Fix rollback invocation after CmdAdd failure in CNI server #5548

antoninbas · 2023-10-05T20:00:06Z

When performing configuration rollback after an error in CmdAdd, we do not invoke CmdDel directly. Instead, we invoke an internal version of it which does not log a "Received CmdDel request" message (the message is confusing otherwise as it implies that we received a new CNI DEL command from the container runtime), and which does not process the network config again (as it was already processed at the beginning of CmdAdd). By not processing the config a second time, we ensure that there are no duplicate CIDRs in the IPAMConfig.
Migrate klog calls in server.go to use structured logging.
Improve unit tests for the CNI server to validate this fix.

antoninbas · 2023-10-05T21:30:29Z

/test-all

antoninbas · 2023-10-06T01:55:37Z

Also tested manually on a K8s cluster, by injecting an error in the CNI server during interface configuration (after IPAM):

I1006 01:47:20.047630       1 server.go:425] "Received CmdAdd request" request="cni_args:{container_id:\"8cba418aee4d4dc7b3f8a842fc5e8194f7708e54995afa9518eceed834b51597\" netns:\"/var/run/netns/cni-5003539e-60c5-d59b-3bbb-98a94f5eadda\" ifname:\"eth0\" args:\"K8S_POD_NAME=toolbox-pttdm;K8S_POD_INFRA_CONTAINER_ID=8cba418aee4d4dc7b3f8a842fc5e8194f7708e54995afa9518eceed834b51597;K8S_POD_UID=42681e73-8601-4d74-b983-6e1035355e55;IgnoreUnknown=1;K8S_POD_NAMESPACE=default\" path:\"/opt/cni/bin\" network_configuration:\"{\\\"cniVersion\\\":\\\"0.3.0\\\",\\\"ipam\\\":{\\\"type\\\":\\\"host-local\\\"},\\\"name\\\":\\\"antrea\\\",\\\"type\\\":\\\"antrea\\\"}\"}"
I1006 01:47:20.053556       1 server.go:494] "Allocated IP addresses" container="8cba418aee4d4dc7b3f8a842fc5e8194f7708e54995afa9518eceed834b51597" result={"cniVersion":"1.0.0","ips":[{"address":"10.10.1.3/24","gateway":"10.10.1.1"}],"dns":{},"VLANID":0}
E1006 01:47:20.053623       1 server.go:503] "FAKING ERROR"
I1006 01:47:20.053643       1 server.go:456] "CmdAdd for container failed, trying to rollback" container="8cba418aee4d4dc7b3f8a842fc5e8194f7708e54995afa9518eceed834b51597"
I1006 01:47:20.058455       1 server.go:560] "Deleted IP addresses for container" container="8cba418aee4d4dc7b3f8a842fc5e8194f7708e54995afa9518eceed834b51597"
I1006 01:47:20.058505       1 server.go:566] "CmdDel for container succeeded" container="8cba418aee4d4dc7b3f8a842fc5e8194f7708e54995afa9518eceed834b51597"
I1006 01:47:20.083692       1 server.go:581] "Received CmdDel request" request="cni_args:{container_id:\"8cba418aee4d4dc7b3f8a842fc5e8194f7708e54995afa9518eceed834b51597\" netns:\"/var/run/netns/cni-5003539e-60c5-d59b-3bbb-98a94f5eadda\" ifname:\"eth0\" args:\"IgnoreUnknown=1;K8S_POD_NAMESPACE=default;K8S_POD_NAME=toolbox-pttdm;K8S_POD_INFRA_CONTAINER_ID=8cba418aee4d4dc7b3f8a842fc5e8194f7708e54995afa9518eceed834b51597;K8S_POD_UID=42681e73-8601-4d74-b983-6e1035355e55\" path:\"/opt/cni/bin\" network_configuration:\"{\\\"cniVersion\\\":\\\"0.3.0\\\",\\\"ipam\\\":{\\\"type\\\":\\\"host-local\\\"},\\\"name\\\":\\\"antrea\\\",\\\"type\\\":\\\"antrea\\\"}\"}"
I1006 01:47:20.087014       1 server.go:560] "Deleted IP addresses for container" container="8cba418aee4d4dc7b3f8a842fc5e8194f7708e54995afa9518eceed834b51597"
I1006 01:47:20.087059       1 server.go:566] "CmdDel for container succeeded" container="8cba418aee4d4dc7b3f8a842fc5e8194f7708e54995afa9518eceed834b51597"

Notice how the rollback is now working (we release the IP address right away).
We still log the "CmdDel for container succeeded" message for the rollback. This is convenient to check for rollback success. If reviewers find it confusing, I can use a different success message for rollback vs. actual request.

wenyingd · 2023-10-06T09:19:43Z

pkg/agent/cniserver/server.go

-				if _, err := s.CmdDel(ctx, request); err != nil {
-					klog.Warningf("Failed to rollback after CNI add failure: %v", err)
+				klog.InfoS("CmdAdd for container failed, trying to rollback", "container", cniConfig.ContainerId)
+				if _, err := s.cmdDel(ctx, cniConfig); err != nil {


I agree that it is an improvement to directly use the existing "cniConfig" in cmdDel to avoid unnecessary call on loadNetworkConfig.

