godot/core/object/worker_thread_pool.cpp

/**************************************************************************/
/*  worker_thread_pool.cpp                                                */
/**************************************************************************/
/*                         This file is part of:                          */
/*                             GODOT ENGINE                               */
/*                        https://godotengine.org                         */
/**************************************************************************/
/* Copyright (c) 2014-present Godot Engine contributors (see AUTHORS.md). */
/* Copyright (c) 2007-2014 Juan Linietsky, Ariel Manzur.                  */
/*                                                                        */
/* Permission is hereby granted, free of charge, to any person obtaining  */
/* a copy of this software and associated documentation files (the        */
/* "Software"), to deal in the Software without restriction, including    */
/* without limitation the rights to use, copy, modify, merge, publish,    */
/* distribute, sublicense, and/or sell copies of the Software, and to     */
/* permit persons to whom the Software is furnished to do so, subject to  */
/* the following conditions:                                              */
/*                                                                        */
/* The above copyright notice and this permission notice shall be         */
/* included in all copies or substantial portions of the Software.        */
/*                                                                        */
/* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,        */
/* EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF     */
/* MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. */
/* IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY   */
/* CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,   */
/* TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE      */
/* SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.                 */
/**************************************************************************/

#include "worker_thread_pool.h"

#include "core/object/script_language.h"
#include "core/os/os.h"
#include "core/os/thread_safe.h"
#include "core/templates/command_queue_mt.h"

void WorkerThreadPool::Task::free_template_userdata() {
	ERR_FAIL_NULL(template_userdata);
	ERR_FAIL_NULL(native_func_userdata);
	BaseTemplateUserdata *btu = (BaseTemplateUserdata *)native_func_userdata;
	memdelete(btu);
}

WorkerThreadPool *WorkerThreadPool::singleton = nullptr;

thread_local CommandQueueMT *WorkerThreadPool::flushing_cmd_queue = nullptr;

void WorkerThreadPool::_process_task(Task *p_task) {
#ifdef THREADS_ENABLED
	int pool_thread_index = thread_ids[Thread::get_caller_id()];
	ThreadData &curr_thread = threads[pool_thread_index];
	Task *prev_task = nullptr; // In case this is recursively called.
	bool safe_for_nodes_backup = is_current_thread_safe_for_nodes();

	{
		// Tasks must start with this unset. They are free to set-and-forget otherwise.
		set_current_thread_safe_for_nodes(false);
		// Since the WorkerThreadPool is started before the script server,
		// its pre-created threads can't have ScriptServer::thread_enter() called on them early.
		// Therefore, we do it late at the first opportunity, so in case the task
		// about to be run uses scripting, guarantees are held.
		if (!curr_thread.ready_for_scripting && ScriptServer::are_languages_initialized()) {
			ScriptServer::thread_enter();
			curr_thread.ready_for_scripting = true;
		}
		task_mutex.lock();
		p_task->pool_thread_index = pool_thread_index;
		prev_task = curr_thread.current_task;
		curr_thread.current_task = p_task;
		task_mutex.unlock();
	}
#endif

	if (p_task->group) {
		// Handling a group
		bool do_post = false;

		while (true) {
			uint32_t work_index = p_task->group->index.postincrement();

			if (work_index >= p_task->group->max) {
				break;
			}
			if (p_task->native_group_func) {
				p_task->native_group_func(p_task->native_func_userdata, work_index);
			} else if (p_task->template_userdata) {
				p_task->template_userdata->callback_indexed(work_index);
			} else {
				p_task->callable.call(work_index);
			}

			// This is the only way to ensure posting is done when all tasks are really complete.
			uint32_t completed_amount = p_task->group->completed_index.increment();

			if (completed_amount == p_task->group->max) {
				do_post = true;
			}
		}

		if (do_post && p_task->template_userdata) {
			memdelete(p_task->template_userdata); // This is no longer needed at this point, so get rid of it.
		}

		if (do_post) {
			p_task->group->done_semaphore.post();
			p_task->group->completed.set_to(true);
		}
		uint32_t max_users = p_task->group->tasks_used + 1; // Add 1 because the thread waiting for it is also user. Read before to avoid another thread freeing task after increment.
		uint32_t finished_users = p_task->group->finished.increment();

		if (finished_users == max_users) {
			// Get rid of the group, because nobody else is using it.
			task_mutex.lock();
			group_allocator.free(p_task->group);
			task_mutex.unlock();
		}

