Re: [PATCH] LoongArch: Make WriteCombine configurable for ioremap()

[Huacai Chen]

On Tue, Mar 14, 2023 at 6:12 PM WANG Xuerui <kernel@xxxxxxxxxx> wrote:
>
> On 2023/3/14 18:02, Huacai Chen wrote:
> > Hi, Ruoyao,
> >
> > On Tue, Mar 14, 2023 at 5:41 PM Xi Ruoyao <xry111@xxxxxxxxxxx> wrote:
> >>
> >> On Tue, 2023-03-14 at 16:54 +0800, Huacai Chen wrote:
> >>> LoongArch maintains cache coherency in hardware, but when works with
> >>> LS7A chipsets the WUC attribute (Weak-ordered UnCached, which is similar
> >>> to WriteCombine) is out of the scope of cache coherency machanism for
> >>> PCIe devices (this is a PCIe protocol violation, may be fixed in newer
> >>> chipsets).
> >>
> >> IIUC all launched LS7A models (7A1000 and 7A2000) suffers this issue?
> > Yes, very unfortunately, but this issue is only observed in the amdgpu
> > driver now.
>
> It's PCIe protocol violation after all, and we can never be sure about
> the vast amount of hardware untested on loongarch after all. Miserable
> as the performance hit may get, we don't really have another choice,
> unfortunately. Someone needs to lecture the LS7A team real hard!
>
> >>
> >>> This means WUC can only used for write-only memory regions now, so this
> >>> option is disabled by default (means WUC falls back to SUC for ioremap).
> >>> You can enable this option if the kernel is ensured to run on bug-free
> >>> hardwares.
> >>
> >> Hmm, is it possible to make a PCI quirk so SUC/WUC will be decided
> >> automatically from the vendor:device ID of the PCI root controller?
> >> Then we don't need to rely on the user or distro maintainer to select an
> >> option.  I see there is already many architecture-dependant #if
> >> directives in drivers/pci/quirks.c so I guess such a quirk is acceptable
> >> in PCI tree...
> > Not a good idea, pci quirks need too long a time to review, and we
> > don't know when this issue can be fixed in hardware.
> >
> >>
> >> If a PCI quirk is not possible, then is it possible to make a kernel
> >> command line option, leaving this CONFIG as the default value of the
> >> option?  I guess in the future many LoongArch users will just install a
> >> binary distro, then it would be much easier to edit grub.cfg than
> >> rebuilding the kernel when they finally buy a compliant PCIe controller.
> > If we use command line parameter, we can remove this Kconfig option.
>
> An option is still useful as specifying the compile-time default for
> such a kernel parameter, IMO.
I will update commit messages and add a kernel parameter, thanks.

Huacai
>
> --
> WANG "xen0n" Xuerui
>
> Linux/LoongArch mailing list: https://lore.kernel.org/loongarch/
>
>