BIRD patch for select() -> poll()

gistfile1.diff

--- bird-1.3.9/sysdep/unix/io.c  2013-01-11 05:53:16.000000000 -0800
+++ /home/hkim/bird-1.3.9/sysdep/unix/io.c	2013-05-20 02:34:36.780698000 -0700
@@ -14,6 +14,7 @@
 #include <stdio.h>
 #include <stdlib.h>
 #include <time.h>
+#include <sys/poll.h>
 #include <sys/time.h>
 #include <sys/types.h>
 #include <sys/socket.h>
@@ -48,6 +49,10 @@
    this to gen small latencies */
 #define MAX_RX_STEPS 4
 
+/* Lets start simple with 8k fds for polling instead of dynamically
+ * allocate pollfds by the number of bird sockets */
+#define MAX_POLLPDFS 1024 * 8
+
 /*
  *	Tracked Files
  */
@@ -1330,7 +1335,7 @@ sk_rx_ready(sock *s)
 
  redo:
   rv = select(s->fd+1, &rd, &wr, NULL, &timo);
-  
+  debug("DEBUG: SK_RX_READY, SELECTOR FULL [%d]\n", rv);
   if ((rv < 0) && (errno == EINTR || errno == EAGAIN))
     goto redo;
 
@@ -1540,16 +1545,19 @@ static int short_loops = 0;
 void
 io_loop(void)
 {
-  fd_set rd, wr;
   struct timeval timo;
   time_t tout;
   int hi, events;
   sock *s;
   node *n;
 
+  /* lets start simple using array for poll fds and use dynamic
+   * allocation by the number of available bird sockets later */
+  int nfds;
+  struct pollfd pollfds[MAX_POLLPDFS];
+
   sock_recalc_fdsets_p = 1;
-  for(;;)
-    {
+  for(;;) {
       events = ev_run_list(&global_event_list);
       update_times();
       tout = tm_first_shot();
@@ -1558,36 +1566,35 @@ io_loop(void)
 	  tm_shot();
 	  continue;
 	}
+
+    // TODO: poll uses milliseconds instead seconds
       timo.tv_sec = events ? 0 : MIN(tout - now, 3);
       timo.tv_usec = 0;
 
-      if (sock_recalc_fdsets_p)
-	{
+    // TODO: figure out whether we need to reset every file
+    //       descriptors.
+    if (sock_recalc_fdsets_p) {
 	  sock_recalc_fdsets_p = 0;
-	  FD_ZERO(&rd);
-	  FD_ZERO(&wr);
 	}
 
       hi = 0;
-      WALK_LIST(n, sock_list)
-	{
+    nfds = 0;
+    WALK_LIST(n, sock_list) {
 	  s = SKIP_BACK(sock, n, n);
-	  if (s->rx_hook)
-	    {
-	      FD_SET(s->fd, &rd);
-	      if (s->fd > hi)
-		hi = s->fd;
+
+      if (s->rx_hook) {
+        pollfds[nfds].fd = s->fd;
+        pollfds[nfds].events = POLLIN;
 	    }
-	  else
-	    FD_CLR(s->fd, &rd);
-	  if (s->tx_hook && s->ttx != s->tpos)
-	    {
-	      FD_SET(s->fd, &wr);
-	      if (s->fd > hi)
-		hi = s->fd;
+
+      if (s->tx_hook && s->ttx != s->tpos) {
+        pollfds[nfds].fd = s->fd;
+        pollfds[nfds].events = POLLOUT;
 	    }
-	  else
-	    FD_CLR(s->fd, &wr);
+
+      /* TODO: this is too naive approach, there might be a bird
+       * socket with no rx/tx hook. */
+      nfds++;
 	}
 
       /*
@@ -1614,47 +1621,71 @@ io_loop(void)
 	  continue;
 	}
 
-      /* And finally enter select() to find active sockets */
-      hi = select(hi+1, &rd, &wr, NULL, &timo);
+    // TODO: decide appropriate timeout value
+    hi = poll(pollfds, nfds + 1, 1000);
+    debug("DEBUG: # of poll events occured [%d] out of [%d] sockets\n", hi, nfds);
 
       if (hi < 0)
-	{
-	  if (errno == EINTR || errno == EAGAIN)
-	    continue;
-	  die("select: %m");
-	}
-      if (hi)
-	{
+      debug("DEBUG: POLL ERROR\n");
+
+    if (hi) {
 	  /* guaranteed to be non-empty */
 	  current_sock = SKIP_BACK(sock, n, HEAD(sock_list));
 
