From: Kirill Isakov Date: Sun, 22 May 2022 07:28:03 +0000 (+0600) Subject: Split event.c into per-API files X-Git-Url: https://tinc-vpn.org/git/browse?a=commitdiff_plain;h=021293e0d03de8d29b22104a8f9bef625b135640;p=tinc Split event.c into per-API files --- diff --git a/src/bsd/darwin/meson.build b/src/bsd/darwin/meson.build index 5672fcd0..f7dc99e2 100644 --- a/src/bsd/darwin/meson.build +++ b/src/bsd/darwin/meson.build @@ -7,3 +7,6 @@ if dep_tunemu.found() and dep_pcap.found() cdata.set('ENABLE_TUNEMU', 1) endif +# macOS apparently doesn't support kqueue with TAP devices +src_tincd += src_event_select + diff --git a/src/bsd/event.c b/src/bsd/event.c new file mode 100644 index 00000000..a1d57429 --- /dev/null +++ b/src/bsd/event.c @@ -0,0 +1,170 @@ +/* + event.c -- kqueue support for the BSD family + Copyright (C) 2012-2022 Guus Sliepen + + This program is free software; you can redistribute it and/or modify + it under the terms of the GNU General Public License as published by + the Free Software Foundation; either version 2 of the License, or + (at your option) any later version. + + This program is distributed in the hope that it will be useful, + but WITHOUT ANY WARRANTY; without even the implied warranty of + MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the + GNU General Public License for more details. + + You should have received a copy of the GNU General Public License along + with this program; if not, write to the Free Software Foundation, Inc., + 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA. +*/ + +#include "../system.h" + +#include + +#include "../event.h" +#include "../utils.h" +#include "../net.h" + +static bool running = false; +static int kq = 0; + +static inline void event_init(void) { + if(!kq) { + kq = kqueue(); + + if(kq == -1) { + logger(DEBUG_ALWAYS, LOG_EMERG, "Could not initialize kqueue: %s", strerror(errno)); + abort(); + } + } +} + +static void event_deinit(void) { + if(kq) { + close(kq); + kq = 0; + } +} + +void io_add(io_t *io, io_cb_t cb, void *data, int fd, int flags) { + if(io->cb) { + return; + } + + io->fd = fd; + io->cb = cb; + io->data = data; + io->node.data = io; + + io_set(io, flags); +} + +void io_set(io_t *io, int flags) { + event_init(); + + if(flags == io->flags) { + return; + } + + io->flags = flags; + + if(io->fd == -1) { + return; + } + + 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(kq, 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++; + } +} + +void io_del(io_t *io) { + if(io->cb) { + io_set(io, 0); + io->cb = NULL; + } +} + +bool event_loop(void) { + event_init(); + running = true; + + while(running) { + struct timeval diff; + struct timeval *tv = timeout_execute(&diff); + 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(kq, NULL, 0, events, MAX_EVENTS_PER_LOOP, &ts); + + if(n < 0) { + if(sockwouldblock(sockerrno)) { + continue; + } else { + return false; + } + } + + if(!n) { + continue; + } + + unsigned int curgen = io_tree.generation; + + 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; + } + } + } + + event_deinit(); + return true; +} + +void event_exit(void) { + running = false; +} diff --git a/src/bsd/meson.build b/src/bsd/meson.build index 91bc5507..cc294792 100644 --- a/src/bsd/meson.build +++ b/src/bsd/meson.build @@ -6,11 +6,6 @@ check_headers += [ 'net/tun/if_tun.h', ] -# macOS apparently doesn't support kqueue with TAP devices -if os_name != 'darwin' - check_headers += 'sys/event.h' -endif - check_functions += [ 'devname', 'fdevname', @@ -18,6 +13,10 @@ check_functions += [ src_tincd += files('device.c') +if os_name != 'darwin' + src_tincd += files('event.c') +endif + if os_name in ['openbsd', 'darwin'] subdir(os_name) endif diff --git a/src/event.c b/src/event.c index c373eaca..56d8fa97 100644 --- a/src/event.c +++ b/src/event.c @@ -1,5 +1,5 @@ /* - event.c -- I/O, timeout