. /src /ex /thread_pool.cpp
100.00% Lines (140/140) 100.00% Functions (29/29)
[ ] prev/next file n p prev/next uncovered line
TLA Baseline Branch
Line Hits Code Line Hits Code
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/continuation.hpp> 12   #include <boost/capy/continuation.hpp>
13   #include <boost/capy/detail/thread_local_ptr.hpp> 13   #include <boost/capy/detail/thread_local_ptr.hpp>
14   #include <boost/capy/ex/frame_allocator.hpp> 14   #include <boost/capy/ex/frame_allocator.hpp>
15   #include <boost/capy/test/thread_name.hpp> 15   #include <boost/capy/test/thread_name.hpp>
16   #include <algorithm> 16   #include <algorithm>
17   #include <atomic> 17   #include <atomic>
18   #include <condition_variable> 18   #include <condition_variable>
19   #include <cstdio> 19   #include <cstdio>
20   #include <mutex> 20   #include <mutex>
21   #include <thread> 21   #include <thread>
22   #include <vector> 22   #include <vector>
23   23  
24   /* 24   /*
25   Thread pool implementation using a shared work queue. 25   Thread pool implementation using a shared work queue.
26   26  
27   Work items are continuations linked via their intrusive next pointer, 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 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 29   allocation: the continuation is owned by the caller and linked
30   directly. Worker threads wait on a condition_variable until work 30   directly. Worker threads wait on a condition_variable until work
31   is available or stop is requested. 31   is available or stop is requested.
32   32  
33   Threads are started lazily on first post() via std::call_once to avoid 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 34   spawning threads for pools that are constructed but never used. Each
35   thread is named with a configurable prefix plus index for debugger 35   thread is named with a configurable prefix plus index for debugger
36   visibility. 36   visibility.
37   37  
38   Work tracking: on_work_started/on_work_finished maintain the atomic 38   Work tracking: on_work_started/on_work_finished maintain the atomic
39   outstanding_work_ counter. on_work_started is lock-free; the worker 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 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 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 42   consistent even if work is started in between. join() blocks until
43   this counter reaches zero, then signals workers to stop and joins 43   this counter reaches zero, then signals workers to stop and joins
44   threads. 44   threads.
45   45  
46   Two shutdown paths: 46   Two shutdown paths:
47   - join(): waits for outstanding work to drain, then stops workers. 47   - join(): waits for outstanding work to drain, then stops workers.
48   - stop(): immediately signals workers to exit; queued work is abandoned. 48   - stop(): immediately signals workers to exit; queued work is abandoned.
49   - Destructor: stop() then join() (abandon + wait for threads). 49   - Destructor: stop() then join() (abandon + wait for threads).
50   */ 50   */
51   51  
52   namespace boost { 52   namespace boost {
53   namespace capy { 53   namespace capy {
54   54  
55   //------------------------------------------------------------------------------ 55   //------------------------------------------------------------------------------
56   56  
57   class thread_pool::impl 57   class thread_pool::impl
58   { 58   {
59   // Identifies the pool owning the current worker thread, or 59   // Identifies the pool owning the current worker thread, or
60   // nullptr if the calling thread is not a pool worker. Checked 60   // nullptr if the calling thread is not a pool worker. Checked
61   // by dispatch() to decide between symmetric transfer (inline 61   // by dispatch() to decide between symmetric transfer (inline
62   // resume) and post. 62   // resume) and post.
63   static inline detail::thread_local_ptr<impl const> current_; 63   static inline detail::thread_local_ptr<impl const> current_;
64   64  
65   // Intrusive queue of continuations: the next link is stored in 65   // Intrusive queue of continuations: the next link is stored in
66   // continuation::reserved (typed continuation* round-tripped through 66   // continuation::reserved (typed continuation* round-tripped through
67   // void*). No per-post allocation: the continuation is owned by the caller. 67   // void*). No per-post allocation: the continuation is owned by the caller.
