[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 void bt_index_inc(unsigned int *index)
{
	*index = (*index + 1) & (BT_WAIT_QUEUES - 1);
}

/*
 * 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 = 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);

		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;

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

	bs = &bt->bs[hctx->wait_index];
	bt_index_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 = 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)) {
			if (wake_index != bt->wake_index)
				bt->wake_index = wake_index;

			return bs;
		}

		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;

	/*
	 * 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 && atomic_dec_and_test(&bs->wait_cnt)) {
		atomic_set(&bs->wait_cnt, bt->wake_cnt);
		bt_index_inc(&bt->wake_index);
		wake_up(&bs->wait);
	}
}

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
0d2602ca JA	46	static inline void bt_index_inc(unsigned int *index)
	47	{
	48	index = (index + 1) & (BT_WAIT_QUEUES - 1);
	49	}
	50
	51	/*
	52	* If a previously inactive queue goes active, bump the active user count.
	53	*/
	54	bool __blk_mq_tag_busy(struct blk_mq_hw_ctx *hctx)
	55	{
	56	if (!test_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state) &&
	57	!test_and_set_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state))
	58	atomic_inc(&hctx->tags->active_queues);
	59
	60	return true;
	61	}
	62
	63	/*
e3a2b3f9	64	* Wakeup all potentially sleeping on normal (non-reserved) tags
0d2602ca	65	*/
e3a2b3f9	66	static void blk_mq_tag_wakeup_all(struct blk_mq_tags *tags)
0d2602ca	67	{
0d2602ca JA	68	struct blk_mq_bitmap_tags *bt;
	69	int i, wake_index;
	70
0d2602ca JA	71	bt = &tags->bitmap_tags;
	72	wake_index = bt->wake_index;
	73	for (i = 0; i < BT_WAIT_QUEUES; i++) {
	74	struct bt_wait_state *bs = &bt->bs[wake_index];
	75
	76	if (waitqueue_active(&bs->wait))
	77	wake_up(&bs->wait);
	78
	79	bt_index_inc(&wake_index);
	80	}
	81	}
	82
e3a2b3f9 JA	83	/*
	84	* If a previously busy queue goes inactive, potential waiters could now
	85	* be allowed to queue. Wake them up and check.
	86	*/
	87	void __blk_mq_tag_idle(struct blk_mq_hw_ctx *hctx)
	88	{
	89	struct blk_mq_tags *tags = hctx->tags;
	90
	91	if (!test_and_clear_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state))
	92	return;
	93
	94	atomic_dec(&tags->active_queues);
	95
	96	blk_mq_tag_wakeup_all(tags);
	97	}
	98
0d2602ca JA	99	/*
	100	* For shared tag users, we track the number of currently active users
	101	* and attempt to provide a fair share of the tag depth for each of them.
	102	*/
	103	static inline bool hctx_may_queue(struct blk_mq_hw_ctx *hctx,
	104	struct blk_mq_bitmap_tags *bt)
	105	{
	106	unsigned int depth, users;
	107
	108	if (!hctx \|\| !(hctx->flags & BLK_MQ_F_TAG_SHARED))
	109	return true;
	110	if (!test_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state))
	111	return true;
	112
	113	/*
	114	* Don't try dividing an ant
	115	*/
	116	if (bt->depth == 1)
	117	return true;
	118
	119	users = atomic_read(&hctx->tags->active_queues);
	120	if (!users)
	121	return true;
	122
	123	/*
	124	* Allow at least some tags
	125	*/
	126	depth = max((bt->depth + users - 1) / users, 4U);
	127	return atomic_read(&hctx->nr_active) < depth;
	128	}
	129
e93ecf60	130	static int __bt_get_word(struct blk_align_bitmap *bm, unsigned int last_tag)
4bb659b1 JA	131	{
	132	int tag, org_last_tag, end;
	133
59d13bf5	134	org_last_tag = last_tag;
4bb659b1 JA	135	end = bm->depth;
	136	do {
	137	restart:
	138	tag = find_next_zero_bit(&bm->word, end, last_tag);
	139	if (unlikely(tag >= end)) {
	140	/*
	141	* We started with an offset, start from 0 to
	142	* exhaust the map.
