Question

I wrote a program which reads the APERF/MPERF counters on an Intel chip (page 2 on http://www.intel.com/content/dam/doc/manual/64-ia-32-architectures-software-developer-vol-3b-part-2-manual.pdf).

These counters are readable/writable via the readmsr/writemsr instructions, and I’m currently simply reading them at a regular interval via a device driver in Windows 7. The counters are 64 bits, and increment approximately with each processor clock, so you’d expect them to overflow in a very long amount of time, but when I read the counters, their value jumps around as if they are being reset by another program.

Is there any way to track down what program would be resetting the counters? Could something else be causing incorrect values to be read? The relevant assembly and corresponding C functions I’m using are attached below. The 64-bit result from rdmsr is saved into eax:edx, so to make sure I wasn’t missing any numbers in the r_x registers, I run the command multiple times to check them all.

C:

long long test1, test2, test3, test4;
test1 = TST1();
test2 = TST2();
test3 = TST3();
test4 = TST4();
status = RtlStringCbPrintfA(buffer, sizeof(buffer), "Value: %llu %llu %llu %llu\n", test1, test2, test3, test4);

Assembly:

;;;;;;;;;;;;;;;;;;;
PUBLIC TST1 
TST1 proc
    mov ecx, 231 ; 0xE7
    rdmsr
    ret ; returns rax
TST1 endp

;;;;;;;;;;;;;;;;;;;
PUBLIC TST2 
TST2 proc
    mov ecx, 231 ; 0xE7
    rdmsr
    mov rax, rbx
    ret ; returns rax
TST2 endp

;;;;;;;;;;;;;;;;;;;
PUBLIC TST3 
TST3 proc
    mov ecx, 231 ; 0xE7
    rdmsr
    mov rax, rcx
    ret ; returns rax
TST3 endp

;;;;;;;;;;;;;;;;;;;
PUBLIC TST4 
TST4  proc
    mov ecx, 231 ; 0xE7
    rdmsr
    mov rax, rdx
    ret ; returns rax
TST4 endp

The result that prints out is something like below, but the only register which ever changes is the rax register, and it doesn’t increase monotonically (can jump around):

Value: 312664 37 231 0
Value: 252576 37 231 0
Value: 1051857 37 231 0

Answer 1

I was not able to figure out what was resetting my counters, but I was able to determine the frequency. The Intel docs state that when one counter overflows, the other counter also will. So even though the counters are constantly resetting, the ratio of aperf and mperf still does represent the processor’s frequency.