68   continuation* head_ = nullptr; 68   continuation* head_ = nullptr;
69   continuation* tail_ = nullptr; 69   continuation* tail_ = nullptr;
70   70  
HITCBC 71   17580 void push(continuation* c) noexcept 71   17834 void push(continuation* c) noexcept
72   { 72   {
HITCBC 73   17580 c->reserved = nullptr; 73   17834 c->reserved = nullptr;
HITCBC 74   17580 if(tail_) 74   17834 if(tail_)
HITCBC 75   4956 tail_->reserved = c; 75   4118 tail_->reserved = c;
76   else 76   else
HITCBC 77   12624 head_ = c; 77   13716 head_ = c;
HITCBC 78   17580 tail_ = c; 78   17834 tail_ = c;
HITCBC 79   17580 } 79   17834 }
80   80  
HITCBC 81   17965 continuation* pop() noexcept 81   18222 continuation* pop() noexcept
82   { 82   {
HITCBC 83   17965 if(!head_) 83   18222 if(!head_)
HITCBC 84   385 return nullptr; 84   388 return nullptr;
HITCBC 85   17580 continuation* c = head_; 85   17834 continuation* c = head_;
HITCBC 86   17580 head_ = static_cast<continuation*>(head_->reserved); 86   17834 head_ = static_cast<continuation*>(head_->reserved);
HITCBC 87   17580 if(!head_) 87   17834 if(!head_)
HITCBC 88   12624 tail_ = nullptr; 88   13716 tail_ = nullptr;
HITCBC 89   17580 return c; 89   17834 return c;
90   } 90   }
91   91  
HITCBC 92   30196 bool empty() const noexcept 92   31661 bool empty() const noexcept
93   { 93   {
HITCBC 94   30196 return head_ == nullptr; 94   31661 return head_ == nullptr;
95   } 95   }
96   96  
97   std::mutex mutex_; 97   std::mutex mutex_;
98   std::condition_variable work_cv_; 98   std::condition_variable work_cv_;
99   std::condition_variable done_cv_; 99   std::condition_variable done_cv_;
100   std::vector<std::thread> threads_; 100   std::vector<std::thread> threads_;
101   std::atomic<std::size_t> outstanding_work_{0}; 101   std::atomic<std::size_t> outstanding_work_{0};
102   bool stop_{false}; 102   bool stop_{false};
103   bool joined_{false}; 103   bool joined_{false};
104   std::size_t num_threads_; 104   std::size_t num_threads_;
105   char thread_name_prefix_[13]{}; // 12 chars max + null terminator 105   char thread_name_prefix_[13]{}; // 12 chars max + null terminator
106   std::once_flag start_flag_; 106   std::once_flag start_flag_;
107   107  
108   public: 108   public:
HITCBC 109   385 ~impl() = default; 109   388 ~impl() = default;
110   110  
111   bool 111   bool
HITCBC 112   673 running_in_this_thread() const noexcept 112   676 running_in_this_thread() const noexcept
113   { 113   {
HITCBC 114   673 return current_.get() == this; 114   676 return current_.get() == this;
115   } 115   }
116   116  
117   // Destroy abandoned coroutine frames. Must be called 117   // Destroy abandoned coroutine frames. Must be called
118   // before execution_context::shutdown()/destroy() so 118   // before execution_context::shutdown()/destroy() so
119   // that suspended-frame destructors touching services 119   // that suspended-frame destructors touching services
120   // (e.g. cancelling registrations) run while those 120   // (e.g. cancelling registrations) run while those
121   // services are still valid. 121   // services are still valid.
122   void 122   void
HITCBC 123   385 drain_abandoned() noexcept 123   388 drain_abandoned() noexcept
124   { 124   {
HITCBC 125   572 while(auto* c = pop()) 125   573 while(auto* c = pop())
126   { 126   {
HITCBC 127   187 auto h = c->h; 127   185 auto h = c->h;
HITCBC 128   187 if(h && h != std::noop_coroutine()) 128   185 if(h && h != std::noop_coroutine())
HITCBC 129   136 h.destroy(); 129   134 h.destroy();
HITCBC 130   187 } 130   185 }
HITCBC 131   385 } 131   388 }
132   132  
HITCBC 133   385 impl(std::size_t num_threads, std::string_view thread_name_prefix) 133   388 impl(std::size_t num_threads, std::string_view thread_name_prefix)
HITCBC 134   385 : num_threads_(num_threads) 134   388 : num_threads_(num_threads)
135   { 135   {
HITCBC 136   385 if(num_threads_ == 0) 136   388 if(num_threads_ == 0)
HITCBC 137   8 num_threads_ = std::max( 137   8 num_threads_ = std::max(
HITCBC 138   4 std::thread::hardware_concurrency(), 1u); 138   4 std::thread::hardware_concurrency(), 1u);
139   139  
140   // Truncate prefix to 12 chars, leaving room for up to 3-digit index. 140   // Truncate prefix to 12 chars, leaving room for up to 3-digit index.
