[CVE-2025-21892][MODERATE REGULAR] RDMA/mlx5: Fix the recovery flow of the UMR QP [ Upstream

[AL-KERNEL <alkernel@kernelcve.org>]

CVE: CVE-2025-21892
Priority: MODERATE REGULAR
AL-KERNEL base severity: MODERATE
KPANIC flag: NO
Patch: RDMA/mlx5: Fix the recovery flow of the UMR QP [ Upstream
Commit: 3e3bf255992cc02404e9d209b127c1c9944239cf
Upstream patch: https://git.kernel.org/pub/scm/linux/kernel/git/stable/linux.git/commit/?id=3e3bf255992cc02404e9d209b127c1c9944239cf
Original CVE announcement: https://lore.kernel.org/linux-cve-announce/?q=CVE-2025-21892
Analysis date: Mon, 12 Oct 2026 21:54:57 -0400
ActionableScore: 2
ActionableScore lower bound: 1
Actionable bucket: Ordinary Moderate
Manual review required: NO

Summary:
A race in mlx5 UMR QP recovery can discard outstanding completion events during reset and leave RDMA memory-region management tasks blocked indefinitely, causing a local availability failure without an established memory-corruption or privilege-escalation primitive.

======================================================================
ABOUT THIS REPORT
======================================================================

The original Linux kernel CVE announcement for CVE-2025-21892 is available here:
https://lore.kernel.org/linux-cve-announce/?q=CVE-2025-21892

The original announcement does not normally provide a security severity
estimate, CVSS assessment, or enough information to determine whether the
reported kernel bug represents a practically relevant security issue.

This report was generated by AL-KERNEL, an AI-assisted Linux kernel
vulnerability analysis system developed by Alexander Larkin. It combines
an autonomous classifier with LLM-assisted technical analysis and a
separate ActionableScore mechanism.

The purpose of this report is to prioritize Linux kernel CVEs before
manual review, identify cases that require prompt investigation, and
support automatic closure of issues that are unlikely to have meaningful
security impact.

Published priority for this report: MODERATE REGULAR
Manual review required: NO

A detailed explanation of the methodology and priority rules is included
at the end of this message.

======================================================================
AL-KERNEL CLASSIFICATION RESULT
======================================================================

CVE-2025-21892	MODERATE	https://www.kernelcve.org/list/?q=CVE-2025-21892 CHECK WITH IMPACT FROM ORIG NN MODERATE | AEGIS IMPACT MODERATE AEGIS_AS=1 LEGACY=MODERATE RESULT=SAME (legacy and Aegis reconciliation agree; DECISION: both MODERATE)	Maybe valid. Check manually. Hints by AL-KERNEL: The best (paranoid) CVSS is 'AV:L/AC:H/PR:L/UI:N/S:U/C:N/I:N/A:H';CWE-362;CWE-667;CWE-755;Other CVSS 'AV:L/AC:H/PR:L/UI:N/S:U/C:N/I:N/A:L';BEST CVSS score: '4.7';DESCR 'A race in the RDMA mlx5 UMR QP recovery flow can move the QP to RESET before all outstanding work requests have completed. Firmware may then discard flushed completion events, leaving a task blocked indefinitely in wait_for_completion during operations such as memory region deregistration through ib_uverbs. For the CVSS the PR:L is selected because a local process needs access to the RDMA uverbs device and MR operations, while full administrative privileges are not normally required. AC:H applies because reliable triggering depends on a QP error, the recovery path, and a timing window. The issue is not directly network reachable based on the analyzed path. The demonstrated impact is denial of service through blocked RDMA tasks. The paranoid availability impact is High because a shared UMR QP failure may stall multiple RDMA operations or make the RDMA service unavailable until device recovery. No confidentiality or integrity impact is assigned because there is no UAF, out of bounds access, stale pointer, or other memory corruption primitive in the analyzed flow.';YES REQUIRES MANUAL CHECK; ,and ActionableScore result is Ordinary Moderate (with actual score 2)	MAYBE	LOCK DANGER INIT RACE NETWORK LINUS  DECREASED_TO_MODERATE_REGULAR_FROM_MODERATE7_BASED_ON_GUESSCVSS_AND_CIANNH	-	-	checked

======================================================================
ACTIONABLESCORE ANALYSIS
======================================================================

ActionableScore=2
ActionableScoreLower=1
PrimitiveClass=CLASS_C
RealUAF=NO
HostAdminRequired=UNKNOWN
NamespaceScopedPrivilege=UNKNOWN
ImportantCandidate=NO
ManualReview=RECOMMENDED
AutoDowngradeAllowed=YES
RequiresHostUID0=UNKNOWN
AuthorizationDecisionBypass=NO
ProtectedResourceImpact=NONE
B4PendingReview=NO
B4PendingConfirmed=NO

