Cloud Hypervisor v55.0 Released!

Posted November 19, 2026 by Cloud Hypervisor Team ‐ 7 min read

This release has been tracked in v55.0 group of our roadmap project.

Security Fixes

This release fixes a use-after-free vulnerability in virtio-pmem during VM shutdown (#9326), which is also fixed in v53.1 and v54.1. See GHSA-3r3x-6hgg-2m38 for details.

AArch64 Nested Virtualization

AArch64 guests on KVM can now run VMs with --cpus nested=on. This requires hardware with FEAT_NV2 and a host kernel in VHE mode booted with kvm-arm.mode=nested. Snapshot/restore and live migration are not supported for such guests yet (#8921).

VFIO Peer-to-Peer DMA with dma-buf

Peer-to-peer DMA between VFIO devices now uses VFIO dma-buf with iommufd where the host kernel supports it, and keeps using VFIO_DMA_MAP otherwise (#9141).

CPU Hotplug for SEV-SNP Guests on KVM

SEV-SNP guests on KVM now support CPU hotplug. They now also require ACPI tables supplied through fw_cfg (#9111).

vsock SEQPACKET Support

The vsock device now supports SOCK_SEQPACKET connections in addition to SOCK_STREAM (#9377).

Higher vCPU Limit on AArch64

The vCPU limit on KVM/AArch64 has been raised from 255 to 990 (#9125, #9225).

pvmemcontrol Improvements

The pvmemcontrol device now supports MADV_HUGEPAGE, MADV_NOHUGEPAGE, MADV_DONTNEED_LOCKED (also in batches) and MLOCK_ONFAULT requests from the guest (#9067).

Snapshot/Restore and Live Migration Improvements

  • Postcopy migration now supports vhost-user devices, including the reference vhost_user_block and vhost_user_net backends, and restores ballooned pages as zero pages (#9094, #9391, #9039).
  • Snapshots of VMs restored with memory_restore_mode=copyonwrite only write the pages changed since the restore, on filesystems with reflink support (#9000).
  • A new memory_restore_mode=ondemand_no_background_fetch_pages restores on demand without fetching the remaining pages in the background (#9435).
  • Live migration is faster and no longer hangs on unresponsive peers (#8944, #9204, #9219).
  • The output of the serial port and virtio-console can be changed on restore and receive-migration (#9358).
  • Add CPUID compatibility checks for AMD-specific leaves, which may reject migrations that appeared to work before (#9350, #9364, #9418).

New Command-Line and API Options

  • restore and receive-migration gained an external_fds option, on the command line and in the API, that takes all replacement file descriptors in one list (#8833).
  • --guest-shutdown-action and --guest-reset-action control what the VMM does when the guest shuts down or resets (#9161).
  • --serial and --console accept tcp=<ip>:<port> to listen on a loopback TCP address (#8874, #9367).
  • CPU profiles can be loaded from a file with --cpus profile=path:<file>, and CPUID leaves can be overridden with x-cpuid-overrides, to migrate VMs between hosts with different CPUs. Both are experimental and x86_64 only (#9076, #9078).
  • The rate limiter supports bounded bursts through the new {bw,ops}_burst_rate, {bw,ops}_burst_seconds and {bw,ops}_burst_recharge_seconds options (#9135).
  • --cpus kvm_hint_realtime=on advertises KVM_HINTS_REALTIME to guests whose vCPUs run on dedicated host CPUs (#9156).
  • The D-Bus API gained disk resize and NMI requests (#9319).
  • --landlock no longer requires a VM payload (#9121).
  • vm.counters now reports a device_type string for each device, next to its integer counters (#9434).

Notable Performance Improvements

  • vsock receive buffers are filled in batches, and a starved backend is no longer polled (#9306).
  • VHDX reads run concurrently and scale with the number of queues (#9412).
  • Serial output is faster (#9189).
  • virtio-iommu UNMAP requests no longer scan all mappings of a domain, which speeds up passthrough devices with many mappings (#9431).

Deprecations and Removals

  • Host kernels older than 6.8 are no longer supported. The previous minimum was 5.13 (#9106, #9109).
  • Intel TDX support has been removed, pending a new implementation (#9112).
  • The --platform options serial_number and uuid and the generic vhost-user option virtio_id have been removed (#9096).
  • The ignored file option of --memory has been removed (#9331).
  • Legacy VFIO (container/group) is deprecated in favor of the vfio-cdev mode with iommufd (#9402).
  • ACPI memory hotplug is deprecated in favor of --memory hotplug_method=virtio-mem (#9443).
  • --no-shutdown is deprecated in favor of --guest-shutdown-action=shutdown-vm (#9161).
  • The net_fds, vfio_fds and iommufd_fd options of restore and receive-migration are deprecated in favor of external_fds. Unused file descriptor entries are now rejected (#8833).
  • Restoring a VM with VFIO devices through their saved device paths is no longer supported. restore now requires their file descriptors through external_fds, like receive-migration (#9281).

