Re: [PATCH v2 0/3] x86/PCI: Log E820 clipping

[Bjorn Helgaas <helgaas@kernel.org>]

On Tue, Apr 19, 2022 at 05:16:44PM +0200, Hans de Goede wrote:
> Hi,
> 
> On 4/19/22 17:03, Bjorn Helgaas wrote:
> > On Tue, Apr 19, 2022 at 11:59:17AM +0200, Hans de Goede wrote:
> >> On 1/1/70 01:00, Bjorn Helgaas wrote:
> >>> This is still work-in-progress on the issue of PNP0A03 _CRS methods that
> >>> are buggy or not interpreted correctly by Linux.
> >>>
> >>> The previous try at:
> >>>   https://lore.kernel.org/r/20220304035110.988712-1-helgaas@kernel.org
> >>> caused regressions on some Chromebooks:
> >>>   https://lore.kernel.org/r/Yjyv03JsetIsTJxN@sirena.org.uk
> >>>
> >>> This v2 drops the commit that caused the Chromebook regression, so it also
> >>> doesn't fix the issue we were *trying* to fix on Lenovo Yoga and Clevo
> >>> Barebones.
> >>>
> >>> The point of this v2 update is to split the logging patch into (1) a pure
> >>> logging addition and (2) the change to only clip PCI windows, which was
> >>> previously hidden inside the logging patch and not well documented.
> >>>
> >>> Bjorn Helgaas (3):
> >>>   x86/PCI: Eliminate remove_e820_regions() common subexpressions
> >>>   x86: Log resource clipping for E820 regions
> >>>   x86/PCI: Clip only host bridge windows for E820 regions
> >>
> >> Thanks, the entire series looks good to me:
> >>
> >> Reviewed-by: Hans de Goede <hdegoede@redhat.com>
> > 
> > Thank you!
> > 
> >> So what is the plan to actually fix the issue seen on some Lenovo models
> >> and Clevo Barebones ?   As I mentioned previously I think that since all
> >> our efforts have failed so far that we should maybe reconsider just
> >> using DMI quirks to ignore the E820 reservation windows for host bridges
> >> on affected models ?
> > 
> > I have been resisting DMI quirks but I'm afraid there's no other way.
> 
> Well there is the first match adjacent windows returned by _CRS and
> only then do the "covers whole region" exception check. I still
> think that would work at least for the chromebook regression...

Without a crystal clear strategy, I think we're going to be tweaking
the algorithm forever as the _CRS/E820 mix changes.  That's why I
think that in the long term, a "use _CRS only, with quirks for
exceptions" strategy will be simplest.

> So do you want me to give that a try; or shall I write a patch
> using DMI quirks. And if we go the DMI quirks, what about
> matching cmdline arguments?  If we add matching cmdline arguments,
> which seems to be the sensible thing to do then to allow users
> to test if they need the quirk, then we basically end up with my
> first attempt at fixing this from 6 months ago:
> 
> https://lore.kernel.org/linux-pci/20211005150956.303707-1-hdegoede@redhat.com/

So I think we should go ahead with DMI quirks instead of trying to
make the algorithm smarter, and yes, I think we will need commandline
arguments, probably one to force E820 clipping for future machines,
and one to disable it for old machines.

> > I think the web we've gotten into, where vendors have used E820 to
> > interact with _CRS in incompatible and undocumented ways, is not
> > sustainable.
> > 
> > I'm not aware of any spec that says the OS should use E820 to clip
> > things out of _CRS, so I think the long term plan should be to
> > decouple them by default.
> 
> Right and AFAICT the reason Windows is getting away with this is
> the same as with the original Dell _CRS has overlap with
> physical RAM issue (1), Linux assigns address to unassigneds BAR-s
> starting with the lowest available address in the bridge window,
> where as Windows assigns addresses from the highest available
> address in the window.

Right, I agree.  I'm guessing Chromebooks don't get tested with
Windows at all, so we don't even have that level of testing to help.

> So the real fix here might very well be
> to rework the BAR assignment code to switch to fill the window
> from the top rather then from the bottom. AFAICT all issues where
> excluding _E820 reservations have helped are with _E820 - bridge
> window overlaps at the bottom of the window.
> 
> IOW these are really all bugs in the _CRS method for the bridge,
> which Windows does not hit because it never actually uses
> the lowest address(es) of the _CRS returned window.

Yes.  We actually did try this
(https://git.kernel.org/linus/1af3c2e45e7a), but unfortunately we had
to revert it.  Even more unfortunately, the revert
(https://git.kernel.org/linus/5e52f1c5e85f) doesn't have any details
about what went wrong.

> 1) At least I read in either a bugzilla, or email thread about
> this that Windows allocating bridge window space from the top
> was assumed to be why Windows was not impacted.
> 
> > Straw man:
> > 
> >   - Disable E820 clipping by default.
> > 
> >   - Add a quirk to enable E820 clipping for machines older than X,
> >     e.g., 2023, to avoid breaking machines that currently work.
> > 
> >   - Add quirks to disable E820 clipping for individual machines like
> >     the Lenovo and Clevos that predate X, but E820 clipping breaks
> >     them.
> > 
> >   - Add quirks to enable E820 clipping for individual machines like
> >     the Chromebooks (and probably machines we don't know about yet)
> >     that have devices that consume part of _CRS but are not
> >     enumerable.
> > 
> >   - Communicate this to OEMs to try to prevent future machines that
> >     need quirks.
> > 
> > Bjorn
> > 
>