X-Git-Url: https://tinc-vpn.org/git/browse?a=blobdiff_plain;f=src%2Fevent.c;h=899b62a077abb435e5327b359bacc41402ccaaf1;hb=28b7a53b693f6b4e70218a926e68a36ece54cda1;hp=5edd3a6d0b7cf8efbef2100ba43012085ac5639d;hpb=f6e87ab476a0faf8b124ecaaa27f967d825e6457;p=tinc

diff --git a/src/event.c b/src/event.c
index 5edd3a6d..899b62a0 100644
--- a/src/event.c
+++ b/src/event.c
@@ -1,6 +1,6 @@
 /*
     event.c -- I/O, timeout and signal event handling
-    Copyright (C) 2012-2013 Guus Sliepen <guus@tinc-vpn.org>
+    Copyright (C) 2012-2022 Guus Sliepen <guus@tinc-vpn.org>
 
     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
@@ -19,17 +19,27 @@
 
 #include "system.h"
 
+#include <assert.h>
+
+#ifdef HAVE_SYS_EPOLL_H
+#include <sys/epoll.h>
+#endif
+
 #include "dropin.h"
 #include "event.h"
-#include "net.h"
 #include "utils.h"
-#include "xalloc.h"
+#include "net.h"
 
 struct timeval now;
-
 #ifndef HAVE_MINGW
+
+#ifdef HAVE_SYS_EPOLL_H
+static int epollset = 0;
+#else
 static fd_set readfds;
 static fd_set writefds;
+#endif
+
 #else
 static const long READ_EVENTS = FD_READ | FD_ACCEPT | FD_CLOSE;
 static const long WRITE_EVENTS = FD_WRITE | FD_CONNECT;
@@ -37,11 +47,31 @@ static DWORD event_count = 0;
 #endif
 static bool running;
 
+#ifdef HAVE_SYS_EPOLL_H
+static inline int event_epoll_init(void) {
+	/* 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.
+	*/
+	return epoll_create(1024);
+}
+#endif
+
 static int io_compare(const io_t *a, const io_t *b) {
 #ifndef HAVE_MINGW
 	return a->fd - b->fd;
 #else
-	return a->event - b->event;
+
+	if(a->event < b->event) {
+		return -1;
+	}
+
+	if(a->event > b->event) {
+		return 1;
+	}
+
+	return 0;
 #endif
 }
 
@@ -103,9 +133,13 @@ void io_add(io_t *io, io_cb_t cb, void *data, int fd, int flags) {
 
 	io_set(io, flags);
 
+#ifndef HAVE_SYS_EPOLL_H
+
 	if(!splay_insert_node(&io_tree, &io->node)) {
 		abort();
 	}
+
+#endif
 }
 
 #ifdef HAVE_MINGW
@@ -116,6 +150,14 @@ void io_add_event(io_t *io, io_cb_t cb, void *data, WSAEVENT event) {
 #endif
 
 void io_set(io_t *io, int flags) {
+#ifdef HAVE_SYS_EPOLL_H
+
+	if(!epollset) {
+		epollset = event_epoll_init();
+	}
+
+#endif
+
 	if(flags == io->flags) {
 		return;
 	}
@@ -127,6 +169,30 @@ void io_set(io_t *io, int flags) {
 	}
 
 #ifndef HAVE_MINGW
+#ifdef HAVE_SYS_EPOLL_H
+	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) {
+		perror("epoll_ctl_add");
+	}
+
+#else
 
 	if(flags & IO_READ) {
 		FD_SET(io->fd, &readfds);
@@ -140,6 +206,7 @@ void io_set(io_t *io, int flags) {
 		FD_CLR(io->fd, &writefds);
 	}
 
+#endif
 #else
 	long events = 0;
 
@@ -173,7 +240,9 @@ void io_del(io_t *io) {
 	event_count--;
 #endif
 
+#ifndef HAVE_SYS_EPOLL_H
 	splay_unlink_node(&io_tree, &io->node);
+#endif
 	io->cb = NULL;
 }
 
