firecracker-microvm/firecracker v1.16.2

Firecracker v1.16.2

on GitHub

Changed

• #6172: Gate CLIRD_EL1 override to only happen on host kernels equal or newer than 6.10 release. This aliviates the guest performance regression seen on host kernels in range [6.3..6.10) correlated to incorrect cpu cache topology presented to the guest.
• #6100: Bumped the snapshot format version. Snapshots created by this version cannot be loaded by 1.16.0 or 1.16.1, and snapshots created by those versions cannot be loaded by this one.

Fixed

• #6100: Fixed the vsock device permanently suppressing RX (host-to-guest) delivery after a bare pause/resume cycle (PATCH /vm with Paused then Resumed, without a snapshot). The resume kick armed the TRANSPORT_RESET RX gate even though no reset event had been sent, so the guest could never acknowledge it and every new host-initiated connection made after the resume hung forever. The gate is now part of the persisted device state and the resume kick respects it instead of arming it.
• #6174: Fixed virtio-mem leaving unplugged memory writable by the VMM, and potentially unmapped, on microVMs restored from a snapshot memory file. Firecracker now maps each slot straight to the protection it should have, with the MAP_NORESERVE flag, and aborts if the re-map fails.
• #6174: Fixed virtio-mem discarding the whole hotpluggable region on every UNPLUG_ALL request, even when nothing was plugged. The discard is now skipped when the range has no plugged blocks.
• #6176: Fixed virtio-mem leaving its block accounting inconsistent with the KVM memory slots if a plug or unplug request failed part way through. Firecracker now commits each slot's state only after its KVM update succeeds, so a partial failure leaves the block state and the KVM slots reflecting exactly the slots that were updated.
• #5956: Fixed a TOCTOU race in the aarch64 jailer when setting ownership of the CPU cache and MIDR_EL1 information files copied into the chroot.
• #6076: Fixed memory hotplug sizes silently wrapping when converted from MiB to bytes. requested_size_mib on PATCH /hotplug/memory and total_size_mib, block_size_mib and slot_size_mib on PUT /hotplug/memory are now bounded to 32 bits. Values above that previously wrapped to a smaller byte count, so a large enough request was accepted as a 0-byte region instead of being rejected.
• #6031, #6041, #6077: Fixed the vsock device re-arming its host-stream EPOLLIN interest while received data was still awaiting a guest RX buffer, which could busy-spin the event thread until the guest posted buffers. The device now drains the host stream fully across successive RX operations instead of one packet per event loop iteration, reducing the host-side CPU cost per gigabit of host-to-guest traffic by up to ~50% and improving host-to-guest throughput by ~44% (median) in most configurations. On single-vcpu microVMs on the newest Intel hosts (m7i, m8i), host-to-guest throughput may instead decrease by ~15% (median), where the single guest vCPU rather than the host becomes the bottleneck. Terminating a connection now also discards its TX buffer, so the device stops advertising EPOLLOUT for a host stream it will never write to again, which could otherwise busy-spin the event thread indefinitely.
• #6086, #6143: Fixed a deadlock in the logger: a signal handler that logs while the interrupted thread is already logging would hang the VMM and its API socket. This is reachable whenever a SIGPIPE is raised by Firecracker's own log write, for example when logging to stdout and the reader of that pipe exits. The logger now uses an RwLock, and sigpipe_handler only increments the signals.sigpipe metric instead of logging, so it no longer emits a Received signal 13, code 0. line.
• #6120: Fixed a bug caused by a KVM behavior change introduced in Linux 6.13, which requires guest CPUID to be set before userspace reads CPUID-dependent MSRs. On x86_64 with Linux 6.18 host kernels, Firecracker read the base values of those MSRs before setting guest CPUID, so KVM returned zero for such MSRs and CPU templates consequently used zero for bits configured for passthrough. Firecracker now sets guest CPUID before reading MSRs to be modified by CPU templates, so passthrough bits retain their KVM-provided values and features such as eIBRS remain exposed to the guest.