This allows Apple platforms to override the default stack size of
a thread in the WorkerThreadPool, which is 512KiB by default.
This must be increased, as SPIRV-Cross, used by the Metal driver, can
use deeply nested stacks, as can debug builds.
- The main thread function and the collaborative wait functions have a much more similar structure than earlier, which yields (pun intended) better maintainability.
- Also, there are not assertions anymore about the reason for ending a wait being valid, because spurious awakes can happen and so the assert would fail without that indicating an issue.
Benefits:
- Simpler code. The main load function is renamed so it's apparent that it's not just a thread entry point anymore.
- Cache and thread modes of the original task are honored. A beautiful consequence of this is that, unlike formerly, re-issued loads can use the resource cache, which makes this mechanism much more performant.
- The newly added getter for caller task id in WorkerThreadPool allows to remove the custom tracking of that in ResourceLoader.
- The check to replace a cached resource and the replacement itself happen atomically. That fixes deadlock prevention leading to multiple resource instances of the same one on disk. As a side effect, it also makes the regular check for replace load mode more robust.
ResourceLoader:
- Fix invalid tokens being returned.
- Remove no longer written `ThreadLoadTask::dependent_path` and the code reading from it.
- Clear deadlock hazard by keeping the mutex unlocked during userland polling.
WorkerThreadPool:
- Include thread call queue override in the thread state reset set, which allows to simplify the code that handled that (imperfectly) in the ResourceLoader.
- Handle the mutex type correctly on entering an allowance zone.
CommandQueueMT:
- Handle the additional possibility of command buffer reallocation that mutex unlock allowance introduces.
Just a little optimization.
**NOTE:**
With `RID_Owner` we could replace each pair of `PagedAllocator` and
`HashMap`-of-ids-to-pointers. However, that would force us to expose
`RID` as the task/group id, instead of `int`, which would break the
API. Too bad. Let's wait until Godot 5.0.
This commit lets CommandQueueMT play nicely with the WorkerThreadPool to avoid
non-progressable situations caused by an interdependence between both. While a
command queue is being flushed, it allows the WTP to release its lock while tasks
are being awaited so they can make progress in case they need in turn to post
to the command queue.
This commits rewrites the sync logic in a way that the
`use_system_threads_for_low_priority_tasks` setting, which was added due to
the lack of a cross-platform wait-for-multiple-objects functionality, can be
removed (it's as if it was effectively hardcoded to `false`).
With the new implementation, we have the best of both worlds: threads don't
have to poll, plus no bespoke threads are used.
In addition, regarding deadlock prevention, since not every possible case of
wait-deadlock could be avoided, this commits removes the current best-effort
avoidance mechanisms and keeps only a simple, pessimistic way of detection.
It turns out that the only current user of deadlock prevention, ResourceLoader,
works fine with it and so every possible situation in resource loading is now
properly handled, with no possibilities of deadlocking. There's a comment in
the code with further details.
Lastly, a potential for load tasks never being awaited/disposed is cleared.
- Fix project settings being ignored.
- Made usages of `native_thread_allocator` thread-safe.
- Remove redundant thread-safety from `low_priority_threads_used`, `exit_threads`.
- Fix deadlock due to unintended extra lock of `task_mutex`.