@@ -214,29 +283,35 @@ void timeout_del(timeout_t *timeout) {
 }
 
 #ifndef HAVE_MINGW
-static int signal_compare(const signal_t *a, const signal_t *b) {
-	return a->signum - b->signum;
-}
+
+// From Matz's Ruby
+#ifndef NSIG
+# define NSIG (_SIGMAX + 1)      /* For QNX */
+#endif
+
 
 static io_t signalio;
 static int pipefd[2] = {-1, -1};
-static splay_tree_t signal_tree = {.compare = (splay_compare_t)signal_compare};
+static signal_t *signal_handle[NSIG + 1] = {NULL};
 
 static void signal_handler(int signum) {
 	unsigned char num = signum;
-	write(pipefd[1], &num, 1);
+
+	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 = splay_search(&signal_tree, &((signal_t) {
-		.signum = signum
-	}));
+	signal_t *sig = signal_handle[signum];
 
 	if(sig) {
 		sig->cb(sig->data);
@@ -254,20 +329,17 @@ void signal_add(signal_t *sig, signal_cb_t cb, void *data, int signum) {
 		return;
 	}
 
+	sig->signum = signum;
 	sig->cb = cb;
 	sig->data = data;
-	sig->signum = signum;
-	sig->node.data = sig;
 
 	if(pipefd[0] == -1) {
 		pipe_init();
 	}
 
-	signal(sig->signum, signal_handler);
+	signal(signum, signal_handler);
 
-	if(!splay_insert_node(&signal_tree, &sig->node)) {
-		abort();
-	}
+	signal_handle[signum] = sig;
 }
 
 void signal_del(signal_t *sig) {
@@ -277,12 +349,12 @@ void signal_del(signal_t *sig) {
 
 	signal(sig->signum, SIG_DFL);
 
-	splay_unlink_node(&signal_tree, &sig->node);
+	signal_handle[sig->signum] = NULL;
 	sig->cb = NULL;
 }
 #endif
 
-static struct timeval *get_time_remaining(struct timeval *diff) {
+static struct timeval *timeout_execute(struct timeval *diff) {
 	gettimeofday(&now, NULL);
 	struct timeval *tv = NULL;
 
@@ -309,23 +381,46 @@ bool event_loop(void) {
 	running = true;
 
 #ifndef HAVE_MINGW
+
+#ifdef HAVE_SYS_EPOLL_H
+
+	if(!epollset) {
+		epollset = event_epoll_init();
+	}
+
+#else
 	fd_set readable;
 	fd_set writable;
+#endif
 
 	while(running) {
 		struct timeval diff;
-		struct timeval *tv = get_time_remaining(&diff);
+		struct timeval *tv = timeout_execute(&diff);
+#ifndef HAVE_SYS_EPOLL_H
 		memcpy(&readable, &readfds, sizeof(readable));
 		memcpy(&writable, &writefds, sizeof(writable));
+#endif
+
 
-		int fds = 0;
+#ifdef HAVE_SYS_EPOLL_H
+		struct epoll_event events[EPOLL_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, EPOLL_MAX_EVENTS_PER_LOOP, (int)timeout);
+#else
+		int maxfds =  0;
 
 		if(io_tree.tail) {
 			io_t *last = io_tree.tail->data;
-			fds = last->fd + 1;
+			maxfds = last->fd + 1;
 		}
 
-		int n = select(fds, &readable, &writable, NULL, tv);
+		int n = select(maxfds, &readable, &writable, NULL, tv);
+#endif
 
 		if(n < 0) {
 			if(sockwouldblock(sockerrno)) {
@@ -339,6 +434,33 @@ bool event_loop(void) {
 			continue;
 		}
 
