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);
}
}
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
mainloop(char **arg)
{
static uint64_t ovfl_count; /* static to avoid setjmp issue */
struct pollfd pollfds[1];
sigset_t bmask;
int go[2], ready[2];
size_t pgsz;
size_t map_size = 0;
pid_t pid;
int status, ret;
int i;
char buf;
if (pfm_initialize() != PFM_SUCCESS)
errx(1, "libpfm initialization failed\n");
pgsz = sysconf(_SC_PAGESIZE);
map_size = (options.mmap_pages+1)*pgsz;
/*
* does allocate fds
*/
ret = perf_setup_list_events(options.events, &fds, &num_fds);
if (ret || !num_fds)
errx(1, "cannot setup event list");
memset(pollfds, 0, sizeof(pollfds));
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\n");
/* 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));
}
*/
int tid;
pid = getpid();
tid = gettid();
printf("From C - pid : %d, tid : %d\n",pid, tid);
close(ready[1]);
close(go[0]);
if (read(ready[0], &buf, 1) == -1)
err(1, "unable to read child_ready_pipe");
close(ready[0]);
fds[0].fd = -1;
fds[0].hw.sample_period=10000;
if (!fds[0].hw.sample_period)
errx(1, "need to set sampling period or freq on first event, use :period= or :freq=");
for(i=0; i < num_fds; i++) {
if (i == 0) {
fds[i].hw.disabled = 1;
fds[i].hw.enable_on_exec = 1; /* start immediately */
} else
fds[i].hw.disabled = 0;
if (options.opt_inherit)
fds[i].hw.inherit = 1;
if (fds[i].hw.sample_period) {
/*
* set notification threshold to be halfway through the buffer
*/
fds[i].hw.wakeup_watermark = (options.mmap_pages*pgsz) / 2;
fds[i].hw.watermark = 1;
fds[i].hw.sample_type = PERF_SAMPLE_IP|PERF_SAMPLE_TID|PERF_SAMPLE_READ|PERF_SAMPLE_TIME|PERF_SAMPLE_PERIOD;
/*
* if we have more than one event, then record event identifier to help with parsing
*/
if (num_fds > 1)
fds[i].hw.sample_type |= PERF_SAMPLE_IDENTIFIER;
fprintf(options.output_file,"%s period=%"PRIu64" freq=%d\n", fds[i].name, fds[i].hw.sample_period, fds[i].hw.freq);
fds[i].hw.read_format = PERF_FORMAT_SCALE;
if (fds[i].hw.freq)
fds[i].hw.sample_type |= PERF_SAMPLE_PERIOD;
fds[i].hw.sample_period=10000;
if (options.mem_mode)
fds[i].hw.sample_type |= PERF_SAMPLE_WEIGHT | PERF_SAMPLE_DATA_SRC | PERF_SAMPLE_ADDR;
if (options.branch_mode) {
fds[i].hw.sample_type |= PERF_SAMPLE_BRANCH_STACK;
fds[i].hw.branch_sample_type = PERF_SAMPLE_BRANCH_ANY;
}
}
/*
* we are grouping the events, so there may be a limit
*/
fds[i].fd = perf_event_open(&fds[i].hw, tid+1, options.cpu, fds[0].fd, 0);
if (fds[i].fd == -1) {
if (fds[i].hw.precise_ip)
err(1, "cannot attach event %s: precise mode may not be supported", fds[i].name);
err(1, "cannot attach event %s", fds[i].name);
}
}
/*
* kernel adds the header page to the size of the mmapped region
*/
fds[0].buf = mmap(NULL, map_size, PROT_READ|PROT_WRITE, MAP_SHARED, fds[0].fd, 0);
if (fds[0].buf == MAP_FAILED)
err(1, "cannot mmap buffer");
/* does not include header page */
fds[0].pgmsk = (options.mmap_pages*pgsz)-1;
/*
