ofproto-dpif: Fix tag caching for learned flows.
authorBen Pfaff <blp@nicira.com>
Tue, 20 Mar 2012 22:26:57 +0000 (15:26 -0700)
committerBen Pfaff <blp@nicira.com>
Wed, 21 Mar 2012 16:26:26 +0000 (09:26 -0700)
commit69ccab125a23c9d1d1ee523cedf7df9dca1dd700
treeafb1d4e5fb4709a99415696f9c96562e705e12e3
parent19fc14ad08b4bc79515a52cc1a89b6f6e92225a3
ofproto-dpif: Fix tag caching for learned flows.

This code in xlate_table_action() is supposed to tag flows in tables that
have special forms so that changes do not require revalidating every flow.
When rule->tag is nonzero, its value can be used, because we know in this
case that rule->cr.wc is the same as table->other_table->wc and that thus
rule->tag caches the return value of the rule_calculate_tag() expression.
When rule->tag is zero (a "catchall" rule) we need to calculate the tag
manually because we have no way to cache it in that case.

I discovered this bug by running an "hping3" between a couple of VMs plus
the following commands on OVS in the middle:

    ovs-ofctl del-flows br0
    ovs-ofctl add-flow br0 "table=0 actions=learn(table=1, \
              idle_timeout=600, NXM_OF_VLAN_TCI[0..11], \
              NXM_OF_ETH_DST[]=NXM_OF_ETH_SRC[], \
              output:NXM_OF_IN_PORT[], fin_idle_timeout=10), resubmit(,1)"
    ovs-ofctl add-flow br0 "table=1 priority=0 actions=flood"

Without this patch, flows don't get properly invalidated upon initial MAC
learning, so one sees warnings like the following:

    in_port(2),eth(src=50:54:00:00:00:05,dst=50:54:00:00:00:07),
    eth_type(0x0800),ipv4(src=192.168.0.1,dst=192.168.0.2,proto=6,tos=0,
    ttl=64,frag=no),tcp(src=13966,dst=0): inconsistency in subfacet
    (actions were: 3,0,1) (correct actions: 1)

This patch fixes the problem and thus avoids these warnings.

Signed-off-by: Ben Pfaff <blp@nicira.com>
ofproto/ofproto-dpif.c