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
;
79 * Make sure all changes prior to this are visible from other CPUs.
82 bt
= &tags
->bitmap_tags
;
83 wake_index
= atomic_read(&bt
->wake_index
);
84 for (i
= 0; i
< BT_WAIT_QUEUES
; i
++) {
85 struct bt_wait_state
*bs
= &bt
->bs
[wake_index
];
87 if (waitqueue_active(&bs
->wait
))
90 wake_index
= bt_index_inc(wake_index
);
93 if (include_reserve
) {
94 bt
= &tags
->breserved_tags
;
95 if (waitqueue_active(&bt
->bs
[0].wait
))
96 wake_up(&bt
->bs
[0].wait
);
101 * If a previously busy queue goes inactive, potential waiters could now
102 * be allowed to queue. Wake them up and check.
104 void __blk_mq_tag_idle(struct blk_mq_hw_ctx
*hctx
)
106 struct blk_mq_tags
*tags
= hctx
->tags
;
108 if (!test_and_clear_bit(BLK_MQ_S_TAG_ACTIVE
, &hctx
->state
))
111 atomic_dec(&tags
->active_queues
);
113 blk_mq_tag_wakeup_all(tags
, false);
117 * For shared tag users, we track the number of currently active users
118 * and attempt to provide a fair share of the tag depth for each of them.
120 static inline bool hctx_may_queue(struct blk_mq_hw_ctx
*hctx
,
121 struct blk_mq_bitmap_tags
*bt
)
123 unsigned int depth
, users
;
125 if (!hctx
|| !(hctx
->flags
& BLK_MQ_F_TAG_SHARED
))
127 if (!test_bit(BLK_MQ_S_TAG_ACTIVE
, &hctx
->state
))
131 * Don't try dividing an ant
136 users
= atomic_read(&hctx
->tags
->active_queues
);
141 * Allow at least some tags
143 depth
= max((bt
->depth
+ users
- 1) / users
, 4U);
144 return atomic_read(&hctx
->nr_active
) < depth
;
147 static int __bt_get_word(struct blk_align_bitmap
*bm
, unsigned int last_tag
,
150 int tag
, org_last_tag
= last_tag
;
153 tag
= find_next_zero_bit(&bm
->word
, bm
->depth
, last_tag
);
154 if (unlikely(tag
>= bm
->depth
)) {
156 * We started with an offset, and we didn't reset the
157 * offset to 0 in a failure case, so start from 0 to
160 if (org_last_tag
&& last_tag
&& !nowrap
) {
161 last_tag
= org_last_tag
= 0;
167 if (!test_and_set_bit(tag
, &bm
->word
))
171 if (last_tag
>= bm
->depth
- 1)
178 #define BT_ALLOC_RR(tags) (tags->alloc_policy == BLK_TAG_ALLOC_RR)
181 * Straight forward bitmap tag implementation, where each bit is a tag
182 * (cleared == free, and set == busy). The small twist is using per-cpu
183 * last_tag caches, which blk-mq stores in the blk_mq_ctx software queue
184 * contexts. This enables us to drastically limit the space searched,
185 * without dirtying an extra shared cacheline like we would if we stored
186 * the cache value inside the shared blk_mq_bitmap_tags structure. On top
187 * of that, each word of tags is in a separate cacheline. This means that
188 * multiple users will tend to stick to different cachelines, at least
189 * until the map is exhausted.
191 static int __bt_get(struct blk_mq_hw_ctx
*hctx
, struct blk_mq_bitmap_tags
*bt
,
192 unsigned int *tag_cache
, struct blk_mq_tags
*tags
)
194 unsigned int last_tag
, org_last_tag
;
197 if (!hctx_may_queue(hctx
, bt
))
200 last_tag
= org_last_tag
= *tag_cache
;
201 index
= TAG_TO_INDEX(bt
, last_tag
);
203 for (i
= 0; i
< bt
->map_nr
; i
++) {
204 tag
= __bt_get_word(&bt
->map
[index
], TAG_TO_BIT(bt
, last_tag
),
207 tag
+= (index
<< bt
->bits_per_word
);
212 * Jump to next index, and reset the last tag to be the
213 * first tag of that index
216 last_tag
= (index
<< bt
->bits_per_word
);
218 if (index
>= bt
->map_nr
) {
228 * Only update the cache from the allocation path, if we ended
229 * up using the specific cached tag.
