GPU VulnDB

Database/Kernel, userspace & hypervisor

Linux kernel nvme-pci: dma_vec allocations leak on every peer-to-peer I/O completion

CVSS 7.5CVE-2026-64020Kernel, userspace & hypervisorcurated

Impact

P2P memory is never unmapped, so the driver did not need to track it, but the dma_vec allocation was still made and never freed on completion. Each P2PDMA NVMe I/O therefore leaks kernel memory. This path is exactly what GPUDirect Storage style setups exercise - NVMe transfers staged directly into device memory - so a node doing sustained dataset streaming leaks continuously rather than occasionally, ending in memory exhaustion and an unusable node. The record gives no evidence of memory disclosure or code execution; the realistic outcome is availability loss on the node.

Who can reach it

Local workload on the node that drives NVMe I/O over peer-to-peer DMA. No special privilege beyond the ability to run storage I/O on a node configured for P2PDMA; nodes without P2PDMA in use are not exposed.

What to do

Pick up the stable kernel carrying the nvme-pci dma_vec fix and reboot each affected node. There is no runtime mitigation other than disabling the P2PDMA path, which on a GPUDirect Storage fleet means giving up the performance the hardware was bought for. Reclaiming the already-leaked memory requires a reboot regardless.

References

Related entries

All Kernel, userspace & hypervisor 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.