HITCBC 141   385 auto n = thread_name_prefix.copy(thread_name_prefix_, 12); 141   388 auto n = thread_name_prefix.copy(thread_name_prefix_, 12);
HITCBC 142   385 thread_name_prefix_[n] = '\0'; 142   388 thread_name_prefix_[n] = '\0';
HITCBC 143   385 } 143   388 }
144   144  
145   void 145   void
HITCBC 146   17580 post(continuation& c) 146   17834 post(continuation& c)
147   { 147   {
HITCBC 148   17580 ensure_started(); 148   17834 ensure_started();
149   { 149   {
HITCBC 150   17580 std::lock_guard<std::mutex> lock(mutex_); 150   17834 std::lock_guard<std::mutex> lock(mutex_);
HITCBC 151   17580 push(&c); 151   17834 push(&c);
152   // Under the lock so the pool cannot drain, join, and 152   // Under the lock so the pool cannot drain, join, and
153   // destroy the condition variable mid-signal. 153   // destroy the condition variable mid-signal.
HITCBC 154   17580 work_cv_.notify_one(); 154   17834 work_cv_.notify_one();
HITCBC 155   17580 } 155   17834 }
HITCBC 156   17580 } 156   17834 }
157   157  
158   void 158   void
HITCBC 159   672 on_work_started() noexcept 159   675 on_work_started() noexcept
160   { 160   {
HITCBC 161   672 outstanding_work_.fetch_add(1, std::memory_order_acq_rel); 161   675 outstanding_work_.fetch_add(1, std::memory_order_acq_rel);
HITCBC 162   672 } 162   675 }
163   163  
164   void 164   void
HITCBC 165   672 on_work_finished() noexcept 165   675 on_work_finished() noexcept
166   { 166   {
HITCBC 167   672 if(outstanding_work_.fetch_sub( 167   675 if(outstanding_work_.fetch_sub(
HITCBC 168   672 1, std::memory_order_acq_rel) == 1) 168   675 1, std::memory_order_acq_rel) == 1)
169   { 169   {
170   // fetch_sub's result can be stale: a concurrent 170   // fetch_sub's result can be stale: a concurrent
171   // on_work_started() may raise the count before we take the 171   // on_work_started() may raise the count before we take the
172   // lock, so re-read it here rather than trust the decrement. 172   // lock, so re-read it here rather than trust the decrement.
HITCBC 173   309 std::lock_guard<std::mutex> lock(mutex_); 173   312 std::lock_guard<std::mutex> lock(mutex_);
HITCBC 174   309 if(outstanding_work_.load( 174   312 if(outstanding_work_.load(
HITCBC 175   309 std::memory_order_acquire) == 0 && joined_ && !stop_) 175   312 std::memory_order_acquire) == 0 && joined_ && !stop_)
176   { 176   {
HITCBC 177   164 stop_ = true; 177   171 stop_ = true;
HITCBC 178   164 done_cv_.notify_all(); 178   171 done_cv_.notify_all();
HITCBC 179   164 work_cv_.notify_all(); 179   171 work_cv_.notify_all();
180   } 180   }
HITCBC 181   309 } 181   312 }
HITCBC 182   672 } 182   675 }
183   183  
184   void 184   void
HITCBC 185   636 join() noexcept 185   642 join() noexcept
186   { 186   {
187   { 187   {
HITCBC 188   636 std::unique_lock<std::mutex> lock(mutex_); 188   642 std::unique_lock<std::mutex> lock(mutex_);
HITCBC 189   636 if(joined_) 189   642 if(joined_)
HITCBC 190   251 return; 190   254 return;
HITCBC 191   385 joined_ = true; 191   388 joined_ = true;
192   192  
HITCBC 193   385 if(outstanding_work_.load( 193   388 if(outstanding_work_.load(
HITCBC 194   385 std::memory_order_acquire) == 0) 194   388 std::memory_order_acquire) == 0)
195   { 195   {
HITCBC 196   170 stop_ = true; 196   164 stop_ = true;
HITCBC 197   170 work_cv_.notify_all(); 197   164 work_cv_.notify_all();
198   } 198   }
199   else 199   else
200   { 200   {
HITCBC 201   215 done_cv_.wait(lock, [this]{ 201   224 done_cv_.wait(lock, [this]{
HITCBC 202   380 return stop_; 202   396 return stop_;
203   }); 203   });
204   } 204   }
HITCBC 205   636 } 205   642 }
206   206  
HITCBC 207   824 for(auto& t : threads_) 207   830 for(auto& t : threads_)