232 if (tag
== org_last_tag
|| unlikely(BT_ALLOC_RR(tags
))) {
234 if (last_tag
>= bt
->depth
- 1)
237 *tag_cache
= last_tag
;
243 static struct bt_wait_state
*bt_wait_ptr(struct blk_mq_bitmap_tags
*bt
,
244 struct blk_mq_hw_ctx
*hctx
)
246 struct bt_wait_state
*bs
;
252 wait_index
= atomic_read(&hctx
->wait_index
);
253 bs
= &bt
->bs
[wait_index
];
254 bt_index_atomic_inc(&hctx
->wait_index
);
258 static int bt_get(struct blk_mq_alloc_data
*data
,
259 struct blk_mq_bitmap_tags
*bt
,
260 struct blk_mq_hw_ctx
*hctx
,
261 unsigned int *last_tag
, struct blk_mq_tags
*tags
)
263 struct bt_wait_state
*bs
;
267 tag
= __bt_get(hctx
, bt
, last_tag
, tags
);
271 if (!gfpflags_allow_blocking(data
->gfp
))
274 bs
= bt_wait_ptr(bt
, hctx
);
276 prepare_to_wait(&bs
->wait
, &wait
, TASK_UNINTERRUPTIBLE
);
278 tag
= __bt_get(hctx
, bt
, last_tag
, tags
);
283 * We're out of tags on this hardware queue, kick any
284 * pending IO submits before going to sleep waiting for
285 * some to complete. Note that hctx can be NULL here for
286 * reserved tag allocation.
289 blk_mq_run_hw_queue(hctx
, false);
292 * Retry tag allocation after running the hardware queue,
293 * as running the queue may also have found completions.
295 tag
= __bt_get(hctx
, bt
, last_tag
, tags
);
299 blk_mq_put_ctx(data
->ctx
);
303 data
->ctx
= blk_mq_get_ctx(data
->q
);
304 data
->hctx
= data
->q
->mq_ops
->map_queue(data
->q
,
306 if (data
->reserved
) {
307 bt
= &data
->hctx
->tags
->breserved_tags
;
309 last_tag
= &data
->ctx
->last_tag
;
311 bt
= &hctx
->tags
->bitmap_tags
;
313 finish_wait(&bs
->wait
, &wait
);
314 bs
= bt_wait_ptr(bt
, hctx
);
317 finish_wait(&bs
->wait
, &wait
);
321 static unsigned int __blk_mq_get_tag(struct blk_mq_alloc_data
*data
)
325 tag
= bt_get(data
, &data
->hctx
->tags
->bitmap_tags
, data
->hctx
,
326 &data
->ctx
->last_tag
, data
->hctx
->tags
);
328 return tag
+ data
->hctx
->tags
->nr_reserved_tags
;
330 return BLK_MQ_TAG_FAIL
;
333 static unsigned int __blk_mq_get_reserved_tag(struct blk_mq_alloc_data
*data
)
337 if (unlikely(!data
->hctx
->tags
->nr_reserved_tags
)) {
339 return BLK_MQ_TAG_FAIL
;
342 tag
= bt_get(data
, &data
->hctx
->tags
->breserved_tags
, NULL
, &zero
,
345 return BLK_MQ_TAG_FAIL
;
350 unsigned int blk_mq_get_tag(struct blk_mq_alloc_data
*data
)
353 return __blk_mq_get_tag(data
);
355 return __blk_mq_get_reserved_tag(data
);
358 static struct bt_wait_state
*bt_wake_ptr(struct blk_mq_bitmap_tags
*bt
)
362 wake_index
= atomic_read(&bt
->wake_index
);
363 for (i
= 0; i
< BT_WAIT_QUEUES
; i
++) {
364 struct bt_wait_state
*bs
= &bt
->bs
[wake_index
];
366 if (waitqueue_active(&bs
->wait
)) {
367 int o
= atomic_read(&bt
->wake_index
