On Wed, Mar 1, 2017 at 9:25 PM, Cong Wang <xiyou.wangc...@gmail.com> wrote: > On Wed, Mar 1, 2017 at 3:15 PM, Eric Dumazet <eduma...@google.com> wrote: >> On Wed, Mar 1, 2017 at 3:09 PM, Cong Wang <xiyou.wangc...@gmail.com> wrote: >> >>> >>> But I doubt skb_orphan() is the solution here, shouldn't we just >>> update sk->sk_wmem_alloc with skb->truesize changes? >> >> Is it worth it ? Apart from syszkaller I mean... >> >> We started with something that had a real impact on real workloads. >> >> 158f323b9868b59967ad96957c4ca388161be321 net: adjust skb->truesize in >> pskb_expand_head() >> >> Note that auditing the stack took me a while. > > I don't know how sk refcnt could work correctly without making > sk_wmem_alloc correctly. We certainly could just call skb_orphan() > is we don't need skb->sk any more, probably like the frag case, > but for this case, the neigh one, the skb's sitting in neigh->arp_queue > are not going to be freed unless in failed case, therefore skb->sk > should not be orphaned so early.
There is absolutely no issue in arp/nd case. Many skbs can sit there and it is fine. Same with skbs sitting a long time in a qdisc. Of course we try to not call skb_orphan() unless really needed. tcp_gso_segment() tries very hard to propagate skb ownership to the segments, but even something apparently easy like that took some patches before being done right. (for details : 0d08c42cf9a71530fef5ebcfe368f38f2dd0476f "tcp: gso: fix truesize tracking") conntrack reasm is mostly used in forwarding workloads, where skb->sk is already NULL. Are you thinking of a real workload where skb->sk _needs_ to be kept in ipv6 reasm ?
