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 return __sbitmap_queue_get(bt
);
102 unsigned int blk_mq_get_tag(struct blk_mq_alloc_data
*data
)
104 struct blk_mq_tags
*tags
= blk_mq_tags_from_data(data
);
105 struct sbitmap_queue
*bt
;
106 struct sbq_wait_state
*ws
;
108 unsigned int tag_offset
;
112 if (data
->flags
& BLK_MQ_REQ_RESERVED
) {
113 if (unlikely(!tags
->nr_reserved_tags
)) {
115 return BLK_MQ_TAG_FAIL
;
117 bt
= &tags
->breserved_tags
;
120 bt
= &tags
->bitmap_tags
;
121 tag_offset
= tags
->nr_reserved_tags
;
124 tag
= __blk_mq_get_tag(data
, bt
);
128 if (data
->flags
& BLK_MQ_REQ_NOWAIT
)
129 return BLK_MQ_TAG_FAIL
;
131 ws
= bt_wait_ptr(bt
, data
->hctx
);
132 drop_ctx
= data
->ctx
== NULL
;
134 prepare_to_wait(&ws
->wait
, &wait
, TASK_UNINTERRUPTIBLE
);
136 tag
= __blk_mq_get_tag(data
, bt
);
141 * We're out of tags on this hardware queue, kick any
142 * pending IO submits before going to sleep waiting for
145 blk_mq_run_hw_queue(data
->hctx
, false);
148 * Retry tag allocation after running the hardware queue,
149 * as running the queue may also have found completions.
151 tag
= __blk_mq_get_tag(data
, bt
);
156 blk_mq_put_ctx(data
->ctx
);
160 data
->ctx
= blk_mq_get_ctx(data
->q
);
161 data
->hctx
= blk_mq_map_queue(data
->q
, data
->ctx
->cpu
);
162 tags
= blk_mq_tags_from_data(data
);
163 if (data
->flags
& BLK_MQ_REQ_RESERVED
)
164 bt
= &tags
->breserved_tags
;
166 bt
= &tags
->bitmap_tags
;
168 finish_wait(&ws
->wait
, &wait
);
169 ws
= bt_wait_ptr(bt
, data
->hctx
);
172 if (drop_ctx
&& data
->ctx
)
173 blk_mq_put_ctx(data
->ctx
);
175 finish_wait(&ws
->wait
, &wait
);
178 return tag
+ tag_offset
;
181 void blk_mq_put_tag(struct blk_mq_hw_ctx
*hctx
, struct blk_mq_tags
*tags
,
182 struct blk_mq_ctx
*ctx
, unsigned int tag
)
184 if (!blk_mq_tag_is_reserved(tags
, tag
)) {
185 const int real_tag
= tag
- tags
->nr_reserved_tags
;
187 BUG_ON(real_tag
>= tags
->nr_tags
);
188 sbitmap_queue_clear(&tags
->bitmap_tags
, real_tag
, ctx
->cpu
);
190 BUG_ON(tag
>= tags
->nr_reserved_tags
);
191 sbitmap_queue_clear(&tags
->breserved_tags
, tag
, ctx
->cpu
);
195 struct bt_iter_data
{
196 struct blk_mq_hw_ctx
*hctx
;
202 static bool bt_iter(struct sbitmap
*bitmap
, unsigned int bitnr
, void *data
)
204 struct bt_iter_data
*iter_data
= data
;
205 struct blk_mq_hw_ctx
*hctx
= iter_data
->hctx
;
206 struct blk_mq_tags
*tags
= hctx
->tags
;
207 bool reserved
= iter_data
->reserved
;
211 bitnr
+= tags
->nr_reserved_tags
;
212 rq
= tags
->rqs
[bitnr
];
214 if (rq
->q
== hctx
->queue
)
215 iter_data
->fn(hctx
, rq
, iter_data
->data
, reserved
);
219 static void bt_for_each(struct blk_mq_hw_ctx
*hctx
, struct sbitmap_queue
*bt
,
220 busy_iter_fn
*fn
, void *data
, bool reserved
)
