On 10/26/24 5:33 PM, Cong Wang wrote:
On Tue, Oct 22, 2024 at 10:52:31PM +0200, Daniel Borkmann wrote:
On 10/8/24 7:33 AM, Cong Wang wrote:
From: Cong Wang <cong.wang@xxxxxxxxxxxxx>
skb_transport_offset() and skb_transport_offset() can be negative when
nit: I presume the 2nd one is skb_network_offset?
they are called after we pull the transport header, for example, when
we use eBPF sockmap (aka at the point of ->sk_data_ready()).
__bpf_skb_min_len() uses an unsigned int to get these offsets, this
leads to a very large number which causes bpf_skb_change_tail() failed
unexpectedly.
Fix this by using a signed int to get these offsets and test them
against zero.
Fixes: 5293efe62df8 ("bpf: add bpf_skb_change_tail helper")
Cc: Daniel Borkmann <daniel@xxxxxxxxxxxxx>
Signed-off-by: Cong Wang <cong.wang@xxxxxxxxxxxxx>
Is there any chance you could also extend the sockmap BPF selftest with
this case you're hitting so that BPF CI can run this regularly?
Yes, my colleague Zijian (Cc'ed) is working on a selftest to cover this case.
Please let me know if you prefer to send it together with the selftest,
technically it would make backporting this fix harder, but I am open to
any suggestion here.
Ideally this is a 2-patch series, first one is the fix itself and the second
one contains the BPF selftest so that CI can run it too, and this way it also
does not hurt backporting the fix.
net/core/filter.c | 21 +++++++++++++++------
1 file changed, 15 insertions(+), 6 deletions(-)
diff --git a/net/core/filter.c b/net/core/filter.c
index 4e3f42cc6611..10ef27639a5d 100644
--- a/net/core/filter.c
+++ b/net/core/filter.c
@@ -3737,13 +3737,22 @@ static const struct bpf_func_proto bpf_skb_adjust_room_proto = {
static u32 __bpf_skb_min_len(const struct sk_buff *skb)
{
- u32 min_len = skb_network_offset(skb);
+ int offset = skb_network_offset(skb);
+ u32 min_len = 0;
- if (skb_transport_header_was_set(skb))
- min_len = skb_transport_offset(skb);
- if (skb->ip_summed == CHECKSUM_PARTIAL)
- min_len = skb_checksum_start_offset(skb) +
- skb->csum_offset + sizeof(__sum16);
+ if (offset > 0)
+ min_len = offset;
+ if (skb_transport_header_was_set(skb)) {
+ offset = skb_transport_offset(skb);
+ if (offset > 0)
+ min_len = offset;
+ }
+ if (skb->ip_summed == CHECKSUM_PARTIAL) {
+ offset = skb_checksum_start_offset(skb) +
+ skb->csum_offset + sizeof(__sum16);
+ if (offset > 0)
+ min_len = offset;
+ }
return min_len;
I'll let John chime in, but does this mean in case of sockmap min_len always ends
up at 0? I just wonder whether we should pass a custom __bpf_skb_min_len to
__bpf_skb_change_tail for bpf_skb_change_tail vs sk_skb_change_tail assuming the
compiler is able to inlining all this (instead of indirect call).
Yes, in case of sockmap skb->data is already past TCP header, so all the
offsets here are negative. And since the 'new_len' of bpf_skb_change_tail()
is unsigned (too late to change), min_len should be zero.
Ok, so hopefully this can be further cleaned up/simplified a bit more then.
Thanks,
Daniel
