The Mystery of the Vanishing Message's Dead Lock

[Michael Truog]

On 7/23/20 1:52 PM, Fred Hebert wrote:
> The release notes for ERTS-10.7.1 mention:
>
>     Fixed a bug in a receive optimization. This could cause a receive
>     not to match even though a matching message was present in the
>     message queue. This bug was introduced in ERTS version 10.6 (OTP
>     22.2).
>     Own Id: OTP-16572 Aux Id: ERL-1199, OTP-16269 
>
>
> We were on ERTS-10.6.2 (OTP-22.1.??). Today we reproduced it again and 
> I knew just what to look for so:
> (rabbit@REDACTED)16> lists:filter(fun({_Pid, _Ref, {sync_notify, _}}) -> false; (_) -> true end,
> (rabbit@REDACTED)16>              element(2,recon:info(rabbit_log_connection_lager_event, messages))).
> [{#Ref<0.229770800.3298820110.79067>,ok}]
> And where we have it. Out of 85,000 messages or so in that worker, 
> only one of them wasn't a blocked sync_notify, and it was a message of 
> the form {Ref, ok} which entirely matches the compiler bug. I can't 
> say for sure it's the right response, but assuming things are 
> synchronous there's no reason we'd have another one either. We're 
> upgrading to Erlang/OTP-22.3.1 or newer (which contains ERTS-10.7.1). 
> I'm very surprised that the compiler bug had an effect on gen:call() 
> rather than just the branching pattern in socket(but somewhat relieved 
> because it would otherwise imply another bug). Out of caution, we'll 
> have engineers try to reproduce it once more to see if that fixes it.
The erts-10.6.2 release was included in Erlang/OTP 22.2.2 [1], so it 
must be that release or a custom release based on it.  The erts-10.7.1 
release with Erlang/OTP 22.3.1 [2] does have an important fix (OTP-16572 
is the Ericsson identifier and ERL-1199 is the bugs.erlang.org 
identifier) since that should impact all the OTP behaviors that create 
synchronous messages (due to taking advantage of the {Ref, _} 
optimization).  Good to know, to make sure development is using a 
version >= 22.3.1.

[1] https://erlang.org/pipermail/erlang-questions/2020-January/098996.html
[2] http://erlang.org/pipermail/erlang-questions/2020-April/099386.html
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://erlang.org/pipermail/erlang-questions/attachments/20200723/14adeffb/attachment.htm>