Re: [PATCH] virtio-net: make copy len check in xdp_linearize_page

[Jason Wang <jasowang@redhat.com>]

On Mon, Dec 13, 2021 at 5:14 PM 孙蒙恩 <mengensun8801@gmail.com> wrote:
>
> Jason Wang <jasowang@redhat.com> 于2021年12月13日周一 15:49写道：
> >
> > On Mon, Dec 13, 2021 at 12:50 PM <mengensun8801@gmail.com> wrote:
> > >
> > > From: mengensun <mengensun@tencent.com>
> > >
> > > xdp_linearize_page asume ring elem size is smaller then page size
> > > when copy the first ring elem, but, there may be a elem size bigger
> > > then page size.
> > >
> > > add_recvbuf_mergeable may add a hole to ring elem, the hole size is
> > > not sure, according EWMA.
> >
> > The logic is to try to avoid dropping packets in this case, so I
> > wonder if it's better to "fix" the add_recvbuf_mergeable().
>

Adding lists back.

> turn to XDP generic is so difficulty for me, here can "fix" the
> add_recvbuf_mergeable link follow:
> --- a/drivers/net/virtio_net.c
> +++ b/drivers/net/virtio_net.c
> diff --git a/drivers/net/virtio_net.c b/drivers/net/virtio_net.c
> index 36a4b7c195d5..06ce8bb10b47 100644
> --- a/drivers/net/virtio_net.c
> +++ b/drivers/net/virtio_net.c
> @@ -1315,6 +1315,7 @@ static int add_recvbuf_mergeable(struct virtnet_info *vi,
>                 alloc_frag->offset += hole;
>         }
> +       len = min(len, PAGE_SIZE - room);
>         sg_init_one(rq->sg, buf, len);
>         ctx = mergeable_len_to_ctx(len, headroom);

Then the truesize here is wrong.

>         err = virtqueue_add_inbuf_ctx(rq->vq, rq->sg, 1, buf, ctx, gfp);
>
> it seems a rule that, length of elem giving to vring is away smaller
> or equall then PAGE_SIZE

It aims to be consistent to what EWMA tries to do:

        len = hdr_len + clamp_t(unsigned int, ewma_pkt_len_read(avg_pkt_len),
                        rq->min_buf_len, PAGE_SIZE - hdr_len);

Thanks

>
> >
> > Or another idea is to switch to use XDP generic here where we can use
> > skb_linearize() which should be more robust and we can drop the
> > xdp_linearize_page() logic completely.
> >
> > Thanks
> >
> > >
> > > so, fix it by check copy len,if checked failed, just dropped the
> > > whole frame, not make the memory dirty after the page.
> > >
> > > Signed-off-by: mengensun <mengensun@tencent.com>
> > > Reviewed-by: MengLong Dong <imagedong@tencent.com>
> > > Reviewed-by: ZhengXiong Jiang <mungerjiang@tencent.com>
> > > ---
> > >  drivers/net/virtio_net.c | 6 +++++-
> > >  1 file changed, 5 insertions(+), 1 deletion(-)
> > >
> > > diff --git a/drivers/net/virtio_net.c b/drivers/net/virtio_net.c
> > > index 36a4b7c195d5..844bdbd67ff7 100644
> > > --- a/drivers/net/virtio_net.c
> > > +++ b/drivers/net/virtio_net.c
> > > @@ -662,8 +662,12 @@ static struct page *xdp_linearize_page(struct receive_queue *rq,
> > >                                        int page_off,
> > >                                        unsigned int *len)
> > >  {
> > > -       struct page *page = alloc_page(GFP_ATOMIC);
> > > +       struct page *page;
> > >
> > > +       if (*len > PAGE_SIZE - page_off)
> > > +               return NULL;
> > > +
> > > +       page = alloc_page(GFP_ATOMIC);
> > >         if (!page)
> > >                 return NULL;
> > >
> > > --
> > > 2.27.0
> > >
> >
>