2 * Fast and scalable bitmap tagging variant. Uses sparser bitmaps spread
3 * over multiple cachelines to avoid ping-pong between multiple submitters
4 * or submitter and completer. Uses rolling wakeups to avoid falling of
5 * the scaling cliff when we run out of tags and have to start putting
8 * Uses active queue tracking to support fairer distribution of tags
9 * between multiple submitters when a shared tag map is used.
11 * Copyright (C) 2013-2014 Jens Axboe
13 #include <linux/kernel.h>
14 #include <linux/module.h>
15 #include <linux/random.h>
17 #include <linux/blk-mq.h>
20 #include "blk-mq-tag.h"
22 static bool bt_has_free_tags(struct blk_mq_bitmap_tags
*bt
)
26 for (i
= 0; i
< bt
->map_nr
; i
++) {
27 struct blk_align_bitmap
*bm
= &bt
->map
[i
];
30 ret
= find_first_zero_bit(&bm
->word
, bm
->depth
);
38 bool blk_mq_has_free_tags(struct blk_mq_tags
*tags
)
43 return bt_has_free_tags(&tags
->bitmap_tags
);
46 static inline int bt_index_inc(int index
)
48 return (index
+ 1) & (BT_WAIT_QUEUES
- 1);
51 static inline void bt_index_atomic_inc(atomic_t
*index
)
53 int old
= atomic_read(index
);
54 int new = bt_index_inc(old
);
55 atomic_cmpxchg(index
, old
, new);
59 * If a previously inactive queue goes active, bump the active user count.
61 bool __blk_mq_tag_busy(struct blk_mq_hw_ctx
*hctx
)
63 if (!test_bit(BLK_MQ_S_TAG_ACTIVE
, &hctx
->state
) &&
64 !test_and_set_bit(BLK_MQ_S_TAG_ACTIVE
, &hctx
->state
))
65 atomic_inc(&hctx
->tags
->active_queues
);
71 * Wakeup all potentially sleeping on tags
73 void blk_mq_tag_wakeup_all(struct blk_mq_tags
*tags
, bool include_reserve
)
75 struct blk_mq_bitmap_tags
*bt
;
78 bt
= &tags
->bitmap_tags
;
79 wake_index
= atomic_read(&bt
->wake_index
);
80 for (i
= 0; i
< BT_WAIT_QUEUES
; i
++) {
81 struct bt_wait_state
*bs
= &bt
->bs
[wake_index
];
83 if (waitqueue_active(&bs
->wait
))
86 wake_index
= bt_index_inc(wake_index
);
89 if (include_reserve
) {
90 bt
= &tags
->breserved_tags
;
91 if (waitqueue_active(&bt
->bs
[0].wait
))
92 wake_up(&bt
->bs
[0].wait
);
97 * If a previously busy queue goes inactive, potential waiters could now
98 * be allowed to queue. Wake them up and check.
100 void __blk_mq_tag_idle(struct blk_mq_hw_ctx
*hctx
)
102 struct blk_mq_tags
*tags
= hctx
->tags
;
104 if (!test_and_clear_bit(BLK_MQ_S_TAG_ACTIVE
, &hctx
->state
))
107 atomic_dec(&tags
->active_queues
);
109 blk_mq_tag_wakeup_all(tags
, false);
113 * For shared tag users, we track the number of currently active users
114 * and attempt to provide a fair share of the tag depth for each of them.
