[mirror_ubuntu-artful-kernel.git] / block / blk-mq-tag.c

#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"

void blk_mq_wait_for_tags(struct blk_mq_hw_ctx *hctx, bool reserved)
{
	int tag, zero = 0;

	tag = blk_mq_get_tag(hctx, &zero, __GFP_WAIT, reserved);
	blk_mq_put_tag(hctx, tag, &zero);
}

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_bitmap_tags *bt, struct blk_mq_hw_ctx *hctx,
		  unsigned int *last_tag, gfp_t gfp)
{
	struct bt_wait_state *bs;
	DEFINE_WAIT(wait);
	int tag;

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

	if (!(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);

		io_schedule();
	} while (1);

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

static unsigned int __blk_mq_get_tag(struct blk_mq_tags *tags,
				     struct blk_mq_hw_ctx *hctx,
				     unsigned int *last_tag, gfp_t gfp)
{
	int tag;

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

	return BLK_MQ_TAG_FAIL;
}

static unsigned int __blk_mq_get_reserved_tag(struct blk_mq_tags *tags,
					      gfp_t gfp)
{
	int tag, zero = 0;

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

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

	return tag;
}

unsigned int blk_mq_get_tag(struct blk_mq_hw_ctx *hctx, unsigned int *last_tag,
			    gfp_t gfp, bool reserved)
{
	if (!reserved)
		return __blk_mq_get_tag(hctx->tags, hctx, last_tag, gfp);

