2019-09-05 19:59:52

by Rafael J. Wysocki

[permalink] [raw]
Subject: Re: [Nouveau] [PATCH 1/7] Revert "ACPI / OSI: Add OEM _OSI string to enable dGPU direct output"

On Thursday, September 5, 2019 5:51:23 PM CEST Karol Herbst wrote:
> is there any update on the testing with my patches? On the hardware I
> had access to those patches helped, but I can't know if it also helped
> on the hardware for which those workarounds where actually added.

Alex Hung and Mario need to answer this question I think.

> On Mon, Aug 19, 2019 at 11:52 AM Rafael J. Wysocki <[email protected]> wrote:
> >
> > On Thursday, August 15, 2019 12:47:35 AM CEST Dave Airlie wrote:
> > > On Thu, 15 Aug 2019 at 07:31, Karol Herbst <[email protected]> wrote:
> > > >
> > > > This reverts commit 28586a51eea666d5531bcaef2f68e4abbd87242c.
> > > >
> > > > The original commit message didn't even make sense. AMD _does_ support it and
> > > > it works with Nouveau as well.
> > > >
> > > > Also what was the issue being solved here? No references to any bugs and not
> > > > even explaining any issue at all isn't the way we do things.
> > > >
> > > > And even if it means a muxed design, then the fix is to make it work inside the
> > > > driver, not adding some hacky workaround through ACPI tricks.
> > > >
> > > > And what out of tree drivers do or do not support we don't care one bit anyway.
> > > >
> > >
> > > I think the reverts should be merged via Rafael's tree as the original
> > > patches went in via there, and we should get them in asap.
> > >
> > > Acked-by: Dave Airlie <[email protected]>
> >
> > The _OSI strings are to be dropped when all of the needed support is there in
> > drivers, so they should go away along with the requisite driver changes.
> >
>
> that goes beside the point. firmware level workarounds for GPU driver
> issues were pushed without consulting with upstream GPU developers.
> That's something which shouldn't have happened in the first place. And
> yes, I am personally annoyed by the fact, that people know about
> issues, but instead of contacting the proper persons and working on a
> proper fix, we end up with stupid firmware level workarounds. I can't
> see why we ever would have wanted such workarounds in the first place.
>
> And I would be much happier if the next time something like that comes
> up, that the drm mailing list will be contacted as well or somebody
> involved.
>
> We could have also just disable the feature inside the driver (and
> probably we should have done that a long time ago, so that is
> essentially our fault, but still....)
>
> > I'm all for dropping then when that's the case, so please feel free to add ACKs
> > from me to the patches in question at that point.
> >
> > Cheers,
> > Rafael
> >
> >
> >
>





2019-09-10 09:40:11

by Alex Hung

[permalink] [raw]
Subject: Re: [Nouveau] [PATCH 1/7] Revert "ACPI / OSI: Add OEM _OSI string to enable dGPU direct output"

On Thu, Sep 5, 2019 at 5:26 PM Rafael J. Wysocki <[email protected]> wrote:
>
> On Thursday, September 5, 2019 5:51:23 PM CEST Karol Herbst wrote:
> > is there any update on the testing with my patches? On the hardware I
> > had access to those patches helped, but I can't know if it also helped
> > on the hardware for which those workarounds where actually added.
>
> Alex Hung and Mario need to answer this question I think.

Sorry for taking a long time. I don't have full testing results yet
but we found at least a regression occurred with _OSI string removed -
it is not on nVidia hardware but on AMD PX one.

I will try to collect and share more details.

>
> > On Mon, Aug 19, 2019 at 11:52 AM Rafael J. Wysocki <[email protected]> wrote:
> > >
> > > On Thursday, August 15, 2019 12:47:35 AM CEST Dave Airlie wrote:
> > > > On Thu, 15 Aug 2019 at 07:31, Karol Herbst <[email protected]> wrote:
> > > > >
> > > > > This reverts commit 28586a51eea666d5531bcaef2f68e4abbd87242c.
> > > > >
> > > > > The original commit message didn't even make sense. AMD _does_ support it and
> > > > > it works with Nouveau as well.
> > > > >
> > > > > Also what was the issue being solved here? No references to any bugs and not
> > > > > even explaining any issue at all isn't the way we do things.
> > > > >
> > > > > And even if it means a muxed design, then the fix is to make it work inside the
> > > > > driver, not adding some hacky workaround through ACPI tricks.
> > > > >
> > > > > And what out of tree drivers do or do not support we don't care one bit anyway.
> > > > >
> > > >
> > > > I think the reverts should be merged via Rafael's tree as the original
> > > > patches went in via there, and we should get them in asap.
> > > >
> > > > Acked-by: Dave Airlie <[email protected]>
> > >
> > > The _OSI strings are to be dropped when all of the needed support is there in
> > > drivers, so they should go away along with the requisite driver changes.
> > >
> >
> > that goes beside the point. firmware level workarounds for GPU driver
> > issues were pushed without consulting with upstream GPU developers.
> > That's something which shouldn't have happened in the first place. And
> > yes, I am personally annoyed by the fact, that people know about
> > issues, but instead of contacting the proper persons and working on a
> > proper fix, we end up with stupid firmware level workarounds. I can't
> > see why we ever would have wanted such workarounds in the first place.
> >
> > And I would be much happier if the next time something like that comes
> > up, that the drm mailing list will be contacted as well or somebody
> > involved.
> >
> > We could have also just disable the feature inside the driver (and
> > probably we should have done that a long time ago, so that is
> > essentially our fault, but still....)
> >
> > > I'm all for dropping then when that's the case, so please feel free to add ACKs
> > > from me to the patches in question at that point.
> > >
> > > Cheers,
> > > Rafael
> > >
> > >
> > >
> >
>
>
>
>


