When mounting an ext4 filesystem, the on-disk superblock's s_mount_opts
field (which stores default mount options set via tune2fs) is read and
parsed. Unlike userspace-provided mount options which are validated by
the VFS layer before reaching the filesystem, the on-disk s_mount_opts
is read directly from the disk buffer without NUL-termination validation.

The two option paths use the same parser but arrive differently:

  Userspace mount options:
    VFS -> ext4_parse_param()

  On-disk default options:
    parse_apply_sb_mount_options() -> parse_options() -> ext4_parse_param()

When s_mount_opts lacks NUL-termination, strscpy_pad()'s internal
fortified strnlen() detects reading beyond the 64-byte field, triggering
an Oops:

  strnlen: detected buffer overflow: 65 byte read of buffer size 64
  WARNING: CPU: 0 PID: 179 at lib/string_helpers.c:1035 __fortify_report+0x5a/0x100
  ...
  Call Trace:
   strnlen+0x71/0xa0 lib/string.c:155
   sized_strscpy+0x48/0x2a0 lib/string.c:298
   parse_apply_sb_mount_options+0x94/0x4a0 fs/ext4/super.c:2486
   __ext4_fill_super+0x31d6/0x51b0 fs/ext4/super.c:5306
   ext4_fill_super+0x3972/0xaf70 fs/ext4/super.c:5736
   get_tree_bdev_flags+0x38c/0x620 fs/super.c:1698
   vfs_get_tree+0x8e/0x340 fs/super.c:1758
   fc_mount fs/namespace.c:1199
   do_new_mount fs/namespace.c:3718
   path_mount+0x7b9/0x23a0 fs/namespace.c:4028
   ...

Reject the mount with an error instead. While there is an existing similar
check in the ioctl path (which validates userspace input before writing
TO disk), this check validates data read FROM disk before parsing.

The painful history here is:

8b67f04ab9de ("ext4: Add mount options in superblock")
  s_mount_opts is created and treated as __nonstring: kstrndup is used to
  make sure all 64 potential characters are available for use (i.e. up
  to 65 bytes may be allocated).

04a91570ac67 ("ext4: implemet new ioctls to set and get superblock parameters")
  Created ext4_tune_sb_params::mount_opts as 64 bytes in size but
  incorrectly treated it and s_mount_opts as a C strings (it used
  strscpy_pad() to copy between them).

8ecb790ea8c3 ("ext4: avoid potential buffer over-read in parse_apply_sb_mount_options()")
  As a prerequisite to the ioctl treating s_mount_opts as a C string, this
  attempted to switch to using strscpy_pad() with a 65 byte destination
  for the case of an unterminated s_mount_opts. (But strscpy_pad() will
  fail due to the over-read of s_mount_opts by strnlen().)

3db63d2c2d1d ("ext4: check if mount_opts is NUL-terminated in ext4_ioctl_set_tune_sb()")
  As a continuation of trying to solve the 64/65 mismatch, this started
  enforcing a 63 character limit (i.e. 64 bytes total) to incoming values
  from userspace to the ioctl API. (But did not check s_mount_opts coming
  from disk.)

ee5a977b4e77 ("ext4: fix string copying in parse_apply_sb_mount_options()")
  Notices the loud failures of strscpy_pad() introduced by 8ecb790ea8c3,
  and attempted to silence them by making the destination 64 and rejecting
  too-long strings from the on-disk copy of s_mount_opts, but didn't
  actually solve it at all, since the problem was always the over-read
  of the source seen by strnlen(). (Note that the report quoted in this
  commit exactly matches the report today.)

The other option is to go back in time and mark both s_mount_opts and
mount_opts as __nonstring and switch to using memcpy_and_pad() to copy
them around between userspace and kernel and disk instead of making them
C strings.

What do the ext4 regression tests expect for s_mount_opts? Is there a
test for a non-terminated s_mount_opts in an image?

Reported-by: 李龙兴 <coregee2000@gmail.com>
Closes: https://lore.kernel.org/lkml/CAHPqNmzBb2LruMA6jymoHXQRsoiAKMFZ1wVEz8JcYKg4U6TBbw@mail.gmail.com/
Fixes: ee5a977b4e77 ("ext4: fix string copying in parse_apply_sb_mount_options()")
Signed-off-by: Kees Cook <kees@kernel.org>
---
Cc: "Theodore Ts'o" <tytso@mit.edu>
Cc: Andreas Dilger <adilger.kernel@dilger.ca>
Cc: Andy Shevchenko <andriy.shevchenko@intel.com>
Cc: <linux-ext4@vger.kernel.org>
---
 fs/ext4/super.c | 7 +++++++
 1 file changed, 7 insertions(+)

diff --git a/fs/ext4/super.c b/fs/ext4/super.c
index 87205660c5d0..9ad6005615d8 100644
--- a/fs/ext4/super.c
+++ b/fs/ext4/super.c
@@ -2485,6 +2485,13 @@ static int parse_apply_sb_mount_options(struct super_block *sb,
 	if (!sbi->s_es->s_mount_opts[0])
 		return 0;
 
