Re: [PATCH v2 resend] vfs: new O_NODE open flag

[ebiederm@xmission.com (Eric W. Biederman)]

Miklos Szeredi <miklos@szeredi.hu> writes:

> On Sat, 07 Nov 2009, ebiederm@xmission.com (Eric W. Biederman wrote:
>> Miklos Szeredi <miklos@szeredi.hu> writes:
>> 
>> > On Fri, 06 Nov 2009, ebiederm@xmission.com (Eric W. Biederman wrote:
>> >> So far no one who believes this to be a security hole has found it
>> >> worth their while to look at nd->intent.open in proc_pid_follow_link
>> >> and write a patch.
>> >
>> > A rather disgusting patch that would be.  The fact is, checking
>> > permissions on follow_link makes little to no sense.  Consider
>> > truncate(2), for example.  Will we add another intent for that?  I
>> > really hope not
>> 
>> No.  I was just thinking we have the open intent that is there for
>> combining lookup and open. We can look test for LOOKUP_OPEN and do
>> exactly what we need.
>
> No, because you just covered half the cases.  truncate(2) will still
> work fine on the /proc/PID/fd/FD belonging to a O_RDONLY file
> descriptor.

Good point as truncate is as bad as opening read-write.  Shrug.

>> > I'm more and more convinced, that the current behavior is the right
>> > one.
>> 
>> I think the 15 or so years we have had the current behavior without
>> problems is persuasive.
>> 
>> I think it is an interesting puzzle on how to get dup instead of
>> reopen as there are cases where that could be useful behavior as well.
>
> Probably doable with ptrace().
>
>> The usefulness of an O_NONE flag increases significantly if you can
>> open the reference file later with more permissions.  Essentially
>> making a hardlink into a running program.  Hmm.  Weird cases do seem
>> to show up when the last dir entry is removed.  
>
> Why are they more weird than files opened without O_NODE?

Because as I recall O_NODE skips all permission checks.  Which would
mean you can limit who holds a reference to your inode without O_NODE.

> The only really weird case Alan spotted is device nodes, where the
> actual device registered to a major/minor pair changes over time,
> possibly allowing a re-open to access a device it otherwise was not
> meant to.  BTW if the device number reuse happens really quickly, this
> could even be a race for a plain open.  Real solution might actually
> be in udev: when deregistering a device, change mode bits to all-zero
> before removing the device node.

Devices nodes specifically were the case I was thinking of.

Changing the mode bits to all-zero at the final unlink would be a lot
more reliable and certain in the kernel.

>> I wonder if we want a rule that you can't open a file with link count
>> of 0.  Reasoning may get truly strange otherwise.
>
> Again, don't see why this would be different for O_NODE as for
> non-O_NODE files descriptors.


Eric