However, I didn't understand why the original code can append a duplicate PodCIDR range in the IPAM configuration. When we call function loadNetworkConfig in CmdDel/CmdAdd/CmdCheck, it has generated a new CNIConfig object, and the cniConfig.NetworkConfig is unmarshaled from the json stringrequest.CniArgs.NetworkConfiguration. So even if we call CmdDel in CmdAdd rollback, the cniConfig passed to IPAM plugin from a new call with loadNetworkConfig is supposed to equal to the values existing in CmdAdd.

The original code was calling validateRequest for the rollback (because CmdDel calls validateRequest), using the same request object. validateRequest actually mutates the request but it is not obvious when looking at the code:

loadNetworkConfig has the following assignment:

antrea/pkg/agent/cniserver/server.go

Line 198 in bbec9d0

cniConfig.CniCmdArgs = request.CniArgs

It means that mutating the contents of cniConfig.CniCmdArgs (a pointer to a protobuf struct) will mutate the request.

updateLocalIPAMSubnet has the following assignment:

antrea/pkg/agent/cniserver/server.go

Line 275 in bbec9d0

cniConfig.NetworkConfiguration, _ = json.Marshal(cniConfig.NetworkConfig)

cniConfig.NetworkConfiguration is actually the same as cniConfig.CniCmdArgs.NetworkConfiguration (it is a byte slice). So at this point we have mutated the request. The next time we call validateRequestMessage (which we no longer do with my patch), the network configuration of the request already includes the Node subnets in the IPAM section.

I confirmed all of this with my unit test.

Note that I don't know why we are mutating the request in the first place; I don't know if it's intentional, but I think we have been doing it for 4 years. Since using the existing cniConfig seemed like the right thing to do, that's what I did in this patch. We could consider a follow-up patch to clean up existing logic, but I didn't want to risk introducing a new bug.

To clarify the statement above:

cniConfig.NetworkConfiguration is actually the same as cniConfig.CniCmdArgs.NetworkConfiguration

This is because CniCmdArgs is an embedded field:

antrea/pkg/agent/cniserver/server.go

Lines 127 to 136 in bbec9d0

type CNIConfig struct {

*types.NetworkConfig

// AntreaIPAM for an interface not managed by Antrea CNI.

secondaryNetworkIPAM bool

// CniCmdArgs received from the CNI plugin. IPAM data in CniCmdArgs can be updated with the

// Node's Pod CIDRs for NodeIPAM.

*cnipb.CniCmdArgs

// K8s CNI_ARGS passed to the CNI plugin.

*types.K8sArgs

}

Got it, thanks for the clarification.

wenyingd

LGTM

jianjuns

Thanks for the fix.

jianjuns · 2023-10-09T18:19:37Z

pkg/agent/cniserver/server.go

@@ -453,12 +453,12 @@ func (s *CNIServer) CmdAdd(ctx context.Context, request *cnipb.CniCmdRequest) (*
 		// Rollback to delete configurations once ADD is failure.


Not your code, but change "is failure" to "fails"?

* When performing configuration rollback after an error in CmdAdd, we do not invoke CmdDel directly. Instead, we invoke an internal version of it which does not log a "Received CmdDel request" message (the message is confusing otherwise as it implies that we received a new CNI DEL command from the container runtime), and which does not process the network config again (as it was already processed at the beginning of CmdAdd). By not processing the config a second time, we ensure that there are no duplicate CIDRs in the IPAMConfig. * Migrate klog calls in server.go to use structured logging. * Improve unit tests for the CNI server to validate this fix. Fixes antrea-io#5547 Signed-off-by: Antonin Bas <[email protected]>

Signed-off-by: Antonin Bas <[email protected]>

antoninbas · 2023-10-09T18:32:21Z

/test-all

antoninbas · 2023-10-09T20:17:47Z

/test-e2e
/test-conformance

…failure in CNI (#5558) * When performing configuration rollback after an error in CmdAdd, we do not invoke CmdDel directly. Instead, we invoke an internal version of it which does not log a "Received CmdDel request" message (the message is confusing otherwise as it implies that we received a new CNI DEL command from the container runtime), and which does not process the network config again (as it was already processed at the beginning of CmdAdd). By not processing the config a second time, we ensure that there are no duplicate CIDRs in the IPAMConfig. * Migrate klog calls in server.go to use structured logging. * Improve unit tests for the CNI server to validate this fix. Fixes #5547 Signed-off-by: Antonin Bas <[email protected]>

