Re: [PATCH v12 bpf-next 03/18] net: mvneta: update mb bit before passing the xdp buffer to eBPF layer

From: Lorenzo Bianconi <lorenzo.bianconi@redhat.com>
To: John Fastabend <john.fastabend@gmail.com>
Cc: Lorenzo Bianconi <lorenzo@kernel.org>,
	BPF-dev-list <bpf@vger.kernel.org>,
	Network Development <netdev@vger.kernel.org>,
	"David S. Miller" <davem@davemloft.net>,
	Jakub Kicinski <kuba@kernel.org>,
	Alexei Starovoitov <ast@kernel.org>,
	Daniel Borkmann <daniel@iogearbox.net>,
	"Agroskin, Shay" <shayagr@amazon.com>,
	David Ahern <dsahern@kernel.org>,
	Jesper Brouer <brouer@redhat.com>,
	Eelco Chaudron <echaudro@redhat.com>,
	Jason Wang <jasowang@redhat.com>,
	Alexander Duyck <alexander.duyck@gmail.com>,
	Saeed Mahameed <saeed@kernel.org>,
	Maciej Fijalkowski <maciej.fijalkowski@intel.com>,
	"Karlsson, Magnus" <magnus.karlsson@intel.com>,
	"Sarkar, Tirthendu" <tirthendu.sarkar@intel.com>,
	Toke Hoiland Jorgensen <toke@redhat.com>
Subject: Re: [PATCH v12 bpf-next 03/18] net: mvneta: update mb bit before passing the xdp buffer to eBPF layer
Date: Fri, 3 Sep 2021 19:23:01 +0200	[thread overview]
Message-ID: <CAJ0CqmU6cFP3WrvqVqbVKt8R-yrKTKp2ZHm65O-w5fjEJVvemA@mail.gmail.com> (raw)
In-Reply-To: <612eba0cd2678_6b872081d@john-XPS-13-9370.notmuch>

On Wed, Sep 1, 2021 at 1:24 AM John Fastabend <john.fastabend@gmail.com> wrote:
>
> Lorenzo Bianconi wrote:
> > Update multi-buffer bit (mb) in xdp_buff to notify XDP/eBPF layer and
> > XDP remote drivers if this is a "non-linear" XDP buffer. Access
> > skb_shared_info only if xdp_buff mb is set in order to avoid possible
> > cache-misses.
> >
> > Signed-off-by: Lorenzo Bianconi <lorenzo@kernel.org>
> > ---
> >  drivers/net/ethernet/marvell/mvneta.c | 23 ++++++++++++++++++-----
> >  1 file changed, 18 insertions(+), 5 deletions(-)
> >
> > diff --git a/drivers/net/ethernet/marvell/mvneta.c b/drivers/net/ethernet/marvell/mvneta.c
> > index 5d1007e1b5c9..9f4858e35566 100644
> > --- a/drivers/net/ethernet/marvell/mvneta.c
> > +++ b/drivers/net/ethernet/marvell/mvneta.c
> > @@ -2037,9 +2037,14 @@ mvneta_xdp_put_buff(struct mvneta_port *pp, struct mvneta_rx_queue *rxq,
> >  {
> >       int i;
> >
> > +     if (likely(!xdp_buff_is_mb(xdp)))
> > +             goto out;
> > +
>
> Wouldn't nr_frags = 0 in the !xdp_buff_is_mb case? Is the
> xdp_buff_is_mb check with goto really required?

if xdp_buff_is_mb is false, nr_frags will not be initialized otherwise
we will trigger a cache-miss for the single-buffer use case (where
initializing skb_shared_info is not required).

>
> >       for (i = 0; i < sinfo->nr_frags; i++)
> >               page_pool_put_full_page(rxq->page_pool,
> >                                       skb_frag_page(&sinfo->frags[i]), true);
> > +
> > +out:
> >       page_pool_put_page(rxq->page_pool, virt_to_head_page(xdp->data),
> >                          sync_len, true);
> >  }
> > @@ -2241,7 +2246,6 @@ mvneta_swbm_rx_frame(struct mvneta_port *pp,
> >       int data_len = -MVNETA_MH_SIZE, len;
> >       struct net_device *dev = pp->dev;
> >       enum dma_data_direction dma_dir;
> > -     struct skb_shared_info *sinfo;
> >
> >       if (*size > MVNETA_MAX_RX_BUF_SIZE) {
> >               len = MVNETA_MAX_RX_BUF_SIZE;
> > @@ -2261,11 +2265,9 @@ mvneta_swbm_rx_frame(struct mvneta_port *pp,
> >
> >       /* Prefetch header */
> >       prefetch(data);
> > +     xdp_buff_clear_mb(xdp);
> >       xdp_prepare_buff(xdp, data, pp->rx_offset_correction + MVNETA_MH_SIZE,
> >                        data_len, false);
> > -
> > -     sinfo = xdp_get_shared_info_from_buff(xdp);
> > -     sinfo->nr_frags = 0;
> >  }
> >
> >  static void
> > @@ -2299,6 +2301,9 @@ mvneta_swbm_add_rx_fragment(struct mvneta_port *pp,
> >               skb_frag_off_set(frag, pp->rx_offset_correction);
> >               skb_frag_size_set(frag, data_len);
> >               __skb_frag_set_page(frag, page);
> > +
> > +             if (!xdp_buff_is_mb(xdp))
> > +                     xdp_buff_set_mb(xdp);
> >       } else {
> >               page_pool_put_full_page(rxq->page_pool, page, true);
> >       }
> > @@ -2320,8 +2325,12 @@ mvneta_swbm_build_skb(struct mvneta_port *pp, struct page_pool *pool,
> >                     struct xdp_buff *xdp, u32 desc_status)
> >  {
> >       struct skb_shared_info *sinfo = xdp_get_shared_info_from_buff(xdp);
> > -     int i, num_frags = sinfo->nr_frags;
> >       struct sk_buff *skb;
> > +     u8 num_frags;
> > +     int i;
> > +
> > +     if (unlikely(xdp_buff_is_mb(xdp)))
> > +             num_frags = sinfo->nr_frags;
> >
> >       skb = build_skb(xdp->data_hard_start, PAGE_SIZE);
> >       if (!skb)
> > @@ -2333,6 +2342,9 @@ mvneta_swbm_build_skb(struct mvneta_port *pp, struct page_pool *pool,
> >       skb_put(skb, xdp->data_end - xdp->data);
> >       skb->ip_summed = mvneta_rx_csum(pp, desc_status);
> >
> > +     if (likely(!xdp_buff_is_mb(xdp)))
> > +             goto out;
> > +
>
> Not that I care much, but couldn't you just init num_frags = 0 and
> avoid the goto?

same here.

Regards,
Lorenzo

>
> Anyways its not my driver so no need to change it if you like it better
> the way it is. Mostly just checking my understanding.
>
> >       for (i = 0; i < num_frags; i++) {
> >               skb_frag_t *frag = &sinfo->frags[i];
> >
> > @@ -2341,6 +2353,7 @@ mvneta_swbm_build_skb(struct mvneta_port *pp, struct page_pool *pool,
> >                               skb_frag_size(frag), PAGE_SIZE);
> >       }
> >
> > +out:
> >       return skb;
> >  }
> >
> > --
> > 2.31.1
> >
>
>