RFR: 8373118: Test java/lang/Thread/virtual/Starvation.java timed out [v26]

[Doug Lea]

On Thu, 15 Jan 2026 12:39:43 GMT, Viktor Klang <vklang at openjdk.org> wrote:

>> Doug Lea has updated the pull request incrementally with one additional commit since the last revision:
>> 
>>   Another set of contend vs deactivate vs park tradeoffs
>
> src/java.base/share/classes/java/util/concurrent/ForkJoinPool.java line 1861:
> 
>> 1859:                 break;
>> 1860:         }
>> 1861:     }
> 
> The following might be a potential alternative encoding of signalWork—it would be interesting to hear if it makes any difference on your testing array, Doug:
> 
> 
>     final void signalWork(WorkQueue src, int base) {
>         int pc = parallelism, i, sp; // rely on caller sync for initial reads
>         long c = U.getLong(this, CTL);
>         WorkQueue[] qs;
>         while ((short)(c >>> RC_SHIFT) < pc && (qs = queues) != null &&
>                qs.length > (i = (sp = (int)c) & SMASK) && (src == null || src.base - base < 1)) {
>             if (i == 0) {
>                 if ((short)(c >>> TC_SHIFT) >= pc)
>                     break;
>                 if (c == (c = U.compareAndExchangeLong(this, CTL, c, ((c + TC_UNIT) & TC_MASK) | ((c + RC_UNIT) & RC_MASK)))) {
>                     createWorker();
>                     break;
>                 }
>             }
>             else {
>                 WorkQueue v;
>                 if ((v = qs[i]) == null)
>                     break;
>                 if (c == (c = U.compareAndExchangeLong(this, CTL, c, (v.stackPred & LMASK) | ((c + RC_UNIT) & UMASK)))) {
>                     v.phase = sp;
>                     if (v.parking != 0)
>                         U.unpark(v.owner);
>                     break;
>                 }
>             }
>         }
>     }

Thanks for prodding me to instrument current version. I saw that retries are no longer very common -- max seen across various contention-prone tests was 17, but in most never more than 4 except more during warmup before code is compiled. Versions that did filter check before CAS systematically have more retries though. And as I noticed in some previous versions, separating the branches with different CAes invites the compiler to rearrange code in a worse way to only have one CAS. at least in the surprisingly common case where it inlines signalWork in callers.  So net change from all this so far was a couple of cosmetic tweaks.

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

PR Review Comment: https://git.openjdk.org/jdk/pull/28797#discussion_r2702638385