1   
//
1   
//
2   
// Copyright (c) 2025 Vinnie Falco (vinnie.falco@gmail.com)
2   
// Copyright (c) 2025 Vinnie Falco (vinnie.falco@gmail.com)
3   
// Copyright (c) 2026 Michael Vandeberg
3   
// Copyright (c) 2026 Michael Vandeberg
4   
//
4   
//
5   
// Distributed under the Boost Software License, Version 1.0. (See accompanying
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)
6   
// file LICENSE_1_0.txt or copy at http://www.boost.org/LICENSE_1_0.txt)
7   
//
7   
//
8   
// Official repository: https://github.com/boostorg/capy
8   
// Official repository: https://github.com/boostorg/capy
9   
//
9   
//
10  
10  
11   
#include <boost/capy/ex/thread_pool.hpp>
11   
#include <boost/capy/ex/thread_pool.hpp>
12   
#include <boost/capy/detail/intrusive.hpp>
12   
#include <boost/capy/detail/intrusive.hpp>
13   
#include <boost/capy/test/thread_name.hpp>
13   
#include <boost/capy/test/thread_name.hpp>
 
14 + 
#include <atomic>
14   
#include <condition_variable>
15   
#include <condition_variable>
15   
#include <cstdio>
16   
#include <cstdio>
16   
#include <mutex>
17   
#include <mutex>
17   
#include <thread>
18   
#include <thread>
18   
#include <vector>
19   
#include <vector>
19  
20  
20   
/*
21   
/*
21   
    Thread pool implementation using a shared work queue.
22   
    Thread pool implementation using a shared work queue.
22  
23  
23   
    Work items are coroutine handles wrapped in intrusive list nodes, stored
24   
    Work items are coroutine handles wrapped in intrusive list nodes, stored
24   
    in a single queue protected by a mutex. Worker threads wait on a
25   
    in a single queue protected by a mutex. Worker threads wait on a
25   
    condition_variable until work is available or stop is requested.
26   
    condition_variable until work is available or stop is requested.
26  
27  
27   
    Threads are started lazily on first post() via std::call_once to avoid
28   
    Threads are started lazily on first post() via std::call_once to avoid
28   
    spawning threads for pools that are constructed but never used. Each
29   
    spawning threads for pools that are constructed but never used. Each
29   
    thread is named with a configurable prefix plus index for debugger
30   
    thread is named with a configurable prefix plus index for debugger
30   
    visibility.
31   
    visibility.
31  
32  
32   
    Shutdown sequence: stop() sets the stop flag and notifies all threads,
33   
    Shutdown sequence: stop() sets the stop flag and notifies all threads,
33   
    then the destructor joins threads and destroys any remaining queued
34   
    then the destructor joins threads and destroys any remaining queued
34   
    work without executing it.
35   
    work without executing it.
35   
*/
36   
*/
36  
37  
37   
namespace boost {
38   
namespace boost {
38   
namespace capy {
39   
namespace capy {
39  
40  
40   
//------------------------------------------------------------------------------
41   
//------------------------------------------------------------------------------
41  
42  
42   
class thread_pool::impl
43   
class thread_pool::impl
43   
{
44   
{
44   
    struct work : detail::intrusive_queue<work>::node
45   
    struct work : detail::intrusive_queue<work>::node
45   
    {
46   
    {
46   
        std::coroutine_handle<> h_;
47   
        std::coroutine_handle<> h_;
47  
48  
48   
        explicit work(std::coroutine_handle<> h) noexcept
49   
        explicit work(std::coroutine_handle<> h) noexcept
49   
            : h_(h)
50   
            : h_(h)
50   
        {
51   
        {
51   
        }
52   
        }
52  
53  
53   
        void run()
54   
        void run()
54   
        {
55   
        {
55   
            auto h = h_;
56   
            auto h = h_;
56   
