. /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   19718 void push(continuation* c) noexcept 71   18745 void push(continuation* c) noexcept
72   { 72   {
HITCBC 73   19718 c->reserved = nullptr; 73   18745 c->reserved = nullptr;
HITCBC 74   19718 if(tail_) 74   18745 if(tail_)
HITCBC 75   5135 tail_->reserved = c; 75   6897 tail_->reserved = c;
76   else 76   else
HITCBC 77   14583 head_ = c; 77   11848 head_ = c;
HITCBC 78   19718 tail_ = c; 78   18745 tail_ = c;
HITCBC 79   19718 } 79   18745 }
80   80  
HITCBC 81   19998 continuation* pop() noexcept 81   19020 continuation* pop() noexcept
82   { 82   {
HITCBC 83   19998 if(!head_) 83   19020 if(!head_)
HITCBC 84   280 return nullptr; 84   275 return nullptr;
HITCBC 85   19718 continuation* c = head_; 85   18745 continuation* c = head_;
HITCBC 86   19718 head_ = static_cast<continuation*>(head_->reserved); 86   18745 head_ = static_cast<continuation*>(head_->reserved);
HITCBC 87   19718 if(!head_) 87   18745 if(!head_)
HITCBC 88   14583 tail_ = nullptr; 88   11848 tail_ = nullptr;
HITCBC 89   19718 return c; 89   18745 return c;
90   } 90   }
91   91  
HITCBC 92   37320 bool empty() const noexcept 92   34447 bool empty() const noexcept
93   { 93   {
HITCBC 94   37320 return head_ == nullptr; 94   34447 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   280 ~impl() = default; 109   275 ~impl() = default;
110   110  
111   bool 111   bool
HITCBC 112   463 running_in_this_thread() const noexcept 112   458 running_in_this_thread() const noexcept
113   { 113   {
HITCBC 114   463 return current_.get() == this; 114   458 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   280 drain_abandoned() noexcept 123   275 drain_abandoned() noexcept
124   { 124   {
HITCBC 125   482 while(auto* c = pop()) 125   461 while(auto* c = pop())
126   { 126   {
HITCBC 127   202 auto h = c->h; 127   186 auto h = c->h;
HITCBC 128   202 if(h && h != std::noop_coroutine()) 128   186 if(h && h != std::noop_coroutine())
HITCBC 129   151 h.destroy(); 129   135 h.destroy();
HITCBC 130   202 } 130   186 }
HITCBC 131   280 } 131   275 }
132   132  
HITCBC 133   280 impl(std::size_t num_threads, std::string_view thread_name_prefix) 133   275 impl(std::size_t num_threads, std::string_view thread_name_prefix)
HITCBC 134   280 : num_threads_(num_threads) 134   275 : num_threads_(num_threads)
135   { 135   {
HITCBC 136   280 if(num_threads_ == 0) 136   275 if(num_threads_ == 0)
HITCBC 137   4 num_threads_ = std::max( 137   4 num_threads_ = std::max(
HITCBC 138   2 std::thread::hardware_concurrency(), 1u); 138   2 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   280 auto n = thread_name_prefix.copy(thread_name_prefix_, 12); 141   275 auto n = thread_name_prefix.copy(thread_name_prefix_, 12);
HITCBC 142   280 thread_name_prefix_[n] = '\0'; 142   275 thread_name_prefix_[n] = '\0';
HITCBC 143   280 } 143   275 }
144   144  
145   void 145   void
HITCBC 146   19718 post(continuation& c) 146   18745 post(continuation& c)
147   { 147   {
HITCBC 148   19718 ensure_started(); 148   18745 ensure_started();
149   { 149   {
HITCBC 150   19718 std::lock_guard<std::mutex> lock(mutex_); 150   18745 std::lock_guard<std::mutex> lock(mutex_);
HITCBC 151   19718 push(&c); 151   18745 push(&c);
HITCBC 152   19718 } 152   18745 }
HITCBC 153   19718 work_cv_.notify_one(); 153   18745 work_cv_.notify_one();
HITCBC 154   19718 } 154   18745 }
155   155  
156   void 156   void
HITCBC 157   463 on_work_started() noexcept 157   458 on_work_started() noexcept
158   { 158   {
HITCBC 159   463 outstanding_work_.fetch_add(1, std::memory_order_acq_rel); 159   458 outstanding_work_.fetch_add(1, std::memory_order_acq_rel);
HITCBC 160   463 } 160   458 }
161   161  
162   void 162   void
HITCBC 163   463 on_work_finished() noexcept 163   458 on_work_finished() noexcept
164   { 164   {
HITCBC 165   463 if(outstanding_work_.fetch_sub( 165   458 if(outstanding_work_.fetch_sub(
HITCBC 166   463 1, std::memory_order_acq_rel) == 1) 166   458 1, std::memory_order_acq_rel) == 1)
167   { 167   {
168   // fetch_sub's result can be stale: a concurrent 168   // fetch_sub's result can be stale: a concurrent
169   // on_work_started() may raise the count before we take the 169   // on_work_started() may raise the count before we take the
170   // lock, so re-read it here rather than trust the decrement. 170   // lock, so re-read it here rather than trust the decrement.
