CVE-2026-89708

In the Linux kernel, the following vulnerability has been resolved:<br /> <br /> nfsd: RCU-protect cl_cb_session to fix use-after-free on session teardown<br /> <br /> After a DESTROY_SESSION the per-session teardown path can free a<br /> session while rpciod still holds an inflight callback rpc_task that<br /> dereferences clp-&gt;cl_cb_session. nfsd4_probe_callback_sync() flushes<br /> cl_callback_wq, but once nfsd4_run_cb_work() has called<br /> rpc_call_async() the rpc_task lives on rpciod; flushing the workqueue<br /> does not wait for it. rpc_shutdown_client() does drain rpciod tasks,<br /> but uses a 1-second wait_event_timeout — tasks stuck in rpc_delay()<br /> (e.g. 2-second NFS4ERR_DELAY retries) can outlive the drain.<br /> <br /> destroy path rpciod<br /> ------------ ------<br /> unhash_session(ses)<br /> nfsd4_probe_callback_sync(clp)<br /> flush_workqueue(cl_callback_wq)<br /> /* returns; rpc_task still live */<br /> nfsd4_put_session_locked(ses)<br /> free_session(ses) -&gt; kfree(ses)<br /> nfsd4_cb_sequence_done()<br /> reads cb_clp-&gt;cl_cb_session<br /> /* freed slab */<br /> <br /> A second window exists in nfsd4_process_cb_update(). When<br /> __nfsd4_find_backchannel() returns NULL because unhash_session() has<br /> already removed the destroyed session from cl_sessions,<br /> setup_callback_client() takes the v4.1 early return so<br /> clp-&gt;cl_cb_session = ses never fires and the field retains a pointer<br /> to the about-to-be-freed session.<br /> <br /> Fix both by converting cl_cb_session to an RCU-protected pointer:<br /> <br /> - Move the cl_cb_session = ses assignment in setup_callback_client()<br /> to after rpc_create() succeeds, so it is only published when a<br /> working backchannel exists. Clear cl_cb_session on the error<br /> return in nfsd4_process_cb_update(). Both stores use<br /> rcu_assign_pointer().<br /> <br /> - Annotate cl_cb_session with __rcu. All rpciod-side readers use<br /> rcu_read_lock()/rcu_dereference() and check for NULL, bailing to<br /> the appropriate error or requeue path:<br /> encode_cb_sequence4args(), decode_cb_sequence4resok(),<br /> nfsd41_cb_get_slot(), nfsd41_cb_release_slot(),<br /> nfsd4_cb_prepare(), and nfsd4_cb_sequence_done().<br /> <br /> - Switch __free_session() from kfree() to kfree_rcu() so the<br /> session slab is not reclaimed until after an RCU grace period,<br /> guaranteeing that rpciod readers inside rcu_read_lock() never<br /> dereference freed memory.<br /> <br /> - Pass the session pointer to the nfsd_cb_seq_status and<br /> nfsd_cb_free_slot tracepoints instead of having them re-read<br /> cl_cb_session.<br /> <br /> - nfsd4_cb_prepare() calls rpc_exit() when the session is NULL,<br /> routing through the done/release path to requeue the callback.