TLA Line data 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_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 <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 : struct select_descriptor_state;
48 :
49 : /** POSIX scheduler using select() for I/O multiplexing.
50 :
51 : This scheduler implements the scheduler interface using the POSIX select()
52 : call for I/O event notification. It inherits the shared reactor threading
53 : model from reactor_scheduler_base: signal state machine, inline completion
54 : budget, work counting, and the do_one event loop.
55 :
56 : The design mirrors epoll_scheduler for behavioral consistency:
57 : - Same single-reactor thread coordination model
58 : - Same deferred I/O pattern (reactor marks ready; workers do I/O)
59 : - Same timer integration pattern
60 :
61 : Known Limitations:
62 : - FD_SETSIZE (~1024) limits maximum concurrent connections
63 : - O(n) scanning: rebuilds fd_sets each iteration
64 : - Level-triggered only (no edge-triggered mode)
65 :
66 : @par Thread Safety
67 : All public member functions are thread-safe.
68 : */
69 : class BOOST_COROSIO_DECL select_scheduler final : public reactor_scheduler_base
70 : {
71 : public:
72 : /** Construct the scheduler.
73 :
74 : Creates a self-pipe for reactor interruption.
75 :
76 : @param ctx Reference to the owning execution_context.
77 : @param concurrency_hint Hint for expected thread count (unused).
78 : */
79 : select_scheduler(capy::execution_context& ctx, int concurrency_hint = -1);
80 :
81 : /// Destroy the scheduler.
82 : ~select_scheduler() override;
83 :
84 : select_scheduler(select_scheduler const&) = delete;
85 : select_scheduler& operator=(select_scheduler const&) = delete;
86 :
87 : /// Shut down the scheduler, draining pending operations.
88 : void shutdown() override;
89 :
90 : /** Return the maximum file descriptor value supported.
91 :
92 : Returns FD_SETSIZE - 1, the maximum fd value that can be
93 : monitored by select(). Operations with fd >= FD_SETSIZE
94 : will fail with EINVAL.
95 :
96 : @return The maximum supported file descriptor value.
97 : */
98 : static constexpr int max_fd() noexcept
99 : {
100 : return FD_SETSIZE - 1;
101 : }
102 :
103 : /** Register a descriptor for persistent monitoring.
104 :
105 : The fd is added to the registered_descs_ map and will be
106 : included in subsequent select() calls. The reactor is
107 : interrupted so a blocked select() rebuilds its fd_sets.
108 :
109 : @param fd The file descriptor to register.
110 : @param desc Pointer to descriptor state for this fd.
111 : */
112 : void register_descriptor(int fd, select_descriptor_state* desc) const;
113 :
114 : /** Deregister a persistently registered descriptor.
115 :
116 : @param fd The file descriptor to deregister.
117 : */
118 : void deregister_descriptor(int fd) const;
119 :
120 : /** Interrupt the reactor so it rebuilds its fd_sets.
121 :
122 : Called when a write or connect op is registered after
123 : the reactor's snapshot was taken. Without this, select()
124 : may block not watching for writability on the fd.
125 : */
126 : void notify_reactor() const;
127 :
128 : private:
129 : void
130 : run_task(lock_type& lock, context_type* ctx) 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, select_descriptor_state*> registered_descs_;
139 : mutable int max_fd_ = -1;
140 : };
141 :
142 HIT 195 : inline select_scheduler::select_scheduler(capy::execution_context& ctx, int)
143 195 : : pipe_fds_{-1, -1}
144 195 : , max_fd_(-1)
145 : {
146 195 : if (::pipe(pipe_fds_) < 0)
147 MIS 0 : detail::throw_system_error(make_err(errno), "pipe");
148 :
149 HIT 585 : for (int i = 0; i < 2; ++i)
150 : {
151 390 : int flags = ::fcntl(pipe_fds_[i], F_GETFL, 0);
152 390 : if (flags == -1)
153 : {
154 MIS 0 : int errn = errno;
155 0 : ::close(pipe_fds_[0]);
156 0 : ::close(pipe_fds_[1]);
157 0 : detail::throw_system_error(make_err(errn), "fcntl F_GETFL");
158 : }
159 HIT 390 : if (::fcntl(pipe_fds_[i], F_SETFL, flags | O_NONBLOCK) == -1)
160 : {
161 MIS 0 : int errn = errno;
162 0 : ::close(pipe_fds_[0]);
163 0 : ::close(pipe_fds_[1]);
164 0 : detail::throw_system_error(make_err(errn), "fcntl F_SETFL");
165 : }
166 HIT 390 : if (::fcntl(pipe_fds_[i], F_SETFD, FD_CLOEXEC) == -1)
167 : {
168 MIS 0 : int errn = errno;
169 0 : ::close(pipe_fds_[0]);
