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_hw_ctx
*hctx
, struct sbitmap_queue
*bt
)
95 if (!hctx_may_queue(hctx
, bt
))
97 return __sbitmap_queue_get(bt
);
100 unsigned int blk_mq_get_tag(struct blk_mq_alloc_data
*data
)
102 struct blk_mq_tags
*tags
= blk_mq_tags_from_data(data
);
103 struct sbitmap_queue
*bt
;
104 struct sbq_wait_state
*ws
;
106 unsigned int tag_offset
;
109 if (data
->flags
& BLK_MQ_REQ_RESERVED
) {
110 if (unlikely(!tags
->nr_reserved_tags
)) {
112 return BLK_MQ_TAG_FAIL
;
114 bt
= &tags
->breserved_tags
;
117 bt
= &tags
->bitmap_tags
;
118 tag_offset
= tags
->nr_reserved_tags
;
121 tag
= __blk_mq_get_tag(data
->hctx
, bt
);
125 if (data
->flags
& BLK_MQ_REQ_NOWAIT
)
126 return BLK_MQ_TAG_FAIL
;
128 ws
= bt_wait_ptr(bt
, data
->hctx
);
130 prepare_to_wait(&ws
->wait
, &wait
, TASK_UNINTERRUPTIBLE
);
132 tag
= __blk_mq_get_tag(data
->hctx
, bt
);
137 * We're out of tags on this hardware queue, kick any
138 * pending IO submits before going to sleep waiting for
141 blk_mq_run_hw_queue(data
->hctx
, false);
144 * Retry tag allocation after running the hardware queue,
145 * as running the queue may also have found completions.
147 tag
= __blk_mq_get_tag(data
->hctx
, bt
);
151 blk_mq_put_ctx(data
->ctx
);
155 data
->ctx
= blk_mq_get_ctx(data
->q
);
156 data
->hctx
= blk_mq_map_queue(data
->q
, data
->ctx
->cpu
);
157 tags
= blk_mq_tags_from_data(data
);
158 if (data
->flags
& BLK_MQ_REQ_RESERVED
)
159 bt
= &tags
->breserved_tags
;
161 bt
= &tags
->bitmap_tags
;
163 finish_wait(&ws
->wait
, &wait
);
164 ws
= bt_wait_ptr(bt
, data
->hctx
);
167 finish_wait(&ws
->wait
, &wait
);
170 return tag
+ tag_offset
;
173 void blk_mq_put_tag(struct blk_mq_hw_ctx
*hctx
, struct blk_mq_tags
*tags
,
174 struct blk_mq_ctx
*ctx
, unsigned int tag
)
176 if (tag
>= tags
->nr_reserved_tags
) {
177 const int real_tag
= tag
- tags
->nr_reserved_tags
;
179 BUG_ON(real_tag
>= tags
->nr_tags
);
180 sbitmap_queue_clear(&tags
->bitmap_tags
, real_tag
, ctx
->cpu
);
182 BUG_ON(tag
>= tags
->nr_reserved_tags
);
183 sbitmap_queue_clear(&tags
->breserved_tags
, tag
, ctx
->cpu
);
187 struct bt_iter_data
{
188 struct blk_mq_hw_ctx
*hctx
;
194 static bool bt_iter(struct sbitmap
*bitmap
, unsigned int bitnr
, void *data
)
196 struct bt_iter_data
*iter_data
= data
;
197 struct blk_mq_hw_ctx
*hctx
= iter_data
->hctx
;
198 struct blk_mq_tags
*tags
= hctx
->tags
;
199 bool reserved
= iter_data
->reserved
;
203 bitnr
+= tags
->nr_reserved_tags
;
204 rq
= tags
->rqs
[bitnr
];
206 if (rq
->q
== hctx
->queue
)
207 iter_data
->fn(hctx
, rq
, iter_data
->data
, reserved
);
211 static void bt_for_each(struct blk_mq_hw_ctx
*hctx
, struct sbitmap_queue
*bt
,
212 busy_iter_fn
*fn
, void *data
, bool reserved
)
214 struct bt_iter_data iter_data
= {
218 .reserved
= reserved
,
221 sbitmap_for_each_set(&bt
->sb
, bt_iter
, &iter_data
);
224 struct bt_tags_iter_data
{
225 struct blk_mq_tags
*tags
;
226 busy_tag_iter_fn
*fn
;
231 static bool bt_tags_iter(struct sbitmap
