RFR(S): 8200374: Add ThreadsSMRSupport::verify_hazard_pointer_scanned() to verify threads_do()

[Erik Österlund]

Hi Dan,

This looks great. I will sleep much better at night with this 1:1 
verification. :)

Thanks,
/Erik


On 2018-03-29 03:52, Daniel D. Daugherty wrote:
> Greetings,
>
> I have a (mostly) small enhancement for Thread-SMR:
>
> JDK-8200374 Add ThreadsSMRSupport::verify_hazard_pointer_scanned() to 
> verify threads_do()
> https://bugs.openjdk.java.net/browse/JDK-8200374
>
> Here's the webrev URL:
>
> http://cr.openjdk.java.net/~dcubed/8200374-webrev/0_for_jdk_hs_open/
>
> This is Erik O's improvement on the assertion added by the following
> bug fix (with some minor tweaking done by me):
>
> JDK-8199813 SIGSEGV in ThreadsList::includes()
> https://bugs.openjdk.java.net/browse/JDK-8199813
>
> Summary of the changes:
>
> - Replace the assertion that I added in JDK-8199813 with a closure
>   based function that verifies the threads_do() contract depended
>   on by Thread-SMR.
> - Add ThreadsSMRSupport::verify_hazard_pointer_scanned() to verify
>   that the calling thread's hazard pointer is scanned by threads_do().
>   The new function is called from both ThreadsSMRSupport::acquire_*()
>   functions.
> - Refactor the non-JavaThread part of Threads::threads_do() into
>   Threads::non_java_threads_do() so that the non-JavaThread part
>   can also be called by other threads_do() functions. Yes, the
>   Threads::threads_do() contract is still to scan every thread in
>   the system.
> - Add hooks for a "tracing sampler thread" to be optionally scanned
>   by Threads::non_java_threads_do().
>
> This fix has gone thru a couple of Mach5 builds-tier1, jdk-tier[1-3],
> and hs-tier[1-3] runs. I've also started my usual 24+ hour Thread-SMR
> stress testing run on my Solaris-X64 server.
>
> Just to be extra sure, I backed out the fix from JDK-8199813 to
> JavaThread::verify_not_published() (which started this round of
> Thread-SMR fixes) and we catch the issue in the new function
> ThreadsSMRSupport::verify_hazard_pointer_scanned().
>
> Thanks, in advance, for any comments, suggestions, or feedback.
>
> Dan
>