. /include/boost/corosio /native /detail /epoll /epoll_scheduler.hpp
97.74% Lines (173/177) 100.00% Functions (14/14)
[ ] prev/next file n p prev/next uncovered line
TLA Baseline Branch
Line Hits Code Line Hits Code
1   // 1   //
2   // Copyright (c) 2026 Steve Gerbino 2   // Copyright (c) 2026 Steve Gerbino
3   // Copyright (c) 2026 Michael Vandeberg 3   // Copyright (c) 2026 Michael Vandeberg
4   // 4   //
5   // Distributed under the Boost Software License, Version 1.0. (See accompanying 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) 6   // file LICENSE_1_0.txt or copy at http://www.boost.org/LICENSE_1_0.txt)
7   // 7   //
8   // Official repository: https://github.com/cppalliance/corosio 8   // Official repository: https://github.com/cppalliance/corosio
9   // 9   //
10   10  
11   #ifndef BOOST_COROSIO_NATIVE_DETAIL_EPOLL_EPOLL_SCHEDULER_HPP 11   #ifndef BOOST_COROSIO_NATIVE_DETAIL_EPOLL_EPOLL_SCHEDULER_HPP
12   #define BOOST_COROSIO_NATIVE_DETAIL_EPOLL_EPOLL_SCHEDULER_HPP 12   #define BOOST_COROSIO_NATIVE_DETAIL_EPOLL_EPOLL_SCHEDULER_HPP
13   13  
14   #include <boost/corosio/detail/platform.hpp> 14   #include <boost/corosio/detail/platform.hpp>
15   15  
16   #if BOOST_COROSIO_HAS_EPOLL 16   #if BOOST_COROSIO_HAS_EPOLL
17   17  
18   #include <boost/corosio/detail/config.hpp> 18   #include <boost/corosio/detail/config.hpp>
19   #include <boost/capy/ex/execution_context.hpp> 19   #include <boost/capy/ex/execution_context.hpp>
20   20  
21   #include <boost/corosio/native/detail/reactor/reactor_scheduler.hpp> 21   #include <boost/corosio/native/detail/reactor/reactor_scheduler.hpp>
22   #include <boost/corosio/native/detail/reactor/reactor_signal_pipe.hpp> 22   #include <boost/corosio/native/detail/reactor/reactor_signal_pipe.hpp>
23   23  
24   #include <boost/corosio/native/detail/epoll/epoll_traits.hpp> 24   #include <boost/corosio/native/detail/epoll/epoll_traits.hpp>
25   #include <boost/corosio/detail/timer_service.hpp> 25   #include <boost/corosio/detail/timer_service.hpp>
26   #include <boost/corosio/native/detail/make_err.hpp> 26   #include <boost/corosio/native/detail/make_err.hpp>
27   27  
28   #include <boost/corosio/detail/except.hpp> 28   #include <boost/corosio/detail/except.hpp>
29   29  
30   #include <atomic> 30   #include <atomic>
31   #include <chrono> 31   #include <chrono>
32   #include <cstdint> 32   #include <cstdint>
33   #include <limits> 33   #include <limits>
34   #include <mutex> 34   #include <mutex>
35   #include <vector> 35   #include <vector>
36   36  
37   #include <errno.h> 37   #include <errno.h>
38   #include <sys/epoll.h> 38   #include <sys/epoll.h>
39   #include <sys/eventfd.h> 39   #include <sys/eventfd.h>
40   #include <sys/timerfd.h> 40   #include <sys/timerfd.h>
41   #include <unistd.h> 41   #include <unistd.h>
42   42  
43   namespace boost::corosio::detail { 43   namespace boost::corosio::detail {
44   44  
45   /** Linux scheduler using epoll for I/O multiplexing. 45   /** Linux scheduler using epoll for I/O multiplexing.
46   46  
47   This scheduler implements the scheduler interface using Linux epoll 47   This scheduler implements the scheduler interface using Linux epoll
48   for efficient I/O event notification. It uses a single reactor model 48   for efficient I/O event notification. It uses a single reactor model
49   where one thread runs epoll_wait while other threads 49   where one thread runs epoll_wait while other threads
50   wait on a condition variable for handler work. This design provides: 50   wait on a condition variable for handler work. This design provides:
51   51  
52   - Handler parallelism: N posted handlers can execute on N threads 52   - Handler parallelism: N posted handlers can execute on N threads
53   - No thundering herd: condition_variable wakes exactly one thread 53   - No thundering herd: condition_variable wakes exactly one thread
54   - IOCP parity: Behavior matches Windows I/O completion port semantics 54   - IOCP parity: Behavior matches Windows I/O completion port semantics
55   55  
56   When threads call run(), they first try to execute queued handlers. 56   When threads call run(), they first try to execute queued handlers.
57   If the queue is empty and no reactor is running, one thread becomes 57   If the queue is empty and no reactor is running, one thread becomes
58   the reactor and runs epoll_wait. Other threads wait on a condition 58   the reactor and runs epoll_wait. Other threads wait on a condition
59   variable until handlers are available. 59   variable until handlers are available.
