Re: [PATCH bpf-next] net: bpf: support direct packet access in tracing program

[Menglong Dong]

On Thu, Apr 28, 2022 at 4:08 AM Andrii Nakryiko
<andrii.nakryiko@xxxxxxxxx> wrote:
>
> On Wed, Apr 27, 2022 at 12:08 AM <menglong8.dong@xxxxxxxxx> wrote:
> >
> > From: Menglong Dong <imagedong@xxxxxxxxxxx>
> >
> > For now, eBPF program of type TRACING is able to access the arguments
> > of the function or raw_tracepoint directly, which makes data access
> > easier and efficient. And we can also output the raw data in skb to
> > user space in tracing program by bpf_skb_output() helper.
> >
> > However, we still can't access the packet data in 'struct sk_buff'
> > directly and have to use the helper bpf_probe_read_kernel() to analyse
> > packet data.
> >
> > Network tools, which based on eBPF TRACING, often do packet analyse
> > works in tracing program for filtering, statistics, etc. For example,
> > we want to trace abnormal skb free through 'kfree_skb' tracepoint with
> > special ip address or tcp port.
> >
> > In this patch, 2 helpers are introduced: bpf_skb_get_header() and
> > bpf_skb_get_end(). The pointer returned by bpf_skb_get_header() has
> > the same effect with the 'data' in 'struct __sk_buff', and
> > bpf_skb_get_end() has the same effect with the 'data_end'.
> >
> > Therefore, we can now access packet data in tracing program in this
> > way:
> >
> >   SEC("fentry/icmp_rcv")
> >   int BPF_PROG(tracing_open, struct sk_buff* skb)
> >   {
> >         void *data, *data_end;
> >         struct ethhdr *eth;
> >
> >         data = bpf_skb_get_header(skb, BPF_SKB_HEADER_MAC);
> >         data_end = bpf_skb_get_end(skb);
> >
> >         if (!data || !data_end)
> >                 return 0;
> >
> >         if (data + sizeof(*eth) > data_end)
> >                 return 0;
> >
> >         eth = data;
> >         bpf_printk("proto:%d\n", bpf_ntohs(eth->h_proto));
> >
> >         return 0;
> >   }
> >
> > With any positive reply, I'll complete the selftests programs.
>
> See bpf_dynptr patches that Joanne is working on. That's an
> alternative mechanism for data/data_end and is going to be easier and
> more flexible to work with. It is the plan that once basic bpf_dynptr
> functionality lands, we'll have dynptr "constructors" for xdp_buff and
> sk_buff. I think it's a better path forward.
>

Yeah, bpf_dynptr seems to be a nice feature, which I didn't notice :/
Thanks for your explanation.

Menglong Dong

> >
> > Reviewed-by: Jiang Biao <benbjiang@xxxxxxxxxxx>
> > Reviewed-by: Hao Peng <flyingpeng@xxxxxxxxxxx>
> > Signed-off-by: Menglong Dong <imagedong@xxxxxxxxxxx>
> > ---
> >  include/linux/bpf.h      |  4 +++
> >  include/uapi/linux/bpf.h | 29 ++++++++++++++++++++
> >  kernel/bpf/verifier.c    |  6 +++++
> >  kernel/trace/bpf_trace.c | 58 ++++++++++++++++++++++++++++++++++++++++
> >  4 files changed, 97 insertions(+)
> >
>
> [...]