222 struct bt_iter_data iter_data
= {
226 .reserved
= reserved
,
229 sbitmap_for_each_set(&bt
->sb
, bt_iter
, &iter_data
);
232 struct bt_tags_iter_data
{
233 struct blk_mq_tags
*tags
;
234 busy_tag_iter_fn
*fn
;
239 static bool bt_tags_iter(struct sbitmap
*bitmap
, unsigned int bitnr
, void *data
)
241 struct bt_tags_iter_data
*iter_data
= data
;
242 struct blk_mq_tags
*tags
= iter_data
->tags
;
243 bool reserved
= iter_data
->reserved
;
247 bitnr
+= tags
->nr_reserved_tags
;
248 rq
= tags
->rqs
[bitnr
];
250 iter_data
->fn(rq
, iter_data
->data
, reserved
);
254 static void bt_tags_for_each(struct blk_mq_tags
*tags
, struct sbitmap_queue
*bt
,
255 busy_tag_iter_fn
*fn
, void *data
, bool reserved
)
257 struct bt_tags_iter_data iter_data
= {
261 .reserved
= reserved
,
265 sbitmap_for_each_set(&bt
->sb
, bt_tags_iter
, &iter_data
);
268 static void blk_mq_all_tag_busy_iter(struct blk_mq_tags
*tags
,
269 busy_tag_iter_fn
*fn
, void *priv
)
271 if (tags
->nr_reserved_tags
)
272 bt_tags_for_each(tags
, &tags
->breserved_tags
, fn
, priv
, true);
273 bt_tags_for_each(tags
, &tags
->bitmap_tags
, fn
, priv
, false);
276 void blk_mq_tagset_busy_iter(struct blk_mq_tag_set
*tagset
,
277 busy_tag_iter_fn
*fn
, void *priv
)
281 for (i
= 0; i
< tagset
->nr_hw_queues
; i
++) {
282 if (tagset
->tags
&& tagset
->tags
[i
])
283 blk_mq_all_tag_busy_iter(tagset
->tags
[i
], fn
, priv
);
286 EXPORT_SYMBOL(blk_mq_tagset_busy_iter
);
288 int blk_mq_reinit_tagset(struct blk_mq_tag_set
*set
)
292 if (!set
->ops
->reinit_request
)
295 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
296 struct blk_mq_tags
*tags
= set
->tags
[i
];
298 for (j
= 0; j
< tags
->nr_tags
; j
++) {
299 if (!tags
->static_rqs
[j
])
302 ret
= set
->ops
->reinit_request(set
->driver_data
,
303 tags
->static_rqs
[j
]);
312 EXPORT_SYMBOL_GPL(blk_mq_reinit_tagset
);
314 void blk_mq_queue_tag_busy_iter(struct request_queue
*q
, busy_iter_fn
*fn
,
317 struct blk_mq_hw_ctx
*hctx
;
321 queue_for_each_hw_ctx(q
, hctx
, i
) {
322 struct blk_mq_tags
*tags
= hctx
->tags
;
325 * If not software queues are currently mapped to this
326 * hardware queue, there's nothing to check
328 if (!blk_mq_hw_queue_mapped(hctx
))
331 if (tags
->nr_reserved_tags
)
332 bt_for_each(hctx
, &tags
->breserved_tags
, fn
, priv
, true);
333 bt_for_each(hctx
, &tags
->bitmap_tags
, fn
, priv
, false);
338 static int bt_alloc(struct sbitmap_queue
*bt
, unsigned int depth
,
339 bool round_robin
, int node
)
341 return sbitmap_queue_init_node(bt
, depth
, -1, round_robin
, GFP_KERNEL
,
345 static struct blk_mq_tags
*blk_mq_init_bitmap_tags(struct blk_mq_tags
*tags
,
346 int node
, int alloc_policy
)
348 unsigned int depth
= tags
->nr_tags
- tags
->nr_reserved_tags
;
349 bool round_robin
