OpenSSL should clear error queue before each call

[rdp]

Appears the "best practice" for OpenSSL is to call ERR_clear_error before doing any call, basically "just in case" an earlier SSL_xx call didn't handle its error return value(s) right.

https://stackoverflow.com/questions/18179128/how-to-manage-the-error-queue-in-openssl-ssl-get-error-and-err-get-error

related discussion:

#8103 (comment)

[rdp]

I would say that this helps it "be better for multi threaded" as well, but, on closer examination, for SSL_Accept it calls underlying #read multiple times underneath. That means if there are multiple fibers doing SSL calls within the same thread, the thread local error queue could get "cross threaded" with failure messages. I would propose some kind of thread local mutex so only one fiber can be within an openSSL "anyting" at a time but that kind of ruins the point...which makes me wonder if OpenSSL is a good paradigm for the M:N fibers to threads paradigm that Crystal is moving toward has...but anyway this will be slightly better :)

[rdp]

I think what should happen is that when there is a fiber context switch, it should "save away" any current openssl failures (and restore from the incoming fiber). Also OpenSSL_Shutdown should clear the queue. FWIW...the latter being the easier of the two, I'll hopefully do a PR for it sometime :)

[RX14]

@rdp at the moment, fibers can only change threads at specific points. As long as none of these occur between setting an error and reading it, it should be fine.

[rdp]

Yeah typically that's enough. Unfortunately sometimes OpenSSL makes several calls to underlying socket #read and #write which all add up to "one failure" and it can store multiple failures messages on the stack (each #read or #write I assume is a context switch point) so if multiple were to fail at the same time they could get cross twisted...

ref: https://stackoverflow.com/q/18179128/32453 comment

Now that you mention it, if we were to make these "thread local" messages basically "fiber local" then...we could call ERR_clear_error before every call to guarantee a clear stack. The original request :)