CVE-2026-99212

*** Pendiente de traducción *** In the Linux kernel, the following vulnerability has been resolved:<br /> <br /> net/rds: don&amp;#39;t let the rdma_cm destroy an id RDS still owns on route failure<br /> <br /> rds_rdma_cm_event_handler_cmn() hands the return value of<br /> rdma_resolve_route() straight back to the rdma_cm on<br /> RDMA_CM_EVENT_ADDR_RESOLVED. A synchronous failure there - -ENOMEM<br /> from the route work allocation, an SA query that cannot be set up, a<br /> RoCE route with no usable device - makes addr_handler() destroy the<br /> id the event was delivered on. That id is ic-&gt;i_cm_id, and nothing<br /> clears the pointer: the connection sits in RDS_CONN_CONNECTING with a<br /> freed id until its shutdown calls rdma_disconnect() and<br /> rdma_destroy_id() on it.<br /> <br /> The ROUTE_RESOLVED case already avoids this: rds_ib_cm_initiate_connect()<br /> forces a zero return while ic-&gt;i_cm_id == cm_id, as the comment there<br /> explains. Do the same here - drop the connection and return 0 - so<br /> the id stays RDS&amp;#39;s to destroy from the shutdown, and the reconnect<br /> gets a fresh one.