RFR (S) 8043070: nmethod::verify_interrupt_point() shouldn't enter safepoint
igor.veresov at oracle.com
Fri Nov 17 08:52:12 UTC 2017
> On Nov 16, 2017, at 9:42 PM, dean.long at oracle.com wrote:
> On 11/16/17 6:32 PM, Igor Veresov wrote:
>>> On Nov 16, 2017, at 3:30 PM, dean.long at oracle.com wrote:
>>> Hi Igor. The changes look good.
>>> Thanks for not changing the values for the other states. It would have broken AOT :-)
>> Yes, we’d like in_use to be 0 so it could be in BSS, right?
> And because we compare against 0 right now:
> 3440: 48 8b 05 39 bb 21 00 mov 0x21bb39(%rip),%rax # 21ef80 <A.meth.state>
> 3447: 48 85 c0 test %rax,%rax
> 344a: 0f 85 20 b1 00 00 jne e570 <plt._aot_handle_wrong_method_stub>
Good point, yes.
>>> Are there any situations where make_in_use() might need a compiler/memory barrier?
>> I don’t think so, is_in_use() actually returns true for not_installed (now it’s _state <= in_use). So as far as other threads are concerned nothing changed.
>> There is an exotic possibility that the initial store of not_installed might be reordered with the store of in_use on non-TSO, but there is a mutex unlock between those, so I’m not really concerned about that.
>> Are you thinking about any other scenarios?
> I was thinking about make_in_use() happening before the verify() call, but that would mean an invalid optimization by a broken C++ compiler, right?
Yes, I think that would mean it doesn’t track memory effects correctly. No reordering should happen from the point of view of the same thread.
>>> I was thinking about an alternative approach, at least for this case. We could have NoSafepointVerififer always set _thread->_allow_safepoint_count (right now it doesn't for product builds), then verify_interrupt_point() could check that flag before taking the path that can safepoint. Or how about just using nmethodLocker as suggested in 8028001? Or does having a new state help solve other problems?
>> nmethodLocker approach seemed less tidy because I’d have to lock in nmethod::nmethod() and unlock in ciEnv::register_method().
>> We could probably introduce sort of an nmethodHandle that we’d return from new_nmethod() but then we’d have to do reference counting of sorts. Anyway, it felt like an additional state is less trouble.
> OK, a new state does look like less trouble than nmethodLocker or nmethodHandle.
>>> On 11/16/17 11:08 AM, Igor Veresov wrote:
>>>> The problems here is that the sweeper can transition a newly created nmethod to not_entrant state before it’s installed. This breaks the logic in nmethod::verify_interrupt_point() called at the end of the installation process. As a result locks might be taken in that can safepoint and safepoints are forbidden during the method install.
>>>> The solution is to introduce an new nmethod state: not_installed. The goal is to be able to prevent certain things like sweeping or parts of the verification happening during the nmethod installation.
>>>> The change passed the failing test and the internal mach5 pre-intergration testing.
>>>> Webrev: http://cr.openjdk.java.net/~iveresov/8043070/webrev.00/
>>>> JBS: https://bugs.openjdk.java.net/browse/JDK-8043070
>>>> Note, this also fixes: https://bugs.openjdk.java.net/browse/JDK-8028001
More information about the hotspot-compiler-dev