include/boost/corosio/native/detail/select/select_scheduler.hpp

87.9% Lines (145/165) 100.0% List of functions (10/10)

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

select_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_SELECT_SELECT_SCHEDULER_HPP
11		#define BOOST_COROSIO_NATIVE_DETAIL_SELECT_SELECT_SCHEDULER_HPP
12
13		#include <boost/corosio/detail/platform.hpp>
14
15		#if BOOST_COROSIO_HAS_SELECT
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/select/select_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 <sys/select.h>
33		#include <unistd.h>
34		#include <errno.h>
35		#include <fcntl.h>
36
37		#include <atomic>
38		#include <chrono>
39		#include <cstdint>
40		#include <limits>
41		#include <mutex>
42		#include <unordered_map>
43
44		namespace boost::corosio::detail {
45
46		struct select_op;
47
48		/** POSIX scheduler using select() for I/O multiplexing.
49
50		This scheduler implements the scheduler interface using the POSIX select()
51		call for I/O event notification. It inherits the shared reactor threading
52		model from reactor_scheduler: signal state machine, inline completion
53		budget, work counting, and the do_one event loop.
54
55		The design mirrors epoll_scheduler for behavioral consistency:
56		- Same single-reactor thread coordination model
57		- Same deferred I/O pattern (reactor marks ready; workers do I/O)
58		- Same timer integration pattern
59
60		Known Limitations:
61		- FD_SETSIZE (~1024) limits maximum concurrent connections
62		- O(n) scanning: rebuilds fd_sets each iteration
63		- Level-triggered only (no edge-triggered mode)
64
65		@par Thread Safety
66		All public member functions are thread-safe.
67		*/
68		class BOOST_COROSIO_DECL select_scheduler final : public reactor_scheduler
69		{
70		public:
71		/** Construct the scheduler.
72
73		Creates a self-pipe for reactor interruption.
74
75		@param ctx Reference to the owning execution_context.
76		@param concurrency_hint Hint for expected thread count (unused).
77		*/
78		select_scheduler(capy::execution_context& ctx, int concurrency_hint = -1);
79
80		/// Destroy the scheduler.
81		~select_scheduler() override;
82
83		select_scheduler(select_scheduler const&) = delete;
84		select_scheduler& operator=(select_scheduler const&) = delete;
85
86		/// Shut down the scheduler, draining pending operations.
87		void shutdown() override;
88
89		/** Return the maximum file descriptor value supported.
90
91		Returns FD_SETSIZE - 1, the maximum fd value that can be
92		monitored by select(). Operations with fd >= FD_SETSIZE
93		will fail with EINVAL.
94
95		@return The maximum supported file descriptor value.
96		*/
97		static constexpr int max_fd() noexcept
98		{
99		return FD_SETSIZE - 1;
100		}
101
102		/** Register a descriptor for persistent monitoring.
103
104		The fd is added to the registered_descs_ map and will be
105		included in subsequent select() calls. The reactor is
106		interrupted so a blocked select() rebuilds its fd_sets.
107
108		@param fd The file descriptor to register.
109		@param desc Pointer to descriptor state for this fd.
110		*/
111		void register_descriptor(int fd, reactor_descriptor_state* desc) const;
112
113		/** Deregister a persistently registered descriptor.
114
115		@param fd The file descriptor to deregister.
116		*/
117		void deregister_descriptor(int fd) const;
118
119		/** Interrupt the reactor so it rebuilds its fd_sets.
120
121		Called when a write or connect op is registered after
122		the reactor's snapshot was taken. Without this, select()
123		may block not watching for writability on the fd.
124		*/
125		void notify_reactor() const;
126
127		private:
128		void
129		run_task(lock_type& lock, context_type* ctx,
130		long timeout_us) override;
131		void interrupt_reactor() const override;
132		long calculate_timeout(long requested_timeout_us) const;
133
134		// Self-pipe for interrupting select()
135		int pipe_fds_[2]; // [0]=read, [1]=write
136
137		// Per-fd tracking for fd_set building
138		mutable std::unordered_map<int, reactor_descriptor_state*> registered_descs_;
139		mutable int max_fd_ = -1;
140		};
141
142	229x	inline select_scheduler::select_scheduler(capy::execution_context& ctx, int)
143	229x	: pipe_fds_{-1, -1}
144	229x	, max_fd_(-1)
145		{
146	229x	if (::pipe(pipe_fds_) < 0)
147	✗	detail::throw_system_error(make_err(errno), "pipe");
148
149	687x	for (int i = 0; i < 2; ++i)
150		{
151	458x	int flags = ::fcntl(pipe_fds_[i], F_GETFL, 0);
152	458x	if (flags == -1)
153		{
154	✗	int errn = errno;
155	✗	::close(pipe_fds_[0]);
