On Mon, May 06, 2024 at 07:50:57PM GMT, Laurent Pinchart wrote:
> On Mon, May 06, 2024 at 10:57:17AM -0400, Sean Anderson wrote:
> > On 5/6/24 03:35, Laurent Pinchart wrote:
> > > On Mon, May 06, 2024 at 09:29:36AM +0200, Maxime Ripard wrote:
> > >> Hi Laurent, Sean,
> > >> 
> > >> On Sat, May 04, 2024 at 03:21:18PM GMT, Laurent Pinchart wrote:
> > >> > On Fri, May 03, 2024 at 05:54:32PM -0400, Sean Anderson wrote:
> > >> > > I have discovered a bug in the displayport driver on drm-misc-next. To
> > >> > > trigger it, run
> > >> > > 
> > >> > > echo fd4a0000.display > /sys/bus/platform/drivers/zynqmp-dpsub/unbind
> > >> > > 
> > >> > > The system will become unresponsive and (after a bit) splat with a hard
> > >> > > LOCKUP. One core will be unresponsive at the first zynqmp_dp_read in
> > >> > > zynqmp_dp_bridge_detect.
> > >> > > 
> > >> > > I believe the issue is due the registers being unmapped and the block
> > >> > > put into reset in zynqmp_dp_remove instead of zynqmp_dpsub_release.
> > >> > 
> > >> > That is on purpose. Drivers are not allowed to access the device at all
> > >> > after .remove() returns.
> > >> 
> > >> It's not "on purpose" no. Drivers indeed are not allowed to access the
> > >> device after remove, but the kernel shouldn't crash. This is exactly
> > >> why we have drm_dev_enter / drm_dev_exit.
> > > 
> > > I didn't mean the crash was on purpose :-) It's the registers being
> > > unmapped that is, as nothing should touch those registers after
> > > .remove() returns.
> > 
> > OK, so then we need to have some kind of flag in the driver or in the drm
> > subsystem so we know not to access those registers.
> 
> To avoid race conditions, the .remove() function should mark the device
> as removed, wait for all ongoing access from userspace to be complete,
> and then proceed to unmapping registers and doing other cleanups.
> Userspace may still have open file descriptors to the device at that
> point. Any new userspace access should be disallowed (by checking the
> removed flag), with the only userspace-initiated operations that still
> need to run being the release-related operations (unmapping memory,
> closing file descriptors, ...).

And for the record, this is exactly what drm_dev_unplug and
drm_dev_enter/drm_dev_exit does.

Maxime