Set HistogramEnumerationFunction even when Partitions are initialized

Source/wtf/Partitions.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 20 matching lines @@
 #include "config.h"
 #include "wtf/Partitions.h"
 
 #include "wtf/DefaultAllocator.h"
 #include "wtf/FastMalloc.h"
 #include "wtf/MainThread.h"
 
 namespace WTF {
 
 bool Partitions::s_initialized;
+int Partitions::s_lock = 0;
 
 PartitionAllocatorGeneric Partitions::m_fastMallocAllocator;
 PartitionAllocatorGeneric Partitions::m_bufferAllocator;
 SizeSpecificPartitionAllocator<3328> Partitions::m_objectModelAllocator;
 SizeSpecificPartitionAllocator<1024> Partitions::m_layoutAllocator;
 HistogramEnumerationFunction Partitions::m_histogramEnumeration = nullptr;
 
-void Partitions::initialize(HistogramEnumerationFunction histogramEnumeration)
+void Partitions::initialize()
 {
-    static int lock = 0;
     // Guard against two threads hitting here in parallel.
-    spinLockLock(&lock);
+    spinLockLock(&s_lock);
     if (!s_initialized) {
         m_fastMallocAllocator.init();
         m_bufferAllocator.init();
         m_objectModelAllocator.init();
         m_layoutAllocator.init();
-        m_histogramEnumeration = histogramEnumeration;
         s_initialized = true;
     }
-    spinLockUnlock(&lock);
+    spinLockUnlock(&s_lock);
+}
+
+void Partitions::setHistogramEnumeration(HistogramEnumerationFunction histogramEnumeration)
+{
+    spinLockLock(&s_lock);

[haraken, 2015/06/17 05:24:20]

I don't think this lock is needed. Partitions::setHistogramEnumeration can be
called only from the main thread. Also it is guaranteed that WTF::initialize()
is called before the m_histogramEnumeration gets used. Thus there is no
threading contention on the m_histogramEnumeration.

[bashi, 2015/06/17 05:25:27]

Hmm, the same can be said for initialize()? Can we remove the lock totally?

[haraken, 2015/06/17 05:29:37]

I think your understanding is correct. Any Blink thread won't start before the
main thread finishes WTF::initialize(), it doesn't happen that multiple threads
contend on s_initialized. (If that happens, it will already be detected as a
threading contention because Partitions::bufferPartition() is touching
s_initialized without acquiring a lock.) So I think it should be safe to remove
the lock.

[bashi, 2015/06/17 05:49:41]

Removed. We can't add ASSERT(isMainThread()) because isMainThread() doesn't work
at this point. (mainThreadIdendifier isn't initialized at this point)

+    ASSERT(!m_histogramEnumeration);
+    m_histogramEnumeration = histogramEnumeration;
+    spinLockUnlock(&s_lock);
 }
 
 void Partitions::shutdown()
 {
     // We could ASSERT here for a memory leak within the partition, but it leads
     // to very hard to diagnose ASSERTs, so it's best to leave leak checking for
     // the valgrind and heapcheck bots, which run without partitions.
     (void) m_layoutAllocator.shutdown();
     (void) m_objectModelAllocator.shutdown();
     (void) m_bufferAllocator.shutdown();
@@ skipping 38 matching lines @@
     // accessed only on the main thread.
     ASSERT(isMainThread());
 
     partitionDumpStatsGeneric(fastMallocPartition(), "fast_malloc_partition", partitionStatsDumper);
     partitionDumpStatsGeneric(bufferPartition(), "buffer_partition", partitionStatsDumper);
     partitionDumpStats(objectModelPartition(), "object_model_partition", partitionStatsDumper);
     partitionDumpStats(layoutPartition(), "layout_partition", partitionStatsDumper);
 }
 
 } // namespace WTF