RFR: Workaround degen-traversal bugs with evac-assists enabled

[Roman Kennke]

Ok

Am 21. August 2018 11:15:48 MESZ schrieb Aleksey Shipilev <shade at redhat.com>:
>I have had a head-scratcher bug during backports testing, which boils
>down to bad degen-traversal
>and evac-assists interaction. Let's turn that off for traversal, until
>we come up with the proper fix:
>
>diff -r dbd712ba015c
>src/hotspot/share/gc/shenandoah/heuristics/shenandoahTraversalHeuristics.cpp
>---
>a/src/hotspot/share/gc/shenandoah/heuristics/shenandoahTraversalHeuristics.cpp
>     Mon Aug 20
>12:32:27 2018 -0400
>+++
>b/src/hotspot/share/gc/shenandoah/heuristics/shenandoahTraversalHeuristics.cpp
>     Tue Aug 21
>11:14:05 2018 +0200
>@@ -47,10 +47,32 @@
>   // Adjust class unloading settings only if globally enabled.
>   if (ClassUnloadingWithConcurrentMark) {
> SHENANDOAH_ERGO_OVERRIDE_DEFAULT(ShenandoahUnloadClassesFrequency, 1);
>   }
>
>+  // Workaround the bug in degen-traversal that evac assists expose.
>+  //
>+  // During traversal cycle, WB can evacuate something from region R1
>(CS) to R2 (R).
>+  // Then degen-traversal happens, drops the cset, and finishes up the
>fixups.
>+  // Then next cycle happens to put both R1 and R2 into CS, and then
>R2 evacuates to R3.
>+  // It creates the double forwarding for that object: R1 (CS) -> R2
>(CS) -> R3 (R).
>+  //
>+  // It is likely at that point that no references to R1 object are
>left after the degen,
>+  // so this double forwarding is not exposed. Unless we have evac
>assists, that touch
>+  // the adjacent objects while evacuating live objects from R1, step
>on "bad" R1 copy,
>+  // and fail the internal asserts when getting oop sizes to walk the
>heap, or touching
>+  // its fwdptrs. The same thing would probably happen if we do
>size-based iteration
>+  // somewhere else.
>+  //
>+  // AllocHumongousFragment test exposes it nicely, always running
>into degens.
>+  //
>+  // TODO: Fix this properly
>+  // There are two alternatives: fix it in degen so that it never
>leaves double forwarding,
>+  // or make sure we only use raw accessors in evac assist path when
>getting oop_size,
>+  // including all exotic shapes like instanceMirrorKlass, and
>touching fwdptrs. The second
>+  // option is partly done in jdk12, but not in earlier jdks.
>+  FLAG_SET_DEFAULT(ShenandoahEvacAssist, 0);
> }
>
> bool ShenandoahTraversalHeuristics::should_start_normal_gc() const {
>   return false;
> }
>diff -r dbd712ba015c src/hotspot/share/gc/shenandoah/shenandoahHeap.cpp
>--- a/src/hotspot/share/gc/shenandoah/shenandoahHeap.cpp        Mon Aug
>20 12:32:27 2018 -0400
>+++ b/src/hotspot/share/gc/shenandoah/shenandoahHeap.cpp        Tue Aug
>21 11:14:05 2018 +0200
>@@ -1888,10 +1888,14 @@
>// but enters it at different points, depending on which concurrent
>phase had
>     // degenerated.
>
>     case _degenerated_traversal:
>       {
>+        // Drop the collection set. Note: this leaves some already
>forwarded objects
>+        // behind, which may be problematic, see comments for
>ShenandoahEvacAssist
>+        // workarounds in ShenandoahTraversalHeuristics.
>+
>         ShenandoahHeapLocker locker(lock());
>         collection_set()->clear_current_index();
>         for (size_t i = 0; i < collection_set()->count(); i++) {
>           ShenandoahHeapRegion* r = collection_set()->next();
>           r->make_regular_bypass();
>
>
>Testing: tier3_gc_shenandoah
>
>Thanks,
>-Aleksey

-- 
Diese Nachricht wurde von meinem Android-Gerät mit K-9 Mail gesendet.