add perf_sample_with_counters to stdlib

Android.bp

@@ -14801,6 +14801,7 @@ filegroup {
         "src/trace_processor/importers/proto/heap_graph_tracker.cc",
         "src/trace_processor/importers/proto/jit_tracker.cc",
         "src/trace_processor/importers/proto/metadata_module.cc",
+        "src/trace_processor/importers/proto/perf_counter_set_tracker.cc",
         "src/trace_processor/importers/proto/pigweed_detokenizer.cc",
         "src/trace_processor/importers/proto/pixel_modem_module.cc",
         "src/trace_processor/importers/proto/pixel_modem_parser.cc",
@@ -15399,6 +15400,7 @@ filegroup {
         "src/trace_processor/perfetto_sql/intrinsics/functions/layout_functions.cc",
         "src/trace_processor/perfetto_sql/intrinsics/functions/math.cc",
         "src/trace_processor/perfetto_sql/intrinsics/functions/package_lookup.cc",
+        "src/trace_processor/perfetto_sql/intrinsics/functions/perf_counter.cc",
         "src/trace_processor/perfetto_sql/intrinsics/functions/pprof_functions.cc",
         "src/trace_processor/perfetto_sql/intrinsics/functions/replace_numbers_function.cc",
         "src/trace_processor/perfetto_sql/intrinsics/functions/sqlite3_str_split.cc",
@@ -15732,6 +15734,7 @@ genrule {
         "src/trace_processor/perfetto_sql/stdlib/linux/memory/general.sql",
         "src/trace_processor/perfetto_sql/stdlib/linux/memory/high_watermark.sql",
         "src/trace_processor/perfetto_sql/stdlib/linux/memory/process.sql",
+        "src/trace_processor/perfetto_sql/stdlib/linux/perf/counters.sql",
         "src/trace_processor/perfetto_sql/stdlib/linux/perf/samples.sql",
         "src/trace_processor/perfetto_sql/stdlib/linux/perf/spe.sql",
         "src/trace_processor/perfetto_sql/stdlib/linux/threads.sql",

BUILD

@@ -2629,6 +2629,8 @@ perfetto_filegroup(
         "src/trace_processor/importers/proto/jit_tracker.h",
         "src/trace_processor/importers/proto/metadata_module.cc",
         "src/trace_processor/importers/proto/metadata_module.h",
+        "src/trace_processor/importers/proto/perf_counter_set_tracker.cc",
+        "src/trace_processor/importers/proto/perf_counter_set_tracker.h",
         "src/trace_processor/importers/proto/pigweed_detokenizer.cc",
         "src/trace_processor/importers/proto/pigweed_detokenizer.h",
         "src/trace_processor/importers/proto/pixel_modem_module.cc",
@@ -3169,6 +3171,8 @@ perfetto_filegroup(
         "src/trace_processor/perfetto_sql/intrinsics/functions/math.h",
         "src/trace_processor/perfetto_sql/intrinsics/functions/package_lookup.cc",
         "src/trace_processor/perfetto_sql/intrinsics/functions/package_lookup.h",
+        "src/trace_processor/perfetto_sql/intrinsics/functions/perf_counter.cc",
+        "src/trace_processor/perfetto_sql/intrinsics/functions/perf_counter.h",
         "src/trace_processor/perfetto_sql/intrinsics/functions/pprof_functions.cc",
         "src/trace_processor/perfetto_sql/intrinsics/functions/pprof_functions.h",
         "src/trace_processor/perfetto_sql/intrinsics/functions/replace_numbers_function.cc",
@@ -3616,6 +3620,7 @@ perfetto_filegroup(
 perfetto_filegroup(
     name = "src_trace_processor_perfetto_sql_stdlib_linux_perf_perf",
     srcs = [
+        "src/trace_processor/perfetto_sql/stdlib/linux/perf/counters.sql",
         "src/trace_processor/perfetto_sql/stdlib/linux/perf/samples.sql",
         "src/trace_processor/perfetto_sql/stdlib/linux/perf/spe.sql",
     ],

src/trace_processor/importers/perf/perf_counter.cc

@@ -23,15 +23,15 @@
 
 namespace perfetto::trace_processor::perf_importer {
 
-void PerfCounter::AddDelta(int64_t ts, double delta) {
+CounterId PerfCounter::AddDelta(int64_t ts, double delta) {
   last_count_ += delta;
-  counter_table_.Insert({ts, track_id_, last_count_});
+  return counter_table_.Insert({ts, track_id_, last_count_}).id;
 }
 
