Re: [PATCH v5 6/6] seccomp: add SECCOMP_EXT_ACT_TSYNC and SECCOMP_FILTER_TSYNC

From: Andy Lutomirski
Date: Tue May 27 2014 - 14:41:03 EST


On Tue, May 27, 2014 at 11:24 AM, Kees Cook <keescook@xxxxxxxxxxxx> wrote:
> On Mon, May 26, 2014 at 12:27 PM, Andy Lutomirski <luto@xxxxxxxxxxxxxx> wrote:
>> On Fri, May 23, 2014 at 10:05 AM, Kees Cook <keescook@xxxxxxxxxxxx> wrote:
>>> On Thu, May 22, 2014 at 4:11 PM, Andy Lutomirski <luto@xxxxxxxxxxxxxx> wrote:
>>>> On Thu, May 22, 2014 at 4:05 PM, Kees Cook <keescook@xxxxxxxxxxxx> wrote:
>>>>> Applying restrictive seccomp filter programs to large or diverse
>>>>> codebases often requires handling threads which may be started early in
>>>>> the process lifetime (e.g., by code that is linked in). While it is
>>>>> possible to apply permissive programs prior to process start up, it is
>>>>> difficult to further restrict the kernel ABI to those threads after that
>>>>> point.
>>>>>
>>>>> This change adds a new seccomp extension action for synchronizing thread
>>>>> group seccomp filters and a prctl() for accessing that functionality,
>>>>> as well as a flag for SECCOMP_EXT_ACT_FILTER to perform sync at filter
>>>>> installation time.
>>>>>
>>>>> When calling prctl(PR_SECCOMP_EXT, SECCOMP_EXT_ACT, SECCOMP_EXT_ACT_FILTER,
>>>>> flags, filter) with flags containing SECCOMP_FILTER_TSYNC, or when calling
>>>>> prctl(PR_SECCOMP_EXT, SECCOMP_EXT_ACT, SECCOMP_EXT_ACT_TSYNC, 0, 0), it
>>>>> will attempt to synchronize all threads in current's threadgroup to its
>>>>> seccomp filter program. This is possible iff all threads are using a filter
>>>>> that is an ancestor to the filter current is attempting to synchronize to.
>>>>> NULL filters (where the task is running as SECCOMP_MODE_NONE) are also
>>>>> treated as ancestors allowing threads to be transitioned into
>>>>> SECCOMP_MODE_FILTER. If prctrl(PR_SET_NO_NEW_PRIVS, ...) has been set on the
>>>>> calling thread, no_new_privs will be set for all synchronized threads too.
>>>>> On success, 0 is returned. On failure, the pid of one of the failing threads
>>>>> will be returned, with as many filters installed as possible.
>>>>
>>>> Is there a use case for adding a filter and synchronizing filters
>>>> being separate operations? If not, I think this would be easier to
>>>> understand and to use if there was just a single operation.
>>>
>>> Yes: if the other thread's lifetime is not well controlled, it's good
>>> to be able to have a distinct interface to retry the thread sync that
>>> doesn't require adding "no-op" filters.
>>
>> Wouldn't this still be solved by:
>>
>> seccomp_add_filter(final_filter, SECCOMP_FILTER_ALL_THREADS);
>>
>> the idea would be that, if seccomp_add_filter fails, then you give up
>> and, if it succeeds, then you're done. It shouldn't fail unless out
>> of memory or you've nested too deeply.
>
> I wanted to keep the case of being able to to wait for non-ancestor
> threads to finish. For example, 2 threads start and set separate
> filters. 1 does work and exits, 2 starts another thread (3) which adds
> filters, does work, and then waits for 1 to finish by calling TSYNC.
> Once 1 dies, TSYNC succeeds. In the case of not having direct control
> over thread lifetime (say, when using third-party libraries), I'd like
> to retain the flexibility of being able to do TSYNC without needing a
> filter being attached to it.

I must admit this strikes me as odd. What's the point of having a
thread set a filter if it intends to be a short-lived thread?

In any cast, I must have missed the ability for TSYNC to block. Hmm.
That seems complicated, albeit potentially useful.

>
>>>> If you did that, you'd have to decide whether to continue requiring
>>>> that all the other threads have a filter that's an ancestor of the
>>>> current thread's filter.
>>>
>>> This is required no matter what to make sure there is no way to
>>> replace a filter tree with a different one (allowing accidental
>>> bypasses, misbehavior, etc).
>>
>> What I mean is: should the add-new-filter-to-all-threads operation
>> add the new filter to all threads, regardless of what their current
>> state is, or should it fail if any thread has a filter that isn't an
>> ancestor of the current thread's filter? Either version should be
>> safe.
>
> It should fail -- we don't want to run the risk of effectively
> replacing a filter out from under a thread. Adding additional
> restrictions is safe as long as we retain the nnp from the caller .

The other option would be to add the new filter. The threads wouldn't
end up being ancestors of each other, but that could still be okay.

In any case, I don't have a strong feeling about this particular issue.

--Andy
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/