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_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 : #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 <sys/select.h>
35 : #include <unistd.h>
36 : #include <errno.h>
37 : #include <fcntl.h>
38 :
39 : #include <atomic>
40 : #include <chrono>
41 : #include <cstdint>
42 : #include <limits>
43 : #include <mutex>
44 : #include <new>
45 : #include <unordered_map>
46 :
47 : namespace boost::corosio::detail {
48 :
49 : struct select_op;
50 :
51 : /** POSIX scheduler using select() for I/O multiplexing.
52 :
53 : This scheduler implements the scheduler interface using the POSIX select()
54 : call for I/O event notification. It inherits the shared reactor threading
55 : model from reactor_scheduler: signal state machine, inline completion
56 : budget, work counting, and the do_one event loop.
57 :
58 : The design mirrors epoll_scheduler for behavioral consistency:
59 : - Same single-reactor thread coordination model
60 : - Same deferred I/O pattern (reactor marks ready; workers do I/O)
61 : - Same timer integration pattern
62 :
63 : Known Limitations:
64 : - FD_SETSIZE (~1024) limits maximum concurrent connections
65 : - O(n) scanning: rebuilds fd_sets each iteration
66 : - Level-triggered only (no edge-triggered mode)
67 :
68 : @par Thread Safety
69 : All public member functions are thread-safe.
70 : */
71 : class BOOST_COROSIO_DECL select_scheduler final : public reactor_scheduler
72 : {
73 : public:
74 : /** Construct the scheduler.
75 :
76 : Creates a self-pipe for reactor interruption.
77 :
78 : @param ctx Reference to the owning execution_context.
79 : @param concurrency_hint Hint for expected thread count (unused).
80 : */
81 : select_scheduler(capy::execution_context& ctx, int concurrency_hint = -1);
82 :
83 : /// Destroy the scheduler.
84 : ~select_scheduler() override;
85 :
86 : select_scheduler(select_scheduler const&) = delete;
87 : select_scheduler& operator=(select_scheduler const&) = delete;
88 :
89 : /// Shut down the scheduler, draining pending operations.
90 : void shutdown() override;
91 :
92 : /** Return the maximum file descriptor value supported.
93 :
94 : Returns FD_SETSIZE - 1, the maximum fd value that can be
95 : monitored by select(). Operations with fd >= FD_SETSIZE
96 : will fail with EINVAL.
97 :
98 : @return The maximum supported file descriptor value.
99 : */
100 : static constexpr int max_fd() noexcept
101 : {
102 : return FD_SETSIZE - 1;
103 : }
104 :
105 : /** Register a descriptor for persistent monitoring.
106 :
107 : The fd is added to the registered_descs_ map and will be
108 : included in subsequent select() calls. The reactor is
109 : interrupted so a blocked select() rebuilds its fd_sets.
110 :
111 : @param fd The file descriptor to register.
112 : @param desc Pointer to descriptor state for this fd.
113 :
114 : @return The error if the fd cannot be tracked, otherwise a
115 : default constructed error code.
116 : */
117 : std::error_code
118 : register_descriptor(int fd, reactor_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 : /** Interrupt the reactor so it rebuilds its fd_sets.
127 :
128 : Called when a write, connect, or write-wait op is registered
129 : after the reactor's snapshot was taken. Without this,
130 : select() may block not watching for writability on the fd.
131 : */
132 : void notify_reactor() const;
133 :
134 : /// Watch the read end of the POSIX signal self-pipe (see scheduler.hpp).
135 HIT 55 : [[nodiscard]] std::error_code register_signal_reader(int read_fd) override
136 : {
137 55 : return register_descriptor(read_fd, signal_pipe_reader_.arm());
138 : }
139 :
140 : private:
141 : void run_task(lock_type& lock, context_type& ctx, long timeout_us) override;
142 : void interrupt_reactor() const override;
143 : long calculate_timeout(long requested_timeout_us) const;
144 :
145 : // Watches the global signal self-pipe's read end (armed lazily by
146 : // register_signal_reader on the first signal registration).
