bcache in userspace; userspace fsck

1 files changed, 2514 insertions, 0 deletions

diff --git a/libbcache/extents.c b/libbcache/extents.c
new file mode 100644
index 0000000..45fa220
--- /dev/null
+++ b/libbcache/extents.c
@@ -0,0 +1,2514 @@
+/*
+ * Copyright (C) 2010 Kent Overstreet <kent.overstreet@gmail.com>
+ *
+ * Code for managing the extent btree and dynamically updating the writeback
+ * dirty sector count.
+ */
+
+#include "bcache.h"
+#include "bkey_methods.h"
+#include "btree_gc.h"
+#include "btree_update.h"
+#include "debug.h"
+#include "dirent.h"
+#include "error.h"
+#include "extents.h"
+#include "inode.h"
+#include "journal.h"
+#include "super.h"
+#include "writeback.h"
+#include "xattr.h"
+
+#include <trace/events/bcache.h>
+
+static bool __bch_extent_normalize(struct cache_set *, struct bkey_s, bool);
+static enum merge_result bch_extent_merge(struct cache_set *, struct btree *,
+					  struct bkey_i *, struct bkey_i *);
+
+static void sort_key_next(struct btree_node_iter *iter,
+			  struct btree *b,
+			  struct btree_node_iter_set *i)
+{
+	i->k += __btree_node_offset_to_key(b, i->k)->u64s;
+
+	if (i->k == i->end)
+		*i = iter->data[--iter->used];
+}
+
+/*
+ * 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.
+ */
+#define key_sort_cmp(l, r)						\
+({									\
+	int _c = bkey_cmp_packed(b,					\
+				 __btree_node_offset_to_key(b, (l).k),	\
+				 __btree_node_offset_to_key(b, (r).k));	\
+									\
+	_c ? _c > 0 : (l).k > (r).k;					\
+})
+
+static inline bool should_drop_next_key(struct btree_node_iter *iter,
+					struct btree *b)
+{
+	struct btree_node_iter_set *l = iter->data, *r = iter->data + 1;
+	struct bkey_packed *k = __btree_node_offset_to_key(b, l->k);
+
+	if (bkey_whiteout(k))
+		return true;
+
+	if (iter->used < 2)
+		return false;
+
+	if (iter->used > 2 &&
+	    key_sort_cmp(r[0], r[1]))
+		r++;
+
+	/*
+	 * key_sort_cmp() ensures that when keys compare equal the older key
+	 * comes first; so if l->k compares equal to r->k then l->k is older and
+	 * should be dropped.
+	 */
+	return !bkey_cmp_packed(b,
+				__btree_node_offset_to_key(b, l->k),
+				__btree_node_offset_to_key(b, r->k));
+}
+
+struct btree_nr_keys bch_key_sort_fix_overlapping(struct bset *dst,
+						  struct btree *b,
+						  struct btree_node_iter *iter)
+{
+	struct bkey_packed *out = dst->start;
+	struct btree_nr_keys nr;
+
+	memset(&nr, 0, sizeof(nr));
+
+	heap_resort(iter, key_sort_cmp);
+
+	while (!bch_btree_node_iter_end(iter)) {
+		if (!should_drop_next_key(iter, b)) {
+			struct bkey_packed *k =
+				__btree_node_offset_to_key(b, iter->data->k);
+
+			bkey_copy(out, k);
+			btree_keys_account_key_add(&nr, 0, out);
+			out = bkey_next(out);
+		}
+
+		sort_key_next(iter, b, iter->data);
+		heap_sift(iter, 0, key_sort_cmp);
+	}
+
+	dst->u64s = cpu_to_le16((u64 *) out - dst->_data);
+	return nr;
+}
+
+/* Common among btree and extent ptrs */
+
+bool bch_extent_has_device(struct bkey_s_c_extent e, unsigned dev)
+{
+	const struct bch_extent_ptr *ptr;
+
+	extent_for_each_ptr(e, ptr)
+		if (ptr->dev == dev)
+			return true;
+
+	return false;
+}
+
+unsigned bch_extent_nr_ptrs_from(struct bkey_s_c_extent e,
+				 const struct bch_extent_ptr *start)
+{
+	const struct bch_extent_ptr *ptr;
+	unsigned nr_ptrs = 0;
+
+	extent_for_each_ptr_from(e, ptr, start)
+		nr_ptrs++;
+
+	return nr_ptrs;
+}
+
+unsigned bch_extent_nr_ptrs(struct bkey_s_c_extent e)
+{
+	return bch_extent_nr_ptrs_from(e, &e.v->start->ptr);
+}
+
+/* returns true if equal */
+static bool crc_cmp(union bch_extent_crc *l, union bch_extent_crc *r)
+{
+	return extent_crc_type(l) == extent_crc_type(r) &&
+		!memcmp(l, r, extent_entry_bytes(to_entry(l)));
+}
+
+/* Increment pointers after @crc by crc's offset until the next crc entry: */
+void bch_extent_crc_narrow_pointers(struct bkey_s_extent e, union bch_extent_crc *crc)
+{
+	union bch_extent_entry *entry;
+
+	extent_for_each_entry_from(e, entry, extent_entry_next(to_entry(crc))) {
+		if (!extent_entry_is_ptr(entry))
+			return;
+
+		entry->ptr.offset += crc_offset(crc);
+	}
+}
+
+/*
+ * We're writing another replica for this extent, so while we've got the data in
+ * memory we'll be computing a new checksum for the currently live data.
+ *
+ * If there are other replicas we aren't moving, and they are checksummed but
+ * not compressed, we can modify them to point to only the data that is
+ * currently live (so that readers won't have to bounce) while we've got the
+ * checksum we need:
+ *
+ * XXX: to guard against data being corrupted while in memory, instead of
+ * recomputing the checksum here, it would be better in the read path to instead
+ * of computing the checksum of the entire extent:
+ *
+ * | extent                              |
+ *
+ * compute the checksums of the live and dead data separately
+ * | dead data || live data || dead data |
+ *
+ * and then verify that crc_dead1 + crc_live + crc_dead2 == orig_crc, and then
+ * use crc_live here (that we verified was correct earlier)
+ */
+void bch_extent_narrow_crcs(struct bkey_s_extent e)
+{
+	union bch_extent_crc *crc;
+	bool have_wide = false, have_narrow = false;
+	u64 csum = 0;
+	unsigned csum_type = 0;
+
+	extent_for_each_crc(e, crc) {
+		if (crc_compression_type(crc))
+			continue;
+
+		if (crc_uncompressed_size(e.k, crc) != e.k->size) {
+			have_wide = true;
+		} else {
+			have_narrow = true;
+			csum = crc_csum(crc);
+			csum_type = crc_csum_type(crc);
+		}
+	}
+
+	if (!have_wide || !have_narrow)
+		return;
+
+	extent_for_each_crc(e, crc) {
+		if (crc_compression_type(crc))
+			continue;
+
+		if (crc_uncompressed_size(e.k, crc) != e.k->size) {
+			switch (extent_crc_type(crc)) {
+			case BCH_EXTENT_CRC_NONE:
+				BUG();
+			case BCH_EXTENT_CRC32:
+				if (bch_crc_size[csum_type] > sizeof(crc->crc32.csum))
+					continue;
+
+				bch_extent_crc_narrow_pointers(e, crc);
+				crc->crc32.compressed_size	= e.k->size;
+				crc->crc32.uncompressed_size	= e.k->size;
+				crc->crc32.offset		= 0;
+				crc->crc32.csum_type		= csum_type;
+				crc->crc32.csum			= csum;
+				break;
+			case BCH_EXTENT_CRC64:
+				if (bch_crc_size[csum_type] > sizeof(crc->crc64.csum))
+					continue;
+
+				bch_extent_crc_narrow_pointers(e, crc);
+				crc->crc64.compressed_size	= e.k->size;
+				crc->crc64.uncompressed_size	= e.k->size;
+				crc->crc64.offset		= 0;
+				crc->crc64.csum_type		= csum_type;
+				crc->crc64.csum			= csum;
+				break;
+			}
+		}
+	}
+}
+
+void bch_extent_drop_redundant_crcs(struct bkey_s_extent e)
+{
+	union bch_extent_entry *entry = e.v->start;
+	union bch_extent_crc *crc, *prev = NULL;
+
+	while (entry != extent_entry_last(e)) {
+		union bch_extent_entry *next = extent_entry_next(entry);
+		size_t crc_u64s = extent_entry_u64s(entry);
+
+		if (!extent_entry_is_crc(entry))
+			goto next;
+
+		crc = entry_to_crc(entry);
+
+		if (next == extent_entry_last(e)) {
+			/* crc entry with no pointers after it: */
+			goto drop;
+		}
+
+		if (extent_entry_is_crc(next)) {
+			/* no pointers before next crc entry: */
+			goto drop;
+		}
+
+		if (prev && crc_cmp(crc, prev)) {
+			/* identical to previous crc entry: */
+			goto drop;
+		}
+
+		if (!prev &&
+		    !crc_csum_type(crc) &&
+		    !crc_compression_type(crc)) {
+			/* null crc entry: */
+			bch_extent_crc_narrow_pointers(e, crc);
+			goto drop;
+		}
+
+		prev = crc;
+next:
+		entry = next;
+		continue;
+drop:
+		memmove_u64s_down(crc, next,
+				  (u64 *) extent_entry_last(e) - (u64 *) next);
+		e.k->u64s -= crc_u64s;
+	}
+
+	EBUG_ON(bkey_val_u64s(e.k) && !bch_extent_nr_ptrs(e.c));
+}
+
+static bool should_drop_ptr(const struct cache_set *c,
+			    struct bkey_s_c_extent e,
+			    const struct bch_extent_ptr *ptr)
+{
+	struct cache *ca;
+
+	return (ca = PTR_CACHE(c, ptr)) && ptr_stale(ca, ptr);
+}
+
+static void bch_extent_drop_stale(struct cache_set *c, struct bkey_s_extent e)
+{
+	struct bch_extent_ptr *ptr = &e.v->start->ptr;
+	bool dropped = false;
+
+	/*
+	 * We don't want to change which pointers are considered cached/dirty,
+	 * so don't remove pointers that are considered dirty:
+	 */
+	rcu_read_lock();
+	while ((ptr = extent_ptr_next(e, ptr)) &&
+	       !bch_extent_ptr_is_dirty(c, e.c, ptr))
+		if (should_drop_ptr(c, e.c, ptr)) {
+			__bch_extent_drop_ptr(e, ptr);
+			dropped = true;
+		} else
+			ptr++;
+	rcu_read_unlock();
+
+	if (dropped)
+		bch_extent_drop_redundant_crcs(e);
+}
+
+static bool bch_ptr_normalize(struct cache_set *c, struct btree *bk,
+			      struct bkey_s k)
+{
+	return __bch_extent_normalize(c, k, false);
+}
+
+static void bch_ptr_swab(const struct bkey_format *f, struct bkey_packed *k)
