On Wed, 14 Aug 2024 19:34:26 GMT, Leonid Mesnik <lmes...@openjdk.org> wrote:
> The summary of the problem. Test is intermittently failing because can't get > expected field watch event. > The test is failing to get event in the 'setfmodw001b' thread with Xcomp only. > I verified that frame with the method 'run' of setfmodw001b is compiled when > line > 118: setfmodw001.setWatch(4); > is executed, however the thread is in interp_only mode. The watch events are > supported by interpreter only and just ignored by compiled code. > > The reason of failure is race beteween setting interp_only mode in line > > https://github.com/openjdk/jdk/blob/master/test/hotspot/jtreg/vmTestbase/nsk/jvmti/SetFieldModificationWatch/setfmodw001.java#L75 > > and calling method call_helper while > the method run() > https://github.com/openjdk/jdk/blob/master/test/hotspot/jtreg/vmTestbase/nsk/jvmti/SetFieldModificationWatch/setfmodw001.java#L116 > > in newly created thread 'setfmodw001b' is invoked. > > The javaCalls:call are used to invoke methods from hotspot, so it might be > rare issues. But still, synchronization might be improved. > The > void JavaCalls::call_helper(JavaValue* result, const methodHandle& method, > JavaCallArguments* args, TRAPS) > > checks if interp_only mode is set and use 'Method::from_interpreted_entry()' > if not. However the interp_only might be set later before compiled method is > called (or enter first safe point?). This might happens in safepoint during > transition via handshake. > So the running thread is in interp_only mode however the run() method is > compiled and executed already and never going to be deoptimized. > > The additional setWatch calls don't try to deptimize method that are already > in interp_only mode. > > BTW, the when JVMCI is enabled and verified adapter exists it is also will be > loaded even in interp_only mode set. There is not race here, just ignoring of > interp_only mode. > > I run failing test with Xcomp ~1000 times and tiers1-5. I was able to reproduce failure in "-Xcomp +ZGC'' once in a couple of several hundred runs. Not reproduced anymore with thousands of executions. Run tier1-5 to ensure all svc tests passed. There is no easy way to develop regression test. It should provoke call_helper for compiled method with breakpoint and setting the only single setWatch. Not sure it can find anything in reasonable time. The call_helper is used only to call methods from hotspot directly, like <clinit> , run() for Thread.start and similar not very common methods. I think to review any other possible cases and and thread stack verification that check that interp_only thread don't contain compiled frames on the stack. So we could find similar issues using our testing. Need to find the good place to inject this self-check. There are also a couple of places for improvements in this events handling. They are not directly rtelated to this bug. Will file them separately. ------------- PR Comment: https://git.openjdk.org/jdk/pull/20587#issuecomment-2299317622
