x86/xstate: Elide redundant writes in set_xcr0()

XSETBV is an expensive instruction as, amongst other things, it involves
reconfiguring the instruction decode at the frontend of the pipeline.

We have several paths which reconfigure %xcr0 in quick succession (the context
switch path has 5, including the fpu save/restore helpers), and only a single
caller takes any care to try to skip redundant writes.

Update set_xcr0() to perform amortisation automatically, and simplify the
__context_switch() path as a consequence.

Signed-off-by: Andrew Cooper <andrew.cooper3@citrix.com>
Reviewed-by: Jan Beulich <jbeulich@suse.com>

diff --git a/xen/arch/x86/domain.c b/xen/arch/x86/domain.c
index 4dc27f798e74bee69d164f3f26fdebf4e3cbc909..50a27197b53e2b8920c7f323e0af2ef807813567 100644 (file)
--- a/xen/arch/x86/domain.c
+++ b/xen/arch/x86/domain.c
@@ -1977,9 +1977,7 @@ static void __context_switch(void)
         memcpy(stack_regs, &n->arch.user_regs, CTXT_SWITCH_STACK_BYTES);
         if ( cpu_has_xsave )
         {
-            u64 xcr0 = n->arch.xcr0 ?: XSTATE_FP_SSE;
-
-            if ( xcr0 != get_xcr0() && !set_xcr0(xcr0) )
+            if ( !set_xcr0(n->arch.xcr0 ?: XSTATE_FP_SSE) )
                 BUG();
 
             if ( cpu_has_xsaves && is_hvm_vcpu(n) )

diff --git a/xen/arch/x86/xstate.c b/xen/arch/x86/xstate.c
index d49413b745c8dc3059d008ea01e1d564382a81d4..65c720a26a368b62635b4b5b496b947141680523 100644 (file)
--- a/xen/arch/x86/xstate.c
+++ b/xen/arch/x86/xstate.c
@@ -55,11 +55,18 @@ static inline bool xsetbv(u32 index, u64 xfeatures)
     return lo != 0;
 }
 
-bool set_xcr0(u64 xfeatures)
+bool set_xcr0(u64 val)
 {
-    if ( !xsetbv(XCR_XFEATURE_ENABLED_MASK, xfeatures) )
-        return false;
-    this_cpu(xcr0) = xfeatures;
+    uint64_t *this_xcr0 = &this_cpu(xcr0);
+
+    if ( *this_xcr0 != val )
+    {
+        if ( !xsetbv(XCR_XFEATURE_ENABLED_MASK, val) )
+            return false;
+
+        *this_xcr0 = val;
+    }
+
     return true;
 }