On Wed, 2016-10-26 at 18:53 +0200, Abdelrhman Ahmed wrote:
> I think it's at the right place as the current one is a little different from
> the
> commit e1f165032c8bade3a6bdf546f8faf61fda4dd01c.
>
> In the next lines, skb_push is called after copying the hardware header and
> there
> is no change to the data pointer inside the retry loop. We only need to reset
> before this loop.
>
> __skb_pull(skb, skb_network_offset(skb));
>
> do {
> seq = read_seqbegin(&hh->hh_lock);
> hh_len = hh->hh_len;
> if (likely(hh_len <= HH_DATA_MOD)) {
> /* this is inlined by gcc */
> memcpy(skb->data - HH_DATA_MOD, hh->hh_data, HH_DATA_MOD);
> } else {
> int hh_alen = HH_DATA_ALIGN(hh_len);
>
> memcpy(skb->data - hh_alen, hh->hh_data, hh_alen);
> }
> } while (read_seqretry(&hh->hh_lock, seq));
>
> skb_push(skb, hh_len);
>
> In the commit e1f165032c8bade3a6bdf546f8faf61fda4dd01c, dev_hard_header which
> calls create method for adding hardware header (uses skb_push) so it was
> required to reset to network header in the beginning of the retry loop.
Right you are, thanks for the clarification !
Back to the cause of the bug then.
If netfilter is the only case this might be needed, can't this be fixed
in netfilter ?
neigh_hh_output() is in fast path, it is quite annoying adding this
operation.
