include/boost/corosio/native/detail/select/select_scheduler.hpp
72.7% Lines (261/359)
81.8% Functions (27/33)
include/boost/corosio/native/detail/select/select_scheduler.hpp
| Line | Hits | 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/native_scheduler.hpp> | |
| 21 | #include <boost/corosio/detail/scheduler_op.hpp> | |
| 22 | ||
| 23 | #include <boost/corosio/native/detail/select/select_op.hpp> | |
| 24 | #include <boost/corosio/detail/timer_service.hpp> | |
| 25 | #include <boost/corosio/detail/make_err.hpp> | |
| 26 | #include <boost/corosio/native/detail/posix/posix_resolver_service.hpp> | |
| 27 | #include <boost/corosio/native/detail/posix/posix_signal_service.hpp> | |
| 28 | ||
| 29 | #include <boost/corosio/detail/except.hpp> | |
| 30 | #include <boost/corosio/detail/thread_local_ptr.hpp> | |
| 31 | ||
| 32 | #include <sys/select.h> | |
| 33 | #include <sys/socket.h> | |
| 34 | #include <unistd.h> | |
| 35 | #include <errno.h> | |
| 36 | #include <fcntl.h> | |
| 37 | ||
| 38 | #include <algorithm> | |
| 39 | #include <atomic> | |
| 40 | #include <chrono> | |
| 41 | #include <condition_variable> | |
| 42 | #include <cstddef> | |
| 43 | #include <limits> | |
| 44 | #include <mutex> | |
| 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 uses a single reactor model | |
| 55 | where one thread runs select() while other threads wait on a condition | |
| 56 | variable for handler work. This design provides: | |
| 57 | ||
| 58 | - Handler parallelism: N posted handlers can execute on N threads | |
| 59 | - No thundering herd: condition_variable wakes exactly one thread | |
| 60 | - Portability: Works on all POSIX systems | |
| 61 | ||
| 62 | The design mirrors epoll_scheduler for behavioral consistency: | |
| 63 | - Same single-reactor thread coordination model | |
| 64 | - Same work counting semantics | |
| 65 | - Same timer integration pattern | |
| 66 | ||
| 67 | Known Limitations: | |
| 68 | - FD_SETSIZE (~1024) limits maximum concurrent connections | |
| 69 | - O(n) scanning: rebuilds fd_sets each iteration | |
| 70 | - Level-triggered only (no edge-triggered mode) | |
| 71 | ||
| 72 | @par Thread Safety | |
| 73 | All public member functions are thread-safe. | |
| 74 | */ | |
| 75 | class BOOST_COROSIO_DECL select_scheduler final | |
| 76 | : public native_scheduler | |
| 77 | , public capy::execution_context::service | |
| 78 | { | |
| 79 | public: | |
| 80 | using key_type = scheduler; | |
| 81 | ||
| 82 | /** Construct the scheduler. | |
| 83 | ||
| 84 | Creates a self-pipe for reactor interruption. | |
| 85 | ||
| 86 | @param ctx Reference to the owning execution_context. | |
| 87 | @param concurrency_hint Hint for expected thread count (unused). | |
| 88 | */ | |
| 89 | select_scheduler(capy::execution_context& ctx, int concurrency_hint = -1); | |
| 90 | ||
| 91 | ~select_scheduler() override; | |
| 92 | ||
| 93 | select_scheduler(select_scheduler const&) = delete; | |
| 94 | select_scheduler& operator=(select_scheduler const&) = delete; | |
| 95 | ||
| 96 | void shutdown() override; | |
| 97 | void post(std::coroutine_handle<> h) const override; | |
| 98 | void post(scheduler_op* h) const override; | |
| 99 | bool running_in_this_thread() const noexcept override; | |
| 100 | void stop() override; | |
| 101 | bool stopped() const noexcept override; | |
| 102 | void restart() override; | |
| 103 | std::size_t run() override; | |
| 104 | std::size_t run_one() override; | |
| 105 | std::size_t wait_one(long usec) override; | |
| 106 | std::size_t poll() override; | |
| 107 | std::size_t poll_one() override; | |
| 108 | ||
| 109 | /** Return the maximum file descriptor value supported. | |
| 110 | ||
| 111 | Returns FD_SETSIZE - 1, the maximum fd value that can be | |
| 112 | monitored by select(). Operations with fd >= FD_SETSIZE | |
| 113 | will fail with EINVAL. | |
| 114 | ||