);
369 atomic_cmpxchg(&bt
->wake_index
, o
, wake_index
);
374 wake_index
= bt_index_inc(wake_index
);
380 static void bt_clear_tag(struct blk_mq_bitmap_tags
*bt
, unsigned int tag
)
382 const int index
= TAG_TO_INDEX(bt
, tag
);
383 struct bt_wait_state
*bs
;
386 clear_bit(TAG_TO_BIT(bt
, tag
), &bt
->map
[index
].word
);
388 /* Ensure that the wait list checks occur after clear_bit(). */
391 bs
= bt_wake_ptr(bt
);
395 wait_cnt
= atomic_dec_return(&bs
->wait_cnt
);
396 if (unlikely(wait_cnt
< 0))
397 wait_cnt
= atomic_inc_return(&bs
->wait_cnt
);
399 atomic_add(bt
->wake_cnt
, &bs
->wait_cnt
);
400 bt_index_atomic_inc(&bt
->wake_index
);
405 void blk_mq_put_tag(struct blk_mq_hw_ctx
*hctx
, unsigned int tag
,
406 unsigned int *last_tag
)
408 struct blk_mq_tags
*tags
= hctx
->tags
;
410 if (tag
>= tags
->nr_reserved_tags
) {
411 const int real_tag
= tag
- tags
->nr_reserved_tags
;
413 BUG_ON(real_tag
>= tags
->nr_tags
);
414 bt_clear_tag(&tags
->bitmap_tags
, real_tag
);
415 if (likely(tags
->alloc_policy
== BLK_TAG_ALLOC_FIFO
))
416 *last_tag
= real_tag
;
418 BUG_ON(tag
>= tags
->nr_reserved_tags
);
419 bt_clear_tag(&tags
->breserved_tags
, tag
);
423 static void bt_for_each(struct blk_mq_hw_ctx
*hctx
,
424 struct blk_mq_bitmap_tags
*bt
, unsigned int off
,
425 busy_iter_fn
*fn
, void *data
, bool reserved
)
430 for (i
= 0; i
< bt
->map_nr
; i
++) {
431 struct blk_align_bitmap
*bm
= &bt
->map
[i
];
433 for (bit
= find_first_bit(&bm
->word
, bm
->depth
);
435 bit
= find_next_bit(&bm
->word
, bm
->depth
, bit
+ 1)) {
436 rq
= hctx
->tags
->rqs
[off
+ bit
];
437 if (rq
->q
== hctx
->queue
)
438 fn(hctx
, rq
, data
, reserved
);
441 off
+= (1 << bt
->bits_per_word
);
445 static void bt_tags_for_each(struct blk_mq_tags
*tags
,
446 struct blk_mq_bitmap_tags
*bt
, unsigned int off
,
447 busy_tag_iter_fn
*fn
, void *data
, bool reserved
)
454 for (i
= 0; i
< bt
->map_nr
; i
++) {
455 struct blk_align_bitmap
*bm
= &bt
->map
[i
];
457 for (bit
= find_first_bit(&bm
->word
, bm
->depth
);
459 bit
= find_next_bit(&bm
->word
, bm
->depth
, bit
+ 1)) {
460 rq
= tags
->rqs
[off
+ bit
];
461 fn(rq
, data
, reserved
);
464 off
+= (1 << bt
->bits_per_word
);
468 void blk_mq_all_tag_busy_iter(struct blk_mq_tags
*tags
, busy_tag_iter_fn
*fn
,
471 if (tags
->nr_reserved_tags
)
472 bt_tags_for_each(tags
, &tags
->breserved_tags
, 0, fn
, priv
, true);
473 bt_tags_for_each(tags
, &tags
->bitmap_tags
, tags
->nr_reserved_tags
, fn
, priv
,
476 EXPORT_SYMBOL(blk_mq_all_tag_busy_iter
);
478 void blk_mq_queue_tag_busy_iter(struct request_queue
*q
, busy_iter_fn
*fn
,
481 struct blk_mq_hw_ctx
*hctx
;
485 * Avoid potential races with things like queue removal.
