X-Git-Url: https://git.sesse.net/?a=blobdiff_plain;f=libbcachefs%2Fdirent.c;h=d37bd07afbfe4088ebb9b92feb56ff86127ae1ab;hb=HEAD;hp=056715bc3ebb912e189a105ecbd6f6adfc0d5f6c;hpb=c35fbbc025c6099969befb4dfaf065215cf40cf3;p=bcachefs-tools-debian

diff --git a/libbcachefs/dirent.c b/libbcachefs/dirent.c
index 056715b..d37bd07 100644
--- a/libbcachefs/dirent.c
+++ b/libbcachefs/dirent.c
@@ -1,5 +1,7 @@
+// SPDX-License-Identifier: GPL-2.0
 
 #include "bcachefs.h"
+#include "bkey_buf.h"
 #include "bkey_methods.h"
 #include "btree_update.h"
 #include "extents.h"
@@ -7,22 +9,29 @@
 #include "fs.h"
 #include "keylist.h"
 #include "str_hash.h"
+#include "subvolume.h"
 
 #include <linux/dcache.h>
 
-unsigned bch2_dirent_name_bytes(struct bkey_s_c_dirent d)
+static unsigned bch2_dirent_name_bytes(struct bkey_s_c_dirent d)
 {
-	unsigned len = bkey_val_bytes(d.k) - sizeof(struct bch_dirent);
-
-	while (len && !d.v->d_name[len - 1])
-		--len;
-
-	return len;
+	unsigned bkey_u64s = bkey_val_u64s(d.k);
+	unsigned bkey_bytes = bkey_u64s * sizeof(u64);
+	u64 last_u64 = ((u64*)d.v)[bkey_u64s - 1];
+#if CPU_BIG_ENDIAN
+	unsigned trailing_nuls = last_u64 ? __builtin_ctzll(last_u64) / 8 : 64 / 8;
+#else
+	unsigned trailing_nuls = last_u64 ? __builtin_clzll(last_u64) / 8 : 64 / 8;
+#endif
+
+	return bkey_bytes -
+		offsetof(struct bch_dirent, d_name) -
+		trailing_nuls;
 }
 
-static unsigned dirent_val_u64s(unsigned len)
+struct qstr bch2_dirent_get_name(struct bkey_s_c_dirent d)
 {
-	return DIV_ROUND_UP(sizeof(struct bch_dirent) + len, sizeof(u64));
+	return (struct qstr) QSTR_INIT(d.v->d_name, bch2_dirent_name_bytes(d));
 }
 
 static u64 bch2_dirent_hash(const struct bch_hash_info *info,
@@ -45,7 +54,7 @@ static u64 dirent_hash_key(const struct bch_hash_info *info, const void *key)
 static u64 dirent_hash_bkey(const struct bch_hash_info *info, struct bkey_s_c k)
 {
 	struct bkey_s_c_dirent d = bkey_s_c_to_dirent(k);
-	struct qstr name = QSTR_INIT(d.v->d_name, bch2_dirent_name_bytes(d));
+	struct qstr name = bch2_dirent_get_name(d);
 
 	return bch2_dirent_hash(info, &name);
 }
@@ -53,134 +62,189 @@ static u64 dirent_hash_bkey(const struct bch_hash_info *info, struct bkey_s_c k)
 static bool dirent_cmp_key(struct bkey_s_c _l, const void *_r)
 {
 	struct bkey_s_c_dirent l = bkey_s_c_to_dirent(_l);
-	int len = bch2_dirent_name_bytes(l);
-	const struct qstr *r = _r;
+	const struct qstr l_name = bch2_dirent_get_name(l);
+	const struct qstr *r_name = _r;
 
-	return len - r->len ?: memcmp(l.v->d_name, r->name, len);
+	return !qstr_eq(l_name, *r_name);
 }
 
 static bool dirent_cmp_bkey(struct bkey_s_c _l, struct bkey_s_c _r)
 {
 	struct bkey_s_c_dirent l = bkey_s_c_to_dirent(_l);
 	struct bkey_s_c_dirent r = bkey_s_c_to_dirent(_r);
-	int l_len = bch2_dirent_name_bytes(l);
-	int r_len = bch2_dirent_name_bytes(r);
+	const struct qstr l_name = bch2_dirent_get_name(l);
+	const struct qstr r_name = bch2_dirent_get_name(r);
+
+	return !qstr_eq(l_name, r_name);
+}
 
