CVE Database

In the Linux kernel, the following vulnerability has been resolved: nexthop: take nh->lock for f6i_list walks in replace check and notify fib6_check_nh_list() and __nexthop_replace_notify() walk nh->f6i_list during an RTNL-serialized nexthop replace without holding nh->lock. IPv6 RTM_NEWROUTE/RTM_DELROUTE run without RTNL and mutate that list under nh->lock (fib6_add_rt2node_nh(), fib6_purge_rt()), so both walks race a concurrent route delete that unlinks and frees a fib6_info: BUG: KASAN: slab-use-after-free in rt6_fill_node.isra.0 (net/ipv6/route.c:5799) Read of size 4 at addr ffff888014607e64 by task exploit/143 rt6_fill_node.isra.0 (net/ipv6/route.c:5799) fib6_rt_update (net/ipv6/route.c:6412) __nexthop_replace_notify (net/ipv4/nexthop.c:2542) rtm_new_nexthop (net/ipv4/nexthop.c:2554) rtnetlink_rcv_msg (net/core/rtnetlink.c:7076) BUG: KASAN: slab-use-after-free in fib6_check_nh_list (net/ipv4/nexthop.c:1605) Read of size 8 at addr ffff888014a7d068 by task exploit/142 fib6_check_nh_list (net/ipv4/nexthop.c:1605) rtm_new_nexthop (net/ipv4/nexthop.c:2575) rtnetlink_rcv_msg (net/core/rtnetlink.c:7076) Both walks only read the entries and take no tb6_lock, so protect them with nh->lock; fib6_rt_update() uses gfp_any(), which returns GFP_ATOMIC under the lock.

In the Linux kernel, the following vulnerability has been resolved: nexthop: avoid unlocked f6i_list walk in nh_rt_cache_flush nh_rt_cache_flush() walks nh->f6i_list during an RTNL-serialized nexthop replace without holding nh->lock, racing the unlocked IPv6 route add/delete that mutate the list under nh->lock and free fib6_info entries (nh_rt_cache_flush() is inlined into rtm_new_nexthop()): BUG: KASAN: slab-use-after-free in nh_rt_cache_flush (net/ipv4/nexthop.c:2243) Read of size 8 at addr ffff888012953e18 by task exploit/146 nh_rt_cache_flush (net/ipv4/nexthop.c:2243) replace_nexthop (net/ipv4/nexthop.c:2610) rtm_new_nexthop (net/ipv4/nexthop.c:3323) rtnetlink_rcv_msg (net/core/rtnetlink.c:7076) Unlike the other f6i_list walks, this one bumps each route's sernum via fib6_update_sernum_upto_root(), which needs tb6_lock; taking nh->lock around it would invert the established tb6_lock -> nh->lock order and deadlock. As the only purpose is to invalidate cached dsts, bump the IPv6 sernum for the whole netns with rt_genid_bump_ipv6() instead, mirroring the rt_cache_flush() already done for IPv4 just above.

In the Linux kernel, the following vulnerability has been resolved: scsi: libiscsi: Fix stale-data leak into the SCSI sense buffer iscsi_scsi_cmd_rsp() copies the sense data of a SCSI Response from the target-supplied data segment. The segment carries a 2-byte sense length followed by the sense bytes, so it must hold 2 + senselen bytes, but the bounds check only requires datalen >= senselen: senselen = get_unaligned_be16(data); if (datalen < senselen) goto invalid_datalen; memcpy(sc->sense_buffer, data + 2, min_t(uint16_t, senselen, SCSI_SENSE_BUFFERSIZE)); A target that returns a SCSI Response whose datalen equals senselen (with senselen <= SCSI_SENSE_BUFFERSIZE) makes the memcpy() from data + 2 read up to two bytes past the received data. Those bytes are stale conn->data contents and end up in the command's sense buffer, which is returned to userspace. Account for the 2-byte sense length prefix in the check.

