VMX: don't crash processing 'd' debug key

There's a window during scheduling where "current" and the active VMCS
may disagree: The former gets set much earlier than the latter. Since
both vmx_vmcs_enter() and vmx_vmcs_exit() immediately return when the
subject vCPU is "current", accessing VMCS fields would, depending on
whether there is any currently active VMCS, either read wrong data, or
cause a crash.

Going forward we might want to consider reducing the window during
which vmx_vmcs_enter() might fail (e.g. doing a plain __vmptrld() when
v->arch.hvm_vmx.vmcs != this_cpu(current_vmcs) but arch_vmx->active_cpu
== -1), but that would add complexities (acquiring and - more
importantly - properly dropping v->arch.hvm_vmx.vmcs_lock) that don't
look worthwhile adding right now.

Signed-off-by: Jan Beulich <jbeulich@xxxxxxxx>
Reviewed-by: Andrew Cooper <andrew.cooper3@xxxxxxxxxx>
---
v3: Introduce vmx_vmcs_try_enter() (as suggested by Tim).

--- a/xen/arch/x86/hvm/vmx/vmcs.c
+++ b/xen/arch/x86/hvm/vmx/vmcs.c
@@ -601,16 +601,16 @@ struct foreign_vmcs {
 };
 static DEFINE_PER_CPU(struct foreign_vmcs, foreign_vmcs);
 
-void vmx_vmcs_enter(struct vcpu *v)
+bool_t vmx_vmcs_try_enter(struct vcpu *v)
 {
     struct foreign_vmcs *fv;
 
     /*
      * NB. We must *always* run an HVM VCPU on its own VMCS, except for
-     * vmx_vmcs_enter/exit critical regions.
+     * vmx_vmcs_enter/exit and scheduling tail critical regions.
      */
     if ( likely(v == current) )
-        return;
+        return v->arch.hvm_vmx.vmcs == this_cpu(current_vmcs);
 
     fv = &this_cpu(foreign_vmcs);
 
@@ -633,6 +633,15 @@ void vmx_vmcs_enter(struct vcpu *v)
     }
 
     fv->count++;
+
+    return 1;
+}
+
+void vmx_vmcs_enter(struct vcpu *v)
+{
+    bool_t okay = vmx_vmcs_try_enter(v);
+
+    ASSERT(okay);
 }
 
 void vmx_vmcs_exit(struct vcpu *v)
--- a/xen/arch/x86/hvm/vmx/vmx.c
+++ b/xen/arch/x86/hvm/vmx/vmx.c
@@ -675,7 +675,27 @@ void vmx_get_segment_register(struct vcp
 {
     unsigned long attr = 0, sel = 0, limit;
 
-    vmx_vmcs_enter(v);
+    /*
+     * We may get here in the context of dump_execstate(), which may have
+     * interrupted context switching between setting "current" and
+     * vmx_do_resume() reaching the end of vmx_load_vmcs(). That would make
+     * all the VMREADs below fail if we don't bail right away.
+     */
+    if ( unlikely(!vmx_vmcs_try_enter(v)) )
+    {
+        static bool_t warned;
+
+        if ( !warned )
+        {
+            warned = 1;
+            printk(XENLOG_WARNING "Segment register inaccessible for d%dv%d\n"
+                   "(If you see this outside of debugging activity,"
+                   " please report to xen-devel@xxxxxxxxxxxxxxxxxxxx)\n",
+                   v->domain->domain_id, v->vcpu_id);
+        }
+        memset(reg, 0, sizeof(*reg));
+        return;
+    }
 
     switch ( seg )
     {
--- a/xen/include/asm-x86/hvm/vmx/vmcs.h
+++ b/xen/include/asm-x86/hvm/vmx/vmcs.h
@@ -144,6 +144,7 @@ struct arch_vmx_struct {
 int vmx_create_vmcs(struct vcpu *v);
 void vmx_destroy_vmcs(struct vcpu *v);
 void vmx_vmcs_enter(struct vcpu *v);
+bool_t __must_check vmx_vmcs_try_enter(struct vcpu *v);
 void vmx_vmcs_exit(struct vcpu *v);
 
 #define CPU_BASED_VIRTUAL_INTR_PENDING        0x00000004