Merge with 1f431b384d bcachefs: Refactor trans_(get|update)_key

1 files changed, 615 insertions, 332 deletions

diff --git a/fs/bcachefs/fsck.c b/fs/bcachefs/fsck.c
index c554a987f3aa..e3738757b6a0 100644
--- a/fs/bcachefs/fsck.c
+++ b/fs/bcachefs/fsck.c
@@ -1,3 +1,4 @@
+// SPDX-License-Identifier: GPL-2.0
 
 #include "bcachefs.h"
 #include "btree_update.h"
@@ -15,9 +16,31 @@
 
 #define QSTR(n) { { { .len = strlen(n) } }, .name = n }
 
-static int remove_dirent(struct bch_fs *c, struct btree_iter *iter,
+static s64 bch2_count_inode_sectors(struct btree_trans *trans, u64 inum)
+{
+	struct btree_iter *iter;
+	struct bkey_s_c k;
+	u64 sectors = 0;
+	int ret;
+
+	for_each_btree_key(trans, iter, BTREE_ID_EXTENTS,
+			   POS(inum, 0), 0, k, ret) {
+		if (k.k->p.inode != inum)
+			break;
+
+		if (bkey_extent_is_allocation(k.k))
+			sectors += k.k->size;
+	}
+
+	bch2_trans_iter_free(trans, iter);
+
+	return ret ?: sectors;
+}
+
+static int remove_dirent(struct btree_trans *trans,
 			 struct bkey_s_c_dirent dirent)
 {
+	struct bch_fs *c = trans->c;
 	struct qstr name;
 	struct bch_inode_unpacked dir_inode;
 	struct bch_hash_info dir_hash_info;
@@ -34,8 +57,8 @@ static int remove_dirent(struct bch_fs *c, struct btree_iter *iter,
 	buf[name.len] = '\0';
 	name.name = buf;
 
-	/* Unlock iter so we don't deadlock, after copying name: */
-	bch2_btree_iter_unlock(iter);
+	/* Unlock so we don't deadlock, after copying name: */
+	bch2_trans_unlock(trans);
 
 	ret = bch2_inode_find_by_inum(c, dir_inum, &dir_inode);
 	if (ret) {
@@ -72,8 +95,9 @@ static int reattach_inode(struct bch_fs *c,
 	bch2_inode_pack(&packed, lostfound_inode);
 
 	ret = bch2_btree_insert(c, BTREE_ID_INODES, &packed.inode.k_i,
-			       NULL, NULL, NULL,
-			       BTREE_INSERT_NOFAIL);
+				NULL, NULL,
+				BTREE_INSERT_NOFAIL|
+				BTREE_INSERT_LAZY_RW);
 	if (ret) {
 		bch_err(c, "error %i reattaching inode %llu while updating lost+found",
 			ret, inum);
@@ -83,7 +107,8 @@ static int reattach_inode(struct bch_fs *c,
 	ret = bch2_dirent_create(c, lostfound_inode->bi_inum,
 				 &lostfound_hash_info,
 				 DT_DIR, &name, inum, NULL,
-				 BTREE_INSERT_NOFAIL);
+				 BTREE_INSERT_NOFAIL|
+				 BTREE_INSERT_LAZY_RW);
 	if (ret) {
 		bch_err(c, "error %i reattaching inode %llu while creating new dirent",
 			ret, inum);
@@ -107,18 +132,21 @@ static struct inode_walker inode_walker_init(void)
 	};
 }
 
-static int walk_inode(struct bch_fs *c, struct inode_walker *w, u64 inum)
+static int walk_inode(struct btree_trans *trans,
+		      struct inode_walker *w, u64 inum)
 {
-	w->first_this_inode	= inum != w->cur_inum;
-	w->cur_inum		= inum;
-
-	if (w->first_this_inode) {
-		int ret = bch2_inode_find_by_inum(c, inum, &w->inode);
+	if (inum != w->cur_inum) {
+		int ret = bch2_inode_find_by_inum_trans(trans, inum,
+							&w->inode);
 
 		if (ret && ret != -ENOENT)
 			return ret;
 
-		w->have_inode = !ret;
+		w->have_inode	= !ret;
+		w->cur_inum	= inum;
+		w->first_this_inode = true;
+	} else {
+		w->first_this_inode = false;
 	}
 
 	return 0;
@@ -126,27 +154,37 @@ static int walk_inode(struct bch_fs *c, struct inode_walker *w, u64 inum)
 
 struct hash_check {
 	struct bch_hash_info	info;
-	struct btree_iter	chain;
-	struct btree_iter	iter;
-	u64			next;
+
+	/* start of current chain of hash collisions: */
+	struct btree_iter	*chain;
+
+	/* next offset in current chain of hash collisions: */
+	u64			chain_end;
 };
 
-static void hash_check_init(const struct bch_hash_desc desc,
-			    struct hash_check *h, struct bch_fs *c)
+static void hash_check_init(struct hash_check *h)
+{
+	h->chain = NULL;
+}
+
+static void hash_stop_chain(struct btree_trans *trans,
+			    struct hash_check *h)
 {
-	bch2_btree_iter_init(&h->chain, c, desc.btree_id, POS_MIN, 0);
-	bch2_btree_iter_init(&h->iter, c, desc.btree_id, POS_MIN, 0);
+	if (h->chain)
+		bch2_trans_iter_free(trans, h->chain);
+	h->chain = NULL;
 }
 
-static void hash_check_set_inode(struct hash_check *h, struct bch_fs *c,
+static void hash_check_set_inode(struct btree_trans *trans,
+				 struct hash_check *h,
 				 const struct bch_inode_unpacked *bi)
 {
-	h->info = bch2_hash_info_init(c, bi);
-	h->next = -1;
+	h->info = bch2_hash_info_init(trans->c, bi);
+	hash_stop_chain(trans, h);
 }
 
 static int hash_redo_key(const struct bch_hash_desc desc,
-			 struct hash_check *h, struct bch_fs *c,
+			 struct btree_trans *trans, struct hash_check *h,
 			 struct btree_iter *k_iter, struct bkey_s_c k,
 			 u64 hashed)
 {
@@ -159,54 +197,142 @@ static int hash_redo_key(const struct bch_hash_desc desc,
 
 	bkey_reassemble(tmp, k);
 
-	ret = bch2_btree_delete_at(k_iter, 0);
+	ret = bch2_btree_delete_at(trans, k_iter, 0);
 	if (ret)
 		goto err;
 
-	bch2_btree_iter_unlock(k_iter);
-
-	bch2_hash_set(desc, &h->info, c, k_iter->pos.inode, NULL, tmp,
-		      BTREE_INSERT_NOFAIL|
-		      BCH_HASH_SET_MUST_CREATE);
+	bch2_hash_set(trans, desc, &h->info, k_iter->pos.inode,
+		      tmp, BCH_HASH_SET_MUST_CREATE);
+	ret = bch2_trans_commit(trans, NULL, NULL,
+				BTREE_INSERT_NOFAIL|
+				BTREE_INSERT_LAZY_RW);
 err:
 	kfree(tmp);
 	return ret;
 }
 