116 static inline bool hctx_may_queue(struct blk_mq_hw_ctx
*hctx
,
117 struct blk_mq_bitmap_tags
*bt
)
119 unsigned int depth
, users
;
121 if (!hctx
|| !(hctx
->flags
& BLK_MQ_F_TAG_SHARED
))
123 if (!test_bit(BLK_MQ_S_TAG_ACTIVE
, &hctx
->state
))
127 * Don't try dividing an ant
132 users
= atomic_read(&hctx
->tags
->active_queues
);
137 * Allow at least some tags
139 depth
= max((bt
->depth
+ users
- 1) / users
, 4U);
140 return atomic_read(&hctx
->nr_active
) < depth
;
143 static int __bt_get_word(struct blk_align_bitmap
*bm
, unsigned int last_tag
,
146 int tag
, org_last_tag
= last_tag
;
149 tag
= find_next_zero_bit(&bm
->word
, bm
->depth
, last_tag
);
150 if (unlikely(tag
>= bm
->depth
)) {
152 * We started with an offset, and we didn't reset the
153 * offset to 0 in a failure case, so start from 0 to
156 if (org_last_tag
&& last_tag
&& !nowrap
) {
157 last_tag
= org_last_tag
= 0;
163 if (!test_and_set_bit(tag
, &bm
->word
))
167 if (last_tag
>= bm
->depth
- 1)
174 #define BT_ALLOC_RR(tags) (tags->alloc_policy == BLK_TAG_ALLOC_RR)
177 * Straight forward bitmap tag implementation, where each bit is a tag
178 * (cleared == free, and set == busy). The small twist is using per-cpu
179 * last_tag caches, which blk-mq stores in the blk_mq_ctx software queue
180 * contexts. This enables us to drastically limit the space searched,
181 * without dirtying an extra shared cacheline like we would if we stored
182 * the cache value inside the shared blk_mq_bitmap_tags structure. On top
183 * of that, each word of tags is in a separate cacheline. This means that
184 * multiple users will tend to stick to different cachelines, at least
185 * until the map is exhausted.
187 static int __bt_get(struct blk_mq_hw_ctx
*hctx
, struct blk_mq_bitmap_tags
*bt
,
188 unsigned int *tag_cache
, struct blk_mq_tags
*tags
)
190 unsigned int last_tag
, org_last_tag
;
193 if (!hctx_may_queue(hctx
, bt
))
196 last_tag
= org_last_tag
= *tag_cache
;
197 index
= TAG_TO_INDEX(bt
, last_tag
);
199 for (i
= 0; i
< bt
->map_nr
; i
++) {
200 tag
= __bt_get_word(&bt
->map
[index
], TAG_TO_BIT(bt
, last_tag
),
203 tag
+= (index
<< bt
->bits_per_word
);
208 * Jump to next index, and reset the last tag to be the
209 * first tag of that index
212 last_tag
= (index
<< bt
->bits_per_word
);
214 if (index
>= bt
->map_nr
) {
224 * Only update the cache from the allocation path, if we ended
225 * up using the specific cached tag.
228 if (tag
== org_last_tag
|| unlikely(BT_ALLOC_RR(tags
))) {
230 if (last_tag
>= bt
->depth
- 1)
233 *tag_cache
= last_tag
;
239 static struct bt_wait_state
*bt_wait_ptr(struct blk_mq_bitmap_tags
*bt
,
240 struct blk_mq_hw_ctx
*hctx
)
242 struct bt_wait_state
*bs
;
248 wait_index
= atomic_read(&hctx
->wait_index
);
249 bs
= &bt
->bs
[wait_index
];
250 bt_index_atomic_inc(&hctx
->wait_index
);
254 static int bt_get(struct blk_mq_alloc_data
*data
,
255 struct blk_mq_bitmap_tags
*bt
,
256 struct blk_mq_hw_ctx
*hctx
,
257 unsigned int *last_tag
, struct blk_mq_tags
*tags
)
259 struct bt_wait_state
*bs
;
263 tag
= __bt_get(hctx
, bt
, last_tag
, tags
);
267 if (!(data
->gfp
& __GFP_WAIT
))
270 bs
= bt_wait_ptr(bt
, hctx
);
272 prepare_to_wait(&bs
->wait
, &wait
, TASK_UNINTERRUPTIBLE
);
274 tag
= __bt_get(hctx
, bt
, last_tag
, tags
);
279 * We're out of tags on this hardware queue, kick any
280 * pending IO submits before going to sleep waiting for
281 * some to complete. Note that hctx can be NULL here for
282 * reserved tag allocation.
285 blk_mq_run_hw_queue(hctx
, false);
288 * Retry tag allocation after running the hardware queue,
289 * as running the queue may also have found completions.