HITCBC 208   439 if(t.joinable()) 208   442 if(t.joinable())
HITCBC 209   439 t.join(); 209   442 t.join();
210   } 210   }
211   211  
212   void 212   void
HITCBC 213   387 stop() noexcept 213   390 stop() noexcept
214   { 214   {
215   { 215   {
HITCBC 216   387 std::lock_guard<std::mutex> lock(mutex_); 216   390 std::lock_guard<std::mutex> lock(mutex_);
HITCBC 217   387 stop_ = true; 217   390 stop_ = true;
HITCBC 218   387 } 218   390 }
HITCBC 219   387 work_cv_.notify_all(); 219   390 work_cv_.notify_all();
HITCBC 220   387 done_cv_.notify_all(); 220   390 done_cv_.notify_all();
HITCBC 221   387 } 221   390 }
222   222  
223   private: 223   private:
224   void 224   void
HITCBC 225   17580 ensure_started() 225   17834 ensure_started()
226   { 226   {
HITCBC 227   17580 std::call_once(start_flag_, [this]{ 227   17834 std::call_once(start_flag_, [this]{
HITCBC 228   335 threads_.reserve(num_threads_); 228   338 threads_.reserve(num_threads_);
HITCBC 229   774 for(std::size_t i = 0; i < num_threads_; ++i) 229   780 for(std::size_t i = 0; i < num_threads_; ++i)
HITCBC 230   878 threads_.emplace_back([this, i]{ run(i); }); 230   884 threads_.emplace_back([this, i]{ run(i); });
HITCBC 231   335 }); 231   338 });
HITCBC 232   17580 } 232   17834 }
233   233  
234   void 234   void
HITCBC 235   439 run(std::size_t index) 235   442 run(std::size_t index)
236   { 236   {
237   // Build name; set_current_thread_name truncates to platform limits. 237   // Build name; set_current_thread_name truncates to platform limits.
238   char name[16]; 238   char name[16];
HITCBC 239   439 std::snprintf(name, sizeof(name), "%s%zu", thread_name_prefix_, index); 239   442 std::snprintf(name, sizeof(name), "%s%zu", thread_name_prefix_, index);
HITCBC 240   439 set_current_thread_name(name); 240   442 set_current_thread_name(name);
241   241  
242   // Mark this thread as a worker of this pool so dispatch() 242   // Mark this thread as a worker of this pool so dispatch()
243   // can symmetric-transfer when called from within pool work. 243   // can symmetric-transfer when called from within pool work.
244   struct scoped_pool 244   struct scoped_pool
245   { 245   {
HITCBC 246   439 scoped_pool(impl const* p) noexcept { current_.set(p); } 246   442 scoped_pool(impl const* p) noexcept { current_.set(p); }
HITCBC 247   439 ~scoped_pool() noexcept { current_.set(nullptr); } 247   442 ~scoped_pool() noexcept { current_.set(nullptr); }
HITCBC 248   439 } guard(this); 248   442 } guard(this);
249   249  
250   for(;;) 250   for(;;)
251   { 251   {
HITCBC 252   17832 continuation* c = nullptr; 252   18091 continuation* c = nullptr;
253   { 253   {
HITCBC 254   17832 std::unique_lock<std::mutex> lock(mutex_); 254   18091 std::unique_lock<std::mutex> lock(mutex_);
HITCBC 255   17832 work_cv_.wait(lock, [this]{ 255   18091 work_cv_.wait(lock, [this]{
HITCBC 256   42906 return !empty() || 256   45582 return !empty() ||
HITCBC 257   42906 stop_; 257   45582 stop_;
258   }); 258   });
HITCBC 259   17832 if(stop_) 259   18091 if(stop_)
HITCBC 260   878 return; 260   884 return;
HITCBC 261   17393 c = pop(); 261   17649 c = pop();
HITCBC 262   17832 } 262   18091 }
HITCBC 263   17393 if(c) 263   17649 if(c)
HITCBC 264   17393 safe_resume(c->h); 264   17649 safe_resume(c->h);
HITCBC 265   17393 } 265   17649 }
HITCBC 266   439 } 266   442 }
267   }; 267   };
268   268  
269   //------------------------------------------------------------------------------ 269   //------------------------------------------------------------------------------
270   270  
HITCBC 271   385 thread_pool:: 271   388 thread_pool::
