使用performance counter读取硬件或软件Event

从Linux Kernel2.6.31版本开始，Linux内核开始提供一个叫__NR_perf_counter_open(最新的版本里叫__NR_perf_event_open)的系统调用。使用这个系统调用我们可以像使用文件一样打开一个Performance counter，通过设置不同的参数让这个Performance Counter统计不同的软件或硬件事件，然后就可以向读文件一样来读取这些事件的统计结果。比如我可以打开一个Performance Counter统计某一个进程的CPU Cache Miss次数。关于如何传递参数构造Performance Counter来统计不同的事件可以看这篇日志： http://tblog29.appspot.com/blog/1004

 

下面是我写的一个小程序，它为每个CPU和每个进程开一个Performance Counter，统计每个CPU上的Cache miss和每个进程上的Cache miss（不能统计每个进程在单个CPU上的事件，详见上边那篇日志）。本代码参考了 perf 的stat部分。运行需要CAP_SYS_ADMIN权限

 

1 mperf.h

/*
 * eperf.h
 *
 *  Created on: Jan 28, 2010
 *      Author: hchen
 */

#ifndef EPERF_H_
#define EPERF_H_
#include <time.h>
#include <asm/unistd.h>
#include "perf_event.h"

#define MAX_COUNTERS		256
#define MAX_NR_CPUS			32
#define PROC		"/proc"
/*
 * We define u64 as unsigned long long for every architecture
 * so that we can print it with %Lx without getting warnings.
 */
typedef unsigned long long u64;
typedef signed long long   s64;
typedef unsigned int	   u32;
typedef signed int	   s32;
typedef unsigned short	   u16;
typedef signed short	   s16;
typedef unsigned char	   u8;
typedef signed char	   s8;

static inline int
sys_perf_event_open(struct perf_event_attr *attr,
		      pid_t pid, int cpu, int group_fd,
		      unsigned long flags)
{
	attr->size = sizeof(*attr);
	//This system call is defined in asm/unistd.h, in the latest linux kernel
	//it's name has been changed to __NR_perf_event_open .
	return syscall(__NR_perf_counter_open, attr, pid, cpu, group_fd, flags);
}

#endif /* EPERF_H_ */

 2 eperf.c

#include <time.h>
#include <unistd.h>
#include <stdlib.h>
#include <stdio.h>
#include <asm/unistd.h>
#include <dirent.h>

#include "eperf.h"

unsigned int verbose = 0;
//event to be countered
static struct perf_event_attr attrs[] = {
  { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CACHE_MISSES	}
};

int nr_counters = 0;
static unsigned int	nr_cpus				=  0; // amount of cpus

static int			inherit				=  1;
static int			scale				=  1;
//used to save performance counter
static int			fd[MAX_COUNTERS];

/*
 * Read out the results of a single counter:
 */
static void read_counter(int counter)
{
	u64 single_count[3];
	size_t res, nv;

	if (fd[counter] <= 0)
		return;

	nv = scale ? 3 : 1;
	res = read(fd[counter], single_count, nv * sizeof(u64));

	if(res == nv * sizeof(u64)){
		if(verbose)
			printf("Counter %d: %llu\n", counter, single_count[0]);
	}else{
		fprintf(stderr, "Fail to read counter %d\n", counter);
	}
}

void close_all_counters(){
	int counter, tn;
	tn = nr_cpus + nr_counters;

	for (counter = 0; counter < tn; counter++){
		if (fd[counter] <= 0)
			continue;

		close(fd[counter]);
		fd[counter] = -1;
	}

}

void run_perf_stat()
{
	int counter, tn;
	tn = nr_cpus + nr_counters;

	for (counter = 0; counter < tn; counter++)
		read_counter(counter);
}

static void create_perf_stat_counter(int counter, int pid, int system_wide)
{
	struct perf_event_attr attr; //cache miss
	memcpy(&attr, attrs, sizeof(struct perf_event_attr));

	if (scale)
		attr.read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
				    PERF_FORMAT_TOTAL_TIME_RUNNING;

	if (system_wide) {
		unsigned int cpu;

		for (cpu = 0; cpu < nr_cpus; cpu++) {
			fd[cpu] = sys_perf_event_open(&attr, -1, cpu, -1, 0);
		}
	} else {
		attr.inherit	     = inherit;
		attr.disabled	     = 0;
		attr.enable_on_exec = 1;

		fd[counter + nr_cpus] = sys_perf_event_open(&attr, pid, -1, -1, 0);
	}
}

int main(int argc, const char **argv)
{
	if(argc > 1)
		verbose = atoi(argv[1]);

	DIR *dir;
	struct dirent *drp;
	int run_count, p, pid;
	struct timespec tim, tim2;
	tim.tv_sec = 1; tim.tv_nsec = 0;

	nr_cpus = sysconf(_SC_NPROCESSORS_ONLN);//the the number of CPU

	int counter = 0;
	/* Open /proc directory */
	if ((dir = opendir(PROC)) == NULL) {
		perror("opendir /proc error!");
		exit(4);
	}

	//create counters for each CPU
	create_perf_stat_counter(-1, 1, 1);

	p = 0;
	while ((p++) < 254) {
		/* Get directory entries */
		while ((drp = readdir(dir)) != NULL) {
			if (isdigit(drp->d_name[0]))
				break;
		}

		if (drp) {
			pid = atoi(drp->d_name);
			create_perf_stat_counter(counter, pid, 0);
			if(fd[counter] != -1)
				counter++;
		}

	}

	nr_counters = counter - 1;
	/* Close /proc directory */
	closedir(dir);

	run_count = 100;
//	for (run_idx = 0; run_idx < run_count; run_idx++) {
	while (1) {
		nanosleep(&tim , &tim2);
		run_perf_stat();
	}

	close_all_counters();

	return 1;
}