Merge "Perfprofd: Add process name to sampling"

6 files changed, 124 insertions, 36 deletions

diff --git a/perfprofd/perf_data_converter.cc b/perfprofd/perf_data_converter.cc
index f9a245d6..9d35d337 100644
--- a/perfprofd/perf_data_converter.cc
+++ b/perfprofd/perf_data_converter.cc
@@ -84,8 +84,17 @@ RawPerfDataToAndroidPerfProfile(const string &perf_file,
   std::unique_ptr<wireless_android_play_playlog::AndroidPerfProfile> ret(
       new wireless_android_play_playlog::AndroidPerfProfile());
 
-  typedef map<string, BinaryProfile> ModuleProfileMap;
-  typedef map<string, ModuleProfileMap> ProgramProfileMap;
+  using ModuleProfileMap = std::map<string, BinaryProfile>;
+  using Program = std::pair<uint32_t /* index into process name table, or uint32_t max */,
+                            std::string /* program name = comm of thread */>;
+  using ProgramProfileMap = std::map<Program, ModuleProfileMap>;
+
+  struct ProcessNameTable {
+    std::vector<std::string> names;
+    std::unordered_map<std::string, uint32_t> index_lookup;
+  };
+  constexpr uint32_t kNoProcessNameTableEntry = std::numeric_limits<uint32_t>::max();
+  ProcessNameTable process_name_table;
 
   // Note: the callchain_count_map member in BinaryProfile contains
   // pointers into callchains owned by "parser" above, meaning
@@ -116,20 +125,38 @@ RawPerfDataToAndroidPerfProfile(const string &perf_file,
       continue;
     }
     string dso_name = event.dso_and_offset.dso_name();
-    string program_name = event.command();
-    const string kernel_name = "[kernel.kallsyms]";
-    if (android::base::StartsWith(dso_name, kernel_name)) {
-      dso_name = kernel_name;
-      if (program_name == "") {
-        program_name = "kernel";
+    Program program_id;
+    {
+      std::string program_name = event.command();
+      const std::string kernel_name = "[kernel.kallsyms]";
+      if (android::base::StartsWith(dso_name, kernel_name)) {
+        dso_name = kernel_name;
+        if (program_name == "") {
+          program_name = "kernel";
+        }
+      } else if (program_name == "") {
+        if (is_kernel_dso(dso_name)) {
+          program_name = "kernel";
+        } else {
+          program_name = "unknown_program";
+        }
       }
-    } else if (program_name == "") {
-      if (is_kernel_dso(dso_name)) {
-        program_name = "kernel";
-      } else {
-        program_name = "unknown_program";
+      std::string process_name = event.process_command();
+      uint32_t process_name_index = kNoProcessNameTableEntry;
+      if (!process_name.empty()) {
+        auto name_iter = process_name_table.index_lookup.find(process_name);
+        if (name_iter == process_name_table.index_lookup.end()) {
+          process_name_index = process_name_table.names.size();
+          process_name_table.index_lookup.emplace(process_name, process_name_index);
+          process_name_table.names.push_back(process_name);
+        } else {
+          process_name_index = name_iter->second;
+        }
       }
+      program_id = std::make_pair(process_name_index, program_name);
     }
+    ModuleProfileMap& module_profile_map = name_profile_map[program_id];
+
     total_samples++;
     // We expect to see either all callchain events, all branch stack
     // events, or all flat sample events, not a mix. For callchains,
@@ -140,14 +167,14 @@ RawPerfDataToAndroidPerfProfile(const string &perf_file,
       CHECK(!seen_branch_stack && "examining callchain");
       seen_callchain = true;
       const callchain *cc = &event.callchain;
-      name_profile_map[program_name][dso_name].callchain_count_map[cc]++;
+      module_profile_map[dso_name].callchain_count_map[cc]++;
     } else if (!event.branch_stack.empty()) {
       CHECK(!seen_callchain && "examining branch stack");
       seen_branch_stack = true;
-      name_profile_map[program_name][dso_name].address_count_map[
+      module_profile_map[dso_name].address_count_map[
           event.dso_and_offset.offset()]++;
     } else {
-      name_profile_map[program_name][dso_name].address_count_map[
+      module_profile_map[dso_name].address_count_map[
           event.dso_and_offset.offset()]++;
     }
     for (size_t i = 1; i < event.branch_stack.size(); i++) {
@@ -160,7 +187,7 @@ RawPerfDataToAndroidPerfProfile(const string &perf_file,
           LOG(WARNING) << "Bogus LBR data: " << start << "->" << end;
           continue;
         }
-        name_profile_map[program_name][dso_name].range_count_map[
+        module_profile_map[dso_name].range_count_map[
             RangeTarget(start, end, to)]++;
       }
     }
@@ -244,7 +271,11 @@ RawPerfDataToAndroidPerfProfile(const string &perf_file,
 
   for (const auto &program_profile : name_profile_map) {
     auto program = ret->add_programs();
-    program->set_name(program_profile.first);
+    const Program& program_id = program_profile.first;
+    program->set_name(program_id.second);
+    if (program_id.first != kNoProcessNameTableEntry) {
+      program->set_process_name_id(program_id.first);
+    }
     for (const auto &module_profile : program_profile.second) {
       ModuleData& module_data = name_data_map[module_profile.first];
       int32 module_id = module_data.index;
@@ -289,6 +320,13 @@ RawPerfDataToAndroidPerfProfile(const string &perf_file,
     }
   }
 