+	if (strnlen(sbi->s_es->s_mount_opts, sizeof(sbi->s_es->s_mount_opts)) ==
+	    sizeof(sbi->s_es->s_mount_opts)) {
+		ext4_msg(sb, KERN_ERR,
+			 "Mount options in superblock are not NUL-terminated");
+		return -EINVAL;
+	}
+
 	if (strscpy_pad(s_mount_opts, sbi->s_es->s_mount_opts) < 0)
 		return -E2BIG;
 
-- 
2.34.1

On Fri, Feb 06, 2026 at 01:27:03PM -0800, Kees Cook wrote:
> When mounting an ext4 filesystem, the on-disk superblock's s_mount_opts
> field (which stores default mount options set via tune2fs) is read and
> parsed. Unlike userspace-provided mount options which are validated by
> the VFS layer before reaching the filesystem, the on-disk s_mount_opts
> is read directly from the disk buffer without NUL-termination validation.
> 
> The two option paths use the same parser but arrive differently:
> 
>   Userspace mount options:
>     VFS -> ext4_parse_param()
> 
>   On-disk default options:
>     parse_apply_sb_mount_options() -> parse_options() -> ext4_parse_param()
> 
> When s_mount_opts lacks NUL-termination, strscpy_pad()'s internal
> fortified strnlen() detects reading beyond the 64-byte field, triggering
> an Oops:
> 
>   strnlen: detected buffer overflow: 65 byte read of buffer size 64
>   WARNING: CPU: 0 PID: 179 at lib/string_helpers.c:1035 __fortify_report+0x5a/0x100
>   ...
>   Call Trace:
>    strnlen+0x71/0xa0 lib/string.c:155
>    sized_strscpy+0x48/0x2a0 lib/string.c:298
>    parse_apply_sb_mount_options+0x94/0x4a0 fs/ext4/super.c:2486
>    __ext4_fill_super+0x31d6/0x51b0 fs/ext4/super.c:5306
>    ext4_fill_super+0x3972/0xaf70 fs/ext4/super.c:5736
>    get_tree_bdev_flags+0x38c/0x620 fs/super.c:1698
>    vfs_get_tree+0x8e/0x340 fs/super.c:1758
>    fc_mount fs/namespace.c:1199
>    do_new_mount fs/namespace.c:3718
>    path_mount+0x7b9/0x23a0 fs/namespace.c:4028
>    ...
> 
> Reject the mount with an error instead. While there is an existing similar
> check in the ioctl path (which validates userspace input before writing
> TO disk), this check validates data read FROM disk before parsing.
> 
> The painful history here is:
> 
> 8b67f04ab9de ("ext4: Add mount options in superblock")
>   s_mount_opts is created and treated as __nonstring: kstrndup is used to
>   make sure all 64 potential characters are available for use (i.e. up
>   to 65 bytes may be allocated).
> 
> 04a91570ac67 ("ext4: implemet new ioctls to set and get superblock parameters")
>   Created ext4_tune_sb_params::mount_opts as 64 bytes in size but
>   incorrectly treated it and s_mount_opts as a C strings (it used
>   strscpy_pad() to copy between them).
> 
> 8ecb790ea8c3 ("ext4: avoid potential buffer over-read in parse_apply_sb_mount_options()")
>   As a prerequisite to the ioctl treating s_mount_opts as a C string, this
>   attempted to switch to using strscpy_pad() with a 65 byte destination
>   for the case of an unterminated s_mount_opts. (But strscpy_pad() will
>   fail due to the over-read of s_mount_opts by strnlen().)
> 
> 3db63d2c2d1d ("ext4: check if mount_opts is NUL-terminated in ext4_ioctl_set_tune_sb()")
>   As a continuation of trying to solve the 64/65 mismatch, this started
>   enforcing a 63 character limit (i.e. 64 bytes total) to incoming values
>   from userspace to the ioctl API. (But did not check s_mount_opts coming
>   from disk.)
> 
> ee5a977b4e77 ("ext4: fix string copying in parse_apply_sb_mount_options()")
>   Notices the loud failures of strscpy_pad() introduced by 8ecb790ea8c3,
>   and attempted to silence them by making the destination 64 and rejecting
>   too-long strings from the on-disk copy of s_mount_opts, but didn't
>   actually solve it at all, since the problem was always the over-read
>   of the source seen by strnlen(). (Note that the report quoted in this
>   commit exactly matches the report today.)