In the Linux kernel, the following vulnerability has been resolved: wifi: ath12k: fix out-of-bounds clear_bit in ath12k_mac_dp_peer_cleanup() ath12k_mac_dp_peer_cleanup() clears the ML peer ID slot on the free_ml_peer_id_map bitmap by indexing it with dp_peer->peer_id. That is wrong: dp_peer->peer_id for an MLO peer always carries the ATH12K_PEER_ML_ID_VALID bit (BIT(13)), so clear_bit() is invoked with index >= 0x2000, which is far outside the bitmap of ATH12K_MAX_MLO_PEERS (256) bits and corrupts memory adjacent to ah->free_ml_peer_id_map. The intended bitmap entry also never gets cleared, so subsequent ath12k_peer_ml_alloc() calls eventually run out of IDs. The ID without the VALID bit is what ath12k_peer_ml_alloc() returned and is stored in ahsta->ml_peer_id. Use that instead. While there, also reset ahsta->ml_peer_id to ATH12K_MLO_PEER_ID_INVALID so the bitmap and ahsta->ml_peer_id stay in sync. Tested-on: WCN7850 hw2.0 PCI WLAN.HMT.1.1.c5-00302-QCAHMTSWPL_V1.0_V2.0_SILICONZ-1.115823.3

In the Linux kernel, the following vulnerability has been resolved: hwmon: (nzxt-smart2) DMA-align output buffer Sashiko reports: When send_output_report() calls hid_hw_output_report(), the underlying USB HID core calls usb_interrupt_msg() which maps this buffer directly for DMA. When the DMA mapping flushes or invalidates the cacheline, it will corrupt the adjacent variables (mutex, update_interval) that were modified concurrently by the CPU. This causes memory corruption due to cacheline sharing on non-coherent CPU architectures (such as ARM or MIPS). The DMA API debugging tool (CONFIG_DMA_API_DEBUG) will trigger runtime warnings for this violation. Any operation that triggers send_output_report() (like setting a fan speed or updating the interval) causes the USB DMA mapping. On systems with non-coherent caches, this structural bug causes immediate and deterministic memory corruption. Align the output buffer to ARCH_DMA_MINALIGN to fix the problem.

In the Linux kernel, the following vulnerability has been resolved: net: do not send ICMP/NDISC Redirects when peer allocation fails When inet_getpeer_v4() or inet_getpeer_v6() fails to allocate a peer entry under memory pressure or tree size caps, redirect handlers previously fell back to sending un-rate-limited ICMP/NDISC Redirect messages. In IPv4, ip_rt_send_redirect() called icmp_send() directly when peer == NULL. In IPv6, ip6_forward() and ndisc_send_redirect() passed a NULL peer into inet_peer_xrlim_allow(), which returned true when peer == NULL. Because ICMP/NDISC Redirects are not part of the default global rate limit mask (sysctl_icmp_ratemask), sending redirects when peer == NULL creates an un-rate-limited ICMP packet storm. Fix this by failing closed in ip_rt_send_redirect(), ip6_forward(), and ndisc_send_redirect() when peer is NULL.

In the Linux kernel, the following vulnerability has been resolved: hwmon: (nct6775-core) Prevent access to unsupported weight registers Sashiko reports: During initialization of the nct6116 chip, the driver sets data->pwm_num to 5. However, it assigns several NCT6106 register arrays (such as NCT6106_REG_WEIGHT_DUTY_STEP, NCT6106_REG_WEIGHT_TEMP_SEL, and NCT6106_REG_WEIGHT_TEMP_*) to data->REG_PWM and data->REG_WEIGHT_TEMP. These arrays only contain 3 elements. In nct6775_update_pwm(), the driver iterates up to data->pwm_num. If data->has_pwm has bits 3 or 4 set (which is structurally possible for nct6116), the loop attempts to read elements at index 3 and 4 from these 3-element arrays. This results in a global out-of-bounds read, which can be caught by KASAN. Furthermore, the driver uses these garbage out-of-bounds values as hardware register addresses for subsequent read and write operations. This leads to invalid hardware register access, potentially causing hardware misconfiguration or system crashes. The underlying problem is that the chip does support up to five fan control channels, but only the first three support weight control. Fix the problem by extending the affected weight register arrays with zeroed fields. The driver uses zeroed register addresses to determine if a register is supported or not, and skips accesses for unsupported registers.