-void PerfCounter::AddCount(int64_t ts, double count) {
+CounterId PerfCounter::AddCount(int64_t ts, double count) {
   PERFETTO_CHECK(count >= last_count_);
   last_count_ = count;
-  counter_table_.Insert({ts, track_id_, last_count_});
+  return counter_table_.Insert({ts, track_id_, last_count_}).id;
 }
 
 }  // namespace perfetto::trace_processor::perf_importer

src/trace_processor/importers/perf/perf_counter.h

@@ -19,11 +19,14 @@
 
 #include <cstdint>
 
+#include "src/trace_processor/storage/trace_storage.h"
 #include "src/trace_processor/tables/counter_tables_py.h"
 #include "src/trace_processor/tables/track_tables_py.h"
 
 namespace perfetto::trace_processor::perf_importer {
 
+using trace_processor::CounterId;
+
 // Helper class to keep track of perf counters and convert delta values found in
 // perf files to absolute values needed for the perfetto counter table.
 class PerfCounter {
@@ -37,8 +40,8 @@ class PerfCounter {
 
   bool is_timebase() const { return is_timebase_; }
 
-  void AddDelta(int64_t ts, double delta);
-  void AddCount(int64_t ts, double count);
+  CounterId AddDelta(int64_t ts, double delta);
+  CounterId AddCount(int64_t ts, double count);
 
  private:
   tables::CounterTable& counter_table_;

src/trace_processor/importers/perf/record_parser.cc

@@ -26,6 +26,7 @@
 #include "perfetto/base/logging.h"
 #include "perfetto/base/status.h"
 #include "perfetto/ext/base/status_macros.h"
+#include "perfetto/ext/base/status_or.h"
 #include "perfetto/ext/base/string_view.h"
 #include "perfetto/public/compiler.h"
 #include "perfetto/trace_processor/ref_counted.h"
@@ -183,14 +184,29 @@ base::Status RecordParser::InternSample(Sample sample) {
   std::optional<CallsiteId> callsite_id = InternCallchain(
       upid, sample.callchain, sample.perf_invocation->needs_pc_adjustment());
 
+  // Update counters and create counter set.
+  ASSIGN_OR_RETURN(std::vector<CounterId> counter_ids, UpdateCounters(sample));
+
+  std::optional<uint32_t> counter_set_id;
+  if (!counter_ids.empty()) {
+    auto* table = context_->storage->mutable_perf_counter_set_table();
+    counter_set_id = static_cast<uint32_t>(table->row_count());
+    for (CounterId counter_id : counter_ids) {
+      tables::PerfCounterSetTable::Row row;
+      row.perf_counter_set_id = *counter_set_id;
+      row.counter_id = counter_id;
+      table->Insert(row);
+    }
+  }
+
   auto session_id = sample.attr->perf_session_id();
   context_->storage->mutable_perf_sample_table()->Insert(
       {sample.trace_ts, utid, sample.cpu,
        context_->storage->InternString(
            ProfilePacketUtils::StringifyCpuMode(sample.cpu_mode)),
-       callsite_id, std::nullopt, session_id});
+       callsite_id, std::nullopt, session_id, counter_set_id});
 
-  return UpdateCounters(sample);
+  return base::OkStatus();
 }
 
 std::optional<CallsiteId> RecordParser::InternCallchain(
@@ -318,7 +334,8 @@ UniquePid RecordParser::GetUpid(const CommonMmapRecordFields& fields) const {
   return *upid;
 }
 