HITCBC 171   211 std::lock_guard<std::mutex> lock(mutex_); 171   206 std::lock_guard<std::mutex> lock(mutex_);
HITCBC 172   211 if(outstanding_work_.load( 172   206 if(outstanding_work_.load(
HITCBC 173   211 std::memory_order_acquire) == 0 && joined_ && !stop_) 173   206 std::memory_order_acquire) == 0 && joined_ && !stop_)
174   { 174   {
HITCBC 175   77 stop_ = true; 175   73 stop_ = true;
HITCBC 176   77 done_cv_.notify_all(); 176   73 done_cv_.notify_all();
HITCBC 177   77 work_cv_.notify_all(); 177   73 work_cv_.notify_all();
178   } 178   }
HITCBC 179   211 } 179   206 }
HITCBC 180   463 } 180   458 }
181   181  
182   void 182   void
HITCBC 183   430 join() noexcept 183   420 join() noexcept
184   { 184   {
185   { 185   {
HITCBC 186   430 std::unique_lock<std::mutex> lock(mutex_); 186   420 std::unique_lock<std::mutex> lock(mutex_);
HITCBC 187   430 if(joined_) 187   420 if(joined_)
HITCBC 188   150 return; 188   145 return;
HITCBC 189   280 joined_ = true; 189   275 joined_ = true;
190   190  
HITCBC 191   280 if(outstanding_work_.load( 191   275 if(outstanding_work_.load(
HITCBC 192   280 std::memory_order_acquire) == 0) 192   275 std::memory_order_acquire) == 0)
193   { 193   {
HITCBC 194   149 stop_ = true; 194   148 stop_ = true;
HITCBC 195   149 work_cv_.notify_all(); 195   148 work_cv_.notify_all();
196   } 196   }
197   else 197   else
198   { 198   {
HITCBC 199   131 done_cv_.wait(lock, [this]{ 199   127 done_cv_.wait(lock, [this]{
HITCBC 200   209 return stop_; 200   201 return stop_;
201   }); 201   });
202   } 202   }
HITCBC 203   430 } 203   420 }
204   204  
HITCBC 205   594 for(auto& t : threads_) 205   584 for(auto& t : threads_)
HITCBC 206   314 if(t.joinable()) 206   309 if(t.joinable())
HITCBC 207   314 t.join(); 207   309 t.join();
208   } 208   }
209   209  
210   void 210   void
HITCBC 211   282 stop() noexcept 211   277 stop() noexcept
212   { 212   {
213   { 213   {
HITCBC 214   282 std::lock_guard<std::mutex> lock(mutex_); 214   277 std::lock_guard<std::mutex> lock(mutex_);
HITCBC 215   282 stop_ = true; 215   277 stop_ = true;
HITCBC 216   282 } 216   277 }
HITCBC 217   282 work_cv_.notify_all(); 217   277 work_cv_.notify_all();
HITCBC 218   282 done_cv_.notify_all(); 218   277 done_cv_.notify_all();
HITCBC 219   282 } 219   277 }
220   220  
221   private: 221   private:
222   void 222   void
HITCBC 223   19718 ensure_started() 223   18745 ensure_started()
224   { 224   {
HITCBC 225   19718 std::call_once(start_flag_, [this]{ 225   18745 std::call_once(start_flag_, [this]{
HITCBC 226   233 threads_.reserve(num_threads_); 226   228 threads_.reserve(num_threads_);
HITCBC 227   547 for(std::size_t i = 0; i < num_threads_; ++i) 227   537 for(std::size_t i = 0; i < num_threads_; ++i)
HITCBC 228   628 threads_.emplace_back([this, i]{ run(i); }); 228   618 threads_.emplace_back([this, i]{ run(i); });
HITCBC 229   233 }); 229   228 });
HITCBC 230   19718 } 230   18745 }
231   231  
232   void 232   void
