Re: [PATCH v5 1/2] x86/PCI: Ignore E820 reservations for bridge windows on newer systems

[Hans de Goede]

Hi Bjorn,

On 11/10/21 09:45, Hans de Goede wrote:
> Hi Bjorn,
> 
> On 11/9/21 23:07, Bjorn Helgaas wrote:
>> On Sat, Nov 06, 2021 at 11:15:07AM +0100, Hans de Goede wrote:
>>> On 10/20/21 23:14, Bjorn Helgaas wrote:
>>>> On Wed, Oct 20, 2021 at 12:23:26PM +0200, Hans de Goede wrote:
>>>>> On 10/19/21 23:52, Bjorn Helgaas wrote:
>>>>>> On Thu, Oct 14, 2021 at 08:39:42PM +0200, Hans de Goede wrote:
>>>>>>> Some BIOS-es contain a bug where they add addresses which map to system
>>>>>>> RAM in the PCI host bridge window returned by the ACPI _CRS method, see
>>>>>>> commit 4dc2287c1805 ("x86: avoid E820 regions when allocating address
>>>>>>> space").
>>>>>>>
>>>>>>> To work around this bug Linux excludes E820 reserved addresses when
>>>>>>> allocating addresses from the PCI host bridge window since 2010.
>>>>>>> ...
>>>>
>>>>>> I haven't seen anybody else eager to merge this, so I guess I'll stick
>>>>>> my neck out here.
>>>>>>
>>>>>> I applied this to my for-linus branch for v5.15.
>>>>>
>>>>> Thank you, and sorry about the build-errors which the lkp
>>>>> kernel-test-robot found.
>>>>>
>>>>> I've just send out a patch which fixes these build-errors
>>>>> (verified with both .config-s from the lkp reports).
>>>>> Feel free to squash this into the original patch (or keep
>>>>> them separate, whatever works for you).
>>>>
>>>> Thanks, I squashed the fix in.
>>>>
>>>> HOWEVER, I think it would be fairly risky to push this into v5.15.
>>>> We would be relying on the assumption that current machines have all
>>>> fixed the BIOS defect that 4dc2287c1805 addressed, and we have little
>>>> evidence for that.
>>>>
>>>> I'm not sure there's significant benefit to having this in v5.15.
>>>> Yes, the mainline v5.15 kernel would work on the affected machines,
>>>> but I suspect most people with those machines are running distro
>>>> kernels, not mainline kernels.
>>>
>>> I understand that you were reluctant to add this to 5.15 so close
>>> near the end of the 5.15 cycle, but can we please get this into
>>> 5.16 now ?
>>>
>>> I know you ultimately want to see if there is a better fix,
>>> but this is hitting a *lot* of users right now and if we come up
>>> with a better fix we can always use that to replace this one
>>> later.
>>
>> I don't know whether there's a "better" fix, but I do know that if we
>> merge what we have right now, nobody will be looking for a better
>> one.
>>
>> We're in the middle of the merge window, so the v5.16 development
>> cycle is over.  The v5.17 cycle is just starting, so we have time to
>> hit that.  Obviously a fix can be backported to older kernels as
>> needed.
>>
>>> So can we please just go with this fix now, so that we can
>>> fix the issues a lot of users are seeing caused by the current
>>> *wrong* behavior of taking the e820 reservations into account ?
>>
>> I think the fix on the table is "ignore E820 for BIOS date >= 2018"
>> plus the obvious parameters to force it both ways.
> 
> Correct.
> 
>> The thing I don't like is that this isn't connected at all to the
>> actual BIOS defect.  We have no indication that current BIOSes have
>> fixed the defect,
> 
> We also have no indication that that defect from 10 years ago, from
> pre UEFI firmware is still present in modern day UEFI firmware which
> is basically an entire different code-base.
> 
> And even 10 years ago the problem was only happening to a single
> family of laptop models (Dell Precision laptops) so this clearly
> was a bug in that specific implementation and not some generic
> issue which is likely to be carried forward.
> 
>> and we have no assurance that future ones will not
>> have the defect.  It would be better if we had some algorithmic way of
>> figuring out what to do.
> 
> You yourself have said that in hindsight taking E820 reservations
> into account for PCI bridge host windows was a mistake. So what
> the "ignore E820 for BIOS date >= 2018" is doing is letting the
> past be the past (without regressing on older models) while fixing
> that mistake on any hardware going forward.
> 
> In the unlikely case that we hit that BIOS bug again on 1 or 2 models,
> we can simply DMI quirk those models, as we do for countless other
> BIOS issues.
> 
>> Thank you very much for chasing down the dmesg log archive
>> (https://github.com/linuxhw/Dmesg; see
>> https://lore.kernel.org/r/82035130-d810-9f0b-259e-61280de1d81f@xxxxxxxxxx).
>> Unfortunately I haven't had time to look through it myself, and I
>> haven't heard of anybody else doing it either.
> 
> Right, I'm afraid that I already have spend way too much time on this
> myself. Note that I've been working with users on this bug on and off
> for over a year now.
> 
> This is hitting many users and now that we have a viable fix, this
> really needs to be fixed now.
> 
> I believe that the "ignore E820 for BIOS date >= 2018" fix is good
> enough and that you are letting perfect be the enemy of good here.
> 
> As an upstream kernel maintainer myself, I'm sorry to say this,
> but if we don't get some fix for this merged soon you are leaving
> my no choice but to add my fix to the Fedora kernels as a downstream
> patch (and to advise other distros to do the same).
> 
> Note that if you are still afraid of regressions going the downstream
> route is also an opportunity, Fedora will start testing moving users
> to 5.15.y soon, so I could add the patch to Fedora's 5.15.y builds and
> see how that goes ?

So I've discussed this with the Fedora kernel maintainers and they have
agreed to add the patch to the Fedora 5.15 kernels, which we will ask
our users to start testing soon (we first run some voluntary testing
before eventually moving all users over).

This will provide us with valuable feedback wrt this patch causing
regressions as you are worried about, or not.

Assuming no regressions show up I hope that this will give you
some assurance that there the patch causes no regressions and that
you will then be willing to pick this up later during the 5.16
cycle so that Fedora only deviates from upstream for 1 cycle.

Regards,

Hans