Re: [Patch bpf-next v6 4/8] skmsg: move sk_redir from TCP_SKB_CB to skb

[Jakub Sitnicki <jakub@cloudflare.com>]

On Mon, Feb 22, 2021 at 08:27 PM CET, Cong Wang wrote:
> On Mon, Feb 22, 2021 at 4:20 AM Jakub Sitnicki <jakub@cloudflare.com> wrote:
>>
>> On Sat, Feb 20, 2021 at 06:29 AM CET, Cong Wang wrote:
>> > From: Cong Wang <cong.wang@bytedance.com>
>> >
>> > Currently TCP_SKB_CB() is hard-coded in skmsg code, it certainly
>> > does not work for any other non-TCP protocols. We can move them to
>> > skb ext, but it introduces a memory allocation on fast path.
>> >
>> > Fortunately, we only need to a word-size to store all the information,
>> > because the flags actually only contains 1 bit so can be just packed
>> > into the lowest bit of the "pointer", which is stored as unsigned
>> > long.
>> >
>> > Inside struct sk_buff, '_skb_refdst' can be reused because skb dst is
>> > no longer needed after ->sk_data_ready() so we can just drop it.
>> >
>> > Cc: Daniel Borkmann <daniel@iogearbox.net>
>> > Cc: Jakub Sitnicki <jakub@cloudflare.com>
>> > Cc: Lorenz Bauer <lmb@cloudflare.com>
>> > Acked-by: John Fastabend <john.fastabend@gmail.com>
>> > Signed-off-by: Cong Wang <cong.wang@bytedance.com>
>> > ---
>>
>> LGTM. I have some questions (below) that would help me confirm if I
>> understand the changes, and what could be improved, if anything.
>>
>> Acked-by: Jakub Sitnicki <jakub@cloudflare.com>
>>
>> >  include/linux/skbuff.h |  3 +++
>> >  include/linux/skmsg.h  | 35 +++++++++++++++++++++++++++++++++++
>> >  include/net/tcp.h      | 19 -------------------
>> >  net/core/skmsg.c       | 32 ++++++++++++++++++++------------
>> >  net/core/sock_map.c    |  8 ++------
>> >  5 files changed, 60 insertions(+), 37 deletions(-)
>> >
>> > diff --git a/include/linux/skbuff.h b/include/linux/skbuff.h
>> > index 6d0a33d1c0db..bd84f799c952 100644
>> > --- a/include/linux/skbuff.h
>> > +++ b/include/linux/skbuff.h
>> > @@ -755,6 +755,9 @@ struct sk_buff {
>> >                       void            (*destructor)(struct sk_buff *skb);
>> >               };
>> >               struct list_head        tcp_tsorted_anchor;
>> > +#ifdef CONFIG_NET_SOCK_MSG
>> > +             unsigned long           _sk_redir;
>> > +#endif
>> >       };
>> >
>> >  #if defined(CONFIG_NF_CONNTRACK) || defined(CONFIG_NF_CONNTRACK_MODULE)
>> > diff --git a/include/linux/skmsg.h b/include/linux/skmsg.h
>> > index e3bb712af257..fc234d507fd7 100644
>> > --- a/include/linux/skmsg.h
>> > +++ b/include/linux/skmsg.h
>> > @@ -459,4 +459,39 @@ static inline bool sk_psock_strp_enabled(struct sk_psock *psock)
>> >               return false;
>> >       return !!psock->saved_data_ready;
>> >  }
>> > +
>> > +#if IS_ENABLED(CONFIG_NET_SOCK_MSG)
>> > +static inline bool skb_bpf_ingress(const struct sk_buff *skb)
>> > +{
>> > +     unsigned long sk_redir = skb->_sk_redir;
>> > +
>> > +     return sk_redir & BPF_F_INGRESS;
>> > +}
>> > +
>> > +static inline void skb_bpf_set_ingress(struct sk_buff *skb)
>> > +{
>> > +     skb->_sk_redir |= BPF_F_INGRESS;
>> > +}
>> > +
>> > +static inline void skb_bpf_set_redir(struct sk_buff *skb, struct sock *sk_redir,
>> > +                                  bool ingress)
>> > +{
>> > +     skb->_sk_redir = (unsigned long)sk_redir;
>> > +     if (ingress)
>> > +             skb->_sk_redir |= BPF_F_INGRESS;
>> > +}
>> > +
>> > +static inline struct sock *skb_bpf_redirect_fetch(const struct sk_buff *skb)
>> > +{
>> > +     unsigned long sk_redir = skb->_sk_redir;
>> > +
>> > +     sk_redir &= ~0x1UL;
>>
>> We're using the enum when setting the bit flag, but a hardcoded constant
>> when masking it. ~BPF_F_INGRESS would be more consistent here.
>
> Well, here we need a mask, not a bit, but we don't have a mask yet,
> hence I just use hard-coded 0x1. Does #define BPF_F_MASK 0x1UL
> look any better?

