Re: [PATCH v3 15/55] ip, udp: Support MSG_SPLICE_PAGES

[David Howells]

Willem de Bruijn <willemdebruijn.kernel@xxxxxxxxx> wrote:

> > +	} else if ((flags & MSG_SPLICE_PAGES) && length) {
> > +		if (inet->hdrincl)
> > +			return -EPERM;
> > +		if (rt->dst.dev->features & NETIF_F_SG)
> > +			/* We need an empty buffer to attach stuff to */
> > +			initial_length = transhdrlen;
> 
> I still don't entirely understand what initial_length means.
> 
> More importantly, transhdrlen can be zero. If not called for UDP
> but for RAW. Or if this is a subsequent call to a packet that is
> being held with MSG_MORE.
> 
> This works fine for existing use-cases, which go to alloc_new_skb.
> Not sure how this case would be different. But the comment alludes
> that it does.

The problem is that in the non-MSG_ZEROCOPY case, __ip_append_data() assumes
that it's going to copy the data it is given and will allocate sufficient
space in the skb in advance to hold it - but I don't want to do that because I
want to splice in the pages holding the data instead.  However, I do need to
allocate space to hold the transport header.

Maybe I should change 'initial_length' to 'initial_alloc'?  It represents the
amount I think we should allocate.  Or maybe I should have a separate
allocation clause for MSG_SPLICE_PAGES?

I also wonder if __ip_append_data() really needs two places that call
getfrag().

David