-	return l_len - r_len ?: memcmp(l.v->d_name, r.v->d_name, l_len);
+static bool dirent_is_visible(subvol_inum inum, struct bkey_s_c k)
+{
+	struct bkey_s_c_dirent d = bkey_s_c_to_dirent(k);
+
+	if (d.v->d_type == DT_SUBVOL)
+		return le32_to_cpu(d.v->d_parent_subvol) == inum.subvol;
+	return true;
 }
 
 const struct bch_hash_desc bch2_dirent_hash_desc = {
-	.btree_id	= BTREE_ID_DIRENTS,
-	.key_type	= BCH_DIRENT,
-	.whiteout_type	= BCH_DIRENT_WHITEOUT,
+	.btree_id	= BTREE_ID_dirents,
+	.key_type	= KEY_TYPE_dirent,
 	.hash_key	= dirent_hash_key,
 	.hash_bkey	= dirent_hash_bkey,
 	.cmp_key	= dirent_cmp_key,
 	.cmp_bkey	= dirent_cmp_bkey,
+	.is_visible	= dirent_is_visible,
 };
 
-static const char *bch2_dirent_invalid(const struct bch_fs *c,
-				       struct bkey_s_c k)
+int bch2_dirent_invalid(struct bch_fs *c, struct bkey_s_c k,
+			enum bkey_invalid_flags flags,
+			struct printbuf *err)
 {
-	struct bkey_s_c_dirent d;
-	unsigned len;
-
-	switch (k.k->type) {
-	case BCH_DIRENT:
-		if (bkey_val_bytes(k.k) < sizeof(struct bch_dirent))
-			return "value too small";
-
-		d = bkey_s_c_to_dirent(k);
-		len = bch2_dirent_name_bytes(d);
+	struct bkey_s_c_dirent d = bkey_s_c_to_dirent(k);
+	struct qstr d_name = bch2_dirent_get_name(d);
+	int ret = 0;
 
-		if (!len)
-			return "empty name";
+	bkey_fsck_err_on(!d_name.len, c, err,
+			 dirent_empty_name,
+			 "empty name");
 
-		if (bkey_val_u64s(k.k) > dirent_val_u64s(len))
-			return "value too big";
+	bkey_fsck_err_on(bkey_val_u64s(k.k) > dirent_val_u64s(d_name.len), c, err,
+			 dirent_val_too_big,
+			 "value too big (%zu > %u)",
+			 bkey_val_u64s(k.k), dirent_val_u64s(d_name.len));
 
-		if (len > NAME_MAX)
-			return "dirent name too big";
+	/*
+	 * Check new keys don't exceed the max length
+	 * (older keys may be larger.)
+	 */
+	bkey_fsck_err_on((flags & BKEY_INVALID_COMMIT) && d_name.len > BCH_NAME_MAX, c, err,
+			 dirent_name_too_long,
+			 "dirent name too big (%u > %u)",
+			 d_name.len, BCH_NAME_MAX);
+
+	bkey_fsck_err_on(d_name.len != strnlen(d_name.name, d_name.len), c, err,
+			 dirent_name_embedded_nul,
+			 "dirent has stray data after name's NUL");
+
+	bkey_fsck_err_on((d_name.len == 1 && !memcmp(d_name.name, ".", 1)) ||
+			 (d_name.len == 2 && !memcmp(d_name.name, "..", 2)), c, err,
+			 dirent_name_dot_or_dotdot,
+			 "invalid name");
+
+	bkey_fsck_err_on(memchr(d_name.name, '/', d_name.len), c, err,
+			 dirent_name_has_slash,
+			 "name with /");
+
+	bkey_fsck_err_on(d.v->d_type != DT_SUBVOL &&
+			 le64_to_cpu(d.v->d_inum) == d.k->p.inode, c, err,
+			 dirent_to_itself,
+			 "dirent points to own directory");
+fsck_err:
+	return ret;
+}
 
-		if (memchr(d.v->d_name, '/', len))
-			return "dirent name has invalid characters";
+void bch2_dirent_to_text(struct printbuf *out, struct bch_fs *c, struct bkey_s_c k)
+{
+	struct bkey_s_c_dirent d = bkey_s_c_to_dirent(k);
+	struct qstr d_name = bch2_dirent_get_name(d);
 
-		return NULL;
-	case BCH_DIRENT_WHITEOUT:
-		return bkey_val_bytes(k.k) != 0
-			? "value size should be zero"
-			: NULL;
+	prt_printf(out, "%.*s -> ", d_name.len, d_name.name);
 
-	default:
-		return "invalid type";
-	}
-}
+	if (d.v->d_type != DT_SUBVOL)
+		prt_printf(out, "%llu", le64_to_cpu(d.v->d_inum));
+	else
+		prt_printf(out, "%u -> %u",
+			   le32_to_cpu(d.v->d_parent_subvol),
+			   le32_to_cpu(d.v->d_child_subvol));
 
