[PATCH v2] proc: allow restricting /proc/pid/mem writes

[Christian Brauner]

On Tue, Mar 05, 2024 at 10:58:31AM +0100, Christian Brauner wrote:
> On Tue, Mar 05, 2024 at 01:41:29AM -0800, Kees Cook wrote:
> > On Tue, Mar 05, 2024 at 09:59:47AM +0100, Christian Brauner wrote:
> > > > > Uhm, this will break the seccomp notifier, no? So you can't turn on
> > > > > SECURITY_PROC_MEM_RESTRICT_WRITE when you want to use the seccomp
> > > > > notifier to do system call interception and rewrite memory locations of
> > > > > the calling task, no? Which is very much relied upon in various
> > > > > container managers and possibly other security tools.
> > > > > 
> > > > > Which means that you can't turn this on in any of the regular distros.
> > > > 
> > > > FWIW, it's a run-time toggle, but yes, let's make sure this works
> > > > correctly.
> > > > 
> > > > > So you need to either account for the calling task being a seccomp
> > > > > supervisor for the task whose memory it is trying to access or you need
> > > > > to provide a migration path by adding an api that let's caller's perform
> > > > > these writes through the seccomp notifier.
> > > > 
> > > > How do seccomp supervisors that use USER_NOTIF do those kinds of
> > > > memory writes currently? I thought they were actually using ptrace?
> > > > Everything I'm familiar with is just using SECCOMP_IOCTL_NOTIF_ADDFD,
> > > > and not doing fancy memory pokes.
> > > 
> > > For example, incus has a seccomp supervisor such that each container
> > > gets it's own goroutine that is responsible for handling system call
> > > interception.
> > > 
> > > If a container is started the container runtime connects to an AF_UNIX
> > > socket to register with the seccomp supervisor. It stays connected until
> > > it stops. Everytime a system call is performed that is registered in the
> > > seccomp notifier filter the container runtime will send a AF_UNIX
> > > message to the seccomp supervisor. This will include the following fds:
> > > 
> > > - the pidfd of the task that performed the system call (we should
> > >   actually replace this with SO_PEERPIDFD now that we have that)
> > > - the fd of the task's memory to /proc/<pid>/mem
> > > 
> > > The seccomp supervisor will then perform the system call interception
> > > including the required memory reads and writes.
> > 
> > Okay, so the patch would very much break that. Some questions, though:
> > - why not use process_vm_writev()?
> 
> Because it's inherently racy as I've explained in an earlier mail in
> this thread. Opening /proc/<pid>/mem we can guard via:
> 
> // Assume we hold @pidfd for supervised process
> 
> int fd_mem = open("/proc/$pid/mem", O_RDWR);:
> 
> if (pidfd_send_signal(pidfd, 0, ...) == 0)
>         write(fd_mem, ...);
> 
> But we can't exactly do:
> 
> process_vm_writev(pid, WRITE_TO_MEMORY, ...);
> if (pidfd_send_signal(pidfd, 0, ...) == 0)
>         write(fd_mem, ...);
> 
> That's always racy. The process might have been reaped before we even
> call pidfd_send_signal() and we're writing to some random process
> memory.
> 
> If we wanted to support this we'd need to implement a proposal I had a
> while ago:
> 
> #define PROCESS_VM_RW_PIDFD (1 << 0)
> 
> process_vm_readv(pidfd,  ..., PROCESS_VM_RW_PIDFD);
> process_vm_writev(pidfd, ..., PROCESS_VM_RW_PIDFD);
> 
> which is similar to what we did for waitid(pidfd, P_PIDFD, ...)
> 
> That would make it possible to use a pidfd instead of a pid in the two
> system calls. Then we can get rid of the raciness and actually use those
> system calls. As they are now, we can't.

What btw, is the Linux sandbox on Chromium doing? Did they finally move
away from SECCOMP_RET_TRAP to SECCOMP_RET_USER_NOTIF? I see:

https://issues.chromium.org/issues/40145101

What ever became of this?