+{
+	u64 *d = (u64 *) bkeyp_val(f, k);
+	unsigned i;
+
+	for (i = 0; i < bkeyp_val_u64s(f, k); i++)
+		d[i] = swab64(d[i]);
+}
+
+static const char *extent_ptr_invalid(struct bkey_s_c_extent e,
+				      const struct cache_member_rcu *mi,
+				      const struct bch_extent_ptr *ptr,
+				      unsigned size_ondisk)
+{
+	const struct bch_extent_ptr *ptr2;
+	const struct cache_member_cpu *m = mi->m + ptr->dev;
+
+	if (ptr->dev > mi->nr_in_set || !m->valid)
+		return "pointer to invalid device";
+
+	extent_for_each_ptr(e, ptr2)
+		if (ptr != ptr2 && ptr->dev == ptr2->dev)
+			return "multiple pointers to same device";
+
+	if (ptr->offset + size_ondisk > m->bucket_size * m->nbuckets)
+		return "offset past end of device";
+
+	if (ptr->offset < m->bucket_size * m->first_bucket)
+		return "offset before first bucket";
+
+	if ((ptr->offset & (m->bucket_size - 1)) + size_ondisk > m->bucket_size)
+		return "spans multiple buckets";
+
+	return NULL;
+}
+
+static size_t extent_print_ptrs(struct cache_set *c, char *buf,
+				size_t size, struct bkey_s_c_extent e)
+{
+	char *out = buf, *end = buf + size;
+	const union bch_extent_entry *entry;
+	const union bch_extent_crc *crc;
+	const struct bch_extent_ptr *ptr;
+	struct cache *ca;
+	bool first = true;
+
+#define p(...)	(out += scnprintf(out, end - out, __VA_ARGS__))
+
+	rcu_read_lock();
+	extent_for_each_entry(e, entry) {
+		if (!first)
+			p(" ");
+
+		switch (__extent_entry_type(entry)) {
+		case BCH_EXTENT_ENTRY_crc32:
+		case BCH_EXTENT_ENTRY_crc64:
+			crc = entry_to_crc(entry);
+			p("crc: c_size %u size %u offset %u csum %u compress %u",
+			  crc_compressed_size(e.k, crc),
+			  crc_uncompressed_size(e.k, crc),
+			  crc_offset(crc), crc_csum_type(crc),
+			  crc_compression_type(crc));
+			break;
+		case BCH_EXTENT_ENTRY_ptr:
+			ptr = &entry->ptr;
+			p("ptr: %u:%llu gen %u%s", ptr->dev,
+			  (u64) ptr->offset, ptr->gen,
+			  (ca = PTR_CACHE(c, ptr)) && ptr_stale(ca, ptr)
+			  ? " stale" : "");
+			break;
+		default:
+			p("(invalid extent entry %.16llx)", *((u64 *) entry));
+			goto out;
+		}
+
+		first = false;
+	}
+out:
+	rcu_read_unlock();
+
+	if (bkey_extent_is_cached(e.k))
+		p(" cached");
+#undef p
+	return out - buf;
+}
+
+/* Btree ptrs */
+
+static const char *bch_btree_ptr_invalid(const struct cache_set *c,
+					 struct bkey_s_c k)
+{
+	if (bkey_extent_is_cached(k.k))
+		return "cached";
+
+	if (k.k->size)
+		return "nonzero key size";
+
+	if (bkey_val_u64s(k.k) > BKEY_BTREE_PTR_VAL_U64s_MAX)
+		return "value too big";
+
+	switch (k.k->type) {
+	case BCH_EXTENT: {
+		struct bkey_s_c_extent e = bkey_s_c_to_extent(k);
+		const union bch_extent_entry *entry;
+		const struct bch_extent_ptr *ptr;
+		const union bch_extent_crc *crc;
+		struct cache_member_rcu *mi;
+		const char *reason;
+
+		extent_for_each_entry(e, entry)
+			if (__extent_entry_type(entry) >= BCH_EXTENT_ENTRY_MAX)
+				return "invalid extent entry type";
+
+		mi = cache_member_info_get(c);
+
+		extent_for_each_ptr_crc(e, ptr, crc) {
+			reason = extent_ptr_invalid(e, mi, ptr,
+						c->sb.btree_node_size);
+
+			if (reason) {
+				cache_member_info_put();
+				return reason;
+			}
+		}
+
+		cache_member_info_put();
+
+		if (crc)
+			return "has crc field";
+
+		return NULL;
+	}
+
+	default:
+		return "invalid value type";
+	}
+}
+
+static void btree_ptr_debugcheck(struct cache_set *c, struct btree *b,
+				 struct bkey_s_c k)
+{
+	struct bkey_s_c_extent e = bkey_s_c_to_extent(k);
+	const struct bch_extent_ptr *ptr;
+	unsigned seq;
+	const char *err;
+	char buf[160];
+	struct bucket *g;
+	struct cache *ca;
+	unsigned replicas = 0;
+	bool bad;
+
+	rcu_read_lock();
+
+	extent_for_each_online_device(c, e, ptr, ca) {
+		replicas++;
+
+		if ((ca = PTR_CACHE(c, ptr))) {
+			g = PTR_BUCKET(ca, ptr);
+
+			err = "stale";
+			if (ptr_stale(ca, ptr))
+				goto err;
+
+			do {
+				seq = read_seqcount_begin(&c->gc_pos_lock);
+				bad = gc_pos_cmp(c->gc_pos, gc_pos_btree_node(b)) > 0 &&
+				       !g->mark.is_metadata;
+			} while (read_seqcount_retry(&c->gc_pos_lock, seq));
+
+			err = "inconsistent";
+			if (bad)
+				goto err;
+		}
+	}
+
+	rcu_read_unlock();
+
+	if (replicas < c->sb.meta_replicas_have) {
+		bch_bkey_val_to_text(c, btree_node_type(b),
+				     buf, sizeof(buf), k);
+		cache_set_bug(c,
+			"btree key bad (too few replicas, %u < %u): %s",
+			replicas, c->sb.meta_replicas_have, buf);
+		return;
+	}
+
+	return;
+err:
+	bch_bkey_val_to_text(c, btree_node_type(b), buf, sizeof(buf), k);
+	cache_set_bug(c, "%s btree pointer %s: bucket %zi prio %i "
+		      "gen %i last_gc %i mark %08x",
+		      err, buf, PTR_BUCKET_NR(ca, ptr),
+		      g->read_prio, PTR_BUCKET(ca, ptr)->mark.gen,
+		      ca->oldest_gens[PTR_BUCKET_NR(ca, ptr)],
+		      (unsigned) g->mark.counter);
+	rcu_read_unlock();
+}
+
+static void bch_btree_ptr_to_text(struct cache_set *c, char *buf,
+				  size_t size, struct bkey_s_c k)
+{
+	char *out = buf, *end = buf + size;
+	const char *invalid;
+
+#define p(...)	(out += scnprintf(out, end - out, __VA_ARGS__))
+
+	if (bkey_extent_is_data(k.k))
+		out += extent_print_ptrs(c, buf, size, bkey_s_c_to_extent(k));
+
+	invalid = bch_btree_ptr_invalid(c, k);
+	if (invalid)
+		p(" invalid: %s", invalid);
+#undef p
+}
+
+struct extent_pick_ptr
+bch_btree_pick_ptr(struct cache_set *c, const struct btree *b)
+{
+	struct bkey_s_c_extent e = bkey_i_to_s_c_extent(&b->key);
+	const union bch_extent_crc *crc;
+	const struct bch_extent_ptr *ptr;
+	struct cache *ca;
+
+	rcu_read_lock();
+
+	extent_for_each_online_device_crc(c, e, crc, ptr, ca) {
+		struct btree *root = btree_node_root(c, b);
+
+		if (cache_set_inconsistent_on(crc, c,
+				"btree node pointer with crc at btree %u level %u/%u bucket %zu",
+				b->btree_id, b->level, root ? root->level : -1,
+				PTR_BUCKET_NR(ca, ptr)))
+			break;
+
+		if (cache_inconsistent_on(ptr_stale(ca, ptr), ca,
+				"stale btree node pointer at btree %u level %u/%u bucket %zu",
+				b->btree_id, b->level, root ? root->level : -1,
+				PTR_BUCKET_NR(ca, ptr)))
+			continue;
+
+		percpu_ref_get(&ca->ref);
+		rcu_read_unlock();
+
+		return (struct extent_pick_ptr) { .ptr = *ptr, .ca = ca };
+	}
+
+	rcu_read_unlock();
+
+	return (struct extent_pick_ptr) { .ca = NULL, };
+}
+
+const struct bkey_ops bch_bkey_btree_ops = {
+	.key_invalid	= bch_btree_ptr_invalid,
+	.key_debugcheck	= btree_ptr_debugcheck,
+	.val_to_text	= bch_btree_ptr_to_text,
+	.swab		= bch_ptr_swab,
+};
+
+/* Extents */
+
+static bool __bch_cut_front(struct bpos where, struct bkey_s k)
+{
+	u64 len = 0;
+
+	if (bkey_cmp(where, bkey_start_pos(k.k)) <= 0)
+		return false;
+
+	EBUG_ON(bkey_cmp(where, k.k->p) > 0);
+
+	len = k.k->p.offset - where.offset;
+
+	BUG_ON(len > k.k->size);
+
+	/*
+	 * Don't readjust offset if the key size is now 0, because that could
+	 * cause offset to point to the next bucket:
+	 */
+	if (!len)
+		__set_bkey_deleted(k.k);
+	else if (bkey_extent_is_data(k.k)) {
+		struct bkey_s_extent e = bkey_s_to_extent(k);
+		struct bch_extent_ptr *ptr;
+		union bch_extent_crc *crc, *prev_crc = NULL;
+
+		extent_for_each_ptr_crc(e, ptr, crc) {
+			switch (extent_crc_type(crc)) {
+			case BCH_EXTENT_CRC_NONE:
+				ptr->offset += e.k->size - len;
+				break;
+			case BCH_EXTENT_CRC32:
+				if (prev_crc != crc)
+					crc->crc32.offset += e.k->size - len;
+				break;
+			case BCH_EXTENT_CRC64:
+				if (prev_crc != crc)
+					crc->crc64.offset += e.k->size - len;
+				break;
+			}
+			prev_crc = crc;
+		}
+	}
+
+	k.k->size = len;
+
+	return true;
+}
+
+bool bch_cut_front(struct bpos where, struct bkey_i *k)
+{
+	return __bch_cut_front(where, bkey_i_to_s(k));
+}
+
+bool bch_cut_back(struct bpos where, struct bkey *k)
+{
+	u64 len = 0;
+
+	if (bkey_cmp(where, k->p) >= 0)
+		return false;
+
+	EBUG_ON(bkey_cmp(where, bkey_start_pos(k)) < 0);
+
+	len = where.offset - bkey_start_offset(k);
+
+	BUG_ON(len > k->size);
+
+	k->p = where;
+	k->size = len;
+
+	if (!len)
+		__set_bkey_deleted(k);
+
+	return true;
+}
+
+/**
+ * bch_key_resize - adjust size of @k
+ *
+ * bkey_start_offset(k) will be preserved, modifies where the extent ends
+ */
+void bch_key_resize(struct bkey *k,
+		    unsigned new_size)
+{
+	k->p.offset -= k->size;
+	k->p.offset += new_size;
+	k->size = new_size;
+}
+
+/*
+ * In extent_sort_fix_overlapping(), insert_fixup_extent(),
+ * extent_merge_inline() - we're modifying keys in place that are packed. To do
+ * that we have to unpack the key, modify the unpacked key - then this
+ * copies/repacks the unpacked to the original as necessary.
