[RFC 0/3] seccomp trap to userspace

[Andy Lutomirski]

> On Mar 16, 2018, at 7:47 AM, Christian Brauner <christian.brauner at mailbox.org> wrote:
> 
>> On Fri, Mar 16, 2018 at 12:46:55AM +0000, Andy Lutomirski wrote:


I bet I confused everyone with a blatant typo:

>> 
>> Hmm, I think we have to be very careful to avoid nasty races.  I think
>> the correct approach is to notice the signal and send a message to the
>> listener that a signal is pending but to take no additional action.
>> If the handler ends up completing the syscall with a successful
>> return, we don't want to replace it with -EINTR.  IOW the code looks
>> kind of like:
>> 
>> send_to_listener("hey I got a signal");

That should be “hey I got a syscall”.   D’oh!

>> wait_ret = wait_interruptible for the listener to reply;
>> if (wait_ret == -EINTR) {
> 
> Hm, so from the pseudo-code it looks like: The handler would inform the
> listener that it received a signal (either from the syscall requester or
> from somewhere else) and then wait for the listener to reply to that
> message.  This would allow the listener to decide what action it wants
> the handler to take based on the signal, i.e. either cancel the request
> or retry?  The comment makes it sound like that the handler doesn't
> really wait on the listener when it receives a signal it simply moves
> on.

It keeps waiting killably but not interruptibly. 

> So no "taking no additional action" here means not have the handler
> decide to abort but the listener?

If by “handler” you mean kernel, then yes. 

There’s no userspace syscall handler involved. From the kernel’s perspective, a syscall is never still in progress when a signal handler is invoked — we only actually invoke syscall handlers in prepare_exit_to_usermode() or the non-x86 equivalent and the functions it calls. While a syscall is running, the kernel might notice that a signal is pending and do one of a few things:

1. Just keep going. Not all syscalls can be interrupted. 

2. Try to finish early. If a send() call has already sent some but not all data, it can stop waiting and return the number of bytes sent.

3. Abort with -EINTR.

4. Abort with -ERESTARTSYS or one of its relatives. These fiddle with user registers in a somewhat unpleasant way to pretend that the syscall never actually happened.  This works for syscalls that wait with an absolute timeout, for example. 

5. Set up restart_syscall() magic, rewrite regs so it looks like the user was about to call restart_syscall() when the signal happened, and abort. 

In all cases, the signal is dealt with afterwards. This could result in changing regs to call the handler or in simply returning. 

1-3 should work fully in seccomp. The only issue is that the kernel doesn’t know *which* to do, nor can the kernel force the listener to abort cleanly, so I think we have  no real choice but to let the listener decide. 

4 could be supported just like 1-3. 5 is awful, and I don’t think we should support it for user listeners.