-static void bch2_dirent_to_text(struct bch_fs *c, char *buf,
-				size_t size, struct bkey_s_c k)
-{
-	struct bkey_s_c_dirent d;
-	size_t n = 0;
-
-	switch (k.k->type) {
-	case BCH_DIRENT:
-		d = bkey_s_c_to_dirent(k);
-
-		n += bch_scnmemcpy(buf + n, size - n, d.v->d_name,
-				   bch2_dirent_name_bytes(d));
-		n += scnprintf(buf + n, size - n, " -> %llu", d.v->d_inum);
-		break;
-	case BCH_DIRENT_WHITEOUT:
-		scnprintf(buf, size, "whiteout");
-		break;
-	}
+	prt_printf(out, " type %s", bch2_d_type_str(d.v->d_type));
 }
 
-const struct bkey_ops bch2_bkey_dirent_ops = {
-	.key_invalid	= bch2_dirent_invalid,
-	.val_to_text	= bch2_dirent_to_text,
-};
-
-static struct bkey_i_dirent *dirent_create_key(u8 type,
+static struct bkey_i_dirent *dirent_create_key(struct btree_trans *trans,
+				subvol_inum dir, u8 type,
 				const struct qstr *name, u64 dst)
 {
 	struct bkey_i_dirent *dirent;
 	unsigned u64s = BKEY_U64s + dirent_val_u64s(name->len);
 
-	dirent = kmalloc(u64s * sizeof(u64), GFP_NOFS);
-	if (!dirent)
-		return NULL;
+	if (name->len > BCH_NAME_MAX)
+		return ERR_PTR(-ENAMETOOLONG);
+
+	BUG_ON(u64s > U8_MAX);
+
+	dirent = bch2_trans_kmalloc(trans, u64s * sizeof(u64));
+	if (IS_ERR(dirent))
+		return dirent;
 
 	bkey_dirent_init(&dirent->k_i);
 	dirent->k.u64s = u64s;
-	dirent->v.d_inum = cpu_to_le64(dst);
+
+	if (type != DT_SUBVOL) {
+		dirent->v.d_inum = cpu_to_le64(dst);
+	} else {
+		dirent->v.d_parent_subvol = cpu_to_le32(dir.subvol);
+		dirent->v.d_child_subvol = cpu_to_le32(dst);
+	}
+
 	dirent->v.d_type = type;
 
 	memcpy(dirent->v.d_name, name->name, name->len);
 	memset(dirent->v.d_name + name->len, 0,
 	       bkey_val_bytes(&dirent->k) -
-	       (sizeof(struct bch_dirent) + name->len));
+	       offsetof(struct bch_dirent, d_name) -
+	       name->len);
 
 	EBUG_ON(bch2_dirent_name_bytes(dirent_i_to_s_c(dirent)) != name->len);
 
 	return dirent;
 }
 
-int bch2_dirent_create(struct bch_fs *c, u64 dir_inum,
+int bch2_dirent_create_snapshot(struct btree_trans *trans,
+			u32 dir_subvol, u64 dir, u32 snapshot,
+			const struct bch_hash_info *hash_info,
+			u8 type, const struct qstr *name, u64 dst_inum,
+			u64 *dir_offset,
+			bch_str_hash_flags_t str_hash_flags)
+{
+	subvol_inum dir_inum = { .subvol = dir_subvol, .inum = dir };
+	struct bkey_i_dirent *dirent;
+	int ret;
+
+	dirent = dirent_create_key(trans, dir_inum, type, name, dst_inum);
+	ret = PTR_ERR_OR_ZERO(dirent);
+	if (ret)
+		return ret;
+
+	dirent->k.p.inode	= dir;
+	dirent->k.p.snapshot	= snapshot;
+
+	ret = bch2_hash_set_in_snapshot(trans, bch2_dirent_hash_desc, hash_info,
+					dir_inum, snapshot,
+					&dirent->k_i, str_hash_flags,
+					BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE);
+	*dir_offset = dirent->k.p.offset;
+
+	return ret;
+}
+
+int bch2_dirent_create(struct btree_trans *trans, subvol_inum dir,
 		       const struct bch_hash_info *hash_info,
 		       u8 type, const struct qstr *name, u64 dst_inum,
-		       u64 *journal_seq, int flags)
+		       u64 *dir_offset,
+		       bch_str_hash_flags_t str_hash_flags)
 {
 	struct bkey_i_dirent *dirent;
 	int ret;
 
-	dirent = dirent_create_key(type, name, dst_inum);
-	if (!dirent)
-		return -ENOMEM;
+	dirent = dirent_create_key(trans, dir, type, name, dst_inum);
+	ret = PTR_ERR_OR_ZERO(dirent);
+	if (ret)
+		return ret;
 
