xref: /linux/drivers/md/bcache/bset.c (revision d56d000a1f424aa77538bd5aad18b43037ed20cc)

/*
 * Code for working with individual keys, and sorted sets of keys with in a
 * btree node
 *
 * Copyright 2012 Google, Inc.
 */

#include "bcache.h"
#include "btree.h"
#include "debug.h"

#include <linux/random.h>
#include <linux/prefetch.h>

/* Keylists */

int bch_keylist_realloc(struct keylist *l, int nptrs, struct cache_set *c)
{
	size_t oldsize = bch_keylist_nkeys(l);
	size_t newsize = oldsize + 2 + nptrs;
	uint64_t *old_keys = l->keys_p == l->inline_keys ? NULL : l->keys_p;
	uint64_t *new_keys;

	/* The journalling code doesn't handle the case where the keys to insert
	 * is bigger than an empty write: If we just return -ENOMEM here,
	 * bio_insert() and bio_invalidate() will insert the keys created so far
	 * and finish the rest when the keylist is empty.
	 */
	if (newsize * sizeof(uint64_t) > block_bytes(c) - sizeof(struct jset))
		return -ENOMEM;

	newsize = roundup_pow_of_two(newsize);

	if (newsize <= KEYLIST_INLINE ||
	    roundup_pow_of_two(oldsize) == newsize)
		return 0;

	new_keys = krealloc(old_keys, sizeof(uint64_t) * newsize, GFP_NOIO);

	if (!new_keys)
		return -ENOMEM;

	if (!old_keys)
		memcpy(new_keys, l->inline_keys, sizeof(uint64_t) * oldsize);

	l->keys_p = new_keys;
	l->top_p = new_keys + oldsize;

	return 0;
}

struct bkey *bch_keylist_pop(struct keylist *l)
{
	struct bkey *k = l->keys;

	if (k == l->top)
		return NULL;

	while (bkey_next(k) != l->top)
		k = bkey_next(k);

	return l->top = k;
}

void bch_keylist_pop_front(struct keylist *l)
{
	l->top_p -= bkey_u64s(l->keys);

	memmove(l->keys,
		bkey_next(l->keys),
		bch_keylist_bytes(l));
}

/* Pointer validation */

static bool __ptr_invalid(struct cache_set *c, const struct bkey *k)
{
	unsigned i;

	for (i = 0; i < KEY_PTRS(k); i++)
		if (ptr_available(c, k, i)) {
			struct cache *ca = PTR_CACHE(c, k, i);
			size_t bucket = PTR_BUCKET_NR(c, k, i);
			size_t r = bucket_remainder(c, PTR_OFFSET(k, i));

			if (KEY_SIZE(k) + r > c->sb.bucket_size ||
			    bucket <  ca->sb.first_bucket ||
			    bucket >= ca->sb.nbuckets)
				return true;
		}

	return false;
}

bool bch_btree_ptr_invalid(struct cache_set *c, const struct bkey *k)
{
	char buf[80];

	if (!KEY_PTRS(k) || !KEY_SIZE(k) || KEY_DIRTY(k))
		goto bad;

	if (__ptr_invalid(c, k))
		goto bad;

	return false;
bad:
	bch_bkey_to_text(buf, sizeof(buf), k);
	cache_bug(c, "spotted btree ptr %s: %s", buf, bch_ptr_status(c, k));
	return true;
}

bool bch_extent_ptr_invalid(struct cache_set *c, const struct bkey *k)
{
	char buf[80];

	if (!KEY_SIZE(k))
		return true;

	if (KEY_SIZE(k) > KEY_OFFSET(k))
		goto bad;

	if (__ptr_invalid(c, k))
		goto bad;

	return false;
bad:
	bch_bkey_to_text(buf, sizeof(buf), k);
	cache_bug(c, "spotted extent %s: %s", buf, bch_ptr_status(c, k));
	return true;
}

static bool ptr_bad_expensive_checks(struct btree *b, const struct bkey *k,
				     unsigned ptr)
{
	struct bucket *g = PTR_BUCKET(b->c, k, ptr);
	char buf[80];

	if (mutex_trylock(&b->c->bucket_lock)) {
		if (b->level) {
			if (KEY_DIRTY(k) ||
			    g->prio != BTREE_PRIO ||
			    (b->c->gc_mark_valid &&
			     GC_MARK(g) != GC_MARK_METADATA))
				goto err;

