Re: [PATCH 1/2] procfs: restore 0400 permissions on /proc/*/{syscall,stack,personality}

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

Kees Cook <keescook@chromium.org> writes:

> On Wed, Aug 28, 2013 at 6:08 PM, Eric W. Biederman
> <ebiederm@xmission.com> wrote:
>> Kees Cook <keescook@chromium.org> writes:
>>
>>> On Wed, Aug 28, 2013 at 5:26 PM, Eric W. Biederman
>>> <ebiederm@xmission.com> wrote:
>>>> Can someome please state what they are worried about in simple language
>>>> step by step?
>>>> [...]
>>>> The closest I saw in the thread was people were worried about ASLR being
>>>> defeated.  All I see are kernel addresses and we don't have much if any
>>>> runtime or even load time randomization of where code is located in the
>>>> kernel address map on x86_64.  So I don't understand the concern.
>>>
>>> I showed the output of "syscall", since that contains user-space
>>> addresses and shows a leak of ASLR from a privileged process to an
>>> unprivileged process.
>>>
>>> The flaw as I see it is that an unprivileged process opens
>>> /proc/$priv_pid/syscall and passes it to a setuid process which is
>>> able to read it, and provides those contents to the unprivileged
>>> process.
>>>
>>> The unprivileged process should not be able to the open the file in
>>> the first place.
>>
>> I see so the complaint is that we don't give read permission but we do
>> give open permission.    Which is a variant of: the permissions used to
>> open are not the permission used to access the file.
>>
>> This does seem to be a legitimate concern.
>>
>> I think there are several discussions that have been going on lately
>> with respect to this class of problems in proc files.
>>
>> Given the existence of suid exec we can not in general prevent this
>> class of bugs with a check at open time.
>
> I'm not suggesting removing the read check -- that's certainly needed.
>
>> I believe the solution needs to be to enhance the ptrace_may_access
>> checks to verify that both the creds of the current task and the creds
>> of the opening process would have allowed the access.
>
> As in, DAC perms are insufficient?

As in any checks at open time are insufficient.

Roughly what we need is to use the credentials that are present at open
time (file->f_cred) in the ptrace_may_access call instead of or
in addition to current_cred().

Eric