+ */
+static bool __extent_save(struct btree *b, struct btree_node_iter *iter,
+			  struct bkey_packed *dst, struct bkey *src)
+{
+	struct bkey_format *f = &b->format;
+	struct bkey_i *dst_unpacked;
+	bool ret;
+
+	if ((dst_unpacked = packed_to_bkey(dst))) {
+		dst_unpacked->k = *src;
+		ret = true;
+	} else {
+		ret = bkey_pack_key(dst, src, f);
+	}
+
+	if (ret && iter)
+		bch_verify_key_order(b, iter, dst);
+
+	return ret;
+}
+
+static void extent_save(struct btree *b, struct btree_node_iter *iter,
+			struct bkey_packed *dst, struct bkey *src)
+{
+	BUG_ON(!__extent_save(b, iter, dst, src));
+}
+
+/*
+ * Returns true if l > r - unless l == r, in which case returns true if l is
+ * older than r.
+ *
+ * Necessary for sort_fix_overlapping() - if there are multiple keys that
+ * compare equal in different sets, we have to process them newest to oldest.
+ */
+#define extent_sort_cmp(l, r)						\
+({									\
+	struct bkey _ul = bkey_unpack_key(b,				\
+				__btree_node_offset_to_key(b, (l).k));	\
+	struct bkey _ur = bkey_unpack_key(b,				\
+				__btree_node_offset_to_key(b, (r).k));	\
+									\
+	int _c = bkey_cmp(bkey_start_pos(&_ul), bkey_start_pos(&_ur));	\
+	_c ? _c > 0 : (l).k < (r).k;					\
+})
+
+static inline void extent_sort_sift(struct btree_node_iter *iter,
+				    struct btree *b, size_t i)
+{
+	heap_sift(iter, i, extent_sort_cmp);
+}
+
+static inline void extent_sort_next(struct btree_node_iter *iter,
+				    struct btree *b,
+				    struct btree_node_iter_set *i)
+{
+	sort_key_next(iter, b, i);
+	heap_sift(iter, i - iter->data, extent_sort_cmp);
+}
+
+static void extent_sort_append(struct cache_set *c,
+			       struct btree *b,
+			       struct btree_nr_keys *nr,
+			       struct bkey_packed *start,
+			       struct bkey_packed **prev,
+			       struct bkey_packed *k)
+{
+	struct bkey_format *f = &b->format;
+	BKEY_PADDED(k) tmp;
+
+	if (bkey_whiteout(k))
+		return;
+
+	bkey_unpack(b, &tmp.k, k);
+
+	if (*prev &&
+	    bch_extent_merge(c, b, (void *) *prev, &tmp.k))
+		return;
+
+	if (*prev) {
+		bkey_pack(*prev, (void *) *prev, f);
+
+		btree_keys_account_key_add(nr, 0, *prev);
+		*prev = bkey_next(*prev);
+	} else {
+		*prev = start;
+	}
+
+	bkey_copy(*prev, &tmp.k);
+}
+
+struct btree_nr_keys bch_extent_sort_fix_overlapping(struct cache_set *c,
+					struct bset *dst,
+					struct btree *b,
+					struct btree_node_iter *iter)
+{
+	struct bkey_format *f = &b->format;
+	struct btree_node_iter_set *_l = iter->data, *_r;
+	struct bkey_packed *prev = NULL, *out, *lk, *rk;
+	struct bkey l_unpacked, r_unpacked;
+	struct bkey_s l, r;
+	struct btree_nr_keys nr;
+
+	memset(&nr, 0, sizeof(nr));
+
+	heap_resort(iter, extent_sort_cmp);
+
+	while (!bch_btree_node_iter_end(iter)) {
+		lk = __btree_node_offset_to_key(b, _l->k);
+
+		if (iter->used == 1) {
+			extent_sort_append(c, b, &nr, dst->start, &prev, lk);
+			extent_sort_next(iter, b, _l);
+			continue;
+		}
+
+		_r = iter->data + 1;
+		if (iter->used > 2 &&
+		    extent_sort_cmp(_r[0], _r[1]))
+			_r++;
+
+		rk = __btree_node_offset_to_key(b, _r->k);
+
+		l = __bkey_disassemble(b, lk, &l_unpacked);
+		r = __bkey_disassemble(b, rk, &r_unpacked);
+
+		/* If current key and next key don't overlap, just append */
+		if (bkey_cmp(l.k->p, bkey_start_pos(r.k)) <= 0) {
+			extent_sort_append(c, b, &nr, dst->start, &prev, lk);
+			extent_sort_next(iter, b, _l);
+			continue;
+		}
+
+		/* Skip 0 size keys */
+		if (!r.k->size) {
+			extent_sort_next(iter, b, _r);
+			continue;
+		}
+
+		/*
+		 * overlap: keep the newer key and trim the older key so they
+		 * don't overlap. comparing pointers tells us which one is
+		 * newer, since the bsets are appended one after the other.
+		 */
+
+		/* can't happen because of comparison func */
+		BUG_ON(_l->k < _r->k &&
+		       !bkey_cmp(bkey_start_pos(l.k), bkey_start_pos(r.k)));
+
+		if (_l->k > _r->k) {
+			/* l wins, trim r */
+			if (bkey_cmp(l.k->p, r.k->p) >= 0) {
+				sort_key_next(iter, b, _r);
+			} else {
+				__bch_cut_front(l.k->p, r);
+				extent_save(b, NULL, rk, r.k);
+			}
+
+			extent_sort_sift(iter, b, _r - iter->data);
+		} else if (bkey_cmp(l.k->p, r.k->p) > 0) {
+			BKEY_PADDED(k) tmp;
+
+			/*
+			 * r wins, but it overlaps in the middle of l - split l:
+			 */
+			bkey_reassemble(&tmp.k, l.s_c);
+			bch_cut_back(bkey_start_pos(r.k), &tmp.k.k);
+
+			__bch_cut_front(r.k->p, l);
+			extent_save(b, NULL, lk, l.k);
+
+			extent_sort_sift(iter, b, 0);
+
+			extent_sort_append(c, b, &nr, dst->start, &prev,
+					   bkey_to_packed(&tmp.k));
+		} else {
+			bch_cut_back(bkey_start_pos(r.k), l.k);
+			extent_save(b, NULL, lk, l.k);
+		}
+	}
+
+	if (prev) {
+		bkey_pack(prev, (void *) prev, f);
+		btree_keys_account_key_add(&nr, 0, prev);
+		out = bkey_next(prev);
+	} else {
+		out = dst->start;
+	}
+
+	dst->u64s = cpu_to_le16((u64 *) out - dst->_data);
+	return nr;
+}
+
+struct extent_insert_state {
+	struct btree_insert		*trans;
+	struct btree_insert_entry	*insert;
+	struct bpos			committed;
+	struct bucket_stats_cache_set	stats;
+
+	/* for deleting: */
+	struct bkey_i			whiteout;
+	bool				do_journal;
+	bool				deleting;
+};
+
+static void bch_add_sectors(struct extent_insert_state *s,
+			    struct bkey_s_c k, u64 offset, s64 sectors)
+{
+	struct cache_set *c = s->trans->c;
+	struct btree *b = s->insert->iter->nodes[0];
+
+	EBUG_ON(bkey_cmp(bkey_start_pos(k.k), b->data->min_key) < 0);
+
+	if (!sectors)
+		return;
+
+	bch_mark_key(c, k, sectors, false, gc_pos_btree_node(b),
+		     &s->stats, s->trans->journal_res.seq);
+
+	if (bkey_extent_is_data(k.k) &&
+	    !bkey_extent_is_cached(k.k))
+		bcache_dev_sectors_dirty_add(c, k.k->p.inode, offset, sectors);
+}
+
+static void bch_subtract_sectors(struct extent_insert_state *s,
+				 struct bkey_s_c k, u64 offset, s64 sectors)
+{
+	bch_add_sectors(s, k, offset, -sectors);
+}
+
+/* These wrappers subtract exactly the sectors that we're removing from @k */
+static void bch_cut_subtract_back(struct extent_insert_state *s,
+				  struct bpos where, struct bkey_s k)
+{
+	bch_subtract_sectors(s, k.s_c, where.offset,
+			     k.k->p.offset - where.offset);
+	bch_cut_back(where, k.k);
+}
+
+static void bch_cut_subtract_front(struct extent_insert_state *s,
+				   struct bpos where, struct bkey_s k)
+{
+	bch_subtract_sectors(s, k.s_c, bkey_start_offset(k.k),
+			     where.offset - bkey_start_offset(k.k));
+	__bch_cut_front(where, k);
+}
+
+static void bch_drop_subtract(struct extent_insert_state *s, struct bkey_s k)
+{
+	if (k.k->size)
+		bch_subtract_sectors(s, k.s_c,
+				     bkey_start_offset(k.k), k.k->size);
+	k.k->size = 0;
+	__set_bkey_deleted(k.k);
+}
+
+/*
+ * Note: If this returns true because only some pointers matched,
+ * we can lose some caching that had happened in the interim.
+ * Because cache promotion only promotes the part of the extent
+ * actually read, and not the whole extent, and due to the key
+ * splitting done in bch_extent_insert_fixup, preserving such
+ * caching is difficult.
+ */
+static bool bch_extent_cmpxchg_cmp(struct bkey_s_c l, struct bkey_s_c r)
+{
+	struct bkey_s_c_extent le, re;
+	const struct bch_extent_ptr *lp, *rp;
+	s64 offset;
+
+	BUG_ON(!l.k->size || !r.k->size);
+
+	if (l.k->type != r.k->type ||
+	    l.k->version != r.k->version)
+		return false;
+
+	switch (l.k->type) {
+	case KEY_TYPE_COOKIE:
+		return !memcmp(bkey_s_c_to_cookie(l).v,
+			       bkey_s_c_to_cookie(r).v,
+			       sizeof(struct bch_cookie));
+
+	case BCH_EXTENT:
+	case BCH_EXTENT_CACHED:
+		le = bkey_s_c_to_extent(l);
+		re = bkey_s_c_to_extent(r);
+
+		/*
+		 * bkey_cmpxchg() handles partial matches - when either l or r
+		 * has been trimmed - so we need just to handle l or r not
+		 * starting at the same place when checking for a match here.
+		 *
+		 * If the starts of the keys are different, we just apply that
+		 * offset to the device pointer offsets when checking those -
+		 * matching how bch_cut_front() adjusts device pointer offsets
+		 * when adjusting the start of a key:
+		 */
+		offset = bkey_start_offset(l.k) - bkey_start_offset(r.k);
+
+		/*
+		 * XXX: perhaps we only raced with copygc or tiering replacing
+		 * one of the pointers: it should suffice to find _any_ matching
+		 * pointer
+		 */
+
+		if (bkey_val_u64s(le.k) != bkey_val_u64s(re.k))
+			return false;
+
+		extent_for_each_ptr(le, lp) {
+			const union bch_extent_entry *entry =
+				bkey_idx(re.v, (u64 *) lp - le.v->_data);
+
+			if (!extent_entry_is_ptr(entry))
+				return false;
+
+			rp = &entry->ptr;
+
+			if (lp->offset	!= rp->offset + offset ||
+			    lp->dev	!= rp->dev ||
+			    lp->gen	!= rp->gen)
+				return false;
+		}
+
+		return true;
+	default:
+		return false;
+	}
+
+}
+
+/*
+ * Returns true on success, false on failure (and false means @new no longer
+ * overlaps with @k)
+ *
+ * If returned true, we may have inserted up to one key in @b.