		} else {
			if (g->prio == BTREE_PRIO)
				goto err;

			if (KEY_DIRTY(k) &&
			    b->c->gc_mark_valid &&
			    GC_MARK(g) != GC_MARK_DIRTY)
				goto err;
		}
		mutex_unlock(&b->c->bucket_lock);
	}

	return false;
err:
	mutex_unlock(&b->c->bucket_lock);
	bch_bkey_to_text(buf, sizeof(buf), k);
	btree_bug(b,
"inconsistent pointer %s: bucket %zu pin %i prio %i gen %i last_gc %i mark %llu gc_gen %i",
		  buf, PTR_BUCKET_NR(b->c, k, ptr), atomic_read(&g->pin),
		  g->prio, g->gen, g->last_gc, GC_MARK(g), g->gc_gen);
	return true;
}

bool bch_ptr_bad(struct btree *b, const struct bkey *k)
{
	struct bucket *g;
	unsigned i, stale;

	if (!bkey_cmp(k, &ZERO_KEY) ||
	    !KEY_PTRS(k) ||
	    bch_ptr_invalid(b, k))
		return true;

	for (i = 0; i < KEY_PTRS(k); i++)
		if (!ptr_available(b->c, k, i))
			return true;

	if (!expensive_debug_checks(b->c) && KEY_DIRTY(k))
		return false;

	for (i = 0; i < KEY_PTRS(k); i++) {
		g = PTR_BUCKET(b->c, k, i);
		stale = ptr_stale(b->c, k, i);

		btree_bug_on(stale > 96, b,
			     "key too stale: %i, need_gc %u",
			     stale, b->c->need_gc);

		btree_bug_on(stale && KEY_DIRTY(k) && KEY_SIZE(k),
			     b, "stale dirty pointer");

		if (stale)
			return true;

		if (expensive_debug_checks(b->c) &&
		    ptr_bad_expensive_checks(b, k, i))
			return true;
	}

	return false;
}

/* Key/pointer manipulation */

void bch_bkey_copy_single_ptr(struct bkey *dest, const struct bkey *src,
			      unsigned i)
{
	BUG_ON(i > KEY_PTRS(src));

	/* Only copy the header, key, and one pointer. */
	memcpy(dest, src, 2 * sizeof(uint64_t));
	dest->ptr[0] = src->ptr[i];
	SET_KEY_PTRS(dest, 1);
	/* We didn't copy the checksum so clear that bit. */
	SET_KEY_CSUM(dest, 0);
}

bool __bch_cut_front(const struct bkey *where, struct bkey *k)
{
	unsigned i, len = 0;

	if (bkey_cmp(where, &START_KEY(k)) <= 0)
		return false;

	if (bkey_cmp(where, k) < 0)
		len = KEY_OFFSET(k) - KEY_OFFSET(where);
	else
		bkey_copy_key(k, where);

	for (i = 0; i < KEY_PTRS(k); i++)
		SET_PTR_OFFSET(k, i, PTR_OFFSET(k, i) + KEY_SIZE(k) - len);

	BUG_ON(len > KEY_SIZE(k));
	SET_KEY_SIZE(k, len);
	return true;
}

bool __bch_cut_back(const struct bkey *where, struct bkey *k)
{
	unsigned len = 0;

	if (bkey_cmp(where, k) >= 0)
		return false;

	BUG_ON(KEY_INODE(where) != KEY_INODE(k));

	if (bkey_cmp(where, &START_KEY(k)) > 0)
		len = KEY_OFFSET(where) - KEY_START(k);

	bkey_copy_key(k, where);

	BUG_ON(len > KEY_SIZE(k));
	SET_KEY_SIZE(k, len);
	return true;
}

static uint64_t merge_chksums(struct bkey *l, struct bkey *r)
{
	return (l->ptr[KEY_PTRS(l)] + r->ptr[KEY_PTRS(r)]) &
		~((uint64_t)1 << 63);
}

/* Tries to merge l and r: l should be lower than r
 * Returns true if we were able to merge. If we did merge, l will be the merged
 * key, r will be untouched.
 */
bool bch_bkey_try_merge(struct btree *b, struct bkey *l, struct bkey *r)
{
	unsigned i;

	if (key_merging_disabled(b->c))
		return false;

	if (KEY_PTRS(l) != KEY_PTRS(r) ||
	    KEY_DIRTY(l) != KEY_DIRTY(r) ||
	    bkey_cmp(l, &START_KEY(r)))
		return false;

	for (i = 0; i < KEY_PTRS(l); i++)
		if (l->ptr[i] + PTR(0, KEY_SIZE(l), 0) != r->ptr[i] ||
		    PTR_BUCKET_NR(b->c, l, i) != PTR_BUCKET_NR(b->c, r, i))
			return false;