156	✗	::close(pipe_fds_[1]);
157	✗	detail::throw_system_error(make_err(errn), "fcntl F_GETFL");
158		}
159	458x	if (::fcntl(pipe_fds_[i], F_SETFL, flags \| O_NONBLOCK) == -1)
160		{
161	✗	int errn = errno;
162	✗	::close(pipe_fds_[0]);
163	✗	::close(pipe_fds_[1]);
164	✗	detail::throw_system_error(make_err(errn), "fcntl F_SETFL");
165		}
166	458x	if (::fcntl(pipe_fds_[i], F_SETFD, FD_CLOEXEC) == -1)
167		{
168	✗	int errn = errno;
169	✗	::close(pipe_fds_[0]);
170	✗	::close(pipe_fds_[1]);
171	✗	detail::throw_system_error(make_err(errn), "fcntl F_SETFD");
172		}
173		}
174
175	229x	timer_svc_ = &get_timer_service(ctx, *this);
176	229x	timer_svc_->set_on_earliest_changed(
177	3985x	timer_service::callback(this, [](void* p) {
178	3756x	static_cast<select_scheduler*>(p)->interrupt_reactor();
179	3756x	}));
180
181	229x	get_resolver_service(ctx, *this);
182	229x	get_signal_service(ctx, *this);
183	229x	get_stream_file_service(ctx, *this);
184	229x	get_random_access_file_service(ctx, *this);
185
186	229x	completed_ops_.push(&task_op_);
187	229x	}
188
189	458x	inline select_scheduler::~select_scheduler()
190		{
191	229x	if (pipe_fds_[0] >= 0)
192	229x	::close(pipe_fds_[0]);
193	229x	if (pipe_fds_[1] >= 0)
194	229x	::close(pipe_fds_[1]);
195	458x	}
196
197		inline void
198	229x	select_scheduler::shutdown()
199		{
200	229x	shutdown_drain();
201
202	229x	if (pipe_fds_[1] >= 0)
203	229x	interrupt_reactor();
204	229x	}
205
206		inline void
207	7212x	select_scheduler::register_descriptor(
208		int fd, reactor_descriptor_state* desc) const
209		{
210	7212x	if (fd < 0 \|\| fd >= FD_SETSIZE)
211	✗	detail::throw_system_error(make_err(EINVAL), "select: fd out of range");
212
213	7212x	desc->registered_events = reactor_event_read \| reactor_event_write;
214	7212x	desc->fd = fd;
215	7212x	desc->scheduler_ = this;
216	7212x	desc->mutex.set_enabled(!single_threaded_);
217	7212x	desc->ready_events_.store(0, std::memory_order_relaxed);
218
219		{
220	7212x	conditionally_enabled_mutex::scoped_lock lock(desc->mutex);
221	7212x	desc->impl_ref_.reset();
222	7212x	desc->read_ready = false;
223	7212x	desc->write_ready = false;
224	7212x	}
225
226		{
227	7212x	mutex_type::scoped_lock lock(mutex_);
228	7212x	registered_descs_[fd] = desc;
229	7212x	if (fd > max_fd_)
230	7208x	max_fd_ = fd;
231	7212x	}
232
233	7212x	interrupt_reactor();
234	7212x	}
235
236		inline void
237	7212x	select_scheduler::deregister_descriptor(int fd) const
238		{
239	7212x	mutex_type::scoped_lock lock(mutex_);
240
241	7212x	auto it = registered_descs_.find(fd);
242	7212x	if (it == registered_descs_.end())
243	✗	return;
244
245	7212x	registered_descs_.erase(it);
246
247	7212x	if (fd == max_fd_)
248		{
249	7150x	max_fd_ = pipe_fds_[0];
250	14191x	for (auto& [registered_fd, state] : registered_descs_)
251		{
252	7041x	if (registered_fd > max_fd_)
253	7031x	max_fd_ = registered_fd;
254		}
255		}
256	7212x	}
257
258		inline void
259	3532x	select_scheduler::notify_reactor() const
260		{
261	3532x	interrupt_reactor();
262	3532x	}
263
264		inline void
265	14884x	select_scheduler::interrupt_reactor() const
266		{
267	14884x	char byte = 1;
268	14884x	[[maybe_unused]] auto r = ::write(pipe_fds_[1], &byte, 1);
269	14884x	}
270
271		inline long
272	127796x	select_scheduler::calculate_timeout(long requested_timeout_us) const
273		{
274	127796x	if (requested_timeout_us == 0)
275	✗	return 0;
276
277	127796x	auto nearest = timer_svc_->nearest_expiry();
278	127796x	if (nearest == timer_service::time_point::max())
279	46x	return requested_timeout_us;
280
281	127750x	auto now = std::chrono::steady_clock::now();
282	127750x	if (nearest <= now)
283	870x	return 0;
284
285		auto timer_timeout_us =
286	126880x	std::chrono::duration_cast<std::chrono::microseconds>(nearest - now)
287	126880x	.count();
288
289	126880x	constexpr auto long_max =
290		static_cast<long long>((std::numeric_limits<long>::max)());
291		auto capped_timer_us =
292	126880x	(std::min)((std::max)(static_cast<long long>(timer_timeout_us),
293	126880x	static_cast<long long>(0)),
294	126880x	long_max);
295
296	126880x	if (requested_timeout_us < 0)
297	126874x	return static_cast<long>(capped_timer_us);
298
299		return static_cast<long>(
300	6x	(std::min)(static_cast<long long>(requested_timeout_us),
301	6x	capped_timer_us));
302		}
303
304		inline void
305	153991x	select_scheduler::run_task(
306		lock_type& lock, context_type* ctx, long timeout_us)
307		{
308		long effective_timeout_us =
309	153991x	task_interrupted_ ? 0 : calculate_timeout(timeout_us);
310
311		// Snapshot registered descriptors while holding lock.
