Skip to contentSkip to conversation

netfilter: conntrack: remove flowtable early-drop test

PATCH In review Florian Westphal 2 participants 5 messages Apr 23, 2024 – Apr 24, 2024

Subsystems (2)
lore
Share & tools
lore

Florian Westphal

author maintainer
#0 lore raw

PATCH 1/1 netfilter: conntrack: remove flowtable early-drop test

Not sure why this special case exists. Early drop logic (which kicks in when conntrack table is full) should be independent of flowtable offload and only consider assured bit (i.e., two-way traffic was seen).

flowtable entries hold a reference to the conntrack entry (struct nf_conn) that has been offloaded. The conntrack use count is not decremented until after the entry is free'd.

This change therefore will not result in exceeding the conntrack table limit. It does allow early-drop of tcp flows even when they've been offloaded, but only if they have been offloaded before syn-ack was received or after at least one peer has sent a fin.

Currently 'fin' packet reception already stops offloading, so this should not impact offloading either.

---
Vlad, do you remember why you added this test?

For reference, this came in
df25455e5a48 ("netfilter: nf_conntrack: allow early drop of offloaded UDP conns") and maybe was just a 'move-it-around' from the check in early_drop_list, which would mean this was there from the
beginning. Doesn't change "i don't understand why this test
exists" though :-)

 net/netfilter/nf_conntrack_core.c | 2 --
 1 file changed, 2 deletions(-)
net/netfilter/nf_conntrack_core.c +0 -2
index c63868666bd9..43629e79067d 100644
@@ -1440,8 +1440,6 @@ static bool gc_worker_can_early_drop(const struct nf_conn *ct)
 	const struct nf_conntrack_l4proto *l4proto;
 	u8 protonum = nf_ct_protonum(ct);
 
-	if (test_bit(IPS_OFFLOAD_BIT, &ct->status) && protonum != IPPROTO_UDP)
-		return false;
 	if (!test_bit(IPS_ASSURED_BIT, &ct->status))
 		return true;
 
Signature · 1 line
2.43.2
#1 lore raw
Quoted 14 lines · Florian Westphal

Not sure why this special case exists. Early drop logic (which kicks in when conntrack table is full) should be independent of flowtable offload and only consider assured bit (i.e., two-way traffic was seen).

flowtable entries hold a reference to the conntrack entry (struct nf_conn) that has been offloaded. The conntrack use count is not decremented until after the entry is free'd.

This change therefore will not result in exceeding the conntrack table limit. It does allow early-drop of tcp flows even when they've been offloaded, but only if they have been offloaded before syn-ack was received or after at least one peer has sent a fin.

Quoted 7 lines

Currently 'fin' packet reception already stops offloading, so this should not impact offloading either.

---
Vlad, do you remember why you added this test?

I added it when I introduced UDP NEW connection offload. As far as I remember the concern was that since at the time early drop algorithm completely ignored all offloaded connections malicious user could fill the whole table by just sending a single packet per range of distinct 5 tuples and none of the resulting connections would be early dropped until they expire.

Quoted 23 lines

For reference, this came in
df25455e5a48 ("netfilter: nf_conntrack: allow early drop of offloaded UDP conns") and maybe was just a 'move-it-around' from the check in early_drop_list, which would mean this was there from the
beginning. Doesn't change "i don't understand why this test
exists" though :-)

 net/netfilter/nf_conntrack_core.c | 2 --
 1 file changed, 2 deletions(-)
net/netfilter/nf_conntrack_core.c +0 -2
index c63868666bd9..43629e79067d 100644
@@ -1440,8 +1440,6 @@ static bool gc_worker_can_early_drop(const struct nf_conn *ct)
 	const struct nf_conntrack_l4proto *l4proto;
 	u8 protonum = nf_ct_protonum(ct);
 
-	if (test_bit(IPS_OFFLOAD_BIT, &ct->status) && protonum != IPPROTO_UDP)
-		return false;
 	if (!test_bit(IPS_ASSURED_BIT, &ct->status))
 		return true;
author maintainer
#2 lore raw
Quoted 9 lines · Vlad Buslov
Quoted 2 lines

---
Vlad, do you remember why you added this test?

I added it when I introduced UDP NEW connection offload. As far as I remember the concern was that since at the time early drop algorithm completely ignored all offloaded connections malicious user could fill the whole table by just sending a single packet per range of distinct 5 tuples and none of the resulting connections would be early dropped until they expire.

Ok, so it was indeed this:

Quoted 4 lines
Quoted 4 lines

and maybe was just a 'move-it-around' from the check in early_drop_list, which would mean this was there from the beginning. Doesn't change "i don't understand why this test exists" though :-)

In this case I think this change is fine, ie. remove offload special treatment, its not needed.

Thanks for checking!

#3 lore raw
Quoted 13 lines · Florian Westphal
Quoted 9 lines · Vlad Buslov
Quoted 2 lines

---
Vlad, do you remember why you added this test?

I added it when I introduced UDP NEW connection offload. As far as I remember the concern was that since at the time early drop algorithm completely ignored all offloaded connections malicious user could fill the whole table by just sending a single packet per range of distinct 5 tuples and none of the resulting connections would be early dropped until they expire.

Ok, so it was indeed this:

Quoted 7 lines
Quoted 4 lines
Quoted 4 lines

and maybe was just a 'move-it-around' from the check in early_drop_list, which would mean this was there from the beginning. Doesn't change "i don't understand why this test exists" though :-)

In this case I think this change is fine, ie. remove offload special treatment, its not needed.

The change will also enable early dropping offloaded non-ASSURED connections for all other protocols though.

Quoted 2 lines

Thanks for checking!

author maintainer
#4 lore raw
Quoted 2 lines · Vlad Buslov

The change will also enable early dropping offloaded non-ASSURED connections for all other protocols though.

But I don't think we support that.
act_ct supports offload of non-assured (unidirectional) UDP.

But TCP needs to have seen two-way communication.
So, AFAICT the check isn't needed.