Skip to main content
Vulnerability Database/CVE-2026-98020

CVE-2026-98020: Linux Kernel Race Condition Vulnerability

CVE-2026-98020 is a race condition flaw in the Linux kernel pds_core driver that can cause NULL pointer dereference during concurrent FLR and firmware updates. This article covers technical details, impact analysis, and mitigation strategies.

Published:

Disclaimer: This content was generated using AI. While we strive for accuracy, please verify critical information with official sources.

Sources: National Vulnerability Database, CISA KEV, Exploit-DB

CVE-2026-98020 Overview

CVE-2026-98020 is a race condition in the Linux kernel's pds_core driver for AMD Pensando Distributed Services (PDS) devices. The flaw allows concurrent execution of a Function Level Reset (FLR) and a devlink firmware flash operation to unmap the cmd_regs memory-mapped I/O (MMIO) region while an in-flight device command (devcmd) still references it. The result is a NULL pointer dereference or a write to unmapped MMIO, destabilizing the host kernel.

The affected code paths include pdsc_reset_prepare(), pdsc_reset_done()'s pdsc_map_bars() error handler, and the download loop in pdsc_legacy_firmware_update(), all of which manipulate cmd_regs without properly holding devcmd_lock.

Critical Impact

Concurrent FLR and devlink flash operations can trigger kernel-level NULL dereferences or MMIO writes to unmapped regions, leading to denial of service on hosts running affected pds_core drivers.

Affected Products

  • Linux kernel pds_core driver (AMD Pensando DSC)
  • Upstream Linux kernel versions prior to the fix commits referenced in stable trees
  • Stable kernel branches receiving backports 09f831bf, 2cc69756, 7980325b, and fa31bd14

Discovery Timeline

  • 2026-09-25 - CVE-2026-98020 published to NVD
  • 2026-09-25 - Last updated in NVD database

Technical Details for CVE-2026-98020

Vulnerability Analysis

The pds_core driver manages command submission to PDS devices through an MMIO region mapped via pdsc_map_bars() and referenced through the cmd_regs pointer. Access to the device command interface is serialized by devcmd_lock. The vulnerable code paths violate this synchronization contract.

When pdsc_reset_prepare() executes or when pdsc_reset_done() hits the pdsc_map_bars() error path, the driver clears and iounmaps cmd_regs without first acquiring devcmd_lock. Separately, pdsc_legacy_firmware_update() drops devcmd_lock during its download loop and re-acquires it without re-validating that cmd_regs still points to a mapped BAR.

An FLR running concurrently with a devlink dev flash operation can therefore tear down the BAR while a firmware download is mid-flight. The in-flight devcmd then dereferences a NULL or stale pointer, or writes into MMIO that is no longer mapped.

Root Cause

The root cause is a missing-lock race condition [CWE-362] between BAR unmap/remap on reset and in-flight device command processing. The code assumes cmd_regs remains stable across the firmware download loop's lock drop, but reset paths can invalidate the pointer at any time without holding devcmd_lock.

Attack Vector

Exploitation requires a local privileged actor with the ability to trigger both a Function Level Reset and a devlink firmware flash operation on the same PF concurrently. The attacker does not need to craft payloads; the race itself produces the kernel fault. VFs do not map cmd_regs or issue devcmds, so only PF-exposed configurations are impacted.

The fix takes devcmd_lock across the BAR unmap and remap sequence, re-checks cmd_regs after each lock re-acquisition in the download loop, and skips the unmap entirely on VFs. Resets that complete inside the unlocked window remain safe because the device clears its update session and rejects resumed downloads, returning PDS_RC_BAD_FW instead of activating a corrupted image.

See the upstream commits for the authoritative patch logic:

Detection Methods for CVE-2026-98020

Indicators of Compromise

  • Kernel oops or panic traces referencing pdsc_devcmd, pdsc_legacy_firmware_update, or pdsc_map_bars on the call stack.
  • BUG: unable to handle NULL pointer dereference or MMIO fault messages logged by the pds_core module during or shortly after a reset event.
  • Devlink flash operations that fail with PDS_RC_BAD_FW immediately following an FLR on the same PF.

Detection Strategies

  • Monitor dmesg and journalctl -k for pds_core warnings or faults that correlate with devlink dev flash commands.
  • Correlate PCIe FLR events for Pensando DSC devices with concurrent firmware update activity to surface the vulnerable race window.
  • Audit kernel versions on hosts using AMD Pensando NICs to identify systems missing the referenced stable backports.

Monitoring Recommendations

  • Collect kernel ring buffer logs centrally and alert on pds_core stack traces or MMIO access faults.
  • Track devlink and PCIe reset administrative activity to detect operational patterns that could trigger the race.
  • Inventory running kernel versions against the fixed stable branches and flag drift.

How to Mitigate CVE-2026-98020

Immediate Actions Required

  • Apply the stable kernel update containing commits 09f831bf, 2cc69756, 7980325b, or fa31bd14 as appropriate for your kernel branch.
  • Avoid issuing devlink dev flash on pds_core PFs while any reset-inducing operation is in progress.
  • Restrict administrative access to devlink and PCIe reset interfaces to trusted operators.

Patch Information

The fix is upstream in the Linux kernel and backported to stable trees via the four referenced commits. The patch takes devcmd_lock across the BAR unmap and remap, re-validates cmd_regs after lock reacquisition in the firmware download loop, and skips the unmap on VFs. Rebuild or install a distribution kernel that includes these backports.

Workarounds

  • Serialize firmware flashing and reset workflows at the orchestration layer so they cannot overlap on the same PF.
  • Temporarily disable automated FLR triggers (for example from management agents) during planned firmware updates.
  • Where feasible, perform firmware updates during maintenance windows with the device otherwise quiesced.
bash
# Verify the running kernel includes one of the fix commits
uname -r
zcat /proc/config.gz 2>/dev/null | grep CONFIG_PDS_CORE
dmesg | grep -i pds_core

Disclaimer: This content was generated using AI. While we strive for accuracy, please verify critical information with official sources.

Sources: National Vulnerability Database, CISA KEV, Exploit-DB

Experience the Most Advanced Cybersecurity Platform

See how the world’s most intelligent, autonomous cybersecurity platform can protect your organization today and into the future.