What google/gvisor shipped

GVisor 20261102.0 is out. This week: a new runsc/client Go package, CLONE_NEWNS now requires CAP_SYS_ADMIN, FUSE FLUSH ENOSYS treated as success, aio-max-nr and aio-nr in /proc/sys/fs, and TCP_INFO reporting timestamps and SACK.

GVisor 20261102.0 is released, with binaries also available from the apt repository and release bucket. This week's changes include a new runsc/client Go package for driving sandboxes and containers without spawning runsc subprocesses, a CAP_SYS_ADMIN requirement for clone with CLONE_NEWNS, and treating an ENOSYS reply to FUSE FLUSH as success so close no longer fails on such mounts. /proc/sys/fs now exposes aio-max-nr and aio-nr, TCP_INFO reports negotiated timestamps and SACK, and route dumps follow Linux FIB order. Half-closed TCP sockets also stay readable across restore.

GVisor 20261026.0 is out. This week: CLOCK_MONOTONIC no longer counts checkpointed time, memory.oom.group is supported, shim waits are bounded so runsc kill recovers, and FUTEX_WAKE_OP, netlink dump acks and proc/sysfs mount checks now match Linux.

GVisor 20261026.0 is released, with binaries also available from the apt repository and the release bucket. This week also lands behavior fixes for people running real workloads. CLOCK_MONOTONIC no longer counts time spent checkpointed, so timers no longer all expire at once on restore. cgroup2fs adds memory.oom.group, which kubelet sets on cgroup v2. The shim bounds runtime waits and recovers explicit teardown, so a wedged Sentry no longer leaves runsc kill blocked. Several behaviors now match Linux: mount_too_revealing checks for new proc and sysfs mounts, FUTEX_WAKE_OP sign extension and shift masking, no ack for successful netlink dumps, and FUSE device watches that stay usable after a connection abort.

GVisor this week: runsc reads the OCI spec once into a sealed memfd, exec rejects out-of-range user and group IDs, restore reschedules NDP jobs and keeps moved TUN devices, procfs rejects fd/03, and CPUID stops advertising AMX.

This week in gVisor: runsc reads the OCI spec once into a sealed memfd and never re-reads it after create, so edits such as CPU limits are kept. Exec now rejects out-of-range user and group IDs instead of wrapping them to 0, and device paths are resolved before the gofer creates their mount points. Restore reschedules pending NDP jobs at their deadlines and preserves moved TUN devices, and restore readers share the existing CPU budget. Procfs rejects noncanonical descriptor names such as fd/03, netfilter query sizes and permissions are validated more like Linux, TCP loss probes no longer skew RTT samples, and CPUID no longer advertises AMX. Remote fsync errors from the gofer are now propagated, and nvproxy supports /dev/nvidia-uvm-tools.

GVisor 20261012.0 is out. This week: faster delete on restored sandboxes, iptables conntrack match, open_tree and move_mount on files, seccomp profile caching, and fixes for a gofer deadlock and a lisafs overflow.

GVisor 20261012.0 is now available. This week the project enabled pinring for restored sandboxes, cutting delete time; added the iptables conntrack match extension with strict validation; allowed open_tree and move_mount on non-directories; and cached OCI seccomp programs by profile hash. Fixes include a gofer deadlock when HandleDeletion fires inside checkCachingLocked, a uint32 overflow in the lisafs PReadHandler response size check, and zero-sized interior pointer resolution in save.

Accepted TCP endpoints now inherit the listener's congestion control. Reno and CUBIC keep ACK credit straight. EROFS reads no longer stop at the inline tail. systrap survives cgroup freezer cycles. /proc/sys/fs/file-nr works. gVisor 20261005.0 is out.

This week's gVisor changes focus on networking correctness and sandbox resilience. Accepted TCP endpoints now inherit the congestion control algorithm selected on the listener, including SYN-cookie handshakes. Reno consumes ACK credit before growing its congestion window, and CUBIC preserves growth credit so windows grow at the right rate after fast recovery. EROFS reads that span full blocks and tail-packed inline data now return the complete requested range, and oversized exec arguments return E2BIG as on Linux. On systrap, the seccomp notify path retries on ENOENT so freezing and thawing with the cgroup v2 freezer no longer kills the application. procfs also gained /proc/sys/fs/file-nr and file-max, and /proc/[pid]/stat reports PF_EXITING. Release 20261005.0 is available.

GVisor 20260928.0 released: MASQUERADE target for netfilter, SECCOMP_RET_KILL_PROCESS fix, improved cgroup cleanup, and enhanced lock contract enforcement.

GVisor 20260928.0 is now available. This release includes the iptables MASQUERADE target with strict validation, a critical fix for SECCOMP_RET_KILL_PROCESS handling, improved container creation cleanup, and enhanced lock contract checking across package boundaries. These changes improve compatibility with Linux behavior and system reliability.

GVisor 20260921.0 ships with strace thread tracking, iptables comment matches, /proc signal masks, user namespace limits, and 2-4% syscall speedup via lock-free forwarder.

GVisor 20260921.0 improves system call debugging with per-thread strace events, adds netfilter comment match extension and nftables checksum validation, exposes signal masks in /proc/[pid]/status, enforces /proc/sys/user/max_user_namespaces limits, fixes file capabilities under no_new_privs to match Linux, and makes the interrupt forwarder lock-free for 2-4% syscall latency reduction.

GVisor 20260914.0 is out. Better FUSE file support, systemd compatibility, GPU restoration, and seccomp performance improvements.

GVisor 20260914.0 released with major improvements: FUSE now handles files >2GB, seccomp initialization 8ms faster per sandbox, systemd 60+ compatibility via STATX_ATTR_MOUNT_ROOT, GPU device restoration fixes, and socket options up to 1MB. Check the changelog for full details.