272   ~thread_pool() 272   ~thread_pool()
273   { 273   {
HITCBC 274   385 impl_->stop(); 274   388 impl_->stop();
HITCBC 275   385 impl_->join(); 275   388 impl_->join();
HITCBC 276   385 impl_->drain_abandoned(); 276   388 impl_->drain_abandoned();
HITCBC 277   385 shutdown(); 277   388 shutdown();
HITCBC 278   385 destroy(); 278   388 destroy();
HITCBC 279   385 delete impl_; 279   388 delete impl_;
HITCBC 280   385 } 280   388 }
281   281  
HITCBC 282   385 thread_pool:: 282   388 thread_pool::
HITCBC 283   385 thread_pool(std::size_t num_threads, std::string_view thread_name_prefix) 283   388 thread_pool(std::size_t num_threads, std::string_view thread_name_prefix)
HITCBC 284   385 : impl_(new impl(num_threads, thread_name_prefix)) 284   388 : impl_(new impl(num_threads, thread_name_prefix))
285   { 285   {
HITCBC 286   385 this->set_frame_allocator(std::allocator<void>{}); 286   388 this->set_frame_allocator(std::allocator<void>{});
HITCBC 287   385 } 287   388 }
288   288  
289   void 289   void
HITCBC 290   251 thread_pool:: 290   254 thread_pool::
291   join() noexcept 291   join() noexcept
292   { 292   {
HITCBC 293   251 impl_->join(); 293   254 impl_->join();
HITCBC 294   251 } 294   254 }
295   295  
296   void 296   void
HITCBC 297   2 thread_pool:: 297   2 thread_pool::
298   stop() noexcept 298   stop() noexcept
299   { 299   {
HITCBC 300   2 impl_->stop(); 300   2 impl_->stop();
HITCBC 301   2 } 301   2 }
302   302  
303   //------------------------------------------------------------------------------ 303   //------------------------------------------------------------------------------
304   304  
305   thread_pool::executor_type 305   thread_pool::executor_type
HITCBC 306   11797 thread_pool:: 306   11800 thread_pool::
307   get_executor() const noexcept 307   get_executor() const noexcept
308   { 308   {
HITCBC 309   11797 return executor_type( 309   11800 return executor_type(
HITCBC 310   11797 const_cast<thread_pool&>(*this)); 310   11800 const_cast<thread_pool&>(*this));
311   } 311   }
312   312  
313   void 313   void
HITCBC 314   672 thread_pool::executor_type:: 314   675 thread_pool::executor_type::
315   on_work_started() const noexcept 315   on_work_started() const noexcept
316   { 316   {
HITCBC 317   672 pool_->impl_->on_work_started(); 317   675 pool_->impl_->on_work_started();
HITCBC 318   672 } 318   675 }
319   319  
320   void 320   void
HITCBC 321   672 thread_pool::executor_type:: 321   675 thread_pool::executor_type::
322   on_work_finished() const noexcept 322   on_work_finished() const noexcept
323   { 323   {
HITCBC 324   672 pool_->impl_->on_work_finished(); 324   675 pool_->impl_->on_work_finished();
HITCBC 325   672 } 325   675 }
326   326  
327   void 327   void
HITCBC 328   16914 thread_pool::executor_type:: 328   17165 thread_pool::executor_type::
329   post(continuation& c) const 329   post(continuation& c) const
330   { 330   {
HITCBC 331   16914 pool_->impl_->post(c); 331   17165 pool_->impl_->post(c);
HITCBC 332   16914 } 332   17165 }
333   333  
334   std::coroutine_handle<> 334   std::coroutine_handle<>
HITCBC 335   673 thread_pool::executor_type:: 335   676 thread_pool::executor_type::
336   dispatch(continuation& c) const 336   dispatch(continuation& c) const
337   { 337   {
HITCBC 338   673 if(pool_->impl_->running_in_this_thread()) 338   676 if(pool_->impl_->running_in_this_thread())
HITCBC 339   7 return c.h; 339   7 return c.h;
HITCBC 340   666 pool_->impl_->post(c); 340   669 pool_->impl_->post(c);
HITCBC 341   666 return std::noop_coroutine(); 341   669 return std::noop_coroutine();
342   } 342   }
343   343  
344   } // capy 344   } // capy
345   } // boost 345   } // boost