--
Cheers,
Alex Hung

2019-10-21 02:16:07

by Alex Hung

[permalink] [raw]
Subject: Re: [Nouveau] [PATCH 1/7] Revert "ACPI / OSI: Add OEM _OSI string to enable dGPU direct output"

We have done some tests on three of Intel + nVidia configuration
systems with OEM _OSI strings removed - while some bugs are still
observed, ex. one out of three has suspend/resume issues, no system
crashes were observed - the biggest issue that worries us.

The positive results give us confident to ack the removal of the OEM
_OSI strings. While our tests were not able to cover all possible I+N
systems, we are sure we can fix issues along the way. If there aren't
systems that cannot be fixed without these OEM _OSI strings, these
strings should probably enable with DMI quirks (possible future
patches) so they won't affect others.

Acked-by: Alex Hung <[email protected]>




On Thu, Sep 5, 2019 at 10:26 AM Rafael J. Wysocki <[email protected]> wrote:
>
> On Thursday, September 5, 2019 5:51:23 PM CEST Karol Herbst wrote:
> > is there any update on the testing with my patches? On the hardware I
> > had access to those patches helped, but I can't know if it also helped
> > on the hardware for which those workarounds where actually added.
>
> Alex Hung and Mario need to answer this question I think.
>
> > On Mon, Aug 19, 2019 at 11:52 AM Rafael J. Wysocki <[email protected]> wrote:
> > >
> > > On Thursday, August 15, 2019 12:47:35 AM CEST Dave Airlie wrote:
> > > > On Thu, 15 Aug 2019 at 07:31, Karol Herbst <[email protected]> wrote:
> > > > >
> > > > > This reverts commit 28586a51eea666d5531bcaef2f68e4abbd87242c.
> > > > >
> > > > > The original commit message didn't even make sense. AMD _does_ support it and
> > > > > it works with Nouveau as well.
> > > > >
> > > > > Also what was the issue being solved here? No references to any bugs and not
> > > > > even explaining any issue at all isn't the way we do things.
> > > > >
> > > > > And even if it means a muxed design, then the fix is to make it work inside the
> > > > > driver, not adding some hacky workaround through ACPI tricks.
> > > > >
> > > > > And what out of tree drivers do or do not support we don't care one bit anyway.
> > > > >
> > > >
> > > > I think the reverts should be merged via Rafael's tree as the original
> > > > patches went in via there, and we should get them in asap.
> > > >
> > > > Acked-by: Dave Airlie <[email protected]>
> > >
> > > The _OSI strings are to be dropped when all of the needed support is there in
> > > drivers, so they should go away along with the requisite driver changes.
> > >
> >
> > that goes beside the point. firmware level workarounds for GPU driver
> > issues were pushed without consulting with upstream GPU developers.
> > That's something which shouldn't have happened in the first place. And
> > yes, I am personally annoyed by the fact, that people know about
> > issues, but instead of contacting the proper persons and working on a
> > proper fix, we end up with stupid firmware level workarounds. I can't
> > see why we ever would have wanted such workarounds in the first place.
> >
> > And I would be much happier if the next time something like that comes
> > up, that the drm mailing list will be contacted as well or somebody
> > involved.
> >
> > We could have also just disable the feature inside the driver (and
> > probably we should have done that a long time ago, so that is
> > essentially our fault, but still....)
> >
> > > I'm all for dropping then when that's the case, so please feel free to add ACKs
> > > from me to the patches in question at that point.
> > >
> > > Cheers,
> > > Rafael
> > >
> > >
> > >
> >
>
>
>
>