170 0 : ::close(pipe_fds_[1]);
171 0 : detail::throw_system_error(make_err(errn), "fcntl F_SETFD");
172 : }
173 : }
174 :
175 HIT 195 : timer_svc_ = &get_timer_service(ctx, *this);
176 195 : timer_svc_->set_on_earliest_changed(
177 3959 : timer_service::callback(this, [](void* p) {
178 3764 : static_cast<select_scheduler*>(p)->interrupt_reactor();
179 3764 : }));
180 :
181 195 : get_resolver_service(ctx, *this);
182 195 : get_signal_service(ctx, *this);
183 195 : get_stream_file_service(ctx, *this);
184 195 : get_random_access_file_service(ctx, *this);
185 :
186 195 : completed_ops_.push(&task_op_);
187 195 : }
188 :
189 390 : inline select_scheduler::~select_scheduler()
190 : {
191 195 : if (pipe_fds_[0] >= 0)
192 195 : ::close(pipe_fds_[0]);
193 195 : if (pipe_fds_[1] >= 0)
194 195 : ::close(pipe_fds_[1]);
195 390 : }
196 :
197 : inline void
198 195 : select_scheduler::shutdown()
199 : {
200 195 : shutdown_drain();
201 :
202 195 : if (pipe_fds_[1] >= 0)
203 195 : interrupt_reactor();
204 195 : }
205 :
206 : inline void
207 7182 : select_scheduler::register_descriptor(
208 : int fd, select_descriptor_state* desc) const
209 : {
210 7182 : if (fd < 0 || fd >= FD_SETSIZE)
211 MIS 0 : detail::throw_system_error(make_err(EINVAL), "select: fd out of range");
212 :
213 HIT 7182 : desc->registered_events = reactor_event_read | reactor_event_write;
214 7182 : desc->fd = fd;
215 7182 : desc->scheduler_ = this;
216 7182 : desc->mutex.set_enabled(!single_threaded_);
217 7182 : desc->ready_events_.store(0, std::memory_order_relaxed);
218 :
219 : {
220 7182 : conditionally_enabled_mutex::scoped_lock lock(desc->mutex);
221 7182 : desc->impl_ref_.reset();
222 7182 : desc->read_ready = false;
223 7182 : desc->write_ready = false;
224 7182 : }
225 :
226 : {
227 7182 : mutex_type::scoped_lock lock(mutex_);
228 7182 : registered_descs_[fd] = desc;
229 7182 : if (fd > max_fd_)
230 7178 : max_fd_ = fd;
231 7182 : }
232 :
233 7182 : interrupt_reactor();
234 7182 : }
235 :
236 : inline void
237 7182 : select_scheduler::deregister_descriptor(int fd) const
238 : {
239 7182 : mutex_type::scoped_lock lock(mutex_);
240 :
241 7182 : auto it = registered_descs_.find(fd);
242 7182 : if (it == registered_descs_.end())
243 MIS 0 : return;
244 :
245 HIT 7182 : registered_descs_.erase(it);
246 :
247 7182 : if (fd == max_fd_)
248 : {
249 7124 : max_fd_ = pipe_fds_[0];
250 14163 : for (auto& [registered_fd, state] : registered_descs_)
251 : {
252 7039 : if (registered_fd > max_fd_)
253 7030 : max_fd_ = registered_fd;
254 : }
255 : }
256 7182 : }
257 :
258 : inline void
259 26057 : select_scheduler::notify_reactor() const
260 : {
261 26057 : interrupt_reactor();
262 26057 : }
263 :
264 : inline void
265 37341 : select_scheduler::interrupt_reactor() const
266 : {
267 37341 : char byte = 1;
268 37341 : [[maybe_unused]] auto r = ::write(pipe_fds_[1], &byte, 1);
269 37341 : }
270 :
271 : inline long
272 126526 : select_scheduler::calculate_timeout(long requested_timeout_us) const
273 : {
274 126526 : if (requested_timeout_us == 0)
275 MIS 0 : return 0;
276 :
277 HIT 126526 : auto nearest = timer_svc_->nearest_expiry();
278 126526 : if (nearest == timer_service::time_point::max())
279 46 : return requested_timeout_us;
280 :
281 126480 : auto now = std::chrono::steady_clock::now();
282 126480 : if (nearest <= now)
283 363 : return 0;
284 :
285 : auto timer_timeout_us =
286 126117 : std::chrono::duration_cast<std::chrono::microseconds>(nearest - now)
287 126117 : .count();
288 :
289 126117 : constexpr auto long_max =
290 : static_cast<long long>((std::numeric_limits<long>::max)());
291 : auto capped_timer_us =
292 126117 : (std::min)((std::max)(static_cast<long long>(timer_timeout_us),
293 126117 : static_cast<long long>(0)),
294 126117 : long_max);
295 :
296 126117 : if (requested_timeout_us < 0)
297 126117 : return static_cast<long>(capped_timer_us);
298 :
299 : return static_cast<long>(
300 MIS 0 : (std::min)(static_cast<long long>(requested_timeout_us),
301 0 : capped_timer_us));
302 : }
303 :
304 : inline void
305 HIT 151727 : select_scheduler::run_task(
306 : lock_type& lock, context_type* ctx)
307 : {
308 151727 : long effective_timeout_us = task_interrupted_ ? 0 : calculate_timeout(-1);
309 :
310 : // Snapshot registered descriptors while holding lock.
