Added isolate + thread high watermark tracking to Observatory

runtime/vm/zone.cc

Index: runtime/vm/zone.cc
diff --git a/runtime/vm/zone.cc b/runtime/vm/zone.cc
index 36e962260dd30ed433c67ff89812392ece46ec04..8dafe15f92211f040ad9706523b9bab7b968b687 100644
--- a/runtime/vm/zone.cc
+++ b/runtime/vm/zone.cc
@@ -44,6 +44,9 @@ class Zone::Segment {
 void Zone::Segment::DeleteSegmentList(Segment* head) {
   Segment* current = head;
   while (current != NULL) {
+    if (Thread::Current() != NULL) {
+      Thread::Current()->DecrementThreadMemoryUsage(current->size());
+    }

[siva, 2016/12/27 22:09:26]

Please hoist Thread:Current() outside the loop into a local variable
instead of accessing thread local storage at every iteration of the loop.

Thread* thread = Thread::Current();
while (...)
  if (thread != NULL) {
    thread->Decrement....(...);
  }
  ....
}

When is Thread::Current() null? (is it just unit tests?)

[bkonyi, 2016/12/28 00:09:58]

Done.

There's a case when running the unit tests where it seems a zone is allocated
and not associated with a thread, so we get a segfault when we delete segments
from that zone.

[siva, 2016/12/28 17:52:35]

It can be Null when the native IO isolate is allocating memory, see
Dart_ScopeAllocate in dart_api_impl.cc

We need to account for this in a special category called native IO isolate, the
recent OOM issue that we ran into during the Fuchsia port was from this entity.

[bkonyi, 2016/12/28 19:00:49]

I've added a TODO to handle the special case when the native IO isolate in a
separate change. 

     Segment* next = current->next();
 #ifdef DEBUG
     // Zap the entire current segment (including the header).
@@ -68,6 +71,9 @@ Zone::Segment* Zone::Segment::New(intptr_t size, Zone::Segment* next) {
 #endif
   result->next_ = next;
   result->size_ = size;
+  if (Thread::Current() != NULL) {
+    Thread::Current()->IncrementThreadMemoryUsage(size);
+  }
   return result;
 }