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

99.4% Lines (173 / 174, 1 excl) 100.0% Functions (12 / 12)

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

select_scheduler.hpp

f(x) Functions (12)

Function Calls Lines Blocks

Line	Hits	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_SELECT_SELECT_SCHEDULER_HPP
12		#define BOOST_COROSIO_NATIVE_DETAIL_SELECT_SELECT_SCHEDULER_HPP
13
14		#include <boost/corosio/detail/platform.hpp>
15
16		#if BOOST_COROSIO_HAS_SELECT
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/select/select_traits.hpp>
25		#include <boost/corosio/detail/timer_service.hpp>
26		#include <boost/corosio/native/detail/make_err.hpp>
27
28		#include <boost/corosio/detail/except.hpp>
29
30		#include <sys/select.h>
31		#include <unistd.h>
32		#include <errno.h>
33		#include <fcntl.h>
34
35		#include <atomic>
36		#include <chrono>
37		#include <cstdint>
38		#include <limits>
39		#include <mutex>
40		#include <new>
41		#include <unordered_map>
42
43		namespace boost::corosio::detail {
44
45		struct select_op;
46
47		/** POSIX scheduler using select() for I/O multiplexing.
48
49		This scheduler implements the scheduler interface using the POSIX select()
50		call for I/O event notification. It inherits the shared reactor threading
51		model from reactor_scheduler: signal state machine, inline completion
52		budget, work counting, and the do_one event loop.
53
54		The design mirrors epoll_scheduler for behavioral consistency:
55		- Same single-reactor thread coordination model
56		- Same deferred I/O pattern (reactor marks ready; workers do I/O)
57		- Same timer integration pattern
58
59		Known Limitations:
60		- FD_SETSIZE (~1024) limits maximum concurrent connections
61		- O(n) scanning: rebuilds fd_sets each iteration
62		- Level-triggered only (no edge-triggered mode)
63
64		@par Thread Safety
65		All public member functions are thread-safe.
66		*/
67		class BOOST_COROSIO_DECL select_scheduler final : public reactor_scheduler
68		{
69		public:
70		/** Construct the scheduler.
71
72		Creates a self-pipe for reactor interruption.
73
74		@param ctx Reference to the owning execution_context.
75		@param concurrency_hint Hint for expected thread count (unused).
76		*/
77		select_scheduler(capy::execution_context& ctx, int concurrency_hint = -1);
78
79		/// Destroy the scheduler.
80		~select_scheduler() override;
81
82		select_scheduler(select_scheduler const&) = delete;
83		select_scheduler& operator=(select_scheduler const&) = delete;
84
85		/// Shut down the scheduler, draining pending operations.
86		void shutdown() override;
87
88		/** Return the maximum file descriptor value supported.
89
90		Returns FD_SETSIZE - 1, the maximum fd value that can be
91		monitored by select(). Operations with fd >= FD_SETSIZE
92		will fail with EINVAL.
93
94		@return The maximum supported file descriptor value.
95		*/
96		static constexpr int max_fd() noexcept
97		{
98		return FD_SETSIZE - 1;
99		}
100
101		/** Register a descriptor for persistent monitoring.
102
103		The fd is added to the registered_descs_ map and will be
104		included in subsequent select() calls. The reactor is
105		interrupted so a blocked select() rebuilds its fd_sets.
106
107		@param fd The file descriptor to register.
108		@param desc Pointer to descriptor state for this fd.
109
110		@return The error if the fd cannot be tracked, otherwise a
111		default constructed error code.
112		*/
113		std::error_code
114		register_descriptor(int fd, reactor_descriptor_state* desc) const;
115
116		/// No-op: write readiness is watched from registration on.
117		std::error_code
118	2330x	ensure_write_registered(int, reactor_descriptor_state*) const noexcept
119		{
120	2330x	return {};
121		}
122
123		/** Deregister a persistently registered descriptor.
124
125		@param fd The file descriptor to deregister.
126		*/
127		void deregister_descriptor(int fd) const;
128
129		/** Interrupt the reactor so it rebuilds its fd_sets.
130
131		Called when a write, connect, or write-wait op is registered
132		after the reactor's snapshot was taken. Without this,
133		select() may block not watching for writability on the fd.
134		*/
135		void notify_reactor() const;
136
137		/// Watch the read end of the POSIX signal self-pipe (see scheduler.hpp).
