Re: [PATCH] x86/cpu: Add a VMX flag to enumerate 5-level EPT support to userspace

[Tao Su <tao1.su@linux.intel.com>]

On Wed, Jan 10, 2024 at 08:26:25AM -0800, Sean Christopherson wrote:
> On Wed, Jan 10, 2024, Chao Gao wrote:
> > On Tue, Jan 09, 2024 at 04:23:40PM -0800, Sean Christopherson wrote:
> > >Add a VMX flag in /proc/cpuinfo, ept_5level, so that userspace can query
> > >whether or not the CPU supports 5-level EPT paging.  EPT capabilities are
> > >enumerated via MSR, i.e. aren't accessible to userspace without help from
> > >the kernel, and knowing whether or not 5-level EPT is supported is sadly
> > >necessary for userspace to correctly configure KVM VMs.
> > 
> > This assumes procfs is enabled in Kconfig and userspace has permission to
> > access /proc/cpuinfo. But it isn't always true. So, I think it is better to
> > advertise max addressable GPA via KVM ioctls.
> 
> Hrm, so the help for PROC_FS says:
> 
>   Several programs depend on this, so everyone should say Y here.
> 
> Given that this is working around something that is borderline an erratum, I'm
> inclined to say that userspace shouldn't simply assume the worst if /proc isn't
> available.  Practically speaking, I don't think a "real" VM is likely to be
> affected; AFAIK, there's no reason for QEMU or any other VMM to _need_ to expose
> a memslot at GPA[51:48] unless the VM really has however much memory that is
> (hundreds of terabytes?).  And a if someone is trying to run such a massive VM on
> such a goofy CPU...

It is unusual to assign a huge RAM to guest, but passthrough a device also may trigger
this issue which we have met, i.e. alloc memslot for the 64bit BAR which can set
bits[51:48]. BIOS can control the BAR address, e.g. seabios moved 64bit pci window
to end of address space by using advertised physical bits[1].

[1] https://gitlab.com/qemu-project/seabios/-/commit/bcfed7e270776ab5595cafc6f1794bea0cae1c6c

> 
> I don't think it's unreasonable for KVM selftests to require access to
> /proc/cpuinfo.  Or actually, they can probably do the same thing and self-limit
> to 48-bit addresses if /proc/cpuinfo isn't available.
> 
> I'm not totally opposed to adding a more programmatic way for userspace to query
> 5-level EPT support, it just seems unnecessary.  E.g. unlike CPUID, userspace
> can't directly influence whether or not KVM uses 5-level EPT.  Even in hindsight,
> I'm not entirely sure KVM should expose such a knob, as it raises questions around
> interactions guest.MAXPHYADDR and memslots that I would rather avoid.
> 
> And even if we do add such uAPI, enumerating 5-level EPT in /proc/cpuinfo is
> definitely worthwhile, the only thing that would need to be tweaked is the
> justification in the changelog.
> 
> One thing we can do irrespective of feature enumeration is have kvm_mmu_page_fault()
> exit to userspace with an explicit error if the guest faults ona GPA that KVM
> knows it can't map, i.e. exit with KVM_EXIT_INTERNAL_ERROR or maybe even
> KVM_EXIT_MEMORY_FAULT instead of looping indefinitely.

If KVM does report guest.MAXPHYADDR=host.MAXPHYADDR, it is not reasonable to kill the
guest directly. And just reporting that it does not support 5-level EPT in /proc/cpuinfo
will make it difficult for users to realize that physical-bits needs to be forcibly
limited in the command. But advertising max addressable GPA via ioctl and this patch do
not conflict.

Thanks,
Tao