| 115 | @return The maximum supported file descriptor value. | |
| 116 | */ | |
| 117 | static constexpr int max_fd() noexcept | |
| 118 | { | |
| 119 | return FD_SETSIZE - 1; | |
| 120 | } | |
| 121 | ||
| 122 | /** Register a file descriptor for monitoring. | |
| 123 | ||
| 124 | @param fd The file descriptor to register. | |
| 125 | @param op The operation associated with this fd. | |
| 126 | @param events Event mask: 1 = read, 2 = write, 3 = both. | |
| 127 | */ | |
| 128 | void register_fd(int fd, select_op* op, int events) const; | |
| 129 | ||
| 130 | /** Unregister a file descriptor from monitoring. | |
| 131 | ||
| 132 | @param fd The file descriptor to unregister. | |
| 133 | @param events Event mask to remove: 1 = read, 2 = write, 3 = both. | |
| 134 | */ | |
| 135 | void deregister_fd(int fd, int events) const; | |
| 136 | ||
| 137 | void work_started() noexcept override; | |
| 138 | void work_finished() noexcept override; | |
| 139 | ||
| 140 | // Event flags for register_fd/deregister_fd | |
| 141 | static constexpr int event_read = 1; | |
| 142 | static constexpr int event_write = 2; | |
| 143 | ||
| 144 | private: | |
| 145 | std::size_t do_one(long timeout_us); | |
| 146 | void run_reactor(std::unique_lock<std::mutex>& lock); | |
| 147 | void wake_one_thread_and_unlock(std::unique_lock<std::mutex>& lock) const; | |
| 148 | void interrupt_reactor() const; | |
| 149 | long calculate_timeout(long requested_timeout_us) const; | |
| 150 | ||
| 151 | // Self-pipe for interrupting select() | |
| 152 | int pipe_fds_[2]; // [0]=read, [1]=write | |
| 153 | ||
| 154 | mutable std::mutex mutex_; | |
| 155 | mutable std::condition_variable wakeup_event_; | |
| 156 | mutable op_queue completed_ops_; | |
| 157 | mutable std::atomic<long> outstanding_work_; | |
| 158 | std::atomic<bool> stopped_; | |
| 159 | bool shutdown_; | |
| 160 | ||
| 161 | // Per-fd state for tracking registered operations | |
| 162 | struct fd_state | |
| 163 | { | |
| 164 | select_op* read_op = nullptr; | |
| 165 | select_op* write_op = nullptr; | |
| 166 | }; | |
| 167 | mutable std::unordered_map<int, fd_state> registered_fds_; | |
| 168 | mutable int max_fd_ = -1; | |
| 169 | ||
| 170 | // Single reactor thread coordination | |
| 171 | mutable bool reactor_running_ = false; | |
| 172 | mutable bool reactor_interrupted_ = false; | |
| 173 | mutable int idle_thread_count_ = 0; | |
| 174 | ||
| 175 | // Sentinel operation for interleaving reactor runs with handler execution. | |
| 176 | // Ensures the reactor runs periodically even when handlers are continuously | |
| 177 | // posted, preventing timer starvation. | |
| 178 | struct task_op final : scheduler_op | |
| 179 | { | |
| 180 | ✗ | void operator()() override {} |
| 181 | ✗ | void destroy() override {} |
| 182 | }; | |
| 183 | task_op task_op_; | |
| 184 | }; | |
| 185 | ||
| 186 | /* | |
| 187 | select Scheduler - Single Reactor Model | |
| 188 | ======================================= | |
| 189 | ||
| 190 | This scheduler mirrors the epoll_scheduler design but uses select() instead | |
| 191 | of epoll for I/O multiplexing. The thread coordination strategy is identical: | |
| 192 | one thread becomes the "reactor" while others wait on a condition variable. | |
| 193 | ||
| 194 | Thread Model | |
| 195 | ------------ | |
| 196 | - ONE thread runs select() at a time (the reactor thread) | |
| 197 | - OTHER threads wait on wakeup_event_ (condition variable) for handlers | |
| 198 | - When work is posted, exactly one waiting thread wakes via notify_one() | |
| 199 | ||
| 200 | Key Differences from epoll | |
| 201 | -------------------------- | |
| 202 | - Uses self-pipe instead of eventfd for interruption (more portable) | |
| 203 | - fd_set rebuilding each iteration (O(n) vs O(1) for epoll) | |
| 204 | - FD_SETSIZE limit (~1024 fds on most systems) | |
| 205 | - Level-triggered only (no edge-triggered mode) | |