## 1. ActionableScore

* Raw conservative score: 1
* Conservative score: 1
* Raw paranoid score: 2
* Paranoid score: 2
* No floor or cap was applied.
* Final recommended bucket: **Ordinary Moderate**::

## 2. Signal breakdown

Conservative signals:

* +1 Availability impact. Resetting the UMR QP before outstanding WR completion can discard CQEs and leave a kernel task blocked indefinitely in `wait_for_completion()`.
* No negative race penalty. The patch establishes a race, but does not prove that its timing window is narrow, one-shot, or materially uncontrollable.
* No hardware rarity penalty. mlx5 RDMA hardware is specialized, but the evidence does not establish rare or physical-attacker-only exposure.

Additional paranoid signal:

* +1 Broader availability impact. The UMR QP is shared device state, so losing completions may block multiple UMR or MR-management operations and require device recovery. This broader service impact is plausible but not directly demonstrated by the single hung-task trace.

Not awarded:

* No UAF, OOB access, heap overwrite, stale pointer, double release, or refcount mutation is established.
* No controlled reclaim, arbitrary write, disclosure, privilege escalation, or authorization bypass is established.
* No local-unprivileged bonus because access to `ib_uverbs` is visible, but the ability of that actor to force the required UMR QP failure and recovery sequence is unresolved.
* No direct network-reachability bonus. The vulnerable operation is an internal mlx5 UMR recovery path, not direct packet processing.
* NEW52 B5-R does not apply because there is no retry-accumulated heap OOB write or later consumption of corrupted heap state.
* NEW53 B3-PENDING does not apply because no shared-backing write is present.
* NEW54 B4-PENDING does not apply because no DENY-to-ALLOW authorization decision changes.

## 3. Reachability analysis

The observed call path starts from an `ib_uverbs` ioctl used to destroy or deregister an MR, so a local process with access to the RDMA userspace device can reach the surrounding operation. However, the complete trigger also requires the internal UMR QP to enter an error and recovery state while WRs remain outstanding. The supplied evidence does not establish whether an ordinary RDMA user can reliably induce that condition.

Therefore, effective host-administrator privilege is unresolved rather than positively required or positively unnecessary. No namespace-scoped capability path is shown. An mlx5 RDMA device, loaded driver, exposed uverbs interface, and configured RDMA environment are prerequisites. No remote-only trigger is established.

The race is real, but its attacker repeatability is unresolved. No timing penalty is applied because the patch does not establish a narrow or uncontrollable race window.

## 4. Primitive and severity interpretation

The exact defect is an ordering and synchronization failure in UMR QP recovery. The old code could transition the QP to RESET before all outstanding WRs had produced completion events. Firmware was then permitted to discard flushed CQEs, leaving waiters permanently blocked.

This is not stale-storage access or an active structured-object UAF. There is no first lifetime-ending operation followed by a second mutation, and no repeated release or refcount signal applies. The primitive is therefore CLASS_C: lost completion events and invalid recovery ordering causing denial of service.

The demonstrated consequence is a blocked task. Broader RDMA service disruption is plausible because the UMR QP is shared, but host-wide failure is not established. There is no patch-supported confidentiality, integrity, LPE, or memory-corruption consequence.

`ImportantCandidate=NO` because no independently useful corruption or security-boundary primitive exists. `AutoDowngradeAllowed=YES` because the evidence supports a constrained availability defect, although manual confirmation of trigger permissions and device-wide impact remains useful.

## 5. One-sentence report phrase

A race in mlx5 UMR QP recovery can discard outstanding completion events during reset and leave RDMA memory-region management tasks blocked indefinitely, causing a local availability failure without an established memory-corruption or privilege-escalation primitive.

## 6. Manual review recommendation

MANUAL CHECK RECOMMENDED

Confirm whether an ordinary process with uverbs access can reliably force the required UMR QP error state and whether one lost CQE can stall only one task or the shared RDMA service.

## 7. Call-site confidence

Call-site confidence: high

======================================================================
UPSTREAM PATCH SUMMARY
======================================================================

Patch: RDMA/mlx5: Fix the recovery flow of the UMR QP [ Upstream
Commit: 3e3bf255992cc02404e9d209b127c1c9944239cf
Upstream URL: https://git.kernel.org/pub/scm/linux/kernel/git/stable/linux.git/commit/?id=3e3bf255992cc02404e9d209b127c1c9944239cf

Changed files:
  drivers/infiniband/hw/mlx5/umr.c

Diff excerpt:

Not included in this email. See the upstream URL for the full patch.