* send samples for all events to first event's buffer
*/
for (i = 1; i < num_fds; i++) {
if (!fds[i].hw.sample_period)
continue;
ret = ioctl(fds[i].fd, PERF_EVENT_IOC_SET_OUTPUT, fds[0].fd);
if (ret)
err(1, "cannot redirect sampling output");
}
if (num_fds > 1 && fds[0].fd > -1) {
for(i = 0; i < num_fds; i++) {
/*
* read the event identifier using ioctl
* new method replaced the trick with PERF_FORMAT_GROUP + PERF_FORMAT_ID + read()
*/
ret = ioctl(fds[i].fd, PERF_EVENT_IOC_ID, &fds[i].id);
if (ret == -1)
err(1, "cannot read ID");
fprintf(options.output_file,"ID %"PRIu64" %s\n", fds[i].id, fds[i].name);
}
}
pollfds[0].fd = fds[0].fd;
pollfds[0].events = POLLIN;
for(i=0; i < num_fds; i++) {
ret = ioctl(fds[i].fd, PERF_EVENT_IOC_ENABLE, 0);
if (ret)
err(1, "cannot enable event %s\n", fds[i].name);
}
signal(SIGCHLD, cld_handler);
close(go[1]);
if (setjmp(jbuf) == 1)
goto terminate_session;
sigemptyset(&bmask);
sigaddset(&bmask, SIGCHLD);
/*
* core loop
*/
for(;;) {
ret = poll(pollfds, 1, -1);
if (ret < 0 && errno == EINTR)
break;
ovfl_count++;
ret = sigprocmask(SIG_SETMASK, &bmask, NULL);
if (ret)
err(1, "setmask");
process_smpl_buf(&fds[0]);
ret = sigprocmask(SIG_UNBLOCK, &bmask, NULL);
if (ret)
err(1, "unblock");
}
printf("How was you day??\n");
terminate_session:
/*
* cleanup child
*/
wait4(pid, &status, 0, NULL);
for(i=0; i < num_fds; i++)
close(fds[i].fd);
/* check for partial event buffer */
process_smpl_buf(&fds[0]);
munmap(fds[0].buf, map_size);
perf_free_fds(fds, num_fds);
fprintf(options.output_file,
"%"PRIu64" samples collected in %"PRIu64" poll events, %"PRIu64" lost samples\n",
collected_samples,
ovfl_count, lost_samples);
/* free libpfm resources cleanly */
pfm_terminate();
fclose(options.output_file);
return 0;
}
int
setup_cpu(int cpu, int fd)
{
uint64_t *val;
int ret, flags;
int i;
/*
* does allocate fds
*/
ret = perf_setup_list_events(options.events, &fds, &num_fds);
if (ret || !num_fds)
errx(1, "cannot setup event list");
if (!fds[0].hw.sample_period)
errx(1, "need to set sampling period or freq on first event, use :period= or :freq=");
fds[0].fd = -1;
for(i=0; i < num_fds; i++) {
fds[i].hw.disabled = !i; /* start immediately */
if (options.cgroup) {
flags = PERF_FLAG_PID_CGROUP;
} else {
flags = 0;
}
if (options.pin)
fds[i].hw.pinned = 1;
if (fds[i].hw.sample_period) {
/*
* set notification threshold to be halfway through the buffer
*/
if (fds[i].hw.sample_period) {
fds[i].hw.wakeup_watermark = (options.mmap_pages*pgsz) / 2;
fds[i].hw.watermark = 1;
}
fds[i].hw.sample_type = PERF_SAMPLE_IP|PERF_SAMPLE_TID|PERF_SAMPLE_READ|PERF_SAMPLE_TIME|PERF_SAMPLE_PERIOD|PERF_SAMPLE_STREAM_ID|PERF_SAMPLE_CPU;
printf("%s period=%"PRIu64" freq=%d\n", fds[i].name, fds[i].hw.sample_period, fds[i].hw.freq);
fds[i].hw.read_format = PERF_FORMAT_SCALE;
if (num_fds > 1)
fds[i].hw.read_format |= PERF_FORMAT_GROUP|PERF_FORMAT_ID;
if (fds[i].hw.freq)
fds[i].hw.sample_type |= PERF_SAMPLE_PERIOD;
}
fds[i].fd = perf_event_open(&fds[i].hw, -1, cpu, fds[0].fd, flags);
