In the Linux kernel, the following vulnerability has been resolved:
RDMA/irdma: Replace waitqueue and flag with completion
The driver previously used a waitqueue along with an explicit
request_done flag, but without proper barriers around request_done.
An earlier patch by Gui-Dong Han hanguidong02@gmail.com attempted
to fix this by adding the missing memory barriers. Rather than
adding the barriers, this patch replaces the waitqueue+flag with
a completion, which is designed for this exact purpose.
CVSS Vector: CVSS:3.1/AV:N/AC:L/PR:N/UI:N/S:U/C:H/I:H/A:H
CVSS Score: 9.8
AV:N - Intel irdma CQP completion/wait synchronization is exercised from remote iWARP/RoCE connection management (SYN/listener/IEQ/AEQ) and deferred-completion events, not only local ibverbs; network peers drive the completion side of the race on internet-facing RDMA/NFS/SMB-Direct servers.
AC:L - Attacker controls both sides by flooding concurrent connections/CQP ops remotely or via concurrent RDMA verbs locally; missing release/acquire ordering on request_done vs compl_info is triggerable at high operation rates without luck-dependent layout.
PR:N - Remote CM and packet-processing paths issue/wait on CQP before MPA/application auth; local exploitation needs only standard unprivileged /dev/infiniband/uverbs* access (0666 under typical rdma-core udev rules), not init-namespace capabilities.
UI:N - Exploitation needs only attacker-driven network or verbs traffic against an already-bound RDMA device/listener; no additional victim user action at attack time beyond normal server role.
S:U - Impact stays within kernel/driver security domain (CQP request objects, NIC PF reset); no VM escape or cross-authority boundary.
C:H - Without proper barriers the waiter can observe request_done before compl_info stores, return success with stale op_ret_val/error, and recycle the cqp_request while hardware still references scratch → UAF/read primitives.
I:H - Premature recycle/double-put of dynamic cqp_requests enables attacker-influenced heap reuse; stale callbacks on reused objects and invalid kfree paths provide kernel write/control-flow corruption primitives.
A:H - Lost request_done causes CQP timeout → ice_rdma_request_reset PF reset; wrong completion state can leave mod_qp_waitq in uninterruptible wait; UAF/panic paths cause kernel crash/hang.
| Attack Vector |
Network |
Scope |
Unchanged |
| Attack Complexity |
Low |
Confidentiality Impact |
High |
| Privileges Required |
None |
Integrity Impact |
High |
| User Interaction |
None |
Availability Impact |
High |
AV:N - Intel irdma CQP completion/wait synchronization is exercised from remote iWARP/RoCE connection management (SYN/listener/IEQ/AEQ) and deferred-completion events, not only local ibverbs; network peers drive the completion side of the race on internet-facing RDMA/NFS/SMB-Direct servers.
AC:L - Attacker controls both sides by flooding concurrent connections/CQP ops remotely or via concurrent RDMA verbs locally; missing release/acquire ordering on request_done vs compl_info is triggerable at high operation rates without luck-dependent layout.
PR:N - Remote CM and packet-processing paths issue/wait on CQP before MPA/application auth; local exploitation needs only standard unprivileged /dev/infiniband/uverbs* access (0666 under typical rdma-core udev rules), not init-namespace capabilities.
UI:N - Exploitation needs only attacker-driven network or verbs traffic against an already-bound RDMA device/listener; no additional victim user action at attack time beyond normal server role.
S:U - Impact stays within kernel/driver security domain (CQP request objects, NIC PF reset); no VM escape or cross-authority boundary.
C:H - Without proper barriers the waiter can observe request_done before compl_info stores, return success with stale op_ret_val/error, and recycle the cqp_request while hardware still references scratch → UAF/read primitives.
I:H - Premature recycle/double-put of dynamic cqp_requests enables attacker-influenced heap reuse; stale callbacks on reused objects and invalid kfree paths provide kernel write/control-flow corruption primitives.
A:H - Lost request_done causes CQP timeout → ice_rdma_request_reset PF reset; wrong completion state can leave mod_qp_waitq in uninterruptible wait; UAF/panic paths cause kernel crash/hang.
CVSS 3.1