[linux-2.6-block.git] / block / blk-mq-tag.c

/*
 * Fast and scalable bitmap tagging variant. Uses sparser bitmaps spread
 * over multiple cachelines to avoid ping-pong between multiple submitters
 * or submitter and completer. Uses rolling wakeups to avoid falling of
 * the scaling cliff when we run out of tags and have to start putting
 * submitters to sleep.
 *
 * Uses active queue tracking to support fairer distribution of tags
 * between multiple submitters when a shared tag map is used.
 *
 * Copyright (C) 2013-2014 Jens Axboe
 */
#include <linux/kernel.h>
#include <linux/module.h>
#include <linux/random.h>

#include <linux/blk-mq.h>
#include "blk.h"
#include "blk-mq.h"
#include "blk-mq-tag.h"

static bool bt_has_free_tags(struct blk_mq_bitmap_tags *bt)
{
	int i;

	for (i = 0; i < bt->map_nr; i++) {
		struct blk_align_bitmap *bm = &bt->map[i];
		int ret;

		ret = find_first_zero_bit(&bm->word, bm->depth);
		if (ret < bm->depth)
			return true;
	}

	return false;
}

bool blk_mq_has_free_tags(struct blk_mq_tags *tags)
{
	if (!tags)
		return true;

	return bt_has_free_tags(&tags->bitmap_tags);
}

static inline int bt_index_inc(int index)
{
	return (index + 1) & (BT_WAIT_QUEUES - 1);
}

static inline void bt_index_atomic_inc(atomic_t *index)
{
	int old = atomic_read(index);
	int new = bt_index_inc(old);
	atomic_cmpxchg(index, old, new);
}

/*
 * If a previously inactive queue goes active, bump the active user count.
 */
bool __blk_mq_tag_busy(struct blk_mq_hw_ctx *hctx)
{
	if (!test_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state) &&
	    !test_and_set_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state))
		atomic_inc(&hctx->tags->active_queues);

	return true;
}

/*
 * Wakeup all potentially sleeping on normal (non-reserved) tags
 */
static void blk_mq_tag_wakeup_all(struct blk_mq_tags *tags)
{
	struct blk_mq_bitmap_tags *bt;
	int i, wake_index;

	bt = &tags->bitmap_tags;
	wake_index = atomic_read(&bt->wake_index);
	for (i = 0; i < BT_WAIT_QUEUES; i++) {
		struct bt_wait_state *bs = &bt->bs[wake_index];

		if (waitqueue_active(&bs->wait))
			wake_up(&bs->wait);

		wake_index = bt_index_inc(wake_index);
	}
}

/*
 * If a previously busy queue goes inactive, potential waiters could now
 * be allowed to queue. Wake them up and check.
 */
void __blk_mq_tag_idle(struct blk_mq_hw_ctx *hctx)
{
	struct blk_mq_tags *tags = hctx->tags;

	if (!test_and_clear_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state))
		return;

	atomic_dec(&tags->active_queues);

	blk_mq_tag_wakeup_all(tags);
}

/*
 * For shared tag users, we track the number of currently active users
 * and attempt to provide a fair share of the tag depth for each of them.
 */
static inline bool hctx_may_queue(struct blk_mq_hw_ctx *hctx,
				  struct blk_mq_bitmap_tags *bt)
{
	unsigned int depth, users;

	if (!hctx || !(hctx->flags & BLK_MQ_F_TAG_SHARED))
		return true;
	if (!test_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state))
		return true;

	/*
	 * Don't try dividing an ant
	 */
	if (bt->depth == 1)
		return true;

	users = atomic_read(&hctx->tags->active_queues);
	if (!users)
		return true;

	/*
	 * Allow at least some tags
	 */
	depth = max((bt->depth + users - 1) / users, 4U);
	return atomic_read(&hctx->nr_active) < depth;
}

static int __bt_get_word(struct blk_align_bitmap *bm, unsigned int last_tag)
{
	int tag, org_last_tag, end;

	org_last_tag = last_tag;
	end = bm->depth;
	do {
restart:
		tag = find_next_zero_bit(&bm->word, end, last_tag);
		if (unlikely(tag >= end)) {
			/*
			 * We started with an offset, start from 0 to
			 * exhaust the map.
			 */
			if (org_last_tag && last_tag) {
				end = last_tag;
				last_tag = 0;
				goto restart;
			}
			return -1;
		}
		last_tag = tag + 1;
	} while (test_and_set_bit_lock(tag, &bm->word));

	return tag;
}

/*
 * Straight forward bitmap tag implementation, where each bit is a tag
 * (cleared == free, and set == busy). The small twist is using per-cpu
 * last_tag caches, which blk-mq stores in the blk_mq_ctx software queue
 * contexts. This enables us to drastically limit the space searched,
 * without dirtying an extra shared cacheline like we would if we stored
 * the cache value inside the shared blk_mq_bitmap_tags structure. On top
 * of that, each word of tags is in a separate cacheline. This means that
 * multiple users will tend to stick to different cachelines, at least
 * until the map is exhausted.
 */
static int __bt_get(struct blk_mq_hw_ctx *hctx, struct blk_mq_bitmap_tags *bt,
		    unsigned int *tag_cache)
{
	unsigned int last_tag, org_last_tag;
	int index, i, tag;

	if (!hctx_may_queue(hctx, bt))
		return -1;

	last_tag = org_last_tag = *tag_cache;
	index = TAG_TO_INDEX(bt, last_tag);

	for (i = 0; i < bt->map_nr; i++) {
		tag = __bt_get_word(&bt->map[index], TAG_TO_BIT(bt, last_tag));
		if (tag != -1) {
			tag += (index << bt->bits_per_word);
			goto done;
		}

		last_tag = 0;
		if (++index >= bt->map_nr)
			index = 0;
	}

	*tag_cache = 0;
	return -1;