-base::Status RecordParser::UpdateCounters(const Sample& sample) {
+base::StatusOr<std::vector<CounterId>> RecordParser::UpdateCounters(
+    const Sample& sample) {
   if (!sample.read_groups.empty()) {
     return UpdateCountersInReadGroups(sample);
   }
@@ -329,23 +346,28 @@ base::Status RecordParser::UpdateCounters(const Sample& sample) {
 
   uint64_t period = sample.period.has_value() ? *sample.period
                                               : *sample.attr->sample_period();
-  sample.attr->GetOrCreateCounter(sample.cpu)
-      .AddDelta(sample.trace_ts, static_cast<double>(period));
-  return base::OkStatus();
+  CounterId counter_id =
+      sample.attr->GetOrCreateCounter(sample.cpu)
+          .AddDelta(sample.trace_ts, static_cast<double>(period));
+  return std::vector<CounterId>{counter_id};
 }
 
-base::Status RecordParser::UpdateCountersInReadGroups(const Sample& sample) {
+base::StatusOr<std::vector<CounterId>> RecordParser::UpdateCountersInReadGroups(
+    const Sample& sample) {
+  std::vector<CounterId> counter_ids;
   for (const auto& entry : sample.read_groups) {
     RefPtr<PerfEventAttr> attr =
         sample.perf_invocation->FindAttrForEventId(*entry.event_id);
     if (PERFETTO_UNLIKELY(!attr)) {
       return base::ErrStatus("No perf_event_attr for id %" PRIu64,
                              *entry.event_id);
     }
-    attr->GetOrCreateCounter(sample.cpu)
-        .AddCount(sample.trace_ts, static_cast<double>(entry.value));
+    CounterId counter_id =
+        attr->GetOrCreateCounter(sample.cpu)
+            .AddCount(sample.trace_ts, static_cast<double>(entry.value));
+    counter_ids.push_back(counter_id);
   }
-  return base::OkStatus();
+  return counter_ids;
 }
 
 DummyMemoryMapping* RecordParser::GetDummyMapping(UniquePid upid) {

src/trace_processor/importers/perf/record_parser.h

@@ -23,6 +23,7 @@
 
 #include "perfetto/base/status.h"
 #include "perfetto/ext/base/flat_hash_map.h"
+#include "perfetto/ext/base/status_or.h"
 #include "src/trace_processor/importers/perf/mmap_record.h"
 #include "src/trace_processor/importers/perf/perf_tracker.h"
 #include "src/trace_processor/importers/perf/record.h"
@@ -59,8 +60,9 @@ class RecordParser : public TraceSorter::Sink<Record, RecordParser> {
 
   base::Status InternSample(Sample sample);
 
-  base::Status UpdateCounters(const Sample& sample);
-  static base::Status UpdateCountersInReadGroups(const Sample& sample);
+  base::StatusOr<std::vector<CounterId>> UpdateCounters(const Sample& sample);
+  static base::StatusOr<std::vector<CounterId>> UpdateCountersInReadGroups(
+      const Sample& sample);
 
   std::optional<CallsiteId> InternCallchain(
       UniquePid upid,

src/trace_processor/importers/proto/profile_module.cc

@@ -56,6 +56,25 @@
 
 namespace perfetto::trace_processor {
 
+namespace {
+
+// Adds a counter set containing the given counter IDs.
+// Returns the set ID that can be stored in PerfSampleTable.
+uint32_t AddCounterSet(TraceProcessorContext* context,
+                       const std::vector<CounterId>& counter_ids) {
+  auto* table = context->storage->mutable_perf_counter_set_table();
+  uint32_t set_id = static_cast<uint32_t>(table->row_count());
+  for (CounterId counter_id : counter_ids) {
+    tables::PerfCounterSetTable::Row row;
+    row.perf_counter_set_id = set_id;
+    row.counter_id = counter_id;
+    table->Insert(row);
+  }
+  return set_id;
+}
+
+}  // namespace
+
 using perfetto::protos::pbzero::TracePacket;
 using protozero::ConstBytes;
 
