Re: OOPS: Multipath routing 2.4.17

[Julian Anastasov]

	Hello,

Kain wrote:

> impossible 888
> divide error: 0000

> > > EIP; c024c5ea <fib_select_multipath+5a/a0>   <=====
> Trace; c02232e8 <ip_route_output_slow+318/670>

	There is no write locking in fib_select_multipath,
combined with high rate of route resolutions and ... boom,
fi->fib_power is 0:

w = jiffies % fi->fib_power;

	What about a different algorithm to apply weighted
round robin (idea mostly from LVS), something like
this code (entirely not tested) where fi->fib_power is not used
and where fib_sync_up and fib_sync_down don't need to play
with nh_power on nh_flags change:

void fib_select_multipath(const struct rt_key *key, struct fib_result *res)
{
	struct fib_info *fi = res->fi;
	int w = -1, sel = 0;

	write_lock(&fib_info_lock);

	repeat:

	change_nexthops(fi) {
		if (nh->nh_power > w && !(nh->nh_flags&RTNH_F_DEAD)) {
			w = nh->nh_power;
			sel = nhsel;
		}
	} endfor_nexthops(fi);
	if (w > 0) {
		fi->fib_nh[sel].nh_power--;
		write_unlock(&fib_info_lock);
		res->nh_sel = sel;
		return;
	}

	if (!w) {
		change_nexthops(fi) {
			if (!(nh->nh_flags&RTNH_F_DEAD)) {
				nh->nh_power = nh->nh_weight;
			}
		} endfor_nexthops(fi);
		w = -1;
		goto repeat;
	}

	write_unlock(&fib_info_lock);

#if 1
	printk(KERN_CRIT "impossible 888\n");
#endif
	return;
}

Regards

--
Julian Anastasov <[e-mail]>