Linux kernel InfiniBand uverbs (ib_uverbs / ib_umem_get, drivers/infiniband/core/umem.c): The canonical RDMA isolation

Impact

The canonical RDMA isolation break. Any process that can open /dev/infiniband/uverbsN registers a memory region whose start+length overflows the page-aligned end computation in ib_umem_get(). The HCA then holds a valid rkey/lkey for physical memory the caller never owned, and the tenant reads and writes it at line rate via ordinary RDMA verbs - other tenants' pages, the page cache, and kernel text all included. Because the DMA is issued by the HCA rather than the CPU, it bypasses page tables entirely: no MMU check, no KASLR, no SMAP/SMEP. On a shared GPU node where the uverbs device is passed into tenant containers so jobs can use NCCL/UCX, this is a full read/write primitive over host memory from inside an unprivileged container.

Who can reach it

Local, unprivileged - requires only read/write access to an /dev/infiniband/uverbsN character device. In practice every RDMA-enabled tenant container has exactly that, because the device node must be exposed for NCCL, UCX, MPI or GPUDirect to work. No fabric access and no root needed.

What to do

Patch the kernel to include commit 8494057ab5e40df590ef6ef7d66324d3ae33356b (IB/uverbs: prevent integer overflow in ib_umem_get address arithmetic) plus the follow-up 66578b0b2f69659f00b6169e6fe7377c4b100d18 that re-permits legitimate registrations starting at 0x0; RHEL 6 kernel-2.6.32-504.12.2 and later carry it. Not live-patchable in practice - the change is in the memory-region registration path, so the fleet needs a rolling kernel upgrade and reboot, one node drained at a time. Where a reboot cannot be scheduled, the only real mitigation is to stop exposing /dev/infiniband/uverbs* to untrusted workloads, which means turning off RDMA for those tenants entirely.

References