-static int hash_check_key(const struct bch_hash_desc desc,
-			  struct hash_check *h, struct bch_fs *c,
-			  struct btree_iter *k_iter, struct bkey_s_c k)
+static int fsck_hash_delete_at(struct btree_trans *trans,
+			       const struct bch_hash_desc desc,
+			       struct bch_hash_info *info,
+			       struct btree_iter *iter)
+{
+	int ret;
+retry:
+	ret   = bch2_hash_delete_at(trans, desc, info, iter) ?:
+		bch2_trans_commit(trans, NULL, NULL,
+				  BTREE_INSERT_ATOMIC|
+				  BTREE_INSERT_NOFAIL|
+				  BTREE_INSERT_LAZY_RW);
+	if (ret == -EINTR) {
+		ret = bch2_btree_iter_traverse(iter);
+		if (!ret)
+			goto retry;
+	}
+
+	return ret;
+}
+
+static int hash_check_duplicates(struct btree_trans *trans,
+			const struct bch_hash_desc desc, struct hash_check *h,
+			struct btree_iter *k_iter, struct bkey_s_c k)
 {
+	struct bch_fs *c = trans->c;
+	struct btree_iter *iter;
+	struct bkey_s_c k2;
 	char buf[200];
-	u64 hashed;
 	int ret = 0;
 
-	if (k.k->type != desc.whiteout_type &&
-	    k.k->type != desc.key_type)
+	if (!bkey_cmp(h->chain->pos, k_iter->pos))
 		return 0;
 
-	if (k.k->p.offset != h->next) {
-		if (!btree_iter_linked(&h->chain)) {
-			bch2_btree_iter_link(k_iter, &h->chain);
-			bch2_btree_iter_link(k_iter, &h->iter);
+	iter = bch2_trans_copy_iter(trans, h->chain);
+	BUG_ON(IS_ERR(iter));
+
+	for_each_btree_key_continue(iter, 0, k2) {
+		if (bkey_cmp(k2.k->p, k.k->p) >= 0)
+			break;
+
+		if (fsck_err_on(k2.k->type == desc.key_type &&
+				!desc.cmp_bkey(k, k2), c,
+				"duplicate hash table keys:\n%s",
+				(bch2_bkey_val_to_text(&PBUF(buf), c,
+						       k), buf))) {
+			ret = fsck_hash_delete_at(trans, desc, &h->info, k_iter);
+			if (ret)
+				return ret;
+			ret = 1;
+			break;
 		}
-		bch2_btree_iter_copy(&h->chain, k_iter);
 	}
-	h->next = k.k->p.offset + 1;
+fsck_err:
+	bch2_trans_iter_free(trans, iter);
+	return ret;
+}
+
+static void hash_set_chain_start(struct btree_trans *trans,
+			const struct bch_hash_desc desc,
+			struct hash_check *h,
+			struct btree_iter *k_iter, struct bkey_s_c k)
+{
+	bool hole = (k.k->type != KEY_TYPE_whiteout &&
+		     k.k->type != desc.key_type);
+
+	if (hole || k.k->p.offset > h->chain_end + 1)
+		hash_stop_chain(trans, h);
+
+	if (!hole) {
+		if (!h->chain) {
+			h->chain = bch2_trans_copy_iter(trans, k_iter);
+			BUG_ON(IS_ERR(h->chain));
+		}
+
+		h->chain_end = k.k->p.offset;
+	}
+}
+
+static bool key_has_correct_hash(struct btree_trans *trans,
+			const struct bch_hash_desc desc,
+			struct hash_check *h,
+			struct btree_iter *k_iter, struct bkey_s_c k)
+{
+	u64 hash;
+
+	hash_set_chain_start(trans, desc, h, k_iter, k);
+
+	if (k.k->type != desc.key_type)
+		return true;
+
+	hash = desc.hash_bkey(&h->info, k);
+
+	return hash >= h->chain->pos.offset &&
+		hash <= k.k->p.offset;
+}
+
+static int hash_check_key(struct btree_trans *trans,
+			const struct bch_hash_desc desc, struct hash_check *h,
+			struct btree_iter *k_iter, struct bkey_s_c k)
+{
+	struct bch_fs *c = trans->c;
+	char buf[200];
+	u64 hashed;
+	int ret = 0;
+
+	hash_set_chain_start(trans, desc, h, k_iter, k);
 
 	if (k.k->type != desc.key_type)
 		return 0;
 
 	hashed = desc.hash_bkey(&h->info, k);
 
-	if (fsck_err_on(hashed < h->chain.pos.offset ||
+	if (fsck_err_on(hashed < h->chain->pos.offset ||
 			hashed > k.k->p.offset, c,
-			"hash table key at wrong offset: %llu, "
+			"hash table key at wrong offset: btree %u, %llu, "
 			"hashed to %llu chain starts at %llu\n%s",
-			k.k->p.offset, hashed, h->chain.pos.offset,
-			(bch2_bkey_val_to_text(c, bkey_type(0, desc.btree_id),
-					       buf, sizeof(buf), k), buf))) {
-		ret = hash_redo_key(desc, h, c, k_iter, k, hashed);
+			desc.btree_id, k.k->p.offset,
+			hashed, h->chain->pos.offset,
+			(bch2_bkey_val_to_text(&PBUF(buf), c,
+					       k), buf))) {
+		ret = hash_redo_key(desc, trans, h, k_iter, k, hashed);
 		if (ret) {
 			bch_err(c, "hash_redo_key err %i", ret);
 			return ret;
@@ -214,27 +340,101 @@ static int hash_check_key(const struct bch_hash_desc desc,
 		return 1;
 	}
 
-	if (!bkey_cmp(h->chain.pos, k_iter->pos))
+	ret = hash_check_duplicates(trans, desc, h, k_iter, k);
+fsck_err:
+	return ret;
+}
+
+static int check_dirent_hash(struct btree_trans *trans, struct hash_check *h,
+			     struct btree_iter *iter, struct bkey_s_c *k)
+{
+	struct bch_fs *c = trans->c;
+	struct bkey_i_dirent *d = NULL;
+	int ret = -EINVAL;
+	char buf[200];
+	unsigned len;
+	u64 hash;
+
+	if (key_has_correct_hash(trans, bch2_dirent_hash_desc, h, iter, *k))
 		return 0;
 
-	bch2_btree_iter_copy(&h->iter, &h->chain);
-	while (bkey_cmp(h->iter.pos, k_iter->pos) < 0) {
-		struct bkey_s_c k2 = bch2_btree_iter_peek(&h->iter);
+	len = bch2_dirent_name_bytes(bkey_s_c_to_dirent(*k));
+	BUG_ON(!len);
 
