On Mon, Mar 15, 2021 at 05:46:43PM +0800, Yongji Xie wrote:

On Mon, Mar 15, 2021 at 5:08 PM Christoph Hellwig [off-list ref] wrote:

quoted

On Mon, Mar 15, 2021 at 01:37:11PM +0800, Xie Yongji wrote:

quoted

Export __receive_fd() so that some modules can use
it to pass file descriptor between processes.

I really don't think any non-core code should do that, especilly not
modular mere driver code.

Do you see any issue? Now I think we're able to do that with the help
of get_unused_fd_flags() and fd_install() in modules. But we may miss
some security stuff in this way. So I try to export __receive_fd() and
use it instead.

The __receive_fd() helper was added for core-kernel code only and we
mainly did it for the seccomp notifier (and scm rights). The "__" prefix
was intended to convey that message.
And I agree with Christoph that we should probably keep it that way
since __receive_fd() allows a few operations that no driver should
probably do.
I can see it being kinda ok to export a variant that really only
receives and installs an fd, i.e. if we were to export what's currently
available as an inline helper:

static inline int receive_fd(struct file *file, unsigned int o_flags)

but definitely none of the fd replacement stuff; that shold be
off-limits. The seccomp notifier is the only codepath that should even
think about fd replacement since it's about managing the syscalls of
another task. Drivers swapping out fds doesn't sound like a good idea to
me.

Christian