This reverts commit decbe7f6bd.
Using level triggered events resulted in 100% CPU usage, because sockets receive a constant write-ready event. They register themselves with the event loop only once in the beginning for efficiency reasons and thus may receive these events without any active waiters.
In case of kqueue it would be possible to register dynamically only when waiting, without a big performance impact, but in case of the epoll API this would result in am excessive growth of kernel calls.
Was reporting IOStatus.ok for a connection that was disconnected, possibly resulting in an endless loop (or recursion) of trying to read another (apparently zero sized) packet.
- Drops all events that happen before any changes are made (FSEventStream reports events from the past to a certain degree)
- Tests the same file twice in a row, since FSEventStream might coalesce events and report bogus changes
- Allows "modified" instead of "added" events, because FSEventStream does not allow to distinguish in some cases
This enables efficient integration of the kqueue based I/O processing with Apple OS based UI apps.
On top of that, an FSEvent based directory watcher can now be implemented to replace the inefficient generic watcher that is used on macOS right now.
Since file descriptors are now identified using their validation counter, it is a valid operation to invalidate a slot that is still referenced.
This fixes a related test failure in vibe-core.
isDir can be a huge performance issue when watching a network based directory and can effectively block the thread almost completely in case of frequent file changes. This does mean than high-level code now needs to perform the check manually, if required, and the free information provided by inotify goes unused.
Introduces a "validationCounter" field for all handle types that gets incremented (at least) whenever an OS file descriptor/handle gets invalidated or re-allocated. This way, an old eventcore handle to a reused OS handle can always be distinguished from the current one to avoid interference.
Ensure that only the connect or the connection error callback is ever triggered.
merged-on-behalf-of: Leonid Kramer <l-kramer@users.noreply.github.com>
On macOS it could happen that both, onConnect and onConnectError, were triggered, resulting in seemingly overlapping connection attempts when they really were sequential. This in turn triggered a connection error leak test in vibe-core.
Now using only the write-ready flag plus the reported socket error status to determine failed connections, guaranteeing a single call back.
Malloc should not force '@safe' or '@nogc' on constructors which are not.
For example, TaskPool's ctor is not '@nogc' but was assumed as such
thanks to the delegate cast happening in malloc.
Likewise, the ctor or the arguments might not be '@safe',
or any other attributes, but they were be mistakenly marked as such.
