- Unsupported members should throw NotSupportedException instead of
NotImplementedException per MSDN docs for CanRead/CanSeek/CanWrite.
- Position should throw NotSupportedException when CanSeek is false.
- FrameRequestStream.Flush/FlushAsync should not throw
NotImplementedException.
- Use expression-bodied members for CanRead/CanSeek/CanWrite on
FrameRequestStream to match FrameResponseStream.
- Provide no-op override of LibuvStream.FlushAsync to match Flush.
- If we're done before the client sends a FIN, force a FIN into the raw
SocketInput so the task in FileteredStreamAdapter finishes gracefully
and we dispose everything in proper order.
- If there's an error while writing to a stream (like ObjectDisposedException),
log it once and prevent further write attempts. This means the client closed
the connection while we were still writing output.
- This also fixes a bug related to the point above, where memory blocks were
being leaked instead of returned to the pool (because we weren't catching
the exception from Write()).
We need to attempt to consume start lines and headers even after
SocketInput.RemoteIntakeFin is set to true to ensure we don't close a
connection without giving the application a chance to respond to a request
sent immediately before the a FIN from the client.
- Without the _sync lock, if new data was produced as ConsumingComplete
was called, the next "await SocketInput" might never complete despite not
all data being examined.
- If more data is produced afterward, the stall would be prevented, but this
isn't always the case such as during the end of the request.
- This means connections become untracked sooner than before and not all blocks will
necessarily be returned.
- The assertion in the MemoryPoolBlock2 finalizer was weakened because FilteredStreamAdapter
will continue to use blocks after libuv stops tracking the associated connection.
- Make 100% sure we don't accept new connections after we dispose the listen socket by using a flag.
- Add a (currently unused) AllowStop method to KestrelThread. This is meant to be called from
listeners when we stop accepting new connections, but needs investigation to prevent flakiness.
- Even when safe handles are disposed explicitly, ReleaseHandle is sometimes
called on another thread which breaks uv_close.
- Ensure we close the UvAsyncHandle the uv loop so that the second call
to uv_run always completes without a timeout/Thread.Abort.
- Re-enable some tests. Add skip conditions for those that aren't passing.
- Make the time given for requests to complete gracefully configurable.
- Complete all async reads so calling code can re-check whether to stop
request processing and exit if in between requests.
- Don't wait for a FIN from the client since some browsers (e.g. IE & Chrome)
will take a long time to send one.
- Ensure all ConnectionFilters complete before the memory pool is disposed.
- Ensure blocks get returned even when a ConnectionFilter produces a failed read
moozzyk
Cesar Blum Silveira
Justin Van Patten
Mike Harder
Stephen Halter
Doug Bunting
Ben Adams
Chris R
Brennan