This reverts commit 40ffcdef46 in PR #209.
Unfortunately, Swift's behaviour around imports tends to be somewhat
"leaky". In this case, the leak is that the dependency on the
Package.swift allows downstream projects to import `NIOFoundationCompat`
without needing to actually specify the package dependency. This has
affected Hummingbird, which we noticed on our internal integration
testing functionality:
```
hummingbird/Sources/Hummingbird/Codable/JSON/JSONCoding.swift:18:8: error: no such module 'NIOFoundationCompat'
import NIOFoundationCompat
^
```
cc @Joannis @adam-fowler for the Hummingbird report.
Unfortunately, we can't make this change until we're willing to use a
semver major to achieve it. There may be some argument for doing that
now, as the semver major will be very cheap to adopt across the
ecosystem. But we'll need to do this in a considered way.
@Cyberbeni please feel free to reopen your PR targetting main, where we
can discuss whether this is worth issuing a semver major for.
Motivation:
In some cases users may want to access APIs we haven't exposed
in NIOTransportServices. We should have a fallback that allows users
to do this.
Modifications:
- Add ChannelOptions for getting NWConnection and NWListener.
Result:
Users have an escape hatch
Motivation:
Looks like when we previously added syncOptions support to our
channels, we had a few issues. The listeners had code added, but
the code never worked. This is because the code was defined in
subclasses, but the protocol conformance comes from the parent class,
and that parent class did not have a customized protocol witness.
The datagram channel was also entirely missing its support.
Modifications:
- Added the missing unit tests for sync options.
- Added syncOptions to StateManagedListenerChannel base class.
- Added overrides to the listener subclasses.
- Added syncOptions to datagram channel
Result:
Sync options actually work across the board.
Motivation:
We're centralizing on the Swift code of conduct, so we'll x-reference
that instead of holding our own.
Modifications:
Hyperlink out to Swift.
Result:
Shared CoC across the projects.
Motivation:
We try to support building our docs on Linux too, but that means we
can't explicitly reference any symbols from `index.md`.
Modifications:
Enable CI-ing docs.
Result:
Docs will render on Linux too.
Motivation
As we've rolled out support for multicast on Linux, it makes sense to
add equivalent configuration for swift-nio-transport-services. The two
features aren't one-to-one, as Network.framework has substantially more
capability than the Linux functionality.
Modifications
- Expose a multipathServiceType channel option
- Add a test to confirm we use it properly
Result
Multicast service types are available.
Motivation:
Lock has been deprecated in favour of NIOLock. Warnings aren't great.
Modifications:
- Replace Lock with NIOLock
- Update the minimum required NIO version to 2.42.0.
Result:
Everything builds cleanly
Motivation
Documentation is nice, and we can help support users by providing useful
clear docs.
Modifications
Add Docc to 5.6 and later builds
Make sure symbol references work
Add overview docs
Result
Nice rendering docs
Motivation:
Right now we're playing a little fast and loose with the lifetimes of
the sec_protocol_metadata_t. As a practical matter it is highly likely
that this is owned (and so kept alive by) the NWConnection, but rather
than risk that we should tighten up the lifetime management.
Modifications:
Use withExtendedLifetime to extend the lifetime.
Result:
Better lifetime management.
Motivation:
When we're waiting for connectivity, the user might tell us that they
aren't interested in waiting. In that case we take advantage of the
signal and close early.
However, the code as-written had a bug: we didn't care whether the user
told us they _didn't_ want to wait, or they _did_: we just closed
because they had an opinion! That's no good! We should only close if
they don't want to wait.
Modifications:
- Only close if the user doesn't want to wait.
- Add tests
Result:
We won't close if users don't want us to.
Motivation:
Private Relay kicks in to secure unencrypted data transfers, and
connections to port 80 are a good signal. As a result, the current
construction of this test will invoke Private Relay and attempt to route
over it.
While Private Relay is in beta this is a bad outcome, as it risks
causing bugs in unusual network scenarios. So let's tell Private Relay
we don't need it.
Modifications:
Pretend we're gonna connect to 443 instead.
Result:
More reliable test.
Motivation:
With NIO 2.32.0 we broke the core NIO module up into modules that split
apart the POSIX layer and the core abstractions. As a result, this
package no longer needs to express a hard dependency on the POSIX layer.
Modifications:
- Rewrote imports of NIO to NIOCore.
- Added NIOEmbedded imports where necessary in tests.
- Note that the main package still _depends_ on NIO, which is necessary
for backwards-compatibility reasons. This dependency is unused.
