TLA Line data Source code
1 : //
2 : // Copyright (c) 2026 Steve Gerbino
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/cppalliance/corosio
9 : //
10 :
11 : #ifndef BOOST_COROSIO_NATIVE_DETAIL_EPOLL_EPOLL_SCHEDULER_HPP
12 : #define BOOST_COROSIO_NATIVE_DETAIL_EPOLL_EPOLL_SCHEDULER_HPP
13 :
14 : #include <boost/corosio/detail/platform.hpp>
15 :
16 : #if BOOST_COROSIO_HAS_EPOLL
17 :
18 : #include <boost/corosio/detail/config.hpp>
19 : #include <boost/capy/ex/execution_context.hpp>
20 :
21 : #include <boost/corosio/native/detail/reactor/reactor_scheduler.hpp>
22 : #include <boost/corosio/native/detail/reactor/reactor_signal_pipe.hpp>
23 :
24 : #include <boost/corosio/native/detail/epoll/epoll_traits.hpp>
25 : #include <boost/corosio/detail/timer_service.hpp>
26 : #include <boost/corosio/native/detail/make_err.hpp>
27 : #include <boost/corosio/native/detail/posix/posix_resolver_service.hpp>
28 : #include <boost/corosio/native/detail/posix/posix_signal_service.hpp>
29 : #include <boost/corosio/native/detail/posix/posix_stream_file_service.hpp>
30 : #include <boost/corosio/native/detail/posix/posix_random_access_file_service.hpp>
31 :
32 : #include <boost/corosio/detail/except.hpp>
33 :
34 : #include <atomic>
35 : #include <chrono>
36 : #include <cstdint>
37 : #include <mutex>
38 : #include <vector>
39 :
40 : #include <errno.h>
41 : #include <sys/epoll.h>
42 : #include <sys/eventfd.h>
43 : #include <sys/timerfd.h>
44 : #include <unistd.h>
45 :
46 : namespace boost::corosio::detail {
47 :
48 : /** Linux scheduler using epoll for I/O multiplexing.
49 :
50 : This scheduler implements the scheduler interface using Linux epoll
51 : for efficient I/O event notification. It uses a single reactor model
52 : where one thread runs epoll_wait while other threads
53 : wait on a condition variable for handler work. This design provides:
54 :
55 : - Handler parallelism: N posted handlers can execute on N threads
56 : - No thundering herd: condition_variable wakes exactly one thread
57 : - IOCP parity: Behavior matches Windows I/O completion port semantics
58 :
59 : When threads call run(), they first try to execute queued handlers.
60 : If the queue is empty and no reactor is running, one thread becomes
61 : the reactor and runs epoll_wait. Other threads wait on a condition
62 : variable until handlers are available.
63 :
64 : @par Thread Safety
65 : All public member functions are thread-safe.
66 : */
67 : class BOOST_COROSIO_DECL epoll_scheduler final : public reactor_scheduler
68 : {
69 : public:
70 : /** Construct the scheduler.
71 :
72 : Creates an epoll instance, eventfd for reactor interruption,
73 : and timerfd for kernel-managed timer expiry.
74 :
75 : @param ctx Reference to the owning execution_context.
76 : @param concurrency_hint Hint for expected thread count (unused).
77 : */
78 : epoll_scheduler(capy::execution_context& ctx, int concurrency_hint = -1);
79 :
80 : /// Destroy the scheduler.
81 : ~epoll_scheduler() override;
82 :
83 : epoll_scheduler(epoll_scheduler const&) = delete;
84 : epoll_scheduler& operator=(epoll_scheduler const&) = delete;
85 :
86 : /// Shut down the scheduler, draining pending operations.
87 : void shutdown() override;
88 :
89 : /// Apply runtime configuration, resizing the event buffer.
90 : void configure_reactor(
91 : unsigned max_events,
92 : unsigned budget_init,
93 : unsigned budget_max,
94 : unsigned unassisted) override;
95 :
96 : /** Return the epoll file descriptor.
97 :
98 : Used by socket services to register file descriptors
99 : for I/O event notification.
100 :
101 : @return The epoll file descriptor.
102 : */
103 : int epoll_fd() const noexcept
104 : {
105 : return epoll_fd_;
106 : }
107 :
108 : /** Register a descriptor for persistent monitoring.
