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	19973x	void push(continuation* c) noexcept
72		{
73	19973x	c->reserved = nullptr;
74	19973x	if(tail_)
75	4821x	tail_->reserved = c;
76		else
77	15152x	head_ = c;
78	19973x	tail_ = c;
79	19973x	}
80
81	20367x	continuation* pop() noexcept
82		{
83	20367x	if(!head_)
84	394x	return nullptr;
85	19973x	continuation* c = head_;
86	19973x	head_ = static_cast<continuation*>(head_->reserved);
87	19973x	if(!head_)
88	15152x	tail_ = nullptr;
89	19973x	return c;
90		}
91
92	35573x	bool empty() const noexcept
93		{
94	35573x	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	394x	~impl() = default;
110
111		bool
112	682x	running_in_this_thread() const noexcept
113		{
114	682x	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	394x	drain_abandoned() noexcept
124		{
125	553x	while(auto* c = pop())
126		{
127	159x	auto h = c->h;
128	159x	if(h && h != std::noop_coroutine())
129	108x	h.destroy();
130	159x	}
131	394x	}
132
133	394x	impl(std::size_t num_threads, std::string_view thread_name_prefix)
134	394x	: num_threads_(num_threads)
135		{
136	394x	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	394x	auto n = thread_name_prefix.copy(thread_name_prefix_, 12);
142	394x	thread_name_prefix_[n] = '\0';
143	394x	}
144
145		void
146	19973x	post(continuation& c)
147		{
148	19973x	ensure_started();
149		{
150	19973x	std::lock_guard<std::mutex> lock(mutex_);
151	19973x	push(&c);
152		// Under the lock so the pool cannot drain, join, and
153		// destroy the condition variable mid-signal.
154	19973x	work_cv_.notify_one();
155	19973x	}
156	19973x	}
157
158		void
159	681x	on_work_started() noexcept
160		{
161	681x	outstanding_work_.fetch_add(1, std::memory_order_acq_rel);
162	681x	}
163
164		void
165	681x	on_work_finished() noexcept
166		{
167	681x	if(outstanding_work_.fetch_sub(
168	681x	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	318x	std::lock_guard<std::mutex> lock(mutex_);
174	318x	if(outstanding_work_.load(
175	318x	std::memory_order_acquire) == 0 && joined_ && !stop_)
176		{
177	180x	stop_ = true;
178	180x	done_cv_.notify_all();
179	180x	work_cv_.notify_all();
180		}
181	318x	}
182	681x	}
183
184		void
185	654x	join() noexcept
186		{
187		{
188	654x	std::unique_lock<std::mutex> lock(mutex_);
189	654x	if(joined_)
190	260x	return;
191	394x	joined_ = true;
192
193	394x	if(outstanding_work_.load(
194	394x	std::memory_order_acquire) == 0)
195		{
196	164x	stop_ = true;
197	164x	work_cv_.notify_all();
198		}
199		else
200		{
201	230x	done_cv_.wait(lock, [this]{
202	411x	return stop_;
203		});
204		}
205	654x	}
206
207	842x	for(auto& t : threads_)
208	448x	if(t.joinable())
209	448x	t.join();
210		}
211
212		void
213	396x	stop() noexcept
214		{
215		{
216	396x	std::lock_guard<std::mutex> lock(mutex_);
217	396x	stop_ = true;
218	396x	}
219	396x	work_cv_.notify_all();
220	396x	done_cv_.notify_all();
221	396x	}
222
223		private:
224		void
225	19973x	ensure_started()
226		{
227	19973x	std::call_once(start_flag_, [this]{
228	344x	threads_.reserve(num_threads_);
229	792x	for(std::size_t i = 0; i < num_threads_; ++i)
230	896x	threads_.emplace_back([this, i]{ run(i); });
231	344x	});
232	19973x	}
233
234		void
235	448x	run(std::size_t index)
236		{
237		// Build name; set_current_thread_name truncates to platform limits.
238		char name[16];
239	448x	std::snprintf(name, sizeof(name), "%s%zu", thread_name_prefix_, index);
240	448x	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	448x	scoped_pool(impl const* p) noexcept { current_.set(p); }
247	448x	~scoped_pool() noexcept { current_.set(nullptr); }
248	448x	} guard(this);
249
250		for(;;)
251		{
252	20262x	continuation* c = nullptr;
253		{
254	20262x	std::unique_lock<std::mutex> lock(mutex_);
255	20262x	work_cv_.wait(lock, [this]{
256	51253x	return !empty() \|\|
257	51253x	stop_;
258		});
259	20262x	if(stop_)
260	896x	return;
261	19814x	c = pop();
262	20262x	}
263	19814x	if(c)
264	19814x	safe_resume(c->h);
265	19814x	}
266	448x	}
267		};
268
269		//------------------------------------------------------------------------------
270
271	394x	thread_pool::
272		~thread_pool()
273		{
274	394x	impl_->stop();
275	394x	impl_->join();
276	394x	impl_->drain_abandoned();
277	394x	shutdown();
278	394x	destroy();
279	394x	delete impl_;
280	394x	}
281
282	394x	thread_pool::
283	394x	thread_pool(std::size_t num_threads, std::string_view thread_name_prefix)
284	394x	: impl_(new impl(num_threads, thread_name_prefix))
285		{
286	394x	this->set_frame_allocator(std::allocator<void>{});
287	394x	}
288
289		void
290	260x	thread_pool::
291		join() noexcept
292		{
293	260x	impl_->join();
294	260x	}
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	11806x	thread_pool::
307		get_executor() const noexcept
308		{
309	11806x	return executor_type(
310	11806x	const_cast<thread_pool&>(*this));
311		}
312
313		void
314	681x	thread_pool::executor_type::
315		on_work_started() const noexcept
316		{
317	681x	pool_->impl_->on_work_started();
318	681x	}
319
320		void
321	681x	thread_pool::executor_type::
322		on_work_finished() const noexcept
323		{
324	681x	pool_->impl_->on_work_finished();
325	681x	}
326
327		void
328	19298x	thread_pool::executor_type::
329		post(continuation& c) const
330		{
331	19298x	pool_->impl_->post(c);
332	19298x	}
333
334		std::coroutine_handle<>
335	682x	thread_pool::executor_type::
336		dispatch(continuation& c) const
337		{
338	682x	if(pool_->impl_->running_in_this_thread())
339	7x	return c.h;
340	675x	pool_->impl_->post(c);
341	675x	return std::noop_coroutine();
342		}
343
344		} // capy
345		} // boost
346