	/* Keys with no pointers aren't restricted to one bucket and could
	 * overflow KEY_SIZE
	 */
	if (KEY_SIZE(l) + KEY_SIZE(r) > USHRT_MAX) {
		SET_KEY_OFFSET(l, KEY_OFFSET(l) + USHRT_MAX - KEY_SIZE(l));
		SET_KEY_SIZE(l, USHRT_MAX);

		bch_cut_front(l, r);
		return false;
	}

	if (KEY_CSUM(l)) {
		if (KEY_CSUM(r))
			l->ptr[KEY_PTRS(l)] = merge_chksums(l, r);
		else
			SET_KEY_CSUM(l, 0);
	}

	SET_KEY_OFFSET(l, KEY_OFFSET(l) + KEY_SIZE(r));
	SET_KEY_SIZE(l, KEY_SIZE(l) + KEY_SIZE(r));

	return true;
}

/* Binary tree stuff for auxiliary search trees */

static unsigned inorder_next(unsigned j, unsigned size)
{
	if (j * 2 + 1 < size) {
		j = j * 2 + 1;

		while (j * 2 < size)
			j *= 2;
	} else
		j >>= ffz(j) + 1;

	return j;
}

static unsigned inorder_prev(unsigned j, unsigned size)
{
	if (j * 2 < size) {
		j = j * 2;

		while (j * 2 + 1 < size)
			j = j * 2 + 1;
	} else
		j >>= ffs(j);

	return j;
}

/* I have no idea why this code works... and I'm the one who wrote it
 *
 * However, I do know what it does:
 * Given a binary tree constructed in an array (i.e. how you normally implement
 * a heap), it converts a node in the tree - referenced by array index - to the
 * index it would have if you did an inorder traversal.
 *
 * Also tested for every j, size up to size somewhere around 6 million.
 *
 * The binary tree starts at array index 1, not 0
 * extra is a function of size:
 *   extra = (size - rounddown_pow_of_two(size - 1)) << 1;
 */
static unsigned __to_inorder(unsigned j, unsigned size, unsigned extra)
{
	unsigned b = fls(j);
	unsigned shift = fls(size - 1) - b;

	j  ^= 1U << (b - 1);
	j <<= 1;
	j  |= 1;
	j <<= shift;

	if (j > extra)
		j -= (j - extra) >> 1;

	return j;
}

static unsigned to_inorder(unsigned j, struct bset_tree *t)
{
	return __to_inorder(j, t->size, t->extra);
}

static unsigned __inorder_to_tree(unsigned j, unsigned size, unsigned extra)
{
	unsigned shift;

	if (j > extra)
		j += j - extra;

	shift = ffs(j);

	j >>= shift;
	j  |= roundup_pow_of_two(size) >> shift;

	return j;
}

static unsigned inorder_to_tree(unsigned j, struct bset_tree *t)
{
	return __inorder_to_tree(j, t->size, t->extra);
}

#if 0
void inorder_test(void)
{
	unsigned long done = 0;
	ktime_t start = ktime_get();

	for (unsigned size = 2;
	     size < 65536000;
	     size++) {
		unsigned extra = (size - rounddown_pow_of_two(size - 1)) << 1;
		unsigned i = 1, j = rounddown_pow_of_two(size - 1);

		if (!(size % 4096))
			printk(KERN_NOTICE "loop %u, %llu per us\n", size,
			       done / ktime_us_delta(ktime_get(), start));

		while (1) {
			if (__inorder_to_tree(i, size, extra) != j)
				panic("size %10u j %10u i %10u", size, j, i);

			if (__to_inorder(j, size, extra) != i)
				panic("size %10u j %10u i %10u", size, j, i);

			if (j == rounddown_pow_of_two(size) - 1)
				break;

			BUG_ON(inorder_prev(inorder_next(j, size), size) != j);

			j = inorder_next(j, size);
			i++;
		}

		done += size - 1;
	}
}
#endif

/*
 * Cacheline/offset <-> bkey pointer arithmetic:
 *
 * t->tree is a binary search tree in an array; each node corresponds to a key
 * in one cacheline in t->set (BSET_CACHELINE bytes).
 *
 * This means we don't have to store the full index of the key that a node in
 * the binary tree points to; to_inorder() gives us the cacheline, and then
 * bkey_float->m gives us the offset within that cacheline, in units of 8 bytes.
 *
 * cacheline_to_bkey() and friends abstract out all the pointer arithmetic to
 * make this work.
 *
 * To construct the bfloat for an arbitrary key we need to know what the key
 * immediately preceding it is: we have to check if the two keys differ in the
 * bits we're going to store in bkey_float->mantissa. t->prev[j] stores the size
 * of the previous key so we can walk backwards to it from t->tree[j]'s key.
 */