	143	*/
	144	if (org_last_tag && last_tag) {
	145	end = last_tag;
	146	last_tag = 0;
	147	goto restart;
	148	}
	149	return -1;
	150	}
	151	last_tag = tag + 1;
	152	} while (test_and_set_bit_lock(tag, &bm->word));
	153
	154	return tag;
	155	}
	156
	157	/*
	158	* Straight forward bitmap tag implementation, where each bit is a tag
	159	* (cleared == free, and set == busy). The small twist is using per-cpu
	160	* last_tag caches, which blk-mq stores in the blk_mq_ctx software queue
	161	* contexts. This enables us to drastically limit the space searched,
	162	* without dirtying an extra shared cacheline like we would if we stored
	163	* the cache value inside the shared blk_mq_bitmap_tags structure. On top
	164	* of that, each word of tags is in a separate cacheline. This means that
	165	* multiple users will tend to stick to different cachelines, at least
	166	* until the map is exhausted.
	167	*/
0d2602ca JA	168	static int __bt_get(struct blk_mq_hw_ctx hctx, struct blk_mq_bitmap_tags bt,
0d2602ca JA	169	unsigned int *tag_cache)
4bb659b1 JA	170	{
	171	unsigned int last_tag, org_last_tag;
	172	int index, i, tag;
	173
0d2602ca JA	174	if (!hctx_may_queue(hctx, bt))
	175	return -1;
	176
4bb659b1	177	last_tag = org_last_tag = *tag_cache;
59d13bf5	178	index = TAG_TO_INDEX(bt, last_tag);
4bb659b1 JA	179
4bb659b1 JA	180	for (i = 0; i < bt->map_nr; i++) {
59d13bf5	181	tag = __bt_get_word(&bt->map[index], TAG_TO_BIT(bt, last_tag));
4bb659b1	182	if (tag != -1) {
59d13bf5	183	tag += (index << bt->bits_per_word);
4bb659b1 JA	184	goto done;
	185	}
	186
	187	last_tag = 0;
	188	if (++index >= bt->map_nr)
	189	index = 0;
	190	}
	191
	192	*tag_cache = 0;
	193	return -1;
	194
	195	/*
	196	* Only update the cache from the allocation path, if we ended
	197	* up using the specific cached tag.
	198	*/
	199	done:
	200	if (tag == org_last_tag) {
	201	last_tag = tag + 1;
	202	if (last_tag >= bt->depth - 1)
	203	last_tag = 0;
	204
	205	*tag_cache = last_tag;
	206	}
	207
	208	return tag;
	209	}
	210
4bb659b1 JA	211	static struct bt_wait_state bt_wait_ptr(struct blk_mq_bitmap_tags bt,
	212	struct blk_mq_hw_ctx *hctx)
	213	{
	214	struct bt_wait_state *bs;
	215
	216	if (!hctx)
	217	return &bt->bs[0];
	218
	219	bs = &bt->bs[hctx->wait_index];
	220	bt_index_inc(&hctx->wait_index);
	221	return bs;
320ae51f JA	222	}
320ae51f JA	223
cb96a42c ML	224	static int bt_get(struct blk_mq_alloc_data *data,
	225	struct blk_mq_bitmap_tags *bt,
	226	struct blk_mq_hw_ctx *hctx,
	227	unsigned int *last_tag)
320ae51f	228	{
4bb659b1 JA	229	struct bt_wait_state *bs;
4bb659b1 JA	230	DEFINE_WAIT(wait);
320ae51f JA	231	int tag;
320ae51f JA	232
0d2602ca	233	tag = __bt_get(hctx, bt, last_tag);