	return __blk_mq_get_reserved_tag(hctx->tags, gfp);
}

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
320ae51f JA	1	#include <linux/kernel.h>
320ae51f JA	2	#include <linux/module.h>
4bb659b1	3	#include <linux/random.h>
320ae51f JA	4
	5	#include <linux/blk-mq.h>
	6	#include "blk.h"
	7	#include "blk-mq.h"
	8	#include "blk-mq-tag.h"
	9
0d2602ca	10	void blk_mq_wait_for_tags(struct blk_mq_hw_ctx *hctx, bool reserved)
320ae51f	11	{
4bb659b1 JA	12	int tag, zero = 0;
4bb659b1 JA	13
0d2602ca JA	14	tag = blk_mq_get_tag(hctx, &zero, __GFP_WAIT, reserved);
0d2602ca JA	15	blk_mq_put_tag(hctx, tag, &zero);
4bb659b1 JA	16	}
	17
	18	static bool bt_has_free_tags(struct blk_mq_bitmap_tags *bt)
	19	{
	20	int i;
	21
	22	for (i = 0; i < bt->map_nr; i++) {
e93ecf60	23	struct blk_align_bitmap *bm = &bt->map[i];
4bb659b1 JA	24	int ret;
	25
	26	ret = find_first_zero_bit(&bm->word, bm->depth);
	27	if (ret < bm->depth)
	28	return true;
	29	}
	30
	31	return false;
320ae51f JA	32	}
	33
	34	bool blk_mq_has_free_tags(struct blk_mq_tags *tags)
	35	{
4bb659b1 JA	36	if (!tags)
	37	return true;
	38
	39	return bt_has_free_tags(&tags->bitmap_tags);
	40	}
	41
0d2602ca JA	42	static inline void bt_index_inc(unsigned int *index)
	43	{
	44	index = (index + 1) & (BT_WAIT_QUEUES - 1);
	45	}
	46
	47	/*
	48	* If a previously inactive queue goes active, bump the active user count.
	49	*/
	50	bool __blk_mq_tag_busy(struct blk_mq_hw_ctx *hctx)
	51	{
	52	if (!test_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state) &&
	53	!test_and_set_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state))
	54	atomic_inc(&hctx->tags->active_queues);
	55
	56	return true;
	57	}
	58
	59	/*
e3a2b3f9	60	* Wakeup all potentially sleeping on normal (non-reserved) tags
0d2602ca	61	*/
e3a2b3f9	62	static void blk_mq_tag_wakeup_all(struct blk_mq_tags *tags)
0d2602ca	63	{
0d2602ca JA	64	struct blk_mq_bitmap_tags *bt;
	65	int i, wake_index;
	66
0d2602ca JA	67	bt = &tags->bitmap_tags;
	68	wake_index = bt->wake_index;
	69	for (i = 0; i < BT_WAIT_QUEUES; i++) {
	70	struct bt_wait_state *bs = &bt->bs[wake_index];
	71
	72	if (waitqueue_active(&bs->wait))
	73	wake_up(&bs->wait);
	74
	75	bt_index_inc(&wake_index);
	76	}
	77	}
	78
e3a2b3f9 JA	79	/*
	80	* If a previously busy queue goes inactive, potential waiters could now
	81	* be allowed to queue. Wake them up and check.
	82	*/
	83	void __blk_mq_tag_idle(struct blk_mq_hw_ctx *hctx)
	84	{
	85	struct blk_mq_tags *tags = hctx->tags;
	86
	87	if (!test_and_clear_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state))
	88	return;
	89
	90	atomic_dec(&tags->active_queues);
	91
	92	blk_mq_tag_wakeup_all(tags);
	93	}
	94
0d2602ca JA	95	/*
	96	* For shared tag users, we track the number of currently active users
	97	* and attempt to provide a fair share of the tag depth for each of them.
	98	*/
	99	static inline bool hctx_may_queue(struct blk_mq_hw_ctx *hctx,
	100	struct blk_mq_bitmap_tags *bt)
	101	{
	102	unsigned int depth, users;
	103
	104	if (!hctx \|\| !(hctx->flags & BLK_MQ_F_TAG_SHARED))
	105	return true;
	106	if (!test_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state))
	107	return true;
	108
	109	/*
	110	* Don't try dividing an ant
	111	*/
	112	if (bt->depth == 1)
	113	return true;
	114
	115	users = atomic_read(&hctx->tags->active_queues);
	116	if (!users)
	117	return true;
