. /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   20549 void push(continuation* c) noexcept 71   19442 void push(continuation* c) noexcept
72   { 72   {
HITCBC 73   20549 c->reserved = nullptr; 73   19442 c->reserved = nullptr;
HITCBC 74   20549 if(tail_) 74   19442 if(tail_)
HITCBC 75   3515 tail_->reserved = c; 75   5396 tail_->reserved = c;
76   else 76   else
HITCBC 77   17034 head_ = c; 77   14046 head_ = c;
HITCBC 78   20549 tail_ = c; 78   19442 tail_ = c;
HITCBC 79   20549 } 79   19442 }
80   80  
HITCBC 81   20949 continuation* pop() noexcept 81   19837 continuation* pop() noexcept
82   { 82   {
HITCBC 83   20949 if(!head_) 83   19837 if(!head_)
HITCBC 84   400 return nullptr; 84   395 return nullptr;
HITCBC 85   20549 continuation* c = head_; 85   19442 continuation* c = head_;
HITCBC 86   20549 head_ = static_cast<continuation*>(head_->reserved); 86   19442 head_ = static_cast<continuation*>(head_->reserved);
HITCBC 87   20549 if(!head_) 87   19442 if(!head_)
HITCBC 88   17034 tail_ = nullptr; 88   14046 tail_ = nullptr;
HITCBC 89   20549 return c; 89   19442 return c;
90   } 90   }
91   91  
HITCBC 92   37945 bool empty() const noexcept 92   33804 bool empty() const noexcept
93   { 93   {
HITCBC 94   37945 return head_ == nullptr; 94   33804 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   400 ~impl() = default; 109   395 ~impl() = default;
110   110  
111   bool 111   bool
HITCBC 112   688 running_in_this_thread() const noexcept 112   679 running_in_this_thread() const noexcept
113   { 113   {
HITCBC 114   688 return current_.get() == this; 114   679 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   400 drain_abandoned() noexcept 123   395 drain_abandoned() noexcept
124   { 124   {
HITCBC 125   580 while(auto* c = pop()) 125   582 while(auto* c = pop())
126   { 126   {
HITCBC 127   180 auto h = c->h; 127   187 auto h = c->h;
HITCBC 128   180 if(h && h != std::noop_coroutine()) 128   187 if(h && h != std::noop_coroutine())
HITCBC 129   129 h.destroy(); 129   136 h.destroy();
HITCBC 130   180 } 130   187 }
HITCBC 131   400 } 131   395 }
132   132  
HITCBC 133   400 impl(std::size_t num_threads, std::string_view thread_name_prefix) 133   395 impl(std::size_t num_threads, std::string_view thread_name_prefix)
HITCBC 134   400 : num_threads_(num_threads) 134   395 : num_threads_(num_threads)
135   { 135   {
HITCBC 136   400 if(num_threads_ == 0) 136   395 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   400 auto n = thread_name_prefix.copy(thread_name_prefix_, 12); 141   395 auto n = thread_name_prefix.copy(thread_name_prefix_, 12);
HITCBC 142   400 thread_name_prefix_[n] = '\0'; 142   395 thread_name_prefix_[n] = '\0';
HITCBC 143   400 } 143   395 }
144   144  
145   void 145   void
HITCBC 146   20549 post(continuation& c) 146   19442 post(continuation& c)
147   { 147   {
HITCBC 148   20549 ensure_started(); 148   19442 ensure_started();
149   { 149   {
HITCBC 150   20549 std::lock_guard<std::mutex> lock(mutex_); 150   19442 std::lock_guard<std::mutex> lock(mutex_);
HITCBC 151   20549 push(&c); 151   19442 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   20549 work_cv_.notify_one(); 154   19442 work_cv_.notify_one();
HITCBC 155   20549 } 155   19442 }
HITCBC 156   20549 } 156   19442 }
157   157  
158   void 158   void
HITCBC 159   687 on_work_started() noexcept 159   678 on_work_started() noexcept
160   { 160   {
HITCBC 161   687 outstanding_work_.fetch_add(1, std::memory_order_acq_rel); 161   678 outstanding_work_.fetch_add(1, std::memory_order_acq_rel);
HITCBC 162   687 } 162   678 }
163   163  
164   void 164   void
HITCBC 165   687 on_work_finished() noexcept 165   678 on_work_finished() noexcept
166   { 166   {
HITCBC 167   687 if(outstanding_work_.fetch_sub( 167   678 if(outstanding_work_.fetch_sub(
HITCBC 168   687 1, std::memory_order_acq_rel) == 1) 168   678 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   324 std::lock_guard<std::mutex> lock(mutex_); 173   316 std::lock_guard<std::mutex> lock(mutex_);