4bb659b1 JA	234	if (tag != -1)
	235	return tag;
	236
cb96a42c	237	if (!(data->gfp & __GFP_WAIT))
4bb659b1 JA	238	return -1;
	239
	240	bs = bt_wait_ptr(bt, hctx);
	241	do {
	242	bool was_empty;
	243
	244	was_empty = list_empty(&wait.task_list);
	245	prepare_to_wait(&bs->wait, &wait, TASK_UNINTERRUPTIBLE);
	246
0d2602ca	247	tag = __bt_get(hctx, bt, last_tag);
4bb659b1 JA	248	if (tag != -1)
	249	break;
	250
	251	if (was_empty)
	252	atomic_set(&bs->wait_cnt, bt->wake_cnt);
	253
cb96a42c ML	254	blk_mq_put_ctx(data->ctx);
cb96a42c ML	255
4bb659b1	256	io_schedule();
cb96a42c ML	257
	258	data->ctx = blk_mq_get_ctx(data->q);
	259	data->hctx = data->q->mq_ops->map_queue(data->q,
	260	data->ctx->cpu);
	261	if (data->reserved) {
	262	bt = &data->hctx->tags->breserved_tags;
	263	} else {
	264	last_tag = &data->ctx->last_tag;
	265	hctx = data->hctx;
	266	bt = &hctx->tags->bitmap_tags;
	267	}
	268	finish_wait(&bs->wait, &wait);
	269	bs = bt_wait_ptr(bt, hctx);
4bb659b1 JA	270	} while (1);
	271
	272	finish_wait(&bs->wait, &wait);
	273	return tag;
	274	}
	275
cb96a42c	276	static unsigned int __blk_mq_get_tag(struct blk_mq_alloc_data *data)
4bb659b1 JA	277	{
	278	int tag;
	279
cb96a42c ML	280	tag = bt_get(data, &data->hctx->tags->bitmap_tags, data->hctx,
cb96a42c ML	281	&data->ctx->last_tag);
4bb659b1	282	if (tag >= 0)
cb96a42c	283	return tag + data->hctx->tags->nr_reserved_tags;
4bb659b1 JA	284
4bb659b1 JA	285	return BLK_MQ_TAG_FAIL;
320ae51f JA	286	}
320ae51f JA	287
cb96a42c	288	static unsigned int __blk_mq_get_reserved_tag(struct blk_mq_alloc_data *data)
320ae51f	289	{
4bb659b1	290	int tag, zero = 0;
320ae51f	291
cb96a42c	292	if (unlikely(!data->hctx->tags->nr_reserved_tags)) {
320ae51f JA	293	WARN_ON_ONCE(1);
	294	return BLK_MQ_TAG_FAIL;
	295	}
	296
cb96a42c	297	tag = bt_get(data, &data->hctx->tags->breserved_tags, NULL, &zero);
320ae51f JA	298	if (tag < 0)
320ae51f JA	299	return BLK_MQ_TAG_FAIL;
4bb659b1	300
320ae51f JA	301	return tag;
	302	}
	303
cb96a42c	304	unsigned int blk_mq_get_tag(struct blk_mq_alloc_data *data)
320ae51f	305	{
cb96a42c ML	306	if (!data->reserved)
cb96a42c ML	307	return __blk_mq_get_tag(data);
320ae51f	308
cb96a42c	309	return __blk_mq_get_reserved_tag(data);
320ae51f JA	310	}
320ae51f JA	311
4bb659b1 JA	312	static struct bt_wait_state bt_wake_ptr(struct blk_mq_bitmap_tags bt)
	313	{
	314	int i, wake_index;
	315
	316	wake_index = bt->wake_index;
	317	for (i = 0; i < BT_WAIT_QUEUES; i++) {
	318	struct bt_wait_state *bs = &bt->bs[wake_index];
	319
	320	if (waitqueue_active(&bs->wait)) {
	321	if (wake_index != bt->wake_index)
	322	bt->wake_index = wake_index;
	323
	324	return bs;
	325	}