if (fds[i].fd == -1) {
if (fds[i].hw.precise_ip)
err(1, "cannot attach event %s: precise mode may not be supported", fds[i].name);
err(1, "cannot attach event %s", fds[i].name);
}
}
/*
* kernel adds the header page to the size of the mmapped region
*/
fds[0].buf = mmap(NULL, map_size, PROT_READ|PROT_WRITE, MAP_SHARED, fds[0].fd, 0);
if (fds[0].buf == MAP_FAILED)
err(1, "cannot mmap buffer");
/* does not include header page */
fds[0].pgmsk = (options.mmap_pages*pgsz)-1;
/*
* send samples for all events to first event's buffer
*/
for (i = 1; i < num_fds; i++) {
if (!fds[i].hw.sample_period)
continue;
ret = ioctl(fds[i].fd, PERF_EVENT_IOC_SET_OUTPUT, fds[0].fd);
if (ret)
err(1, "cannot redirect sampling output");
}
/*
* we are using PERF_FORMAT_GROUP, therefore the structure
* of val is as follows:
*
* { u64 nr;
* { u64 time_enabled; } && PERF_FORMAT_ENABLED
* { u64 time_running; } && PERF_FORMAT_RUNNING
* { u64 value;
* { u64 id; } && PERF_FORMAT_ID
* } cntr[nr];
* }
* We are skipping the first 3 values (nr, time_enabled, time_running)
* and then for each event we get a pair of values.
*/
if (num_fds > 1) {
sz = (3+2*num_fds)*sizeof(uint64_t);
val = malloc(sz);
if (!val)
err(1, "cannot allocated memory");
ret = read(fds[0].fd, val, sz);
if (ret == -1)
err(1, "cannot read id %zu", sizeof(val));
for(i=0; i < num_fds; i++) {
fds[i].id = val[2*i+1+3];
printf("%"PRIu64" %s\n", fds[i].id, fds[i].name);
}
free(val);
}
return 0;
}
static void
measure(void)
{
perf_event_desc_t *fds = NULL;
int num_fds = 0;
uint64_t values[3];
ssize_t n;
int i, ret;
int pr[2], pw[2];
pid_t pid;
char cc = '0';
ret = pfm_initialize();
if (ret != PFM_SUCCESS)
err(1, "cannot initialize libpfm");
if (options.cpu == -1) {
srandom(getpid());
options.cpu = random() % sysconf(_SC_NPROCESSORS_ONLN);
}
ret = pipe(pr);
if (ret)
err(1, "cannot create read pipe");
ret = pipe(pw);
if (ret)
err(1, "cannot create write pipe");
ret = perf_setup_list_events(options.events, &fds, &num_fds);
if (ret || !num_fds)
exit(1);
for(i=0; i < num_fds; i++) {
fds[i].hw.disabled = 1;
fds[i].hw.read_format = PERF_FORMAT_SCALE;
fds[i].fd = perf_event_open(&fds[i].hw, 0, -1, -1, 0);
if (fds[i].fd == -1)
err(1, "cannot open event %d", i);
}
/*
* Pin to CPU0, inherited by child process. That will enforce
* the ping-pionging and thus stress the PMU context switch
* which is what we want
*/
ret = pin_cpu(getpid(), options.cpu);
if (ret)
err(1, "cannot pin to CPU%d", options.cpu);
printf("Both processes pinned to CPU%d, running for %d seconds\n", options.cpu, options.delay);
/*
* create second process which is not monitoring at the moment
*/
switch(pid=fork()) {
case -1:
err(1, "cannot create child\n");
case 0:
/* do not inherit session fd */
for(i=0; i < num_fds; i++)
close(fds[i].fd);
/* pr[]: write master, read child */
/* pw[]: read master, write child */
close(pr[1]); close(pw[0]);
do_child(pr[0], pw[1]);
exit(1);
}
close(pr[0]);
close(pw[1]);