| 206 | ||
| 207 | Self-Pipe Pattern | |
| 208 | ----------------- | |
| 209 | To interrupt a blocking select() call (e.g., when work is posted or a timer | |
| 210 | expires), we write a byte to pipe_fds_[1]. The read end pipe_fds_[0] is | |
| 211 | always in the read_fds set, so select() returns immediately. We drain the | |
| 212 | pipe to clear the readable state. | |
| 213 | ||
| 214 | fd-to-op Mapping | |
| 215 | ---------------- | |
| 216 | We use an unordered_map<int, fd_state> to track which operations are | |
| 217 | registered for each fd. This allows O(1) lookup when select() returns | |
| 218 | ready fds. Each fd can have at most one read op and one write op registered. | |
| 219 | */ | |
| 220 | ||
| 221 | namespace select { | |
| 222 | ||
| 223 | struct BOOST_COROSIO_SYMBOL_VISIBLE scheduler_context | |
| 224 | { | |
| 225 | select_scheduler const* key; | |
| 226 | scheduler_context* next; | |
| 227 | }; | |
| 228 | ||
| 229 | inline thread_local_ptr<scheduler_context> context_stack; | |
| 230 | ||
| 231 | struct thread_context_guard | |
| 232 | { | |
| 233 | scheduler_context frame_; | |
| 234 | ||
| 235 | 124 | explicit thread_context_guard(select_scheduler const* ctx) noexcept |
| 236 | 124 | : frame_{ctx, context_stack.get()} |
| 237 | { | |
| 238 | 124 | context_stack.set(&frame_); |
| 239 | 124 | } |
| 240 | ||
| 241 | 124 | ~thread_context_guard() noexcept |
| 242 | { | |
| 243 | 124 | context_stack.set(frame_.next); |
| 244 | 124 | } |
| 245 | }; | |
| 246 | ||
| 247 | struct work_guard | |
| 248 | { | |
| 249 | select_scheduler* self; | |
| 250 | 156773 | ~work_guard() |
| 251 | { | |
| 252 | 156773 | self->work_finished(); |
| 253 | 156773 | } |
| 254 | }; | |
| 255 | ||
| 256 | } // namespace select | |
| 257 | ||
| 258 | 135 | inline select_scheduler::select_scheduler(capy::execution_context& ctx, int) |
| 259 | 135 | : pipe_fds_{-1, -1} |
| 260 | 135 | , outstanding_work_(0) |
| 261 | 135 | , stopped_(false) |
| 262 | 135 | , shutdown_(false) |
| 263 | 135 | , max_fd_(-1) |
| 264 | 135 | , reactor_running_(false) |
| 265 | 135 | , reactor_interrupted_(false) |
| 266 | 270 | , idle_thread_count_(0) |
| 267 | { | |
| 268 | // Create self-pipe for interrupting select() | |
| 269 | 135 | if (::pipe(pipe_fds_) < 0) |
| 270 | ✗ | detail::throw_system_error(make_err(errno), "pipe"); |
| 271 | ||
| 272 | // Set both ends to non-blocking and close-on-exec | |
| 273 | 405 | for (int i = 0; i < 2; ++i) |
| 274 | { | |
| 275 | 270 | int flags = ::fcntl(pipe_fds_[i], F_GETFL, 0); |
| 276 | 270 | if (flags == -1) |
| 277 | { | |
| 278 | ✗ | int errn = errno; |
| 279 | ✗ | ::close(pipe_fds_[0]); |
| 280 | ✗ | ::close(pipe_fds_[1]); |
| 281 | ✗ | detail::throw_system_error(make_err(errn), "fcntl F_GETFL"); |
| 282 | } | |
| 283 | 270 | if (::fcntl(pipe_fds_[i], F_SETFL, flags | O_NONBLOCK) == -1) |
| 284 | { | |
| 285 | ✗ | int errn = errno; |
| 286 | ✗ | ::close(pipe_fds_[0]); |
| 287 | ✗ | ::close(pipe_fds_[1]); |
| 288 | ✗ | detail::throw_system_error(make_err(errn), "fcntl F_SETFL"); |
| 289 | } | |
| 290 | 270 | if (::fcntl(pipe_fds_[i], F_SETFD, FD_CLOEXEC) == -1) |
| 291 | { | |
| 292 | ✗ | int errn = errno; |
| 293 | ✗ | ::close(pipe_fds_[0]); |
| 294 | ✗ | ::close(pipe_fds_[1]); |
| 295 | ✗ | detail::throw_system_error(make_err(errn), "fcntl F_SETFD"); |
| 296 | } | |
| 297 | } | |
| 298 | ||
| 299 | 135 | timer_svc_ = &get_timer_service(ctx, *this); |
| 300 | 135 | timer_svc_->set_on_earliest_changed( |
| 301 | 3792 | timer_service::callback(this, [](void* p) { |
| 302 | 3657 | static_cast<select_scheduler*>(p)->interrupt_reactor(); |
| 303 | 3657 | })); |
| 304 | ||
| 305 | // Initialize resolver service | |
| 306 | 135 | get_resolver_service(ctx, *this); |
| 307 | ||
| 308 | // Initialize signal service | |
| 309 | 135 | get_signal_service(ctx, *this); |
| 310 | ||
