static void
print_counts(perf_event_desc_t *fds, int num)
{
int i;
read_groups(fds, num);
for(i=0; i < num; i++) {
double ratio;
uint64_t val;
val = fds[i].value - fds[i].prev_value;
ratio = 0.0;
if (fds[i].enabled)
ratio = 1.0 * fds[i].running / fds[i].enabled;
/* separate groups */
if (perf_is_group_leader(fds, i))
putchar('\n');
if (fds[i].value < fds[i].prev_value) {
printf("inconsistent scaling %s (cur=%'"PRIu64" : prev=%'"PRIu64")\n", fds[i].name, fds[i].value, fds[i].prev_value);
continue;
}
printf("%'20"PRIu64" %s (%.2f%% scaling, ena=%'"PRIu64", run=%'"PRIu64")\n",
val,
fds[i].name,
(1.0-ratio)*100.0,
fds[i].enabled,
fds[i].running);
}
}
static void
print_counts(perf_event_desc_t *fds, int num)
{
double ratio;
uint64_t val, delta;
int i;
read_groups(fds, num);
for(i=0; i < num; i++) {
val = perf_scale(fds[i].values);
delta = perf_scale_delta(fds[i].values, fds[i].prev_values);
ratio = perf_scale_ratio(fds[i].values);
/* separate groups */
if (perf_is_group_leader(fds, i))
putchar('\n');
if (options.print)
printf("%'20"PRIu64" %'20"PRIu64" %s (%.2f%% scaling, ena=%'"PRIu64", run=%'"PRIu64")\n",
val,
delta,
fds[i].name,
(1.0-ratio)*100.0,
fds[i].values[1],
fds[i].values[2]);
else
printf("%'20"PRIu64" %s (%.2f%% scaling, ena=%'"PRIu64", run=%'"PRIu64")\n",
val,
fds[i].name,
(1.0-ratio)*100.0,
fds[i].values[1],
fds[i].values[2]);
fds[i].prev_values[0] = fds[i].values[0];
fds[i].prev_values[1] = fds[i].values[1];
fds[i].prev_values[2] = fds[i].values[2];
}
}
int
parent(char **arg)
{
perf_event_desc_t *fds = NULL;
int status, ret, i, num_fds = 0, grp, group_fd;
int ready[2], go[2];
char buf;
pid_t pid;
// output_file = fopen(file_name, "w");
//if (chroot("/home/pittdvs/cpu2006/benchspec/CPU2006/464.h264ref/run/run_base_ref_amd64-m64-gcc42-nn.0000"))
// err(1, "err on chroot()");
if (pfm_initialize() != PFM_SUCCESS)
errx(1, "libpfm initialization failed");
for (grp = 0; grp < options.num_groups; grp++) {
int ret;
ret = perf_setup_list_events(options.events[grp], &fds, &num_fds);
if (ret || !num_fds)
exit(1);
}
pid = options.pid;
if (!pid) {
ret = pipe(ready);
if (ret)
err(1, "cannot create pipe ready");
ret = pipe(go);
if (ret)
err(1, "cannot create pipe go");
/*
* Create the child task
*/
if ((pid=fork()) == -1)
err(1, "Cannot fork process");
/*
* and launch the child code
*
* The pipe is used to avoid a race condition
* between for() and exec(). We need the pid
* of the new tak but we want to start measuring
* at the first user level instruction. Thus we
* need to prevent exec until we have attached
* the events.
*/
if (pid == 0) {
close(ready[0]);
close(go[1]);
/*
* let the parent know we exist
*/
close(ready[1]);
if (read(go[0], &buf, 1) == -1)
err(1, "unable to read go_pipe");
exit(child(arg));
}
close(ready[1]);
close(go[0]);
if (read(ready[0], &buf, 1) == -1)
err(1, "unable to read child_ready_pipe");
close(ready[0]);
}
for(i=0; i < num_fds; i++) {
int is_group_leader; /* boolean */
is_group_leader = perf_is_group_leader(fds, i);
if (is_group_leader) {
/* this is the group leader */
group_fd = -1;
} else {
group_fd = fds[fds[i].group_leader].fd;
}
/*
* create leader disabled with enable_on-exec
*/
if (!options.pid) {
fds[i].hw.disabled = is_group_leader;
fds[i].hw.enable_on_exec = is_group_leader;
}
fds[i].hw.read_format = PERF_FORMAT_SCALE;
/* request timing information necessary for scaling counts */
if (is_group_leader && options.format_group)
fds[i].hw.read_format |= PERF_FORMAT_GROUP;
if (options.inherit)
fds[i].hw.inherit = 1;
if (options.pin && is_group_leader)
fds[i].hw.pinned = 1;
fds[i].fd = perf_event_open(&fds[i].hw, pid, -1, group_fd, 0);
if (fds[i].fd == -1) {
warn("cannot attach event%d %s", i, fds[i].name);
goto error;
}
}
if (!options.pid)
close(go[1]);
if (options.print) {
if (!options.pid) {
while(waitpid(pid, &status, WNOHANG) == 0 && quit == 0) {
//sleep(1);
//usleep(10000);
msleep(100);
//better_sleep(0.05); // 50ms
print_counts(fds, num_fds);
}
} else {
while(quit == 0) {
sleep(1);
print_counts(fds, num_fds);
}
}
} else {
if (!options.pid)
waitpid(pid, &status, 0);
else
pause();
print_counts(fds, num_fds);
}
for(i=0; i < num_fds; i++)
close(fds[i].fd);
free(fds);
/* free libpfm resources cleanly */
pfm_terminate();
// fclose(output_file);
return 0;
error:
free(fds);
if (!options.pid)
kill(SIGKILL, pid);
/* free libpfm resources cleanly */
pfm_terminate();
// fclose(output_file);
return -1;
}
void
setup_cpu(int cpu, int cfd)
{
perf_event_desc_t *fds = NULL;
int old_total, total = 0, num;
int i, j, n, ret, is_lead, group_fd;
unsigned long flags;
pid_t pid;
for(i=0, j=0; i < options.num_groups; i++) {
old_total = total;
ret = perf_setup_list_events(options.events[i], &fds, &total);
if (ret)
errx(1, "cannot setup events\n");
all_fds[cpu] = fds;
num = total - old_total;
options.nevents[i] = num;
for(n=0; n < num; n++, j++) {
is_lead = perf_is_group_leader(fds, j);
if (is_lead) {
fds[j].hw.disabled = 1;
group_fd = -1;
} else {
fds[j].hw.disabled = 0;
group_fd = fds[fds[j].group_leader].fd;
}
fds[j].hw.size = sizeof(struct perf_event_attr);
if (options.cgroup_name) {
flags = PERF_FLAG_PID_CGROUP;
pid = cfd;
//fds[j].hw.cgroup = 1;
//fds[j].hw.cgroup_fd = cfd;
} else {
flags = 0;
pid = -1;
}
if (options.pin && is_lead)
fds[j].hw.pinned = 1;
if (options.excl && is_lead)
fds[j].hw.exclusive = 1;
/* request timing information necessary for scaling counts */
fds[j].hw.read_format = PERF_FORMAT_SCALE;
fds[j].fd = perf_event_open(&fds[j].hw, pid, cpu, group_fd, flags);
if (fds[j].fd == -1) {
if (errno == EACCES)
err(1, "you need to be root to run system-wide on this machine");
warn("cannot attach event %s to CPU%ds, skipping it", fds[j].name, cpu);
goto error;
}
}
}
return;
error:
for (i=0; i < j; i++) {
close(fds[i].fd);
fds[i].fd = -1;
}
}