/*
- event.c -- I/O, timeout and signal event handling
+ event.c -- I/O, timeout, and event handling
Copyright (C) 2012-2022 Guus Sliepen <guus@tinc-vpn.org>
This program is free software; you can redistribute it and/or modify
#include "system.h"
-#ifdef HAVE_WINDOWS
-# include <assert.h>
-#else
-# if defined(HAVE_SYS_EPOLL_H)
-# include <sys/epoll.h>
-# define HAVE_EPOLL 1
-# elif defined(HAVE_SYS_EVENT_H)
-# include <sys/event.h>
-# define HAVE_KQUEUE 1
-# else
-# define HAVE_SELECT 1
-# endif
-#endif
-
#include "event.h"
-#include "utils.h"
-#include "net.h"
struct timeval now;
-static bool running;
-
-#if defined(HAVE_EPOLL) || defined(HAVE_KQUEUE)
-static int event_fd = 0;
-#elif defined(HAVE_SELECT)
-static fd_set readfds;
-static fd_set writefds;
-#elif defined(HAVE_WINDOWS)
-static const long READ_EVENTS = FD_READ | FD_ACCEPT | FD_CLOSE;
-static const long WRITE_EVENTS = FD_WRITE | FD_CONNECT;
-static DWORD event_count = 0;
-#endif
-
-static inline void event_init(void) {
-#if defined(HAVE_EPOLL) || defined(HAVE_KQUEUE)
-
- if(!event_fd) {
-#if defined(HAVE_EPOLL)
- /* NOTE: 1024 limit is only used on ancient (pre 2.6.27) kernels.
- Decent kernels will ignore this value making it unlimited.
- epoll_create1 might be better, but these kernels would not be supported
- in that case. */
- event_fd = epoll_create(1024);
-#else
- event_fd = kqueue();
-#endif
-
- if(event_fd == -1) {
- logger(DEBUG_ALWAYS, LOG_EMERG, "Could not initialize events: %s", strerror(errno));
- abort();
- }
- }
-
-#endif
-}
-
-static void event_deinit(void) {
-#if defined(HAVE_EPOLL) || defined(HAVE_KQUEUE)
-
- if(event_fd > 0) {
- close(event_fd);
- event_fd = 0;
- }
-
-#endif
-}
static int io_compare(const io_t *a, const io_t *b) {
#ifndef HAVE_WINDOWS
return 0;
}
-static splay_tree_t io_tree = {.compare = (splay_compare_t)io_compare};
-static splay_tree_t timeout_tree = {.compare = (splay_compare_t)timeout_compare};
-
-void io_add(io_t *io, io_cb_t cb, void *data, int fd, int flags) {
- if(io->cb) {
- return;
- }
-
- io->fd = fd;
-#ifdef HAVE_WINDOWS
-
- if(io->fd != -1) {
- io->event = WSACreateEvent();
-
- if(io->event == WSA_INVALID_EVENT) {
- abort();
- }
- }
-
- event_count++;
-#endif
- io->cb = cb;
- io->data = data;
- io->node.data = io;
-
- io_set(io, flags);
-
-#ifdef HAVE_SELECT
-
- if(!splay_insert_node(&io_tree, &io->node)) {
- abort();
- }
-
-#endif
-}
-
-#ifdef HAVE_WINDOWS
-void io_add_event(io_t *io, io_cb_t cb, void *data, WSAEVENT event) {
- io->event = event;
- io_add(io, cb, data, -1, 0);
-}
-#endif
-
-void io_set(io_t *io, int flags) {
- event_init();
-
- if(flags == io->flags) {
- return;
- }
-
- io->flags = flags;
-
