Message ID | 20140222104419.6b4daa44@vostro |
---|---|
State | RFC, archived |
Delegated to: | David Miller |
Headers | show |
From: Timo Teras <timo.teras@iki.fi> Date: Sat, 22 Feb 2014 10:44:19 +0200 > When a stale or delayed neigh entry is being re-validated the entry > goes to NUD_PROBE state. At the moment only unicast probes are sent. > This is basically because neigh_max_probes() limits the probe amount so. > > Now, opennhrp intentionally configures UCAST_PROBES and MCAST_PROBES to > zero and APP_PROBES to something meaningful. The idea is that opennhrp > replaces arp completely with NHRP implemented in userland. > > Due to this it seems there is a very small time window, when the > NUD_PROBE times out and the neighbour entry gets invalidated, and > packets get lost. > > To remedy this, I would like to have these NUD_PROBE validations sent > via netlink too. > > First choice is to change to just use both unicast and application > probes: So basically, the generic and protocol-specific neighbour code work together to implement a cascading priority based probing scheme using a per-neigh counter and three limits. It seems that neigh->probes is zero when we enter the probing state. On each solicit we increment neigh->probes. Then we have this very funny logic in the protocol specific implementations of solicit: probes = atomic_read(&neigh->probes); probes -= UCAST_PROBES; if (probes < 0) { send unicast probe; } else { probes -= APP_PROBES; if (probes < 0) { trigger application based probe via netlink } else { send multicast probe; } } As an example, for ipv4 arp, UCAST_PROBES defaults to 3 and APP_PROBES defaults (as you say) to zero. If neigh_max_probes() evaluates to UCAST_PROBES, we'll do 3 unicast probes then fail the neigh, for example. I took a look at iproute2's arpd, and I suggest you take a gander over there as well. If given the '-a N' option it will set app_probes to N and send it's own ARP requests out in certain situations. It might depend upon the current behavior of neigh_max_probes(). -- To unsubscribe from this list: send the line "unsubscribe netdev" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On Tue, 25 Feb 2014 18:18:44 -0500 (EST) David Miller <davem@davemloft.net> wrote: > From: Timo Teras <timo.teras@iki.fi> > Date: Sat, 22 Feb 2014 10:44:19 +0200 > > > When a stale or delayed neigh entry is being re-validated the entry > > goes to NUD_PROBE state. At the moment only unicast probes are sent. > > This is basically because neigh_max_probes() limits the probe > > amount so. > > > > Now, opennhrp intentionally configures UCAST_PROBES and > > MCAST_PROBES to zero and APP_PROBES to something meaningful. The > > idea is that opennhrp replaces arp completely with NHRP implemented > > in userland. > > > > Due to this it seems there is a very small time window, when the > > NUD_PROBE times out and the neighbour entry gets invalidated, and > > packets get lost. > > > > To remedy this, I would like to have these NUD_PROBE validations > > sent via netlink too. > > > > First choice is to change to just use both unicast and application > > probes: > > So basically, the generic and protocol-specific neighbour code work > together to implement a cascading priority based probing scheme using > a per-neigh counter and three limits. > > It seems that neigh->probes is zero when we enter the probing state. > > On each solicit we increment neigh->probes. > > Then we have this very funny logic in the protocol specific > implementations of solicit: > > probes = atomic_read(&neigh->probes); > > probes -= UCAST_PROBES; > if (probes < 0) { > send unicast probe; > } else { > probes -= APP_PROBES; > if (probes < 0) { > trigger application based probe via netlink > } else { > send multicast probe; > } > } > > As an example, for ipv4 arp, UCAST_PROBES defaults to 3 and APP_PROBES > defaults (as you say) to zero. Right, the idea is that on NUD_INVALID it does sequence of: UCAST_PROBES protocol specific unicasts APP_PROBE netlink requeusts MCAST_PROBES protocol specific multicasts And in NUD_PROBE currently only the unicast probes. > If neigh_max_probes() evaluates to UCAST_PROBES, we'll do 3 unicast > probes then fail the neigh, for example. > > I took a look at iproute2's arpd, and I suggest you take a gander > over there as well. > > If given the '-a N' option it will set app_probes to N and send it's > own ARP requests out in certain situations. > > It might depend upon the current behavior of neigh_max_probes(). In fact, it seems that the arpd is expecting to get app probe in NUD_PROBE. There's code like: if (ndm->ndm_state&NUD_PROBE) { /* If we get this, kernel still has some valid * address, but unicast probing failed and host * is either dead or changed its mac address. * Kernel is going to initiate broadcast resolution. * OK, we invalidate our information as well. */ if (dbdat.data && !IS_NEG(dbdat.data)) stats.app_neg++; dbase->del(dbase, &dbkey, 0); On the other hand opennhrp, needs the NUD_PROBE app queries to confirm the existing entries, as opennhrp intentionally turns UCAST_PROBES to zero. So I'll submit the first variant soon, where app probes are added to be done for NUD_PROBE. - Timo -- To unsubscribe from this list: send the line "unsubscribe netdev" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
diff --git a/net/core/neighbour.c b/net/core/neighbour.c index b9e9e0d..36d3f8c 100644 --- a/net/core/neighbour.c +++ b/net/core/neighbour.c @@ -836,10 +836,10 @@ out: static __inline__ int neigh_max_probes(struct neighbour *n) { struct neigh_parms *p = n->parms; - return (n->nud_state & NUD_PROBE) ? - NEIGH_VAR(p, UCAST_PROBES) : - NEIGH_VAR(p, UCAST_PROBES) + NEIGH_VAR(p, APP_PROBES) + - NEIGH_VAR(p, MCAST_PROBES); + int max_probes = NEIGH_VAR(p, UCAST_PROBES) + NEIGH_VAR(p, APP_PROBES); + if (!(n->nud_state & NUD_PROBE)) + max_probes += NEIGH_VAR(p, MCAST_PROBES); + return max_probes; } static void neigh_invalidate(struct neighbour *neigh) On default configuration there is no behaviour change, as APP_PROBES defaults zero. I'm not sure if other ARPD programs than opennhrp are currently commonly used. If that feels risky, alternative would be: diff --git a/net/core/neighbour.c b/net/core/neighbour.c index b9e9e0d..8bb320b 100644 --- a/net/core/neighbour.c +++ b/net/core/neighbour.c @@ -836,9 +836,11 @@ out: static __inline__ int neigh_max_probes(struct neighbour *n) { struct neigh_parms *p = n->parms; - return (n->nud_state & NUD_PROBE) ? - NEIGH_VAR(p, UCAST_PROBES) : - NEIGH_VAR(p, UCAST_PROBES) + NEIGH_VAR(p, APP_PROBES) + + + if (n->nud_state & NUD_PROBE) + return NEIGH_VAR(p, UCAST_PROBES) ? : NEIGH_VAR(p, APP_PROBES); + + return NEIGH_VAR(p, UCAST_PROBES) + NEIGH_VAR(p, APP_PROBES) + NEIGH_VAR(p, MCAST_PROBES); }