HITCBC 174   324 if(outstanding_work_.load( 174   316 if(outstanding_work_.load(
HITCBC 175   324 std::memory_order_acquire) == 0 && joined_ && !stop_) 175   316 std::memory_order_acquire) == 0 && joined_ && !stop_)
176   { 176   {
HITCBC 177   175 stop_ = true; 177   185 stop_ = true;
HITCBC 178   175 done_cv_.notify_all(); 178   185 done_cv_.notify_all();
HITCBC 179   175 work_cv_.notify_all(); 179   185 work_cv_.notify_all();
180   } 180   }
HITCBC 181   324 } 181   316 }
HITCBC 182   687 } 182   678 }
183   183  
184   void 184   void
HITCBC 185   666 join() noexcept 185   652 join() noexcept
186   { 186   {
187   { 187   {
HITCBC 188   666 std::unique_lock<std::mutex> lock(mutex_); 188   652 std::unique_lock<std::mutex> lock(mutex_);
HITCBC 189   666 if(joined_) 189   652 if(joined_)
HITCBC 190   266 return; 190   257 return;
HITCBC 191   400 joined_ = true; 191   395 joined_ = true;
192   192  
HITCBC 193   400 if(outstanding_work_.load( 193   395 if(outstanding_work_.load(
HITCBC 194   400 std::memory_order_acquire) == 0) 194   395 std::memory_order_acquire) == 0)
195   { 195   {
HITCBC 196   176 stop_ = true; 196   157 stop_ = true;
HITCBC 197   176 work_cv_.notify_all(); 197   157 work_cv_.notify_all();
198   } 198   }
199   else 199   else
200   { 200   {
HITCBC 201   224 done_cv_.wait(lock, [this]{ 201   238 done_cv_.wait(lock, [this]{
HITCBC 202   400 return stop_; 202   424 return stop_;
203   }); 203   });
204   } 204   }
HITCBC 205   666 } 205   652 }
206   206  
HITCBC 207   854 for(auto& t : threads_) 207   840 for(auto& t : threads_)
HITCBC 208   454 if(t.joinable()) 208   445 if(t.joinable())
HITCBC 209   454 t.join(); 209   445 t.join();
210   } 210   }
211   211  
212   void 212   void
HITCBC 213   402 stop() noexcept 213   397 stop() noexcept
214   { 214   {
215   { 215   {
HITCBC 216   402 std::lock_guard<std::mutex> lock(mutex_); 216   397 std::lock_guard<std::mutex> lock(mutex_);
HITCBC 217   402 stop_ = true; 217   397 stop_ = true;
HITCBC 218   402 } 218   397 }
HITCBC 219   402 work_cv_.notify_all(); 219   397 work_cv_.notify_all();
HITCBC 220   402 done_cv_.notify_all(); 220   397 done_cv_.notify_all();
HITCBC 221   402 } 221   397 }
222   222  
223   private: 223   private:
224   void 224   void
HITCBC 225   20549 ensure_started() 225   19442 ensure_started()
226   { 226   {
HITCBC 227   20549 std::call_once(start_flag_, [this]{ 227   19442 std::call_once(start_flag_, [this]{
HITCBC 228   350 threads_.reserve(num_threads_); 228   341 threads_.reserve(num_threads_);
HITCBC 229   804 for(std::size_t i = 0; i < num_threads_; ++i) 229   786 for(std::size_t i = 0; i < num_threads_; ++i)
HITCBC 230   908 threads_.emplace_back([this, i]{ run(i); }); 230   890 threads_.emplace_back([this, i]{ run(i); });
HITCBC 231   350 }); 231   341 });
HITCBC 232   20549 } 232   19442 }
233   233  
234   void 234   void
HITCBC 235   454 run(std::size_t index) 235   445 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   454 std::snprintf(name, sizeof(name), "%s%zu", thread_name_prefix_, index); 239   445 std::snprintf(name, sizeof(name), "%s%zu", thread_name_prefix_, index);
HITCBC 240   454 set_current_thread_name(name); 240   445 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   454 scoped_pool(impl const* p) noexcept { current_.set(p); } 246   445 scoped_pool(impl const* p) noexcept { current_.set(p); }
HITCBC 247   454 ~scoped_pool() noexcept { current_.set(nullptr); } 247   445 ~scoped_pool() noexcept { current_.set(nullptr); }
HITCBC 248   454 } guard(this); 248   445 } guard(this);
249   249  
250   for(;;) 250   for(;;)
251   { 251   {
HITCBC 252   20823 continuation* c = nullptr; 252   19700 continuation* c = nullptr;
253   { 253   {
HITCBC 254   20823 std::unique_lock<std::mutex> lock(mutex_); 254   19700 std::unique_lock<std::mutex> lock(mutex_);
HITCBC 255   20823 work_cv_.wait(lock, [this]{ 255   19700 work_cv_.wait(lock, [this]{