= alloc_policy
== BLK_TAG_ALLOC_RR
;
351 if (bt_alloc(&tags
->bitmap_tags
, depth
, round_robin
, node
))
353 if (bt_alloc(&tags
->breserved_tags
, tags
->nr_reserved_tags
, round_robin
,
355 goto free_bitmap_tags
;
359 sbitmap_queue_free(&tags
->bitmap_tags
);
365 struct blk_mq_tags
*blk_mq_init_tags(unsigned int total_tags
,
366 unsigned int reserved_tags
,
367 int node
, int alloc_policy
)
369 struct blk_mq_tags
*tags
;
371 if (total_tags
> BLK_MQ_TAG_MAX
) {
372 pr_err("blk-mq: tag depth too large\n");
376 tags
= kzalloc_node(sizeof(*tags
), GFP_KERNEL
, node
);
380 tags
->nr_tags
= total_tags
;
381 tags
->nr_reserved_tags
= reserved_tags
;
383 return blk_mq_init_bitmap_tags(tags
, node
, alloc_policy
);
386 void blk_mq_free_tags(struct blk_mq_tags
*tags
)
388 sbitmap_queue_free(&tags
->bitmap_tags
);
389 sbitmap_queue_free(&tags
->breserved_tags
);
393 int blk_mq_tag_update_depth(struct blk_mq_hw_ctx
*hctx
,
394 struct blk_mq_tags
**tagsptr
, unsigned int tdepth
,
397 struct blk_mq_tags
*tags
= *tagsptr
;
399 if (tdepth
<= tags
->nr_reserved_tags
)
402 tdepth
-= tags
->nr_reserved_tags
;
405 * If we are allowed to grow beyond the original size, allocate
406 * a new set of tags before freeing the old one.
408 if (tdepth
> tags
->nr_tags
) {
409 struct blk_mq_tag_set
*set
= hctx
->queue
->tag_set
;
410 struct blk_mq_tags
*new;
417 * We need some sort of upper limit, set it high enough that
418 * no valid use cases should require more.
420 if (tdepth
> 16 * BLKDEV_MAX_RQ
)
423 new = blk_mq_alloc_rq_map(set
, hctx
->queue_num
, tdepth
, 0);
426 ret
= blk_mq_alloc_rqs(set
, new, hctx
->queue_num
, tdepth
);
428 blk_mq_free_rq_map(new);
432 blk_mq_free_rqs(set
, *tagsptr
, hctx
->queue_num
);
433 blk_mq_free_rq_map(*tagsptr
);
437 * Don't need (or can't) update reserved tags here, they
438 * remain static and should never need resizing.
440 sbitmap_queue_resize(&tags
->bitmap_tags
, tdepth
);
447 * blk_mq_unique_tag() - return a tag that is unique queue-wide
448 * @rq: request for which to compute a unique tag
450 * The tag field in struct request is unique per hardware queue but not over
451 * all hardware queues. Hence this function that returns a tag with the
452 * hardware context index in the upper bits and the per hardware queue tag in
455 * Note: When called for a request that is queued on a non-multiqueue request
456 * queue, the hardware context index is set to zero.
458 u32
blk_mq_unique_tag(struct request
*rq
)
460 struct request_queue
*q
= rq
->q
;
461 struct blk_mq_hw_ctx
*hctx
;
465 hctx
= blk_mq_map_queue(q
, rq
->mq_ctx
->cpu
);
466 hwq
= hctx
->queue_num
;
469 return (hwq
<< BLK_MQ_UNIQUE_TAG_BITS
) |
470 (rq
->tag
& BLK_MQ_UNIQUE_TAG_MASK
);
472 EXPORT_SYMBOL(blk_mq_unique_tag
);