In the Linux kernel, the following vulnerability has been resolved: forcedeth: fix UAF of txrx_stats in nv_remove nv_remove() frees the per-CPU txrx_stats before unregister_netdev(). Until unregister completes, ndo_get_stats64, the NAPI/xmit data path, and nv_close()/drain may still access txrx_stats, leading to a use-after-free. Free the stats only after unregister_netdev().

In the Linux kernel, the following vulnerability has been resolved: net/sched: cls_u32: validate offshift to prevent shift-out-of-bounds u32_change() copies the user-provided tc_u32_sel.offshift (unsigned char, 0-255) into the kernel knode object without bounds validation. When a packet later hits u32_classify() with TC_U32_VAROFFSET set, it evaluates `ntohs(offmask & *data) >> offshift` where the left operand is a 16-bit value promoted to a 32-bit int. Any offshift >= 32 is undefined behavior per C11 6.5.7p3, triggerable by an unprivileged user via user/network namespaces. UBSAN: shift-out-of-bounds in net/sched/cls_u32.c:236:43 shift exponent 32 is too large for 32-bit type int Fix this by rejecting offshift >= 16 during filter creation in u32_change().

In the Linux kernel, the following vulnerability has been resolved: Bluetooth: ISO: clear iso_data always when detaching conn from hcon When setting conn->hcon = NULL, also conn->hcon->iso_data = NULL is necessary, otherwise later iso_conn_free() will UAF. Fix clearing of iso_data in iso_sock_disconn() Fixes KASAN: slab-use-after-free in iso_conn_hold_unless_zero on iso_sock_release() followed by hci_abort_conn_sync().

In the Linux kernel, the following vulnerability has been resolved: Bluetooth: L2CAP: fix UAF in l2cap_le_connect_rsp l2cap_le_connect_rsp() obtains a channel via __l2cap_get_chan_by_ident() but neither holds a reference nor uses l2cap_chan_hold_unless_zero() before locking and operating on it. A concurrent l2cap_chan_del() triggered by a remote disconnect can free the channel between the lookup and l2cap_chan_lock(), causing a use-after-free. The BR/EDR counterpart l2cap_connect_rsp() and the sibling handler l2cap_le_command_rej() already use l2cap_chan_hold_unless_zero() to safely hold a reference, but l2cap_le_connect_rsp() was left unprotected. Fix by adding l2cap_chan_hold_unless_zero() after the ident lookup and l2cap_chan_put() on the exit path, consistent with other L2CAP response handlers.

In the Linux kernel, the following vulnerability has been resolved: Bluetooth: ISO: lock sk in iso_sock_getname Accessing iso_pi(sk)->conn requires lock_sock, which is not held here. Fix by adding the lock/release.

In the Linux kernel, the following vulnerability has been resolved: Bluetooth: ISO: lock sk in iso_connect_ind Accessing iso_pi(sk)->conn requires lock_sock, which is not taken in the "ev3" part of iso_connect_ind. It may also be NULL if socket has transitioned away from the LISTEN/CONNECT states before locking. Fix by adding lock/release. Recheck hcon is valid after lock acquire where needed.

In the Linux kernel, the following vulnerability has been resolved: Bluetooth: ISO: hold sk properly in iso_conn_ready sk deref in iso_conn_ready must be done either under conn->lock, or holding a refcount, to avoid concurrent close. conn->sk is currently accessed without either: [Task 1] [Task 2] iso_sock_release iso_conn_ready sk = conn->sk lock_sock(sk) conn->sk = NULL lock_sock(sk) release_sock(sk) iso_sock_kill(sk) UAF on sk deref Fix possible UAF by holding sk refcount in iso_conn_ready(). Also recheck after lock_sock that the socket is still valid. Adjust locking so conn->sk is cleared only under lock_sock.