-		if (fsck_err_on(k2.k->type == desc.key_type &&
-				!desc.cmp_bkey(k, k2), c,
-				"duplicate hash table keys:\n%s",
-				(bch2_bkey_val_to_text(c, bkey_type(0, desc.btree_id),
-						       buf, sizeof(buf), k), buf))) {
-			ret = bch2_hash_delete_at(desc, &h->info, &h->iter, NULL);
-			if (ret)
-				return ret;
-			return 1;
-		}
-		bch2_btree_iter_next(&h->iter);
+	memcpy(buf, bkey_s_c_to_dirent(*k).v->d_name, len);
+	buf[len] = '\0';
+
+	d = kmalloc(bkey_bytes(k->k), GFP_KERNEL);
+	if (!d) {
+		bch_err(c, "memory allocation failure");
+		return -ENOMEM;
+	}
+
+	bkey_reassemble(&d->k_i, *k);
+
+	do {
+		--len;
+		if (!len)
+			goto err_redo;
+
+		d->k.u64s = BKEY_U64s + dirent_val_u64s(len);
+
+		BUG_ON(bkey_val_bytes(&d->k) <
+		       offsetof(struct bch_dirent, d_name) + len);
+
+		memset(d->v.d_name + len, 0,
+		       bkey_val_bytes(&d->k) -
+		       offsetof(struct bch_dirent, d_name) - len);
+
+		hash = bch2_dirent_hash_desc.hash_bkey(&h->info,
+						bkey_i_to_s_c(&d->k_i));
+	} while (hash < h->chain->pos.offset ||
+		 hash > k->k->p.offset);
+
+	if (fsck_err(c, "dirent with junk at end, was %s (%zu) now %s (%u)",
+		     buf, strlen(buf), d->v.d_name, len)) {
+		bch2_trans_update(trans, BTREE_INSERT_ENTRY(iter, &d->k_i));
+
+		ret = bch2_trans_commit(trans, NULL, NULL,
+					BTREE_INSERT_NOFAIL|
+					BTREE_INSERT_LAZY_RW);
+		if (ret)
+			goto err;
+
+		*k = bch2_btree_iter_peek(iter);
+
+		BUG_ON(k->k->type != KEY_TYPE_dirent);
 	}
+err:
 fsck_err:
+	kfree(d);
 	return ret;
+err_redo:
+	hash = bch2_dirent_hash_desc.hash_bkey(&h->info, *k);
+
+	if (fsck_err(c, "cannot fix dirent by removing trailing garbage %s (%zu)\n"
+		     "hash table key at wrong offset: btree %u, offset %llu, "
+		     "hashed to %llu chain starts at %llu\n%s",
+		     buf, strlen(buf), BTREE_ID_DIRENTS,
+		     k->k->p.offset, hash, h->chain->pos.offset,
+		     (bch2_bkey_val_to_text(&PBUF(buf), c,
+					    *k), buf))) {
+		ret = hash_redo_key(bch2_dirent_hash_desc, trans,
+				    h, iter, *k, hash);
+		if (ret)
+			bch_err(c, "hash_redo_key err %i", ret);
+		else
+			ret = 1;
+	}
+
+	goto err;
+}
+
+static int bch2_inode_truncate(struct bch_fs *c, u64 inode_nr, u64 new_size)
+{
+	return bch2_btree_delete_range(c, BTREE_ID_EXTENTS,
+			POS(inode_nr, round_up(new_size, block_bytes(c)) >> 9),
+			POS(inode_nr + 1, 0), NULL);
 }
 
 /*
@@ -245,17 +445,21 @@ noinline_for_stack
 static int check_extents(struct bch_fs *c)
 {
 	struct inode_walker w = inode_walker_init();
-	struct btree_iter iter;
+	struct btree_trans trans;
+	struct btree_iter *iter;
 	struct bkey_s_c k;
 	u64 i_sectors;
 	int ret = 0;
 
-	for_each_btree_key(&iter, c, BTREE_ID_EXTENTS,
-			   POS(BCACHEFS_ROOT_INO, 0), 0, k) {
-		if (k.k->type == KEY_TYPE_DISCARD)
-			continue;
+	bch2_trans_init(&trans, c, BTREE_ITER_MAX, 0);
+
+	bch_verbose(c, "checking extents");
 
-		ret = walk_inode(c, &w, k.k->p.inode);
+	iter = bch2_trans_get_iter(&trans, BTREE_ID_EXTENTS,
+				   POS(BCACHEFS_ROOT_INO, 0), 0);
+retry:
+	for_each_btree_key_continue(iter, 0, k) {
+		ret = walk_inode(&trans, &w, k.k->p.inode);
 		if (ret)
 			break;
 
@@ -266,9 +470,9 @@ static int check_extents(struct bch_fs *c)
 			!S_ISREG(w.inode.bi_mode) && !S_ISLNK(w.inode.bi_mode), c,
 			"extent type %u for non regular file, inode %llu mode %o",
 			k.k->type, k.k->p.inode, w.inode.bi_mode)) {
-			bch2_btree_iter_unlock(&iter);
+			bch2_trans_unlock(&trans);
 
-			ret = bch2_inode_truncate(c, k.k->p.inode, 0, NULL, NULL);
+			ret = bch2_inode_truncate(c, k.k->p.inode, 0);
 			if (ret)
 				goto err;
 			continue;
@@ -278,44 +482,40 @@ static int check_extents(struct bch_fs *c)
 			w.have_inode &&
 			!(w.inode.bi_flags & BCH_INODE_I_SECTORS_DIRTY) &&
 			w.inode.bi_sectors !=
-			(i_sectors = bch2_count_inode_sectors(c, w.cur_inum)),
+			(i_sectors = bch2_count_inode_sectors(&trans, w.cur_inum)),
 			c, "i_sectors wrong: got %llu, should be %llu",
 			w.inode.bi_sectors, i_sectors)) {
 			struct bkey_inode_buf p;
 
 			w.inode.bi_sectors = i_sectors;
 
-			bch2_btree_iter_unlock(&iter);
+			bch2_trans_unlock(&trans);
 
 			bch2_inode_pack(&p, &w.inode);
 
 			ret = bch2_btree_insert(c, BTREE_ID_INODES,
-						&p.inode.k_i,
-						NULL,
-						NULL,
-						NULL,
-						BTREE_INSERT_NOFAIL);
+						&p.inode.k_i, NULL, NULL,
+						BTREE_INSERT_NOFAIL|
+						BTREE_INSERT_LAZY_RW);
 			if (ret) {
-				bch_err(c, "error in fs gc: error %i "
-					"updating inode", ret);
+				bch_err(c, "error in fsck: error %i updating inode", ret);
 				goto err;
 			}
 
 			/* revalidate iterator: */
-			k = bch2_btree_iter_peek(&iter);
+			k = bch2_btree_iter_peek(iter);
 		}
 
 		if (fsck_err_on(w.have_inode &&
 			!(w.inode.bi_flags & BCH_INODE_I_SIZE_DIRTY) &&
-			k.k->type != BCH_RESERVATION &&
+			k.k->type != KEY_TYPE_reservation &&
 			k.k->p.offset > round_up(w.inode.bi_size, PAGE_SIZE) >> 9, c,
 			"extent type %u offset %llu past end of inode %llu, i_size %llu",
 			k.k->type, k.k->p.offset, k.k->p.inode, w.inode.bi_size)) {
-			bch2_btree_iter_unlock(&iter);
+			bch2_trans_unlock(&trans);
 
 			ret = bch2_inode_truncate(c, k.k->p.inode,
-					round_up(w.inode.bi_size, PAGE_SIZE) >> 9,
-					NULL, NULL);
+						  w.inode.bi_size);
 			if (ret)
 				goto err;
 			continue;
@@ -323,7 +523,9 @@ static int check_extents(struct bch_fs *c)
 	}
 err:
 fsck_err:
-	return bch2_btree_iter_unlock(&iter) ?: ret;
+	if (ret == -EINTR)
+		goto retry;
+	return bch2_trans_exit(&trans) ?: ret;
 }
 
 /*
@@ -335,53 +537,62 @@ static int check_dirents(struct bch_fs *c)
 {
 	struct inode_walker w = inode_walker_init();
 	struct hash_check h;
-	struct btree_iter iter;
+	struct btree_trans trans;
+	struct btree_iter *iter;
 	struct bkey_s_c k;
 	unsigned name_len;
 	char buf[200];
 	int ret = 0;
 
-	hash_check_init(bch2_dirent_hash_desc, &h, c);
+	bch_verbose(c, "checking dirents");
+
+	bch2_trans_init(&trans, c, BTREE_ITER_MAX, 0);
+
+	hash_check_init(&h);
 
