[ovs-dev] [PATCH v3 1/2] ofproto: Allow xlate_actions() to fail.

Jarno Rajahalme jarno at ovn.org
Wed Nov 25 23:06:17 UTC 2015


> On Nov 25, 2015, at 2:58 PM, Joe Stringer <joe at ovn.org> wrote:
> 
> On 25 November 2015 at 11:23, Jarno Rajahalme <jarno at ovn.org <mailto:jarno at ovn.org>> wrote:
>> 
>> On Nov 25, 2015, at 11:11 AM, Jarno Rajahalme <jarno at ovn.org> wrote:
>> 
>> 
>> On Nov 25, 2015, at 10:52 AM, Joe Stringer <joe at ovn.org> wrote:
>> 
>> On 25 November 2015 at 10:31, Jarno Rajahalme <jarno at ovn.org> wrote:
>> 
>> 
>> On Nov 24, 2015, at 5:02 PM, Joe Stringer <joe at ovn.org> wrote:
>> 
>> On 24 November 2015 at 13:41, Jarno Rajahalme <jarno at ovn.org> wrote:
>> 
>> Sometimes xlate_actions() fails due to too deep recursion, too many
>> MPLS labels, or missing recirculation context.  Make xlate_actions()
>> clear out the produced odp actions in these cases to make it easy for
>> the caller to install a drop flow (instead or installing a flow with
>> partially translated actions).  Also, return a specific error code, so
>> that the error can be properly propagated where meaningful.
>> 
>> Before this patch it was possible that the revalidation installed a
>> flow with a recirculation ID with an invalid recirc ID (== 0), due to
>> the introduction of in-place modification in commit 43b2f131a229
>> (ofproto: Allow in-place modifications of datapath flows).
>> 
>> Signed-off-by: Jarno Rajahalme <jarno at ovn.org>
>> 
>> 
>> Should this also set the error when receiving packets on a mirror port
>> in xlate_actions()? Or when receiving tagged VLAN traffic that doesn't
>> correspond to the port's vlan tag? Or when a group has no live bucket?
>> Are there any other cases that should also be covered? (I just scanned
>> across ofproto/ofproto-dpif-xlate.c looking for cases where we're
>> already logging that we drop the packet, but maybe there's a reasoning
>> behind not including these - if so, please enlighten me)
>> 
>> 
>> No reasoning for missing those, I just did not notice them. Thanks for
>> pointing them out.
>> 
>> 
>> OK, I thought it may have been something like "expected errors" vs.
>> "unexpected errors".
>> 
>> 
>> Looking into these I noticed this to be the case. Must discern whether to
>> fail just the individual action v.s. the whole pipeline.
>> 
>> 
>> How about this incremental to cover two cases here (rest are “expected
>> errors” IMO):
>> 
>> diff --git a/ofproto/ofproto-dpif-xlate.c b/ofproto/ofproto-dpif-xlate.c
>> index 36a6fbc..2908339 100644
>> --- a/ofproto/ofproto-dpif-xlate.c
>> +++ b/ofproto/ofproto-dpif-xlate.c
>> @@ -336,6 +336,10 @@ const char *xlate_strerror(enum xlate_error error)
>>         return "Recirculation conflict";
>>     case XLATE_TOO_MANY_MPLS_LABELS:
>>         return "Too many MPLS labels";
>> +    case XLATE_BUCKET_CHAINING_TOO_DEEP:
>> +        return "Bucket chaining too deep";
>> +    case XLATE_NO_INPUT_BUNDLE:
>> +        return "No input bundle";
>>     }
>>     return "Unknown error";
>> }
>> @@ -1444,10 +1448,9 @@ bucket_is_alive(const struct xlate_ctx *ctx,
>>                 struct ofputil_bucket *bucket, int depth)
>> {
>>     if (depth >= MAX_LIVENESS_RECURSION) {
>> -        static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 1);
>> -
>> -        VLOG_WARN_RL(&rl, "bucket chaining exceeded %d links",
>> -                     MAX_LIVENESS_RECURSION);
>> +        XLATE_REPORT_ERROR(ctx, "bucket chaining exceeded %d links",
>> +                           MAX_LIVENESS_RECURSION);
>> +        ctx->error = XLATE_BUCKET_CHAINING_TOO_DEEP;
>>         return false;
>>     }
>> 
>> @@ -2323,7 +2326,8 @@ xlate_normal(struct xlate_ctx *ctx)
>>     in_xbundle = lookup_input_bundle(ctx->xbridge, flow->in_port.ofp_port,
>>                                      ctx->xin->packet != NULL, &in_port);
>>     if (!in_xbundle) {
>> -        xlate_report(ctx, "no input bundle, dropping");
>> +        XLATE_REPORT_ERROR(ctx, "no input bundle, dropping");
>> +        ctx->error = XLATE_NO_INPUT_BUNDLE;
>>         return;
>>     }
>> 
>> 
>> The last one is debatable, as setting the error fails the whole translation
>> rather than just the normal action. But this is most likely an configuration
>> error, so maybe failing the whole pipeline (and installing a drop flow) is
>> the right thing to do here?
> 
> Jarno and I discussed this offline, and I'll try to summarise here.
> Broadly speaking, we're talking about the decision between failing an
> individual (piece of an) action or completely failing the action
> processing for the flow. And I think arguably the approach should be
> that if it is a serious error such as running out of resources or an
> internal conflict of recirc IDs, then we should fail the entire action
> processing. In this case it will have two user-visible effects:
> 1) ofproto/trace will tell the user which serious condition is being
> triggered that causes dropping of the flow
> 2) OpenFlow controllers attempting packet_out could be notified that
> the error occurred (rather than silently failing like currently)
> 
> However, in the two cases in the incremental patch here, the actions
> inherently have some ambiguity as to whether they successfully execute
> (eg output) or not. The more obvious case is in the bucket_is_alive()
> logic, where recursion will cause a bucket to be not used. If a bucket
> is not live in the spec, this doesn't mean that the entire flow should
> stop processing. In the case of normal, I'd argue it's very similar in
> that 'normal' doesn't specifically attempt to output to a particular
> port; sending packets out to different ports may fail for different
> reasons, but this shouldn't prevent later actions in the actions list
> from being executed.
> 
> I think the latter cases should be reported for ofproto/trace, though.
> 
> Looking back across this thread, it looks not far off your reasoning
> described earlier so I think we're converging on the same view. Does
> this sound like a fair approach?
> 
> --
> 
> In the mirror case, the point is moot because do_xlate_actions() isn't
> even called in that case, so it's purely a matter of whether we want
> to return the error up the stack or not. Maybe that should be reported
> for ofproto/trace as well.
> 
> I didn't see any other cases that might need handling through this.

So the only ask I see here is that more of the cases of individual actions bailing out should have xlate_report() calls on them. To me this sounds like a different patch, not directly related to erroring out the whole translation. As such I hope to get an Ack on the original patch of this now lengthy discussion…

  Jarno





More information about the dev mailing list