include/boost/corosio/native/detail/epoll/epoll_scheduler.hpp

85.3% Lines (128/150) 100.0% List of functions (10/10)

[ ] prev/next file n p prev/next uncovered line

epoll_scheduler.hpp

f(x) Functions (10)

Function Calls Lines Blocks

Line	Hits	Source Code
1		//
2		// Copyright (c) 2026 Steve Gerbino
3		//
4		// Distributed under the Boost Software License, Version 1.0. (See accompanying
5		// file LICENSE_1_0.txt or copy at http://www.boost.org/LICENSE_1_0.txt)
6		//
7		// Official repository: https://github.com/cppalliance/corosio
8		//
9
10		#ifndef BOOST_COROSIO_NATIVE_DETAIL_EPOLL_EPOLL_SCHEDULER_HPP
11		#define BOOST_COROSIO_NATIVE_DETAIL_EPOLL_EPOLL_SCHEDULER_HPP
12
13		#include <boost/corosio/detail/platform.hpp>
14
15		#if BOOST_COROSIO_HAS_EPOLL
16
17		#include <boost/corosio/detail/config.hpp>
18		#include <boost/capy/ex/execution_context.hpp>
19
20		#include <boost/corosio/native/detail/reactor/reactor_scheduler.hpp>
21
22		#include <boost/corosio/native/detail/epoll/epoll_traits.hpp>
23		#include <boost/corosio/detail/timer_service.hpp>
24		#include <boost/corosio/native/detail/make_err.hpp>
25		#include <boost/corosio/native/detail/posix/posix_resolver_service.hpp>
26		#include <boost/corosio/native/detail/posix/posix_signal_service.hpp>
27		#include <boost/corosio/native/detail/posix/posix_stream_file_service.hpp>
28		#include <boost/corosio/native/detail/posix/posix_random_access_file_service.hpp>
29
30		#include <boost/corosio/detail/except.hpp>
31
32		#include <atomic>
33		#include <chrono>
34		#include <cstdint>
35		#include <mutex>
36		#include <vector>
37
38		#include <errno.h>
39		#include <sys/epoll.h>
40		#include <sys/eventfd.h>
41		#include <sys/timerfd.h>
42		#include <unistd.h>
43
44		namespace boost::corosio::detail {
45
46		/** Linux scheduler using epoll for I/O multiplexing.
47
48		This scheduler implements the scheduler interface using Linux epoll
49		for efficient I/O event notification. It uses a single reactor model
50		where one thread runs epoll_wait while other threads
51		wait on a condition variable for handler work. This design provides:
52
53		- Handler parallelism: N posted handlers can execute on N threads
54		- No thundering herd: condition_variable wakes exactly one thread
55		- IOCP parity: Behavior matches Windows I/O completion port semantics
56
57		When threads call run(), they first try to execute queued handlers.
58		If the queue is empty and no reactor is running, one thread becomes
59		the reactor and runs epoll_wait. Other threads wait on a condition
60		variable until handlers are available.
61
62		@par Thread Safety
63		All public member functions are thread-safe.
64		*/
65		class BOOST_COROSIO_DECL epoll_scheduler final : public reactor_scheduler
66		{
67		public:
68		/** Construct the scheduler.
69
70		Creates an epoll instance, eventfd for reactor interruption,
71		and timerfd for kernel-managed timer expiry.
72
73		@param ctx Reference to the owning execution_context.
74		@param concurrency_hint Hint for expected thread count (unused).
75		*/
76		epoll_scheduler(capy::execution_context& ctx, int concurrency_hint = -1);
77
78		/// Destroy the scheduler.
79		~epoll_scheduler() override;
80
81		epoll_scheduler(epoll_scheduler const&) = delete;
82		epoll_scheduler& operator=(epoll_scheduler const&) = delete;
83
84		/// Shut down the scheduler, draining pending operations.
85		void shutdown() override;
86
87		/// Apply runtime configuration, resizing the event buffer.
88		void configure_reactor(
89		unsigned max_events,
90		unsigned budget_init,
91		unsigned budget_max,
92		unsigned unassisted) override;
93
94		/** Return the epoll file descriptor.