and signal event handling + event.c -- I/O, timeout, and event handling Copyright (C) 2012-2022 Guus Sliepen This program is free software; you can redistribute it and/or modify @@ -19,71 +19,9 @@ #include "system.h" -#ifdef HAVE_WINDOWS -# include -#else -# if defined(HAVE_SYS_EPOLL_H) -# include -# define HAVE_EPOLL 1 -# elif defined(HAVE_SYS_EVENT_H) -# include -# 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 @@ -133,178 +71,8 @@ static int timeout_compare(const timeout_t *a, const timeout_t *b) { 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; @@ -342,79 +110,7 @@ void timeout_del(timeout_t *timeout) { }; } -#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; @@ -436,258 +132,3 @@ static struct timeval *timeout_execute(struct timeval *diff) { 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; -} diff --git a/src/event.h b/src/event.h index 1443ebca..386c0b51 100644 --- a/src/event.h +++ b/src/event.h @@ -56,6 +56,9 @@ typedef struct signal_t { extern struct timeval now; +extern splay_tree_t io_tree; +extern splay_tree_t timeout_tree; + extern void io_add(io_t *io, io_cb_t cb, void *data, int fd, int flags); #ifdef HAVE_WINDOWS extern void io_add_event(io_t *io, io_cb_t cb, void *data, WSAEVENT event); @@ -66,6 +69,7 @@ extern void io_set(io_t *io, int flags); extern void timeout_add(timeout_t *timeout, timeout_cb_t cb, void *data, const struct timeval *tv); extern void timeout_del(timeout_t *timeout); extern void timeout_set(timeout_t *timeout, const struct timeval *tv); +extern struct timeval *timeout_execute(struct timeval *diff); extern void signal_add(signal_t *sig, signal_cb_t cb, void *data, int signum); extern void signal_del(signal_t *sig); diff --git a/src/event_select.c b/src/event_select.c new file mode 100644 index 00000000..cac61e39 --- /dev/null +++ b/src/event_select.c @@ -0,0 +1,140 @@ +/* + event_select.c -- select(2) support + Copyright (C) 2012-2022 Guus Sliepen + + This program is free software; you can redistribute it and/or modify + it under the terms of the GNU General Public License as published by + the Free Software Foundation; either version 2 of the License, or + (at your option) any later version. + + This program is distributed in the hope that it will be useful, + but WITHOUT ANY WARRANTY; without even the implied warranty of + MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the + GNU General Public License for more details. + + You should have received a copy of the GNU General Public License along + with this program; if not, write to the Free Software Foundation, Inc., + 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA. +*/ + +#include "system.h" + +#include "event.h" +#include "utils.h" + +static bool running = false; +static fd_set readfds; +static fd_set writefds; + +void io_add(io_t *io, io_cb_t cb, void *data, int fd, int flags) { + if(io->cb) { + return; + } + + io->fd = fd; + io->cb = cb; + io->data = data; + io->node.data = io; + + io_set(io, flags); + + if(!splay_insert_node(&io_tree, &io->node)) { + abort(); + } +} + +void io_set(io_t *io, int flags) { + if(flags == io->flags) { + return; + } + + io->flags = flags; + + if(io->fd == -1) { + return; + } + + 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); + } +} + +void io_del(io_t *io) { + if(io->cb) { + io_set(io, 0); + splay_unlink_node(&io_tree, &io->node); + io->cb = NULL; + } +} + +bool event_loop(void) { + running = true; + + fd_set readable; + fd_set writable; + + while(running) { + struct timeval diff; + struct timeval *tv = timeout_execute(&diff); + + memcpy(&readable, &readfds, sizeof(readable)); + memcpy(&writable, &writefds, sizeof(writable)); + + 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); + + if(n < 0) { + if(sockwouldblock(sockerrno)) { + continue; + } else { + return false; + } + } + + if(!n) { + continue; + } + + unsigned