RFR: 8288904: Incorrect memory ordering in UL [v2]

[Johan Sjölén]

On Wed, 22 Jun 2022 21:57:59 GMT, David Holmes <dholmes at openjdk.org> wrote:

>> @dholmes-ora,
>> 
>> The implementation is lock-free and uses RCU.
>> 
>> This is approximately how it works in pseudo-code:
>> 
>> 
>> class LogOutputList {
>>   LogOutputNode* entry; // All threads gain entry to the linked list through this pointer
>> }
>> 
>> void LogOutputList::mutate_output_list() {
>>   // Save the entry pointer
>>   LogOutputNode* cur = entry;
>>   // Remove the entry pointer, now no new readers can access this list
>>   entry = NULL;
>>   // OK, we can now wait for all readers to exit
>>   wait_until_no_readers();
>>   // ... Do things we need to do to cur ...
>>   // Restore pointer, new readers can occur.
>>   entry = cur;
>> } 
>> 
>> 
>> There are some details here which I have omitted, but this is the gist of it.
>
>> There are some details here which I have omitted, but this is the gist of it.
> 
> @jdksjolen  - it is the actual details I have concerns about. More seems to be done than just a pointer switch to unlink, and the unlink itself is not an atomic operation - are we guaranteed only a single writer is possible? And even so, as others have noted, without atomics or barriers there is no guarantee readers will see the unlink has happened. I'm a bit surprised as this code has been extensively examined in the past.

This PR is still looking for a sponsor. @dholmes-ora, would you mind sponsoring this?

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

PR: https://git.openjdk.org/jdk/pull/9225