Linux: use epoll to handle great numbers of peers with less overhead

4 files changed, 179 insertions, 50 deletions

diff --git a/cmake/config.cmake b/cmake/config.cmake
index cbdc775..dbd5668 100644
--- a/cmake/config.cmake
+++ b/cmake/config.cmake
@@ -6,6 +6,7 @@ endif()
 
 
 set(USE_BINDTODEVICE ${LINUX})
+set(USE_EPOLL ${LINUX})
 set(USE_FREEBIND ${LINUX})
 set(USE_PMTU ${LINUX})
 set(USE_PKTINFO ${LINUX})
diff --git a/src/fastd.h b/src/fastd.h
index 8fcd247..f1ae2d3 100644
--- a/src/fastd.h
+++ b/src/fastd.h
@@ -241,10 +241,14 @@ struct fastd_context {
 	fastd_peer_group_t *peer_group;
 
 	VECTOR(fastd_peer_t*) peers;
-	VECTOR(struct pollfd) pollfds;
-
 	VECTOR(fastd_peer_t*) peers_temp;
 
+#ifdef USE_EPOLL
+	int epoll_fd;
+#else
+	VECTOR(struct pollfd) pollfds;
+#endif
+
 	uint32_t peer_addr_ht_seed;
 	VECTOR(fastd_peer_t*) *peer_addr_ht;
 
diff --git a/src/fastd_config.h.in b/src/fastd_config.h.in
index 8394853..0f9522b 100644
--- a/src/fastd_config.h.in
+++ b/src/fastd_config.h.in
@@ -32,6 +32,7 @@
 
 
 #cmakedefine USE_BINDTODEVICE
+#cmakedefine USE_EPOLL
 #cmakedefine USE_FREEBIND
 #cmakedefine USE_PMTU
 #cmakedefine USE_PKTINFO
diff --git a/src/poll.c b/src/poll.c
index 31e69bf..150f2ca 100644
--- a/src/poll.c
+++ b/src/poll.c
@@ -28,6 +28,174 @@
 #include "async.h"
 #include "peer.h"
 
+#ifdef USE_EPOLL
+
+#include <sys/epoll.h>
+
+#endif
+
+
+static inline bool handle_tun_tap(fastd_context_t *ctx, fastd_buffer_t buffer) {
+	if (ctx->conf->mode != MODE_TAP)
+		return false;
+
+	if (buffer.len < ETH_HLEN) {
+		pr_debug(ctx, "truncated packet on tap interface");
+		fastd_buffer_free(buffer);
+		return true;
+	}
+
+	fastd_eth_addr_t dest_addr = fastd_get_dest_address(ctx, buffer);
+	if (!fastd_eth_addr_is_unicast(dest_addr))
+		return false;
+
+	fastd_peer_t *peer = fastd_peer_find_by_eth_addr(ctx, dest_addr);
+
+	if (!peer)
+		return false;
+
+	ctx->conf->protocol->send(ctx, peer, buffer);
+	return true;
+}
+
+static void handle_tun(fastd_context_t *ctx) {
+	fastd_buffer_t buffer = fastd_tuntap_read(ctx);
+	if (!buffer.len)
+		return;
+
+	if (handle_tun_tap(ctx, buffer))
+		return;
+
+	/* TUN mode or multicast packet */
+	fastd_send_all(ctx, NULL, buffer);
+}
+
+static inline int handshake_timeout(fastd_context_t *ctx) {
+	if (!ctx->handshake_queue.next)
+		return -1;
+
+	fastd_peer_t *peer = container_of(ctx->handshake_queue.next, fastd_peer_t, handshake_entry);
+
+	int diff_msec = timespec_diff(&peer->next_handshake, &ctx->now);
+	if (diff_msec < 0)
+		return 0;
+	else
+		return diff_msec;
+}
+
+
+#ifdef USE_EPOLL
+
+#include <fcntl.h>
+#include <sys/epoll.h>
+
+
+void fastd_poll_init(fastd_context_t *ctx) {
+	ctx->epoll_fd = epoll_create1(EPOLL_CLOEXEC);
+	if (ctx->epoll_fd < 0)
+		exit_errno(ctx, "epoll_create1");
+
+	struct epoll_event event = {
+		.events = EPOLLIN,
+		.data.ptr = &ctx->async_rfd,
+	};
+	if (epoll_ctl(ctx->epoll_fd, EPOLL_CTL_ADD, ctx->async_rfd, &event) < 0)
+		exit_errno(ctx, "epoll_ctl");
+}
+
+void fastd_poll_free(fastd_context_t *ctx) {
+	if (close(ctx->epoll_fd))
+		pr_warn_errno(ctx, "closing EPOLL: close");
+}
+
+
+void fastd_poll_set_fd_tuntap(fastd_context_t *ctx) {
+	struct epoll_event event = {
+		.events = EPOLLIN,
+		.data.ptr = &ctx->tunfd,
+	};
+	if (epoll_ctl(ctx->epoll_fd, EPOLL_CTL_ADD, ctx->tunfd, &event) < 0)
+		exit_errno(ctx, "epoll_ctl");
+}
+
+void fastd_poll_set_fd_sock(fastd_context_t *ctx, size_t i) {
+	struct epoll_event event = {
+		.events = EPOLLIN,
+		.data.ptr = &ctx->socks[i],
+	};
+	if (epoll_ctl(ctx->epoll_fd, EPOLL_CTL_ADD, ctx->socks[i].fd, &event) < 0)
+		exit_errno(ctx, "epoll_ctl");
+}
+
+void fastd_poll_set_fd_peer(fastd_context_t *ctx, size_t i) {
+	fastd_peer_t *peer = VECTOR_INDEX(ctx->peers, i);
+
+	if (!peer->sock || !fastd_peer_is_socket_dynamic(peer))
+		return;
+
+	struct epoll_event event = {
+		.events = EPOLLIN,
+		.data.ptr = peer->sock,
+	};
+	if (epoll_ctl(ctx->epoll_fd, EPOLL_CTL_ADD, peer->sock->fd, &event) < 0)
+		exit_errno(ctx, "epoll_ctl");
+}
+
+void fastd_poll_add_peer(fastd_context_t *ctx UNUSED) {
+}
+
+void fastd_poll_delete_peer(fastd_context_t *ctx UNUSED, size_t i UNUSED) {
+}
+
+
+void fastd_poll_handle(fastd_context_t *ctx) {
+	int maintenance_timeout = timespec_diff(&ctx->next_maintenance, &ctx->now);
+
+	if (maintenance_timeout < 0)
+		maintenance_timeout = 0;
+
+	int timeout = handshake_timeout(ctx);
+	if (timeout < 0 || timeout > maintenance_timeout)
+		timeout = maintenance_timeout;
+
+	fastd_update_time(ctx);
+
+	struct epoll_event events[16];
+	int ret = epoll_wait(ctx->epoll_fd, events, 16, timeout);
+	if (ret < 0) {
+		if (errno == EINTR)
+			return;
+
+		exit_errno(ctx, "epoll_wait");
+	}
+
+	size_t i;
+	for (i = 0; i < (size_t)ret; i++) {
+		if (events[i].data.ptr == &ctx->tunfd) {
+			if (events[i].events & EPOLLIN)
+				handle_tun(ctx);
+		}
+		else if (events[i].data.ptr == &ctx->async_rfd) {
+			if (events[i].events & EPOLLIN)
+				fastd_async_handle(ctx);
+		}
+		else {
+			fastd_socket_t *sock = events[i].data.ptr;
+
+			if (events[i].events & (EPOLLERR|EPOLLHUP)) {
+				if (sock->peer)
+					fastd_peer_reset_socket(ctx, sock->peer);
+				else
+					fastd_socket_error(ctx, sock);
+			}
+			else if (events[i].events & EPOLLIN) {
+				fastd_receive(ctx, sock);
+			}
+		}
+	}
+}
+
+#else
 
