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-2026-93224 Overview
CVE-2026-93224 is a Linux kernel vulnerability in the svcrdma (Server-Side Remote Direct Memory Access) component of the SUNRPC subsystem. The flaw occurs when svc_rdma_accept() takes the error path before rpcrdma_rn_register() has succeeded. The cleanup block unconditionally calls rpcrdma_rn_unregister(), causing an unmatched kref_put() on another caller's device reference and corrupting a shared xarray slot. The defect also leaks the svcxprt_rdma structure and its pinned network namespace tracker on every failed accept.
Critical Impact
Attackers who can trigger repeated RDMA accept failures may corrupt kernel reference counts and leak kernel memory, enabling denial of service or follow-on memory corruption attacks against NFS/RDMA servers.
Affected Products
- Linux kernel versions containing the svcrdma transport with rpcrdma_rn_register() logic
- Systems using NFS over RDMA server functionality
- Distributions shipping stable kernels prior to the fix commits
Discovery Timeline
- 2026-09-24 - CVE-2026-93224 published to NVD
- 2026-09-25 - Last updated in NVD database
Technical Details for CVE-2026-93224
Vulnerability Analysis
The vulnerability resides in the server-side RDMA transport layer used by the Linux kernel SUNRPC subsystem. When svc_rdma_accept() fails and jumps to its errout label before rpcrdma_rn_register() has completed, the cleanup code assumes registration succeeded. Because svcxprt_rdma is allocated via kzalloc, both sc_rn.rn_index and sc_rn.rn_done remain zero or NULL. The cleanup then calls rpcrdma_rn_unregister(dev, &newxprt->sc_rn) which performs an xa_erase() against slot 0, destroying registration state belonging to a different, unrelated caller.
The same errout path also brackets cleanup with svc_xprt_get() and svc_xprt_put() around the kref_init() birth reference. The resulting kref sequence transitions 1 → 2 → 1 and never reaches zero. The transport object and the network namespace tracker it holds are leaked on every failed accept, producing cumulative kernel resource exhaustion.
Root Cause
The root cause is a missing "never registered" sentinel check in rpcrdma_rn_unregister(). The register routine writes rn->rn_done last, only after both xa_alloc() and kref_get() succeed, making a NULL rn_done a reliable indicator that registration never occurred. The original unregister path lacked that guard, so a failed accept would execute full cleanup against uninitialized state. This is a reference count mismanagement issue combined with cross-caller xarray slot corruption.
Attack Vector
An attacker with network access to an RDMA-enabled NFS server can repeatedly initiate RDMA connections that fail during the accept phase. Each failed accept leaks a svcxprt_rdma structure and the pinned net/ns_tracker, while also decrementing another connection's device reference counter. Over time this can crash the kernel through use-after-free on the rpcrdma_device or exhaust memory. The fix adds an early return in rpcrdma_rn_unregister() when rn_done is NULL, clears rn_done before the matching xa_erase(), and takes a replacement module reference so module_put() calls remain balanced.
See the upstream commits referenced in the patch information section for the full fix.
Detection Methods for CVE-2026-93224
Indicators of Compromise
- Unexplained growth in kernel memory consumption on NFS/RDMA server hosts, particularly in kmalloc slabs holding svcxprt_rdma allocations
- Repeated failed RDMA accept events in kernel logs, often paired with device probe or memory region allocation errors
- Kernel warnings or oops messages referencing rpcrdma_rn_unregister, xa_erase, or kref_put on the rpcrdma_device path
- Rising counts of network namespace trackers that are never released after client disconnects
Detection Strategies
- Monitor /proc/slabinfo for anomalous growth in SUNRPC-related slab caches over time
- Enable kernel KASAN or KFENCE in test environments to catch cross-caller xarray slot corruption triggered by failed accepts
- Correlate RDMA connection failure spikes with host memory pressure using telemetry from endpoint and workload sensors
- Alert on repeated svc_rdma error paths captured through dmesg or journalctl -k ingestion pipelines
Monitoring Recommendations
- Forward kernel ring buffer and rpc_debug output to a centralized log platform for correlation across NFS/RDMA fleet nodes
- Track per-process and per-namespace reference count metrics using bpftrace or perf probes on kref_get and kref_put
- Baseline normal RDMA accept failure rates and alert when the rate deviates significantly from baseline
How to Mitigate CVE-2026-93224
Immediate Actions Required
- Inventory all hosts running NFS over RDMA server workloads and identify kernels missing the fix commits
- Apply vendor-provided kernel updates containing the rpcrdma_rn_unregister() NULL-sentinel guard
- Restrict network access to RDMA listener ports to trusted client subnets using host firewalls or network ACLs
- Reboot affected hosts after patching to ensure the kernel module is reloaded with the fixed code
Patch Information
The fix is available in the following upstream Linux kernel commits:
- Linux Kernel Commit 0335800
- Linux Kernel Commit 2619039
- Linux Kernel Commit 45a444a
- Linux Kernel Commit 5aabe0
The patch guards rpcrdma_rn_unregister() with an early return when rn_done is NULL, clears rn_done before the matching xa_erase(), and acquires a replacement module reference so module_put() calls remain balanced across the error path.
Workarounds
- Disable the svcrdma kernel module on NFS servers that do not require RDMA transport until patches are applied
- Block access to RDMA transport ports from untrusted networks using host-based firewall rules
- Limit NFS/RDMA exports to trusted client IP ranges to reduce the attack surface for triggering failed accepts
# Temporarily disable svcrdma module if RDMA transport is not required
sudo rmmod svcrdma
echo "blacklist svcrdma" | sudo tee /etc/modprobe.d/blacklist-svcrdma.conf
# Verify the module is not loaded
lsmod | grep svcrdma
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

