src/ex/thread_pool.cpp

100.0% Lines (140 / 140) 100.0% Functions (29 / 29)

[ ] prev/next file

thread_pool.cpp

f(x) Functions (29)

Function Calls Lines Blocks

Line	Hits	Source Code
1		//
2		// Copyright (c) 2025 Vinnie Falco (vinnie.falco@gmail.com)
3		// Copyright (c) 2026 Michael Vandeberg
4		//
5		// Distributed under the Boost Software License, Version 1.0. (See accompanying
6		// file LICENSE_1_0.txt or copy at http://www.boost.org/LICENSE_1_0.txt)
7		//
8		// Official repository: https://github.com/boostorg/capy
9		//
10
11		#include <boost/capy/ex/thread_pool.hpp>
12		#include <boost/capy/continuation.hpp>
13		#include <boost/capy/detail/thread_local_ptr.hpp>
14		#include <boost/capy/ex/frame_allocator.hpp>
15		#include <boost/capy/test/thread_name.hpp>
16		#include <algorithm>
17		#include <atomic>
18		#include <condition_variable>
19		#include <cstdio>
20		#include <mutex>
21		#include <thread>
22		#include <vector>
23
24		/*
25		Thread pool implementation using a shared work queue.
26
27		Work items are continuations linked via their intrusive next pointer,
28		stored in a single queue protected by a mutex. No per-post heap
29		allocation: the continuation is owned by the caller and linked
30		directly. Worker threads wait on a condition_variable until work
31		is available or stop is requested.
32
33		Threads are started lazily on first post() via std::call_once to avoid
34		spawning threads for pools that are constructed but never used. Each
35		thread is named with a configurable prefix plus index for debugger
36		visibility.
37
38		Work tracking: on_work_started/on_work_finished maintain the atomic
39		outstanding_work_ counter. on_work_started is lock-free; the worker
40		that drives the count to zero takes mutex_ and re-reads the count
41		before deciding to stop, so the count and the stop decision stay
42		consistent even if work is started in between. join() blocks until
43		this counter reaches zero, then signals workers to stop and joins
44		threads.
45
46		Two shutdown paths:
47		- join(): waits for outstanding work to drain, then stops workers.
48		- stop(): immediately signals workers to exit; queued work is abandoned.
49		- Destructor: stop() then join() (abandon + wait for threads).
50		*/
51
52		namespace boost {
53		namespace capy {
54
55		//------------------------------------------------------------------------------
56
57		class thread_pool::impl
58		{
59		// Identifies the pool owning the current worker thread, or
60		// nullptr if the calling thread is not a pool worker. Checked
61		// by dispatch() to decide between symmetric transfer (inline
62		// resume) and post.
63		static inline detail::thread_local_ptr<impl const> current_;
64
65		// Intrusive queue of continuations: the next link is stored in
66		// continuation::reserved (typed continuation* round-tripped through
67		// void*). No per-post allocation: the continuation is owned by the caller.
68		continuation* head_ = nullptr;
69		continuation* tail_ = nullptr;
70
71	19464x	void push(continuation* c) noexcept
72		{
73	19464x	c->reserved = nullptr;
74	19464x	if(tail_)
75	1510x	tail_->reserved = c;
76		else
77	17954x	head_ = c;
78	19464x	tail_ = c;
79	19464x	}
80
81	19860x	continuation* pop() noexcept
82		{
83	19860x	if(!head_)
84	396x	return nullptr;
85	19464x	continuation* c = head_;
86	19464x	head_ = static_cast<continuation*>(head_->reserved);
87	19464x	if(!head_)
88	17954x	tail_ = nullptr;
89	19464x	return c;
90		}
91
92	38160x	bool empty() const noexcept
93		{
94	38160x	return head_ == nullptr;
95		}
96
97		std::mutex mutex_;
98		std::condition_variable work_cv_;
99		std::condition_variable done_cv_;
100		std::vector<std::thread> threads_;
101		std::atomic<std::size_t> outstanding_work_{0};
102		bool stop_{false};
103		bool joined_{false};
104		std::size_t num_threads_;
105		char thread_name_prefix_[13]{}; // 12 chars max + null terminator
106		std::once_flag start_flag_;
107
108		public:
109	396x	~impl() = default;
110
111		bool
112	684x	running_in_this_thread() const noexcept
113		{
114	684x	return current_.get() == this;
115		}
116
117		// Destroy abandoned coroutine frames. Must be called
118		// before execution_context::shutdown()/destroy() so
119		// that suspended-frame destructors touching services
120		// (e.g. cancelling registrations) run while those
121		// services are still valid.
122		void
123	396x	drain_abandoned() noexcept
124		{
125	571x	while(auto* c = pop())
126		{
127	175x	auto h = c->h;
128	175x	if(h && h != std::noop_coroutine())
129	124x	h.destroy();
130	175x	}
131	396x	}
132
133	396x	impl(std::size_t num_threads, std::string_view thread_name_prefix)
134	396x	: num_threads_(num_threads)
135		{
136	396x	if(num_threads_ == 0)
137	8x	num_threads_ = std::max(
138	4x	std::thread::hardware_concurrency(), 1u);
139
140		// Truncate prefix to 12 chars, leaving room for up to 3-digit index.
