RFR: 8265753: Remove manual JavaThread transitions to blocked [v3]

[Richard Reingruber]

On Wed, 12 May 2021 08:04:47 GMT, Robbin Ehn <rehn at openjdk.org> wrote:

>> Please consider this change which removes the manual transitions to blocked.
>> This adds a preprocess template/functor which is executed in the destructor of 'ThreadBlockInVM' if we are going to do any processing.
>> This gives us a way to backout of the object/raw monitor before suspend or other processing, such as a safepoint.
>> 
>> The object monitor code could be straight forward changed to use this instead of manual transitions.
>> 
>> Raw monitors on the other hand are a bit more complicated due to 'implicit' rules (consequences of the specs).
>> Added a comment in hpp trying to explain it; we cannot simply transition with a raw monitor held.
>> This caused the change in the destructor ~ThreadInVMfromNative() (this specific change have also been tested in unrelated exploration of transition), now this RAII does the same as we do when going to native from Java, just setting the state.
>> Since we are going from an unsafe state, in VM, to a safe state, in native, we do not need to check the poll.
>> That made it possible to careful use ThreadInVMfromNative in raw monitors.
>> 
>> I also remove the early CAS in raw_enter.
>> We lock a lock to do a CAS, in the uncontended case means CAS on lock then CAS raw monitor.
>> Now we instead do a transitions, in the uncontended case means fence, CAS raw monitor, fence.
>> (multiple fence (CAS is also a fence) very close to each other have little additional performance impact on contemporary hardware)
>> 
>> Passes t1-t7 and manual stressing relevant test groups.
>
> Robbin Ehn has updated the pull request incrementally with one additional commit since the last revision:
> 
>   Fixes for Dan

src/hotspot/share/prims/jvmtiRawMonitor.cpp line 395:

> 393:           ThreadBlockInVMPreprocess<ExitOnSuspend> tbivmp(jt, eos);
> 394:           simple_enter(jt);
> 395:           _recursions = save;

I think you should restore `_recursions` only when finally leaving the loop because if you do it here and exit the monitor again because of a suspend then another thread could hit the `guarantee(_recursions == 0, "invariant")` at L349. This is a preexisting issue but since you're modifying this I reckon it makes sense to correct it.

-------------

PR: https://git.openjdk.java.net/jdk/pull/3875