Skip to contentSkip to conversation

neigh: support smaller retrans_time settting

PATCH Superseded Hangbin Liu 3 participants 6 messages Mar 31, 2020 – Apr 3, 2020

Subsystems (3)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 neigh: support smaller retrans_time settting

Currently, we limited the retrans_time to be greater than HZ/2. For example, if the HZ = 1000, setting retrans_time less than 500ms will not work. This makes the user unable to achieve a more accurate control for bonding arp fast failover.

But remove the sanity check would make the retransmission immediately if user set retrans_time to 0 by wrong config. So I hard code the sanity check to 10, which is 10ms if HZ is 1000 and 100ms if HZ is 100. This number should be enough for user to get more accurate neigh control.

---
 net/core/neighbour.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
net/core/neighbour.c +1 -1
index 789a73aa7bd8..245bb5bd30e2 100644
@@ -1125,7 +1125,7 @@ int __neigh_event_send(struct neighbour *neigh, struct sk_buff *skb)
 			neigh->nud_state     = NUD_INCOMPLETE;
 			neigh->updated = now;
 			next = now + max(NEIGH_VAR(neigh->parms, RETRANS_TIME),
-					 HZ/2);
+					 10);
 			neigh_add_timer(neigh, next);
 			immediate_probe = true;
 		} else {
Signature · 1 line
2.19.2
maintainer
#1 lore raw

From: Hangbin Liu <[email protected]>
Date: Tue, 31 Mar 2020 11:33:56 +0800

Quoted 5 lines

Currently, we limited the retrans_time to be greater than HZ/2. For example, if the HZ = 1000, setting retrans_time less than 500ms will not work. This makes the user unable to achieve a more accurate control for bonding arp fast failover.

Quoted 7 lines

But remove the sanity check would make the retransmission immediately if user set retrans_time to 0 by wrong config. So I hard code the sanity check to 10, which is 10ms if HZ is 1000 and 100ms if HZ is 100. This number should be enough for user to get more accurate neigh control.

It only makes sense to make the safety limit be constant (time wise), and therefore relative to HZ.

#2 lore raw

PATCH v2 1/1 neigh: support smaller retrans_time settting

In reply to Hangbin Liu

Currently, we limited the retrans_time to be greater than HZ/2. i.e. setting retrans_time less than 500ms will not work. This makes the user unable to achieve a more accurate control for bonding arp fast failover.

Update the sanity check to HZ/100, which is 10ms, to let users have more ability on the retrans_time control.

v2: use HZ instead of hard code number

---
 net/core/neighbour.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
net/core/neighbour.c +1 -1
index 5bf8d22a47ec..46a5611a9f3d 100644
@@ -1125,7 +1125,7 @@ int __neigh_event_send(struct neighbour *neigh, struct sk_buff *skb)
 			neigh->nud_state     = NUD_INCOMPLETE;
 			neigh->updated = now;
 			next = now + max(NEIGH_VAR(neigh->parms, RETRANS_TIME),
-					 HZ/2);
+					 HZ/100);
 			neigh_add_timer(neigh, next);
 			immediate_probe = true;
 		} else {
Signature · 1 line
2.19.2
#3 lore raw
Quoted 20 lines · Hangbin Liu

Currently, we limited the retrans_time to be greater than HZ/2. i.e. setting retrans_time less than 500ms will not work. This makes the user unable to achieve a more accurate control for bonding arp fast failover.

Update the sanity check to HZ/100, which is 10ms, to let users have more ability on the retrans_time control.

v2: use HZ instead of hard code number

---
 net/core/neighbour.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
net/core/neighbour.c +0 -0
index 5bf8d22a47ec..46a5611a9f3d 100644

net/core/neighbour.c int __neigh_event_send(struct neighbour *neigh, struct sk_buff *skb) · line 1131

Quoted 8 lines
net/core/neighbour.c +1 -1
@@ -1125,7 +1125,7 @@ int __neigh_event_send(struct neighbour *neigh, struct sk_buff *skb)
 			neigh->nud_state     = NUD_INCOMPLETE;
 			neigh->updated = now;
 			next = now + max(NEIGH_VAR(neigh->parms, RETRANS_TIME),
-					 HZ/2);
+					 HZ/100);
 			neigh_add_timer(neigh, next);
 			immediate_probe = true;
 		} else {

Note that IPv6 has a different limit (HZ/10)

It would be nice to have converged values.

#4 lore raw

PATCH v3 1/1 neigh: support smaller retrans_time settting

In reply to Hangbin Liu

Currently, we limited the retrans_time to be greater than HZ/2. i.e. setting retrans_time less than 500ms will not work. This makes the user unable to achieve a more accurate control for bonding arp fast failover.