95
96		Used by socket services to register file descriptors
97		for I/O event notification.
98
99		@return The epoll file descriptor.
100		*/
101		int epoll_fd() const noexcept
102		{
103		return epoll_fd_;
104		}
105
106		/** Register a descriptor for persistent monitoring.
107
108		The fd is registered once and stays registered until explicitly
109		deregistered. Events are dispatched via reactor_descriptor_state which
110		tracks pending read/write/connect operations.
111
112		@param fd The file descriptor to register.
113		@param desc Pointer to descriptor data (stored in epoll_event.data.ptr).
114		*/
115		void register_descriptor(int fd, reactor_descriptor_state* desc) const;
116
117		/** Deregister a persistently registered descriptor.
118
119		@param fd The file descriptor to deregister.
120		*/
121		void deregister_descriptor(int fd) const;
122
123		private:
124		void
125		run_task(lock_type& lock, context_type* ctx,
126		long timeout_us) override;
127		void interrupt_reactor() const override;
128		void update_timerfd() const;
129
130		int epoll_fd_;
131		int event_fd_;
132		int timer_fd_;
133
134		// Edge-triggered eventfd state
135		mutable std::atomic<bool> eventfd_armed_{false};
136
137		// Set when the earliest timer changes; flushed before epoll_wait
138		mutable std::atomic<bool> timerfd_stale_{false};
139
140		// Event buffer sized from max_events_per_poll_ (set at construction,
141		// resized by configure_reactor via io_context_options).
142		std::vector<epoll_event> event_buffer_;
143		};
144
145	591x	inline epoll_scheduler::epoll_scheduler(capy::execution_context& ctx, int)
146	591x	: epoll_fd_(-1)
147	591x	, event_fd_(-1)
148	591x	, timer_fd_(-1)
149	1182x	, event_buffer_(max_events_per_poll_)
150		{
151	591x	epoll_fd_ = ::epoll_create1(EPOLL_CLOEXEC);
152	591x	if (epoll_fd_ < 0)
153	✗	detail::throw_system_error(make_err(errno), "epoll_create1");
154
155	591x	event_fd_ = ::eventfd(0, EFD_NONBLOCK \| EFD_CLOEXEC);
156	591x	if (event_fd_ < 0)
157		{
158	✗	int errn = errno;
159	✗	::close(epoll_fd_);
160	✗	detail::throw_system_error(make_err(errn), "eventfd");
161		}
162
163	591x	timer_fd_ = ::timerfd_create(CLOCK_MONOTONIC, TFD_NONBLOCK \| TFD_CLOEXEC);
164	591x	if (timer_fd_ < 0)
165		{
166	✗	int errn = errno;
167	✗	::close(event_fd_);
168	✗	::close(epoll_fd_);
169	✗	detail::throw_system_error(make_err(errn), "timerfd_create");
170		}
171
172	591x	epoll_event ev{};
173	591x	ev.events = EPOLLIN \| EPOLLET;
174	591x	ev.data.ptr = nullptr;
175	591x	if (::epoll_ctl(epoll_fd_, EPOLL_CTL_ADD, event_fd_, &ev) < 0)
176		{
177	✗	int errn = errno;
178	✗	::close(timer_fd_);
179	✗	::close(event_fd_);
180	✗	::close(epoll_fd_);
181	✗	detail::throw_system_error(make_err(errn), "epoll_ctl");
182		}
183
184	591x	epoll_event timer_ev{};
185	591x	timer_ev.events = EPOLLIN \| EPOLLERR;
186	591x	timer_ev.data.ptr = &timer_fd_;
187	591x	if (::epoll_ctl(epoll_fd_, EPOLL_CTL_ADD, timer_fd_, &timer_ev) < 0)
188		{
189	✗	int errn = errno;
190	✗	::close(timer_fd_);
191	✗	::close(event_fd_);
192	✗	::close(epoll_fd_);
193	✗	detail::throw_system_error(make_err(errn), "epoll_ctl (timerfd)");