-	ret = bch2_hash_set(bch2_dirent_hash_desc, hash_info, c, dir_inum,
-			   journal_seq, &dirent->k_i, flags);
-	kfree(dirent);
+	ret = bch2_hash_set(trans, bch2_dirent_hash_desc, hash_info,
+			    dir, &dirent->k_i, str_hash_flags);
+	*dir_offset = dirent->k.p.offset;
 
 	return ret;
 }
@@ -192,253 +256,355 @@ static void dirent_copy_target(struct bkey_i_dirent *dst,
 	dst->v.d_type = src.v->d_type;
 }
 
-static struct bpos bch2_dirent_pos(struct bch_inode_info *ei,
-				   const struct qstr *name)
+int bch2_dirent_read_target(struct btree_trans *trans, subvol_inum dir,
+			    struct bkey_s_c_dirent d, subvol_inum *target)
 {
-	return POS(ei->vfs_inode.i_ino, bch2_dirent_hash(&ei->str_hash, name));
+	struct bch_subvolume s;
+	int ret = 0;
+
+	if (d.v->d_type == DT_SUBVOL &&
+	    le32_to_cpu(d.v->d_parent_subvol) != dir.subvol)
+		return 1;
+
+	if (likely(d.v->d_type != DT_SUBVOL)) {
+		target->subvol	= dir.subvol;
+		target->inum	= le64_to_cpu(d.v->d_inum);
+	} else {
+		target->subvol	= le32_to_cpu(d.v->d_child_subvol);
+
+		ret = bch2_subvolume_get(trans, target->subvol, true, BTREE_ITER_CACHED, &s);
+
+		target->inum	= le64_to_cpu(s.inode);
+	}
+
+	return ret;
 }
 
-int bch2_dirent_rename(struct bch_fs *c,
-		       struct inode *src_dir, const struct qstr *src_name,
-		       struct inode *dst_dir, const struct qstr *dst_name,
-		       u64 *journal_seq, enum bch_rename_mode mode)
+int bch2_dirent_rename(struct btree_trans *trans,
+		subvol_inum src_dir, struct bch_hash_info *src_hash,
+		subvol_inum dst_dir, struct bch_hash_info *dst_hash,
+		const struct qstr *src_name, subvol_inum *src_inum, u64 *src_offset,
+		const struct qstr *dst_name, subvol_inum *dst_inum, u64 *dst_offset,
+		enum bch_rename_mode mode)
 {
-	struct bch_inode_info *src_ei = to_bch_ei(src_dir);
-	struct bch_inode_info *dst_ei = to_bch_ei(dst_dir);
-	struct btree_iter src_iter, dst_iter, whiteout_iter;
-	struct bkey_s_c old_src, old_dst;
-	struct bkey delete;
+	struct btree_iter src_iter = { NULL };
+	struct btree_iter dst_iter = { NULL };
+	struct bkey_s_c old_src, old_dst = bkey_s_c_null;
 	struct bkey_i_dirent *new_src = NULL, *new_dst = NULL;
-	struct bpos src_pos = bch2_dirent_pos(src_ei, src_name);
-	struct bpos dst_pos = bch2_dirent_pos(dst_ei, dst_name);
-	bool need_whiteout;
-	int ret = -ENOMEM;
+	struct bpos dst_pos =
+		POS(dst_dir.inum, bch2_dirent_hash(dst_hash, dst_name));
+	unsigned src_update_flags = 0;
+	bool delete_src, delete_dst;
+	int ret = 0;
 
-	bch2_btree_iter_init(&src_iter, c, BTREE_ID_DIRENTS, src_pos,
-			     BTREE_ITER_INTENT);
-	bch2_btree_iter_init(&dst_iter, c, BTREE_ID_DIRENTS, dst_pos,
-			     BTREE_ITER_INTENT);
-	bch2_btree_iter_link(&src_iter, &dst_iter);
+	memset(src_inum, 0, sizeof(*src_inum));
+	memset(dst_inum, 0, sizeof(*dst_inum));
 
-	bch2_btree_iter_init(&whiteout_iter, c, BTREE_ID_DIRENTS, src_pos, 0);
-	bch2_btree_iter_link(&src_iter, &whiteout_iter);
+	/* Lookup src: */
+	ret = bch2_hash_lookup(trans, &src_iter, bch2_dirent_hash_desc,
+			       src_hash, src_dir, src_name,
+			       BTREE_ITER_INTENT);
+	if (ret)
+		goto out;
 