| 311 | // Push task sentinel to interleave reactor runs with handler execution | |
| 312 | 135 | completed_ops_.push(&task_op_); |
| 313 | 135 | } |
| 314 | ||
| 315 | 270 | inline select_scheduler::~select_scheduler() |
| 316 | { | |
| 317 | 135 | if (pipe_fds_[0] >= 0) |
| 318 | 135 | ::close(pipe_fds_[0]); |
| 319 | 135 | if (pipe_fds_[1] >= 0) |
| 320 | 135 | ::close(pipe_fds_[1]); |
| 321 | 270 | } |
| 322 | ||
| 323 | inline void | |
| 324 | 135 | select_scheduler::shutdown() |
| 325 | { | |
| 326 | { | |
| 327 | 135 | std::unique_lock lock(mutex_); |
| 328 | 135 | shutdown_ = true; |
| 329 | ||
| 330 | 270 | while (auto* h = completed_ops_.pop()) |
| 331 | { | |
| 332 | 135 | if (h == &task_op_) |
| 333 | 135 | continue; |
| 334 | ✗ | lock.unlock(); |
| 335 | ✗ | h->destroy(); |
| 336 | ✗ | lock.lock(); |
| 337 | 135 | } |
| 338 | 135 | } |
| 339 | ||
| 340 | 135 | outstanding_work_.store(0, std::memory_order_release); |
| 341 | ||
| 342 | 135 | if (pipe_fds_[1] >= 0) |
| 343 | 135 | interrupt_reactor(); |
| 344 | ||
| 345 | 135 | wakeup_event_.notify_all(); |
| 346 | 135 | } |
| 347 | ||
| 348 | inline void | |
| 349 | 4001 | select_scheduler::post(std::coroutine_handle<> h) const |
| 350 | { | |
| 351 | struct post_handler final : scheduler_op | |
| 352 | { | |
| 353 | std::coroutine_handle<> h_; | |
| 354 | ||
| 355 | 4001 | explicit post_handler(std::coroutine_handle<> h) : h_(h) {} |
| 356 | ||
| 357 | 8002 | ~post_handler() override = default; |
| 358 | ||
| 359 | 4001 | void operator()() override |
| 360 | { | |
| 361 | 4001 | auto h = h_; |
| 362 | 4001 | delete this; |
| 363 | 4001 | h.resume(); |
| 364 | 4001 | } |
| 365 | ||
| 366 | ✗ | void destroy() override |
| 367 | { | |
| 368 | ✗ | delete this; |
| 369 | ✗ | } |
| 370 | }; | |
| 371 | ||
| 372 | 4001 | auto ph = std::make_unique<post_handler>(h); |
| 373 | 4001 | outstanding_work_.fetch_add(1, std::memory_order_relaxed); |
| 374 | ||
| 375 | 4001 | std::unique_lock lock(mutex_); |
| 376 | 4001 | completed_ops_.push(ph.release()); |
| 377 | 4001 | wake_one_thread_and_unlock(lock); |
| 378 | 4001 | } |
| 379 | ||
| 380 | inline void | |
| 381 | 145764 | select_scheduler::post(scheduler_op* h) const |
| 382 | { | |
| 383 | 145764 | outstanding_work_.fetch_add(1, std::memory_order_relaxed); |
| 384 | ||
| 385 | 145764 | std::unique_lock lock(mutex_); |
| 386 | 145764 | completed_ops_.push(h); |
| 387 | 145764 | wake_one_thread_and_unlock(lock); |
| 388 | 145764 | } |
| 389 | ||
| 390 | inline bool | |
| 391 | 558 | select_scheduler::running_in_this_thread() const noexcept |
| 392 | { | |
| 393 | 558 | for (auto* c = select::context_stack.get(); c != nullptr; c = c->next) |
| 394 | 368 | if (c->key == this) |
| 395 | 368 | return true; |
| 396 | 190 | return false; |
| 397 | } | |
| 398 | ||
| 399 | inline void | |
| 400 | 103 | select_scheduler::stop() |
| 401 | { | |
| 402 | 103 | bool expected = false; |
| 403 | 103 | if (stopped_.compare_exchange_strong( |
| 404 | expected, true, std::memory_order_release, | |
| 405 | std::memory_order_relaxed)) | |
| 406 | { | |
| 407 | // Wake all threads so they notice stopped_ and exit | |
| 408 | { | |
| 409 | 103 | std::lock_guard lock(mutex_); |
| 410 | 103 | wakeup_event_.notify_all(); |
| 411 | 103 | } |
| 412 | 103 | interrupt_reactor(); |
| 413 | } | |
| 414 | 103 | } |
| 415 | ||
| 416 | inline bool | |
| 417 | 3 | select_scheduler::stopped() const noexcept |
| 418 | { | |
| 419 | 3 | return stopped_.load(std::memory_order_acquire); |
| 420 | } | |
| 421 | ||
| 422 | inline void | |
| 423 | 37 | select_scheduler::restart() |
| 424 | { | |
| 425 | 37 | stopped_.store(false, std::memory_order_release); |
| 426 | 37 | } |
| 427 | ||
| 428 | inline std::size_t | |
| 429 | 100 | select_scheduler::run() |
| 430 | { | |
| 431 | 100 | if (stopped_.load(std::memory_order_acquire)) |
| 432 | ✗ | return 0; |
| 433 | ||