		// For groups, tasks get rid of themselves.

		task_mutex.lock();
		task_allocator.free(p_task);
	} else {
		if (p_task->native_func) {
			p_task->native_func(p_task->native_func_userdata);
		} else if (p_task->template_userdata) {
			p_task->template_userdata->callback();
			memdelete(p_task->template_userdata);
		} else {
			p_task->callable.call();
		}

		task_mutex.lock();
		p_task->completed = true;
		p_task->pool_thread_index = -1;
		if (p_task->waiting_user) {
			p_task->done_semaphore.post(p_task->waiting_user);
		}
		// Let awaiters know.
		for (uint32_t i = 0; i < threads.size(); i++) {
			if (threads[i].awaited_task == p_task) {
				threads[i].cond_var.notify_one();
				threads[i].signaled = true;
			}
		}
	}

#ifdef THREADS_ENABLED
	{
		curr_thread.current_task = prev_task;
		if (p_task->low_priority) {
			low_priority_threads_used--;

			if (_try_promote_low_priority_task()) {
				if (prev_task) { // Otherwise, this thread will catch it.
					_notify_threads(&curr_thread, 1, 0);
				}
			}
		}

		task_mutex.unlock();
	}

	set_current_thread_safe_for_nodes(safe_for_nodes_backup);
#endif
}

void WorkerThreadPool::_thread_function(void *p_user) {
	ThreadData *thread_data = (ThreadData *)p_user;
	while (true) {
		Task *task_to_process = nullptr;
		{
			MutexLock lock(singleton->task_mutex);
			if (singleton->exit_threads) {
				return;
			}
			thread_data->signaled = false;

			if (singleton->task_queue.first()) {
				task_to_process = singleton->task_queue.first()->self();
				singleton->task_queue.remove(singleton->task_queue.first());
			} else {
				thread_data->cond_var.wait(lock);
				DEV_ASSERT(singleton->exit_threads || thread_data->signaled);
			}
		}

		if (task_to_process) {
			singleton->_process_task(task_to_process);
		}
	}
}

void WorkerThreadPool::_post_tasks_and_unlock(Task **p_tasks, uint32_t p_count, bool p_high_priority) {
	// Fall back to processing on the calling thread if there are no worker threads.
	// Separated into its own variable to make it easier to extend this logic
	// in custom builds.
	bool process_on_calling_thread = threads.size() == 0;
	if (process_on_calling_thread) {
		task_mutex.unlock();
		for (uint32_t i = 0; i < p_count; i++) {
			_process_task(p_tasks[i]);
		}
		return;
	}

	uint32_t to_process = 0;
	uint32_t to_promote = 0;

	ThreadData *caller_pool_thread = thread_ids.has(Thread::get_caller_id()) ? &threads[thread_ids[Thread::get_caller_id()]] : nullptr;

	for (uint32_t i = 0; i < p_count; i++) {
		p_tasks[i]->low_priority = !p_high_priority;
		if (p_high_priority || low_priority_threads_used < max_low_priority_threads) {
			task_queue.add_last(&p_tasks[i]->task_elem);
			if (!p_high_priority) {
				low_priority_threads_used++;
			}
			to_process++;
		} else {
			// Too many threads using low priority, must go to queue.
			low_priority_task_queue.add_last(&p_tasks[i]->task_elem);
			to_promote++;
		}
	}

	_notify_threads(caller_pool_thread, to_process, to_promote);

	task_mutex.unlock();
}

void WorkerThreadPool::_notify_threads(const ThreadData *p_current_thread_data, uint32_t p_process_count, uint32_t p_promote_count) {
	uint32_t to_process = p_process_count;
	uint32_t to_promote = p_promote_count;

