[patch] tcp_tw in 2.4

[Andrew Morton <andrewm@xxxxxxxxxx>]

A couple of problems here..


* tcp_init() wants to set sysctl_tcp_max_tw_buckets to 180,000.  This
seems too high (Andi says 22 megs).  I think the patch here is more
consistent.

* tcp_twkill() can consume a huge amount of time if it has enough
connections to deal with.  When running lmbench I have observed it
killing 2,500 connections in a single pass, which means we spend 15
milliseconds in the timer handler.  This is crazy.

So I just kill a hundred and then reschedule the timer to run in a
couple of jiffies time.  The downside: this limits the tw reaping to
2,500 connections per second.  But I don't see a DoS opportunity here.



--- linux-2.4.0-test9-pre7/net/ipv4/tcp.c       Tue Sep 26 21:45:33 2000
+++ linux-akpm/net/ipv4/tcp.c   Mon Oct  2 02:44:59 2000
@@ -2438,7 +2438,7 @@
        if (order > 4) {
                sysctl_local_port_range[0] = 32768;
                sysctl_local_port_range[1] = 61000;
-               sysctl_tcp_max_tw_buckets = 180000;
+               sysctl_tcp_max_tw_buckets <<= 1;
                sysctl_tcp_max_orphans = 4096<<(order-4);
                sysctl_max_syn_backlog = 1024;
        } else if (order < 3) {
--- linux-2.4.0-test9-pre7/net/ipv4/tcp_minisocks.c     Tue Sep 26 21:45:33 2000
+++ linux-akpm/net/ipv4/tcp_minisocks.c Mon Oct  2 02:15:02 2000
@@ -434,6 +434,7 @@
 {
        struct tcp_tw_bucket *tw;
        int killed = 0;
+       int max_killed = 0;
 
        /* NOTE: compare this to previous version where lock
         * was released after detaching chain. It was racy,
@@ -447,6 +448,13 @@
                goto out;
 
        while((tw = tcp_tw_death_row[tcp_tw_death_row_slot]) != NULL) {
+
+               /* This loop takes ~6 usecs per iteration. */
+               if (killed > 100) {
+                       max_killed = 1;
+                       break;
+               }
+
                tcp_tw_death_row[tcp_tw_death_row_slot] = tw->next_death;
                tw->pprev_death = NULL;
                spin_unlock(&tw_death_lock);
@@ -457,12 +465,19 @@
                killed++;
 
                spin_lock(&tw_death_lock);
+
+       }
+
+       if (max_killed) {       /* More to do: do it soon */
+               mod_timer(&tcp_tw_timer, jiffies+2);
+       } else {
+               tcp_tw_death_row_slot =
+                       ((tcp_tw_death_row_slot + 1) & (TCP_TWKILL_SLOTS - 1));
+       
+               if ((tcp_tw_count -= killed) != 0)
+                       mod_timer(&tcp_tw_timer, jiffies+TCP_TWKILL_PERIOD);
        }
-       tcp_tw_death_row_slot =
-               ((tcp_tw_death_row_slot + 1) & (TCP_TWKILL_SLOTS - 1));
 
-       if ((tcp_tw_count -= killed) != 0)
-               mod_timer(&tcp_tw_timer, jiffies+TCP_TWKILL_PERIOD);
        net_statistics[smp_processor_id()*2].TimeWaited += killed;
 out:
        spin_unlock(&tw_death_lock);