[linux-2.6-block.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"

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