*bitmap
, unsigned int bitnr
, void *data
)
233 struct bt_tags_iter_data
*iter_data
= data
;
234 struct blk_mq_tags
*tags
= iter_data
->tags
;
235 bool reserved
= iter_data
->reserved
;
239 bitnr
+= tags
->nr_reserved_tags
;
240 rq
= tags
->rqs
[bitnr
];
242 iter_data
->fn(rq
, iter_data
->data
, reserved
);
246 static void bt_tags_for_each(struct blk_mq_tags
*tags
, struct sbitmap_queue
*bt
,
247 busy_tag_iter_fn
*fn
, void *data
, bool reserved
)
249 struct bt_tags_iter_data iter_data
= {
253 .reserved
= reserved
,
257 sbitmap_for_each_set(&bt
->sb
, bt_tags_iter
, &iter_data
);
260 static void blk_mq_all_tag_busy_iter(struct blk_mq_tags
*tags
,
261 busy_tag_iter_fn
*fn
, void *priv
)
263 if (tags
->nr_reserved_tags
)
264 bt_tags_for_each(tags
, &tags
->breserved_tags
, fn
, priv
, true);
265 bt_tags_for_each(tags
, &tags
->bitmap_tags
, fn
, priv
, false);
268 void blk_mq_tagset_busy_iter(struct blk_mq_tag_set
*tagset
,
269 busy_tag_iter_fn
*fn
, void *priv
)
273 for (i
= 0; i
< tagset
->nr_hw_queues
; i
++) {
274 if (tagset
->tags
&& tagset
->tags
[i
])
275 blk_mq_all_tag_busy_iter(tagset
->tags
[i
], fn
, priv
);
278 EXPORT_SYMBOL(blk_mq_tagset_busy_iter
);
280 int blk_mq_reinit_tagset(struct blk_mq_tag_set
*set
)
284 if (!set
->ops
->reinit_request
)
287 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
288 struct blk_mq_tags
*tags
= set
->tags
[i
];
290 for (j
= 0; j
< tags
->nr_tags
; j
++) {
291 if (!tags
->static_rqs
[j
])
294 ret
= set
->ops
->reinit_request(set
->driver_data
,
295 tags
->static_rqs
[j
]);
304 EXPORT_SYMBOL_GPL(blk_mq_reinit_tagset
);
306 void blk_mq_queue_tag_busy_iter(struct request_queue
*q
, busy_iter_fn
*fn
,
309 struct blk_mq_hw_ctx
*hctx
;
313 queue_for_each_hw_ctx(q
, hctx
, i
) {
314 struct blk_mq_tags
*tags
= hctx
->tags
;
317 * If not software queues are currently mapped to this
318 * hardware queue, there's nothing to check
320 if (!blk_mq_hw_queue_mapped(hctx
))
323 if (tags
->nr_reserved_tags
)
324 bt_for_each(hctx
, &tags
->breserved_tags
, fn
, priv
, true);
325 bt_for_each(hctx
, &tags
->bitmap_tags
, fn
, priv
, false);
330 static unsigned int bt_unused_tags(const struct sbitmap_queue
*bt
)
332 return bt
->sb
.depth
- sbitmap_weight(&bt
->sb
);
335 static int bt_alloc(struct sbitmap_queue
*bt
, unsigned int depth
,
336 bool round_robin
, int node
)
338 return sbitmap_queue_init_node(bt
, depth
, -1, round_robin
, GFP_KERNEL
,
342 static struct blk_mq_tags
*blk_mq_init_bitmap_tags(struct blk_mq_tags
*tags
,
343 int node
, int alloc_policy
)
345 unsigned int depth
= tags
->nr_tags
- tags
->nr_reserved_tags
;
346 bool round_robin
= alloc_policy
== BLK_TAG_ALLOC_RR
;
348 if (bt_alloc(&tags
->bitmap_tags
, depth
, round_robin
, node
))
350 if (bt_alloc(&tags
->breserved_tags
, tags
->nr_reserved_tags
, round_robin
,
352 goto free_bitmap_tags
;
356 sbitmap_queue_free(&tags
->bitmap_tags
);
362 struct blk_mq_tags
*blk_mq_init_tags(unsigned int total_tags
,
363 unsigned int reserved_tags