147 : reactor_signal_pipe_reader signal_pipe_reader_;
148 :
149 : // Self-pipe for interrupting select()
150 : int pipe_fds_[2]; // [0]=read, [1]=write
151 :
152 : // Per-fd tracking for fd_set building
153 : mutable std::unordered_map<int, reactor_descriptor_state*>
154 : registered_descs_;
155 : mutable int max_fd_ = -1;
156 : };
157 :
158 886 : inline select_scheduler::select_scheduler(capy::execution_context& ctx, int)
159 886 : : pipe_fds_{-1, -1}
160 886 : , max_fd_(-1)
161 : {
162 886 : if (::pipe(pipe_fds_) < 0)
163 1 : detail::throw_system_error(make_err(errno), "pipe");
164 :
165 2646 : for (int i = 0; i < 2; ++i)
166 : {
167 1767 : int flags = ::fcntl(pipe_fds_[i], F_GETFL, 0);
168 1767 : if (flags == -1)
169 : {
170 2 : int errn = errno;
171 2 : ::close(pipe_fds_[0]);
172 2 : ::close(pipe_fds_[1]);
173 2 : detail::throw_system_error(make_err(errn), "fcntl F_GETFL");
174 : }
175 1765 : if (::fcntl(pipe_fds_[i], F_SETFL, flags | O_NONBLOCK) == -1)
176 : {
177 2 : int errn = errno;
178 2 : ::close(pipe_fds_[0]);
179 2 : ::close(pipe_fds_[1]);
180 2 : detail::throw_system_error(make_err(errn), "fcntl F_SETFL");
181 : }
182 1763 : if (::fcntl(pipe_fds_[i], F_SETFD, FD_CLOEXEC) == -1)
183 : {
184 2 : int errn = errno;
185 2 : ::close(pipe_fds_[0]);
186 2 : ::close(pipe_fds_[1]);
187 2 : detail::throw_system_error(make_err(errn), "fcntl F_SETFD");
188 : }
189 : }
190 :
191 879 : timer_svc_ = &get_timer_service(ctx, *this);
192 879 : timer_svc_->set_on_earliest_changed(
193 3628 : timer_service::callback(this, [](void* p) {
194 2749 : static_cast<select_scheduler*>(p)->interrupt_reactor();
195 2749 : }));
196 :
197 879 : get_resolver_service(ctx, *this);
198 879 : get_signal_service(ctx, *this);
199 879 : get_stream_file_service(ctx, *this);
200 879 : get_random_access_file_service(ctx, *this);
201 :
202 879 : completed_ops_.push(&task_op_);
203 900 : }
204 :
205 1758 : inline select_scheduler::~select_scheduler()
206 : {
207 879 : if (pipe_fds_[0] >= 0)
208 879 : ::close(pipe_fds_[0]);
209 879 : if (pipe_fds_[1] >= 0)
210 879 : ::close(pipe_fds_[1]);
211 1758 : }
212 :
213 : inline void
214 879 : select_scheduler::shutdown()
215 : {
216 879 : shutdown_drain();
217 :
218 879 : if (pipe_fds_[1] >= 0)
219 879 : interrupt_reactor();
220 879 : }
221 :
222 : inline std::error_code
223 4894 : select_scheduler::register_descriptor(
224 : int fd, reactor_descriptor_state* desc) const
225 : {
226 4894 : if (fd < 0 || fd >= FD_SETSIZE)
227 1 : return make_err(EMFILE);
228 :
229 4893 : desc->registered_events = reactor_event_read | reactor_event_write;
230 4893 : desc->fd = fd;
231 4893 : desc->scheduler_ = this;
232 4893 : desc->mutex.set_enabled(reactor_io_locking_);
233 4893 : desc->ready_events_.store(0, std::memory_order_relaxed);
234 :
235 : {
236 4893 : conditionally_enabled_mutex::scoped_lock lock(desc->mutex);
237 4893 : desc->impl_ref_.reset();
238 4893 : desc->read_ready = false;
239 4893 : desc->write_ready = false;
240 4893 : }
241 :
242 : {
243 4893 : mutex_type::scoped_lock lock(mutex_);
244 : try
245 : {