Update the sanity check to HZ/100, which is 10ms, to let users have more ability on the retrans_time control.

v3: sync the behavior with IPv6 and update all the timer handler
v2: use HZ instead of hard code number

---
 net/core/neighbour.c | 10 ++++++----
 net/ipv6/addrconf.c  |  7 ++++---
 net/ipv6/ndisc.c     |  4 ++--
 3 files changed, 12 insertions(+), 9 deletions(-)
net/core/neighbour.c +6 -4
index 5bf8d22a47ec..39d37d0ef575 100644
@@ -1065,11 +1065,12 @@ static void neigh_timer_handler(struct timer_list *t)
 			neigh->updated = jiffies;
 			atomic_set(&neigh->probes, 0);
 			notify = 1;
-			next = now + NEIGH_VAR(neigh->parms, RETRANS_TIME);
+			next = now + max(NEIGH_VAR(neigh->parms, RETRANS_TIME),
+					 HZ/100);
 		}
 	} else {
 		/* NUD_PROBE|NUD_INCOMPLETE */
-		next = now + NEIGH_VAR(neigh->parms, RETRANS_TIME);
+		next = now + max(NEIGH_VAR(neigh->parms, RETRANS_TIME), HZ/100);
 	}
 
 	if ((neigh->nud_state & (NUD_INCOMPLETE | NUD_PROBE)) &&
@@ -1125,7 +1126,7 @@ int __neigh_event_send(struct neighbour *neigh, struct sk_buff *skb)
 			neigh->nud_state     = NUD_INCOMPLETE;
 			neigh->updated = now;
 			next = now + max(NEIGH_VAR(neigh->parms, RETRANS_TIME),
-					 HZ/2);
+					 HZ/100);
 			neigh_add_timer(neigh, next);
 			immediate_probe = true;
 		} else {
@@ -1427,7 +1428,8 @@ void __neigh_set_probe_once(struct neighbour *neigh)
 	neigh->nud_state = NUD_INCOMPLETE;
 	atomic_set(&neigh->probes, neigh_max_probes(neigh));
 	neigh_add_timer(neigh,
-			jiffies + NEIGH_VAR(neigh->parms, RETRANS_TIME));
+			jiffies + max(NEIGH_VAR(neigh->parms, RETRANS_TIME),
+				      HZ/100));
 }
 EXPORT_SYMBOL(__neigh_set_probe_once);
 
net/ipv6/addrconf.c +4 -3
index a11fd4d67832..89aa6da8661d 100644
@@ -1357,7 +1357,7 @@ static int ipv6_create_tempaddr(struct inet6_ifaddr *ifp,
 
 	regen_advance = idev->cnf.regen_max_retry *
 			idev->cnf.dad_transmits *
-			NEIGH_VAR(idev->nd_parms, RETRANS_TIME) / HZ;
+			max(NEIGH_VAR(idev->nd_parms, RETRANS_TIME), HZ/100) / HZ;
 
 	/* recalculate max_desync_factor each time and update
 	 * idev->desync_factor if it's larger
@@ -4117,7 +4117,8 @@ static void addrconf_dad_work(struct work_struct *w)
 
 	ifp->dad_probes--;
 	addrconf_mod_dad_work(ifp,
-			      NEIGH_VAR(ifp->idev->nd_parms, RETRANS_TIME));
+			      max(NEIGH_VAR(ifp->idev->nd_parms, RETRANS_TIME),
+				  HZ/100));
 	spin_unlock(&ifp->lock);
 	write_unlock_bh(&idev->lock);
 
@@ -4523,7 +4524,7 @@ static void addrconf_verify_rtnl(void)
 				   !(ifp->flags&IFA_F_TENTATIVE)) {
 				unsigned long regen_advance = ifp->idev->cnf.regen_max_retry *
 					ifp->idev->cnf.dad_transmits *
-					NEIGH_VAR(ifp->idev->nd_parms, RETRANS_TIME) / HZ;
+					max(NEIGH_VAR(ifp->idev->nd_parms, RETRANS_TIME), HZ/100) / HZ;
 
 				if (age >= ifp->prefered_lft - regen_advance) {
 					struct inet6_ifaddr *ifpub = ifp->ifpub;
net/ipv6/ndisc.c +2 -2
index 6ffa153e5166..1ecd4e9b0bdf 100644
@@ -1359,8 +1359,8 @@ static void ndisc_router_discovery(struct sk_buff *skb)
 
 		if (rtime && rtime/1000 < MAX_SCHEDULE_TIMEOUT/HZ) {
 			rtime = (rtime*HZ)/1000;
-			if (rtime < HZ/10)
-				rtime = HZ/10;
+			if (rtime < HZ/100)
+				rtime = HZ/100;
 			NEIGH_VAR_SET(in6_dev->nd_parms, RETRANS_TIME, rtime);
 			in6_dev->tstamp = jiffies;
 			send_ifinfo_notify = true;
Signature · 1 line
2.19.2
maintainer Applied
#5 lore raw

From: Hangbin Liu <[email protected]>
Date: Wed, 1 Apr 2020 14:46:20 +0800

Quoted 11 lines

Currently, we limited the retrans_time to be greater than HZ/2. i.e. setting retrans_time less than 500ms will not work. This makes the user unable to achieve a more accurate control for bonding arp fast failover.

Update the sanity check to HZ/100, which is 10ms, to let users have more ability on the retrans_time control.

v3: sync the behavior with IPv6 and update all the timer handler
v2: use HZ instead of hard code number

Applied, I like the consistency between ipv4 and ipv6 now.

Thanks.