Re: [PATCH] ocfs2: validate inline dir size in ocfs2_dir_foreach_blk_id

[ZhengYuan Huang]

On Fri, Apr 10, 2026 at 6:03 PM Joseph Qi <joseph.qi@xxxxxxxxxxxxxxxxx> wrote:
> On 4/10/26 2:22 PM, ZhengYuan Huang wrote:
> > [BUG]
> > A crafted inline-data directory can set i_size larger than id_count.
> > readdir then walks past data->id_data and KASAN reports:
> >
> > BUG: KASAN: use-after-free in ocfs2_check_dir_entry.isra.0+0x31f/0x370 fs/ocfs2/dir.c:305
> > Read of size 2 at addr ffff8880088f0008 by task syz.0.1936/4656
> > Call Trace:
> >  ...
> >  ocfs2_check_dir_entry.isra.0+0x31f/0x370 fs/ocfs2/dir.c:305
> >  ocfs2_dir_foreach_blk_id+0x203/0xa70 fs/ocfs2/dir.c:1805
> >  ocfs2_dir_foreach_blk fs/ocfs2/dir.c:1933 [inline]
> >  ocfs2_readdir+0x4ba/0x520 fs/ocfs2/dir.c:1977
> >  wrap_directory_iterator+0x9c/0xe0 fs/readdir.c:65
> >  shared_ocfs2_readdir+0x29/0x40 fs/ocfs2/file.c:2822
> >  iterate_dir+0x276/0x9e0 fs/readdir.c:108
> >  __do_sys_getdents64 fs/readdir.c:410 [inline]
> >  __se_sys_getdents64 fs/readdir.c:396 [inline]
> >  __x64_sys_getdents64+0x143/0x2a0 fs/readdir.c:396
> >  ...
> >
> > [CAUSE]
> > ocfs2_dir_foreach_blk_id() uses i_size_read(inode) as the loop bound
> > after reading the inode block. Inline directories are only valid while
> > i_size <= le16_to_cpu(data->id_count), but that invariant is never
> > checked on read. Once ctx->pos reaches id_count, data->id_data +
> > ctx->pos points past the inode block and can land in a freed neighbor
> > page.
> >
>
> ocfs2_read_inode_block() has already validated dinode->i_size.
> So how it happens for the corrupted in-memory i_size?
>
> Thanks,
> Joseph

Thanks, my previous description was inaccurate.

The issue is not that inode->i_size becomes corrupted in memory after
ocfs2_read_inode_block() validates the dinode. The issue is that the
current dinode validation does not check the inline-directory
invariant i_size <= id_count.

ocfs2_validate_inode_block() does basic dinode validation, but it does
not verify that inline directory size fits within id_count. Then
ocfs2_populate_inode() / ocfs2_refresh_inode() /
ocfs2_refresh_inode_from_lvb() copy that unchecked size into the VFS
inode.

So the actual problem is an unchecked corrupted on-disk i_size, not a
post-validation in-memory corruption.

Once such an inode is loaded, readdir can legally reach
ocfs2_dir_foreach_blk_id(), which uses i_size_read(inode) as the upper
bound.

Based on your comment, I think the cleaner fix is to reject such
inline directories when the inode is loaded or refreshed, rather than
only checking in ocfs2_dir_foreach_blk_id(). I will respin the patch
accordingly.

Thanks,
ZhengYuan Huang