# KubeVirt v1.10.0

2026-11-18 · [Original release notes](https://github.com/kubevirt/kubevirt/releases/tag/v1.10.0) · [Full summary](https://ratatosk.io/en/releases/kubevirt/v1.10.0)

Security 7 · Breaking changes 38 · Deprecations 1 · Important fixes 74 · Enhancements 39 · Dependency updates 7 · Other changes 31

## Check before upgrading

### Known issues

- **Client-side apply of the KubeVirt operator manifest fails**: The `kubevirts.kubevirt.io` CRD exceeds the 256 KiB annotation limit. The note names `kubectl apply --server-side` and `kubectl create` as alternatives.
  > > [!IMPORTANT]
  > > Installing the `kubevirt-operator.yaml` with client-side `kubectl apply -f kubevirt-operator.yaml` fails because the kubevirts.kubevirt.io CRD now exceeds the 256 KiB annotation limit. Use `kubectl apply --server-side` (or `kubectl create`) for `kubevirt-operator.yaml` instead. See #19980.

## Security (7)

### [medium] virt-api port-forward and SSH traffic redirection fixed

- Affects: VMIs with a pod network accessed via `virtctl port-forward` or `virtctl ssh`
- [CVE-2026-13318](https://www.cve.org/CVERecord?id=CVE-2026-13318)

When the launcher pod IP could not be resolved, virt-api used an unvalidated interface IP reported by the guest agent, allowing traffic to be redirected to arbitrary reachable endpoints. Port-forward and SSH now dial only the launcher pod IP, and VMIs without a pod network are rejected before websocket upgrade or dialing.

> - [PR #18753][orelmisan] Fix SSRF vulnerability (CVE-2026-13318) in virt-api port-forward that allowed a malicious QEMU guest agent to redirect virtctl port-forward and ssh traffic to arbitrary network endpoints. virt-api now dials only the launcher pod IP and never the guest-agent-reported interface IP.

### [medium] virt-api port-forward protocol checks block SSRF

- Affects: virt-api port-forward and SSH traffic (Inferred)
- [CVE-2026-13318](https://www.cve.org/CVERecord?id=CVE-2026-13318), [GHSA-cgm6-wvp9-4g5h](https://github.com/advisories/GHSA-cgm6-wvp9-4g5h)

A malicious QEMU guest agent could redirect `virtctl` port-forward and SSH traffic to arbitrary network endpoints because the forwarded protocol was passed to `net.Dial` unchecked. Forwarded protocols other than `tcp` and `udp` now receive a `BadRequest` error.

> - [PR #18753][orelmisan] Fix SSRF vulnerability (CVE-2026-13318) in virt-api port-forward that allowed a malicious QEMU guest agent to redirect virtctl port-forward and ssh traffic to arbitrary network endpoints. virt-api now dials only the launcher pod IP and never the guest-agent-reported interface IP.

### [medium] virt-api port-forward target selection fixed

- Affects: VMIs connected to the pod network
- [CVE-2026-13318](https://www.cve.org/CVERecord?id=CVE-2026-13318), [GHSA-cgm6-wvp9-4g5h](https://github.com/advisories/GHSA-cgm6-wvp9-4g5h)

When the virt-launcher pod is multihomed or its binding leaves no address on the pod-side link, the reported interface IP can be unreachable from virt-api, preventing the port-forward tunnel from being established. Port-forward now dials the launcher pod IP when available, falling back to the interface-status IP; existing core bindings have no observable change, and VMIs without a pod network retain the prior behavior.

> - [PR #19019][lllamnyp] virt-api: the port-forward subresource (used by `virtctl port-forward` and `virtctl ssh`) now dials the virt-launcher pod IP for VMIs connected to the pod network, instead of the first IP reported in the VMI interface status. The virt-api cluster role gains the `watch` verb on pods.

### Migration identity sharing limited to proxy-enabled setups

- Affects: KubeVirt setups with the migration proxy disabled

Previously, the synchronization controller pod received the migration identity even when the migration proxy was disabled. It now receives that identity only when the proxy is enabled.

> - [PR #19490][awels] Only add virt-handler migration identity to synchronization controller if proxy is enabled.

### [Security status unclear] VM admission blocks restricted `ContainerPath` paths

- Affects: VM/VMI `ContainerPath` volumes

Submissions using paths matching `/run/kubevirt*` or `/run/libvirt*` are rejected at admission. These restrictions match the existing `/var/run` restrictions.

> - [PR #19251][mhenriks] Check ContainerPathVolume feature gate at admission time.  Reject /run/kubevirt* and /run/libvirt* paths.

### [Security status unclear] VMI name label added to GPU info metric

- Affects: installations scraping or querying `kubevirt_vmi_gpu_info` (Inferred)

Before this release, `kubevirt_vmi_gpu_info` did not include a VMI name label.

> - [PR #18902][machadovilaca] Add name label to kubevirt_vmi_gpu_info

### Export server token timing side channel fixed

- Affects: setups using the export server token check

The token check did not use a constant-time comparison, creating a timing side channel. It now uses a constant-time comparison.

> - [PR #18200][naruto-lgtm] Fixed a timing side channel in the export server token check by switching to a constant-time comparison.

## Breaking changes (38)

### Memory-dump PVC excluded from VM snapshots and exports

- Affects: VM snapshots or exports with a memory-dump PVC attached (Inferred)

Restoring a snapshot or export did not restore the dump PVC even before this change.

Evidence: “Snapshots and exports no longer include the dump PVC (restore never restored it anyway).” https://github.com/kubevirt/kubevirt/pull/19550

> - [PR #19550][Acedus] Memory dumps are attached as utility volumes and require the UtilityVolumes feature gate, which is promoted to Beta and enabled by default. VM snapshots and exports no longer include the memory dump PVC.

### `VirtualMachineBackup` type reporting moves to per-volume status

- Affects: checkpoint-based `VirtualMachineBackup`s with disks missing the base checkpoint bitmap

Previously, one disk missing its base checkpoint bitmap caused the entire backup to fail, including disks with valid bitmaps. Now that disk is backed up in full while the others retain their incremental deltas.

> - [PR #19056][Acedus] VirtualMachineBackup no longer fails when a disk has lost the base checkpoint's bitmap; that disk is backed up in full and type is now reported per volume in `status.includedVolumes[].type`, replacing the removed `status.type`.

### VirtualMachineBackup backup type reported per volume

- Affects: clients or automation reading `VirtualMachineBackup` status (Inferred)

Previously, if a disk in a checkpoint-based backup had lost the base checkpoint's bitmap, it failed the entire backup. That disk is now backed up in full instead.

> - [PR #19056][Acedus] VirtualMachineBackup no longer fails when a disk has lost the base checkpoint's bitmap; that disk is backed up in full and type is now reported per volume in `status.includedVolumes[].type`, replacing the removed `status.type`.

### `io.kubevirt.disk.name` now uses VM volume names

- Affects: exported OCI artifacts with disk layers (Inferred)

Previously, the annotation held a sanitized PVC-derived mount-directory name rather than the VM-spec volume name, so it did not identify the VM volume. For example, a `rootdisk` volume backed by `fedora-vm-rootdisk` is now labeled `rootdisk`.

Evidence: “The volume name is passed to the exporter pod and the layer is annotated with it. A volume `rootdisk` backed by claim `fedora-vm-rootdisk` is annotated `rootdisk`.” https://github.com/kubevirt/kubevirt/pull/19066

> - [PR #19066][0xFelix] The io.kubevirt.disk.name annotation of exported OCI artifacts now carries the volume name from the VM instead of the PersistentVolumeClaim name

### User-defined HostDisk volumes excluded from CBT

- Affects: CBT with user-defined HostDisk volumes

When a missing `DiskOrCreate` image was silently recreated, its CBT overlay could keep describing a disk that no longer existed. HostDisk volumes no longer count toward CBT memory overhead.

Evidence: “Only PVC and DataVolume volumes are eligible for CBT. User-defined HostDisk volumes are no longer tracked, and no longer count towards CBT's memory overhead.” https://github.com/kubevirt/kubevirt/pull/19410

> - [PR #19410][Acedus] HostDisk volumes are no longer eligible for changed block tracking.

### `virt-launcher-monitor` no longer honors `VIRT_LAUNCHER_LOG_VERBOSITY`

- Affects: setups that configure `VIRT_LAUNCHER_LOG_VERBOSITY` for `virt-launcher-monitor` (Inferred)

Previously, a valid integer value set the Go monitor’s log verbosity. The rewritten C monitor does not use the configured value to set its verbosity.

Evidence: “- no `VIRT_LAUNCHER_LOG_VERBOSITY` in the monitor” https://github.com/kubevirt/kubevirt/pull/19240

> - [PR #19240][vladikr] virt-launcher-monitor is rewritten in C to reduce memory footprint to about 2Mi

### Persistent-reservation LUNs now default to SCSI and reject non-SCSI buses

- Affects: LUN disks with persistent reservation enabled

Evidence: “Default bus to SCSI when reservation is true and bus is unset, and reject with a validation error when reservation is true with an explicitly non-SCSI bus.” https://github.com/kubevirt/kubevirt/pull/18267

> - [PR #18267][akalenyu] Default LUN bus to SCSI when persistent reservation is enabled

### Software emulation now requires an explicit policy

- Affects: KubeVirt clusters using Software Emulation or emulation policies for VMs

Previously, software emulation was controlled by a cluster-wide on/off behavior; it is now used only when selected through an emulation policy. Existing behavior is retained when the KubeVirt CR sets `spec.configuration.emulationPolicy: Software`.