| 434 | 200 | if (outstanding_work_.load(std::memory_order_acquire) == 0) |
| 435 | { | |
| 436 | ✗ | stop(); |
| 437 | ✗ | return 0; |
| 438 | } | |
| 439 | ||
| 440 | 100 | select::thread_context_guard ctx(this); |
| 441 | ||
| 442 | 100 | std::size_t n = 0; |
| 443 | 156849 | while (do_one(-1)) |
| 444 | 156749 | if (n != (std::numeric_limits<std::size_t>::max)()) |
| 445 | 156749 | ++n; |
| 446 | 100 | return n; |
| 447 | 100 | } |
| 448 | ||
| 449 | inline std::size_t | |
| 450 | ✗ | select_scheduler::run_one() |
| 451 | { | |
| 452 | ✗ | if (stopped_.load(std::memory_order_acquire)) |
| 453 | ✗ | return 0; |
| 454 | ||
| 455 | ✗ | if (outstanding_work_.load(std::memory_order_acquire) == 0) |
| 456 | { | |
| 457 | ✗ | stop(); |
| 458 | ✗ | return 0; |
| 459 | } | |
| 460 | ||
| 461 | ✗ | select::thread_context_guard ctx(this); |
| 462 | ✗ | return do_one(-1); |
| 463 | ✗ | } |
| 464 | ||
| 465 | inline std::size_t | |
| 466 | 27 | select_scheduler::wait_one(long usec) |
| 467 | { | |
| 468 | 27 | if (stopped_.load(std::memory_order_acquire)) |
| 469 | 3 | return 0; |
| 470 | ||
| 471 | 48 | if (outstanding_work_.load(std::memory_order_acquire) == 0) |
| 472 | { | |
| 473 | ✗ | stop(); |
| 474 | ✗ | return 0; |
| 475 | } | |
| 476 | ||
| 477 | 24 | select::thread_context_guard ctx(this); |
| 478 | 24 | return do_one(usec); |
| 479 | 24 | } |
| 480 | ||
| 481 | inline std::size_t | |
| 482 | ✗ | select_scheduler::poll() |
| 483 | { | |
| 484 | ✗ | if (stopped_.load(std::memory_order_acquire)) |
| 485 | ✗ | return 0; |
| 486 | ||
| 487 | ✗ | if (outstanding_work_.load(std::memory_order_acquire) == 0) |
| 488 | { | |
| 489 | ✗ | stop(); |
| 490 | ✗ | return 0; |
| 491 | } | |
| 492 | ||
| 493 | ✗ | select::thread_context_guard ctx(this); |
| 494 | ||
| 495 | ✗ | std::size_t n = 0; |
| 496 | ✗ | while (do_one(0)) |
| 497 | ✗ | if (n != (std::numeric_limits<std::size_t>::max)()) |
| 498 | ✗ | ++n; |
| 499 | ✗ | return n; |
| 500 | ✗ | } |
| 501 | ||
| 502 | inline std::size_t | |
| 503 | ✗ | select_scheduler::poll_one() |
| 504 | { | |
| 505 | ✗ | if (stopped_.load(std::memory_order_acquire)) |
| 506 | ✗ | return 0; |
| 507 | ||
| 508 | ✗ | if (outstanding_work_.load(std::memory_order_acquire) == 0) |
| 509 | { | |
| 510 | ✗ | stop(); |
| 511 | ✗ | return 0; |
| 512 | } | |
| 513 | ||
| 514 | ✗ | select::thread_context_guard ctx(this); |
| 515 | ✗ | return do_one(0); |
| 516 | ✗ | } |
| 517 | ||
| 518 | inline void | |
| 519 | 7170 | select_scheduler::register_fd(int fd, select_op* op, int events) const |
| 520 | { | |
| 521 | // Validate fd is within select() limits | |
| 522 | 7170 | if (fd < 0 || fd >= FD_SETSIZE) |
| 523 | ✗ | detail::throw_system_error(make_err(EINVAL), "select: fd out of range"); |
| 524 | ||
| 525 | { | |
| 526 | 7170 | std::lock_guard lock(mutex_); |
| 527 | ||
| 528 | 7170 | auto& state = registered_fds_[fd]; |
| 529 | 7170 | if (events & event_read) |
| 530 | 3726 | state.read_op = op; |
| 531 | 7170 | if (events & event_write) |
| 532 | 3444 | state.write_op = op; |
| 533 | ||
| 534 | 7170 | if (fd > max_fd_) |
| 535 | 231 | max_fd_ = fd; |
| 536 | 7170 | } |
| 537 | ||
| 538 | // Wake the reactor so a thread blocked in select() rebuilds its fd_sets | |
| 539 | // with the newly registered fd. | |
| 540 | 7170 | interrupt_reactor(); |
| 541 | 7170 | } |
| 542 | ||
| 543 | inline void | |
| 544 | 7104 | select_scheduler::deregister_fd(int fd, int events) const |
| 545 | { | |
| 546 | 7104 | std::lock_guard lock(mutex_); |
| 547 | ||
| 548 | 7104 | auto it = registered_fds_.find(fd); |
| 549 | 7104 | if (it == registered_fds_.end()) |
| 550 | 6942 | return; |
| 551 | ||
| 552 | 162 | if (events & event_read) |
| 553 | 162 | it->second.read_op = nullptr; |
| 554 | 162 | if (events & event_write) |
| 555 | ✗ | it->second.write_op = nullptr; |
| 556 | ||
| 557 | // Remove entry if both are null | |
| 558 | 162 | if (!it->second.read_op && !it->second.write_op) |