- if(io->fd == -1) {
- return;
- }
-
-#ifndef HAVE_WINDOWS
-#ifdef HAVE_EPOLL
- epoll_ctl(event_fd, EPOLL_CTL_DEL, io->fd, NULL);
-
- struct epoll_event ev = {
- .events = 0,
- .data.ptr = io,
- };
-
- if(flags & IO_READ) {
- ev.events |= EPOLLIN;
- }
-
- if(flags & IO_WRITE) {
- ev.events |= EPOLLOUT;
- } else if(ev.events == 0) {
- io_tree.generation++;
- return;
- }
-
- if(epoll_ctl(event_fd, EPOLL_CTL_ADD, io->fd, &ev) < 0) {
- perror("epoll_ctl_add");
- }
-
-#endif
-
-#ifdef HAVE_KQUEUE
- const struct kevent change[] = {
- {
- .ident = io->fd,
- .filter = EVFILT_READ,
- .flags = EV_RECEIPT | (flags & IO_READ ? EV_ADD : EV_DELETE),
- .udata = io,
- },
- {
- .ident = io->fd,
- .filter = EVFILT_WRITE,
- .flags = EV_RECEIPT | (flags & IO_WRITE ? EV_ADD : EV_DELETE),
- .udata = io,
- },
- };
- struct kevent result[2];
-
- if(kevent(event_fd, change, 2, result, 2, NULL) < 0) {
- logger(DEBUG_ALWAYS, LOG_EMERG, "kevent failed: %s", strerror(errno));
- abort();
- }
-
- int rerr = (int)result[0].data;
- int werr = (int)result[1].data;
-
- if((rerr && rerr != ENOENT) || (werr && werr != ENOENT)) {
- logger(DEBUG_ALWAYS, LOG_EMERG, "kevent errors: %s, %s", strerror(rerr), strerror(werr));
- abort();
- }
-
- if(!flags) {
- io_tree.generation++;
- }
-
-#endif
-
-#ifdef HAVE_SELECT
-
- if(flags & IO_READ) {
- FD_SET(io->fd, &readfds);
- } else {
- FD_CLR(io->fd, &readfds);
- }
-
- if(flags & IO_WRITE) {
- FD_SET(io->fd, &writefds);
- } else {
- FD_CLR(io->fd, &writefds);
- }
-
-#endif
-
-#else
- long events = 0;
-
- if(flags & IO_WRITE) {
- events |= WRITE_EVENTS;
- }
-
- if(flags & IO_READ) {
- events |= READ_EVENTS;
- }
-
- if(WSAEventSelect(io->fd, io->event, events) != 0) {
- abort();
- }
-
-#endif
-}
-
-void io_del(io_t *io) {
- if(!io->cb) {
- return;
- }
-
- io_set(io, 0);
-#ifdef HAVE_WINDOWS
-
- if(io->fd != -1 && WSACloseEvent(io->event) == FALSE) {
- abort();
- }
-
- event_count--;
-#endif
-
-#if HAVE_SELECT
- splay_unlink_node(&io_tree, &io->node);
-#endif
- io->cb = NULL;
-}
+splay_tree_t io_tree = {.compare = (splay_compare_t)io_compare};
+splay_tree_t timeout_tree = {.compare = (splay_compare_t)timeout_compare};
void timeout_add(timeout_t *timeout, timeout_cb_t cb, void *data, const struct timeval *tv) {
timeout->cb = cb;
};
}
-#ifndef HAVE_WINDOWS
-
-// From Matz's Ruby
-#ifndef NSIG
-# define NSIG (_SIGMAX + 1) /* For QNX */
-#endif
-
-
-static io_t signalio;
-static int pipefd[2] = {-1, -1};
-static signal_t *signal_handle[NSIG + 1] = {NULL};
-
-static void signal_handler(int signum) {
- unsigned char num = signum;
-
- if(write(pipefd[1], &num, 1) != 1) {
- // Pipe full or broken, nothing we can do about it.
