-
Bug
-
Resolution: Fixed
-
P2
-
14, 15
-
b28
-
aarch64
| Issue | Fix Version | Assignee | Priority | Status | Resolution | Resolved In Build |
|---|---|---|---|---|---|---|
| JDK-8236216 | 15 | Nick Gasson | P2 | Resolved | Fixed | b02 |
I occasionally see this test crash on AArch64 with the following assertion failure:
# Internal Error (/home/ent-user/ci-scripts/jdk_src/src/hotspot/share/oops/compressedOops.inline.hpp:84), pid=85650, tid=85656
# assert(!is_null(v)) failed: narrow klass value can never be zero
See the attached full crash log.
Looks like there's a missing store-store barrier in parallal GC MoveAndUpdateShadowClosure::complete_region, as the stores from set_shadow_region() and mark_filled() can be reordered on a non-TSO processor.
I'll do some more testing and send out a patch.
# Internal Error (/home/ent-user/ci-scripts/jdk_src/src/hotspot/share/oops/compressedOops.inline.hpp:84), pid=85650, tid=85656
# assert(!is_null(v)) failed: narrow klass value can never be zero
See the attached full crash log.
Looks like there's a missing store-store barrier in parallal GC MoveAndUpdateShadowClosure::complete_region, as the stores from set_shadow_region() and mark_filled() can be reordered on a non-TSO processor.
I'll do some more testing and send out a patch.
- backported by
-
JDK-8236216 AArch64: Insufficient memory barriers in shadow region algorithm
-
- Resolved
-
- relates to
-
JDK-8220465 Use shadow regions for faster ParallelGC full GCs
-
- Resolved
-