Bug 90568 – [GTK] Enable parallel GC

• Component: JavaScriptCore
• Status: RESOLVED FIXED

[Andy Wingo]

The patch to be attached adds GTK+ to the parallel GC party.

[Andy Wingo]

Created attachment 150819 [details]
Patch

[WebKit Review Bot]

Comment on attachment 150819 [details]
Patch

Clearing flags on attachment: 150819

Committed r121869: <http://trac.webkit.org/changeset/121869>

[WebKit Review Bot]

All reviewed patches have been landed.  Closing bug.

[Zan Dobersek]

Parallel GC is currently causing flaky crashes in JavaScriptCore. See bug #90957, comment #22 there describes the nature of issues for the GTK port.

I believe the parallel GC should be disabled at least before branching out for the next series of stable releases if the crashes are not fixed by then.

[Martin Robinson]

One option is to disable this shortly after branching.

[Carlos Garcia Campos]

Should this be reopened?

[Martin Robinson]

Disabling it here: https://bugs.webkit.org/show_bug.cgi?id=94823

[Peng Xinchao]

Hello, GC experts.

Recently I met a GC crash when PARRALLEL GC is enabled. Crash is like:

#0  0x00007ffff72a1671 in isJSString (this=0x7fffeaec0cd8) at /home/oszi/WebKit/Source/JavaScriptCore/runtime/JSString.h:501
#1  visitChildren (this=0x7fffeaec0cd8) at /home/oszi/WebKit/Source/JavaScriptCore/heap/MarkStack.cpp:351
#2  JSC::SlotVisitor::drain (this=0x7fffeaec0cd8) at /home/oszi/WebKit/Source/JavaScriptCore/heap/MarkStack.cpp:405
#3  0x00007ffff72a19e4 in JSC::SlotVisitor::drainFromShared (this=0x7fffeaec0cd8, sharedDrainMode=JSC::SlotVisitor::MasterDrain) at /home/oszi/WebKit/Source/JavaScriptCore/heap/MarkStack.cpp:498
#4  0x00007ffff729dd38 in JSC::Heap::markRoots (this=0x7fffeaec0050, fullGC=<value optimized out>) at /home/oszi/WebKit/Source/JavaScriptCore/heap/Heap.cpp:555
#5  0x00007ffff729df8b in JSC::Heap::collect (this=0x7fffeaec0050, sweepToggle=JSC::Heap::DoNotSweep) at /home/oszi/WebKit/Source/JavaScriptCore/heap/Heap.cpp:717
#6  0x00007ffff72a410c in JSC::MarkedAllocator::allocateSlowCase (this=0x7fffeaec0158) at /home/oszi/WebKit/Source/JavaScriptCore/heap/MarkedAllocator.cpp:75
#7  0x00007ffff72e84ba in JSC::MarkedAllocator::allocate (exec=<value optimized out>) at /home/oszi/WebKit/Source/JavaScriptCore/heap/MarkedAllocator.h:77
#8  JSC::MarkedSpace::allocateWithDestructor (exec=<value optimized out>) at /home/oszi/WebKit/Source/JavaScriptCore/heap/MarkedSpace.h:191
#9  JSC::Heap::allocateWithDestructor (exec=<value optimized out>) at /home/oszi/WebKit/Source/JavaScriptCore/heap/Heap.h:362
#10 allocateCell<JSC::JSFinalObject> (exec=<value optimized out>) at /home/oszi/WebKit/Source/JavaScriptCore/runtime/JSCell.h:340
#11 JSC::JSFinalObject::create (exec=<value optimized out>) at /home/oszi/WebKit/Source/JavaScriptCore/runtime/JSObject.h:439
#12 constructEmptyObject (exec=<value optimized out>) at /home/oszi/WebKit/Source/JavaScriptCore/runtime/JSObject.h:515
#13 constructEmptyObject (exec=<value optimized out>) at /home/oszi/WebKit/Source/JavaScriptCore/runtime/JSGlobalObject.h:431
#14 constructEmptyObject (exec=<value optimized out>) at /home/oszi/WebKit/Source/JavaScriptCore/runtime/JSGlobalObject.h:436
#15 operationNewObject (exec=<value optimized out>) at /home/oszi/WebKit/Source/JavaScriptCore/dfg/DFGOperations.cpp:305
#16 0x00007fffaaf8880d in ?? ()
#17 0x0000000000000000 in ?? ()


I checked the code of JSCore GC.

void SlotVisitor::drain()
{
    StackStats::probe();
    ASSERT(m_isInParallelMode);

#if ENABLE(PARALLEL_GC)
    if (Options::numberOfGCMarkers() > 1) {
        while (!m_stack.isEmpty()) {
            m_stack.refill();
            for (unsigned countdown = Options::minimumNumberOfScansBetweenRebalance(); m_stack.canRemoveLast() && countdown--;)
                visitChildren(*this, m_stack.removeLast());
            donateKnownParallel();
        }

        mergeOpaqueRootsIfNecessary();
        return;
    }
#endif

    while (!m_stack.isEmpty()) {
        m_stack.refill();
        while (m_stack.canRemoveLast())
            visitChildren(*this, m_stack.removeLast());
    }
}

Why is m_shared.m_markingLock not added before m_stack.refill() and m_stack.removeLast(), just like inside the function void SlotVisitor::donateKnownParallel() ?

It seems that m_stack is operated unsafely when PARRELLEL GC is enabled.

What do you think?