In the Linux kernel, the following vulnerability has been resolved: Bluetooth: ISO: avoid deadlocks in iso_sock_timeout iso_sock_timeout() takes lock_sock, so sync disabling the timer while holding that lock may deadlock. iso_sock_timeout() may also run concurrently with iso_conn_del(), which leads to UAF [Task 1] [Task hdev->workqueue] iso_sock_timeout iso_conn_del iso_conn_hold_unless_zero iso_chan_del `------------> iso_conn_put caller frees hcon iso_conn_put iso_conn_free conn->hcon->iso_data = NULL; /* UAF */ Fix the deadlock by removing the disable from the lock_sock sections. Move the timer from iso_conn to iso_pinfo to decouple it from iso_conn which may need to be freed in lock_sock section. Convert some of the clear_timer to disable_timer.

In the Linux kernel, the following vulnerability has been resolved: Bluetooth: ISO: fix refcounting of iso_conn iso_conn_del() and iso_chan_del() have a race that results to double-put of iso_conn: [Task hdev->workqueue] [Task 2] iso_conn_del iso_chan_del iso_conn_hold_unless_zero iso_conn_lock iso_conn_lock conn->sk = NULL iso_conn_unlock sk = iso_sock_hold(conn) <---------´ if (!sk) iso_conn_put iso_conn_put iso_conn_put /* UAF */ The extra put for !sk in iso_conn_del() is currently required since failing iso_chan_add() may leave iso_conn not associated with any sk. Fix by having iso_pi(sk)->conn own refcount when non-NULL, so iso_conn_del does not need to put it. Adjust the iso_conn_add() refcounting so that conn is put if it does not get associated with an sk.

In the Linux kernel, the following vulnerability has been resolved: Bluetooth: ISO: fix race of kfree vs kref_get_unless_zero hci_conn::iso_data is accessed and modified without lock or RCU. This leads to a race [Task hdev->workqueue] [Task 2] iso_recv iso_conn_put(conn) conn = LOAD hcon->iso_data iso_conn_free(conn) iso_conn_hold_unless_zero(conn) hcon->iso_data = NULL kfree(conn) kref_get_unless_zero(&conn->ref) /* UAF */ and also to races in iso_conn_add() vs. iso_conn_free(). Fix by adding spinlock hci_conn::proto_lock and using it to guard hci_conn::iso_data.

In the Linux kernel, the following vulnerability has been resolved: Bluetooth: hci_conn: hold conn reference in abort_conn_sync() There is theoretical UAF if the conn is freed while the hci_sync task is running. Hold refcount to avoid that.

In the Linux kernel, the following vulnerability has been resolved: Bluetooth: hci_sync: hold conn in hci_connect_big_sync() callback There is theoretical UAF if the conn is freed while the hci_sync task is running. Hold refcount to avoid that. Handle NULL hcon, return 0 + do nothing to match the previous behavior.

In the Linux kernel, the following vulnerability has been resolved: Bluetooth: hci_sync: hold conn in hci_connect_pa_sync() callback There is theoretical UAF if the conn is freed while the hci_sync task is running. Hold refcount to avoid that.

In the Linux kernel, the following vulnerability has been resolved: Bluetooth: hci_sync: hold conn in hci_past_sync() callback Avoids giving freed pointers to hci_conn_valid(), which kmalloc may have reused. Hold refcount to avoid that.

In the Linux kernel, the following vulnerability has been resolved: octeontx2-af: Block VFs from clobbering special CGX PKIND state PF and VF NIX LFs that share a CGX LMAC reuse the same hardware PKIND programming. When HiGig2 or EDSA parsing is enabled, a VF NIX LF alloc must not reset the LMAC RX PKIND or default TX parse config over the PF setup. Add cgx_get_pkind() and rvu_cgx_is_pkind_config_permitted() so VFs skip cgx_set_pkind(), rvu_npc_set_pkind(), and NIX_AF_LFX_TX_PARSE_CFG updates when the LMAC is using NPC_RX_HIGIG_PKIND or NPC_RX_EDSA_PKIND.

