[Xen-devel] [V2] x86/xsaves: fix overwriting between non-lazy/lazy xsaves

[Shuai Ruan <shuai.ruan@xxxxxxxxxxxxxxx>]

The offset at which components xsaved by xsave[sc] are not fixed.
So when when a save with v->fpu_dirtied set is followed by one
with v->fpu_dirtied clear, non-lazy xsave[sc] may overwriting data
written by the lazy one.

The solution is when xsave[sc] is enabled and taking xcr0_accum into
consideration, if guest has ever used lazy states, vcpu_xsave_mask will
return XSTATE_ALL. Otherwise vcpu_xsave_mask will return XSTATE_NONLAZY.
This may cause overhead save on lazy states which will cause performance
impact. As xsavec support code will be cleaned up(reason is list below),
so the patch only take xsaves into consideration.

After doing some performance test on xsavec and xsaveopt(suggested by jan),
the result show xsaveopt performs better than xsavec. So next step clean up
xsavec suppot code in xen and disable xsaves (xsaves should be used when
supervised state is introduced).

Signed-off-by: Shuai Ruan <shuai.ruan@xxxxxxxxxxxxxxx>
Reported-by: Jan Beulich <jbeulich@xxxxxxxx>
---

 V2: Address comments from Jan:
 Add performance impact and future steps in description.

 xen/arch/x86/i387.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/xen/arch/x86/i387.c b/xen/arch/x86/i387.c
index c29d0fa..0084578 100644
--- a/xen/arch/x86/i387.c
+++ b/xen/arch/x86/i387.c
@@ -118,7 +118,8 @@ static inline uint64_t vcpu_xsave_mask(const struct vcpu *v)
     if ( v->fpu_dirtied )
         return v->arch.nonlazy_xstate_used ? XSTATE_ALL : XSTATE_LAZY;
 
-    return v->arch.nonlazy_xstate_used ? XSTATE_NONLAZY : 0;
+    return cpu_has_xsaves && (v->arch.xcr0_accum & XSTATE_LAZY & 
~XSTATE_FP_SSE) ?
+           XSTATE_ALL : XSTATE_NONLAZY;
 }
 
 /* Save x87 extended state */
-- 
1.9.1


_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxx
http://lists.xen.org/xen-devel