. /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   19254 void push(continuation* c) noexcept 71   19335 void push(continuation* c) noexcept
72   { 72   {
HITCBC 73   19254 c->reserved = nullptr; 73   19335 c->reserved = nullptr;
HITCBC 74   19254 if(tail_) 74   19335 if(tail_)
HITCBC 75   1740 tail_->reserved = c; 75   1727 tail_->reserved = c;
76   else 76   else
HITCBC 77   17514 head_ = c; 77   17608 head_ = c;
HITCBC 78   19254 tail_ = c; 78   19335 tail_ = c;
HITCBC 79   19254 } 79   19335 }
80   80  
HITCBC 81   19655 continuation* pop() noexcept 81   19739 continuation* pop() noexcept
82   { 82   {
HITCBC 83   19655 if(!head_) 83   19739 if(!head_)
HITCBC 84   401 return nullptr; 84   404 return nullptr;
HITCBC 85   19254 continuation* c = head_; 85   19335 continuation* c = head_;
HITCBC 86   19254 head_ = static_cast<continuation*>(head_->reserved); 86   19335 head_ = static_cast<continuation*>(head_->reserved);
HITCBC 87   19254 if(!head_) 87   19335 if(!head_)
HITCBC 88   17514 tail_ = nullptr; 88   17608 tail_ = nullptr;
HITCBC 89   19254 return c; 89   19335 return c;
90   } 90   }
91   91  
HITCBC 92   37750 bool empty() const noexcept 92   37592 bool empty() const noexcept
93   { 93   {
HITCBC 94   37750 return head_ == nullptr; 94   37592 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   404 ~impl() = default;
110   110  
111   bool 111   bool
HITCBC 112   689 running_in_this_thread() const noexcept 112   692 running_in_this_thread() const noexcept
113   { 113   {
HITCBC 114   689 return current_.get() == this; 114   692 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   404 drain_abandoned() noexcept
124   { 124   {
HITCBC 125   595 while(auto* c = pop()) 125   606 while(auto* c = pop())
126   { 126   {
HITCBC 127   194 auto h = c->h; 127   202 auto h = c->h;
HITCBC 128   194 if(h && h != std::noop_coroutine()) 128   202 if(h && h != std::noop_coroutine())
HITCBC 129   143 h.destroy(); 129   151 h.destroy();
HITCBC 130   194 } 130   202 }
HITCBC 131   401 } 131   404 }
132   132  
HITCBC 133   401 impl(std::size_t num_threads, std::string_view thread_name_prefix) 133   404 impl(std::size_t num_threads, std::string_view thread_name_prefix)
HITCBC 134   401 : num_threads_(num_threads) 134   404 : num_threads_(num_threads)
135   { 135   {
HITCBC 136   401 if(num_threads_ == 0) 136   404 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   404 auto n = thread_name_prefix.copy(thread_name_prefix_, 12);
HITCBC 142   401 thread_name_prefix_[n] = '\0'; 142   404 thread_name_prefix_[n] = '\0';
HITCBC 143   401 } 143   404 }
144   144  
145   void 145   void
HITCBC 146   19254 post(continuation& c) 146   19335 post(continuation& c)
147   { 147   {
HITCBC 148   19254 ensure_started(); 148   19335 ensure_started();
149   { 149   {
HITCBC 150   19254 std::lock_guard<std::mutex> lock(mutex_); 150   19335 std::lock_guard<std::mutex> lock(mutex_);
HITCBC 151   19254 push(&c); 151   19335 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   19254 work_cv_.notify_one(); 154   19335 work_cv_.notify_one();
HITCBC 155   19254 } 155   19335 }
HITCBC 156   19254 } 156   19335 }
157   157  
158   void 158   void
HITCBC 159   688 on_work_started() noexcept 159   691 on_work_started() noexcept
160   { 160   {
HITCBC 161   688 outstanding_work_.fetch_add(1, std::memory_order_acq_rel); 161   691 outstanding_work_.fetch_add(1, std::memory_order_acq_rel);
HITCBC 162   688 } 162   691 }
163   163  
164   void 164   void
HITCBC 165   688 on_work_finished() noexcept 165   691 on_work_finished() noexcept
166   { 166   {
HITCBC 167   688 if(outstanding_work_.fetch_sub( 167   691 if(outstanding_work_.fetch_sub(