+ * If returned false, we may have inserted up to two keys in @b.
+ *
+ * On return, there is room in @res for at least one more key of the same size
+ * as @new.
+ */
+enum extent_insert_hook_ret bch_extent_cmpxchg(struct extent_insert_hook *hook,
+					       struct bpos committed_pos,
+					       struct bpos next_pos,
+					       struct bkey_s_c k,
+					       const struct bkey_i *new)
+{
+	struct bch_replace_info *replace = container_of(hook,
+					struct bch_replace_info, hook);
+	struct bkey_i *old = &replace->key;
+
+	EBUG_ON(bkey_cmp(committed_pos, bkey_start_pos(&new->k)) < 0);
+
+	/* must have something to compare against */
+	EBUG_ON(!bkey_val_u64s(&old->k));
+
+	/* new must be a subset of old */
+	EBUG_ON(bkey_cmp(new->k.p, old->k.p) > 0 ||
+		bkey_cmp(bkey_start_pos(&new->k), bkey_start_pos(&old->k)) < 0);
+
+	if (k.k && bch_extent_cmpxchg_cmp(k, bkey_i_to_s_c(old))) {
+		replace->successes++;
+		return BTREE_HOOK_DO_INSERT;
+	} else {
+		replace->failures++;
+		return BTREE_HOOK_NO_INSERT;
+	}
+}
+
+static bool bch_extent_merge_inline(struct cache_set *,
+				    struct btree_iter *,
+				    struct bkey_packed *,
+				    struct bkey_packed *,
+				    bool);
+
+#define MAX_LOCK_HOLD_TIME	(5 * NSEC_PER_MSEC)
+
+static enum btree_insert_ret
+extent_insert_should_stop(struct extent_insert_state *s)
+{
+	struct btree *b = s->insert->iter->nodes[0];
+
+	/*
+	 * Check if we have sufficient space in both the btree node and the
+	 * journal reservation:
+	 *
+	 * Each insert checks for room in the journal entry, but we check for
+	 * room in the btree node up-front. In the worst case, bkey_cmpxchg()
+	 * will insert two keys, and one iteration of this room will insert one
+	 * key, so we need room for three keys.
+	 */
+	if (!bch_btree_node_insert_fits(s->trans->c, b, s->insert->k->k.u64s))
+		return BTREE_INSERT_BTREE_NODE_FULL;
+	else if (!journal_res_insert_fits(s->trans, s->insert))
+		return BTREE_INSERT_JOURNAL_RES_FULL; /* XXX worth tracing */
+	else
+		return BTREE_INSERT_OK;
+}
+
+static void extent_bset_insert(struct cache_set *c, struct btree_iter *iter,
+			       struct bkey_i *insert)
+{
+	struct btree *b = iter->nodes[0];
+	struct btree_node_iter *node_iter = &iter->node_iters[0];
+	struct bset_tree *t = bset_tree_last(b);
+	struct bkey_packed *where =
+		bch_btree_node_iter_bset_pos(node_iter, b, t);
+	struct bkey_packed *prev = bkey_prev(b, t, where);
+	struct bkey_packed *next_live_key = where;
+	unsigned clobber_u64s;
+
+	if (prev)
+		where = bkey_next(prev);
+
+	while (next_live_key != btree_bkey_last(b, t) &&
+	       bkey_deleted(next_live_key))
+		next_live_key = bkey_next(next_live_key);
+
+	/*
+	 * Everything between where and next_live_key is now deleted keys, and
+	 * is overwritten:
+	 */
+	clobber_u64s = (u64 *) next_live_key - (u64 *) where;
+
+	if (prev &&
+	    bch_extent_merge_inline(c, iter, prev, bkey_to_packed(insert), true))
+		goto drop_deleted_keys;
+
+	if (next_live_key != btree_bkey_last(b, t) &&
+	    bch_extent_merge_inline(c, iter, bkey_to_packed(insert),
+				    next_live_key, false))
+		goto drop_deleted_keys;
+
+	bch_bset_insert(b, node_iter, where, insert, clobber_u64s);
+	bch_btree_node_iter_fix(iter, b, node_iter, t, where,
+				clobber_u64s, where->u64s);
+	return;
+drop_deleted_keys:
+	bch_bset_delete(b, where, clobber_u64s);
+	bch_btree_node_iter_fix(iter, b, node_iter, t, where, clobber_u64s, 0);
+}
+
+static void extent_insert_committed(struct extent_insert_state *s)
+{
+	struct cache_set *c = s->trans->c;
+	struct btree_iter *iter = s->insert->iter;
+	struct bkey_i *insert = !s->deleting
+		? s->insert->k
+		: &s->whiteout;
+	BKEY_PADDED(k) split;
+
+	EBUG_ON(bkey_cmp(insert->k.p, s->committed) < 0);
+	EBUG_ON(bkey_cmp(s->committed, bkey_start_pos(&insert->k)) < 0);
+
+	if (!bkey_cmp(s->committed, bkey_start_pos(&insert->k)))
+		return;
+
+	if (s->deleting && !s->do_journal) {
+		bch_cut_front(s->committed, insert);
+		goto done;
+	}
+
+	EBUG_ON(bkey_deleted(&insert->k) || !insert->k.size);
+
+	bkey_copy(&split.k, insert);
+
+	if (!(s->trans->flags & BTREE_INSERT_JOURNAL_REPLAY) &&
+	    bkey_cmp(s->committed, insert->k.p) &&
+	    bkey_extent_is_compressed(c, bkey_i_to_s_c(insert))) {
+		/* XXX: possibly need to increase our reservation? */
+		bch_cut_subtract_back(s, s->committed,
+				      bkey_i_to_s(&split.k));
+		bch_cut_front(s->committed, insert);
+		bch_add_sectors(s, bkey_i_to_s_c(insert),
+				bkey_start_offset(&insert->k),
+				insert->k.size);
+	} else {
+		bch_cut_back(s->committed, &split.k.k);
+		bch_cut_front(s->committed, insert);
+	}
+
+	if (debug_check_bkeys(c))
+		bkey_debugcheck(c, iter->nodes[iter->level],
+				bkey_i_to_s_c(&split.k));
+
+	bch_btree_journal_key(s->trans, iter, &split.k);
+
+	if (!s->deleting)
+		extent_bset_insert(c, iter, &split.k);
+done:
+	bch_btree_iter_set_pos_same_leaf(iter, s->committed);
+
+	insert->k.needs_whiteout	= false;
+	s->do_journal			= false;
+	s->trans->did_work		= true;
+}
+
+static enum extent_insert_hook_ret
+__extent_insert_advance_pos(struct extent_insert_state *s,
+			    struct bpos next_pos,
+			    struct bkey_s_c k)
+{
+	struct extent_insert_hook *hook = s->trans->hook;
+	enum extent_insert_hook_ret ret;
+
+	if (k.k && k.k->size &&
+	    s->insert->k->k.version &&
+	    k.k->version > s->insert->k->k.version)
+		ret = BTREE_HOOK_NO_INSERT;
+	else if (hook)
+		ret = hook->fn(hook, s->committed, next_pos, k, s->insert->k);
+	else
+		ret = BTREE_HOOK_DO_INSERT;
+
+	EBUG_ON(bkey_deleted(&s->insert->k->k) || !s->insert->k->k.size);
+
+	switch (ret) {
+	case BTREE_HOOK_DO_INSERT:
+		break;
+	case BTREE_HOOK_NO_INSERT:
+		extent_insert_committed(s);
+		bch_cut_subtract_front(s, next_pos, bkey_i_to_s(s->insert->k));
+
+		bch_btree_iter_set_pos_same_leaf(s->insert->iter, next_pos);
+		break;
+	case BTREE_HOOK_RESTART_TRANS:
+		return ret;
+	}
+
+	s->committed = next_pos;
+	return ret;
+}
+
+/*
+ * Update iter->pos, marking how much of @insert we've processed, and call hook
+ * fn:
+ */
+static enum extent_insert_hook_ret
+extent_insert_advance_pos(struct extent_insert_state *s, struct bkey_s_c k)
+{
+	struct btree *b = s->insert->iter->nodes[0];
+	struct bpos next_pos = bpos_min(s->insert->k->k.p,
+					k.k ? k.k->p : b->key.k.p);
+
+	/* hole? */
+	if (k.k && bkey_cmp(s->committed, bkey_start_pos(k.k)) < 0) {
+		bool have_uncommitted = bkey_cmp(s->committed,
+				bkey_start_pos(&s->insert->k->k)) > 0;
+
+		switch (__extent_insert_advance_pos(s, bkey_start_pos(k.k),
+						    bkey_s_c_null)) {
+		case BTREE_HOOK_DO_INSERT:
+			break;
+		case BTREE_HOOK_NO_INSERT:
+			/*
+			 * we had to split @insert and insert the committed
+			 * part - need to bail out and recheck journal
+			 * reservation/btree node before we advance pos past @k:
+			 */
+			if (have_uncommitted)
+				return BTREE_HOOK_NO_INSERT;
+			break;
+		case BTREE_HOOK_RESTART_TRANS:
+			return BTREE_HOOK_RESTART_TRANS;
+		}
+	}
+
+	/* avoid redundant calls to hook fn: */
+	if (!bkey_cmp(s->committed, next_pos))
+		return BTREE_HOOK_DO_INSERT;
+
+	return __extent_insert_advance_pos(s, next_pos, k);
+}
+
+static enum btree_insert_ret
+extent_insert_check_split_compressed(struct extent_insert_state *s,
+				     struct bkey_s_c k,
+				     enum bch_extent_overlap overlap)
+{
+	struct cache_set *c = s->trans->c;
+	unsigned sectors;
+
+	if (overlap == BCH_EXTENT_OVERLAP_MIDDLE &&
+	    (sectors = bkey_extent_is_compressed(c, k))) {
+		int flags = BCH_DISK_RESERVATION_BTREE_LOCKS_HELD;
+
+		if (s->trans->flags & BTREE_INSERT_NOFAIL)
+			flags |= BCH_DISK_RESERVATION_NOFAIL;
+
+		switch (bch_disk_reservation_add(c,
+				s->trans->disk_res,
+				sectors, flags)) {
+		case 0:
+			break;
+		case -ENOSPC:
+			return BTREE_INSERT_ENOSPC;
+		case -EINTR:
+			return BTREE_INSERT_NEED_GC_LOCK;
+		default:
+			BUG();
+		}
+	}
+
+	return BTREE_INSERT_OK;
+}
+
+static enum btree_insert_ret
+extent_squash(struct extent_insert_state *s, struct bkey_i *insert,
+	      struct bset_tree *t, struct bkey_packed *_k, struct bkey_s k,
+	      enum bch_extent_overlap overlap)
+{
+	struct cache_set *c = s->trans->c;
+	struct btree_iter *iter = s->insert->iter;
+	struct btree *b = iter->nodes[0];
+	struct btree_node_iter *node_iter = &iter->node_iters[0];
+
+	switch (overlap) {
+	case BCH_EXTENT_OVERLAP_FRONT:
+		/* insert overlaps with start of k: */
+		bch_cut_subtract_front(s, insert->k.p, k);
+		BUG_ON(bkey_deleted(k.k));
+		extent_save(b, node_iter, _k, k.k);
+		break;
+
+	case BCH_EXTENT_OVERLAP_BACK:
+		/* insert overlaps with end of k: */
+		bch_cut_subtract_back(s, bkey_start_pos(&insert->k), k);
+		BUG_ON(bkey_deleted(k.k));
+		extent_save(b, node_iter, _k, k.k);
+
+		/*
+		 * As the auxiliary tree is indexed by the end of the
+		 * key and we've just changed the end, update the
+		 * auxiliary tree.