            delete this;
57   
            delete this;
57   
            h.resume();
58   
            h.resume();
58   
        }
59   
        }
59  
60  
60   
        void destroy()
61   
        void destroy()
61   
        {
62   
        {
62   
            delete this;
63   
            delete this;
63   
        }
64   
        }
64   
    };
65   
    };
65  
66  
66   
    std::mutex mutex_;
67   
    std::mutex mutex_;
67   
    std::condition_variable cv_;
68   
    std::condition_variable cv_;
68   
    detail::intrusive_queue<work> q_;
69   
    detail::intrusive_queue<work> q_;
69   
    std::vector<std::thread> threads_;
70   
    std::vector<std::thread> threads_;
70 - 
    bool stop_{false};
71 + 
    std::atomic<bool> stop_{false};
71   
    std::size_t num_threads_;
72   
    std::size_t num_threads_;
72   
    char thread_name_prefix_[13]{};  // 12 chars max + null terminator
73   
    char thread_name_prefix_[13]{};  // 12 chars max + null terminator
73   
    std::once_flag start_flag_;
74   
    std::once_flag start_flag_;
74  
75  
75   
public:
76   
public:
76   
    ~impl()
77   
    ~impl()
77   
    {
78   
    {
78   
        stop();
79   
        stop();
79   
        for(auto& t : threads_)
80   
        for(auto& t : threads_)
80   
            if(t.joinable())
81   
            if(t.joinable())
81   
                t.join();
82   
                t.join();
82  
83  
83   
        while(auto* w = q_.pop())
84   
        while(auto* w = q_.pop())
84   
            w->destroy();
85   
            w->destroy();
85   
    }
86   
    }
86  
87  
87   
    impl(std::size_t num_threads, std::string_view thread_name_prefix)
88   
    impl(std::size_t num_threads, std::string_view thread_name_prefix)
88   
        : num_threads_(num_threads)
89   
        : num_threads_(num_threads)
89   
    {
90   
    {
90   
        if(num_threads_ == 0)
91   
        if(num_threads_ == 0)
91   
            num_threads_ = std::thread::hardware_concurrency();
92   
            num_threads_ = std::thread::hardware_concurrency();
92   
        if(num_threads_ == 0)
93   
        if(num_threads_ == 0)
93   
            num_threads_ = 1;
94   
            num_threads_ = 1;
94  
95  
95   
        // Truncate prefix to 12 chars, leaving room for up to 3-digit index.
96   
        // Truncate prefix to 12 chars, leaving room for up to 3-digit index.
96   
        auto n = thread_name_prefix.copy(thread_name_prefix_, 12);
97   
        auto n = thread_name_prefix.copy(thread_name_prefix_, 12);
97   
        thread_name_prefix_[n] = '\0';
98   
        thread_name_prefix_[n] = '\0';
98   
    }
99   
    }
99  
100  
100   
    void
101   
    void
101   
    post(std::coroutine_handle<> h)
102   
    post(std::coroutine_handle<> h)
102   
    {
103   
    {
103   
        ensure_started();
104   
        ensure_started();
104   
        auto* w = new work(h);
105   
        auto* w = new work(h);
105   
        {
106   
        {
106   
            std::lock_guard<std::mutex> lock(mutex_);
107   
            std::lock_guard<std::mutex> lock(mutex_);
107   
            q_.push(w);
108   
            q_.push(w);
108   
        }
109   
        }
109   
        cv_.notify_one();
110   
        cv_.notify_one();
110   
    }
111   
    }
111  
112  
112 - 
    join() noexcept
 
113 - 
    {
 
114 - 
        stop();
 
115 - 
        for(auto& t : threads_)
 
116 - 
            if(t.joinable())
 
117 - 
                t.join();
 
118 - 
    }
 
119 -
 
120 - 
    void
 
121   
    void
113   
    void
122   
    stop() noexcept
114   
    stop() noexcept
123   
    {
115   
    {
124 - 
        {
116 + 
        stop_.store(true, std::memory_order_release);
125 - 
            std::lock_guard<std::mutex> lock(mutex_);
 
126 - 
            stop_ = true;
 
