On 4/29/18 7:13 PM, David Ahern wrote:

The idea here is to fast pass packets that fit a supported profile and
are to be forwarded. Everything else should continue up the stack as it
has wider capabilities. The helper and XDP programs should make no
assumptions on what the broader kernel and userspace might be monitoring
or want to do with packets that can not be forwarded in the fast path.
This is very similar to hardware forwarding when it punts packets to the
CPU for control plane assistance.

Thinking about this some more and how to return more information to the
bpf program about the FIB lookup.

bpf_fib_lookup struct is 64-bytes. It can not be expanded without
hurting performance. I could do another union on an input parameter and
return flags indicating why the returned index is 0. Something like this:

diff --git a/include/uapi/linux/bpf.h b/include/uapi/linux/bpf.h
index 360a1168c353..75591522444c 100644
--- a/include/uapi/linux/bpf.h
+++ b/include/uapi/linux/bpf.h

@@ -2314,6 +2314,12 @@ struct bpf_raw_tracepoint_args {
 #define BPF_FIB_LOOKUP_DIRECT  BIT(0)
 #define BPF_FIB_LOOKUP_OUTPUT  BIT(1)

+#define BPF_FIB_LKUP_RET_NO_FWD      BIT(0)  /* pkt is not fwded */
+#define BPF_FIB_LKUP_RET_UNSUPP_LWT  BIT(1)  /* fwd requires unsupp

encap */
+#define BPF_FIB_LKUP_RET_NO_NHDEV    BIT(2)  /* nh device does not exist */
+#define BPF_FIB_LKUP_RET_NO_NEIGH    BIT(3)  /* no neigh entry for nh */
+#define BPF_FIB_LKUP_RET_FRAG_NEEDED BIT(4)  /* pkt too big to fwd */
+
 struct bpf_fib_lookup {
        /* input */
        __u8    family;   /* network family, AF_INET, AF_INET6, AF_MPLS */

@@ -2325,7 +2331,11 @@ struct bpf_fib_lookup {

        /* total length of packet from network header - used for MTU

check */
        __u16   tot_len;
-       __u32   ifindex;  /* L3 device index for lookup */
+
+       union {
+               __u32   ifindex;   /* in: L3 device index for lookup */
+               __u32   ret_flags; /* out: BPF_FIB_LOOKUP_RET flags */
+       }

        union {
                /* inputs to lookup */


Similarly for the fib result, it could be returned with a union on say
family:
    union {
        __u8 family;   /* in: network family, AF_INET, AF_INET6, AF_MPLS */
        __u8 rt_type;  /* out: FIB lookup route type */
    };

Then if the fib result is -EINVAL/-EHOSTUNREACH/-EACCES, rt_type is set
to RTN_BLACKHOLE/RTN_UNREACHABLE/RTN_PROHIBIT allowing the XDP program
to make an informed decision on dropping the packet.

To avoid performance hits on the forwarding path, these return values
would *only* set if the ifindex returned is 0.