bool PerfSource::prepare() {
DynBuf printb;
DynBuf b1;
long long schedSwitchId;
long long cpuIdleId;
const uint64_t currTime = getTime();
// Reread cpuinfo since cores may have changed since startup
gSessionData->readCpuInfo();
if (0
|| !mMonitor.init()
|| !mUEvent.init()
|| !mMonitor.add(mUEvent.getFd())
|| (schedSwitchId = PerfDriver::getTracepointId(SCHED_SWITCH, &printb)) < 0
|| !sendTracepointFormat(currTime, &mBuffer, SCHED_SWITCH, &printb, &b1)
|| (cpuIdleId = PerfDriver::getTracepointId(CPU_IDLE, &printb)) < 0
|| !sendTracepointFormat(currTime, &mBuffer, CPU_IDLE, &printb, &b1)
// Only want RAW but not IP on sched_switch and don't want TID on SAMPLE_ID
|| !mCountersGroup.add(currTime, &mBuffer, 100/**/, PERF_TYPE_TRACEPOINT, schedSwitchId, 1, PERF_SAMPLE_RAW, PERF_GROUP_MMAP | PERF_GROUP_COMM | PERF_GROUP_TASK | PERF_GROUP_SAMPLE_ID_ALL | PERF_GROUP_PER_CPU)
|| !mIdleGroup.add(currTime, &mBuffer, 101/**/, PERF_TYPE_TRACEPOINT, cpuIdleId, 1, PERF_SAMPLE_RAW, PERF_GROUP_PER_CPU)
// Only want TID and IP but not RAW on timer
|| (gSessionData->mSampleRate > 0 && !gSessionData->mIsEBS && !mCountersGroup.add(currTime, &mBuffer, 102/**/, PERF_TYPE_SOFTWARE, PERF_COUNT_SW_CPU_CLOCK, 1000000000UL / gSessionData->mSampleRate, PERF_SAMPLE_TID | PERF_SAMPLE_IP, PERF_GROUP_PER_CPU))
|| !gSessionData->perf.enable(currTime, &mCountersGroup, &mBuffer)
|| 0) {
logg->logMessage("%s(%s:%i): perf setup failed, are you running Linux 3.4 or later?", __FUNCTION__, __FILE__, __LINE__);
return false;
}
for (int cpu = 0; cpu < gSessionData->mCores; ++cpu) {
const int result = mCountersGroup.prepareCPU(cpu, &mMonitor);
if ((result != PG_SUCCESS) && (result != PG_CPU_OFFLINE)) {
logg->logError(__FILE__, __LINE__, "PerfGroup::prepareCPU on mCountersGroup failed");
handleException();
}
}
for (int cpu = 0; cpu < gSessionData->mCores; ++cpu) {
const int result = mIdleGroup.prepareCPU(cpu, &mMonitor);
if ((result != PG_SUCCESS) && (result != PG_CPU_OFFLINE)) {
logg->logError(__FILE__, __LINE__, "PerfGroup::prepareCPU on mIdleGroup failed");
handleException();
}
}
int numEvents = 0;
for (int cpu = 0; cpu < gSessionData->mCores; ++cpu) {
numEvents += mCountersGroup.onlineCPU(currTime, cpu, false, &mBuffer);
}
for (int cpu = 0; cpu < gSessionData->mCores; ++cpu) {
numEvents += mIdleGroup.onlineCPU(currTime, cpu, false, &mBuffer);
}
if (numEvents <= 0) {
logg->logMessage("%s(%s:%i): PerfGroup::onlineCPU failed on all cores", __FUNCTION__, __FILE__, __LINE__);
return false;
}
// Send the summary right before the start so that the monotonic delta is close to the start time
if (!gSessionData->perf.summary(&mSummary)) {
logg->logError(__FILE__, __LINE__, "PerfDriver::summary failed", __FUNCTION__, __FILE__, __LINE__);
handleException();
}
// Start the timer thread to used to sync perf and monotonic raw times
pthread_t syncThread;
if (pthread_create(&syncThread, NULL, syncFunc, NULL)) {
logg->logError(__FILE__, __LINE__, "pthread_create failed", __FUNCTION__, __FILE__, __LINE__);