	118
	119	/*
	120	* Allow at least some tags
	121	*/
	122	depth = max((bt->depth + users - 1) / users, 4U);
	123	return atomic_read(&hctx->nr_active) < depth;
	124	}
	125
e93ecf60	126	static int __bt_get_word(struct blk_align_bitmap *bm, unsigned int last_tag)
4bb659b1 JA	127	{
	128	int tag, org_last_tag, end;
	129
59d13bf5	130	org_last_tag = last_tag;
4bb659b1 JA	131	end = bm->depth;
	132	do {
	133	restart:
	134	tag = find_next_zero_bit(&bm->word, end, last_tag);
	135	if (unlikely(tag >= end)) {
	136	/*
	137	* We started with an offset, start from 0 to
	138	* exhaust the map.
	139	*/
	140	if (org_last_tag && last_tag) {
	141	end = last_tag;
	142	last_tag = 0;
	143	goto restart;
	144	}
	145	return -1;
	146	}
	147	last_tag = tag + 1;
	148	} while (test_and_set_bit_lock(tag, &bm->word));
	149
	150	return tag;
	151	}
	152
	153	/*
	154	* Straight forward bitmap tag implementation, where each bit is a tag
	155	* (cleared == free, and set == busy). The small twist is using per-cpu
	156	* last_tag caches, which blk-mq stores in the blk_mq_ctx software queue
	157	* contexts. This enables us to drastically limit the space searched,
	158	* without dirtying an extra shared cacheline like we would if we stored
	159	* the cache value inside the shared blk_mq_bitmap_tags structure. On top
	160	* of that, each word of tags is in a separate cacheline. This means that
	161	* multiple users will tend to stick to different cachelines, at least
	162	* until the map is exhausted.
	163	*/
0d2602ca JA	164	static int __bt_get(struct blk_mq_hw_ctx hctx, struct blk_mq_bitmap_tags bt,
0d2602ca JA	165	unsigned int *tag_cache)
4bb659b1 JA	166	{
	167	unsigned int last_tag, org_last_tag;
	168	int index, i, tag;
	169
0d2602ca JA	170	if (!hctx_may_queue(hctx, bt))
	171	return -1;
	172
4bb659b1	173	last_tag = org_last_tag = *tag_cache;
59d13bf5	174	index = TAG_TO_INDEX(bt, last_tag);
4bb659b1 JA	175
4bb659b1 JA	176	for (i = 0; i < bt->map_nr; i++) {
59d13bf5	177	tag = __bt_get_word(&bt->map[index], TAG_TO_BIT(bt, last_tag));
4bb659b1	178	if (tag != -1) {
59d13bf5	179	tag += (index << bt->bits_per_word);
4bb659b1 JA	180	goto done;
	181	}
	182
	183	last_tag = 0;
	184	if (++index >= bt->map_nr)
	185	index = 0;
	186	}
	187
	188	*tag_cache = 0;
	189	return -1;
	190
	191	/*
	192	* Only update the cache from the allocation path, if we ended
	193	* up using the specific cached tag.
	194	*/
	195	done:
	196	if (tag == org_last_tag) {
	197	last_tag = tag + 1;
	198	if (last_tag >= bt->depth - 1)
	199	last_tag = 0;
	200
	201	*tag_cache = last_tag;
	202	}
	203
	204	return tag;
	205	}
	206
4bb659b1 JA	207	static struct bt_wait_state bt_wait_ptr(struct blk_mq_bitmap_tags bt,
	208	struct blk_mq_hw_ctx *hctx)
	209	{
	210	struct bt_wait_state *bs;
	211
	212	if (!hctx)
	213	return &bt->bs[0];
	214
	215	bs = &bt->bs[hctx->wait_index];
	216	bt_index_inc(&hctx->wait_index);
	217	return bs;
320ae51f JA	218	}
320ae51f JA	219
4bb659b1 JA	220	static int bt_get(struct blk_mq_bitmap_tags bt, struct blk_mq_hw_ctx hctx,
4bb659b1 JA	221	unsigned int *last_tag, gfp_t gfp)
320ae51f	222	{
4bb659b1 JA	223	struct bt_wait_state *bs;
4bb659b1 JA	224	DEFINE_WAIT(wait);
320ae51f JA	225	int tag;
320ae51f JA	226
0d2602ca	227	tag = __bt_get(hctx, bt, last_tag);
4bb659b1 JA	228	if (tag != -1)
	229	return tag;
	230
	231	if (!(gfp & __GFP_WAIT))
	232	return -1;
