Re: [PATCH v1] misc: vmw_vmci: Leak queue pair endpoint if detach fails on exit
From: Vishnu Dasa
Date: Wed Nov 18 2026 - 16:13:40 EST
On Thu, Nov 5, 2026 at 7:44 PM Yuho Choi <oss.patchbox@xxxxxxxxx> wrote:
>
> When the VMCI guest device is removed, vmci_qp_guest_endpoints_exit()
> asks the hypervisor to detach each remaining queue pair and then frees
> its queue pages with qp_guest_endpoint_destroy(), even if the detach
> hypercall failed.
>
> If the detach fails, the hypervisor may still have the queue pair, and
> the peer may keep writing to the queue pages after they have gone back
> to the page allocator. The device reset that follows in
> vmci_guest_remove_device() comes only after the pages have been freed.
>
> qp_detatch_guest_work() already handles the same failure by keeping the
> entry, because the peer "might still be accessing the shared memory",
> and leaves the cleanup to this exit path, which then frees it anyway.
>
> Leak the endpoint instead if the detach fails. This also keeps its
> handle in the resource table, so the handle is not given to a new queue
> pair while the hypervisor may still have the old one.
>
> Fixes: 06164d2b72aa ("VMCI: queue pairs implementation.")
> Signed-off-by: Yuho Choi <oss.patchbox@xxxxxxxxx>
Thanks for the fix. It makes sense.
> ---
> Compile-tested only (x86_64 with VMWARE_VMCI, W=1); I do not have
> a VMware setup to test it on.
>
> drivers/misc/vmw_vmci/vmci_queue_pair.c | 15 +++++++++++----
> 1 file changed, 11 insertions(+), 4 deletions(-)
>
> diff --git a/drivers/misc/vmw_vmci/vmci_queue_pair.c b/drivers/misc/vmw_vmci/vmci_queue_pair.c
> index e1a11820be01..57a4f0ae4f49 100644
> --- a/drivers/misc/vmw_vmci/vmci_queue_pair.c
> +++ b/drivers/misc/vmw_vmci/vmci_queue_pair.c
> @@ -2391,28 +2391,35 @@ int vmci_qp_broker_unmap(struct vmci_handle handle,
> /*
> * Destroys all guest queue pair endpoints. If active guest queue
> * pairs still exist, hypercalls to attempt detach from these
> - * queue pairs will be made. Any failure to detach is silently
> - * ignored.
> + * queue pairs will be made. If a detach fails, the endpoint is
> + * leaked, as the peer may still access its queues.
IMHO, this comment could be a little clearer. For example:
"Destroys all guest queue pair endpoints. If active guest queue pairs
still exist, hypercalls to attempt detach from these queue pairs will
be made.
If a detach hypercall fails, the hypervisor may still have the queue
pair's pages mapped. Freeing them can cause a use-after-free bug if
the peer continues to access the pages. The endpoint is therefore
intentionally leaked instead."
> */
> void vmci_qp_guest_endpoints_exit(void)
> {
> struct qp_entry *entry;
> struct qp_guest_endpoint *ep;
> + int result;
>
> mutex_lock(&qp_guest_endpoints.mutex);
>
> while ((entry = qp_list_get_head(&qp_guest_endpoints))) {
> ep = (struct qp_guest_endpoint *)entry;
> + result = VMCI_SUCCESS;
>
> /* Don't make a hypercall for local queue_pairs. */
> if (!(entry->flags & VMCI_QPFLAG_LOCAL))
> - qp_detatch_hypercall(entry->handle);
> + result = qp_detatch_hypercall(entry->handle);
>
> /* We cannot fail the exit, so let's reset ref_count. */
> entry->ref_count = 0;
> qp_list_remove_entry(&qp_guest_endpoints, entry);
>
> - qp_guest_endpoint_destroy(ep);
> + /*
> + * If the detach failed, the peer may still access the
> + * queues, so leak the endpoint.
> + */
> + if (result >= VMCI_SUCCESS)
> + qp_guest_endpoint_destroy(ep);
Could you add an else branch that does a pr_warn() with the handle and
the detach result, so a leaked queue pair can be matched to a cause?
Otherwise, this intentional leak will be silent.
> }
>
> mutex_unlock(&qp_guest_endpoints.mutex);
>
> base-commit: 6628e7d4c6b25c8d18f1345e3a20d21829aae7e4
> --
> 2.43.0
>
Attachment:
smime.p7s
Description: S/MIME Cryptographic Signature