. /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   19828 void push(continuation* c) noexcept 71   18319 void push(continuation* c) noexcept
72   { 72   {
HITCBC 73   19828 c->reserved = nullptr; 73   18319 c->reserved = nullptr;
HITCBC 74   19828 if(tail_) 74   18319 if(tail_)
HITCBC 75   5040 tail_->reserved = c; 75   4179 tail_->reserved = c;
76   else 76   else
HITCBC 77   14788 head_ = c; 77   14140 head_ = c;
HITCBC 78   19828 tail_ = c; 78   18319 tail_ = c;
HITCBC 79   19828 } 79   18319 }
80   80  
HITCBC 81   20229 continuation* pop() noexcept 81   18725 continuation* pop() noexcept
82   { 82   {
HITCBC 83   20229 if(!head_) 83   18725 if(!head_)
HITCBC 84   401 return nullptr; 84   406 return nullptr;
HITCBC 85   19828 continuation* c = head_; 85   18319 continuation* c = head_;
HITCBC 86   19828 head_ = static_cast<continuation*>(head_->reserved); 86   18319 head_ = static_cast<continuation*>(head_->reserved);
HITCBC 87   19828 if(!head_) 87   18319 if(!head_)
HITCBC 88   14788 tail_ = nullptr; 88   14140 tail_ = nullptr;
HITCBC 89   19828 return c; 89   18319 return c;
90   } 90   }
91   91  
HITCBC 92   34964 bool empty() const noexcept 92   32554 bool empty() const noexcept
93   { 93   {
HITCBC 94   34964 return head_ == nullptr; 94   32554 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   401 ~impl() = default; 109   406 ~impl() = default;
110   110  
111   bool 111   bool
HITCBC 112   689 running_in_this_thread() const noexcept 112   694 running_in_this_thread() const noexcept
113   { 113   {
HITCBC 114   689 return current_.get() == this; 114   694 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   401 drain_abandoned() noexcept 123   406 drain_abandoned() noexcept
124   { 124   {
HITCBC 125   604 while(auto* c = pop()) 125   584 while(auto* c = pop())
126   { 126   {
HITCBC 127   203 auto h = c->h; 127   178 auto h = c->h;
HITCBC 128   203 if(h && h != std::noop_coroutine()) 128   178 if(h && h != std::noop_coroutine())
HITCBC 129   152 h.destroy(); 129   127 h.destroy();
HITCBC 130   203 } 130   178 }
HITCBC 131   401 } 131   406 }
132   132  
HITCBC 133   401 impl(std::size_t num_threads, std::string_view thread_name_prefix) 133   406 impl(std::size_t num_threads, std::string_view thread_name_prefix)
HITCBC 134   401 : num_threads_(num_threads) 134   406 : num_threads_(num_threads)
135   { 135   {
HITCBC 136   401 if(num_threads_ == 0) 136   406 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   401 auto n = thread_name_prefix.copy(thread_name_prefix_, 12); 141   406 auto n = thread_name_prefix.copy(thread_name_prefix_, 12);
HITCBC 142   401 thread_name_prefix_[n] = '\0'; 142   406 thread_name_prefix_[n] = '\0';
HITCBC 143   401 } 143   406 }
144   144  
145   void 145   void
HITCBC 146   19828 post(continuation& c) 146   18319 post(continuation& c)
147   { 147   {
HITCBC 148   19828 ensure_started(); 148   18319 ensure_started();
149   { 149   {
HITCBC 150   19828 std::lock_guard<std::mutex> lock(mutex_); 150   18319 std::lock_guard<std::mutex> lock(mutex_);
HITCBC 151   19828 push(&c); 151   18319 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   19828 work_cv_.notify_one(); 154   18319 work_cv_.notify_one();
HITCBC 155   19828 } 155   18319 }
HITCBC 156   19828 } 156   18319 }
157   157  
158   void 158   void
HITCBC 159   688 on_work_started() noexcept 159   693 on_work_started() noexcept
160   { 160   {
HITCBC 161   688 outstanding_work_.fetch_add(1, std::memory_order_acq_rel); 161   693 outstanding_work_.fetch_add(1, std::memory_order_acq_rel);
HITCBC 162   688 } 162   693 }
163   163  
164   void 164   void
HITCBC 165   688 on_work_finished() noexcept 165   693 on_work_finished() noexcept
166   { 166   {
HITCBC 167   688 if(outstanding_work_.fetch_sub( 167   693 if(outstanding_work_.fetch_sub(
HITCBC 168   688 1, std::memory_order_acq_rel) == 1) 168   693 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   326 std::lock_guard<std::mutex> lock(mutex_); 173   330 std::lock_guard<std::mutex> lock(mutex_);
