[Intel-gfx] [PATCH] drm: Add fake controlD* symlinks for backwards compat

[David Herrmann]

Hi Greg

On Fri, Dec 9, 2016 at 11:53 AM, Greg Kroah-Hartman
<gregkh at linuxfoundation.org> wrote:
> On Fri, Dec 09, 2016 at 11:42:01AM +0100, Daniel Vetter wrote:
>> We thought that no userspace is using them, but oops libdrm is using
>> them to figure out whether a driver supports modesetting. Check out
>> drmCheckModesettingSupported but maybe don't because it's horrible and
>> totally runs counter to where we want to go with libdrm device
>> handling. The function looks in the device hierarchy for whether
>> controlD* exist using the following format string:
>>
>> /sys/bus/pci/devices/%04x:%02x:%02x.%d/drm/controlD%d
>>
>> The "/drm" subdirectory is the glue directory from the sysfs class
>> stuff, and the only way to get at it seems to through
>> kdev->kobj.parent (when kdev is represents e.g. the card0 chardev
>> instance in sysfs). Git grep says we're not the only ones touching
>> that, so I hope it's ok we dig into such internals - I couldn't find a
>> proper interface for getting at the glue directory.
>>
>> Quick git grep shows that at least -amdgpu, -ati and UXA in -intel are
>> using this. -modesetting and SNA in -intel do not, which is why this
>> didn't blow up earlier.
>>
>> Oh well, we need to keep it working, and the simplest way is to add a
>> symlink at the right place in debugfs from controlD* to card*.
>
> In debugfs?  This patch seems to be for sysfs.

Yes, typo. It is meant to be sysfs.

>> Fixes: 8a357d10043c ("drm: Nerf DRM_CONTROL nodes")
>> Cc: Dave Airlie <airlied at gmail.com>
>> Reported-by: Alex Deucher <alexander.deucher at amd.com>
>> Cc: Alex Deucher <alexander.deucher at amd.com>
>> Cc: David Herrmann <dh.herrmann at gmail.com>
>> Cc: Greg Kroah-Hartman <gregkh at linuxfoundation.org>
>> Signed-off-by: Daniel Vetter <daniel.vetter at intel.com>
>> ---
>>  drivers/gpu/drm/drm_drv.c | 47 +++++++++++++++++++++++++++++++++++++++++++++++
>>  1 file changed, 47 insertions(+)
>>
>> diff --git a/drivers/gpu/drm/drm_drv.c b/drivers/gpu/drm/drm_drv.c
>> index 4ec61ac27477..5baec7202558 100644
>> --- a/drivers/gpu/drm/drm_drv.c
>> +++ b/drivers/gpu/drm/drm_drv.c
>> @@ -650,6 +650,47 @@ void drm_dev_unref(struct drm_device *dev)
>>  }
>>  EXPORT_SYMBOL(drm_dev_unref);
>>
>> +static int create_compat_control_link(struct drm_device *dev)
>> +{
>> +     struct drm_minor *minor;
>> +     char *name;
>> +     int ret;
>> +
>> +     if (!drm_core_check_feature(dev, DRIVER_MODESET))
>> +             return 0;
>> +
>> +     minor = *drm_minor_get_slot(dev, DRM_MINOR_PRIMARY);
>> +     name = kasprintf(GFP_KERNEL, "controlD%d", minor->index);
>> +     if (!name)
>> +             return -ENOMEM;
>> +
>> +     ret = sysfs_create_link(minor->kdev->kobj.parent,
>> +                             &minor->kdev->kobj,
>> +                             name);
>
> Ick.  Dropping down to a sysfs call and a kobject isn't nice in driver
> code, but I guess the driver core doesn't provide symlinks.  We do
> provide the "class_compat" functionality, but that doesn't really apply
> here.
>
> So, what happened, you just added a "middle layer" device object for
> these control files?
>
> If you need these compatibility symlinks, why even do the original code
> at all?  That kind of implies it shouldn't have been made one layer
> deeper if it was going to break userspace somehow.
>
> If you add these, what's the odds that they ever can be removed
> (probably never, right?)

Yes. The alternative would be to create a dummy "struct device" and
register it, but do not include any information on it. I don't think
we need the symlink-behavior. Daniel? But the symlink would at least
be kinda useful. The dummy device, on the other hand, would just make
sure the readdir() calls of legacy stuff sees the control nodes (even
though they never use it).

Thanks
David