60   60  
61   @par Thread Safety 61   @par Thread Safety
62   All public member functions are thread-safe. 62   All public member functions are thread-safe.
63   */ 63   */
64   class BOOST_COROSIO_DECL epoll_scheduler final : public reactor_scheduler 64   class BOOST_COROSIO_DECL epoll_scheduler final : public reactor_scheduler
65   { 65   {
66   public: 66   public:
67   /** Construct the scheduler. 67   /** Construct the scheduler.
68   68  
69   Creates an epoll instance and an eventfd for reactor 69   Creates an epoll instance and an eventfd for reactor
70   interruption. The timerfd for timer expiry is created with the 70   interruption. The timerfd for timer expiry is created with the
71   context's first timer. 71   context's first timer.
72   72  
73   @param ctx Reference to the owning execution_context. 73   @param ctx Reference to the owning execution_context.
74   @param concurrency_hint Hint for expected thread count (unused). 74   @param concurrency_hint Hint for expected thread count (unused).
75   */ 75   */
76   epoll_scheduler(capy::execution_context& ctx, int concurrency_hint = -1); 76   epoll_scheduler(capy::execution_context& ctx, int concurrency_hint = -1);
77   77  
78   /// Destroy the scheduler. 78   /// Destroy the scheduler.
79   ~epoll_scheduler() override; 79   ~epoll_scheduler() override;
80   80  
81   epoll_scheduler(epoll_scheduler const&) = delete; 81   epoll_scheduler(epoll_scheduler const&) = delete;
82   epoll_scheduler& operator=(epoll_scheduler const&) = delete; 82   epoll_scheduler& operator=(epoll_scheduler const&) = delete;
83   83  
84   /// Shut down the scheduler, draining pending operations. 84   /// Shut down the scheduler, draining pending operations.
85   void shutdown() override; 85   void shutdown() override;
86   86  
87   /// Apply runtime configuration, resizing the event buffer. 87   /// Apply runtime configuration, resizing the event buffer.
88   void configure_reactor( 88   void configure_reactor(
89   unsigned max_events, 89   unsigned max_events,
90   unsigned budget_init, 90   unsigned budget_init,
91   unsigned budget_max, 91   unsigned budget_max,
92   unsigned unassisted) override; 92   unsigned unassisted) override;
93   93  
94   /** Return the epoll file descriptor. 94   /** Return the epoll file descriptor.
95   95  
96   Used by socket services to register file descriptors 96   Used by socket services to register file descriptors
97   for I/O event notification. 97   for I/O event notification.
98   98  
99   @return The epoll file descriptor. 99   @return The epoll file descriptor.
100   */ 100   */
101   int epoll_fd() const noexcept 101   int epoll_fd() const noexcept
102   { 102   {
103   return epoll_fd_; 103   return epoll_fd_;
104   } 104   }
105   105  
106   /** Register a descriptor for persistent monitoring. 106   /** Register a descriptor for persistent monitoring.
107   107  
108   The fd is registered once and stays registered until explicitly 108   The fd is registered once and stays registered until explicitly
109   deregistered. Events are dispatched via reactor_descriptor_state which 109   deregistered. Events are dispatched via reactor_descriptor_state which
110   tracks pending read/write/connect operations. 110   tracks pending read/write/connect operations.
111   111  
112   @param fd The file descriptor to register. 112   @param fd The file descriptor to register.
113   @param desc Pointer to descriptor data (stored in epoll_event.data.ptr). 113   @param desc Pointer to descriptor data (stored in epoll_event.data.ptr).
114   114  
115   @return The error if registration fails, otherwise a default 115   @return The error if registration fails, otherwise a default
116   constructed error code. 116   constructed error code.
117   */ 117   */
118   std::error_code 118   std::error_code
119   register_descriptor(int fd, reactor_descriptor_state* desc) const; 119   register_descriptor(int fd, reactor_descriptor_state* desc) const;
120   120  
121   /// No-op: write readiness is watched from registration on. 121   /// No-op: write readiness is watched from registration on.
122   std::error_code 122   std::error_code
HITCBC 123   5499 ensure_write_registered(int, reactor_descriptor_state*) const noexcept 123   5469 ensure_write_registered(int, reactor_descriptor_state*) const noexcept
124   { 124   {
HITCBC 125   5499 return {}; 125   5469 return {};
126   } 126   }
127   127  
128   /** Deregister a persistently registered descriptor. 128   /** Deregister a persistently registered descriptor.
129   129  
130   @param fd The file descriptor to deregister. 130   @param fd The file descriptor to deregister.
131   */ 131   */
132   void deregister_descriptor(int fd) const; 132   void deregister_descriptor(int fd) const;
133   133  
134   /// Watch the read end of the POSIX signal self-pipe (see scheduler.hpp). 134   /// Watch the read end of the POSIX signal self-pipe (see scheduler.hpp).