	233
	234	bs = bt_wait_ptr(bt, hctx);
	235	do {
	236	bool was_empty;
	237
	238	was_empty = list_empty(&wait.task_list);
	239	prepare_to_wait(&bs->wait, &wait, TASK_UNINTERRUPTIBLE);
	240
0d2602ca	241	tag = __bt_get(hctx, bt, last_tag);
4bb659b1 JA	242	if (tag != -1)
	243	break;
	244
	245	if (was_empty)
	246	atomic_set(&bs->wait_cnt, bt->wake_cnt);
	247
	248	io_schedule();
	249	} while (1);
	250
	251	finish_wait(&bs->wait, &wait);
	252	return tag;
	253	}
	254
	255	static unsigned int __blk_mq_get_tag(struct blk_mq_tags *tags,
	256	struct blk_mq_hw_ctx *hctx,
	257	unsigned int *last_tag, gfp_t gfp)
	258	{
	259	int tag;
	260
	261	tag = bt_get(&tags->bitmap_tags, hctx, last_tag, gfp);
	262	if (tag >= 0)
	263	return tag + tags->nr_reserved_tags;
	264
	265	return BLK_MQ_TAG_FAIL;
320ae51f JA	266	}
	267
	268	static unsigned int __blk_mq_get_reserved_tag(struct blk_mq_tags *tags,
	269	gfp_t gfp)
	270	{
4bb659b1	271	int tag, zero = 0;
320ae51f JA	272
	273	if (unlikely(!tags->nr_reserved_tags)) {
	274	WARN_ON_ONCE(1);
	275	return BLK_MQ_TAG_FAIL;
	276	}
	277
4bb659b1	278	tag = bt_get(&tags->breserved_tags, NULL, &zero, gfp);
320ae51f JA	279	if (tag < 0)
320ae51f JA	280	return BLK_MQ_TAG_FAIL;
4bb659b1	281
320ae51f JA	282	return tag;
	283	}
	284
0d2602ca	285	unsigned int blk_mq_get_tag(struct blk_mq_hw_ctx hctx, unsigned int last_tag,
4bb659b1	286	gfp_t gfp, bool reserved)
320ae51f JA	287	{
320ae51f JA	288	if (!reserved)
0d2602ca	289	return __blk_mq_get_tag(hctx->tags, hctx, last_tag, gfp);
320ae51f	290
0d2602ca	291	return __blk_mq_get_reserved_tag(hctx->tags, gfp);
320ae51f JA	292	}
320ae51f JA	293
4bb659b1 JA	294	static struct bt_wait_state bt_wake_ptr(struct blk_mq_bitmap_tags bt)
	295	{
	296	int i, wake_index;
	297
	298	wake_index = bt->wake_index;
	299	for (i = 0; i < BT_WAIT_QUEUES; i++) {
	300	struct bt_wait_state *bs = &bt->bs[wake_index];
	301
	302	if (waitqueue_active(&bs->wait)) {
	303	if (wake_index != bt->wake_index)
	304	bt->wake_index = wake_index;
	305
	306	return bs;
	307	}
	308
	309	bt_index_inc(&wake_index);
	310	}
	311
	312	return NULL;
	313	}
	314
	315	static void bt_clear_tag(struct blk_mq_bitmap_tags *bt, unsigned int tag)
	316	{
59d13bf5	317	const int index = TAG_TO_INDEX(bt, tag);
4bb659b1 JA	318	struct bt_wait_state *bs;
4bb659b1 JA	319
0289b2e1 ML	320	/*
	321	* The unlock memory barrier need to order access to req in free
	322	* path and clearing tag bit
	323	*/
	324	clear_bit_unlock(TAG_TO_BIT(bt, tag), &bt->map[index].word);
4bb659b1 JA	325
	326	bs = bt_wake_ptr(bt);
	327	if (bs && atomic_dec_and_test(&bs->wait_cnt)) {
4bb659b1 JA	328	atomic_set(&bs->wait_cnt, bt->wake_cnt);
	329	bt_index_inc(&bt->wake_index);
	330	wake_up(&bs->wait);
	331	}
	332	}
	333
320ae51f JA	334	static void __blk_mq_put_tag(struct blk_mq_tags *tags, unsigned int tag)
	335	{
	336	BUG_ON(tag >= tags->nr_tags);
	337
4bb659b1	338	bt_clear_tag(&tags->bitmap_tags, tag);
320ae51f JA	339	}
	340
	341	static void __blk_mq_put_reserved_tag(struct blk_mq_tags *tags,
	342	unsigned int tag)
	343	{
	344	BUG_ON(tag >= tags->nr_reserved_tags);
	345
4bb659b1	346	bt_clear_tag(&tags->breserved_tags, tag);
320ae51f JA	347	}
320ae51f JA	348
0d2602ca	349	void blk_mq_put_tag(struct blk_mq_hw_ctx *hctx, unsigned int tag,
4bb659b1	350	unsigned int *last_tag)
320ae51f	351	{
0d2602ca JA	352	struct blk_mq_tags *tags = hctx->tags;