487 if (!percpu_ref_tryget(&q
->q_usage_counter
))
490 queue_for_each_hw_ctx(q
, hctx
, i
) {
491 struct blk_mq_tags
*tags
= hctx
->tags
;
494 * If not software queues are currently mapped to this
495 * hardware queue, there's nothing to check
497 if (!blk_mq_hw_queue_mapped(hctx
))
500 if (tags
->nr_reserved_tags
)
501 bt_for_each(hctx
, &tags
->breserved_tags
, 0, fn
, priv
, true);
502 bt_for_each(hctx
, &tags
->bitmap_tags
, tags
->nr_reserved_tags
, fn
, priv
,
508 static unsigned int bt_unused_tags(struct blk_mq_bitmap_tags
*bt
)
510 unsigned int i
, used
;
512 for (i
= 0, used
= 0; i
< bt
->map_nr
; i
++) {
513 struct blk_align_bitmap
*bm
= &bt
->map
[i
];
515 used
+= bitmap_weight(&bm
->word
, bm
->depth
);
518 return bt
->depth
- used
;
521 static void bt_update_count(struct blk_mq_bitmap_tags
*bt
,
524 unsigned int tags_per_word
= 1U << bt
->bits_per_word
;
525 unsigned int map_depth
= depth
;
530 for (i
= 0; i
< bt
->map_nr
; i
++) {
531 bt
->map
[i
].depth
= min(map_depth
, tags_per_word
);
532 map_depth
-= bt
->map
[i
].depth
;
536 bt
->wake_cnt
= BT_WAIT_BATCH
;
537 if (bt
->wake_cnt
> depth
/ BT_WAIT_QUEUES
)
538 bt
->wake_cnt
= max(1U, depth
/ BT_WAIT_QUEUES
);
543 static int bt_alloc(struct blk_mq_bitmap_tags
*bt
, unsigned int depth
,
544 int node
, bool reserved
)
548 bt
->bits_per_word
= ilog2(BITS_PER_LONG
);
551 * Depth can be zero for reserved tags, that's not a failure
555 unsigned int nr
, tags_per_word
;
557 tags_per_word
= (1 << bt
->bits_per_word
);
560 * If the tag space is small, shrink the number of tags
561 * per word so we spread over a few cachelines, at least.
562 * If less than 4 tags, just forget about it, it's not
563 * going to work optimally anyway.