Evidence: “Action Required: Add `spec.configuration.emulationPolicy: Software` to kubevirt CR to keep existing behaviour if using Software Emulation” https://github.com/kubevirt/kubevirt/pull/19244

> - [PR #19244][jschintag] Add configuration option for cluster wide and per VM emulation (Alpha Stage).

### Istio proxy termination removed from virt-launcher-monitor

- Affects: KubeVirt setups using Istio sidecars
- Action required: Istio 1.27 or later, or native sidecar mode.

With legacy Istio injection, Envoy ran as a regular container and could keep the pod alive after the main container exited. virt-launcher-monitor no longer terminates the proxy during shutdown.

> - [PR #19239][akalenyu] Istio sidecar proxy termination is no longer handled by virt-launcher-monitor; Istio >= 1.27 is required with kubevirt (alternatively, use native sidecar mode)

### Separate pull-mode endpoint links in `VirtualMachineBackup` status

- Affects: `VirtualMachineBackup` status for pull-mode backups

When a pull-mode `VirtualMachineExport` has both internal and external links, backup status previously exposed only one set, losing the other. The old `EndpointCert`, `DataEndpoint`, and `MapEndpoint` fields are removed; each new link entry has its own CA certificate and per-volume data and map URLs.

Evidence: “The existing fields (`EndpointCert`, `DataEndpoint`, `MapEndpoint`) are removed. Consumers of pull-mode backups should use `status.links.internal` and/or `status.links.external` instead.” https://github.com/kubevirt/kubevirt/pull/18537

> - [PR #18537][davmarro] Add structured `links` field to `VirtualMachineBackup` status exposing both internal and external export endpoints for pull-mode backups. The existing endpoint fields are removed.

### `VSOCK` feature gate graduated to Beta and enabled by default

- Affects: installations using or configuring the `VSOCK` feature gate, and workloads using VSOCK (Inferred)

> - [PR #19155][0xFelix] The VSOCK feature gate graduated to Beta and is now enabled by default.

### KubeVirt stops using the legacy `cpumanager` node label

- Affects: automation or scheduling rules that use the legacy `cpumanager` node label on nodes updated by virt-handler (Inferred)

The `kubevirt.io/cpumanager` label remains in use for Pods.

> - [PR #19100][xpivarc] Label "cpumanager" is no longer used

### CBT dirty-bitmap overhead included in VM memory calculations

- Affects: VMs with CBT enabled

Previously, their memory calculations omitted the memory needed for CBT dirty bitmaps, understating the requirement. The calculation now includes overhead for each included disk, accounting for CBT-eligible volumes, active backup trackers, and a buffer for dynamic growth.

> - [PR #18416][Acedus] VMs with CBT now correctly factor in the required memory overhead for each of their included disks.

### Custom seccomp profile installation no longer gated

- Affects: clusters using the `KubevirtSeccompProfile` gate or relying on KubeVirt's custom seccomp profile

Before this release, KubeVirt installed the profile only when `KubevirtSeccompProfile` was explicitly enabled; it is now installed even when the gate is disabled.

Evidence: “**Before:** The `KubevirtSeccompProfile` feature gate is at Beta. The custom seccomp profile is only installed when the feature gate is explicitly enabled, and operator tests verify configuration statically (file presence, pod annotation). **After:** The `KubevirtSeccompProfile` feature gate graduates to GA. The custom seccomp profile is always installed, the feature gate check is removed, and three hardening improvements are included:” https://github.com/kubevirt/kubevirt/pull/18888

> - [PR #18888][dankenigsberg] Graduate the KubevirtSeccompProfile feature gate to GA. The custom seccomp profile is now always installed. Tighten profile file permissions to 0600, and write it atomically.

### Network resource injection is always on through Network Resources Injector

- Affects: installations using network resource injection for `NetworkAttachmentDefinition`-backed interfaces

The `ExternalNetResourceInjection` feature gate had been enabled by default since v1.9, but installations that disabled it can no longer opt out. `virt-controller` no longer queries `NetworkAttachmentDefinition` objects or injects network resource requests into launcher pods, and its `ClusterRole` no longer grants get access to `k8s.cni.cncf.io network-attachment-definitions`.

Evidence: “The `ExternalNetResourceInjection` FG has been Beta since v1.8.0 and enabled by default since v1.9. It disables virt-controller's query of `NetworkAttachmentDefinition` objects and the deployment of NAD RBAC rules by virt-operator. This PR graduates it to GA, making this behavior always on. virt-controller no longer queries `NetworkAttachmentDefinition` objects, injects network resource requests into the launcher pod, or holds the `k8s.cni.cncf.io network-attachment-definitions` get permission.” https://github.com/kubevirt/kubevirt/pull/18497

> - [PR #18497][orelmisan] ExternalNetResourceInjection graduated to GA. virt-controller no longer queries NetworkAttachmentDefinition objects or injects network resource requests into the launcher pod. The k8s.cni.cncf.io network-attachment-definitions get rule is

### VM migration now uses hashed interface names

- Affects: VMs with ordinal network or tap interface names, and automation using `--upgrade-ordinal-ifaces`
- Action required: Affected VMs are converted during live migration; no manual intervention is required.

Previously, migration targets could retain the source pod’s ordinal interface names, and the pre-migration hook was gated. The `--upgrade-ordinal-ifaces` flag is no longer available.

Evidence: “Migration targets will no longer preserve ordinal pod interface names from the source pod - the hashed scheme is always used. The pre-migration hook that converts ordinal tap names to hashed equivalents now runs unconditionally” https://github.com/kubevirt/kubevirt/pull/18317

> - [PR #18317][orelmisan] Graduate the PodSecondaryInterfaceNamingUpgrade feature gate to GA. The pre-migration hook that converts ordinal network interface names to hashed equivalents now runs unconditionally.

### QEMU overhead estimate increased for VMI memory requests

- Affects: VMI pods using the KVM or MSHV hypervisor backend

The QEMU process overhead estimate rises from 30Mi to 50Mi, adding 20Mi to each VMI pod memory request.

> - [PR #18388][dasionov] Bumped QEMU process memory overhead from 30Mi to 50Mi, increasing VMI pod memory requests by 20Mi to reflect actual QEMU RSS requirements with current QEMU versions.

### Node hooks now honor plugin-level conditions

- Affects: node hooks configured with a plugin-level `Spec.Condition` (Inferred)

This brings their condition handling in line with domain hooks.

> - [PR #18414][iholder101] Fix CallNodeHooks to evaluate plugin-level Spec.Condition, ensuring consistent behavior between domain hooks and node hooks.

### Global `-v` registration removed from `kubevirt.io/client-go/log`

- Affects: code importing `kubevirt.io/client-go/log`

Importing the package used to panic when another imported package, such as glog or klog, had already registered `-v` on `flag.CommandLine`. The flag is now kept in a package-private `FlagSet`, so `flag.CommandLine.Lookup("v")` no longer returns it; `log.VerbosityFlag()` exposes it instead.

Evidence: “The `-v` flag is registered on a package-private `FlagSet` instead of the global `flag.CommandLine`. A new `VerbosityFlag()` function is exported so that consumers can bridge it into pflag without going through `flag.CommandLine`. All 8 internal call sites are updated. Importing `kubevirt.io/client-go/log` no longer registers any flags on the global `flag.CommandLine`, eliminating the conflict entirely regardless of `init()` ordering.” https://github.com/kubevirt/kubevirt/pull/18294

> - [PR #18294][guzalv] client-go/log: The `-v` verbosity flag is no longer registered on `flag.CommandLine` during `init()`. External consumers that used `flag.CommandLine.Lookup("v")` to obtain the verbosity flag must switch to `log.VerbosityFlag()`. This eliminates panics when importing `kubevirt.io/client-go` alongside packages that also register a `-v` flag (e.g. glog, klog).

### Memory dumps now use VMI utility volumes

- Affects: setups using memory dumps

Previously, a dump added a `MemoryDump` volume to both the VM template and `vmi.spec.volumes`; it now attaches its PVC to the VMI without changing the template. If the VMI stops or restarts during a dump, the dump fails rather than continuing on the new VMI; dumps already in progress during the upgrade finish and are then detached.

Evidence: “A memory dump request attaches its PVC to the VMI as a utility volume of type `MemoryDump`, nothing is written to the VM template. A dump still in progress when its VMI stops or restarts fails instead of dumping the new VMI. Dumps that were in progress on it during the upgrade finish, then it's detached.” https://github.com/kubevirt/kubevirt/pull/19550

> - [PR #19550][Acedus] Memory dumps are attached as utility volumes and require the UtilityVolumes feature gate, which is promoted to Beta and enabled by default. VM snapshots and exports no longer include the memory dump PVC.

