Re: [PATCH] Revert "drm/radeon: handle PCIe root ports with addressing limitations"

[Christoph Hellwig <hch@lst.de>]

On Wed, Sep 16, 2020 at 06:16:25PM -0400, Alex Deucher wrote:
> On Wed, Sep 16, 2020 at 3:04 AM Christoph Hellwig <hch@lst.de> wrote:
> >
> > On Tue, Sep 15, 2020 at 02:46:07PM -0400, Alex Deucher wrote:
> > > This change breaks tons of systems.
> >
> > Did you do at least some basic root causing on why?  Do GPUs get
> > fed address they can't deal with?  Any examples?
> >
> > Bug 1 doesn't seem to contain any analysis and was reported against
> > a very old kernel that had all kind of fixes since.
> >
> > Bug 2 seems to imply a drm kthread is accessing some structure it
> > shouldn't, which would imply a mismatch between pools used by radeon
> > now and those actually provided by the core.  Something that should
> > be pretty to trivial to fix for someone understanding the whole ttm
> > pool maze.
> >
> > Bug 3: same as 1, but an even older kernel.
> >
> > Bug 4: looks like 1 and 3, and actually verified to work properly
> > in 5.9-rc.  Did you try to get the other reporters test this as well?
> 
> It would appear that the change in 5.9 to disable AGP on radeon fixed
> the issue.  I'm following up on the other tickets to see if I can get
> confirmation.  On another thread[1], the user was able to avoid the
> issue by disabling HIMEM.  Looks like some issue with HIMEM and/or
> AGP.

Thanks.  I'll try to spend some time to figure out what could be
highmem related.  I'd much rather get this fixed properly.