+		 */
+		bch_bset_fix_invalidated_key(b, t, _k);
+		bch_btree_node_iter_fix(iter, b, node_iter, t,
+					_k, _k->u64s, _k->u64s);
+		break;
+
+	case BCH_EXTENT_OVERLAP_ALL: {
+		struct bpos orig_pos = k.k->p;
+
+		/* The insert key completely covers k, invalidate k */
+		if (!bkey_whiteout(k.k))
+			btree_keys_account_key_drop(&b->nr,
+						t - b->set, _k);
+
+		bch_drop_subtract(s, k);
+		k.k->p = bkey_start_pos(&insert->k);
+		if (!__extent_save(b, node_iter, _k, k.k)) {
+			/*
+			 * Couldn't repack: we aren't necessarily able
+			 * to repack if the new key is outside the range
+			 * of the old extent, so we have to split
+			 * @insert:
+			 */
+			k.k->p = orig_pos;
+			extent_save(b, node_iter, _k, k.k);
+
+			if (extent_insert_advance_pos(s, k.s_c) ==
+			    BTREE_HOOK_RESTART_TRANS)
+				return BTREE_INSERT_NEED_TRAVERSE;
+
+			extent_insert_committed(s);
+			/*
+			 * We split and inserted upto at k.k->p - that
+			 * has to coincide with iter->pos, so that we
+			 * don't have anything more we have to insert
+			 * until we recheck our journal reservation:
+			 */
+			EBUG_ON(bkey_cmp(s->committed, k.k->p));
+		} else {
+			bch_bset_fix_invalidated_key(b, t, _k);
+			bch_btree_node_iter_fix(iter, b, node_iter, t,
+						_k, _k->u64s, _k->u64s);
+		}
+
+		break;
+	}
+	case BCH_EXTENT_OVERLAP_MIDDLE: {
+		BKEY_PADDED(k) split;
+		/*
+		 * The insert key falls 'in the middle' of k
+		 * The insert key splits k in 3:
+		 * - start only in k, preserve
+		 * - middle common section, invalidate in k
+		 * - end only in k, preserve
+		 *
+		 * We update the old key to preserve the start,
+		 * insert will be the new common section,
+		 * we manually insert the end that we are preserving.
+		 *
+		 * modify k _before_ doing the insert (which will move
+		 * what k points to)
+		 */
+		bkey_reassemble(&split.k, k.s_c);
+		split.k.k.needs_whiteout |= bset_written(b, bset(b, t));
+
+		bch_cut_back(bkey_start_pos(&insert->k), &split.k.k);
+		BUG_ON(bkey_deleted(&split.k.k));
+
+		bch_cut_subtract_front(s, insert->k.p, k);
+		BUG_ON(bkey_deleted(k.k));
+		extent_save(b, node_iter, _k, k.k);
+
+		bch_add_sectors(s, bkey_i_to_s_c(&split.k),
+				bkey_start_offset(&split.k.k),
+				split.k.k.size);
+		extent_bset_insert(c, iter, &split.k);
+		break;
+	}
+	}
+
+	return BTREE_INSERT_OK;
+}
+
+static enum btree_insert_ret
+bch_delete_fixup_extent(struct extent_insert_state *s)
+{
+	struct cache_set *c = s->trans->c;
+	struct btree_iter *iter = s->insert->iter;
+	struct btree *b = iter->nodes[0];
+	struct btree_node_iter *node_iter = &iter->node_iters[0];
+	struct bkey_packed *_k;
+	struct bkey unpacked;
+	struct bkey_i *insert = s->insert->k;
+	enum btree_insert_ret ret = BTREE_INSERT_OK;
+
+	EBUG_ON(bkey_cmp(iter->pos, bkey_start_pos(&insert->k)));
+
+	s->whiteout	= *insert;
+	s->do_journal	= false;
+
+	while (bkey_cmp(s->committed, insert->k.p) < 0 &&
+	       (ret = extent_insert_should_stop(s)) == BTREE_INSERT_OK &&
+	       (_k = bch_btree_node_iter_peek_all(node_iter, b))) {
+		struct bset_tree *t = bch_bkey_to_bset(b, _k);
+		struct bkey_s k = __bkey_disassemble(b, _k, &unpacked);
+		enum bch_extent_overlap overlap;
+
+		EBUG_ON(bkey_cmp(iter->pos, bkey_start_pos(&insert->k)));
+		EBUG_ON(bkey_cmp(iter->pos, k.k->p) >= 0);
+
+		if (bkey_cmp(bkey_start_pos(k.k), insert->k.p) >= 0)
+			break;
+
+		if (bkey_whiteout(k.k)) {
+			s->committed = bpos_min(insert->k.p, k.k->p);
+			goto next;
+		}
+
+		overlap = bch_extent_overlap(&insert->k, k.k);
+
+		ret = extent_insert_check_split_compressed(s, k.s_c, overlap);
+		if (ret != BTREE_INSERT_OK)
+			goto stop;
+
+		switch (extent_insert_advance_pos(s, k.s_c)) {
+		case BTREE_HOOK_DO_INSERT:
+			break;
+		case BTREE_HOOK_NO_INSERT:
+			continue;
+		case BTREE_HOOK_RESTART_TRANS:
+			ret = BTREE_INSERT_NEED_TRAVERSE;
+			goto stop;
+		}
+
+		s->do_journal = true;
+
+		if (overlap == BCH_EXTENT_OVERLAP_ALL) {
+			btree_keys_account_key_drop(&b->nr,
+						t - b->set, _k);
+			bch_subtract_sectors(s, k.s_c,
+					     bkey_start_offset(k.k), k.k->size);
+			_k->type = KEY_TYPE_DISCARD;
+			reserve_whiteout(b, t, _k);
+		} else if (k.k->needs_whiteout ||
+			   bset_written(b, bset(b, t))) {
+			struct bkey_i discard = *insert;
+
+			switch (overlap) {
+			case BCH_EXTENT_OVERLAP_FRONT:
+				bch_cut_front(bkey_start_pos(k.k), &discard);
+				break;
+			case BCH_EXTENT_OVERLAP_BACK:
+				bch_cut_back(k.k->p, &discard.k);
+				break;
+			default:
+				break;
+			}
+
+			discard.k.needs_whiteout = true;
+
+			ret = extent_squash(s, insert, t, _k, k, overlap);
+			BUG_ON(ret != BTREE_INSERT_OK);
+
+			extent_bset_insert(c, iter, &discard);
+		} else {
+			ret = extent_squash(s, insert, t, _k, k, overlap);
+			BUG_ON(ret != BTREE_INSERT_OK);
+		}
+next:
+		bch_cut_front(s->committed, insert);
+		bch_btree_iter_set_pos_same_leaf(iter, s->committed);
+	}
+
+	if (bkey_cmp(s->committed, insert->k.p) < 0 &&
+	    ret == BTREE_INSERT_OK &&
+	    extent_insert_advance_pos(s, bkey_s_c_null) == BTREE_HOOK_RESTART_TRANS)
+		ret = BTREE_INSERT_NEED_TRAVERSE;
+stop:
+	extent_insert_committed(s);
+
+	bch_cache_set_stats_apply(c, &s->stats, s->trans->disk_res,
+				  gc_pos_btree_node(b));
+
+	EBUG_ON(bkey_cmp(iter->pos, s->committed));
+	EBUG_ON((bkey_cmp(iter->pos, b->key.k.p) == 0) != iter->at_end_of_leaf);
+
+	bch_cut_front(iter->pos, insert);
+
+	if (insert->k.size && iter->at_end_of_leaf)
+		ret = BTREE_INSERT_NEED_TRAVERSE;
+
+	EBUG_ON(insert->k.size && ret == BTREE_INSERT_OK);
+
+	return ret;
+}
+
+/**
+ * bch_extent_insert_fixup - insert a new extent and deal with overlaps
+ *
+ * this may result in not actually doing the insert, or inserting some subset
+ * of the insert key. For cmpxchg operations this is where that logic lives.
+ *
+ * All subsets of @insert that need to be inserted are inserted using
+ * bch_btree_insert_and_journal(). If @b or @res fills up, this function
+ * returns false, setting @iter->pos for the prefix of @insert that actually got
+ * inserted.
+ *
+ * BSET INVARIANTS: this function is responsible for maintaining all the
+ * invariants for bsets of extents in memory. things get really hairy with 0
+ * size extents
+ *
+ * within one bset:
+ *
+ * bkey_start_pos(bkey_next(k)) >= k
+ * or bkey_start_offset(bkey_next(k)) >= k->offset
+ *
+ * i.e. strict ordering, no overlapping extents.
+ *
+ * multiple bsets (i.e. full btree node):
+ *
+ * ∀ k, j
+ *   k.size != 0 ∧ j.size != 0 →
+ *     ¬ (k > bkey_start_pos(j) ∧ k < j)
+ *
+ * i.e. no two overlapping keys _of nonzero size_
+ *
+ * We can't realistically maintain this invariant for zero size keys because of
+ * the key merging done in bch_btree_insert_key() - for two mergeable keys k, j
+ * there may be another 0 size key between them in another bset, and it will
+ * thus overlap with the merged key.
+ *
+ * In addition, the end of iter->pos indicates how much has been processed.
+ * If the end of iter->pos is not the same as the end of insert, then
+ * key insertion needs to continue/be retried.