0d2602ca JA	353
4bb659b1 JA	354	if (tag >= tags->nr_reserved_tags) {
	355	const int real_tag = tag - tags->nr_reserved_tags;
	356
	357	__blk_mq_put_tag(tags, real_tag);
	358	*last_tag = real_tag;
	359	} else
320ae51f JA	360	__blk_mq_put_reserved_tag(tags, tag);
	361	}
	362
4bb659b1 JA	363	static void bt_for_each_free(struct blk_mq_bitmap_tags *bt,
4bb659b1 JA	364	unsigned long *free_map, unsigned int off)
320ae51f	365	{
4bb659b1 JA	366	int i;
	367
	368	for (i = 0; i < bt->map_nr; i++) {
e93ecf60	369	struct blk_align_bitmap *bm = &bt->map[i];
4bb659b1 JA	370	int bit = 0;
	371
	372	do {
	373	bit = find_next_zero_bit(&bm->word, bm->depth, bit);
	374	if (bit >= bm->depth)
	375	break;
	376
	377	__set_bit(bit + off, free_map);
	378	bit++;
	379	} while (1);
	380
59d13bf5	381	off += (1 << bt->bits_per_word);
4bb659b1	382	}
320ae51f JA	383	}
	384
	385	void blk_mq_tag_busy_iter(struct blk_mq_tags *tags,
	386	void (fn)(void , unsigned long ), void data)
	387	{
	388	unsigned long *tag_map;
	389	size_t map_size;
	390
	391	map_size = ALIGN(tags->nr_tags, BITS_PER_LONG) / BITS_PER_LONG;
	392	tag_map = kzalloc(map_size * sizeof(unsigned long), GFP_ATOMIC);
	393	if (!tag_map)
	394	return;
	395
4bb659b1	396	bt_for_each_free(&tags->bitmap_tags, tag_map, tags->nr_reserved_tags);
320ae51f	397	if (tags->nr_reserved_tags)
4bb659b1	398	bt_for_each_free(&tags->breserved_tags, tag_map, 0);
320ae51f JA	399
	400	fn(data, tag_map);
	401	kfree(tag_map);
	402	}
edf866b3	403	EXPORT_SYMBOL(blk_mq_tag_busy_iter);
320ae51f	404
4bb659b1 JA	405	static unsigned int bt_unused_tags(struct blk_mq_bitmap_tags *bt)
	406	{
	407	unsigned int i, used;
	408
	409	for (i = 0, used = 0; i < bt->map_nr; i++) {
e93ecf60	410	struct blk_align_bitmap *bm = &bt->map[i];
4bb659b1 JA	411
	412	used += bitmap_weight(&bm->word, bm->depth);
	413	}
	414
	415	return bt->depth - used;
	416	}
	417
e3a2b3f9 JA	418	static void bt_update_count(struct blk_mq_bitmap_tags *bt,
	419	unsigned int depth)
	420	{
	421	unsigned int tags_per_word = 1U << bt->bits_per_word;
	422	unsigned int map_depth = depth;
	423
	424	if (depth) {
	425	int i;
	426
	427	for (i = 0; i < bt->map_nr; i++) {
	428	bt->map[i].depth = min(map_depth, tags_per_word);
	429	map_depth -= bt->map[i].depth;
	430	}
	431	}
	432
	433	bt->wake_cnt = BT_WAIT_BATCH;
	434	if (bt->wake_cnt > depth / 4)
	435	bt->wake_cnt = max(1U, depth / 4);
	436
	437	bt->depth = depth;
	438	}
	439
4bb659b1 JA	440	static int bt_alloc(struct blk_mq_bitmap_tags *bt, unsigned int depth,
	441	int node, bool reserved)
	442	{
	443	int i;
	444
59d13bf5 JA	445	bt->bits_per_word = ilog2(BITS_PER_LONG);
59d13bf5 JA	446
4bb659b1 JA	447	/*
	448	* Depth can be zero for reserved tags, that's not a failure
	449	* condition.
	450	*/
	451	if (depth) {
e3a2b3f9	452	unsigned int nr, tags_per_word;
59d13bf5 JA	453
	454	tags_per_word = (1 << bt->bits_per_word);
	455
	456	/*
	457	* If the tag space is small, shrink the number of tags
	458	* per word so we spread over a few cachelines, at least.
	459	* If less than 4 tags, just forget about it, it's not
	460	* going to work optimally anyway.
	461	*/
	462	if (depth >= 4) {
	463	while (tags_per_word * 4 > depth) {
	464	bt->bits_per_word--;
	465	tags_per_word = (1 << bt->bits_per_word);
	466	}
	467	}
4bb659b1	468
59d13bf5	469	nr = ALIGN(depth, tags_per_word) / tags_per_word;
