On 6/10/26 9:58 AM, sun jian wrote:
On Wed, Jun 10, 2026 at 9:21 AM Menglong Dong <[email protected]> wrote:
On 2026/6/10 08:06 Emil Tsalapatis <[email protected]> write:
On Tue Jun 9, 2026 at 7:06 AM EDT, Menglong Dong wrote:
On 2026/6/9 18:02 Sun Jian <[email protected]> write:
[...]
Hi, Jian.
This sounds like a good idea in this case. When I have a look at
bpf_clone_redirect(),
I found that it use skb_clone() too, which means it has the same problem. The
data can be modified by other xdp prog in the destination NIC if we use
bpf_clone_redirect().
So maybe this is the default logic, and I'm not sure if this patch can break the
existing users :/
I think for use cases where we are using bpf_clone_redirect() to use
one clone for inspection this would add an unnecessary copy. Maybe
adding *_copy() variants instead of changing the *_clone() would be
better? That way we wouldn't be changing the behavior for existing
consumers and the naming would be consistent with the skb_* methods.
Agree. It's not a good idea to change the logic of the existing API. Or
maybe we can add a BPF_F_CLONE flag for the existing API.
But more importantly, is there an actual use case for the kind of API
that the modified selftest requires? Nobody until now has considered
the existing behavior to be a problem.
Agree too. Obviously, this is not a bug. For the use case in the commit
log, it's something that can be fixed by the user themself. If we need
modify the MAC, we'd better attach a BPF program for all the egress
device in the devmap.
[...]
Thanks for all the comments.
I agree this should not be treated as a generic skb_clone() issue, and I
understand the concern about changing existing clone/shared-data semantics.
The case I was trying to address is narrower: generic XDP devmap
broadcast/multi redirect with per-devmap-entry egress programs. The use case is
not newly introduced by this patch. The existing xdp_veth_egress test already
attaches xdp_devmap_prog to all egress devmap entries, and that program rewrites
eth->h_source based on ctx->egress_ifindex. My change only strengthens the test
from checking that the observed MAC is not equal to a single magic MAC to
checking that each destination observes the MAC selected for its own egress
ifindex.
So attaching an egress program to all devmap entries is already what this test
does. The SKB_MODE failure happens because the cloned skbs can still share the
packet data, and a later per-egress rewrite can be observed by another
destination.
That said, I see the concern that this may need a clearer semantic boundary
between clone and copy behavior. I will also check the last_dst path pointed out
by Sashiko before deciding whether this should be handled as a bug fix, or
whether it needs a separate explicit copy/isolated redirect semantic instead.
I agree there's a real concern here. For context, native XDP already
makes a full copy
of the frame for every broadcast destination:
dev_map_enqueue_multi -> dev_map_enqueue_clone -> xdpf_clone
so on the native path each destination gets its own independent buffer.
The generic path is the odd one out, because skb_clone() shares the
underlying data.
I think this rationale should be spelled out in the commit message.
Regarding performance: generic XDP is already a slow path — the ingress
side already
linearizes / expands the skb (pskb_expand_head() in
netif_receive_generic_xdp())
so the extra copy here is not a concern.
However, there is one problem with the patch as it stands.
The last destination is transmitted using the original skb directly,
without unsharing it.
So with two destinations where the first one has no egress XDP program
and the second (i.e. the last one) does,
the egress modification on the second destination will corrupt the
packet sent to the first.
This could be addressed by doing the skb_unshare() inside
dev_map_generic_redirect() instead ?
For the non-broadcast (existing) path the skb is not cloned, so
skb_unshare() simply returns it unchanged.
