On Thu, Apr 13, 2017 at 11:38:10AM -0400, David Miller wrote: > From: Johannes Berg <johan...@sipsolutions.net> > Date: Thu, 13 Apr 2017 08:10:56 +0200 > > > On Wed, 2017-04-12 at 21:20 -0700, Alexei Starovoitov wrote: > > > >> > + if (skb_linearize(skb)) > >> > + goto do_drop; > >> > >> when we discussed supporting jumbo frames in XDP, the idea > >> was that the program would need to look at first 3+k bytes only > >> and the rest of the packet will be in non-contiguous pages. > >> If we do that, it means that XDP program would have to assume > >> that the packet is more than [data, data_end] and this range > >> only covers linear part. > >> If that's the future, we don't need to linearize the skb here > >> and can let the program access headlen only. > > > > I'm not sure how you think that would work - at least with our (wifi) > > driver, the headlen should be maybe ETH_HLEN or so at this point. We'd > > let the program know that it can only look at so much, but then the > > program can't do anything at all with those frames. At some point then > > we go back to bpf_skb_load_bytes() being necessary in one form or > > another, no? > > Agreed, this is completely unusable. Some wired ethernet drivers do the > same exact thing.
ahh. i thought all drivers do at least copy-break (256) bytes or copy get_headlen or build_skb the whole thing. Since wireless does eth_hlen, then yeah, skb_linearize() is the only way. I guess any driver that would care about XDP performance would either implement in-driver XDP or make sure that skb_linearize() doesn't happen in generic XDP by doing build_skb() with the whole packet. The driver can be smart and avoid doing copy-break if generic XDP is on.
