Move some global variables in more appropriate places.

Add XXX comments to mark places which need to be taken care of
if we want to remove this part of the kernel from Giant.

Add a comment on a potential performance problem with ip_forward()

1 files changed, 28 insertions, 3 deletions

diff --git a/sys/netinet/ip_input.c b/sys/netinet/ip_input.c
index 2780622..0778146 100644
--- a/sys/netinet/ip_input.c
+++ b/sys/netinet/ip_input.c
@@ -87,8 +87,6 @@
 #endif
 
 int rsvp_on = 0;
-static int ip_rsvp_on;
-struct socket *ip_rsvpd;
 
 int	ipforwarding = 0;
 SYSCTL_INT(_net_inet_ip, IPCTL_FORWARDING, forwarding, CTLFLAG_RW,
@@ -193,6 +191,12 @@ ip_dn_io_t *ip_dn_io_ptr;
 
 
 /*
+ * XXX this is ugly -- the following two global variables are
+ * used to store packet state while it travels through the stack.
+ * Note that the code even makes assumptions on the size and
+ * alignment of fields inside struct ip_srcrt so e.g. adding some
+ * fields will break the code. This needs to be fixed.
+ *
  * We need to save the IP options in case a protocol wants to respond
  * to an incoming packet over the same route if the packet got here
  * using IP source routing.  This allows connection establishment and
@@ -256,7 +260,11 @@ ip_init()
 	register_netisr(NETISR_IP, ipintr);
 }
 
-static struct	sockaddr_in ipaddr = { sizeof(ipaddr), AF_INET };
+/*
+ * XXX watch out this one. It is perhaps used as a cache for
+ * the most recently used route ? it is cleared in in_addroute()
+ * when a new route is successfully created.
+ */
 struct	route ipforward_rt;
 
 /*
@@ -1167,6 +1175,7 @@ ip_dooptions(struct mbuf *m, int pass, struct sockaddr_in *next_hop)
 	int opt, optlen, cnt, off, code, type = ICMP_PARAMPROB, forward = 0;
 	struct in_addr *sin, dst;
 	n_time ntime;
+	struct	sockaddr_in ipaddr = { sizeof(ipaddr), AF_INET };
 
 	dst = ip->ip_dst;
 	cp = (u_char *)(ip + 1);
@@ -1646,6 +1655,13 @@ ip_forward(struct mbuf *m, int srcrt, struct sockaddr_in *next_hop)
 	 * Save the IP header and at most 8 bytes of the payload,
 	 * in case we need to generate an ICMP message to the src.
 	 *
+	 * XXX this can be optimized a lot by saving the data in a local
+	 * buffer on the stack (72 bytes at most), and only allocating the
+	 * mbuf if really necessary. The vast majority of the packets
+	 * are forwarded without having to send an ICMP back (either
+	 * because unnecessary, or because rate limited), so we are
+	 * really we are wasting a lot of work here.
+	 *
 	 * We don't use m_copy() because it might return a reference
 	 * to a shared cluster. Both this function and ip_output()
 	 * assume exclusive access to the IP header in `m', so any
@@ -1884,6 +1900,15 @@ makedummy:
 	}
 }
 
+/*
+ * XXX these routines are called from the upper part of the kernel.
+ * They need to be locked when we remove Giant.
+ *
+ * They could also be moved to ip_mroute.c, since all the RSVP
+ *  handling is done there already.
+ */
+static int ip_rsvp_on;
+struct socket *ip_rsvpd;
 int
 ip_rsvp_init(struct socket *so)
 {