	/*
	 * Only update the cache from the allocation path, if we ended
	 * up using the specific cached tag.
	 */
done:
	if (tag == org_last_tag) {
		last_tag = tag + 1;
		if (last_tag >= bt->depth - 1)
			last_tag = 0;

		*tag_cache = last_tag;
	}

	return tag;
}

static struct bt_wait_state *bt_wait_ptr(struct blk_mq_bitmap_tags *bt,
					 struct blk_mq_hw_ctx *hctx)
{
	struct bt_wait_state *bs;
	int wait_index;

	if (!hctx)
		return &bt->bs[0];

	wait_index = atomic_read(&hctx->wait_index);
	bs = &bt->bs[wait_index];
	bt_index_atomic_inc(&hctx->wait_index);
	return bs;
}

static int bt_get(struct blk_mq_alloc_data *data,
		struct blk_mq_bitmap_tags *bt,
		struct blk_mq_hw_ctx *hctx,
		unsigned int *last_tag)
{
	struct bt_wait_state *bs;
	DEFINE_WAIT(wait);
	int tag;

	tag = __bt_get(hctx, bt, last_tag);
	if (tag != -1)
		return tag;

	if (!(data->gfp & __GFP_WAIT))
		return -1;

	bs = bt_wait_ptr(bt, hctx);
	do {
		bool was_empty;

		was_empty = list_empty(&wait.task_list);
		prepare_to_wait(&bs->wait, &wait, TASK_UNINTERRUPTIBLE);

		tag = __bt_get(hctx, bt, last_tag);
		if (tag != -1)
			break;

		if (was_empty)
			atomic_set(&bs->wait_cnt, bt->wake_cnt);

		blk_mq_put_ctx(data->ctx);

		io_schedule();

		data->ctx = blk_mq_get_ctx(data->q);
		data->hctx = data->q->mq_ops->map_queue(data->q,
				data->ctx->cpu);
		if (data->reserved) {
			bt = &data->hctx->tags->breserved_tags;
		} else {
			last_tag = &data->ctx->last_tag;
			hctx = data->hctx;
			bt = &hctx->tags->bitmap_tags;
		}
		finish_wait(&bs->wait, &wait);
		bs = bt_wait_ptr(bt, hctx);
	} while (1);

	finish_wait(&bs->wait, &wait);
	return tag;
}

static unsigned int __blk_mq_get_tag(struct blk_mq_alloc_data *data)
{
	int tag;

	tag = bt_get(data, &data->hctx->tags->bitmap_tags, data->hctx,
			&data->ctx->last_tag);
	if (tag >= 0)
		return tag + data->hctx->tags->nr_reserved_tags;

	return BLK_MQ_TAG_FAIL;
}

static unsigned int __blk_mq_get_reserved_tag(struct blk_mq_alloc_data *data)
{
	int tag, zero = 0;

	if (unlikely(!data->hctx->tags->nr_reserved_tags)) {
		WARN_ON_ONCE(1);
		return BLK_MQ_TAG_FAIL;
	}

	tag = bt_get(data, &data->hctx->tags->breserved_tags, NULL, &zero);
	if (tag < 0)
		return BLK_MQ_TAG_FAIL;

	return tag;
}

unsigned int blk_mq_get_tag(struct blk_mq_alloc_data *data)
{
	if (!data->reserved)
		return __blk_mq_get_tag(data);

	return __blk_mq_get_reserved_tag(data);
}

static struct bt_wait_state *bt_wake_ptr(struct blk_mq_bitmap_tags *bt)
{
	int i, wake_index;

	wake_index = atomic_read(&bt->wake_index);
	for (i = 0; i < BT_WAIT_QUEUES; i++) {
		struct bt_wait_state *bs = &bt->bs[wake_index];

		if (waitqueue_active(&bs->wait)) {
			int o = atomic_read(&bt->wake_index);
			if (wake_index != o)
				atomic_cmpxchg(&bt->wake_index, o, wake_index);

			return bs;
		}

		wake_index = bt_index_inc(wake_index);
	}

	return NULL;
}

static void bt_clear_tag(struct blk_mq_bitmap_tags *bt, unsigned int tag)
{
	const int index = TAG_TO_INDEX(bt, tag);
	struct bt_wait_state *bs;
	int wait_cnt;

	/*
	 * The unlock memory barrier need to order access to req in free
	 * path and clearing tag bit
	 */
	clear_bit_unlock(TAG_TO_BIT(bt, tag), &bt->map[index].word);

	bs = bt_wake_ptr(bt);
	if (!bs)
		return;

	wait_cnt = atomic_dec_return(&bs->wait_cnt);
	if (wait_cnt == 0) {
wake:
		atomic_add(bt->wake_cnt, &bs->wait_cnt);
		bt_index_atomic_inc(&bt->wake_index);
		wake_up(&bs->wait);
	} else if (wait_cnt < 0) {
		wait_cnt = atomic_inc_return(&bs->wait_cnt);
		if (!wait_cnt)
			goto wake;
	}
}

static void __blk_mq_put_tag(struct blk_mq_tags *tags, unsigned int tag)
{
	BUG_ON(tag >= tags->nr_tags);

	bt_clear_tag(&tags->bitmap_tags, tag);
}

static void __blk_mq_put_reserved_tag(struct blk_mq_tags *tags,
				      unsigned int tag)
{
	BUG_ON(tag >= tags->nr_reserved_tags);

	bt_clear_tag(&tags->breserved_tags, tag);
}

void blk_mq_put_tag(struct blk_mq_hw_ctx *hctx, unsigned int tag,
		    unsigned int *last_tag)
{
	struct blk_mq_tags *tags = hctx->tags;

	if (tag >= tags->nr_reserved_tags) {
		const int real_tag = tag - tags->nr_reserved_tags;

