Skip to contentSkip to conversation

net/ipv6: always honour route mtu during forwarding

PATCH In review Maciej Żenczykowski 3 participants 5 messages Oct 8, 2020

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

PATCH 1/2 net/ipv6: always honour route mtu during forwarding

From: Maciej Żenczykowski <[email protected]>

This matches the new ipv4 behaviour as of commit:
commit 02a1b175b0e92d9e0fa5df3957ade8d733ceb6a0
Author: Maciej Żenczykowski <[email protected]>
Date: Wed Sep 23 13:18:15 2020 -0700

net/ipv4: always honour route mtu during forwarding

The reasoning is similar: There doesn't seem to be any reason why you would want to ignore route mtu.

There are two potential sources of ipv6 route mtu:
- manually configured by NET_ADMIN, since you configured a route mtu explicitly you probably know best...
- derived from mtu information from RA messages, but this is the network telling you what will work, again presumably whatever network admin configured the RA content knows best what the network conditions are.

One could argue that RAs can be spoofed, but if we get spoofed RAs we're *already* screwed, and erroneous mtu information is less dangerous then the erroneous routes themselves... (The proper place to do RA filtering is in the switch/router)

Additionally, a reduction from 1500 to 1280 (min ipv6 mtu) is not very noticable on performance (especially with gro/gso/tso), while packets getting lost (due to rx buffer overruns) or generating icmpv6 packet too big errors and needing to be retransmitted is very noticable (guaranteed impact of full rtt)

It is pretty common to have a higher device mtu to allow receiving large (jumbo) frames, while having some routes via that interface (potentially including the default route to the internet) specify a lower mtu.

There might also be use cases around xfrm/ipsec/tunnels. Especially for something like sit/6to4/6rd, where you may have one sit device, but traffic through it will flow over different underlying paths and thus is per subnet and not per device.

(Note that this function does not honour pmtu, which can be spoofed via icmpv6 messages, but see also ip6_mtu_from_fib6() which honours pmtu for ipv6 'locked mtu' routes)

Cc Eric Dumazet <[email protected]>
Cc Willem de Bruijn <[email protected]>
Cc Lorenzo Colitti <[email protected]>
Cc Sunmeet Gill (Sunny) <[email protected]>
Cc Vinay Paradkar <[email protected]>
Cc Tyler Wear <[email protected]>
Cc David Ahern <[email protected]>
---
 include/net/ip6_route.h | 14 ++++----------
 1 file changed, 4 insertions(+), 10 deletions(-)
include/net/ip6_route.h +4 -10
index 2a5277758379..598415743f46 100644
@@ -311,19 +311,13 @@ static inline bool rt6_duplicate_nexthop(struct fib6_info *a, struct fib6_info *
 static inline unsigned int ip6_dst_mtu_forward(const struct dst_entry *dst)
 {
 	struct inet6_dev *idev;
-	unsigned int mtu;
+	unsigned int mtu = dst_metric_raw(dst, RTAX_MTU);
+	if (mtu)
+		return mtu;
 
-	if (dst_metric_locked(dst, RTAX_MTU)) {
-		mtu = dst_metric_raw(dst, RTAX_MTU);
-		if (mtu)
-			return mtu;
-	}
-
-	mtu = IPV6_MIN_MTU;
 	rcu_read_lock();
 	idev = __in6_dev_get(dst->dev);
-	if (idev)
-		mtu = idev->cnf.mtu6;
+	mtu = idev ? idev->cnf.mtu6 : IPV6_MIN_MTU;
 	rcu_read_unlock();
 
 	return mtu;
Signature · 1 line
2.28.0.806.g8561365e88-goog
#1 lore raw

PATCH 2/2 net/ipv6: ensure ip6_dst_mtu_forward() returns at least IPV6_MIN_MTU

From: Maciej Żenczykowski <[email protected]>

This is basically just a refactor.

But it does affect (a presumably buggy) call site in:
net/netfilter/nf_flow_table_core.c flow_offload_fill_route()

---
 include/net/ip6_route.h | 4 ++--
 net/ipv6/ip6_output.c   | 2 --
 2 files changed, 2 insertions(+), 4 deletions(-)
include/net/ip6_route.h +2 -2
index 598415743f46..25c113dd88ea 100644
@@ -313,14 +313,14 @@ static inline unsigned int ip6_dst_mtu_forward(const struct dst_entry *dst)
 	struct inet6_dev *idev;
 	unsigned int mtu = dst_metric_raw(dst, RTAX_MTU);
 	if (mtu)
-		return mtu;
+		return max(mtu, (unsigned)IPV6_MIN_MTU);
 
 	rcu_read_lock();
 	idev = __in6_dev_get(dst->dev);
 	mtu = idev ? idev->cnf.mtu6 : IPV6_MIN_MTU;
 	rcu_read_unlock();
 