-	for_each_btree_key(&iter, c, BTREE_ID_DIRENTS,
-			   POS(BCACHEFS_ROOT_INO, 0), 0, k) {
+	iter = bch2_trans_get_iter(&trans, BTREE_ID_DIRENTS,
+				   POS(BCACHEFS_ROOT_INO, 0), 0);
+retry:
+	for_each_btree_key_continue(iter, 0, k) {
 		struct bkey_s_c_dirent d;
 		struct bch_inode_unpacked target;
 		bool have_target;
 		u64 d_inum;
 
-		ret = walk_inode(c, &w, k.k->p.inode);
+		ret = walk_inode(&trans, &w, k.k->p.inode);
 		if (ret)
 			break;
 
 		if (fsck_err_on(!w.have_inode, c,
 				"dirent in nonexisting directory:\n%s",
-				(bch2_bkey_val_to_text(c, BTREE_ID_DIRENTS,
-						       buf, sizeof(buf), k), buf)) ||
+				(bch2_bkey_val_to_text(&PBUF(buf), c,
+						       k), buf)) ||
 		    fsck_err_on(!S_ISDIR(w.inode.bi_mode), c,
 				"dirent in non directory inode type %u:\n%s",
 				mode_to_type(w.inode.bi_mode),
-				(bch2_bkey_val_to_text(c, BTREE_ID_DIRENTS,
-						       buf, sizeof(buf), k), buf))) {
-			ret = bch2_btree_delete_at(&iter, 0);
+				(bch2_bkey_val_to_text(&PBUF(buf), c,
+						       k), buf))) {
+			ret = bch2_btree_delete_at(&trans, iter, 0);
 			if (ret)
 				goto err;
 			continue;
 		}
 
 		if (w.first_this_inode && w.have_inode)
-			hash_check_set_inode(&h, c, &w.inode);
+			hash_check_set_inode(&trans, &h, &w.inode);
 
-		ret = hash_check_key(bch2_dirent_hash_desc, &h, c, &iter, k);
+		ret = check_dirent_hash(&trans, &h, iter, &k);
 		if (ret > 0) {
 			ret = 0;
 			continue;
 		}
+		if (ret)
+			goto fsck_err;
 
 		if (ret)
 			goto fsck_err;
 
-		if (k.k->type != BCH_DIRENT)
+		if (k.k->type != KEY_TYPE_dirent)
 			continue;
 
 		d = bkey_s_c_to_dirent(k);
@@ -395,8 +606,13 @@ static int check_dirents(struct bch_fs *c)
 				". dirent") ||
 		    fsck_err_on(name_len == 2 &&
 				!memcmp(d.v->d_name, "..", 2), c,
-				".. dirent")) {
-			ret = remove_dirent(c, &iter, d);
+				".. dirent") ||
+		    fsck_err_on(name_len == 2 &&
+				!memcmp(d.v->d_name, "..", 2), c,
+				".. dirent") ||
+		    fsck_err_on(memchr(d.v->d_name, '/', name_len), c,
+				"dirent name has invalid chars")) {
+			ret = remove_dirent(&trans, d);
 			if (ret)
 				goto err;
 			continue;
@@ -404,15 +620,15 @@ static int check_dirents(struct bch_fs *c)
 
 		if (fsck_err_on(d_inum == d.k->p.inode, c,
 				"dirent points to own directory:\n%s",
-				(bch2_bkey_val_to_text(c, BTREE_ID_DIRENTS,
-						       buf, sizeof(buf), k), buf))) {
-			ret = remove_dirent(c, &iter, d);
+				(bch2_bkey_val_to_text(&PBUF(buf), c,
+						       k), buf))) {
+			ret = remove_dirent(&trans, d);
 			if (ret)
 				goto err;
 			continue;
 		}
 
-		ret = bch2_inode_find_by_inum(c, d_inum, &target);
+		ret = bch2_inode_find_by_inum_trans(&trans, d_inum, &target);
 		if (ret && ret != -ENOENT)
 			break;
 
@@ -421,9 +637,9 @@ static int check_dirents(struct bch_fs *c)
 
 		if (fsck_err_on(!have_target, c,
 				"dirent points to missing inode:\n%s",
-				(bch2_bkey_val_to_text(c, BTREE_ID_DIRENTS,
-						       buf, sizeof(buf), k), buf))) {
-			ret = remove_dirent(c, &iter, d);
+				(bch2_bkey_val_to_text(&PBUF(buf), c,
+						       k), buf))) {
+			ret = remove_dirent(&trans, d);
 			if (ret)
 				goto err;
 			continue;
@@ -434,8 +650,8 @@ static int check_dirents(struct bch_fs *c)
 				mode_to_type(target.bi_mode), c,
 				"incorrect d_type: should be %u:\n%s",
 				mode_to_type(target.bi_mode),
-				(bch2_bkey_val_to_text(c, BTREE_ID_DIRENTS,
-						       buf, sizeof(buf), k), buf))) {
+				(bch2_bkey_val_to_text(&PBUF(buf), c,
+						       k), buf))) {
 			struct bkey_i_dirent *n;
 
 			n = kmalloc(bkey_bytes(d.k), GFP_KERNEL);
@@ -447,20 +663,26 @@ static int check_dirents(struct bch_fs *c)
 			bkey_reassemble(&n->k_i, d.s_c);
 			n->v.d_type = mode_to_type(target.bi_mode);
 
-			ret = bch2_btree_insert_at(c, NULL, NULL, NULL,
-					BTREE_INSERT_NOFAIL,
-					BTREE_INSERT_ENTRY(&iter, &n->k_i));
+			bch2_trans_update(&trans,
+				BTREE_INSERT_ENTRY(iter, &n->k_i));
+
+			ret = bch2_trans_commit(&trans, NULL, NULL,
+						BTREE_INSERT_NOFAIL|
+						BTREE_INSERT_LAZY_RW);
 			kfree(n);
 			if (ret)
 				goto err;
 
 		}
 	}
+
+	hash_stop_chain(&trans, &h);
 err:
 fsck_err:
-	bch2_btree_iter_unlock(&h.chain);
-	bch2_btree_iter_unlock(&h.iter);
-	return bch2_btree_iter_unlock(&iter) ?: ret;
+	if (ret == -EINTR)
+		goto retry;
+
+	return bch2_trans_exit(&trans) ?: ret;
 }
 
 /*
@@ -471,39 +693,47 @@ static int check_xattrs(struct bch_fs *c)
 {
 	struct inode_walker w = inode_walker_init();
 	struct hash_check h;
-	struct btree_iter iter;
+	struct btree_trans trans;
+	struct btree_iter *iter;
 	struct bkey_s_c k;
 	int ret = 0;
 
-	hash_check_init(bch2_xattr_hash_desc, &h, c);
+	bch_verbose(c, "checking xattrs");
 
-	for_each_btree_key(&iter, c, BTREE_ID_XATTRS,
-			   POS(BCACHEFS_ROOT_INO, 0), 0, k) {
-		ret = walk_inode(c, &w, k.k->p.inode);
+	hash_check_init(&h);
+
+	bch2_trans_init(&trans, c, BTREE_ITER_MAX, 0);
+
+	iter = bch2_trans_get_iter(&trans, BTREE_ID_XATTRS,
+				   POS(BCACHEFS_ROOT_INO, 0), 0);
+retry:
+	for_each_btree_key_continue(iter, 0, k) {
+		ret = walk_inode(&trans, &w, k.k->p.inode);
 		if (ret)
 			break;
 
 		if (fsck_err_on(!w.have_inode, c,
 				"xattr for missing inode %llu",
 				k.k->p.inode)) {
-			ret = bch2_btree_delete_at(&iter, 0);
+			ret = bch2_btree_delete_at(&trans, iter, 0);
 			if (ret)
 				goto err;
 			continue;
 		}
 
 		if (w.first_this_inode && w.have_inode)
-			hash_check_set_inode(&h, c, &w.inode);
+			hash_check_set_inode(&trans, &h, &w.inode);
 