+ */
+enum btree_insert_ret
+bch_insert_fixup_extent(struct btree_insert *trans,
+			struct btree_insert_entry *insert)
+{
+	struct cache_set *c = trans->c;
+	struct btree_iter *iter = insert->iter;
+	struct btree *b = iter->nodes[0];
+	struct btree_node_iter *node_iter = &iter->node_iters[0];
+	struct bkey_packed *_k;
+	struct bkey unpacked;
+	enum btree_insert_ret ret = BTREE_INSERT_OK;
+
+	struct extent_insert_state s = {
+		.trans		= trans,
+		.insert		= insert,
+		.committed	= insert->iter->pos,
+		.deleting	= bkey_whiteout(&insert->k->k),
+	};
+
+	EBUG_ON(iter->level);
+	EBUG_ON(bkey_deleted(&insert->k->k) || !insert->k->k.size);
+
+	if (s.deleting)
+		return bch_delete_fixup_extent(&s);
+
+	/*
+	 * As we process overlapping extents, we advance @iter->pos both to
+	 * signal to our caller (btree_insert_key()) how much of @insert->k has
+	 * been inserted, and also to keep @iter->pos consistent with
+	 * @insert->k and the node iterator that we're advancing:
+	 */
+	EBUG_ON(bkey_cmp(iter->pos, bkey_start_pos(&insert->k->k)));
+
+	if (!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY))
+		bch_add_sectors(&s, bkey_i_to_s_c(insert->k),
+				bkey_start_offset(&insert->k->k),
+				insert->k->k.size);
+
+	while (bkey_cmp(s.committed, insert->k->k.p) < 0 &&
+	       (ret = extent_insert_should_stop(&s)) == BTREE_INSERT_OK &&
+	       (_k = bch_btree_node_iter_peek_all(node_iter, b))) {
+		struct bset_tree *t = bch_bkey_to_bset(b, _k);
+		struct bkey_s k = __bkey_disassemble(b, _k, &unpacked);
+		enum bch_extent_overlap overlap;
+
+		EBUG_ON(bkey_cmp(iter->pos, bkey_start_pos(&insert->k->k)));
+		EBUG_ON(bkey_cmp(iter->pos, k.k->p) >= 0);
+
+		if (bkey_cmp(bkey_start_pos(k.k), insert->k->k.p) >= 0)
+			break;
+
+		overlap = bch_extent_overlap(&insert->k->k, k.k);
+
+		ret = extent_insert_check_split_compressed(&s, k.s_c, overlap);
+		if (ret != BTREE_INSERT_OK)
+			goto stop;
+
+		if (!k.k->size)
+			goto squash;
+
+		/*
+		 * Only call advance pos & call hook for nonzero size extents:
+		 * If hook returned BTREE_HOOK_NO_INSERT, @insert->k no longer
+		 * overlaps with @k:
+		 */
+		switch (extent_insert_advance_pos(&s, k.s_c)) {
+		case BTREE_HOOK_DO_INSERT:
+			break;
+		case BTREE_HOOK_NO_INSERT:
+			continue;
+		case BTREE_HOOK_RESTART_TRANS:
+			ret = BTREE_INSERT_NEED_TRAVERSE;
+			goto stop;
+		}
+
+		if (k.k->size &&
+		    (k.k->needs_whiteout || bset_written(b, bset(b, t))))
+			insert->k->k.needs_whiteout = true;
+
+		if (overlap == BCH_EXTENT_OVERLAP_ALL &&
+		    bkey_whiteout(k.k) &&
+		    k.k->needs_whiteout) {
+			unreserve_whiteout(b, t, _k);
+			_k->needs_whiteout = false;
+		}
+squash:
+		ret = extent_squash(&s, insert->k, t, _k, k, overlap);
+		if (ret != BTREE_INSERT_OK)
+			goto stop;
+	}
+
+	if (bkey_cmp(s.committed, insert->k->k.p) < 0 &&
+	    ret == BTREE_INSERT_OK &&
+	    extent_insert_advance_pos(&s, bkey_s_c_null) == BTREE_HOOK_RESTART_TRANS)
+		ret = BTREE_INSERT_NEED_TRAVERSE;
+stop:
+	extent_insert_committed(&s);
+	/*
+	 * Subtract any remaining sectors from @insert, if we bailed out early
+	 * and didn't fully insert @insert:
+	 */
+	if (insert->k->k.size &&
+	    !(trans->flags & BTREE_INSERT_JOURNAL_REPLAY))
+		bch_subtract_sectors(&s, bkey_i_to_s_c(insert->k),
+				     bkey_start_offset(&insert->k->k),
+				     insert->k->k.size);
+
+	bch_cache_set_stats_apply(c, &s.stats, trans->disk_res,
+				  gc_pos_btree_node(b));
+
+	EBUG_ON(bkey_cmp(iter->pos, bkey_start_pos(&insert->k->k)));
+	EBUG_ON(bkey_cmp(iter->pos, s.committed));
+	EBUG_ON((bkey_cmp(iter->pos, b->key.k.p) == 0) != iter->at_end_of_leaf);
+
+	if (insert->k->k.size && iter->at_end_of_leaf)
+		ret = BTREE_INSERT_NEED_TRAVERSE;
+
+	EBUG_ON(insert->k->k.size && ret == BTREE_INSERT_OK);
+
+	return ret;
+}
+
+static const char *bch_extent_invalid(const struct cache_set *c,
+				      struct bkey_s_c k)
+{
+	if (bkey_val_u64s(k.k) > BKEY_EXTENT_VAL_U64s_MAX)
+		return "value too big";
+
+	if (!k.k->size)
+		return "zero key size";
+
+	switch (k.k->type) {
+	case BCH_EXTENT:
+	case BCH_EXTENT_CACHED: {
+		struct bkey_s_c_extent e = bkey_s_c_to_extent(k);
+		const union bch_extent_entry *entry;
+		const union bch_extent_crc *crc;
+		struct cache_member_rcu *mi = cache_member_info_get(c);
+		unsigned size_ondisk = e.k->size;
+		const char *reason;
+
+		extent_for_each_entry(e, entry) {
+			reason = "invalid extent entry type";
+			if (__extent_entry_type(entry) >= BCH_EXTENT_ENTRY_MAX)
+				goto invalid;
+
+			switch (extent_entry_type(entry)) {
+			case BCH_EXTENT_ENTRY_crc32:
+			case BCH_EXTENT_ENTRY_crc64:
+				crc = entry_to_crc(entry);
+
+				reason = "checksum offset + key size > uncompressed size";
+				if (crc_offset(crc) + e.k->size >
+				    crc_uncompressed_size(e.k, crc))
+					goto invalid;
+
+				size_ondisk = crc_compressed_size(e.k, crc);
+
+				reason = "invalid checksum type";
+				if (crc_csum_type(crc) >= BCH_CSUM_NR)
+					goto invalid;
+
+				reason = "invalid compression type";
+				if (crc_compression_type(crc) >= BCH_COMPRESSION_NR)
+					goto invalid;
+				break;
+			case BCH_EXTENT_ENTRY_ptr:
+				reason = extent_ptr_invalid(e, mi,
+						&entry->ptr, size_ondisk);
+				if (reason)
+					goto invalid;
+				break;
+			}
+		}
+
+		cache_member_info_put();
+		return NULL;
+invalid:
+		cache_member_info_put();
+		return reason;
+	}
+
+	case BCH_RESERVATION:
+		return NULL;
+
+	default:
+		return "invalid value type";
+	}
+}
+
+static void bch_extent_debugcheck_extent(struct cache_set *c, struct btree *b,
+					 struct bkey_s_c_extent e)
+{
+	const struct bch_extent_ptr *ptr;
+	struct cache_member_rcu *mi;
+	struct cache *ca;
+	struct bucket *g;
+	unsigned seq, stale;
+	char buf[160];
+	bool bad;
+	unsigned ptrs_per_tier[CACHE_TIERS];
+	unsigned tier, replicas = 0;
+
+	/*
+	 * XXX: we should be doing most/all of these checks at startup time,
+	 * where we check bkey_invalid() in btree_node_read_done()
+	 *
+	 * But note that we can't check for stale pointers or incorrect gc marks
+	 * until after journal replay is done (it might be an extent that's
+	 * going to get overwritten during replay)
+	 */
+
+	memset(ptrs_per_tier, 0, sizeof(ptrs_per_tier));
+
+	mi = cache_member_info_get(c);
+
+	extent_for_each_ptr(e, ptr) {
+		bool dirty = bch_extent_ptr_is_dirty(c, e, ptr);
+
+		replicas++;
+
+		if (ptr->dev >= mi->nr_in_set)
+			goto bad_device;
+
+		/*
+		 * If journal replay hasn't finished, we might be seeing keys
+		 * that will be overwritten by the time journal replay is done:
+		 */
+		if (!test_bit(JOURNAL_REPLAY_DONE, &c->journal.flags))
+			continue;
+
+		if (!mi->m[ptr->dev].valid)
+			goto bad_device;
+
+		tier = mi->m[ptr->dev].tier;
+		ptrs_per_tier[tier]++;
+
+		stale = 0;
+
+		if ((ca = PTR_CACHE(c, ptr))) {
+			g = PTR_BUCKET(ca, ptr);
+
+			do {
+				struct bucket_mark mark;
+
+				seq = read_seqcount_begin(&c->gc_pos_lock);
+				mark = READ_ONCE(g->mark);
+
+				/* between mark and bucket gen */
+				smp_rmb();
+
+				stale = ptr_stale(ca, ptr);
+
+				cache_set_bug_on(stale && dirty, c,
+						 "stale dirty pointer");
+
+				cache_set_bug_on(stale > 96, c,
+						 "key too stale: %i",
+						 stale);
+
+				if (stale)
+					break;
+
+				bad = (mark.is_metadata ||
+				       (gc_pos_cmp(c->gc_pos, gc_pos_btree_node(b)) > 0 &&
+					!mark.owned_by_allocator &&
+					!(dirty
+					  ? mark.dirty_sectors
+					  : mark.cached_sectors)));
+			} while (read_seqcount_retry(&c->gc_pos_lock, seq));
+
+			if (bad)
+				goto bad_ptr;
+		}
+	}
+	cache_member_info_put();
+
+	if (replicas > BCH_REPLICAS_MAX) {
+		bch_bkey_val_to_text(c, btree_node_type(b), buf,
+				     sizeof(buf), e.s_c);
+		cache_set_bug(c,
+			"extent key bad (too many replicas: %u): %s",
+			replicas, buf);
+		return;
+	}
+
+	if (!bkey_extent_is_cached(e.k) &&
+	    replicas < c->sb.data_replicas_have) {
+		bch_bkey_val_to_text(c, btree_node_type(b), buf,
+				     sizeof(buf), e.s_c);
+		cache_set_bug(c,
+			"extent key bad (too few replicas, %u < %u): %s",
+			replicas, c->sb.data_replicas_have, buf);
+		return;
+	}
+
+	return;
+
+bad_device:
