On 7/10/26 00:43, Link Lin wrote:
> During system power management freeze (e.g. ACPI S3 suspend or S4
> hibernation), virtballoon_freeze() calls remove_common() to reset the
> virtio device and delete all virtqueues via vdev->config->del_vqs().
> However, unlike virtballoon_remove(), virtballoon_freeze() fails to call
> page_reporting_unregister(&vb->pr_dev_info).
>
> The comment in virtballoon_freeze() states:
> /*
> * The workqueue is already frozen by the PM core before this
> * function is called.
> */
>
> While this comment was accurate in 2011 for balloon-internal workqueues
> (such as balloon_wq, which was created with WQ_FREEZABLE and is paused
> by the PM freezer), it is invalid for Free Page Reporting.
>
> Free Page Reporting (mm/page_reporting.c) schedules its delayed work
> (prdev->work) on the global system_wq. Because system_wq lacks the
> WQ_FREEZABLE flag, the PM freezer (freeze_workqueues_busy()) explicitly
> skips it. Consequently, page_reporting_process() on system_wq remains
> active and unfrozen throughout device suspend.
>
> If memory is freed into the buddy allocator or a delayed work timer
> expires while the device is being frozen, page_reporting_process() fires
> on system_wq and calls virtballoon_free_page_report(). This function
> passes vb->reporting_vq into virtqueue_add_inbuf() / virtqueue_add_split().
> Because the virtqueues were already destroyed by del_vqs(), this results
> in a Use-After-Free / General Protection Fault:
>
> [ 250.709271] general protection fault, probably for non-canonical
> address 0x7f728084daf08d5e: 0000 [#1] SMP PTI
> [ 250.732967] CPU: 2 PID: 38 Comm: kworker/2:1 Not tainted
> 5.10.0-44-cloud-amd64 #1 Debian 5.10.257-1
> [ 250.751575] Workqueue: events page_reporting_process
> [ 250.756665] RIP: 0010:virtqueue_add_split+0x233/0x4c0 [virtio_ring]
> ...
> [ 250.867678] virtballoon_free_page_report+0x3a/0xe0 [virtio_balloon]
> [ 250.883446] page_reporting_process+0x225/0x4f0
>
> (Note: The OOM Notifier and Shrinker/Free Page Hinting features suffer
> from an identical lifecycle flaw and are also vulnerable to UAFs during
> S4 hibernation when memory pressure spikes. This patch focuses on Free
> Page Reporting, which runs periodically, to ensure clean backports to
> stable kernels).
>
> Fix this by:
> 1. Unregistering page reporting in virtballoon_freeze() prior to calling
> remove_common(). This clears the RCU pr_dev_info pointer and flushes/
> cancels prdev->work on system_wq via cancel_delayed_work_sync().
> 2. Re-registering page reporting in virtballoon_restore() after the
> virtqueues are re-initialized and virtio_device_ready() has been called.
> 3. Unwinding virtqueue initialization via remove_common() in
> virtballoon_restore() if page_reporting_register() fails.
>
> Fixes: 924a663f75e2 ("virtio-balloon: Reporting free page reservations")
> Cc: [email protected]
> Cc: [email protected]
> Cc: [email protected]
> Cc: Ammar Faizi <[email protected]>
> Cc: [email protected]
> Cc: [email protected]
> Cc: Greg Thelen <[email protected]>
> Cc: Alexander Duyck <[email protected]>
> Signed-off-by: Link Lin <[email protected]>
> ---
> drivers/virtio/virtio_balloon.c | 11 +++++++++++
> 1 file changed, 11 insertions(+)
>
> diff --git a/drivers/virtio/virtio_balloon.c b/drivers/virtio/virtio_balloon.c
> index a1b2c3d4e5f6..45a90fb3abf8 100640
> --- a/drivers/virtio/virtio_balloon.c
> +++ b/drivers/virtio/virtio_balloon.c
> @@ -1055,6 +1055,9 @@ static int virtballoon_freeze(struct virtio_device
> *vdev)
> * The workqueue is already frozen by the PM core before this
> * function is called.
> */
> + if (virtio_has_feature(vb->vdev, VIRTIO_BALLOON_F_REPORTING))
> + page_reporting_unregister(&vb->pr_dev_info);
> +
> remove_common(vb);
> return 0;
> }
>
> static int virtballoon_restore(struct virtio_device *vdev)
> {
> struct virtio_balloon *vb = vdev->priv;
> int ret;
>
> ret = init_vqs(vdev->priv);
> if (ret)
> return ret;
>
> virtio_device_ready(vdev);
>
> + if (virtio_has_feature(vb->vdev, VIRTIO_BALLOON_F_REPORTING)) {
> + ret = page_reporting_register(&vb->pr_dev_info);
> + if (ret)
> + goto out_remove_vqs;
> + }
Hm, that failure handling is rather nasty.
In virtballoon_freeze() we document:
"The workqueue is already frozen by the PM core before this function is called"
Your report states:
"Workqueue: events page_reporting_process"
I assume that workqueue is not frozen yet because ... it's not freezable :)
So could we queue to system_freezable_wq instead, or define our own freezable
workqueue there? Then a driver doesn't have to worry about that.
--
Cheers,
David