246 4893 : registered_descs_[fd] = desc;
247 : }
248 1 : catch (std::bad_alloc const&)
249 : {
250 1 : return make_err(ENOMEM);
251 1 : }
252 4892 : if (fd > max_fd_)
253 4838 : max_fd_ = fd;
254 4893 : }
255 :
256 4892 : interrupt_reactor();
257 4892 : return {};
258 : }
259 :
260 : inline void
261 4838 : select_scheduler::deregister_descriptor(int fd) const
262 : {
263 4838 : mutex_type::scoped_lock lock(mutex_);
264 :
265 4838 : auto it = registered_descs_.find(fd);
266 4838 : if (it == registered_descs_.end())
267 MIS 0 : return;
268 :
269 HIT 4838 : registered_descs_.erase(it);
270 :
271 4838 : if (fd == max_fd_)
272 : {
273 4501 : max_fd_ = pipe_fds_[0];
274 8600 : for (auto& [registered_fd, state] : registered_descs_)
275 : {
276 4099 : if (registered_fd > max_fd_)
277 4006 : max_fd_ = registered_fd;
278 : }
279 : }
280 4838 : }
281 :
282 : inline void
283 2179 : select_scheduler::notify_reactor() const
284 : {
285 2179 : interrupt_reactor();
286 2179 : }
287 :
288 : inline void
289 11496 : select_scheduler::interrupt_reactor() const
290 : {
291 11496 : char byte = 1;
292 11496 : [[maybe_unused]] auto r = ::write(pipe_fds_[1], &byte, 1);
293 11496 : }
294 :
295 : inline long
296 291727 : select_scheduler::calculate_timeout(long requested_timeout_us) const
297 : {
298 291727 : if (requested_timeout_us == 0)
299 : return 0; // LCOV_EXCL_LINE run_task passes 0 via task_interrupted_, never through this argument
300 :
301 291727 : auto nearest = timer_svc_->nearest_expiry();
302 291727 : if (nearest == timer_service::time_point::max())
303 728 : return requested_timeout_us;
304 :
305 290999 : auto now = std::chrono::steady_clock::now();
306 290999 : if (nearest <= now)
307 530 : return 0;
308 :
309 : auto timer_timeout_us =
310 290469 : std::chrono::duration_cast<std::chrono::microseconds>(nearest - now)
311 290469 : .count();
312 :
313 290469 : constexpr auto long_max =
314 : static_cast<long long>((std::numeric_limits<long>::max)());
315 : auto capped_timer_us =
316 290469 : (std::min)((std::max)(static_cast<long long>(timer_timeout_us),
317 290469 : static_cast<long long>(0)),
318 290469 : long_max);
319 :
320 290469 : if (requested_timeout_us < 0)
321 290467 : return static_cast<long>(capped_timer_us);
322 :
323 : return static_cast<long>(
324 2 : (std::min)(static_cast<long long>(requested_timeout_us),
325 2 : capped_timer_us));
326 : }
327 :
328 : inline void
329 315410 : select_scheduler::run_task(lock_type& lock, context_type& ctx, long timeout_us)
330 : {
331 : long effective_timeout_us =
332 315410 : task_interrupted_ ? 0 : calculate_timeout(timeout_us);
333 :
334 : // Snapshot registered descriptors while holding lock.
335 : // Record which fds need write monitoring to avoid a hot loop:
336 : // select is level-triggered so writable sockets (nearly always
337 : // writable) would cause select() to return immediately every
338 : // iteration if unconditionally added to write_fds. Membership
339 : // stays opt-in: a parked write wait opts in the same way a
340 : // parked write or connect op does.
