Hello Thomas,
On 5/5/22 10:05, Thomas Zimmermann wrote:
[snip]
>>
>> In other words, in most cases (i.e: only fbcon bound to the fbdev)
>> the driver's removal/ device unbind and the memory release will be
>> at the same time.
>>
>
> We're one the same page here, but it's still sort of a mystery to me why
> this works in practice.
>
> I'm specifically talking about pci_request_regions() in vmwgfx [1]. IIRC
> this would fail if simplefb still owns the framebuffer region. Lots of
> systems run Plymouth during boot and this should result in failures
> occasionally. Still, we never heard about anything.
>
Yes, I think is because Plymouth IIUC waits for a /dev/dri/card? to be
present and only uses a /dev/fb? as a fallback if a timeout expires.
At least in Fedora (even before the efifb -> simpledrm change) it will
use KMS/DRM since the DRM kernel module for the graphics device in the
machine would be in the intird.
So efifb was only used for fbcon and plymouth would only use DRM/KMS
and not its fbdev backend.
This seems to be sort of a corner case when you have {efi,simple}fb
in the early boot but the real DRM module only in the rootfs after the
initrd has done a pivot_root(2).
> Of course, it's always been broken (even long before real fbdev
> hotunplugging). Switching to simpledrm resolves the problem.
>
Indeed. My opinion after dealing with these fbdev problems is that we
shouldn't try to fix all possible corner cases and just try to get rid
of fbdev as soon as possible.
--
Best regards,
Javier Martinez Canillas
Linux Engineering
Red Hat
