diff options
author | Daniel Borkmann <[email protected]> | 2022-01-06 01:46:06 +0100 |
---|---|---|
committer | David S. Miller <[email protected]> | 2022-01-06 13:49:54 +0000 |
commit | 710ad98c363a66a0cd8526465426c5c5f8377ee0 (patch) | |
tree | af6727f8d4f07e9495f6a606ad338b99f7df58b9 | |
parent | c288bc0db2d1938691ef283ce61ae6122e562bc3 (diff) |
veth: Do not record rx queue hint in veth_xmit
Laurent reported that they have seen a significant amount of TCP retransmissions
at high throughput from applications residing in network namespaces talking to
the outside world via veths. The drops were seen on the qdisc layer (fq_codel,
as per systemd default) of the phys device such as ena or virtio_net due to all
traffic hitting a _single_ TX queue _despite_ multi-queue device. (Note that the
setup was _not_ using XDP on veths as the issue is generic.)
More specifically, after edbea9220251 ("veth: Store queue_mapping independently
of XDP prog presence") which made it all the way back to v4.19.184+,
skb_record_rx_queue() would set skb->queue_mapping to 1 (given 1 RX and 1 TX
queue by default for veths) instead of leaving at 0.
This is eventually retained and callbacks like ena_select_queue() will also pick
single queue via netdev_core_pick_tx()'s ndo_select_queue() once all the traffic
is forwarded to that device via upper stack or other means. Similarly, for others
not implementing ndo_select_queue() if XPS is disabled, netdev_pick_tx() might
call into the skb_tx_hash() and check for prior skb_rx_queue_recorded() as well.
In general, it is a _bad_ idea for virtual devices like veth to mess around with
queue selection [by default]. Given dev->real_num_tx_queues is by default 1,
the skb->queue_mapping was left untouched, and so prior to edbea9220251 the
netdev_core_pick_tx() could do its job upon __dev_queue_xmit() on the phys device.
Unbreak this and restore prior behavior by removing the skb_record_rx_queue()
from veth_xmit() altogether.
If the veth peer has an XDP program attached, then it would return the first RX
queue index in xdp_md->rx_queue_index (unless configured in non-default manner).
However, this is still better than breaking the generic case.
Fixes: edbea9220251 ("veth: Store queue_mapping independently of XDP prog presence")
Fixes: 638264dc9022 ("veth: Support per queue XDP ring")
Reported-by: Laurent Bernaille <[email protected]>
Signed-off-by: Daniel Borkmann <[email protected]>
Cc: Maciej Fijalkowski <[email protected]>
Cc: Toshiaki Makita <[email protected]>
Cc: Eric Dumazet <[email protected]>
Cc: Paolo Abeni <[email protected]>
Cc: John Fastabend <[email protected]>
Cc: Willem de Bruijn <[email protected]>
Acked-by: John Fastabend <[email protected]>
Reviewed-by: Eric Dumazet <[email protected]>
Acked-by: Toshiaki Makita <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
-rw-r--r-- | drivers/net/veth.c | 1 |
1 files changed, 0 insertions, 1 deletions
diff --git a/drivers/net/veth.c b/drivers/net/veth.c index d21dd25f429e..354a963075c5 100644 --- a/drivers/net/veth.c +++ b/drivers/net/veth.c @@ -335,7 +335,6 @@ static netdev_tx_t veth_xmit(struct sk_buff *skb, struct net_device *dev) */ use_napi = rcu_access_pointer(rq->napi) && veth_skb_is_eligible_for_gro(dev, rcv, skb); - skb_record_rx_queue(skb, rxq); } skb_tx_timestamp(skb); |