-		ret = hash_check_key(bch2_xattr_hash_desc, &h, c, &iter, k);
+		ret = hash_check_key(&trans, bch2_xattr_hash_desc,
+				     &h, iter, k);
 		if (ret)
 			goto fsck_err;
 	}
 err:
 fsck_err:
-	bch2_btree_iter_unlock(&h.chain);
-	bch2_btree_iter_unlock(&h.iter);
-	return bch2_btree_iter_unlock(&iter) ?: ret;
+	if (ret == -EINTR)
+		goto retry;
+	return bch2_trans_exit(&trans) ?: ret;
 }
 
 /* Get root directory, create if it doesn't exist: */
@@ -512,6 +742,8 @@ static int check_root(struct bch_fs *c, struct bch_inode_unpacked *root_inode)
 	struct bkey_inode_buf packed;
 	int ret;
 
+	bch_verbose(c, "checking root directory");
+
 	ret = bch2_inode_find_by_inum(c, BCACHEFS_ROOT_INO, root_inode);
 	if (ret && ret != -ENOENT)
 		return ret;
@@ -534,7 +766,9 @@ create_root:
 	bch2_inode_pack(&packed, root_inode);
 
 	return bch2_btree_insert(c, BTREE_ID_INODES, &packed.inode.k_i,
-				 NULL, NULL, NULL, BTREE_INSERT_NOFAIL);
+				 NULL, NULL,
+				 BTREE_INSERT_NOFAIL|
+				 BTREE_INSERT_LAZY_RW);
 }
 
 /* Get lost+found, create if it doesn't exist: */
@@ -549,6 +783,8 @@ static int check_lostfound(struct bch_fs *c,
 	u64 inum;
 	int ret;
 
+	bch_verbose(c, "checking lost+found");
+
 	inum = bch2_dirent_lookup(c, BCACHEFS_ROOT_INO, &root_hash_info,
 				 &lostfound);
 	if (!inum) {
@@ -576,7 +812,9 @@ create_lostfound:
 	bch2_inode_pack(&packed, root_inode);
 
 	ret = bch2_btree_insert(c, BTREE_ID_INODES, &packed.inode.k_i,
-				NULL, NULL, NULL, BTREE_INSERT_NOFAIL);
+				NULL, NULL,
+				BTREE_INSERT_NOFAIL|
+				BTREE_INSERT_LAZY_RW);
 	if (ret)
 		return ret;
 
@@ -590,7 +828,8 @@ create_lostfound:
 
 	ret = bch2_dirent_create(c, BCACHEFS_ROOT_INO, &root_hash_info, DT_DIR,
 				 &lostfound, lostfound_inode->bi_inum, NULL,
-				 BTREE_INSERT_NOFAIL);
+				 BTREE_INSERT_NOFAIL|
+				 BTREE_INSERT_LAZY_RW);
 	if (ret)
 		return ret;
 
@@ -668,13 +907,18 @@ static int check_directory_structure(struct bch_fs *c,
 	struct inode_bitmap dirs_done = { NULL, 0 };
 	struct pathbuf path = { 0, 0, NULL };
 	struct pathbuf_entry *e;
-	struct btree_iter iter;
+	struct btree_trans trans;
+	struct btree_iter *iter;
 	struct bkey_s_c k;
 	struct bkey_s_c_dirent dirent;
 	bool had_unreachable;
 	u64 d_inum;
 	int ret = 0;
 
+	bch2_trans_init(&trans, c, BTREE_ITER_MAX, 0);
+
+	bch_verbose(c, "checking directory structure");
+
 	/* DFS: */
 restart_dfs:
 	had_unreachable = false;
@@ -686,9 +930,8 @@ restart_dfs:
 	}
 
 	ret = path_down(&path, BCACHEFS_ROOT_INO);
-	if (ret) {
-		return ret;
-	}
+	if (ret)
+		goto err;
 
 	while (path.nr) {
 next:
@@ -697,14 +940,14 @@ next:
 		if (e->offset == U64_MAX)
 			goto up;
 
-		for_each_btree_key(&iter, c, BTREE_ID_DIRENTS,
-				   POS(e->inum, e->offset + 1), 0, k) {
+		for_each_btree_key(&trans, iter, BTREE_ID_DIRENTS,
+				   POS(e->inum, e->offset + 1), 0, k, ret) {
 			if (k.k->p.inode != e->inum)
 				break;
 
 			e->offset = k.k->p.offset;
 
-			if (k.k->type != BCH_DIRENT)
+			if (k.k->type != KEY_TYPE_dirent)
 				continue;
 
 			dirent = bkey_s_c_to_dirent(k);
@@ -717,7 +960,7 @@ next:
 			if (fsck_err_on(inode_bitmap_test(&dirs_done, d_inum), c,
 					"directory %llu has multiple hardlinks",
 					d_inum)) {
-				ret = remove_dirent(c, &iter, dirent);
+				ret = remove_dirent(&trans, dirent);
 				if (ret)
 					goto err;
 				continue;
@@ -734,10 +977,14 @@ next:
 				goto err;
 			}
 
-			bch2_btree_iter_unlock(&iter);
+			ret = bch2_trans_iter_free(&trans, iter);
+			if (ret) {
+				bch_err(c, "btree error %i in fsck", ret);
+				goto err;
+			}
 			goto next;
 		}
-		ret = bch2_btree_iter_unlock(&iter);
+		ret = bch2_trans_iter_free(&trans, iter) ?: ret;
 		if (ret) {
 			bch_err(c, "btree error %i in fsck", ret);
 			goto err;
@@ -746,15 +993,25 @@ up:
 		path.nr--;
 	}
 
-	for_each_btree_key(&iter, c, BTREE_ID_INODES, POS_MIN, 0, k) {
-		if (k.k->type != BCH_INODE_FS ||
-		    !S_ISDIR(le16_to_cpu(bkey_s_c_to_inode(k).v->bi_mode)))
+	iter = bch2_trans_get_iter(&trans, BTREE_ID_INODES, POS_MIN, 0);
+retry:
+	for_each_btree_key_continue(iter, 0, k) {
+		if (k.k->type != KEY_TYPE_inode)
+			continue;
+
+		if (!S_ISDIR(le16_to_cpu(bkey_s_c_to_inode(k).v->bi_mode)))
+			continue;
+
+		ret = bch2_empty_dir_trans(&trans, k.k->p.inode);
+		if (ret == -EINTR)
+			goto retry;
+		if (!ret)
 			continue;
 
 		if (fsck_err_on(!inode_bitmap_test(&dirs_done, k.k->p.inode), c,
 				"unreachable directory found (inum %llu)",
 				k.k->p.inode)) {
-			bch2_btree_iter_unlock(&iter);
+			bch2_trans_unlock(&trans);
 
 			ret = reattach_inode(c, lostfound_inode, k.k->p.inode);
 			if (ret) {
@@ -764,7 +1021,7 @@ up:
 			had_unreachable = true;
 		}
 	}
-	ret = bch2_btree_iter_unlock(&iter);
+	ret = bch2_trans_iter_free(&trans, iter);
 	if (ret)
 		goto err;
 