HITCBC 135   287 [[nodiscard]] std::error_code register_signal_reader(int read_fd) override 135   287 [[nodiscard]] std::error_code register_signal_reader(int read_fd) override
136   { 136   {
HITCBC 137   287 return register_descriptor(read_fd, signal_pipe_reader_.arm()); 137   287 return register_descriptor(read_fd, signal_pipe_reader_.arm());
138   } 138   }
139   139  
140   private: 140   private:
141   void run_task(lock_type& lock, context_type& ctx, long timeout_us) override; 141   void run_task(lock_type& lock, context_type& ctx, long timeout_us) override;
142   void interrupt_reactor() const override; 142   void interrupt_reactor() const override;
143   void update_timerfd() const; 143   void update_timerfd() const;
144   static int timer_timeout_ms(timer_service& svc, int timeout_ms) noexcept; 144   static int timer_timeout_ms(timer_service& svc, int timeout_ms) noexcept;
145   void on_timer_service_attached(timer_service& svc) override; 145   void on_timer_service_attached(timer_service& svc) override;
146   146  
147   int epoll_fd_; 147   int epoll_fd_;
148   int event_fd_; 148   int event_fd_;
149   int timer_fd_; 149   int timer_fd_;
150   150  
151   // Watches the global signal self-pipe's read end (armed lazily by 151   // Watches the global signal self-pipe's read end (armed lazily by
152   // register_signal_reader on the first signal registration). 152   // register_signal_reader on the first signal registration).
153   reactor_signal_pipe_reader signal_pipe_reader_; 153   reactor_signal_pipe_reader signal_pipe_reader_;
154   154  
155   // Edge-triggered eventfd state 155   // Edge-triggered eventfd state
156   mutable std::atomic<bool> eventfd_armed_{false}; 156   mutable std::atomic<bool> eventfd_armed_{false};
157   157  
158   // Set when the earliest timer changes; flushed before epoll_wait 158   // Set when the earliest timer changes; flushed before epoll_wait
159   mutable std::atomic<bool> timerfd_stale_{false}; 159   mutable std::atomic<bool> timerfd_stale_{false};
160   160  
161   // Event buffer sized from max_events_per_poll_ (set at construction, 161   // Event buffer sized from max_events_per_poll_ (set at construction,
162   // resized by configure_reactor via io_context_options). 162   // resized by configure_reactor via io_context_options).
163   std::vector<epoll_event> event_buffer_; 163   std::vector<epoll_event> event_buffer_;
164   }; 164   };
165   165  
HITCBC 166   2527 inline epoll_scheduler::epoll_scheduler(capy::execution_context&, int) 166   2527 inline epoll_scheduler::epoll_scheduler(capy::execution_context&, int)
HITCBC 167   2527 : epoll_fd_(-1) 167   2527 : epoll_fd_(-1)
HITCBC 168   2527 , event_fd_(-1) 168   2527 , event_fd_(-1)
HITCBC 169   2527 , timer_fd_(-1) 169   2527 , timer_fd_(-1)
HITCBC 170   5054 , event_buffer_(max_events_per_poll_) 170   5054 , event_buffer_(max_events_per_poll_)
171   { 171   {
HITCBC 172   2527 epoll_fd_ = ::epoll_create1(EPOLL_CLOEXEC); 172   2527 epoll_fd_ = ::epoll_create1(EPOLL_CLOEXEC);
HITCBC 173   2527 if (epoll_fd_ < 0) 173   2527 if (epoll_fd_ < 0)
HITCBC 174   3 detail::throw_system_error(make_err(errno), "epoll_create1"); 174   3 detail::throw_system_error(make_err(errno), "epoll_create1");
175   175  
HITCBC 176   2524 event_fd_ = ::eventfd(0, EFD_NONBLOCK | EFD_CLOEXEC); 176   2524 event_fd_ = ::eventfd(0, EFD_NONBLOCK | EFD_CLOEXEC);
HITCBC 177   2524 if (event_fd_ < 0) 177   2524 if (event_fd_ < 0)
178   { 178   {
HITCBC 179   3 int errn = errno; 179   3 int errn = errno;
HITCBC 180   3 ::close(epoll_fd_); 180   3 ::close(epoll_fd_);
HITCBC 181   3 detail::throw_system_error(make_err(errn), "eventfd"); 181   3 detail::throw_system_error(make_err(errn), "eventfd");
182   } 182   }
183   183  
HITCBC 184   2521 epoll_event ev{}; 184   2521 epoll_event ev{};
HITCBC 185   2521 ev.events = EPOLLIN | EPOLLET; 185   2521 ev.events = EPOLLIN | EPOLLET;
HITCBC 186   2521 ev.data.ptr = nullptr; 186   2521 ev.data.ptr = nullptr;
HITCBC 187   2521 if (::epoll_ctl(epoll_fd_, EPOLL_CTL_ADD, event_fd_, &ev) < 0) 187   2521 if (::epoll_ctl(epoll_fd_, EPOLL_CTL_ADD, event_fd_, &ev) < 0)
188   { 188   {
HITCBC 189   3 int errn = errno; 189   3 int errn = errno;
HITCBC 190   3 ::close(event_fd_); 190   3 ::close(event_fd_);
HITCBC 191   3 ::close(epoll_fd_); 191   3 ::close(epoll_fd_);
HITCBC 192   3 detail::throw_system_error(make_err(errn), "epoll_ctl"); 192   3 detail::throw_system_error(make_err(errn), "epoll_ctl");
193   } 193   }
194   194  
HITCBC 195   2518 completed_ops_.push(&task_op_); 195   2518 completed_ops_.push(&task_op_);
HITCBC 196   2545 } 196   2545 }
197   197  
198   inline void 198   inline void
HITCBC 199   532 epoll_scheduler::on_timer_service_attached(timer_service& svc) 199   536 epoll_scheduler::on_timer_service_attached(timer_service& svc)
200   { 200   {
201   // Without a timerfd the epoll_wait timeout drives the timers: the 201   // Without a timerfd the epoll_wait timeout drives the timers: the
202   // first timer is often made where a throw can only terminate. 202   // first timer is often made where a throw can only terminate.