	326
	327	bt_index_inc(&wake_index);
	328	}
	329
	330	return NULL;
	331	}
	332
	333	static void bt_clear_tag(struct blk_mq_bitmap_tags *bt, unsigned int tag)
	334	{
59d13bf5	335	const int index = TAG_TO_INDEX(bt, tag);
4bb659b1 JA	336	struct bt_wait_state *bs;
4bb659b1 JA	337
0289b2e1 ML	338	/*
	339	* The unlock memory barrier need to order access to req in free
	340	* path and clearing tag bit
	341	*/
	342	clear_bit_unlock(TAG_TO_BIT(bt, tag), &bt->map[index].word);
4bb659b1 JA	343
	344	bs = bt_wake_ptr(bt);
	345	if (bs && atomic_dec_and_test(&bs->wait_cnt)) {
4bb659b1 JA	346	atomic_set(&bs->wait_cnt, bt->wake_cnt);
	347	bt_index_inc(&bt->wake_index);
	348	wake_up(&bs->wait);
	349	}
	350	}
	351
320ae51f JA	352	static void __blk_mq_put_tag(struct blk_mq_tags *tags, unsigned int tag)
	353	{
	354	BUG_ON(tag >= tags->nr_tags);
	355
4bb659b1	356	bt_clear_tag(&tags->bitmap_tags, tag);
320ae51f JA	357	}
	358
	359	static void __blk_mq_put_reserved_tag(struct blk_mq_tags *tags,
	360	unsigned int tag)
	361	{
	362	BUG_ON(tag >= tags->nr_reserved_tags);
	363
4bb659b1	364	bt_clear_tag(&tags->breserved_tags, tag);
320ae51f JA	365	}
320ae51f JA	366
0d2602ca	367	void blk_mq_put_tag(struct blk_mq_hw_ctx *hctx, unsigned int tag,
4bb659b1	368	unsigned int *last_tag)
320ae51f	369	{
0d2602ca JA	370	struct blk_mq_tags *tags = hctx->tags;
0d2602ca JA	371
4bb659b1 JA	372	if (tag >= tags->nr_reserved_tags) {
	373	const int real_tag = tag - tags->nr_reserved_tags;
	374
	375	__blk_mq_put_tag(tags, real_tag);
	376	*last_tag = real_tag;
	377	} else
320ae51f JA	378	__blk_mq_put_reserved_tag(tags, tag);
	379	}
	380
4bb659b1 JA	381	static void bt_for_each_free(struct blk_mq_bitmap_tags *bt,
4bb659b1 JA	382	unsigned long *free_map, unsigned int off)
320ae51f	383	{
4bb659b1 JA	384	int i;
	385
	386	for (i = 0; i < bt->map_nr; i++) {
e93ecf60	387	struct blk_align_bitmap *bm = &bt->map[i];
4bb659b1 JA	388	int bit = 0;
	389
	390	do {
	391	bit = find_next_zero_bit(&bm->word, bm->depth, bit);
	392	if (bit >= bm->depth)
	393	break;
	394
	395	__set_bit(bit + off, free_map);
	396	bit++;
	397	} while (1);
	398
59d13bf5	399	off += (1 << bt->bits_per_word);
4bb659b1	400	}
320ae51f JA	401	}
	402
	403	void blk_mq_tag_busy_iter(struct blk_mq_tags *tags,
	404	void (fn)(void , unsigned long ), void data)
	405	{
	406	unsigned long *tag_map;
	407	size_t map_size;
	408
	409	map_size = ALIGN(tags->nr_tags, BITS_PER_LONG) / BITS_PER_LONG;
	410	tag_map = kzalloc(map_size * sizeof(unsigned long), GFP_ATOMIC);
	411	if (!tag_map)
	412	return;
	413
4bb659b1	414	bt_for_each_free(&tags->bitmap_tags, tag_map, tags->nr_reserved_tags);
320ae51f	415	if (tags->nr_reserved_tags)
