On Tue, Jul 18, 2023 at 4:06 PM Jakub Kicinski <kuba@xxxxxxxxxx> wrote:
>
> On Tue, 18 Jul 2023 13:34:06 -0700 Alexei Starovoitov wrote:
> > > Direct packet access via skb->data is there for those who want high
> > > speed 🤷️
> >
> > skb->data/data_end approach unfortunately doesn't work that well.
> > Too much verifier fighting. That's why dynptr was introduced.
>
> I wish Daniel told us more about the use case.
>
> > > My worry is that people will think that whether the buffer is needed or
> > > not depends on _their program_, rather than on the underlying platform.
> > > So if it works in testing without the buffer - the buffer must not be
> > > required for their use case.
> >
> > Are you concerned about bpf progs breaking this way?
>
> Both, BPF progs breaking and netdev code doing things which don't make
> sense. But I won't argue too hard about the former, i.e. the BPF API.
>
> > I thought you're worried about the driver misusing
> > skb_header_pointer() with buffer==NULL.
> >
> > We can remove !buffer check as in the attached patch,
> > but I don't quite see how it would improve driver quality.
>
> The drivers may not be pretty but they aren't buggy AFAICT.
>
> > [0001-bpf-net-Introduce-skb_pointer_if_linear.patch application/octet-stream (2873 bytes)]
>
> Or we can simply pretend we don't have the skb:
>
> diff --git a/include/linux/skbuff.h b/include/linux/skbuff.h
> index 91ed66952580..217447f01d56 100644
> --- a/include/linux/skbuff.h
> +++ b/include/linux/skbuff.h
> @@ -4023,7 +4023,7 @@ __skb_header_pointer(const struct sk_buff *skb, int offset, int len,
> if (likely(hlen - offset >= len))
> return (void *)data + offset;
>
> - if (!skb || !buffer || unlikely(skb_copy_bits(skb, offset, buffer, len) < 0))
> + if (!skb || unlikely(skb_copy_bits(skb, offset, buffer, len) < 0))
> return NULL;
>
> return buffer;
> diff --git a/kernel/bpf/helpers.c b/kernel/bpf/helpers.c
> index 9e80efa59a5d..8bc4622cc1df 100644
> --- a/kernel/bpf/helpers.c
> +++ b/kernel/bpf/helpers.c
> @@ -2239,7 +2239,13 @@ __bpf_kfunc void *bpf_dynptr_slice(const struct bpf_dynptr_kern *ptr, u32 offset
> case BPF_DYNPTR_TYPE_RINGBUF:
> return ptr->data + ptr->offset + offset;
> case BPF_DYNPTR_TYPE_SKB:
> - return skb_header_pointer(ptr->data, ptr->offset + offset, len, buffer__opt);
> + {
> + const struct sk_buff *skb = ptr->data;
> +
> + return __skb_header_pointer(NULL, ptr->offset + offset, len,
> + skb->data, skb_headlen(skb),
> + buffer__opt);
> + }
Which would encourage bnxt-like hacks.
I don't like it tbh.
At least skb_pointer_if_linear() has a clear meaning.
It's more run-time overhead, since buffer__opt is checked early,
but that's ok.