+  if (!process_name_table.names.empty()) {
+    wireless_android_play_playlog::ProcessNames* process_names = ret->mutable_process_names();
+    for (const std::string& name : process_name_table.names) {
+      process_names->add_name(name);
+    }
+  }
+
   return ret.release();
 }
 
diff --git a/perfprofd/perf_profile.proto b/perfprofd/perf_profile.proto
index 8981eadc..65c9c39a 100644
--- a/perfprofd/perf_profile.proto
+++ b/perfprofd/perf_profile.proto
@@ -63,6 +63,11 @@ message LoadModuleSamples {
   repeated RangeSample range_samples = 3;
 }
 
+// A table of program names.
+message ProcessNames {
+  repeated string name = 1;
+}
+
 // All samples for a program.
 message ProgramSamples {
   // Name of the program.
@@ -70,6 +75,9 @@ message ProgramSamples {
 
   // Load module profiles.
   repeated LoadModuleSamples modules = 2;
+
+  // Index into ProcessNames for the name of the process.
+  optional uint32 process_name_id = 3;
 }
 
 // A compressed representation of a perf profile, which contains samples from
@@ -95,6 +103,9 @@ message AndroidPerfProfile {
   // List of all load modules.
   repeated LoadModule load_modules = 4;
 
+  // Table of process names.
+  optional ProcessNames process_names = 11;
+
   // is device screen on at point when profile is collected?
   optional bool display_on = 5;
 
diff --git a/perfprofd/quipper/perf_parser.cc b/perfprofd/quipper/perf_parser.cc
index c9ec1891..c529d25b 100644
--- a/perfprofd/quipper/perf_parser.cc
+++ b/perfprofd/quipper/perf_parser.cc
@@ -265,6 +265,13 @@ bool PerfParser::MapSampleEvent(ParsedEvent* parsed_event) {
   if (comm_iter != pidtid_to_comm_map_.end()) {
     parsed_event->set_command(comm_iter->second);
   }
+  {
+    PidTid pidtid_main = std::make_pair(sample_info.pid, sample_info.pid);
+    const auto main_iter = pidtid_to_comm_map_.find(pidtid_main);
+    if (main_iter != pidtid_to_comm_map_.end()) {
+      parsed_event->set_process_command(main_iter->second);
+    }
+  }
 
   const uint64_t unmapped_event_ip = sample_info.ip;
 
diff --git a/perfprofd/quipper/perf_parser.h b/perfprofd/quipper/perf_parser.h
index bb66de20..28644832 100644
--- a/perfprofd/quipper/perf_parser.h
+++ b/perfprofd/quipper/perf_parser.h
@@ -50,6 +50,9 @@ struct ParsedEvent {
   // Command associated with this sample.
   const string* command_;
 
+  // Process' name (command where pid = tid for this thread's pid).
+  const string* process_command_;
+
   // Accessor for command string.
   const string command() const {
     if (command_)
@@ -61,6 +64,17 @@ struct ParsedEvent {
     command_ = command;
   }
 
+  // Accessor for process_command string.
+  const string process_command() const {
+    if (process_command_)
+      return *process_command_;
+    return string();
+  }
+
+  void set_process_command(const string* command) {
+    process_command_ = command;
+  }
+
   // A struct that contains a DSO + offset pair.
   struct DSOAndOffset {
     const DSOInfo* dso_info_;
diff --git a/perfprofd/scripts/perf_proto_stack.py b/perfprofd/scripts/perf_proto_stack.py
index 4a576b47..03693589 100644
--- a/perfprofd/scripts/perf_proto_stack.py
+++ b/perfprofd/scripts/perf_proto_stack.py
@@ -47,10 +47,13 @@ def indent(txt, stops = 1):
     return '\n'.join('  ' * stops + line for line in txt.splitlines())
 
 
-def print_samples(module_list, programs, counters):
+def print_samples(module_list, programs, process_names, counters):
     print 'Samples:'
     for program in programs:
-        print indent(program.name, 1)
+        process_name = '?'
+        if program.HasField('process_name_id'):
+            process_name = process_names[program.process_name_id]
+        print indent('%s (%s)' % (program.name, process_name), 1)
         for module in program.modules:
             if module.HasField('load_module_id'):
                 module_descr = module_list[module.load_module_id]
@@ -118,6 +121,17 @@ def print_modules(module_list):
         for symbol in module.symbol:
             print indent(symbol, 3)
 
