GPU VulnDB

Database/Kernel, userspace & hypervisor

Linux kernel SUNRPC client: use-after-free on the upper rpc_clnt when a TLS mount handshake fails fatally

CVSS 9.8CVE-2026-72317Kernel, userspace & hypervisorcurated

Impact

xs_connect() stores task->tk_client in sock_xprt::clnt as a raw pointer and queues the delayed connect worker. For TLS transports that worker is xs_tcp_tls_setup_socket(), which reads cl_timeout, cl_program, cl_prog, cl_vers, cl_cred and cl_stats out of that pointer. Nothing pins the rpc_clnt: the xprt cannot be destroyed until the client releases it, so cancellation in xs_destroy() never runs in time. A fatal TLS failure - the reported case is an mTLS mount whose client certificate does not match the server - wakes the task with -EACCES, the mount caller runs rpc_shutdown_client(), and the still-queued worker then dereferences freed memory, producing a refcount_t underflow. This bites GPU nodes mounting NFS-over-TLS datasets: a certificate rotation gone wrong turns a failed mount into kernel memory corruption on the compute node rather than a clean error.

Who can reach it

Local - triggered on the client node by a mount attempt whose TLS handshake fails fatally. Any user able to initiate such a mount, or an operator whose certificates have expired or been rotated, reaches it; an NFS server that rejects client certificates can induce it on every node that mounts from it.

What to do

Apply the stable-kernel fix, which adds rpc_hold_client() in xs_connect() for TLS transports and releases it in the connect worker's exit path. Kernel update plus reboot on each client node - that means draining GPU jobs node by node across the fleet. Until then, keep client certificates valid and avoid repeated failing mTLS mount attempts. No vendor advisory beyond the kernel stable commits is present in the record.

References

Related entries

All Kernel, userspace & hypervisor entries

This entry is curated: imported from vendor advisories with machine assistance, not yet individually verified. Confirm against your vendor's advisory before acting, and report anything wrong.