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