C++ perf_event_open函数代码示例

void Performance::_add(perf_type_id type, long long config) {
    struct perf_event_attr pe;
    memset(&pe, 0, sizeof(struct perf_event_attr));
    pe.size = sizeof(struct perf_event_attr);
    pe.exclude_kernel = 1;
    pe.exclude_hv = 1;

    pe.type = type;
    pe.config = config;


    order[events].type = type;
    order[events].config = config;
    events++;

    if (fd == -1) {
        /* if changed: change calculations in get */
        pe.read_format = PERF_FORMAT_GROUP;
        pe.disabled = 1;

        fd = perf_event_open(&pe, 0, -1, -1, 0);
        if (fd < 0) {
            fprintf(stderr, "Error opening leader for %llx / %llx\n", (long long) type, config);
            exit(EXIT_FAILURE);
        }
    } else {
        int ret = perf_event_open(&pe, 0, -1, fd, 0);
        if (ret < 0) {
            fprintf(stderr, "Error connecting to leader for %llx / %llx\n", (long long) type, config);
            exit(EXIT_FAILURE);
        }
    }
}

int *fd_group_init(int num_events, char **events) {
  struct perf_event_attr *attr;
  int fd, ret, leader_fd, i;
  int *fds = malloc(num_events * sizeof(int));
  if (!fds) err(1, "could not allocate memory");
  attr = calloc(num_events, sizeof(*attr));
  // do refs, then misses
  for (i = 0; i < num_events; i++) {
    /*
     * 1st argument: event string
     * 2nd argument: default privilege level (used if not specified in the event string)
     * 3rd argument: the perf_event_attr to initialize
     */
    ret = pfm_get_perf_event_encoding(events[i], PFM_PLM3, &attr[i], NULL, NULL);
    if (ret != PFM_SUCCESS) errx(1, "evt %d: cannot find encoding: %s", i, pfm_strerror(ret));
    printf("Using encoding %lx for event %s\n", attr[i].config, events[i]);

    attr[i].inherit = 0; // inheritance currently doesn't work with FORMAT_GROUP

    /*
     * request timing information because event may be multiplexed
     * and thus it may not count all the time. The scaling information
     * will be used to scale the raw count as if the event had run all
     * along
     */
    attr[i].read_format = PERF_FORMAT_TOTAL_TIME_ENABLED|PERF_FORMAT_TOTAL_TIME_RUNNING
                          | PERF_FORMAT_GROUP;

    /* do not start immediately after perf_event_open() */
    attr[i].disabled = 1;

    /*
     * create the event and attach to self
     * Note that it attaches only to the main thread, there is no inheritance
     * to threads that may be created subsequently.
     *
     * if mulithreaded, then getpid() must be replaced by gettid()
     */
    if (i == 0) {
      fd = perf_event_open(&attr[i], 0, -1, -1, 0);
      fds[i] = fd;
      leader_fd = fd;
    } else {
      fd = perf_event_open(&attr[i], 0, -1, leader_fd, 0);
      fds[i] = fd;
    }
    if (fd < 0) {
      warn("warning: evt %d: cannot create event", i);
      free(attr);
      free(fds);
      return NULL;
    }
  }
  free(attr);
  return fds;
}

static void setup(void)
{
	int i;
	struct perf_event_attr tsk_event, hw_event;

	/*
	 * According to perf_event_open's manpage, the official way of
	 * knowing if perf_event_open() support is enabled is checking for
	 * the existence of the file /proc/sys/kernel/perf_event_paranoid.
	 */
	if (access("/proc/sys/kernel/perf_event_paranoid", F_OK) == -1)
		tst_brkm(TCONF, NULL, "Kernel doesn't have perf_event support");

	tst_sig(NOFORK, DEF_HANDLER, cleanup);