Full patch:
https://git.kernel.org/pub/scm/linux/kernel/git/stable/linux.git/commit/?id=3e3bf255992cc02404e9d209b127c1c9944239cf

======================================================================
DETAILED REPORT METHODOLOGY
======================================================================

The original Linux kernel CVE announcement for CVE-2025-21892 can be found here:
https://lore.kernel.org/linux-cve-announce/?q=CVE-2025-21892

The original CVE announcement normally does not include a security-level
estimate. In particular, it may not contain a CVSS assessment, an impact
level, or enough information to determine whether the reported bug is a
practically relevant security issue. One purpose of this parallel CVE list
is to provide that missing technical and prioritization information.

The original goal of the AL-KERNEL project was to prioritize Linux kernel
CVE analysis automatically before manual review. The system can also help
identify non-security issues that may be suitable for automatic closure.

This report was generated by AL-KERNEL, an AI-assisted Linux kernel
vulnerability analysis system developed by Alexander Larkin.

The first analysis stage combines an autonomous classifier with additional
LLM-based analysis. The autonomous classifier runs locally on a CPU and is
based on a backpropagation neural network. Together, these mechanisms
produce a technical vulnerability description, identify likely weakness
types, estimate CVSS severity, and provide input for ActionableScore.

Two CVSS estimates are retained because incomplete kernel vulnerability
information often permits more than one defensible interpretation:

  Conservative CVSS vector: AV:L/AC:H/PR:L/UI:N/S:U/C:N/I:N/A:L
  The Best / paranoid CVSS vector: AV:L/AC:H/PR:L/UI:N/S:U/C:N/I:N/A:H
  The Best / paranoid CVSS score: 4.7

The conservative vector represents a lower-impact interpretation.
The Best/paranoid vector intentionally represents a plausible upper-bound
interpretation and should not automatically be treated as demonstrated
real-world impact.

CVSS may also need to be adjusted for a particular Linux deployment,
because actual reachability, privileges, enabled kernel configuration,
hardware, namespaces, exposed device nodes, and other environmental
conditions can differ significantly between systems.

A separate ActionableScore mechanism evaluates practical remediation
urgency. Its analysis may include reachability, attack prerequisites,
subsystem exposure, memory-corruption characteristics, denial-of-service
reliability, and possible confidentiality, integrity, or
privilege-escalation impact.

  Conservative ActionableScore: 1
  Paranoid ActionableScore: 2

The final base severity is taken directly from the second tab-separated
field of the AL-KERNEL classification result. ActionableScore does not
replace or independently override that final AL-KERNEL decision, and
if ActionableScore adjusted impact level of ALKERNEL, then you would see
self-readable flags above like INCREASED_TO_HIGH_BASED_ON_ACTIONABLESCOREHIGHEREQTHAN7.

For an AL-KERNEL result of MODERATE, this report uses the following
additional presentation split:

  ActionableScore below 5   -> MODERATE REGULAR
  ActionableScore 5 or more -> MODERATE 7.0

The distinction between MODERATE REGULAR and MODERATE 7.0 makes it
possible to identify Moderate issues that should receive manual analysis
and fixes before lower-priority MODERATE REGULAR issues. In many cases,
MODERATE REGULAR fixes may wait for a later rebase or routine update.

There is one override in which MODERATE REGULAR becomes MODERATE 7.0
even when the ActionableScore is below 5. When the AL-KERNEL result
contains the KPANIC flag, a MODERATE result is always presented as
MODERATE 7.0. The KPANIC flag selected with few regexps without
usage of AI at all, so it helps to detect cases when Kernel Crash happens
and similar (to filter False-Negative results from the LLM usage).

KPANIC indicates that a reliable kernel crash, kernel panic, or similarly
serious kernel availability impact was identified by the classification
workflow.

AL-KERNEL base severity for this report: MODERATE
KPANIC detected for this report: NO
Published priority for this report (same as in Subject): MODERATE REGULAR

These results are intended to support engineering triage. They are
machine-generated estimates, and cases marked for manual review should
be validated by a human security engineer before final disposition.
For more info read docs linked from here: https://kernelcve.org/
(and you can submit you own patch there to generate such a report
for non-existant CVE-id yet).

Note that in many cases this AI tool selects higher severity, than
real is (means you can expect Importants instead of Moderate 7.0 or
Moderates 7.0 instead of regular Moderates). If you see such cases,
please use reply email interface to add additional manual analyses
info to this particular CVE.
And please, please, let me know when you see Lows instead of Importants
or Important instead of Low (because particular for such cases I
need to tune this AI tool to make it better for this one and next similar).
My contact email for such notifications is alexanjelausa@gmail.com (and both
send reply to CVE record itself too and see "reply" button below for howto reply).