On Wed, Mar 25, 2020 at 3:29 AM Lorenz Bauer <lmb@xxxxxxxxxxxxxx> wrote:
>
> On Wed, 25 Mar 2020 at 05:58, Joe Stringer <joe@xxxxxxxxxxx> wrote:
> >
> > Avoid taking a reference on listen sockets by checking the socket type
> > in the sk_assign and in the corresponding skb_steal_sock() code in the
> > the transport layer, and by ensuring that the prefetch free (sock_pfree)
> > function uses the same logic to check whether the socket is refcounted.
> >
> > Suggested-by: Martin KaFai Lau <kafai@xxxxxx>
> > Signed-off-by: Joe Stringer <joe@xxxxxxxxxxx>
> > ---
> > v2: Initial version
> > ---
> > include/net/sock.h | 25 +++++++++++++++++--------
> > net/core/filter.c | 6 +++---
> > net/core/sock.c | 3 ++-
> > 3 files changed, 22 insertions(+), 12 deletions(-)
> >
> > diff --git a/include/net/sock.h b/include/net/sock.h
> > index 1ca2e808cb8e..3ec1865f173e 100644
> > --- a/include/net/sock.h
> > +++ b/include/net/sock.h
> > @@ -2533,6 +2533,21 @@ skb_sk_is_prefetched(struct sk_buff *skb)
> > return skb->destructor == sock_pfree;
> > }
> >
> > +/* This helper checks if a socket is a full socket,
> > + * ie _not_ a timewait or request socket.
> > + */
> > +static inline bool sk_fullsock(const struct sock *sk)
> > +{
> > + return (1 << sk->sk_state) & ~(TCPF_TIME_WAIT | TCPF_NEW_SYN_RECV);
> > +}
> > +
> > +static inline bool
> > +sk_is_refcounted(struct sock *sk)
> > +{
> > + /* Only full sockets have sk->sk_flags. */
> > + return !sk_fullsock(sk) || !sock_flag(sk, SOCK_RCU_FREE);
> > +}
> > +
> > /**
> > * skb_steal_sock
> > * @skb to steal the socket from
> > @@ -2545,6 +2560,8 @@ skb_steal_sock(struct sk_buff *skb, bool *refcounted)
> > struct sock *sk = skb->sk;
> >
> > *refcounted = true;
> > + if (skb_sk_is_prefetched(skb))
> > + *refcounted = sk_is_refcounted(sk);
> > skb->destructor = NULL;
> > skb->sk = NULL;
> > return sk;
> > @@ -2553,14 +2570,6 @@ skb_steal_sock(struct sk_buff *skb, bool *refcounted)
> > return NULL;
> > }
> >
> > -/* This helper checks if a socket is a full socket,
> > - * ie _not_ a timewait or request socket.
> > - */
> > -static inline bool sk_fullsock(const struct sock *sk)
> > -{
> > - return (1 << sk->sk_state) & ~(TCPF_TIME_WAIT | TCPF_NEW_SYN_RECV);
> > -}
> > -
> > /* Checks if this SKB belongs to an HW offloaded socket
> > * and whether any SW fallbacks are required based on dev.
> > * Check decrypted mark in case skb_orphan() cleared socket.
> > diff --git a/net/core/filter.c b/net/core/filter.c
> > index 0fada7fe9b75..997b8606167e 100644
> > --- a/net/core/filter.c
> > +++ b/net/core/filter.c
> > @@ -5343,8 +5343,7 @@ static const struct bpf_func_proto bpf_sk_lookup_udp_proto = {
> >
> > BPF_CALL_1(bpf_sk_release, struct sock *, sk)
> > {
> > - /* Only full sockets have sk->sk_flags. */
> > - if (!sk_fullsock(sk) || !sock_flag(sk, SOCK_RCU_FREE))
> > + if (sk_is_refcounted(sk))
> > sock_gen_put(sk);
> > return 0;
> > }
> > @@ -5870,7 +5869,8 @@ BPF_CALL_3(bpf_sk_assign, struct sk_buff *, skb, struct sock *, sk, u64, flags)
> > return -ESOCKTNOSUPPORT;
> > if (unlikely(dev_net(skb->dev) != sock_net(sk)))
> > return -ENETUNREACH;
> > - if (unlikely(!refcount_inc_not_zero(&sk->sk_refcnt)))
> > + if (sk_is_refcounted(sk) &&
> > + unlikely(!refcount_inc_not_zero(&sk->sk_refcnt)))
> > return -ENOENT;
> >
> > skb_orphan(skb);
> > diff --git a/net/core/sock.c b/net/core/sock.c
> > index cfaf60267360..a2ab79446f59 100644
> > --- a/net/core/sock.c
> > +++ b/net/core/sock.c
> > @@ -2076,7 +2076,8 @@ EXPORT_SYMBOL(sock_efree);
> > */
> > void sock_pfree(struct sk_buff *skb)
> > {
> > - sock_edemux(skb);
> > + if (sk_is_refcounted(skb->sk))
> > + sock_edemux(skb);
>
> sock_edemux calls sock_gen_put, which is also called by
> bpf_sk_release. Is it worth teaching sock_gen_put about
> sk_fullsock, and dropping the other helpers? I was considering this
> when fixing up sk_release, but then forgot
> about it.
I like the idea, but I'm concerned about breaking things outside the
focus of this new helper if the skb_sk_is_prefetched() function from
patch 1 is allowed to return true for sockets other than the ones
assigned from the bpf_sk_assign() helper. At a glance there's users of
sock_efree (which sock_edemux can be defined to) like netem_enqueue()
which may inadvertently trigger unexpected paths here. I think it's
more explicit so more obviously correct if the destructor pointer used
in this series is unique compared to other paths, even if the
underlying code is the same.