-	if (mode == BCH_RENAME_EXCHANGE) {
-		new_src = dirent_create_key(0, src_name, 0);
-		if (!new_src)
-			goto err;
+	old_src = bch2_btree_iter_peek_slot(&src_iter);
+	ret = bkey_err(old_src);
+	if (ret)
+		goto out;
+
+	ret = bch2_dirent_read_target(trans, src_dir,
+			bkey_s_c_to_dirent(old_src), src_inum);
+	if (ret)
+		goto out;
+
+	/* Lookup dst: */
+	if (mode == BCH_RENAME) {
+		/*
+		 * Note that we're _not_ checking if the target already exists -
+		 * we're relying on the VFS to do that check for us for
+		 * correctness:
+		 */
+		ret = bch2_hash_hole(trans, &dst_iter, bch2_dirent_hash_desc,
+				     dst_hash, dst_dir, dst_name);
+		if (ret)
+			goto out;
 	} else {
-		new_src = (void *) &delete;
+		ret = bch2_hash_lookup(trans, &dst_iter, bch2_dirent_hash_desc,
+				       dst_hash, dst_dir, dst_name,
+				       BTREE_ITER_INTENT);
+		if (ret)
+			goto out;
+
+		old_dst = bch2_btree_iter_peek_slot(&dst_iter);
+		ret = bkey_err(old_dst);
+		if (ret)
+			goto out;
+
+		ret = bch2_dirent_read_target(trans, dst_dir,
+				bkey_s_c_to_dirent(old_dst), dst_inum);
+		if (ret)
+			goto out;
 	}
 
-	new_dst = dirent_create_key(0, dst_name, 0);
-	if (!new_dst)
-		goto err;
-retry:
-	/*
-	 * Note that on -EINTR/dropped locks we're not restarting the lookup
-	 * from the original hashed position (like we do when creating dirents,
-	 * in bch_hash_set) -  we never move existing dirents to different slot:
-	 */
-	old_src = bch2_hash_lookup_at(bch2_dirent_hash_desc,
-				     &src_ei->str_hash,
-				     &src_iter, src_name);
-	if ((ret = btree_iter_err(old_src)))
-		goto err;
+	if (mode != BCH_RENAME_EXCHANGE)
+		*src_offset = dst_iter.pos.offset;
 
-	ret = bch2_hash_needs_whiteout(bch2_dirent_hash_desc,
-				&src_ei->str_hash,
-				&whiteout_iter, &src_iter);
-	if (ret < 0)
-		goto err;
-	need_whiteout = ret;
+	/* Create new dst key: */
+	new_dst = dirent_create_key(trans, dst_dir, 0, dst_name, 0);
+	ret = PTR_ERR_OR_ZERO(new_dst);
+	if (ret)
+		goto out;
 
-	/*
-	 * Note that in BCH_RENAME mode, we're _not_ checking if
-	 * the target already exists - we're relying on the VFS
-	 * to do that check for us for correctness:
-	 */
-	old_dst = mode == BCH_RENAME
-		? bch2_hash_hole_at(bch2_dirent_hash_desc, &dst_iter)
-		: bch2_hash_lookup_at(bch2_dirent_hash_desc,
-				     &dst_ei->str_hash,
-				     &dst_iter, dst_name);
-	if ((ret = btree_iter_err(old_dst)))
-		goto err;
+	dirent_copy_target(new_dst, bkey_s_c_to_dirent(old_src));
+	new_dst->k.p = dst_iter.pos;
+
+	/* Create new src key: */
+	if (mode == BCH_RENAME_EXCHANGE) {
+		new_src = dirent_create_key(trans, src_dir, 0, src_name, 0);
+		ret = PTR_ERR_OR_ZERO(new_src);
+		if (ret)
+			goto out;
+
+		dirent_copy_target(new_src, bkey_s_c_to_dirent(old_dst));
+		new_src->k.p = src_iter.pos;
+	} else {
+		new_src = bch2_trans_kmalloc(trans, sizeof(struct bkey_i));
+		ret = PTR_ERR_OR_ZERO(new_src);
+		if (ret)
+			goto out;
 
-	switch (mode) {
-	case BCH_RENAME:
 		bkey_init(&new_src->k);
-		dirent_copy_target(new_dst, bkey_s_c_to_dirent(old_src));
+		new_src->k.p = src_iter.pos;
 