/*
* Let's roll now
*/
prctl(PR_TASK_PERF_EVENTS_ENABLE);
signal(SIGALRM, sig_handler);
alarm(options.delay);
/*
* ping pong loop
*/
while(!quit) {
n = write(pr[1], "c", 1);
if (n < 1)
err(1, "write failed");
n = read(pw[0], &cc, 1);
if (n < 1)
err(1, "read failed");
}
prctl(PR_TASK_PERF_EVENTS_DISABLE);
for(i=0; i < num_fds; i++) {
uint64_t val;
double ratio;
ret = read(fds[i].fd, values, sizeof(values));
if (ret == -1)
err(1,"pfm_read error");
if (ret != sizeof(values))
errx(1, "did not read correct amount %d", ret);
val = perf_scale(values);
ratio = perf_scale_ratio(values);
if (ratio == 1.0)
printf("%20"PRIu64" %s\n", val, fds[i].name);
else
if (ratio == 0.0)
printf("%20"PRIu64" %s (did not run: competing session)\n", val, fds[i].name);
else
printf("%20"PRIu64" %s (scaled from %.2f%% of time)\n", val, fds[i].name, ratio*100.0);
}
/*
* kill child process
*/
kill(SIGKILL, pid);
/*
* close pipes
*/
close(pr[1]);
close(pw[0]);
/*
* and destroy our session
*/
for(i=0; i < num_fds; i++)
close(fds[i].fd);
perf_free_fds(fds, num_fds);
/* free libpfm resources cleanly */
pfm_terminate();
}
int
mainloop(char **arg)
{
static uint64_t ovfl_count; /* static to avoid setjmp issue */
struct pollfd pollfds[1];
sigset_t bmask;
int go[2], ready[2];
uint64_t *val;
size_t sz, pgsz;
size_t map_size = 0;
pid_t pid;
int status, ret;
int i;
char buf;
if (pfm_initialize() != PFM_SUCCESS)
errx(1, "libpfm initialization failed\n");
pgsz = sysconf(_SC_PAGESIZE);
map_size = (options.mmap_pages+1)*pgsz;
/*
* does allocate fds
*/
ret = perf_setup_list_events(options.events, &fds, &num_fds);
if (ret || !num_fds)
errx(1, "cannot setup event list");
memset(pollfds, 0, sizeof(pollfds));
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\n");
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]);
fds[0].fd = -1;
if (!fds[0].hw.sample_period)
errx(1, "need to set sampling period or freq on first event, use :period= or :freq=");
for(i=0; i < num_fds; i++) {
if (i == 0) {
fds[i].hw.disabled = 1;
fds[i].hw.enable_on_exec = 1; /* start immediately */
} else
fds[i].hw.disabled = 0;
if (options.opt_inherit)
fds[i].hw.inherit = 1;
if (fds[i].hw.sample_period) {
/*
* set notification threshold to be halfway through the buffer
*/
fds[i].hw.wakeup_watermark = (options.mmap_pages*pgsz) / 2;
fds[i].hw.watermark = 1;
fds[i].hw.sample_type = PERF_SAMPLE_IP|PERF_SAMPLE_TID|PERF_SAMPLE_READ|PERF_SAMPLE_TIME|PERF_SAMPLE_PERIOD|PERF_SAMPLE_STREAM_ID;
fprintf(options.output_file,"%s period=%"PRIu64" freq=%d\n", fds[i].name, fds[i].hw.sample_period, fds[i].hw.freq);
fds[i].hw.read_format = PERF_FORMAT_SCALE;
if (num_fds > 1)
fds[i].hw.read_format |= PERF_FORMAT_GROUP|PERF_FORMAT_ID;
if (fds[i].hw.freq)
fds[i].hw.sample_type |= PERF_SAMPLE_PERIOD;
}
fds[i].fd = perf_event_open(&fds[i].hw, pid, options.cpu, fds[0].fd, 0);
if (fds[i].fd == -1) {
if (fds[i].hw.precise_ip)
err(1, "cannot attach event %s: precise mode may not be supported", fds[i].name);
