performance - rdpmc in user mode does not work even with PCE set -

based on wikipedia entry intel manual, rdpmc should available user-mode processes long bit 8 of cr4 set. however, still running general protection error when trying run rdpmc userspace bit set.

i running on 8-core intel x3470 on kernel 2.6.32-279.el6.x86_64.

here user-mode program trying execute:

#define _gnu_source  #include <stdio.h> #include <stdlib.h> #include <inttypes.h>  #include <sched.h> #include <assert.h>  uint64_t read_pmc(int ecx) {     unsigned int a, d;     __asm __volatile("rdpmc" : "=a"(a), "=d"(d) : "c"(ecx));     return ((uint64_t)a) | (((uint64_t)d) << 32); }  int main(int ac, char **av) {     uint64_t start, end;     cpu_set_t cpuset;     unsigned int c;     int i;      if (ac != 3) {         fprintf(stderr, "usage: %s cpu-id pmc-num\n", av[0]);         exit(exit_failure);     }      = atoi(av[1]);     c = atoi(av[2]);      cpu_zero(&cpuset);         cpu_set(i, &cpuset);         assert(sched_setaffinity(0, sizeof(cpuset), &cpuset) == 0);      printf("%lu\n", read_pmc(c));     return 0; } 

here kernel module sets bit , reads out cr4 can manually verify bit has been set.

/*    *  enable pmc in user mode.  */ #include <linux/module.h> #include <linux/kernel.h>    int init_module(void) {     typedef long unsigned int uint64_t;     uint64_t output;     // set cr4, bit 8  enable pmc __asm__("push   %rax\n\t"                 "mov    %cr4,%rax;\n\t"                 "or     $(1 << 7),%rax;\n\t"                 "mov    %rax,%cr4;\n\t"                 "wbinvd\n\t"                 "pop    %rax"     );      // read cr4 check bit.     __asm__("\t mov %%cr4,%0" : "=r"(output));     printk(kern_info "%lu", output);      return 0; } void cleanup_module(void) { __asm__("push   %rax\n\t"         "push   %rbx\n\t"                 "mov    %cr4,%rax;\n\t"                 "mov  $(1 << 7), %rbx\n\t"                 "not  %rbx\n\t"                 "and   %rbx, %rax;\n\t"                 "mov    %rax,%cr4;\n\t"                 "wbinvd\n\t"                 "pop    %rbx\n\t"                 "pop    %rax\n\t"     );  } 

apparently, when intel says bit 8, referring 9th bit right, since indexing begins @ 0. replacing $(1 << 7) $(1 << 8) globally resolves issue, , allows rdpmc called user mode.

here updated kernel module, using on_each_cpu make sure set on every core.

/*    *  read pmc in kernel mode.  */ #include <linux/module.h>   /* needed modules */ #include <linux/kernel.h>   /* needed kern_info */  static void printc4(void) {     typedef long unsigned int uint64_t;     uint64_t output;     // read cr4 check bit.     __asm__("\t mov %%cr4,%0" : "=r"(output));     printk(kern_info "%lu", output); }  static void setc4b8(void * info) {     // set cr4, bit 8 (9th bit right)  enable __asm__("push   %rax\n\t"                 "mov    %cr4,%rax;\n\t"                 "or     $(1 << 8),%rax;\n\t"                 "mov    %rax,%cr4;\n\t"                 "wbinvd\n\t"                 "pop    %rax"     );      // check cpu on:     printk(kern_info "ran on processor %d", smp_processor_id());     printc4(); }  static void clearc4b8(void * info) {     printc4(); __asm__("push   %rax\n\t"         "push   %rbx\n\t"                 "mov    %cr4,%rax;\n\t"                 "mov  $(1 << 8), %rbx\n\t"                 "not  %rbx\n\t"                 "and   %rbx, %rax;\n\t"                 "mov    %rax,%cr4;\n\t"                 "wbinvd\n\t"                 "pop    %rbx\n\t"                 "pop    %rax\n\t"     );     printk(kern_info "ran on processor %d", smp_processor_id()); }    int init_module(void) {     on_each_cpu(setc4b8, null, 0);     return 0; } void cleanup_module(void) {     on_each_cpu(clearc4b8, null, 0); }