[PATCH] drm: pl111: enable render node

[Eric Anholt]

On Mon, Apr 27, 2020 at 7:45 AM Emil Velikov <emil.l.velikov at gmail.com> wrote:
>
> On Fri, 24 Apr 2020 at 19:54, Peter Collingbourne <pcc at google.com> wrote:
> >
> > On Fri, Apr 24, 2020 at 4:11 AM Emil Velikov <emil.l.velikov at gmail.com> wrote:
> > >
> > > On Thu, 23 Apr 2020 at 23:51, Peter Collingbourne <pcc at google.com> wrote:
> > > >
> > > > The render node is required by Android which does not support the legacy
> > > > drmAuth authentication process.
> > > >
> > > > Signed-off-by: Peter Collingbourne <pcc at google.com>
> > > > ---
> > > >  drivers/gpu/drm/pl111/pl111_drv.c | 2 +-
> > > >  1 file changed, 1 insertion(+), 1 deletion(-)
> > > >
> > > The summary talks about drmAuth, yet exposes a render node. Even
> > > through there's no rendering engine in the HW, as mentioned by Eric.
> > >
> > > AFAICT the only way drmAuth is relevant with pl111 is when you want to
> > > export/import dma bufs.
> > > Although that is handled in core and the artificial DRM_AUTH
> > > restriction has been lifted with commit [1].
> > >
> > > -Emil
> > >
> > > [1] https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?h=v5.7-rc2&id=30a958526d2cc6df38347336a602479d048d92e7
> >
> > Okay, most likely drmAuth is irrelevant here (I don't know much about
> > it to be honest; I know that Android uses render nodes, so I figured
> > that drmAuth must therefore be the thing that it doesn't use). Sorry
> > for the confusion. Here is a better explanation of why I needed this
> > change.
> >
> > Android has a composer process that opens the primary node and uses
> > DRM_IOCTL_MODE_ATOMIC to switch between frame buffers, and a renderer
> > process (surfaceflinger) that opens the render node, prepares frame
> > buffers and sends them to the composer. One idea for adapting this
> > architecture to devices without render nodes is to have the renderer
> > process open the primary node instead. But this runs into a problem:
> > suppose that the renderer process starts before the composer process.
> > In this case, the kernel makes the renderer the DRM master, so the
> > composer can't change the frame buffer. Render nodes don't have this
> > problem because opening them doesn't affect the master.
> >
> > I considered fixing this by having the composer issue
> > DRM_IOCTL_SET_MASTER, but this requires root privileges. If we require
> > drivers to provide render nodes and control access to the primary node
> > while opening up the render node, we can ensure that only authorized
> > processes can become the master without requiring them to be root.
> >
> I think that the crux, is trying to open a _primary_ node for
> _rendering_ purposes.
> While that may work - as you outline above - it's usually a bad idea.
>
> To step back a bit, in practise we have three SoC combinations:
>  - complete lack of render device - then you default to software rendering [1]
>  - display and render device are one and the same - no change needed,
> things should just work
>  - display and render devices are separate - surfaceflinger should
> open the correct _rendering_ device node.
>
> [1] Mesa's libEGL (don't recall exact name on Android) can open VGEM
> render node, to work with the kms-swrast_dri.so software rasteriser
> module.
>
> While I did not try [1] with Android, it was working fine with CrOS. I
> suggest giving it a try and reporting bugs.

VGEM is not a solution, because it doesn't coordinate caching behavior
with your display node and so you get corruption if you try to to
share dma-bufs between them.  In cros it's used only for some tests as
far as I've heard, and it's been a source of pain.

If we want to go the route of "kms-only nodes also provide a render
node interface for doing swrast on", I think that would be a good
idea, but we should do this at a core DRM level (probably "does this
driver expose dma-buf? then also expose render nodes") rather than per
kms driver.