Handle threads in PerfListener

teqwve · teqwve · commit eca9f7baf9a6 · 2019-07-30T22:00:30.000+02:00
diff --git a/src/perf/PerfListener.cc b/src/perf/PerfListener.cc
@@ -32,8 +32,12 @@ namespace perf {
 
 const Feature PerfListener::feature = Feature::PERF;
 
-PerfListener::PerfListener(uint64_t instructionCountLimit)
-        : perfFd_(-1), instructionCountLimit_(instructionCountLimit) {}
+PerfListener::PerfListener(
+        uint64_t instructionCountLimit,
+        uint64_t samplingFactor)
+        : instructionCountLimit_(instructionCountLimit)
+        , samplingFactor_{std::max<uint64_t>(1ULL, samplingFactor)}
+        , perfFd_{-1} {}
 
 PerfListener::~PerfListener() {
     // TODO: handle closing perfFd in move assignement / constructor as well
@@ -72,7 +76,6 @@ void PerfListener::onPostForkParent(pid_t childPid) {
     TRACE();
 
     childPid_ = childPid;
-    // TODO: fix this // What fix?
     struct perf_event_attr attrs {};
     memset(&attrs, 0, sizeof(attrs));
     attrs.type = PERF_TYPE_HARDWARE;
@@ -83,8 +86,9 @@ void PerfListener::onPostForkParent(pid_t childPid) {
     attrs.exclude_hv = 1;
     attrs.disabled = 1;
     attrs.enable_on_exec = 1;
+    attrs.inherit = 1;
     if (instructionCountLimit_ != 0) {
-        attrs.sample_period = instructionCountLimit_;
+        attrs.sample_period = instructionCountLimit_ / samplingFactor_;
         attrs.wakeup_events = 1;
     }
     // Apparently older (3.13) kernel versions doesn't support
diff --git a/src/perf/PerfListener.h b/src/perf/PerfListener.h
@@ -11,7 +11,7 @@ class PerfListener
         : public executor::ExecuteEventListener
         , public printer::OutputSource {
 public:
-    PerfListener(uint64_t instructionCountLimit);
+    PerfListener(uint64_t instructionCountLimit, uint64_t samplingFactor);
     ~PerfListener();
 
     void onPreFork() override;
@@ -25,8 +25,9 @@ class PerfListener
 private:
     uint64_t getInstructionsUsed();
 
+    const uint64_t instructionCountLimit_;
+    const uint64_t samplingFactor_;
     int perfFd_;
-    uint64_t instructionCountLimit_;
     pid_t childPid_{};
 
     // Barrier used for synchronization
diff --git a/src/s2japp/Application.cc b/src/s2japp/Application.cc
@@ -72,8 +72,15 @@ Application::ExitCode Application::handleRun() {
             settings_.threadsLimit >= 0);
 
     auto traceExecutor = createListener<tracer::TraceExecutor>();
-    auto perfListener =
-            createListener<perf::PerfListener>(settings_.instructionCountLimit);
+
+
+    uint64_t perfSamplingFactor = settings_.perfOversamplingFactor;
+    if (settings_.threadsLimit > 0) {
+        perfSamplingFactor *= static_cast<uint64_t>(settings_.threadsLimit);
+    }
+
+    auto perfListener = createListener<perf::PerfListener>(
+            settings_.instructionCountLimit, perfSamplingFactor);
     auto userNsListener = createListener<ns::UserNamespaceListener>();
     auto utsNsListener = createListener<ns::UTSNamespaceListener>();
     auto ipcNsListener = createListener<ns::IPCNamespaceListener>();
diff --git a/src/s2japp/ApplicationSettings.cc b/src/s2japp/ApplicationSettings.cc
@@ -1,6 +1,6 @@
-#include "ApplicationSettings.h"
 #include "ApplicationArguments.h"
 #include "ApplicationException.h"
+#include "ApplicationSettings.h"
 
 #include "common/Utils.h"
 #include "printer/AugmentedOIOutputBuilder.h"
@@ -312,6 +312,16 @@ ApplicationSettings::ApplicationSettings(int argc, const char* argv[])
                 "count",
                 cmd);
 
+        TCLAP::ValueArg<uint32_t> argPerfOversamplingFactor(
+                "w",
+                "perf-oversampling-factor",
+                "Additional number of perf wakupus during each base period "
+                "(wakeups number equals threads limit * oversampling factor)",
+                false,
+                2,
+                "factor",
+                cmd);
+
 
         TCLAP::UnlabeledValueArg<std::string> argProgramName(
                 "path", "Name of program to run", true, "", "path", cmd);
@@ -397,6 +407,7 @@ ApplicationSettings::ApplicationSettings(int argc, const char* argv[])
         suppressStderr = !argShowStderr.getValue();
         resultsFD = argResultsFD.getValue();
         threadsLimit = argThreadsLimit.getValue();
+        perfOversamplingFactor = argPerfOversamplingFactor.getValue();
     }
     catch (const TCLAP::ArgException& ex) {
         outputGenerator.failure(ex);
diff --git a/src/s2japp/ApplicationSettings.h b/src/s2japp/ApplicationSettings.h
@@ -45,6 +45,7 @@ struct ApplicationSettings : public ns::MountNamespaceListener::Settings {
 
     int resultsFD{};
     int threadsLimit{};
+    uint32_t perfOversamplingFactor{};
 
     std::string parsingError;
     std::string helpMessage;