291 tag
= __bt_get(hctx
, bt
, last_tag
, tags
);
295 blk_mq_put_ctx(data
->ctx
);
299 data
->ctx
= blk_mq_get_ctx(data
->q
);
300 data
->hctx
= data
->q
->mq_ops
->map_queue(data
->q
,
302 if (data
->reserved
) {
303 bt
= &data
->hctx
->tags
->breserved_tags
;
305 last_tag
= &data
->ctx
->last_tag
;
307 bt
= &hctx
->tags
->bitmap_tags
;
309 finish_wait(&bs
->wait
, &wait
);
310 bs
= bt_wait_ptr(bt
, hctx
);
313 finish_wait(&bs
->wait
, &wait
);
317 static unsigned int __blk_mq_get_tag(struct blk_mq_alloc_data
*data
)
321 tag
= bt_get(data
, &data
->hctx
->tags
->bitmap_tags
, data
->hctx
,
322 &data
->ctx
->last_tag
, data
->hctx
->tags
);
324 return tag
+ data
->hctx
->tags
->nr_reserved_tags
;
326 return BLK_MQ_TAG_FAIL
;
329 static unsigned int __blk_mq_get_reserved_tag(struct blk_mq_alloc_data
*data
)
333 if (unlikely(!data
->hctx
->tags
->nr_reserved_tags
)) {
335 return BLK_MQ_TAG_FAIL
;
338 tag
= bt_get(data
, &data
->hctx
->tags
->breserved_tags
, NULL
, &zero
,
341 return BLK_MQ_TAG_FAIL
;
346 unsigned int blk_mq_get_tag(struct blk_mq_alloc_data
*data
)
349 return __blk_mq_get_tag(data
);
351 return __blk_mq_get_reserved_tag(data
);
354 static struct bt_wait_state
*bt_wake_ptr(struct blk_mq_bitmap_tags
*bt
)
358 wake_index
= atomic_read(&bt
->wake_index
);
359 for (i
= 0; i
< BT_WAIT_QUEUES
; i
++) {
360 struct bt_wait_state
*bs
= &bt
->bs
[wake_index
];
362 if (waitqueue_active(&bs
->wait
)) {
363 int o
= atomic_read(&bt
->wake_index
);
365 atomic_cmpxchg(&bt
->wake_index
, o
, wake_index
);
370 wake_index
= bt_index_inc(wake_index
);
376 static void bt_clear_tag(struct blk_mq_bitmap_tags
*bt
, unsigned int tag
)
378 const int index
= TAG_TO_INDEX(bt
, tag
);
379 struct bt_wait_state
*bs
;
382 clear_bit(TAG_TO_BIT(bt
, tag
), &bt
->map
[index
].word
);
384 /* Ensure that the wait list checks occur after clear_bit(). */
387 bs
= bt_wake_ptr(bt
);
391 wait_cnt
= atomic_dec_return(&bs
->wait_cnt
);
392 if (unlikely(wait_cnt
< 0))
393 wait_cnt
= atomic_inc_return(&bs
->wait_cnt
);
395 atomic_add(bt
->wake_cnt
, &bs
->wait_cnt
);
396 bt_index_atomic_inc(&bt
->wake_index
);
401 void blk_mq_put_tag(struct blk_mq_hw_ctx
*hctx
, unsigned int tag
,
402 unsigned int *last_tag
)
404 struct blk_mq_tags
*tags
= hctx
->tags
;
406 if (tag
>= tags
->nr_reserved_tags
) {
407 const int real_tag
= tag
- tags
->nr_reserved_tags
;
409 BUG_ON(real_tag
>= tags
->nr_tags
);
410 bt_clear_tag(&tags
->bitmap_tags
, real_tag
);
411 if (likely(tags
->alloc_policy
== BLK_TAG_ALLOC_FIFO
))
412 *last_tag
= real_tag
;
414 BUG_ON(tag
>= tags
->nr_reserved_tags
);
415 bt_clear_tag(&tags
->breserved_tags
, tag
);
419 static void bt_for_each(struct blk_mq_hw_ctx
*hctx
,
420 struct blk_mq_bitmap_tags
*bt
, unsigned int off
,
421 busy_iter_fn
*fn
, void *data
, bool reserved
)
426 for (i
= 0; i
< bt
->map_nr
; i
++) {
427 struct blk_align_bitmap
*bm
= &bt
