RFR: 8305819: LogConfigurationTest intermittently fails on AArch64

[David Holmes]

On Thu, 4 May 2023 09:23:53 GMT, Andrew Haley <aph at openjdk.org> wrote:

>> The PR was never approved and was somewhat contentious. I would want to see how UL uses the proposed RWL before considering the integration of both changes. This is not a trivial amount of work I know.
>> 
>> Buried on the other PR was a suggestion to switch the broken UL RCU mechanism for one of the other VM RCU-like mechanisms, but I'm not familiar enough with those other mechanisms to say if that is viable. But perhaps the UL RCU mechanism can be examined against those other ones to see if any more bugs are evident?
>> 
>> Or we just proceed with a suitably commented storeStore barrier and hope that this is the last RCU bug.
>
>> Or we just proceed with a suitably commented storeStore barrier and hope that this is the last RCU bug.
> 
> We can't do only that, because it's still buggy. For it to be correct portable code we need at least a `memory_order::consume` load on the reading side. Given that `consume` is C++20 and we're still on C++14, something like the kernel's `READ_ONCE` would probably do.

I assume that is the memory barrier that RCU is supposed to be implicitly implementing? If this is missing from this RCU-like thing can we just use release/acquire as suggested much earlier?

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

PR Review Comment: https://git.openjdk.org/jdk/pull/13421#discussion_r1184798055