Looking through the history of tune2fs, I see v1.42 has commit
9a976ac732a7e9 ("tune2fs: Fix mount_opts handling"), which always set
byte 63 to a null, which implies that s_mount_opts is supposed to be a
64-byte field that contains a null-terminated string.  It also doesn't
appear that slack space after the null terminator is set to any
particular value.

In e2fsprogs v1.41.13, commit 9345f02671fd39 ("tune2fs, debugfs,
libext2fs: Add support for ext4 default mount options") added the
ability to set s_mount_opts, and it rejects any string whose length
is greater than or equal to 64 bytes.  This validation is still in
place.

My guess is that the ext4 code should go back to requiring a null
terminated string, and the __nonstring annotations dropped.

I guess the hard part is, were there any filesystems written out with a
64-byte unterminated s_mount_opts?  It looks like 04a91570ac67 used
strscpy_pad to copy the string to the superblock and that function
always zero-terminates the destination, so perhaps we can stick with
"the ondisk superblock s_mount_opts is always a null terminated string"?

--D

> The other option is to go back in time and mark both s_mount_opts and
> mount_opts as __nonstring and switch to using memcpy_and_pad() to copy
> them around between userspace and kernel and disk instead of making them
> C strings.
> 
> What do the ext4 regression tests expect for s_mount_opts? Is there a
> test for a non-terminated s_mount_opts in an image?
> 
> Reported-by: 李龙兴 <coregee2000@gmail.com>
> Closes: https://lore.kernel.org/lkml/CAHPqNmzBb2LruMA6jymoHXQRsoiAKMFZ1wVEz8JcYKg4U6TBbw@mail.gmail.com/
> Fixes: ee5a977b4e77 ("ext4: fix string copying in parse_apply_sb_mount_options()")
> Signed-off-by: Kees Cook <kees@kernel.org>
> ---
> Cc: "Theodore Ts'o" <tytso@mit.edu>
> Cc: Andreas Dilger <adilger.kernel@dilger.ca>
> Cc: Andy Shevchenko <andriy.shevchenko@intel.com>
> Cc: <linux-ext4@vger.kernel.org>
> ---
>  fs/ext4/super.c | 7 +++++++
>  1 file changed, 7 insertions(+)
> 
> diff --git a/fs/ext4/super.c b/fs/ext4/super.c
> index 87205660c5d0..9ad6005615d8 100644
> --- a/fs/ext4/super.c
> +++ b/fs/ext4/super.c
> @@ -2485,6 +2485,13 @@ static int parse_apply_sb_mount_options(struct super_block *sb,
>  	if (!sbi->s_es->s_mount_opts[0])
>  		return 0;
>  
> +	if (strnlen(sbi->s_es->s_mount_opts, sizeof(sbi->s_es->s_mount_opts)) ==
> +	    sizeof(sbi->s_es->s_mount_opts)) {
> +		ext4_msg(sb, KERN_ERR,
> +			 "Mount options in superblock are not NUL-terminated");
> +		return -EINVAL;
> +	}
> +
>  	if (strscpy_pad(s_mount_opts, sbi->s_es->s_mount_opts) < 0)
>  		return -E2BIG;
>  
> -- 
> 2.34.1
> 
> 

