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
6 * Copyright (C) 2013-2014 Jens Axboe
8 #include <linux/kernel.h>
9 #include <linux/module.h>
11 #include <linux/blk-mq.h>
14 #include "blk-mq-tag.h"
16 bool blk_mq_has_free_tags(struct blk_mq_tags
*tags
)
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
);
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
);
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
))
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
))
73 if (!test_bit(BLK_MQ_S_TAG_ACTIVE
, &hctx
->state
))
77 * Don't try dividing an ant
79 if (bt
->sb
.depth
== 1)
82 users
= atomic_read(&hctx
->tags
->active_queues
);
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
))
99 if (data
->shallow_depth
)
100 return __sbitmap_queue_get_shallow(bt
, data
->shallow_depth
);
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
;
111 unsigned int tag_offset
;
115 if (data
->flags
& BLK_MQ_REQ_RESERVED
) {
116 if (unlikely(!tags
->nr_reserved_tags
)) {
118 return BLK_MQ_TAG_FAIL
;
120 bt
= &tags
->breserved_tags
;
123 bt
= &tags
->bitmap_tags
;
124 tag_offset
= tags
->nr_reserved_tags
;
127 tag
= __blk_mq_get_tag(data
, bt
);
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
;
137 prepare_to_wait(&ws
->wait
, &wait
, TASK_UNINTERRUPTIBLE
);
139 tag
= __blk_mq_get_tag(data
, bt
);
144 * We're out of tags on this hardware queue, kick any
145 * pending IO submits before going to sleep waiting for
148 blk_mq_run_hw_queue(data
->hctx
, false);
151 * Retry tag allocation after running the hardware queue,
152 * as running the queue may also have found completions.
154 tag
= __blk_mq_get_tag(data
, bt
);
159 blk_mq_put_ctx(data
->ctx
);
163 data
->ctx
= blk_mq_get_ctx(data
->q
);
164 data
->hctx
= blk_mq_map_queue(data
->q
, data
->ctx
->cpu
);
165 tags
= blk_mq_tags_from_data(data
);
166 if (data
->flags
& BLK_MQ_REQ_RESERVED
)
167 bt
= &tags
->breserved_tags
;
169 bt
= &tags
->bitmap_tags
;
171 finish_wait(&ws
->wait
, &wait
);
172 ws
= bt_wait_ptr(bt
, data
->hctx
);
175 if (drop_ctx
&& data
->ctx
)
176 blk_mq_put_ctx(data
->ctx
);
178 finish_wait(&ws
->wait
, &wait
);
181 return tag
+ tag_offset
;
184 void blk_mq_put_tag(struct blk_mq_hw_ctx
*hctx
, struct blk_mq_tags
*tags
,
185 struct blk_mq_ctx
*ctx
, unsigned int tag
)
187 if (!blk_mq_tag_is_reserved(tags
, tag
)) {
188 const int real_tag
= tag
- tags
->nr_reserved_tags
;
190 BUG_ON(real_tag
>= tags
->nr_tags
);
191 sbitmap_queue_clear(&tags
->bitmap_tags
, real_tag
, ctx
->cpu
);
193 BUG_ON(tag
>= tags
->nr_reserved_tags
);
194 sbitmap_queue_clear(&tags
->breserved_tags
, tag
, ctx
->cpu
);
198 struct bt_iter_data
{
199 struct blk_mq_hw_ctx
*hctx
;
205 static bool bt_iter(struct sbitmap
*bitmap
, unsigned int bitnr
, void *data
)
207 struct bt_iter_data
*iter_data
= data
;
208 struct blk_mq_hw_ctx
*hctx
= iter_data
->hctx
;
209 struct blk_mq_tags
*tags
= hctx
->tags
;
210 bool reserved
= iter_data
->reserved
;
214 bitnr
+= tags
->nr_reserved_tags
;
215 rq
= tags
->rqs
[bitnr
];
218 * We can hit rq == NULL here, because the tagging functions
219 * test and set the bit before assining ->rqs[].
