-
Marco Angaroni authored
When using OPS mode in conjunction with SIP persistent-engine, packets originating from the same ip-address/port could be balanced to different real servers, and (to properly handle SIP responses) OPS connections are created in the in-out direction too, where ip_vs_update_conntrack() is called to modify the reply tuple. As a result, there can be collision of conntrack tuples, causing random packet drops, as explained below: conntrack1: orig=CIP->VIP, reply=RIP1->CIP conntrack2: orig=RIP2->CIP, reply=CIP->VIP Tuple CIP->VIP is both in orig of conntrack1 and reply of conntrack2. The collision triggers packet drop inside nf_conntrack processing. In addition, the current implementation deletes the conntrack object at every expire of an OPS connection (once every forwarded packet), to have it recreated from scratch at next packet traversing IPVS. Since in OPS mode, by definition, we don't expect any associated response, the choices implemented in this patch are: a) don't call nf_conntrack_alter_reply() for OPS connections inside ip_vs_update_conntrack(). b) don't delete the conntrack object at OPS connection expire. The result is that created conntrack objects for each tuple CIP->VIP, RIP-N->CIP, etc. are left in UNREPLIED state and not modified by IPVS OPS connection management. This eliminates packet drops and leaves a single conntrack object for each tuple packets are sent from. Signed-off-by:
Marco Angaroni <marcoangaroni@gmail.com> Signed-off-by:
Julian Anastasov <ja@ssi.bg> Signed-off-by:
Simon Horman <horms@verge.net.au>
8fb04d9f