int curgen = io_tree.generation; + + 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; + } + } + } + + return true; +} + +void event_exit(void) { + running = false; +} diff --git a/src/linux/event.c b/src/linux/event.c new file mode 100644 index 00000000..ebcccfe7 --- /dev/null +++ b/src/linux/event.c @@ -0,0 +1,169 @@ +/* + event.c -- epoll support for Linux + Copyright (C) 2012-2022 Guus Sliepen + + This program is free software; you can redistribute it and/or modify + it under the terms of the GNU General Public License as published by + the Free Software Foundation; either version 2 of the License, or + (at your option) any later version. + + This program is distributed in the hope that it will be useful, + but WITHOUT ANY WARRANTY; without even the implied warranty of + MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the + GNU General Public License for more details. + + You should have received a copy of the GNU General Public License along + with this program; if not, write to the Free Software Foundation, Inc., + 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA. +*/ + +#include "../system.h" + +#include + +#include "../event.h" +#include "../utils.h" +#include "../net.h" + +static bool running = false; +static int epollset = 0; + +/* 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. */ +static inline void event_init(void) { + if(!epollset) { + epollset = epoll_create(1024); + + if(epollset == -1) { + logger(DEBUG_ALWAYS, LOG_EMERG, "Could not initialize epoll: %s", strerror(errno)); + abort(); + } + } +} + +static void event_deinit(void) { + if(epollset) { + close(epollset); + epollset = 0; + } +} + +void io_add(io_t *io, io_cb_t cb, void *data, int fd, int flags) { + if(io->cb) { + return; + } + + io->fd = fd; + io->cb = cb; + io->data = data; + io->node.data = io; + + io_set(io, flags); +} + +void io_set(io_t *io, int flags) { + event_init(); + + if(flags == io->flags) { + return; + } + + io->flags = flags; + + if(io->fd == -1) { + return; + } + + epoll_ctl(epollset, 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(epollset, EPOLL_CTL_ADD, io->fd, &ev) < 0) { + logger(DEBUG_ALWAYS, LOG_EMERG, "epoll_ctl failed: %s", strerror(errno)); + abort(); + } +} + +void io_del(io_t *io) { + if(io->cb) { + io_set(io, 0); + io->cb = NULL; + } +} + +bool event_loop(void) { + event_init(); + running = true; + + while(running) { + struct timeval diff; + struct timeval *tv = timeout_execute(&diff); + + 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(epollset, events, MAX_EVENTS_PER_LOOP, (int)timeout); + + if(n < 0) { + if(sockwouldblock(sockerrno)) { + continue; + } else { + return false; + } + } + + if(!n) { + continue; + } + + unsigned int curgen = io_tree.generation; + + 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; + } + } + + } + + event_deinit(); + return true; +} + +void event_exit(void) { + running = false; +} diff --git a/src/linux/meson.build b/src/linux/meson.build index 1b94f95c..5c5a273b 100644 --- a/src/linux/meson.build +++ b/src/linux/meson.build @@ -1,12 +1,14 @@ check_headers += [ 'linux/if_tun.h', 'netpacket/packet.h', - 'sys/epoll.h', ] check_functions += 'recvmmsg' -src_tincd += files('device.c') +src_tincd += files( + 'device.c', + 'event.c', +) dep_libsystemd = dependency('libsystemd', required: opt_systemd) if dep_libsystemd.found() diff --git a/src/meson.build b/src/meson.build index d9f7b14b..cc646288 100644 --- a/src/meson.build +++ b/src/meson.build @@ -174,6 +174,12 @@ src_tincd = [ 'subnet.c', ] +src_event_select = files('event_select.c') + +if os_name != 'windows' + src_tincd += 'signal.c' +endif + cc_flags_tinc = cc_flags cc_flags_tincd = cc_flags diff --git a/src/signal.c b/src/signal.c new file mode 100644 index 00000000..73f72191 --- /dev/null +++ b/src/signal.c @@ -0,0 +1,90 @@ +/* + signal.c -- signal handling + Copyright (C) 2012-2022 Guus Sliepen + + This program is free software; you can redistribute it and/or modify + it under the terms of the GNU General Public License as published by + the Free Software Foundation; either version 2 of the License, or + (at your option) any later version. + + This program is distributed in the hope that it will be useful, + but WITHOUT ANY WARRANTY; without even the implied warranty of + MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the + GNU General Public License for more details. + + You should have received a copy of the GNU General Public License along + with this program; if not, write to the Free Software Foundation, Inc., + 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA. +*/ + +#include "system.h" + +#include "event.h" + +// 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; +} diff --git a/src/solaris/meson.build b/src/solaris/meson.build index 5d071ff7..6247af20 100644 --- a/src/solaris/meson.build +++ b/src/solaris/meson.build @@ -1,4 +1,7 @@ -src_tincd += files('device.c') +src_tincd += [ + files('device.c'), + src_event_select, +] deps_common += cc.find_library('libsocket') diff --git a/src/windows/event.c b/src/windows/event.c new file mode 100644 index 00000000..a4688d0e --- /dev/null +++ b/src/windows/event.c @@ -0,0 +1,223 @@ +/* + event.c -- event support for Windows + Copyright (C) 2012-2022 Guus Sliepen + + This program is free software; you can redistribute it and/or modify + it under the terms of the GNU General Public License as published by + the Free Software Foundation; either version 2 of the License, or + (at your option) any later version. + + This program is distributed in the hope that it will be useful, + but WITHOUT ANY WARRANTY; without even the implied warranty of + MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the + GNU General Public License for more details. + + You should have received a copy of the GNU General Public License along + with this program; if not, write to the Free Software Foundation, Inc., + 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA. +*/ + +#include "../system.h" + +#include + +#include "../event.h" +#include "../utils.h" +#include "../net.h" + +static bool running = false; +static DWORD event_count = 0; + +static const long READ_EVENTS = FD_READ | FD_ACCEPT | FD_CLOSE; +static const long WRITE_EVENTS = FD_WRITE | FD_CONNECT; + +void io_add(io_t *io, io_cb_t cb, void *data, int fd, int flags) { + if(io->cb) { + return; + } + + io->fd = fd; + + if(io->fd != -1) { + io->event = WSACreateEvent(); + + if(io->event == WSA_INVALID_EVENT) { + abort(); + } + } + + event_count++; + + io->cb = cb; + io->data = data; + io->node.data = io; + + io_set(io, flags); + + if(!splay_insert_node(&io_tree, &io->node)) { + abort(); + } +} + +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); +} + +void io_set(io_t *io, int flags) { + if(flags == io->flags) { + return; + } + + io->flags = flags; + + if(io->fd == -1) { + return; + } + + 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(); + } +} + +void io_del(io_t *io) { + if(!io->cb) { + return; + } + + io_set(io, 0); + + if(io->fd != -1 && WSACloseEvent(io->event) == FALSE) { + abort(); + } + + event_count--; + splay_unlink_node(&io_tree, &io->node); + io->cb = NULL; +} + +bool event_loop(void) { + running = true; + + 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; + } + } + + return true; +} + +void event_exit(void) { + running = false; +} diff --git a/src/windows/meson.build b/src/windows/meson.build index b78b901d..40f7013e 100644 --- a/src/windows/meson.build +++ b/src/windows/meson.build @@ -15,5 +15,9 @@ foreach libname : win_common_libs endforeach src_lib_common += files('random.c') -src_tincd += files('device.c') + +src_tincd += files( + 'device.c', + 'event.c', +)