HITCBC 203   532 int fd = ::timerfd_create(CLOCK_MONOTONIC, TFD_NONBLOCK | TFD_CLOEXEC); 203   536 int fd = ::timerfd_create(CLOCK_MONOTONIC, TFD_NONBLOCK | TFD_CLOEXEC);
HITCBC 204   532 if (fd >= 0) 204   536 if (fd >= 0)
205   { 205   {
206   // Set before registration, which orders it for every thread 206   // Set before registration, which orders it for every thread
207   // the timerfd's events can wake. 207   // the timerfd's events can wake.
HITCBC 208   531 timer_fd_ = fd; 208   535 timer_fd_ = fd;
HITCBC 209   531 epoll_event timer_ev{}; 209   535 epoll_event timer_ev{};
HITCBC 210   531 timer_ev.events = EPOLLIN | EPOLLERR; 210   535 timer_ev.events = EPOLLIN | EPOLLERR;
HITCBC 211   531 timer_ev.data.ptr = &timer_fd_; 211   535 timer_ev.data.ptr = &timer_fd_;
HITCBC 212   531 if (::epoll_ctl(epoll_fd_, EPOLL_CTL_ADD, fd, &timer_ev) < 0) 212   535 if (::epoll_ctl(epoll_fd_, EPOLL_CTL_ADD, fd, &timer_ev) < 0)
213   { 213   {
HITCBC 214   1 ::close(fd); 214   1 ::close(fd);
HITCBC 215   1 timer_fd_ = -1; 215   1 timer_fd_ = -1;
216   } 216   }
217   } 217   }
HITCBC 218   532 svc.set_on_earliest_changed(timer_service::callback(this, [](void* p) { 218   536 svc.set_on_earliest_changed(timer_service::callback(this, [](void* p) {
HITCBC 219   5021 auto* self = static_cast<epoll_scheduler*>(p); 219   4893 auto* self = static_cast<epoll_scheduler*>(p);
HITCBC 220   5021 self->timerfd_stale_.store(true, std::memory_order_release); 220   4893 self->timerfd_stale_.store(true, std::memory_order_release);
HITCBC 221   5021 self->interrupt_reactor(); 221   4893 self->interrupt_reactor();
HITCBC 222   5021 })); 222   4893 }));
HITCBC 223   532 } 223   536 }
224   224  
HITCBC 225   5036 inline epoll_scheduler::~epoll_scheduler() 225   5036 inline epoll_scheduler::~epoll_scheduler()
226   { 226   {
HITCBC 227   2518 if (timer_fd_ >= 0) 227   2518 if (timer_fd_ >= 0)
HITCBC 228   530 ::close(timer_fd_); 228   534 ::close(timer_fd_);
HITCBC 229   2518 if (event_fd_ >= 0) 229   2518 if (event_fd_ >= 0)
HITCBC 230   2518 ::close(event_fd_); 230   2518 ::close(event_fd_);
HITCBC 231   2518 if (epoll_fd_ >= 0) 231   2518 if (epoll_fd_ >= 0)
HITCBC 232   2518 ::close(epoll_fd_); 232   2518 ::close(epoll_fd_);
HITCBC 233   5036 } 233   5036 }
234   234  
235   inline void 235   inline void
HITCBC 236   2518 epoll_scheduler::shutdown() 236   2518 epoll_scheduler::shutdown()
237   { 237   {
HITCBC 238   2518 shutdown_drain(); 238   2518 shutdown_drain();
239   239  
HITCBC 240   2518 if (event_fd_ >= 0) 240   2518 if (event_fd_ >= 0)
HITCBC 241   2518 interrupt_reactor(); 241   2518 interrupt_reactor();
HITCBC 242   2518 } 242   2518 }
243   243  
244   inline void 244   inline void
HITCBC 245   30 epoll_scheduler::configure_reactor( 245   30 epoll_scheduler::configure_reactor(
246   unsigned max_events, 246   unsigned max_events,
247   unsigned budget_init, 247   unsigned budget_init,
248   unsigned budget_max, 248   unsigned budget_max,
249   unsigned unassisted) 249   unsigned unassisted)
250   { 250   {
HITCBC 251   30 reactor_scheduler::configure_reactor( 251   30 reactor_scheduler::configure_reactor(
252   max_events, budget_init, budget_max, unassisted); 252   max_events, budget_init, budget_max, unassisted);
HITCBC 253   30 event_buffer_.resize(max_events_per_poll_); 253   30 event_buffer_.resize(max_events_per_poll_);
HITCBC 254   30 } 254   30 }
255   255  
256   inline std::error_code 256   inline std::error_code
HITCBC 257   13559 epoll_scheduler::register_descriptor( 257   13502 epoll_scheduler::register_descriptor(
258   int fd, reactor_descriptor_state* desc) const 258   int fd, reactor_descriptor_state* desc) const
259   { 259   {
HITCBC 260   13559 epoll_event ev{}; 260   13502 epoll_event ev{};
HITCBC 261   13559 ev.events = EPOLLIN | EPOLLOUT | EPOLLET | EPOLLERR | EPOLLHUP; 261   13502 ev.events = EPOLLIN | EPOLLOUT | EPOLLET | EPOLLERR | EPOLLHUP;
HITCBC 262   13559 ev.data.ptr = desc; 262   13502 ev.data.ptr = desc;
263   263  
264   // Initialized before the add: another run thread's reactor can 264   // Initialized before the add: another run thread's reactor can
265   // dispatch an event for the descriptor as soon as it is added. 265   // dispatch an event for the descriptor as soon as it is added.