HITCBC 233   314 run(std::size_t index) 233   309 run(std::size_t index)
234   { 234   {
235   // Build name; set_current_thread_name truncates to platform limits. 235   // Build name; set_current_thread_name truncates to platform limits.
236   char name[16]; 236   char name[16];
HITCBC 237   314 std::snprintf(name, sizeof(name), "%s%zu", thread_name_prefix_, index); 237   309 std::snprintf(name, sizeof(name), "%s%zu", thread_name_prefix_, index);
HITCBC 238   314 set_current_thread_name(name); 238   309 set_current_thread_name(name);
239   239  
240   // Mark this thread as a worker of this pool so dispatch() 240   // Mark this thread as a worker of this pool so dispatch()
241   // can symmetric-transfer when called from within pool work. 241   // can symmetric-transfer when called from within pool work.
242   struct scoped_pool 242   struct scoped_pool
243   { 243   {
HITCBC 244   314 scoped_pool(impl const* p) noexcept { current_.set(p); } 244   309 scoped_pool(impl const* p) noexcept { current_.set(p); }
HITCBC 245   314 ~scoped_pool() noexcept { current_.set(nullptr); } 245   309 ~scoped_pool() noexcept { current_.set(nullptr); }
HITCBC 246   314 } guard(this); 246   309 } guard(this);
247   247  
248   for(;;) 248   for(;;)
249   { 249   {
HITCBC 250   19830 continuation* c = nullptr; 250   18868 continuation* c = nullptr;
251   { 251   {
HITCBC 252   19830 std::unique_lock<std::mutex> lock(mutex_); 252   18868 std::unique_lock<std::mutex> lock(mutex_);
HITCBC 253   19830 work_cv_.wait(lock, [this]{ 253   18868 work_cv_.wait(lock, [this]{
HITCBC 254   55024 return !empty() || 254   50244 return !empty() ||
HITCBC 255   55024 stop_; 255   50244 stop_;
256   }); 256   });
HITCBC 257   19830 if(stop_) 257   18868 if(stop_)
HITCBC 258   628 return; 258   618 return;
HITCBC 259   19516 c = pop(); 259   18559 c = pop();
HITCBC 260   19830 } 260   18868 }
HITCBC 261   19516 if(c) 261   18559 if(c)
HITCBC 262   19516 safe_resume(c->h); 262   18559 safe_resume(c->h);
HITCBC 263   19516 } 263   18559 }
HITCBC 264   314 } 264   309 }
265   }; 265   };
266   266  
267   //------------------------------------------------------------------------------ 267   //------------------------------------------------------------------------------
268   268  
HITCBC 269   280 thread_pool:: 269   275 thread_pool::
270   ~thread_pool() 270   ~thread_pool()
271   { 271   {
HITCBC 272   280 impl_->stop(); 272   275 impl_->stop();
HITCBC 273   280 impl_->join(); 273   275 impl_->join();
HITCBC 274   280 impl_->drain_abandoned(); 274   275 impl_->drain_abandoned();
HITCBC 275   280 shutdown(); 275   275 shutdown();
HITCBC 276   280 destroy(); 276   275 destroy();
HITCBC 277   280 delete impl_; 277   275 delete impl_;
HITCBC 278   280 } 278   275 }
279   279  
HITCBC 280   280 thread_pool:: 280   275 thread_pool::
HITCBC 281   280 thread_pool(std::size_t num_threads, std::string_view thread_name_prefix) 281   275 thread_pool(std::size_t num_threads, std::string_view thread_name_prefix)