4bb659b1	416	bt_for_each_free(&tags->breserved_tags, tag_map, 0);
320ae51f JA	417
	418	fn(data, tag_map);
	419	kfree(tag_map);
	420	}
edf866b3	421	EXPORT_SYMBOL(blk_mq_tag_busy_iter);
320ae51f	422
4bb659b1 JA	423	static unsigned int bt_unused_tags(struct blk_mq_bitmap_tags *bt)
	424	{
	425	unsigned int i, used;
	426
	427	for (i = 0, used = 0; i < bt->map_nr; i++) {
e93ecf60	428	struct blk_align_bitmap *bm = &bt->map[i];
4bb659b1 JA	429
	430	used += bitmap_weight(&bm->word, bm->depth);
	431	}
	432
	433	return bt->depth - used;
	434	}
	435
e3a2b3f9 JA	436	static void bt_update_count(struct blk_mq_bitmap_tags *bt,
	437	unsigned int depth)
	438	{
	439	unsigned int tags_per_word = 1U << bt->bits_per_word;
	440	unsigned int map_depth = depth;
	441
	442	if (depth) {
	443	int i;
	444
	445	for (i = 0; i < bt->map_nr; i++) {
	446	bt->map[i].depth = min(map_depth, tags_per_word);
	447	map_depth -= bt->map[i].depth;
	448	}
	449	}
	450
	451	bt->wake_cnt = BT_WAIT_BATCH;
	452	if (bt->wake_cnt > depth / 4)
	453	bt->wake_cnt = max(1U, depth / 4);
	454
	455	bt->depth = depth;
	456	}
	457
4bb659b1 JA	458	static int bt_alloc(struct blk_mq_bitmap_tags *bt, unsigned int depth,
	459	int node, bool reserved)
	460	{
	461	int i;
	462
59d13bf5 JA	463	bt->bits_per_word = ilog2(BITS_PER_LONG);
59d13bf5 JA	464
4bb659b1 JA	465	/*
	466	* Depth can be zero for reserved tags, that's not a failure
	467	* condition.
	468	*/
	469	if (depth) {
e3a2b3f9	470	unsigned int nr, tags_per_word;
59d13bf5 JA	471
	472	tags_per_word = (1 << bt->bits_per_word);
	473
	474	/*
	475	* If the tag space is small, shrink the number of tags
	476	* per word so we spread over a few cachelines, at least.
	477	* If less than 4 tags, just forget about it, it's not
	478	* going to work optimally anyway.
	479	*/
	480	if (depth >= 4) {
	481	while (tags_per_word * 4 > depth) {
	482	bt->bits_per_word--;
	483	tags_per_word = (1 << bt->bits_per_word);
	484	}
	485	}
4bb659b1	486
59d13bf5	487	nr = ALIGN(depth, tags_per_word) / tags_per_word;
e93ecf60	488	bt->map = kzalloc_node(nr * sizeof(struct blk_align_bitmap),
4bb659b1 JA	489	GFP_KERNEL, node);
	490	if (!bt->map)
	491	return -ENOMEM;
	492
	493	bt->map_nr = nr;
4bb659b1 JA	494	}
	495
	496	bt->bs = kzalloc(BT_WAIT_QUEUES * sizeof(*bt->bs), GFP_KERNEL);
	497	if (!bt->bs) {
	498	kfree(bt->map);
	499	return -ENOMEM;
	500	}
	501
	502	for (i = 0; i < BT_WAIT_QUEUES; i++)
	503	init_waitqueue_head(&bt->bs[i].wait);
	504
e3a2b3f9	505	bt_update_count(bt, depth);
4bb659b1 JA	506	return 0;
	507	}
	508
	509	static void bt_free(struct blk_mq_bitmap_tags *bt)
	510	{
	511	kfree(bt->map);
	512	kfree(bt->bs);
	513	}
	514
