Re: [PATCH bpf-next] net: xdp: introduce xdp_init_buff utility routine

[Saeed Mahameed <saeed@kernel.org>]

On Thu, 2020-12-10 at 20:28 +0100, Lorenzo Bianconi wrote:
> > On Thu, 2020-12-10 at 18:59 +0100, Lorenzo Bianconi wrote:
> > > On Dec 10, Maciej Fijalkowski wrote:
> > > > On Thu, Dec 10, 2020 at 05:32:41PM +0100, Lorenzo Bianconi
> > > > wrote:
> > > > > > On Thu, Dec 10, 2020 at 04:50:42PM +0100, Lorenzo Bianconi
> > > > > > wrote:
> > > > > > > Introduce xdp_init_buff utility routine to initialize
> > > > > > > xdp_buff data
> > > > > > > structure. Rely on xdp_init_buff in all XDP capable
> > > > > > > drivers.
> > > > > > 
> > > > > > Hm, Jesper was suggesting two helpers, one that you
> > > > > > implemented
> > > > > > for things
> > > > > > that are set once per NAPI and the other that is set per
> > > > > > each
> > > > > > buffer.
> > > > > > 
> > > > > > Not sure about the naming for a second one -
> > > > > > xdp_prepare_buff ?
> > > > > > xdp_init_buff that you have feels ok.
> > > > > 
> > > > > ack, so we can have xdp_init_buff() for initialization done
> > > > > once
> > > > > per NAPI run and 
> > > > > xdp_prepare_buff() for per-NAPI iteration initialization,
> > > > > e.g.
> > > > > 
> > > > > static inline void
> > > > > xdp_prepare_buff(struct xdp_buff *xdp, unsigned char
> > > > > *hard_start,
> > > > > 		 int headroom, int data_len)
> > > > > {
> > > > > 	xdp->data_hard_start = hard_start;
> > > > > 	xdp->data = hard_start + headroom;
> > > > > 	xdp->data_end = xdp->data + data_len;
> > > > > 	xdp_set_data_meta_invalid(xdp);
> > > > > }
> > > > 
> > > > I think we should allow for setting the data_meta as well.
> > > > x64 calling convention states that first four args are placed
> > > > onto
> > > > registers, so to keep it fast maybe have a third helper:
> > > > 
> > > > static inline void
> > > > xdp_prepare_buff_meta(struct xdp_buff *xdp, unsigned char
> > > > *hard_start,
> > > > 		      int headroom, int data_len)
> > > > {
> > > > 	xdp->data_hard_start = hard_start;
> > > > 	xdp->data = hard_start + headroom;
> > > > 	xdp->data_end = xdp->data + data_len;
> > > > 	xdp->data_meta = xdp->data;
> > > > }
> > > > 
> > > > Thoughts?
> > > 
> > > ack, I am fine with it. Let's wait for some feedback.
> > > 
> > > Do you prefer to have xdp_prepare_buff/xdp_prepare_buff_meta in
> > > the
> > > same series
> > > of xdp_buff_init() or is it ok to address it in a separate patch?
> > > 
> > 
> > you only need 2
> > why do you need xpd_prepare_buff_meta? that's exactly
> > what xdp_set_data_meta_invalid(xdp) is all about.
> 
> IIUC what Maciej means is to avoid to overwrite xdp->data_meta with
> xdp_set_data_meta_invalid() after setting it to xdp->data in
> xdp_prepare_buff_meta().
> I guess setting xdp->data_meta to xdp->data is valid, it means an
> empty meta
> area.
> Anyway I guess we can set xdp->data_meta to xdp->data wherever we
> need and just
> keep xdp_prepare_buff(). Agree?
> 

hmm, i agree, but I would choose a default that is best for common use
case performance, so maybe do xd->data_meta = xdp->data by default and
drivers can override it, as they are already doing today if they don't
support it.

> Regards,
> Lorenzo
> 
> >