static struct bkey *cacheline_to_bkey(struct bset_tree *t, unsigned cacheline,
				      unsigned offset)
{
	return ((void *) t->data) + cacheline * BSET_CACHELINE + offset * 8;
}

static unsigned bkey_to_cacheline(struct bset_tree *t, struct bkey *k)
{
	return ((void *) k - (void *) t->data) / BSET_CACHELINE;
}

static unsigned bkey_to_cacheline_offset(struct bkey *k)
{
	return ((size_t) k & (BSET_CACHELINE - 1)) / sizeof(uint64_t);
}

static struct bkey *tree_to_bkey(struct bset_tree *t, unsigned j)
{
	return cacheline_to_bkey(t, to_inorder(j, t), t->tree[j].m);
}

static struct bkey *tree_to_prev_bkey(struct bset_tree *t, unsigned j)
{
	return (void *) (((uint64_t *) tree_to_bkey(t, j)) - t->prev[j]);
}

/*
 * For the write set - the one we're currently inserting keys into - we don't
 * maintain a full search tree, we just keep a simple lookup table in t->prev.
 */
static struct bkey *table_to_bkey(struct bset_tree *t, unsigned cacheline)
{
	return cacheline_to_bkey(t, cacheline, t->prev[cacheline]);
}

static inline uint64_t shrd128(uint64_t high, uint64_t low, uint8_t shift)
{
	low >>= shift;
	low  |= (high << 1) << (63U - shift);
	return low;
}

static inline unsigned bfloat_mantissa(const struct bkey *k,
				       struct bkey_float *f)
{
	const uint64_t *p = &k->low - (f->exponent >> 6);
	return shrd128(p[-1], p[0], f->exponent & 63) & BKEY_MANTISSA_MASK;
}

static void make_bfloat(struct bset_tree *t, unsigned j)
{
	struct bkey_float *f = &t->tree[j];
	struct bkey *m = tree_to_bkey(t, j);
	struct bkey *p = tree_to_prev_bkey(t, j);

	struct bkey *l = is_power_of_2(j)
		? t->data->start
		: tree_to_prev_bkey(t, j >> ffs(j));

	struct bkey *r = is_power_of_2(j + 1)
		? node(t->data, t->data->keys - bkey_u64s(&t->end))
		: tree_to_bkey(t, j >> (ffz(j) + 1));

	BUG_ON(m < l || m > r);
	BUG_ON(bkey_next(p) != m);

	if (KEY_INODE(l) != KEY_INODE(r))
		f->exponent = fls64(KEY_INODE(r) ^ KEY_INODE(l)) + 64;
	else
		f->exponent = fls64(r->low ^ l->low);

	f->exponent = max_t(int, f->exponent - BKEY_MANTISSA_BITS, 0);

	/*
	 * Setting f->exponent = 127 flags this node as failed, and causes the
	 * lookup code to fall back to comparing against the original key.
	 */

	if (bfloat_mantissa(m, f) != bfloat_mantissa(p, f))
		f->mantissa = bfloat_mantissa(m, f) - 1;
	else
		f->exponent = 127;
}

static void bset_alloc_tree(struct btree *b, struct bset_tree *t)
{
	if (t != b->sets) {
		unsigned j = roundup(t[-1].size,
				     64 / sizeof(struct bkey_float));

		t->tree = t[-1].tree + j;
		t->prev = t[-1].prev + j;
	}

	while (t < b->sets + MAX_BSETS)
		t++->size = 0;
}

static void bset_build_unwritten_tree(struct btree *b)
{
	struct bset_tree *t = b->sets + b->nsets;

	bset_alloc_tree(b, t);

	if (t->tree != b->sets->tree + bset_tree_space(b)) {
		t->prev[0] = bkey_to_cacheline_offset(t->data->start);
		t->size = 1;
	}
}

static void bset_build_written_tree(struct btree *b)
{
	struct bset_tree *t = b->sets + b->nsets;
	struct bkey *k = t->data->start;
	unsigned j, cacheline = 1;

	bset_alloc_tree(b, t);

	t->size = min_t(unsigned,
			bkey_to_cacheline(t, end(t->data)),
			b->sets->tree + bset_tree_space(b) - t->tree);

	if (t->size < 2) {
		t->size = 0;
		return;
	}

	t->extra = (t->size - rounddown_pow_of_two(t->size - 1)) << 1;

	/* First we figure out where the first key in each cacheline is */
	for (j = inorder_next(0, t->size);
	     j;
	     j = inorder_next(j, t->size)) {
		while (bkey_to_cacheline(t, k) != cacheline)
			k = bkey_next(k);

		t->prev[j] = bkey_u64s(k);
		k = bkey_next(k);
		cacheline++;
		t->tree[j].m = bkey_to_cacheline_offset(k);
	}

	while (bkey_next(k) != end(t->data))
		k = bkey_next(k);

	t->end = *k;