In the Linux kernel, the following vulnerability has been resolved: qede: sync udp_tunnel ports outside qede_lock in the recovery path A TX timeout on a qede NIC that has VXLAN/GENEVE tunnel ports configured wedges the rtnetlink control plane of the whole machine: NETDEV WATCHDOG: ens6f1 (qede): transmit queue 2 timed out 10226 ms [qede_tx_timeout:586(ens6f1)]TX timeout on queue 2! [qede_recovery_handler:2665(ens6f0)]Starting a recovery process The recovery path deadlocks on the driver's own mutex: qede_sp_task rtnl_lock() mutex_lock(&edev->qede_lock) <- taken qede_recovery_handler qede_load udp_tunnel_nic_reset_ntf __udp_tunnel_nic_device_sync info->sync_table == qede_udp_tunnel_sync mutex_lock(&edev->qede_lock) <- same task: deadlock The mutex is not recursive, so the kworker blocks on itself with rtnl_lock held, and neither lock is ever released. Every task that calls rtnl_lock() afterwards (ip, ovs-vswitchd, lldpad, IPv6 addrconf, sshd) blocks forever while the node still answers ping. In a vmcore from an affected production node rtnl_mutex.owner decodes to the very kworker blocked at the innermost mutex_lock() above. Re-sync the tunnel ports from qede_sp_task() after the internal lock is dropped, still under rtnl_lock as the udp_tunnel API requires. This mirrors qede_open(), which calls udp_tunnel_nic_reset_ntf() under rtnl without the internal lock. qede_recovery_handler() now returns whether it has successfully reloaded an open device, and the caller re-syncs the ports only in that case. This keeps the old gating exactly: a device that was down or a failed recovery returns false, as those paths never reached the udp_tunnel_nic_reset_ntf() call before either. This was the only user of the qede_lock()/qede_unlock() helpers, so remove them.

In the Linux kernel, the following vulnerability has been resolved: ksmbd: fix use-after-free in __close_file_table_ids() A ksmbd_file can remain alive after logical close while another session holds a temporary reference obtained through ksmbd_lookup_fd_inode(). ksmbd_close_fd() currently marks the file closed and drops the idr-owned reference, but leaves the pointer published in the closing session's idr until the final reference is dropped. If the foreign holder performs the final ksmbd_fd_put(), __put_fd_final() supplies the foreign session's file table to __ksmbd_close_fd(). The object is then freed without being removed from its owner's idr, and the owner session later dereferences the stale pointer during file-table teardown. Remove the volatile id from the owner's idr while ksmbd_close_fd() still holds that table's lock, and clear volatile_id before dropping the idr-owned reference. A later foreign final put then only performs physical destruction and cannot remove the object from the wrong table.

In the Linux kernel, the following vulnerability has been resolved: iommu/iommufd: Fix IOPF group ownership UAF iopf_group_alloc() links each last-page IOPF group into the generic IOPF pending list before invoking the domain fault handler. iommufd_fault_iopf_handler() also queued an accepted group in the IOMMUFD deliver list without removing it from the generic pending list. When detach or HWPT replacement drops the device's IOPF reference count to zero, an IOMMU driver may call iopf_queue_remove_device(). That function responds to and frees groups through the generic pending list without removing the same groups from IOMMUFD's deliver list or response xarray. A later read, response, or cleanup can then access the freed group and cause a UAF. Fix this by dequeuing an accepted group from the generic pending list before IOMMUFD queues it for userspace response. Make iopf_group_response() send a response regardless of pending-list membership, so the dequeued group can still be completed by IOMMUFD.

