Patrick McHardy wrote:

Joonwoo Park wrote:

quoted

Patrick,
Do you mind inserting a new flag to indicate vlan_tag is just for
af_packet or not into sk_buff?
I think if flag exists, we can pass vlan_tag to af_packet with just
modifications of vlan_get_tag(), vlan_put_tag() and
vlan_tag_present(),
without patching for invalidation.

I've considered this, but that would mean that we report
the vlan tag for all devices layered on top, which seems
wrong.

For now I think I'm going to temporarily do something like
that so I can continue integrating the remaining patches
and fix it properly later.

The entire idea didn't work out too well. Your original patch
passed packets for unknown VLANs to netif_receive_skb again,
but since they don't have a VLAN header, they are treated like
a normal non-VLAN packet and might end up in a protocol handler
instead of getting dropped. It would also have left one
inconsistency, for locally configured VLANs the packet is still
not visible on the lower device.

This patch (on top of my local tree with other changes, I'll
post the entire series soon) takes a different approach.
__vlan_hwaccel_rx() is supposed to behave similar to
netif_receive_skb(), so the easiest fix is to make it behave
more like netif_receive_skb() and deliver packets to ptype_all
handlers manually.