		__blk_mq_put_tag(tags, real_tag);
		*last_tag = real_tag;
	} else
		__blk_mq_put_reserved_tag(tags, tag);
}

static void bt_for_each_free(struct blk_mq_bitmap_tags *bt,
			     unsigned long *free_map, unsigned int off)
{
	int i;

	for (i = 0; i < bt->map_nr; i++) {
		struct blk_align_bitmap *bm = &bt->map[i];
		int bit = 0;

		do {
			bit = find_next_zero_bit(&bm->word, bm->depth, bit);
			if (bit >= bm->depth)
				break;

			__set_bit(bit + off, free_map);
			bit++;
		} while (1);

		off += (1 << bt->bits_per_word);
	}
}

void blk_mq_tag_busy_iter(struct blk_mq_tags *tags,
			  void (*fn)(void *, unsigned long *), void *data)
{
	unsigned long *tag_map;
	size_t map_size;

	map_size = ALIGN(tags->nr_tags, BITS_PER_LONG) / BITS_PER_LONG;
	tag_map = kzalloc(map_size * sizeof(unsigned long), GFP_ATOMIC);
	if (!tag_map)
		return;

	bt_for_each_free(&tags->bitmap_tags, tag_map, tags->nr_reserved_tags);
	if (tags->nr_reserved_tags)
		bt_for_each_free(&tags->breserved_tags, tag_map, 0);

	fn(data, tag_map);
	kfree(tag_map);
}
EXPORT_SYMBOL(blk_mq_tag_busy_iter);

static unsigned int bt_unused_tags(struct blk_mq_bitmap_tags *bt)
{
	unsigned int i, used;

	for (i = 0, used = 0; i < bt->map_nr; i++) {
		struct blk_align_bitmap *bm = &bt->map[i];

		used += bitmap_weight(&bm->word, bm->depth);
	}

	return bt->depth - used;
}

static void bt_update_count(struct blk_mq_bitmap_tags *bt,
			    unsigned int depth)
{
	unsigned int tags_per_word = 1U << bt->bits_per_word;
	unsigned int map_depth = depth;

	if (depth) {
		int i;

		for (i = 0; i < bt->map_nr; i++) {
			bt->map[i].depth = min(map_depth, tags_per_word);
			map_depth -= bt->map[i].depth;
		}
	}

	bt->wake_cnt = BT_WAIT_BATCH;
	if (bt->wake_cnt > depth / 4)
		bt->wake_cnt = max(1U, depth / 4);

	bt->depth = depth;
}

static int bt_alloc(struct blk_mq_bitmap_tags *bt, unsigned int depth,
			int node, bool reserved)
{
	int i;

	bt->bits_per_word = ilog2(BITS_PER_LONG);

	/*
	 * Depth can be zero for reserved tags, that's not a failure
	 * condition.
	 */
	if (depth) {
		unsigned int nr, tags_per_word;

		tags_per_word = (1 << bt->bits_per_word);

		/*
		 * If the tag space is small, shrink the number of tags
		 * per word so we spread over a few cachelines, at least.
		 * If less than 4 tags, just forget about it, it's not
		 * going to work optimally anyway.
		 */
		if (depth >= 4) {
			while (tags_per_word * 4 > depth) {
				bt->bits_per_word--;
				tags_per_word = (1 << bt->bits_per_word);
			}
		}

		nr = ALIGN(depth, tags_per_word) / tags_per_word;
		bt->map = kzalloc_node(nr * sizeof(struct blk_align_bitmap),
						GFP_KERNEL, node);
		if (!bt->map)
			return -ENOMEM;

		bt->map_nr = nr;
	}

	bt->bs = kzalloc(BT_WAIT_QUEUES * sizeof(*bt->bs), GFP_KERNEL);
	if (!bt->bs) {
		kfree(bt->map);
		return -ENOMEM;
	}

	for (i = 0; i < BT_WAIT_QUEUES; i++)
		init_waitqueue_head(&bt->bs[i].wait);

	bt_update_count(bt, depth);
	return 0;
}

static void bt_free(struct blk_mq_bitmap_tags *bt)
{
	kfree(bt->map);
	kfree(bt->bs);
}

static struct blk_mq_tags *blk_mq_init_bitmap_tags(struct blk_mq_tags *tags,
						   int node)
{
	unsigned int depth = tags->nr_tags - tags->nr_reserved_tags;

	if (bt_alloc(&tags->bitmap_tags, depth, node, false))
		goto enomem;
	if (bt_alloc(&tags->breserved_tags, tags->nr_reserved_tags, node, true))
		goto enomem;

	return tags;
enomem:
	bt_free(&tags->bitmap_tags);
	kfree(tags);
	return NULL;
}

struct blk_mq_tags *blk_mq_init_tags(unsigned int total_tags,
				     unsigned int reserved_tags, int node)
{
	struct blk_mq_tags *tags;