Security Hardening

  • Trap the PCI configuration space mirror in BAR0 of NVIDIA GPUs instead of mapping it into the guest (#9365).
  • Enforce mapping permissions in virtio-iommu (#9397).
  • Enforce the virtio PCI queue configuration lifecycle (#9099), reject feature writes after FEATURES_OK (#9324), and require VIRTIO_F_ACCESS_PLATFORM to be accepted when offered (#9136).
  • Reject virtio-net offloads that were not negotiated (#9323, #9327), and validate RX and control queue descriptors (#9147, #9155).
  • Validate virtio-block request headers (#9184) and virtio-rtc buffers (#9400).
  • Bound pvmemcontrol hole punching to guest memory (#9117), don’t expose uninitialized bytes (#9123), and check pvpanic access sizes (#9190).
  • Restrict the io_uring operations of the block backends (#9129).
  • Narrow the Landlock rules for read-only disks and the vsock socket directory (#9119).
  • Harden QCOW2 parsing (#9073, #9097, #9172, #9249), and reject encrypted images and writable images with internal snapshots (#9170, #9159).
  • Harden VHDX BAT validation (#9148, #9157, #9185, #9198), and reject images with an active log (#9199).

Notable Bug Fixes

  • Various fixes for NVIDIA Grace-Blackwell (GB) platforms (#9105, #9098, #9114, #9110).
  • Return zeros for punched QCOW2 clusters with a backing file (#9162), and don’t leak preallocated zero clusters (#9192).
  • Fix a stack overflow in the refcount handling of large QCOW2 images (#9441).
  • Fix reads and writes beyond 4 GiB in VHDX images. Data that earlier releases wrote there is not found at the corrected location (#9328).
  • Report I/O errors for rejected virtio-block requests and short completions (#9268, #9255), fail writethrough writes whose fsync fails (#9231), and pad disk serials to 20 bytes (#9164).
  • Further robustness fixes for virtio-block and virtio-net (#9197).
  • Fix stale memory after a migration with vhost-user devices (#9131), fix snapshots on hosts with nested virtualization disabled (#9116), and don’t resume devices on the migration source during teardown (#9221).
  • Fix late guest timers and RCU stall warnings after a restore (#9300, #9273, #9351).
  • Save and restore more device state, such as platform devices, hotplug controllers, and the AArch64 GIC (#9363, #9339, #9415, #9425, #9417, #9318, #9312, #9270, #9305, #9337).
  • Handle devices that the guest already ejected in remove-device and on restore (#9293).
  • Send the host RARP announcement only on the first TAP queue after a restore or migration, avoiding duplicates and a host kernel warning (#9436).
  • Fix vhost-user devices staying stopped after a failed backend state save (#9362).
  • Drop features that a driver no longer accepts on renegotiation, fixing OpenBSD guests with offload_tso=on (#9384).
  • Fix hangs when DRIVER_OK is written over MMCONFIG (#9361) and when a device thread misses a pause request (#9308).
  • Honor the window offset of the virtio PCI configuration access capability (#9088), and return correct virtio-iommu request statuses (#9246).
  • Align PCI BARs to their size (#9389), and fix phantom 64-bit BAR reprogramming (#9087) and byte writes to the MSI-X message control (#9396).
  • Fix device hotplug failing when it races with a guest BAR move, and failed hotplugs blocking a retry with the same ID (#9428, #9419, #9440).
  • Fix vfio-user MSI-X emulation when relocation enlarged the BAR (#9416).
  • Populate the FDT iommu-map for all PCI segments on AArch64 (#9314).
  • Use the QEMU subsystem ID for pvpanic and ivshmem, so the Windows pvpanic driver binds (#9075).
  • Fix 16550 interrupt handling, which stopped the console output of FreeBSD guests (#9187, #9263, #9347).
  • Don’t fail memory hotplug with a disconnected vhost-user backend (#9205).
  • Fix a deadlock between memory resize and device eject (#9201), and validate resize requests against the VM configuration (#9248).
  • Release guest memory after a failed VFIO attachment and on SEV-SNP teardown (#9297).
  • Fix Landlock with VFIO devices (#9405) and with TAP devices passed by name (#9413).
  • Fix a race on the API socket lock file (#9203).
  • Honor the system timezone in local log timestamps (#9257).
  • Don’t warn about an existing TAP device without ip= (#9093).
  • Reject zero vCPUs, zero queues, invalid max_phys_bits and overflowing CPU topologies (#9317, #9330, #9336, #9321).
  • Return errors instead of panicking in several places (#9214, #9215, #9176, #9138, #9140).
  • Fix x86 emulator bugs on MSHV (#9084, #9193, #9207, #9346, #9433, #9445), and validate the IGVM CPUID page (#9158).
  • On KVM, only finalize SVE when available (#9171), handle capability query errors (#9149), and fail early without SEV-SNP capabilities (#9142).
  • OpenAPI specification fixes (#9370, #9160).
  • Documentation improvements and fixes (#9368, #9372, #9373, #9385, #9357, #9383).

Contributors

Many thanks to everyone who has contributed to our release:

Download

See the GitHub Release for the release assets.