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 :
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 HIT 2709 : ensure_write_registered(int, reactor_descriptor_state*) const noexcept
119 : {
120 2709 : 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).
138 273 : [[nodiscard]] std::error_code register_signal_reader(int read_fd) override
139 : {
140 273 : 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 :
147 : /// Rebuild the poll set a blocked select() was given.
148 467 : void notify_reparked() const override
149 : {
150 467 : interrupt_reactor();
151 467 : }
152 : long calculate_timeout(long requested_timeout_us) const;
153 :
154 : // Watches the global signal self-pipe's read end (armed lazily by
155 : // register_signal_reader on the first signal registration).
156 : reactor_signal_pipe_reader signal_pipe_reader_;
157 :
158 : // Self-pipe for interrupting select()
159 : int pipe_fds_[2]; // [0]=read, [1]=write
160 :
161 : // Per-fd tracking for fd_set building
162 : mutable std::unordered_map<int, reactor_descriptor_state*>
163 : registered_descs_;
164 : mutable int max_fd_ = -1;
165 : };
166 :
167 1958 : inline select_scheduler::select_scheduler(capy::execution_context&, int)
168 1958 : : pipe_fds_{-1, -1}
169 1958 : , max_fd_(-1)
170 : {
171 1958 : if (::pipe(pipe_fds_) < 0)
172 1 : detail::throw_system_error(make_err(errno), "pipe");
173 :
174 5862 : for (int i = 0; i < 2; ++i)
175 : {
176 3911 : int flags = ::fcntl(pipe_fds_[i], F_GETFL, 0);
177 3911 : if (flags == -1)
178 : {
179 2 : int errn = errno;
180 2 : ::close(pipe_fds_[0]);
181 2 : ::close(pipe_fds_[1]);
182 2 : detail::throw_system_error(make_err(errn), "fcntl F_GETFL");
183 : }
184 3909 : if (::fcntl(pipe_fds_[i], F_SETFL, flags | O_NONBLOCK) == -1)
185 : {
186 2 : int errn = errno;
187 2 : ::close(pipe_fds_[0]);
188 2 : ::close(pipe_fds_[1]);
189 2 : detail::throw_system_error(make_err(errn), "fcntl F_SETFL");
190 : }
191 3907 : if (::fcntl(pipe_fds_[i], F_SETFD, FD_CLOEXEC) == -1)
192 : {
193 2 : int errn = errno;
194 2 : ::close(pipe_fds_[0]);
195 2 : ::close(pipe_fds_[1]);
196 2 : detail::throw_system_error(make_err(errn), "fcntl F_SETFD");
197 : }
198 : }
199 :
200 1951 : completed_ops_.push(&task_op_);
201 1972 : }
202 :
203 3902 : inline select_scheduler::~select_scheduler()
204 : {
205 1951 : if (pipe_fds_[0] >= 0)
206 1951 : ::close(pipe_fds_[0]);
207 1951 : if (pipe_fds_[1] >= 0)
208 1951 : ::close(pipe_fds_[1]);
209 3902 : }
210 :
211 : inline void
212 1951 : select_scheduler::shutdown()
213 : {
214 1951 : shutdown_drain();
215 :
216 1951 : if (pipe_fds_[1] >= 0)
217 1951 : interrupt_reactor();
218 1951 : }
219 :
220 : inline std::error_code
221 9846 : select_scheduler::register_descriptor(
222 : int fd, reactor_descriptor_state* desc) const
223 : {
224 9846 : if (fd < 0 || fd >= FD_SETSIZE)
225 1 : return make_err(EMFILE);
226 :
227 : // A completion queued by an earlier registration of the same state
228 : // may still be running, so the fields it reads change under its
229 : // mutex, and the mutex mode is only written when it differs.
230 9845 : if (desc->mutex.enabled() != reactor_io_locking_)
231 3 : desc->mutex.set_enabled(reactor_io_locking_);
232 : {
233 9845 : conditionally_enabled_mutex::scoped_lock lock(desc->mutex);
234 9845 : desc->registered_events = reactor_event_read | reactor_event_write;
235 9845 : desc->unpollable = false; // the state is reused across adoptions
236 9845 : desc->fd = fd;
237 9845 : desc->scheduler_ = this;
238 9845 : desc->ready_events_.store(0, std::memory_order_release);
239 : // object_ref_ is not touched: a queued invocation from an
240 : // earlier registration still owns the reference in it.