HITCBC 168   688 1, std::memory_order_acq_rel) == 1) 168   691 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   325 std::lock_guard<std::mutex> lock(mutex_); 173   328 std::lock_guard<std::mutex> lock(mutex_);
HITCBC 174   325 if(outstanding_work_.load( 174   328 if(outstanding_work_.load(
HITCBC 175   325 std::memory_order_acquire) == 0 && joined_ && !stop_) 175   328 std::memory_order_acquire) == 0 && joined_ && !stop_)
176   { 176   {
HITCBC 177   186 stop_ = true; 177   184 stop_ = true;
HITCBC 178   186 done_cv_.notify_all(); 178   184 done_cv_.notify_all();
HITCBC 179   186 work_cv_.notify_all(); 179   184 work_cv_.notify_all();
180   } 180   }
HITCBC 181   325 } 181   328 }
HITCBC 182   688 } 182   691 }
183   183  
184   void 184   void
HITCBC 185   668 join() noexcept 185   674 join() noexcept
186   { 186   {
187   { 187   {
HITCBC 188   668 std::unique_lock<std::mutex> lock(mutex_); 188   674 std::unique_lock<std::mutex> lock(mutex_);
HITCBC 189   668 if(joined_) 189   674 if(joined_)
HITCBC 190   267 return; 190   270 return;
HITCBC 191   401 joined_ = true; 191   404 joined_ = true;
192   192  
HITCBC 193   401 if(outstanding_work_.load( 193   404 if(outstanding_work_.load(
HITCBC 194   401 std::memory_order_acquire) == 0) 194   404 std::memory_order_acquire) == 0)
195   { 195   {
HITCBC 196   162 stop_ = true; 196   168 stop_ = true;
HITCBC 197   162 work_cv_.notify_all(); 197   168 work_cv_.notify_all();
198   } 198   }
199   else 199   else
200   { 200   {
HITCBC 201   239 done_cv_.wait(lock, [this]{ 201   236 done_cv_.wait(lock, [this]{
HITCBC 202   426 return stop_; 202   421 return stop_;
203   }); 203   });
204   } 204   }
HITCBC 205   668 } 205   674 }
206   206  
HITCBC 207   856 for(auto& t : threads_) 207   862 for(auto& t : threads_)
HITCBC 208   455 if(t.joinable()) 208   458 if(t.joinable())
HITCBC 209   455 t.join(); 209   458 t.join();
210   } 210   }
211   211  
212   void 212   void
HITCBC 213   403 stop() noexcept 213   406 stop() noexcept
214   { 214   {
215   { 215   {
HITCBC 216   403 std::lock_guard<std::mutex> lock(mutex_); 216   406 std::lock_guard<std::mutex> lock(mutex_);
HITCBC 217   403 stop_ = true; 217   406 stop_ = true;
HITCBC 218   403 } 218   406 }
HITCBC 219   403 work_cv_.notify_all(); 219   406 work_cv_.notify_all();
HITCBC 220   403 done_cv_.notify_all(); 220   406 done_cv_.notify_all();
HITCBC 221   403 } 221   406 }
222   222  
223   private: 223   private:
224   void 224   void
HITCBC 225   19254 ensure_started() 225   19335 ensure_started()
226   { 226   {
HITCBC 227   19254 std::call_once(start_flag_, [this]{ 227   19335 std::call_once(start_flag_, [this]{
HITCBC 228   351 threads_.reserve(num_threads_); 228   354 threads_.reserve(num_threads_);
HITCBC 229   806 for(std::size_t i = 0; i < num_threads_; ++i) 229   812 for(std::size_t i = 0; i < num_threads_; ++i)
HITCBC 230   910 threads_.emplace_back([this, i]{ run(i); }); 230   916 threads_.emplace_back([this, i]{ run(i); });
HITCBC 231   351 }); 231   354 });
HITCBC 232   19254 } 232   19335 }
233   233  
234   void 234   void
HITCBC 235   455 run(std::size_t index) 235   458 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   458 std::snprintf(name, sizeof(name), "%s%zu", thread_name_prefix_, index);
HITCBC 240   455 set_current_thread_name(name); 240   458 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   458 scoped_pool(impl const* p) noexcept { current_.set(p); }
HITCBC 247   455 ~scoped_pool() noexcept { current_.set(nullptr); } 247   458 ~scoped_pool() noexcept { current_.set(nullptr); }
HITCBC 248   455 } guard(this); 248   458 } guard(this);
249   249  
250   for(;;) 250   for(;;)
251   { 251   {