err(1, "cannot attach event %s", fds[i].name);
}
}
/*
* kernel adds the header page to the size of the mmapped region
*/
fds[0].buf = mmap(NULL, map_size, PROT_READ|PROT_WRITE, MAP_SHARED, fds[0].fd, 0);
if (fds[0].buf == MAP_FAILED)
err(1, "cannot mmap buffer");
/* does not include header page */
fds[0].pgmsk = (options.mmap_pages*pgsz)-1;
/*
* send samples for all events to first event's buffer
*/
for (i = 1; i < num_fds; i++) {
if (!fds[i].hw.sample_period)
continue;
ret = ioctl(fds[i].fd, PERF_EVENT_IOC_SET_OUTPUT, fds[0].fd);
if (ret)
err(1, "cannot redirect sampling output");
}
/*
* we are using PERF_FORMAT_GROUP, therefore the structure
* of val is as follows:
*
* { u64 nr;
* { u64 time_enabled; } && PERF_FORMAT_ENABLED
* { u64 time_running; } && PERF_FORMAT_RUNNING
* { u64 value;
* { u64 id; } && PERF_FORMAT_ID
* } cntr[nr];
* We are skipping the first 3 values (nr, time_enabled, time_running)
* and then for each event we get a pair of values.
*/
if (num_fds > 1) {
sz = (3+2*num_fds)*sizeof(uint64_t);
val = malloc(sz);
if (!val)
err(1, "cannot allocate memory");
ret = read(fds[0].fd, val, sz);
if (ret == -1)
err(1, "cannot read id %zu", sizeof(val));
for(i=0; i < num_fds; i++) {
fds[i].id = val[2*i+1+3];
fprintf(options.output_file,"%"PRIu64" %s\n", fds[i].id, fds[i].name);
}
free(val);
}
pollfds[0].fd = fds[0].fd;
pollfds[0].events = POLLIN;
for(i=0; i < num_fds; i++) {
ret = ioctl(fds[i].fd, PERF_EVENT_IOC_ENABLE, 0);
if (ret)
err(1, "cannot enable event %s\n", fds[i].name);
}
signal(SIGCHLD, cld_handler);
close(go[1]);
if (setjmp(jbuf) == 1)
goto terminate_session;
sigemptyset(&bmask);
sigaddset(&bmask, SIGCHLD);
/*
* core loop
*/
for(;;) {
ret = poll(pollfds, 1, -1);
if (ret < 0 && errno == EINTR)
break;
ovfl_count++;
ret = sigprocmask(SIG_SETMASK, &bmask, NULL);
if (ret)
err(1, "setmask");
process_smpl_buf(&fds[0]);
ret = sigprocmask(SIG_UNBLOCK, &bmask, NULL);
if (ret)
err(1, "unblock");
}
terminate_session:
/*
* cleanup child
*/
wait4(pid, &status, 0, NULL);
for(i=0; i < num_fds; i++)
close(fds[i].fd);
/* check for partial event buffer */
process_smpl_buf(&fds[0]);
munmap(fds[0].buf, map_size);
perf_free_fds(fds, num_fds);
fprintf(options.output_file,
"%"PRIu64" samples collected in %"PRIu64" poll events, %"PRIu64" lost samples\n",
collected_samples,
ovfl_count, lost_samples);
/* free libpfm resources cleanly */
pfm_terminate();
fclose(options.output_file);
return 0;
}
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;
}
}
int
main(int argc, char **argv)
{
struct sigaction act;
uint64_t *val;
size_t sz, pgsz;
int ret, i;
setlocale(LC_ALL, "");
ret = pfm_initialize();
if (ret != PFM_SUCCESS)
errx(1, "Cannot initialize library: %s", pfm_strerror(ret));
pgsz = sysconf(_SC_PAGESIZE);
/*
* Install the signal handler (SIGIO)
* need SA_SIGINFO because we need the fd
* in the signal handler
*/
memset(&act, 0, sizeof(act));
act.sa_sigaction = sigio_handler;
act.sa_flags = SA_SIGINFO;
sigaction (SIGIO, &act, 0);
