- Cloud-controller-manager: routes controller should not depend on --allocate-node-cidrs ([#97029](https://github.com/kubernetes/kubernetes/pull/97029), [@andrewsykim](https://github.com/andrewsykim)) [SIG Cloud Provider and Testing]

* Fixed race conditions around devicemanager Allocate() and endpoint deletion. ([#60856](https://github.com/kubernetes/kubernetes/pull/60856), [@jiayingz](https://github.com/jiayingz))

    }

    @Override
    public void forEach(Consumer<? super K> action) {
      checkNotNull(action);
      // avoids entry allocation for those maps that allocate entries on iteration
      map.forEach((k, v) -> action.accept(k));
    }

    @Override
    public int size() {
      return map().size();
    }

    @Override

- Image garbage collection no longer fails for images with only one tag but more than one repository associated. ([#70647](https://github.com/kubernetes/kubernetes/pull/70647), [@corvus-ch](https://github.com/corvus-ch))
- Re-issued Allocate grpc calls before starting a container that requests device-plugin resources if the cached state is missing. ([#73824](https://github.com/kubernetes/kubernetes/pull/73824), [@jiayingz](https://github.com/jiayingz))

    
- Kubeadm: improve the validation of serviceSubnet and podSubnet.
  ServiceSubnet has to be limited in size, due to implementation details, and the mask can not allocate more than 20 bits.
  PodSubnet validates against the corresponding cluster "--node-cidr-mask-size" of the kube-controller-manager, it fail if the values are not compatible.

- Fixed a bug that caused the `apiserver` to panic when trying to allocate a Service with a dynamic `ClusterIP` and was configured with Service CIDRs with a /28 mask for IPv4 and a /124 mask for IPv6 ([#115322](https://github.com/kubernetes/kubernetes/pull/115322), [@aojea](https://github.com/aojea))

### Documentation

- DRA: When the prioritized list feature was used in a request and the resulting number of allocated devices exceeded the number of allowed devices per claim, the scheduler aborted the attempt to allocate devices early. Previously, it tried to many different combinations, which could take a long time. ([#130593](https://github.com/kubernetes/kubernetes/pull/130593), [@mortent](https://github.com/mortent)) [SIG Apps, Node, Scheduling and Testing]

- Fix a bug that caused to panic the apiserver when trying to allocate a Service with a dynamic ClusterIP and it has been configured with Service CIDRs with a /28 mask for IPv4 and a /124 mask for IPv6 ([#115333](https://github.com/kubernetes/kubernetes/pull/115333), [@aojea](https://github.com/aojea)) [SIG Testing]...

- Dynamic resource allocation: when a claim uses "wait for first consumer" allocation (the default), then it will now get deallocated after it was used by a pod. That ensures that the next pod isn't affected by previous scheduling decision and that resources are not kept allocated unless really needed. If keeping a claim allocated is desired, use "immediate allocation." ([#11...