--
Cheers,
Alex Hung

2019-10-21 08:16:57

by Rafael J. Wysocki

[permalink] [raw]
Subject: Re: [Nouveau] [PATCH 1/7] Revert "ACPI / OSI: Add OEM _OSI string to enable dGPU direct output"

On Mon, Oct 21, 2019 at 4:14 AM Alex Hung <[email protected]> wrote:
>
> We have done some tests on three of Intel + nVidia configuration
> systems with OEM _OSI strings removed - while some bugs are still
> observed, ex. one out of three has suspend/resume issues, no system
> crashes were observed - the biggest issue that worries us.
>
> The positive results give us confident to ack the removal of the OEM
> _OSI strings. While our tests were not able to cover all possible I+N
> systems, we are sure we can fix issues along the way. If there aren't
> systems that cannot be fixed without these OEM _OSI strings, these
> strings should probably enable with DMI quirks (possible future
> patches) so they won't affect others.
>
> Acked-by: Alex Hung <[email protected]>

OK, thanks!

I can queue this up or if it's better to route it through the DRM
tree, please do that (and let me know).

2019-10-21 08:49:43

by Karol Herbst

[permalink] [raw]
Subject: Re: [Nouveau] [PATCH 1/7] Revert "ACPI / OSI: Add OEM _OSI string to enable dGPU direct output"

fyi: I decided to go for a different workaround to fix the runpm
issues observed with nvidia gpus with nouveau in the "pci: prevent
putting nvidia GPUs into lower device states on certain intel bridges"
thread

that's on the pci and pm mailing list. Maybe it makes sense to wait
for that to land before actually removing the ACPI workarounds here?
The workaround I had in this series didn't seem to be reliable enough,
so I ditched that approached.

On Mon, Oct 21, 2019 at 10:14 AM Rafael J. Wysocki <[email protected]> wrote:
>
> On Mon, Oct 21, 2019 at 4:14 AM Alex Hung <[email protected]> wrote:
> >
> > We have done some tests on three of Intel + nVidia configuration
> > systems with OEM _OSI strings removed - while some bugs are still
> > observed, ex. one out of three has suspend/resume issues, no system
> > crashes were observed - the biggest issue that worries us.
> >
> > The positive results give us confident to ack the removal of the OEM
> > _OSI strings. While our tests were not able to cover all possible I+N
> > systems, we are sure we can fix issues along the way. If there aren't
> > systems that cannot be fixed without these OEM _OSI strings, these
> > strings should probably enable with DMI quirks (possible future
> > patches) so they won't affect others.
> >
> > Acked-by: Alex Hung <[email protected]>
>
> OK, thanks!
>
> I can queue this up or if it's better to route it through the DRM
> tree, please do that (and let me know).

2019-10-21 08:55:38

by Rafael J. Wysocki

[permalink] [raw]
Subject: Re: [Nouveau] [PATCH 1/7] Revert "ACPI / OSI: Add OEM _OSI string to enable dGPU direct output"

On Mon, Oct 21, 2019 at 10:48 AM Karol Herbst <[email protected]> wrote:
>
> fyi: I decided to go for a different workaround to fix the runpm
> issues observed with nvidia gpus with nouveau in the "pci: prevent
> putting nvidia GPUs into lower device states on certain intel bridges"
> thread

OK, I've seen that.

> that's on the pci and pm mailing list. Maybe it makes sense to wait
> for that to land before actually removing the ACPI workarounds here?

Sounds reasonable.

> The workaround I had in this series didn't seem to be reliable enough,
> so I ditched that approached.

OK, please let me know when the _OSI string in question can be dropped.

> On Mon, Oct 21, 2019 at 10:14 AM Rafael J. Wysocki <[email protected]> wrote:
> >
> > On Mon, Oct 21, 2019 at 4:14 AM Alex Hung <[email protected]> wrote:
> > >
> > > We have done some tests on three of Intel + nVidia configuration
> > > systems with OEM _OSI strings removed - while some bugs are still
> > > observed, ex. one out of three has suspend/resume issues, no system
> > > crashes were observed - the biggest issue that worries us.
> > >
> > > The positive results give us confident to ack the removal of the OEM
> > > _OSI strings. While our tests were not able to cover all possible I+N
> > > systems, we are sure we can fix issues along the way. If there aren't
> > > systems that cannot be fixed without these OEM _OSI strings, these
> > > strings should probably enable with DMI quirks (possible future
> > > patches) so they won't affect others.
> > >
> > > Acked-by: Alex Hung <[email protected]>
> >
> > OK, thanks!
> >
> > I can queue this up or if it's better to route it through the DRM
> > tree, please do that (and let me know).