127 - 
        }
 
128   
        cv_.notify_all();
117   
        cv_.notify_all();
129   
    }
118   
    }
130  
119  
131   
private:
120   
private:
132   
    void
121   
    void
133   
    ensure_started()
122   
    ensure_started()
134   
    {
123   
    {
135   
        std::call_once(start_flag_, [this]{
124   
        std::call_once(start_flag_, [this]{
136   
            threads_.reserve(num_threads_);
125   
            threads_.reserve(num_threads_);
137   
            for(std::size_t i = 0; i < num_threads_; ++i)
126   
            for(std::size_t i = 0; i < num_threads_; ++i)
138   
                threads_.emplace_back([this, i]{ run(i); });
127   
                threads_.emplace_back([this, i]{ run(i); });
139   
        });
128   
        });
140   
    }
129   
    }
141  
130  
142   
    void
131   
    void
143   
    run(std::size_t index)
132   
    run(std::size_t index)
144   
    {
133   
    {
145   
        // Build name; set_current_thread_name truncates to platform limits.
134   
        // Build name; set_current_thread_name truncates to platform limits.
146   
        char name[16];
135   
        char name[16];
147   
        std::snprintf(name, sizeof(name), "%s%zu", thread_name_prefix_, index);
136   
        std::snprintf(name, sizeof(name), "%s%zu", thread_name_prefix_, index);
148   
        set_current_thread_name(name);
137   
        set_current_thread_name(name);
149  
138  
150   
        for(;;)
139   
        for(;;)
151   
        {
140   
        {
152   
            work* w = nullptr;
141   
            work* w = nullptr;
153   
            {
142   
            {
154   
                std::unique_lock<std::mutex> lock(mutex_);
143   
                std::unique_lock<std::mutex> lock(mutex_);
155   
                cv_.wait(lock, [this]{
144   
                cv_.wait(lock, [this]{
156   
                    return !q_.empty() ||
145   
                    return !q_.empty() ||
157 - 
                        stop_;
146 + 
                        stop_.load(std::memory_order_acquire);
158   
                });
147   
                });
159 - 
                if(stop_ && q_.empty())
148 + 
                if(stop_.load(std::memory_order_acquire) && q_.empty())
160   
                    return;
149   
                    return;
161   
                w = q_.pop();
150   
                w = q_.pop();
162   
            }
151   
            }
163   
            if(w)
152   
            if(w)
164   
                w->run();
153   
                w->run();
165   
        }
154   
        }
166   
    }
155   
    }
167   
};
156   
};
168  
157  
169   
//------------------------------------------------------------------------------
158   
//------------------------------------------------------------------------------
170  
159  
171   
thread_pool::
160   
thread_pool::
172   
~thread_pool()
161   
~thread_pool()
173 - 
    impl_->join();
 
174   
{
162   
{
175   
    shutdown();
163   
    shutdown();
176   
    destroy();
164   
    destroy();
177   
    delete impl_;
165   
    delete impl_;
178   
}
166   
}
179  
167  
180   
thread_pool::
168   
thread_pool::
181   
thread_pool(std::size_t num_threads, std::string_view thread_name_prefix)
169   
thread_pool(std::size_t num_threads, std::string_view thread_name_prefix)
182   
    : impl_(new impl(num_threads, thread_name_prefix))
170   
    : impl_(new impl(num_threads, thread_name_prefix))
183   
{
171   
{
184   
    this->set_frame_allocator(std::allocator<void>{});
172   
    this->set_frame_allocator(std::allocator<void>{});
185   
}
173   
}
186  
174  
187   
void
175   
void
188   
thread_pool::
176   
thread_pool::
189   
stop() noexcept
177   
stop() noexcept
190   
{
178   
{
191   
    impl_->stop();
179   
    impl_->stop();
192   
}
180   
}
193  
181  
194 -
 
195 - 
thread_pool::executor_type
 
196 - 
thread_pool::
 
197 - 
get_executor() const noexcept
 
198 - 
{
 
199 - 
    return executor_type(
 
200 - 
        const_cast<thread_pool&>(*this));
 
201 - 
}
 
202   
//------------------------------------------------------------------------------
182   
//------------------------------------------------------------------------------
203  
183  
204   
void
184   
void
205   
thread_pool::executor_type::
185   
thread_pool::executor_type::
206   
post(std::coroutine_handle<> h) const
186   
post(std::coroutine_handle<> h) const
207   
{
187   
{
208   
    pool_->impl_->post(h);
188   
    pool_->impl_->post(h);
209   
}
189   
}
210  
190  
211   
} // capy
191   
} // capy
212   
} // boost
192   
} // boost