Result:
No need to use NIOPosix.
Motivation:
We shouldn't crash if the user gives us a bad port.
Modifications:
- Check the ports are in-band before we move on.
Result:
We won't crash.
Motivation:
The zero-length writes handler will crash if it receives channelInactive
before channelActive. Sadly, that's totally possible if a channel is
closed from the connect promise.
Modifications:
- Tone back the preconditions a bit.
- Add a test that triggers this path.
Result:
Channel handler behaves better on early close
Motivation:
Network.framework will crash when we attempt to connect to the host
string "". That's not ideal, so we should detect that case and avoid it.
Modifications:
- Add code to detect the empty host string.
Result:
No crashes when accidentally connecting to "".
Motivation:
In a patch where we added some metadata options, we accidentally set the
availability wrong. This is a compile-time error, so we broke watchOS
compilation.
Modifications:
Update the availability of some types on watchOS.
Result:
Users can build on watchOS again.
Motivation:
On Apple's devices it is considered a best practice to wait for network
connectivity in cases when a connection request cannot immediately be
satisfied, rather than erroring out. This reflects the dynamic and fluid
network environment on Apple devices, with their many network interfaces
and complex interactions between radios, VPNs, and network devices.
While NIOTS supports this model of interaction (and indeed uses it out
of the box), and supports configuring it (by setting
`NIOTSChannelOptions.waitForActivity`), a key pillar is missing:
observability. Right now we don't actually _tell_ the user when we're
waiting for connectivity. This makes it difficult to act on this
information.
Modifications:
- Added a user event, `NIOTSNetworkEvents.WaitingForConnectivity` that
is fired when connectivity could not be established, but the situation
may change in future.
Result:
Our users can tell their users when they're waiting for something!
Motivation:
Syncing onto the event loop queues is essentially never acceptable: we
can't appropriately guard against it going terribly wrong.
Modifications:
- Use locks instead, locks are good.
Result:
Removes some crashes.
Resolves#70.
Motivation:
We deprecated Atomic because we made a new, better one. We should use
it.
Modifications:
Replaced all use of Atomic with NIOAtomic.
Result:
Better performance, fewer compiler warnings.
Motivation:
Sadly the previous NIO warnings patch failed to notice that the tests
had a bunch of errors, despite the fact that those tests were
(repeatedly) run.
Modifications:
Directly apply the fixits.
Result:
Fewer warnings
Motivation:
We can remove a couple of unnecessary flatMap { $0 } operations by way
of using flatSubmit, added in NIO 2.9.0.
Modifications:
- Replace flatMap { $0 } with flatSubmit().
Result:
Remove some code noise.
Motivation:
In NIO 2.9.0 we moved all first-party `ChannelOption` types into their
own namespace. That was a good idea, but NIOTS uses them too and so it
now encounters build warnings. Additionally, NIOTS defines its own
`ChannelOption`s, and so should namespace those as well.
Modifications:
- Updated the code to use the namespaced `ChannelOption` types.
- Namespaced our own.
Result:
More namespacing.
Motivation:
In Network.framework it is possible for a bind operation to take a while
to complete if it ends up waiting for appropriate network conditions. As
a result, unlike in the POSIX case, we need to give users the ability
to configure the maximum amount of time they'd be willing to wait for a
bind call to succeed.
Modifications:
- Add a bind timeout.
Result:
Users won't have to wait forever.
Motivation:
When autoRead is on, the pipeline must observe the read calls in order
to be able to exert backpressure. Otherwise, autoRead is a
zero-backpressure mode, which isn't great.
Correctly call pipeline.read instead of self.read0 to avoid this.
Modifications:
- Updated NIOTSConnectionChannel to call pipeline.read().
Result:
Backpressure can be exerted.
Motivation:
macOS Catalina brings a .url endpoint to Network.framework, which is
currently causing build warnings. We should tolerate it and silence
those warnings.
Modifications:
Do the right thing when we receive a .url endpoint.
Result:
No build warnings.
Motivation:
Networking software like SwiftNIO that always has explicit flushes
usually does not benefit from having TCP_NODELAY switched off. The
benefits of having it turned on are usually quite substantial and yet we
forced our users for the longest time to enable it manually.
Quite a bit of engineering time has been lost finding performance
problems and it turns out switching TCP_NODELAY on solves them
magically.
Netty has made the switch to TCP_NODELAY on by default, SwiftNIO should
follow. This patch is the equivalent of apple/swift-nio#1020.
Modifications:
Enable TCP_NODELAY by default.