Based on what I've seen around, mask for sanitizing tagged pointers is
usually derived from the flag(s). For instance:

#define SKB_DST_NOREF	1UL
#define SKB_DST_PTRMASK	~(SKB_DST_NOREF)

#define SK_USER_DATA_NOCOPY	1UL
#define SK_USER_DATA_BPF	2UL	/* Managed by BPF */
#define SK_USER_DATA_PTRMASK	~(SK_USER_DATA_NOCOPY | SK_USER_DATA_BPF)

Using ~(BPF_F_INGRESS) expression would be like substituting mask
definition.

[..]

>> > diff --git a/include/net/tcp.h b/include/net/tcp.h
>> > index 947ef5da6867..075de26f449d 100644
>> > --- a/include/net/tcp.h
>> > +++ b/include/net/tcp.h
>> > @@ -883,30 +883,11 @@ struct tcp_skb_cb {
>> >                       struct inet6_skb_parm   h6;
>> >  #endif
>> >               } header;       /* For incoming skbs */
>> > -             struct {
>> > -                     __u32 flags;
>> > -                     struct sock *sk_redir;
>> > -             } bpf;
>> >       };
>> >  };
>> >
>> >  #define TCP_SKB_CB(__skb)    ((struct tcp_skb_cb *)&((__skb)->cb[0]))
>> >
>> > -static inline bool tcp_skb_bpf_ingress(const struct sk_buff *skb)
>> > -{
>> > -     return TCP_SKB_CB(skb)->bpf.flags & BPF_F_INGRESS;
>> > -}
>> > -
>> > -static inline struct sock *tcp_skb_bpf_redirect_fetch(struct sk_buff *skb)
>> > -{
>> > -     return TCP_SKB_CB(skb)->bpf.sk_redir;
>> > -}
>> > -
>> > -static inline void tcp_skb_bpf_redirect_clear(struct sk_buff *skb)
>> > -{
>> > -     TCP_SKB_CB(skb)->bpf.sk_redir = NULL;
>> > -}
>> > -
>> >  extern const struct inet_connection_sock_af_ops ipv4_specific;
>> >
>> >  #if IS_ENABLED(CONFIG_IPV6)
>> > diff --git a/net/core/skmsg.c b/net/core/skmsg.c
>> > index 2d8bbb3fd87c..05b5af09ff42 100644
>> > --- a/net/core/skmsg.c
>> > +++ b/net/core/skmsg.c
>> > @@ -494,6 +494,8 @@ static int sk_psock_skb_ingress_self(struct sk_psock *psock, struct sk_buff *skb
>> >  static int sk_psock_handle_skb(struct sk_psock *psock, struct sk_buff *skb,
>> >                              u32 off, u32 len, bool ingress)
>> >  {
>> > +     skb_bpf_redirect_clear(skb);
>>
>> This is called to avoid leaking state in skb->_skb_refdst. Correct?
>
> This is to teach kfree_skb() not to consider it as a valid _skb_refdst.

