On Sat, Sep 26, 2020 at 08:30:58PM +0300, Vladimir Oltean wrote:
> For all DSA formats that don't use tail tags, it looks like behind the
> obscure number crunching they're all doing the same thing: locating the
> real EtherType behind the DSA tag. Nonetheless, this is not immediately
> obvious, so create a generic helper for those DSA taggers that put the
> header before the EtherType.
This is interesting, and opens up the possibility of an optimisation on
the hot path.
net/core/flow_dissector.c:
bool __skb_flow_dissect(const struct net *net,
const struct sk_buff *skb,
struct flow_dissector *flow_dissector,
void *target_container,
void *data, __be16 proto, int nhoff, int hlen,
unsigned int flags)
{
...
#if IS_ENABLED(CONFIG_NET_DSA)
if (unlikely(skb->dev && netdev_uses_dsa(skb->dev) &&
proto == htons(ETH_P_XDSA))) {
const struct dsa_device_ops *ops;
int offset = 0;
ops = skb->dev->dsa_ptr->tag_ops;
if (ops->flow_dissect &&
!ops->flow_dissect(skb, &proto, &offset)) {
hlen -= offset;
nhoff += offset;
}
}
#endif
Given spectra and meltdown etc, jumping through a pointer is expensive
and we try to avoid it on the hot path. Given most of the taggers are
going to use the generic version, maybe add a test here, is
ops->flow_dissect the generic version, and if so, call it directly,
rather than go through the pointer. Or only set ops->flow_dissect if
the generic version cannot be used.
Andrew
