Re: Linux 5.12-rc7

[Eric Dumazet]

On Tue, Apr 13, 2021 at 2:57 PM Michael S. Tsirkin <mst@xxxxxxxxxx> wrote:
>
> On Mon, Apr 12, 2021 at 06:47:07PM +0200, Eric Dumazet wrote:
> > On Mon, Apr 12, 2021 at 6:31 PM Eric Dumazet <edumazet@xxxxxxxxxx> wrote:
> > >
> > > On Mon, Apr 12, 2021 at 6:28 PM Linus Torvalds
> > > <torvalds@xxxxxxxxxxxxxxxxxxxx> wrote:
> > > >
> > > > On Sun, Apr 11, 2021 at 10:14 PM Guenter Roeck <linux@xxxxxxxxxxxx> wrote:
> > > > >
> > > > > Qemu test results:
> > > > >         total: 460 pass: 459 fail: 1
> > > > > Failed tests:
> > > > >         sh:rts7751r2dplus_defconfig:ata:net,virtio-net:rootfs
> > > > >
> > > > > The failure bisects to commit 0f6925b3e8da ("virtio_net: Do not pull payload in
> > > > > skb->head"). It is a spurious problem - the test passes roughly every other
> > > > > time. When the failure is seen, udhcpc fails to get an IP address and aborts
> > > > > with SIGTERM. So far I have only seen this with the "sh" architecture.
> > > >
> > > > Hmm. Let's add in some more of the people involved in that commit, and
> > > > also netdev.
> > > >
> > > > Nothing in there looks like it should have any interaction with
> > > > architecture, so that "it happens on sh" sounds odd, but maybe it's
> > > > some particular interaction with the qemu environment.
> > >
> > > Yes, maybe.
> > >
> > > I spent few hours on this, and suspect a buggy memcpy() implementation
> > > on SH, but this was not conclusive.
> > >
> > > By pulling one extra byte, the problem goes away.
> > >
> > > Strange thing is that the udhcpc process does not go past sendto().
> >
> > This is the patch working around the issue. Unfortunately I was not
> > able to root-cause it (I really suspect something on SH)
> >
> > diff --git a/drivers/net/virtio_net.c b/drivers/net/virtio_net.c
> > index 0824e6999e49957f7aaf7c990f6259792d42f32b..fd890a951beea03bdf24406809042666eb972655
> > 100644
> > --- a/drivers/net/virtio_net.c
> > +++ b/drivers/net/virtio_net.c
> > @@ -408,11 +408,17 @@ static struct sk_buff *page_to_skb(struct
> > virtnet_info *vi,
> >
> >         /* Copy all frame if it fits skb->head, otherwise
> >          * we let virtio_net_hdr_to_skb() and GRO pull headers as needed.
> > +        *
> > +        * Apparently, pulling only the Ethernet Header triggers a bug
> > on qemu-system-sh4.
> > +        * Since GRO aggregation really cares of IPv4/IPv6, pull 20 bytes
> > +        * more to work around this bug : These 20 bytes can not belong
> > +        * to UDP/TCP payload.
> > +        * As a bonus, this makes GRO slightly faster for IPv4 (one less copy).
> >          */
>
> Question: do we still want to do this for performance reasons?
> We also have the hdr_len coming from the device which is
> just skb_headlen on the host.

Well, putting 20 bytes in skb->head will disable frag0 optimization.

The change would only benefit to sh architecture :)

About hdr_len, I suppose we could try it, with appropriate safety checks.

>
> >         if (len <= skb_tailroom(skb))
> >                 copy = len;
> >         else
> > -               copy = ETH_HLEN + metasize;
> > +               copy = ETH_HLEN + sizeof(struct iphdr) + metasize;
> >         skb_put_data(skb, p, copy);
> >
> >         if (metasize) {
>