Re: Capture xdp packets in an fentry BPF hook

[Alexei Starovoitov <alexei.starovoitov@gmail.com>]

On Wed, Feb 19, 2020 at 03:38:40PM +0100, Eelco Chaudron wrote:
> Hi Alexei at al.,
> 
> I'm getting closer to finally have an xdpdump tool that uses the bpf
> fentry/fexit tracepoints, but I ran into a final hurdle...
> 
> To stuff the packet into a perf ring I'll need to use the
> bpf_perf_event_output(), but unfortunately, this is a program of trace type,
> and not XDP so the packet data is not added automatically :(
> 
> Secondly even trying to pass the actual packet data as a reference to
> bpf_perf_event_output() will not work as the verifier wants the data to be
> on the fp.
> 
> Even worse, the trace program gets the XDP info not thought the ctx, but
> trough the fentry/fexit input value, i.e.:
> 
> 	SEC("fentry/func")
> 	int BPF_PROG(trace_on_entry, struct xdp_buff *xdp)...
> 
> 	struct net_device {
> 	    int ifindex;
> 	} __attribute__((preserve_access_index));
> 
> 	struct xdp_rxq_info {
> 	    struct net_device *dev;
> 	    __u32 queue_index;
> 	} __attribute__((preserve_access_index));
> 
> 	struct xdp_buff {
> 	    void *data;
> 	    void *data_end;
> 	    void *data_meta;
> 	    void *data_hard_start;
> 	    unsigned long handle;
> 	    struct xdp_rxq_info *rxq;
> 	} __attribute__((preserve_access_index));
> 
> Hence even trying to copy in bytes to a local buffer is not allowed by the
> verifier, i.e. __u8 *data = (u8 *)(long)xdp->data;
> 
> Can you let me know how you envisioned a BPF entry hook to capture packets
> from XDP. Am I missing something, or is there something missing from the
> infrastructure?

Tracing of XDP is missing a helper similar to bpf_skb_output() for skb.
Its first arg will be 'struct xdp_buff *' and .arg1_type = ARG_PTR_TO_BTF_ID
then it will work similar to bpf_skb_output() in progs/kfree_skb.c.