This might help address #3015
This only affects rate timeouts. Normal fixed timeouts might deserve the same treatment, but that would require some additional locking to ensure we don't modify the sentinel value.
Prior to this, only the response body counted toward the HTTP/2 response data rate. This PR aligns the HTTP/2 logic closer to the HTTP/1.x logic and measures the rate for all HTTP/2 response data.
This PR also accounts for all response bytes written, not just those that immediately induced backpressure.
- Improve test reliability of tests verifying the RequestAborted token gets tripped
- Once the response body is completed, don't fire the token for that request even if it is accessed later on.
* Don't preserve the ExecutionContext when dispatching callbacks
- The layers up stack already preserve and restore the ExecutionContext, there's no need to capture and run for the scheduler callback and cancellation token callback
This change makes the handling of graceful shutdown work for more than just http scenarios. This should allow us to move TLS further out and should also allow us to start moving things to connection middleware instead of connection adapters.
Summary of the things changed/added:
- Added IConnectionLifetimeNotificationFeature that represents an attempt to gracefully close the connection that isn't being aborted. This feels pretty awful but we may have to do it.
- Moved connection management to the ConnectionDispatcher and out of the HttpConnectionMiddleware
- Removed Http from the names of the ConnectionManager and Heartbeat
- Include hosting logs in some tests that where previously missing them
- Prevent duplicate logs from Mock CallBase an CompositeKestrelTrace
- Log ports used by transport functional tests
- Add file logging to HTTP/2 "unit" tests
Stephen Halter
John Luo
Chris Ross (ASP.NET)
Chris Ross
Simon Cropp
tiago-soczek
David Fowler