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

82.7% Lines (124/150) 90.0% List of functions (9/10)

[ ] 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_op.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		struct epoll_op;
47		struct descriptor_state;
48
49		/** Linux scheduler using epoll for I/O multiplexing.
50
51		This scheduler implements the scheduler interface using Linux epoll
52		for efficient I/O event notification. It uses a single reactor model
53		where one thread runs epoll_wait while other threads
54		wait on a condition variable for handler work. This design provides:
55
56		- Handler parallelism: N posted handlers can execute on N threads
57		- No thundering herd: condition_variable wakes exactly one thread
58		- IOCP parity: Behavior matches Windows I/O completion port semantics
59
60		When threads call run(), they first try to execute queued handlers.
61		If the queue is empty and no reactor is running, one thread becomes
62		the reactor and runs epoll_wait. Other threads wait on a condition
63		variable until handlers are available.
64
65		@par Thread Safety
66		All public member functions are thread-safe.
67		*/
68		class BOOST_COROSIO_DECL epoll_scheduler final : public reactor_scheduler
69		{
70		public:
71		/** Construct the scheduler.
72
73		Creates an epoll instance, eventfd for reactor interruption,
74		and timerfd for kernel-managed timer expiry.
75
76		@param ctx Reference to the owning execution_context.
77		@param concurrency_hint Hint for expected thread count (unused).
78		*/
79		epoll_scheduler(capy::execution_context& ctx, int concurrency_hint = -1);
80
81		/// Destroy the scheduler.
82		~epoll_scheduler() override;
83
84		epoll_scheduler(epoll_scheduler const&) = delete;
85		epoll_scheduler& operator=(epoll_scheduler const&) = delete;
86
87		/// Shut down the scheduler, draining pending operations.
88		void shutdown() override;
89
90		/// Apply runtime configuration, resizing the event buffer.
91		void configure_reactor(
92		unsigned max_events,
93		unsigned budget_init,
94		unsigned budget_max,
95		unsigned unassisted) override;
96
97		/** Return the epoll file descriptor.
98
99		Used by socket services to register file descriptors
100		for I/O event notification.
101
102		@return The epoll file descriptor.
103		*/
104		int epoll_fd() const noexcept
105		{
106		return epoll_fd_;
107		}
108
109		/** Register a descriptor for persistent monitoring.
110
111		The fd is registered once and stays registered until explicitly
112		deregistered. Events are dispatched via descriptor_state which
113		tracks pending read/write/connect operations.
114
115		@param fd The file descriptor to register.
116		@param desc Pointer to descriptor data (stored in epoll_event.data.ptr).
117		*/
118		void register_descriptor(int fd, descriptor_state* desc) const;
119
120		/** Deregister a persistently registered descriptor.
121
122		@param fd The file descriptor to deregister.
123		*/
124		void deregister_descriptor(int fd) const;
125
126		private:
127		void
128		run_task(lock_type& lock, context_type* ctx,
129		long timeout_us) override;
130		void interrupt_reactor() const override;
131		void update_timerfd() const;
132
133		int epoll_fd_;
134		int event_fd_;
135		int timer_fd_;
136
137		// Edge-triggered eventfd state
138		mutable std::atomic<bool> eventfd_armed_{false};
139
140		// Set when the earliest timer changes; flushed before epoll_wait
141		mutable std::atomic<bool> timerfd_stale_{false};
142
143		// Event buffer sized from max_events_per_poll_ (set at construction,
144		// resized by configure_reactor via io_context_options).