138	61x	[[nodiscard]] std::error_code register_signal_reader(int read_fd) override
139		{
140	61x	return register_descriptor(read_fd, signal_pipe_reader_.arm());
141		}
142
143		private:
144		void run_task(lock_type& lock, context_type& ctx, long timeout_us) override;
145		void interrupt_reactor() const override;
146		long calculate_timeout(long requested_timeout_us) const;
147
148		// Watches the global signal self-pipe's read end (armed lazily by
149		// register_signal_reader on the first signal registration).
150		reactor_signal_pipe_reader signal_pipe_reader_;
151
152		// Self-pipe for interrupting select()
153		int pipe_fds_[2]; // [0]=read, [1]=write
154
155		// Per-fd tracking for fd_set building
156		mutable std::unordered_map<int, reactor_descriptor_state*>
157		registered_descs_;
158		mutable int max_fd_ = -1;
159		};
160
161	1237x	inline select_scheduler::select_scheduler(capy::execution_context& ctx, int)
162	1237x	: pipe_fds_{-1, -1}
163	1237x	, max_fd_(-1)
164		{
165	1237x	if (::pipe(pipe_fds_) < 0)
166	1x	detail::throw_system_error(make_err(errno), "pipe");
167
168	3699x	for (int i = 0; i < 2; ++i)
169		{
170	2469x	int flags = ::fcntl(pipe_fds_[i], F_GETFL, 0);
171	2469x	if (flags == -1)
172		{
173	2x	int errn = errno;
174	2x	::close(pipe_fds_[0]);
175	2x	::close(pipe_fds_[1]);
176	2x	detail::throw_system_error(make_err(errn), "fcntl F_GETFL");
177		}
178	2467x	if (::fcntl(pipe_fds_[i], F_SETFL, flags \| O_NONBLOCK) == -1)
179		{
180	2x	int errn = errno;
181	2x	::close(pipe_fds_[0]);
182	2x	::close(pipe_fds_[1]);
183	2x	detail::throw_system_error(make_err(errn), "fcntl F_SETFL");
184		}
185	2465x	if (::fcntl(pipe_fds_[i], F_SETFD, FD_CLOEXEC) == -1)
186		{
187	2x	int errn = errno;
188	2x	::close(pipe_fds_[0]);
189	2x	::close(pipe_fds_[1]);
190	2x	detail::throw_system_error(make_err(errn), "fcntl F_SETFD");
191		}
192		}
193
194	1230x	timer_svc_ = &get_timer_service(ctx, *this);
195	1230x	timer_svc_->set_on_earliest_changed(
196	4113x	timer_service::callback(this, [](void* p) {
197	2883x	static_cast<select_scheduler*>(p)->interrupt_reactor();
198	2883x	}));
199
200	1230x	completed_ops_.push(&task_op_);
201	1251x	}
202
203	2460x	inline select_scheduler::~select_scheduler()
204		{
205	1230x	if (pipe_fds_[0] >= 0)
206	1230x	::close(pipe_fds_[0]);
207	1230x	if (pipe_fds_[1] >= 0)
208	1230x	::close(pipe_fds_[1]);
209	2460x	}
210
211		inline void
212	1230x	select_scheduler::shutdown()
213		{
214	1230x	shutdown_drain();
215
216	1230x	if (pipe_fds_[1] >= 0)
217	1230x	interrupt_reactor();
218	1230x	}
219
220		inline std::error_code
221	5254x	select_scheduler::register_descriptor(
222		int fd, reactor_descriptor_state* desc) const
223		{
224	5254x	if (fd < 0 \|\| fd >= FD_SETSIZE)
225	1x	return make_err(EMFILE);
226
227	5253x	desc->registered_events = reactor_event_read \| reactor_event_write;
228	5253x	desc->unpollable = false; // the state is reused across adoptions
229	5253x	desc->fd = fd;
230	5253x	desc->scheduler_ = this;
231	5253x	desc->mutex.set_enabled(reactor_io_locking_);
232	5253x	desc->ready_events_.store(0, std::memory_order_relaxed);
233
234		{
235	5253x	conditionally_enabled_mutex::scoped_lock lock(desc->mutex);
236	5253x	desc->impl_ref_.reset();
237	5253x	desc->read_ready = false;
