Bug#770492: [RFC PATCH RESEND] vfs: Move security_inode_killpriv() after permission checks

[Casey Schaufler <casey@schaufler-ca.com>]

On 1/20/2015 3:17 PM, James Morris wrote:
> On Sat, 17 Jan 2015, Ben Hutchings wrote:
>
>> chown() and write() should clear all privilege attributes on
>> a file - setuid, setgid, setcap and any other extended
>> privilege attributes.
>>
>> However, any attributes beyond setuid and setgid are managed by the
>> LSM and not directly by the filesystem, so they cannot be set along
>> with the other attributes.
>>
>> Currently we call security_inode_killpriv() in notify_change(),
>> but in case of a chown() this is too early - we have not called
>> inode_change_ok() or made any filesystem-specific permission/sanity
>> checks.
>>
>> Add a new function setattr_killpriv() which calls
>> security_inode_killpriv() if necessary, and change the setattr()
>> implementation to call this in each filesystem that supports xattrs.
>> This assumes that extended privilege attributes are always stored in
>> xattrs.
> It'd be useful to get some input from LSM module maintainers on this.

I've already chimed in.

Clearing the Smack label on a file because someone writes to it
makes no sense whatsoever. The same with chown. The Smack label is
attached to the object, which is a container of data, not the data
itself. Smack labels are Mandatory Access Control labels, not Information
labels. If that doesn't mean anything to the reader, check out the
P1003.1e/2c (withdrawn) DRAFT.

The proposed implementation does not correctly handle either
Mandatory Access Control labels or Information labels. The MAC
label is *very different* from the setuid bit.

>
> e.g. doesn't SELinux already handle this via policy directives?
>
>