Thread (8 messages) 8 messages, 4 authors, 2025-07-14

Re: [PATCH v13 nf-next 1/3] netfilter: utils: nf_checksum(_partial) correct data!=networkheader

From: Florian Westphal <fw@strlen.de>
Date: 2025-07-04 19:39:15
Also in: bridge, netfilter-devel

Eric Woudstra [off-list ref] wrote:
quoted hunk ↗ jump to hunk
In the conntrack hook it may not always be the case that:
skb_network_header(skb) == skb->data.

This is problematic when L4 function nf_conntrack_handle_packet()
is accessing L3 data. This function uses thoff and ip_hdr()
to finds it's data. But it also calculates the checksum.
nf_checksum() and nf_checksum_partial() both use lower skb-checksum
functions that are based on using skb->data.

When skb_network_header(skb) != skb->data, adjust accordingly,
so that the checksum is calculated correctly.

Signed-off-by: Eric Woudstra <redacted>
---
 net/netfilter/utils.c | 22 ++++++++++++++++------
 1 file changed, 16 insertions(+), 6 deletions(-)
diff --git a/net/netfilter/utils.c b/net/netfilter/utils.c
index 008419db815a..daee035c25b8 100644
--- a/net/netfilter/utils.c
+++ b/net/netfilter/utils.c
@@ -124,16 +124,21 @@ __sum16 nf_checksum(struct sk_buff *skb, unsigned int hook,
 		    unsigned int dataoff, u8 protocol,
 		    unsigned short family)
 {
+	unsigned int nhpull = skb_network_header(skb) - skb->data;
 	__sum16 csum = 0;
 
+	if (!pskb_may_pull(skb, nhpull))
+		return -ENOMEM;
Hmm.  Not sure about this.  We should really audit all conntrack users
to make sure the network header is in the linear area, i.e.
ip_hdr() and friends return the right value, even though skb->data !=
skb_network_header().

Such may_pull, in case of skb->head reallocation, invalidate a pointer
to e.g. ethernet header in the caller.

No idea if we have callers that do this, I did not check, but such
"hidden" pulls tend to cause hard to spot bugs.

Maybe use
       if (WARN_ON_ONCE(skb_pointer_if_linear())
		return 0;

instead?  That allows to track down any offenders.  Given conntrack
takes presence of the l3 header in the linear area for granted, I don't
see how this can ever trigger.  You could also use
DEBUG_NET_WARN_ON_ONCE if you prefer, given this condition should never
be true anyway.
Keyboard shortcuts
hback out one level
jnext message in thread
kprevious message in thread
ldrill in
Escclose help / fold thread tree
?toggle this help