Re: [PATCH v3 2/2] staging: ion: create one device entry per heap
From: Greg KH
Date: Tue Sep 26 2017 - 02:56:22 EST
On Tue, Sep 26, 2017 at 07:09:14AM +0200, Daniel Vetter wrote:
> On Mon, Sep 25, 2017 at 11:26:27AM -0700, Laura Abbott wrote:
> > On 09/20/2017 01:45 AM, Benjamin Gaignard wrote:
> > > Instead a getting one common device "/dev/ion" for
> > > all the heaps this patch allow to create one device
> > > entry ("/dev/ionX") per heap.
> > > Getting an entry per heap could allow to set security rules
> > > per heap and global ones for all heaps.
> > >
> > > Allocation requests will be only allowed if the mask_id
> > > match with device minor.
> > > Query request could be done on any of the devices.
> > > Deivce node major will also change and that may impact init scripts.
> > >
> >
> > We should start Cc linux-api for future changes since we're going
> > to have more than just /dev/ion.
> >
> > Thinking about this some more, I think we need to allow backwards
> > compatibility. It's just not feasible to keep throwing workarounds
> > into userspace if we can avoid it. I'd propose keeping the old /dev/ion
> > misc interface available under a Kconfig and then creating the new
> > split heaps in parallel.
> >
> > On a somewhat related note, there was some interest in possibly
> > having a sysfs interface for Ion long term. I don't think this
> > needs to happen right now but I'd like whatever we do to not
> > make adding that harder.
>
> Not sure sysfs is a good idea for allocating buffers. The backlight
> interface is in sysfs, which defacto makes it a root-only interface.
> Distros really hate that, because it makes unpriviledged X/wayland really
> hard to pull of. Passing a device file otoh from logind to the compositor
> is dead simple (and how X et al get at e.g. the drm/input devices
> already).
sysfs is not a good idea for allocating buffers. We already have some
out-of-tree drm drivers doing horrid things in sysfs in ways that
totally abuse that api, and vendors have to do crazy things with selinux
rules to try to lock it down because of that. A device node is fine, we
are used to that for graphics stuff :)
thanks,
greg k-h