Re: [PATCH net-next 1/8] virtio_net: mergeable xdp: put old page immediately

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



On Thu, Mar 23, 2023 at 9:43 AM Xuan Zhuo <xuanzhuo@xxxxxxxxxxxxxxxxx> wrote:
>
> On Wed, 22 Mar 2023 16:22:18 +0800, Yunsheng Lin <linyunsheng@xxxxxxxxxx> wrote:
> > On 2023/3/22 11:03, Xuan Zhuo wrote:
> > > In the xdp implementation of virtio-net mergeable, it always checks
> > > whether two page is used and a page is selected to release. This is
> > > complicated for the processing of action, and be careful.
> > >
> > > In the entire process, we have such principles:
> > > * If xdp_page is used (PASS, TX, Redirect), then we release the old
> > >   page.
> > > * If it is a drop case, we will release two. The old page obtained from
> > >   buf is release inside err_xdp, and xdp_page needs be relased by us.
> > >
> > > But in fact, when we allocate a new page, we can release the old page
> > > immediately. Then just one is using, we just need to release the new
> > > page for drop case. On the drop path, err_xdp will release the variable
> > > "page", so we only need to let "page" point to the new xdp_page in
> > > advance.
> > >
> > > Signed-off-by: Xuan Zhuo <xuanzhuo@xxxxxxxxxxxxxxxxx>
> > > ---
> > >  drivers/net/virtio_net.c | 15 ++++++---------
> > >  1 file changed, 6 insertions(+), 9 deletions(-)
> > >
> > > diff --git a/drivers/net/virtio_net.c b/drivers/net/virtio_net.c
> > > index e2560b6f7980..4d2bf1ce0730 100644
> > > --- a/drivers/net/virtio_net.c
> > > +++ b/drivers/net/virtio_net.c
> > > @@ -1245,6 +1245,9 @@ static struct sk_buff *receive_mergeable(struct net_device *dev,
> > >                     if (!xdp_page)
> > >                             goto err_xdp;
> > >                     offset = VIRTIO_XDP_HEADROOM;
> > > +
> > > +                   put_page(page);
> >
> > the error handling of xdp_linearize_page() does not seems self contained.
> > Does it not seem better:
> > 1. if xdp_linearize_page() succesed, call put_page() for first buffer just
> >    as put_page() is call for other buffer
> > 2. or call virtqueue_get_buf() and put_page() for all the buffer of the packet
> >    so the error handling is not needed outside the virtqueue_get_buf().
> >
> > In that case, it seems we can just do below without xdp_page:
> > page = xdp_linearize_page(rq, num_buf, page, ...);
>
>
> This does look better.
>
> In fact, we already have vq reset, we can load XDP based on vq reset.
> In this way, we can run without xdp_linearize_page.

The goal is to try our best not to drop packets, so I think it's
better to keep it.

Thanks

>
>
> >
> >
> > > +                   page = xdp_page;
> > >             } else if (unlikely(headroom < virtnet_get_headroom(vi))) {
> > >                     xdp_room = SKB_DATA_ALIGN(VIRTIO_XDP_HEADROOM +
> > >                                               sizeof(struct skb_shared_info));
> > > @@ -1259,6 +1262,9 @@ static struct sk_buff *receive_mergeable(struct net_device *dev,
> > >                            page_address(page) + offset, len);
> > >                     frame_sz = PAGE_SIZE;
> > >                     offset = VIRTIO_XDP_HEADROOM;
> > > +
> > > +                   put_page(page);
> > > +                   page = xdp_page;
> >
> > It seems we can limit the scope of xdp_page in this "else if" block.
> >
> > >             } else {
> > >                     xdp_page = page;
> > >             }
> >
> > It seems the above else block is not needed anymore.
>
> Yes, the follow-up patch has this optimization.
>
>
> >
> > > @@ -1278,8 +1284,6 @@ static struct sk_buff *receive_mergeable(struct net_device *dev,
> > >                     if (unlikely(!head_skb))
> > >                             goto err_xdp_frags;
> > >
> > > -                   if (unlikely(xdp_page != page))
> > > -                           put_page(page);
> > >                     rcu_read_unlock();
> > >                     return head_skb;
> > >             case XDP_TX:
> > > @@ -1297,8 +1301,6 @@ static struct sk_buff *receive_mergeable(struct net_device *dev,
> > >                             goto err_xdp_frags;
> > >                     }
> > >                     *xdp_xmit |= VIRTIO_XDP_TX;
> > > -                   if (unlikely(xdp_page != page))
> > > -                           put_page(page);
> > >                     rcu_read_unlock();
> > >                     goto xdp_xmit;
> > >             case XDP_REDIRECT:
> > > @@ -1307,8 +1309,6 @@ static struct sk_buff *receive_mergeable(struct net_device *dev,
> > >                     if (err)
> > >                             goto err_xdp_frags;
> > >                     *xdp_xmit |= VIRTIO_XDP_REDIR;
> > > -                   if (unlikely(xdp_page != page))
> > > -                           put_page(page);
> > >                     rcu_read_unlock();
> > >                     goto xdp_xmit;
> > >             default:
> > > @@ -1321,9 +1321,6 @@ static struct sk_buff *receive_mergeable(struct net_device *dev,
> > >                     goto err_xdp_frags;
> > >             }
> > >  err_xdp_frags:
> > > -           if (unlikely(xdp_page != page))
> > > -                   __free_pages(xdp_page, 0);
> >
> > It seems __free_pages() and put_page() is used interchangeably here.
> > Perhaps using __free_pages() have performance reason? As the comment below:
> >
> > https://elixir.bootlin.com/linux/v6.3-rc3/source/net/core/page_pool.c#L500
>
>
> Yes, but now we don't seem to be very good to distinguish it. But I think
> it doesn't matter. This logic is rare under actual situation.
>
> Thanks.
>
>
> >
> > > -
> > >             if (xdp_buff_has_frags(&xdp)) {
> > >                     shinfo = xdp_get_shared_info_from_buff(&xdp);
> > >                     for (i = 0; i < shinfo->nr_frags; i++) {
> > >
>





[Index of Archives]     [Linux Samsung SoC]     [Linux Rockchip SoC]     [Linux Actions SoC]     [Linux for Synopsys ARC Processors]     [Linux NFS]     [Linux NILFS]     [Linux USB Devel]     [Video for Linux]     [Linux Audio Users]     [Yosemite News]     [Linux Kernel]     [Linux SCSI]


  Powered by Linux