	/* Then we build the tree */
	for (j = inorder_next(0, t->size);
	     j;
	     j = inorder_next(j, t->size))
		make_bfloat(t, j);
}

void bch_bset_fix_invalidated_key(struct btree *b, struct bkey *k)
{
	struct bset_tree *t;
	unsigned inorder, j = 1;

	for (t = b->sets; t <= &b->sets[b->nsets]; t++)
		if (k < end(t->data))
			goto found_set;

	BUG();
found_set:
	if (!t->size || !bset_written(b, t))
		return;

	inorder = bkey_to_cacheline(t, k);

	if (k == t->data->start)
		goto fix_left;

	if (bkey_next(k) == end(t->data)) {
		t->end = *k;
		goto fix_right;
	}

	j = inorder_to_tree(inorder, t);

	if (j &&
	    j < t->size &&
	    k == tree_to_bkey(t, j))
fix_left:	do {
			make_bfloat(t, j);
			j = j * 2;
		} while (j < t->size);

	j = inorder_to_tree(inorder + 1, t);

	if (j &&
	    j < t->size &&
	    k == tree_to_prev_bkey(t, j))
fix_right:	do {
			make_bfloat(t, j);
			j = j * 2 + 1;
		} while (j < t->size);
}

void bch_bset_fix_lookup_table(struct btree *b, struct bkey *k)
{
	struct bset_tree *t = &b->sets[b->nsets];
	unsigned shift = bkey_u64s(k);
	unsigned j = bkey_to_cacheline(t, k);

	/* We're getting called from btree_split() or btree_gc, just bail out */
	if (!t->size)
		return;

	/* k is the key we just inserted; we need to find the entry in the
	 * lookup table for the first key that is strictly greater than k:
	 * it's either k's cacheline or the next one
	 */
	if (j < t->size &&
	    table_to_bkey(t, j) <= k)
		j++;

	/* Adjust all the lookup table entries, and find a new key for any that
	 * have gotten too big
	 */
	for (; j < t->size; j++) {
		t->prev[j] += shift;

		if (t->prev[j] > 7) {
			k = table_to_bkey(t, j - 1);

			while (k < cacheline_to_bkey(t, j, 0))
				k = bkey_next(k);

			t->prev[j] = bkey_to_cacheline_offset(k);
		}
	}

	if (t->size == b->sets->tree + bset_tree_space(b) - t->tree)
		return;

	/* Possibly add a new entry to the end of the lookup table */

	for (k = table_to_bkey(t, t->size - 1);
	     k != end(t->data);
	     k = bkey_next(k))
		if (t->size == bkey_to_cacheline(t, k)) {
			t->prev[t->size] = bkey_to_cacheline_offset(k);
			t->size++;
		}
}

void bch_bset_init_next(struct btree *b)
{
	struct bset *i = write_block(b);

	if (i != b->sets[0].data) {
		b->sets[++b->nsets].data = i;
		i->seq = b->sets[0].data->seq;
	} else
		get_random_bytes(&i->seq, sizeof(uint64_t));

	i->magic	= bset_magic(&b->c->sb);
	i->version	= 0;
	i->keys		= 0;

	bset_build_unwritten_tree(b);
}

struct bset_search_iter {
	struct bkey *l, *r;
};

static struct bset_search_iter bset_search_write_set(struct btree *b,
						     struct bset_tree *t,
						     const struct bkey *search)
{
	unsigned li = 0, ri = t->size;

	BUG_ON(!b->nsets &&
	       t->size < bkey_to_cacheline(t, end(t->data)));

	while (li + 1 != ri) {
		unsigned m = (li + ri) >> 1;

		if (bkey_cmp(table_to_bkey(t, m), search) > 0)
			ri = m;
		else
			li = m;
	}

	return (struct bset_search_iter) {
		table_to_bkey(t, li),
		ri < t->size ? table_to_bkey(t, ri) : end(t->data)
	};
}

static struct bset_search_iter bset_search_tree(struct btree *b,
						struct bset_tree *t,
						const struct bkey *search)
{
	struct bkey *l, *r;
	struct bkey_float *f;
	unsigned inorder, j, n = 1;

	do {
		unsigned p = n << 4;
		p &= ((int) (p - t->size)) >> 31;

		prefetch(&t->tree[p]);

		j = n;
		f = &t->tree[j];