@@ -776,15 +1033,12 @@ up:
 		memset(&path, 0, sizeof(path));
 		goto restart_dfs;
 	}
-
-out:
+err:
+fsck_err:
+	ret = bch2_trans_exit(&trans) ?: ret;
 	kfree(dirs_done.bits);
 	kfree(path.entries);
 	return ret;
-err:
-fsck_err:
-	ret = bch2_btree_iter_unlock(&iter) ?: ret;
-	goto out;
 }
 
 struct nlink {
@@ -805,7 +1059,7 @@ static void inc_link(struct bch_fs *c, nlink_table *links,
 
 	link = genradix_ptr_alloc(links, inum - range_start, GFP_KERNEL);
 	if (!link) {
-		bch_verbose(c, "allocation failed during fs gc - will need another pass");
+		bch_verbose(c, "allocation failed during fsck - will need another pass");
 		*range_end = inum;
 		return;
 	}
@@ -820,17 +1074,20 @@ noinline_for_stack
 static int bch2_gc_walk_dirents(struct bch_fs *c, nlink_table *links,
 			       u64 range_start, u64 *range_end)
 {
-	struct btree_iter iter;
+	struct btree_trans trans;
+	struct btree_iter *iter;
 	struct bkey_s_c k;
 	struct bkey_s_c_dirent d;
 	u64 d_inum;
 	int ret;
 
+	bch2_trans_init(&trans, c, BTREE_ITER_MAX, 0);
+
 	inc_link(c, links, range_start, range_end, BCACHEFS_ROOT_INO, false);
 
-	for_each_btree_key(&iter, c, BTREE_ID_DIRENTS, POS_MIN, 0, k) {
+	for_each_btree_key(&trans, iter, BTREE_ID_DIRENTS, POS_MIN, 0, k, ret) {
 		switch (k.k->type) {
-		case BCH_DIRENT:
+		case KEY_TYPE_dirent:
 			d = bkey_s_c_to_dirent(k);
 			d_inum = le64_to_cpu(d.v->d_inum);
 
@@ -844,115 +1101,169 @@ static int bch2_gc_walk_dirents(struct bch_fs *c, nlink_table *links,
 			break;
 		}
 
-		bch2_btree_iter_cond_resched(&iter);
+		bch2_trans_cond_resched(&trans);
 	}
-	ret = bch2_btree_iter_unlock(&iter);
+	ret = bch2_trans_exit(&trans) ?: ret;
 	if (ret)
-		bch_err(c, "error in fs gc: btree error %i while walking dirents", ret);
+		bch_err(c, "error in fsck: btree error %i while walking dirents", ret);
 
 	return ret;
 }
 
-s64 bch2_count_inode_sectors(struct bch_fs *c, u64 inum)
+static int check_inode_nlink(struct bch_fs *c,
+			     struct bch_inode_unpacked *lostfound_inode,
+			     struct bch_inode_unpacked *u,
+			     struct nlink *link,
+			     bool *do_update)
 {
-	struct btree_iter iter;
-	struct bkey_s_c k;
-	u64 sectors = 0;
+	u32 i_nlink = u->bi_flags & BCH_INODE_UNLINKED
+		? 0
+		: u->bi_nlink + nlink_bias(u->bi_mode);
+	u32 real_i_nlink =
+		link->count * nlink_bias(u->bi_mode) +
+		link->dir_count;
+	int ret = 0;
 
-	for_each_btree_key(&iter, c, BTREE_ID_EXTENTS, POS(inum, 0), 0, k) {
-		if (k.k->p.inode != inum)
-			break;
+	/*
+	 * These should have been caught/fixed by earlier passes, we don't
+	 * repair them here:
+	 */
+	if (S_ISDIR(u->bi_mode) && link->count > 1) {
+		need_fsck_err(c, "directory %llu with multiple hardlinks: %u",
+			      u->bi_inum, link->count);
+		return 0;
+	}
 
-		if (bkey_extent_is_allocation(k.k))
-			sectors += k.k->size;
+	if (S_ISDIR(u->bi_mode) && !link->count) {
+		need_fsck_err(c, "unreachable directory found (inum %llu)",
+			      u->bi_inum);
+		return 0;
+	}
+
+	if (!S_ISDIR(u->bi_mode) && link->dir_count) {
+		need_fsck_err(c, "non directory with subdirectories",
+			      u->bi_inum);
+		return 0;
+	}
+
+	if (!link->count &&
+	    !(u->bi_flags & BCH_INODE_UNLINKED) &&
+	    (c->sb.features & (1 << BCH_FEATURE_ATOMIC_NLINK))) {
+		if (fsck_err(c, "unreachable inode %llu not marked as unlinked (type %u)",
+			     u->bi_inum, mode_to_type(u->bi_mode)) ==
+		    FSCK_ERR_IGNORE)
+			return 0;
+
+		ret = reattach_inode(c, lostfound_inode, u->bi_inum);
+		if (ret)
+			return ret;
+
+		link->count = 1;
+		real_i_nlink = nlink_bias(u->bi_mode) + link->dir_count;
+		goto set_i_nlink;
+	}
+
+	if (i_nlink < link->count) {
+		if (fsck_err(c, "inode %llu i_link too small (%u < %u, type %i)",
+			     u->bi_inum, i_nlink, link->count,
+			     mode_to_type(u->bi_mode)) == FSCK_ERR_IGNORE)
+			return 0;
+		goto set_i_nlink;
+	}
+
+	if (i_nlink != real_i_nlink &&
+	    c->sb.clean) {
+		if (fsck_err(c, "filesystem marked clean, "
+			     "but inode %llu has wrong i_nlink "
+			     "(type %u i_nlink %u, should be %u)",
+			     u->bi_inum, mode_to_type(u->bi_mode),
+			     i_nlink, real_i_nlink) == FSCK_ERR_IGNORE)
+			return 0;
+		goto set_i_nlink;
 	}
 
-	return bch2_btree_iter_unlock(&iter) ?: sectors;
+	if (i_nlink != real_i_nlink &&
+	    (c->sb.features & (1 << BCH_FEATURE_ATOMIC_NLINK))) {
+		if (fsck_err(c, "inode %llu has wrong i_nlink "
+			     "(type %u i_nlink %u, should be %u)",
+			     u->bi_inum, mode_to_type(u->bi_mode),
+			     i_nlink, real_i_nlink) == FSCK_ERR_IGNORE)
+			return 0;
+		goto set_i_nlink;
+	}
+
+	if (real_i_nlink && i_nlink != real_i_nlink)
+		bch_verbose(c, "setting inode %llu nlink from %u to %u",
+			    u->bi_inum, i_nlink, real_i_nlink);
+set_i_nlink:
+	if (i_nlink != real_i_nlink) {
+		if (real_i_nlink) {
+			u->bi_nlink = real_i_nlink - nlink_bias(u->bi_mode);
+			u->bi_flags &= ~BCH_INODE_UNLINKED;
+		} else {
+			u->bi_nlink = 0;
+			u->bi_flags |= BCH_INODE_UNLINKED;
+		}
+
+		*do_update = true;
+	}
+fsck_err:
+	return ret;
 }
 
-static int bch2_gc_do_inode(struct bch_fs *c,
-			   struct bch_inode_unpacked *lostfound_inode,
-			   struct btree_iter *iter,
-			   struct bkey_s_c_inode inode, struct nlink link)
+static int check_inode(struct btree_trans *trans,
+		       struct bch_inode_unpacked *lostfound_inode,
+		       struct btree_iter *iter,
+		       struct bkey_s_c_inode inode,
+		       struct nlink *link)
 {
+	struct bch_fs *c = trans->c;
 	struct bch_inode_unpacked u;
-	int ret = 0;
-	u32 i_nlink, real_i_nlink;
 	bool do_update = false;
+	int ret = 0;
 
 	ret = bch2_inode_unpack(inode, &u);
+
+	bch2_trans_unlock(trans);
+
 	if (bch2_fs_inconsistent_on(ret, c,
 			 "error unpacking inode %llu in fsck",
 			 inode.k->p.inode))
 		return ret;
 