141	396x	auto n = thread_name_prefix.copy(thread_name_prefix_, 12);
142	396x	thread_name_prefix_[n] = '\0';
143	396x	}
144
145		void
146	19464x	post(continuation& c)
147		{
148	19464x	ensure_started();
149		{
150	19464x	std::lock_guard<std::mutex> lock(mutex_);
151	19464x	push(&c);
152		// Under the lock so the pool cannot drain, join, and
153		// destroy the condition variable mid-signal.
154	19464x	work_cv_.notify_one();
155	19464x	}
156	19464x	}
157
158		void
159	683x	on_work_started() noexcept
160		{
161	683x	outstanding_work_.fetch_add(1, std::memory_order_acq_rel);
162	683x	}
163
164		void
165	683x	on_work_finished() noexcept
166		{
167	683x	if(outstanding_work_.fetch_sub(
168	683x	1, std::memory_order_acq_rel) == 1)
169		{
170		// fetch_sub's result can be stale: a concurrent
171		// on_work_started() may raise the count before we take the
172		// lock, so re-read it here rather than trust the decrement.
173	320x	std::lock_guard<std::mutex> lock(mutex_);
174	320x	if(outstanding_work_.load(
175	320x	std::memory_order_acquire) == 0 && joined_ && !stop_)
176		{
177	188x	stop_ = true;
178	188x	done_cv_.notify_all();
179	188x	work_cv_.notify_all();
180		}
181	320x	}
182	683x	}
183
184		void
185	658x	join() noexcept
186		{
187		{
188	658x	std::unique_lock<std::mutex> lock(mutex_);
189	658x	if(joined_)
190	262x	return;
191	396x	joined_ = true;
192
193	396x	if(outstanding_work_.load(
194	396x	std::memory_order_acquire) == 0)
195		{
196	157x	stop_ = true;
197	157x	work_cv_.notify_all();
198		}
199		else
200		{
201	239x	done_cv_.wait(lock, [this]{
202	428x	return stop_;
203		});
204		}
205	658x	}
206
207	846x	for(auto& t : threads_)
208	450x	if(t.joinable())
209	450x	t.join();
210		}
211
212		void
213	398x	stop() noexcept
214		{
215		{
216	398x	std::lock_guard<std::mutex> lock(mutex_);
217	398x	stop_ = true;
218	398x	}
219	398x	work_cv_.notify_all();
220	398x	done_cv_.notify_all();
221	398x	}
222
223		private:
224		void
225	19464x	ensure_started()
226		{
227	19464x	std::call_once(start_flag_, [this]{
228	346x	threads_.reserve(num_threads_);
229	796x	for(std::size_t i = 0; i < num_threads_; ++i)
230	900x	threads_.emplace_back([this, i]{ run(i); });
231	346x	});
232	19464x	}
233
234		void
235	450x	run(std::size_t index)
236		{
237		// Build name; set_current_thread_name truncates to platform limits.
238		char name[16];
239	450x	std::snprintf(name, sizeof(name), "%s%zu", thread_name_prefix_, index);
240	450x	set_current_thread_name(name);
241
242		// Mark this thread as a worker of this pool so dispatch()
243		// can symmetric-transfer when called from within pool work.
244		struct scoped_pool
245		{
246	450x	scoped_pool(impl const* p) noexcept { current_.set(p); }
247	450x	~scoped_pool() noexcept { current_.set(nullptr); }
248	450x	} guard(this);
249
250		for(;;)
251		{
252	19739x	continuation* c = nullptr;
253		{
254	19739x	std::unique_lock<std::mutex> lock(mutex_);
255	19739x	work_cv_.wait(lock, [this]{
256	56944x	return !empty() \|\|
257	56944x	stop_;
258		});
259	19739x	if(stop_)
260	900x	return;
261	19289x	c = pop();
262	19739x	}
263	19289x	if(c)
264	19289x	safe_resume(c->h);
265	19289x	}
266	450x	}
267		};
268
269		//------------------------------------------------------------------------------
270
271	396x	thread_pool::
272		~thread_pool()
273		{
274	396x	impl_->stop();
275	396x	impl_->join();
276	396x	impl_->drain_abandoned();
277	396x	shutdown();
278	396x	destroy();
279	396x	delete impl_;
280	396x	}
281
282	396x	thread_pool::
283	396x	thread_pool(std::size_t num_threads, std::string_view thread_name_prefix)
284	396x	: impl_(new impl(num_threads, thread_name_prefix))
285		{
286	396x	this->set_frame_allocator(std::allocator<void>{});
287	396x	}
288
289		void
290	262x	thread_pool::
291		join() noexcept
292		{
293	262x	impl_->join();
294	262x	}
295
296		void
297	2x	thread_pool::
298		stop() noexcept
299		{
300	2x	impl_->stop();
301	2x	}
302
303		//------------------------------------------------------------------------------
304
305		thread_pool::executor_type
306	11808x	thread_pool::
307		get_executor() const noexcept
308		{
309	11808x	return executor_type(
310	11808x	const_cast<thread_pool&>(*this));
311		}
312
313		void
314	683x	thread_pool::executor_type::
315		on_work_started() const noexcept
316		{
317	683x	pool_->impl_->on_work_started();
318	683x	}
319
320		void
321	683x	thread_pool::executor_type::
322		on_work_finished() const noexcept
323		{
324	683x	pool_->impl_->on_work_finished();
325	683x	}
326
327		void
328	18787x	thread_pool::executor_type::
329		post(continuation& c) const
330		{
331	18787x	pool_->impl_->post(c);
332	18787x	}
333
334		std::coroutine_handle<>
335	684x	thread_pool::executor_type::
336		dispatch(continuation& c) const
337		{
338	684x	if(pool_->impl_->running_in_this_thread())
339	7x	return c.h;
340	677x	pool_->impl_->post(c);
341	677x	return std::noop_coroutine();
342		}
343
344		} // capy
345		} // boost
346