	TEST_PAUSE;

	nhw = count_hardware_counters();
	n = nhw + 4;

	memset(&hw_event, 0, sizeof(struct perf_event_attr));
	memset(&tsk_event, 0, sizeof(struct perf_event_attr));

	tsk_event.type =  PERF_TYPE_SOFTWARE;
	tsk_event.size = sizeof(struct perf_event_attr);
	tsk_event.disabled = 1;
	tsk_event.config = PERF_COUNT_SW_TASK_CLOCK;

	hw_event.type = PERF_TYPE_HARDWARE;
	hw_event.size = sizeof(struct perf_event_attr);
	hw_event.disabled = 1;
	hw_event.config =  PERF_COUNT_HW_INSTRUCTIONS;

	tsk0 = perf_event_open(&tsk_event, 0, -1, -1, 0);
	if (tsk0 == -1) {
		tst_brkm(TBROK | TERRNO, cleanup, "perf_event_open failed");
	} else {
		tsk_event.disabled = 0;
		for (i = 0; i < n; ++i) {
			hwfd[i] = perf_event_open(&hw_event, 0, -1, -1, 0);
			tskfd[i] = perf_event_open(&tsk_event, 0, -1,
						   hwfd[i], 0);
			if (tskfd[i] == -1 || hwfd[i] == -1) {
				tst_brkm(TBROK | TERRNO, cleanup,
					 "perf_event_open failed");
			}
		}
	}
}

struct mmap_pages *
perf_event__open(struct event_open *e_open)
{
	int fd = -1;

	e_open->attr.disabled 		= 0;
	e_open->attr.size 		= sizeof(e_open->attr);
	e_open->attr.sample_period	= 1;
	e_open->attr.inherit		= 1;
	e_open->attr.sample_type 	= PERF_SAMPLE_TIME 	|
					  PERF_SAMPLE_RAW  	|
					  PERF_SAMPLE_CPU	|
					  PERF_SAMPLE_TID  	;
	e_open->attr.__reserved_1	= 0;
	e_open->attr.read_format	= PERF_FORMAT_TOTAL_TIME_ENABLED |
					  PERF_FORMAT_TOTAL_TIME_RUNNING ;

	fd = perf_event_open(&e_open->attr,e_open->pid,e_open->cpu	\
				,e_open->group_id,e_open->flags);

	if(fd < 0) {
		fprintf(stderr,"Failed to open event FD\n");
		return NULL;
	}

	struct mmap_pages *mmap_pages = NULL;
	mmap_pages = mmap_pages__new(fd,8);
	if(mmap_pages != NULL) {
		mmap_pages->attr = e_open->attr;
		fcntl(fd, F_SETFL, O_NONBLOCK);
	}
	return mmap_pages;
}

int main(int ac, char **av)
{
	int test = 0;
	int ret = 0;

	while (*++av) {
		if (!strcmp(*av, "--test")) {
			test = 1;
			continue;
		}

		struct perf_event_attr attr;
		if (resolve_event(*av, &attr) < 0) {
			fprintf(stderr, "Cannot resolve %s\n", *av); 
			ret = 1;
			continue;
		}
		char *ev = format_raw_event(&attr, *av);
		printf("%s\n", ev);
		free(ev);
		if (test) {
			if (perf_event_open(&attr, 0, -1, -1, 0) < 0)
				perror("perf_event_open");
		}
	}
	return ret;
}

void perf_start(const char* kname) {
  if (!enabled) return;

  char* prof_envvar = getenv("MXPA_PROFILE");

  int tids[32];
  int ntid;
  get_tids(tids, &ntid);
  g_nthreads = ntid;

  int n;
  for (n = 0; n < ntid; n++) {
    int ret;
    ret = perf_setup_list_events(prof_envvar, &(g_fds[n]), &num_fds);
    perf_event_desc_t *fds = g_fds[n];
    int cpu = -1;
    int group_fd = -1;
    int pid = tids[n];
    fds[0].fd = -1;
    int i;
    for(i=0; i < num_fds; i++) {
      fds[i].hw.read_format = PERF_FORMAT_SCALE;
      fds[i].hw.disabled = 1; /* do not start now */
      fds[i].hw.inherit = 1;  /* XXX child process will inherit, when forked only? */

      /* each event is in an independent group (multiplexing likely) */
      fds[i].fd = perf_event_open(&fds[i].hw, pid, cpu, group_fd, 0);
      if (fds[i].fd == -1) {
        fprintf(stderr, "cannot open event %d\n", i);
        exit(2);
      }
    }
  }
  prctl(PR_TASK_PERF_EVENTS_ENABLE);
}

void perfevent_initialize(char*events) {
  int i, ret;
  perfevent_events = strdup(events);

