bpf: restore skb->sk before pskb_trim() call

While testing a fix [1] in ___pskb_trim(), addressing the WARN_ON_ONCE()
in skb_try_coalesce() reported by Andrey, I found that we had an skb
with skb->sk set but no skb->destructor.

This invalidated heuristic found in commit 158f323b9868 ("net: adjust
skb->truesize in pskb_expand_head()") and in cited patch.

Considering the BUG_ON(skb->sk) we have in skb_orphan(), we should
restrain the temporary setting to a minimal section.

[1] https://patchwork.ozlabs.org/patch/755570/
    net: adjust skb->truesize in ___pskb_trim()

Fixes: 8f917bba0042 ("bpf: pass sk to helper functions")
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Willem de Bruijn <willemb@google.com>
Cc: Andrey Konovalov <andreyknvl@google.com>
Acked-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Alexei Starovoitov <ast@kernel.org>
Acked-by: Willem de Bruijn <willemb@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

net/core/filter.c

@@ -98,8 +98,8 @@ int sk_filter_trim_cap(struct sock *sk, struct sk_buff *skb, unsigned int cap)
 
 		skb->sk = sk;
 		pkt_len = bpf_prog_run_save_cb(filter->prog, skb);
-		err = pkt_len ? pskb_trim(skb, max(cap, pkt_len)) : -EPERM;
 		skb->sk = save_sk;
+		err = pkt_len ? pskb_trim(skb, max(cap, pkt_len)) : -EPERM;
 	}
 	rcu_read_unlock();