CVE-2025-37861

Quoted from the kernel CNA's own scoring file, word for word. A highlighted metric carries the most severe value CVSS v3.1 defines for it, which is what pushes the score up. source

AV:LAttack vector, local

The vulnerable code is a Broadcom MPI3 SAS/NVMe HBA driver path reached only through local I/O submission, SCSI error handling, or the /dev/bsg management node; there is no remote or adjacent-network reachability.

AC:LAttack complexity, low

The attacker controls both sides of the race — sustained I/O to a device on the controller keeps threads inside mpi3mr_process_op_reply_q() (threaded IRQ poll, blk-mq poll, or SCSI-EH task-management polling), while the same load storm drives the timeout/EH escalation and reset paths that call mpi3mr_memset_buffers(); the window can be retried indefinitely with no attacker-uncontrollable precondition.

PR:LPrivileges required, low

An unprivileged local user with access to any block device backed by the mpi3mr HBA can drive the polling side and induce the timeout-driven error-recovery side; the higher-privileged BSG reset path (CAP_SYS_RAWIO) only makes triggering more deterministic and is not required.

UI:NUser interaction, none

The race is triggered entirely by the attacker's own I/O and error-recovery activity; no action by another user or administrator is needed.

S:UScope, unchanged

The corruption is confined to the kernel's own memory and MMIO mappings within the same security authority; no VM, IOMMU, or sandbox boundary is crossed.

C:HConfidentiality impact, high

The zeroed num_replies makes reply_ci grow without wrapping, so mpi3mr_get_reply_desc() reads segment pointers past the q_segments array and dereferences them, giving unbounded out-of-bounds kernel reads whose contents flow into reply/sense processing and can reach userspace.

I:HIntegrity impact, high

With qid/request_queue_id zeroed, WRITE_ONCE(mrioc->req_qinfo[0xFFFF].ci, ...) performs an out-of-bounds write roughly 4.5 MB beyond the allocation into unrelated kernel heap, and writel() targets addresses ~512 KB outside the ioremap()ed BAR — memory corruption usable to build a write primitive.

A:HAvailability impact, high

The commit explicitly states the invalid 0xFFFF queue ID points to unallocated memory and causes a crash; the writel() to oper_queue_indexes[0xFFFF] hits an unmapped kernel virtual address, producing a kernel oops and loss of the storage controller.