  /*
   * Initialize pfm library (required before we can use it)
   */
  ret = pfm_initialize();
  if (ret != PFM_SUCCESS)
    errx(1, "Cannot initialize library: %s", pfm_strerror(ret));

  ret = perf_setup_list_events(events, &perfevent_fds, &perfevent_num_fds);
  if (ret || !perfevent_num_fds)
    errx(1, "cannot setup events");

  perfevent_fds[0].fd = -1;
  for(i=0; i < perfevent_num_fds; i++) {
    /* request timing information necessary for scaling */
    perfevent_fds[i].hw.read_format = PERF_FORMAT_SCALE;

    perfevent_fds[i].hw.disabled = (i == 0); /* do not start now */
    perfevent_fds[i].hw.inherit = 1; /* pass on to child threads */

    /* each event is in an independent group (multiplexing likely) */
    perfevent_fds[i].fd = perf_event_open(&perfevent_fds[i].hw, 0, -1, perfevent_fds[0].fd, 0);
    if (perfevent_fds[i].fd == -1)
      err(1, "cannot open event %d", i);
  }
}

void parent(int sock) {

	ssize_t size;
	int fd,i;
	struct perf_event_attr pe;

	memset(&pe,0,sizeof(struct perf_event_attr));
	pe.type=PERF_TYPE_HARDWARE;
	pe.size=sizeof(struct perf_event_attr);
	pe.config=PERF_COUNT_HW_INSTRUCTIONS;
	pe.disabled=1;
	pe.exclude_kernel=1;
	pe.exclude_hv=1;

	arch_adjust_domain(&pe,quiet);

	fd=perf_event_open(&pe,0,-1,-1,0);
	if (fd<0) {
		fprintf(stderr,"Error opening leader %llx %s\n",pe.config,strerror(errno));
		test_fail(test_string);
	}

	size = sock_fd_write(sock, "1", 1, fd);
	if (!quiet) printf ("wrote fd %d (size %d)\n", fd, (int)size);

	sleep(1);

	for(i=0;i<20;i++) instructions_million();

}

int setup_event(struct event *e, int cpu, struct event *leader,
		bool measure_all, int measure_pid)
{
	e->attr.inherit = 1;
	if (!measure_all) {
		e->attr.disabled = 1;
		e->attr.enable_on_exec = 1;
	}
	e->attr.read_format |= PERF_FORMAT_TOTAL_TIME_ENABLED |
				PERF_FORMAT_TOTAL_TIME_RUNNING;

	e->efd[cpu].fd = perf_event_open(&e->attr,
			measure_all ? -1 : measure_pid,
			cpu,
			leader ? leader->efd[cpu].fd : -1,
			0);

	if (e->efd[cpu].fd < 0) {
		/* Handle offline CPU */
		if (errno == EINVAL && !cpu_online(cpu))
			return 0;

		fprintf(stderr, "Cannot open perf event for %s/%d: %s\n",
				e->event, cpu, strerror(errno));
		return -1;
	}
	return 0;
}

int main(int argc, char **argv) {

	int fd;

	struct perf_event_attr pe;

	int i;
	int result;
	long long counts[1],prev=0;

	long long total=0,average,max=0,min=0x7ffffffffffffffULL;

	memset(&pe,0,sizeof(struct perf_event_attr));
	pe.type=PERF_TYPE_HARDWARE;
	pe.size=sizeof(struct perf_event_attr);
	pe.config=PERF_COUNT_HW_INSTRUCTIONS;
	pe.disabled=1;
	pe.exclude_kernel=1;
	pe.exclude_hv=1;

	fd=perf_event_open(&pe,0,-1,-1,0);
	if (fd<0) {
     		fprintf(stderr,"Error opening leader %llx %s\n",pe.config,strerror(errno));
		exit(1);
	}

	ioctl(fd, PERF_EVENT_IOC_RESET, 0);
	ioctl(fd, PERF_EVENT_IOC_ENABLE,0);

	for(i=0;i<NUM_RUNS;i++) {

		result=instructions_million();

		result=read(fd,&counts,sizeof(long long));

		results[i]=counts[0]-prev;
		prev=counts[0];