- }
-}
-
-static void signalio_handler(void *data, int flags) {
- (void)data;
- (void)flags;
- unsigned char signum;
-
- if(read(pipefd[0], &signum, 1) != 1) {
- return;
- }
-
- signal_t *sig = signal_handle[signum];
-
- if(sig) {
- sig->cb(sig->data);
- }
-}
-
-static void pipe_init(void) {
- if(!pipe(pipefd)) {
- io_add(&signalio, signalio_handler, NULL, pipefd[0], IO_READ);
- }
-}
-
-void signal_add(signal_t *sig, signal_cb_t cb, void *data, int signum) {
- if(sig->cb) {
- return;
- }
-
- sig->signum = signum;
- sig->cb = cb;
- sig->data = data;
-
- if(pipefd[0] == -1) {
- pipe_init();
- }
-
- signal(signum, signal_handler);
-
- signal_handle[signum] = sig;
-}
-
-void signal_del(signal_t *sig) {
- if(!sig->cb) {
- return;
- }
-
- signal(sig->signum, SIG_DFL);
-
- signal_handle[sig->signum] = NULL;
- sig->cb = NULL;
-}
-#endif
-
-static struct timeval *timeout_execute(struct timeval *diff) {
+struct timeval *timeout_execute(struct timeval *diff) {
gettimeofday(&now, NULL);
struct timeval *tv = NULL;
return tv;
}
-
-bool event_loop(void) {
- event_init();
- running = true;
-
-#ifndef HAVE_WINDOWS
-
-#ifdef HAVE_SELECT
- fd_set readable;
- fd_set writable;
-#endif
-
- while(running) {
- struct timeval diff;
- struct timeval *tv = timeout_execute(&diff);
-
-#ifdef HAVE_SELECT
- memcpy(&readable, &readfds, sizeof(readable));
- memcpy(&writable, &writefds, sizeof(writable));
-#endif
-
-#ifdef HAVE_EPOLL
- struct epoll_event events[MAX_EVENTS_PER_LOOP];
- long timeout = (tv->tv_sec * 1000) + (tv->tv_usec / 1000);
-
- if(timeout > INT_MAX) {
- timeout = INT_MAX;
- }
-
- int n = epoll_wait(event_fd, events, MAX_EVENTS_PER_LOOP, (int)timeout);
-#endif
-
-#ifdef HAVE_KQUEUE
- struct kevent events[MAX_EVENTS_PER_LOOP];
-
- const struct timespec ts = {
- .tv_sec = tv->tv_sec,
- .tv_nsec = tv->tv_usec * 1000,
- };
-
- int n = kevent(event_fd, NULL, 0, events, MAX_EVENTS_PER_LOOP, &ts);
-#endif
-
-#ifdef HAVE_SELECT
- int maxfds = 0;
-
- if(io_tree.tail) {
- io_t *last = io_tree.tail->data;
- maxfds = last->fd + 1;
- }
-
- int n = select(maxfds, &readable, &writable, NULL, tv);
-#endif
-
- if(n < 0) {
- if(sockwouldblock(sockerrno)) {
- continue;
- } else {
- return false;
- }
- }
-
- if(!n) {
- continue;
- }
-
- unsigned int curgen = io_tree.generation;
-
-
-#ifdef HAVE_EPOLL
-
- for(int i = 0; i < n; i++) {
- io_t *io = events[i].data.ptr;
-
- if(events[i].events & EPOLLOUT && io->flags & IO_WRITE) {
- io->cb(io->data, IO_WRITE);
- }
-
- if(curgen != io_tree.generation) {
- break;
- }
-
- if(events[i].events & EPOLLIN && io->flags & IO_READ) {
- io->cb(io->data, IO_READ);
- }
-
- if(curgen != io_tree.generation) {
- break;
- }
- }
-
-#endif
-
-#ifdef HAVE_KQUEUE
-
- for(int i = 0; i < n; i++) {
- const struct kevent *evt = &events[i];
- const io_t *io = evt->udata;
-
- if(evt->filter == EVFILT_WRITE) {
- io->cb(io->data, IO_WRITE);
- } else if(evt->filter == EVFILT_READ) {
- io->cb(io->data, IO_READ);
- } else {
- continue;
- }
-
- if(curgen != io_tree.generation) {
- break;
- }
- }
-
-#endif
-
-#ifdef HAVE_SELECT
-
- for splay_each(io_t, io, &io_tree) {
- if(FD_ISSET(io->fd, &writable)) {
- io->cb(io->data, IO_WRITE);
- } else if(FD_ISSET(io->fd, &readable)) {
- io->cb(io->data, IO_READ);
- } else {
- continue;
- }
-
- /*
- There are scenarios in which the callback will remove another io_t from the tree
- (e.g. closing a double connection). Since splay_each does not support that, we
- need to exit the loop if that happens. That's okay, since any remaining events will
- get picked up by the next select() call.