194		}
195
196	591x	timer_svc_ = &get_timer_service(ctx, *this);
197	591x	timer_svc_->set_on_earliest_changed(
198	5659x	timer_service::callback(this, [](void* p) {
199	5068x	auto* self = static_cast<epoll_scheduler*>(p);
200	5068x	self->timerfd_stale_.store(true, std::memory_order_release);
201	5068x	self->interrupt_reactor();
202	5068x	}));
203
204	591x	get_resolver_service(ctx, *this);
205	591x	get_signal_service(ctx, *this);
206	591x	get_stream_file_service(ctx, *this);
207	591x	get_random_access_file_service(ctx, *this);
208
209	591x	completed_ops_.push(&task_op_);
210	591x	}
211
212	1182x	inline epoll_scheduler::~epoll_scheduler()
213		{
214	591x	if (timer_fd_ >= 0)
215	591x	::close(timer_fd_);
216	591x	if (event_fd_ >= 0)
217	591x	::close(event_fd_);
218	591x	if (epoll_fd_ >= 0)
219	591x	::close(epoll_fd_);
220	1182x	}
221
222		inline void
223	591x	epoll_scheduler::shutdown()
224		{
225	591x	shutdown_drain();
226
227	591x	if (event_fd_ >= 0)
228	591x	interrupt_reactor();
229	591x	}
230
231		inline void
232	8x	epoll_scheduler::configure_reactor(
233		unsigned max_events,
234		unsigned budget_init,
235		unsigned budget_max,
236		unsigned unassisted)
237		{
238	8x	reactor_scheduler::configure_reactor(
239		max_events, budget_init, budget_max, unassisted);
240	7x	event_buffer_.resize(max_events_per_poll_);
241	7x	}
242
243		inline void
244	9984x	epoll_scheduler::register_descriptor(int fd, reactor_descriptor_state* desc) const
245		{
246	9984x	epoll_event ev{};
247	9984x	ev.events = EPOLLIN \| EPOLLOUT \| EPOLLET \| EPOLLERR \| EPOLLHUP;
248	9984x	ev.data.ptr = desc;
249
250	9984x	if (::epoll_ctl(epoll_fd_, EPOLL_CTL_ADD, fd, &ev) < 0)
251	✗	detail::throw_system_error(make_err(errno), "epoll_ctl (register)");
252
253	9984x	desc->registered_events = ev.events;
254	9984x	desc->fd = fd;
255	9984x	desc->scheduler_ = this;
256	9984x	desc->mutex.set_enabled(!single_threaded_);
257	9984x	desc->ready_events_.store(0, std::memory_order_relaxed);
258
259	9984x	conditionally_enabled_mutex::scoped_lock lock(desc->mutex);
260	9984x	desc->impl_ref_.reset();
261	9984x	desc->read_ready = false;
262	9984x	desc->write_ready = false;
263	9984x	}
264
265		inline void
266	9984x	epoll_scheduler::deregister_descriptor(int fd) const
267		{
268	9984x	::epoll_ctl(epoll_fd_, EPOLL_CTL_DEL, fd, nullptr);
269	9984x	}
270
271		inline void
272	6113x	epoll_scheduler::interrupt_reactor() const
273		{
274	6113x	bool expected = false;
275	6113x	if (eventfd_armed_.compare_exchange_strong(
276		expected, true, std::memory_order_release,
277		std::memory_order_relaxed))
278		{
279	5763x	std::uint64_t val = 1;
280	5763x	[[maybe_unused]] auto r = ::write(event_fd_, &val, sizeof(val));
281		}
282	6113x	}
283
284		inline void
285	10093x	epoll_scheduler::update_timerfd() const
286		{
287	10093x	auto nearest = timer_svc_->nearest_expiry();
288
289	10093x	itimerspec ts{};
290	10093x	int flags = 0;
291
292	10093x	if (nearest == timer_service::time_point::max())
293		{
294		// No timers — disarm by setting to 0 (relative)
295		}
296		else
297		{
298	9995x	auto now = std::chrono::steady_clock::now();