### `VGPULiveMigration` renamed to `MDevVGPULiveMigration` for mdev-based vGPU live migration

- Affects: setups enabling the Alpha feature gate for mdev-based vGPU live migration

Evidence: “Mdev-based vGPU live migration uses `MDevVGPULiveMigration`.” https://github.com/kubevirt/kubevirt/pull/19546

> - [PR #19546][dasionov] Renamed the alpha feature gate for mdev-based vGPU live migration from VGPULiveMigration to MDevVGPULiveMigration.

### Removal of `domainSpec` from launcher hook conditions

- Affects: launcher hook `condition` expressions that reference `domainSpec` (Inferred)

Previously, hook conditions could reference both the VMI and `domainSpec`. Conditional domain mutation can still use CEL's `has()` operator.

Evidence: “In addition, Launcher hook `condition` expressions could also reference `domainSpec` in addition to the VMI, which both complicated the code and was wrong because on certain situations the domain did not exist yet, it is not relevant for most hook points and it's a low-level libvirt-coupled variable which we don't want to commit to.” https://github.com/kubevirt/kubevirt/pull/19403

> - [PR #19403][iholder101] Add PreBoot and PreMigrationSource sidecar launcher hook points for plugins.

### Preparation-time live migration aborts now cancel the migration

- Affects: live migrations

Previously, an abort during domain preparation or a `PreMigrationSource` hook reported `AbortFailed` and the migration continued; it now cancels the migration and reports `AbortSucceeded`. During the short interval after hand-off but before libvirt's migration job is observable, aborts still report `AbortFailed` and virt-handler retries.

Evidence: “Also, there's a small behavior change for all migrations (with or without plugins): an abort that arrives while the migration is being prepared (domain preparation or `PreMigrationSource` hooks) now cancels it and reports `AbortSucceeded`, where it previously reported `AbortFailed` and the migration continued. A short window remains between the hand-off and libvirt's migration job becoming observable; an abort landing there still reports `AbortFailed`, and virt-handler retries.” https://github.com/kubevirt/kubevirt/pull/19403

> - [PR #19403][iholder101] Add PreBoot and PreMigrationSource sidecar launcher hook points for plugins.

### VM/VMI admission rejection of new `ContainerPath` use with `ContainerPathVolume` disabled

- Affects: VM/VMI workloads using `ContainerPath` while the `ContainerPathVolume` feature gate is disabled

Before this change, VM/VMI admission did not check the gate for new usage, though a controller-side check remained for VMIs admitted before the gate was disabled. Updates to objects already using `ContainerPath` remain allowed.

Evidence: “Rejects new ContainerPath usage at VM/VMI admission when the feature gate is disabled, while allowing updates to objects already using it.” https://github.com/kubevirt/kubevirt/pull/19251

> - [PR #19251][mhenriks] Check ContainerPathVolume feature gate at admission time.  Reject /run/kubevirt* and /run/libvirt* paths.

### DRA `ResourceClaim` VMIs marked non-migratable

- Affects: VMIs with DRA networks backed by `ResourceClaim` while `NetworkDevicesWithDRA` is enabled

A `ResourceClaim` is bound to a specific allocation, which the migration target pod cannot reserve again.

Evidence: “VerifyVMIMigratable now rejects the former. The check is guarded by the NetworkDevicesWithDRA feature gate, threaded in from virt-handler, so behavior is unchanged when the gate is disabled.” https://github.com/kubevirt/kubevirt/pull/19468

> - [PR #19468][orelmisan] VMI whose DRA network is backed by a ResourceClaim (rather than a ResourceClaimTemplate) are now reported as non-migratable.

### KubeVirt infrastructure containers now use read-only root filesystems

- Affects: KubeVirt infrastructure containers other than virt-handler and the virt-launcher init container
- Action required: No writable `/tmp` volume is required; certificates are loaded directly from mounted Secret paths.

Before this release, the included containers ran without `readOnlyRootFilesystem: true`, and virt-api created an unused temporary certificate directory. Certificates are loaded directly from mounted Secret paths, and no writable `/tmp` volume is added.

Evidence: “KubeVirt containers (`virt-operator`, `virt-api`, `virt-controller`, `virt-exportproxy`, and `virt-synchronization-controller`) run without `readOnlyRootFilesystem: true`, triggering compliance scan findings.” https://github.com/kubevirt/kubevirt/pull/18853

> - [PR #18853][davmarro] Enable readOnlyRootFilesystem for all KubeVirt infrastructure containers

### Deprecated `virtctl create vm` volume flags removed

- Affects: Invocations of `virtctl create vm` using `--volume-blank`, `--volume-datasource`, or `--volume-clone-pvc` (Inferred)

Volume creation now uses `--volume-import` with `type:blank`, `type:ds`, or `type:pvc` in place of the removed flags.

> - [PR #19362][NoamBerkoLaor] Removed deprecated virtctl create vm flags --volume-blank, --volume-datasource, and --volume-clone-pvc. Use --volume-import with type:blank, type:ds, or type:pvc instead.

### New virt-export Services use headless port 8443

- Affects: setups creating new VMExports

Links for new exports use `:8443`, while existing ClusterIP exports continue to use `:443`. Clients that hardcode `https://virt-export-*.svc:443` break for new exports; clients that follow `status.links` continue to work.

Evidence: “New exports are not ClusterIP on `:443`. Clients that hardcode `https://virt-export-*.svc:443` will break for those. Clients that follow `status.links` should be fine. Leftover ClusterIP exports stay on `:443`.” https://github.com/kubevirt/kubevirt/pull/18606

> - [PR #18606][OOsemka] Make new virt-export Services headless on port 8443 so virt-exportproxy can reach exporters across isolated networks. Existing ClusterIP/443 export Services are left in place. Internal VMExport links include the Service dial port (:8443 or :443).

### Plugin hooks move to `spec.launcherHooks`

- Affects: Plugin objects using `spec.domainHooks` or `PreVMStop` (Inferred)

Previously, `spec.domainHooks` did not require explicit hook points, and the stop hook point was named `PreVMStop`; launcher hooks now require explicit hook points and use `OnVMStop`.

> - [PR #19160][iholder101] Plugin's spec.domainHooks is renamed to spec.launcherHooks. Launcher hooks now require explicit hook points. PreVMStop is renamed to OnVMStop.

### virt-handler `/v1/vmstats` switches from GET to POST

- Affects: clients calling virt-handler's `/v1/vmstats` endpoint (Inferred)

Previously, one `VMStatsRequest` applied to every VMI on the node, with stats categories selected through query parameters. Requests now use a JSON body to specify stats categories for each VMI.

Evidence: “virt-handler GetVMStats accepts a per-VMI VMStatsRequest body, allowing the caller to filter which VMIs to collect from and specify the stats categories it needs on a per-VMI basis, rather than one set applied to every VMI on the node, so it can collect different data for different VMIs in a single call.” https://github.com/kubevirt/kubevirt/pull/18938

> - [PR #18938][machadovilaca] Accept a per-VMI VMStatsRequest body in virt-handler GetVMStats

### GPU memory-copy recording rule renamed to `vmi:kubevirt_vmi_gpu_mem_copy_utilization_ratio:sum`

- Affects: monitoring queries and dashboards using `vmi:kubevirt_vmi_gpu_mem_copy_util:sum` (Inferred)

The previous recording-rule name was `vmi:kubevirt_vmi_gpu_mem_copy_util:sum`.

Evidence: “-| vmi:kubevirt_vmi_gpu_mem_copy_util:sum | Recording rule | Gauge | Memory utilization (ratio, 0-1) of the GPU passed through to a virtual machine instance. |” https://github.com/kubevirt/kubevirt/pull/19144

> - [PR #19144][machadovilaca] Add GPU utilization time and active engine time recording rules

### client-go KubeVirt subresource client helpers removed

- Affects: client-go applications or tools calling `GetKubevirtSubresourceClient` or `GetKubevirtSubresourceClientFromFlags` (Inferred)

The functions provided a way to construct and return a KubeVirt subresource client.

> - [PR #18874][xpivarc] client-go: GetKubevirtSubresourceClient is removed

### DRA networks reject SR-IOV binding

- Affects: DRA networks configured with SR-IOV interfaces

Previously, SR-IOV binding with DRA networks could be used, with virt-launcher resolving PCI addresses from DRA device metadata for passthrough. These interfaces must now use binding plugins.

Evidence: “It is no longer possible to use SR-IOV binding with DRA networks - the network admitter rejects them and only allows interfaces using binding plugins.” https://github.com/kubevirt/kubevirt/pull/18451

> - [PR #18451][orelmisan] SR-IOV interfaces are no longer supported with DRA networks.

### `LibvirtHooksServerAndClient` graduates to GA and is always enabled

- Affects: setups configuring `LibvirtHooksServerAndClient` or using its domain-mutation hooks for live migration

Previously, the domain-mutation hook mechanism was a Beta feature that had to be explicitly enabled; the hooks adjust tap device names during migration.

Evidence: “Graduate the `LibvirtHooksServerAndClient` feature gate (VEP-141) to GA.” https://github.com/kubevirt/kubevirt/pull/18687

> - [PR #18687][Barakmor1] Graduate LibvirtHooksServerAndClient feature gate to GA

### Stall detector API field types changed

- Affects: clusters upgrading from v1.9 with `MigrationStallDetection` enabled, and SDKs importing v1.9 `StallDetectorOptions`

The four factor fields change from `*string` to `*resource.Quantity`, while `StallProgressTimeout` and `SwitchoverTimeout` change from `*uint64` to `*int64`. These changes can break cluster upgrades with `MigrationStallDetection` enabled and cause compile-time errors in SDKs importing the v1.9 type.

