[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"

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

	tag = blk_mq_get_tag(tags, hctx, &zero, __GFP_WAIT, reserved);
	blk_mq_put_tag(tags, 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_mq_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 int __bt_get_word(struct blk_mq_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_bitmap_tags *bt, unsigned int *tag_cache)
{
	unsigned int last_tag, org_last_tag;
	int index, i, tag;

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

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

	return __blk_mq_get_reserved_tag(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_tags *tags, unsigned int tag,
		    unsigned int *last_tag)
{
	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_mq_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);
}

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_mq_bitmap *bm = &bt->map[i];

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

	return bt->depth - used;
}

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, i, map_depth, 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_mq_bitmap),
						GFP_KERNEL, node);
		if (!bt->map)
			return -ENOMEM;

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

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

	bt->depth = 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;
}

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);

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