[vm, gc] Report RSS as timeline counter events, sampled before and after GC.

Catapult will display a nice graph for this.

TEST=view timeline
Change-Id: I6b6fdede463b37dbb10841cfa7c1e08c95e7cb62
Reviewed-on: https://dart-review.googlesource.com/c/sdk/+/427060
Commit-Queue: Ryan Macnak <rmacnak@google.com>
Reviewed-by: Slava Egorov <vegorov@google.com>

pkg/vm_service/test/get_vm_timeline_rpc_test.dart

@@ -111,6 +111,10 @@ void allEventsHaveIsolateNumber(List<TimelineEvent> events) {
       // Skip API category events which sometimes don't have an isolate.
       continue;
     }
+    if (event['name'] == 'RSS' && event['ph'] == 'C') {
+      // Skip RSS events, which don't have an isolate or isolate group.
+      continue;
+    }
     if (event['cat'] == 'Embedder' &&
         (event['name'] == 'DFE::ReadScript' ||
             event['name'] == 'CreateIsolateGroupAndSetupHelper' ||

runtime/observatory/tests/service/get_vm_timeline_rpc_test.dart

@@ -103,6 +103,10 @@ void allEventsHaveIsolateNumber(List events) {
       // Skip API category events which sometimes don't have an isolate.
       continue;
     }
+    if (event['name'] == 'RSS' && event['ph'] == 'C') {
+      // Skip RSS events, which don't have an isolate or isolate group.
+      continue;
+    }
     if (event['cat'] == 'Embedder' &&
         (event['name'] == 'DFE::ReadScript' ||
             event['name'] == 'CreateIsolateGroupAndSetupHelper')) {

runtime/vm/heap/heap.cc

@@ -1002,6 +1002,19 @@ void Heap::PrintMemoryUsageJSON(JSONObject* jsobj) const {
 }
 #endif  // PRODUCT
 
+static void RecordRSS() {
+#if defined(SUPPORT_TIMELINE)
+  TimelineEvent* event = Timeline::GetGCStream()->StartEvent();
+  if (event != nullptr) {
+    event->Counter("RSS");
+    event->SetNumArguments(1);
+    event->FormatArgument(0, "value", "%" Pd, OS::CurrentRSS());
+    event->ClearIsolateGroupId();  // This value is per-process.
+    event->Complete();
+  }
+#endif
+}
+
 void Heap::RecordBeforeGC(GCType type, GCReason reason) {
   stats_.num_++;
   stats_.type_ = type;
@@ -1010,6 +1023,7 @@ void Heap::RecordBeforeGC(GCType type, GCReason reason) {
   stats_.before_.new_ = new_space_.GetCurrentUsage();
   stats_.before_.old_ = old_space_.GetCurrentUsage();
   stats_.before_.store_buffer_ = isolate_group_->store_buffer()->Size();
+  RecordRSS();
 }
 
 void Heap::RecordAfterGC(GCType type) {
@@ -1025,6 +1039,7 @@ void Heap::RecordAfterGC(GCType type) {
   stats_.after_.new_ = new_space_.GetCurrentUsage();
   stats_.after_.old_ = old_space_.GetCurrentUsage();
   stats_.after_.store_buffer_ = isolate_group_->store_buffer()->Size();
+  RecordRSS();
 #ifndef PRODUCT
   // For now we'll emit the same GC events on all isolates.
   if (Service::gc_stream.enabled()) {

runtime/vm/os.h

@@ -68,6 +68,10 @@ class OS {
   // Returns number of available processor cores.
   static int NumberOfAvailableProcessors();
 
+  // Returns the current resident set size in bytes, or 0 if it could not be
+  // determined.
+  static uintptr_t CurrentRSS();
+
   // Sleep the currently executing thread for millis ms.
   static void Sleep(int64_t millis);
 

runtime/vm/os_android.cc

@@ -204,6 +204,23 @@ int OS::NumberOfAvailableProcessors() {
   return sysconf(_SC_NPROCESSORS_ONLN);
 }
 
+uintptr_t OS::CurrentRSS() {
+  // The second value in /proc/self/statm is the current RSS in pages.
+  // It is not possible to use getrusage() because the interested fields are not
+  // implemented by the linux kernel.
+  FILE* statm = fopen("/proc/self/statm", "r");
+  if (statm == nullptr) {
+    return 0;
+  }
+  int64_t current_rss_pages = 0;
+  int matches = fscanf(statm, "%*s%" Pd64 "", &current_rss_pages);
+  fclose(statm);
+  if (matches != 1) {
+    return 0;
+  }
+  return current_rss_pages * getpagesize();
+}
+
 void OS::Sleep(int64_t millis) {
   int64_t micros = millis * kMicrosecondsPerMillisecond;
   SleepMicros(micros);

runtime/vm/os_fuchsia.cc

@@ -488,6 +488,18 @@ int OS::NumberOfAvailableProcessors() {
   return sysconf(_SC_NPROCESSORS_CONF);
 }
 