Evidence: “**Backward compatibility note:** These changes are NOT strictly backward compatible. At the API level, these changes break API SDK since anyone importing kubevirt.io/api/core/v1 from v1.9 and using StallDetectorOptions would see compile-time breakage when upgrading (e.g. *string -> *resource.Quantity). But at the serialization/upgrade level for API they ARE compatible. On the wire, MigrationOptions changes are breaking and thus this will break cluster upgrades if someone has the FG enabled when upgrading 1.9->1.10.” https://github.com/kubevirt/kubevirt/pull/18303

> - [PR #18303][Aseeef] BREAKING CHANGE: Modifies internal on-the-wire API fields for the stall detector feature; MigrationStallDetection FG must be disabled for cluster upgrades to succeed.

### PCI root ports for non-NIC devices made non-hotpluggable

- Affects: VMs with PCI root ports hosting non-NIC devices

Guest OS hotplug mechanisms, such as Windows’ “Safely Remove Hardware” menu, could eject critical devices including disks, the memory balloon, RNG, and watchdog. NIC ports and empty ports reserved for future hotplug remain available.

> - [PR #17605][dasionov] Bug fix: Prevent guest OS from ejecting critical PCI devices (disks, memory balloon, RNG, watchdog) via hotplug mechanisms like the Windows "Safely Remove Hardware" tray menu. PCI root ports hosting non-NIC devices are now marked as non-hotpluggable, while NIC ports and empty ports reserved for future hotplug remain available.

### Slirp network binding plugin sidecar image removed

- Affects: setups referencing or using the slirp network binding plugin sidecar image

> - [PR #18613][orelmisan] Removed the slirp network binding plugin sidecar image. The plugin had no known consumers.

### RamFB display validation rejects templates with multiple vGPUs

- Affects: `VirtualMachine` and `VirtualMachineInstanceReplicaSet` templates with vGPUs whose `ramFB` is present and `enabled` is unset

Previously, a template with more than one such vGPU could be accepted at creation and fail later when its VMI was created.

> - [PR #18542][thc1006] RamFB display validation now treats a vGPU whose ramFB is present with its enabled field unset as an enabled display, matching the API default. A VirtualMachine or VirtualMachineInstanceReplicaSet template with more than one such vGPU is now rejected at creation instead of being accepted and failing later when its VMI is created.

## Deprecations (1)

### Legacy guest-agent SSH-key setting deprecated

- Affects: VMs using the legacy guest-agent `guest-file-write` flow to set SSH keys in `~/.ssh/authorized_keys`
- Removed in: not stated

Previously, this flow wrote the SSH key directly to `~/.ssh/authorized_keys` without a deprecation warning. In this release, using it raises a warning in VMI conditions and an event.

> - [PR #13881][akrejcir] Deprecated old method to set SSH key in a VM.

## Important fixes (74)

### SMMUv3 IOMMU `ssidsize` handling fixed

- Affects: Grace IOVirtualization VMIs for which guest definition plugin hooks run

Guest definition plugin hooks now preserve these VMIs' SSID size and IOMMU driver attributes.

> - [PR #19862][kubevirt-bot] Bump libvirtxml to v1.12007.0 and fix the smmuv3 IOMMU ssidsize attribute, so Grace IOVirtualization VMIs get their SSID size and keep their IOMMU driver attributes when guest definition plugin hooks run.

### CBT overlays now distinguish PVCs per volume

- Affects: setups using CBT incremental backups

Previously, when a volume pointed at a different PVC, its overlay could be reused and produce an incorrect incremental backup. In that situation, the backup is now Full instead.

> - [PR #19688][kubevirt-bot] CBT overlays are now keyed by volume and PVC name. Pointing a volume at a different PVC triggers a Full backup instead of an incorrect incremental.

### OCI export archives preserve Image Indexes on registry push

- Affects: KubeVirt OCI export archives

When exported archives were pushed with `skopeo copy --multi-arch=all`, the registry could receive a plain manifest that lost platform information; `oras copy --from-oci-layout vm.oci.tar:latest` could also fail to resolve `latest`. Archives now store the Image Index as a blob and reference it from `index.json` with `ref.name: latest`, so pushes produce an Image Index in the registry.

> - [PR #19485][0xFelix] Exported OCI archives now contain the Image Index as a blob referenced from index.json with ref.name "latest", so pushing them creates an Image Index in the registry

### [Silent failure] Dedicated-CPU VM host CPU reservation overcount fixed

- Affects: Dedicated-CPU VMIs combining `supplementalPool` IO threads with `EmulatorThreadCompleteToEvenParity`

An odd IO-thread count could silently reserve one extra host CPU; the example requested 11 instead of the correct 10. The reservation now counts guest vCPUs, IO threads, and the emulator thread once.

> - [PR #19228][Sreeja1725] Alpha: Add CPUsWithDRA to provision dedicated VMI CPUs via a DRA CPU driver instead of kubelet's CPU manager.

### virt-handler preserves externally managed mediated devices

- Affects: setups with externally created mediated devices, including DRA-owned `vfio-ap` devices and devices supplied by a GPU operator

When `mediatedDeviceTypes` was empty, virt-handler's periodic sweep could delete unknown mediated devices, leaving the guest AP bus empty while the VMI stayed Ready. It now preserves externally created devices, while clearing `mediatedDeviceTypes` still removes KubeVirt-created types.

> - [PR #19099][HarshithaMS005] virt-handler no longer deletes mediated devices it did not create, so DRA-owned vfio-ap devices are not torn out from under running VMs

### [Silent failure] Incremental pull-mode backup maps include previously skipped changed blocks

- Affects: incremental pull-mode backups on fragmented disks

When block-allocation and dirty-bitmap ranges ended at different offsets, the merge could skip changed extents between them, leaving backups incomplete without reporting an error. The map-generation loop now resumes from the offset it emitted, so those skipped ranges are requested and included.

> - [PR #19315][Acedus] Fix: incremental pull-mode backups no longer silently leave out changed blocks from the backup map on fragmented disks.

### Common-instancetypes version label preserved in captured ControllerRevisions

- Affects: referenced instancetype or preference objects with the `instancetype.kubevirt.io/common-instancetypes-version` label

When one of these objects was captured into a `ControllerRevision`, its label was dropped. Captured revisions now retain the label, including across upgrades.

> - [PR #19006][lyarwood] Propagate the `instancetype.kubevirt.io/common-instancetypes-version` label from referenced instancetype and preference objects to their captured `ControllerRevision` metadata labels.

### virt-launcher resets guest-agent VMStats caches on domain reboot

- Affects: VMStats collection with the `VMStatsCollector` feature gate and guest-agent data

After a guest OS reboot, VMStats could return stale hostname, OS, or network-interface data until the cache TTL expired. When libvirt reports a domain reboot, the next VMStats request fetches fresh guest-agent data.

> - [PR #19169][0xFelix] VMStats guest agent data is refreshed after a guest OS reboot instead of being served from cache until its TTL expires

### Invalid container disk image errors surfaced on VMs and VMIs

- Affects: VMs configured with `containerDisk.image`

When `containerDisk.image` has an incorrect image reference and the virt-launcher init container enters `InvalidImageName`, the VM previously stayed in `Starting` without reflecting the error on the VMI. The VMI now reports reason `InvalidImageName`, the VM displays `ErrImagePull`, and VM pool autoheal may apply in this condition.

> - [PR #18872][ameya-keskar] Report InvalidImageName container disk errors on the VMI and surface ErrImagePull on the VM instead of leaving the VM stuck in Starting. Pool autoheal for VM's may apply, on the error condition, where it did not before.

### Guest clock synchronization fixed for repeated unpauses and migrations

- Affects: VMIs that undergo repeated pause/unpause cycles or migration finalization

The one-time `sync.Once` guard meant `setGuestTime()` ran only once per VMI lifetime, leaving later unpauses or migration finalization without time synchronization.

> - [PR #19089][jean-edouard] Guest clock now properly syncs after multiple pause/unpause

### Single attachment pod for utility and hotplug volumes

- Affects: VMs using both utility volumes and hotplug volumes

Using both volume types could create two hotplug attachment pods. The single pod contains the appropriate mounts and devices for the volumes used.

> - [PR #18843][awels] BugFix: Prevent duplicate attachment pods when both utility volumes and hotplug volumes are used in a single VM

### `VMBackup` status loss on finalizer addition fixed

- Affects: `VMBackup` resources whose controller adds a finalizer (Inferred)

> - [PR #18843][awels] BugFix: Prevent duplicate attachment pods when both utility volumes and hotplug volumes are used in a single VM

### [Silent failure] Migration failure status race fixed

- Affects: migrations during source ownership transfer

When `SourceNode` was set before `TargetNode`, a race could mark a migration as failed even though it was fine. The ownership-transfer logic now waits for the domain-migrated condition before taking the unknown-host failure path.

> - [PR #18604][awels] BugFix: Fixed race condition where target node and source fields were racing which could cause a migration failed.

