until now, any thread crash would show as a generic crash since we aren't able to get the trace from the crashed thread directly. This uses some dirty tricks to export a partially serialized stack trace to the main thread, where it can be written into a crashdump.
This enables us to see proper crash information for async tasks in the crash archive (finally!!!) as well as being able to capture RakLib errors properly.
closes#5854
Cancelling task runs doesn't make any sense.
- It breaks sequential task execution - later tasks might depend on state from earlier tasks
- It doesn't actually cancel the task - at best, it prevents it from running, but cannot interrupt it (though interrupting a task does not make sense either)
We don't use this "feature" in the core anymore since 22b5e5db5e, as this was causing unexpected behaviour for plugins anyway, along with the occasional shutdown crash due to inconsistent worker states.
Since task execution depends on tasks executing sequentially on a particular worker in some cases (e.g. PopulationTask must be preceded by GeneratorRegisterTask), it doesn't make sense to continue task execution if an error occurs.
Moreover, a task crashing may render the whole server unstable, as it leaves the server in an undefined state. This is the same kind of problem we fixed with scheduled tasks in PM3.
In versions past, pthreads was unreliable enough that random tasks would crash without an obvious reason, forcing us to accommodate this. I still don't know the origin or frequency of said issues, but I think it's time to rip the band-aid off and solve these problems for real.
this is only possible since pthreads 5.1 and pmmpthread
the performance cost of this one ThreadSafeArray allocation is 30% of the total cost of allocating an AsyncTask object on Windows, which is enormous.
In past versions we couldn't lazily initialize it, because the object might get destroyed before the main thread had a chance to dereference it, leading to a crash when collecting completed tasks. This is no longer an issue thanks to object rescue behaviour implemented in pthreads 5.1.
I think this is probably OK in terms of thread-safety, as only one thread writes the property.
the only use for this class is to facilitate random runtime plugin loading, and it's not complete even for that purpose.
Since nothing but PM uses pocketmine/classloader anyway, it doesn't make sense to have it outside the core. As with LogPthreads, it's just adding more maintenance work.
this was previously part of the abandoned package pocketmine/spl. It had to be separated in the PM3 days, because RakLib depended on it.
Since RakLib 0.13, RakLib stopped being dependent on or aware of pthreads, so it no longer depends on any thread-related packages.
It's also possible to absorb pocketmine/snooze and pocketmine/classloader back into the core with this in mind.
because of the way async tasks are processed, we might collect results from a task before it's able to be collected by collect(). Therefore, the queue might be empty, even though there are still tasks to be collected.
regardless of how long an async task takes to run, it will take a multiple of 50ms to get the result processed. This delay causes issues in some cases for stuff like generation, which causes locking of adjacent chunks, and async packet compression, which experiences elevated latency because of this problem.
This is not an ideal solution for packet compression since it will cause the sleeper handler to get hammered, but since it's already getting hammered by every packet from RakLib, I don't think that's a big problem.
