Fix a long standing bug in VMXCTX_GUEST_RESTORE().

There was an assumption by the "callers" of this macro that on "return" the
%rsp will be pointing to the 'vmxctx'. The macro was not doing this and thus
when trying to restore host state on an error from "vmlaunch" or "vmresume"
we were treating the memory locations on the host stack as 'struct vmxctx'.
This led to all sorts of weird bugs like double faults or invalid instruction
faults.

This bug is exposed by the -O2 option used to compile the kernel module. With
the -O2 flag the compiler will optimize the following piece of code:

	int loopstart = 1;
	...
	if (loopstart) {
		loopstart = 0;
		vmx_launch();
	} else
		vmx_resume();

into this:

	vmx_launch();

Since vmx_launch() and vmx_resume() are declared to be __dead2 functions the
compiler is free to do this. The compiler has no way to know that the
functions return indirectly through vmx_setjmp(). This optimization in turn
leads us to trigger the bug in VMXCTX_GUEST_RESTORE().

With this change we can boot a 8.1 guest on a 9.0 host.

Reported by: jhb@


git-svn-id: svn://svn.freebsd.org/base/projects/bhyve@222112 ccf9f872-aa2e-dd11-9fc8-001c23d0bc1f

sys/amd64/vmm/intel/vmx.c

@@ -1189,7 +1189,7 @@ vmx_exit_process(struct vmx *vmx, int vcpu, struct vm_exit *vmexit)
 static int
 vmx_run(void *arg, int vcpu, register_t rip, struct vm_exit *vmexit)
 {
-	int error, vie, rc, handled, astpending, loopstart;
+	int error, vie, rc, handled, astpending;
 	uint32_t exit_reason;
 	struct vmx *vmx;
 	struct vmxctx *vmxctx;
@@ -1198,7 +1198,7 @@ vmx_run(void *arg, int vcpu, register_t rip, struct vm_exit *vmexit)
 	vmx = arg;
 	vmcs = &vmx->vmcs[vcpu];
 	vmxctx = &vmx->ctx[vcpu];
-	loopstart = 1;
+	vmxctx->launched = 0;
 
 	/*
 	 * XXX Can we avoid doing this every time we do a vm run?
@@ -1232,8 +1232,8 @@ vmx_run(void *arg, int vcpu, register_t rip, struct vm_exit *vmexit)
 #endif
 		switch (rc) {
 		case VMX_RETURN_DIRECT:
-			if (loopstart) {
-				loopstart = 0;
+			if (vmxctx->launched == 0) {
+				vmxctx->launched = 1;
 				vmx_launch(vmxctx);
 			} else
 				vmx_resume(vmxctx);

sys/amd64/vmm/intel/vmx.h

@@ -34,6 +34,9 @@
 #define	GUEST_MSR_MAX_ENTRIES	64		/* arbitrary */
 
 struct vmxctx {
+	register_t	tmpstk[32];		/* vmx_return() stack */
+	register_t	tmpstktop;
+
 	register_t	guest_rdi;		/* Guest state */
 	register_t	guest_rsi;
 	register_t	guest_rdx;
@@ -63,6 +66,7 @@ struct vmxctx {
 	 * XXX todo debug registers and fpu state
 	 */
 
+	int		launched;		/* vmcs launch state */
 	int		launch_error;
 };
 

sys/amd64/vmm/intel/vmx_genassym.c

@@ -43,6 +43,7 @@ __FBSDID("$FreeBSD$");
 #include "vmx.h"
 #include "vmx_cpufunc.h"
 
+ASSYM(VMXCTX_TMPSTKTOP, offsetof(struct vmxctx, tmpstktop));
 ASSYM(VMXCTX_GUEST_RDI, offsetof(struct vmxctx, guest_rdi));
 ASSYM(VMXCTX_GUEST_RSI, offsetof(struct vmxctx, guest_rsi));
 ASSYM(VMXCTX_GUEST_RDX, offsetof(struct vmxctx, guest_rdx));

sys/amd64/vmm/intel/vmx_support.S

@@ -31,15 +31,23 @@
 #include "vmx_assym.s"
 
 /*
- * Assumes that %rdi holds a pointer to the 'vmxctx'
+ * Assumes that %rdi holds a pointer to the 'vmxctx'.
+ *
+ * On "return" all registers are updated to reflect guest state. The two
+ * exceptions are %rip and %rsp. These registers are atomically switched
+ * by hardware from the guest area of the vmcs.
+ *
+ * We modify %rsp to point to the 'vmxctx' so we can use it to restore
+ * host context in case of an error with 'vmlaunch' or 'vmresume'.
  */
 #define	VMX_GUEST_RESTORE						\
 	/*								\
-	 * Make sure that interrupts are disabled before restoring CR2.	\
-	 * Otherwise there could be a page fault during the interrupt	\
-	 * handler execution that would end up trashing CR2.		\
+	 * Disable interrupts before updating %rsp. The location that	\
+	 * %rsp points to is a 'vmxctx' and not a real stack so we	\
+	 * don't want an interrupt handler to trash it.			\
 	 */								\
 	cli;								\
+	movq	%rdi,%rsp;						\
 	movq	VMXCTX_GUEST_CR2(%rdi),%rsi;				\
 	movq	%rsi,%cr2;						\
 	movq	VMXCTX_GUEST_RSI(%rdi),%rsi;				\
@@ -148,6 +156,8 @@ ENTRY(vmx_longjmp)
 
 	movq	%rsp,%rdi
 	movq	$VMX_RETURN_LONGJMP,%rsi
+
+	addq	$VMXCTX_TMPSTKTOP,%rsp
 	callq	vmx_return
 END(vmx_longjmp)
 
@@ -174,6 +184,8 @@ ENTRY(vmx_resume)
 	/* Return via vmx_setjmp with return value of VMX_RETURN_VMRESUME */
 	movq	%rsp,%rdi
 	movq	$VMX_RETURN_VMRESUME,%rsi
+
+	addq	$VMXCTX_TMPSTKTOP,%rsp
 	callq	vmx_return
 END(vmx_resume)
 
@@ -200,5 +212,7 @@ ENTRY(vmx_launch)
 	/* Return via vmx_setjmp with return value of VMX_RETURN_VMLAUNCH */
 	movq	%rsp,%rdi
 	movq	$VMX_RETURN_VMLAUNCH,%rsi
+
+	addq	$VMXCTX_TMPSTKTOP,%rsp
 	callq	vmx_return
 END(vmx_launch)