### virt-handler honors `--kubelet-root` for CPU Manager detection

- Affects: clusters with a non-default kubelet root directory, including k3s, k0s, and MicroK8s

Previously, `virt-handler` ignored `--kubelet-root` when checking whether CPU Manager was enabled and looked for `cpu_manager_state` at the default path, causing incorrect node labeling.

> - [PR #16346][ram2valar] Fixed a bug where virt-handler ignored the --kubelet-root flag when detecting whether CPU Manager is enabled, causing incorrect node labeling on clusters with a non-default kubelet root directory (e.g. k3s, k0s, MicroK8s).```

### Suspected orphaned storage recording rule metric reference fixed

- Affects: installations using the `pvc:kubevirt_suspected_orphaned_storage_bytes:info` recording rule

The rule referenced `kube_pod_spec_volumes_persistentvolumeclaims`, which does not exist, so that input could not contribute to its result. It now uses `kube_pod_spec_volumes_persistentvolumeclaims_info` to calculate suspected orphaned storage bytes.

> - [PR #18986][Ronilerr] Changing recording rule expression

### virt-handler unresponsive socket handling fixed

- Affects: virt-handler's unresponsive domain socket handling (Inferred)

Multiple unresponsive sockets can now be cleaned up in one pass, including when sockets switch between responsive and unresponsive states. Untracking one unresponsive socket no longer blocks another's expiration.

> - [PR #18970][xpivarc] Bug fix: unresponsive sockets are not correctly accounted

### Same-name VM and VMI time-series reuse fixed with UID labels

- Affects: installations scraping or querying `kubevirt_vm_info` or `kubevirt_vmi_info` (Inferred)

Previously, deleting and recreating an object with the same name and namespace reused its Prometheus time series. The `uid` label now makes the recreated object a distinct series; VM UIDs stay stable across stop/start, while VMI UIDs change on each start.

> - [PR #18854][sradco] Add uid label (Kubernetes object UID) to kubevirt_vm_info and kubevirt_vmi_info

### Stopped VM status corrected for `RunStrategyRerunOnFailure`

- Affects: VMs using `RunStrategyRerunOnFailure`

When one of these VMs was manually stopped while its PVC was unbound, it could show `WaitingForVolumeBinding` even though the controller had not created a VMI. It now reports `Stopped` regardless of PVC state.

> - [PR #18787][davmarro] Fixed a bug where a VM with RunStrategyRerunOnFailure incorrectly showed WaitingForVolumeBinding status after being manually stopped when its PVC was unbound.

### Host NUMA distances passed through to guest NUMA cells

- Affects: VMs on multi-NUMA hosts with host distance data available (Inferred)

For VMs spanning three or more NUMA nodes, remote nodes previously appeared equidistant, limiting scheduling decisions. The guest OS now sees the host NUMA distance matrix, with host node IDs mapped to sequential guest IDs.

> - [PR #18829][omastama-spec] Added NUMA node distance passthrough from host topology to guest VMs, enabling performance-sensitive applications to make optimal scheduling decisions on multi-NUMA systems.

### VM alert evaluation restored to cluster/node scope

- Affects: the `OrphanedVirtualMachineInstances` or `VirtLauncherPodsStuckFailed` alert rules

Previously, `OrphanedVirtualMachineInstances` could fire once per VM namespace even when `virt-handler` was ready, and `VirtLauncherPodsStuckFailed` applied its 200-launcher threshold within a single namespace. The namespace label now identifies the KubeVirt install namespace instead of coming from virt-launcher series.

> - [PR #18822][sradco] Fix OrphanedVirtualMachineInstances false positives and VirtLauncherPodsStuckFailed per-namespace threshold by restoring cluster-scoped evaluation with a static install-namespace label

### [Silent failure] Sidecar hooks preserve QEMU command-line arguments across XML round-trips

- Affects: sidecar hooks carrying QEMU `qemu:commandline` arguments through XML round-trips (Inferred)

During this round-trip, the arguments could be silently dropped.

> - [PR #18460][dasionov] Fix sidecar hooks silently dropping QEMU command-line arguments (qemu:commandline) during XML round-trip

### Preferred panic device creation fixed

- Affects: VMs using a `VirtualMachinePreference` with `PreferredPanicDeviceModel` set

Previously, the preference was silently ignored when the VMI had no panic device, leaving it without the device it was meant to ensure. KubeVirt now creates a panic device with the preferred model in that situation.

> - [PR #18314][lyarwood] Bug fix: PreferredPanicDeviceModel now creates a panic device when none exist on the VMI, instead of silently ignoring the preference.

### SR-IOV, GPU and PCI host-device VM start failure fixed

- Affects: VMIs using SR-IOV, GPU or PCI host devices in clusters with the `Plugins` feature gate enabled and at least one Plugin

Before this fix, those VMIs failed to start in that setup. They can now start, with `subsystem` mode added to the generated host-device entries.

> - [PR #19771][kubevirt-bot] Fix VMIs with SR-IOV, GPU or PCI host devices failing to start when the Plugins feature gate is enabled and any Plugin exists in the cluster.

### Node hooks inherit the Plugin-level `failureStrategy`

- Affects: Plugins with node hooks that omit their own `failureStrategy`

When a Plugin has `spec.failureStrategy: Ignore`, a failing node hook without its own strategy no longer blocks VM start, stop, or migration. If neither level sets a strategy, the hook defaults to `Fail`; a hook-level strategy takes precedence.

> - [PR #19863][kubevirt-bot] Node hooks now fall back to the Plugin-level failureStrategy when they do not set their own, like launcher hooks. Plugins with spec.failureStrategy: Ignore no longer block VM start, stop or migration when one of their node hooks without its own failureStrategy fails.

### virt-handler panic on device-plugin restart fixed

- Affects: nodes with host devices

Restarting a device plugin could panic virt-handler; on nodes with host devices, this happened on every kubelet restart.

> - [PR #19743][kubevirt-bot] Fixed a virt-handler panic when a device plugin is restarted, which on nodes with host devices happened on every kubelet restart.

### Live migration targeting fixed for virt-handler upgrades

- Affects: live migrations during rolling upgrades (Inferred)

A migration from an updated virt-handler node could be scheduled onto a node that had not yet been updated and get stuck. The fix restricts targets to nodes running the desired virt-handler image.

> - [PR #19732][stu-gott] Live migrations are no longer scheduled from a node running an updated virt-handler onto a node that has not yet been updated, preventing migrations from getting stuck during a rolling upgrade.

### virt-handler gRPC connection leak fixed

- Affects: virt-handler client initialization (Inferred)

When `newClientWithInfoClient` failed during `NewClient` initialization, the gRPC connection could remain open. The connection is now closed on that path, while the error is still returned.

> - [PR #19492][caxu-rh] Fixed a bug that caused a gRPC connection to leak when virt-handler client initialization fails

### VM startup recovers overlays after External snapshot crashes

- Affects: VMs using External snapshots

If virt-launcher crashes during an External snapshot and leaves the VM's disks on overlays, startup previously stopped with `ManualRecoveryRequired`; recovering writes required manual overlay commits or restoring a snapshot. On start, the replacement VMI now boots on the remaining overlays and commits their writes live, clearing the startup block and releasing the scratch volume after the commit is verified; recovery is still refused if the volume or scratch state fails checks.

> - [PR #19574][noamasu] Guest writes are recovered when a VM starts after a virt-launcher crash during an External mode snapshot.

### Hotplug mount teardown now preserves PVC files

- Affects: VM workloads using hotplug volumes mounted through the hotplug mount root (Inferred)

When kubelet cleanup ran before virt-handler unmounted a utility volume, it could delete files on that volume’s PVC. Teardown now unmounts the memory-backed root instead, preserving files on mounted PVCs.

> - [PR #19520][noamasu] Use a memory-backed emptyDir for the hotplug mount root to support mount propagation.

### Orphaned CBT overlay cleanup on disable and disk hot-unplug

- Affects: setups using CBT with retained backend PVCs or hot-unplugged disks

Hot-unplugging a disk in certain circumstances or disabling CBT could leave orphaned overlays. These overlays are now removed; after CBT is re-enabled, the next backup falls back to a full backup instead of failing checkpoint redefinition.

> - [PR #19435][awels] Cleanup CBT overlays when disabling CBT

### VirtualMachineExport volume-name collisions fixed

- Affects: VirtualMachineExports for VMs with claims whose names differ only by dots or collide with fixed exporter-pod volume names

Sanitizing claim names made dot-differing names such as `my.disk` and `my-disk` produce duplicate volume names, so the API server rejected the exporter pod and the export never completed. The pod now uses unique index-based volume names, allowing the export to complete.

> - [PR #19064][0xFelix] Fixed a bug where a VirtualMachineExport of a VM with two claims whose names differ only in dots would never complete

### `InjectNotify` closes the listener before returning errors

- Affects: use of `InjectNotify` with a Unix socket

When the socket path was invalid or setting its file ownership failed, `InjectNotify` returned an error but could leave the listener open.

> - [PR #19534][caxu-rh] Fixed a bug where a Unix socket listener may be left open if an error occurs while setting file ownership on the socket

### Backups now fail promptly when their VMI is deleted

- Affects: backups of VMIs

When a VMI was deleted during a backup without a status change, the backup previously remained pending until its TTL. It now fails with `SourceLost`, and pull-mode exports are removed.

