On 17 May 2017 at 17:54, Joe Stringer <[email protected]> wrote:
> On 17 May 2017 at 16:26, Darrell Ball <[email protected]> wrote:
>>
>>
>> On 5/17/17, 2:19 PM, "Joe Stringer" <[email protected]> wrote:
>>
>>     On 16 May 2017 at 21:01, Darrell Ball <[email protected]> wrote:
>>     >
>>     >
>>     > On 5/15/17, 6:00 AM, "[email protected] on behalf of 
>> Bodireddy, Bhanuprakash" <[email protected] on behalf of 
>> [email protected]> wrote:
>>     >
>>     >     >Commit daf4d3c18da4("odp: Support conntrack orig tuple key.") 
>> introduced
>>     >     >new fields in struct 'pkt_metadata'.  pkt_metadata_init() is 
>> called for every
>>     >     >packet in the userspace datapath.  When testing a simple single 
>> flow case with
>>     >     >DPDK, we observe a lower throughput after the above commit (it 
>> was 14.88
>>     >     >Mpps before, it is 13 Mpps after).
>>     >     >
>>     >     >This patch skips initializing ct_orig_tuple in 
>> pkt_metadata_init().
>>     >     >It should be enough to initialize ct_state, because nobody should 
>> look at
>>     >     >ct_orig_tuple unless ct_state is != 0.
>>     >     >
>>     >     >CC: Jarno Rajahalme <[email protected]>
>>     >     >Signed-off-by: Daniele Di Proietto <[email protected]>
>>     >     >---
>>     >     >I'm sending this as an RFC because I didn't check very carefully 
>> if we can really
>>     >     >avoid initializing ct_orig_tuple.
>>     >     >
>>     >     >Maybe there are better solutions to this problem.
>>     >     >---
>>     >     > lib/packets.h | 2 +-
>>     >     > 1 file changed, 1 insertion(+), 1 deletion(-)
>>     >     >
>>     >     >diff --git a/lib/packets.h b/lib/packets.h index 
>> a5a483bc8..6f1791c7a 100644
>>     >     >--- a/lib/packets.h
>>     >     >+++ b/lib/packets.h
>>     >     >@@ -129,7 +129,7 @@ pkt_metadata_init(struct pkt_metadata *md,
>>     >     >odp_port_t port)
>>     >     >     /* It can be expensive to zero out all of the tunnel 
>> metadata. However,
>>     >     >      * we can just zero out ip_dst and the rest of the data will 
>> never be
>>     >     >      * looked at. */
>>     >     >-    memset(md, 0, offsetof(struct pkt_metadata, in_port));
>>     >     >+    memset(md, 0, offsetof(struct pkt_metadata, ct_orig_tuple));
>>     >     >     md->tunnel.ip_dst = 0;
>>     >     >     md->tunnel.ipv6_dst = in6addr_any;
>>     >     >
>>     >
>>     >     It's been a while this RFC patch has been submitted to fix 
>> performance drop on Master. This indeed fixes the OvS-DPDK performance drop 
>> that was introduced by the conntrack commit.
>>     >     Is there a better fix than what is suggested above?
>>     >
>>     >
>>     > This affects both kernel and userspace.
>>     > I tested this on both datapaths.
>>     > LGTM
>>
>>     How do we make sure that ct_orig_tuple isn't used uninitialized? Do we
>>     need to zero out the ct_orig_tuple proto?
>>
>> There are a couple places where we explicitly set or reset the pkt metadata 
>> ct_orig_tuple;
>> one is in pkt_metadata_from_flow().
>>
>> I know there is a check for ct_orig_tuple proto in 
>> odp_key_from_pkt_metadata()
>> Can you find a case where can run this code without a set or reset of 
>> ct_orig_tuple pkt md
>> or you are not sure ?
>
> I wasn't sure, and I didn't see a response to the question that
> Daniele asked below the dashes in the original submission.
>
> Is miniflow_extract() safe wrt this? Seems like plausibly if the
> recirc_id is nonzero but there is no CT state (because, eg, bonds, or
> MPLS pop to IP, etc) it could end up populating the miniflow with
> garbage. In particular the path from emc_processing() seems suspect.

Perhaps this is actually OK, because the ct_state would be initialized
to zero in pkt_metadata_init(), and the classifier would never look at
ct_orig_tuple unless the ct_state is valid (according to the
ovs-fields(7) documentation). In which case, it shouldn't(?) matter
that garbage is written to the miniflow. Alternatively if md_is_valid
(eg, because it was recirculated), then it's up to the other callers
to ensure that md.ct_state was properly initialized. Effectively this
is done by the write_ct_md() function.

I wonder though if actually the miniflow_extract() "ct_nw_proto_p"
initialization should only occur if the md->ct_state is valid. That
could save a few extra cycles in a similar vein to how this patch
proposes.
_______________________________________________
dev mailing list
[email protected]
https://mail.openvswitch.org/mailman/listinfo/ovs-dev

Reply via email to