Expose IO#timeout and IO#timeout= if available

[casperisfine]

Ref: https://bugs.ruby-lang.org/issues/18630

FYI: @ioquatix

[casperisfine]

Actually, might not be that simple, as the buffering layer breaks on a timeout:

RedisClient::SSLConnectionTest#test_tcp_downstream_latency:
TypeError: no implicit conversion of Integer into Hash
    /opt/rubies/3.2.2/lib/ruby/3.2.0/openssl/buffering.rb:247:in `>='
    /opt/rubies/3.2.2/lib/ruby/3.2.0/openssl/buffering.rb:247:in `gets'

I'll have to dig deeper.

[casperisfine]

Yeah, definitely doesn't work, OpenSSL buffering uses ossl_ssl_read_internal, which isn't aware of IO#timeout it seems.

[ioquatix]

openssl/ext/openssl/ossl_ssl.c

Line 1744 in fac5cac

rb_io_maybe_wait_readable(errno, io, RUBY_IO_TIMEOUT_DEFAULT);

Maybe we need to set it to nonblock mode? In theory, it should use the default timeout.

[casperisfine]

Maybe we need to set it to nonblock mode?

Perhaps? I'm not quite familiar enough with that code.

[ioquatix]

@rhenium is there a chance you can help look into this? I'm happy to help.

[rhenium]

There seem to be two issues before the buffering layer.

IO::Timeout isn't raised by OpenSSL::SSL at all. rb_io_maybe_wait_readable() and rb_io_maybe_wait_writable() on timeout return 0. This return value is currently discarded.

Also, a single blocking operation on TLS, such as #connect and #sysread, can involve multiple blocking operations on the underlying socket. This would mean we can't rely on RUBY_IO_TIMEOUT_DEFAULT, but have to get rb_io_timeout() and subtract elapsed time before waiting again.

[rhenium]

openssl/ext/openssl/ossl_ssl.c

Line 1744 in fac5cac

rb_io_maybe_wait_readable(errno, io, RUBY_IO_TIMEOUT_DEFAULT);

Maybe we need to set it to nonblock mode? In theory, it should use the default timeout.

The underlying socket is set to non-blocking mode by SSLSocket#initialize. We needed this for implementing SSLSocket#read_nonblock and also because we didn't want blocking IO within libssl while holding the GVL.

[rhenium]

SocketForwarder (in this file, above) is the right place for this definition.

[ioquatix]

Also, a single blocking operation on TLS, such as #connect and #sysread, can involve multiple blocking operations on the underlying socket. This would mean we can't rely on RUBY_IO_TIMEOUT_DEFAULT, but have to get rb_io_timeout() and subtract elapsed time before waiting again.

Thanks for this clarification. IO#timeout was designed with discrete, low level operations in mind. I think we can start with the simplest implementation, which is discrete timeout per operation. The goal of IO#timeout is essentially a safety net. If #connect performs several internal operations, it's reasonable that there are several "timeouts". Other parts of IO behave similarly. e.g. each call to wait_whatever will use a new timeout, therefore, on slow connections, it's possible for things like gets to invoke read several times, and thus wait several times, each with a new timeout.

If users desire application level timeouts, e.g. "This group of operations should take at most X seconds", they should use Timeout.timeout or some similar construct.

[rhenium]

This is superseded by #714 which fixes SSLSocket to correctly raise IO::Timeout.