 	}

	ioctl(fd, PERF_EVENT_IOC_DISABLE,0);

	close(fd);

	for(i=0;i<NUM_RUNS;i++) {
		total+=results[i];
		if (results[i]>max) max=results[i];
		if (results[i]<min) min=results[i];
	}

	average=total/NUM_RUNS;
	printf("Average=%lld max=%lld min=%lld\n",average,max,min);

	(void) result;

	return 0;
}

int
measure(pid_t pid)
{
	perf_event_desc_t *fds = NULL;
	int i, ret, num_fds = 0;
	char fn[32];

	if (pfm_initialize() != PFM_SUCCESS)
		errx(1, "libpfm initialization failed\n");

	ret = perf_setup_list_events(options.events, &fds, &num_fds);
	if (ret || (num_fds == 0))
		exit(1);

	fds[0].fd = -1;
	for(i=0; i < num_fds; i++) {
		fds[i].hw.disabled = 0; /* start immediately */

		/* request timing information necessary for scaling counts */
		fds[i].hw.read_format = PERF_FORMAT_SCALE;
		fds[i].hw.pinned = !i && options.pinned;
		fds[i].fd = perf_event_open(&fds[i].hw, pid, -1, (options.group? fds[0].fd : -1), 0);
		if (fds[i].fd == -1)
			errx(1, "cannot attach event %s", fds[i].name);
	}
	/*
	 * no notification is generated by perf_counters
	 * when the monitored thread exits. Thus we need
	 * to poll /proc/ to detect it has disappeared,
	 * otherwise we have to wait until the end of the
	 * timeout
	 */
	sprintf(fn, "/proc/%d/status", pid);

	while(access(fn, F_OK) == 0 && options.delay) {
		sleep(1);
		options.delay--;
		if (options.print)
			print_counts(fds, num_fds, 1);
	}
	if (options.delay)
		warn("thread %d terminated before timeout", pid);

	if (!options.print)
		print_counts(fds, num_fds, 0);

	for(i=0; i < num_fds; i++)
		close(fds[i].fd);

	free(fds);

	/* free libpfm resources cleanly */
	pfm_terminate();

	return 0;
}

int main()
{
#ifdef __NR_perf_event_open
    struct perf_event_attr pe;
    int fd;
    long long count;

    memset(&pe, 0, sizeof(struct perf_event_attr));
    pe.type = PERF_TYPE_SOFTWARE;
    pe.size = sizeof(struct perf_event_attr);
    pe.config = PERF_COUNT_SW_CPU_CLOCK;
    pe.disabled = 0;
    pe.exclude_kernel = 1;
    pe.exclude_hv = 1;

    fd = perf_event_open(&pe, 0, -1, -1, 0);
    //staptest// perf_event_open (XXXX, 0, -1, -1, 0x0) = NNNN

    sleep(1);
    
    ioctl(fd, PERF_EVENT_IOC_DISABLE, 0);
    read(fd, &count, sizeof(count));
    
    // Limit testing.

    fd = perf_event_open((struct perf_event_attr*)-1, 0, -1, -1, 0);
#ifdef __s390__
    //staptest// perf_event_open (0x[7]?[f]+, 0, -1, -1, 0x0) = -NNNN
#else
    //staptest// perf_event_open (0x[f]+, 0, -1, -1, 0x0) = -NNNN
#endif

    fd = perf_event_open(&pe, -1, -1, -1, 0);
    //staptest// perf_event_open (XXXX, -1, -1, -1, 0x0) = -NNNN

    fd = perf_event_open(&pe, 0, -1, -1, -1);
    //staptest// perf_event_open (XXXX, 0, -1, -1, PERF_FLAG_[^ ]+|XXXX) = -NNNN
    
    close(fd);
#endif
    return 0;
}

int setupHandler(int sig) {
  // Perf event settings
  struct perf_event_attr pe;

	memset(&pe, 0, sizeof(struct perf_event_attr));
	 