241 9845 : desc->read_ready = false;
242 9845 : desc->write_ready = false;
243 9845 : }
244 :
245 : {
246 9845 : mutex_type::scoped_lock lock(mutex_);
247 : try
248 : {
249 9845 : registered_descs_[fd] = desc;
250 : }
251 1 : catch (std::bad_alloc const&)
252 : {
253 1 : return make_err(ENOMEM);
254 1 : }
255 9844 : if (fd > max_fd_)
256 9363 : max_fd_ = fd;
257 9845 : }
258 :
259 9844 : interrupt_reactor();
260 9844 : return {};
261 : }
262 :
263 : inline void
264 9572 : select_scheduler::deregister_descriptor(int fd) const
265 : {
266 9572 : mutex_type::scoped_lock lock(mutex_);
267 :
268 9572 : auto it = registered_descs_.find(fd);
269 9572 : if (it == registered_descs_.end())
270 MIS 0 : return;
271 :
272 HIT 9572 : registered_descs_.erase(it);
273 :
274 9572 : if (fd == max_fd_)
275 : {
276 7658 : max_fd_ = pipe_fds_[0];
277 12714 : for (auto& [registered_fd, state] : registered_descs_)
278 : {
279 5056 : if (registered_fd > max_fd_)
280 4823 : max_fd_ = registered_fd;
281 : }
282 : }
283 9572 : }
284 :
285 : inline void
286 6764 : select_scheduler::notify_reactor() const
287 : {
288 6764 : interrupt_reactor();
289 6764 : }
290 :
291 : inline void
292 25515 : select_scheduler::interrupt_reactor() const
293 : {
294 25515 : char byte = 1;
295 25515 : [[maybe_unused]] auto r = ::write(pipe_fds_[1], &byte, 1);
296 25515 : }
297 :
298 : inline long
299 17846 : select_scheduler::calculate_timeout(long requested_timeout_us) const
300 : {
301 17846 : if (requested_timeout_us == 0)
302 : return 0; // LCOV_EXCL_LINE run_task passes 0 via task_interrupted_, never through this argument
303 :
304 17846 : auto* tsvc = timers();
305 17846 : if (!tsvc)
306 11374 : return requested_timeout_us;
307 6472 : auto nearest = tsvc->nearest_expiry();
308 6472 : if (nearest == timer_service::time_point::max())
309 139 : return requested_timeout_us;
310 :
311 6333 : auto now = std::chrono::steady_clock::now();
312 6333 : if (nearest <= now)
313 155 : return 0;
314 :
315 : auto timer_timeout_us =
316 6178 : std::chrono::duration_cast<std::chrono::microseconds>(nearest - now)
317 6178 : .count();
318 :
319 6178 : constexpr auto long_max =
320 : static_cast<long long>((std::numeric_limits<long>::max)());
321 : auto capped_timer_us =
322 6178 : (std::min)((std::max)(static_cast<long long>(timer_timeout_us),
323 6178 : static_cast<long long>(0)),
324 6178 : long_max);
325 :
326 6178 : if (requested_timeout_us < 0)
327 6172 : return static_cast<long>(capped_timer_us);
328 :
329 : return static_cast<long>(
330 6 : (std::min)(static_cast<long long>(requested_timeout_us),
331 6 : capped_timer_us));
332 : }
333 :
334 : inline void
335 144069 : select_scheduler::run_task(lock_type& lock, context_type& ctx, long timeout_us)
336 : {
337 : long effective_timeout_us =
338 144069 : task_interrupted_ ? 0 : calculate_timeout(timeout_us);
339 :
340 : // Snapshot registered descriptors while holding lock.
341 : // Record which directions each fd needs monitored to avoid a hot
342 : // loop: select is level-triggered, so a writable socket (nearly
343 : // always writable) or an always-readable fd (/dev/zero, a pipe at
344 : // EOF) would return select() immediately every iteration if
345 : // unconditionally added. Membership in both sets is opt-in: a
346 : // parked op or wait in a direction opts that direction in. The
347 : // exceptional set is opt-in too, for any parked op or wait:
348 : // Darwin reports a character device (/dev/zero, /dev/null) as
349 : // exceptional on every call.
350 : struct fd_entry
351 : {
352 : int fd;
353 : reactor_descriptor_state* desc;
354 : std::uint32_t want;
355 : };
356 : fd_entry snapshot[FD_SETSIZE];
357 144069 : int snapshot_count = 0;
358 :
359 : // The caller may have released the lock to hand queued work to a
360 : // peer, and deregister_descriptor() erases from this map under it.
