Re: overlayfs access checks on underlying layers

[Vivek Goyal]

On Tue, Dec 04, 2018 at 10:15:49AM -0500, Vivek Goyal wrote:
> On Tue, Dec 04, 2018 at 09:30:53AM -0500, Stephen Smalley wrote:
> > On 12/4/18 8:32 AM, Miklos Szeredi wrote:
> > > On Thu, Nov 29, 2018 at 10:16 PM Stephen Smalley <sds@xxxxxxxxxxxxx> wrote:
> > > > 
> > > > On 11/29/18 4:03 PM, Stephen Smalley wrote:
> > > > > On 11/29/18 2:47 PM, Miklos Szeredi wrote:
> > > > > > On Thu, Nov 29, 2018 at 5:14 PM Stephen Smalley <sds@xxxxxxxxxxxxx>
> > > > > > wrote:
> > > > > > 
> > > > > > > Possibly I misunderstood you, but I don't think we want to copy-up on
> > > > > > > permission denial, as that would still allow the mounter to read/write
> > > > > > > special files or execute regular files to which it would normally be
> > > > > > > denied access, because the copy would inherit the context specified by
> > > > > > > the mounter in the context mount case.  It still represents an
> > > > > > > escalation of privilege for the mounter.  In contrast, the copy-up on
> > > > > > > write behavior does not allow the mounter to do anything it could not do
> > > > > > > already (i.e. read from the lower, write to the upper).
> > > > > > 
> > > > > > Let's get this straight:  when file is copied up, it inherits label
> > > > > > from context=, not from label of lower file?
> > > > > 
> > > > > That's correct.  The overlay inodes are all assigned the label from the
> > > > > context= mount option, and so are any upper inodes created through the
> > > > > overlay.  At least that's my understanding of how it is supposed to
> > > > > work.  The original use case was for containers with the lower dir
> > > > > labeled with a context that is read-only to the container context and
> > > > > using a context that is writable by the container context for the
> > > > > context= mount.
> > > > > 
> > > > > > Next question: permission to change metadata is tied to permission to
> > > > > > open?  Is it possible that open is denied, but metadata can be
> > > > > > changed?
> > > > > 
> > > > > There is no metadata change occurring here. The overlay, upper, and
> > > > > lower inodes all keep their labels intact for their lifetime (both
> > > > > overlay and upper always have the context= label; upper has whatever its
> > > >                                                     ^^lower^^
> > > > 
> > > > > original label was), unless explicitly relabeled by some process.  And
> > > > > when viewed through the overlay, the file always has the label specified
> > > > > via context=, even before the copy-up.
> > > 
> > > Okay.
> > > 
> > > > > 
> > > > > > DAC model allows this: metadata change is tied to ownership, not mode
> > > > > > bits.   And different capability flag.
> > > > > > 
> > > > > > If the same is true for MAC, then the pre-v4.20-rc1 is already
> > > > > > susceptible to the privilege escalation you describe, right?
> > > > > 
> > > > > Actually, I guess there wouldn't be a privilege escalation if you
> > > > > checked the mounter's ability to create the new file upon copy-up, and
> > > > > checked the mounter's access to the upper inode label upon the
> > > > > subsequent read, write, or execute access.  Then we'd typically block
> > > > > the ability to create the device file and we'd block the ability to
> > > > > execute files with the label from context=.
> > > > > 
> > > > > But copy-up of special files seems undesirable for other reasons (e.g.
> > > > > requiring mounters to be allowed to create device nodes just to permit
> > > > > client's to read/write them, possible implications for nodev/noexec,
> > > > > implications for socket and fifo files).
> > > 
> > > I think you missed my point: opening a device file or executing an
> > > executable wouldn't normally require copy-up.   If
> > > 
> > >   -  permission is granted on overlay to task, and
> > >   -  permission is granted on lower layer to mounter,
> > > 
> > > then copy-up wouldn't be performed.
> > > 
> > > My proposed sequence would be
> > > 
> > > a) check task's creds against overlay inode, fail -> return fail, otherwise:
> > > b) check mounter's creds against lower inode, success -> return
> > > success, otherwise:
> > > c) copy up inode, fail -> return fail, otherwise
> > > d) check mounter's creds against upper inode, return result.
> > > 
> > > So, unlike write access to regular files, write access to special
> > > files don't necessarily result in copy-up.
> > > 
> > > You say this is an escalation of privilege, but I don't get it how.
> > > As DWalsh points out downthread, if mounter cannot create device
> > > files, then the copy-up will simply fail.  If mounter can create
> > > device files, then this is not an escalation of privilege for the
> > > mounter.
> > 
> > Yes, in that case there isn't an escalation of privilege for the mounter (I
> > acknowledged that above).  I'm still not sure copy-up of special files is a
> > good idea though:
> > 
> > - In the case of device files, there is the potential for mischief by the
> > client task in misusing the mounter's privileges to gain access to otherwise
> > unusable device node (nodev lower vs upper?),
> 
> I was thinking about it as well. But client can always bypass permissions
> of lower device inode by first removing device file and then by doing
> a mknod. And that will be equivalent of copy up. IOW, IIUC, we do not deny
> mknod to client and that always creates a way for it to write to device
> file (and it does not matter what are permissions on lower?)

Having said that, this still create little anomaly when mknod to client
is not allowed on context label. So a device file, which is on lower
and client can not open it for read/write on host, it can now be opened
for read/write because mounter will allow access. So why it is different
that regular copy up. Well, in regular copy up, we created a copy of
the original object and allowed writing to that object (cp --preserve=all)
model. But in case of device file, writes will go to same original
object. (And not a separate copy).

Thanks
Vivek