109 :
110 : The fd is registered once and stays registered until explicitly
111 : deregistered. Events are dispatched via reactor_descriptor_state which
112 : tracks pending read/write/connect operations.
113 :
114 : @param fd The file descriptor to register.
115 : @param desc Pointer to descriptor data (stored in epoll_event.data.ptr).
116 : */
117 : void register_descriptor(int fd, reactor_descriptor_state* desc) const;
118 :
119 : /** Deregister a persistently registered descriptor.
120 :
121 : @param fd The file descriptor to deregister.
122 : */
123 : void deregister_descriptor(int fd) const;
124 :
125 : /// Watch the read end of the POSIX signal self-pipe (see scheduler.hpp).
126 HIT 41 : void register_signal_reader(int read_fd) override
127 : {
128 41 : register_descriptor(read_fd, signal_pipe_reader_.arm());
129 41 : }
130 :
131 : private:
132 : void
133 : run_task(lock_type& lock, context_type* ctx,
134 : long timeout_us) override;
135 : void interrupt_reactor() const override;
136 : void update_timerfd() const;
137 :
138 : int epoll_fd_;
139 : int event_fd_;
140 : int timer_fd_;
141 :
142 : // Watches the global signal self-pipe's read end (armed lazily by
143 : // register_signal_reader on the first signal registration).
144 : reactor_signal_pipe_reader signal_pipe_reader_;
145 :
146 : // Edge-triggered eventfd state
147 : mutable std::atomic<bool> eventfd_armed_{false};
148 :
149 : // Set when the earliest timer changes; flushed before epoll_wait
150 : mutable std::atomic<bool> timerfd_stale_{false};
151 :
152 : // Event buffer sized from max_events_per_poll_ (set at construction,
153 : // resized by configure_reactor via io_context_options).
154 : std::vector<epoll_event> event_buffer_;
155 : };
156 :
157 641 : inline epoll_scheduler::epoll_scheduler(capy::execution_context& ctx, int)
158 641 : : epoll_fd_(-1)
159 641 : , event_fd_(-1)
160 641 : , timer_fd_(-1)
161 1282 : , event_buffer_(max_events_per_poll_)
162 : {
163 641 : epoll_fd_ = ::epoll_create1(EPOLL_CLOEXEC);
164 641 : if (epoll_fd_ < 0)
165 MIS 0 : detail::throw_system_error(make_err(errno), "epoll_create1");
166 :
167 HIT 641 : event_fd_ = ::eventfd(0, EFD_NONBLOCK | EFD_CLOEXEC);
168 641 : if (event_fd_ < 0)
169 : {
170 MIS 0 : int errn = errno;
171 0 : ::close(epoll_fd_);
172 0 : detail::throw_system_error(make_err(errn), "eventfd");
173 : }
174 :
175 HIT 641 : timer_fd_ = ::timerfd_create(CLOCK_MONOTONIC, TFD_NONBLOCK | TFD_CLOEXEC);
176 641 : if (timer_fd_ < 0)
177 : {
178 MIS 0 : int errn = errno;
179 0 : ::close(event_fd_);
180 0 : ::close(epoll_fd_);
181 0 : detail::throw_system_error(make_err(errn), "timerfd_create");
182 : }
183 :
184 HIT 641 : epoll_event ev{};
185 641 : ev.events = EPOLLIN | EPOLLET;
186 641 : ev.data.ptr = nullptr;
187 641 : if (::epoll_ctl(epoll_fd_, EPOLL_CTL_ADD, event_fd_, &ev) < 0)
188 : {
189 MIS 0 : int errn = errno;
190 0 : ::close(timer_fd_);
191 0 : ::close(event_fd_);
192 0 : ::close(epoll_fd_);
193 0 : detail::throw_system_error(make_err(errn), "epoll_ctl");
194 : }
195 :
196 HIT 641 : epoll_event timer_ev{};
197 641 : timer_ev.events = EPOLLIN | EPOLLERR;
198 641 : timer_ev.data.ptr = &timer_fd_;
199 641 : if (::epoll_ctl(epoll_fd_, EPOLL_CTL_ADD, timer_fd_, &timer_ev) < 0)
200 : {
201 MIS 0 : int errn = errno;