| 559 | { | |
| 560 | 162 | registered_fds_.erase(it); |
| 561 | ||
| 562 | // Recalculate max_fd_ if needed | |
| 563 | 162 | if (fd == max_fd_) |
| 564 | { | |
| 565 | 161 | max_fd_ = pipe_fds_[0]; // At minimum, the pipe read end |
| 566 | 161 | for (auto& [registered_fd, state] : registered_fds_) |
| 567 | { | |
| 568 | ✗ | if (registered_fd > max_fd_) |
| 569 | ✗ | max_fd_ = registered_fd; |
| 570 | } | |
| 571 | } | |
| 572 | } | |
| 573 | 7104 | } |
| 574 | ||
| 575 | inline void | |
| 576 | 11408 | select_scheduler::work_started() noexcept |
| 577 | { | |
| 578 | 11408 | outstanding_work_.fetch_add(1, std::memory_order_relaxed); |
| 579 | 11408 | } |
| 580 | ||
| 581 | inline void | |
| 582 | 161173 | select_scheduler::work_finished() noexcept |
| 583 | { | |
| 584 | 322346 | if (outstanding_work_.fetch_sub(1, std::memory_order_acq_rel) == 1) |
| 585 | 103 | stop(); |
| 586 | 161173 | } |
| 587 | ||
| 588 | inline void | |
| 589 | 14713 | select_scheduler::interrupt_reactor() const |
| 590 | { | |
| 591 | 14713 | char byte = 1; |
| 592 | 14713 | [[maybe_unused]] auto r = ::write(pipe_fds_[1], &byte, 1); |
| 593 | 14713 | } |
| 594 | ||
| 595 | inline void | |
| 596 | 149765 | select_scheduler::wake_one_thread_and_unlock( |
| 597 | std::unique_lock<std::mutex>& lock) const | |
| 598 | { | |
| 599 | 149765 | if (idle_thread_count_ > 0) |
| 600 | { | |
| 601 | // Idle worker exists - wake it via condvar | |
| 602 | ✗ | wakeup_event_.notify_one(); |
| 603 | ✗ | lock.unlock(); |
| 604 | } | |
| 605 | 149765 | else if (reactor_running_ && !reactor_interrupted_) |
| 606 | { | |
| 607 | // No idle workers but reactor is running - interrupt it | |
| 608 | 3648 | reactor_interrupted_ = true; |
| 609 | 3648 | lock.unlock(); |
| 610 | 3648 | interrupt_reactor(); |
| 611 | } | |
| 612 | else | |
| 613 | { | |
| 614 | // No one to wake | |
| 615 | 146117 | lock.unlock(); |
| 616 | } | |
| 617 | 149765 | } |
| 618 | ||
| 619 | inline long | |
| 620 | 10518 | select_scheduler::calculate_timeout(long requested_timeout_us) const |
| 621 | { | |
| 622 | 10518 | if (requested_timeout_us == 0) |
| 623 | ✗ | return 0; |
| 624 | ||
| 625 | 10518 | auto nearest = timer_svc_->nearest_expiry(); |
| 626 | 10518 | if (nearest == timer_service::time_point::max()) |
| 627 | 37 | return requested_timeout_us; |
| 628 | ||
| 629 | 10481 | auto now = std::chrono::steady_clock::now(); |
| 630 | 10481 | if (nearest <= now) |
| 631 | 154 | return 0; |
| 632 | ||
| 633 | auto timer_timeout_us = | |
| 634 | 10327 | std::chrono::duration_cast<std::chrono::microseconds>(nearest - now) |
| 635 | 10327 | .count(); |
| 636 | ||
| 637 | // Clamp to [0, LONG_MAX] to prevent truncation on 32-bit long platforms | |
| 638 | 10327 | constexpr auto long_max = |
| 639 | static_cast<long long>((std::numeric_limits<long>::max)()); | |
| 640 | auto capped_timer_us = | |
| 641 | 10327 | (std::min)((std::max)(static_cast<long long>(timer_timeout_us), |
| 642 | 10327 | static_cast<long long>(0)), |
| 643 | 10327 | long_max); |
| 644 | ||
| 645 | 10327 | if (requested_timeout_us < 0) |
| 646 | 10327 | return static_cast<long>(capped_timer_us); |
| 647 | ||
| 648 | // requested_timeout_us is already long, so min() result fits in long | |
| 649 | return static_cast<long>( | |
| 650 | ✗ | (std::min)(static_cast<long long>(requested_timeout_us), |
| 651 | ✗ | capped_timer_us)); |
| 652 | } | |
| 653 | ||
| 654 | inline void | |
| 655 | 84242 | select_scheduler::run_reactor(std::unique_lock<std::mutex>& lock) |
| 656 | { | |
| 657 | // Calculate timeout considering timers, use 0 if interrupted | |
| 658 | long effective_timeout_us = | |
| 659 | 84242 | reactor_interrupted_ ? 0 : calculate_timeout(-1); |
| 660 | ||
| 661 | // Build fd_sets from registered_fds_ | |
| 662 | fd_set read_fds, write_fds, except_fds; | |
| 663 | 1432114 | FD_ZERO(&read_fds); |
