Re: [PATCH v3 2/4] xfs: initialize inode security on tmpfile creation

[Stephen Smalley <sds@tycho.nsa.gov>]

On 04/15/2014 04:21 PM, Stephen Smalley wrote:
> On 04/15/2014 04:22 PM, Christoph Hellwig wrote:
>> On Tue, Apr 15, 2014 at 04:04:32PM -0400, Stephen Smalley wrote:
>>> Is there a reason that xfs_init_security() isn't called from the inode
>>> allocation function (e.g. xfs_ialloc), as in ext4 (__ext4_new_inode
>>> calls ext4_init_security and also calls ext4_init_acl)?  That would have
>>> ensured that tmpfile inodes would have been labeled without requiring a
>>> separate change and more generally ensures complete coverage for all inodes.
>>
>> Really just code structuring - we don't like callouts to high level VFS
>> functions from deep down in the guts of the filesystem.
>>
>>> For SELinux, we need the tmpfile inodes to be labeled at creation time,
>>> not just if linked into the namespace, since they may be shared via
>>> local socket IPC or inherited across a label-changing exec and since we
>>> revalidate access on transfer or use.
>>>
>>> Labeling based on the provided directory could be a bit random, although
>>> it will work out with current policy if the provided directory
>>> corresponds to existing tmpfile locations (e.g. /tmp, /var/tmp) and
>>> therefore already has a label associated with temporary files.
>>> Otherwise we might want some indication that it is a tmpfile passed into
>>> security_inode_init_security() so that we can always select a stable
>>> label irrespective of the directory.
>>
>> Just check for I_LINKABLE in i_flags.
> 
> Thanks, that should allow us to handle it cleanly in the security modules!

Maybe I spoke too soon.  IIUC, I_LINKABLE doesn't necessarily
distinguish tmpfiles from other files, as some tmpfiles may be linkable
and others not.  But what we want is a way to identify all tmpfiles when
security_inode_init_security() is called if we are going to label them
independently of the provided dir.

Also, in that situation, we would need to likewise distinguish them
during the create-time checking, i.e. when security_inode_create() is
called (from may_o_create), as we have to determine the label that will
be applied at that point for permission checking.  And there we do not
have the inode yet so we do not even have I_LINKABLE as a distinguisher.

So I think we need __O_TMPFILE or similar flag passed down to
may_o_create() -> security_inode_create() and to
security_inode_init_security() if we are going to label these files
independently of the provided directory.

_______________________________________________
xfs mailing list
xfs@oss.sgi.com
http://oss.sgi.com/mailman/listinfo/xfs