[11u] RFR: 8254185: Fix Code cache sweeper heuristics for JDK 11

[Nils Eliasson]

Hi Goetz,

Yes. With both fixes the sweeper will work fine again.

Best regards,
Nils

On 2020-10-15 12:47, Lindenmaier, Goetz wrote:
> Hi Nils,
>
> Do I understand correctly that after this fix I should also push
> "8244278: Excessive code cache flushes and sweeps" ?
> https://bugs.openjdk.java.net/browse/JDK-8244278
>
> Best regards,
>    Goetz.
>
>> -----Original Message-----
>> From: jdk-updates-dev <jdk-updates-dev-retn at openjdk.java.net> On Behalf
>> Of Nils Eliasson
>> Sent: Friday, October 9, 2020 4:35 PM
>> To: jdk-updates-dev at openjdk.java.net
>> Subject: [11u] RFR: 8254185: Fix Code cache sweeper heuristics for JDK 11
>>
>> Hi,
>>
>> Bug: https://bugs.openjdk.java.net/browse/JDK-8254185
>> Webrev: http://cr.openjdk.java.net/~neliasso/8254185/webrev.01/
>>
>> This fix is a replacement for JDK-8244660 that I consider to risky to
>> back port.
>>
>> This issue can't be seen unless JDK-8254185 is applied. That bug causes
>> the sweeper to trigger on every code cache allocation. When JDK-8254185
>> is applied then no sweeps can be observed at all.
>>
>> The main issue is that the sweeper thread is never woken. And if it
>> isn't woken, it can't check the counters to see if it is time to sweep.
>>
>> This fix include two changes:
>>
>> 1) Notifying waiters on the codecache_lock when a threshold for
>> reclaimed nmethods has been reached. This fix is enough on its own.
>>
>> In JDK 15 I introduced a new lock for the sweeper because a adding a
>> notify in NMethodSweeper::possibly_enable_sweeper conflicted with the
>> compiledMethod_lock in add_osr_method. That lock was added when
>> deoptimization was converted to using handshakes. So in 11 it is still
>> ok to use the codecache_lock.
>>
>> 2) Moving the updates of _total_nof_code_cache_sweeps and _last_sweep
>> inside block that is only run when doing a sweep.
>> This change just makes the counters update as expected. _last_sweep is
>> used to trigger sweeps after a number of safepoints. This is counted by
>> _time_counter. But since updates of _time_counter doesn't wake the
>> sweeper - it still doesn't work.
>>
>> In JDK 15 I removed all of this code since we can't even rely on
>> safepoints happening at all. I don't want to introduce any bugs in JDK
>> 11 by mistake - so I am leaving it as it is.
>>
>> Please review.
>>
>> Best regards,
>> Nils Eliasson