145		std::vector<epoll_event> event_buffer_;
146		};
147
148	334x	inline epoll_scheduler::epoll_scheduler(capy::execution_context& ctx, int)
149	334x	: epoll_fd_(-1)
150	334x	, event_fd_(-1)
151	334x	, timer_fd_(-1)
152	668x	, event_buffer_(max_events_per_poll_)
153		{
154	334x	epoll_fd_ = ::epoll_create1(EPOLL_CLOEXEC);
155	334x	if (epoll_fd_ < 0)
156	✗	detail::throw_system_error(make_err(errno), "epoll_create1");
157
158	334x	event_fd_ = ::eventfd(0, EFD_NONBLOCK \| EFD_CLOEXEC);
159	334x	if (event_fd_ < 0)
160		{
161	✗	int errn = errno;
162	✗	::close(epoll_fd_);
163	✗	detail::throw_system_error(make_err(errn), "eventfd");
164		}
165
166	334x	timer_fd_ = ::timerfd_create(CLOCK_MONOTONIC, TFD_NONBLOCK \| TFD_CLOEXEC);
167	334x	if (timer_fd_ < 0)
168		{
169	✗	int errn = errno;
170	✗	::close(event_fd_);
171	✗	::close(epoll_fd_);
172	✗	detail::throw_system_error(make_err(errn), "timerfd_create");
173		}
174
175	334x	epoll_event ev{};
176	334x	ev.events = EPOLLIN \| EPOLLET;
177	334x	ev.data.ptr = nullptr;
178	334x	if (::epoll_ctl(epoll_fd_, EPOLL_CTL_ADD, event_fd_, &ev) < 0)
179		{
180	✗	int errn = errno;
181	✗	::close(timer_fd_);
182	✗	::close(event_fd_);
183	✗	::close(epoll_fd_);
184	✗	detail::throw_system_error(make_err(errn), "epoll_ctl");
185		}
186
187	334x	epoll_event timer_ev{};
188	334x	timer_ev.events = EPOLLIN \| EPOLLERR;
189	334x	timer_ev.data.ptr = &timer_fd_;
190	334x	if (::epoll_ctl(epoll_fd_, EPOLL_CTL_ADD, timer_fd_, &timer_ev) < 0)
191		{
192	✗	int errn = errno;
193	✗	::close(timer_fd_);
194	✗	::close(event_fd_);
195	✗	::close(epoll_fd_);
196	✗	detail::throw_system_error(make_err(errn), "epoll_ctl (timerfd)");
197		}
198
199	334x	timer_svc_ = &get_timer_service(ctx, *this);
200	334x	timer_svc_->set_on_earliest_changed(
201	4325x	timer_service::callback(this, [](void* p) {
202	3991x	auto* self = static_cast<epoll_scheduler*>(p);
203	3991x	self->timerfd_stale_.store(true, std::memory_order_release);
204	3991x	self->interrupt_reactor();
205	3991x	}));
206
207	334x	get_resolver_service(ctx, *this);
208	334x	get_signal_service(ctx, *this);
209	334x	get_stream_file_service(ctx, *this);
210	334x	get_random_access_file_service(ctx, *this);
211
212	334x	completed_ops_.push(&task_op_);
213	334x	}
214
215	668x	inline epoll_scheduler::~epoll_scheduler()
216		{
217	334x	if (timer_fd_ >= 0)
218	334x	::close(timer_fd_);
219	334x	if (event_fd_ >= 0)
220	334x	::close(event_fd_);
221	334x	if (epoll_fd_ >= 0)
222	334x	::close(epoll_fd_);
223	668x	}
224
225		inline void
226	334x	epoll_scheduler::shutdown()
227		{
228	334x	shutdown_drain();
229
230	334x	if (event_fd_ >= 0)
231	334x	interrupt_reactor();
232	334x	}
233
234		inline void
235	✗	epoll_scheduler::configure_reactor(
236		unsigned max_events,
237		unsigned budget_init,
238		unsigned budget_max,
239		unsigned unassisted)
240		{
241	✗	reactor_scheduler::configure_reactor(
242		max_events, budget_init, budget_max, unassisted);
243	✗	event_buffer_.resize(max_events_per_poll_);
244	✗	}
245
246		inline void
247	7664x	epoll_scheduler::register_descriptor(int fd, descriptor_state* desc) const
248		{
249	7664x	epoll_event ev{};
250	7664x	ev.events = EPOLLIN \| EPOLLOUT \| EPOLLET \| EPOLLERR \| EPOLLHUP;
251	7664x	ev.data.ptr = desc;
252
253	7664x	if (::epoll_ctl(epoll_fd_, EPOLL_CTL_ADD, fd, &ev) < 0)
254	✗	detail::throw_system_error(make_err(errno), "epoll_ctl (register)");
255
256	7664x	desc->registered_events = ev.events;
257	7664x	desc->fd = fd;
258	7664x	desc->scheduler_ = this;
259	7664x	desc->mutex.set_enabled(!single_threaded_);
260	7664x	desc->ready_events_.store(0, std::memory_order_relaxed);
261
262	7664x	conditionally_enabled_mutex::scoped_lock lock(desc->mutex);
263	7664x	desc->impl_ref_.reset();
264	7664x	desc->read_ready = false;
265	7664x	desc->write_ready = false;
266	7664x	}
267
268		inline void
269	7664x	epoll_scheduler::deregister_descriptor(int fd) const
270		{
