Would you be able to provide example code that reproduces this problem or additional logs from when it occurred? Thanks.

It's hard to reproduce, we collect about 17G log to figure out what happen at that points.
I add many debug log to print some variables, It's useless without give you codes, hard to remove log's secret info.

From our limited not completed logs. It's happen in HTTP2 POST, retrying. content-length == 0

session.as_mut().is_body_done() == false
// false || true
buffer.is_some() || session.as_mut().is_body_empty()
// false || true
let end_of_body = end_of_body || data.is_none();

Ok, that’s helpful, I’ll spend some more time looking into this.

It's happen in HTTP2 POST, retrying. content-length == 0

I'm able to reproduce this race condition locally, this was a good hint. The session.as_mut().is_body_done() depends on the receive half of the stream to be EOS in order to be true which is different than http/1.1.

Returning the request_done here works but it's only because data.is_none() evaluates to true in send_body_to_pipe, there's probably a better way to handle this.

yes, end of stream has different means than end of body.

I know that's not the best solution but after patch it, it's never happened again in our prod.

you can give some hint how I could change pr.

We'll want to fix is_body_done for h2 to check is_body_empty so it's consistent with the http/1.1 implementation, that makes the proxy logic consistent between protocols.

It's also a good idea in the tx.reserve() select! arm to check if the tx is closed to be defensive. I don't think any changes need to be made in proxy_h2.

This is a bit more involved, we can fix this internally if you don't have time to amend the PR. Thanks again for filing an issue for this.

Yes, we can revert proxy_h2 changes and also we haven't use h2 upstream yet.