- Make sure we always await the last flush task before calling FlushAsync
again instead of preemptively calling FlushAsync and checking to see
if the ValueTask is incomplete before bothering to acquire the _flushLock
- This now acquires the _flushLock fore every call to Response.Body.Write
whereas this only happened for truly async writes before.
- I don't think this is a big concern since this should normally be uncontested,
and DefaultPipeWriter.FlushAsync/GetResult already acquire a lock.
This is a massive set of changes to Kestrel to remove the existing pubternal transport layer and implement a public facing API for listeners and clients, see the details here #10308.
This change only has the server side pieces of the story as I don't want to add the client APIs without having ported SignalR to use them. Here are the highlights:
- Transport.Abstractions is empty (will be removed in a separate PR as it requires removing it from a ton of places)
- TransportConnection has been moved to Connection.Abstractions (we can decide if we need to consolidate with DefaultConnectionContext in a later PR)
- Added FileHandleEndPoint which allows binding to a file handle (could be a pipe or tcp handle)
ListenOptions has been gutted for most pubternal API and returns various types of binding information . The source of truth is the EndPoint instance.
- Cleaned up a bunch of libuv tests decoupling them from Kestrel.Core
## Breaking changes
- Removing pubternal API is itself a breaking change but one that we already planned to do.
- We've removed the ability to set the scheduling mode on Kestrel
- DisposeAsync was added to ConnectionContext
- NoDelay was removed from ListenOptions. This has been moved to each of the transports. One major difference though is that it's no longer localized per endpoint but is global. We'd need a derived EndPoint type (or maybe extend IPEndPoint) to store both the socket options and the binding information.
Changes:
* Remove obsolete targets which are unnecessary now that this repo no longer builds git submodules in a separate build process
* Remove the need for static analysis of 'ArtifactInfo' items
* Simplify how the code signing task is configured
* Remove unused repo tasks
* Remove duplicate lists of external dependencies and packages to be produced
* Remove obsolete build definition
* Remove obsolete build script parameters
* Add VisualStudioSetupOutputPath
Stephen Halter
David Fowler
Ben Adams
Justin Kotalik
Nate McMaster
Nate McMaster