Re: [PATCH net] gso: fix GSO_DODGY bit handling for related protocols

From: Paolo Abeni
Date: Thu Jul 13 2023 - 03:14:13 EST

On Wed, 2023-07-12 at 21:58 -0500, Yan Zhai wrote:
> On Wed, Jul 12, 2023 at 9:11 PM Jason Wang <jasowang@xxxxxxxxxx> wrote:
> >
> > On Thu, Jul 13, 2023 at 9:55 AM Yan Zhai <yan@xxxxxxxxxxxxxx> wrote:
> > >
> > > SKB_GSO_DODGY bit indicates a GSO packet comes from an untrusted source.
> > > The canonical way is to recompute the gso_segs to avoid device driver
> > > issues. Afterwards, the DODGY bit can be removed to avoid re-check at the
> > > egress of later devices, e.g. packets can egress to a vlan device backed
> > > by a real NIC.
> > >
> > > Commit 1fd54773c267 ("udp: allow header check for dodgy GSO_UDP_L4
> > > packets.") checks DODGY bit for UDP, but for packets that can be fed
> > > directly to the device after gso_segs reset, it actually falls through
> > > to fragmentation [1].
> > >
> > > Commit 90017accff61 ("sctp: Add GSO support") and commit 3820c3f3e417
> > > ("[TCP]: Reset gso_segs if packet is dodgy") both didn't remove the DODGY
> > > bit after recomputing gso_segs.
> >
> > If we try to fix two issues, we'd better use separate patches.
> >
> > >
> > > This change fixes the GSO_UDP_L4 handling case, and remove the DODGY bit
> > > at other places.
> > >
> > > Fixes: 90017accff61 ("sctp: Add GSO support")
> > > Fixes: 3820c3f3e417 ("[TCP]: Reset gso_segs if packet is dodgy")
> > > Fixes: 1fd54773c267 ("udp: allow header check for dodgy GSO_UDP_L4 packets.")
> > > Signed-off-by: Yan Zhai <yan@xxxxxxxxxxxxxx>
> > >
> > > ---
> > > [1]:
> > > https://lore.kernel.org/all/CAJPywTKDdjtwkLVUW6LRA2FU912qcDmQOQGt2WaDo28KzYDg+A@xxxxxxxxxxxxxx/
> > >
> > > ---
> > > net/ipv4/tcp_offload.c | 1 +
> > > net/ipv4/udp_offload.c | 19 +++++++++++++++----
> > > net/ipv6/udp_offload.c | 19 +++++++++++++++----
> > > net/sctp/offload.c | 2 ++
> > > 4 files changed, 33 insertions(+), 8 deletions(-)
> > >
> > > diff --git a/net/ipv4/tcp_offload.c b/net/ipv4/tcp_offload.c
> > > index 8311c38267b5..f9b93708c22e 100644
> > > --- a/net/ipv4/tcp_offload.c
> > > +++ b/net/ipv4/tcp_offload.c
> > > @@ -87,6 +87,7 @@ struct sk_buff *tcp_gso_segment(struct sk_buff *skb,
> > > /* Packet is from an untrusted source, reset gso_segs. */
> > >
> > > skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss);
> > > + skb_shinfo(skb)->gso_type &= ~SKB_GSO_DODGY;
> > >
> > > segs = NULL;
> > > goto out;
> > > diff --git a/net/ipv4/udp_offload.c b/net/ipv4/udp_offload.c
> > > index 75aa4de5b731..bd29cf19bb6b 100644
> > > --- a/net/ipv4/udp_offload.c
> > > +++ b/net/ipv4/udp_offload.c
> > > @@ -388,11 +388,22 @@ static struct sk_buff *udp4_ufo_fragment(struct sk_buff *skb,
> > > if (!pskb_may_pull(skb, sizeof(struct udphdr)))
> > > goto out;
> > >
> > > - if (skb_shinfo(skb)->gso_type & SKB_GSO_UDP_L4 &&
> > > - !skb_gso_ok(skb, features | NETIF_F_GSO_ROBUST))
> > > - return __udp_gso_segment(skb, features, false);
> > > -
> > > mss = skb_shinfo(skb)->gso_size;
> > > +
> > > + if (skb_shinfo(skb)->gso_type & SKB_GSO_UDP_L4) {
> > > + if (skb_gso_ok(skb, features | NETIF_F_GSO_ROBUST)) {
> > > + /* Packet is from an untrusted source, reset actual gso_segs */
> > > + skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len - sizeof(*uh),
> > > + mss);
> > > + skb_shinfo(skb)->gso_type &= ~SKB_GSO_DODGY;
> > > +
> > > + segs = NULL;
> > > + goto out;
> > > + } else {
> > > + return __udp_gso_segment(skb, features, false);
> >
> > I think it's better and cleaner to move those changes in
> > __udp_gso_segment() as Willem suggests.
> >
> > > + }
> > > + }
> > > +
> > > if (unlikely(skb->len <= mss))
> > > goto out;
> > >
> > > diff --git a/net/ipv6/udp_offload.c b/net/ipv6/udp_offload.c
> > > index ad3b8726873e..6857d9f7bd06 100644
> > > --- a/net/ipv6/udp_offload.c
> > > +++ b/net/ipv6/udp_offload.c
> > > @@ -43,11 +43,22 @@ static struct sk_buff *udp6_ufo_fragment(struct sk_buff *skb,
> > > if (!pskb_may_pull(skb, sizeof(struct udphdr)))
> > > goto out;
> > >
> > > - if (skb_shinfo(skb)->gso_type & SKB_GSO_UDP_L4 &&
> > > - !skb_gso_ok(skb, features | NETIF_F_GSO_ROBUST))
> > > - return __udp_gso_segment(skb, features, true);
> > > -
> > > mss = skb_shinfo(skb)->gso_size;
> > > +
> > > + if (skb_shinfo(skb)->gso_type & SKB_GSO_UDP_L4) {
> > > + if (skb_gso_ok(skb, features | NETIF_F_GSO_ROBUST)) {
> > > + /* Packet is from an untrusted source, reset actual gso_segs */
> > > + skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len - sizeof(*uh),
> > > + mss);
> > > + skb_shinfo(skb)->gso_type &= ~SKB_GSO_DODGY;
> >
> > Any reason you want to remove the DODGY here? Is this an optimization?
> > We will lose the chance to recognize/validate it elsewhere.
> >
> It is intended as a small optimization. And this is in fact the piece
> I am not fully confident about: after validating the gso_segs at a
> trusted location (i.e. assuming the kernel is the trusted computing
> base), do we need to validate it somewhere else? For example, in our
> scenario, we have a tun/tap device in a net namespace, so the packet
> going out will enter from the tap, get forwarded through an veth, and
> then a vlan backed by a real ethernet interface. If the bit is carried
> over, then at each egress of these devices, we need to enter the GSO
> code, which feels pretty redundant as long as the packet does not
> leave kernel space. WDYT?

As an optimization, I think it should land on a different (net-next)
patch. Additionally I think it should be possible to get a greater gain
adding the ROBUST feature to virtual devices (but I'm not sure if
syzkaller will be able to use that in nasty ways).

Cheers,

Paolo