HITCBC 256   55436 return !empty() || 256   48266 return !empty() ||
HITCBC 257   55436 stop_; 257   48266 stop_;
258   }); 258   });
HITCBC 259   20823 if(stop_) 259   19700 if(stop_)
HITCBC 260   908 return; 260   890 return;
HITCBC 261   20369 c = pop(); 261   19255 c = pop();
HITCBC 262   20823 } 262   19700 }
HITCBC 263   20369 if(c) 263   19255 if(c)
HITCBC 264   20369 safe_resume(c->h); 264   19255 safe_resume(c->h);
HITCBC 265   20369 } 265   19255 }
HITCBC 266   454 } 266   445 }
267   }; 267   };
268   268  
269   //------------------------------------------------------------------------------ 269   //------------------------------------------------------------------------------
270   270  
HITCBC 271   400 thread_pool:: 271   395 thread_pool::
272   ~thread_pool() 272   ~thread_pool()
273   { 273   {
HITCBC 274   400 impl_->stop(); 274   395 impl_->stop();
HITCBC 275   400 impl_->join(); 275   395 impl_->join();
HITCBC 276   400 impl_->drain_abandoned(); 276   395 impl_->drain_abandoned();
HITCBC 277   400 shutdown(); 277   395 shutdown();
HITCBC 278   400 destroy(); 278   395 destroy();
HITCBC 279   400 delete impl_; 279   395 delete impl_;
HITCBC 280   400 } 280   395 }
281   281  
HITCBC 282   400 thread_pool:: 282   395 thread_pool::
HITCBC 283   400 thread_pool(std::size_t num_threads, std::string_view thread_name_prefix) 283   395 thread_pool(std::size_t num_threads, std::string_view thread_name_prefix)
HITCBC 284   400 : impl_(new impl(num_threads, thread_name_prefix)) 284   395 : impl_(new impl(num_threads, thread_name_prefix))
285   { 285   {
HITCBC 286   400 this->set_frame_allocator(std::allocator<void>{}); 286   395 this->set_frame_allocator(std::allocator<void>{});
HITCBC 287   400 } 287   395 }
288   288  
289   void 289   void
HITCBC 290   266 thread_pool:: 290   257 thread_pool::
291   join() noexcept 291   join() noexcept
292   { 292   {
HITCBC 293   266 impl_->join(); 293   257 impl_->join();
HITCBC 294   266 } 294   257 }
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   11812 thread_pool:: 306   11809 thread_pool::
307   get_executor() const noexcept 307   get_executor() const noexcept
308   { 308   {
HITCBC 309   11812 return executor_type( 309   11809 return executor_type(
HITCBC 310   11812 const_cast<thread_pool&>(*this)); 310   11809 const_cast<thread_pool&>(*this));
311   } 311   }
312   312  
313   void 313   void
HITCBC 314   687 thread_pool::executor_type:: 314   678 thread_pool::executor_type::
315   on_work_started() const noexcept 315   on_work_started() const noexcept
316   { 316   {
HITCBC 317   687 pool_->impl_->on_work_started(); 317   678 pool_->impl_->on_work_started();
HITCBC 318   687 } 318   678 }
319   319  
320   void 320   void
HITCBC 321   687 thread_pool::executor_type:: 321   678 thread_pool::executor_type::
322   on_work_finished() const noexcept 322   on_work_finished() const noexcept
323   { 323   {
HITCBC 324   687 pool_->impl_->on_work_finished(); 324   678 pool_->impl_->on_work_finished();
HITCBC 325   687 } 325   678 }
326   326  
327   void 327   void
HITCBC 328   19868 thread_pool::executor_type:: 328   18770 thread_pool::executor_type::
329   post(continuation& c) const 329   post(continuation& c) const
330   { 330   {
HITCBC 331   19868 pool_->impl_->post(c); 331   18770 pool_->impl_->post(c);
HITCBC 332   19868 } 332   18770 }
333   333  
334   std::coroutine_handle<> 334   std::coroutine_handle<>
HITCBC 335   688 thread_pool::executor_type:: 335   679 thread_pool::executor_type::
336   dispatch(continuation& c) const 336   dispatch(continuation& c) const
337   { 337   {
HITCBC 338   688 if(pool_->impl_->running_in_this_thread()) 338   679 if(pool_->impl_->running_in_this_thread())
HITCBC 339   7 return c.h; 339   7 return c.h;
HITCBC 340   681 pool_->impl_->post(c); 340   672 pool_->impl_->post(c);
HITCBC 341   681 return std::noop_coroutine(); 341   672 return std::noop_coroutine();
342   } 342   }
343   343  
344   } // capy 344   } // capy
345   } // boost 345   } // boost