Re: [PATCH] net: core: sk_buff: zero-fill skb->data in __alloc_skb function

From: Eric Dumazet
Date: Sat Apr 10 2021 - 07:00:57 EST


On Sat, Apr 10, 2021 at 12:12 PM Eric Dumazet <edumazet@xxxxxxxxxx> wrote:
>
> On Sat, Apr 10, 2021 at 11:51 AM Phillip Potter <phil@xxxxxxxxxxxxxxxx> wrote:
> >
> > Zero-fill skb->data in __alloc_skb function of net/core/skbuff.c,
> > up to start of struct skb_shared_info bytes. Fixes a KMSAN-found
> > uninit-value bug reported by syzbot at:
> > https://syzkaller.appspot.com/bug?id=abe95dc3e3e9667fc23b8d81f29ecad95c6f106f
> >
> > Reported-by: syzbot+2e406a9ac75bb71d4b7a@xxxxxxxxxxxxxxxxxxxxxxxxx
> > Signed-off-by: Phillip Potter <phil@xxxxxxxxxxxxxxxx>
> > ---
> > net/core/skbuff.c | 1 +
> > 1 file changed, 1 insertion(+)
> >
> > diff --git a/net/core/skbuff.c b/net/core/skbuff.c
> > index 785daff48030..9ac26cdb5417 100644
> > --- a/net/core/skbuff.c
> > +++ b/net/core/skbuff.c
> > @@ -215,6 +215,7 @@ struct sk_buff *__alloc_skb(unsigned int size, gfp_t gfp_mask,
> > * to allow max possible filling before reallocation.
> > */
> > size = SKB_WITH_OVERHEAD(ksize(data));
> > + memset(data, 0, size);
> > prefetchw(data + size);
>
>
> Certainly not.
>
> There is a difference between kmalloc() and kzalloc()
>
> Here you are basically silencing KMSAN and make it useless.
>
> Please fix the real issue, or stop using KMSAN if it bothers you.

My understanding of the KMSAN bug (when I released it months ago) was
that it was triggered by some invalid assumptions in geneve_xmit()

The syzbot repro sends a packet with a very small size (Ethernet
header only) and no IP/IPv6 header

Fix for ipv4 part (sorry, not much time during week end to test all this)

diff --git a/drivers/net/geneve.c b/drivers/net/geneve.c
index e3b2375ac5eb55f544bbc1f309886cc9be189fd1..0a72779bc74bc50c20c34c05b2c525cca829f33c
100644
--- a/drivers/net/geneve.c
+++ b/drivers/net/geneve.c
@@ -892,6 +892,9 @@ static int geneve_xmit_skb(struct sk_buff *skb,
struct net_device *dev,
__be16 sport;
int err;

+ if (!pskb_network_may_pull(skb, sizeof(struct iphdr))
+ return -EINVAL;
+
sport = udp_flow_src_port(geneve->net, skb, 1, USHRT_MAX, true);
rt = geneve_get_v4_rt(skb, dev, gs4, &fl4, info,
geneve->cfg.info.key.tp_dst, sport);