Hello,
On Wed, 30 Sep 2020, Eric Dumazet wrote:
> On 9/29/20 7:35 PM, Stephen Hemminger wrote:
> >
> >
> > then I noticed that in some cases skb->tstamp is equal to real ts whereas in
> > the regular cases where a packet pass through it's time since kernel boot. This
> > doesn't make any sense for me as this condition is satisfied constantly
> >
> > net/sched/sch_fq.c:439
> > static bool fq_packet_beyond_horizon(const struct sk_buff *skb,
> > const struct fq_sched_data *q)
> > {
> > return unlikely((s64)skb->tstamp > (s64)(q->ktime_cache + q->horizon));
> > }
> >
> > Any ideas on what it can be?
> >
> Thanks for the detailed report !
>
> I suspect ipvs or bridge code needs something similar to the fixes done in
>
> https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=de20900fbe1c4fd36de25a7a5a43223254ecf0d0
>
> https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=41d1c8839e5f8cb781cc635f12791decee8271b7
>
> The reason for that is that skb->tstamp can get a timestamp in input path,
> with a base which is not CLOCK_MONOTONIC, unfortunately.
>
> Whenever a packet is forwarded, its tstamp must be cleared.
>
> Can you try :
>
> diff --git a/net/netfilter/ipvs/ip_vs_xmit.c b/net/netfilter/ipvs/ip_vs_xmit.c
> index b00866d777fe0e9ed8018087ebc664c56f29b5c9..11e8ccdae358a89067046efa62ed40308b9e06f9 100644
> --- a/net/netfilter/ipvs/ip_vs_xmit.c
> +++ b/net/netfilter/ipvs/ip_vs_xmit.c
> @@ -952,6 +952,8 @@ ip_vs_prepare_tunneled_skb(struct sk_buff *skb, int skb_af,
>
> ip_vs_drop_early_demux_sk(skb);
>
> + skb->tstamp = 0;
> +
Should be after all skb_forward_csum() calls in ip_vs_xmit.c
> if (skb_headroom(skb) < max_headroom || skb_cloned(skb)) {
> new_skb = skb_realloc_headroom(skb, max_headroom);
> if (!new_skb)
Regards
--
Julian Anastasov <ja@xxxxxx>
