[9] RFR(S): 8152947: VM crash with assert(!removed || is_in_use()) failed: unused osr nmethod should be invalidated

[Vladimir Ivanov]

Reviewed.

Best regards,
Vladimir Ivanov

On 3/29/16 6:56 PM, Tobias Hartmann wrote:
> Hi Vladimir,
>
> On 29.03.2016 17:42, Vladimir Ivanov wrote:
>> Does it still work when method() == NULL?
>>
>> +#ifdef ASSERT
>> +    // Make sure osr nmethod is invalidated, i.e. not on the list
>> +    bool found = method()->method_holder()->remove_osr_nmethod(this);
>> +    assert(!found, "osr nmethod should have been invalidated");
>> +#endif
>
> Yes, you are right. We should check for method() != NULL. New webrev:
> http://cr.openjdk.java.net/~thartmann/8152947/webrev.01
>
> Thanks,
> Tobias
>
>>
>> Best regards,
>> Vladimir Ivanov
>>
>> On 3/29/16 6:28 PM, Tobias Hartmann wrote:
>>> Hi Dmitry,
>>>
>>> On 29.03.2016 15:05, Dmitry Dmitriev wrote:
>>>> Hi Tobias,
>>>>
>>>> I think that "#ifdef ASSERT" must be instead of "#ifndef ASSERT" on lines 1338 and 1516. And it seems that the whole fix should be retested in this case.
>>>
>>> Oh, thanks for catching this!
>>>
>>> Here is the new webrev:
>>> http://cr.openjdk.java.net/~thartmann/8152947/webrev.01
>>>
>>> I verified that this (still) fixes the problem and re-submitted the tests.
>>>
>>> Thanks,
>>> Tobias
>>>
>>>>
>>>> Thanks,
>>>> Dmitry
>>>>
>>>> On 29.03.2016 15:01, Tobias Hartmann wrote:
>>>>> Hi,
>>>>>
>>>>> please review the following patch:
>>>>>
>>>>> https://bugs.openjdk.java.net/browse/JDK-8152947
>>>>> http://cr.openjdk.java.net/~thartmann/8152947/webrev.00/
>>>>>
>>>>> JDK-8023191 added an assert to check if osr nmethods are invalidated (i.e. removed from the osr list). The assert is too strong because nmethod::invalidate_osr_method() may be executed by multiple threads that compete for making an osr nmethod not-entrant:
>>>>>
>>>>> Thread 1: Invalidates the osr nmethod and blocks just before executing the assert in line 1397.
>>>>> Thread 2: nmethod is already invalidated. Acquires patching lock and sets the nmethod to not-entrant.
>>>>> Thread 1: Continues execution and hits the assert because the nmethod was on the list and is (now) not-entrant.
>>>>>
>>>>> Thread 1 should ignore this and return. See also comment in nmethod::make_not_entrant_or_zombie():
>>>>>      // another thread already performed this transition so nothing
>>>>>      // to do, but return false to indicate this.
>>>>>
>>>>> I changed the verification code to check for osr invalidation *after* the state of the nmethod was set.
>>>>>
>>>>> Tested with failing test, JPRT and RBT (in progress).
>>>>>
>>>>> Thanks,
>>>>> Tobias
>>>>