x86 RDTSCP instruction test code 1
Posted by Anonymous on Mon 17th Oct 2022 17:49
raw | new post
modification of post by Anonymous (view diff)

/*
 * x86_rdtscp1.c - test x86 RDTSCP instruction to get the current
 * CPU number
 *
 * Compile with:
 * $ gcc -std=gnu17 -m64 -Wall x86_rdtscp1.c -lpthread
 *
 * Written by Roland Mainz <roland.mainz@nrubsig.org>
 *
 */

#define _XOPEN_SOURCE 700
#define _GNU_SOURCE 1

#include <stdio.h>
#include <stdlib.h>
#include <errno.h>

#include <pthread.h>
#include <sched.h>

#include <x86intrin.h>


int main(int ac, char *av[])
{
        (void)puts("#start.");

        unsigned int A = 0;
        unsigned long long retval = 0;
        int cpu;

        /*
         * The x86 rdtscp instruction and |sched_getcpu()| syscall can
         * return different results when the scheduler reschedules the
         * running thread, usually at the syscall boundary.
         *
         * Notes:
         * - Using the FIFO scheduler can reduce the instances where
         * this happens, but even then the scheduler can get active:
         * $ time chrt --fifo 20 ksh93 -c 'while true ; do \
         *      x="$(/home/rmainz/tmp/x86_rdtscp/a.out)" ; \
         *      [[ "$x" == *match* ]] && break ; done ; \
         *      printf "res=%s\n" "$x"'
         * res=#start.
         * _rdtscp:        A=7, retval=43db24d4804f4e
         * sched_getcpu(): cpu=3
         * # mismatch A!=cpu
         * #done.
         * real    5m27.782s
         * user    3m53.786s
         * sys     1m40.138s
         *
         * - The only useable solution is to pin the thread to one
         * cpu strand, as shown in the |USE_FIXED_CPU_NUMBER| codepath
         * below...
         *
         * - Read https://unix4lyfe.org/benchmarking/
         *
         * - On some machines the cpu (strand) number is always |0|,
         * independent from which CPU/strand the thread is really
         * executing on... ;-(
         * Current (2022-10-17) working theory is that this is the CPU ID
         * of timer (the RDTSCP instruction returns a timer value!!) and
         * the specific CPU only has one timer...
         */

#define USE_FIXED_CPU_NUMBER 1
#if USE_FIXED_CPU_NUMBER
        cpu_set_t cpuset;
        pthread_t thread;

        thread = pthread_self();
        cpu = sched_getcpu();

        CPU_ZERO(&cpuset);
        CPU_SET(cpu, &cpuset);

        if (pthread_setaffinity_np(thread, sizeof(cpuset), &cpuset) == 0) {
                (void)printf("# pthread_setaffinity_np(), "
                        "thread fixed to cpu %d\n",
                        cpu);
        }
        else {
                /*
                 * If you specify a CPU number beyond the maximum number
                 * of CPUs, then |pthread_setaffinity_np()| can have a
                 * |errno == 0|
                 */
                perror("pthread_setaffinity_np");
        }
#endif /* USE_FIXED_CPU_NUMBER */

        retval = _rdtscp(&A);
        cpu = sched_getcpu();

        (void)printf("_rdtscp:\tA=%x, ", A);
        (void)printf("retval=%llx\n", retval);
        (void)printf("sched_getcpu():\tcpu=%x\n", cpu);

        if (cpu != A)
                (void)puts("# mismatch A!=cpu");

        (void)puts("#done.");

        return EXIT_SUCCESS;
}