Jakub Sitnicki wrote:
> On Sat, Sep 26, 2020 at 06:27 AM CEST, John Fastabend wrote:
> > This implements a new helper skb_adjust_room() so users can push/pop
> > extra bytes from a BPF_SK_SKB_STREAM_VERDICT program.
> >
> > Some protocols may include headers and other information that we may
> > not want to include when doing a redirect from a BPF_SK_SKB_STREAM_VERDICT
> > program. One use case is to redirect TLS packets into a receive socket
> > that doesn't expect TLS data. In TLS case the first 13B or so contain the
> > protocol header. With KTLS the payload is decrypted so we should be able
> > to redirect this to a receiving socket, but the receiving socket may not
> > be expecting to receive a TLS header and discard the data. Using the
> > above helper we can pop the header off and put an appropriate header on
> > the payload. This allows for creating a proxy between protocols without
> > extra hops through the stack or userspace.
>
> This is useful stuff. Apart from the TLS use-case, you might want to pop
> off proxy headers like PROXY v1/v2 (CC Marek):
>
> https://www.haproxy.org/download/1.8/doc/proxy-protocol.txt
Great!
>
> >
> > So in order to fix this case add skb_adjust_room() so users can strip the
> > header. After this the user can strip the header and an unmodified receiver
> > thread will work correctly when data is redirected into the ingress path
> > of a sock.
> >
> > Signed-off-by: John Fastabend <john.fastab...@gmail.com>
> > ---
> > net/core/filter.c | 51
> > +++++++++++++++++++++++++++++++++++++++++++++++++++
> > 1 file changed, 51 insertions(+)
> >
> > diff --git a/net/core/filter.c b/net/core/filter.c
> > index 4d8dc7a31a78..d232358f1dcd 100644
> > --- a/net/core/filter.c
> > +++ b/net/core/filter.c
> > @@ -76,6 +76,7 @@
> > #include <net/bpf_sk_storage.h>
> > #include <net/transp_v6.h>
> > #include <linux/btf_ids.h>
> > +#include <net/tls.h>
> >
> > static const struct bpf_func_proto *
> > bpf_sk_base_func_proto(enum bpf_func_id func_id);
> > @@ -3218,6 +3219,53 @@ static u32 __bpf_skb_max_len(const struct sk_buff
> > *skb)
> > SKB_MAX_ALLOC;
> > }
> >
> > +BPF_CALL_4(sk_skb_adjust_room, struct sk_buff *, skb, s32, len_diff,
> > + u32, mode, u64, flags)
> > +{
> > + unsigned int len_diff_abs = abs(len_diff);
> > + bool shrink = len_diff < 0;
> > + int ret = 0;
> > +
> > + if (unlikely(flags))
> > + return -EINVAL;
> > + if (unlikely(len_diff_abs > 0xfffU))
> > + return -EFAULT;
> > +
> > + if (!shrink) {
> > + unsigned int grow = len_diff;
> > +
> > + ret = skb_cow(skb, grow);
> > + if (likely(!ret)) {
> > + __skb_push(skb, len_diff_abs);
> > + memset(skb->data, 0, len_diff_abs);
> > + }
> > + } else {
> > + /* skb_ensure_writable() is not needed here, as we're
> > + * already working on an uncloned skb.
> > + */
>
> I'm trying to digest the above comment. What if:
I'll delete the comment its not accurate. We happily write headers
from verdict programs today. Do you have a specific concern or
just noticing I was a bit careless and cut'n'pasted an incorrect
comment around.
>
> static int __strp_recv(…)
> {
> …
> while (eaten < orig_len) {
> /* Always clone since we will consume something */
> skb = skb_clone(orig_skb, GFP_ATOMIC);
> …
> head = strp->skb_head;
> if (!head) {
> head = skb;
> …
> } else {
> …
> }
> …
> /* Give skb to upper layer */
> strp->cb.rcv_msg(strp, head); // → sk_psock_init_strp
> …
> }
> …
> }
>
> That looks like a code path where we pass a cloned SKB.
Right but its there to cover the sk_eat_skb in tcp_read_sock()
otherwise
sk_eat_skb() -> __kfree_skb() -> skb_release_all()
would go all the way to page_frag_free().
>
> > + if (unlikely(!pskb_may_pull(skb, len_diff_abs)))
> > + return -ENOMEM;
> > + __skb_pull(skb, len_diff_abs);
> > + }
> > + bpf_compute_data_end_sk_skb(skb);
> > + if (tls_sw_has_ctx_rx(skb->sk)) {
> > + struct strp_msg *rxm = strp_msg(skb);
> > +
> > + rxm->full_len += len_diff;
> > + }
> > + return ret;
> > +}
>
> [...]
