Jakub Kicinski <kuba@xxxxxxxxxx> wrote: > On Tue, 7 Jul 2020 12:55:08 +0800 wenxu@xxxxxxxxx wrote: > > From: wenxu <wenxu@xxxxxxxxx> > > > > Add nf_ct_frag_gather and Make nf_ct_frag6_gather elide the CB clear > > when packets are defragmented by connection tracking. This can make > > each subsystem such as br_netfilter, openvswitch, act_ct do defrag > > without restore the CB. > > This also avoid serious crashes and problems in ct subsystem. > > Because Some packet schedulers store pointers in the qdisc CB private > > area and parallel accesses to the SKB. > > > > This series following up > > http://patchwork.ozlabs.org/project/netdev/patch/1593422178-26949-1-git-send-email-wenxu@xxxxxxxxx/ > > > > patch1: add nf_ct_frag_gather elide the CB clear > > patch2: make nf_ct_frag6_gather elide the CB clear > > patch3: fix clobber qdisc_skb_cb in act_ct with defrag > > > > v2: resue some ip_defrag function in patch1 > > Florian, Cong - are you willing to venture an ack on these? Anyone? Nope, sorry. Reason is that I can't figure out the need for this series. Taking a huge step back: http://patchwork.ozlabs.org/project/netdev/patch/1593422178-26949-1-git-send-email-wenxu@xxxxxxxxx/ That patch looks ok to me: I understand the problem statement/commit message and I can see how its addressed. I don't understand why the CB clearing must be avoided. defrag assumes skb ownership -- e.g. it may realloc skb->data (calls pskb_may_pull), it calls skb_orphan(), etc. AFAICS, tcf_classify makes same assumption -- exclusive ownership and no parallel skb accesses. So, if in fact the "only" problem is the loss of qdisc_skb_cb(skb)->pkt_len, then the other patch looks ok to me. If we indeed have parallel access, then I do not understand how avoiding the memsets in the defrag path makes things any better (see above wrt. skb pull and the like). As for these patches here: - if (!(IPCB(skb)->flags & IPSKB_FRAG_COMPLETE) && + if ((ignore_skb_cb || !(IPCB(skb)->flags & IPSKB_FRAG_COMPLETE)) && This is very questionable, we take different code path depending on call site. Why is it okay to unconditionally take this branch for act_ct case (ignore_skb_cb set)?