266   // A completion queued by an earlier registration of the same state 266   // A completion queued by an earlier registration of the same state
267   // may still be running, so the fields it reads change under its 267   // may still be running, so the fields it reads change under its
268   // mutex, and the mutex mode is only written when it differs. 268   // mutex, and the mutex mode is only written when it differs.
HITCBC 269   13559 if (desc->mutex.enabled() != reactor_io_locking_) 269   13502 if (desc->mutex.enabled() != reactor_io_locking_)
HITCBC 270   3 desc->mutex.set_enabled(reactor_io_locking_); 270   3 desc->mutex.set_enabled(reactor_io_locking_);
271   { 271   {
HITCBC 272   13559 conditionally_enabled_mutex::scoped_lock lock(desc->mutex); 272   13502 conditionally_enabled_mutex::scoped_lock lock(desc->mutex);
HITCBC 273   13559 desc->registered_events = ev.events; 273   13502 desc->registered_events = ev.events;
HITCBC 274   13559 desc->unpollable = false; 274   13502 desc->unpollable = false;
HITCBC 275   13559 desc->fd = fd; 275   13502 desc->fd = fd;
HITCBC 276   13559 desc->scheduler_ = this; 276   13502 desc->scheduler_ = this;
HITCBC 277   13559 desc->ready_events_.store(0, std::memory_order_release); 277   13502 desc->ready_events_.store(0, std::memory_order_release);
278   // object_ref_ is not touched: a queued invocation from an 278   // object_ref_ is not touched: a queued invocation from an
279   // earlier registration still owns the reference in it. 279   // earlier registration still owns the reference in it.
HITCBC 280   13559 desc->read_ready = false; 280   13502 desc->read_ready = false;
HITCBC 281   13559 desc->write_ready = false; 281   13502 desc->write_ready = false;
HITCBC 282   13559 } 282   13502 }
283   283  
HITCBC 284   13559 if (::epoll_ctl(epoll_fd_, EPOLL_CTL_ADD, fd, &ev) < 0) 284   13502 if (::epoll_ctl(epoll_fd_, EPOLL_CTL_ADD, fd, &ev) < 0)
285   { 285   {
HITCBC 286   12 int const err = errno; 286   12 int const err = errno;
HITCBC 287   12 conditionally_enabled_mutex::scoped_lock lock(desc->mutex); 287   12 conditionally_enabled_mutex::scoped_lock lock(desc->mutex);
HITCBC 288   12 desc->registered_events = 0; 288   12 desc->registered_events = 0;
289   // EPERM: a file type epoll cannot watch. Its I/O does not 289   // EPERM: a file type epoll cannot watch. Its I/O does not
290   // block, so adopt it unwatched, as asio does. 290   // block, so adopt it unwatched, as asio does.
HITCBC 291   12 if (err != EPERM) 291   12 if (err != EPERM)
HITCBC 292   7 return make_err(err); 292   7 return make_err(err);
HITCBC 293   5 desc->unpollable = true; 293   5 desc->unpollable = true;
HITCBC 294   12 } 294   12 }
HITCBC 295   13552 return {}; 295   13495 return {};
296   } 296   }
297   297  
298   inline void 298   inline void
HITCBC 299   13261 epoll_scheduler::deregister_descriptor(int fd) const 299   13204 epoll_scheduler::deregister_descriptor(int fd) const
300   { 300   {
HITCBC 301   13261 ::epoll_ctl(epoll_fd_, EPOLL_CTL_DEL, fd, nullptr); 301   13204 ::epoll_ctl(epoll_fd_, EPOLL_CTL_DEL, fd, nullptr);
HITCBC 302   13261 } 302   13204 }
303   303  
304   inline void 304   inline void
HITCBC 305   13235 epoll_scheduler::interrupt_reactor() const 305   13188 epoll_scheduler::interrupt_reactor() const
306   { 306   {
HITCBC 307   13235 bool expected = false; 307   13188 bool expected = false;
HITCBC 308   13235 if (eventfd_armed_.compare_exchange_strong( 308   13188 if (eventfd_armed_.compare_exchange_strong(
309   expected, true, std::memory_order_release, 309   expected, true, std::memory_order_release,
310   std::memory_order_relaxed)) 310   std::memory_order_relaxed))
311   { 311   {
HITCBC 312   10800 std::uint64_t val = 1; 312   10780 std::uint64_t val = 1;
HITCBC 313   10800 if (::write(event_fd_, &val, sizeof(val)) < 0) 313   10780 if (::write(event_fd_, &val, sizeof(val)) < 0)
314   { 314   {
315   // The flag is what coalesces later interrupts into a byte 315   // The flag is what coalesces later interrupts into a byte
316   // already in the eventfd; a write that failed put no byte 316   // already in the eventfd; a write that failed put no byte
317   // there, so leaving it armed would swallow every interrupt 317   // there, so leaving it armed would swallow every interrupt
318   // that follows. Disarming keeps the cost to the interrupts 318   // that follows. Disarming keeps the cost to the interrupts
319   // already in flight -- the next one arms and writes again, 319   // already in flight -- the next one arms and writes again,
320   // instead of every one after this coalescing into a byte 320   // instead of every one after this coalescing into a byte
321   // that does not exist. 321   // that does not exist.
