On Fri, Feb 14, 2020 at 08:59:39AM -0800, Dave Hansen wrote:

On 2/13/20 11:23 PM, Joerg Roedel wrote:

quoted

Yes, investigating this is on the list for future optimizations (besides
caching CPUID results). My idea is to use alternatives patching for
this. But the exception handling is needed anyway because #VC
exceptions happen very early already, basically the first thing after
setting up a stack is calling verify_cpu(), which uses CPUID.

Ahh, bummer.  How does a guest know that it's running under SEV-ES?
What's the enumeration mechanism if CPUID doesn't "work"?

There are two ways a guest can find out:

	1) Read the SEV_STATUS_MSR and check for the SEV-ES bit
	2) If a #VC exception is raised it also knows it runs as an
	   SEV-ES guest

This patch-set implements both ways at the appropriate stages of the
boot process. Very early it just installs a #VC handler without checking
whether it is running under SEV-ES and handles the exceptions when they
are raised.

Later in the boot process it also reads the SEV_STATUS_MSR and sets a
cpu_feature flag to do alternative patching based on its value.

quoted

The other reason is that things like MMIO and IOIO instructions can't be
easily patched by alternatives. Those would work with the runtime
checking you showed above, though.

Is there a reason we can't make a rule that you *must* do MMIO through
an accessor function so we *can* patch them?  I know random drivers
might break the rule, but are SEV-ES guests going to be running random
drivers?  I would think that they mostly if not all want to use
virtio.

Yeah, there are already defined accessor functions for MMIO, like
read/write[bwlq] and memcpy_toio/memcpy_fromio. It is probably worth
testing what performance overhead is involved in overloading these to
call directly into the paravirt path when SEV-ES is enabled. With
alternatives patching it would still add a couple of NOPS for the
non-SEV-ES case.

But all that does not remove the need for the #VC exception handler, as
#VC exceptions can also be triggered by user-space, and the instruction
emulation for MMIO will be needed to allow MMIO in user-space (the
patch-set currently does not allow that, but it could be needed in the
future).

Regards,

	Joerg