handleException();
}
struct sched_param param;
param.sched_priority = sched_get_priority_max(SCHED_FIFO);
if (pthread_setschedparam(syncThread, SCHED_FIFO | SCHED_RESET_ON_FORK, ¶m) != 0) {
logg->logError(__FILE__, __LINE__, "pthread_setschedparam failed");
handleException();
}
mBuffer.commit(currTime);
return true;
}
bool PerfSource::prepare() {
DynBuf printb;
DynBuf b1;
DynBuf b2;
DynBuf b3;
long long schedSwitchId;
// Reread cpuinfo since cores may have changed since startup
gSessionData->readCpuInfo();
if (0
|| !mMonitor.init()
|| !mUEvent.init()
|| !mMonitor.add(mUEvent.getFd())
|| (schedSwitchId = PerfDriver::getTracepointId(SCHED_SWITCH, &printb)) < 0
|| !sendTracepointFormat(&mBuffer, SCHED_SWITCH, &printb, &b1)
// Only want RAW but not IP on sched_switch and don't want TID on SAMPLE_ID
|| !mCountersGroup.add(&mBuffer, 100/**/, PERF_TYPE_TRACEPOINT, schedSwitchId, 1, PERF_SAMPLE_RAW, PERF_GROUP_MMAP | PERF_GROUP_COMM | PERF_GROUP_TASK | PERF_GROUP_SAMPLE_ID_ALL | PERF_GROUP_PER_CPU)
// Only want TID and IP but not RAW on timer
|| (gSessionData->mSampleRate > 0 && !gSessionData->mIsEBS && !mCountersGroup.add(&mBuffer, 99/**/, PERF_TYPE_SOFTWARE, PERF_COUNT_SW_CPU_CLOCK, 1000000000UL / gSessionData->mSampleRate, PERF_SAMPLE_TID | PERF_SAMPLE_IP, PERF_GROUP_PER_CPU))
|| !gSessionData->perf.enable(&mCountersGroup, &mBuffer)
|| 0) {
logg->logMessage("%s(%s:%i): perf setup failed, are you running Linux 3.4 or later?", __FUNCTION__, __FILE__, __LINE__);
return false;
}
if (!gSessionData->perf.summary(&mSummary)) {
logg->logMessage("%s(%s:%i): PerfDriver::summary failed", __FUNCTION__, __FILE__, __LINE__);
return false;
}
{
// Run prepareCPU in parallel as perf_event_open can take more than 1 sec in some cases
pthread_t threads[NR_CPUS];
PrepareParallelArgs args[NR_CPUS];
for (int cpu = 0; cpu < gSessionData->mCores; ++cpu) {
args[cpu].pg = &mCountersGroup;
args[cpu].cpu = cpu;
if (pthread_create(&threads[cpu], NULL, prepareParallel, &args[cpu]) != 0) {
logg->logMessage("%s(%s:%i): pthread_create failed", __FUNCTION__, __FILE__, __LINE__);
return false;
}
}
for (int cpu = 0; cpu < gSessionData->mCores; ++cpu) {
if (pthread_join(threads[cpu], NULL) != 0) {
logg->logMessage("%s(%s:%i): pthread_join failed", __FUNCTION__, __FILE__, __LINE__);
return false;
}
}
}
int numEvents = 0;
for (int cpu = 0; cpu < gSessionData->mCores; ++cpu) {
numEvents += mCountersGroup.onlineCPU(cpu, false, &mBuffer, &mMonitor);
}
if (numEvents <= 0) {
logg->logMessage("%s(%s:%i): PerfGroup::onlineCPU failed on all cores", __FUNCTION__, __FILE__, __LINE__);
return false;
}
// Start events before reading proc to avoid race conditions
if (!mCountersGroup.start()) {
logg->logMessage("%s(%s:%i): PerfGroup::start failed", __FUNCTION__, __FILE__, __LINE__);
return false;
}
if (!readProc(&mBuffer, true, &printb, &b1, &b2, &b3)) {
logg->logMessage("%s(%s:%i): readProc failed", __FUNCTION__, __FILE__, __LINE__);
return false;
}
mBuffer.commit(1);
return true;
}