 void fastd_poll_init(fastd_context_t *ctx) {
 	VECTOR_ALLOC(ctx->pollfds, 2 + ctx->n_socks);
@@ -91,54 +259,6 @@ void fastd_poll_delete_peer(fastd_context_t *ctx, size_t i) {
 }
 
 
-static inline bool handle_tun_tap(fastd_context_t *ctx, fastd_buffer_t buffer) {
-	if (ctx->conf->mode != MODE_TAP)
-		return false;
-
-	if (buffer.len < ETH_HLEN) {
-		pr_debug(ctx, "truncated packet on tap interface");
-		fastd_buffer_free(buffer);
-		return true;
-	}
-
-	fastd_eth_addr_t dest_addr = fastd_get_dest_address(ctx, buffer);
-	if (!fastd_eth_addr_is_unicast(dest_addr))
-		return false;
-
-	fastd_peer_t *peer = fastd_peer_find_by_eth_addr(ctx, dest_addr);
-
-	if (!peer)
-		return false;
-
-	ctx->conf->protocol->send(ctx, peer, buffer);
-	return true;
-}
-
-static void handle_tun(fastd_context_t *ctx) {
-	fastd_buffer_t buffer = fastd_tuntap_read(ctx);
-	if (!buffer.len)
-		return;
-
-	if (handle_tun_tap(ctx, buffer))
-		return;
-
-	/* TUN mode or multicast packet */
-	fastd_send_all(ctx, NULL, buffer);
-}
-
-static inline int handshake_timeout(fastd_context_t *ctx) {
-	if (!ctx->handshake_queue.next)
-		return -1;
-
-	fastd_peer_t *peer = container_of(ctx->handshake_queue.next, fastd_peer_t, handshake_entry);
-
-	int diff_msec = timespec_diff(&peer->next_handshake, &ctx->now);
-	if (diff_msec < 0)
-		return 0;
-	else
-		return diff_msec;
-}
-
 void fastd_poll_handle(fastd_context_t *ctx) {
 	int maintenance_timeout = timespec_diff(&ctx->next_maintenance, &ctx->now);
 
@@ -186,3 +306,6 @@ void fastd_poll_handle(fastd_context_t *ctx) {
 		}
 	}
 }
+
+#endif
+