,
364 int node
, int alloc_policy
)
366 struct blk_mq_tags
*tags
;
368 if (total_tags
> BLK_MQ_TAG_MAX
) {
369 pr_err("blk-mq: tag depth too large\n");
373 tags
= kzalloc_node(sizeof(*tags
), GFP_KERNEL
, node
);
377 tags
->nr_tags
= total_tags
;
378 tags
->nr_reserved_tags
= reserved_tags
;
380 return blk_mq_init_bitmap_tags(tags
, node
, alloc_policy
);
383 void blk_mq_free_tags(struct blk_mq_tags
*tags
)
385 sbitmap_queue_free(&tags
->bitmap_tags
);
386 sbitmap_queue_free(&tags
->breserved_tags
);
390 int blk_mq_tag_update_depth(struct blk_mq_hw_ctx
*hctx
,
391 struct blk_mq_tags
**tagsptr
, unsigned int tdepth
,
394 struct blk_mq_tags
*tags
= *tagsptr
;
396 if (tdepth
<= tags
->nr_reserved_tags
)
399 tdepth
-= tags
->nr_reserved_tags
;
402 * If we are allowed to grow beyond the original size, allocate
403 * a new set of tags before freeing the old one.
405 if (tdepth
> tags
->nr_tags
) {
406 struct blk_mq_tag_set
*set
= hctx
->queue
->tag_set
;
407 struct blk_mq_tags
*new;
414 * We need some sort of upper limit, set it high enough that
415 * no valid use cases should require more.
417 if (tdepth
> 16 * BLKDEV_MAX_RQ
)
420 new = blk_mq_alloc_rq_map(set
, hctx
->queue_num
, tdepth
, 0);
423 ret
= blk_mq_alloc_rqs(set
, new, hctx
->queue_num
, tdepth
);
425 blk_mq_free_rq_map(new);
429 blk_mq_free_rqs(set
, *tagsptr
, hctx
->queue_num
);
430 blk_mq_free_rq_map(*tagsptr
);
434 * Don't need (or can't) update reserved tags here, they
435 * remain static and should never need resizing.
437 sbitmap_queue_resize(&tags
->bitmap_tags
, tdepth
);
444 * blk_mq_unique_tag() - return a tag that is unique queue-wide
445 * @rq: request for which to compute a unique tag
447 * The tag field in struct request is unique per hardware queue but not over
448 * all hardware queues. Hence this function that returns a tag with the
449 * hardware context index in the upper bits and the per hardware queue tag in
452 * Note: When called for a request that is queued on a non-multiqueue request
453 * queue, the hardware context index is set to zero.
455 u32
blk_mq_unique_tag(struct request
*rq
)
457 struct request_queue
*q
= rq
->q
;
458 struct blk_mq_hw_ctx
*hctx
;
462 hctx
= blk_mq_map_queue(q
, rq
->mq_ctx
->cpu
);
463 hwq
= hctx
->queue_num
;
466 return (hwq
<< BLK_MQ_UNIQUE_TAG_BITS
) |
467 (rq
->tag
& BLK_MQ_UNIQUE_TAG_MASK
);
469 EXPORT_SYMBOL(blk_mq_unique_tag
);
471 ssize_t
blk_mq_tag_sysfs_show(struct blk_mq_tags
*tags
, char *page
)
473 char *orig_page
= page
;
474 unsigned int free
, res
;
479 page
+= sprintf(page
, "nr_tags=%u, reserved_tags=%u, "
480 "bits_per_word=%u\n",
481 tags
->nr_tags
, tags
->nr_reserved_tags
,
482 1U << tags
->bitmap_tags
.sb
.shift
);
484 free
= bt_unused_tags(&tags
->bitmap_tags
);
485 res
= bt_unused_tags(&tags
->breserved_tags
);
487 page
+= sprintf(page
, "nr_free=%u, nr_reserved=%u\n", free
, res
);
488 page
+= sprintf(page
, "active_queues=%u\n", atomic_read(&tags
->active_queues
));
490 return page
- orig_page
;