	if (total_tags > BLK_MQ_TAG_MAX) {
		pr_err("blk-mq: tag depth too large\n");
		return NULL;
	}

	tags = kzalloc_node(sizeof(*tags), GFP_KERNEL, node);
	if (!tags)
		return NULL;

	tags->nr_tags = total_tags;
	tags->nr_reserved_tags = reserved_tags;

	return blk_mq_init_bitmap_tags(tags, node);
}

void blk_mq_free_tags(struct blk_mq_tags *tags)
{
	bt_free(&tags->bitmap_tags);
	bt_free(&tags->breserved_tags);
	kfree(tags);
}

void blk_mq_tag_init_last_tag(struct blk_mq_tags *tags, unsigned int *tag)
{
	unsigned int depth = tags->nr_tags - tags->nr_reserved_tags;

	*tag = prandom_u32() % depth;
}

int blk_mq_tag_update_depth(struct blk_mq_tags *tags, unsigned int tdepth)
{
	tdepth -= tags->nr_reserved_tags;
	if (tdepth > tags->nr_tags)
		return -EINVAL;

	/*
	 * Don't need (or can't) update reserved tags here, they remain
	 * static and should never need resizing.
	 */
	bt_update_count(&tags->bitmap_tags, tdepth);
	blk_mq_tag_wakeup_all(tags);
	return 0;
}

ssize_t blk_mq_tag_sysfs_show(struct blk_mq_tags *tags, char *page)
{
	char *orig_page = page;
	unsigned int free, res;

	if (!tags)
		return 0;

	page += sprintf(page, "nr_tags=%u, reserved_tags=%u, "
			"bits_per_word=%u\n",
			tags->nr_tags, tags->nr_reserved_tags,
			tags->bitmap_tags.bits_per_word);

	free = bt_unused_tags(&tags->bitmap_tags);
	res = bt_unused_tags(&tags->breserved_tags);

	page += sprintf(page, "nr_free=%u, nr_reserved=%u\n", free, res);
	page += sprintf(page, "active_queues=%u\n", atomic_read(&tags->active_queues));