	// This is where which threads are awaken is decided according to the workload.
	// Threads that will anyway have a chance to check the situation and process/promote tasks
	// are excluded from being notified. Others will be tried anyway to try to distribute load.
	// The current thread, if is a pool thread, is also excluded depending on the promoting/processing
	// needs because it will anyway loop again. However, it will contribute to decreasing the count,
	// which helps reducing sync traffic.

	uint32_t thread_count = threads.size();

	// First round:
	// 1. For processing: notify threads that are not running tasks, to keep the stacks as shallow as possible.
	// 2. For promoting: since it's exclusive with processing, we fin threads able to promote low-prio tasks now.
	for (uint32_t i = 0;
			i < thread_count && (to_process || to_promote);
			i++, notify_index = (notify_index + 1) % thread_count) {
		ThreadData &th = threads[notify_index];

		if (th.signaled) {
			continue;
		}
		if (th.current_task) {
			// Good thread for promoting low-prio?
			if (to_promote && th.awaited_task && th.current_task->low_priority) {
				if (likely(&th != p_current_thread_data)) {
					th.cond_var.notify_one();
				}
				th.signaled = true;
				to_promote--;
			}
		} else {
			if (to_process) {
				if (likely(&th != p_current_thread_data)) {
					th.cond_var.notify_one();
				}
				th.signaled = true;
				to_process--;
			}
		}
	}

	// Second round:
	// For processing: if the first round wasn't enough, let's try now with threads processing tasks but currently awaiting.
	for (uint32_t i = 0;
			i < thread_count && to_process;
			i++, notify_index = (notify_index + 1) % thread_count) {
		ThreadData &th = threads[notify_index];

		if (th.signaled) {
			continue;
		}
		if (th.awaited_task) {
			if (likely(&th != p_current_thread_data)) {
				th.cond_var.notify_one();
			}
			th.signaled = true;
			to_process--;
		}
	}
}

bool WorkerThreadPool::_try_promote_low_priority_task() {
	if (low_priority_task_queue.first()) {
		Task *low_prio_task = low_priority_task_queue.first()->self();
		low_priority_task_queue.remove(low_priority_task_queue.first());
		task_queue.add_last(&low_prio_task->task_elem);
		low_priority_threads_used++;
		return true;
	} else {
		return false;
	}
}

WorkerThreadPool::TaskID WorkerThreadPool::add_native_task(void (*p_func)(void *), void *p_userdata, bool p_high_priority, const String &p_description) {
	return _add_task(Callable(), p_func, p_userdata, nullptr, p_high_priority, p_description);
}

WorkerThreadPool::TaskID WorkerThreadPool::_add_task(const Callable &p_callable, void (*p_func)(void *), void *p_userdata, BaseTemplateUserdata *p_template_userdata, bool p_high_priority, const String &p_description) {
	task_mutex.lock();
	// Get a free task
	Task *task = task_allocator.alloc();
	TaskID id = last_task++;
	task->self = id;
	task->callable = p_callable;
	task->native_func = p_func;
	task->native_func_userdata = p_userdata;
	task->description = p_description;
	task->template_userdata = p_template_userdata;
	tasks.insert(id, task);

	_post_tasks_and_unlock(&task, 1, p_high_priority);

	return id;
}

WorkerThreadPool::TaskID WorkerThreadPool::add_task(const Callable &p_action, bool p_high_priority, const String &p_description) {
	return _add_task(p_action, nullptr, nullptr, nullptr, p_high_priority, p_description);
}

bool WorkerThreadPool::is_task_completed(TaskID p_task_id) const {
	task_mutex.lock();
	const Task *const *taskp = tasks.getptr(p_task_id);
	if (!taskp) {
		task_mutex.unlock();
		ERR_FAIL_V_MSG(false, "Invalid Task ID"); // Invalid task
	}

	bool completed = (*taskp)->completed;
	task_mutex.unlock();

