| Message ID | 20230307052254.198305-1-edumazet@google.com |
|---|---|
| State | Accepted |
| Delegated to: | Pablo Neira |
| Headers | show |
| Series | [net] netfilter: conntrack: | expand |
On Tue, Mar 7, 2023 at 6:22 AM Eric Dumazet <edumazet@google.com> wrote: > > Customers using GKE 1.25 and 1.26 are facing conntrack issues > root caused to commit c9c3b6811f74 ("netfilter: conntrack: make > max chain length random"). > > Even if we assume Uniform Hashing, a bucket often reachs 8 chained > items while the load factor of the hash table is smaller than 0.5 > Sorry for the messed patch title. This should have been: netfilter: conntrack: adopt safer max chain length Florian, Pablo, let me know if you want me to send a v2, thanks !
On Tue, Mar 07, 2023 at 06:27:37AM +0100, Eric Dumazet wrote: > On Tue, Mar 7, 2023 at 6:22 AM Eric Dumazet <edumazet@google.com> wrote: > > > > Customers using GKE 1.25 and 1.26 are facing conntrack issues > > root caused to commit c9c3b6811f74 ("netfilter: conntrack: make > > max chain length random"). > > > > Even if we assume Uniform Hashing, a bucket often reachs 8 chained > > items while the load factor of the hash table is smaller than 0.5 > > > > Sorry for the messed patch title. > > This should have been: > > netfilter: conntrack: adopt safer max chain length > > Florian, Pablo, let me know if you want me to send a v2, thanks ! Applied, thanks
diff --git a/net/netfilter/nf_conntrack_core.c b/net/netfilter/nf_conntrack_core.c index 7250082e7de56c77298b0d3b62c4f0dff95b77cc..c6a6a6099b4e2200951367ccd90c12c0ac800a6a 100644 --- a/net/netfilter/nf_conntrack_core.c +++ b/net/netfilter/nf_conntrack_core.c @@ -96,8 +96,8 @@ static DEFINE_MUTEX(nf_conntrack_mutex); #define GC_SCAN_MAX_DURATION msecs_to_jiffies(10) #define GC_SCAN_EXPIRED_MAX (64000u / HZ) -#define MIN_CHAINLEN 8u -#define MAX_CHAINLEN (32u - MIN_CHAINLEN) +#define MIN_CHAINLEN 50u +#define MAX_CHAINLEN (80u - MIN_CHAINLEN) static struct conntrack_gc_work conntrack_gc_work;
Customers using GKE 1.25 and 1.26 are facing conntrack issues root caused to commit c9c3b6811f74 ("netfilter: conntrack: make max chain length random"). Even if we assume Uniform Hashing, a bucket often reachs 8 chained items while the load factor of the hash table is smaller than 0.5 With a limit of 16, we reach load factors of 3. With a limit of 32, we reach load factors of 11. With a limit of 40, we reach load factors of 15. With a limit of 50, we reach load factors of 24. This patch changes MIN_CHAINLEN to 50, to minimize risks. Ideally, we could in the future add a cushion based on expected load factor (2 * nf_conntrack_max / nf_conntrack_buckets), because some setups might expect unusual values. Fixes: c9c3b6811f74 ("netfilter: conntrack: make max chain length random") Signed-off-by: Eric Dumazet <edumazet@google.com> --- net/netfilter/nf_conntrack_core.c | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-)