Reduce number of per connection sys calls

4 files changed, 30 insertions, 5 deletions

diff --git a/ot_event.c b/ot_event.c
index 1542bf5..8f55817 100644
--- a/ot_event.c
+++ b/ot_event.c
@@ -71,6 +71,10 @@ int event_add(int fd, int what, uint32_t generation) { return event_ctl(EPOLL_CT
 int event_modify(int fd, int what, uint32_t generation) { return event_ctl(EPOLL_CTL_MOD, fd, what, generation); }
 int event_del(int fd) { return epoll_ctl(g_epoll, EPOLL_CTL_DEL, fd, NULL); }
 
+/* A descriptor leaves the interest list when the last reference to its open
+   file description is closed, and the tracker never duplicates one */
+int event_close(int fd) { return close(fd); }
+
 int event_wait(ot_event *out, int max, int timeout_ms) {
   int count, i;
 
@@ -171,6 +175,9 @@ int event_add(int fd, int what, uint32_t generation) { return event_change(fd, w
 int event_modify(int fd, int what, uint32_t generation) { return event_change(fd, what, generation, 0); }
 int event_del(int fd) { return event_change(fd, 0, 0, 1); }
 
+/* Closing a descriptor removes every kevent that references it */
+int event_close(int fd) { return close(fd); }
+
 int event_wait(ot_event *out, int max, int timeout_ms) {
   struct timespec  timeout;
   struct timespec *deadline = NULL;
@@ -311,6 +318,13 @@ int event_del(int fd) {
   return 0;
 }
 
+/* Nothing here lives in the kernel, so the entry has to come out by hand
+   before the descriptor goes, or poll reports POLLNVAL for it forever */
+int event_close(int fd) {
+  event_del(fd);
+  return close(fd);
+}
+
 int event_wait(ot_event *out, int max, int timeout_ms) {
   int ready, found = 0, i;
 
diff --git a/ot_event.h b/ot_event.h
index 6023134..aa007d0 100644
--- a/ot_event.h
+++ b/ot_event.h
@@ -29,6 +29,11 @@ int  event_add(int fd, int what, uint32_t generation);
 int  event_modify(int fd, int what, uint32_t generation);
 int  event_del(int fd);
 
+/* Stop watching a descriptor and close it. Where the kernel drops a closed
+   descriptor from its set by itself, this is one system call where doing the
+   two separately would be two. */
+int  event_close(int fd);
+
 /* Waits at most timeout_ms, or indefinitely when that is negative, and fills
    out with up to max ready descriptors. Returns how many, or -1. */
 int  event_wait(ot_event *out, int max, int timeout_ms);
diff --git a/ot_http.c b/ot_http.c
index c2f7e5e..d4c7363 100644
--- a/ot_http.c
+++ b/ot_http.c
@@ -56,8 +56,7 @@ void http_close(const int64_t sock) {
     free(cookie);
   }
 
-  event_del((int)sock);
-  close((int)sock);
+  event_close((int)sock);
 }
 
 static void http_senddata(const int64_t sock, struct ot_workstruct *ws) {
diff --git a/ot_socket.c b/ot_socket.c
index 158459f..47ed26d 100644
--- a/ot_socket.c
+++ b/ot_socket.c
@@ -4,6 +4,7 @@
    $id$ */
 
 /* System */
+#define _GNU_SOURCE /* accept4 on glibc */
 #include <errno.h>
 #include <fcntl.h>
 #include <netinet/in.h>
@@ -170,19 +171,25 @@ int socket_is_connected(int s) {
   return 1;
 }
 
+/* Linux hands over a blocking socket whatever the listening socket was, the
+   BSDs and macOS inherit the listener's flags, so the accepted socket has to
+   be told either way. accept4 carries the flag itself, which saves the two
+   fcntl calls it otherwise takes. Darwin has neither accept4 nor the flag. */
 int socket_accept(int s, ot_ip6 ip, uint16_t *port) {
   struct sockaddr_storage ss;
   socklen_t               len = sizeof(ss);
   int                     fd;
 
   memset(&ss, 0, sizeof(ss));
+#ifdef SOCK_NONBLOCK
+  if ((fd = accept4(s, (struct sockaddr *)&ss, &len, SOCK_NONBLOCK)) == -1)
+    return -1;
+#else
   if ((fd = accept(s, (struct sockaddr *)&ss, &len)) == -1)
     return -1;
-
-  /* Linux hands over a blocking socket whatever the listening socket was, the
-     BSDs and macOS inherit the listener's flags, so say which one we want. */
   if (set_nonblocking(fd) == -1)
     return close_and_fail(fd);
+#endif
 
   ip_from_sockaddr(ip, port, &ss);
   return fd;