238	5253x	desc->write_ready = false;
239	5253x	}
240
241		{
242	5253x	mutex_type::scoped_lock lock(mutex_);
243		try
244		{
245	5253x	registered_descs_[fd] = desc;
246		}
247	1x	catch (std::bad_alloc const&)
248		{
249	1x	return make_err(ENOMEM);
250	1x	}
251	5252x	if (fd > max_fd_)
252	5203x	max_fd_ = fd;
253	5253x	}
254
255	5252x	interrupt_reactor();
256	5252x	return {};
257		}
258
259		inline void
260	5192x	select_scheduler::deregister_descriptor(int fd) const
261		{
262	5192x	mutex_type::scoped_lock lock(mutex_);
263
264	5192x	auto it = registered_descs_.find(fd);
265	5192x	if (it == registered_descs_.end())
266	✗	return;
267
268	5192x	registered_descs_.erase(it);
269
270	5192x	if (fd == max_fd_)
271		{
272	4843x	max_fd_ = pipe_fds_[0];
273	9226x	for (auto& [registered_fd, state] : registered_descs_)
274		{
275	4383x	if (registered_fd > max_fd_)
276	4285x	max_fd_ = registered_fd;
277		}
278		}
279	5192x	}
280
281		inline void
282	5092x	select_scheduler::notify_reactor() const
283		{
284	5092x	interrupt_reactor();
285	5092x	}
286
287		inline void
288	16623x	select_scheduler::interrupt_reactor() const
289		{
290	16623x	char byte = 1;
291	16623x	[[maybe_unused]] auto r = ::write(pipe_fds_[1], &byte, 1);
292	16623x	}
293
294		inline long
295	9175x	select_scheduler::calculate_timeout(long requested_timeout_us) const
296		{
297	9175x	if (requested_timeout_us == 0)
298	−	return 0; // LCOV_EXCL_LINE run_task passes 0 via task_interrupted_, never through this argument
299
300	9175x	auto nearest = timer_svc_->nearest_expiry();
301	9175x	if (nearest == timer_service::time_point::max())
302	2597x	return requested_timeout_us;
303
304	6578x	auto now = std::chrono::steady_clock::now();
305	6578x	if (nearest <= now)
306	362x	return 0;
307
308		auto timer_timeout_us =
309	6216x	std::chrono::duration_cast<std::chrono::microseconds>(nearest - now)
310	6216x	.count();
311
312	6216x	constexpr auto long_max =
313		static_cast<long long>((std::numeric_limits<long>::max)());
314		auto capped_timer_us =
315	6216x	(std::min)((std::max)(static_cast<long long>(timer_timeout_us),
316	6216x	static_cast<long long>(0)),
317	6216x	long_max);
318
319	6216x	if (requested_timeout_us < 0)
320	6214x	return static_cast<long>(capped_timer_us);
321
322		return static_cast<long>(
323	2x	(std::min)(static_cast<long long>(requested_timeout_us),
324	2x	capped_timer_us));
325		}
326
327		inline void
328	38602x	select_scheduler::run_task(lock_type& lock, context_type& ctx, long timeout_us)
329		{
330		long effective_timeout_us =
331	38602x	task_interrupted_ ? 0 : calculate_timeout(timeout_us);
332
333		// Snapshot registered descriptors while holding lock.
334		// Record which directions each fd needs monitored to avoid a hot
335		// loop: select is level-triggered, so a writable socket (nearly
336		// always writable) or an always-readable fd (/dev/zero, a pipe at
337		// EOF) would return select() immediately every iteration if
338		// unconditionally added. Membership in both sets is opt-in: a
339		// parked op or wait in a direction opts that direction in. The
340		// exceptional set is opt-in too, for any parked op or wait:
341		// Darwin reports a character device (/dev/zero, /dev/null) as
342		// exceptional on every call.