-		if (bkey_cmp(dst_pos, src_iter.pos) <= 0 &&
-		    bkey_cmp(src_iter.pos, dst_iter.pos) < 0) {
+		if (bkey_le(dst_pos, src_iter.pos) &&
+		    bkey_lt(src_iter.pos, dst_iter.pos)) {
 			/*
-			 * If we couldn't insert new_dst at its hashed
-			 * position (dst_pos) due to a hash collision,
-			 * and we're going to be deleting in
-			 * between the hashed position and first empty
-			 * slot we found - just overwrite the pos we
-			 * were going to delete:
-			 *
-			 * Note: this is a correctness issue, in this
-			 * situation bch2_hash_needs_whiteout() could
-			 * return false when the whiteout would have
-			 * been needed if we inserted at the pos
-			 * __dirent_find_hole() found
+			 * We have a hash collision for the new dst key,
+			 * and new_src - the key we're deleting - is between
+			 * new_dst's hashed slot and the slot we're going to be
+			 * inserting it into - oops.  This will break the hash
+			 * table if we don't deal with it:
 			 */
-			new_dst->k.p = src_iter.pos;
-			ret = bch2_btree_insert_at(c, NULL, NULL,
-					journal_seq,
-					BTREE_INSERT_ATOMIC,
-					BTREE_INSERT_ENTRY(&src_iter,
-							   &new_dst->k_i));
-			goto err;
+			if (mode == BCH_RENAME) {
+				/*
+				 * If we're not overwriting, we can just insert
+				 * new_dst at the src position:
+				 */
+				new_src = new_dst;
+				new_src->k.p = src_iter.pos;
+				goto out_set_src;
+			} else {
+				/* If we're overwriting, we can't insert new_dst
+				 * at a different slot because it has to
+				 * overwrite old_dst - just make sure to use a
+				 * whiteout when deleting src:
+				 */
+				new_src->k.type = KEY_TYPE_hash_whiteout;
+			}
+		} else {
+			/* Check if we need a whiteout to delete src: */
+			ret = bch2_hash_needs_whiteout(trans, bch2_dirent_hash_desc,
+						       src_hash, &src_iter);
+			if (ret < 0)
+				goto out;
+
+			if (ret)
+				new_src->k.type = KEY_TYPE_hash_whiteout;
 		}
+	}
 
-		if (need_whiteout)
-			new_src->k.type = BCH_DIRENT_WHITEOUT;
-		break;
-	case BCH_RENAME_OVERWRITE:
-		bkey_init(&new_src->k);
-		dirent_copy_target(new_dst, bkey_s_c_to_dirent(old_src));
+	if (new_dst->v.d_type == DT_SUBVOL)
+		new_dst->v.d_parent_subvol = cpu_to_le32(dst_dir.subvol);
 
-		if (bkey_cmp(dst_pos, src_iter.pos) <= 0 &&
-		    bkey_cmp(src_iter.pos, dst_iter.pos) < 0) {
-			/*
-			 * Same case described above -
-			 * bch_hash_needs_whiteout could spuriously
-			 * return false, but we have to insert at
-			 * dst_iter.pos because we're overwriting
-			 * another dirent:
-			 */
-			new_src->k.type = BCH_DIRENT_WHITEOUT;
-		} else if (need_whiteout)
-			new_src->k.type = BCH_DIRENT_WHITEOUT;
-		break;
-	case BCH_RENAME_EXCHANGE:
-		dirent_copy_target(new_src, bkey_s_c_to_dirent(old_dst));
-		dirent_copy_target(new_dst, bkey_s_c_to_dirent(old_src));
-		break;
+	if ((mode == BCH_RENAME_EXCHANGE) &&
+	    new_src->v.d_type == DT_SUBVOL)
+		new_src->v.d_parent_subvol = cpu_to_le32(src_dir.subvol);
+
+	ret = bch2_trans_update(trans, &dst_iter, &new_dst->k_i, 0);
+	if (ret)
+		goto out;
+out_set_src:
+	/*
+	 * If we're deleting a subvolume we need to really delete the dirent,
+	 * not just emit a whiteout in the current snapshot - there can only be
+	 * single dirent that points to a given subvolume.
+	 *
+	 * IOW, we don't maintain multiple versions in different snapshots of
+	 * dirents that point to subvolumes - dirents that point to subvolumes
+	 * are only visible in one particular subvolume so it's not necessary,
+	 * and it would be particularly confusing for fsck to have to deal with.
+	 */
+	delete_src = bkey_s_c_to_dirent(old_src).v->d_type == DT_SUBVOL &&
+		new_src->k.p.snapshot != old_src.k->p.snapshot;
+
+	delete_dst = old_dst.k &&
+		bkey_s_c_to_dirent(old_dst).v->d_type == DT_SUBVOL &&
+		new_dst->k.p.snapshot != old_dst.k->p.snapshot;
+
+	if (!delete_src || !bkey_deleted(&new_src->k)) {
+		ret = bch2_trans_update(trans, &src_iter, &new_src->k_i, src_update_flags);
+		if (ret)
+			goto out;
 	}
 