202 0 : ::close(timer_fd_);
203 0 : ::close(event_fd_);
204 0 : ::close(epoll_fd_);
205 0 : detail::throw_system_error(make_err(errn), "epoll_ctl (timerfd)");
206 : }
207 :
208 HIT 641 : timer_svc_ = &get_timer_service(ctx, *this);
209 641 : timer_svc_->set_on_earliest_changed(
210 5213 : timer_service::callback(this, [](void* p) {
211 4572 : auto* self = static_cast<epoll_scheduler*>(p);
212 4572 : self->timerfd_stale_.store(true, std::memory_order_release);
213 4572 : self->interrupt_reactor();
214 4572 : }));
215 :
216 641 : get_resolver_service(ctx, *this);
217 641 : get_signal_service(ctx, *this);
218 641 : get_stream_file_service(ctx, *this);
219 641 : get_random_access_file_service(ctx, *this);
220 :
221 641 : completed_ops_.push(&task_op_);
222 641 : }
223 :
224 1282 : inline epoll_scheduler::~epoll_scheduler()
225 : {
226 641 : if (timer_fd_ >= 0)
227 641 : ::close(timer_fd_);
228 641 : if (event_fd_ >= 0)
229 641 : ::close(event_fd_);
230 641 : if (epoll_fd_ >= 0)
231 641 : ::close(epoll_fd_);
232 1282 : }
233 :
234 : inline void
235 641 : epoll_scheduler::shutdown()
236 : {
237 641 : shutdown_drain();
238 :
239 641 : if (event_fd_ >= 0)
240 641 : interrupt_reactor();
241 641 : }
242 :
243 : inline void
244 16 : epoll_scheduler::configure_reactor(
245 : unsigned max_events,
246 : unsigned budget_init,
247 : unsigned budget_max,
248 : unsigned unassisted)
249 : {
250 16 : reactor_scheduler::configure_reactor(
251 : max_events, budget_init, budget_max, unassisted);
252 15 : event_buffer_.resize(max_events_per_poll_);
253 15 : }
254 :
255 : inline void
256 7804 : epoll_scheduler::register_descriptor(int fd, reactor_descriptor_state* desc) const
257 : {
258 7804 : epoll_event ev{};
259 7804 : ev.events = EPOLLIN | EPOLLOUT | EPOLLET | EPOLLERR | EPOLLHUP;
260 7804 : ev.data.ptr = desc;
261 :
262 7804 : if (::epoll_ctl(epoll_fd_, EPOLL_CTL_ADD, fd, &ev) < 0)
263 MIS 0 : detail::throw_system_error(make_err(errno), "epoll_ctl (register)");
264 :
265 HIT 7804 : desc->registered_events = ev.events;
266 7804 : desc->fd = fd;
267 7804 : desc->scheduler_ = this;
268 7804 : desc->mutex.set_enabled(reactor_io_locking_);
269 7804 : desc->ready_events_.store(0, std::memory_order_relaxed);
270 :
271 7804 : conditionally_enabled_mutex::scoped_lock lock(desc->mutex);
272 7804 : desc->impl_ref_.reset();
273 7804 : desc->read_ready = false;
274 7804 : desc->write_ready = false;
275 7804 : }
276 :
277 : inline void
278 7763 : epoll_scheduler::deregister_descriptor(int fd) const
279 : {
280 7763 : ::epoll_ctl(epoll_fd_, EPOLL_CTL_DEL, fd, nullptr);
281 7763 : }
282 :
283 : inline void
284 5735 : epoll_scheduler::interrupt_reactor() const
285 : {
286 5735 : bool expected = false;
287 5735 : if (eventfd_armed_.compare_exchange_strong(
288 : expected, true, std::memory_order_release,
289 : std::memory_order_relaxed))
290 : {
291 4696 : std::uint64_t val = 1;
292 4696 : [[maybe_unused]] auto r = ::write(event_fd_, &val, sizeof(val));
293 : }
294 5735 : }
295 :
296 : inline void
297 7794 : epoll_scheduler::update_timerfd() const
298 : {
299 7794 : auto nearest = timer_svc_->nearest_expiry();
300 :
301 7794 : itimerspec ts{};
302 7794 : int flags = 0;
303 :
304 7794 : if (nearest == timer_service::time_point::max())
305 : {
306 : // No timers — disarm by setting to 0 (relative)
