TLA Line data Source code
1 : //
2 : // Copyright (c) 2025 Vinnie Falco (vinnie.falco@gmail.com)
3 : // Copyright (c) 2026 Michael Vandeberg
4 : //
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)
7 : //
8 : // Official repository: https://github.com/boostorg/capy
9 : //
10 :
11 : #include <boost/capy/ex/thread_pool.hpp>
12 : #include <boost/capy/continuation.hpp>
13 : #include <boost/capy/detail/thread_local_ptr.hpp>
14 : #include <boost/capy/ex/frame_allocator.hpp>
15 : #include <boost/capy/test/thread_name.hpp>
16 : #include <algorithm>
17 : #include <atomic>
18 : #include <condition_variable>
19 : #include <cstdio>
20 : #include <mutex>
21 : #include <thread>
22 : #include <vector>
23 :
24 : /*
25 : Thread pool implementation using a shared work queue.
26 :
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
29 : allocation: the continuation is owned by the caller and linked
30 : directly. Worker threads wait on a condition_variable until work
31 : is available or stop is requested.
32 :
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
35 : thread is named with a configurable prefix plus index for debugger
36 : visibility.
37 :
38 : Work tracking: on_work_started/on_work_finished maintain the atomic
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
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
43 : this counter reaches zero, then signals workers to stop and joins
44 : threads.
45 :
46 : Two shutdown paths:
47 : - join(): waits for outstanding work to drain, then stops workers.
48 : - stop(): immediately signals workers to exit; queued work is abandoned.
49 : - Destructor: stop() then join() (abandon + wait for threads).
50 : */
51 :
52 : namespace boost {
53 : namespace capy {
54 :
55 : //------------------------------------------------------------------------------
56 :
57 : class thread_pool::impl
58 : {
59 : // Identifies the pool owning the current worker thread, or
60 : // nullptr if the calling thread is not a pool worker. Checked
61 : // by dispatch() to decide between symmetric transfer (inline
62 : // resume) and post.
63 : static inline detail::thread_local_ptr<impl const> current_;
64 :
65 : // Intrusive queue of continuations: the next link is stored in
66 : // continuation::reserved (typed continuation* round-tripped through
67 : // void*). No per-post allocation: the continuation is owned by the caller.
68 : continuation* head_ = nullptr;
69 : continuation* tail_ = nullptr;
70 :
71 HIT 16806 : void push(continuation* c) noexcept
72 : {
73 16806 : c->reserved = nullptr;
74 16806 : if(tail_)
75 5918 : tail_->reserved = c;
76 : else
77 10888 : head_ = c;
78 16806 : tail_ = c;
79 16806 : }
80 :
81 17071 : continuation* pop() noexcept
82 : {
83 17071 : if(!head_)
84 265 : return nullptr;
85 16806 : continuation* c = head_;
86 16806 : head_ = static_cast<continuation*>(head_->reserved);
87 16806 : if(!head_)
88 10888 : tail_ = nullptr;
89 16806 : return c;
90 : }
91 :
92 28926 : bool empty() const noexcept
93 : {
94 28926 : return head_ == nullptr;
95 : }
96 :
97 : std::mutex mutex_;
98 : std::condition_variable work_cv_;
99 : std::condition_variable done_cv_;
100 : std::vector<std::thread> threads_;
101 : std::atomic<std::size_t> outstanding_work_{0};
102 : bool stop_{false};
103 : bool joined_{false};
104 : std::size_t num_threads_;
105 : char thread_name_prefix_[13]{}; // 12 chars max + null terminator
106 : std::once_flag start_flag_;
107 :
108 : public:
109 265 : ~impl() = default;
110 :
111 : bool
112 448 : running_in_this_thread() const noexcept
113 : {
114 448 : return current_.get() == this;
115 : }
116 :
117 : // Destroy abandoned coroutine frames. Must be called
118 : // before execution_context::shutdown()/destroy() so
119 : // that suspended-frame destructors touching services
120 : // (e.g. cancelling registrations) run while those
121 : // services are still valid.
122 : void
123 265 : drain_abandoned() noexcept
124 : {
125 496 : while(auto* c = pop())
126 : {
127 231 : auto h = c->h;
128 231 : if(h && h != std::noop_coroutine())
129 180 : h.destroy();
130 231 : }
131 265 : }
132 :
133 265 : impl(std::size_t num_threads, std::string_view thread_name_prefix)
134 265 : : num_threads_(num_threads)
135 : {
136 265 : if(num_threads_ == 0)
137 4 : num_threads_ = std::max(
138 2 : std::thread::hardware_concurrency(), 1u);
139 :
140 : // Truncate prefix to 12 chars, leaving room for up to 3-digit index.
141 265 : auto n = thread_name_prefix.copy(thread_name_prefix_, 12);
142 265 : thread_name_prefix_[n] = '\0';
143 265 : }
144 :
145 : void
146 16806 : post(continuation& c)
147 : {
148 16806 : ensure_started();
149 : {
150 16806 : std::lock_guard<std::mutex> lock(mutex_);
151 16806 : push(&c);
152 16806 : }
153 16806 : work_cv_.notify_one();
154 16806 : }
155 :
156 : void
157 448 : on_work_started() noexcept
158 : {
159 448 : outstanding_work_.fetch_add(1, std::memory_order_acq_rel);
160 448 : }
161 :
162 : void
163 448 : on_work_finished() noexcept
164 : {
165 448 : if(outstanding_work_.fetch_sub(
166 448 : 1, std::memory_order_acq_rel) == 1)
167 : {
168 : // fetch_sub's result can be stale: a concurrent
169 : // on_work_started() may raise the count before we take the
170 : // lock, so re-read it here rather than trust the decrement.