-	new_src->k.p = src_iter.pos;
-	new_dst->k.p = dst_iter.pos;
-	ret = bch2_btree_insert_at(c, NULL, NULL, journal_seq,
-			BTREE_INSERT_ATOMIC,
-			BTREE_INSERT_ENTRY(&src_iter, &new_src->k_i),
-			BTREE_INSERT_ENTRY(&dst_iter, &new_dst->k_i));
-err:
-	if (ret == -EINTR)
-		goto retry;
+	if (delete_src) {
+		bch2_btree_iter_set_snapshot(&src_iter, old_src.k->p.snapshot);
+		ret =   bch2_btree_iter_traverse(&src_iter) ?:
+			bch2_btree_delete_at(trans, &src_iter, BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE);
+		if (ret)
+			goto out;
+	}
 
-	bch2_btree_iter_unlock(&whiteout_iter);
-	bch2_btree_iter_unlock(&dst_iter);
-	bch2_btree_iter_unlock(&src_iter);
+	if (delete_dst) {
+		bch2_btree_iter_set_snapshot(&dst_iter, old_dst.k->p.snapshot);
+		ret =   bch2_btree_iter_traverse(&dst_iter) ?:
+			bch2_btree_delete_at(trans, &dst_iter, BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE);
+		if (ret)
+			goto out;
+	}
 
-	if (new_src != (void *) &delete)
-		kfree(new_src);
-	kfree(new_dst);
+	if (mode == BCH_RENAME_EXCHANGE)
+		*src_offset = new_src->k.p.offset;
+	*dst_offset = new_dst->k.p.offset;
+out:
+	bch2_trans_iter_exit(trans, &src_iter);
+	bch2_trans_iter_exit(trans, &dst_iter);
 	return ret;
 }
 
-int bch2_dirent_delete(struct bch_fs *c, u64 dir_inum,
-		       const struct bch_hash_info *hash_info,
-		       const struct qstr *name,
-		       u64 *journal_seq)
+int bch2_dirent_lookup_trans(struct btree_trans *trans,
+			     struct btree_iter *iter,
+			     subvol_inum dir,
+			     const struct bch_hash_info *hash_info,
+			     const struct qstr *name, subvol_inum *inum,
+			     unsigned flags)
 {
-	return bch2_hash_delete(bch2_dirent_hash_desc, hash_info,
-			       c, dir_inum, journal_seq, name);
+	int ret = bch2_hash_lookup(trans, iter, bch2_dirent_hash_desc,
+				   hash_info, dir, name, flags);
+	if (ret)
+		return ret;
+
+	struct bkey_s_c k = bch2_btree_iter_peek_slot(iter);
+	ret = bkey_err(k);
+	if (ret)
+		goto err;
+
+	ret = bch2_dirent_read_target(trans, dir, bkey_s_c_to_dirent(k), inum);
+	if (ret > 0)
+		ret = -ENOENT;
+err:
+	if (ret)
+		bch2_trans_iter_exit(trans, iter);
+	return ret;
 }
 
-u64 bch2_dirent_lookup(struct bch_fs *c, u64 dir_inum,
+u64 bch2_dirent_lookup(struct bch_fs *c, subvol_inum dir,
 		       const struct bch_hash_info *hash_info,
-		       const struct qstr *name)
+		       const struct qstr *name, subvol_inum *inum)
 {
-	struct btree_iter iter;
-	struct bkey_s_c k;
-	u64 inum;
+	struct btree_trans *trans = bch2_trans_get(c);
+	struct btree_iter iter = { NULL };
 
-	k = bch2_hash_lookup(bch2_dirent_hash_desc, hash_info, c,
-			    dir_inum, &iter, name);
-	if (IS_ERR(k.k)) {
-		bch2_btree_iter_unlock(&iter);
-		return 0;
-	}
-
-	inum = le64_to_cpu(bkey_s_c_to_dirent(k).v->d_inum);
-	bch2_btree_iter_unlock(&iter);
-
-	return inum;
+	int ret = lockrestart_do(trans,
+		bch2_dirent_lookup_trans(trans, &iter, dir, hash_info, name, inum, 0));
+	bch2_trans_iter_exit(trans, &iter);
+	bch2_trans_put(trans);
+	return ret;
 }
 
-int bch2_empty_dir(struct bch_fs *c, u64 dir_inum)
+int bch2_empty_dir_snapshot(struct btree_trans *trans, u64 dir, u32 subvol, u32 snapshot)
 {
 	struct btree_iter iter;
 	struct bkey_s_c k;
-	int ret = 0;
-
-	for_each_btree_key(&iter, c, BTREE_ID_DIRENTS, POS(dir_inum, 0), 0, k) {
-		if (k.k->p.inode > dir_inum)
-			break;
+	int ret;
 