221 if (rq
&& rq
->q
== hctx
->queue
)
222 iter_data
->fn(hctx
, rq
, iter_data
->data
, reserved
);
226 static void bt_for_each(struct blk_mq_hw_ctx
*hctx
, struct sbitmap_queue
*bt
,
227 busy_iter_fn
*fn
, void *data
, bool reserved
)
229 struct bt_iter_data iter_data
= {
233 .reserved
= reserved
,
236 sbitmap_for_each_set(&bt
->sb
, bt_iter
, &iter_data
);
239 struct bt_tags_iter_data
{
240 struct blk_mq_tags
*tags
;
241 busy_tag_iter_fn
*fn
;
246 static bool bt_tags_iter(struct sbitmap
*bitmap
, unsigned int bitnr
, void *data
)
248 struct bt_tags_iter_data
*iter_data
= data
;
249 struct blk_mq_tags
*tags
= iter_data
->tags
;
250 bool reserved
= iter_data
->reserved
;
254 bitnr
+= tags
->nr_reserved_tags
;
257 * We can hit rq == NULL here, because the tagging functions
258 * test and set the bit before assining ->rqs[].
260 rq
= tags
->rqs
[bitnr
];
262 iter_data
->fn(rq
, iter_data
->data
, reserved
);
267 static void bt_tags_for_each(struct blk_mq_tags
*tags
, struct sbitmap_queue
*bt
,
268 busy_tag_iter_fn
*fn
, void *data
, bool reserved
)
270 struct bt_tags_iter_data iter_data
= {
274 .reserved
= reserved
,
278 sbitmap_for_each_set(&bt
->sb
, bt_tags_iter
, &iter_data
);
281 static void blk_mq_all_tag_busy_iter(struct blk_mq_tags
*tags
,
282 busy_tag_iter_fn
*fn
, void *priv
)
284 if (tags
->nr_reserved_tags
)
285 bt_tags_for_each(tags
, &tags
->breserved_tags
, fn
, priv
, true);
286 bt_tags_for_each(tags
, &tags
->bitmap_tags
, fn
, priv
, false);
289 void blk_mq_tagset_busy_iter(struct blk_mq_tag_set
*tagset
,
290 busy_tag_iter_fn
*fn
, void *priv
)
294 for (i
= 0; i
< tagset
->nr_hw_queues
; i
++) {
295 if (tagset
->tags
&& tagset
->tags
[i
])
296 blk_mq_all_tag_busy_iter(tagset
->tags
[i
], fn
, priv
);
299 EXPORT_SYMBOL(blk_mq_tagset_busy_iter
);
301 int blk_mq_tagset_iter(struct blk_mq_tag_set
*set
, void *data
,
302 int (fn
)(void *, struct request
*))
306 if (WARN_ON_ONCE(!fn
))
309 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
310 struct blk_mq_tags
*tags
= set
->tags
[i
];
315 for (j
= 0; j
< tags
->nr_tags
; j
++) {
316 if (!tags
->static_rqs
[j
])
319 ret
= fn(data
, tags
->static_rqs
[j
]);
328 EXPORT_SYMBOL_GPL(blk_mq_tagset_iter
);
330 void blk_mq_queue_tag_busy_iter(struct request_queue
*q
, busy_iter_fn
*fn
,
333 struct blk_mq_hw_ctx
*hctx
;
337 queue_for_each_hw_ctx(q
, hctx
, i
) {
338 struct blk_mq_tags
*tags
= hctx
->tags
;
341 * If not software queues are currently mapped to this
342 * hardware queue, there's nothing to check
344 if (!blk_mq_hw_queue_mapped(hctx
))
347 if (tags
->nr_reserved_tags
)
348 bt_for_each(hctx
, &tags
->breserved_tags
, fn
, priv
, true);
349 bt_for_each(hctx
, &tags
->bitmap_tags
, fn
, priv
, false);
354 static int bt_alloc(struct sbitmap_queue
*bt
, unsigned int depth
,
355 bool round_robin
, int node
)
357 return sbitmap_queue_init_node(bt
, depth
, -1, round_robin
, GFP_KERNEL
,
361 static struct blk_mq_tags
