags;
242 busy_tag_iter_fn *fn;
243 void *data;
244 bool reserved;
247 static bool bt_tags_iter(struct sbitmap *bitmap, unsigned int bitnr, void *data)
249 struct bt_tags_iter_data *iter_data = data;
250 struct blk_mq_tags *tags = iter_data->tags;
251 bool reserved = iter_data->reserved;
252 struct request *rq;
254 if (!reserved)
255 bitnr += tags->nr_reserved_tags;
258 * We can hit rq == NULL here, because the tagging functions
259 * test and set the bit before assining ->rqs[].
261 rq = tags->rqs[bitnr];
262 if (rq)
263 iter_data->fn(rq, iter_data->data, reserved);
265 return true;
268 static void bt_tags_for_each(struct blk_mq_tags *tags, struct sbitmap_queue *bt,
269 busy_tag_iter_fn *fn, void *data, bool reserved)
271 struct bt_tags_iter_data iter_data = {
272 .tags = tags,
273 .fn = fn,
274 .data = data,
275 .reserved = reserved,
278 if (tags->rqs)
279 sbitmap_for_each_set(&bt->sb, bt_tags_iter, &iter_data);
282 static void blk_mq_all_tag_busy_iter(struct blk_mq_tags *tags,
283 busy_tag_iter_fn *fn, void *priv)
285 if (tags->nr_reserved_tags)
286 bt_tags_for_each(tags, &tags->breserved_tags, fn, priv, true);
287 bt_tags_for_each(tags, &tags->bitmap_tags, fn, priv, false);
290 void blk_mq_tagset_busy_iter(struct blk_mq_tag_set *tagset,
291 busy_tag_iter_fn *fn, void *priv)
293 int i;
295 for (i = 0; i < tagset->nr_hw_queues; i++) {
296 if (tagset->tags && tagset->tags[i])
297 blk_mq_all_tag_busy_iter(tagset->tags[i], fn, priv);
300 EXPORT_SYMBOL(blk_mq_tagset_busy_iter);
302 int blk_mq_tagset_iter(struct blk_mq_tag_set *set, void *data,
303 int (fn)(void *, struct request *))
305 int i, j, ret = 0;
307 if (WARN_ON_ONCE(!fn))
308 goto out;
310 for (i = 0; i < set->nr_hw_queues; i++) {
311 struct blk_mq_tags *tags = set->tags[i];
313 if (!tags)
314 continue;
316 for (j = 0; j < tags->nr_tags; j++) {
317 if (!tags->static_rqs[j])
318 continue;
320 ret = fn(data, tags->static_rqs[j]);
321 if (ret)
322 goto out;
326 out:
327 return ret;
329 EXPORT_SYMBOL_GPL(blk_mq_tagset_iter);
331 void blk_mq_queue_tag_busy_iter(struct request_queue *q, busy_iter_fn *fn,
332 void *priv)
334 struct blk_mq_hw_ctx *hctx;
335 int i;
338 queue_for_each_hw_ctx(q, hctx, i) {
339 struct blk_mq_tags *tags = hctx->tags;
342 * If not software queues are currently mapped to this
343 * hardware queue, there's nothing to check
345 if (!blk_mq_hw_queue_mapped(hctx))
346 continue;
348 if (tags->nr_reserved_tags)
349 bt_for_each(hctx, &tags->breserved_tags, fn, priv, true);
350 bt_for_each(hctx, &tags->bitmap_tags, fn, priv, false);
355 static int bt_alloc(struct sbitmap_queue *bt, unsigned int depth,
356 bool round_robin, int node)
358 return sbitmap_queue_init_node(bt, depth, -1, round_robin, GFP_KERNEL,
359 node);
362 static struct blk_mq_tags *blk_mq_init_bitmap_tags(struct blk_mq_tags *tags,
363 int node, int alloc_policy)
365 unsigned int depth = tags->nr_tags - tags->nr_reserved_tags;
366 bool round_robin = alloc_policy == BLK_TAG_ALLOC_RR;
368 if (bt_alloc(&tags->bitmap_tags, depth, round_robin, node))