311 : // Record which fds need write monitoring to avoid a hot loop:
312 : // select is level-triggered so writable sockets (nearly always
313 : // writable) would cause select() to return immediately every
314 : // iteration if unconditionally added to write_fds.
315 : struct fd_entry
316 : {
317 : int fd;
318 : select_descriptor_state* desc;
319 : bool needs_write;
320 : };
321 : fd_entry snapshot[FD_SETSIZE];
322 151727 : int snapshot_count = 0;
323 :
324 449307 : for (auto& [fd, desc] : registered_descs_)
325 : {
326 297580 : if (snapshot_count < FD_SETSIZE)
327 : {
328 297580 : conditionally_enabled_mutex::scoped_lock desc_lock(desc->mutex);
329 297580 : snapshot[snapshot_count].fd = fd;
330 297580 : snapshot[snapshot_count].desc = desc;
331 297580 : snapshot[snapshot_count].needs_write =
332 297580 : (desc->write_op || desc->connect_op);
333 297580 : ++snapshot_count;
334 297580 : }
335 : }
336 :
337 151727 : if (lock.owns_lock())
338 126526 : lock.unlock();
339 :
340 151727 : task_cleanup on_exit{this, &lock, ctx};
341 :
342 : fd_set read_fds, write_fds, except_fds;
343 2579359 : FD_ZERO(&read_fds);
344 2579359 : FD_ZERO(&write_fds);
345 2579359 : FD_ZERO(&except_fds);
346 :
347 151727 : FD_SET(pipe_fds_[0], &read_fds);
348 151727 : int nfds = pipe_fds_[0];
349 :
350 449307 : for (int i = 0; i < snapshot_count; ++i)
351 : {
352 297580 : int fd = snapshot[i].fd;
353 297580 : FD_SET(fd, &read_fds);
354 297580 : if (snapshot[i].needs_write)
355 3538 : FD_SET(fd, &write_fds);
356 297580 : FD_SET(fd, &except_fds);
357 297580 : if (fd > nfds)
358 151462 : nfds = fd;
359 : }
360 :
361 : struct timeval tv;
362 151727 : struct timeval* tv_ptr = nullptr;
363 151727 : if (effective_timeout_us >= 0)
364 : {
365 151681 : tv.tv_sec = effective_timeout_us / 1000000;
366 151681 : tv.tv_usec = effective_timeout_us % 1000000;
367 151681 : tv_ptr = &tv;
368 : }
369 :
370 151727 : int ready = ::select(nfds + 1, &read_fds, &write_fds, &except_fds, tv_ptr);
371 :
372 : // EINTR: signal interrupted select(), just retry.
373 : // EBADF: an fd was closed between snapshot and select(); retry
374 : // with a fresh snapshot from registered_descs_.
375 151727 : if (ready < 0)
376 : {
377 MIS 0 : if (errno == EINTR || errno == EBADF)
378 0 : return;
379 0 : detail::throw_system_error(make_err(errno), "select");
380 : }
381 :
382 : // Process timers outside the lock
383 HIT 151727 : timer_svc_->process_expired();
384 :
385 151727 : op_queue local_ops;
386 :
387 151727 : if (ready > 0)
388 : {
389 135723 : if (FD_ISSET(pipe_fds_[0], &read_fds))
390 : {
391 : char buf[256];
392 39200 : while (::read(pipe_fds_[0], buf, sizeof(buf)) > 0)
393 : {
394 : }
395 : }
396 :
397 408437 : for (int i = 0; i < snapshot_count; ++i)
398 : {
399 272714 : int fd = snapshot[i].fd;
400 272714 : select_descriptor_state* desc = snapshot[i].desc;
401 :
402 272714 : std::uint32_t flags = 0;
403 272714 : if (FD_ISSET(fd, &read_fds))
404 128627 : flags |= reactor_event_read;
405 272714 : if (FD_ISSET(fd, &write_fds))
406 3538 : flags |= reactor_event_write;
407 272714 : if (FD_ISSET(fd, &except_fds))
408 MIS 0 : flags |= reactor_event_error;
409 :
410 HIT 272714 : if (flags == 0)
411 140552 : continue;
412 :
413 132162 : desc->add_ready_events(flags);
414 :
415 132162 : bool expected = false;
416 132162 : if (desc->is_enqueued_.compare_exchange_strong(
417 : expected, true, std::memory_order_release,
418 : std::memory_order_relaxed))
419 : {
420 132162 : local_ops.push(desc);
421 : }
422 : }
423 : }
424 :
425 151727 : lock.lock();
426 :
427 151727 : if (!local_ops.empty())
428 128627 : completed_ops_.splice(local_ops);
429 151727 : }
430 :
431 : } // namespace boost::corosio::detail
432 :
433 : #endif // BOOST_COROSIO_HAS_SELECT
434 :
435 : #endif // BOOST_COROSIO_NATIVE_DETAIL_SELECT_SELECT_SCHEDULER_HPP
|