HITCBC 174   326 if(outstanding_work_.load( 174   330 if(outstanding_work_.load(
HITCBC 175   326 std::memory_order_acquire) == 0 && joined_ && !stop_) 175   330 std::memory_order_acquire) == 0 && joined_ && !stop_)
176   { 176   {
HITCBC 177   177 stop_ = true; 177   184 stop_ = true;
HITCBC 178   177 done_cv_.notify_all(); 178   184 done_cv_.notify_all();
HITCBC 179   177 work_cv_.notify_all(); 179   184 work_cv_.notify_all();
180   } 180   }
HITCBC 181   326 } 181   330 }
HITCBC 182   688 } 182   693 }
183   183  
184   void 184   void
HITCBC 185   668 join() noexcept 185   678 join() noexcept
186   { 186   {
187   { 187   {
HITCBC 188   668 std::unique_lock<std::mutex> lock(mutex_); 188   678 std::unique_lock<std::mutex> lock(mutex_);
HITCBC 189   668 if(joined_) 189   678 if(joined_)
HITCBC 190   267 return; 190   272 return;
HITCBC 191   401 joined_ = true; 191   406 joined_ = true;
192   192  
HITCBC 193   401 if(outstanding_work_.load( 193   406 if(outstanding_work_.load(
HITCBC 194   401 std::memory_order_acquire) == 0) 194   406 std::memory_order_acquire) == 0)
195   { 195   {
HITCBC 196   174 stop_ = true; 196   174 stop_ = true;
HITCBC 197   174 work_cv_.notify_all(); 197   174 work_cv_.notify_all();
198   } 198   }
199   else 199   else
200   { 200   {
HITCBC 201   227 done_cv_.wait(lock, [this]{ 201   232 done_cv_.wait(lock, [this]{
HITCBC 202   405 return stop_; 202   417 return stop_;
203   }); 203   });
204   } 204   }
HITCBC 205   668 } 205   678 }
206   206  
HITCBC 207   856 for(auto& t : threads_) 207   866 for(auto& t : threads_)
HITCBC 208   455 if(t.joinable()) 208   460 if(t.joinable())
HITCBC 209   455 t.join(); 209   460 t.join();
210   } 210   }
211   211  
212   void 212   void
HITCBC 213   403 stop() noexcept 213   408 stop() noexcept
214   { 214   {
215   { 215   {
HITCBC 216   403 std::lock_guard<std::mutex> lock(mutex_); 216   408 std::lock_guard<std::mutex> lock(mutex_);
HITCBC 217   403 stop_ = true; 217   408 stop_ = true;
HITCBC 218   403 } 218   408 }
HITCBC 219   403 work_cv_.notify_all(); 219   408 work_cv_.notify_all();
HITCBC 220   403 done_cv_.notify_all(); 220   408 done_cv_.notify_all();
HITCBC 221   403 } 221   408 }
222   222  
223   private: 223   private:
224   void 224   void
HITCBC 225   19828 ensure_started() 225   18319 ensure_started()
226   { 226   {
HITCBC 227   19828 std::call_once(start_flag_, [this]{ 227   18319 std::call_once(start_flag_, [this]{
HITCBC 228   351 threads_.reserve(num_threads_); 228   356 threads_.reserve(num_threads_);
HITCBC 229   806 for(std::size_t i = 0; i < num_threads_; ++i) 229   816 for(std::size_t i = 0; i < num_threads_; ++i)
HITCBC 230   910 threads_.emplace_back([this, i]{ run(i); }); 230   920 threads_.emplace_back([this, i]{ run(i); });
HITCBC 231   351 }); 231   356 });
HITCBC 232   19828 } 232   18319 }
233   233  
234   void 234   void
HITCBC 235   455 run(std::size_t index) 235   460 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   455 std::snprintf(name, sizeof(name), "%s%zu", thread_name_prefix_, index); 239   460 std::snprintf(name, sizeof(name), "%s%zu", thread_name_prefix_, index);
HITCBC 240   455 set_current_thread_name(name); 240   460 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   455 scoped_pool(impl const* p) noexcept { current_.set(p); } 246   460 scoped_pool(impl const* p) noexcept { current_.set(p); }
HITCBC 247   455 ~scoped_pool() noexcept { current_.set(nullptr); } 247   460 ~scoped_pool() noexcept { current_.set(nullptr); }
HITCBC 248   455 } guard(this); 248   460 } guard(this);
249   249  
250   for(;;) 250   for(;;)
251   { 251   {
HITCBC 252   20080 continuation* c = nullptr; 252   18601 continuation* c = nullptr;
253   { 253   {
HITCBC 254   20080 std::unique_lock<std::mutex> lock(mutex_); 254   18601 std::unique_lock<std::mutex> lock(mutex_);
HITCBC 255   20080 work_cv_.wait(lock, [this]{ 255   18601 work_cv_.wait(lock, [this]{
