Re: [RFC PATCH v1 01/12] Revert "drm/sysfs: Link DRM connectors to corresponding Type-C connectors"

[Laurent Pinchart]

On Wed, Sep 06, 2023 at 03:48:35PM +0300, Dmitry Baryshkov wrote:
> On Wed, 6 Sept 2023 at 15:44, Heikki Krogerus wrote:
> > On Tue, Sep 05, 2023 at 01:56:59PM +0300, Dmitry Baryshkov wrote:
> > > On Tue, 5 Sept 2023 at 11:50, Heikki Krogerus wrote:
> > > > On Mon, Sep 04, 2023 at 12:41:39AM +0300, Dmitry Baryshkov wrote:
> > > > > The kdev->fwnode pointer is never set in drm_sysfs_connector_add(), so
> > > > > dev_fwnode() checks never succeed, making the respective commit NOP.
> > > >
> > > > That's not true. The dev->fwnode is assigned when the device is
> > > > created on ACPI platforms automatically. If the drm_connector fwnode
> > > > member is assigned before the device is registered, then that fwnode
> > > > is assigned also to the device - see drm_connector_acpi_find_companion().
> > > >
> > > > But please note that even if drm_connector does not have anything in
> > > > its fwnode member, the device may still be assigned fwnode, just based
> > > > on some other logic (maybe in drivers/acpi/acpi_video.c?).
> > > >
> > > > > And if drm_sysfs_connector_add() is modified to set kdev->fwnode, it
> > > > > breaks drivers already using components (as it was pointed at [1]),
> > > > > resulting in a deadlock. Lockdep trace is provided below.
> > > > >
> > > > > Granted these two issues, it seems impractical to fix this commit in any
> > > > > sane way. Revert it instead.
> > > >
> > > > I think there is already user space stuff that relies on these links,
> > > > so I'm not sure you can just remove them like that. If the component
> > > > framework is not the correct tool here, then I think you need to
> > > > suggest some other way of creating them.
> > >
> > > The issue (that was pointed out during review) is that having a
> > > component code in the framework code can lead to lockups. With the
> > > patch #2 in place (which is the only logical way to set kdev->fwnode
> > > for non-ACPI systems) probing of drivers which use components and set
> > > drm_connector::fwnode breaks immediately.
> > >
> > > Can we move the component part to the respective drivers? With the
> > > patch 2 in place, connector->fwnode will be copied to the created
> > > kdev's fwnode pointer.
> > >
> > > Another option might be to make this drm_sysfs component registration optional.
> >
> > You don't need to use the component framework at all if there is
> > a better way of determining the connection between the DP and its
> > Type-C connector (I'm assuming that that's what this series is about).
> > You just need the symlinks, not the component.
> 
> The problem is that right now this component registration has become
> mandatory. And if I set the kdev->fwnode manually (like in the patch
> 2), the kernel hangs inside the component code.
> That's why I proposed to move the components to the place where they
> are really necessary, e.g. i915 and amd drivers.

I'm all for keeping the component framework out of common code. I
dislike that framework with passion, and still haven't lost all hopes of
replacing it with something better.

-- 
Regards,

Laurent Pinchart