e93ecf60	470	bt->map = kzalloc_node(nr * sizeof(struct blk_align_bitmap),
4bb659b1 JA	471	GFP_KERNEL, node);
	472	if (!bt->map)
	473	return -ENOMEM;
	474
	475	bt->map_nr = nr;
4bb659b1 JA	476	}
	477
	478	bt->bs = kzalloc(BT_WAIT_QUEUES * sizeof(*bt->bs), GFP_KERNEL);
	479	if (!bt->bs) {
	480	kfree(bt->map);
	481	return -ENOMEM;
	482	}
	483
	484	for (i = 0; i < BT_WAIT_QUEUES; i++)
	485	init_waitqueue_head(&bt->bs[i].wait);
	486
e3a2b3f9	487	bt_update_count(bt, depth);
4bb659b1 JA	488	return 0;
	489	}
	490
	491	static void bt_free(struct blk_mq_bitmap_tags *bt)
	492	{
	493	kfree(bt->map);
	494	kfree(bt->bs);
	495	}
	496
	497	static struct blk_mq_tags blk_mq_init_bitmap_tags(struct blk_mq_tags tags,
	498	int node)
	499	{
	500	unsigned int depth = tags->nr_tags - tags->nr_reserved_tags;
	501
	502	if (bt_alloc(&tags->bitmap_tags, depth, node, false))
	503	goto enomem;
	504	if (bt_alloc(&tags->breserved_tags, tags->nr_reserved_tags, node, true))
	505	goto enomem;
	506
	507	return tags;
	508	enomem:
	509	bt_free(&tags->bitmap_tags);
	510	kfree(tags);
	511	return NULL;
	512	}
	513
320ae51f JA	514	struct blk_mq_tags *blk_mq_init_tags(unsigned int total_tags,
	515	unsigned int reserved_tags, int node)
	516	{
320ae51f	517	struct blk_mq_tags *tags;
320ae51f JA	518
	519	if (total_tags > BLK_MQ_TAG_MAX) {
	520	pr_err("blk-mq: tag depth too large\n");
	521	return NULL;
	522	}
	523
	524	tags = kzalloc_node(sizeof(*tags), GFP_KERNEL, node);
	525	if (!tags)
	526	return NULL;
	527
320ae51f JA	528	tags->nr_tags = total_tags;
320ae51f JA	529	tags->nr_reserved_tags = reserved_tags;
320ae51f	530
4bb659b1	531	return blk_mq_init_bitmap_tags(tags, node);
320ae51f JA	532	}
	533
	534	void blk_mq_free_tags(struct blk_mq_tags *tags)
	535	{
4bb659b1 JA	536	bt_free(&tags->bitmap_tags);
4bb659b1 JA	537	bt_free(&tags->breserved_tags);
320ae51f JA	538	kfree(tags);
	539	}
	540
4bb659b1 JA	541	void blk_mq_tag_init_last_tag(struct blk_mq_tags tags, unsigned int tag)
	542	{
	543	unsigned int depth = tags->nr_tags - tags->nr_reserved_tags;
	544
9d3d21ae	545	*tag = prandom_u32() % depth;
4bb659b1 JA	546	}
4bb659b1 JA	547
e3a2b3f9 JA	548	int blk_mq_tag_update_depth(struct blk_mq_tags *tags, unsigned int tdepth)
	549	{
	550	tdepth -= tags->nr_reserved_tags;
	551	if (tdepth > tags->nr_tags)
	552	return -EINVAL;
	553
	554	/*
	555	* Don't need (or can't) update reserved tags here, they remain
	556	* static and should never need resizing.
	557	*/
	558	bt_update_count(&tags->bitmap_tags, tdepth);
	559	blk_mq_tag_wakeup_all(tags);
	560	return 0;
	561	}
	562
320ae51f JA	563	ssize_t blk_mq_tag_sysfs_show(struct blk_mq_tags tags, char page)
	564	{
	565	char *orig_page = page;
4bb659b1	566	unsigned int free, res;
320ae51f JA	567
	568	if (!tags)
	569	return 0;
	570
59d13bf5 JA	571	page += sprintf(page, "nr_tags=%u, reserved_tags=%u, "
	572	"bits_per_word=%u\n",
	573	tags->nr_tags, tags->nr_reserved_tags,
	574	tags->bitmap_tags.bits_per_word);
320ae51f	575
4bb659b1 JA	576	free = bt_unused_tags(&tags->bitmap_tags);
4bb659b1 JA	577	res = bt_unused_tags(&tags->breserved_tags);
320ae51f	578
4bb659b1	579	page += sprintf(page, "nr_free=%u, nr_reserved=%u\n", free, res);
0d2602ca	580	page += sprintf(page, "active_queues=%u\n", atomic_read(&tags->active_queues));
320ae51f JA	581
	582	return page - orig_page;
	583	}