Node.js leaks memory when creating classes in a loop.

[hashseed]

Steps to reproduce:

$ cat test.js
'use strict';
for (var i = 0;; ++i) {
  if (i % 100 == 0) gc();
  new class{};
}

$ node --expose-gc test.js & top -p $! && kill -9 $!  # use q to quit

Observe reserved memory slowly grow. This will eventually cause out-of-memory.

This was first observed here. Note that this does not reproduce with d8 (V8's test shell), where the memory use stabilizes fairly quickly.

[bounc3-paradise-on-e]

Thanks for filing this issue.

Just a small correction (to avoid possible confusion): if (i % 100 == 0) - there's no variable i. If we want to call gc every 100th iteration, we can use for (var i = 0;; i = -~i) instead of for (;;)

[hashseed]

Yes, you are right. I fixed the test.

[addaleax]

Any chance the RSS increase is caused by running the GC, rather than the new class part? That’s the only part I see where some Node-specific behaviour kicks in, and the RSS grows slows down if I decrease the GC frequency…

(edit: confirmed.)

[addaleax]

This could do the trick here:

diff --git a/src/node_perf.cc b/src/node_perf.cc
index 5a50223ed593..8aad8ca1a705 100644
--- a/src/node_perf.cc
+++ b/src/node_perf.cc
@@ -272,6 +272,8 @@ void MarkGarbageCollectionEnd(Isolate* isolate,
                               v8::GCCallbackFlags flags,
                               void* data) {
   Environment* env = static_cast<Environment*>(data);
+  if (!env->performance_state()->observers[NODE_PERFORMANCE_ENTRY_TYPE_GC])
+    return;
   GCPerformanceEntry* entry =
       new GCPerformanceEntry(env,
                              static_cast<PerformanceGCKind>(type),

[hashseed]

That might be right :)

[jasnell]

I'll open PR