Recently, when ptrace-ing code, I found that the LBT register ftop is always zero, even thought the ptrace-d code sets ftop to non-zero. So I suspect ftop is not correctly maintained across context switches. I wrote the following code to verify my assumption:


#include <assert.h>
#include <unistd.h>
#include <stdio.h>

int main() {
    int b;
    for(int i = 0;;i++) {
        printf("loop %d\n", i);

        // set
        __asm__("x86mttop 0");

        // get
        __asm__("x86mftop %0" : "=r"(b));
        assert(b == 0);
        usleep(100000);

        // get again
        __asm__("x86mftop %0" : "=r"(b));
        assert(b == 0);

        // set
        __asm__("x86mttop 1");

        // get
        __asm__("x86mftop %0" : "=r"(b));
        assert(b == 1);
        usleep(100000);

        // get again
        __asm__("x86mftop %0" : "=r"(b));
        assert(b == 1);
    }
}


The code snippet sets and gets ftop in a loop, calls to usleep() is added to trigger context switch. The program will fail in several iterations:

loop 0
loop 1
loop 2
ftop: ftop.c:32: main: Assertion `b == 1' failed.
fish: Job 1, './ftop' terminated by signal SIGABRT (Abort)

It might even fail in the first iteration:

loop 0
ftop: ftop.c:32: main: Assertion `b == 1' failed.
fish: Job 1, './ftop' terminated by signal SIGABRT (Abort)

I am using Linux 6.6.0-rc5 with binutils 2.41.0 and gcc 13.2.1.