Re: [PATCH 00/10] rm/bochs: Modernize driver

[Thomas Zimmermann <tzimmermann@suse.de>]

Hi

Am 23.08.24 um 16:34 schrieb Gerd Hoffmann:
>    Hi,
>
>> Patch 8 replaces GEM VRAM with GEM SHMEM. The new memory manager
>> is more reliable and allows for larger resolutions.
> Valid point.
>
>> Display updates were so slow that Gnome was unmanageable with a
>> flickering cursor and single FPS. The new memory management makes
>> Gnome at least useable.
> Hmm?  I'm wondering where this huge improvement comes from?
>
> With enough video memory VRAM performance should be ok.
>
> If video memory is tight and ttm is forced to shuffle around
> framebuffers between vram and system memory on each page flip (touching
> much of vram along the way which causes additional overhead on the qemu
> side), that is obviously very bad for performance.  One of the reasons
> why cirrus uses SHMEM + shadowing since years.
>
> Shadow buffering comes with some overhead too, so the switch isn't an
> obvious win (assuming enough vram).  Hiding the page flips from qemu
> might reduce the work qemu has to do though, especially if the shadowing
> uses dirty tracking and only touches the vram pages which have actually
> changed content.  So there is a fair chance that this outweighs the
> shadowing overhead and ends up being a net win.  I don't expect the
> difference being very big though.  Also different display usage patterns
> might yield different results (fbcon vs. gnome for example).
>
> So this probably makes sense, but I'd like to see a bit more background
> information ...

The difference is in damage handling.

The old code had two BOs in video memory and flipped between them. IDK 
the details of the old rendering, but from the massive flickering of the 
cursor, I assume that X11's internal either copies a full buffer during 
each redraw, or doesn't really handle damage well. It could also happen 
that X didn't use a shadow buffer for rendering. Bochs didn't request 
one. Without, drawing to I/O memory is really slow. If that applies to 
virtual I/O memory as well IDK.

The new driver code only copies areas that have been changed from 
rendering. The flickering is gone and the overall update performance is 
acceptable.

>
> On vram sizes:  The default qemu vram size (16M) should be fine for the
> default display resolution (1280x800).  For FullHD vram size should be
> doubled (-device VGA,vgamem_mb=32).

Right. Bochs never really tested that. So I saw something like 5k by 3k 
resolutions on my test setup with 16 MiB. Now that video-memory 
requirements for each mode can be calculated easily, we can sort out the 
invalid modes.

>
>
> Skimmed the other patches, looks sane overall, but I don't follow drm
> close enough any more to do an full review.  So I leave this here:
>
> Acked-by: Gerd Hoffmann <kraxel@redhat.com>

Thank you so much.

Best regards
Thomas

>
> take care,
>    Gerd
>

-- 
--
Thomas Zimmermann
Graphics Driver Developer
SUSE Software Solutions Germany GmbH
Frankenstrasse 146, 90461 Nuernberg, Germany
GF: Ivo Totev, Andrew Myers, Andrew McDonald, Boudien Moerman
HRB 36809 (AG Nuernberg)