+	bch_bkey_val_to_text(c, btree_node_type(b), buf,
+			     sizeof(buf), e.s_c);
+	cache_set_bug(c, "extent pointer to dev %u missing device: %s",
+		      ptr->dev, buf);
+	cache_member_info_put();
+	return;
+
+bad_ptr:
+	bch_bkey_val_to_text(c, btree_node_type(b), buf,
+			     sizeof(buf), e.s_c);
+	cache_set_bug(c, "extent pointer bad gc mark: %s:\nbucket %zu prio %i "
+		      "gen %i last_gc %i mark 0x%08x",
+		      buf, PTR_BUCKET_NR(ca, ptr),
+		      g->read_prio, PTR_BUCKET(ca, ptr)->mark.gen,
+		      ca->oldest_gens[PTR_BUCKET_NR(ca, ptr)],
+		      (unsigned) g->mark.counter);
+	cache_member_info_put();
+	return;
+}
+
+static void bch_extent_debugcheck(struct cache_set *c, struct btree *b,
+				  struct bkey_s_c k)
+{
+	switch (k.k->type) {
+	case BCH_EXTENT:
+	case BCH_EXTENT_CACHED:
+		bch_extent_debugcheck_extent(c, b, bkey_s_c_to_extent(k));
+	case BCH_RESERVATION:
+		break;
+	default:
+		BUG();
+	}
+}
+
+static void bch_extent_to_text(struct cache_set *c, char *buf,
+			       size_t size, struct bkey_s_c k)
+{
+	char *out = buf, *end = buf + size;
+	const char *invalid;
+
+#define p(...)	(out += scnprintf(out, end - out, __VA_ARGS__))
+
+	if (bkey_extent_is_data(k.k))
+		out += extent_print_ptrs(c, buf, size, bkey_s_c_to_extent(k));
+
+	invalid = bch_extent_invalid(c, k);
+	if (invalid)
+		p(" invalid: %s", invalid);
+#undef p
+}
+
+static unsigned PTR_TIER(struct cache_member_rcu *mi,
+			 const struct bch_extent_ptr *ptr)
+{
+	return ptr->dev < mi->nr_in_set
+		? mi->m[ptr->dev].tier
+		: UINT_MAX;
+}
+
+void bch_extent_entry_append(struct bkey_i_extent *e,
+			     union bch_extent_entry *entry)
+{
+	BUG_ON(bkey_val_u64s(&e->k) + extent_entry_u64s(entry) >
+	       BKEY_EXTENT_VAL_U64s_MAX);
+
+	memcpy_u64s(extent_entry_last(extent_i_to_s(e)),
+		    entry,
+		    extent_entry_u64s(entry));
+	e->k.u64s += extent_entry_u64s(entry);
+}
+
+const unsigned bch_crc_size[] = {
+	[BCH_CSUM_NONE]			= 0,
+	[BCH_CSUM_CRC32C]		= 4,
+	[BCH_CSUM_CRC64]		= 8,
+};
+
+static void bch_extent_crc_init(union bch_extent_crc *crc,
+				unsigned compressed_size,
+				unsigned uncompressed_size,
+				unsigned compression_type,
+				u64 csum, unsigned csum_type)
+{
+	if (bch_crc_size[csum_type] <= 4 &&
+	    uncompressed_size <= CRC32_EXTENT_SIZE_MAX) {
+		crc->crc32 = (struct bch_extent_crc32) {
+			.type = 1 << BCH_EXTENT_ENTRY_crc32,
+			.compressed_size	= compressed_size,
+			.uncompressed_size	= uncompressed_size,
+			.offset			= 0,
+			.compression_type	= compression_type,
+			.csum_type		= csum_type,
+			.csum			= csum,
+		};
+	} else {
+		BUG_ON(uncompressed_size > CRC64_EXTENT_SIZE_MAX);
+
+		crc->crc64 = (struct bch_extent_crc64) {
+			.type = 1 << BCH_EXTENT_ENTRY_crc64,
+			.compressed_size	= compressed_size,
+			.uncompressed_size	= uncompressed_size,
+			.offset			= 0,
+			.compression_type	= compression_type,
+			.csum_type		= csum_type,
+			.csum			= csum,
+		};
+	}
+}
+
+void bch_extent_crc_append(struct bkey_i_extent *e,
+			   unsigned compressed_size,
+			   unsigned uncompressed_size,
+			   unsigned compression_type,
+			   u64 csum, unsigned csum_type)
+{
+	union bch_extent_crc *crc;
+	union bch_extent_crc new;
+
+	BUG_ON(compressed_size > uncompressed_size);
+	BUG_ON(uncompressed_size != e->k.size);
+	BUG_ON(!compressed_size || !uncompressed_size);
+
+	/*
+	 * Look up the last crc entry, so we can check if we need to add
+	 * another:
+	 */
+	extent_for_each_crc(extent_i_to_s(e), crc)
+		;
+
+	switch (extent_crc_type(crc)) {
+	case BCH_EXTENT_CRC_NONE:
+		if (!csum_type && !compression_type)
+			return;
+		break;
+	case BCH_EXTENT_CRC32:
+	case BCH_EXTENT_CRC64:
+		if (crc_compressed_size(&e->k, crc)	== compressed_size &&
+		    crc_uncompressed_size(&e->k, crc)	== uncompressed_size &&
+		    crc_offset(crc)			== 0 &&
+		    crc_compression_type(crc)		== compression_type &&
+		    crc_csum_type(crc)			== csum_type &&
+		    crc_csum(crc)			== csum)
+			return;
+		break;
+	}
+
+	bch_extent_crc_init(&new,
+			    compressed_size,
+			    uncompressed_size,
+			    compression_type,
+			    csum, csum_type);
+	bch_extent_entry_append(e, to_entry(&new));
+}
+
+static void __extent_sort_ptrs(struct cache_member_rcu *mi,
+			       struct bkey_s_extent src)
+{
+	struct bch_extent_ptr *src_ptr, *dst_ptr;
+	union bch_extent_crc *src_crc, *dst_crc;
+	union bch_extent_crc _src;
+	BKEY_PADDED(k) tmp;
+	struct bkey_s_extent dst;
+	size_t u64s, crc_u64s;
+	u64 *p;
+
+	/*
+	 * Insertion sort:
+	 *
+	 * Note: this sort needs to be stable, because pointer order determines
+	 * pointer dirtyness.
+	 */
+
+	tmp.k.k = *src.k;
+	dst = bkey_i_to_s_extent(&tmp.k);
+	set_bkey_val_u64s(dst.k, 0);
+
+	extent_for_each_ptr_crc(src, src_ptr, src_crc) {
+		extent_for_each_ptr_crc(dst, dst_ptr, dst_crc)
+			if (PTR_TIER(mi, src_ptr) < PTR_TIER(mi, dst_ptr))
+				goto found;
+
+		dst_ptr = &extent_entry_last(dst)->ptr;
+		dst_crc = NULL;
+found:
+		/* found insert position: */
+
+		/*
+		 * we're making sure everything has a crc at this point, if
+		 * dst_ptr points to a pointer it better have a crc:
+		 */
+		BUG_ON(dst_ptr != &extent_entry_last(dst)->ptr && !dst_crc);
+		BUG_ON(dst_crc &&
+		       (extent_entry_next(to_entry(dst_crc)) !=
+			to_entry(dst_ptr)));
+
+		if (!src_crc) {
+			bch_extent_crc_init(&_src, src.k->size,
+					    src.k->size, 0, 0, 0);
+			src_crc = &_src;
+		}
+
+		p = dst_ptr != &extent_entry_last(dst)->ptr
+			? (void *) dst_crc
+			: (void *) dst_ptr;
+
+		crc_u64s = extent_entry_u64s(to_entry(src_crc));
+		u64s = crc_u64s + sizeof(*dst_ptr) / sizeof(u64);
+
+		memmove_u64s_up(p + u64s, p,
+				(u64 *) extent_entry_last(dst) - (u64 *) p);
+		set_bkey_val_u64s(dst.k, bkey_val_u64s(dst.k) + u64s);
+
+		memcpy_u64s(p, src_crc, crc_u64s);
+		memcpy_u64s(p + crc_u64s, src_ptr,
+			    sizeof(*src_ptr) / sizeof(u64));
+	}
+
+	/* Sort done - now drop redundant crc entries: */
+	bch_extent_drop_redundant_crcs(dst);
+
+	memcpy_u64s(src.v, dst.v, bkey_val_u64s(dst.k));
+	set_bkey_val_u64s(src.k, bkey_val_u64s(dst.k));
+}
+
+static void extent_sort_ptrs(struct cache_set *c, struct bkey_s_extent e)
+{
+	struct cache_member_rcu *mi;
+	struct bch_extent_ptr *ptr, *prev = NULL;
+	union bch_extent_crc *crc;
+
+	/*
+	 * First check if any pointers are out of order before doing the actual
+	 * sort:
+	 */
+	mi = cache_member_info_get(c);
+
+	extent_for_each_ptr_crc(e, ptr, crc) {
+		if (prev &&
+		    PTR_TIER(mi, ptr) < PTR_TIER(mi, prev)) {
+			__extent_sort_ptrs(mi, e);
+			break;
+		}
+		prev = ptr;
+	}
+
+	cache_member_info_put();
+}
+
+/*
+ * bch_extent_normalize - clean up an extent, dropping stale pointers etc.
+ *
+ * Returns true if @k should be dropped entirely
+ *
+ * For existing keys, only called when btree nodes are being rewritten, not when
+ * they're merely being compacted/resorted in memory.
+ */
+static bool __bch_extent_normalize(struct cache_set *c, struct bkey_s k,
+				   bool sort)
+{
+	struct bkey_s_extent e;
+
+	switch (k.k->type) {
+	case KEY_TYPE_ERROR:
+		return false;
+
+	case KEY_TYPE_DELETED:
+	case KEY_TYPE_COOKIE:
+		return true;
+
+	case KEY_TYPE_DISCARD:
+		return !k.k->version;
+
+	case BCH_EXTENT:
+	case BCH_EXTENT_CACHED:
+		e = bkey_s_to_extent(k);
+
+		bch_extent_drop_stale(c, e);
+
+		if (sort)
+			extent_sort_ptrs(c, e);
+
+		if (!bkey_val_u64s(e.k)) {
+			if (bkey_extent_is_cached(e.k)) {
+				k.k->type = KEY_TYPE_DISCARD;
+				if (!k.k->version)
+					return true;
+			} else {
+				k.k->type = KEY_TYPE_ERROR;
+			}
+		}
+
+		return false;
+	case BCH_RESERVATION:
+		return false;
+	default:
+		BUG();
+	}
+}
+
+bool bch_extent_normalize(struct cache_set *c, struct bkey_s k)
+{
+	return __bch_extent_normalize(c, k, true);
+}
+
+/*
+ * This picks a non-stale pointer, preferabbly from a device other than
+ * avoid.  Avoid can be NULL, meaning pick any.  If there are no non-stale
+ * pointers to other devices, it will still pick a pointer from avoid.
+ * Note that it prefers lowered-numbered pointers to higher-numbered pointers
+ * as the pointers are sorted by tier, hence preferring pointers to tier 0
+ * rather than pointers to tier 1.