299	9995x	if (nearest <= now)
300		{
301		// Use 1ns instead of 0 — zero disarms the timerfd
302	279x	ts.it_value.tv_nsec = 1;
303		}
304		else
305		{
306	9716x	auto nsec = std::chrono::duration_cast<std::chrono::nanoseconds>(
307	9716x	nearest - now)
308	9716x	.count();
309	9716x	ts.it_value.tv_sec = nsec / 1000000000;
310	9716x	ts.it_value.tv_nsec = nsec % 1000000000;
311	9716x	if (ts.it_value.tv_sec == 0 && ts.it_value.tv_nsec == 0)
312	✗	ts.it_value.tv_nsec = 1;
313		}
314		}
315
316	10093x	if (::timerfd_settime(timer_fd_, flags, &ts, nullptr) < 0)
317	✗	detail::throw_system_error(make_err(errno), "timerfd_settime");
318	10093x	}
319
320		inline void
321	41140x	epoll_scheduler::run_task(
322		lock_type& lock, context_type* ctx, long timeout_us)
323		{
324		int timeout_ms;
325	41140x	if (task_interrupted_)
326	26878x	timeout_ms = 0;
327	14262x	else if (timeout_us < 0)
328	14258x	timeout_ms = -1;
329		else
330	4x	timeout_ms = static_cast<int>((timeout_us + 999) / 1000);
331
332	41140x	if (lock.owns_lock())
333	14262x	lock.unlock();
334
335	41140x	task_cleanup on_exit{this, &lock, ctx};
336
337		// Flush deferred timerfd programming before blocking
338	41140x	if (timerfd_stale_.exchange(false, std::memory_order_acquire))
339	5044x	update_timerfd();
340
341	41140x	int nfds = ::epoll_wait(
342		epoll_fd_, event_buffer_.data(),
343	41140x	static_cast<int>(event_buffer_.size()), timeout_ms);
344
345	41140x	if (nfds < 0 && errno != EINTR)
346	✗	detail::throw_system_error(make_err(errno), "epoll_wait");
347
348	41140x	bool check_timers = false;
349	41140x	op_queue local_ops;
350
351	92417x	for (int i = 0; i < nfds; ++i)
352		{
353	51277x	if (event_buffer_[i].data.ptr == nullptr)
354		{
355		std::uint64_t val;
356		// NOLINTNEXTLINE(clang-analyzer-unix.BlockInCriticalSection)
357	5172x	[[maybe_unused]] auto r = ::read(event_fd_, &val, sizeof(val));
358	5172x	eventfd_armed_.store(false, std::memory_order_relaxed);
359	5172x	continue;
360	5172x	}
361
362	46105x	if (event_buffer_[i].data.ptr == &timer_fd_)
363		{
364		std::uint64_t expirations;
365		// NOLINTNEXTLINE(clang-analyzer-unix.BlockInCriticalSection)
366		[[maybe_unused]] auto r =
367	5049x	::read(timer_fd_, &expirations, sizeof(expirations));
368	5049x	check_timers = true;
369	5049x	continue;
370	5049x	}
371
372		auto* desc =
373	41056x	static_cast<reactor_descriptor_state*>(event_buffer_[i].data.ptr);
374	41056x	desc->add_ready_events(event_buffer_[i].events);
375
376	41056x	bool expected = false;
377	41056x	if (desc->is_enqueued_.compare_exchange_strong(
378		expected, true, std::memory_order_release,
379		std::memory_order_relaxed))
380		{
381	41056x	local_ops.push(desc);
382		}
383		}
384
385	41140x	if (check_timers)
386		{
387	5049x	timer_svc_->process_expired();
388	5049x	update_timerfd();
389		}
390
391	41140x	lock.lock();
392
393	41140x	if (!local_ops.empty())
394	26298x	completed_ops_.splice(local_ops);
395	41140x	}
396
397		} // namespace boost::corosio::detail
398
399		#endif // BOOST_COROSIO_HAS_EPOLL
400
401		#endif // BOOST_COROSIO_NATIVE_DETAIL_EPOLL_EPOLL_SCHEDULER_HPP
402