> On Thu, 10 Oct 2019 01:18:35 +0200
> Lorenzo Bianconi <lorenzo@kernel.org> wrote:
> 
> > Add basic XDP support to mvneta driver for devices that rely on software
> > buffer management. Currently supported verdicts are:
> > - XDP_DROP
> > - XDP_PASS
> > - XDP_REDIRECT
> > - XDP_ABORTED
> > 
> > Signed-off-by: Lorenzo Bianconi <lorenzo@kernel.org>
> > ---
> >  drivers/net/ethernet/marvell/mvneta.c | 144 ++++++++++++++++++++++++--
> >  1 file changed, 135 insertions(+), 9 deletions(-)
> > 
> > diff --git a/drivers/net/ethernet/marvell/mvneta.c b/drivers/net/ethernet/marvell/mvneta.c
> > index ba4aa9bbc798..e2795dddbcaf 100644
> > --- a/drivers/net/ethernet/marvell/mvneta.c
> > +++ b/drivers/net/ethernet/marvell/mvneta.c
> [...]
> 
> > @@ -1950,16 +1960,60 @@ int mvneta_rx_refill_queue(struct mvneta_port *pp, struct mvneta_rx_queue *rxq)
> >  	return i;
> >  }
> >  
> > +static int
> > +mvneta_run_xdp(struct mvneta_port *pp, struct bpf_prog *prog,
> > +	       struct xdp_buff *xdp)
> > +{
> > +	u32 ret, act = bpf_prog_run_xdp(prog, xdp);
> > +
> > +	switch (act) {
> > +	case XDP_PASS:
> > +		ret = MVNETA_XDP_PASS;
> > +		break;
> > +	case XDP_REDIRECT: {
> > +		int err;
> > +
> > +		err = xdp_do_redirect(pp->dev, xdp, prog);
> > +		if (err) {
> > +			ret = MVNETA_XDP_CONSUMED;
> > +			xdp_return_buff(xdp);
> 
> > +		} else {
> > +			ret = MVNETA_XDP_REDIR;
> > +		}
> > +		break;
> > +	}
> > +	default:
> > +		bpf_warn_invalid_xdp_action(act);
> > +		/* fall through */
> > +	case XDP_ABORTED:
> > +		trace_xdp_exception(pp->dev, prog, act);
> > +		/* fall through */
> > +	case XDP_DROP:
> > +		ret = MVNETA_XDP_CONSUMED;
> > +		xdp_return_buff(xdp);
> 
> Using xdp_return_buff() here is actually not optimal for performance.
> I can see that others socionext/netsec.c and AF_XDP also use this
> xdp_return_buff().
> 
> I do think code wise it looks a lot nice to use xdp_return_buff(), so
> maybe we should optimize xdp_return_buff(), instead of using
> page_pool_recycle_direct() here?  (That would also help AF_XDP ?)
> 
> The problem with xdp_return_buff() is that it does a "full" lookup from
> the mem.id (xdp_buff->xdp_rxq_info->mem.id) to find the "allocator"
> pointer in this case the page_pool pointer.  Here in the driver we
> already have access to the stable allocator page_pool pointer via
> struct mvneta_rx_queue *rxq->page_pool.

ack, right. Thx for pointing it out. I will fix it in v3

> 
> 
> > +		break;
> > +	}
> > +
> > +	return ret;
> > +}
> > +
> >  static int
> >  mvneta_swbm_rx_frame(struct mvneta_port *pp,
> >  		     struct mvneta_rx_desc *rx_desc,
> >  		     struct mvneta_rx_queue *rxq,
> > -		     struct page *page)
> > +		     struct bpf_prog *xdp_prog,
> > +		     struct page *page, u32 *xdp_ret)
> >  {
> >  	unsigned char *data = page_address(page);
> >  	int data_len = -MVNETA_MH_SIZE, len;
> >  	struct net_device *dev = pp->dev;
> >  	enum dma_data_direction dma_dir;
> > +	struct xdp_buff xdp = {
> > +		.data_hard_start = data,
> > +		.data = data + MVNETA_SKB_HEADROOM + MVNETA_MH_SIZE,
> > +		.rxq = &rxq->xdp_rxq,
> > +	};
> 
> Creating the struct xdp_buff (on call-stack) this way is not optimal
> for performance (IHMO it looks nicer code-wise, but too bad).
> 
> This kind of initialization of only some of the members, cause GCC to
> zero out other members (I observed this on Intel, which use an
> expensive rep-sto operation).  Thus, this cause extra unnecessary memory
> writes.
> 
> A further optimization, is that you can avoid re-assigning:
>  rxq = &rxq->xdp_rxq
> for each frame, as this actually stays the same for all the frames in
> this NAPI cycle.  By instead allocating the xdp_buff on the callers
> stack, and parsing in xdp_buff as a pointer.

ack, will fix it in v3.

Regards,
Lorenzo

> 
> 
> > +	xdp_set_data_meta_invalid(&xdp);
> >  
> >  	if (MVNETA_SKB_SIZE(rx_desc->data_size) > PAGE_SIZE) {
> >  		len = MVNETA_MAX_RX_BUF_SIZE;
> > @@ -1968,13 +2022,27 @@ mvneta_swbm_rx_frame(struct mvneta_port *pp,
> >  		len = rx_desc->data_size;
> >  		data_len += len - ETH_FCS_LEN;
> >  	}
> > +	xdp.data_end = xdp.data + data_len;
> >  
> >  	dma_dir = page_pool_get_dma_dir(rxq->page_pool);
> >  	dma_sync_single_range_for_cpu(dev->dev.parent,
> >  				      rx_desc->buf_phys_addr, 0,
> >  				      len, dma_dir);
> >  
> > -	rxq->skb = build_skb(data, PAGE_SIZE);
> > +	if (xdp_prog) {
> > +		u32 ret;
> > +
> > +		ret = mvneta_run_xdp(pp, xdp_prog, &xdp);
> > +		if (ret != MVNETA_XDP_PASS) {
> > +			mvneta_update_stats(pp, 1, xdp.data_end - xdp.data,
> > +					    false);
> > +			rx_desc->buf_phys_addr = 0;
> > +			*xdp_ret |= ret;
> > +			return ret;
> > +		}
> > +	}
> > +
> > +	rxq->skb = build_skb(xdp.data_hard_start, PAGE_SIZE);
> >  	if (unlikely(!rxq->skb)) {
> >  		netdev_err(dev,
> >  			   "Can't allocate skb on queue %d\n",
> [...]
> 
> -- 
> Best regards,
>   Jesper Dangaard Brouer
>   MSc.CS, Principal Kernel Engineer at Red Hat
>   LinkedIn: http://www.linkedin.com/in/brouer