	return page - orig_page;
}
Commit	Line	Data
75bb4625 JA	1	/*
	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
	6	* submitters to sleep.
	7	*
	8	* Uses active queue tracking to support fairer distribution of tags
	9	* between multiple submitters when a shared tag map is used.
	10	*
	11	* Copyright (C) 2013-2014 Jens Axboe
	12	*/
320ae51f JA	13	#include <linux/kernel.h>
320ae51f JA	14	#include <linux/module.h>
4bb659b1	15	#include <linux/random.h>
320ae51f JA	16
	17	#include <linux/blk-mq.h>
	18	#include "blk.h"
	19	#include "blk-mq.h"
	20	#include "blk-mq-tag.h"
	21
4bb659b1 JA	22	static bool bt_has_free_tags(struct blk_mq_bitmap_tags *bt)
	23	{
	24	int i;
	25
	26	for (i = 0; i < bt->map_nr; i++) {
e93ecf60	27	struct blk_align_bitmap *bm = &bt->map[i];
4bb659b1 JA	28	int ret;
	29
	30	ret = find_first_zero_bit(&bm->word, bm->depth);
	31	if (ret < bm->depth)
	32	return true;
	33	}
	34
	35	return false;
320ae51f JA	36	}
	37
	38	bool blk_mq_has_free_tags(struct blk_mq_tags *tags)
	39	{
4bb659b1 JA	40	if (!tags)
	41	return true;
	42
	43	return bt_has_free_tags(&tags->bitmap_tags);
	44	}
	45
8537b120	46	static inline int bt_index_inc(int index)
0d2602ca	47	{
8537b120 AG	48	return (index + 1) & (BT_WAIT_QUEUES - 1);
	49	}
	50
	51	static inline void bt_index_atomic_inc(atomic_t *index)
	52	{
	53	int old = atomic_read(index);
	54	int new = bt_index_inc(old);
	55	atomic_cmpxchg(index, old, new);
0d2602ca JA	56	}
	57
	58	/*
	59	* If a previously inactive queue goes active, bump the active user count.
	60	*/
	61	bool __blk_mq_tag_busy(struct blk_mq_hw_ctx *hctx)
	62	{
	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);
	66
	67	return true;
	68	}
	69
	70	/*
e3a2b3f9	71	* Wakeup all potentially sleeping on normal (non-reserved) tags
0d2602ca	72	*/
e3a2b3f9	73	static void blk_mq_tag_wakeup_all(struct blk_mq_tags *tags)
0d2602ca	74	{
0d2602ca JA	75	struct blk_mq_bitmap_tags *bt;
	76	int i, wake_index;
	77
0d2602ca	78	bt = &tags->bitmap_tags;
8537b120	79	wake_index = atomic_read(&bt->wake_index);
0d2602ca JA	80	for (i = 0; i < BT_WAIT_QUEUES; i++) {
	81	struct bt_wait_state *bs = &bt->bs[wake_index];
	82
	83	if (waitqueue_active(&bs->wait))
	84	wake_up(&bs->wait);
	85
8537b120	86	wake_index = bt_index_inc(wake_index);
0d2602ca JA	87	}
	88	}
	89
e3a2b3f9 JA	90	/*
	91	* If a previously busy queue goes inactive, potential waiters could now
	92	* be allowed to queue. Wake them up and check.
	93	*/
	94	void __blk_mq_tag_idle(struct blk_mq_hw_ctx *hctx)
	95	{
	96	struct blk_mq_tags *tags = hctx->tags;
	97
	98	if (!test_and_clear_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state))
	99	return;
	100
	101	atomic_dec(&tags->active_queues);
	102
	103	blk_mq_tag_wakeup_all(tags);
	104	}
	105
0d2602ca JA	106	/*
	107	* For shared tag users, we track the number of currently active users
	108	* and attempt to provide a fair share of the tag depth for each of them.
	109	*/
	110	static inline bool hctx_may_queue(struct blk_mq_hw_ctx *hctx,
	111	struct blk_mq_bitmap_tags *bt)
	112	{
	113	unsigned int depth, users;
	114
	115	if (!hctx \|\| !(hctx->flags & BLK_MQ_F_TAG_SHARED))
	116	return true;
	117	if (!test_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state))
	118	return true;
	119
	120	/*
	121	* Don't try dividing an ant
	122	*/
	123	if (bt->depth == 1)
	124	return true;
	125
	126	users = atomic_read(&hctx->tags->active_queues);
	127	if (!users)
	128	return true;
	129
	130	/*
	131	* Allow at least some tags
	132	*/
	133	depth = max((bt->depth + users - 1) / users, 4U);
	134	return atomic_read(&hctx->nr_active) < depth;
	135	}
	136
e93ecf60	137	static int __bt_get_word(struct blk_align_bitmap *bm, unsigned int last_tag)
4bb659b1 JA	138	{
	139	int tag, org_last_tag, end;
	140
59d13bf5	141	org_last_tag = last_tag;
4bb659b1 JA	142	end = bm->depth;
	143	do {
	144	restart:
	145	tag = find_next_zero_bit(&bm->word, end, last_tag);
	146	if (unlikely(tag >= end)) {
	147	/*
	148	* We started with an offset, start from 0 to
	149	* exhaust the map.
	150	*/
	151	if (org_last_tag && last_tag) {
	152	end = last_tag;
	153	last_tag = 0;
	154	goto restart;
	155	}
	156	return -1;
	157	}
	158	last_tag = tag + 1;
	159	} while (test_and_set_bit_lock(tag, &bm->word));
	160
	161	return tag;
	162	}
	163
	164	/*
	165	* Straight forward bitmap tag implementation, where each bit is a tag
	166	* (cleared == free, and set == busy). The small twist is using per-cpu
	167	* last_tag caches, which blk-mq stores in the blk_mq_ctx software queue
