Re: [Xen-devel] [PATCH 10/24] xen: mask XSAVE from cpuid
H. Peter Anvin wrote:
Jeremy Fitzhardinge wrote:
Jan Beulich wrote:
As pointed out on an earlier thread, it seems inappropriate to do probing
like this when there is a cpuid feature flag (osxsave) that can be
determine whether XSAVE can be used. And even without that flag,
simply reading CR4 and checking whether osxsave is set there would
suffice. This is under the assumption that Xen's to-be-done
of XSAVE support would match that of FXSAVE (Xen turns its support on
unconditionally and for all [pv] guests).
I didn't want to make too many assumptions about how Xen's XSAVE support
would look. In particular, I thought it might virtualize the state of
OSXSAVE to give the guest the honour of appearing to enable it. A guest
kernel may get confused if it starts with OSXSAVE set, as it may use it
to control its own init logic.
That wouldn't be an issue if you use the *native* CPUID to look for
OSXSAVE early on, since such virtualization would only be visible though
the PV interface, right?
It seems cleaner than probing, to be sure...
Well, at the moment the problem is that cpuid (both PV and native) show
XSAVE, but Xen prevents cr4.OSXSAVE from being set, crashing the
kernel. There's now a patch in Xen to mask XSAVE in CPUID, so that
guests don't try to use it; the patch in the kernel is just to support
non-bleeding-edge versions of Xen.
There have been some patches floating around for Xen support of XSAVE,
but I think there are some issues with the variable-sized CPU context
and save/restore/migrate, so they've been put on the backburner until
there's a real need for them. I haven't looked at them, but I wouldn't
have assumed that Xen would necessarily set OSXSAVE for itself, or
require guests to do so (if a guest can make do with a simpler CPU
context structure, then that might be simpler for things like
cross-architecture migration, etc). I think that the only safe
assumption is that XSAVE is available iff cpuid.XSAVE is set, modulo the
bug mentioned above.
I guess if we support XSAVE for any vcpu, all the pcpus must have
OSXSAVE set, and we rely on the fact that the XSAVE format is compatible
with FXSAVE where they overlap. But I really don't know what happens
when guests use xsetbv and how that might be virtualized/paravirtualized.
Xen-devel mailing list