Wrong instruction after single-step exception with 'rdtsc' -> fixed in 6.0

[gim]

There was bug 5 years ago (#10947) and was fixed, but in current release still appears. Here slightly modified code with looping 1000000 times around RDTSC call with charged TF. If at least one call does not work correctly, a corresponding message is displayed:

      .586
      .model flat, stdcall
      option casemap :none   ; case sensitive

      include \masm32\include\windows.inc
      include \masm32\include\kernel32.inc
      include \masm32\include\user32.inc
      includelib \masm32\lib\kernel32.lib
      includelib \masm32\lib\user32.lib
     
.data
Flag	dd 0
Address dd 0
Counter dd 0
szRight	db 'Flag Value is right!, address = 0x%lx, counter = %ld',0
szWrong	db 'Flag Value is wrong!, address = 0x%lx, counter = %ld',0
szMessage db 256 dup(0)
    
szInfo	db 'Info:'

.code
start:
	assume fs: nothing
test_loop:
	call @MyCode
	mov     ecx, dword ptr [esp+0Ch]
	mov 	ecx, dword ptr [ecx+0B8h]	;;Ecx = Seh.eip
        mov     Address, ecx
	.if ecx == offset @WrongExceptionEip
		mov Flag,0
	.else
		mov Flag,1
	.endif
	xor     eax, eax
	retn
	@MyCode:
	push    dword ptr fs:[0]
	mov     dword ptr fs:[0], esp
	push 397h             ;;Set Eflags
	popfd
	rdtsc
	@RightExceptionEip:		;;Normally,Seh.eip should be pointed here
	nop
	@WrongExceptionEip:		;;In Guest system,('Without' VT-X/AMD-V),Seh.eip is pointed here.But 'With' VT-X/AMD-V,Seh.eip is right.
        cmp Flag, 1
        jnz flag_wrong
        pop eax
        pop fs:[0]
        inc Counter
        cmp Counter, 1000000
        jnz test_loop
        invoke wsprintf,offset szMessage, offset szRight, Address, Counter
        jmp exit
flag_wrong:
        invoke wsprintf,offset szMessage, offset szWrong, Address, Counter
exit:
        invoke MessageBoxA,0,offset szMessage,offset szInfo,MB_OK
	invoke ExitProcess,0
end start

(compiled sample attached rdtsc.exe)

For example, in the real world, this misbehavior is used by the vmprotect to detect a virtual machine. I hope there is no good program crashing because of this misbehavior...

[gim]

compiled asm code

[gim]

And this problem appears with VT-X/AMD-V too on any Windows OS

[gim]

up

[michaln]

Please provide a VBox.log from a VM showing the problem. It would not hurt to specify what "any" Windows OS is either. Windows 3.1? Windows 95? Windows 10 64-bit?

[gim]

I've attached VBox.log and proof screenshot. But I believe that you could not find any usefully info inside VBox.log without enabling R0-logging or at least some VBOX RELEASE LOGGING flags. The problem probably lies somewhere deeply in VMM.

[gim]

up

[gim]

up

[janitor]

See also #17961.

[gim]

Well, we have been able to understand a little essence of the problem. The problem happens only when between "rdtsc vmexit handler" and "single step exception handler" on this instruction one more another interrupt was received (pended). We think this is because BS flag in VMCS clears during processing these last interrupt (probably because of some "27.3.4 Intel 64 and IA-32 Architectures Software Developers Manual Volume 3C: Part3" reasons) and single step exception will occur only on next instruction (so one instruction was skipped).

We’ve added single step interruption event pending and this solve the problem. Accordingly 32.2.1 "Intel 64 and IA-32 Architectures Software Developers Manual Volume 3C: Part3" it is possible to do so.

pending_fix.patch with this code attached

[Ramshankar Venkataraman]

Thanks for the patch and the testcases.

Unfortunately, your patch does not take into account the case where we single-step using EFLAGS.TF using the hypervisor debugger.

Also, the problem was not about an interrupt pending with RDTSC, it was simply RDTSC interception (which is dynamic) combined with single-stepping in the guest. When RDTSC was not intercepted by VirtualBox, it should work 100% of the time. Without Hyper-V/KVM paravirtualization interface in effect, we would be intercepting RDTSCs occasionally and only in those instances would it hit this issue.

I've implemented a more comprehensive fix which also makes your testcase work. The fix will be included in the upcoming release of VirtualBox 6.0.

I will try provide a test build shortly for you early next week.

In the mean time if you'd like the patch (probably only applies cleanly to the very latest VirtualBox OSE trunk and not 5.2.x branch), I've attached it to this ticket (vmx_singlestep_001.diff).

[Ramshankar Venkataraman]

Single-step patch to VirtualBox r127059