341 : struct fd_entry
342 : {
343 : int fd;
344 : reactor_descriptor_state* desc;
345 : bool needs_write;
346 : };
347 : fd_entry snapshot[FD_SETSIZE];
348 315410 : int snapshot_count = 0;
349 :
350 816233 : for (auto& [fd, desc] : registered_descs_)
351 : {
352 500823 : if (snapshot_count < FD_SETSIZE)
353 : {
354 500823 : conditionally_enabled_mutex::scoped_lock desc_lock(desc->mutex);
355 500823 : snapshot[snapshot_count].fd = fd;
356 500823 : snapshot[snapshot_count].desc = desc;
357 500823 : snapshot[snapshot_count].needs_write =
358 500823 : (desc->write_op || desc->connect_op || desc->wait_write_op);
359 500823 : ++snapshot_count;
360 500823 : }
361 : }
362 :
363 315410 : if (lock.owns_lock())
364 291728 : lock.unlock();
365 :
366 315410 : task_cleanup on_exit{this, &lock, ctx};
367 :
368 : fd_set read_fds, write_fds, except_fds;
369 5361970 : FD_ZERO(&read_fds);
370 5361970 : FD_ZERO(&write_fds);
371 5361970 : FD_ZERO(&except_fds);
372 :
373 315410 : FD_SET(pipe_fds_[0], &read_fds);
374 315410 : int nfds = pipe_fds_[0];
375 :
376 816233 : for (int i = 0; i < snapshot_count; ++i)
377 : {
378 500823 : int fd = snapshot[i].fd;
379 500823 : FD_SET(fd, &read_fds);
380 500823 : if (snapshot[i].needs_write)
381 13218 : FD_SET(fd, &write_fds);
382 500823 : FD_SET(fd, &except_fds);
383 500823 : if (fd > nfds)
384 315019 : nfds = fd;
385 : }
386 :
387 : struct timeval tv;
388 315410 : struct timeval* tv_ptr = nullptr;
389 315410 : if (effective_timeout_us >= 0)
390 : {
391 314700 : tv.tv_sec = effective_timeout_us / 1000000;
392 314700 : tv.tv_usec = effective_timeout_us % 1000000;
393 314700 : tv_ptr = &tv;
394 : }
395 :
396 315410 : int ready = ::select(nfds + 1, &read_fds, &write_fds, &except_fds, tv_ptr);
397 :
398 : // EINTR: signal interrupted select(), just retry.
399 : // EBADF: an fd was closed between snapshot and select(); retry
400 : // with a fresh snapshot from registered_descs_.
401 : // Both fall through with no ready descriptors rather than
402 : // returning: the caller handed this function an owned lock that
403 : // only the epilogue below re-acquires.
404 315410 : if (ready < 0)
405 : {
406 3 : if (errno != EINTR && errno != EBADF)
407 1 : detail::throw_system_error(make_err(errno), "select");
408 2 : ready = 0;
409 : }
410 :
411 : // Process timers outside the lock
412 315409 : timer_svc_->process_expired();
413 :
414 315409 : ready_queue local_ops;
415 :
416 315409 : if (ready > 0)
417 : {
418 299359 : if (FD_ISSET(pipe_fds_[0], &read_fds))
419 : {
420 : char buf[256];
421 10110 : while (::read(pipe_fds_[0], buf, sizeof(buf)) > 0)
422 : {
423 : }
424 : }
425 :
426 754097 : for (int i = 0; i < snapshot_count; ++i)
427 : {
428 454738 : int fd = snapshot[i].fd;
429 454738 : reactor_descriptor_state* desc = snapshot[i].desc;
430 :
431 454738 : std::uint32_t flags = 0;
432 454738 : if (FD_ISSET(fd, &read_fds))
433 298926 : flags |= reactor_event_read;
434 454738 : if (FD_ISSET(fd, &write_fds))
435 2171 : flags |= reactor_event_write;
436 454738 : if (FD_ISSET(fd, &except_fds))
437 16 : flags |= reactor_event_error;
438 :
439 454738 : if (flags == 0)
440 153650 : continue;
441 :
442 301088 : desc->add_ready_events(flags);
443 :
444 301088 : bool expected = false;
445 301088 : if (desc->is_enqueued_.compare_exchange_strong(
446 : expected, true, std::memory_order_release,
447 : std::memory_order_relaxed))
448 : {
449 301088 : local_ops.push(desc);
450 : }
451 : }
452 : }
453 :
454 315409 : lock.lock();
455 :
456 315409 : completed_ops_.splice(local_ops);
457 315410 : }
458 :
459 : } // namespace boost::corosio::detail
460 :
461 : #endif // BOOST_COROSIO_HAS_SELECT
462 :
463 : #endif // BOOST_COROSIO_NATIVE_DETAIL_SELECT_SELECT_SCHEDULER_HPP
|