		/*
		 * n = (f->mantissa > bfloat_mantissa())
		 *	? j * 2
		 *	: j * 2 + 1;
		 *
		 * We need to subtract 1 from f->mantissa for the sign bit trick
		 * to work  - that's done in make_bfloat()
		 */
		if (likely(f->exponent != 127))
			n = j * 2 + (((unsigned)
				      (f->mantissa -
				       bfloat_mantissa(search, f))) >> 31);
		else
			n = (bkey_cmp(tree_to_bkey(t, j), search) > 0)
				? j * 2
				: j * 2 + 1;
	} while (n < t->size);

	inorder = to_inorder(j, t);

	/*
	 * n would have been the node we recursed to - the low bit tells us if
	 * we recursed left or recursed right.
	 */
	if (n & 1) {
		l = cacheline_to_bkey(t, inorder, f->m);

		if (++inorder != t->size) {
			f = &t->tree[inorder_next(j, t->size)];
			r = cacheline_to_bkey(t, inorder, f->m);
		} else
			r = end(t->data);
	} else {
		r = cacheline_to_bkey(t, inorder, f->m);

		if (--inorder) {
			f = &t->tree[inorder_prev(j, t->size)];
			l = cacheline_to_bkey(t, inorder, f->m);
		} else
			l = t->data->start;
	}

	return (struct bset_search_iter) {l, r};
}

struct bkey *__bch_bset_search(struct btree *b, struct bset_tree *t,
			       const struct bkey *search)
{
	struct bset_search_iter i;

	/*
	 * First, we search for a cacheline, then lastly we do a linear search
	 * within that cacheline.
	 *
	 * To search for the cacheline, there's three different possibilities:
	 *  * The set is too small to have a search tree, so we just do a linear
	 *    search over the whole set.
	 *  * The set is the one we're currently inserting into; keeping a full
	 *    auxiliary search tree up to date would be too expensive, so we
	 *    use a much simpler lookup table to do a binary search -
	 *    bset_search_write_set().
	 *  * Or we use the auxiliary search tree we constructed earlier -
	 *    bset_search_tree()
	 */

	if (unlikely(!t->size)) {
		i.l = t->data->start;
		i.r = end(t->data);
	} else if (bset_written(b, t)) {
		/*
		 * Each node in the auxiliary search tree covers a certain range
		 * of bits, and keys above and below the set it covers might
		 * differ outside those bits - so we have to special case the
		 * start and end - handle that here:
		 */

		if (unlikely(bkey_cmp(search, &t->end) >= 0))
			return end(t->data);

		if (unlikely(bkey_cmp(search, t->data->start) < 0))
			return t->data->start;

		i = bset_search_tree(b, t, search);
	} else
		i = bset_search_write_set(b, t, search);

	if (expensive_debug_checks(b->c)) {
		BUG_ON(bset_written(b, t) &&
		       i.l != t->data->start &&
		       bkey_cmp(tree_to_prev_bkey(t,
			  inorder_to_tree(bkey_to_cacheline(t, i.l), t)),
				search) > 0);

		BUG_ON(i.r != end(t->data) &&
		       bkey_cmp(i.r, search) <= 0);
	}

	while (likely(i.l != i.r) &&
	       bkey_cmp(i.l, search) <= 0)
		i.l = bkey_next(i.l);

	return i.l;
}

/* Btree iterator */

/*
 * Returns true if l > r - unless l == r, in which case returns true if l is
 * older than r.
 *
 * Necessary for btree_sort_fixup() - if there are multiple keys that compare
 * equal in different sets, we have to process them newest to oldest.
 */
static inline bool btree_iter_cmp(struct btree_iter_set l,
				  struct btree_iter_set r)
{
	int64_t c = bkey_cmp(&START_KEY(l.k), &START_KEY(r.k));

	return c ? c > 0 : l.k < r.k;
}

static inline bool btree_iter_end(struct btree_iter *iter)
{
	return !iter->used;
}

void bch_btree_iter_push(struct btree_iter *iter, struct bkey *k,
			 struct bkey *end)
{
	if (k != end)
		BUG_ON(!heap_add(iter,
				 ((struct btree_iter_set) { k, end }),
				 btree_iter_cmp));
}

struct bkey *__bch_btree_iter_init(struct btree *b, struct btree_iter *iter,
				   struct bkey *search, struct bset_tree *start)
{
	struct bkey *ret = NULL;
	iter->size = ARRAY_SIZE(iter->data);
	iter->used = 0;