	168	* contexts. This enables us to drastically limit the space searched,
	169	* without dirtying an extra shared cacheline like we would if we stored
	170	* the cache value inside the shared blk_mq_bitmap_tags structure. On top
	171	* of that, each word of tags is in a separate cacheline. This means that
	172	* multiple users will tend to stick to different cachelines, at least
	173	* until the map is exhausted.
	174	*/
0d2602ca JA	175	static int __bt_get(struct blk_mq_hw_ctx hctx, struct blk_mq_bitmap_tags bt,
0d2602ca JA	176	unsigned int *tag_cache)
4bb659b1 JA	177	{
	178	unsigned int last_tag, org_last_tag;
	179	int index, i, tag;
	180
0d2602ca JA	181	if (!hctx_may_queue(hctx, bt))
	182	return -1;
	183
4bb659b1	184	last_tag = org_last_tag = *tag_cache;
59d13bf5	185	index = TAG_TO_INDEX(bt, last_tag);
4bb659b1 JA	186
4bb659b1 JA	187	for (i = 0; i < bt->map_nr; i++) {
59d13bf5	188	tag = __bt_get_word(&bt->map[index], TAG_TO_BIT(bt, last_tag));
4bb659b1	189	if (tag != -1) {
59d13bf5	190	tag += (index << bt->bits_per_word);
4bb659b1 JA	191	goto done;
	192	}
	193
	194	last_tag = 0;
	195	if (++index >= bt->map_nr)
	196	index = 0;
	197	}
	198
	199	*tag_cache = 0;
	200	return -1;
	201
	202	/*
	203	* Only update the cache from the allocation path, if we ended
	204	* up using the specific cached tag.
	205	*/
	206	done:
	207	if (tag == org_last_tag) {
	208	last_tag = tag + 1;
	209	if (last_tag >= bt->depth - 1)
	210	last_tag = 0;
	211
	212	*tag_cache = last_tag;
	213	}
	214
	215	return tag;
	216	}
	217
4bb659b1 JA	218	static struct bt_wait_state bt_wait_ptr(struct blk_mq_bitmap_tags bt,
	219	struct blk_mq_hw_ctx *hctx)
	220	{
	221	struct bt_wait_state *bs;
8537b120	222	int wait_index;
4bb659b1 JA	223
	224	if (!hctx)
	225	return &bt->bs[0];
	226
8537b120 AG	227	wait_index = atomic_read(&hctx->wait_index);
	228	bs = &bt->bs[wait_index];
	229	bt_index_atomic_inc(&hctx->wait_index);
4bb659b1	230	return bs;
320ae51f JA	231	}
320ae51f JA	232
cb96a42c ML	233	static int bt_get(struct blk_mq_alloc_data *data,
	234	struct blk_mq_bitmap_tags *bt,
	235	struct blk_mq_hw_ctx *hctx,
	236	unsigned int *last_tag)
320ae51f	237	{
4bb659b1 JA	238	struct bt_wait_state *bs;
4bb659b1 JA	239	DEFINE_WAIT(wait);
320ae51f JA	240	int tag;
320ae51f JA	241
0d2602ca	242	tag = __bt_get(hctx, bt, last_tag);
4bb659b1 JA	243	if (tag != -1)
	244	return tag;
	245
cb96a42c	246	if (!(data->gfp & __GFP_WAIT))
4bb659b1 JA	247	return -1;
	248
	249	bs = bt_wait_ptr(bt, hctx);
	250	do {
	251	bool was_empty;
	252
	253	was_empty = list_empty(&wait.task_list);
	254	prepare_to_wait(&bs->wait, &wait, TASK_UNINTERRUPTIBLE);
	255
0d2602ca	256	tag = __bt_get(hctx, bt, last_tag);
4bb659b1 JA	257	if (tag != -1)
	258	break;
	259
	260	if (was_empty)
	261	atomic_set(&bs->wait_cnt, bt->wake_cnt);
	262
cb96a42c ML	263	blk_mq_put_ctx(data->ctx);
cb96a42c ML	264
4bb659b1	265	io_schedule();
cb96a42c ML	266
	267	data->ctx = blk_mq_get_ctx(data->q);
	268	data->hctx = data->q->mq_ops->map_queue(data->q,
	269	data->ctx->cpu);
	270	if (data->reserved) {
	271	bt = &data->hctx->tags->breserved_tags;
	272	} else {
	273	last_tag = &data->ctx->last_tag;
	274	hctx = data->hctx;
	275	bt = &hctx->tags->bitmap_tags;
	276	}
	277	finish_wait(&bs->wait, &wait);
	278	bs = bt_wait_ptr(bt, hctx);
4bb659b1 JA	279	} while (1);
	280
	281	finish_wait(&bs->wait, &wait);
	282	return tag;
	283	}
	284
cb96a42c	285	static unsigned int __blk_mq_get_tag(struct blk_mq_alloc_data *data)
4bb659b1 JA	286	{
	287	int tag;
	288
cb96a42c ML	289	tag = bt_get(data, &data->hctx->tags->bitmap_tags, data->hctx,
cb96a42c ML	290	&data->ctx->last_tag);
4bb659b1	291	if (tag >= 0)
cb96a42c	292	return tag + data->hctx->tags->nr_reserved_tags;
4bb659b1 JA	293
4bb659b1 JA	294	return BLK_MQ_TAG_FAIL;
320ae51f JA	295	}
320ae51f JA	296
cb96a42c	297	static unsigned int __blk_mq_get_reserved_tag(struct blk_mq_alloc_data *data)
320ae51f	298	{
4bb659b1	299	int tag, zero = 0;
320ae51f	300
cb96a42c	301	if (unlikely(!data->hctx->tags->nr_reserved_tags)) {
320ae51f JA	302	WARN_ON_ONCE(1);
	303	return BLK_MQ_TAG_FAIL;
	304	}
	305
cb96a42c	306	tag = bt_get(data, &data->hctx->tags->breserved_tags, NULL, &zero);
320ae51f JA	307	if (tag < 0)
320ae51f JA	308	return BLK_MQ_TAG_FAIL;
4bb659b1	309
320ae51f JA	310	return tag;
	311	}
	312
cb96a42c	313	unsigned int blk_mq_get_tag(struct blk_mq_alloc_data *data)
320ae51f	314	{
cb96a42c ML	315	if (!data->reserved)
