PEP 748: tlslib - context & socket#4960
Conversation
Ease reading the diff of the next commits.
The `connect()` function is poorly named server-side as it actually `bind()` the socket. Server-side the function lacks a `family` parameter to support IPv6 without a DNS lookup. Actually all three `connect()`, `bind()` and `listen()` socket function are pretty low-level. The high-level `create_connection` (for `connect`) and `create_server` (for `bind()` + `listen()`) are more pythonic. Wrap the latter and not the former, also to remove the need of a `listen()` method on the created socket (which is useless client-side).
|
All commit authors signed the Contributor License Agreement. |
Documentation build overview
709 files changed ·
|
|
cc PEP authors @jvdprng and @woodruffw. |
jvdprng
left a comment
There was a problem hiding this comment.
Thanks for writing this up!
I feel like this pulls a lot of socket API surface into the PEP. Couldn't we stick to a minimum API surface that leaves a bunch of socket options to the backends? E.g., something like:
class ClientContext(Protocol):
def connect(
self,
address: tuple[str | None, int],
*,
server_hostname: str | None = None,
timeout: float | None = None,
) -> TLSSocket: ...
class ServerContext(Protocol):
def create_server(
self,
address: tuple[str | None, int],
*,
) -> TLSListener: ...
I'm not sure whether splitting out 'listening sockets' and 'connected sockets' is a good idea, but it would solve some of the current blurriness around listen/accept. In general, I'm hesitant to pull in a bunch of socket stuff without discussing this on the PEP boards first.
If we do want to go down this route, we should probably also specify in the PEP that we want to focus on INET sockets. I've left some comments regarding cleaning up types in that case.
The shutdown changes are definitely needed, but I feel like the current PR overspecifies a little.
| timeout=GLOBAL_DEFAULT, | ||
| source_address=None, | ||
| *, | ||
| all_errors=False, |
| family=AF_INET, | ||
| backlog=None, | ||
| reuse_port=False, | ||
| dualstack_ipv6=False, |
| self, | ||
| address: tuple[str | None, int], | ||
| *, | ||
| family=AF_INET, |
There was a problem hiding this comment.
Do we want to set AF_INET as the default here? We could have None that lets the backend choose based on the address?
|
|
||
| @abstractmethod | ||
| def connect(self, address: tuple[str | None, int]) -> TLSSocket: | ||
| def create_connection( |
There was a problem hiding this comment.
Should we add a server_hostname parameter here to support 'connect by IP, verify as DNS name'?
| received from the other side. | ||
| In either case, this method should return WantWriteError if sending the | ||
| close_notify alert currently fails.""" | ||
| def shutdown(self, how: Literal[0, 1, 2]) -> None: |
There was a problem hiding this comment.
I'm worried that the current shutdown / close descriptions may be overspecified for the PEP. Should we focus on describing the observable behavior?
shutdown(SHUT_WR)gracefully closes the TLS sending sideshutdown(SHUT_RD)andshutdown(SHUT_RDWR)may discard unread peer data and are unsafe unless truncation is not an issueclose()should only fully close the transport when safe, otherwise it should raiseWantReadError/WantWriteErroras appropriate
Also, should we define a specific enum for how? Using literals might be brittle.
| backlog=None, | ||
| reuse_port=False, | ||
| dualstack_ipv6=False, | ||
| ) -> TLSSocket: |
There was a problem hiding this comment.
Do we want stronger typing for the TLS sockets? Both 'listening sockets' and 'connected sockets' are still the same type.
PEP 748: create_connection and create_server instead of connect
Prior discussions at:
That commit is the solution to my struggle binding an Ipv6 localhost address when I first tested tlslib. At the time the library was always binding the socket using
family=AF_INET. The solution we implemented in tlslib is to perform a DNS request on the user-provided address to determine the family to use (family=AF_INET6in my case).I'm still unhappy with this API server-side. "connect" is the wrong verb, it should be "bind". It is not obvious if other sockets are supported (e.g. Unix Domain Socket).
create_connectionandcreate_servermake so much more sense to me, at least as long as we don't have awrap_socketfunction.PEP 748: shutdown(show: 0|1|2) instead of close(force: bool)
Prior discussions:
Work in progress in siotls:
The messages I wrote on the forum explain the problems I have with
close(force: bool), which mainly boils down to "it feels too TLS 1.2-ish". I spent many days exploring ideas to have a clean API to terminate the TLS connection a way that is safe in both TLS 1.2 (synchronous termination) and TLS 1.3 (asynchronous and IMO better). Reusingsocket'sshutdownterminology feels right to me. It works well for TLS 1.3 and is fine for TLS 1.2.There are other things in the current Buffer & Socket spec that I find are leaking openssl's API, and that I expect won't work very well with other TLS libraries (including my own) but I'm not ready at this time to make new proposals.