Re: [RFC bpf-next v1 3/7] bpf: Support pulling non-linear xdp data

[Stanislav Fomichev <stfomichev@gmail.com>]

On 08/25, Amery Hung wrote:
> On Mon, Aug 25, 2025 at 2:29 PM Stanislav Fomichev <stfomichev@gmail.com> wrote:
> >
> > On 08/25, Amery Hung wrote:
> > > Add kfunc, bpf_xdp_pull_data(), to support pulling data from xdp
> > > fragments. Similar to bpf_skb_pull_data(), bpf_xdp_pull_data() makes
> > > the first len bytes of data directly readable and writable in bpf
> > > programs. If the "len" argument is larger than the linear data size,
> > > data in fragments will be copied to the linear region when there
> > > is enough room between xdp->data_end and xdp_data_hard_end(xdp),
> > > which is subject to driver implementation.
> > >
> > > A use case of the kfunc is to decapsulate headers residing in xdp
> > > fragments. It is possible for a NIC driver to place headers in xdp
> > > fragments. To keep using direct packet access for parsing and
> > > decapsulating headers, users can pull headers into the linear data
> > > area by calling bpf_xdp_pull_data() and then pop the header with
> > > bpf_xdp_adjust_head().
> > >
> > > An unused argument, flags is reserved for future extension (e.g.,
> > > tossing the data instead of copying it to the linear data area).
> > >
> > > Signed-off-by: Amery Hung <ameryhung@gmail.com>
> > > ---
> > >  net/core/filter.c | 52 +++++++++++++++++++++++++++++++++++++++++++++++
> > >  1 file changed, 52 insertions(+)
> > >
> > > diff --git a/net/core/filter.c b/net/core/filter.c
> > > index f0ee5aec7977..82d953e077ac 100644
> > > --- a/net/core/filter.c
> > > +++ b/net/core/filter.c
> > > @@ -12211,6 +12211,57 @@ __bpf_kfunc int bpf_sock_ops_enable_tx_tstamp(struct bpf_sock_ops_kern *skops,
> > >       return 0;
> > >  }
> > >
> > > +__bpf_kfunc int bpf_xdp_pull_data(struct xdp_md *x, u32 len, u64 flags)
> > > +{
> > > +     struct xdp_buff *xdp = (struct xdp_buff *)x;
> > > +     struct skb_shared_info *sinfo = xdp_get_shared_info_from_buff(xdp);
> > > +     void *data_end, *data_hard_end = xdp_data_hard_end(xdp);
> > > +     int i, delta, buff_len, n_frags_free = 0, len_free = 0;
> > > +
> > > +     buff_len = xdp_get_buff_len(xdp);
> > > +
> > > +     if (unlikely(len > buff_len))
> > > +             return -EINVAL;
> > > +
> > > +     if (!len)
> > > +             len = xdp_get_buff_len(xdp);
> >
> > Why not return -EINVAL here for len=0?
> >
> 
> I try to mirror the behavior of bpf_skb_pull_data() to reduce confusion here.

Ah, makes sense!

> > > +
> > > +     data_end = xdp->data + len;
> > > +     delta = data_end - xdp->data_end;
> > > +
> > > +     if (delta <= 0)
> > > +             return 0;
> > > +
> > > +     if (unlikely(data_end > data_hard_end))
> > > +             return -EINVAL;
> > > +
> > > +     for (i = 0; i < sinfo->nr_frags && delta; i++) {
> > > +             skb_frag_t *frag = &sinfo->frags[i];
> > > +             u32 shrink = min_t(u32, delta, skb_frag_size(frag));
> > > +
> > > +             memcpy(xdp->data_end + len_free, skb_frag_address(frag), shrink);
> >
> > skb_frag_address can return NULL for unreadable frags.
> 
> Is it safe to assume that drivers will ensure frags to be readable? It
> seems at least mlx5 does.
> 
> I did a quick check and found other xdp kfuncs using
> skb_frag_address() without checking the return.

The unreadable frags will always be unredabale to the host. This is TCP
device memory, the memory on the accelerators that is not mapped onto
the CPU. Any attempts to read that memory should gracefully error out.

Can you also pls fix that other one? (not as part of the series should
be ok)