- */
- if(curgen != io_tree.generation) {
- break;
- }
- }
-
-#endif
- }
-
-#else
- assert(WSA_WAIT_EVENT_0 == 0);
-
- while(running) {
- struct timeval diff;
- struct timeval *tv = timeout_execute(&diff);
- DWORD timeout_ms = tv ? (DWORD)(tv->tv_sec * 1000 + tv->tv_usec / 1000 + 1) : WSA_INFINITE;
-
- if(!event_count) {
- Sleep(timeout_ms);
- continue;
- }
-
- /*
- For some reason, Microsoft decided to make the FD_WRITE event edge-triggered instead of level-triggered,
- which is the opposite of what select() does. In practice, that means that if a FD_WRITE event triggers,
- it will never trigger again until a send() returns EWOULDBLOCK. Since the semantics of this event loop
- is that write events are level-triggered (i.e. they continue firing until the socket is full), we need
- to emulate these semantics by making sure we fire each IO_WRITE that is still writeable.
-
- Note that technically FD_CLOSE has the same problem, but it's okay because user code does not rely on
- this event being fired again if ignored.
- */
- unsigned int curgen = io_tree.generation;
-
- for splay_each(io_t, io, &io_tree) {
- if(io->flags & IO_WRITE && send(io->fd, NULL, 0, 0) == 0) {
- io->cb(io->data, IO_WRITE);
-
- if(curgen != io_tree.generation) {
- break;
- }
- }
- }
-
- if(event_count > WSA_MAXIMUM_WAIT_EVENTS) {
- WSASetLastError(WSA_INVALID_PARAMETER);
- return(false);
- }
-
- WSAEVENT events[WSA_MAXIMUM_WAIT_EVENTS];
- io_t *io_map[WSA_MAXIMUM_WAIT_EVENTS];
- DWORD event_index = 0;
-
- for splay_each(io_t, io, &io_tree) {
- events[event_index] = io->event;
- io_map[event_index] = io;
- event_index++;
- }
-
- /*
- * If the generation number changes due to event addition
- * or removal by a callback we restart the loop.
- */
- curgen = io_tree.generation;
-
- for(DWORD event_offset = 0; event_offset < event_count;) {
- DWORD result = WSAWaitForMultipleEvents(event_count - event_offset, &events[event_offset], FALSE, timeout_ms, FALSE);
-
- if(result == WSA_WAIT_TIMEOUT) {
- break;
- }
-
- if(result >= event_count - event_offset) {
- return false;
- }
-
- /* Look up io in the map by index. */
- event_index = result + event_offset;
- io_t *io = io_map[event_index];
-
- if(io->fd == -1) {
- io->cb(io->data, 0);
-
- if(curgen != io_tree.generation) {
- break;
- }
- } else {
- WSANETWORKEVENTS network_events;
-
- if(WSAEnumNetworkEvents(io->fd, io->event, &network_events) != 0) {
- return(false);
- }
-
- if(network_events.lNetworkEvents & READ_EVENTS) {
- io->cb(io->data, IO_READ);
-
- if(curgen != io_tree.generation) {
- break;
- }
- }
-
- /*
- The fd might be available for write too. However, if we already fired the read callback, that
- callback might have deleted the io (e.g. through terminate_connection()), so we can't fire the
- write callback here. Instead, we loop back and let the writable io loop above handle it.
- */
- }
-
- /* Continue checking the rest of the events. */
- event_offset = event_index + 1;
-
- /* Just poll the next time through. */
- timeout_ms = 0;
- }
- }
-
-#endif
-
- event_deinit();
- return true;
-}
-
-void event_exit(void) {
- running = false;
-}
projects / tinc / blobdiff