-	i_nlink = u.bi_nlink + nlink_bias(u.bi_mode);
-
-	fsck_err_on(i_nlink < link.count, c,
-		    "inode %llu i_link too small (%u < %u, type %i)",
-		    inode.k->p.inode, i_nlink,
-		    link.count, mode_to_type(u.bi_mode));
-
-	/* These should have been caught/fixed by earlier passes: */
-	if (S_ISDIR(u.bi_mode)) {
-		need_fsck_err_on(link.count > 1, c,
-			"directory %llu with multiple hardlinks: %u",
-			inode.k->p.inode, link.count);
-
-		real_i_nlink = link.count * 2 + link.dir_count;
-	} else {
-		need_fsck_err_on(link.dir_count, c,
-			"found dirents for non directory %llu",
-			inode.k->p.inode);
-
-		real_i_nlink = link.count + link.dir_count;
+	if (link) {
+		ret = check_inode_nlink(c, lostfound_inode, &u, link,
+					&do_update);
+		if (ret)
+			return ret;
 	}
 
-	if (!link.count) {
-		fsck_err_on(c->sb.clean, c,
-			    "filesystem marked clean, "
-			    "but found orphaned inode %llu",
-			    inode.k->p.inode);
-
-		if (fsck_err_on(S_ISDIR(u.bi_mode) &&
-				bch2_empty_dir(c, inode.k->p.inode), c,
-				"non empty directory with link count 0, "
-				"inode nlink %u, dir links found %u",
-				i_nlink, link.dir_count)) {
-			ret = reattach_inode(c, lostfound_inode,
-					     inode.k->p.inode);
-			if (ret)
-				return ret;
-		}
-
-		bch_verbose(c, "deleting inode %llu", inode.k->p.inode);
+	if (u.bi_flags & BCH_INODE_UNLINKED &&
+	    (!c->sb.clean ||
+	     fsck_err(c, "filesystem marked clean, but inode %llu unlinked",
+		      u.bi_inum))) {
+		bch_verbose(c, "deleting inode %llu", u.bi_inum);
 
-		ret = bch2_inode_rm(c, inode.k->p.inode);
+		ret = bch2_inode_rm(c, u.bi_inum);
 		if (ret)
-			bch_err(c, "error in fs gc: error %i "
-				"while deleting inode", ret);
+			bch_err(c, "error in fsck: error %i while deleting inode", ret);
 		return ret;
 	}
 
-	if (u.bi_flags & BCH_INODE_I_SIZE_DIRTY) {
-		fsck_err_on(c->sb.clean, c,
-			    "filesystem marked clean, "
-			    "but inode %llu has i_size dirty",
-			    inode.k->p.inode);
-
-		bch_verbose(c, "truncating inode %llu", inode.k->p.inode);
+	if (u.bi_flags & BCH_INODE_I_SIZE_DIRTY &&
+	    (!c->sb.clean ||
+	     fsck_err(c, "filesystem marked clean, but inode %llu has i_size dirty",
+		      u.bi_inum))) {
+		bch_verbose(c, "truncating inode %llu", u.bi_inum);
 
 		/*
 		 * XXX: need to truncate partial blocks too here - or ideally
 		 * just switch units to bytes and that issue goes away
 		 */
 
-		ret = bch2_inode_truncate(c, inode.k->p.inode,
-				round_up(u.bi_size, PAGE_SIZE) >> 9,
-				NULL, NULL);
+		ret = bch2_inode_truncate(c, u.bi_inum, u.bi_size);
 		if (ret) {
-			bch_err(c, "error in fs gc: error %i "
-				"truncating inode", ret);
+			bch_err(c, "error in fsck: error %i truncating inode", ret);
 			return ret;
 		}
 
@@ -966,21 +1277,18 @@ static int bch2_gc_do_inode(struct bch_fs *c,
 		do_update = true;
 	}
 
-	if (u.bi_flags & BCH_INODE_I_SECTORS_DIRTY) {
+	if (u.bi_flags & BCH_INODE_I_SECTORS_DIRTY &&
+	    (!c->sb.clean ||
+	     fsck_err(c, "filesystem marked clean, but inode %llu has i_sectors dirty",
+		      u.bi_inum))) {
 		s64 sectors;
 
-		fsck_err_on(c->sb.clean, c,
-			    "filesystem marked clean, "
-			    "but inode %llu has i_sectors dirty",
-			    inode.k->p.inode);
-
 		bch_verbose(c, "recounting sectors for inode %llu",
-			    inode.k->p.inode);
+			    u.bi_inum);
 
-		sectors = bch2_count_inode_sectors(c, inode.k->p.inode);
+		sectors = bch2_count_inode_sectors(trans, u.bi_inum);
 		if (sectors < 0) {
-			bch_err(c, "error in fs gc: error %i "
-				"recounting inode sectors",
+			bch_err(c, "error in fsck: error %i recounting inode sectors",
 				(int) sectors);
 			return sectors;
 		}
@@ -990,30 +1298,17 @@ static int bch2_gc_do_inode(struct bch_fs *c,
 		do_update = true;
 	}
 
-	if (i_nlink != real_i_nlink) {
-		fsck_err_on(c->sb.clean, c,
-			    "filesystem marked clean, "
-			    "but inode %llu has wrong i_nlink "
-			    "(type %u i_nlink %u, should be %u)",
-			    inode.k->p.inode, mode_to_type(u.bi_mode),
-			    i_nlink, real_i_nlink);
-
-		bch_verbose(c, "setting inode %llu nlinks from %u to %u",
-			    inode.k->p.inode, i_nlink, real_i_nlink);
-		u.bi_nlink = real_i_nlink - nlink_bias(u.bi_mode);
-		do_update = true;
-	}
-
 	if (do_update) {
 		struct bkey_inode_buf p;
 
 		bch2_inode_pack(&p, &u);
+		bch2_trans_update(trans, BTREE_INSERT_ENTRY(iter, &p.inode.k_i));
 
-		ret = bch2_btree_insert_at(c, NULL, NULL, NULL,
-					  BTREE_INSERT_NOFAIL,
-					  BTREE_INSERT_ENTRY(iter, &p.inode.k_i));
+		ret = bch2_trans_commit(trans, NULL, NULL,
+					BTREE_INSERT_NOFAIL|
+					BTREE_INSERT_LAZY_RW);
 		if (ret && ret != -EINTR)
-			bch_err(c, "error in fs gc: error %i "
+			bch_err(c, "error in fsck: error %i "
 				"updating inode", ret);
 	}
 fsck_err:
@@ -1022,29 +1317,33 @@ fsck_err:
 
 noinline_for_stack
 static int bch2_gc_walk_inodes(struct bch_fs *c,
-			      struct bch_inode_unpacked *lostfound_inode,
-			      nlink_table *links,
-			      u64 range_start, u64 range_end)
+			       struct bch_inode_unpacked *lostfound_inode,
+			       nlink_table *links,
+			       u64 range_start, u64 range_end)
 {
-	struct btree_iter iter;
+	struct btree_trans trans;
+	struct btree_iter *iter;
 	struct bkey_s_c k;
 	struct nlink *link, zero_links = { 0, 0 };
 	struct genradix_iter nlinks_iter;
 	int ret = 0, ret2 = 0;
 	u64 nlinks_pos;
 