#ifdef CONFIG_BCACHE_DEBUG
	iter->b = b;
#endif

	for (; start <= &b->sets[b->nsets]; start++) {
		ret = bch_bset_search(b, start, search);
		bch_btree_iter_push(iter, ret, end(start->data));
	}

	return ret;
}

struct bkey *bch_btree_iter_next(struct btree_iter *iter)
{
	struct btree_iter_set unused;
	struct bkey *ret = NULL;

	if (!btree_iter_end(iter)) {
		bch_btree_iter_next_check(iter);

		ret = iter->data->k;
		iter->data->k = bkey_next(iter->data->k);

		if (iter->data->k > iter->data->end) {
			WARN_ONCE(1, "bset was corrupt!\n");
			iter->data->k = iter->data->end;
		}

		if (iter->data->k == iter->data->end)
			heap_pop(iter, unused, btree_iter_cmp);
		else
			heap_sift(iter, 0, btree_iter_cmp);
	}

	return ret;
}

struct bkey *bch_btree_iter_next_filter(struct btree_iter *iter,
					struct btree *b, ptr_filter_fn fn)
{
	struct bkey *ret;

	do {
		ret = bch_btree_iter_next(iter);
	} while (ret && fn(b, ret));

	return ret;
}

/* Mergesort */

static void sort_key_next(struct btree_iter *iter,
			  struct btree_iter_set *i)
{
	i->k = bkey_next(i->k);

	if (i->k == i->end)
		*i = iter->data[--iter->used];
}

static struct bkey *btree_sort_fixup(struct btree_iter *iter, struct bkey *tmp)
{
	while (iter->used > 1) {
		struct btree_iter_set *top = iter->data, *i = top + 1;

		if (iter->used > 2 &&
		    btree_iter_cmp(i[0], i[1]))
			i++;

		if (bkey_cmp(top->k, &START_KEY(i->k)) <= 0)
			break;

		if (!KEY_SIZE(i->k)) {
			sort_key_next(iter, i);
			heap_sift(iter, i - top, btree_iter_cmp);
			continue;
		}

		if (top->k > i->k) {
			if (bkey_cmp(top->k, i->k) >= 0)
				sort_key_next(iter, i);
			else
				bch_cut_front(top->k, i->k);

			heap_sift(iter, i - top, btree_iter_cmp);
		} else {
			/* can't happen because of comparison func */
			BUG_ON(!bkey_cmp(&START_KEY(top->k), &START_KEY(i->k)));

			if (bkey_cmp(i->k, top->k) < 0) {
				bkey_copy(tmp, top->k);

				bch_cut_back(&START_KEY(i->k), tmp);
				bch_cut_front(i->k, top->k);
				heap_sift(iter, 0, btree_iter_cmp);

				return tmp;
			} else {
				bch_cut_back(&START_KEY(i->k), top->k);
			}
		}
	}

	return NULL;
}

static void btree_mergesort(struct btree *b, struct bset *out,
			    struct btree_iter *iter,
			    bool fixup, bool remove_stale)
{
	struct bkey *k, *last = NULL;
	BKEY_PADDED(k) tmp;
	bool (*bad)(struct btree *, const struct bkey *) = remove_stale
		? bch_ptr_bad
		: bch_ptr_invalid;

	while (!btree_iter_end(iter)) {
		if (fixup && !b->level)
			k = btree_sort_fixup(iter, &tmp.k);
		else
			k = NULL;

		if (!k)
			k = bch_btree_iter_next(iter);

		if (bad(b, k))
			continue;

		if (!last) {
			last = out->start;
			bkey_copy(last, k);
		} else if (b->level ||
			   !bch_bkey_try_merge(b, last, k)) {
			last = bkey_next(last);
			bkey_copy(last, k);
		}
	}

	out->keys = last ? (uint64_t *) bkey_next(last) - out->d : 0;

	pr_debug("sorted %i keys", out->keys);
}

static void __btree_sort(struct btree *b, struct btree_iter *iter,
			 unsigned start, unsigned order, bool fixup)
{
	uint64_t start_time;
	bool remove_stale = !b->written;
	struct bset *out = (void *) __get_free_pages(__GFP_NOWARN|GFP_NOIO,
						     order);
	if (!out) {
		mutex_lock(&b->c->sort_lock);
		out = b->c->sort;
		order = ilog2(bucket_pages(b->c));
	}

	start_time = local_clock();

	btree_mergesort(b, out, iter, fixup, remove_stale);
	b->nsets = start;

	if (!fixup && !start && b->written)
		bch_btree_verify(b, out);

	if (!start && order == b->page_order) {
		/*
		 * Our temporary buffer is the same size as the btree node's
		 * buffer, we can just swap buffers instead of doing a big
		 * memcpy()
		 */

		out->magic	= bset_magic(&b->c->sb);
		out->seq	= b->sets[0].data->seq;
		out->version	= b->sets[0].data->version;
		swap(out, b->sets[0].data);

		if (b->c->sort == b->sets[0].data)
			b->c->sort = out;
	} else {
		b->sets[start].data->keys = out->keys;
		memcpy(b->sets[start].data->start, out->start,
		       (void *) end(out) - (void *) out->start);
	}

	if (out == b->c->sort)
		mutex_unlock(&b->c->sort_lock);
	else
		free_pages((unsigned long) out, order);

	if (b->written)
		bset_build_written_tree(b);

	if (!start)
		bch_time_stats_update(&b->c->sort_time, start_time);
}

void bch_btree_sort_partial(struct btree *b, unsigned start)
{
	size_t order = b->page_order, keys = 0;
	struct btree_iter iter;
	int oldsize = bch_count_data(b);

