Hi Dan,
On 11/6/17 12:23 PM, Daniel D. Daugherty wrote:
On 11/3/17 8:25 PM, Chris Plummer wrote:
Hello,
Please review the following:
https://bugs.openjdk.java.net/browse/JDK-8059334
Wow! This bug was quite the adventure...
I know how to pick 'em, huh? :)
http://cr.openjdk.java.net/~cjplummer/8059334/webrev.00/webrev.open/
src/hotspot/share/interpreter/interpreterRuntime.cpp
L924: if (thread->is_interp_only_mode()) {
Perhaps: if (nm != NULL && thread->is_interp_only_mode()) {
Ok, I'll add the NULL check.
I kept trying to decide whether your new check only needs to
be inside the existing block:
L913: if (branch_bcp != NULL && nm != NULL) {
:
L923 }
but I finally convinced myself that you want to check a non-NULL
nm value in either branch_bcp code branch (NULL or non-NULL) so
where you put the fix is just fine.
Of course, the usual question we have to ask in these kinds of
races is what prevents the racy condition from asserting itself
right after the fixed code location. Thanks for including your
last sentence:
If we are not in "interp only" mode at this point (and start
executing
the compiled method) it should not be possible to enter "interp
only"
mode until we reach a safepoint at some later time, and at that
point
the method will be properly deopt so it can execute interpreted.
Nicely done bug hunt!
Thumbs up!
Thanks!
Chris
Dan
The CR is closed, so I'll try to explain the issue here. The very
short explanation is that the JVMTI test was enabling SINGLE STEP and
doing a PopFrame, but the test method managed to get compiled and
started executing compiled after the thread was put in "interp only"
mode (which should never happen) and before the PopFrame was
processed. The cause is a lack of a check for "interp only" mode in
the OSR related compilation policy code.
Details:
The test is testing JVMTI PopFrame support. The test thread has a
small method that sits in a tight loop. It will never exit. The main
thread enables SINGLE STEP on the test thread, and then does a
PopFrame on the test thread to force it out of the looping method.
When the test failed due to a time out, I noticed it was still stuck
in the small method, even though a PopFrame had been requested.
Further, I noticed the method was compiled, so there was no chance
the method would ever detect that it should do a PopFrame. Since
"interp only" mode for SINGLE STEP had been enabled, the method
should not be running compiled, so clearly something went wrong that
allowed it to compile and execute.
When SINGLE STEP is requested, JVMTI will deopt the topmost method
(actually the top 2), put the thread in "interp only" mode, and then
has checks to make sure the thread continues to execute interpreted.
To avoid compilation when a back branch tries to trigger one, there
is a check for "interp only" mode in SimpleThresholdPolicy::event().
If the thread is in "interp only" mode, it will prevent compilation.
SimpleThresholdPolicy::event() is called (indirectly) by
InterpreterRuntime::frequency_counter_overflow(), which is called
from the interpreter when the back branch threshold is reached.
After some debugging I noticed when the test timeout happens, "interp
only" mode is not yet enabled when
InterpreterRuntime::frequency_counter_overflow() is called, but is
enabled by the time InterpreterRuntime::frequency_counter_overflow()
has done the lookup of the nm. So there is a race here allowing the
thread to begin execution in a compiled method even though "interp
only" mode is enabled. I think the reason is because we safepoint
during the compilation, and this allows a SINGLE STEP request to be
processed, which enables "interp only" mode.
I should add that initially I only saw this bug with -Xcomp, but
eventually realized it was caused by disabling BackgroundCompilation.
That makes it much more likely that a SINGLE STEP request will come
in and be processed during the call to
InterpreterRuntime::frequency_counter_overflow() (because it will
block until the compilation completes).
I believe for the fix it is enough just to add an "interp only" mode
check in InterpreterRuntime::frequency_counter_overflow() after the
nm lookup, and set it nm to NULL if we are now in "interp only" mode.
If we are not in "interp only" mode at this point (and start
executing the compiled method) it should not be possible to enter
"interp only" mode until we reach a safepoint at some later time, and
at that point the method will be properly deopt so it can execute
interpreted.
thanks,
Chris