	515	static struct blk_mq_tags blk_mq_init_bitmap_tags(struct blk_mq_tags tags,
	516	int node)
	517	{
	518	unsigned int depth = tags->nr_tags - tags->nr_reserved_tags;
	519
	520	if (bt_alloc(&tags->bitmap_tags, depth, node, false))
	521	goto enomem;
	522	if (bt_alloc(&tags->breserved_tags, tags->nr_reserved_tags, node, true))
	523	goto enomem;
	524
	525	return tags;
	526	enomem:
	527	bt_free(&tags->bitmap_tags);
	528	kfree(tags);
	529	return NULL;
	530	}
	531
320ae51f JA	532	struct blk_mq_tags *blk_mq_init_tags(unsigned int total_tags,
	533	unsigned int reserved_tags, int node)
	534	{
320ae51f	535	struct blk_mq_tags *tags;
320ae51f JA	536
	537	if (total_tags > BLK_MQ_TAG_MAX) {
	538	pr_err("blk-mq: tag depth too large\n");
	539	return NULL;
	540	}
	541
	542	tags = kzalloc_node(sizeof(*tags), GFP_KERNEL, node);
	543	if (!tags)
	544	return NULL;
	545
320ae51f JA	546	tags->nr_tags = total_tags;
320ae51f JA	547	tags->nr_reserved_tags = reserved_tags;
320ae51f	548
4bb659b1	549	return blk_mq_init_bitmap_tags(tags, node);
320ae51f JA	550	}
	551
	552	void blk_mq_free_tags(struct blk_mq_tags *tags)
	553	{
4bb659b1 JA	554	bt_free(&tags->bitmap_tags);
4bb659b1 JA	555	bt_free(&tags->breserved_tags);
320ae51f JA	556	kfree(tags);
	557	}
	558
4bb659b1 JA	559	void blk_mq_tag_init_last_tag(struct blk_mq_tags tags, unsigned int tag)
	560	{
	561	unsigned int depth = tags->nr_tags - tags->nr_reserved_tags;
	562
9d3d21ae	563	*tag = prandom_u32() % depth;
4bb659b1 JA	564	}
4bb659b1 JA	565
e3a2b3f9 JA	566	int blk_mq_tag_update_depth(struct blk_mq_tags *tags, unsigned int tdepth)
	567	{
	568	tdepth -= tags->nr_reserved_tags;
	569	if (tdepth > tags->nr_tags)
	570	return -EINVAL;
	571
	572	/*
	573	* Don't need (or can't) update reserved tags here, they remain
	574	* static and should never need resizing.
	575	*/
	576	bt_update_count(&tags->bitmap_tags, tdepth);
	577	blk_mq_tag_wakeup_all(tags);
	578	return 0;
	579	}
	580
320ae51f JA	581	ssize_t blk_mq_tag_sysfs_show(struct blk_mq_tags tags, char page)
	582	{
	583	char *orig_page = page;
4bb659b1	584	unsigned int free, res;
320ae51f JA	585
	586	if (!tags)
	587	return 0;
	588
59d13bf5 JA	589	page += sprintf(page, "nr_tags=%u, reserved_tags=%u, "
	590	"bits_per_word=%u\n",
	591	tags->nr_tags, tags->nr_reserved_tags,
	592	tags->bitmap_tags.bits_per_word);
320ae51f	593
4bb659b1 JA	594	free = bt_unused_tags(&tags->bitmap_tags);
4bb659b1 JA	595	res = bt_unused_tags(&tags->breserved_tags);
320ae51f	596
4bb659b1	597	page += sprintf(page, "nr_free=%u, nr_reserved=%u\n", free, res);
0d2602ca	598	page += sprintf(page, "active_queues=%u\n", atomic_read(&tags->active_queues));
320ae51f JA	599
	600	return page - orig_page;
	601	}