On Wed, Jun 10, 2020 at 11:47:35AM +0200, Christian Brauner wrote:

On Tue, Jun 09, 2020 at 09:52:12PM -0700, Kees Cook wrote:

quoted

Hi,

This extends the recent work hch did for scm_detach_fds(), and updates
the compat path as well, fixing bugs in the process. Additionally,
an effectively incomplete and open-coded __scm_install_fd() is fixed
in pidfd_getfd().

Since __scm_detach_fds() becomes something that is available outside of
net/* should we provide a static inline wrapper under a different name? The
"socket-level control message" prefix seems a bit odd in pidfd_getfd()
and - once we make use of it there - seccomp.

I'd suggest we do:

static inline int fd_install_received(struct file *file, unsigned int flags)
{
	return __scm_install_fd(file, NULL, flags);
}

which can be called in pidfd_getfd() and once we have other callers that
want the additional put_user() (e.g. seccomp_ in there we simply add:

static inline fd_install_user(struct file *file, unsigned int flags, int __user *ufd)
{
	return __scm_install_fd(file, ufd, flags);
}

and seems the wrappers both could happily live in the fs part of the world?

Yeah, this seems good. I also note that randconfigs are kicking back my
series as broken when CONFIG_NET=n (oops), so this needs some refactoring
before patch 2.

-- 
Kees Cook