  pe.type = PERF_TYPE_HW_CACHE;
  pe.size = sizeof(struct perf_event_attr);
	pe.config = (PERF_COUNT_HW_CACHE_L1D << 0) | (PERF_COUNT_HW_CACHE_OP_READ << 8) | (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16); 
	//pe.config = (PERF_COUNT_HW_CACHE_L1D << 0) | (PERF_COUNT_HW_CACHE_OP_WRITE << 8) | (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16); 
  pe.disabled = 1;
	pe.exclude_kernel = 1;
	pe.exclude_hv = 1;
//	pe.exclude_callchain_kernel = 1;
//  pe.exclude_idle = 1;
  pe.sample_period = 1000;
//	pe.sample_id_all = 1;
	pe.sample_type=sample_type;
//	pe.precise_ip = 0;
 // pe.read_format=read_format;
	// Record those information when trap happens
	//pe.sample_type = PERF_SAMPLE_IP | PERF_SAMPLE_TID | PERF_SAMPLE_ADDR;
 
  // Create the perf_event for this thread on all CPUs with no event group
  int perf_fd = perf_event_open(&pe, 0, -1, -1, 0);
  if(perf_fd == -1) {
    fprintf(stderr, "Failed to open perf event file: %s\n", strerror(errno));
    abort();
  }
 
	// Setting up 9 pages to pass information about a trap
	our_mmap=mmap(NULL, 9*4096,
      PROT_READ|PROT_WRITE, MAP_SHARED, perf_fd, 0);
 
  // Set the perf_event file to async mode
  if(fcntl(perf_fd, F_SETFL, O_RDWR|O_NONBLOCK|O_ASYNC) == -1) {
    fprintf(stderr, "Failed to set perf event file to ASYNC mode: %s\n", strerror(errno));
    abort();
  }
  
  // Tell the file to send a SIGUSR1 when an event occurs
  if(fcntl(perf_fd, F_SETSIG, sig) == -1) {
    fprintf(stderr, "Failed to set perf event file's async signal: %s\n", strerror(errno));
    abort();
  }
  
  // Deliver the signal to this thread
  if(fcntl(perf_fd, F_SETOWN, getpid()) == -1) {
    fprintf(stderr, "Failed to set the owner of the perf event file: %s\n", strerror(errno));
    abort();
  }
  
  return perf_fd;
}

int perf_init(unsigned int num_ctrs, ...) {
    va_list args;
	int ctr, ret = 0;
	number = num_ctrs;

	attr = (struct perf_event_attr*)calloc(num_ctrs, sizeof(struct perf_event_attr));
	fds  = (int*) malloc(sizeof(int) * num_ctrs);
	if (attr == NULL || fds == NULL) {
		perror("Could not allocate space for counter data: ");
		ret = -1;
		goto error;
	}

    //Configure each counter, staring with 1
    va_start(args, num_ctrs);
    for (ctr=0; ctr<num_ctrs; ctr++) {
        unsigned event, conf;
        conf = va_arg(args, unsigned int) & 0xFFFFFF;
        event = conf & 0xFFFF;

		attr[ctr].type = PERF_TYPE_RAW;
		attr[ctr].config = event; /* generic PMU event*/
		attr[ctr].disabled = 0;
		attr[ctr].pinned = 1;
		attr[ctr].inherit = 1;
		fds[ctr] = perf_event_open(&attr[ctr], getpid(), -1, -1, 0);
		if ((conf & PERFMON_EVENTSEL_OS) == 0)
			attr[ctr].exclude_kernel = 1; 
		if ((conf & PERFMON_EVENTSEL_USR) == 0)
			attr[ctr].exclude_user= 1; 

		if (fds[ctr] < 0) {
			perror("Opening performance counter");
			goto error;
		}
    }
    va_end(args);

    return ret;

error:
	number = 0;
	if (attr) {
		free(attr);
		attr = NULL;
	}
	if (fds) {
		free(fds);
		fds = NULL;
	}
	return ret;
}

static int perf_open(void)
{
	struct perf_event_attr attr;

	memset(&attr, 0, sizeof (attr));

	attr.type = i915_type_id();
	if (attr.type == 0)
		return -ENOENT;
	attr.config = I915_PERF_ENERGY;

	attr.read_format = PERF_FORMAT_TOTAL_TIME_ENABLED;
	return perf_event_open(&attr, -1, 0, -1, 0);
}