perf_event_open(2)

Synopsis

#include <linux/perf_event.h>
#include <sys/syscall.h>
#include <unistd.h>

/* no glibc wrapper; call via syscall(2) */
int perf_event_open(struct perf_event_attr *attr, pid_t pid, int cpu,
                    int group_fd, unsigned long flags);

Rod's take

perf stat calls perf_event_open(2). Call it yourself when you need counters inline without a subprocess or perf.data file. Use group_fd to atomic-read multiple counters from the same scheduling window. Always request PERF_FORMAT_TOTAL_TIME_ENABLED|PERF_FORMAT_TOTAL_TIME_RUNNING and scale accordingly when the kernel multiplexes more events than the PMU has physical registers. Set disabled=1 and gate with ioctl(PERF_EVENT_IOC_ENABLE/DISABLE) to control the measurement window precisely.

Flags and arguments that matter

Errors you will actually hit

Example

#include <linux/perf_event.h>
#include <sys/ioctl.h>
#include <sys/syscall.h>
#include <unistd.h>

struct perf_event_attr attr = {
    .type = PERF_TYPE_HARDWARE,
    .size = sizeof(attr),
    .config = PERF_COUNT_HW_INSTRUCTIONS,
    .disabled = 1,
    .exclude_kernel = 1,
};
int fd = syscall(SYS_perf_event_open, &attr, 0, -1, -1, PERF_FLAG_FD_CLOEXEC);
ioctl(fd, PERF_EVENT_IOC_RESET, 0);
ioctl(fd, PERF_EVENT_IOC_ENABLE, 0);
/* ... work ... */
ioctl(fd, PERF_EVENT_IOC_DISABLE, 0);
uint64_t count;
read(fd, &count, sizeof(count));

Trimmed from the full walkthrough in perf_event_open(2): hardware counters without the wrapper.

BSD equivalent

FreeBSD: hwpmc(4) and pmc(3)/pmcstat(8). Allocate via pmc_allocate(3), attach to a process, read via pmc_read(3). No group-atomic read. pmcstat -S instructions -p $pid is the closest equivalent.