	return completed;
}

Error WorkerThreadPool::wait_for_task_completion(TaskID p_task_id) {
	task_mutex.lock();
	Task **taskp = tasks.getptr(p_task_id);
	if (!taskp) {
		task_mutex.unlock();
		ERR_FAIL_V_MSG(ERR_INVALID_PARAMETER, "Invalid Task ID"); // Invalid task
	}
	Task *task = *taskp;

	if (task->completed) {
		if (task->waiting_pool == 0 && task->waiting_user == 0) {
			tasks.erase(p_task_id);
			task_allocator.free(task);
		}
		task_mutex.unlock();
		return OK;
	}

	ThreadData *caller_pool_thread = thread_ids.has(Thread::get_caller_id()) ? &threads[thread_ids[Thread::get_caller_id()]] : nullptr;
	if (caller_pool_thread && p_task_id <= caller_pool_thread->current_task->self) {
		// Deadlock prevention:
		// When a pool thread wants to wait for an older task, the following situations can happen:
		// 1. Awaited task is deep in the stack of the awaiter.
		// 2. A group of awaiter threads end up depending on some tasks buried in the stack
		//    of their worker threads in such a way that progress can't be made.
		// Both would entail a deadlock. Some may be handled here in the WorkerThreadPool
		// with some extra logic and bookkeeping. However, there would still be unavoidable
		// cases of deadlock because of the way waiting threads process outstanding tasks.
		// Taking into account there's no feasible solution for every possible case
		// with the current design, we just simply reject attempts to await on older tasks,
		// with a specific error code that signals the situation so the caller can handle it.
		task_mutex.unlock();
		return ERR_BUSY;
	}

	if (caller_pool_thread) {
		task->waiting_pool++;
	} else {
		task->waiting_user++;
	}

	task_mutex.unlock();

	if (caller_pool_thread) {
		while (true) {
			Task *task_to_process = nullptr;
			{
				MutexLock lock(task_mutex);
				bool was_signaled = caller_pool_thread->signaled;
				caller_pool_thread->signaled = false;

				if (task->completed) {
					// This thread was awaken also for some reason, but it's about to exit.
					// Let's find out what may be pending and forward the requests.
					if (!exit_threads && was_signaled) {
						uint32_t to_process = task_queue.first() ? 1 : 0;
						uint32_t to_promote = caller_pool_thread->current_task->low_priority && low_priority_task_queue.first() ? 1 : 0;
						if (to_process || to_promote) {
							// This thread must be left alone since it won't loop again.
							caller_pool_thread->signaled = true;
							_notify_threads(caller_pool_thread, to_process, to_promote);
						}
					}

					task->waiting_pool--;
					if (task->waiting_pool == 0 && task->waiting_user == 0) {
						tasks.erase(p_task_id);
						task_allocator.free(task);
					}

					break;
				}

				if (!exit_threads) {
					// This is a thread from the pool. It shouldn't just idle.
					// Let's try to process other tasks while we wait.

					if (caller_pool_thread->current_task->low_priority && low_priority_task_queue.first()) {
						if (_try_promote_low_priority_task()) {
							_notify_threads(caller_pool_thread, 1, 0);
						}
					}

					if (singleton->task_queue.first()) {
						task_to_process = task_queue.first()->self();
						task_queue.remove(task_queue.first());
					}

					if (!task_to_process) {
						caller_pool_thread->awaited_task = task;

						if (flushing_cmd_queue) {
							flushing_cmd_queue->unlock();
						}
						caller_pool_thread->cond_var.wait(lock);
						if (flushing_cmd_queue) {
							flushing_cmd_queue->lock();
						}

