[RFC 1/2] vfio/pci: keep the prefetchable attribute of a BAR region in VMA

[Alex Williamson]

On Thu, 29 Apr 2021 14:14:50 -0500
Shanker R Donthineni <sdonthineni at nvidia.com> wrote:

> Thanks Alex for quick reply.
> 
> On 4/29/21 1:28 PM, Alex Williamson wrote:
> > If this were a valid thing to do, it should be done for all
> > architectures, not just ARM64.  However, a prefetchable range only
> > necessarily allows merged writes, which seems like a subset of the
> > semantics implied by a WC attribute, therefore this doesn't seem
> > universally valid.
> >
> > I'm also a bit confused by your problem statement that indicates that
> > without WC you're seeing unaligned accesses, does this suggest that
> > your driver is actually relying on WC semantics to perform merging to
> > achieve alignment?  That seems rather like a driver bug, I'd expect UC
> > vs WC is largely a difference in performance, not a means to enforce
> > proper driver access patterns.  Per the PCI spec, the bridge itself can
> > merge writes to prefetchable areas, presumably regardless of this
> > processor attribute, perhaps that's the feature your driver is relying
> > on that might be missing here.  Thanks,  
> The driver uses WC semantics, It's mapping PCI prefetchable BARS
> using ioremap_wc().  We don't see any issue for x86 architecture,
> driver works fine in the host and guest kernel. The same driver works
> on ARM64 kernel but crashes inside VM. GPU driver uses the
> architecture agnostic function ioremap_wc() like other drivers. This
> limitation applies to all the drivers if they use WC memory and
> follow ARM64 NORMAL-NC access rules.

x86 KVM works for other reasons, KVM will trust the vCPU attributes for
the memory range rather than relying only on the host mapping.

> On ARM64, ioremap_wc() is mapped to non-cacheable memory-type, no
> side effects on reads and unaligned accesses are allowed as per
> ARM-ARM architecture. The driver behavior is different in host vs
> guest on ARM64. 

Per the PCI spec, prefetchable memory only necessarily allows the bridge
to merge writes.  I believe this is only a subset of what WC mappings
allow, therefore I expect this is incompatible with drivers that do not
use WC mappings.
 
> ARM CPU generating alignment faults before transaction reaches the
> PCI-RC/switch/end-point-device.

If an alignment fault is fixed by configuring a WC mapping, doesn't
that suggest that the driver performed an unaligned access itself and
is relying on write combining by the processor to correct that error?
That's wrong.  Fix the driver or please offer another explanation of
how the WC mapping resolves this.  I suspect you could enable tracing
in QEMU, disable MMIO mmaps on the vfio-pci device and find the invalid
access.

> We've two concerns here:
>    - Performance impacts for pass-through devices.
>    - The definition of ioremap_wc() function doesn't match the host
> kernel on ARM64

Performance I can understand, but I think you're also using it to mask
a driver bug which should be resolved first.  Thanks,

Alex