HITCBC 256   50214 return !empty() || 256   46888 return !empty() ||
HITCBC 257   50214 stop_; 257   46888 stop_;
258   }); 258   });
HITCBC 259   20080 if(stop_) 259   18601 if(stop_)
HITCBC 260   910 return; 260   920 return;
HITCBC 261   19625 c = pop(); 261   18141 c = pop();
HITCBC 262   20080 } 262   18601 }
HITCBC 263   19625 if(c) 263   18141 if(c)
HITCBC 264   19625 safe_resume(c->h); 264   18141 safe_resume(c->h);
HITCBC 265   19625 } 265   18141 }
HITCBC 266   455 } 266   460 }
267   }; 267   };
268   268  
269   //------------------------------------------------------------------------------ 269   //------------------------------------------------------------------------------
270   270  
HITCBC 271   401 thread_pool:: 271   406 thread_pool::
272   ~thread_pool() 272   ~thread_pool()
273   { 273   {
HITCBC 274   401 impl_->stop(); 274   406 impl_->stop();
HITCBC 275   401 impl_->join(); 275   406 impl_->join();
HITCBC 276   401 impl_->drain_abandoned(); 276   406 impl_->drain_abandoned();
HITCBC 277   401 shutdown(); 277   406 shutdown();
HITCBC 278   401 destroy(); 278   406 destroy();
HITCBC 279   401 delete impl_; 279   406 delete impl_;
HITCBC 280   401 } 280   406 }
281   281  
HITCBC 282   401 thread_pool:: 282   406 thread_pool::
HITCBC 283   401 thread_pool(std::size_t num_threads, std::string_view thread_name_prefix) 283   406 thread_pool(std::size_t num_threads, std::string_view thread_name_prefix)
HITCBC 284   401 : impl_(new impl(num_threads, thread_name_prefix)) 284   406 : impl_(new impl(num_threads, thread_name_prefix))
285   { 285   {
HITCBC 286   401 this->set_frame_allocator(std::allocator<void>{}); 286   406 this->set_frame_allocator(std::allocator<void>{});
HITCBC 287   401 } 287   406 }
288   288  
289   void 289   void
HITCBC 290   267 thread_pool:: 290   272 thread_pool::
291   join() noexcept 291   join() noexcept
292   { 292   {
HITCBC 293   267 impl_->join(); 293   272 impl_->join();
HITCBC 294   267 } 294   272 }
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   11813 thread_pool:: 306   11818 thread_pool::
307   get_executor() const noexcept 307   get_executor() const noexcept
308   { 308   {
HITCBC 309   11813 return executor_type( 309   11818 return executor_type(
HITCBC 310   11813 const_cast<thread_pool&>(*this)); 310   11818 const_cast<thread_pool&>(*this));
311   } 311   }
312   312  
313   void 313   void
HITCBC 314   688 thread_pool::executor_type:: 314   693 thread_pool::executor_type::
315   on_work_started() const noexcept 315   on_work_started() const noexcept
316   { 316   {
HITCBC 317   688 pool_->impl_->on_work_started(); 317   693 pool_->impl_->on_work_started();
HITCBC 318   688 } 318   693 }
319   319  
320   void 320   void
HITCBC 321   688 thread_pool::executor_type:: 321   693 thread_pool::executor_type::
322   on_work_finished() const noexcept 322   on_work_finished() const noexcept
323   { 323   {
HITCBC 324   688 pool_->impl_->on_work_finished(); 324   693 pool_->impl_->on_work_finished();
HITCBC 325   688 } 325   693 }
326   326  
327   void 327   void
HITCBC 328   19146 thread_pool::executor_type:: 328   17632 thread_pool::executor_type::
329   post(continuation& c) const 329   post(continuation& c) const
330   { 330   {
HITCBC 331   19146 pool_->impl_->post(c); 331   17632 pool_->impl_->post(c);
HITCBC 332   19146 } 332   17632 }
333   333  
334   std::coroutine_handle<> 334   std::coroutine_handle<>
HITCBC 335   689 thread_pool::executor_type:: 335   694 thread_pool::executor_type::
336   dispatch(continuation& c) const 336   dispatch(continuation& c) const
337   { 337   {
HITCBC 338   689 if(pool_->impl_->running_in_this_thread()) 338   694 if(pool_->impl_->running_in_this_thread())
HITCBC 339   7 return c.h; 339   7 return c.h;
HITCBC 340   682 pool_->impl_->post(c); 340   687 pool_->impl_->post(c);
HITCBC 341   682 return std::noop_coroutine(); 341   687 return std::noop_coroutine();
342   } 342   }
343   343  
344   } // capy 344   } // capy
345   } // boost 345   } // boost