+		unsigned int curgen = io_tree.generation;
+
+
+#ifdef HAVE_SYS_EPOLL_H
+
+		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;
+			}
+		}
+
+#else
+
 		for splay_each(io_t, io, &io_tree) {
 			if(FD_ISSET(io->fd, &writable)) {
 				io->cb(io->data, IO_WRITE);
@@ -349,21 +471,26 @@ bool event_loop(void) {
 			}
 
 			/*
-			   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 now. That's okay, since any remaining events will get picked
-			   up by the next select() call.
-			 */
-			break;
+			        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 = get_time_remaining(&diff);
-		DWORD timeout_ms = tv ? (tv->tv_sec * 1000 + tv->tv_usec / 1000 + 1) : WSA_INFINITE;
+		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);
@@ -380,71 +507,87 @@ bool event_loop(void) {
 		   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.
 		*/
-		io_t *writeable_io = NULL;
+		unsigned int curgen = io_tree.generation;
 
-		for splay_each(io_t, io, &io_tree)
+		for splay_each(io_t, io, &io_tree) {
 			if(io->flags & IO_WRITE && send(io->fd, NULL, 0, 0) == 0) {
-				writeable_io = io;
-				break;
+				io->cb(io->data, IO_WRITE);
+
+				if(curgen != io_tree.generation) {
+					break;
+				}
 			}
+		}
 
-		if(writeable_io) {
-			writeable_io->cb(writeable_io->data, IO_WRITE);
-			continue;
+		if(event_count > WSA_MAXIMUM_WAIT_EVENTS) {
+			WSASetLastError(WSA_INVALID_PARAMETER);
+			return(false);
 		}
 
-		WSAEVENT *events = xmalloc(event_count * sizeof(*events));
+		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++;
 		}
 
-		DWORD result = WSAWaitForMultipleEvents(event_count, events, FALSE, timeout_ms, FALSE);
+		/*
+		 * If the generation number changes due to event addition
+		 * or removal by a callback we restart the loop.
+		 */
+		curgen = io_tree.generation;
 
-		WSAEVENT event;
+		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_EVENT_0 && result < WSA_WAIT_EVENT_0 + event_count) {
-			event = events[result - WSA_WAIT_EVENT_0];
-		}
+			if(result == WSA_WAIT_TIMEOUT) {
+				break;
+			}
 
-		free(events);
+			if(result >= event_count - event_offset) {
+				return false;
+			}
 
-		if(result == WSA_WAIT_TIMEOUT) {
-			continue;
-		}
+			/* Look up io in the map by index. */
+			event_index = result + event_offset;
+			io_t *io = io_map[event_index];
 
-		if(result < WSA_WAIT_EVENT_0 || result >= WSA_WAIT_EVENT_0 + event_count) {
-			return false;
-		}
+			if(io->fd == -1) {
+				io->cb(io->data, 0);
 
-		io_t *io = splay_search(&io_tree, &((io_t) {
-			.event = event
-		}));
+				if(curgen != io_tree.generation) {
+					break;
+				}
+			} else {
+				WSANETWORKEVENTS network_events;
 
-		if(!io) {
-			abort();
-		}
+				if(WSAEnumNetworkEvents(io->fd, io->event, &network_events) != 0) {
+					return(false);
+				}
 
-		if(io->fd == -1) {
-			io->cb(io->data, 0);
-		} else {
-			WSANETWORKEVENTS network_events;
+				if(network_events.lNetworkEvents & READ_EVENTS) {
+					io->cb(io->data, IO_READ);
 
-			if(WSAEnumNetworkEvents(io->fd, io->event, &network_events) != 0) {
-				return false;
-			}
+					if(curgen != io_tree.generation) {
+						break;
+					}
+				}
 
-			if(network_events.lNetworkEvents & READ_EVENTS) {
-				io->cb(io->data, IO_READ);
+				/*
+				    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.
+				 */
 			}
 
-			/*
-			    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;
 		}
 	}