	__bch_btree_iter_init(b, &iter, NULL, &b->sets[start]);

	BUG_ON(b->sets[b->nsets].data == write_block(b) &&
	       (b->sets[b->nsets].size || b->nsets));


	if (start) {
		unsigned i;

		for (i = start; i <= b->nsets; i++)
			keys += b->sets[i].data->keys;

		order = roundup_pow_of_two(__set_bytes(b->sets->data,
						       keys)) / PAGE_SIZE;
		if (order)
			order = ilog2(order);
	}

	__btree_sort(b, &iter, start, order, false);

	EBUG_ON(b->written && oldsize >= 0 && bch_count_data(b) != oldsize);
}

void bch_btree_sort_and_fix_extents(struct btree *b, struct btree_iter *iter)
{
	BUG_ON(!b->written);
	__btree_sort(b, iter, 0, b->page_order, true);
}

void bch_btree_sort_into(struct btree *b, struct btree *new)
{
	uint64_t start_time = local_clock();

	struct btree_iter iter;
	bch_btree_iter_init(b, &iter, NULL);

	btree_mergesort(b, new->sets->data, &iter, false, true);

	bch_time_stats_update(&b->c->sort_time, start_time);

	bkey_copy_key(&new->key, &b->key);
	new->sets->size = 0;
}

#define SORT_CRIT	(4096 / sizeof(uint64_t))

void bch_btree_sort_lazy(struct btree *b)
{
	unsigned crit = SORT_CRIT;
	int i;

	/* Don't sort if nothing to do */
	if (!b->nsets)
		goto out;

	/* If not a leaf node, always sort */
	if (b->level) {
		bch_btree_sort(b);
		return;
	}

	for (i = b->nsets - 1; i >= 0; --i) {
		crit *= b->c->sort_crit_factor;

		if (b->sets[i].data->keys < crit) {
			bch_btree_sort_partial(b, i);
			return;
		}
	}

	/* Sort if we'd overflow */
	if (b->nsets + 1 == MAX_BSETS) {
		bch_btree_sort(b);
		return;
	}

out:
	bset_build_written_tree(b);
}

/* Sysfs stuff */

struct bset_stats {
	struct btree_op op;
	size_t nodes;
	size_t sets_written, sets_unwritten;
	size_t bytes_written, bytes_unwritten;
	size_t floats, failed;
};

static int btree_bset_stats(struct btree_op *op, struct btree *b)
{
	struct bset_stats *stats = container_of(op, struct bset_stats, op);
	unsigned i;

	stats->nodes++;

	for (i = 0; i <= b->nsets; i++) {
		struct bset_tree *t = &b->sets[i];
		size_t bytes = t->data->keys * sizeof(uint64_t);
		size_t j;

		if (bset_written(b, t)) {
			stats->sets_written++;
			stats->bytes_written += bytes;

			stats->floats += t->size - 1;

			for (j = 1; j < t->size; j++)
				if (t->tree[j].exponent == 127)
					stats->failed++;
		} else {
			stats->sets_unwritten++;
			stats->bytes_unwritten += bytes;
		}
	}

	return MAP_CONTINUE;
}

int bch_bset_print_stats(struct cache_set *c, char *buf)
{
	struct bset_stats t;
	int ret;

	memset(&t, 0, sizeof(struct bset_stats));
	bch_btree_op_init(&t.op, -1);

	ret = bch_btree_map_nodes(&t.op, c, &ZERO_KEY, btree_bset_stats);
	if (ret < 0)
		return ret;

	return snprintf(buf, PAGE_SIZE,
			"btree nodes:		%zu\n"
			"written sets:		%zu\n"
			"unwritten sets:		%zu\n"
			"written key bytes:	%zu\n"
			"unwritten key bytes:	%zu\n"
			"floats:			%zu\n"
			"failed:			%zu\n",
			t.nodes,
			t.sets_written, t.sets_unwritten,
			t.bytes_written, t.bytes_unwritten,
			t.floats, t.failed);
}