In the Linux kernel, the following vulnerability has been resolved: pinctrl: devicetree: don't free uninitialized dev_name on error path dt_remember_or_free_map() duplicates dev_name for each map entry. If kstrdup_const() fails, dt_free_map() frees dev_name in all num_maps entries, including entries that have not been initialized. Some pinctrl drivers, including pinctrl-imx, allocate the map with kmalloc() and leave dev_name for the core to initialize. The untouched entries therefore contain uninitialized data which is passed to kfree_const(). Reproduced on qemu's mcimx6ul-evk (pinctrl-imx) with failslab injection while binding the pinctrl-consuming device, under KASAN: BUG: KASAN: double-free in dt_free_map+0x34/0xa4 Free of addr c425a900 by task init/1 kfree from dt_free_map+0x34/0xa4 dt_free_map from dt_remember_or_free_map+0x184/0x198 dt_remember_or_free_map from pinctrl_dt_to_map+0x33c/0x4c8 pinctrl_dt_to_map from create_pinctrl+0x9c/0x5c0 Initialize all dev_name fields to NULL before duplicating the device name, making the full-map cleanup safe after a partial failure.

In the Linux kernel, the following vulnerability has been resolved: mm/hugetlb: fix list corruption in allocate_file_region_entries() allocate_file_region_entries() tops up resv->region_cache with freshly allocated file_region descriptors. The allocation uses GFP_KERNEL, so resv->lock is dropped around it: the new entries are gathered on a stack-local list head, allocated_regions, and spliced into resv->region_cache once the lock is re-acquired. The splice used list_splice(), which moves the entries but does not re-initialize the source head, so allocated_regions is left pointing at an entry that now lives on resv->region_cache. The top-up runs in a while loop that re-checks the cache deficit after re-acquiring the lock. For a shared mapping the resv_map is shared by every mapper of the hugetlbfs inode, so a concurrent region_chg()/region_add()/region_del() on the same resv_map can consume cache entries during the unlocked window and force a second iteration. That iteration calls list_add() on the stale head and corrupts the list; with CONFIG_DEBUG_LIST the __list_add_valid() check trips: list_add corruption. next->prev should be prev (ffffc900011ff7f8), but was ffff88814c281460. (next=ffff88814c545640). kernel BUG at lib/list_debug.c:31! allocate_file_region_entries+0x191/0x420 region_chg+0x267/0x300 hugetlb_reserve_pages+0x387/0xc80 hugetlbfs_file_mmap+0x2ce/0x3f0 mmap_region+0x1348/0x1a80 do_mmap+0x85e/0xb90 vm_mmap_pgoff+0x18c/0x330 ksys_mmap_pgoff+0x2a1/0x3e0 do_syscall_64+0xd7/0x420 Without CONFIG_DEBUG_LIST the bad list_add() silently links a kernel-stack address into resv->region_cache, leading to later use-after-free. This was observed as a real host panic on a dense KVM host where a QEMU guest-RAM hugetlbfs file was mapped MAP_SHARED by both QEMU and a separate SPDK/DPDK vhost-user target, generating concurrent region_* traffic on one shared resv_map. Use list_splice_init() so the source head is re-initialized empty after each splice, making the retry loop safe.