cb96a42c ML	316	return __blk_mq_get_tag(data);
320ae51f	317
cb96a42c	318	return __blk_mq_get_reserved_tag(data);
320ae51f JA	319	}
320ae51f JA	320
4bb659b1 JA	321	static struct bt_wait_state bt_wake_ptr(struct blk_mq_bitmap_tags bt)
	322	{
	323	int i, wake_index;
	324
8537b120	325	wake_index = atomic_read(&bt->wake_index);
4bb659b1 JA	326	for (i = 0; i < BT_WAIT_QUEUES; i++) {
	327	struct bt_wait_state *bs = &bt->bs[wake_index];
	328
	329	if (waitqueue_active(&bs->wait)) {
8537b120 AG	330	int o = atomic_read(&bt->wake_index);
	331	if (wake_index != o)
	332	atomic_cmpxchg(&bt->wake_index, o, wake_index);
4bb659b1 JA	333
	334	return bs;
	335	}
	336
8537b120	337	wake_index = bt_index_inc(wake_index);
4bb659b1 JA	338	}
	339
	340	return NULL;
	341	}
	342
	343	static void bt_clear_tag(struct blk_mq_bitmap_tags *bt, unsigned int tag)
	344	{
59d13bf5	345	const int index = TAG_TO_INDEX(bt, tag);
4bb659b1	346	struct bt_wait_state *bs;
2971c35f	347	int wait_cnt;
4bb659b1	348
0289b2e1 ML	349	/*
	350	* The unlock memory barrier need to order access to req in free
	351	* path and clearing tag bit
	352	*/
	353	clear_bit_unlock(TAG_TO_BIT(bt, tag), &bt->map[index].word);
4bb659b1 JA	354
4bb659b1 JA	355	bs = bt_wake_ptr(bt);
2971c35f AG	356	if (!bs)
	357	return;
	358
	359	wait_cnt = atomic_dec_return(&bs->wait_cnt);
	360	if (wait_cnt == 0) {
	361	wake:
	362	atomic_add(bt->wake_cnt, &bs->wait_cnt);
8537b120	363	bt_index_atomic_inc(&bt->wake_index);
4bb659b1	364	wake_up(&bs->wait);
2971c35f AG	365	} else if (wait_cnt < 0) {
	366	wait_cnt = atomic_inc_return(&bs->wait_cnt);
	367	if (!wait_cnt)
	368	goto wake;
4bb659b1 JA	369	}
	370	}
	371
320ae51f JA	372	static void __blk_mq_put_tag(struct blk_mq_tags *tags, unsigned int tag)
	373	{
	374	BUG_ON(tag >= tags->nr_tags);
	375
4bb659b1	376	bt_clear_tag(&tags->bitmap_tags, tag);
320ae51f JA	377	}
	378
	379	static void __blk_mq_put_reserved_tag(struct blk_mq_tags *tags,
	380	unsigned int tag)
	381	{
	382	BUG_ON(tag >= tags->nr_reserved_tags);
	383
4bb659b1	384	bt_clear_tag(&tags->breserved_tags, tag);
320ae51f JA	385	}
320ae51f JA	386
0d2602ca	387	void blk_mq_put_tag(struct blk_mq_hw_ctx *hctx, unsigned int tag,
4bb659b1	388	unsigned int *last_tag)
320ae51f	389	{
0d2602ca JA	390	struct blk_mq_tags *tags = hctx->tags;
0d2602ca JA	391
4bb659b1 JA	392	if (tag >= tags->nr_reserved_tags) {
	393	const int real_tag = tag - tags->nr_reserved_tags;
	394
	395	__blk_mq_put_tag(tags, real_tag);
	396	*last_tag = real_tag;
	397	} else
320ae51f JA	398	__blk_mq_put_reserved_tag(tags, tag);
	399	}
	400
4bb659b1 JA	401	static void bt_for_each_free(struct blk_mq_bitmap_tags *bt,
4bb659b1 JA	402	unsigned long *free_map, unsigned int off)
320ae51f	403	{
4bb659b1 JA	404	int i;
	405
	406	for (i = 0; i < bt->map_nr; i++) {
e93ecf60	407	struct blk_align_bitmap *bm = &bt->map[i];
4bb659b1 JA	408	int bit = 0;
	409
	410	do {
	411	bit = find_next_zero_bit(&bm->word, bm->depth, bit);
	412	if (bit >= bm->depth)
	413	break;
	414
	415	__set_bit(bit + off, free_map);
	416	bit++;
	417	} while (1);
	418
59d13bf5	419	off += (1 << bt->bits_per_word);
4bb659b1	420	}
320ae51f JA	421	}
	422
	423	void blk_mq_tag_busy_iter(struct blk_mq_tags *tags,
	424	void (fn)(void , unsigned long ), void data)
	425	{
	426	unsigned long *tag_map;
	427	size_t map_size;
	428
	429	map_size = ALIGN(tags->nr_tags, BITS_PER_LONG) / BITS_PER_LONG;
	430	tag_map = kzalloc(map_size * sizeof(unsigned long), GFP_ATOMIC);
	431	if (!tag_map)
	432	return;
	433
4bb659b1	434	bt_for_each_free(&tags->bitmap_tags, tag_map, tags->nr_reserved_tags);
320ae51f	435	if (tags->nr_reserved_tags)
4bb659b1	436	bt_for_each_free(&tags->breserved_tags, tag_map, 0);
320ae51f JA	437
	438	fn(data, tag_map);
	439	kfree(tag_map);
	440	}
edf866b3	441	EXPORT_SYMBOL(blk_mq_tag_busy_iter);
320ae51f	442
4bb659b1 JA	443	static unsigned int bt_unused_tags(struct blk_mq_bitmap_tags *bt)
	444	{
	445	unsigned int i, used;
	446
	447	for (i = 0, used = 0; i < bt->map_nr; i++) {
e93ecf60	448	struct blk_align_bitmap *bm = &bt->map[i];
4bb659b1 JA	449
	450	used += bitmap_weight(&bm->word, bm->depth);
	451	}
	452
	453	return bt->depth - used;
	454	}
	455
e3a2b3f9 JA	456	static void bt_update_count(struct blk_mq_bitmap_tags *bt,
	457	unsigned int depth)
	458	{
	459	unsigned int tags_per_word = 1U << bt->bits_per_word;
	460	unsigned int map_depth = depth;
	461
	462	if (depth) {
	463	int i;
	464
	465	for (i = 0; i < bt->map_nr; i++) {
	466	bt->map[i].depth = min(map_depth, tags_per_word);