OK

>
>>
>> I'm wondering why we're doing it every time sk_psock_handle_skb() gets
>> invoked from the do/while loop in sk_psock_backlog(), instead of doing
>> it once after reading ingress flag with skb_bpf_ingress()?
>
> It should also work, I don't see much difference here, as we almost
> always process a full skb, that is, ret == skb->len.

OK

>
>
>>
>> > +
>> >       if (!ingress) {
>> >               if (!sock_writeable(psock->sk))
>> >                       return -EAGAIN;
>> > @@ -525,7 +527,7 @@ static void sk_psock_backlog(struct work_struct *work)
>> >               len = skb->len;
>> >               off = 0;
>> >  start:
>> > -             ingress = tcp_skb_bpf_ingress(skb);
>> > +             ingress = skb_bpf_ingress(skb);
>> >               do {
>> >                       ret = -EIO;
>> >                       if (likely(psock->sk->sk_socket))
>> > @@ -631,7 +633,12 @@ void __sk_psock_purge_ingress_msg(struct sk_psock *psock)
>> >
>> >  static void sk_psock_zap_ingress(struct sk_psock *psock)
>> >  {
>> > -     __skb_queue_purge(&psock->ingress_skb);
>> > +     struct sk_buff *skb;
>> > +
>> > +     while ((skb = __skb_dequeue(&psock->ingress_skb)) != NULL) {
>> > +             skb_bpf_redirect_clear(skb);
>>
>> I believe we clone the skb before enqueuing it psock->ingress_skb.
>> Clone happens either in sk_psock_verdict_recv() or in __strp_recv().
>> There are not other users holding a ref, so clearing the redirect seems
>> unneeded. Unless I'm missing something?
>
> Yes, skb dst is also cloned:
>
>  980 static void __copy_skb_header(struct sk_buff *new, const struct
> sk_buff *old)
>  981 {
>  982         new->tstamp             = old->tstamp;
>  983         /* We do not copy old->sk */
>  984         new->dev                = old->dev;
>  985         memcpy(new->cb, old->cb, sizeof(old->cb));
>  986         skb_dst_copy(new, old);
>
> Also, if without this, dst_release() would complain again. I was not smart
> enough to add it in the beginning, dst_release() taught me this lesson. ;)

OK, I think I follow you now.

Alternatively we could clear _skb_refdest after clone, but before
enqueuing the skb in ingress_skb. And only for when we're redirecting.

I believe that would be in sk_psock_skb_redirect, right before skb_queue_tail.

>
>>
>> > +             kfree_skb(skb);
>> > +     }
>> >       __sk_psock_purge_ingress_msg(psock);
>> >  }
>> >
>> > @@ -752,7 +759,7 @@ static void sk_psock_skb_redirect(struct sk_buff *skb)
>> >       struct sk_psock *psock_other;
>> >       struct sock *sk_other;
>> >
>> > -     sk_other = tcp_skb_bpf_redirect_fetch(skb);
>> > +     sk_other = skb_bpf_redirect_fetch(skb);
>> >       /* This error is a buggy BPF program, it returned a redirect
>> >        * return code, but then didn't set a redirect interface.
>> >        */
>> > @@ -802,9 +809,10 @@ int sk_psock_tls_strp_read(struct sk_psock *psock, struct sk_buff *skb)
>> >                * TLS context.
>> >                */
>> >               skb->sk = psock->sk;
>> > -             tcp_skb_bpf_redirect_clear(skb);
>> > +             skb_dst_drop(skb);
>> > +             skb_bpf_redirect_clear(skb);
>>
>> After skb_dst_drop(), skb->_skb_refdst is clear. So it seems the
>> redirect_clear() is not needed. But I'm guessing it is being invoked
>> to communicate the intention?
>
> Technically true, but I prefer to call them explicitly, not to rely on the
> fact skb->_skb_refdst shares the same storage with skb->_sk_redir,
> which would also require some comments to explain.
>

OK