> - [PR #19484][Acedus] Backups whose VMI is deleted mid-backup now fail immediately, and pull-mode backups remove their export.

### `GetVMStats` preserves successful statistics when a sub-command fails

- Affects: setups that collect VMStats through `GetVMStats` (Inferred)

Previously, one failed sub-command caused virt-handler to discard data returned by the others; failures are now recorded in the `VMStats.Errors` map.

> - [PR #19192][machadovilaca] Keep partial VMStats when a sub-command fails

### Duplicate-PVC export failures now report status

- Affects: `VirtualMachineExport`s from VM, VMSnapshot, or VMTemplate sources that reference the same PVC in multiple volumes

Previously, the exporter pod was rejected for duplicate volume names before status was written, leaving the export with an empty phase, no conditions, and endless backoff. Now it reports `Pending` with a `DuplicatePVC` ready condition naming the claims, and creates no exporter pod until the duplicate references are fixed.

> - [PR #19057][0xFelix] Fixed a bug where a VirtualMachineExport of a source referencing the same PVC from more than one volume would never complete

### SeaBIOS now honors the PCI-hole setting

- Affects: VMs using SeaBIOS with the `kubevirt.io/disablePCIHole64` annotation

Previously, SeaBIOS ignored QEMU’s setting and programmed its own 64-bit PCI hole, so guests with more than 4Gi of memory could still have the hole. It now leaves that hole alone, allowing legacy guests such as Windows Server 2003 to run with 4Gi of memory or more.

> - [PR #18941][0xFelix] The kubevirt.io/disablePCIHole64 annotation now also stops SeaBIOS from programming a 64-bit PCI hole, allowing legacy guests like Windows Server 2003 to run with 4Gi of memory or more.

### Velero post-backup unfreeze timeout extended to 60 seconds

- Affects: Velero backups with post-backup unfreeze hooks

Previously, the internal gRPC timeout could fail unfreeze when Windows guest thaw took longer than 20 seconds, regardless of the post-hook timeout annotation. Unfreeze can now run for up to 60 seconds, and generated post-backup hooks set `post.hook.backup.velero.io/timeout` to `60s`.

> - [PR #19166][alromeros] Bugfix: Increase timeout for velero post-backup hook

### Online file-backed disk expansion now uses PVC capacity

- Affects: VMs with PVC-backed file disks undergoing online expansion (Inferred)

During I/O on filesystem volumes, changing free space could repeatedly trigger `BlockResize` attempts.

> - [PR #19120][akalenyu] BugFix: blockresize spam when doing I/O for filesystem volumes

### Online `BlockResize` now prevents guest disk shrinking

- Affects: VMs using online disk resizing (Inferred)

When the newly observed capacity is less than or equal to the tracked size, KubeVirt skips the resize rather than shrinking the guest disk.

> - [PR #19120][akalenyu] BugFix: blockresize spam when doing I/O for filesystem volumes

### Restored `/dev/urandom` access across cgroup rebuilds

- Affects: VMIs without a virtio-rng device

During block-volume hotplug or live migration, rebuilding the cgroup allow list could remove `/dev/urandom` access; on s390x, this crashed `virtqemud` and could cause sync or migration failures. The default device rules now always allow `/dev/urandom` across cgroup rebuilds on cgroups v1 and v2, regardless of VMI configuration or architecture.

> - [PR #19131][vishnubijukumar] Fixed a bug where hotplugging a block volume or triggering a live migration on s390x crashed virtqemud due to /dev/urandom being removed from the cgroup device allow list. /dev/urandom (char 1:9) is now unconditionally included in the default cgroup device rules, consistent with the OCI runtime specification. This also fixes a latent bug on x86_64 where /dev/urandom was incorrectly revoked but masked by the RDRAND hardware entropy fallback.

### OCI exports skip non-disk backend directories

- Affects: OCI VirtualMachineExports of VMs with persistent TPM or EFI backend storage

The digest pass tried to open `disk.img` in backend state directories, leaving exports `Pending` and serving no disks or manifests. It now skips those directories and serves the export contents.

> - [PR #19085][0xFelix] Fixed a bug where an OCI VirtualMachineExport of a VM with a persistent TPM or EFI never became ready

### Centos Stream 10 socket-path compatibility fixed for KubeVirt images

- Affects: Centos Stream 10-based deployments using the `virt-handler` or `vm-killer` images (Inferred)

Centos Stream 10 creates sockets in `/run/`, while KubeVirt expects them in `/var/run/`; this prevented `vm-killer` from resolving the dbus socket and `virt-handler` from mounting the multipathd socket. The images now include a `/var/run -> ../run` symlink, so `/var/run/` resolves to `/run/`.

> - [PR #18951][dsanatar] image changes for Centos Stream 10 compatibility, notably adds /var/run -> ../run symlink to virt-handler image layer due to how CS10 creates sockets in /run/ but kubevirt expects them to exist in /var/run/

### Vera Rubin GPU passthrough configuration fixed

- Affects: VMs passing through NVIDIA Vera Rubin GPUs (`10de:3041`) with `GraceIOVirtualization`

Previously, `virt-handler` did not recognize the GPU as a Grace GPU, so `virt-launcher` skipped the required domain configuration. The launcher now configures a 2 TiB `pcihole64` aperture and the `arm-smmuv3` IOMMU model for it.

> - [PR #19070][lyarwood] Vera Rubin GPU (10de:3041) is now recognised as a Grace GPU, enabling GraceIOVirtualization for VMs with a passthrough Vera Rubin GPU.

### Terminated launcher pod UIDs pruned from VMI active-pod status

- Affects: VMIs with hotplug volume reconciliation

After successful live migration, terminated source launcher pods in `Succeeded` or `Failed` phase could remain alongside the target UID in `VMI.status.activePods`, causing `findVirtlauncherUID()` to return an empty UID and hotplug reconciliation to loop with `SyncFailed`. Those terminated pods are now excluded, leaving the running target UID so hotplug reconciliation can proceed.

> - [PR #18287][dasionov] Ensure terminated launcher pods are pruned from VMI status.activePods after live migration

### Prometheus EndpointSlice discovery permission fixed

- Affects: Prometheus target discovery through EndpointSlices in the KubeVirt namespace

When the Prometheus CR sets `serviceDiscoveryRole` to `EndpointSlice`, Prometheus can list those EndpointSlices for target discovery.

> - [PR #19041][machadovilaca] Allow Prometheus to list EndpointSlices in the KubeVirt namespace

### Offline PVC expansion now checks actual preallocation status

- Affects: thin-provisioned PVC/DataVolume-backed disks (Inferred)

During offline PVC expansion, a full thin-provisioned disk could be mistaken for preallocated based on its virtual and allocated sizes, causing the expanded space to be unnecessarily preallocated.

> - [PR #18827][germanovm] Fix PVC expansion preallocating new space on full thin provisioned disks

### Stale CA fixed in export manifest `ConfigMap`s

- Affects: `VirtualMachineExport` manifests used for CDI disk imports

After export CA rotation and exporter Pod recreation, an existing manifest `ConfigMap` could retain the old CA, causing CDI imports using it to fail TLS verification. The controller now updates the existing `ConfigMap` with the new CA.

> - [PR #18965][lukashes] Fixed VirtualMachineExport manifests retaining stale CA certificates when exporter Pods are recreated after CA rotation, which could cause CDI disk imports to fail TLS verification

### Source ghost record leak during decentralized live migration fixed

- Affects: decentralized live migration where a later same-name VMI may be scheduled on the same node

On a busy source node, cleanup could miss the brief `Shutoff+Migrated` state, leaving a ghost record that blocked a later same-name VMI from registering and starting and could cause the migration to time out. Once the source is terminal, virt-handler now continues through `sync()` to delete the source ghost record, allowing a same-name, new-UID receiver on that node to register and start.

> - [PR #18988][davmarro] Fixed a source-side ghost record leak on decentralized live migration that could prevent a subsequent same-name VirtualMachineInstance from starting on the same node.

### False `GuestFilesystemAlmostOutOfSpace` alerts for Windows UDF media fixed

- Affects: Windows VMs with a UDF CDROM/ISO attached and monitored by `GuestFilesystemAlmostOutOfSpace`

The QEMU Guest Agent reports attached optical media with 0 bytes free, and the case-sensitive filesystem-type exclusion missed uppercase `UDF`, causing the alert to fire. The exclusion is now case-insensitive, so the alert no longer fires for UDF CDROM/ISO media.

> - [PR #18994][avlitman] Fix GuestFilesystemAlmostOutOfSpace firing on Windows VMs when a UDF CDROM/ISO is attached.

### NoReady alerts corrected to avoid firing with Down alerts

- Affects: KubeVirt VirtController, VirtHandler, or API NoReady and Down monitoring alerts (Inferred)

Previously, the Down and NoReady alerts could fire simultaneously. Now, NoReady alerts fire only for pods that are running but not ready.

> - [PR #18942][Ronilerr] Fix no ready alerts

### PCI passthrough startup failure fixed

- Affects: VMs using PCI host-device passthrough on hosts without `CONFIG_VFIO_DEVICE_CDEV` support

These VMs failed to start even when the `IOMMUFD` feature gate was disabled. On kernels without IOMMUFD support, nonexistent cdev devices are skipped and passthrough uses `/dev/vfio/<group>`; kernels with IOMMUFD support retain cdev behavior.