@@ -253,20 +272,35 @@ void ProfileModule::ParsePerfSample(
 
   // Populate the |perf_sample| table with everything except the recorded
   // counter values, which go to |counter|.
-  context_->event_tracker->PushCounter(
+  // Collect counter IDs for counter set association
+  std::vector<CounterId> counter_ids;
+
+  auto timebase_counter_id = context_->event_tracker->PushCounter(
       ts, static_cast<double>(sample.timebase_count()),
       sampling_stream.timebase_track_id);
+  if (timebase_counter_id) {
+    counter_ids.push_back(*timebase_counter_id);
+  }
 
   if (sample.has_follower_counts()) {
     auto track_it = sampling_stream.follower_track_ids.begin();
     auto track_end = sampling_stream.follower_track_ids.end();
     for (auto it = sample.follower_counts(); it && track_it != track_end;
          ++it, ++track_it) {
-      context_->event_tracker->PushCounter(ts, static_cast<double>(*it),
-                                           *track_it);
+      auto follower_counter_id = context_->event_tracker->PushCounter(
+          ts, static_cast<double>(*it), *track_it);
+      if (follower_counter_id) {
+        counter_ids.push_back(*follower_counter_id);
+      }
     }
   }
 
+  // Create counter set if we have any counter IDs
+  std::optional<uint32_t> counter_set_id;
+  if (!counter_ids.empty()) {
+    counter_set_id = AddCounterSet(context_, counter_ids);
+  }
+
   const UniqueTid utid =
       context_->process_tracker->UpdateThread(sample.tid(), sample.pid());
   const UniquePid upid =
@@ -295,9 +329,9 @@ void ProfileModule::ParsePerfSample(
     unwind_error_id = storage->InternString(
         ProfilePacketUtils::StringifyStackUnwindError(unwind_error));
   }
-  tables::PerfSampleTable::Row sample_row(ts, utid, sample.cpu(), cpu_mode_id,
-                                          cs_id, unwind_error_id,
-                                          sampling_stream.perf_session_id);
+  tables::PerfSampleTable::Row sample_row(
+      ts, utid, sample.cpu(), cpu_mode_id, cs_id, unwind_error_id,
+      sampling_stream.perf_session_id, counter_set_id);
   context_->storage->mutable_perf_sample_table()->Insert(sample_row);
 }
 

src/trace_processor/perfetto_sql/engine/perfetto_sql_engine.h

@@ -440,8 +440,8 @@ base::Status PerfettoSqlEngine::RegisterFunction(
   function_count_++;
   const char* name = args.name ? args.name : Function::kName;
   int argc = args.argc.has_value() ? args.argc.value() : Function::kArgCount;
-  return engine_->RegisterFunction(name, argc, Function::Step, ctx, nullptr,
-                                   args.deterministic);
+  return RegisterFunctionAndAddToRegistry(name, argc, Function::Step, ctx,
+                                          nullptr, args.deterministic);
 }
 
 template <typename Function>
@@ -451,7 +451,7 @@ base::Status PerfettoSqlEngine::RegisterFunction(
   function_count_++;
   const char* name = args.name ? args.name : Function::kName;
   int argc = args.argc.has_value() ? args.argc.value() : Function::kArgCount;
-  return engine_->RegisterFunction(
+  return RegisterFunctionAndAddToRegistry(
       name, argc, Function::Step, ctx.release(),
       [](void* ptr) {
         std::unique_ptr<typename Function::UserData>(

src/trace_processor/perfetto_sql/intrinsics/functions/BUILD.gn

@@ -46,6 +46,8 @@ source_set("functions") {
     "math.h",
     "package_lookup.cc",
     "package_lookup.h",
+    "perf_counter.cc",
+    "perf_counter.h",
     "pprof_functions.cc",
     "pprof_functions.h",
     "replace_numbers_function.cc",