361 144069 : if (!lock.owns_lock())
362 126220 : lock.lock();
363 :
364 398341 : for (auto& [fd, desc] : registered_descs_)
365 : {
366 254272 : if (snapshot_count < FD_SETSIZE)
367 : {
368 254272 : conditionally_enabled_mutex::scoped_lock desc_lock(desc->mutex);
369 254272 : snapshot[snapshot_count].fd = fd;
370 254272 : snapshot[snapshot_count].desc = desc;
371 254272 : snapshot[snapshot_count].want =
372 254272 : ((desc->read_op || desc->wait_read_op) ? reactor_event_read
373 254272 : : 0) |
374 253898 : ((desc->write_op || desc->connect_op || desc->wait_write_op)
375 508170 : ? reactor_event_write
376 254272 : : 0) |
377 254272 : (desc->wait_error_op ? reactor_event_error : 0);
378 254272 : ++snapshot_count;
379 254272 : }
380 : }
381 :
382 144069 : if (lock.owns_lock())
383 144068 : lock.unlock();
384 :
385 144069 : task_cleanup on_exit{this, &lock, ctx};
386 :
387 : fd_set read_fds, write_fds, except_fds;
388 2449173 : FD_ZERO(&read_fds);
389 2449173 : FD_ZERO(&write_fds);
390 2449173 : FD_ZERO(&except_fds);
391 :
392 144069 : FD_SET(pipe_fds_[0], &read_fds);
393 144069 : int nfds = pipe_fds_[0];
394 :
395 398341 : for (int i = 0; i < snapshot_count; ++i)
396 : {
397 254272 : int fd = snapshot[i].fd;
398 254272 : if (snapshot[i].want & reactor_event_read)
399 117409 : FD_SET(fd, &read_fds);
400 254272 : if (snapshot[i].want & reactor_event_write)
401 2973 : FD_SET(fd, &write_fds);
402 254272 : if (snapshot[i].want != 0)
403 120577 : FD_SET(fd, &except_fds);
404 254272 : if (fd > nfds)
405 142229 : nfds = fd;
406 : }
407 :
408 : struct timeval tv;
409 144069 : struct timeval* tv_ptr = nullptr;
410 144069 : if (effective_timeout_us >= 0)
411 : {
412 133204 : tv.tv_sec = effective_timeout_us / 1000000;
413 133204 : tv.tv_usec = effective_timeout_us % 1000000;
414 133204 : tv_ptr = &tv;
415 : }
416 :
417 144069 : int ready = ::select(nfds + 1, &read_fds, &write_fds, &except_fds, tv_ptr);
418 :
419 : // EINTR: signal interrupted select(), just retry.
420 : // EBADF: an fd was closed between snapshot and select(); retry
421 : // with a fresh snapshot from registered_descs_.
422 : // Both fall through with no ready descriptors rather than
423 : // returning: the caller handed this function an owned lock that
424 : // only the epilogue below re-acquires.
425 144069 : if (ready < 0)
426 : {
427 3 : if (errno != EINTR && errno != EBADF)
428 1 : detail::throw_system_error(make_err(errno), "select");
429 2 : ready = 0;
430 : }
431 :
432 : // Process timers outside the lock
433 144068 : if (auto* tsvc = timers())
434 27882 : tsvc->process_expired();
435 :
436 144068 : ready_queue local_ops;
437 :
438 144068 : if (ready > 0)
439 : {
440 18344 : if (FD_ISSET(pipe_fds_[0], &read_fds))
441 : {
442 : char buf[256];
443 22660 : while (::read(pipe_fds_[0], buf, sizeof(buf)) > 0)
444 : {
445 : }
446 : }
447 :
448 105795 : for (int i = 0; i < snapshot_count; ++i)
449 : {
450 87451 : int fd = snapshot[i].fd;
451 87451 : reactor_descriptor_state* desc = snapshot[i].desc;
452 :
453 87451 : std::uint32_t flags = 0;
454 87451 : if (FD_ISSET(fd, &read_fds))
455 9725 : flags |= reactor_event_read;
456 87451 : if (FD_ISSET(fd, &write_fds))
457 2689 : flags |= reactor_event_write;
458 87451 : if (FD_ISSET(fd, &except_fds))
459 10 : flags |= reactor_event_error;
460 :
461 87451 : if (flags == 0)
462 75035 : continue;
463 :
464 12416 : desc->add_ready_events(flags);
465 :
466 12416 : bool expected = false;
467 12416 : if (desc->is_enqueued_.compare_exchange_strong(
468 : expected, true, std::memory_order_release,
469 : std::memory_order_relaxed))
470 : {
471 6215 : local_ops.push(desc);
472 : }
473 : }
474 : }
475 :
476 144068 : lock.lock();
477 :
478 144068 : completed_ops_.splice(local_ops);
479 144069 : }
480 :
481 : } // namespace boost::corosio::detail
482 :
483 : #endif // BOOST_COROSIO_HAS_SELECT
484 :
485 : #endif // BOOST_COROSIO_NATIVE_DETAIL_SELECT_SELECT_SCHEDULER_HPP
|