> - [PR #18905][fossedihelm] Fix PCI passthrough regression on kernels without CONFIG_VFIO_DEVICE_CDEV support. VMs with GPU/PCI host devices now correctly fall back to legacy VFIO group interface when IOMMUFD cdev is unavailable.

### Memory overcommit bypassed for hugepages-enabled VMIs

- Affects: VMIs configured to use hugepages

Previously, applying memory overcommit to these VMIs produced an unaligned memory value. They now use their memory value without overcommit.

> - [PR #18880][jean-edouard] Memory overcommit is now ignored for VMIs with hugepages

### VFIO device ownership error handling fixed

- Affects: setups that use VFIO device ownership preparation

When resolving the launcher mount root failed during VFIO device ownership preparation, a nil pointer dereference could terminate virt-handler and stop VM reconciliation on that node. It now returns errors other than `os.ErrNotExist` to its caller, which propagates them into a `SyncFailed` event instead of terminating virt-handler.

> - [PR #18867][cwedgwood] Fix a virt-handler crash that could occur when preparing VFIO device ownership fails.

### Atomic disk writes for virt-handler ghost and mount records

- Affects: setups where virt-handler writes ghost or mount records to disk

If virt-handler restarted while writing records, the implementation could lose data or produce corrupt records.

> - [PR #18716][xpivarc] Bug fix: Ensuring that records (ghost records, mount records) are written atomically to disk

### Premature domain undefine during slow QEMU startup fixed

- Affects: VMs using hotplug (Inferred)

During slow QEMU startup, premature domain undefine left VMs transient and broke hotplug. After the upgrade, this premature undefine is fixed.

> - [PR #18545][amasolov] Fix premature domain undefine during slow QEMU startup that left VMs transient and broke hotplug

### CBT backend PVC request increased by 1Gi

- Affects: VMs with CBT enabled that require a backend persistent-state PVC

Previously, these PVCs requested only 10Mi, which was insufficient for most CBT use cases.

> - [PR #18470][Acedus] Backend storage PVC size is increased by 1Gi for VMs with CBT enabled.

### Domain stats collection no longer blocks on locked domains

- Affects: live migration with domain stats collection

During live migration, `GetDomainStats()` could wait up to 30 seconds on the migration-held libvirt job lock, time out, log errors, and disrupt upgrade flows monitoring virt-launcher health. It now skips locked domains and serves cached values; stats collection continues for other cases, though skipped domains can still produce an `empty DomainStats` error log about every 3.25 seconds during migration.

> - [PR #18362][iholder101] Bug fix: During live migration, domain stats collection no longer blocks on the libvirt job lock for up to 30 seconds. The NOWAIT flag is used to skip locked domains instead of waiting, avoiding lock contention and disruptive error logs.

### vmexport proxy file descriptor leak fixed

- Affects: setups using the vmexport proxy (Inferred)

Before this fix, the vmexport proxy leaked file descriptors.

> - [PR #18400][awels] BugFix: file descriptor leak in vmexport proxy

### `client-go` `AsyncSubresourceHelper` leak fixed

- Affects: `AsyncSubresourceHelper` `StreamInterface` consumers using `AsConn()`

Using `AsConn()` instead of `Stream()` leaked a goroutine and connection resource. The `AsConn()` path no longer has this leak.

> - [PR #18410][vikin91] Fix a goroutine (and connection resource) leak in client-go's AsyncSubresourceHelper when its StreamInterface is consumed via AsConn() instead of Stream() (affects VSOCK, PortForward TCP/UDP, and similar consumers).

### Running VMIs get a 3-minute launcher reconnection grace period

- Affects: running VMIs

When virt-handler restarted and temporarily lost the launcher socket connection, a running VMI was immediately marked Failed. It now remains within the 3-minute reconnection grace period before it can transition to Failed.

> - [PR #18453][iholder101] Bug fix: Running VMIs are no longer immediately marked as Failed when virt-handler restarts and temporarily loses the launcher socket connection. A 3-minute grace period (matching the existing Scheduled VMI behavior) now allows the launcher client to reconnect before transitioning the VMI to Failed.

### Invalid network interface bindings rejected at admission

- Affects: VMs or VMIs with a network interface configured with no binding or more than one binding

Previously, these interfaces were not rejected at admission and failed silently at runtime.

> - [PR #18374][orelmisan] Fixed missing validation for network interface bindings: interfaces with no binding or multiple bindings are now rejected at admission instead of failing silently at runtime.

### VFIO cdev devices exposed to virt-launcher containers

- Affects: ARM64 systems with SMMUv3 using IOMMUFD-based GPU passthrough

This fixes IOMMUFD-based GPU passthrough on ARM64 systems with SMMUv3.

> - [PR #18300][lyarwood] Bug fix: Expose VFIO cdev devices (/dev/vfio/devices/vfioN) to virt-launcher containers, fixing IOMMUFD-based GPU passthrough on ARM64 systems with SMMUv3.

### VMI owner resolution fallback for hotplug attachment pods

- Affects: hotplug attachment pods

If an attachment-pod event was processed after the virt-launcher pod had been deleted, owner resolution could fail and the VMI was not notified. The attachment pod's owner name and UID annotations now provide a fallback when the owner-reference chain cannot resolve the VMI.

> - [PR #18222][lukashes] Fixed VMI owner resolution when an attachment pod is deleted while the virt-launcher pod is gone

### Hotplug volume detach deadlock fixed

- Affects: hotplug volume detachment

Detaching a hotplug volume that had already been removed could deadlock.

> - [PR #18044][maxwmsft] BugFix: hotplug volume detach deadlock for already removed volumes.

### Hotplug volume detach deadlock fixed

- Affects: VMIs exchanging RWO PVCs between attachment pods

Hotplug volume detachment could deadlock during these exchanges.

> - [PR #18045][maxwmsft] BugFix: hotplug volume detach deadlock when VMIs exchange RWO PVCs between attachment pods.

### Session-mode memlock calculation fixed for memory-locked VMs

- Affects: VMs requiring memory locking for VFIO, realtime, SEV, or explicit memlock, running in session mode

Libvirt's memlock calculation could fail for these VMs. The memlock `prlimit` is now unlimited, so the calculation succeeds.

> - [PR #17857][lyarwood] Set unlimited memlock prlimit for VMs requiring memory locking (VFIO, realtime, SEV, explicit memlock), ensuring libvirt's memlock calculation always succeeds in session mode.

### NAD changes on non-migratable VMs now require restart

- Affects: NAD references on non-migratable VMs

When a NAD reference changed, it was synced to the VMI spec and triggered a `MigrationRequired` condition; because the VM could not migrate, the condition persisted and the change never applied. The changed reference is now withheld from the VMI spec.

> - [PR #18194][frenzyfriday] Require restart for NAD swap for non-migratable VMs

### Low virt count alerts now use the `running` rule

- Affects: setups that consume low virt count alerts (Inferred)

Previously, these alerts used the `up` rule instead of the `running` rule.

> - [PR #18380][Ronilerr] Low virt count alerts are using running rule intead of up rule

### Incorrect restarts of NFS-backed VMs fixed

- Affects: VMs with NFS-backed storage

During temporary NFS unavailability, such as an NFS server failover, these VMs could be restarted incorrectly. That unavailability no longer causes the incorrect restart.

> - [PR #17752][iholder101] Bug fix: VMs with NFS-backed storage are no longer incorrectly restarted during temporary NFS unavailability (e.g., NFS server failover).

### VM restore handling for backend storage fixed

- Affects: VMs using a `VMStateStorageClass` without snapshot support

When `VMStateStorageClass` lacked snapshot support, the backend PVC could not be included in the snapshot, causing restore issues. Restore now skips reconciliation for a backend volume absent from snapshot contents and reuses the existing backend PVC.

> - [PR #17286][alromeros] Bugfix: Fix VM restore when backend storage is not snapshottable

### `kubevirt-plugin-sidecar-subpath-policy` importer-pod rejection fixed

- Affects: installations using the `kubevirt-plugin-sidecar-subpath-policy` (Inferred)

Importer pods could be rejected when the policy’s `matchConditions` evaluated them. The policy now selects virt-launcher pods with an `objectSelector` before CEL evaluation.

> - [PR #18392][akalenyu] BugFix: Importer pod rejected by ValidatingAdmissionPolicy 'kubevirt-plugin-sidecar-subpath-policy'

### `hugetblfs-dir` EmptyDir hugepage backing fixed

- Affects: VMs using the `hugetblfs-dir` EmptyDir (Inferred)

When QEMU needed hugepages via `-mem-path`, this volume masked the parent hugetlbfs mount at `/dev/hugepages`, preventing allocation. The EmptyDir is now explicitly hugepage-backed, fixing that problem.

> - [PR #18272][laxmi-333] Bug fix: Fixed hugetblfs-dir EmptyDir volume masking the parent hugetlbfs mount at /dev/hugepages, which prevented QEMU from allocating hugepages via -mem-path. The EmptyDir is now explicitly hugepage-backed.

### VMI controller retries sync after expectations timeout

- Affects: VM/VMI deletion involving pod, VMI, or PVC expectations

When no pod event woke the controller while expectations were unmet, deletion could wait for the next informer resync, which was hours away by default. The retry now occurs after the expectations timeout, five minutes by default.