HITCBC 252   19515 continuation* c = nullptr; 252   19591 continuation* c = nullptr;
253   { 253   {
HITCBC 254   19515 std::unique_lock<std::mutex> lock(mutex_); 254   19591 std::unique_lock<std::mutex> lock(mutex_);
HITCBC 255   19515 work_cv_.wait(lock, [this]{ 255   19591 work_cv_.wait(lock, [this]{
HITCBC 256   56343 return !empty() || 256   55958 return !empty() ||
HITCBC 257   56343 stop_; 257   55958 stop_;
258   }); 258   });
HITCBC 259   19515 if(stop_) 259   19591 if(stop_)
HITCBC 260   910 return; 260   916 return;
HITCBC 261   19060 c = pop(); 261   19133 c = pop();
HITCBC 262   19515 } 262   19591 }
HITCBC 263   19060 if(c) 263   19133 if(c)
HITCBC 264   19060 safe_resume(c->h); 264   19133 safe_resume(c->h);
HITCBC 265   19060 } 265   19133 }
HITCBC 266   455 } 266   458 }
267   }; 267   };
268   268  
269   //------------------------------------------------------------------------------ 269   //------------------------------------------------------------------------------
270   270  
HITCBC 271   401 thread_pool:: 271   404 thread_pool::
272   ~thread_pool() 272   ~thread_pool()
273   { 273   {
HITCBC 274   401 impl_->stop(); 274   404 impl_->stop();
HITCBC 275   401 impl_->join(); 275   404 impl_->join();
HITCBC 276   401 impl_->drain_abandoned(); 276   404 impl_->drain_abandoned();
HITCBC 277   401 shutdown(); 277   404 shutdown();
HITCBC 278   401 destroy(); 278   404 destroy();
HITCBC 279   401 delete impl_; 279   404 delete impl_;
HITCBC 280   401 } 280   404 }
281   281  
HITCBC 282   401 thread_pool:: 282   404 thread_pool::
HITCBC 283   401 thread_pool(std::size_t num_threads, std::string_view thread_name_prefix) 283   404 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   404 : impl_(new impl(num_threads, thread_name_prefix))
285   { 285   {
HITCBC 286   401 this->set_frame_allocator(std::allocator<void>{}); 286   404 this->set_frame_allocator(std::allocator<void>{});
HITCBC 287   401 } 287   404 }
288   288  
289   void 289   void
HITCBC 290   267 thread_pool:: 290   270 thread_pool::
291   join() noexcept 291   join() noexcept
292   { 292   {
HITCBC 293   267 impl_->join(); 293   270 impl_->join();
HITCBC 294   267 } 294   270 }
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   11816 thread_pool::
307   get_executor() const noexcept 307   get_executor() const noexcept
308   { 308   {
HITCBC 309   11813 return executor_type( 309   11816 return executor_type(
HITCBC 310   11813 const_cast<thread_pool&>(*this)); 310   11816 const_cast<thread_pool&>(*this));
311   } 311   }
312   312  
313   void 313   void
HITCBC 314   688 thread_pool::executor_type:: 314   691 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   691 pool_->impl_->on_work_started();
HITCBC 318   688 } 318   691 }
319   319  
320   void 320   void
HITCBC 321   688 thread_pool::executor_type:: 321   691 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   691 pool_->impl_->on_work_finished();
HITCBC 325   688 } 325   691 }
326   326  
327   void 327   void
HITCBC 328   18572 thread_pool::executor_type:: 328   18650 thread_pool::executor_type::
329   post(continuation& c) const 329   post(continuation& c) const
330   { 330   {
HITCBC 331   18572 pool_->impl_->post(c); 331   18650 pool_->impl_->post(c);
HITCBC 332   18572 } 332   18650 }
333   333  
334   std::coroutine_handle<> 334   std::coroutine_handle<>
HITCBC 335   689 thread_pool::executor_type:: 335   692 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   692 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   685 pool_->impl_->post(c);
HITCBC 341   682 return std::noop_coroutine(); 341   685 return std::noop_coroutine();
342   } 342   }
343   343  
344   } // capy 344   } // capy
345   } // boost 345   } // boost