1 files changed, 35 insertions, 37 deletions

diff --git a/src/knot/server/tcp-handler.c b/src/knot/server/tcp-handler.c
index 1f18abf..3ebdfe4 100644
--- a/src/knot/server/tcp-handler.c
+++ b/src/knot/server/tcp-handler.c
@@ -25,6 +25,9 @@
 #include <netinet/in.h>
 #include <stdio.h>
 #include <stdlib.h>
+#ifdef HAVE_SYS_UIO_H			// struct iovec (OpenBSD)
+#include <sys/uio.h>
+#endif // HAVE_SYS_UIO_H
 #ifdef HAVE_CAP_NG_H
 #include <cap-ng.h>
 #endif /* HAVE_CAP_NG_H */
@@ -403,44 +406,22 @@ static void tcp_loop_free(void *data)
 
 int tcp_send(int fd, uint8_t *msg, size_t msglen)
 {
-
-	/*! \brief TCP corking.
-	 *  \see http://vger.kernel.org/~acme/unbehaved.txt
-	 */
-#ifdef TCP_CORK
-	int cork = 1;
-	int uncork = setsockopt(fd, SOL_TCP, TCP_CORK, &cork, sizeof(cork));
-#endif
-
-	/* Flags. */
-	int flags = 0;
-#ifdef MSG_NOSIGNAL
-	flags |= MSG_NOSIGNAL;
-#endif
-
-	/* Send message size. */
-	unsigned short pktsize = htons(msglen);
-	int sent = send(fd, &pktsize, sizeof(pktsize), flags);
-	if (sent < 0) {
-		return KNOT_ERROR;
-	}
-
-	/* Send message data. */
-	sent = send(fd, msg, msglen, flags);
-	if (sent < 0) {
+	/* Create iovec for gathered write. */
+	struct iovec iov[2];
+	uint16_t pktsize = htons(msglen);
+	iov[0].iov_base = &pktsize;
+	iov[0].iov_len = sizeof(uint16_t);
+	iov[1].iov_base = msg;
+	iov[1].iov_len = msglen;
+
+	/* Send. */
+	int total_len = iov[0].iov_len + iov[1].iov_len;
+	int sent = writev(fd, iov, 2);
+	if (sent != total_len) {
 		return KNOT_ERROR;
 	}
 
-#ifdef TCP_CORK
-	/* Uncork only if corked successfuly. */
-	if (uncork == 0) {
-		cork = 0;
-		if (setsockopt(fd, SOL_TCP, TCP_CORK, &cork, sizeof(cork)) < 0) {
-			dbg_net("tcp: failed to uncork socket\n");
-		}
-	}
-#endif
-	return sent;
+	return msglen; /* Do not count the size prefix. */
 }
 
 int tcp_recv(int fd, uint8_t *buf, size_t len, sockaddr_t *addr)
@@ -543,7 +524,16 @@ int tcp_loop_master(dthread_t *thread)
 		}
 
 		for (unsigned i = 0; nfds > 0 && i < set.n; ++i) {
-			/* Skip inactive. */
+
+			/* Error events. */
+			if (set.pfd[i].revents & (POLLERR|POLLHUP|POLLNVAL)) {
+				socket_close(set.pfd[i].fd);
+				fdset_remove(&set, i);
+				--nfds;   /* Treat error event as activity. */
+				continue; /* Stay on the same index. */
+			}
+
+			/* Accept POLLIN events. */
 			if (!(set.pfd[i].revents & POLLIN))
 				continue;
 
@@ -611,6 +601,15 @@ int tcp_loop_worker(dthread_t *thread)
 		unsigned i = 0;
 		while (nfds > 0 && i < set->n) {
 
+			/* Terminate faulty connections. */
+			int fd = set->pfd[i].fd;
+			if (set->pfd[i].revents & (POLLERR|POLLHUP|POLLNVAL)) {
+				fdset_remove(set, i);
+				close(fd);
+				--nfds;   /* Treat error event as activity. */
+				continue; /* Stay on the same index. */
+			}
+
 			if (!(set->pfd[i].revents & set->pfd[i].events)) {
 				/* Skip inactive. */
 				++i;
@@ -621,7 +620,6 @@ int tcp_loop_worker(dthread_t *thread)
 			}
 
 			/* Register new TCP client or process a query. */
-			int fd = set->pfd[i].fd;
 			if (fd == w->pipe[0]) {
 				tcp_loop_assign(fd, set);
 			} else {