x86/xsave: adjust state management

The initial state for a vCPU is using default values, so there's no
need to force the XRSTOR to read the state from memory. This saves a
couple of thousand restores from memory just during boot of Linux on
my Sandy Bridge system (I didn't try to make further measurements).

The above requires that arch_set_info_guest() updates the state flags
in the save area when valid floating point state got passed in, but
that would really have been needed even before in case XSAVE{,OPT}
decided to clear one or both of the FP and SSE bits.

Furthermore, hvm_vcpu_reset_state() shouldn't just clear out the FPU/
SSE area, but needs to re-initialized MXCSR and FCW.

Signed-off-by: Jan Beulich <jbeulich@suse.com>
Acked-by: Keir Fraser <keir@xen.org>

4 files changed, 17 insertions, 2 deletions

diff --git a/xen/arch/x86/domain.c b/xen/arch/x86/domain.c
index 52c9040f59..874742c3b1 100644
--- a/xen/arch/x86/domain.c
+++ b/xen/arch/x86/domain.c
@@ -712,7 +712,11 @@ int arch_set_info_guest(
     v->arch.vgc_flags = flags;
 
     if ( flags & VGCF_I387_VALID )
+    {
         memcpy(v->arch.fpu_ctxt, &c.nat->fpu_ctxt, sizeof(c.nat->fpu_ctxt));
+        if ( v->arch.xsave_area )
+             v->arch.xsave_area->xsave_hdr.xstate_bv = XSTATE_FP_SSE;
+    }
 
     if ( !compat )
     {
diff --git a/xen/arch/x86/hvm/hvm.c b/xen/arch/x86/hvm/hvm.c
index 43b6d052ba..1fcaed0b83 100644
--- a/xen/arch/x86/hvm/hvm.c
+++ b/xen/arch/x86/hvm/hvm.c
@@ -3466,6 +3466,7 @@ void hvm_vcpu_reset_state(struct vcpu *v, uint16_t cs, uint16_t ip)
 {
     struct domain *d = v->domain;
     struct segment_register reg;
+    typeof(v->arch.xsave_area->fpu_sse) *fpu_ctxt = v->arch.fpu_ctxt;
 
     domain_lock(d);
 
@@ -3479,7 +3480,12 @@ void hvm_vcpu_reset_state(struct vcpu *v, uint16_t cs, uint16_t ip)
         v->arch.guest_table = pagetable_null();
     }
 
-    memset(v->arch.fpu_ctxt, 0, sizeof(v->arch.xsave_area->fpu_sse));
+    memset(fpu_ctxt, 0, sizeof(*fpu_ctxt));
+    fpu_ctxt->fcw = FCW_RESET;
+    fpu_ctxt->mxcsr = MXCSR_DEFAULT;
+    if ( v->arch.xsave_area )
+        v->arch.xsave_area->xsave_hdr.xstate_bv = XSTATE_FP;
+
     v->arch.vgc_flags = VGCF_online;
     memset(&v->arch.user_regs, 0, sizeof(v->arch.user_regs));
     v->arch.user_regs.eflags = 2;
diff --git a/xen/arch/x86/xstate.c b/xen/arch/x86/xstate.c
index d52dbd715c..608accfe48 100644
--- a/xen/arch/x86/xstate.c
+++ b/xen/arch/x86/xstate.c
@@ -204,9 +204,13 @@ int xstate_alloc_save_area(struct vcpu *v)
     if ( save_area == NULL )
         return -ENOMEM;
 
+    /*
+     * Set the memory image to default values, but don't force the context
+     * to be loaded from memory (i.e. keep save_area->xsave_hdr.xstate_bv
+     * clear).
+     */
     save_area->fpu_sse.fcw = FCW_DEFAULT;
     save_area->fpu_sse.mxcsr = MXCSR_DEFAULT;
-    save_area->xsave_hdr.xstate_bv = XSTATE_FP_SSE;
 
     v->arch.xsave_area = save_area;
     v->arch.xcr0 = XSTATE_FP_SSE;
diff --git a/xen/include/asm-x86/xstate.h b/xen/include/asm-x86/xstate.h
index 49168f3293..dc3cb8eb75 100644
--- a/xen/include/asm-x86/xstate.h
+++ b/xen/include/asm-x86/xstate.h
@@ -11,6 +11,7 @@
 #include <xen/types.h>
 
 #define FCW_DEFAULT               0x037f
+#define FCW_RESET                 0x0040
 #define MXCSR_DEFAULT             0x1f80
 
 #define XSTATE_CPUID              0x0000000d