[BUG PROPOSAL]: C++ code that calls JNI_CreateJavaVM can be exited by java
david.holmes at oracle.com
Fri Aug 25 02:10:21 UTC 2017
On 24/08/2017 11:30 PM, Adam Farley8 wrote:
> Hi Alan, David, and Tom,
> First, thanks again for your efforts on this. As a new guy to OpenJDK
> contributions, it means a lot to see so much progress on this so
> quickly. :)
All I see is discussion :) Progress would be something else entirely.
> >On 24/08/2017 07:33, David Holmes wrote:
> >> Hi Adam,
> >> cc'ing hotspot runtime dev as runtime own JNI and the invocation API -
> >> and some of the problematic code resides in the VM.
> >Yeah, the hotspot mailing list would be a better place to discuss this
> >as there are several issues here and several places where HotSpot aborts
> >the process when initialization fails. It's a long standing issue (going
> >back 15+ years) that I think is partly because it's not easy to release
> >all resources and cleanup before CreateJavaVM returns with an error.
> According to the JNI spec, it is not possible (yet) to create a second VM
> in the same thread as the first.
> There is also a bug (dup'd against another bug I don't have the access for)
> which states that even a successful VM creation+destruction won't permit
> a second VM to be created.
> Both of these seem to imply that making a new VM after a failed VM-creation
> (in the same thread) is unsupported behaviour.
> So is it important to release all resources and cleanup, given that we
> be trying to create a new VM in this thread? By "important" I mean "more
> important than exiting with a return code and allowing the user's code
> to finish".
Okay, so if there is no intention of attempting to reload the jvm again,
I'm unclear what the purpose of the hosting process actually is. To me
it is either a customer launcher - in which case the exit calls are
"harmless" (and atexit handlers could be used if the process has its own
clean up) - or it's something multi-purpose part of which is to launch a
VM. In the latter case given the inability to reload a VM, and assuming
the process does not what it's java launching powers to be removed, then
the only real option is to filter out the problematic arguments and
either ignore them or exec a separate process to handle them.
> >> This specific case seems like a bug to me as the logic is assuming it
> >> is only ever called by a launcher which it is okay to terminate.
> >> Though to be honest the very existence of the "help" option seems to
> >> me somewhat misguided in a hosted-VM environment. That said, I see
> >> unified logging in 9 also added a terminating "help" option <sigh>.
> >The agent "help" option case is tricky and would likely need an update
> >to the JVM TI spec and the Agent_OnLoad return value.
> To clarify, the agent "help" option is only an example of this problem.
> There are 19 locations both within and without hotspot that call exit(0)
> directly, plus more places where exit is passed a variable that can be
> 0 (e.g. the aforementioned agent "help", which calls the forceExit function
> with an argument of 0, which calls exit(arg) in turn).
> I understand that your comment was intended as an effort to effect a fix
> for this specific instance of the problem. I wanted to make sure we kept
> sight of the wider problem, as ideally we'd come up with an ideal solution
> that could be applied to all cases.
The fact there are numerous potential process termination points in the
VM and JDK native code, is something we just have to live with. I'm only
considering these kind of "report and terminate" flags to be the problem
cases that should be handled better.
> My thought on this was a unique return code that tells the user's code
> that the VM is not in a usable state, but that no error has occurred. This
> should be a negative code (so the usual x<0 check will prevent the user's
> code from using the VM), but it shouldn't be one of the existing JNI codes;
> all of which seem to indicate either:
> a) The VM is fine and can be used (0).
> b) The VM is not fine because an error occurred (-1 to -6).
> Ideally we need a c) The VM is not fine, but no error has occurred.
It's somewhat debatable how to classify the case where you ask the VM to
load and then perform a one-off action that effectively succeeds but
leaves the VM unusable. Again ideally, to me, the VM would never do that
- such actions would occur as part of VM initialization, the VM would be
usable, but the launcher would do the termination because that is how
the flag is specified. But that is non-trivial to untangle.
> Or is there another solution to the exit(0) problem? Other than putting
> a copy of the rest of your code on the exit hook, I mean.
> >> Options processed by the VM will be recognized, while options
> >> processed by the Java launcher will not be. "-version", "-X", "-help"
> >> and numerous others are launcher options. Pure VM options are -XX
> >> options, but the VM also processes some -X flags and, as a result of
> >> jigsaw, now also processes a bunch of module-related flags that are
> >> simple --foo options.
> >Right because these options need to passed to CreateJavaVM as they are
> >used when initializing the VM. Using system properties would just repeat
> >the issues of past (e.g. java.class.path) and require documenting a slew
> >of system properties (which is complicated at repeating options).
> Unless stated otherwise above:
> IBM United Kingdom Limited - Registered in England and Wales with number
> Registered office: PO Box 41, North Harbour, Portsmouth, Hampshire PO6 3AU
More information about the core-libs-dev