HITCBC 322   2 eventfd_armed_.store(false, std::memory_order_release); 322   2 eventfd_armed_.store(false, std::memory_order_release);
323   } 323   }
324   } 324   }
HITCBC 325   13235 } 325   13188 }
326   326  
327   inline void 327   inline void
HITCBC 328   11882 epoll_scheduler::update_timerfd() const 328   11710 epoll_scheduler::update_timerfd() const
329   { 329   {
HITCBC 330   11882 auto* tsvc = timers(); 330   11710 auto* tsvc = timers();
HITCBC 331   11882 if (!tsvc || timer_fd_ < 0) 331   11710 if (!tsvc || timer_fd_ < 0)
HITCBC 332   6 return; 332   6 return;
HITCBC 333   11876 auto nearest = tsvc->nearest_expiry(); 333   11704 auto nearest = tsvc->nearest_expiry();
334   334  
HITCBC 335   11876 itimerspec ts{}; 335   11704 itimerspec ts{};
HITCBC 336   11876 int flags = 0; 336   11704 int flags = 0;
337   337  
HITCBC 338   11876 if (nearest == timer_service::time_point::max()) 338   11704 if (nearest == timer_service::time_point::max())
339   { 339   {
340   // No timers — disarm by setting to 0 (relative) 340   // No timers — disarm by setting to 0 (relative)
341   } 341   }
342   else 342   else
343   { 343   {
HITCBC 344   10433 auto now = std::chrono::steady_clock::now(); 344   10355 auto now = std::chrono::steady_clock::now();
HITCBC 345   10433 if (nearest <= now) 345   10355 if (nearest <= now)
346   { 346   {
347   // Use 1ns instead of 0 — zero disarms the timerfd 347   // Use 1ns instead of 0 — zero disarms the timerfd
HITCBC 348   987 ts.it_value.tv_nsec = 1; 348   1100 ts.it_value.tv_nsec = 1;
349   } 349   }
350   else 350   else
351   { 351   {
HITCBC 352   9446 auto nsec = std::chrono::duration_cast<std::chrono::nanoseconds>( 352   9255 auto nsec = std::chrono::duration_cast<std::chrono::nanoseconds>(
HITCBC 353   9446 nearest - now) 353   9255 nearest - now)
HITCBC 354   9446 .count(); 354   9255 .count();
HITCBC 355   9446 ts.it_value.tv_sec = nsec / 1000000000; 355   9255 ts.it_value.tv_sec = nsec / 1000000000;
HITCBC 356   9446 ts.it_value.tv_nsec = nsec % 1000000000; 356   9255 ts.it_value.tv_nsec = nsec % 1000000000;
HITCBC 357   9446 if (ts.it_value.tv_sec == 0 && ts.it_value.tv_nsec == 0) 357   9255 if (ts.it_value.tv_sec == 0 && ts.it_value.tv_nsec == 0)
MISUBC 358   ✗ ts.it_value.tv_nsec = 1; 358   ✗ ts.it_value.tv_nsec = 1;
359   } 359   }
360   } 360   }
361   361  
HITCBC 362   11876 if (::timerfd_settime(timer_fd_, flags, &ts, nullptr) < 0) 362   11704 if (::timerfd_settime(timer_fd_, flags, &ts, nullptr) < 0)
HITCBC 363   1 detail::throw_system_error(make_err(errno), "timerfd_settime"); 363   1 detail::throw_system_error(make_err(errno), "timerfd_settime");
364   } 364   }
365   365  
366   // Rounded up, so the wait never wakes before the timer is due. 366   // Rounded up, so the wait never wakes before the timer is due.
