Re: [PATCH 2/2] gpiolib: cdev: export the consumer's PID

[Bartosz Golaszewski]

On Tue, Sep 13, 2022 at 4:12 AM Kent Gibson <warthog618@xxxxxxxxx> wrote:
>
> On Mon, Sep 12, 2022 at 11:56:17AM +0200, Bartosz Golaszewski wrote:
> > On Mon, Sep 12, 2022 at 11:53 AM Kent Gibson <warthog618@xxxxxxxxx> wrote:
> > >
> >
> > [snip]
> >
> > > >
> > > > Using -1 sounds good but I've just realized there's a different
> > > > problem. A process holding a file descriptor may fork and both the
> > > > parent and the child will keep the same file descriptors open. Now
> > > > we'll have two processes (with different PIDs) holding the same GPIO
> > > > lines (specifically holding a file descriptor to the same anonymous
> > > > inode).
> > > >
> > > > This already poses a problem for this patch as we'd need to return an
> > > > array of PIDs which we don't have the space for but also is a
> > > > situation which we haven't discussed previously IIRC - two processes
> > > > keeping the same GPIO lines requested.
> > > >
> > > > I don't have any good idea on how to address this yet. One thing off
> > > > the top of my head is: close the parent's file descriptor from kernel
> > > > space (is it even possible?) on fork() (kind of like the close() on
> > > > exec flag).
> > > >
> > > > I need to think about it more.
> > > >
> > >
> > > I thought the O_CLOEXEC was set on the request fds exactly to prevent this
> > > case - only one process can hold the request fd.
> > >
> >
> > O_CLOEXEC means "close on exec" not "close on fork". When you fork,
> > you inherit all file descriptors from your parent. Only once you call
> > execve() are the fds with this flag closed *in the child*.
> >
>
> Ah, ok.
> You want to pass request fd ownership from parent to child??
> Why not lock ownership to the parent, so O_CLOFORK, were that
> available?
>

Because what if we want to request a line and then daemonize i.e. fork
and exit in parent? It makes much more sense to keep the lines
requested in the child IMO.

During the BoF at Linux Plumbers it was suggested to use
/proc/$PID/fdinfo to expose the information about which lines are
requested but I can't figure out a way to do it elegantly.

Bart