On Fri, Feb 06, 2026 at 01:45:02PM -0800, Darrick J. Wong wrote:
> On Fri, Feb 06, 2026 at 01:27:03PM -0800, Kees Cook wrote:
> > When mounting an ext4 filesystem, the on-disk superblock's s_mount_opts
> > field (which stores default mount options set via tune2fs) is read and
> > parsed. Unlike userspace-provided mount options which are validated by
> > the VFS layer before reaching the filesystem, the on-disk s_mount_opts
> > is read directly from the disk buffer without NUL-termination validation.
> > 
> > The two option paths use the same parser but arrive differently:
> > 
> >   Userspace mount options:
> >     VFS -> ext4_parse_param()
> > 
> >   On-disk default options:
> >     parse_apply_sb_mount_options() -> parse_options() -> ext4_parse_param()
> > 
> > When s_mount_opts lacks NUL-termination, strscpy_pad()'s internal
> > fortified strnlen() detects reading beyond the 64-byte field, triggering
> > an Oops:
> > 
> >   strnlen: detected buffer overflow: 65 byte read of buffer size 64
> >   WARNING: CPU: 0 PID: 179 at lib/string_helpers.c:1035 __fortify_report+0x5a/0x100
> >   ...
> >   Call Trace:
> >    strnlen+0x71/0xa0 lib/string.c:155
> >    sized_strscpy+0x48/0x2a0 lib/string.c:298
> >    parse_apply_sb_mount_options+0x94/0x4a0 fs/ext4/super.c:2486
> >    __ext4_fill_super+0x31d6/0x51b0 fs/ext4/super.c:5306
> >    ext4_fill_super+0x3972/0xaf70 fs/ext4/super.c:5736
> >    get_tree_bdev_flags+0x38c/0x620 fs/super.c:1698
> >    vfs_get_tree+0x8e/0x340 fs/super.c:1758
> >    fc_mount fs/namespace.c:1199
> >    do_new_mount fs/namespace.c:3718
> >    path_mount+0x7b9/0x23a0 fs/namespace.c:4028
> >    ...
> > 
> > Reject the mount with an error instead. While there is an existing similar
> > check in the ioctl path (which validates userspace input before writing
> > TO disk), this check validates data read FROM disk before parsing.
> > 
> > The painful history here is:
> > 
> > 8b67f04ab9de ("ext4: Add mount options in superblock")
> >   s_mount_opts is created and treated as __nonstring: kstrndup is used to
> >   make sure all 64 potential characters are available for use (i.e. up
> >   to 65 bytes may be allocated).
> > 
> > 04a91570ac67 ("ext4: implemet new ioctls to set and get superblock parameters")
> >   Created ext4_tune_sb_params::mount_opts as 64 bytes in size but
> >   incorrectly treated it and s_mount_opts as a C strings (it used
> >   strscpy_pad() to copy between them).
> > 
> > 8ecb790ea8c3 ("ext4: avoid potential buffer over-read in parse_apply_sb_mount_options()")
> >   As a prerequisite to the ioctl treating s_mount_opts as a C string, this
> >   attempted to switch to using strscpy_pad() with a 65 byte destination
> >   for the case of an unterminated s_mount_opts. (But strscpy_pad() will
> >   fail due to the over-read of s_mount_opts by strnlen().)
> > 
> > 3db63d2c2d1d ("ext4: check if mount_opts is NUL-terminated in ext4_ioctl_set_tune_sb()")
> >   As a continuation of trying to solve the 64/65 mismatch, this started
> >   enforcing a 63 character limit (i.e. 64 bytes total) to incoming values
> >   from userspace to the ioctl API. (But did not check s_mount_opts coming
> >   from disk.)
> > 
> > ee5a977b4e77 ("ext4: fix string copying in parse_apply_sb_mount_options()")
> >   Notices the loud failures of strscpy_pad() introduced by 8ecb790ea8c3,
> >   and attempted to silence them by making the destination 64 and rejecting
> >   too-long strings from the on-disk copy of s_mount_opts, but didn't
> >   actually solve it at all, since the problem was always the over-read
> >   of the source seen by strnlen(). (Note that the report quoted in this
> >   commit exactly matches the report today.)
> 
> Looking through the history of tune2fs, I see v1.42 has commit
> 9a976ac732a7e9 ("tune2fs: Fix mount_opts handling"), which always set
> byte 63 to a null, which implies that s_mount_opts is supposed to be a
> 64-byte field that contains a null-terminated string.  It also doesn't
> appear that slack space after the null terminator is set to any
> particular value.

Okay, so the userspace side is treating it as a C string. Then ignore my
"PATCH ALTERNATIVE" I just sent. ;)

> In e2fsprogs v1.41.13, commit 9345f02671fd39 ("tune2fs, debugfs,
> libext2fs: Add support for ext4 default mount options") added the
> ability to set s_mount_opts, and it rejects any string whose length
> is greater than or equal to 64 bytes.  This validation is still in
> place.

Great, thanks for tracking that down. That looks to be Dec 2010.

> My guess is that the ext4 code should go back to requiring a null
> terminated string, and the __nonstring annotations dropped.

It does already; no __nonstring annotations currently exist for
s_mount_opts. The trouble was that it was originally treated as
effectively __nonstring between 8b67f04ab9de (v2.6.36, Oct 2010) and
04a91570ac67 (v6.18, Nov 2025).

> I guess the hard part is, were there any filesystems written out with a
> 64-byte unterminated s_mount_opts?  It looks like 04a91570ac67 used
> strscpy_pad to copy the string to the superblock and that function
> always zero-terminates the destination, so perhaps we can stick with
> "the ondisk superblock s_mount_opts is always a null terminated string"?

Yeah, if the userspace tools always wrote images with a NUL-terminated
s_mount_opts, then it should be perfectly sane to reject mounts that
don't follow that. (i.e. this patch.)

-- 
Kees Cook