netdev - Re: [PATCH] net: ip_finish_output_gso: If skb_gso_network_seglen exceeds MTU, allow segmentation for gre tunneled skbs

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [<thread-prev] [day] [month] [year] [list]

Date:   Fri, 19 Aug 2016 16:40:07 +0300
From:   Shmulik Ladkani <shmulik.ladkani@...il.com>
To:     Hannes Frederic Sowa <hannes@...essinduktion.org>
Cc:     David Miller <davem@...emloft.net>, wenxu@...oud.cn,
        kuznet@....inr.ac.ru, jmorris@...ei.org, kaber@...sh.net,
        yoshfuji@...ux-ipv6.org, netdev@...r.kernel.org,
        wenx05124561@....com
Subject: Re: [PATCH] net: ip_finish_output_gso: If skb_gso_network_seglen
 exceeds MTU, allow segmentation for gre tunneled skbs

Hi,

On Fri, 19 Aug 2016 11:20:40 +0200 Hannes Frederic Sowa <hannes@...essinduktion.org> wrote:
> >> Maybe we can change our criteria in the following manner:
> >>  
> >> -	if (skb_iif && proto == IPPROTO_UDP) {
> >> +	if (skb_iif && !(df & htons(IP_DF))) {
> >> 		IPCB(skb)->flags |= IPSKB_FRAG_SEGS;
> >>
> >> This way, any tunnel explicitly providing DF is NOT allowed to
> >> further fragment the resulting segments (leading to tx segments being
> >> dropped).
> >> Others tunnels, that do not care (e.g. vxlan and geneve, and probably
> >> ovs vport-gre, or other ovs encap vports, in df_default=false mode),
> >> will behave same for gso and non-gso.
> >>
> 
> I am really not sure...
> 
> Probably we have no other choice.

Further diving into this, seems the !IP_DF approach is more correct 
then the IPPROTO_UDP approach (WRT packets/segments arriving from other
interface, that exceed egress mtu):

vxlan/geneve:
  Both set df to zero.
  !IP_DF approach acts same as IPPROTO_UDP approach

vxlan/geneve in collect_md (e.g. OvS):
  They set df according to tun_flags & TUNNEL_DONT_FRAGMENT.
  IPPROTO_UDP approach:
    IPSKB_FRAG_SEGS gets set unconditionally.
    In case TUNNEL_DONT_FRAGMENT requested, non-gso get dropped
    due to IPSTATS_MIB_FRAGFAILS, whereas gso gets segmented+fragmented (!)
  !IP_DF approach:
    Aligned, both non-gso and gso gets dropped for TUNNEL_DONT_FRAGMENT.

ip_gre in collect_md (e.g. OvS):
  Sets df according to tun_flags & TUNNEL_DONT_FRAGMENT.
  IPPROTO_UDP approach:
    IPSKB_FRAG_SEGS is never set.
    Therefore in the case were df is NOT set, non-gso are fragged and
    passed, whereas gso gets dropped (!)
  !IP_DF approach:
    Non-gso vs gso aligned.

ip_gre in nopmtudisc:
  Will pass tnl_update_pmtu checks; Then, df inherrited from inner_iph
  (or stays unset if IFLA_GRE_IGNORE_DF specified).
  IPPROTO_UDP approach:
    IPSKB_FRAG_SEGS never set.
    Therefore in the case were df is NOT set, non-gso are fragged and
    passed, whereas gso gets dropped (!)
  !IP_DF approach:
    Aligned.
    
ip_gre in fou/gue mode in nopmtudisc:
  Assuming they pass tnl_update_pmtu checks; Then, df inherrited from
  inner_iph (or stays unset if IFLA_GRE_IGNORE_DF specified).
  IPPROTO_UDP approach:
    IPSKB_FRAG_SEGS gets always set (since proto==IPPROTO_UDP).
    In the case df is set, non-gso dropped by IPSTATS_MIB_FRAGFAILS,
    whereas gso gets segmented+fragmented (!)
  !IP_DF approach: 
    Aligned.

ip_gre in pmtudisc:
  Sets df to IP_DF.
  Non-gso will fail tnl_update_pmtu checks (gso should pass).
  IPPROTO_UDP approach:
    IPSKB_FRAG_SEGS never set. This leads the gso skbs to be eventually
    dropped. okay.
  !IP_DF approach:
    IPSKB_FRAG_SEGS not set, since IP_DF is true.
    This leads to gso skbs to be eventually dropped. okay.

(truely appreciate if you can review my above analysis)

Therefore using !(df & htons(IP_DF)) actually fixes some oversights of
our former proto==IPPROTO_UDP approach.

I'll send a patch.

Thanks
Shmulik