Regression: drm: Lobotomize set_busid nonsense for !pci drivers (a325725633c2)

[Laszlo Ersek]

On 10/03/16 13:34, Emil Velikov wrote:
> On 30 September 2016 at 18:26, Laszlo Ersek <lersek at redhat.com> wrote:

>> This setup (modulo the kernel of course) was known to work, but now the
>> X server actually segfaults (apparently in the
>> xf86PlatformDeviceCheckBusID() function). Please find the logfile attached.
>>
>> (NB: this is unrelated to upstream commit de9ce6757c2e -- which the
>> pristine FC24 build lacks -- because I don't set AutoAddGPU to "off" --
>> it is left at its default "on" value.)
>>
> Where is this upstream commit again - it shows as unknown for the
> kernel, xserver and libdrm ?

Apologies, that commit hash belongs to a cherry-picked patch on one of
my private branches; the original is

  ea91db4b8331 config: fix GPUDevice fail when AutoAddGPU off + BusID


> So my theory was a bit off - SetVersion is the one responsible to set
> the "BusID", as retrieved by drmGetBusID, regardless if drmOpen or
> open is used.
> 
> Here's a bit of a brain dump from the other day:
> 
>  - The commit mentioned

(Here I assume you mean kernel commit

  a325725633c2 drm: Lobotomize set_busid nonsense for !pci drivers

again.)

>'affects' the drmSetBusid/DRM_IOCTL_SET_UNIQUE
> userspace codepaths.
>  - The latter itself is dri1/legacy (xserver hw/xfree86/dri/) which is
> not functional for platform devices.
> The latter of which seems to be the case for virt-gpu based on the
> kernel module.
>  - The modesetting driver should/cannot reach the above xserver codepath
> 
> That said, it seems that (at least some) userspace expects a PCI
> device despite the kernel module 'advertising' itself as platform one
> :-\
> 
> Going through the xserver layers is a bit inspiring I'm wondering if
> we can not get a strace before/after the xserver commit
> ca8d88e50310a0d440a127c22a0a383cc149f408 ? It will help us track
> things a lot quicker/easier.

This commit:

  ca8d88e50310 xfree86: recognize primary BUS_PCI device in
               xf86IsPrimaryPlatform()

is not relevant when running Xorg on QEMU's "virtio-vga" device, it is
only relevant when the "virtio-gpu-pci" device is used. Aarch64 guests
can't have "virtio-vga", only "virgio-gpu-pci", so that commit was
needed for Aarch64 guests primarily.

In this case however, kernel commit a325725633c2 breaks "virtio-vga" in
x86_64 guests, both with and without xserver commit ca8d88e50310.

Anyway I'll look into capturing a strace and/or following Xorg with gdb.

Thanks!
Laszlo