171 196 : std::lock_guard<std::mutex> lock(mutex_);
172 196 : if(outstanding_work_.load(
173 196 : std::memory_order_acquire) == 0 && joined_ && !stop_)
174 : {
175 90 : stop_ = true;
176 90 : done_cv_.notify_all();
177 90 : work_cv_.notify_all();
178 : }
179 196 : }
180 448 : }
181 :
182 : void
183 400 : join() noexcept
184 : {
185 : {
186 400 : std::unique_lock<std::mutex> lock(mutex_);
187 400 : if(joined_)
188 135 : return;
189 265 : joined_ = true;
190 :
191 265 : if(outstanding_work_.load(
192 265 : std::memory_order_acquire) == 0)
193 : {
194 121 : stop_ = true;
195 121 : work_cv_.notify_all();
196 : }
197 : else
198 : {
199 144 : done_cv_.wait(lock, [this]{
200 235 : return stop_;
201 : });
202 : }
203 400 : }
204 :
205 564 : for(auto& t : threads_)
206 299 : if(t.joinable())
207 299 : t.join();
208 : }
209 :
210 : void
211 267 : stop() noexcept
212 : {
213 : {
214 267 : std::lock_guard<std::mutex> lock(mutex_);
215 267 : stop_ = true;
216 267 : }
217 267 : work_cv_.notify_all();
218 267 : done_cv_.notify_all();
219 267 : }
220 :
221 : private:
222 : void
223 16806 : ensure_started()
224 : {
225 16806 : std::call_once(start_flag_, [this]{
226 218 : threads_.reserve(num_threads_);
227 517 : for(std::size_t i = 0; i < num_threads_; ++i)
228 598 : threads_.emplace_back([this, i]{ run(i); });
229 218 : });
230 16806 : }
231 :
232 : void
233 299 : run(std::size_t index)
234 : {
235 : // Build name; set_current_thread_name truncates to platform limits.
236 : char name[16];
237 299 : std::snprintf(name, sizeof(name), "%s%zu", thread_name_prefix_, index);
238 299 : set_current_thread_name(name);
239 :
240 : // Mark this thread as a worker of this pool so dispatch()
241 : // can symmetric-transfer when called from within pool work.
242 : struct scoped_pool
243 : {
244 299 : scoped_pool(impl const* p) noexcept { current_.set(p); }
245 299 : ~scoped_pool() noexcept { current_.set(nullptr); }
246 299 : } guard(this);
247 :
248 : for(;;)
249 : {
250 16874 : continuation* c = nullptr;
251 : {
252 16874 : std::unique_lock<std::mutex> lock(mutex_);
253 16874 : work_cv_.wait(lock, [this]{
254 41174 : return !empty() ||
255 41174 : stop_;
256 : });
257 16874 : if(stop_)
258 598 : return;
259 16575 : c = pop();
260 16874 : }
261 16575 : if(c)
262 16575 : safe_resume(c->h);
263 16575 : }
264 299 : }
265 : };
266 :
267 : //------------------------------------------------------------------------------
268 :
269 265 : thread_pool::
270 : ~thread_pool()
271 : {
272 265 : impl_->stop();
273 265 : impl_->join();
274 265 : impl_->drain_abandoned();
275 265 : shutdown();
276 265 : destroy();
277 265 : delete impl_;
278 265 : }
279 :
280 265 : thread_pool::
281 265 : thread_pool(std::size_t num_threads, std::string_view thread_name_prefix)
282 265 : : impl_(new impl(num_threads, thread_name_prefix))
283 : {
284 265 : this->set_frame_allocator(std::allocator<void>{});
285 265 : }
286 :
287 : void
288 135 : thread_pool::
289 : join() noexcept
290 : {
291 135 : impl_->join();
292 135 : }
293 :
294 : void
295 2 : thread_pool::
296 : stop() noexcept
297 : {
298 2 : impl_->stop();
299 2 : }
300 :
301 : //------------------------------------------------------------------------------
302 :
303 : thread_pool::executor_type
304 11671 : thread_pool::
305 : get_executor() const noexcept
306 : {
307 11671 : return executor_type(
308 11671 : const_cast<thread_pool&>(*this));
309 : }
310 :
311 : void
312 448 : thread_pool::executor_type::
313 : on_work_started() const noexcept
314 : {
315 448 : pool_->impl_->on_work_started();
316 448 : }
317 :
318 : void
319 448 : thread_pool::executor_type::
320 : on_work_finished() const noexcept
321 : {
322 448 : pool_->impl_->on_work_finished();
323 448 : }
324 :
325 : void
326 16363 : thread_pool::executor_type::
327 : post(continuation& c) const
328 : {
329 16363 : pool_->impl_->post(c);
330 16363 : }
331 :
332 : std::coroutine_handle<>
333 448 : thread_pool::executor_type::
334 : dispatch(continuation& c) const
335 : {
336 448 : if(pool_->impl_->running_in_this_thread())
337 5 : return c.h;
338 443 : pool_->impl_->post(c);
339 443 : return std::noop_coroutine();
340 : }
341 :
342 : } // capy
343 : } // boost
|