271	7664x	::epoll_ctl(epoll_fd_, EPOLL_CTL_DEL, fd, nullptr);
272	7664x	}
273
274		inline void
275	4611x	epoll_scheduler::interrupt_reactor() const
276		{
277	4611x	bool expected = false;
278	4611x	if (eventfd_armed_.compare_exchange_strong(
279		expected, true, std::memory_order_release,
280		std::memory_order_relaxed))
281		{
282	4399x	std::uint64_t val = 1;
283	4399x	[[maybe_unused]] auto r = ::write(event_fd_, &val, sizeof(val));
284		}
285	4611x	}
286
287		inline void
288	7942x	epoll_scheduler::update_timerfd() const
289		{
290	7942x	auto nearest = timer_svc_->nearest_expiry();
291
292	7942x	itimerspec ts{};
293	7942x	int flags = 0;
294
295	7942x	if (nearest == timer_service::time_point::max())
296		{
297		// No timers — disarm by setting to 0 (relative)
298		}
299		else
300		{
301	7889x	auto now = std::chrono::steady_clock::now();
302	7889x	if (nearest <= now)
303		{
304		// Use 1ns instead of 0 — zero disarms the timerfd
305	341x	ts.it_value.tv_nsec = 1;
306		}
307		else
308		{
309	7548x	auto nsec = std::chrono::duration_cast<std::chrono::nanoseconds>(
310	7548x	nearest - now)
311	7548x	.count();
312	7548x	ts.it_value.tv_sec = nsec / 1000000000;
313	7548x	ts.it_value.tv_nsec = nsec % 1000000000;
314	7548x	if (ts.it_value.tv_sec == 0 && ts.it_value.tv_nsec == 0)
315	✗	ts.it_value.tv_nsec = 1;
316		}
317		}
318
319	7942x	if (::timerfd_settime(timer_fd_, flags, &ts, nullptr) < 0)
320	✗	detail::throw_system_error(make_err(errno), "timerfd_settime");
321	7942x	}
322
323		inline void
324	36651x	epoll_scheduler::run_task(
325		lock_type& lock, context_type* ctx, long timeout_us)
326		{
327		int timeout_ms;
328	36651x	if (task_interrupted_)
329	25486x	timeout_ms = 0;
330	11165x	else if (timeout_us < 0)
331	11157x	timeout_ms = -1;
332		else
333	8x	timeout_ms = static_cast<int>((timeout_us + 999) / 1000);
334
335	36651x	if (lock.owns_lock())
336	11165x	lock.unlock();
337
338	36651x	task_cleanup on_exit{this, &lock, ctx};
339
340		// Flush deferred timerfd programming before blocking
341	36651x	if (timerfd_stale_.exchange(false, std::memory_order_acquire))
342	3969x	update_timerfd();
343
344	36651x	int nfds = ::epoll_wait(
345		epoll_fd_, event_buffer_.data(),
346	36651x	static_cast<int>(event_buffer_.size()), timeout_ms);
347
348	36651x	if (nfds < 0 && errno != EINTR)
349	✗	detail::throw_system_error(make_err(errno), "epoll_wait");
350
351	36651x	bool check_timers = false;
352	36651x	op_queue local_ops;
353
354	83263x	for (int i = 0; i < nfds; ++i)
355		{
356	46612x	if (event_buffer_[i].data.ptr == nullptr)
357		{
358		std::uint64_t val;
359		// NOLINTNEXTLINE(clang-analyzer-unix.BlockInCriticalSection)
360	4065x	[[maybe_unused]] auto r = ::read(event_fd_, &val, sizeof(val));
361	4065x	eventfd_armed_.store(false, std::memory_order_relaxed);
362	4065x	continue;
363	4065x	}
364
365	42547x	if (event_buffer_[i].data.ptr == &timer_fd_)
366		{
367		std::uint64_t expirations;
368		// NOLINTNEXTLINE(clang-analyzer-unix.BlockInCriticalSection)
369		[[maybe_unused]] auto r =
370	3973x	::read(timer_fd_, &expirations, sizeof(expirations));
371	3973x	check_timers = true;
372	3973x	continue;
373	3973x	}
374
375		auto* desc =
376	38574x	static_cast<descriptor_state*>(event_buffer_[i].data.ptr);
377	38574x	desc->add_ready_events(event_buffer_[i].events);
378
379	38574x	bool expected = false;
380	38574x	if (desc->is_enqueued_.compare_exchange_strong(
381		expected, true, std::memory_order_release,
382		std::memory_order_relaxed))
383		{
384	38574x	local_ops.push(desc);
385		}
386		}
387
388	36651x	if (check_timers)
389		{
390	3973x	timer_svc_->process_expired();
391	3973x	update_timerfd();
392		}
393
394	36651x	lock.lock();
395
396	36651x	if (!local_ops.empty())
397	24946x	completed_ops_.splice(local_ops);
398	36651x	}
399
400		} // namespace boost::corosio::detail
401
402		#endif // BOOST_COROSIO_HAS_EPOLL
403
404		#endif // BOOST_COROSIO_NATIVE_DETAIL_EPOLL_EPOLL_SCHEDULER_HPP
405