[PATCH] fs: Return raw xattr for security.* if there is size disagreement with LSMs

[Mimi Zohar]

On Thu, 2021-06-17 at 07:09 +0000, Roberto Sassu wrote:
> > From: Stefan Berger [mailto:stefanb at linux.ibm.com]
> > Sent: Wednesday, June 16, 2021 4:40 PM
> > On 6/16/21 9:22 AM, Roberto Sassu wrote:
> > > vfs_getxattr() differs from vfs_setxattr() in the way it obtains the xattr
> > > value. The former gives precedence to the LSMs, and if the LSMs don't
> > > provide a value, obtains it from the filesystem handler. The latter does
> > > the opposite, first invokes the filesystem handler, and if the filesystem
> > > does not support xattrs, passes the xattr value to the LSMs.
> > >
> > > The problem is that not necessarily the user gets the same xattr value that
> > > he set. For example, if he sets security.selinux with a value not
> > > terminated with '\0', he gets a value terminated with '\0' because SELinux
> > > adds it during the translation from xattr to internal representation
> > > (vfs_setxattr()) and from internal representation to xattr
> > > (vfs_getxattr()).
> > >
> > > Normally, this does not have an impact unless the integrity of xattrs is
> > > verified with EVM. The kernel and the user see different values due to the
> > > different functions used to obtain them:
> > >
> > > kernel (EVM): uses vfs_getxattr_alloc() which obtains the xattr value from
> > >                the filesystem handler (raw value);
> > >
> > > user (ima-evm-utils): uses vfs_getxattr() which obtains the xattr value
> > >                        from the LSMs (normalized value).
> > 
> > Maybe there should be another implementation similar to
> > vfs_getxattr_alloc() (or modify it) to behave like vfs_getxattr() but do
> > the memory allocation part so that the kernel sees what user space see
> > rather than modifying it with your patch so that user space now sees
> > something different than what it has been for years (previous
> > NUL-terminated SELinux xattr may not be NUL-terminated anymore)?
> 
> I'm concerned that this would break HMACs/digital signatures
> calculated with raw values.

Which would happen if the LSM is not enabled (e.g. "lsm=" boot command
line option).

> 
> An alternative would be to do the EVM verification twice if the
> first time didn't succeed (with vfs_getxattr_alloc() and with the
> new function that behaves like vfs_getxattr()).

Unfortunately, I don't see an alternative.

thanks,

Mimi