| 664 | 1432114 | FD_ZERO(&write_fds); |
| 665 | 1432114 | FD_ZERO(&except_fds); |
| 666 | ||
| 667 | // Always include the interrupt pipe | |
| 668 | 84242 | FD_SET(pipe_fds_[0], &read_fds); |
| 669 | 84242 | int nfds = pipe_fds_[0]; |
| 670 | ||
| 671 | // Add registered fds | |
| 672 | 101379 | for (auto& [fd, state] : registered_fds_) |
| 673 | { | |
| 674 | 17137 | if (state.read_op) |
| 675 | 13693 | FD_SET(fd, &read_fds); |
| 676 | 17137 | if (state.write_op) |
| 677 | { | |
| 678 | 3444 | FD_SET(fd, &write_fds); |
| 679 | // Also monitor for errors on connect operations | |
| 680 | 3444 | FD_SET(fd, &except_fds); |
| 681 | } | |
| 682 | 17137 | if (fd > nfds) |
| 683 | 13696 | nfds = fd; |
| 684 | } | |
| 685 | ||
| 686 | // Convert timeout to timeval | |
| 687 | struct timeval tv; | |
| 688 | 84242 | struct timeval* tv_ptr = nullptr; |
| 689 | 84242 | if (effective_timeout_us >= 0) |
| 690 | { | |
| 691 | 84205 | tv.tv_sec = effective_timeout_us / 1000000; |
| 692 | 84205 | tv.tv_usec = effective_timeout_us % 1000000; |
| 693 | 84205 | tv_ptr = &tv; |
| 694 | } | |
| 695 | ||
| 696 | 84242 | lock.unlock(); |
| 697 | ||
| 698 | 84242 | int ready = ::select(nfds + 1, &read_fds, &write_fds, &except_fds, tv_ptr); |
| 699 | 84242 | int saved_errno = errno; |
| 700 | ||
| 701 | // Process timers outside the lock | |
| 702 | 84242 | timer_svc_->process_expired(); |
| 703 | ||
| 704 | 84242 | if (ready < 0 && saved_errno != EINTR) |
| 705 | ✗ | detail::throw_system_error(make_err(saved_errno), "select"); |
| 706 | ||
| 707 | // Re-acquire lock before modifying completed_ops_ | |
| 708 | 84242 | lock.lock(); |
| 709 | ||
| 710 | // Drain the interrupt pipe if readable | |
| 711 | 84242 | if (ready > 0 && FD_ISSET(pipe_fds_[0], &read_fds)) |
| 712 | { | |
| 713 | char buf[256]; | |
| 714 | 21758 | while (::read(pipe_fds_[0], buf, sizeof(buf)) > 0) |
| 715 | { | |
| 716 | } | |
| 717 | } | |
| 718 | ||
| 719 | // Process I/O completions | |
| 720 | 84242 | int completions_queued = 0; |
| 721 | 84242 | if (ready > 0) |
| 722 | { | |
| 723 | // Iterate over registered fds (copy keys to avoid iterator invalidation) | |
| 724 | 10879 | std::vector<int> fds_to_check; |
| 725 | 10879 | fds_to_check.reserve(registered_fds_.size()); |
| 726 | 24607 | for (auto& [fd, state] : registered_fds_) |
| 727 | 13728 | fds_to_check.push_back(fd); |
| 728 | ||
| 729 | 24607 | for (int fd : fds_to_check) |
| 730 | { | |
| 731 | 13728 | auto it = registered_fds_.find(fd); |
| 732 | 13728 | if (it == registered_fds_.end()) |
| 733 | ✗ | continue; |
| 734 | ||
| 735 | 13728 | auto& state = it->second; |
| 736 | ||
| 737 | // Check for errors (especially for connect operations) | |
| 738 | 13728 | bool has_error = FD_ISSET(fd, &except_fds); |
| 739 | ||
| 740 | // Process read readiness | |
| 741 | 13728 | if (state.read_op && (FD_ISSET(fd, &read_fds) || has_error)) |
| 742 | { | |
| 743 | 3564 | auto* op = state.read_op; |
| 744 | // Claim the op by exchanging to unregistered. Both registering and | |
| 745 | // registered states mean the op is ours to complete. | |
| 746 | 3564 | auto prev = op->registered.exchange( |
| 747 | select_registration_state::unregistered, | |
| 748 | std::memory_order_acq_rel); | |
| 749 | 3564 | if (prev != select_registration_state::unregistered) |
| 750 | { | |
| 751 | 3564 | state.read_op = nullptr; |
| 752 | ||
| 753 | 3564 | if (has_error) |
| 754 | { | |
| 755 | ✗ | int errn = 0; |
| 756 | ✗ | socklen_t len = sizeof(errn); |
| 757 | ✗ | if (::getsockopt( |
| 758 | ✗ | fd, SOL_SOCKET, SO_ERROR, &errn, &len) < 0) |
| 759 | ✗ | errn = errno; |
| 760 | ✗ | if (errn == 0) |
| 761 | ✗ | errn = EIO; |
| 762 | ✗ | op->complete(errn, 0); |
| 763 | } | |
| 764 | else | |
| 765 | { | |
| 766 | 3564 | op->perform_io(); |
| 767 | } | |
| 768 | ||
| 769 | 3564 | completed_ops_.push(op); |
| 770 | 3564 | ++completions_queued; |
