Fix BlinkGC triggering so it is more sensitive to PartitionAlloc only allocations

third_party/WebKit/Source/platform/heap/ThreadState.cpp

 /*
  * Copyright (C) 2013 Google Inc. All rights reserved.
  *
  * Redistribution and use in source and binary forms, with or without
  * modification, are permitted provided that the following conditions are
  * met:
  *
  *     * Redistributions of source code must retain the above copyright
  * notice, this list of conditions and the following disclaimer.
  *     * Redistributions in binary form must reproduce the above
@@ skipping 485 matching lines @@
     double growingRate = estimatedSize > 0 ? 1.0 * currentSize / estimatedSize : 100;
     TRACE_COUNTER1(TRACE_DISABLED_BY_DEFAULT("blink_gc"), "ThreadState::partitionAllocEstimatedSizeKB", std::min(estimatedSize / 1024, static_cast<size_t>(INT_MAX)));
     TRACE_COUNTER1(TRACE_DISABLED_BY_DEFAULT("blink_gc"), "ThreadState::partitionAllocGrowingRate", static_cast<int>(100 * growingRate));
     return growingRate;
 }
 
 // TODO(haraken): We should improve the GC heuristics. The heuristics affect
 // performance significantly.
 bool ThreadState::judgeGCThreshold(size_t totalMemorySizeThreshold, double heapGrowingRateThreshold)
 {
-    // If the allocated object size or the total memory size is small, don't trigger a GC.
-    if (m_heap->heapStats().allocatedObjectSize() < 100 * 1024 || totalMemorySize() < totalMemorySizeThreshold)
+    // If the total memory size is small, don't trigger a GC.
+    if (totalMemorySize() < totalMemorySizeThreshold)
         return false;
-    // If the growing rate of Oilpan's heap or PartitionAlloc is high enough,
-    // trigger a GC.
-#if PRINT_HEAP_STATS

[haraken, 2016/07/26 14:40:51]

Shall we keep this line?

-    dataLogF("heapGrowingRate=%.1lf, partitionAllocGrowingRate=%.1lf\n", heapGrowingRate(), partitionAllocGrowingRate());
-#endif
-    return heapGrowingRate() >= heapGrowingRateThreshold || partitionAllocGrowingRate() >= heapGrowingRateThreshold;
+
+    // If the growing rate of PartitionAlloc is high enough, trigger a GC.
+    if (partitionAllocGrowingRate() >= heapGrowingRateThreshold)

[haraken, 2016/07/26 14:40:51]

I'm a bit afraid of this change. After this change, won't we end up with
triggering Oilpan's GC when the renderer begins to run? (because at that point
the PartitionAlloc's growing rate would be high but Oilpan's memory usage is
quite low.)

[keishi, 2016/07/26 15:46:13]

Yeah. I'm running the perf tests right now but it probably has a wider effect
than we want.
The html file from the bug is a memory pressure situation caused by partition
alloc so this new PS uses MemoryPressureGC to trigger the GC. I think the
probability of a negative impact is lower with this.

+        return true;
+
+    // If the allocated object size is small, don't trigger a GC.
+    if (m_heap->heapStats().allocatedObjectSize() < 100 * 1024)
+        return false;
+    // If the growing rate of Oilpan's heap is high enough, trigger a GC.
+    return heapGrowingRate() >= heapGrowingRateThreshold;
 }
 
 bool ThreadState::shouldScheduleIdleGC()
 {
     if (gcState() != NoGCScheduled)
         return false;
     return judgeGCThreshold(1024 * 1024, 1.5);
 }
 
 bool ThreadState::shouldScheduleV8FollowupGC()
@@ skipping 1003 matching lines @@
     threadDump->AddScalar("dead_count", "objects", totalDeadCount);
     threadDump->AddScalar("live_size", "bytes", totalLiveSize);
     threadDump->AddScalar("dead_size", "bytes", totalDeadSize);
 
     base::trace_event::MemoryAllocatorDump* heapsDump = BlinkGCMemoryDumpProvider::instance()->createMemoryAllocatorDumpForCurrentGC(heapsDumpName);
     base::trace_event::MemoryAllocatorDump* classesDump = BlinkGCMemoryDumpProvider::instance()->createMemoryAllocatorDumpForCurrentGC(classesDumpName);
     BlinkGCMemoryDumpProvider::instance()->currentProcessMemoryDump()->AddOwnershipEdge(classesDump->guid(), heapsDump->guid());
 }
 
 } // namespace blink