-	return mtu;
+	return max(mtu, (unsigned)IPV6_MIN_MTU);
 }
 
 u32 ip6_mtu_from_fib6(const struct fib6_result *res,
net/ipv6/ip6_output.c +0 -2
index c78e67d7747f..bc85f92adaf9 100644
@@ -540,8 +540,6 @@ int ip6_forward(struct sk_buff *skb)
 	}
 
 	mtu = ip6_dst_mtu_forward(dst);
-	if (mtu < IPV6_MIN_MTU)
-		mtu = IPV6_MIN_MTU;
 
 	if (ip6_pkt_too_big(skb, mtu)) {
 		/* Again, force OUTPUT device used as source address */
Signature · 1 line
2.28.0.806.g8561365e88-goog
#2 lore raw
In reply to Maciej Żenczykowski

include/net/ip6_route.h static inline bool rt6_duplicate_nexthop(struct fib6_info *a, struct fib6_info * · line 316

Quoted 12 lines · Maciej Żenczykowski
include/net/ip6_route.h +3 -1
index 2a5277758379..598415743f46 100644
@@ -311,19 +311,13 @@ static inline bool rt6_duplicate_nexthop(struct fib6_info *a, struct fib6_info *
 static inline unsigned int ip6_dst_mtu_forward(const struct dst_entry *dst)
 {
        struct inet6_dev *idev;
-       unsigned int mtu;
+       unsigned int mtu = dst_metric_raw(dst, RTAX_MTU);
+       if (mtu)
+               return mtu;

What should happen here if mtu is less than idev->cnf.mtu6? Should the code pick the minimum? If not: will picking the higher value work, or will the packet be dropped? I suppose we already have this problem today if the administrator configures a route with a locked MTU.

#3 lore raw
Quoted 11 lines
Quoted 9 lines · Maciej Żenczykowski
include/net/ip6_route.h +0 -1
index 2a5277758379..598415743f46 100644
@@ -311,19 +311,13 @@ static inline bool rt6_duplicate_nexthop(struct fib6_info *a, struct fib6_info *
 static inline unsigned int ip6_dst_mtu_forward(const struct dst_entry *dst)
 {
        struct inet6_dev *idev;
-       unsigned int mtu;
Quoted 8 lines

include/net/ip6_route.h static inline bool rt6_duplicate_nexthop(struct fib6_info *a, struct fib6_info * · line 316

Quoted 3 lines
include/net/ip6_route.h +3 -0
+       unsigned int mtu = dst_metric_raw(dst, RTAX_MTU);
+       if (mtu)
+               return mtu;

What should happen here if mtu is less than idev->cnf.mtu6? Should the code pick the minimum? If not: will picking the higher value work, or will the packet be dropped? I suppose we already have this problem today if the administrator configures a route with a locked MTU.

This feels like a misconfiguration of some sort (ie maybe should be denied at route config time), but honestly it could maybe potentially be useful: for example an ipv6 route out an ipv4 only interface with ebpf doing translation should/could actually be higher (by 20 or even
28) then device mtu. (Note: this is not the Android case, as we translate on ingress, not egress, but a setup could be created that would work, especially if there was no nat44 in the picture and the ipv6 dst address would directly select the end host)

And either way... it's the same for v4, and it already behave this way in other places.

#4 lore raw
In reply to Maciej Żenczykowski
Quoted 8 lines · Maciej Żenczykowski

From: Maciej Żenczykowski <[email protected]>

This matches the new ipv4 behaviour as of commit:
commit 02a1b175b0e92d9e0fa5df3957ade8d733ceb6a0
Author: Maciej Żenczykowski <[email protected]>
Date: Wed Sep 23 13:18:15 2020 -0700

net/ipv4: always honour route mtu during forwarding

just summarize that as:
commit 02a1b175b0e9 ("net/ipv4: always honour route mtu during forwarding")

include/net/ip6_route.h static inline bool rt6_duplicate_nexthop(struct fib6_info *a, struct fib6_info * · line 314

Quoted 10 lines
include/net/ip6_route.h +1 -1
index 2a5277758379..598415743f46 100644
@@ -311,19 +311,13 @@ static inline bool rt6_duplicate_nexthop(struct fib6_info *a, struct fib6_info *
 static inline unsigned int ip6_dst_mtu_forward(const struct dst_entry *dst)
 {
 	struct inet6_dev *idev;
-	unsigned int mtu;
+	unsigned int mtu = dst_metric_raw(dst, RTAX_MTU);

newline here for readability

Quoted 9 lines
include/net/ip6_route.h +2 -5
+	if (mtu)
+		return mtu;
 
-	if (dst_metric_locked(dst, RTAX_MTU)) {
-		mtu = dst_metric_raw(dst, RTAX_MTU);
-		if (mtu)
-			return mtu;
-	}

include/net/ip6_route.h static inline bool rt6_duplicate_nexthop(struct fib6_info *a, struct fib6_info * · line 323

Quoted 10 lines
include/net/ip6_route.h +1 -4
-
-	mtu = IPV6_MIN_MTU;
 	rcu_read_lock();
 	idev = __in6_dev_get(dst->dev);
-	if (idev)
-		mtu = idev->cnf.mtu6;
+	mtu = idev ? idev->cnf.mtu6 : IPV6_MIN_MTU;
 	rcu_read_unlock();
 
 	return mtu;

besides the nit comments, the change looks fine to me. Please add test cases to tools/testing/selftests/net/pmtu.sh for this change.