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

[Kent Gibson]

On Tue, Sep 13, 2022 at 04:35:08PM +0200, Bartosz Golaszewski wrote:
> On Tue, Sep 13, 2022 at 4:28 PM Kent Gibson <warthog618@xxxxxxxxx> wrote:
> >
> > On Tue, Sep 13, 2022 at 10:54:26AM +0200, Bartosz Golaszewski wrote:
> > > 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.
> > >
> >
> > Then you are doing it backwards - daemonize first ;-).
> >
> > Generally speaking, doesn't transfer of resource ownership to the forked
> > child create havoc in multi-threaded apps? i.e. one thread requests a
> > resource, another forks.  The parent thread unknowingly loses ownership,
> > and the forked child process only starts with a replica of the forking
> > thread.
> >
> 
> Yeah, sounds like a bad idea.
> 
> > > 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.
> > >
> >
> > Yeah, missed that :-(.
> >
> > Makes sense.
> >
> > As each request fd can contain multiple lines on a particular chip,
> > you would need to identify the gpiochip and the offsets for that request.
> > So two fields - the gpiochip path, and the list of offsets.
> >
> > Is that already too clunky or am I missing something?
> >
> 
> It's worse than that - we don't know the character device's filesystem
> path in gpiolib. Nor should we, as we can be in a different fs
> namespace when checking it than in which we were when we opened the
> device (which is also another concern for storing the path to the
> character device in struct gpiod_chip - unless we specify explicitly
> that it's the path that was used to open it). Since we don't know it -
> we can only get it from the file descriptor that the requesting
> process got after calling open() on the GPIO device. But this fd may
> have been closed in the meantime. I think I opened a can of worms with
> this one. :)
> 

Forgot that we don't have the path readily available in the kernel -
would device name suffice?

Cheers,
Kent.