> > Which buffer? skb->data..skb->tail? And isn't that, in mac80211's case,
quoted text
> > at least currently all the buffer anyway?
>=20
> Right.
>=20
> For TSO or cases using frag lists of frag page vectors, it gets
> more complicated.

Yeah, but at least right now we don't accept any of those. That might
change, but even then we'll only care about the headroom (and maybe for
TKIP the crypto MIC that's at the end, assuming the hardware does the
encryption which it usually does), but we can stick that elsewhere and
just DMA it on.

quoted text
> Basically, skb_header_cloned()=3D=3Dfalse means that you can modify
> anything covered by skb->data..skb->tail

Ok.

quoted text
> > Interesting point. I'll have to see when socket filters are run,
> > wpa_supplicant could have a tap open I think. Maybe that's why I'm
> > seeing so many cloned packets. I think I'll stacktrace skb_clone and
> > print that out.
>=20
> Note that we can even optimize that case even further if we really
> have to.
>=20
> But try to see why the filter used by WPA Supplicant is not effective.
> If the filter fails to match, the packet is free'd immediately using
> kfree_skb() which should undo clone'age.

Right. I just checked, and realised I remembered wrongly, wpa_supplicant
actually binds the socket to the ethertype it wants so
dev_queue_xmit_nit() shouldn't be called. I'll have to check why I got
so many cloned packets, best see if it still happens with the
header_cloned() change.

johannes