367   inline int 367   inline int
HITCBC 368   4 epoll_scheduler::timer_timeout_ms(timer_service& svc, int timeout_ms) noexcept 368   4 epoll_scheduler::timer_timeout_ms(timer_service& svc, int timeout_ms) noexcept
369   { 369   {
HITCBC 370   4 auto nearest = svc.nearest_expiry(); 370   4 auto nearest = svc.nearest_expiry();
HITCBC 371   4 if (nearest == timer_service::time_point::max()) 371   4 if (nearest == timer_service::time_point::max())
MISUBC 372   ✗ return timeout_ms; 372   ✗ return timeout_ms;
HITCBC 373   4 auto now = std::chrono::steady_clock::now(); 373   4 auto now = std::chrono::steady_clock::now();
HITCBC 374   4 if (nearest <= now) 374   4 if (nearest <= now)
MISUBC 375   ✗ return 0; 375   ✗ return 0;
376   auto ms = 376   auto ms =
HITCBC 377   4 std::chrono::ceil<std::chrono::milliseconds>(nearest - now).count(); 377   4 std::chrono::ceil<std::chrono::milliseconds>(nearest - now).count();
HITCBC 378   8 ms = (std::min)(ms, 378   8 ms = (std::min)(ms,
HITCBC 379   4 static_cast<decltype(ms)>( 379   4 static_cast<decltype(ms)>(
HITCBC 380   4 (std::numeric_limits<int>::max)())); 380   4 (std::numeric_limits<int>::max)()));
HITCBC 381   4 if (timeout_ms >= 0 && timeout_ms < ms) 381   4 if (timeout_ms >= 0 && timeout_ms < ms)
MISUBC 382   ✗ return timeout_ms; 382   ✗ return timeout_ms;
HITCBC 383   4 return static_cast<int>(ms); 383   4 return static_cast<int>(ms);
384   } 384   }
385   385  
386   inline void 386   inline void
HITCBC 387   149616 epoll_scheduler::run_task(lock_type& lock, context_type& ctx, long timeout_us) 387   147102 epoll_scheduler::run_task(lock_type& lock, context_type& ctx, long timeout_us)
388   { 388   {
389   int timeout_ms; 389   int timeout_ms;
HITCBC 390   149616 if (task_interrupted_) 390   147102 if (task_interrupted_)
HITCBC 391   129172 timeout_ms = 0; 391   126464 timeout_ms = 0;
HITCBC 392   20444 else if (timeout_us < 0) 392   20638 else if (timeout_us < 0)
HITCBC 393   19893 timeout_ms = -1; 393   20110 timeout_ms = -1;
394   else 394   else
HITCBC 395   551 timeout_ms = static_cast<int>((timeout_us + 999) / 1000); 395   528 timeout_ms = static_cast<int>((timeout_us + 999) / 1000);
396   396  
HITCBC 397   149616 if (lock.owns_lock()) 397   147102 if (lock.owns_lock())
HITCBC 398   20448 lock.unlock(); 398   20642 lock.unlock();
399   399  
HITCBC 400   149616 task_cleanup on_exit{this, &lock, ctx}; 400   147102 task_cleanup on_exit{this, &lock, ctx};
401   401  
402   // Flush deferred timerfd programming before blocking 402   // Flush deferred timerfd programming before blocking
HITCBC 403   149616 if (timerfd_stale_.exchange(false, std::memory_order_acquire)) 403   147102 if (timerfd_stale_.exchange(false, std::memory_order_acquire))
HITCBC 404   4054 update_timerfd(); 404   3993 update_timerfd();
405   405  
406   // timer_fd_ is read only once the service is seen, after which it 406   // timer_fd_ is read only once the service is seen, after which it
407   // no longer changes. 407   // no longer changes.
HITCBC 408   149615 auto* tsvc = timers(); 408   147101 auto* tsvc = timers();
HITCBC 409   149615 bool const poll_timers = tsvc && timer_fd_ < 0; 409   147101 bool const poll_timers = tsvc && timer_fd_ < 0;
HITCBC 410   149615 if (poll_timers && timeout_ms != 0) 410   147101 if (poll_timers && timeout_ms != 0)
HITCBC 411   4 timeout_ms = timer_timeout_ms(*tsvc, timeout_ms); 411   4 timeout_ms = timer_timeout_ms(*tsvc, timeout_ms);
412   412  
HITCBC 413   149615 int nfds = ::epoll_wait( 413   147101 int nfds = ::epoll_wait(
HITCBC 414   149615 epoll_fd_, event_buffer_.data(), static_cast<int>(event_buffer_.size()), 414   147101 epoll_fd_, event_buffer_.data(), static_cast<int>(event_buffer_.size()),
415   timeout_ms); 415   timeout_ms);
416   416  
HITCBC 417   149615 if (nfds < 0 && errno != EINTR) 417   147101 if (nfds < 0 && errno != EINTR)
HITCBC 418   1 detail::throw_system_error(make_err(errno), "epoll_wait"); 418   1 detail::throw_system_error(make_err(errno), "epoll_wait");
419   419  
HITCBC 420   149614 bool check_timers = poll_timers; 420   147100 bool check_timers = poll_timers;
HITCBC 421   149614 ready_queue local_ops; 421   147100 ready_queue local_ops;
422   422  
HITCBC 423   213325 for (int i = 0; i < nfds; ++i) 423   206575 for (int i = 0; i < nfds; ++i)
424   { 424   {
HITCBC 425   63711 if (event_buffer_[i].data.ptr == nullptr) 425   59475 if (event_buffer_[i].data.ptr == nullptr)
426   { 426   {
427   std::uint64_t val; 427   std::uint64_t val;
428   // NOLINTNEXTLINE(clang-analyzer-unix.BlockInCriticalSection) 428   // NOLINTNEXTLINE(clang-analyzer-unix.BlockInCriticalSection)