> - [PR #18206][lukashes] Fixed a delay in VM/VMI deletion that could happen when the VMI controller missed a pod event.

### `ClusterConfig` CRD delete-handler tombstone panic fixed

- Affects: installations using the `ClusterConfig` CRD

After a missed CRD delete event, the informer could deliver a `cache.DeletedFinalStateUnknown` tombstone instead of the CRD, causing the handler to panic. The handler now extracts the CRD from the tombstone, and logs an error and returns if it cannot unwrap the tombstone or it does not contain a CRD.

> - [PR #17880][Shreesha001] Fix panic in ClusterConfig CRD delete handler when the informer delivers a tombstone object instead of the deleted CRD.

## Enhancements (39)

- `IncrementalBackup` feature gate promoted to Beta
- Backup API `backup.kubevirt.io/v1beta1` introduced
- Incremental backups for stopped VMs (To use: `OfflineIncrementalBackup` feature gate, disabled by default and requiring `IncrementalBackup`)
- External-mode `VirtualMachineSnapshot` snapshots use QCOW2 overlays (To use: The `ExternalVMSnapshot` feature gate (Alpha) must be enabled to use External mode.)
- Alpha UEFI Secure Boot for ARM64 guests (To use: Both `ARM64SecureBoot` and `FirmwareAutoSelection` feature gates, plus CentOS Stream 10 virt-launcher images with suitable QEMU, libvirt, and edk2 firmware.)
- Alpha DRA allocation for dedicated VMI CPUs (To use: Enable the `CPUsWithDRA` feature gate)
- VMI `pod-status` reporting for ready network interfaces (To use: `NetworkDevicesWithDRA` feature gate must be enabled)
- Declarative VM-state PVC selection and creation (To use: `DeclarativeVMState` feature gate)
- Multiple I/O threads for virtio-scsi controllers (To use: The `SCSIMultiIOThread` feature gate enabled and an `IOThreadsPolicy` assigned to the VMI)
- virtio-blk `auto` IOThread policy gains pool allocation (To use: Enable the `MultiIOThreadAutoPolicy` configuration toggle together with the `SCSIMultiIOThread` feature gate.)
- Panic-device label added to `kubevirt_vmi_info` metric
- Sidecar plugin `PreBoot` launcher hook
- Sidecar `PreMigrationSource` hook for libvirt migration parameters
- `pkg/render` adds cluster-free VM-to-Pod rendering
- Configurable TLS key exchange groups for virt pod endpoints (To use: Enable the `TLSGroupPreferences` feature gate and configure `TLSConfiguration.Groups` with the desired groups. The gate is Alpha and disabled by default.)
- Proxy-mode migration endpoint discovery from labeled Routes, Ingresses, and Services (To use: Proxy mode enabled, no cross-cluster network configured, and a Route, Ingress, or Service with the appropriate endpoint-discovery label)
- virt-handler downtime gauge for latest successful VMI migrations
- Container waiting reasons added to `Virt*Down` alerts
- VM guest memory opt-out from kernel merging (To use: Set the `kubevirt.io/mergeable-memory` annotation to `false` on the VMI to opt out of kernel memory merging.)
- GPU utilization and graphics-engine activity recording rules added
- TCP proxy support for decentralized live migration (To use: `DecentralizedLiveMigration` feature gate)
- VMI GPU recording rules for DCGM metrics
- VM disks gain the `directsync` cache mode (To use: Configure `cache=directsync` for the VM disk.)
- VSOCK transport for `virtctl port-forward`, `virtctl ssh`, and `virtctl scp` (To use: Set `--vsock` to use VSOCK transport. For `virtctl port-forward`, `--vsock-tls` controls the TLS handshake and defaults to true.)
- Guest device payload forwarding through `VMStatsCollector`
- `GetVMStats` adds guest filesystem information (To use: `GuestGetFsInfo` in the `GetVMStats` request)
- DRA network claims reach binding sidecars and migration targets
- KubeVirt recording rule for suspected orphaned PVC storage
- Dynamic live-migration downtime tuning (To use: Alpha `MigrationDowntimeTuning` feature gate and opt-in through `.spec.experimental.downtimeTuning` in `MigrationPolicy`)
- ImageVolume containerDisk support for OCI artifacts and root-level disks (To use: For non-container OCI artifacts, set the VMI annotation `kubevirt.io/image-volume-skip-digest-resolution: "true"` to skip incompatible digest-resolving init containers.)
- Namespace labels added to KubeVirt alerts
- Network binding plugin sidecar registration-name inference (To use: `NETWORK_BINDING_PLUGIN_NAME` environment variable)
- KubeVirt emits the `kubevirt_vmi_guest_os_panic_total` metric
- Separate disk and source volume names in `virtctl addvolume` (To use: `--disk-name` on `virtctl addvolume`)
- Snapshot-backed `inferFromVolume` support (To use: `inferFromVolume` configured for a snapshot-backed DataVolume, DataVolumeTemplate, or DataSource)
- KubeVirt adds a leading virt-controller recording rule and alert
- Built-in KubeVirt roles gain VM memory-dump removal permission
- Optional zstd compression for live migration streams (To use: `AdvancedLiveMigration` feature gate)
- Local-namespace VSOCK dialing in `virt-handler` (To use: `ns_mode` set to `local`)

## Dependency updates (7)

### KubeVirt build toolchain upgraded to Go 1.26.10

- [GO-2026-6599](https://osv.dev/vulnerability/GO-2026-6599), [GO-2026-6600](https://osv.dev/vulnerability/GO-2026-6600), [GO-2026-6603](https://osv.dev/vulnerability/GO-2026-6603), [GO-2026-6604](https://osv.dev/vulnerability/GO-2026-6604), [GO-2026-6605](https://osv.dev/vulnerability/GO-2026-6605), [GO-2026-6607](https://osv.dev/vulnerability/GO-2026-6607), [GO-2026-6608](https://osv.dev/vulnerability/GO-2026-6608), [GO-2026-6609](https://osv.dev/vulnerability/GO-2026-6609), [GO-2026-6610](https://osv.dev/vulnerability/GO-2026-6610), [GO-2026-6611](https://osv.dev/vulnerability/GO-2026-6611), [GO-2026-6612](https://osv.dev/vulnerability/GO-2026-6612), [GO-2026-6613](https://osv.dev/vulnerability/GO-2026-6613), [GO-2026-6617](https://osv.dev/vulnerability/GO-2026-6617), [GO-2026-6809](https://osv.dev/vulnerability/GO-2026-6809)

The previous Go 1.26.6 toolchain was affected by 14 published advisories that do not affect Go 1.26.10.

> - [PR #19590][lyarwood] Build KubeVirt with go 1.26.10

6 more, none mentioning security

- `- [PR #19521][kubevirt-bot] Updated common-instancetypes bundles to v1.8.0`
- `- [PR #19388][0xFelix] Updated virt-template to v0.3.0`
- `- [PR #19048][michalskrivanek] Libvirt updated to 12.5.0-4.el10`
- `- [PR #18908][kubevirt-bot] Updated virt-template to v0.2.3`
- `- [PR #18315][alancaldelas] Kubernetes dependencies have been updated to v0.36.4 (Kubernetes 1.36)`
- `- [PR #18761][lentzi90] Build KubeVirt with go 1.26.6`

## Other changes (31)

- `NetworkDevicesWithDRA` feature gate promoted to Beta
- VM-state PVC layout migrated to a canonical, VM-agnostic layout
- ContainerPath volumes graduate to Beta and are enabled by default
- `virt-launcher-monitor` rewritten as a statically linked C supervisor with a 2Mi reserve
- Live migration creates missing CBT overlays from source domain XML
- CBT overlay cleanup on volume unplug
- `VirtualMachineRestore` patch admission accepts valid JSON Patch operations
- virt-operator stops repeated GETs for the `virt-launcher-pod-mutator` webhook
- Hotplugged volumes attach independently
- virt-operator stabilizes duplicate endpoint messages and warnings
- Legacy firmware UUID calculation and synchronization removed
- Windows crash handling changed for `hyperv` panic devices
- VM identity labels added to KubeVirt inventory metrics
- Warning added for amd64 guests with ACPI disabled
- Clearer errors for VMs sharing NFS disk images
- `VMStatsCollector` returns cached empty results for unsupported guest-agent commands
- Custom seccomp profile files now use `0600` permissions
- Atomic writes for KubeVirt custom seccomp profiles
- PVC expansion test added to storage self-certification suite
- Cached shared setup for plugin CEL hook evaluators
- Legacy containerdisk CRC32 checksum computation removed
- `virtctl evacuate-cancel` now supplies the VMI node name
- VMI memory-overhead reporting graduates to GA
- s390x container image build fix with `hermetic_launcher` v0.0.15
- Online disk expansion avoids repeated `GetBlockInfo` calls
- Container-image builds now use `rules_img`
- virt-launcher pods cleaned up with migration objects
- Custom network binding plugin support for network conformance tests
- End-to-end test support for DRA
- Explicit SMMUv3 address capabilities for Grace I/O Virtualization
- DRA GPU and host-device feature gates move to Beta

---

Summaries are produced by AI, so they can be wrong, and they are provided without warranty.
