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(std::unique_lock<std::mutex>& lock, context_type* ctx,
131 : long timeout_us) override;
132 : void interrupt_reactor() const override;
133 : long calculate_timeout(long requested_timeout_us) const;
134 :
135 : // Self-pipe for interrupting select()
136 : int pipe_fds_[2]; // [0]=read, [1]=write
137 :
138 : // Per-fd tracking for fd_set building
139 : mutable std::unordered_map<int, select_descriptor_state*> registered_descs_;
140 : mutable int max_fd_ = -1;
141 : };
142 :
143 HIT 195 : inline select_scheduler::select_scheduler(capy::execution_context& ctx, int)
144 195 : : pipe_fds_{-1, -1}
145 195 : , max_fd_(-1)
146 : {
147 195 : if (::pipe(pipe_fds_) < 0)
148 MIS 0 : detail::throw_system_error(make_err(errno), "pipe");
149 :
150 HIT 585 : for (int i = 0; i < 2; ++i)
151 : {
152 390 : int flags = ::fcntl(pipe_fds_[i], F_GETFL, 0);
153 390 : if (flags == -1)
154 : {
155 MIS 0 : int errn = errno;
156 0 : ::close(pipe_fds_[0]);
157 0 : ::close(pipe_fds_[1]);
158 0 : detail::throw_system_error(make_err(errn), "fcntl F_GETFL");
159 : }
160 HIT 390 : if (::fcntl(pipe_fds_[i], F_SETFL, flags | O_NONBLOCK) == -1)
161 : {
162 MIS 0 : int errn = errno;
163 0 : ::close(pipe_fds_[0]);
164 0 : ::close(pipe_fds_[1]);
165 0 : detail::throw_system_error(make_err(errn), "fcntl F_SETFL");
166 : }
167 HIT 390 : if (::fcntl(pipe_fds_[i], F_SETFD, FD_CLOEXEC) == -1)
168 : {
169 MIS 0 : int errn = errno;
170 0 : ::close(pipe_fds_[0]);
171 0 : ::close(pipe_fds_[1]);
172 0 : detail::throw_system_error(make_err(errn), "fcntl F_SETFD");
173 : }
174 : }
175 :
176 HIT 195 : timer_svc_ = &get_timer_service(ctx, *this);
177 195 : timer_svc_->set_on_earliest_changed(
178 2637 : timer_service::callback(this, [](void* p) {
179 2442 : static_cast<select_scheduler*>(p)->interrupt_reactor();
180 2442 : }));
181 :
182 195 : get_resolver_service(ctx, *this);
183 195 : get_signal_service(ctx, *this);
184 195 : get_stream_file_service(ctx, *this);
185 195 : get_random_access_file_service(ctx, *this);
186 :
187 195 : completed_ops_.push(&task_op_);
188 195 : }
189 :
190 390 : inline select_scheduler::~select_scheduler()
191 : {
192 195 : if (pipe_fds_[0] >= 0)
193 195 : ::close(pipe_fds_[0]);
194 195 : if (pipe_fds_[1] >= 0)
195 195 : ::close(pipe_fds_[1]);
196 390 : }
197 :
198 : inline void
199 195 : select_scheduler::shutdown()
200 : {
201 195 : shutdown_drain();
202 :
203 195 : if (pipe_fds_[1] >= 0)
204 195 : interrupt_reactor();
205 195 : }
206 :
207 : inline void
208 4591 : select_scheduler::register_descriptor(
209 : int fd, select_descriptor_state* desc) const
210 : {
211 4591 : if (fd < 0 || fd >= FD_SETSIZE)
212 MIS 0 : detail::throw_system_error(make_err(EINVAL), "select: fd out of range");
213 :
214 HIT 4591 : desc->registered_events = reactor_event_read | reactor_event_write;
215 4591 : desc->fd = fd;
216 4591 : desc->scheduler_ = this;
217 4591 : desc->ready_events_.store(0, std::memory_order_relaxed);
218 :
219 : {
220 4591 : std::lock_guard lock(desc->mutex);
221 4591 : desc->impl_ref_.reset();
222 4591 : desc->read_ready = false;
223 4591 : desc->write_ready = false;
224 4591 : }
225 :
226 : {
227 4591 : std::lock_guard lock(mutex_);
228 4591 : registered_descs_[fd] = desc;
229 4591 : if (fd > max_fd_)
230 4587 : max_fd_ = fd;
231 4591 : }
232 :
233 4591 : interrupt_reactor();
234 4591 : }
235 :
236 : inline void
237 4591 : select_scheduler::deregister_descriptor(int fd) const
238 : {
239 4591 : std::lock_guard lock(mutex_);
240 :
241 4591 : auto it = registered_descs_.find(fd);
242 4591 : if (it == registered_descs_.end())
243 MIS 0 : return;
244 :
245 HIT 4591 : registered_descs_.erase(it);
246 :
247 4591 : if (fd == max_fd_)
248 : {
249 4534 : max_fd_ = pipe_fds_[0];
250 8983 : for (auto& [registered_fd, state] : registered_descs_)
251 : {
252 4449 : if (registered_fd > max_fd_)
253 4440 : max_fd_ = registered_fd;
254 : }
255 : }
256 4591 : }
257 :
258 : inline void
259 17952 : select_scheduler::notify_reactor() const
260 : {
261 17952 : interrupt_reactor();
262 17952 : }
263 :
264 : inline void
265 25323 : select_scheduler::interrupt_reactor() const
266 : {
267 25323 : char byte = 1;
268 25323 : [[maybe_unused]] auto r = ::write(pipe_fds_[1], &byte, 1);
269 25323 : }
270 :
271 : inline long
272 79141 : select_scheduler::calculate_timeout(long requested_timeout_us) const
273 : {
274 79141 : if (requested_timeout_us == 0)