-	  while (current_sock)
-	    {
+      while (current_sock) {
 	      sock *s = current_sock;
-	      int e;
+	int i, e;
 	      int steps;
 
+        // TODO: lets come up with better data structure so that we can avoid
+        //       lookup fd for corresponding bird socket from the entire poll fds.
+	for (i = 0; i <= nfds; i++) {
+	  if (s->fd == pollfds[i].fd) {
+	    debug("DEBUG: unix socket type [%d] poll type [%d]\n", s->type,
+		  pollfds[i].revents);
 	      steps = MAX_STEPS;
-	      if ((s->type >= SK_MAGIC) && FD_ISSET(s->fd, &rd) && s->rx_hook)
-		do
-		  {
+
+	    if ((s->type >= SK_MAGIC) &&
+		pollfds[i].revents & POLLIN &&
+		s->rx_hook) {
+	      debug("DEBUG: POLLIN event from fd [%d]\n", pollfds[i].fd);
+
+	      do {
 		    steps--;
 		    e = sk_read(s);
-		    if (s != current_sock)
+		if (s != current_sock) {
+                  // TODO: still don't get why we have this. But I've seen this debug output eventually.
+		  debug("DEBUG: wtf POLLIN\n");
 		      goto next;
 		  }
-		while (e && s->rx_hook && steps);
+	      } while (e && s->rx_hook && steps);
+	    }
 
 	      steps = MAX_STEPS;
-	      if (FD_ISSET(s->fd, &wr))
-		do
-		  {
+
+	    if (pollfds[i].revents & POLLOUT) {
+	      debug("DEBUG: POLLOUT event from fd [%d]\n", pollfds[i].fd);
+
+	      do {
 		    steps--;
 		    e = sk_write(s);
-		    if (s != current_sock)
+		if (s != current_sock) {
+                  // TODO: still don't get why we have this. But I've seen this debug output eventually.
+		  debug("DEBUG: wtf POLLOUT\n");
 		      goto next;
 		  }
-		while (e && steps);
+	      } while (e && steps);
+	    }
+
+	  } else if (pollfds[i].revents & POLLNVAL) {
+            // once we allocate pollfds in dynamic manner we will free resource differently.
+            debug("DEBUG: POLLNVAL on socket [%d]. Freeing resource\n", pollfds[i].fd);
+            pollfds[i].fd = NULL;
+            pollfds[i].events = 0;
+            pollfds[i].revents = 0;
+          }
+	}
+
 	      current_sock = sk_next(s);
 	    next: ;
 	    }
@@ -1669,18 +1700,27 @@ io_loop(void)
 	  if (current_sock == NULL)
 	    current_sock = SKIP_BACK(sock, n, HEAD(sock_list));
 
-	  while (current_sock && count < MAX_RX_STEPS)
-	    {
+      while (current_sock && count < MAX_RX_STEPS) {
 	      sock *s = current_sock;
-	      int e;
+	int i, e;
 
-	      if ((s->type < SK_MAGIC) && FD_ISSET(s->fd, &rd) && s->rx_hook)
-		{
+	/* TODO: short loop */
+	for (i = 0; i <= nfds; i++) {
+	  if (s->fd == pollfds[i].fd) {
+	    debug("DEBUG: tcp/udp socket type [%d] poll event [%d]\n", s->type,
+		  pollfds[i].revents);
+	    if ((s->type < SK_MAGIC) &&
+		pollfds[i].revents & POLLIN &&
+		s->rx_hook) {
+	      sk_dump_all();
+	      debug("DEBUG: tcp/udp POLLIN event from fd [%d]\n", pollfds[i].fd);
 		  count++;
 		  e = sk_read(s);
 		  if (s != current_sock)
 		      goto next2;
 		}
+	  }
+	}
 	      current_sock = sk_next(s);
 	    next2: ;
 	    }

아고 bird-1.3.9.tar.gz이었음 죄송