-		if (k.k->type == BCH_DIRENT) {
-			ret = -ENOTEMPTY;
+	for_each_btree_key_upto_norestart(trans, iter, BTREE_ID_dirents,
+			   SPOS(dir, 0, snapshot),
+			   POS(dir, U64_MAX), 0, k, ret)
+		if (k.k->type == KEY_TYPE_dirent) {
+			struct bkey_s_c_dirent d = bkey_s_c_to_dirent(k);
+			if (d.v->d_type == DT_SUBVOL && le32_to_cpu(d.v->d_parent_subvol) != subvol)
+				continue;
+			ret = -BCH_ERR_ENOTEMPTY_dir_not_empty;
 			break;
 		}
-	}
-	bch2_btree_iter_unlock(&iter);
+	bch2_trans_iter_exit(trans, &iter);
 
 	return ret;
 }
 
-int bch2_readdir(struct bch_fs *c, struct file *file,
-		 struct dir_context *ctx)
+int bch2_empty_dir_trans(struct btree_trans *trans, subvol_inum dir)
 {
-	struct inode *inode = file_inode(file);
+	u32 snapshot;
+
+	return bch2_subvolume_get_snapshot(trans, dir.subvol, &snapshot) ?:
+		bch2_empty_dir_snapshot(trans, dir.inum, dir.subvol, snapshot);
+}
+
+int bch2_readdir(struct bch_fs *c, subvol_inum inum, struct dir_context *ctx)
+{
+	struct btree_trans *trans = bch2_trans_get(c);
 	struct btree_iter iter;
 	struct bkey_s_c k;
 	struct bkey_s_c_dirent dirent;
-	unsigned len;
+	subvol_inum target;
+	u32 snapshot;
+	struct bkey_buf sk;
+	struct qstr name;
+	int ret;
 
-	if (!dir_emit_dots(file, ctx))
-		return 0;
+	bch2_bkey_buf_init(&sk);
+retry:
+	bch2_trans_begin(trans);
 
-	pr_debug("listing for %lu from %llu", inode->i_ino, ctx->pos);
+	ret = bch2_subvolume_get_snapshot(trans, inum.subvol, &snapshot);
+	if (ret)
+		goto err;
 
-	for_each_btree_key(&iter, c, BTREE_ID_DIRENTS,
-			   POS(inode->i_ino, ctx->pos), 0, k) {
-		if (k.k->type != BCH_DIRENT)
+	for_each_btree_key_upto_norestart(trans, iter, BTREE_ID_dirents,
+			   SPOS(inum.inum, ctx->pos, snapshot),
+			   POS(inum.inum, U64_MAX), 0, k, ret) {
+		if (k.k->type != KEY_TYPE_dirent)
 			continue;
 
 		dirent = bkey_s_c_to_dirent(k);
 
-		pr_debug("saw %llu:%llu (%s) -> %llu",
-			 k.k->p.inode, k.k->p.offset,
-			 dirent.v->d_name, dirent.v->d_inum);
-
-		if (bkey_cmp(k.k->p, POS(inode->i_ino, ctx->pos)) < 0)
+		ret = bch2_dirent_read_target(trans, inum, dirent, &target);
+		if (ret < 0)
+			break;
+		if (ret)
 			continue;
 
-		if (k.k->p.inode > inode->i_ino)
-			break;
+		/* dir_emit() can fault and block: */
+		bch2_bkey_buf_reassemble(&sk, c, k);
+		dirent = bkey_i_to_s_c_dirent(sk.k);
+		bch2_trans_unlock(trans);
 
-		len = bch2_dirent_name_bytes(dirent);
+		name = bch2_dirent_get_name(dirent);
 
-		pr_debug("emitting %s", dirent.v->d_name);
+		ctx->pos = dirent.k->p.offset;
+		if (!dir_emit(ctx, name.name,
+			      name.len,
+			      target.inum,
+			      vfs_d_type(dirent.v->d_type)))
+			break;
+		ctx->pos = dirent.k->p.offset + 1;
 
 		/*
-		 * XXX: dir_emit() can fault and block, while we're holding
-		 * locks
+		 * read_target looks up subvolumes, we can overflow paths if the
+		 * directory has many subvolumes in it
 		 */
-		if (!dir_emit(ctx, dirent.v->d_name, len,
-			      le64_to_cpu(dirent.v->d_inum),
-			      dirent.v->d_type))
+		ret = btree_trans_too_many_iters(trans);
+		if (ret)
 			break;
-
-		ctx->pos = k.k->p.offset + 1;
 	}
-	bch2_btree_iter_unlock(&iter);
+	bch2_trans_iter_exit(trans, &iter);
+err:
+	if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
+		goto retry;
+
+	bch2_trans_put(trans);
+	bch2_bkey_buf_exit(&sk, c);
 
-	return 0;
+	return ret;
 }