In the Linux kernel, the following vulnerability has been resolved: KVM: SVM: Update x2APIC MSR intercepts if AVIC is inhibited while L2 is active Always update x2APIC MSR intercepts for L1 when AVIC is deactivated, even if L2 is active and KVM is using a separate MSR bitmap to run L2. If AVIC is fully enabled prior to running L2, and is then inhibited while L2 is active (for a VM-scoped inhibit), then KVM will run L1 with AVIC disabled, but with x2APIC MSR intercepts disabled, i.e. will allow L1 to read most of the host's APIC state, send arbitrary interrupts, change task priority, and ultimately trivially DoS the host. E.g. sending a self-IPI in L1 on HYPERV_REENLIGHTENMENT_VECTOR, 0xee, with CONFIG_HYPERV=n in the host kernel as a "safe" PoC, yields: Spurious interrupt (vector 0xee) on CPU#425. Acked And hacking KVM to abuse kvm_set_posted_intr_wakeup_handler() to register a handler and WARN on POSTED_INTR_WAKEUP_VECTOR yields: ------------[ cut here ]------------ WARNING: arch/x86/kvm/svm/svm.c:5594 at pi_wakeup_handler+0x9/0x10 [kvm_amd], CPU#156: nested_x2apic_t/316940 CPU: 156 UID: 0 PID: 316940 Comm: nested_x2apic_t Tainted: G S U Tainted: [S]=CPU_OUT_OF_SPEC, [U]=USER Hardware name: Google Astoria-Turin/astoria, BIOS 0.20260209.0-0 02/09/2026 RIP: 0010:pi_wakeup_handler+0x9/0x10 [kvm_amd] Call Trace: <IRQ> sysvec_kvm_posted_intr_wakeup_ipi+0x64/0x80 </IRQ> <TASK> asm_sysvec_kvm_posted_intr_wakeup_ipi+0x1a/0x20 RIP: 0010:vcpu_run+0x1430/0x1e40 [kvm] kvm_arch_vcpu_ioctl_run+0x2c1/0x600 [kvm] kvm_vcpu_ioctl+0x580/0x6b0 [kvm] __se_sys_ioctl+0x6d/0xb0 do_syscall_64+0x10a/0x480 entry_SYSCALL_64_after_hwframe+0x4b/0x53 RIP: 0033:0x46ff4b </TASK> ---[ end trace 0000000000000000 ]---

In the Linux kernel, the following vulnerability has been resolved: KVM: s390: pci: Reject adapter interrupt forwarding if already enabled The MPCIFC instruction doesn't allow registering adapter interrupts without first unregistering. So reject any request to enable interrupt forwarding if its already enabled for the zPCI device. This also fixes overwriting and thus leaking resources when the ioctl is called multiple times for the same device.

In the Linux kernel, the following vulnerability has been resolved: dibs: fix use-after-free of dmb_node in loopback attach/detach/unregister dibs_lo_attach_dmb(), dibs_lo_detach_dmb() and dibs_lo_unregister_dmb() look up the dmb_node under dmb_ht_lock, drop the lock and only then operate on the node's refcount. Nothing keeps the node alive across that window: __dibs_lo_unregister_dmb() removes the node from the hash table under the write lock and immediately frees it. A concurrent final put can therefore free the node between the lookup and the refcount operation: CPU0 (attach) CPU1 (owner unregisters) read_lock_bh(&dmb_ht_lock) find dmb_node (refcnt == 1) read_unlock_bh(&dmb_ht_lock) refcount_dec_and_test() 1 -> 0 write_lock_bh(&dmb_ht_lock) hash_del(&dmb_node->list) write_unlock_bh(&dmb_ht_lock) kfree(dmb_node) refcount_inc_not_zero(&dmb_node->refcnt) <-- use-after-free The same window exists for the refcount_dec_and_test() calls in the detach and unregister paths. Close the race structurally by making hash table membership and the refcount transitions atomic with respect to each other: - Perform the final refcount_dec_and_test() and hash_del() in a single dmb_ht_lock write-side critical section, in both the unregister and the detach path. Freeing the node still happens after the lock is dropped, which is safe because a node whose refcount reached zero has left the hash table and can no longer be found. - This establishes the invariant that any node found in the hash table holds at least one reference, and that the final reference can only be dropped under the write lock. dibs_lo_attach_dmb() can thus take its reference with a plain refcount_inc() while still holding the read lock; refcount_inc_not_zero() is no longer needed. __dibs_lo_unregister_dmb() no longer touches the hash table and is renamed to dibs_lo_free_dmb() accordingly. Note: commit cc21191b584c ("dibs: Move data path to dibs layer") moved the code to its current location; the race was introduced earlier by commit c3a910f2380f ("net/smc: implement DMB-merged operations of loopback-ism"). Tested SMC-D via ISM and dibs loopback.