Re: [PATCH 1/2] seccomp: notify user trap about unused filter

[Kees Cook <keescook@chromium.org>]

On Wed, May 27, 2020 at 03:52:03PM -0600, Tycho Andersen wrote:
> On Wed, May 27, 2020 at 02:43:49PM -0700, Kees Cook wrote:
> > (While I'm here -- why can there be only one listener per task? The
> > notifications are filter-specific, not task-specific?)
> 
> Not sure what you mean here?

tatic struct file *init_listener(struct seccomp_filter *filter)
{
        struct file *ret = ERR_PTR(-EBUSY);
        struct seccomp_filter *cur;

        for (cur = current->seccomp.filter; cur; cur = cur->prev) {
                if (cur->notif)
                        goto out;
        }

...

        /* Installing a second listener in the chain should EBUSY */
        EXPECT_EQ(user_trap_syscall(__NR_getpid,
                                    SECCOMP_FILTER_FLAG_NEW_LISTENER),
                  -1);
        EXPECT_EQ(errno, EBUSY);


Why does this limit exist? Since the fd is tied to a specific filter,
I don't see conflicts about having multiple USER_NOTIF filters on one
task -- the monitor's response will either fake it or continue it, so
there is no "composition" needed? I must be missing something.

> Maybe it's best to decouple them entirely, and have usage go back to
> just being the number of tasks, and introduce a new counter for
> notification fds.

But isn't that already tracked by the VFS? i.e. there is a one-to-one
mapping from the "struct file *" returned by "anon_inode_getfile()" and
the "struct filter" (though we do not presently save it in the filter)
and the VFS tracks how many userspace fds are attached to that struct
file via ->f_count (i.e. f_count reaching zero is what triggers calling
seccomp_notify_release()).

In trying to write up an example patch for this, though, yeah, I don't
see how to do the locking. There is the "file" part, which is effectively
held by both any task's fd table and by the seccomp filter.

I suspect the issue is that the private_data can't be the
seccomp_filter. The "validity" of the mapping between kernel and user
needs to be tracked externally:

struct seccomp_notification_pipe
{
	struct filter *filter;
	struct file *file;
};

But I still can't see where to put the lock or refcount....

-- 
Kees Cook