[PATCH v2] ext4: clear in-inode xattr space before adding the first xattr

[Jiaming Zhang <r772577952@gmail.com>]

An inode can hold extended attributes in the space that follows its extra
fields, whose size is given by i_extra_isize.  When ext4 loads an inode,
ext4_iget_extra_inode() validates that space only if it starts with the
xattr magic, and only then sets EXT4_STATE_XATTR.  Without the magic, the
space is treated as empty and is not checked.

ext4_xattr_ibody_set() passes the space to ext4_xattr_set_entry() whether
or not EXT4_STATE_XATTR is set.  ext4_xattr_set_entry() walks the entries
there to find min_offs, the lowest value offset.  That offset decides
whether the new value fits and where the value is copied:

	free = min_offs - ((void *)last - s->base) - sizeof(__u32);
	...
	void *val = s->base + min_offs - new_size;

A crafted image can leave non-zero bytes in that space without the magic
where ext4 looks for it, for example by shrinking i_extra_isize.  Those
bytes are parsed as entries, and an entry that claims an in-inode value at
offset 0 sets min_offs to 0.  Because free is a size_t, the subtraction
then wraps around to a huge value, and the free space check passes although
the value does not fit.  The value is copied new_size bytes before the
start of the space, over the inode's own fields and, for a large value, the
memory in front of the inode.

On a kernel without KASAN, the copy corrupts memory: it overwrites the
inode's own fields and, for a large value, also the memory in front of the
inode.  The bytes and their length are decided by the setxattr() caller.
Nothing is freed and reused here, so the use-after-free in the report only
reflects where the write happened to land.

Triggering this issue requires mounting a crafted image, which needs
CAP_SYS_ADMIN or automounting of removable media; after that, any
setxattr() on the file reaches this path.  The issue was found by fuzzing
and has not been seen in production; a reproducer is available in the
linked report.

When EXT4_STATE_XATTR is not set, zero the first 32 bits of the space in
ext4_xattr_ibody_set() before calling ext4_xattr_set_entry().  These are
the bits IS_LAST_ENTRY() reads, so the space is treated as empty and the
walk stops before reading any stale entry.

Fixes: ac27a0ec112a ("[PATCH] ext4: initial copy of files from ext3")
Closes: https://lore.kernel.org/lkml/CANypQFZzxTAH=4QRRDoqH6VAw1daAq88NDWiQKoM7QTOnP40sg@mail.gmail.com/
Cc: stable@vger.kernel.org
Assisted-by: Claude Code:claude-opus-5-5
Signed-off-by: Jiaming Zhang <r772577952@gmail.com>
---
Changes in v2:
- Zero only the first 32 bits space instead of the whole space.
- Add a comment above the check.
- v1: https://lore.kernel.org/lkml/20260916084723.877552-1-r772577952@gmail.com/

 fs/ext4/xattr.c | 8 ++++++++
 1 file changed, 8 insertions(+)

diff --git a/fs/ext4/xattr.c b/fs/ext4/xattr.c
index 5c310747b965..8e1fd26329d3 100644
--- a/fs/ext4/xattr.c
+++ b/fs/ext4/xattr.c
@@ -2274,6 +2274,14 @@ int ext4_xattr_ibody_set(handle_t *handle, struct inode *inode,
 		if (IS_ERR(ea_inode))
 			return PTR_ERR(ea_inode);
 	}
+
+	/*
+	 * Make sure xattr space is properly initialized so that
+	 * IS_LAST_ENTRY() returns true for its first entry
+	 */
+	if (!ext4_test_inode_state(inode, EXT4_STATE_XATTR))
+		*(__u32 *)s->first = 0;
+
 	error = ext4_xattr_set_entry(i, s, handle, inode, ea_inode,
 				     false /* is_block */);
 	if (error) {
-- 
2.43.0