On Mon, Oct 26, 2020 at 10:51 AM Jann Horn <jannh@xxxxxxxxxx> wrote:
> On Mon, Oct 26, 2020 at 1:32 AM Kees Cook <keescook@xxxxxxxxxxxx> wrote:
> > On Thu, Oct 01, 2020 at 03:52:02AM +0200, Jann Horn wrote:
> > > On Thu, Oct 1, 2020 at 1:25 AM Tycho Andersen <tycho@tycho.pizza> wrote:
> > > > On Thu, Oct 01, 2020 at 01:11:33AM +0200, Jann Horn wrote:
> > > > > On Thu, Oct 1, 2020 at 1:03 AM Tycho Andersen <tycho@tycho.pizza> wrote:
> > > > > > On Wed, Sep 30, 2020 at 10:34:51PM +0200, Michael Kerrisk (man-pages) wrote:
> > > > > > > On 9/30/20 5:03 PM, Tycho Andersen wrote:
> > > > > > > > On Wed, Sep 30, 2020 at 01:07:38PM +0200, Michael Kerrisk (man-pages) wrote:
> > > > > > > >> ┌─────────────────────────────────────────────────────┐
> > > > > > > >> │FIXME │
> > > > > > > >> ├─────────────────────────────────────────────────────┤
> > > > > > > >> │From my experiments, it appears that if a SEC‐ │
> > > > > > > >> │COMP_IOCTL_NOTIF_RECV is done after the target │
> > > > > > > >> │process terminates, then the ioctl() simply blocks │
> > > > > > > >> │(rather than returning an error to indicate that the │
> > > > > > > >> │target process no longer exists). │
> > > > > > > >
> > > > > > > > Yeah, I think Christian wanted to fix this at some point,
> > > > > > >
> > > > > > > Do you have a pointer that discussion? I could not find it with a
> > > > > > > quick search.
> > > > > > >
> > > > > > > > but it's a
> > > > > > > > bit sticky to do.
> > > > > > >
> > > > > > > Can you say a few words about the nature of the problem?
> > > > > >
> > > > > > I remembered wrong, it's actually in the tree: 99cdb8b9a573 ("seccomp:
> > > > > > notify about unused filter"). So maybe there's a bug here?
> > > > >
> > > > > That thing only notifies on ->poll, it doesn't unblock ioctls; and
> > > > > Michael's sample code uses SECCOMP_IOCTL_NOTIF_RECV to wait. So that
> > > > > commit doesn't have any effect on this kind of usage.
> > > >
> > > > Yes, thanks. And the ones stuck in RECV are waiting on a semaphore so
> > > > we don't have a count of all of them, unfortunately.
> > > >
> > > > We could maybe look inside the wait_list, but that will probably make
> > > > people angry :)
> > >
> > > The easiest way would probably be to open-code the semaphore-ish part,
> > > and let the semaphore and poll share the waitqueue. The current code
> > > kind of mirrors the semaphore's waitqueue in the wqh - open-coding the
> > > entire semaphore would IMO be cleaner than that. And it's not like
> > > semaphore semantics are even a good fit for this code anyway.
> > >
> > > Let's see... if we didn't have the existing UAPI to worry about, I'd
> > > do it as follows (*completely* untested). That way, the ioctl would
> > > block exactly until either there actually is a request to deliver or
> > > there are no more users of the filter. The problem is that if we just
> > > apply this patch, existing users of SECCOMP_IOCTL_NOTIF_RECV that use
> > > an event loop and don't set O_NONBLOCK will be screwed. So we'd
> >
> > Wait, why? Do you mean a ioctl calling loop (rather than a poll event
> > loop)?
>
> No, I'm talking about poll event loops.
>
> > I think poll would be fine, but a "try calling RECV and expect to
> > return ENOENT" loop would change. But I don't think anyone would do this
> > exactly because it _currently_ acts like O_NONBLOCK, yes?
> >
> > > probably also have to add some stupid counter in place of the
> > > semaphore's counter that we can use to preserve the old behavior of
> > > returning -ENOENT once for each cancelled request. :(
> >
> > I only see this in Debian Code Search:
> > https://sources.debian.org/src/crun/0.15+dfsg-1/src/libcrun/seccomp_notify.c/?hl=166#L166
> > which is using epoll_wait():
> > https://sources.debian.org/src/crun/0.15+dfsg-1/src/libcrun/container.c/?hl=1326#L1326
> >
> > I expect LXC is using it. :)
>
> The problem is the scenario where a process is interrupted while it's
> waiting for the supervisor to reply.
>
> Consider the following scenario (with supervisor "S" and target "T"; S
> wants to wait for events on two file descriptors seccomp_fd and
> other_fd):
>
> S: starts poll() to wait for events on seccomp_fd and other_fd
> T: performs a syscall that's filtered with RET_USER_NOTIF
> S: poll() returns and signals readiness of seccomp_fd
> T: receives signal SIGUSR1
> T: syscall aborts, enters signal handler
> T: signal handler blocks on unfiltered syscall (e.g. write())
> S: starts SECCOMP_IOCTL_NOTIF_RECV
> S: blocks because no syscalls are pending
>
> Depending on what other_fd is, this could in a worst case even lead to
> a deadlock (if e.g. the signal handler wants to write to stdout, but
> the stdout fd is hooked up to other_fd in the supervisor, but the
> supervisor can't consume the data written because it's stuck in
> seccomp handling).
>
> So we have to ensure that when existing code (like that crun code you
> linked to) triggers this case, SECCOMP_IOCTL_NOTIF_RECV returns
> immediately instead of blocking.
Or I guess we could also just set O_NONBLOCK on the fd by default?
Since the one existing user is eventloop-based...
