Linux kernel CephFS client: reclaim during MDS reply handling crashes the kernel via ext4 journal_info

Impact

handle_reply() parks a ceph_mds_request pointer in current->journal_info while populating the inode and dentry cache. An allocation in that window can enter direct reclaim, prune dentries belonging to another filesystem, and dirty an ext4 inode; ext4 then starts a JBD2 transaction, JBD2 reads current->journal_info as a journal handle, and dereferences the Ceph request's r_tid as h_transaction. The record includes the resulting oops in jbd2__journal_start() from the ceph-msgr workqueue. On a GPU node this is an unprivileged-triggerable kernel crash on any host that mounts CephFS alongside an ext4 filesystem and runs under memory pressure - which describes a busy training node with a large page cache - and it takes the whole node and its jobs down.

Who can reach it

Local, no special privilege: any workload doing CephFS metadata activity on a node that also has ext4 mounted, while the node is under enough memory pressure to enter direct reclaim. Not remotely reachable on its own. Nodes without a kernel CephFS mount are unaffected.

What to do

Apply the stable fix, which enters a scoped NOFS allocation context around the region that uses current->journal_info and leaves it after clearing the field. Five stable commits are listed; the record names no fixed release version. Rollout is a patched kernel and a per-node reboot - drain GPU workloads first. There is no configuration mitigation short of not mounting CephFS on the affected hosts.

References

Related entries

• Linux kernel libceph: out-of-bounds read in decode_watchers() from a zero-length struct_lenCVE-2026-80557 · Linux kernel libceph (decode_watchers(), CEPH_OSD_OP_LIST_WATCHERS reply parsing)Unscored
• Linux kernel libceph: unvalidated primary_temp OSD index causes out-of-bounds array readsCVE-2026-80558 · Linux kernel libceph (get_temp_osds(), unvalidated primary_temp OSD index)Unscored
• Linux kernel libceph: unchecked decodes in decode_locker() let a malicious OSD read out of boundsCVE-2026-80561 · Linux kernel libceph (cls_lock_client decode_locker)Unscored
• Microsoft Hyper-V: vmswitch fails to validate guest OID requestsCVE-2021-28476 · Microsoft Hyper-VCritical
• Incus: instance snapshots bypass restricted.containers.lowlevel, giving command execution on the hostCVE-2026-48751 · Incus (instance snapshots ignore restricted.containers.lowlevel)Critical
• VMware ESXi (OpenSLP): Use-after-free in OpenSLP on port 427 - unauthenticated remote code execution on the hypervisorCVE-2020-3992 · VMware ESXi (OpenSLP)Critical