->map
[i
];
429 for (bit
= find_first_bit(&bm
->word
, bm
->depth
);
431 bit
= find_next_bit(&bm
->word
, bm
->depth
, bit
+ 1)) {
432 rq
= hctx
->tags
->rqs
[off
+ bit
];
433 if (rq
->q
== hctx
->queue
)
434 fn(hctx
, rq
, data
, reserved
);
437 off
+= (1 << bt
->bits_per_word
);
441 static void bt_tags_for_each(struct blk_mq_tags
*tags
,
442 struct blk_mq_bitmap_tags
*bt
, unsigned int off
,
443 busy_tag_iter_fn
*fn
, void *data
, bool reserved
)
450 for (i
= 0; i
< bt
->map_nr
; i
++) {
451 struct blk_align_bitmap
*bm
= &bt
->map
[i
];
453 for (bit
= find_first_bit(&bm
->word
, bm
->depth
);
455 bit
= find_next_bit(&bm
->word
, bm
->depth
, bit
+ 1)) {
456 rq
= tags
->rqs
[off
+ bit
];
457 fn(rq
, data
, reserved
);
460 off
+= (1 << bt
->bits_per_word
);
464 void blk_mq_all_tag_busy_iter(struct blk_mq_tags
*tags
, busy_tag_iter_fn
*fn
,
467 if (tags
->nr_reserved_tags
)
468 bt_tags_for_each(tags
, &tags
->breserved_tags
, 0, fn
, priv
, true);
469 bt_tags_for_each(tags
, &tags
->bitmap_tags
, tags
->nr_reserved_tags
, fn
, priv
,
472 EXPORT_SYMBOL(blk_mq_all_tag_busy_iter
);
474 void blk_mq_queue_tag_busy_iter(struct request_queue
*q
, busy_iter_fn
*fn
,
477 struct blk_mq_hw_ctx
*hctx
;
481 queue_for_each_hw_ctx(q
, hctx
, i
) {
482 struct blk_mq_tags
*tags
= hctx
->tags
;
485 * If not software queues are currently mapped to this
486 * hardware queue, there's nothing to check
488 if (!blk_mq_hw_queue_mapped(hctx
))
491 if (tags
->nr_reserved_tags
)
492 bt_for_each(hctx
, &tags
->breserved_tags
, 0, fn
, priv
, true);
493 bt_for_each(hctx
, &tags
->bitmap_tags
, tags
->nr_reserved_tags
, fn
, priv
,
499 static unsigned int bt_unused_tags(struct blk_mq_bitmap_tags
*bt
)
501 unsigned int i
, used
;
503 for (i
= 0, used
= 0; i
< bt
->map_nr
; i
++) {
504 struct blk_align_bitmap
*bm
= &bt
->map
[i
];
506 used
+= bitmap_weight(&bm
->word
, bm
->depth
);
509 return bt
->depth
- used
;
512 static void bt_update_count(struct blk_mq_bitmap_tags
*bt
,
515 unsigned int tags_per_word
= 1U << bt
->bits_per_word
;
516 unsigned int map_depth
= depth
;
521 for (i
= 0; i
< bt
->map_nr
; i
++) {
522 bt
->map
[i
].depth
= min(map_depth
, tags_per_word
);
523 map_depth
-= bt
->map
[i
].depth
;
527 bt
->wake_cnt
= BT_WAIT_BATCH
;
528 if (bt
->wake_cnt
> depth
/ BT_WAIT_QUEUES
)
529 bt
->wake_cnt
= max(1U, depth
/ BT_WAIT_QUEUES
);
534 static int bt_alloc(struct blk_mq_bitmap_tags
*bt
, unsigned int depth
,
535 int node
, bool reserved
)
539 bt
->bits_per_word
= ilog2(BITS_PER_LONG
);
542 * Depth can be zero for reserved tags, that's not a failure
546 unsigned int nr
, tags_per_word
;
548 tags_per_word
= (1 << bt
->bits_per_word
);
551 * If the tag space is small, shrink the number of tags
552 * per word so we spread over a few cachelines, at least.
553 * If less than 4 tags, just forget about it, it's not
554 * going to work optimally anyway.
