Hi,
Could I please get reviews for this SA fix? The issue only happens
intermittently and with -Xcomp. The new regression test reproduces the
issue somewhat reliably. I got 10/10 fails for unpatched, but I've seen
it pass as well.
When the issue happens, PCDesc's getScopeDecodeOffset() returs 0
(DebugInformationRecorder.SERIALIZED_NULL). The current SA code doesn't
handle this case and goes on and tries to read ScopeDesc from the
DebugInfoReadStream at the bogus offset. From then on, bad things
happen. A NPE in StackTrace could be one symptom.
The same code in hotspot deals with serialized null differently. It
doesn't read from the debug info stream, and manually sets up a
reasonable frame. Note decode_body is called from ScopeDesc's
constructor where decode_offset might have been set to 0:
void ScopeDesc::decode_body() {
if (decode_offset() == DebugInformationRecorder::serialized_null) {
// This is a sentinel record, which is only relevant to
// approximate queries. Decode a reasonable frame.
_sender_decode_offset = DebugInformationRecorder::serialized_null;
_method = _code->method();
_bci = InvocationEntryBci;
_locals_decode_offset = DebugInformationRecorder::serialized_null;
_expressions_decode_offset = DebugInformationRecorder::serialized_null;
_monitors_decode_offset = DebugInformationRecorder::serialized_null;
} else {
// decode header
DebugInfoReadStream* stream = stream_at(decode_offset());
_sender_decode_offset = stream->read_int();
_method = stream->read_method();
_bci = stream->read_bci();
// decode offsets for body and sender
_locals_decode_offset = stream->read_int();
_expressions_decode_offset = stream->read_int();
_monitors_decode_offset = stream->read_int();
}
}
The proposed patch handles serialized null scopes similar to the
hotspot side of things, by returning a null scope. CompiledVFrame
already deals with null scopes when in debugging mode.
Bug: https://bugs.openjdk.java.net/browse/JDK-8196969
webrev: http://cr.openjdk.java.net/~sgehwolf/webrevs/JDK-8196969/03/webrev/
Testing: tier 1 tests on Linux x86_64 (release/fastdebug). jdk-submit
and ran various reproducer tests including 1000 interations of the
added regression test. All pass.
Thoughts?
Thanks,
Severin
