Re: [PATCH 3/7] drm: Update drm_addmap and drm_mmap to use PAT WC instead of MTRRs

[Alex Deucher <alexdeucher@gmail.com>]

On Mon, May 6, 2013 at 7:39 PM, Andy Lutomirski <luto@amacapital.net> wrote:
> On Mon, May 6, 2013 at 4:04 PM, Jerome Glisse <j.glisse@gmail.com> wrote:
>> On Mon, May 6, 2013 at 5:22 PM, Andy Lutomirski <luto@amacapital.net> wrote:
>>> On Fri, May 3, 2013 at 4:00 PM, Andy Lutomirski <luto@amacapital.net> wrote:
>>>> Signed-off-by: Andy Lutomirski <luto@amacapital.net>
>>>> ---
>>>>
>>>> This needs careful review.  I don't really know what this code does, nor
>>>> do I have the hardware.  (I don't understand AGP and the associated
>>>> caching implications.)
>>>
>>> This patch is wrong (I didn't update the matching mtrr_del), and I'm
>>> reworking this whole series.  But I may need some help on this one:
>>> why is the mtrr handle of a map (whatever a map is) exported to
>>> userspace via the ADD_MAP and GET_MAP ioctls?  What (if anything) is
>>> userspace supposed to do with it?  Do I need to return a valid MTRR
>>> register number?  Is there any userspace code at all that sets
>>> _DRM_WRITE_COMBINING in DRM_IOCTL_ADD_MAP with appropriate alignment
>>> and needs the MTRR, for which the drm driver doesn't already add the
>>> MTRR?
>>>
>>> --Andy
>>
>> From memory, even on pat system we need mtrr for VRAM is PCI BAR. We
>> cover it with a write combine MTRR. The whole ioctl is use by some ddx
>> or maybe even directly the XServer to do this mtrr mess in userspace.
>
> Egads!  So we have a _DRM_WRITE_COMBINING flag, which will continue to
> work fine, but almost nothing uses it.
>
> I'm amazed this stuff works in the current code, though.  Apparently
> the memory type (WC or UC) of a drm mapping is determined by the mtrr
> the driver set, but if one part of the BAR is textures or the
> framebuffer and another part is an outgoing command ring, won't one of
> them end up with the wrong memory type?

A lot of old chips used to put the registers and framebuffer in the
same BAR.  IIRC, the xserver and later libpciaccess had workarounds to
deal with this.

Alex