On 7/29/25 3:06 AM, Mahe Tardy wrote:
On Mon, Jul 28, 2025 at 06:05:26PM -0700, Martin KaFai Lau wrote:
On 7/28/25 2:43 AM, Mahe Tardy wrote:
This is needed in the context of Tetragon to provide improved feedback
(in contrast to just dropping packets) to east-west traffic when blocked
by policies using cgroup_skb programs.
This reuse concepts from netfilter reject target codepath with the
differences that:
* Packets are cloned since the BPF user can still return SK_PASS from
the cgroup_skb progs and the current skb need to stay untouched
This needs more details. Which field(s) of the skb are changed by the kfunc,
the skb_dst_set in ip[6]_route_reply_fetch_dst() and/or the code path in the
icmp[v6]_send() ?
Okay I can add that: "ip[6]_route_reply_fetch_dst set the dst of the skb
by using the saddr as a daddr and routing it", I don't think
icmp[v6]_send touches the skb?
I also don't think icmp[v6]_send touches the skb. I am still not sure if
ip[6]_route_reply_fetch_dst is needed.
(cgroup_skb hooks only allow read-only skb payload).
* Since cgroup_skb programs are called late in the stack, checksums do
not need to be computed or verified, and IPv4 fragmentation does not
need to be checked (ip_local_deliver should take care of that
earlier).
Signed-off-by: Mahe Tardy <mahe.tardy@xxxxxxxxx>
---
net/core/filter.c | 61 +++++++++++++++++++++++++++++++++++++++++++++++
1 file changed, 61 insertions(+)
diff --git a/net/core/filter.c b/net/core/filter.c
index 7a72f766aacf..050872324575 100644
--- a/net/core/filter.c
+++ b/net/core/filter.c
@@ -85,6 +85,10 @@
#include <linux/un.h>
#include <net/xdp_sock_drv.h>
#include <net/inet_dscp.h>
+#include <linux/icmp.h>
+#include <net/icmp.h>
+#include <net/route.h>
+#include <net/ip6_route.h>
#include "dev.h"
@@ -12148,6 +12152,53 @@ __bpf_kfunc int bpf_sock_ops_enable_tx_tstamp(struct bpf_sock_ops_kern *skops,
return 0;
}
+__bpf_kfunc int bpf_icmp_send_unreach(struct __sk_buff *__skb, int code)
+{
+ struct sk_buff *skb = (struct sk_buff *)__skb;
+ struct sk_buff *nskb;
+
+ switch (skb->protocol) {
+ case htons(ETH_P_IP):
+ if (code < 0 || code > NR_ICMP_UNREACH)
+ return -EINVAL;
+
+ nskb = skb_clone(skb, GFP_ATOMIC);
+ if (!nskb)
+ return -ENOMEM;
+
+ if (ip_route_reply_fetch_dst(nskb) < 0) {
+ kfree_skb(nskb);
+ return -EHOSTUNREACH;
+ }
+
+ icmp_send(nskb, ICMP_DEST_UNREACH, code, 0);
+ kfree_skb(nskb);
+ break;
+#if IS_ENABLED(CONFIG_IPV6)
+ case htons(ETH_P_IPV6):
+ if (code < 0 || code > ICMPV6_REJECT_ROUTE)
+ return -EINVAL;
+
+ nskb = skb_clone(skb, GFP_ATOMIC);
+ if (!nskb)
+ return -ENOMEM;
+
+ if (ip6_route_reply_fetch_dst(nskb) < 0) {
From a very quick look at icmpv6_send(), it does its own route lookup. I
haven't looked at the v4 yet.
I am likely missing some details. Can you explain why it needs to do a
lookup before calling icmpv6_send()?
From my understanding, I need to do this to invert the daddr with the
saddr to send the unreach message back to the sender.
From looking at how fl6.{daddr,saddr} are filled and passed to
icmpv6_route_lookup in icmpv6_send(), the icmpv6_send() should have done the
reverse/invert route lookup. I also don't see icmpv6_send uses the skb_dst() of
the original skb. Did I misread the code? The kfunc does not work without
ip[6]_route_reply_fetch_dst()? Again, I have not checked the v4 icmp_send. fwiw,
the selftest should have both v4 and v6 test.
Note that at cgroup/egress, the skb->_skb_refdst should have been set.
The same should be true for cgroup/ingress for inet proto but it seems
BPF_CGROUP_RUN_PROG_"INET"_INGRESS is not called from INET only now. e.g.
sk_filter() can be called from af_netlink. It seems like there is a bug.
+ kfree_skb(nskb);
+ return -EHOSTUNREACH;
+ }
+
+ icmpv6_send(nskb, ICMPV6_DEST_UNREACH, code, 0);
+ kfree_skb(nskb);
+ break;
+#endif
+ default:
+ return -EPROTONOSUPPORT;
+ }
+
+ return SK_DROP;
+}
+