| 771 | } | |
| 772 | } | |
| 773 | ||
| 774 | // Process write readiness | |
| 775 | 13728 | if (state.write_op && (FD_ISSET(fd, &write_fds) || has_error)) |
| 776 | { | |
| 777 | 3444 | auto* op = state.write_op; |
| 778 | // Claim the op by exchanging to unregistered. Both registering and | |
| 779 | // registered states mean the op is ours to complete. | |
| 780 | 3444 | auto prev = op->registered.exchange( |
| 781 | select_registration_state::unregistered, | |
| 782 | std::memory_order_acq_rel); | |
| 783 | 3444 | if (prev != select_registration_state::unregistered) |
| 784 | { | |
| 785 | 3444 | state.write_op = nullptr; |
| 786 | ||
| 787 | 3444 | if (has_error) |
| 788 | { | |
| 789 | ✗ | int errn = 0; |
| 790 | ✗ | socklen_t len = sizeof(errn); |
| 791 | ✗ | if (::getsockopt( |
| 792 | ✗ | fd, SOL_SOCKET, SO_ERROR, &errn, &len) < 0) |
| 793 | ✗ | errn = errno; |
| 794 | ✗ | if (errn == 0) |
| 795 | ✗ | errn = EIO; |
| 796 | ✗ | op->complete(errn, 0); |
| 797 | } | |
| 798 | else | |
| 799 | { | |
| 800 | 3444 | op->perform_io(); |
| 801 | } | |
| 802 | ||
| 803 | 3444 | completed_ops_.push(op); |
| 804 | 3444 | ++completions_queued; |
| 805 | } | |
| 806 | } | |
| 807 | ||
| 808 | // Clean up empty entries | |
| 809 | 13728 | if (!state.read_op && !state.write_op) |
| 810 | 7008 | registered_fds_.erase(it); |
| 811 | } | |
| 812 | 10879 | } |
| 813 | ||
| 814 | 84242 | if (completions_queued > 0) |
| 815 | { | |
| 816 | 3567 | if (completions_queued == 1) |
| 817 | 126 | wakeup_event_.notify_one(); |
| 818 | else | |
| 819 | 3441 | wakeup_event_.notify_all(); |
| 820 | } | |
| 821 | 84242 | } |
| 822 | ||
| 823 | inline std::size_t | |
| 824 | 156873 | select_scheduler::do_one(long timeout_us) |
| 825 | { | |
| 826 | 156873 | std::unique_lock lock(mutex_); |
| 827 | ||
| 828 | for (;;) | |
| 829 | { | |
| 830 | 241115 | if (stopped_.load(std::memory_order_acquire)) |
| 831 | 100 | return 0; |
| 832 | ||
| 833 | 241015 | scheduler_op* op = completed_ops_.pop(); |
| 834 | ||
| 835 | 241015 | if (op == &task_op_) |
| 836 | { | |
| 837 | 84242 | bool more_handlers = !completed_ops_.empty(); |
| 838 | ||
| 839 | 84242 | if (!more_handlers) |
| 840 | { | |
| 841 | 21036 | if (outstanding_work_.load(std::memory_order_acquire) == 0) |
| 842 | { | |
| 843 | ✗ | completed_ops_.push(&task_op_); |
| 844 | ✗ | return 0; |
| 845 | } | |
| 846 | 10518 | if (timeout_us == 0) |
| 847 | { | |
| 848 | ✗ | completed_ops_.push(&task_op_); |
| 849 | ✗ | return 0; |
| 850 | } | |
| 851 | } | |
| 852 | ||
| 853 | 84242 | reactor_interrupted_ = more_handlers || timeout_us == 0; |
| 854 | 84242 | reactor_running_ = true; |
| 855 | ||
| 856 | 84242 | if (more_handlers && idle_thread_count_ > 0) |
| 857 | ✗ | wakeup_event_.notify_one(); |
| 858 | ||
| 859 | 84242 | run_reactor(lock); |
| 860 | ||
| 861 | 84242 | reactor_running_ = false; |
| 862 | 84242 | completed_ops_.push(&task_op_); |
| 863 | 84242 | continue; |
| 864 | 84242 | } |
| 865 | ||
| 866 | 156773 | if (op != nullptr) |
| 867 | { | |
| 868 | 156773 | lock.unlock(); |
| 869 | 156773 | select::work_guard g{this}; |
| 870 | 156773 | (*op)(); |
| 871 | 156773 | return 1; |
| 872 | 156773 | } |
| 873 | ||
| 874 | ✗ | if (outstanding_work_.load(std::memory_order_acquire) == 0) |
| 875 | ✗ | return 0; |
| 876 | ||
| 877 | ✗ | if (timeout_us == 0) |
| 878 | ✗ | return 0; |
| 879 | ||
| 880 | ✗ | ++idle_thread_count_; |
| 881 | ✗ | if (timeout_us < 0) |
| 882 | ✗ | wakeup_event_.wait(lock); |
| 883 | else | |
| 884 | ✗ | wakeup_event_.wait_for(lock, std::chrono::microseconds(timeout_us)); |
| 885 | ✗ | --idle_thread_count_; |
| 886 | 84242 | } |
| 887 | 156873 | } |
| 888 | ||
| 889 | } // namespace boost::corosio::detail | |
| 890 | ||
| 891 | #endif // BOOST_COROSIO_HAS_SELECT | |
| 892 | ||
| 893 | #endif // BOOST_COROSIO_NATIVE_DETAIL_SELECT_SELECT_SCHEDULER_HPP | |
| 894 |