343		struct fd_entry
344		{
345		int fd;
346		reactor_descriptor_state* desc;
347		std::uint32_t want;
348		};
349		fd_entry snapshot[FD_SETSIZE];
350	38602x	int snapshot_count = 0;
351
352	129403x	for (auto& [fd, desc] : registered_descs_)
353		{
354	90801x	if (snapshot_count < FD_SETSIZE)
355		{
356	90801x	conditionally_enabled_mutex::scoped_lock desc_lock(desc->mutex);
357	90801x	snapshot[snapshot_count].fd = fd;
358	90801x	snapshot[snapshot_count].desc = desc;
359	90801x	snapshot[snapshot_count].want =
360	90801x	((desc->read_op \|\| desc->wait_read_op) ? reactor_event_read
361	90801x	: 0) \|
362	90452x	((desc->write_op \|\| desc->connect_op \|\| desc->wait_write_op)
363	181253x	? reactor_event_write
364	90801x	: 0) \|
365	90801x	(desc->wait_error_op ? reactor_event_error : 0);
366	90801x	++snapshot_count;
367	90801x	}
368		}
369
370	38602x	if (lock.owns_lock())
371	9176x	lock.unlock();
372
373	38602x	task_cleanup on_exit{this, &lock, ctx};
374
375		fd_set read_fds, write_fds, except_fds;
376	656234x	FD_ZERO(&read_fds);
377	656234x	FD_ZERO(&write_fds);
378	656234x	FD_ZERO(&except_fds);
379
380	38602x	FD_SET(pipe_fds_[0], &read_fds);
381	38602x	int nfds = pipe_fds_[0];
382
383	129403x	for (int i = 0; i < snapshot_count; ++i)
384		{
385	90801x	int fd = snapshot[i].fd;
386	90801x	if (snapshot[i].want & reactor_event_read)
387	11335x	FD_SET(fd, &read_fds);
388	90801x	if (snapshot[i].want & reactor_event_write)
389	2581x	FD_SET(fd, &write_fds);
390	90801x	if (snapshot[i].want != 0)
391	14206x	FD_SET(fd, &except_fds);
392	90801x	if (fd > nfds)
393	36287x	nfds = fd;
394		}
395
396		struct timeval tv;
397	38602x	struct timeval* tv_ptr = nullptr;
398	38602x	if (effective_timeout_us >= 0)
399		{
400	36398x	tv.tv_sec = effective_timeout_us / 1000000;
401	36398x	tv.tv_usec = effective_timeout_us % 1000000;
402	36398x	tv_ptr = &tv;
403		}
404
405	38602x	int ready = ::select(nfds + 1, &read_fds, &write_fds, &except_fds, tv_ptr);
406
407		// EINTR: signal interrupted select(), just retry.
408		// EBADF: an fd was closed between snapshot and select(); retry
409		// with a fresh snapshot from registered_descs_.
410		// Both fall through with no ready descriptors rather than
411		// returning: the caller handed this function an owned lock that
412		// only the epilogue below re-acquires.
413	38602x	if (ready < 0)
414		{
415	3x	if (errno != EINTR && errno != EBADF)
416	1x	detail::throw_system_error(make_err(errno), "select");
417	2x	ready = 0;
418		}
419
420		// Process timers outside the lock
421	38601x	timer_svc_->process_expired();
422
423	38601x	ready_queue local_ops;
424
425	38601x	if (ready > 0)
426		{
427	8306x	if (FD_ISSET(pipe_fds_[0], &read_fds))
428		{
429		char buf[256];
430	13776x	while (::read(pipe_fds_[0], buf, sizeof(buf)) > 0)
431		{
432		}
433		}
434
435	19826x	for (int i = 0; i < snapshot_count; ++i)
436		{
437	11520x	int fd = snapshot[i].fd;
438	11520x	reactor_descriptor_state* desc = snapshot[i].desc;
439
440	11520x	std::uint32_t flags = 0;
441	11520x	if (FD_ISSET(fd, &read_fds))
442	3788x	flags \|= reactor_event_read;
443	11520x	if (FD_ISSET(fd, &write_fds))
444	2318x	flags \|= reactor_event_write;
445	11520x	if (FD_ISSET(fd, &except_fds))
446	6x	flags \|= reactor_event_error;
447
448	11520x	if (flags == 0)
449	5412x	continue;
450
451	6108x	desc->add_ready_events(flags);
452
453	6108x	bool expected = false;
454	6108x	if (desc->is_enqueued_.compare_exchange_strong(
455		expected, true, std::memory_order_release,
456		std::memory_order_relaxed))
457		{
458	4944x	local_ops.push(desc);
459		}
460		}
461		}
462
463	38601x	lock.lock();
464
465	38601x	completed_ops_.splice(local_ops);
466	38602x	}
467
468		} // namespace boost::corosio::detail
469
470		#endif // BOOST_COROSIO_HAS_SELECT
471
472		#endif // BOOST_COROSIO_NATIVE_DETAIL_SELECT_SELECT_SCHEDULER_HPP
473