566 while (tags_per_word
* 4 > depth
) {
568 tags_per_word
= (1 << bt
->bits_per_word
);
572 nr
= ALIGN(depth
, tags_per_word
) / tags_per_word
;
573 bt
->map
= kzalloc_node(nr
* sizeof(struct blk_align_bitmap
),
581 bt
->bs
= kzalloc(BT_WAIT_QUEUES
* sizeof(*bt
->bs
), GFP_KERNEL
);
588 bt_update_count(bt
, depth
);
590 for (i
= 0; i
< BT_WAIT_QUEUES
; i
++) {
591 init_waitqueue_head(&bt
->bs
[i
].wait
);
592 atomic_set(&bt
->bs
[i
].wait_cnt
, bt
->wake_cnt
);
598 static void bt_free(struct blk_mq_bitmap_tags
*bt
)
604 static struct blk_mq_tags
*blk_mq_init_bitmap_tags(struct blk_mq_tags
*tags
,
605 int node
, int alloc_policy
)
607 unsigned int depth
= tags
->nr_tags
- tags
->nr_reserved_tags
;
609 tags
->alloc_policy
= alloc_policy
;
611 if (bt_alloc(&tags
->bitmap_tags
, depth
, node
, false))
613 if (bt_alloc(&tags
->breserved_tags
, tags
->nr_reserved_tags
, node
, true))
618 bt_free(&tags
->bitmap_tags
);
623 struct blk_mq_tags
*blk_mq_init_tags(unsigned int total_tags
,
624 unsigned int reserved_tags
,
625 int node
, int alloc_policy
)
627 struct blk_mq_tags
*tags
;
629 if (total_tags
> BLK_MQ_TAG_MAX
) {
630 pr_err("blk-mq: tag depth too large\n");
634 tags
= kzalloc_node(sizeof(*tags
), GFP_KERNEL
, node
);
638 if (!zalloc_cpumask_var(&tags
->cpumask
, GFP_KERNEL
)) {
643 tags
->nr_tags
= total_tags
;
644 tags
->nr_reserved_tags
= reserved_tags
;
646 return blk_mq_init_bitmap_tags(tags
, node
, alloc_policy
);
649 void blk_mq_free_tags(struct blk_mq_tags
*tags
)
651 bt_free(&tags
->bitmap_tags
);
652 bt_free(&tags
->breserved_tags
);
653 free_cpumask_var(tags
->cpumask
);
657 void blk_mq_tag_init_last_tag(struct blk_mq_tags
*tags
, unsigned int *tag
)
659 unsigned int depth
= tags
->nr_tags
- tags
->nr_reserved_tags
;
661 *tag
= prandom_u32() % depth
;
664 int blk_mq_tag_update_depth(struct blk_mq_tags
*tags
, unsigned int tdepth
)
666 tdepth
-= tags
->nr_reserved_tags
;
667 if (tdepth
> tags
->nr_tags
)
671 * Don't need (or can't) update reserved tags here, they remain
672 * static and should never need resizing.
674 bt_update_count(&tags
->bitmap_tags
, tdepth
);
675 blk_mq_tag_wakeup_all(tags
, false);
680 * blk_mq_unique_tag() - return a tag that is unique queue-wide
681 * @rq: request for which to compute a unique tag
683 * The tag field in struct request is unique per hardware queue but not over
684 * all hardware queues. Hence this function that returns a tag with the
685 * hardware context index in the upper bits and the per hardware queue tag in
688 * Note: When called for a request that is queued on a non-multiqueue request
689 * queue, the hardware context index is set to zero.
691 u32
blk_mq_unique_tag(struct request
*rq
)
693 struct request_queue
*q
= rq
->q
;
694 struct blk_mq_hw_ctx
*hctx
;
698 hctx
= q
->mq_ops
->map_queue(q
, rq
->mq_ctx
->cpu
);
699 hwq
= hctx
->queue_num
;
702 return (hwq
<< BLK_MQ_UNIQUE_TAG_BITS
) |
703 (rq
->tag
& BLK_MQ_UNIQUE_TAG_MASK
);
705 EXPORT_SYMBOL(blk_mq_unique_tag
);
707 ssize_t
blk_mq_tag_sysfs_show(struct blk_mq_tags
*tags
, char *page
)
709 char *orig_page
= page
;
710 unsigned int free
, res
;
715 page
+= sprintf(page
, "nr_tags=%u, reserved_tags=%u, "
716 "bits_per_word=%u\n",
717 tags
->nr_tags
, tags
->nr_reserved_tags
,
718 tags
->bitmap_tags
.bits_per_word
);
720 free
= bt_unused_tags(&tags
->bitmap_tags
);
721 res
= bt_unused_tags(&tags
->breserved_tags
);
723 page
+= sprintf(page
, "nr_free=%u, nr_reserved=%u\n", free
, res
);
724 page
+= sprintf(page
, "active_queues=%u\n", atomic_read(&tags
->active_queues
));
726 return page
- orig_page
;