Mercurial > hg > graal-compiler
diff src/share/vm/gc_implementation/g1/concurrentMark.cpp @ 20386:edb5f3b38aab
8054808: Bitmap verification sometimes fails after Full GC aborts concurrent mark.
Summary: The verification code that checked whether no bitmap mark had been found re-read HeapRegion::end() after the check on the bitmap. Concurrent humongous object allocation could have changed HeapRegion::end() in the meantime. Fix this by using the actual end of the region instead of HeapRegion::end() for comparison.
Reviewed-by: brutisso, jmasa
author | tschatzl |
---|---|
date | Thu, 28 Aug 2014 17:05:41 +0200 |
parents | 9337d0e7ea4f |
children | e5668dcf12e9 |
line wrap: on
line diff
--- a/src/share/vm/gc_implementation/g1/concurrentMark.cpp Wed Sep 03 08:44:47 2014 +0000 +++ b/src/share/vm/gc_implementation/g1/concurrentMark.cpp Thu Aug 28 17:05:41 2014 +0200 @@ -887,7 +887,16 @@ } virtual bool doHeapRegion(HeapRegion* r) { - return _bitmap->getNextMarkedWordAddress(r->bottom(), r->end()) != r->end(); + // This closure can be called concurrently to the mutator, so we must make sure + // that the result of the getNextMarkedWordAddress() call is compared to the + // value passed to it as limit to detect any found bits. + // We can use the region's orig_end() for the limit and the comparison value + // as it always contains the "real" end of the region that never changes and + // has no side effects. + // Due to the latter, there can also be no problem with the compiler generating + // reloads of the orig_end() call. + HeapWord* end = r->orig_end(); + return _bitmap->getNextMarkedWordAddress(r->bottom(), end) != end; } };