						DEV_ASSERT(exit_threads || caller_pool_thread->signaled || task->completed);
						caller_pool_thread->awaited_task = nullptr;
					}
				}
			}

			if (task_to_process) {
				_process_task(task_to_process);
			}
		}
	} else {
		task->done_semaphore.wait();
		task_mutex.lock();
		task->waiting_user--;
		if (task->waiting_pool == 0 && task->waiting_user == 0) {
			tasks.erase(p_task_id);
			task_allocator.free(task);
		}
		task_mutex.unlock();
	}

	return OK;
}

WorkerThreadPool::GroupID WorkerThreadPool::_add_group_task(const Callable &p_callable, void (*p_func)(void *, uint32_t), void *p_userdata, BaseTemplateUserdata *p_template_userdata, int p_elements, int p_tasks, bool p_high_priority, const String &p_description) {
	ERR_FAIL_COND_V(p_elements < 0, INVALID_TASK_ID);
	if (p_tasks < 0) {
		p_tasks = MAX(1u, threads.size());
	}

	task_mutex.lock();
	Group *group = group_allocator.alloc();
	GroupID id = last_task++;
	group->max = p_elements;
	group->self = id;

	Task **tasks_posted = nullptr;
	if (p_elements == 0) {
		// Should really not call it with zero Elements, but at least it should work.
		group->completed.set_to(true);
		group->done_semaphore.post();
		group->tasks_used = 0;
		p_tasks = 0;
		if (p_template_userdata) {
			memdelete(p_template_userdata);
		}

	} else {
		group->tasks_used = p_tasks;
		tasks_posted = (Task **)alloca(sizeof(Task *) * p_tasks);
		for (int i = 0; i < p_tasks; i++) {
			Task *task = task_allocator.alloc();
			task->native_group_func = p_func;
			task->native_func_userdata = p_userdata;
			task->description = p_description;
			task->group = group;
			task->callable = p_callable;
			task->template_userdata = p_template_userdata;
			tasks_posted[i] = task;
			// No task ID is used.
		}
	}

	groups[id] = group;

	_post_tasks_and_unlock(tasks_posted, p_tasks, p_high_priority);

	return id;
}

WorkerThreadPool::GroupID WorkerThreadPool::add_native_group_task(void (*p_func)(void *, uint32_t), void *p_userdata, int p_elements, int p_tasks, bool p_high_priority, const String &p_description) {
	return _add_group_task(Callable(), p_func, p_userdata, nullptr, p_elements, p_tasks, p_high_priority, p_description);
}

WorkerThreadPool::GroupID WorkerThreadPool::add_group_task(const Callable &p_action, int p_elements, int p_tasks, bool p_high_priority, const String &p_description) {
	return _add_group_task(p_action, nullptr, nullptr, nullptr, p_elements, p_tasks, p_high_priority, p_description);
}

uint32_t WorkerThreadPool::get_group_processed_element_count(GroupID p_group) const {
	task_mutex.lock();
	const Group *const *groupp = groups.getptr(p_group);
	if (!groupp) {
		task_mutex.unlock();
		ERR_FAIL_V_MSG(0, "Invalid Group ID");
	}
	uint32_t elements = (*groupp)->completed_index.get();
	task_mutex.unlock();
	return elements;
}
bool WorkerThreadPool::is_group_task_completed(GroupID p_group) const {
	task_mutex.lock();
	const Group *const *groupp = groups.getptr(p_group);
	if (!groupp) {
		task_mutex.unlock();
		ERR_FAIL_V_MSG(false, "Invalid Group ID");
	}
	bool completed = (*groupp)->completed.is_set();
	task_mutex.unlock();
	return completed;
}

void WorkerThreadPool::wait_for_group_task_completion(GroupID p_group) {
#ifdef THREADS_ENABLED
	task_mutex.lock();
	Group **groupp = groups.getptr(p_group);
	task_mutex.unlock();
	if (!groupp) {
		ERR_FAIL_MSG("Invalid Group ID");
	}

	{
		Group *group = *groupp;

		if (flushing_cmd_queue) {
			flushing_cmd_queue->unlock();
		}
		group->done_semaphore.wait();
		if (flushing_cmd_queue) {
			flushing_cmd_queue->lock();
		}

		uint32_t max_users = group->tasks_used + 1; // Add 1 because the thread waiting for it is also user. Read before to avoid another thread freeing task after increment.
		uint32_t finished_users = group->finished.increment(); // fetch happens before inc, so increment later.