HITCBC 282   280 : impl_(new impl(num_threads, thread_name_prefix)) 282   275 : impl_(new impl(num_threads, thread_name_prefix))
283   { 283   {
HITCBC 284   280 this->set_frame_allocator(std::allocator<void>{}); 284   275 this->set_frame_allocator(std::allocator<void>{});
HITCBC 285   280 } 285   275 }
286   286  
287   void 287   void
HITCBC 288   150 thread_pool:: 288   145 thread_pool::
289   join() noexcept 289   join() noexcept
290   { 290   {
HITCBC 291   150 impl_->join(); 291   145 impl_->join();
HITCBC 292   150 } 292   145 }
293   293  
294   void 294   void
HITCBC 295   2 thread_pool:: 295   2 thread_pool::
296   stop() noexcept 296   stop() noexcept
297   { 297   {
HITCBC 298   2 impl_->stop(); 298   2 impl_->stop();
HITCBC 299   2 } 299   2 }
300   300  
301   //------------------------------------------------------------------------------ 301   //------------------------------------------------------------------------------
302   302  
303   thread_pool::executor_type 303   thread_pool::executor_type
HITCBC 304   11686 thread_pool:: 304   11681 thread_pool::
305   get_executor() const noexcept 305   get_executor() const noexcept
306   { 306   {
HITCBC 307   11686 return executor_type( 307   11681 return executor_type(
HITCBC 308   11686 const_cast<thread_pool&>(*this)); 308   11681 const_cast<thread_pool&>(*this));
309   } 309   }
310   310  
311   void 311   void
HITCBC 312   463 thread_pool::executor_type:: 312   458 thread_pool::executor_type::
313   on_work_started() const noexcept 313   on_work_started() const noexcept
314   { 314   {
HITCBC 315   463 pool_->impl_->on_work_started(); 315   458 pool_->impl_->on_work_started();
HITCBC 316   463 } 316   458 }
317   317  
318   void 318   void
HITCBC 319   463 thread_pool::executor_type:: 319   458 thread_pool::executor_type::
320   on_work_finished() const noexcept 320   on_work_finished() const noexcept
321   { 321   {
HITCBC 322   463 pool_->impl_->on_work_finished(); 322   458 pool_->impl_->on_work_finished();
HITCBC 323   463 } 323   458 }
324   324  
325   void 325   void
HITCBC 326   19260 thread_pool::executor_type:: 326   18292 thread_pool::executor_type::
327   post(continuation& c) const 327   post(continuation& c) const
328   { 328   {
HITCBC 329   19260 pool_->impl_->post(c); 329   18292 pool_->impl_->post(c);
HITCBC 330   19260 } 330   18292 }
331   331  
332   std::coroutine_handle<> 332   std::coroutine_handle<>
HITCBC 333   463 thread_pool::executor_type:: 333   458 thread_pool::executor_type::
334   dispatch(continuation& c) const 334   dispatch(continuation& c) const
335   { 335   {
HITCBC 336   463 if(pool_->impl_->running_in_this_thread()) 336   458 if(pool_->impl_->running_in_this_thread())
HITCBC 337   5 return c.h; 337   5 return c.h;
HITCBC 338   458 pool_->impl_->post(c); 338   453 pool_->impl_->post(c);
HITCBC 339   458 return std::noop_coroutine(); 339   453 return std::noop_coroutine();
340   } 340   }
341   341  
342   } // capy 342   } // capy
343   } // boost 343   } // boost