HITCBC 429   8280 [[maybe_unused]] auto r = ::read(event_fd_, &val, sizeof(val)); 429   8260 [[maybe_unused]] auto r = ::read(event_fd_, &val, sizeof(val));
HITCBC 430   8280 eventfd_armed_.store(false, std::memory_order_relaxed); 430   8260 eventfd_armed_.store(false, std::memory_order_relaxed);
HITCBC 431   8280 continue; 431   8260 continue;
HITCBC 432   8280 } 432   8260 }
433   433  
HITCBC 434   55431 if (event_buffer_[i].data.ptr == &timer_fd_) 434   51215 if (event_buffer_[i].data.ptr == &timer_fd_)
435   { 435   {
436   std::uint64_t expirations; 436   std::uint64_t expirations;
437   // NOLINTNEXTLINE(clang-analyzer-unix.BlockInCriticalSection) 437   // NOLINTNEXTLINE(clang-analyzer-unix.BlockInCriticalSection)
438   [[maybe_unused]] auto r = 438   [[maybe_unused]] auto r =
HITCBC 439   7824 ::read(timer_fd_, &expirations, sizeof(expirations)); 439   7713 ::read(timer_fd_, &expirations, sizeof(expirations));
HITCBC 440   7824 check_timers = true; 440   7713 check_timers = true;
HITCBC 441   7824 continue; 441   7713 continue;
HITCBC 442   7824 } 442   7713 }
443   443  
444   auto* desc = 444   auto* desc =
HITCBC 445   47607 static_cast<reactor_descriptor_state*>(event_buffer_[i].data.ptr); 445   43502 static_cast<reactor_descriptor_state*>(event_buffer_[i].data.ptr);
446   446  
447   // EPOLLHUP maps to no reactor_event_* bit, so a HUP the widening 447   // EPOLLHUP maps to no reactor_event_* bit, so a HUP the widening
448   // below did not translate would take no branch in 448   // below did not translate would take no branch in
449   // invoke_deferred_io() and, the registration being 449   // invoke_deferred_io() and, the registration being
450   // edge-triggered, never get another chance. 450   // edge-triggered, never get another chance.
451   // 451   //
452   // HUP without OUT means a non-socket: a pipe or tty whose peer 452   // HUP without OUT means a non-socket: a pipe or tty whose peer
453   // closed, with or without data still unread (HUP alone, or 453   // closed, with or without data still unread (HUP alone, or
454   // IN|HUP). Treat it as readable, writable and faulted, the way 454   // IN|HUP). Treat it as readable, writable and faulted, the way
455   // poll() and io_uring report it, so a parked error wait 455   // poll() and io_uring report it, so a parked error wait
456   // completes too. Sockets carry at least OUT with HUP (a fresh 456   // completes too. Sockets carry at least OUT with HUP (a fresh
457   // unconnected stream socket reports HUP|OUT), so they take the 457   // unconnected stream socket reports HUP|OUT), so they take the
458   // plain widening, where forcing IN costs at most one spurious 458   // plain widening, where forcing IN costs at most one spurious
459   // EAGAIN. 459   // EAGAIN.
HITCBC 460   47607 std::uint32_t ev = event_buffer_[i].events; 460   43502 std::uint32_t ev = event_buffer_[i].events;
HITCBC 461   47607 if ((ev & (EPOLLHUP | EPOLLOUT)) == EPOLLHUP) 461   43502 if ((ev & (EPOLLHUP | EPOLLOUT)) == EPOLLHUP)
HITCBC 462   8 ev |= EPOLLIN | EPOLLOUT | EPOLLERR; 462   8 ev |= EPOLLIN | EPOLLOUT | EPOLLERR;
HITCBC 463   47599 else if (ev & EPOLLHUP) 463   43494 else if (ev & EPOLLHUP)
HITCBC 464   2987 ev |= EPOLLIN | EPOLLOUT; 464   2923 ev |= EPOLLIN | EPOLLOUT;
HITCBC 465   47607 desc->add_ready_events(ev); 465   43502 desc->add_ready_events(ev);
466   466  
HITCBC 467   47607 bool expected = false; 467   43502 bool expected = false;
HITCBC 468   47607 if (desc->is_enqueued_.compare_exchange_strong( 468   43502 if (desc->is_enqueued_.compare_exchange_strong(
469   expected, true, std::memory_order_release, 469   expected, true, std::memory_order_release,
470   std::memory_order_relaxed)) 470   std::memory_order_relaxed))
471   { 471   {
HITCBC 472   47603 local_ops.push(desc); 472   43490 local_ops.push(desc);
473   } 473   }
474   } 474   }
475   475  
HITCBC 476   149614 if (check_timers) 476   147100 if (check_timers)
477   { 477   {
HITCBC 478   7828 if (auto* svc = timers()) 478   7717 if (auto* svc = timers())
HITCBC 479   7828 svc->process_expired(); 479   7717 svc->process_expired();
HITCBC 480   7828 update_timerfd(); 480   7717 update_timerfd();
481   } 481   }
482   482  
HITCBC 483   149614 lock.lock(); 483   147100 lock.lock();
484   484  
HITCBC 485   149614 completed_ops_.splice(local_ops); 485   147100 completed_ops_.splice(local_ops);
HITCBC 486   149616 } 486   147102 }
487   487  
488   } // namespace boost::corosio::detail 488   } // namespace boost::corosio::detail
489   489  
490   #endif // BOOST_COROSIO_HAS_EPOLL 490   #endif // BOOST_COROSIO_HAS_EPOLL
491   491  
492   #endif // BOOST_COROSIO_NATIVE_DETAIL_EPOLL_EPOLL_SCHEDULER_HPP 492   #endif // BOOST_COROSIO_NATIVE_DETAIL_EPOLL_EPOLL_SCHEDULER_HPP