+ */
+void bch_extent_pick_ptr_avoiding(struct cache_set *c, struct bkey_s_c k,
+				  struct cache *avoid,
+				  struct extent_pick_ptr *ret)
+{
+	struct bkey_s_c_extent e;
+	const union bch_extent_crc *crc;
+	const struct bch_extent_ptr *ptr;
+	struct cache *ca;
+
+	switch (k.k->type) {
+	case KEY_TYPE_DELETED:
+	case KEY_TYPE_DISCARD:
+	case KEY_TYPE_COOKIE:
+		ret->ca = NULL;
+		return;
+
+	case KEY_TYPE_ERROR:
+		ret->ca = ERR_PTR(-EIO);
+		return;
+
+	case BCH_EXTENT:
+	case BCH_EXTENT_CACHED:
+		e = bkey_s_c_to_extent(k);
+		rcu_read_lock();
+		ret->ca = NULL;
+
+		extent_for_each_online_device_crc(c, e, crc, ptr, ca)
+			if (!ptr_stale(ca, ptr)) {
+				*ret = (struct extent_pick_ptr) {
+					.crc = crc_to_64(e.k, crc),
+					.ptr = *ptr,
+					.ca = ca,
+				};
+
+				if (ca != avoid)
+					break;
+			}
+
+		if (ret->ca)
+			percpu_ref_get(&ret->ca->ref);
+		else if (!bkey_extent_is_cached(e.k))
+			ret->ca = ERR_PTR(-EIO);
+
+		rcu_read_unlock();
+		return;
+
+	case BCH_RESERVATION:
+		ret->ca = NULL;
+		return;
+
+	default:
+		BUG();
+	}
+}
+
+static enum merge_result bch_extent_merge(struct cache_set *c,
+					  struct btree *bk,
+					  struct bkey_i *l, struct bkey_i *r)
+{
+	struct bkey_s_extent el, er;
+	union bch_extent_entry *en_l, *en_r;
+
+	if (key_merging_disabled(c))
+		return BCH_MERGE_NOMERGE;
+
+	/*
+	 * Generic header checks
+	 * Assumes left and right are in order
+	 * Left and right must be exactly aligned
+	 */
+
+	if (l->k.u64s		!= r->k.u64s ||
+	    l->k.type		!= r->k.type ||
+	    l->k.version	!= r->k.version ||
+	    bkey_cmp(l->k.p, bkey_start_pos(&r->k)))
+		return BCH_MERGE_NOMERGE;
+
+	switch (l->k.type) {
+	case KEY_TYPE_DELETED:
+	case KEY_TYPE_DISCARD:
+	case KEY_TYPE_ERROR:
+	case BCH_RESERVATION:
+		/* These types are mergeable, and no val to check */
+		break;
+
+	case BCH_EXTENT:
+	case BCH_EXTENT_CACHED:
+		el = bkey_i_to_s_extent(l);
+		er = bkey_i_to_s_extent(r);
+
+		extent_for_each_entry(el, en_l) {
+			struct bch_extent_ptr *lp, *rp;
+			struct cache_member_cpu *m;
+
+			en_r = bkey_idx(er.v, (u64 *) en_l - el.v->_data);
+
+			if ((extent_entry_type(en_l) !=
+			     extent_entry_type(en_r)) ||
+			    extent_entry_is_crc(en_l))
+				return BCH_MERGE_NOMERGE;
+
+			lp = &en_l->ptr;
+			rp = &en_r->ptr;
+
+			if (lp->offset + el.k->size	!= rp->offset ||
+			    lp->dev			!= rp->dev ||
+			    lp->gen			!= rp->gen)
+				return BCH_MERGE_NOMERGE;
+
+			/* We don't allow extents to straddle buckets: */
+
+			m = cache_member_info_get(c)->m + lp->dev;
+			if ((lp->offset & ~((u64) m->bucket_size - 1)) !=
+			    (rp->offset & ~((u64) m->bucket_size - 1))) {
+				cache_member_info_put();
+				return BCH_MERGE_NOMERGE;
+
+			}
+			cache_member_info_put();
+		}
+
+		break;
+	default:
+		return BCH_MERGE_NOMERGE;
+	}
+
+	l->k.needs_whiteout |= r->k.needs_whiteout;
+
+	/* Keys with no pointers aren't restricted to one bucket and could
+	 * overflow KEY_SIZE
+	 */
+	if ((u64) l->k.size + r->k.size > KEY_SIZE_MAX) {
+		bch_key_resize(&l->k, KEY_SIZE_MAX);
+		bch_cut_front(l->k.p, r);
+		return BCH_MERGE_PARTIAL;
+	}
+
+	bch_key_resize(&l->k, l->k.size + r->k.size);
+
+	return BCH_MERGE_MERGE;
+}
+
+static void extent_i_save(struct btree *b, struct bkey_packed *dst,
+			  struct bkey_i *src)
+{
+	struct bkey_format *f = &b->format;
+	struct bkey_i *dst_unpacked;
+
+	BUG_ON(bkeyp_val_u64s(f, dst) != bkey_val_u64s(&src->k));
+
+	/*
+	 * We don't want the bch_verify_key_order() call in extent_save(),
+	 * because we may be out of order with deleted keys that are about to be
+	 * removed by extent_bset_insert()
+	 */
+
+	if ((dst_unpacked = packed_to_bkey(dst)))
+		bkey_copy(dst_unpacked, src);
+	else
+		BUG_ON(!bkey_pack(dst, src, f));
+}
+
+static bool extent_merge_one_overlapping(struct btree_iter *iter,
+					 struct bpos new_pos,
+					 struct bset_tree *t,
+					 struct bkey_packed *k, struct bkey uk,
+					 bool check, bool could_pack)
+{
+	struct btree *b = iter->nodes[0];
+	struct btree_node_iter *node_iter = &iter->node_iters[0];
+
+	BUG_ON(!bkey_deleted(k));
+
+	if (check) {
+		return !bkey_packed(k) || could_pack;
+	} else {
+		uk.p = new_pos;
+		extent_save(b, node_iter, k, &uk);
+		bch_bset_fix_invalidated_key(b, t, k);
+		bch_btree_node_iter_fix(iter, b, node_iter, t,
+					k, k->u64s, k->u64s);
+		return true;
+	}
+}
+
+static bool extent_merge_do_overlapping(struct btree_iter *iter,
+					struct bkey *m, bool back_merge)
+{
+	struct btree *b = iter->nodes[0];
+	struct btree_node_iter *node_iter = &iter->node_iters[0];
+	struct bset_tree *t;
+	struct bkey_packed *k;
+	struct bkey uk;
+	struct bpos new_pos = back_merge ? m->p : bkey_start_pos(m);
+	bool could_pack = bkey_pack_pos((void *) &uk, new_pos, b);
+	bool check = true;
+
+	/*
+	 * @m is the new merged extent:
+	 *
+	 * The merge took place in the last bset; we know there can't be any 0
+	 * size extents overlapping with m there because if so they would have
+	 * been between the two extents we merged.
+	 *
+	 * But in the other bsets, we have to check for and fix such extents:
+	 */
+do_fixup:
+	for_each_bset(b, t) {
+		if (t == bset_tree_last(b))
+			break;
+
+		/*
+		 * if we don't find this bset in the iterator we already got to
+		 * the end of that bset, so start searching from the end.
+		 */
+		k = bch_btree_node_iter_bset_pos(node_iter, b, t);
+
+		if (k == btree_bkey_last(b, t))
+			k = bkey_prev_all(b, t, k);
+		if (!k)
+			continue;
+
+		if (back_merge) {
+			/*
+			 * Back merge: 0 size extents will be before the key
+			 * that was just inserted (and thus the iterator
+			 * position) - walk backwards to find them
+			 */
+			for (;
+			     k &&
+			     (uk = bkey_unpack_key(b, k),
+			      bkey_cmp(uk.p, bkey_start_pos(m)) > 0);
+			     k = bkey_prev_all(b, t, k)) {
+				if (bkey_cmp(uk.p, m->p) >= 0)
+					continue;
+
+				if (!extent_merge_one_overlapping(iter, new_pos,
+						t, k, uk, check, could_pack))
+					return false;
+			}
+		} else {
+			/* Front merge - walk forwards */
+			for (;
+			     k != btree_bkey_last(b, t) &&
+			     (uk = bkey_unpack_key(b, k),
+			      bkey_cmp(uk.p, m->p) < 0);
+			     k = bkey_next(k)) {
+				if (bkey_cmp(uk.p,
+					     bkey_start_pos(m)) <= 0)
+					continue;
+
+				if (!extent_merge_one_overlapping(iter, new_pos,
+						t, k, uk, check, could_pack))
+					return false;
+			}
+		}
+	}
+
+	if (check) {
+		check = false;
+		goto do_fixup;
+	}
+
+	return true;
+}
+
+/*
+ * When merging an extent that we're inserting into a btree node, the new merged
+ * extent could overlap with an existing 0 size extent - if we don't fix that,
+ * it'll break the btree node iterator so this code finds those 0 size extents
+ * and shifts them out of the way.
+ *
+ * Also unpacks and repacks.
+ */
+static bool bch_extent_merge_inline(struct cache_set *c,
+				    struct btree_iter *iter,
+				    struct bkey_packed *l,
+				    struct bkey_packed *r,
+				    bool back_merge)
+{
+	struct btree *b = iter->nodes[0];
+	struct btree_node_iter *node_iter = &iter->node_iters[0];
+	const struct bkey_format *f = &b->format;
+	struct bset_tree *t = bset_tree_last(b);
+	struct bkey_packed *m;
+	BKEY_PADDED(k) li;
+	BKEY_PADDED(k) ri;
+	struct bkey_i *mi;
+	struct bkey tmp;
+
+	/*
+	 * We need to save copies of both l and r, because we might get a
+	 * partial merge (which modifies both) and then fails to repack
+	 */
+	bkey_unpack(b, &li.k, l);
+	bkey_unpack(b, &ri.k, r);
+
+	m = back_merge ? l : r;
+	mi = back_merge ? &li.k : &ri.k;
+
+	/* l & r should be in last bset: */
+	EBUG_ON(bch_bkey_to_bset(b, m) != t);
+
+	switch (bch_extent_merge(c, b, &li.k, &ri.k)) {
+	case BCH_MERGE_NOMERGE:
+		return false;
+	case BCH_MERGE_PARTIAL:
+		if (bkey_packed(m) && !bkey_pack_key((void *) &tmp, &mi->k, f))
+			return false;
+
+		if (!extent_merge_do_overlapping(iter, &li.k.k, back_merge))
+			return false;
+
+		extent_i_save(b, m, mi);
+		bch_bset_fix_invalidated_key(b, t, m);
+
+		/*
+		 * Update iterator to reflect what we just inserted - otherwise,
+		 * the iter_fix() call is going to put us _before_ the key we
+		 * just partially merged with:
+		 */
+		if (back_merge)
+			bch_btree_iter_set_pos_same_leaf(iter, li.k.k.p);
+
+		bch_btree_node_iter_fix(iter, iter->nodes[0], node_iter,
+					t, m, m->u64s, m->u64s);
+
+		if (!back_merge)
+			bkey_copy(packed_to_bkey(l), &li.k);
+		else
+			bkey_copy(packed_to_bkey(r), &ri.k);
+		return false;
+	case BCH_MERGE_MERGE:
+		if (bkey_packed(m) && !bkey_pack_key((void *) &tmp, &li.k.k, f))
+			return false;
+
+		if (!extent_merge_do_overlapping(iter, &li.k.k, back_merge))
+			return false;
+
+		extent_i_save(b, m, &li.k);
+		bch_bset_fix_invalidated_key(b, t, m);
+
+		bch_btree_node_iter_fix(iter, iter->nodes[0], node_iter,
+					t, m, m->u64s, m->u64s);
+		return true;
+	default:
+		BUG();
+	}
+}
+
+const struct bkey_ops bch_bkey_extent_ops = {
+	.key_invalid	= bch_extent_invalid,
+	.key_debugcheck	= bch_extent_debugcheck,
+	.val_to_text	= bch_extent_to_text,
+	.swab		= bch_ptr_swab,
+	.key_normalize	= bch_ptr_normalize,
+	.key_merge	= bch_extent_merge,
+	.is_extents	= true,
+};