src/ex/thread_pool.cpp

99.3% Lines (150 / 151) 100.0% Functions (28 / 28)

[ ] prev/next file n p prev/next uncovered line

thread_pool.cpp

f(x) Functions (28)

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/abandon.hpp>
15		#include <boost/capy/ex/frame_allocator.hpp>
16		#include <boost/capy/test/thread_name.hpp>
17		#include <algorithm>
18		#include <atomic>
19		#include <condition_variable>
20		#include <cstdio>
21		#include <mutex>
22		#include <thread>
23		#include <vector>
24
25		/*
26		Thread pool implementation using a shared work queue.
27
28		Work items are continuations linked via their intrusive next pointer,
29		stored in a single queue protected by a mutex. No per-post heap
30		allocation: the continuation is owned by the caller and linked
31		directly. Worker threads wait on a condition_variable until work
32		is available or stop is requested.
33
34		Threads are started lazily on first post(), under mutex_, to avoid
35		spawning threads for pools that are constructed but never used. A
36		stopped pool starts none. Each thread is named
37		with a configurable prefix plus index for debugger visibility.
38
39		Work tracking: on_work_started/on_work_finished maintain the atomic
40		outstanding_work_ counter. on_work_started is lock-free, and so is
41		on_work_finished while the count stays above one. The release that
42		may take it to zero is made under mutex_, so join() can't see zero,
43		return and let the pool be destroyed while that call still uses it.
44		join() blocks until this counter reaches zero, then signals workers
45		to stop and joins threads.
46
47		Two shutdown paths:
48		- join(): waits for outstanding work to drain, then stops workers.
49		- stop(): immediately signals workers to exit; queued work stays queued.
50		- Destructor: stop() then join(), then queued work is abandoned.
51		*/
52
53		namespace boost {
54		namespace capy {
55
56		//------------------------------------------------------------------------------
57
58		class thread_pool::impl
59		{
60		// Identifies the pool owning the current worker thread, or
61		// nullptr if the calling thread is not a pool worker. Checked
62		// by dispatch() to decide between symmetric transfer (inline
63		// resume) and post.
64		static inline detail::thread_local_ptr<impl const> current_;
65
66		// Intrusive queue of continuations: the next link is stored in
67		// continuation::reserved (typed continuation* round-tripped through
68		// void*). No per-post allocation: the continuation is owned by the caller.
69		continuation* head_ = nullptr;
70		continuation* tail_ = nullptr;
71
72	24238x	void push(continuation* c) noexcept
73		{
74	24238x	c->reserved = nullptr;
75	24238x	if(tail_)
76	2132x	tail_->reserved = c;
77		else
78	22106x	head_ = c;
79	24238x	tail_ = c;
80	24238x	}
81
82	27086x	continuation* pop() noexcept
83		{
84	27086x	if(!head_)
85	2848x	return nullptr;
86	24238x	continuation* c = head_;
87	24238x	head_ = static_cast<continuation*>(head_->reserved);
88	24238x	if(!head_)
89	22106x	tail_ = nullptr;
90	24238x	return c;
91		}
92
93	46795x	bool empty() const noexcept
94		{
95	46795x	return head_ == nullptr;
96		}
97
98		std::mutex mutex_;
99		std::condition_variable work_cv_;
100		std::condition_variable done_cv_;
101		std::vector<std::thread> threads_;
102		std::atomic<std::size_t> outstanding_work_{0};
103		bool stop_{false};
104		bool joined_{false};
105		std::size_t num_threads_;
106		char thread_name_prefix_[13]{}; // 12 chars max + null terminator
107		std::atomic<bool> started_{false};
108
109		public:
110	1424x	~impl() = default;
111
112		bool
113	3205x	running_in_this_thread() const noexcept
114		{
115	3205x	return current_.get() == this;
116		}
117
118		// Abandon queued coroutines. Each is taken under the lock, as other
119		// threads may still post, and abandoned outside it, as abandoning
120		// can post.
121		void
122	2848x	drain_abandoned() noexcept
123		{
124		for(;;)
125		{
126		continuation* c;
127		{
128	3417x	std::lock_guard<std::mutex> lock(mutex_);
129	3417x	c = pop();
130	3417x	}
131	3417x	if(!c)
132	2848x	return;
133	569x	if(c->h && c->h != std::noop_coroutine())
134	316x	capy::abandon(*c);
135	569x	}
136		}
137
138	1424x	impl(std::size_t num_threads, std::string_view thread_name_prefix)
139	1424x	: num_threads_(num_threads)
140		{
141	1424x	if(num_threads_ == 0)
142	8x	num_threads_ = std::max(
143	4x	std::thread::hardware_concurrency(), 1u);
144
145		// Truncate prefix to 12 chars, leaving room for up to 3-digit index.
146	1424x	auto n = thread_name_prefix.copy(thread_name_prefix_, 12);
147	1424x	thread_name_prefix_[n] = '\0';
148	1424x	}
149
150		void
151	24240x	post(continuation& c)
152		{
153	24240x	if(!started_.load(std::memory_order_acquire))
154		{
155		{
156	589x	std::lock_guard<std::mutex> lock(mutex_);
157	589x	if(stop_)
158		{
159		// Stopped before any thread started: nothing will
160		// run it, and the destructor's drain abandons it.
161	220x	push(&c);
162	220x	return;
163		}
164	589x	}
165	369x	ensure_started();
166		}
167	24018x	std::lock_guard<std::mutex> lock(mutex_);
168	24018x	push(&c);
169		// Under the lock so the pool cannot drain, join, and
170		// destroy the condition variable mid-signal.
171	24018x	work_cv_.notify_one();
172	24018x	}
173
174		void
175	4082x	on_work_started() noexcept
176		{
177	4082x	outstanding_work_.fetch_add(1, std::memory_order_acq_rel);
178	4082x	}
179
180		void
181	3782x	on_work_finished() noexcept
182		{
183	3782x	auto n = outstanding_work_.load(std::memory_order_relaxed);
184	3791x	while(n > 1)
185		{
186	910x	if(outstanding_work_.compare_exchange_weak(
187		n, n - 1, std::memory_order_acq_rel))
188	446x	return;
189		}
190	3336x	std::lock_guard<std::mutex> lock(mutex_);
191	3336x	outstanding_work_.fetch_sub(1, std::memory_order_acq_rel);
192		// Re-read: a concurrent on_work_started() may have raised it.
193	3336x	if(outstanding_work_.load(
194	3336x	std::memory_order_acquire) == 0 && joined_ && !stop_)
195		{
196	709x	stop_ = true;
197	709x	done_cv_.notify_all();
198	709x	work_cv_.notify_all();
199		}
200	3336x	}
201
202		void
203	2705x	join() noexcept
204		{
205		{
206	2705x	std::unique_lock<std::mutex> lock(mutex_);
207	2705x	if(joined_)
208	1281x	return;
209	1424x	joined_ = true;
210
211	1424x	if(outstanding_work_.load(
212	1424x	std::memory_order_acquire) == 0)
213		{
214	359x	stop_ = true;
215	359x	work_cv_.notify_all();
216		}
217		else
218		{
219	1065x	done_cv_.wait(lock, [this]{
220	2075x	return stop_;
221		});
222		}
223	2705x	}
224
225	1918x	for(auto& t : threads_)
226	494x	if(t.joinable())
227	494x	t.join();
228		}
229
230		void
231	1727x	stop() noexcept
232		{
233		// Under the lock: the join() this releases may return and let
234		// the pool be destroyed as soon as the lock is free.
235	1727x	std::lock_guard<std::mutex> lock(mutex_);
236	1727x	stop_ = true;
237	1727x	work_cv_.notify_all();
238	1727x	done_cv_.notify_all();
239	1727x	}
240
241		private:
242		void
243	369x	ensure_started()
244		{
245		// Under the lock, so join() either sees every thread started
246		// here or makes this start none. A joined pool still starts them:
247		// join() may be waiting on work that needs them.
248	369x	std::lock_guard<std::mutex> lock(mutex_);
249	369x	if(started_.load(std::memory_order_relaxed) \|\| stop_)
250	✗	return;
251		// Start at the threads already running: if an earlier call
252		// threw partway, this one adds only the missing ones.
253	369x	threads_.reserve(num_threads_);
254	863x	for(std::size_t i = threads_.size(); i < num_threads_; ++i)
255	990x	threads_.emplace_back([this, i]{ run(i); });
256	367x	started_.store(true, std::memory_order_release);
257	369x	}
258
259		void
260	494x	run(std::size_t index)
261		{
262		// Build name; set_current_thread_name truncates to platform limits.
263		char name[16];
264	494x	std::snprintf(name, sizeof(name), "%s%zu", thread_name_prefix_, index);
265	494x	set_current_thread_name(name);
266
267		// Mark this thread as a worker of this pool so dispatch()
268		// can symmetric-transfer when called from within pool work.
269		struct scoped_pool
270		{
271	494x	scoped_pool(impl const* p) noexcept { current_.set(p); }
272	494x	~scoped_pool() noexcept { current_.set(nullptr); }
273	494x	} guard(this);
274
275		for(;;)
276		{
277	24163x	continuation* c = nullptr;
278		{
279	24163x	std::unique_lock<std::mutex> lock(mutex_);
280	24163x	work_cv_.wait(lock, [this]{
281	69820x	return !empty() \|\|
282	69820x	stop_;
283		});
284	24163x	if(stop_)
285	988x	return;
286	23669x	c = pop();
287	24163x	}
288	23669x	if(c)
289	23669x	safe_resume(c->h);
290	23669x	}
291	494x	}
292		};
293
294		//------------------------------------------------------------------------------
295
296	1424x	thread_pool::
297		~thread_pool()
298		{
299	1424x	impl_->stop();
300	1424x	impl_->join();
301	1424x	shutdown();
302		// Takes what the unwinds post too, and the work of any service they
303		// create, which is shut down as it is created.
304	1424x	impl_->drain_abandoned();
305	1424x	destroy();
306		// Services give up their work in shutdown(), but anything a
307		// destructor posts anyway is still destroyed.
308	1424x	impl_->drain_abandoned();
309	1424x	delete impl_;
310	1424x	}
311
312	1424x	thread_pool::
313	1424x	thread_pool(std::size_t num_threads, std::string_view thread_name_prefix)
314	1424x	: impl_(new impl(num_threads, thread_name_prefix))
315		{
316	1424x	this->set_frame_allocator(std::allocator<void>{});
317	1424x	}
318
319		void
320	1281x	thread_pool::
321		join() noexcept
322		{
323	1281x	impl_->join();
324	1281x	}
325
326		void
327	303x	thread_pool::
328		stop() noexcept
329		{
330	303x	impl_->stop();
331	303x	}
332
333		//------------------------------------------------------------------------------
334
335		thread_pool::executor_type
336	15343x	thread_pool::
337		get_executor() const noexcept
338		{
339	15343x	return executor_type(
340	15343x	const_cast<thread_pool&>(*this));
341		}
342
343		void
344	4082x	thread_pool::executor_type::
345		on_work_started() const noexcept
346		{
347	4082x	pool_->impl_->on_work_started();
348	4082x	}
349
350		void
351	3782x	thread_pool::executor_type::
352		on_work_finished() const noexcept
353		{
354	3782x	pool_->impl_->on_work_finished();
355	3782x	}
356
357		void
358	21042x	thread_pool::executor_type::
359		post(continuation& c) const
360		{
361	21042x	pool_->impl_->post(c);
362	21040x	}
363
364		std::coroutine_handle<>
365	3205x	thread_pool::executor_type::
366		dispatch(continuation& c) const
367		{
368	3205x	if(pool_->impl_->running_in_this_thread())
369	7x	return c.h;
370	3198x	pool_->impl_->post(c);
371	3198x	return std::noop_coroutine();
372		}
373
374		} // capy
375		} // boost
376