OpenSSL 3.0.0 state machine handles unexpected EOFs more strict and requires peers to properly shut down connections. The old OpenSSL 1.1.1 behavior can be get back with SSL_OP_IGNORE_UNEXPECTED_EOF.

I propose to add the option by default until Python's ssl module has better ways to perform one-way shutdown of connections.

https://www.openssl.org/docs/manmaster/man3/SSL_CTX_set_options.html

> Some TLS implementations do not send the mandatory close_notify alert on shutdown. If the application tries to wait for the close_notify alert but the peer closes the connection without sending it, an error is generated. When this option is enabled the peer does not need to send the close_notify alert and a closed connection will be treated as if the close_notify alert was received.

> You should only enable this option if the protocol running over TLS can detect a truncation attack itself, and that the application is checking for that truncation attack.

New changeset 6f37ebc61e9e0d13bcb1a2ddb7fc9723c04b6372 by Christian Heimes in branch 'master':
bpo-43794: OpenSSL 3.0.0: set OP_IGNORE_UNEXPECTED_EOF by default (GH-25309)
https://github.com/python/cpython/commit/6f37ebc61e9e0d13bcb1a2ddb7fc9723c04b6372

New changeset e18ebd9ec546a3647a57c282735350f60a26d66d by Miss Islington (bot) in branch '3.8':
bpo-43794: OpenSSL 3.0.0: set OP_IGNORE_UNEXPECTED_EOF by default (GH-25309)
https://github.com/python/cpython/commit/e18ebd9ec546a3647a57c282735350f60a26d66d

New changeset 54d89a33e0d1b854fd5a72889d6554aeeb4170f0 by Miss Islington (bot) in branch '3.9':
bpo-43794: OpenSSL 3.0.0: set OP_IGNORE_UNEXPECTED_EOF by default (GH-25309)
https://github.com/python/cpython/commit/54d89a33e0d1b854fd5a72889d6554aeeb4170f0

I have updated whatsnew in GH-25817.

OpenSSL 1.1.1 also handled EOFs strictly, but this behavior was generally suppressed in the ssl module through the default setting of suppress_ragged_eofs=True (thus enabling truncation attacks by default). The PR changes the behavior of existing applications in such a way that previously detectable unexpected EOFs are now no longer detectable by default. To make matters worse, EOF errors are not translated to SSLEOFError anymore, and instead I have to match the strerror attribute in SSLError to detect this condition.