*blk_mq_init_bitmap_tags(struct blk_mq_tags
*tags
,
362 int node
, int alloc_policy
)
364 unsigned int depth
= tags
->nr_tags
- tags
->nr_reserved_tags
;
365 bool round_robin
= alloc_policy
== BLK_TAG_ALLOC_RR
;
367 if (bt_alloc(&tags
->bitmap_tags
, depth
, round_robin
, node
))
369 if (bt_alloc(&tags
->breserved_tags
, tags
->nr_reserved_tags
, round_robin
,
371 goto free_bitmap_tags
;
375 sbitmap_queue_free(&tags
->bitmap_tags
);
381 struct blk_mq_tags
*blk_mq_init_tags(unsigned int total_tags
,
382 unsigned int reserved_tags
,
383 int node
, int alloc_policy
)
385 struct blk_mq_tags
*tags
;
387 if (total_tags
> BLK_MQ_TAG_MAX
) {
388 pr_err("blk-mq: tag depth too large\n");
392 tags
= kzalloc_node(sizeof(*tags
), GFP_KERNEL
, node
);
396 tags
->nr_tags
= total_tags
;
397 tags
->nr_reserved_tags
= reserved_tags
;
399 return blk_mq_init_bitmap_tags(tags
, node
, alloc_policy
);
402 void blk_mq_free_tags(struct blk_mq_tags
*tags
)
404 sbitmap_queue_free(&tags
->bitmap_tags
);
405 sbitmap_queue_free(&tags
->breserved_tags
);
409 int blk_mq_tag_update_depth(struct blk_mq_hw_ctx
*hctx
,
410 struct blk_mq_tags
**tagsptr
, unsigned int tdepth
,
413 struct blk_mq_tags
*tags
= *tagsptr
;
415 if (tdepth
<= tags
->nr_reserved_tags
)
418 tdepth
-= tags
->nr_reserved_tags
;
421 * If we are allowed to grow beyond the original size, allocate
422 * a new set of tags before freeing the old one.
424 if (tdepth
> tags
->nr_tags
) {
425 struct blk_mq_tag_set
*set
= hctx
->queue
->tag_set
;
426 struct blk_mq_tags
*new;
433 * We need some sort of upper limit, set it high enough that
434 * no valid use cases should require more.
436 if (tdepth
> 16 * BLKDEV_MAX_RQ
)
439 new = blk_mq_alloc_rq_map(set
, hctx
->queue_num
, tdepth
, 0);
442 ret
= blk_mq_alloc_rqs(set
, new, hctx
->queue_num
, tdepth
);
444 blk_mq_free_rq_map(new);
448 blk_mq_free_rqs(set
, *tagsptr
, hctx
->queue_num
);
449 blk_mq_free_rq_map(*tagsptr
);
453 * Don't need (or can't) update reserved tags here, they
454 * remain static and should never need resizing.
456 sbitmap_queue_resize(&tags
->bitmap_tags
, tdepth
);
463 * blk_mq_unique_tag() - return a tag that is unique queue-wide
464 * @rq: request for which to compute a unique tag
466 * The tag field in struct request is unique per hardware queue but not over
467 * all hardware queues. Hence this function that returns a tag with the
468 * hardware context index in the upper bits and the per hardware queue tag in
471 * Note: When called for a request that is queued on a non-multiqueue request
472 * queue, the hardware context index is set to zero.
474 u32
blk_mq_unique_tag(struct request
*rq
)
476 struct request_queue
*q
= rq
->q
;
477 struct blk_mq_hw_ctx
*hctx
;
481 hctx
= blk_mq_map_queue(q
, rq
->mq_ctx
->cpu
);
482 hwq
= hctx
->queue_num
;
485 return (hwq
<< BLK_MQ_UNIQUE_TAG_BITS
) |
486 (rq
->tag
& BLK_MQ_UNIQUE_TAG_MASK
);
488 EXPORT_SYMBOL(blk_mq_unique_tag
);