Result:
If the user forgot to enable TCP_NODELAY, their software should now be
faster.
Motivation:
With Network.framework becoming available on watchOS, you should now be
able to run your favourite NIO applications on watchOS 6!
Modifications:
Updated availability annotations to allow NIO TS.
Result:
More NIO on more watches
Motivation:
The expectation is that server channels use Channel as their data type,
but the initial data type in NIOTSListenerChannel was actually
NWConnection. This is unnecessary and it makes it hard to interop
between NIOTS and NIO.
Modifications:
- Initialize the NIOTSConnectionChannel in the NIOTSListenerChannel
instead of in AcceptHandler.
- Added some missing @available annotations in the tests.
Result:
More consistency
Motivation:
While we fixed a bug in SwiftNIO about setting multiple options of the same type in
apple/swift-nio#597, we never brought a fix for that issue forward. We did aim to resolve
this in SwiftNIO 2.0 by providing ChannelOptions.Storage, but unfortunately we forgot to
make the initializer for that type public (see apple/swift-nio#988). While SwiftNIO core
has to get its ducks in a row, users of this library should still be able to set more than
one socket option, in my humble opinion.
Modifications:
- Ported over ChannelOptions.Storage until apple/swift-nio#988 is fixed.
- Transitioned our code to use it.
- Tested it works.
Result:
Users can set multiple channel options.
Motivation:
At least in some cases it's possible to have a local or remote address
that we cannot represent in a SocketAddress. In the face of that, we were
unnecessarily crashing. Crashes should probably be avoided there.
Modifications:
- Replace the crashes with throws.
Result:
Fewer crashes, more nils.
Motivation:
When entering convergence we incorrectly specified a branch dependency
for something that is actually a tag.
Modifications:
- Remove branch dependency, use from: instead.
Result:
Users can use this repo!
Motivation:
Once every few thousand years, an event called syzygy (or "convergence")
occurs in our solar system. When these events occur, the fabric of spacetime
becomes thin, allowing reflections of a possible future to appear in our
reality.
This gives a few brave souls the opportunity to experience a future: a
possibility of the way the universe may one day be. But be warned: to
toy with causality is to risk unbearable pain and suffering. Only the
bravest of souls should look too closely at these shadows of the future.
Modifications:
- Chose a firm anchor point in the SwiftNIO timeline.
- Removed the braces that kept time travellers safe.
- Said a quiet prayer.
Result:
Opportunity.
Motivation:
Some users may wish to use peer-to-peer networking with bonjour.
Modifications:
Expose peer-to-peer networking via a new channel option.
Result:
Users will be able to use peer-to-peer networking
(cherry picked from commit e9c1e41fd1)
Motivation:
In some cases, such as when an event loop has been shutdown, channel
registration may fail. In these cases, we would incorrectly attempt to
deregister the channel, which would fail (and in debug builds, assert).
Really, we shouldn't transition into .registered until we know that we have,
in fact, registered.
However, we need to be cautious: we don't want to register unless we
believe we're in an acceptable state to register.
Modifications:
Updated the state enum to perform the registration at the correct
part of the state change function.
Result:
Harder to crash in debug mode
Motivation:
We were inadvertently modifying the registration information of
NIOTSEventLoop objects from multiple threads, which is very not good.
Modifications:
Add an event loop execute to hop loops.
Result:
Thread safe!
Motivation:
Sadly I overlooked the fact that Channel.isActive is supposed to be
safe to call from multiple threads: the implementation here was not.
Modifications:
Store the active state into an Atomic.
Result:
It will be thread-safe to ask if a channel is active.
Motivation:
In some cases users may prefer not to wait for Network.framework to
reattempt connection in the future. Users should be able to opt-out of the
default waiting behaviour in those cases.
Modifications:
- Added WaitForActivity ChannelOption.
Result:
Users can configure channels better.
Motivation:
It is possible to receive the dataReceived callback after we've called
cancel, mostly because there may be such a block in place in the queue
before that cancel call fires. We should tolerate that.
Modifications:
Removed the invalid assertion.
Result:
Fewer failures.
Motivation:
In SwiftNIO 1.11 we shipped improved interfaces for Channels to support
asserting being in the event loop. This is necessary because Dispatch
provides no non-precondition way to 100% guarantee that you are on a
specific queue, which is a requirement for accurate behaviour of
inEventLoop.
Modifications:
- Added an implementation of preconditionInEventLoop.
- Changed all assertions to use new interface.
- Required NIO 1.11.
Result:
Accurate assertions