GPU VulnDB

Database/Firmware, BMC & network fabric

Linux net/rds: use-after-free of the RDMA CM id when rdma_resolve_route() fails synchronously

CVSS 7.0CVE-2026-99212Firmware, BMC & network fabriccurated

Impact

rds_rdma_cm_event_handler_cmn() returns the rdma_resolve_route() error straight to the rdma_cm on ADDR_RESOLVED, which makes addr_handler() destroy the very id the event arrived on. Nothing clears ic->i_cm_id, so the RDS connection sits in CONNECTING holding a freed id until shutdown calls rdma_disconnect() and rdma_destroy_id() on it - a double free and use-after-free in the RDMA path. The triggers named are ordinary transient conditions on a fabric: an -ENOMEM allocating route work, an SA query that cannot be set up, or a RoCE route with no usable device, so a congested or partially misconfigured InfiniBand/RoCE fabric can reach it without an attacker. Affects only nodes that actually load and use RDS over RDMA; a GPU node doing RDMA with verbs or NCCL alone is not exposed.

Who can reach it

Local, requires RDS-over-RDMA in use on the node. Reached through connection setup failures on the IB/RoCE fabric rather than by an authenticated request; anyone able to make route resolution fail (fabric pressure, SA unavailability) can drive it.

What to do

Take the stable kernel that drops the connection and returns 0 so RDS keeps ownership of the id, then reboot the node. If RDS is not needed, blacklisting the rds_rdma/rds modules removes the exposure without a kernel change. No distribution fixed version appears in the record; only the kernel.org commits.

References

Related entries

All Firmware, BMC & network fabric 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.