557 while (tags_per_word
* 4 > depth
) {
559 tags_per_word
= (1 << bt
->bits_per_word
);
563 nr
= ALIGN(depth
, tags_per_word
) / tags_per_word
;
564 bt
->map
= kzalloc_node(nr
* sizeof(struct blk_align_bitmap
),
572 bt
->bs
= kzalloc(BT_WAIT_QUEUES
* sizeof(*bt
->bs
), GFP_KERNEL
);
579 bt_update_count(bt
, depth
);
581 for (i
= 0; i
< BT_WAIT_QUEUES
; i
++) {
582 init_waitqueue_head(&bt
->bs
[i
].wait
);
583 atomic_set(&bt
->bs
[i
].wait_cnt
, bt
->wake_cnt
);
589 static void bt_free(struct blk_mq_bitmap_tags
*bt
)
595 static struct blk_mq_tags
*blk_mq_init_bitmap_tags(struct blk_mq_tags
*tags
,
596 int node
, int alloc_policy
)
598 unsigned int depth
= tags
->nr_tags
- tags
->nr_reserved_tags
;
600 tags
->alloc_policy
= alloc_policy
;
602 if (bt_alloc(&tags
->bitmap_tags
, depth
, node
, false))
604 if (bt_alloc(&tags
->breserved_tags
, tags
->nr_reserved_tags
, node
, true))
609 bt_free(&tags
->bitmap_tags
);
614 struct blk_mq_tags
*blk_mq_init_tags(unsigned int total_tags
,
615 unsigned int reserved_tags
,
616 int node
, int alloc_policy
)
618 struct blk_mq_tags
*tags
;
620 if (total_tags
> BLK_MQ_TAG_MAX
) {
621 pr_err("blk-mq: tag depth too large\n");
625 tags
= kzalloc_node(sizeof(*tags
), GFP_KERNEL
, node
);
629 if (!zalloc_cpumask_var(&tags
->cpumask
, GFP_KERNEL
)) {
634 tags
->nr_tags
= total_tags
;
635 tags
->nr_reserved_tags
= reserved_tags
;
637 return blk_mq_init_bitmap_tags(tags
, node
, alloc_policy
);
640 void blk_mq_free_tags(struct blk_mq_tags
*tags
)
642 bt_free(&tags
->bitmap_tags
);
643 bt_free(&tags
->breserved_tags
);
647 void blk_mq_tag_init_last_tag(struct blk_mq_tags
*tags
, unsigned int *tag
)
649 unsigned int depth
= tags
->nr_tags
- tags
->nr_reserved_tags
;
651 *tag
= prandom_u32() % depth
;
654 int blk_mq_tag_update_depth(struct blk_mq_tags
*tags
, unsigned int tdepth
)
656 tdepth
-= tags
->nr_reserved_tags
;
657 if (tdepth
> tags
->nr_tags
)
661 * Don't need (or can't) update reserved tags here, they remain
662 * static and should never need resizing.
664 bt_update_count(&tags
->bitmap_tags
, tdepth
);
665 blk_mq_tag_wakeup_all(tags
, false);
670 * blk_mq_unique_tag() - return a tag that is unique queue-wide
671 * @rq: request for which to compute a unique tag
673 * The tag field in struct request is unique per hardware queue but not over
674 * all hardware queues. Hence this function that returns a tag with the
675 * hardware context index in the upper bits and the per hardware queue tag in
678 * Note: When called for a request that is queued on a non-multiqueue request
679 * queue, the hardware context index is set to zero.
681 u32
blk_mq_unique_tag(struct request
*rq
)
683 struct request_queue
*q
= rq
->q
;
684 struct blk_mq_hw_ctx
*hctx
;
688 hctx
= q
->mq_ops
->map_queue(q
, rq
->mq_ctx
->cpu
);
689 hwq
= hctx
->queue_num
;
692 return (hwq
<< BLK_MQ_UNIQUE_TAG_BITS
) |
693 (rq
->tag
& BLK_MQ_UNIQUE_TAG_MASK
);
695 EXPORT_SYMBOL(blk_mq_unique_tag
);
697 ssize_t
blk_mq_tag_sysfs_show(struct blk_mq_tags
*tags
, char *page
)
699 char *orig_page
= page
;
700 unsigned int free
, res
;
705 page
+= sprintf(page
, "nr_tags=%u, reserved_tags=%u, "
706 "bits_per_word=%u\n",
707 tags
->nr_tags
, tags
->nr_reserved_tags
,
708 tags
->bitmap_tags
.bits_per_word
);
710 free
= bt_unused_tags(&tags
->bitmap_tags
);
711 res
= bt_unused_tags(&tags
->breserved_tags
);
713 page
+= sprintf(page
, "nr_free=%u, nr_reserved=%u\n", free
, res
);
714 page
+= sprintf(page
, "active_queues=%u\n", atomic_read(&tags
->active_queues
));
716 return page
- orig_page
;