312		// Record which fds need write monitoring to avoid a hot loop:
313		// select is level-triggered so writable sockets (nearly always
314		// writable) would cause select() to return immediately every
315		// iteration if unconditionally added to write_fds.
316		struct fd_entry
317		{
318		int fd;
319		reactor_descriptor_state* desc;
320		bool needs_write;
321		};
322		fd_entry snapshot[FD_SETSIZE];
323	153991x	int snapshot_count = 0;
324
325	455963x	for (auto& [fd, desc] : registered_descs_)
326		{
327	301972x	if (snapshot_count < FD_SETSIZE)
328		{
329	301972x	conditionally_enabled_mutex::scoped_lock desc_lock(desc->mutex);
330	301972x	snapshot[snapshot_count].fd = fd;
331	301972x	snapshot[snapshot_count].desc = desc;
332	301972x	snapshot[snapshot_count].needs_write =
333	301972x	(desc->write_op \|\| desc->connect_op);
334	301972x	++snapshot_count;
335	301972x	}
336		}
337
338	153991x	if (lock.owns_lock())
339	127796x	lock.unlock();
340
341	153991x	task_cleanup on_exit{this, &lock, ctx};
342
343		fd_set read_fds, write_fds, except_fds;
344	2617847x	FD_ZERO(&read_fds);
345	2617847x	FD_ZERO(&write_fds);
346	2617847x	FD_ZERO(&except_fds);
347
348	153991x	FD_SET(pipe_fds_[0], &read_fds);
349	153991x	int nfds = pipe_fds_[0];
350
351	455963x	for (int i = 0; i < snapshot_count; ++i)
352		{
353	301972x	int fd = snapshot[i].fd;
354	301972x	FD_SET(fd, &read_fds);
355	301972x	if (snapshot[i].needs_write)
356	3532x	FD_SET(fd, &write_fds);
357	301972x	FD_SET(fd, &except_fds);
358	301972x	if (fd > nfds)
359	153725x	nfds = fd;
360		}
361
362		struct timeval tv;
363	153991x	struct timeval* tv_ptr = nullptr;
364	153991x	if (effective_timeout_us >= 0)
365		{
366	153945x	tv.tv_sec = effective_timeout_us / 1000000;
367	153945x	tv.tv_usec = effective_timeout_us % 1000000;
368	153945x	tv_ptr = &tv;
369		}
370
371	153991x	int ready = ::select(nfds + 1, &read_fds, &write_fds, &except_fds, tv_ptr);
372
373		// EINTR: signal interrupted select(), just retry.
374		// EBADF: an fd was closed between snapshot and select(); retry
375		// with a fresh snapshot from registered_descs_.
376	153991x	if (ready < 0)
377		{
378	✗	if (errno == EINTR \|\| errno == EBADF)
379	✗	return;
380	✗	detail::throw_system_error(make_err(errno), "select");
381		}
382
383		// Process timers outside the lock
384	153991x	timer_svc_->process_expired();
385
386	153991x	op_queue local_ops;
387
388	153991x	if (ready > 0)
389		{
390	134289x	if (FD_ISSET(pipe_fds_[0], &read_fds))
391		{
392		char buf[256];
393	14754x	while (::read(pipe_fds_[0], buf, sizeof(buf)) > 0)
394		{
395		}
396		}
397
398	381715x	for (int i = 0; i < snapshot_count; ++i)
399		{
400	247426x	int fd = snapshot[i].fd;
401	247426x	reactor_descriptor_state* desc = snapshot[i].desc;
402
403	247426x	std::uint32_t flags = 0;
404	247426x	if (FD_ISSET(fd, &read_fds))
405	130624x	flags \|= reactor_event_read;
406	247426x	if (FD_ISSET(fd, &write_fds))
407	3532x	flags \|= reactor_event_write;
408	247426x	if (FD_ISSET(fd, &except_fds))
409	✗	flags \|= reactor_event_error;
410
411	247426x	if (flags == 0)
412	113272x	continue;
413
414	134154x	desc->add_ready_events(flags);
415
416	134154x	bool expected = false;
417	134154x	if (desc->is_enqueued_.compare_exchange_strong(
418		expected, true, std::memory_order_release,
419		std::memory_order_relaxed))
420		{
421	134154x	local_ops.push(desc);
422		}
423		}
424		}
425
426	153991x	lock.lock();
427
428	153991x	if (!local_ops.empty())
429	130625x	completed_ops_.splice(local_ops);
430	153991x	}
431
432		} // namespace boost::corosio::detail
433
434		#endif // BOOST_COROSIO_HAS_SELECT
435
436		#endif // BOOST_COROSIO_NATIVE_DETAIL_SELECT_SELECT_SCHEDULER_HPP
437