-print_samples(module_list, profile.programs, counters)
+def print_process_names(process_names):
+    print 'Processes:'
+    for proc in process_names:
+        print indent(proc, 1)
+
+if profile.HasField('process_names'):
+    process_names = profile.process_names.name
+else:
+    process_names = []
+
+print_samples(module_list, profile.programs, process_names, counters)
 print_modules(module_list)
 print_histogram(counters, 100)
+print_process_names(process_names)
diff --git a/perfprofd/tests/perfprofd_test.cc b/perfprofd/tests/perfprofd_test.cc
index 1ee1f777..599bb501 100644
--- a/perfprofd/tests/perfprofd_test.cc
+++ b/perfprofd/tests/perfprofd_test.cc
@@ -90,22 +90,26 @@ static std::string JoinTestLog(const char* delimiter) {
 
 class PerfProfdTest : public testing::Test {
  protected:
-  virtual void SetUp() {
+  void SetUp() override {
     InitTestLog();
     android::base::SetLogger(TestLogFunction);
     create_dirs();
   }
 
-  virtual void TearDown() {
+  void TearDown() override {
     android::base::SetLogger(android::base::StderrLogger);
     ClearTestLog();
 
     // TODO: proper management of test files. For now, use old system() code.
-    for (const auto dir : { &dest_dir, &conf_dir }) {
-      std::string cmd("rm -rf ");
-      cmd += *dir;
-      int ret = system(cmd.c_str());
-      CHECK_EQ(0, ret);
+    if (!HasFailure()) {
+      for (const auto dir : { &dest_dir, &conf_dir }) {
+        std::string cmd("rm -rf ");
+        cmd += *dir;
+        int ret = system(cmd.c_str());
+        CHECK_EQ(0, ret);
+      }
+    } else {
+      std::cerr << "Failed test: conf_dir=" << conf_dir << " dest_dir=" << dest_dir;
     }
   }
 
@@ -624,8 +628,8 @@ TEST_F(PerfProfdTest, BasicRunWithCannedPerf)
                      "BasicRunWithCannedPerf",
                      encodedProfile);
 
-  // Expect 45 programs
-  EXPECT_EQ(45, encodedProfile.programs_size());
+  // Expect 48 programs
+  EXPECT_EQ(48, encodedProfile.programs_size());
 
   // Check a couple of load modules
   { const auto &lm0 = encodedProfile.load_modules(0);
@@ -648,7 +652,7 @@ TEST_F(PerfProfdTest, BasicRunWithCannedPerf)
   }
 
   // Examine some of the samples now
-  { const auto &p1 = encodedProfile.programs(0);
+  { const auto &p1 = encodedProfile.programs(9);
     const auto &lm1 = p1.modules(0);
     std::string act_lm1 = encodedModuleSamplesToString(lm1);
     std::string sqact1 = squeezeWhite(act_lm1, "actual for lm1");
@@ -658,7 +662,7 @@ TEST_F(PerfProfdTest, BasicRunWithCannedPerf)
     std::string sqexp1 = squeezeWhite(expected_lm1, "expected_lm1");
     EXPECT_STREQ(sqexp1.c_str(), sqact1.c_str());
   }
-  { const auto &p1 = encodedProfile.programs(2);
+  { const auto &p1 = encodedProfile.programs(11);
     const auto &lm2 = p1.modules(0);
     std::string act_lm2 = encodedModuleSamplesToString(lm2);
     std::string sqact2 = squeezeWhite(act_lm2, "actual for lm2");
@@ -713,7 +717,7 @@ TEST_F(PerfProfdTest, BasicRunWithCannedPerfWithSymbolizer)
                      encodedProfile);
 
   // Expect 45 programs
-  EXPECT_EQ(45, encodedProfile.programs_size());
+  EXPECT_EQ(48, encodedProfile.programs_size());
 
   // Check a couple of load modules
   { const auto &lm0 = encodedProfile.load_modules(0);
@@ -738,7 +742,7 @@ TEST_F(PerfProfdTest, BasicRunWithCannedPerfWithSymbolizer)
   }
 
   // Examine some of the samples now
-  { const auto &p1 = encodedProfile.programs(0);
+  { const auto &p1 = encodedProfile.programs(9);
     const auto &lm1 = p1.modules(0);
     std::string act_lm1 = encodedModuleSamplesToString(lm1);
     std::string sqact1 = squeezeWhite(act_lm1, "actual for lm1");
@@ -748,7 +752,7 @@ TEST_F(PerfProfdTest, BasicRunWithCannedPerfWithSymbolizer)
     std::string sqexp1 = squeezeWhite(expected_lm1, "expected_lm1");
     EXPECT_STREQ(sqexp1.c_str(), sqact1.c_str());
   }
-  { const auto &p1 = encodedProfile.programs(2);
+  { const auto &p1 = encodedProfile.programs(11);
     const auto &lm2 = p1.modules(0);
     std::string act_lm2 = encodedModuleSamplesToString(lm2);
     std::string sqact2 = squeezeWhite(act_lm2, "actual for lm2");
@@ -817,7 +821,7 @@ TEST_F(PerfProfdTest, CallchainRunWithCannedPerf)
   }
 
   // Examine some of the samples now
-  { const auto &p0 = encodedProfile.programs(0);
+  { const auto &p0 = encodedProfile.programs(1);
     const auto &lm1 = p0.modules(0);
     std::string act_lm1 = encodedModuleSamplesToString(lm1);
     std::string sqact1 = squeezeWhite(act_lm1, "actual for lm1");