+uintptr_t OS::CurrentRSS() {
+  zx_info_task_stats_t task_stats;
+  zx_handle_t process = zx_process_self();
+  zx_status_t status =
+      zx_object_get_info(process, ZX_INFO_TASK_STATS, &task_stats,
+                         sizeof(task_stats), nullptr, nullptr);
+  if (status != ZX_OK) {
+    return 0;
+  }
+  return task_stats.mem_private_bytes + task_stats.mem_shared_bytes;
+}
+
 void OS::Sleep(int64_t millis) {
   SleepMicros(millis * kMicrosecondsPerMillisecond);
 }

runtime/vm/os_linux.cc

@@ -521,6 +521,23 @@ int OS::NumberOfAvailableProcessors() {
   return sysconf(_SC_NPROCESSORS_ONLN);
 }
 
+uintptr_t OS::CurrentRSS() {
+  // The second value in /proc/self/statm is the current RSS in pages.
+  // It is not possible to use getrusage() because the interested fields are not
+  // implemented by the linux kernel.
+  FILE* statm = fopen("/proc/self/statm", "r");
+  if (statm == nullptr) {
+    return 0;
+  }
+  int64_t current_rss_pages = 0;
+  int matches = fscanf(statm, "%*s%" Pd64 "", &current_rss_pages);
+  fclose(statm);
+  if (matches != 1) {
+    return 0;
+  }
+  return current_rss_pages * getpagesize();
+}
+
 void OS::Sleep(int64_t millis) {
   int64_t micros = millis * kMicrosecondsPerMillisecond;
   SleepMicros(micros);

runtime/vm/os_macos.cc

@@ -123,6 +123,18 @@ int OS::NumberOfAvailableProcessors() {
   return sysconf(_SC_NPROCESSORS_ONLN);
 }
 
+uintptr_t OS::CurrentRSS() {
+  struct mach_task_basic_info info;
+  mach_msg_type_number_t infoCount = MACH_TASK_BASIC_INFO_COUNT;
+  kern_return_t result =
+      task_info(mach_task_self(), MACH_TASK_BASIC_INFO,
+                reinterpret_cast<task_info_t>(&info), &infoCount);
+  if (result != KERN_SUCCESS) {
+    return 0;
+  }
+  return info.resident_size;
+}
+
 void OS::Sleep(int64_t millis) {
   int64_t micros = millis * kMicrosecondsPerMillisecond;
   SleepMicros(micros);

runtime/vm/os_win.cc

@@ -238,6 +238,22 @@ int OS::NumberOfAvailableProcessors() {
   return info.dwNumberOfProcessors;
 }
 
+uintptr_t OS::CurrentRSS() {
+// Although the documentation at
+// https://docs.microsoft.com/en-us/windows/win32/api/psapi/nf-psapi-getprocessmemoryinfo
+// claims that GetProcessMemoryInfo is UWP compatible, it is actually not
+// hence this function cannot work when compiled in UWP mode.
+#ifdef DART_TARGET_OS_WINDOWS_UWP
+  return 0;
+#else
+  PROCESS_MEMORY_COUNTERS pmc;
+  if (!GetProcessMemoryInfo(GetCurrentProcess(), &pmc, sizeof(pmc))) {
+    return 0;
+  }
+  return pmc.WorkingSetSize;
+#endif
+}
+
 void OS::Sleep(int64_t millis) {
   ::Sleep(millis);
 }

runtime/vm/timeline.cc

@@ -1220,6 +1220,10 @@ bool TimelineEvent::HasIsolateGroupId() const {
   return isolate_group_id_ != ILLEGAL_ISOLATE_GROUP_ID;
 }
 
+void TimelineEvent::ClearIsolateGroupId() {
+  isolate_group_id_ = ILLEGAL_ISOLATE_GROUP_ID;
+}
+
 TimelineTrackMetadata::TimelineTrackMetadata(intptr_t pid,
                                              intptr_t tid,
                                              CStringUniquePtr&& track_name)

runtime/vm/timeline.h

@@ -457,6 +457,7 @@ class TimelineEvent {
 
   bool HasIsolateId() const;
   bool HasIsolateGroupId() const;
+  void ClearIsolateGroupId();
 
   // The lowest time value stored in this event.
   int64_t LowTime() const;