369 goto free_tags;
370 if (bt_alloc(&tags->breserved_tags, tags->nr_reserved_tags, round_robin,
371 node))
372 goto free_bitmap_tags;
374 return tags;
375 free_bitmap_tags:
376 sbitmap_queue_free(&tags->bitmap_tags);
377 free_tags:
378 kfree(tags);
379 return NULL;
382 struct blk_mq_tags *blk_mq_init_tags(unsigned int total_tags,
383 unsigned int reserved_tags,
384 int node, int alloc_policy)
386 struct blk_mq_tags *tags;
388 if (total_tags > BLK_MQ_TAG_MAX) {
389 pr_err("blk-mq: tag depth too large\n");
390 return NULL;
393 tags = kzalloc_node(sizeof(*tags), GFP_KERNEL, node);
394 if (!tags)
395 return NULL;
397 tags->nr_tags = total_tags;
398 tags->nr_reserved_tags = reserved_tags;
400 return blk_mq_init_bitmap_tags(tags, node, alloc_policy);
403 void blk_mq_free_tags(struct blk_mq_tags *tags)
405 sbitmap_queue_free(&tags->bitmap_tags);
406 sbitmap_queue_free(&tags->breserved_tags);
407 kfree(tags);
410 int blk_mq_tag_update_depth(struct blk_mq_hw_ctx *hctx,
411 struct blk_mq_tags **tagsptr, unsigned int tdepth,
412 bool can_grow)
414 struct blk_mq_tags *tags = *tagsptr;
416 if (tdepth <= tags->nr_reserved_tags)
417 return -EINVAL;
419 tdepth -= tags->nr_reserved_tags;
422 * If we are allowed to grow beyond the original size, allocate
423 * a new set of tags before freeing the old one.
425 if (tdepth > tags->nr_tags) {
426 struct blk_mq_tag_set *set = hctx->queue->tag_set;
427 struct blk_mq_tags *new;
428 bool ret;
430 if (!can_grow)
431 return -EINVAL;
434 * We need some sort of upper limit, set it high enough that
435 * no valid use cases should require more.
437 if (tdepth > 16 * BLKDEV_MAX_RQ)
438 return -EINVAL;
440 new = blk_mq_alloc_rq_map(set, hctx->queue_num, tdepth, 0);
441 if (!new)
442 return -ENOMEM;
443 ret = blk_mq_alloc_rqs(set, new, hctx->queue_num, tdepth);
444 if (ret) {
445 blk_mq_free_rq_map(new);
446 return -ENOMEM;
449 blk_mq_free_rqs(set, *tagsptr, hctx->queue_num);
450 blk_mq_free_rq_map(*tagsptr);
451 *tagsptr = new;
452 } else {
454 * Don't need (or can't) update reserved tags here, they
455 * remain static and should never need resizing.
457 sbitmap_queue_resize(&tags->bitmap_tags, tdepth);
460 return 0;
464 * blk_mq_unique_tag() - return a tag that is unique queue-wide
465 * @rq: request for which to compute a unique tag
467 * The tag field in struct request is unique per hardware queue but not over
468 * all hardware queues. Hence this function that returns a tag with the
469 * hardware context index in the upper bits and the per hardware queue tag in
470 * the lower bits.
472 * Note: When called for a request that is queued on a non-multiqueue request
473 * queue, the hardware context index is set to zero.
475 u32 blk_mq_unique_tag(struct request *rq)
477 struct request_queue *q = rq->q;
478 struct blk_mq_hw_ctx *hctx;
479 int hwq = 0;
481 if (q->mq_ops) {
482 hctx = blk_mq_map_queue(q, rq->mq_ctx->cpu);
483 hwq = hctx->queue_num;
486 return (hwq << BLK_MQ_UNIQUE_TAG_BITS) |
487 (rq->tag & BLK_MQ_UNIQUE_TAG_MASK);
489 EXPORT_SYMBOL(blk_mq_unique_tag);