	467	map_depth -= bt->map[i].depth;
	468	}
	469	}
	470
	471	bt->wake_cnt = BT_WAIT_BATCH;
	472	if (bt->wake_cnt > depth / 4)
	473	bt->wake_cnt = max(1U, depth / 4);
	474
	475	bt->depth = depth;
	476	}
	477
4bb659b1 JA	478	static int bt_alloc(struct blk_mq_bitmap_tags *bt, unsigned int depth,
	479	int node, bool reserved)
	480	{
	481	int i;
	482
59d13bf5 JA	483	bt->bits_per_word = ilog2(BITS_PER_LONG);
59d13bf5 JA	484
4bb659b1 JA	485	/*
	486	* Depth can be zero for reserved tags, that's not a failure
	487	* condition.
	488	*/
	489	if (depth) {
e3a2b3f9	490	unsigned int nr, tags_per_word;
59d13bf5 JA	491
	492	tags_per_word = (1 << bt->bits_per_word);
	493
	494	/*
	495	* If the tag space is small, shrink the number of tags
	496	* per word so we spread over a few cachelines, at least.
	497	* If less than 4 tags, just forget about it, it's not
	498	* going to work optimally anyway.
	499	*/
	500	if (depth >= 4) {
	501	while (tags_per_word * 4 > depth) {
	502	bt->bits_per_word--;
	503	tags_per_word = (1 << bt->bits_per_word);
	504	}
	505	}
4bb659b1	506
59d13bf5	507	nr = ALIGN(depth, tags_per_word) / tags_per_word;
e93ecf60	508	bt->map = kzalloc_node(nr * sizeof(struct blk_align_bitmap),
4bb659b1 JA	509	GFP_KERNEL, node);
	510	if (!bt->map)
	511	return -ENOMEM;
	512
	513	bt->map_nr = nr;
4bb659b1 JA	514	}
	515
	516	bt->bs = kzalloc(BT_WAIT_QUEUES * sizeof(*bt->bs), GFP_KERNEL);
	517	if (!bt->bs) {
	518	kfree(bt->map);
	519	return -ENOMEM;
	520	}
	521
	522	for (i = 0; i < BT_WAIT_QUEUES; i++)
	523	init_waitqueue_head(&bt->bs[i].wait);
	524
e3a2b3f9	525	bt_update_count(bt, depth);
4bb659b1 JA	526	return 0;
	527	}
	528
	529	static void bt_free(struct blk_mq_bitmap_tags *bt)
	530	{
	531	kfree(bt->map);
	532	kfree(bt->bs);
	533	}
	534
	535	static struct blk_mq_tags blk_mq_init_bitmap_tags(struct blk_mq_tags tags,
	536	int node)
	537	{
	538	unsigned int depth = tags->nr_tags - tags->nr_reserved_tags;
	539
	540	if (bt_alloc(&tags->bitmap_tags, depth, node, false))
	541	goto enomem;
	542	if (bt_alloc(&tags->breserved_tags, tags->nr_reserved_tags, node, true))
	543	goto enomem;
	544
	545	return tags;
	546	enomem:
	547	bt_free(&tags->bitmap_tags);
	548	kfree(tags);
	549	return NULL;
	550	}
	551
320ae51f JA	552	struct blk_mq_tags *blk_mq_init_tags(unsigned int total_tags,
	553	unsigned int reserved_tags, int node)
	554	{
320ae51f	555	struct blk_mq_tags *tags;
320ae51f JA	556
	557	if (total_tags > BLK_MQ_TAG_MAX) {
	558	pr_err("blk-mq: tag depth too large\n");
	559	return NULL;
	560	}
	561
	562	tags = kzalloc_node(sizeof(*tags), GFP_KERNEL, node);
	563	if (!tags)
	564	return NULL;
	565
320ae51f JA	566	tags->nr_tags = total_tags;
320ae51f JA	567	tags->nr_reserved_tags = reserved_tags;
320ae51f	568
4bb659b1	569	return blk_mq_init_bitmap_tags(tags, node);
320ae51f JA	570	}
	571
	572	void blk_mq_free_tags(struct blk_mq_tags *tags)
	573	{
4bb659b1 JA	574	bt_free(&tags->bitmap_tags);
4bb659b1 JA	575	bt_free(&tags->breserved_tags);
320ae51f JA	576	kfree(tags);
	577	}
	578
4bb659b1 JA	579	void blk_mq_tag_init_last_tag(struct blk_mq_tags tags, unsigned int tag)
	580	{
	581	unsigned int depth = tags->nr_tags - tags->nr_reserved_tags;
	582
9d3d21ae	583	*tag = prandom_u32() % depth;
4bb659b1 JA	584	}
4bb659b1 JA	585
e3a2b3f9 JA	586	int blk_mq_tag_update_depth(struct blk_mq_tags *tags, unsigned int tdepth)
	587	{
	588	tdepth -= tags->nr_reserved_tags;
	589	if (tdepth > tags->nr_tags)
	590	return -EINVAL;
	591
	592	/*
	593	* Don't need (or can't) update reserved tags here, they remain
	594	* static and should never need resizing.
	595	*/
	596	bt_update_count(&tags->bitmap_tags, tdepth);
	597	blk_mq_tag_wakeup_all(tags);
	598	return 0;
	599	}
	600
320ae51f JA	601	ssize_t blk_mq_tag_sysfs_show(struct blk_mq_tags tags, char page)
	602	{
	603	char *orig_page = page;
4bb659b1	604	unsigned int free, res;
320ae51f JA	605
	606	if (!tags)
	607	return 0;
	608
59d13bf5 JA	609	page += sprintf(page, "nr_tags=%u, reserved_tags=%u, "
	610	"bits_per_word=%u\n",
	611	tags->nr_tags, tags->nr_reserved_tags,
	612	tags->bitmap_tags.bits_per_word);
320ae51f	613
4bb659b1 JA	614	free = bt_unused_tags(&tags->bitmap_tags);
4bb659b1 JA	615	res = bt_unused_tags(&tags->breserved_tags);
320ae51f	616
4bb659b1	617	page += sprintf(page, "nr_free=%u, nr_reserved=%u\n", free, res);
0d2602ca	618	page += sprintf(page, "active_queues=%u\n", atomic_read(&tags->active_queues));
320ae51f JA	619
	620	return page - orig_page;
	621	}