275 MIS 0 : return 0;
276 :
277 HIT 79141 : auto nearest = timer_svc_->nearest_expiry();
278 79141 : if (nearest == timer_service::time_point::max())
279 46 : return requested_timeout_us;
280 :
281 79095 : auto now = std::chrono::steady_clock::now();
282 79095 : if (nearest <= now)
283 315 : return 0;
284 :
285 : auto timer_timeout_us =
286 78780 : std::chrono::duration_cast<std::chrono::microseconds>(nearest - now)
287 78780 : .count();
288 :
289 78780 : constexpr auto long_max =
290 : static_cast<long long>((std::numeric_limits<long>::max)());
291 : auto capped_timer_us =
292 78780 : (std::min)((std::max)(static_cast<long long>(timer_timeout_us),
293 78780 : static_cast<long long>(0)),
294 78780 : long_max);
295 :
296 78780 : if (requested_timeout_us < 0)
297 78774 : return static_cast<long>(capped_timer_us);
298 :
299 : return static_cast<long>(
300 6 : (std::min)(static_cast<long long>(requested_timeout_us),
301 6 : capped_timer_us));
302 : }
303 :
304 : inline void
305 91216 : select_scheduler::run_task(
306 : std::unique_lock<std::mutex>& lock, context_type* ctx, long timeout_us)
307 : {
308 : long effective_timeout_us =
309 91216 : 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 : select_descriptor_state* desc;
320 : bool needs_write;
321 : };
322 : fd_entry snapshot[FD_SETSIZE];
323 91216 : int snapshot_count = 0;
324 :
325 281071 : for (auto& [fd, desc] : registered_descs_)
326 : {
327 189855 : if (snapshot_count < FD_SETSIZE)
328 : {
329 189855 : std::lock_guard desc_lock(desc->mutex);
330 189855 : snapshot[snapshot_count].fd = fd;
331 189855 : snapshot[snapshot_count].desc = desc;
332 189855 : snapshot[snapshot_count].needs_write =
333 189855 : (desc->write_op || desc->connect_op);
334 189855 : ++snapshot_count;
335 189855 : }
336 : }
337 :
338 91216 : if (lock.owns_lock())
339 79141 : lock.unlock();
340 :
341 91216 : task_cleanup on_exit{this, &lock, ctx};
342 :
343 : fd_set read_fds, write_fds, except_fds;
344 1550672 : FD_ZERO(&read_fds);
345 1550672 : FD_ZERO(&write_fds);
346 1550672 : FD_ZERO(&except_fds);
347 :
348 91216 : FD_SET(pipe_fds_[0], &read_fds);
349 91216 : int nfds = pipe_fds_[0];
350 :
351 281071 : for (int i = 0; i < snapshot_count; ++i)
352 : {
353 189855 : int fd = snapshot[i].fd;
354 189855 : FD_SET(fd, &read_fds);
355 189855 : if (snapshot[i].needs_write)
356 2242 : FD_SET(fd, &write_fds);
357 189855 : FD_SET(fd, &except_fds);
358 189855 : if (fd > nfds)
359 90953 : nfds = fd;
360 : }
361 :
362 : struct timeval tv;
363 91216 : struct timeval* tv_ptr = nullptr;
364 91216 : if (effective_timeout_us >= 0)
365 : {
366 91170 : tv.tv_sec = effective_timeout_us / 1000000;
367 91170 : tv.tv_usec = effective_timeout_us % 1000000;
368 91170 : tv_ptr = &tv;
369 : }
370 :
371 91216 : 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 91216 : if (ready < 0)
377 : {
378 MIS 0 : if (errno == EINTR || errno == EBADF)
379 0 : return;
380 0 : detail::throw_system_error(make_err(errno), "select");
381 : }
382 :
383 : // Process timers outside the lock
384 HIT 91216 : timer_svc_->process_expired();
385 :
386 91216 : op_queue local_ops;
387 :
388 91216 : if (ready > 0)
389 : {
390 83608 : if (FD_ISSET(pipe_fds_[0], &read_fds))
391 : {
392 : char buf[256];
393 21936 : while (::read(pipe_fds_[0], buf, sizeof(buf)) > 0)
394 : {
395 : }
396 : }
397 :
398 262784 : for (int i = 0; i < snapshot_count; ++i)
399 : {
400 179176 : int fd = snapshot[i].fd;
401 179176 : select_descriptor_state* desc = snapshot[i].desc;
402 :
403 179176 : std::uint32_t flags = 0;
404 179176 : if (FD_ISSET(fd, &read_fds))
405 78082 : flags |= reactor_event_read;
406 179176 : if (FD_ISSET(fd, &write_fds))
407 2242 : flags |= reactor_event_write;
408 179176 : if (FD_ISSET(fd, &except_fds))
409 MIS 0 : flags |= reactor_event_error;
410 :
411 HIT 179176 : if (flags == 0)
412 98854 : continue;
413 :
414 80322 : desc->add_ready_events(flags);
415 :
416 80322 : bool expected = false;
417 80322 : if (desc->is_enqueued_.compare_exchange_strong(
418 : expected, true, std::memory_order_release,
419 : std::memory_order_relaxed))
420 : {
421 80322 : local_ops.push(desc);
422 : }
423 : }
424 : }
425 :
426 91216 : lock.lock();
427 :
428 91216 : if (!local_ops.empty())
429 78082 : completed_ops_.splice(local_ops);
430 91216 : }
431 :
432 : } // namespace boost::corosio::detail
433 :
434 : #endif // BOOST_COROSIO_HAS_SELECT
435 :
436 : #endif // BOOST_COROSIO_NATIVE_DETAIL_SELECT_SELECT_SCHEDULER_HPP
|
