Re: [PATCH v2] x86/PCI: Prefer MMIO over PIO on all hypervisor

From: Michael S. Tsirkin
Date: Mon Oct 10 2022 - 13:05:49 EST


On Tue, Oct 04, 2022 at 06:48:11PM +0000, Nadav Amit wrote:
> On Oct 4, 2022, at 1:22 AM, Alexander Graf <graf@xxxxxxxxxx> wrote:
>
> > ⚠ External Email
> >
> > Hey Nadav,
> >
> > On 03.10.22 19:34, Nadav Amit wrote:
> >> On Oct 3, 2022, at 8:03 AM, Vitaly Kuznetsov <vkuznets@xxxxxxxxxx> wrote:
> >>
> >>> Not my but rather PCI maintainer's call but IMHO dropping 'const' is
> >>> better, introducing a new global var is our 'last resort' and should be
> >>> avoided whenever possible. Alternatively, you can add a
> >>> raw_pci_ext_ops_preferred() function checking somethin within 'struct
> >>> hypervisor_x86' but I'm unsure if it's better.
> >>>
> >>> Also, please check Alex' question/suggestion.
> >> Here is my take (and Ajay knows probably more than me):
> >>
> >> Looking briefly on MCFG, I do not see a clean way of using the ACPI table.
> >> The two options are either to use a reserved field (which who knows, might
> >> be used one day) or some OEM ID. I am also not familiar with
> >> PCI_COMMAND.MEMORY=0, so Ajay can hopefully give some answer about that.
> >>
> >> Anyhow, I understand (although not relate) to the objection for a new global
> >> variable. How about explicitly calling this hardware bug a “bug” and using
> >> the proper infrastructure? Calling it explicitly a bug may even push whoever
> >> can to resolve it.
> >
> >
> > I am a lot more concerned with how we propagate it externally than
> > within Linux. If we hard code that all Linux kernels 6.2+ that are
> > running in VMware prefer ECAM over PIO, we lock ourselves into that
> > stance for better or worse, which means:
> >
> > * All past and future versions of any VMware hypervisor product have to
> > always allow ECAM access for any PCIe config space write
> > * No other hypervisor benefits from any of this without upstream code change
> > * No real hardware platform benefits from this without upstream code change
> >
> > By moving it into MCFG, we can create a path for the outside environment
> > to tell the OS whether it's safe to use ECAM always. This obviously
> > doesn't work with MCFG as it stands today, we'd have to propose an MCFG
> > spec change to the PCI SIG's "PCI Firmware Specification" to add the
> > respective field. Future VMware versions could then always expose the
> > flag - and if you find it broken, remove it again.
> >
> > Putting all of the logic on which system potentially prefers ECAM over
> > PIO config space access into Linux is just a big hack that we should
> > avoid as much as possible.
>
> Thanks Alex. You raise important points. Let me try to break down your
> concerns slightly differently:
>
> 1. Enabling MMIO access should be selective, and potentially controlled by
> the hypervisor. The very least a "chicken-bit” is needed.
>
> 2. PCI SIG would change its specifications to address unclear hardware bug.
>
> I think (1) makes sense and we can discuss different ways of addressing it.
> But (2) would not happen in a reasonable timeline and seems to me as an
> unnecessary complication.
>
> But before we discuss how to address the issue, perhaps we need to first
> understand it better. I am not sure that I understand this MMIO bug, and so
> far nobody was able to provide exact details.
>
> So I went to have a look. It might not be super helpful, but for the record,
> here is what I collected.
>
> First, we have commit d6ece5491ae71d ("i386/x86-64 Correct for broken MCFG
> tables on K8 systems”). It tried to "try to discover all devices on bus 0
> that are unreachable using MM and fallback for them.” Interestingly, it
> seems similar to FreeBSD code (commit 2d10570afe2b3e) that also mentions K8
> and has similar detection logic in FreeBSD’s pcie_cfgregopen().
>
> Then commit a0ca9909609470 ("PCI x86: always use conf1 to access config
> space below 256 bytes”). The correspondence [1] mentions some bugs: ATI
> chipset, VIA chipset, Intel 3 Series Express chipset family and some reports
> on Nvidia. It turned out some devices had problem probing - to figure out if
> MMIO is broken - the way the previous patch did.

There's also a statement by Linus that MCFG might not cover all buses
in that thread. I didn't think the implications through yet ...

> All of these bugs are circa 2008, of course. And note that FreeBSD did not
> take a similar path. The correspondence around Linux patch is endless. I
> admit that I did not understand whether eventually the issues were found to
> be per-bus or per-device.
>
>
> Back to the matter at hand. The benefit of using the MCFG approach that you
> propose is that it can enable native systems to use MMIO as well. However,
> since the list of bugs is unclear and the problems might be device-specific,
> it is not clear what information BIOSes have that Linux doesn’t. In other
> words, the benefit of getting it into the specifications is questionable,
> and the complexity+time is high.
>
> Can we agree that the feature would be enabled explicitly by the hypervisor
> and Linux would enable it based on the hypervisor input (through some
> channel?)
>
> Thanks,
> Nadav
>
> [1] https://lore.kernel.org/all/20080112144030.GA19279@xxxxxxxxxxxxxxxxxxxx/T/#u