…failure in CNI (#5559) * When performing configuration rollback after an error in CmdAdd, we do not invoke CmdDel directly. Instead, we invoke an internal version of it which does not log a "Received CmdDel request" message (the message is confusing otherwise as it implies that we received a new CNI DEL command from the container runtime), and which does not process the network config again (as it was already processed at the beginning of CmdAdd). By not processing the config a second time, we ensure that there are no duplicate CIDRs in the IPAMConfig. * Migrate klog calls in server.go to use structured logging. * Improve unit tests for the CNI server to validate this fix. Fixes #5547 Signed-off-by: Antonin Bas <[email protected]>

…failure in CNI server (#5560) * When performing configuration rollback after an error in CmdAdd, we do not invoke CmdDel directly. Instead, we invoke an internal version of it which does not log a "Received CmdDel request" message (the message is confusing otherwise as it implies that we received a new CNI DEL command from the container runtime), and which does not process the network config again (as it was already processed at the beginning of CmdAdd). By not processing the config a second time, we ensure that there are no duplicate CIDRs in the IPAMConfig. * Migrate klog calls in server.go to use structured logging. * Improve unit tests for the CNI server to validate this fix. Fixes #5547 Signed-off-by: Antonin Bas <[email protected]>

…5548)" This reverts commit f43477f.

antoninbas requested review from wenyingd, tnqn and jianjuns October 5, 2023 21:30

antoninbas added action/backport Indicates a PR that requires backports. action/release-note Indicates a PR that should be included in release notes. labels Oct 5, 2023

wenyingd reviewed Oct 6, 2023

View reviewed changes

wenyingd previously approved these changes Oct 8, 2023

View reviewed changes

jianjuns previously approved these changes Oct 9, 2023

View reviewed changes

antoninbas added 2 commits October 9, 2023 11:24

Fix existing comment

6ab005d

Signed-off-by: Antonin Bas <[email protected]>

antoninbas dismissed stale reviews from jianjuns and wenyingd via 6ab005d October 9, 2023 18:31

antoninbas force-pushed the fix-rollback-after-cni-add-failure branch from 904a6b5 to 6ab005d Compare October 9, 2023 18:31

antoninbas requested a review from jianjuns October 9, 2023 18:32

jianjuns approved these changes Oct 9, 2023

View reviewed changes

antoninbas merged commit f43477f into antrea-io:main Oct 9, 2023

antoninbas deleted the fix-rollback-after-cni-add-failure branch October 9, 2023 23:10

luolanzone added a commit that referenced this pull request Oct 20, 2023

Revert "Fix rollback invocation after CmdAdd failure in CNI server (#…

0ede4ef

…5548)" This reverts commit f43477f.

antoninbas mentioned this pull request Nov 1, 2023

IP may leak after agent restart #4326

Closed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Fix rollback invocation after CmdAdd failure in CNI server #5548

Fix rollback invocation after CmdAdd failure in CNI server #5548

antoninbas commented Oct 5, 2023

antoninbas commented Oct 5, 2023

antoninbas commented Oct 6, 2023

wenyingd Oct 6, 2023

antoninbas Oct 6, 2023

antoninbas Oct 6, 2023

wenyingd Oct 8, 2023

wenyingd left a comment

jianjuns left a comment

jianjuns Oct 9, 2023

antoninbas Oct 9, 2023

antoninbas commented Oct 9, 2023

antoninbas commented Oct 9, 2023

	type CNIConfig struct {
	*types.NetworkConfig
	// AntreaIPAM for an interface not managed by Antrea CNI.
	secondaryNetworkIPAM bool
	// CniCmdArgs received from the CNI plugin. IPAM data in CniCmdArgs can be updated with the
	// Node's Pod CIDRs for NodeIPAM.
	*cnipb.CniCmdArgs
	// K8s CNI_ARGS passed to the CNI plugin.
	*types.K8sArgs
	}

		@@ -453,12 +453,12 @@ func (s CNIServer) CmdAdd(ctx context.Context, request cnipb.CniCmdRequest) (*
		// Rollback to delete configurations once ADD is failure.

Fix rollback invocation after CmdAdd failure in CNI server #5548

Fix rollback invocation after CmdAdd failure in CNI server #5548

Conversation

antoninbas commented Oct 5, 2023

antoninbas commented Oct 5, 2023

antoninbas commented Oct 6, 2023

wenyingd Oct 6, 2023

Choose a reason for hiding this comment

antoninbas Oct 6, 2023

Choose a reason for hiding this comment

antoninbas Oct 6, 2023

Choose a reason for hiding this comment

wenyingd Oct 8, 2023

Choose a reason for hiding this comment

wenyingd left a comment

Choose a reason for hiding this comment

jianjuns left a comment

Choose a reason for hiding this comment

jianjuns Oct 9, 2023

Choose a reason for hiding this comment

antoninbas Oct 9, 2023

Choose a reason for hiding this comment

antoninbas commented Oct 9, 2023

antoninbas commented Oct 9, 2023