		if (finished_users == max_users) {
			// All tasks using this group are gone (finished before the group), so clear the group too.
			task_mutex.lock();
			group_allocator.free(group);
			task_mutex.unlock();
		}
	}

	task_mutex.lock(); // This mutex is needed when Physics 2D and/or 3D is selected to run on a separate thread.
	groups.erase(p_group);
	task_mutex.unlock();
#endif
}

int WorkerThreadPool::get_thread_index() {
	Thread::ID tid = Thread::get_caller_id();
	return singleton->thread_ids.has(tid) ? singleton->thread_ids[tid] : -1;
}

void WorkerThreadPool::thread_enter_command_queue_mt_flush(CommandQueueMT *p_queue) {
	ERR_FAIL_COND(flushing_cmd_queue != nullptr);
	flushing_cmd_queue = p_queue;
}

void WorkerThreadPool::thread_exit_command_queue_mt_flush() {
	ERR_FAIL_NULL(flushing_cmd_queue);
	flushing_cmd_queue = nullptr;
}

void WorkerThreadPool::init(int p_thread_count, float p_low_priority_task_ratio) {
	ERR_FAIL_COND(threads.size() > 0);
	if (p_thread_count < 0) {
		p_thread_count = OS::get_singleton()->get_default_thread_pool_size();
	}

	max_low_priority_threads = CLAMP(p_thread_count * p_low_priority_task_ratio, 1, p_thread_count - 1);

	threads.resize(p_thread_count);

	for (uint32_t i = 0; i < threads.size(); i++) {
		threads[i].index = i;
		threads[i].thread.start(&WorkerThreadPool::_thread_function, &threads[i]);
		thread_ids.insert(threads[i].thread.get_id(), i);
	}
}

void WorkerThreadPool::finish() {
	if (threads.size() == 0) {
		return;
	}

	{
		MutexLock lock(task_mutex);
		SelfList<Task> *E = low_priority_task_queue.first();
		while (E) {
			print_error("Task waiting was never re-claimed: " + E->self()->description);
			E = E->next();
		}
	}

	{
		MutexLock lock(task_mutex);
		exit_threads = true;
	}
	for (ThreadData &data : threads) {
		data.cond_var.notify_one();
	}
	for (ThreadData &data : threads) {
		data.thread.wait_to_finish();
	}

	{
		MutexLock lock(task_mutex);
		for (KeyValue<TaskID, Task *> &E : tasks) {
			task_allocator.free(E.value);
		}
	}

	threads.clear();
}

void WorkerThreadPool::_bind_methods() {
	ClassDB::bind_method(D_METHOD("add_task", "action", "high_priority", "description"), &WorkerThreadPool::add_task, DEFVAL(false), DEFVAL(String()));
	ClassDB::bind_method(D_METHOD("is_task_completed", "task_id"), &WorkerThreadPool::is_task_completed);
	ClassDB::bind_method(D_METHOD("wait_for_task_completion", "task_id"), &WorkerThreadPool::wait_for_task_completion);

	ClassDB::bind_method(D_METHOD("add_group_task", "action", "elements", "tasks_needed", "high_priority", "description"), &WorkerThreadPool::add_group_task, DEFVAL(-1), DEFVAL(false), DEFVAL(String()));
	ClassDB::bind_method(D_METHOD("is_group_task_completed", "group_id"), &WorkerThreadPool::is_group_task_completed);
	ClassDB::bind_method(D_METHOD("get_group_processed_element_count", "group_id"), &WorkerThreadPool::get_group_processed_element_count);
	ClassDB::bind_method(D_METHOD("wait_for_group_task_completion", "group_id"), &WorkerThreadPool::wait_for_group_task_completion);
}

WorkerThreadPool::WorkerThreadPool() {
	singleton = this;
}

WorkerThreadPool::~WorkerThreadPool() {
	finish();
}