-	bch2_btree_iter_init(&iter, c, BTREE_ID_INODES, POS(range_start, 0), 0);
+	bch2_trans_init(&trans, c, BTREE_ITER_MAX, 0);
+
+	iter = bch2_trans_get_iter(&trans, BTREE_ID_INODES,
+				   POS(range_start, 0), 0);
 	nlinks_iter = genradix_iter_init(links, 0);
 
-	while ((k = bch2_btree_iter_peek(&iter)).k &&
-	       !btree_iter_err(k)) {
+	while ((k = bch2_btree_iter_peek(iter)).k &&
+	       !(ret2 = bkey_err(k))) {
 peek_nlinks:	link = genradix_iter_peek(&nlinks_iter, links);
 
-		if (!link && (!k.k || iter.pos.inode >= range_end))
+		if (!link && (!k.k || iter->pos.inode >= range_end))
 			break;
 
 		nlinks_pos = range_start + nlinks_iter.pos;
-		if (iter.pos.inode > nlinks_pos) {
+		if (iter->pos.inode > nlinks_pos) {
 			/* Should have been caught by dirents pass: */
 			need_fsck_err_on(link && link->count, c,
 				"missing inode %llu (nlink %u)",
@@ -1053,25 +1352,15 @@ peek_nlinks:	link = genradix_iter_peek(&nlinks_iter, links);
 			goto peek_nlinks;
 		}
 
-		if (iter.pos.inode < nlinks_pos || !link)
+		if (iter->pos.inode < nlinks_pos || !link)
 			link = &zero_links;
 
-		if (k.k && k.k->type == BCH_INODE_FS) {
-			/*
-			 * Avoid potential deadlocks with iter for
-			 * truncate/rm/etc.:
-			 */
-			bch2_btree_iter_unlock(&iter);
-
-			ret = bch2_gc_do_inode(c, lostfound_inode, &iter,
-					      bkey_s_c_to_inode(k), *link);
-			if (ret == -EINTR)
-				continue;
+		if (k.k && k.k->type == KEY_TYPE_inode) {
+			ret = check_inode(&trans, lostfound_inode, iter,
+					  bkey_s_c_to_inode(k), link);
+			BUG_ON(ret == -EINTR);
 			if (ret)
 				break;
-
-			if (link->count)
-				atomic_long_inc(&c->nr_inodes);
 		} else {
 			/* Should have been caught by dirents pass: */
 			need_fsck_err_on(link->count, c,
@@ -1079,16 +1368,17 @@ peek_nlinks:	link = genradix_iter_peek(&nlinks_iter, links);
 				nlinks_pos, link->count);
 		}
 
-		if (nlinks_pos == iter.pos.inode)
+		if (nlinks_pos == iter->pos.inode)
 			genradix_iter_advance(&nlinks_iter, links);
 
-		bch2_btree_iter_next(&iter);
-		bch2_btree_iter_cond_resched(&iter);
+		bch2_btree_iter_next(iter);
+		bch2_trans_cond_resched(&trans);
 	}
 fsck_err:
-	ret2 = bch2_btree_iter_unlock(&iter);
+	bch2_trans_exit(&trans);
+
 	if (ret2)
-		bch_err(c, "error in fs gc: btree error %i while walking inodes", ret2);
+		bch_err(c, "error in fsck: btree error %i while walking inodes", ret2);
 
 	return ret ?: ret2;
 }
@@ -1101,6 +1391,8 @@ static int check_inode_nlinks(struct bch_fs *c,
 	u64 this_iter_range_start, next_iter_range_start = 0;
 	int ret = 0;
 
+	bch_verbose(c, "checking inode nlinks");
+
 	genradix_init(&links);
 
 	do {
@@ -1131,64 +1423,55 @@ static int check_inode_nlinks(struct bch_fs *c,
  * Checks for inconsistencies that shouldn't happen, unless we have a bug.
  * Doesn't fix them yet, mainly because they haven't yet been observed:
  */
-int bch2_fsck(struct bch_fs *c, bool full_fsck)
+int bch2_fsck_full(struct bch_fs *c)
 {
 	struct bch_inode_unpacked root_inode, lostfound_inode;
-	int ret;
-
-	if (full_fsck) {
-		bch_verbose(c, "checking extents");
-		ret = check_extents(c);
-		if (ret)
-			return ret;
 
-		bch_verbose(c, "checking dirents");
-		ret = check_dirents(c);
-		if (ret)
-			return ret;
+	return  check_extents(c) ?:
+		check_dirents(c) ?:
+		check_xattrs(c) ?:
+		check_root(c, &root_inode) ?:
+		check_lostfound(c, &root_inode, &lostfound_inode) ?:
+		check_directory_structure(c, &lostfound_inode) ?:
+		check_inode_nlinks(c, &lostfound_inode);
+}
 
-		bch_verbose(c, "checking xattrs");
-		ret = check_xattrs(c);
-		if (ret)
-			return ret;
+int bch2_fsck_inode_nlink(struct bch_fs *c)
+{
+	struct bch_inode_unpacked root_inode, lostfound_inode;
 
-		bch_verbose(c, "checking root directory");
-		ret = check_root(c, &root_inode);
-		if (ret)
-			return ret;
+	return  check_root(c, &root_inode) ?:
+		check_lostfound(c, &root_inode, &lostfound_inode) ?:
+		check_inode_nlinks(c, &lostfound_inode);
+}
 
-		bch_verbose(c, "checking lost+found");
-		ret = check_lostfound(c, &root_inode, &lostfound_inode);
-		if (ret)
-			return ret;
+int bch2_fsck_walk_inodes_only(struct bch_fs *c)
+{
+	struct btree_trans trans;
+	struct btree_iter *iter;
+	struct bkey_s_c k;
+	struct bkey_s_c_inode inode;
+	int ret;
 
-		bch_verbose(c, "checking directory structure");
-		ret = check_directory_structure(c, &lostfound_inode);
-		if (ret)
-			return ret;
+	bch2_trans_init(&trans, c, BTREE_ITER_MAX, 0);
 
-		bch_verbose(c, "checking inode nlinks");
-		ret = check_inode_nlinks(c, &lostfound_inode);
-		if (ret)
-			return ret;
-	} else {
-		bch_verbose(c, "checking root directory");
-		ret = check_root(c, &root_inode);
-		if (ret)
-			return ret;
+	for_each_btree_key(&trans, iter, BTREE_ID_INODES, POS_MIN, 0, k, ret) {
+		if (k.k->type != KEY_TYPE_inode)
+			continue;
 
-		bch_verbose(c, "checking lost+found");
-		ret = check_lostfound(c, &root_inode, &lostfound_inode);
-		if (ret)
-			return ret;
+		inode = bkey_s_c_to_inode(k);
 
-		bch_verbose(c, "checking inode nlinks");
-		ret = check_inode_nlinks(c, &lostfound_inode);
-		if (ret)
-			return ret;
+		if (inode.v->bi_flags &
+		    (BCH_INODE_I_SIZE_DIRTY|
+		     BCH_INODE_I_SECTORS_DIRTY|
+		     BCH_INODE_UNLINKED)) {
+			ret = check_inode(&trans, NULL, iter, inode, NULL);
+			BUG_ON(ret == -EINTR);
+			if (ret)
+				break;
+		}
 	}
+	BUG_ON(ret == -EINTR);
 
-	bch2_flush_fsck_errs(c);
-
-	return 0;
+	return bch2_trans_exit(&trans) ?: ret;
 }