diff options
| author | Christoph Paasch <cpaasch@apple.com> | 2024-03-06 19:11:17 +0100 |
|---|---|---|
| committer | Jakub Kicinski <kuba@kernel.org> | 2024-03-08 05:42:13 +0100 |
| commit | 8edbd3960150ee7dd7a2857bd9a5cf14906facc2 (patch) | |
| tree | 85d480f4f73234cf604802a01de1a3df8664b031 /net/mpls | |
| parent | net: dsa: Leverage core stats allocator (diff) | |
| download | linux-8edbd3960150ee7dd7a2857bd9a5cf14906facc2.tar.xz linux-8edbd3960150ee7dd7a2857bd9a5cf14906facc2.zip | |
mpls: Do not orphan the skb
We observed that TCP-pacing was falling back to the TCP-layer pacing
instead of utilizing sch_fq for the pacing. This causes significant
CPU-usage due to the hrtimer running on a per-TCP-connection basis.
The issue is that mpls_xmit() calls skb_orphan() and thus sets
skb->sk to NULL. Which implies that many of the goodies of TCP won't
work. Pacing falls back to TCP-layer pacing. TCP Small Queues does not
work, ...
It is safe to remove this call to skb_orphan() in mpls_xmit() as there
really is not reason for it to be there. It appears that this call to
skb_orphan comes from the very initial implementation of MPLS.
Cc: Roopa Prabhu <roopa@nvidia.com>
Reported-by: Craig Taylor <cmtaylor@apple.com>
Signed-off-by: Christoph Paasch <cpaasch@apple.com>
Link: https://lore.kernel.org/r/20240306181117.77419-1-cpaasch@apple.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Diffstat (limited to 'net/mpls')
| -rw-r--r-- | net/mpls/mpls_iptunnel.c | 2 |
1 files changed, 0 insertions, 2 deletions
diff --git a/net/mpls/mpls_iptunnel.c b/net/mpls/mpls_iptunnel.c index ef59e25dc482..8fc790f2a01b 100644 --- a/net/mpls/mpls_iptunnel.c +++ b/net/mpls/mpls_iptunnel.c @@ -55,8 +55,6 @@ static int mpls_xmit(struct sk_buff *skb) out_dev = dst->dev; net = dev_net(out_dev); - skb_orphan(skb); - if (!mpls_output_possible(out_dev) || !dst->lwtstate || skb_warn_if_lro(skb)) goto drop; |