307 : }
308 : else
309 : {
310 7660 : auto now = std::chrono::steady_clock::now();
311 7660 : if (nearest <= now)
312 : {
313 : // Use 1ns instead of 0 — zero disarms the timerfd
314 540 : ts.it_value.tv_nsec = 1;
315 : }
316 : else
317 : {
318 7120 : auto nsec = std::chrono::duration_cast<std::chrono::nanoseconds>(
319 7120 : nearest - now)
320 7120 : .count();
321 7120 : ts.it_value.tv_sec = nsec / 1000000000;
322 7120 : ts.it_value.tv_nsec = nsec % 1000000000;
323 7120 : if (ts.it_value.tv_sec == 0 && ts.it_value.tv_nsec == 0)
324 MIS 0 : ts.it_value.tv_nsec = 1;
325 : }
326 : }
327 :
328 HIT 7794 : if (::timerfd_settime(timer_fd_, flags, &ts, nullptr) < 0)
329 MIS 0 : detail::throw_system_error(make_err(errno), "timerfd_settime");
330 HIT 7794 : }
331 :
332 : inline void
333 36184 : epoll_scheduler::run_task(
334 : lock_type& lock, context_type* ctx, long timeout_us)
335 : {
336 : int timeout_ms;
337 36184 : if (task_interrupted_)
338 25139 : timeout_ms = 0;
339 11045 : else if (timeout_us < 0)
340 11041 : timeout_ms = -1;
341 : else
342 4 : timeout_ms = static_cast<int>((timeout_us + 999) / 1000);
343 :
344 36184 : if (lock.owns_lock())
345 11047 : lock.unlock();
346 :
347 36184 : task_cleanup on_exit{this, &lock, ctx};
348 :
349 : // Flush deferred timerfd programming before blocking
350 36184 : if (timerfd_stale_.exchange(false, std::memory_order_acquire))
351 3903 : update_timerfd();
352 :
353 36184 : int nfds = ::epoll_wait(
354 : epoll_fd_, event_buffer_.data(),
355 36184 : static_cast<int>(event_buffer_.size()), timeout_ms);
356 :
357 36184 : if (nfds < 0 && errno != EINTR)
358 MIS 0 : detail::throw_system_error(make_err(errno), "epoll_wait");
359 :
360 HIT 36184 : bool check_timers = false;
361 36184 : ready_queue local_ops;
362 :
363 83339 : for (int i = 0; i < nfds; ++i)
364 : {
365 47155 : if (event_buffer_[i].data.ptr == nullptr)
366 : {
367 : std::uint64_t val;
368 : // NOLINTNEXTLINE(clang-analyzer-unix.BlockInCriticalSection)
369 4055 : [[maybe_unused]] auto r = ::read(event_fd_, &val, sizeof(val));
370 4055 : eventfd_armed_.store(false, std::memory_order_relaxed);
371 4055 : continue;
372 4055 : }
373 :
374 43100 : if (event_buffer_[i].data.ptr == &timer_fd_)
375 : {
376 : std::uint64_t expirations;
377 : // NOLINTNEXTLINE(clang-analyzer-unix.BlockInCriticalSection)
378 : [[maybe_unused]] auto r =
379 3891 : ::read(timer_fd_, &expirations, sizeof(expirations));
380 3891 : check_timers = true;
381 3891 : continue;
382 3891 : }
383 :
384 : auto* desc =
385 39209 : static_cast<reactor_descriptor_state*>(event_buffer_[i].data.ptr);
386 39209 : desc->add_ready_events(event_buffer_[i].events);
387 :
388 39209 : bool expected = false;
389 39209 : if (desc->is_enqueued_.compare_exchange_strong(
390 : expected, true, std::memory_order_release,
391 : std::memory_order_relaxed))
392 : {
393 39209 : local_ops.push(desc);
394 : }
395 : }
396 :
397 36184 : if (check_timers)
398 : {
399 3891 : timer_svc_->process_expired();
400 3891 : update_timerfd();
401 : }
402 :
403 36184 : lock.lock();
404 :
405 36184 : completed_ops_.splice(local_ops);
406 36184 : }
407 :
408 : } // namespace boost::corosio::detail
409 :
410 : #endif // BOOST_COROSIO_HAS_EPOLL
411 :
412 : #endif // BOOST_COROSIO_NATIVE_DETAIL_EPOLL_EPOLL_SCHEDULER_HPP
|