/*
* allocates fd for us
*/
ret = perf_setup_list_events("cycles,"
"instructions",
&fds, &num_fds);
if (ret || (num_fds == 0))
exit(1);
fds[0].fd = -1;
for(i=0; i < num_fds; i++) {
/* want a notification for every each added to the buffer */
fds[i].hw.disabled = !i;
if (!i) {
fds[i].hw.wakeup_events = 1;
fds[i].hw.sample_type = PERF_SAMPLE_IP|PERF_SAMPLE_READ|PERF_SAMPLE_PERIOD;
fds[i].hw.sample_period = SMPL_PERIOD;
/* read() returns event identification for signal handler */
fds[i].hw.read_format = PERF_FORMAT_GROUP|PERF_FORMAT_ID|PERF_FORMAT_SCALE;
}
fds[i].fd = perf_event_open(&fds[i].hw, 0, -1, fds[0].fd, 0);
if (fds[i].fd == -1)
err(1, "cannot attach event %s", fds[i].name);
}
sz = (3+2*num_fds)*sizeof(uint64_t);
val = malloc(sz);
if (!val)
err(1, "cannot allocated memory");
/*
* On overflow, the non lead events are stored in the sample.
* However we need some key to figure the order in which they
* were laid out in the buffer. The file descriptor does not
* work for this. Instead, we extract a unique ID for each event.
* That id will be part of the sample for each event value.
* Therefore we will be able to match value to events
*
* PERF_FORMAT_ID: returns unique 64-bit identifier in addition
* to event value.
*/
ret = read(fds[0].fd, val, sz);
if (ret == -1)
err(1, "cannot read id %zu", sizeof(val));
/*
* we are using PERF_FORMAT_GROUP, therefore the structure
* of val is as follows:
*
* { u64 nr;
* { u64 time_enabled; } && PERF_FORMAT_ENABLED
* { u64 time_running; } && PERF_FORMAT_RUNNING
* { u64 value;
* { u64 id; } && PERF_FORMAT_ID
* } cntr[nr];
* We are skipping the first 3 values (nr, time_enabled, time_running)
* and then for each event we get a pair of values.
*/
for(i=0; i < num_fds; i++) {
fds[i].id = val[2*i+1+3];
printf("%"PRIu64" %s\n", fds[i].id, fds[i].name);
}
fds[0].buf = mmap(NULL, (buffer_pages+1)*pgsz, PROT_READ|PROT_WRITE, MAP_SHARED, fds[0].fd, 0);
if (fds[0].buf == MAP_FAILED)
err(1, "cannot mmap buffer");
fds[0].pgmsk = (buffer_pages * pgsz) - 1;
/*
* setup asynchronous notification on the file descriptor
*/
ret = fcntl(fds[0].fd, F_SETFL, fcntl(fds[0].fd, F_GETFL, 0) | O_ASYNC);
if (ret == -1)
err(1, "cannot set ASYNC");
/*
* necessary if we want to get the file descriptor for
* which the SIGIO is sent in siginfo->si_fd.
* SA_SIGINFO in itself is not enough
*/
ret = fcntl(fds[0].fd, F_SETSIG, SIGIO);
if (ret == -1)
err(1, "cannot setsig");
/*
* get ownership of the descriptor
*/
ret = fcntl(fds[0].fd, F_SETOWN, getpid());
if (ret == -1)
err(1, "cannot setown");
/*
* enable the group for one period
*/
ret = ioctl(fds[0].fd, PERF_EVENT_IOC_REFRESH , 1);
if (ret == -1)
err(1, "cannot refresh");
busyloop();
ret = ioctl(fds[0].fd, PERF_EVENT_IOC_DISABLE, 1);
if (ret == -1)
err(1, "cannot disable");
/*
* destroy our session
*/
for(i=0; i < num_fds; i++)
close(fds[i].fd);
perf_free_fds(fds, num_fds);
free(val);
/* free libpfm resources cleanly */
pfm_terminate();
return 0;
}
