On Fri, Mar 07, 2014 at 03:52:58PM -0500, David Miller wrote: > From: Michal Kubecek <mkubecek@xxxxxxx> > Date: Thu, 6 Mar 2014 21:06:58 +0100 > > > On Thu, Mar 06, 2014 at 02:24:25PM -0500, David Miller wrote: > >> > >> > +static void rt6_metrics_to_peer(struct rt6_info *rt) > >> > +{ > >> > + struct inet_peer *peer = rt6_has_peer(rt) ? rt6_peer_ptr(rt) : NULL; > >> > + struct dst_entry *dst = &rt->dst; > >> > + unsigned long old = dst->_metrics; > >> > + > >> > + if (!(rt->dst.flags & DST_HOST) || dst_metrics_read_only(dst)) > >> > + return; > >> > + if (peer && dst_metrics_ptr(dst) == peer->metrics) > >> > + return; > >> > + > >> > + dst->ops->cow_metrics(dst, old); > >> > + if (dst->_metrics != old) { > >> > + u32 *old_p = __DST_METRICS_PTR(old); > >> > + > >> > + memcpy(dst_metrics_ptr(dst), old_p, RTAX_MAX * sizeof(u32)); > >> > + kfree(old_p); > >> > + } > >> > +} > >> > >> Hmmm... if inet_metrics_new() is true then ->cow_metrics() will copy the > >> metrics from old to new. So you therefore shouldn't have to do the copy > >> explicitly here. > > > > If we are replacing an existing host route with metrics, e.g. > > > > ip route add fec0::1 dev eth0 rto_min 1000 > > ip route change fec0::1 dev eth0 rto_min 1500 > > > > then peer will be the existing inetpeer and inet_metrics_new() will be > > false. However, we still need to copy the new metrics from the netlink > > message over the old ones. > > > >> If inet_metrics_new() is not true, you are overwriting non-new metrics. > > > > The only problem with this is IMHO that if inet_metrics_new() is true, > > i.e. when adding a new route with new inetpeer (or old inetpeer whose > > metrics were not used before), the memcpy() is done twice, once in > > ipv6_cow_metrics() and once in rt6_metrics_to_peer(). We are copying the > > same data twice so that the result is correct but it's not efficient and > > it's not nice. > > > > The only way I can see to avoid this (except using own metrics always > > instead of those in struct inetpeer as we do for non-host routes) would > > be not to call ipv6_cow_metrics() at all and write a special function > > for this purpose in net/ipv6/route.c which would duplicate the parts of > > ipv6_cow_metrics() we really need (and add the free()). Do you think > > this is the way to go? > > Thank you for explaining all of this, I would like to think about this > some more. > > My initial suspicion is that the something about the test in cow > metrics might need to be adjusted. > > The conceptual attributes we have built for inetpeer metrics, that of > "newness" and "read-only", might not be built adequately for the task > at hand here. Today I also realized another problem with current code: if we already have inetpeer with metrics and use "ip route change" with metrics, e.g. ip route add fec0::1 dev eth0 rto_min 1000 ip route change fec0::1 dev eth0 hoplimit 10 only the metrics listed in the netlink message (hoplimit here) are modified and the rest (rto_min here) is preserved. This is inconsistent with non-host IPv6 routes and IPv4 routes where the whole set of metrics is replaced. Michal Kubecek -- To unsubscribe from this list: send the line "unsubscribe netdev" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html