2014-11-09 09:38:55

by Fam Zheng

Subject: [PATCH] virtio-pci: Reset device on shutdown

This fixes a hanging issue during guest shutdown.

The device is left enabled even though we removed it and disabled msix
during shutdown. If the virtio device happens to get a new event right
at this point, seeing msix is disabled, it may try to notify us with an
IRQ, which is totally unexpected thus will not be handled. In this
case the guest hangs.

Let's reset the device so that it will not generate any interrupts.

Link: https://bugzilla.redhat.com/show_bug.cgi?id=1128424
Cc: Ulrich Obergfell <[email protected]>
Signed-off-by: Fam Zheng <[email protected]>
---
drivers/virtio/virtio_pci.c | 9 +++++++++
1 file changed, 9 insertions(+)

diff --git a/drivers/virtio/virtio_pci.c b/drivers/virtio/virtio_pci.c
index d34ebfa..cd7be6f 100644
--- a/drivers/virtio/virtio_pci.c
+++ b/drivers/virtio/virtio_pci.c
@@ -743,6 +743,14 @@ out:
return err;
}

+static void virtio_pci_shutdown(struct pci_dev *pci_dev)
+{
+ struct virtio_pci_device *vp_dev = pci_get_drvdata(pci_dev);
+
+ vp_reset(&vp_dev->vdev);
+}
+
+
static void virtio_pci_remove(struct pci_dev *pci_dev)
{
struct virtio_pci_device *vp_dev = pci_get_drvdata(pci_dev);
@@ -794,6 +802,7 @@ static struct pci_driver virtio_pci_driver = {
.id_table = virtio_pci_id_table,
.probe = virtio_pci_probe,
.remove = virtio_pci_remove,
+ .shutdown = virtio_pci_shutdown,
#ifdef CONFIG_PM_SLEEP
.driver.pm = &virtio_pci_pm_ops,
#endif
--
1.9.3

2014-11-09 11:52:33

by Michael S. Tsirkin

Subject: Re: [PATCH] virtio-pci: Reset device on shutdown

On Sun, Nov 09, 2014 at 05:39:05PM +0800, Fam Zheng wrote:
> This fixes a hanging issue during guest shutdown.
>
> The device is left enabled even though we removed it and disabled msix
> during shutdown. If the virtio device happens to get a new event right
> at this point, seeing msix is disabled, it may try to notify us with an
> IRQ, which is totally unexpected thus will not be handled. In this
> case the guest hangs.
>
> Let's reset the device so that it will not generate any interrupts.
>
> Link: https://bugzilla.redhat.com/show_bug.cgi?id=1128424
> Cc: Ulrich Obergfell <[email protected]>
> Signed-off-by: Fam Zheng <[email protected]>

This needs some careful thought.

vp_reset isn't safe unless there's a guarantee that
nothing is accessing the device at the same time,
which normally needs coordination with the function-specific
driver.

For example, for virtio net, you will notice that on remove path we
do unregister_netdev before reset for this reason.

What guarantees it in this case?


> ---
> drivers/virtio/virtio_pci.c | 9 +++++++++
> 1 file changed, 9 insertions(+)
>
> diff --git a/drivers/virtio/virtio_pci.c b/drivers/virtio/virtio_pci.c
> index d34ebfa..cd7be6f 100644
> --- a/drivers/virtio/virtio_pci.c
> +++ b/drivers/virtio/virtio_pci.c
> @@ -743,6 +743,14 @@ out:
> return err;
> }
>
> +static void virtio_pci_shutdown(struct pci_dev *pci_dev)
> +{
> + struct virtio_pci_device *vp_dev = pci_get_drvdata(pci_dev);
> +
> + vp_reset(&vp_dev->vdev);
> +}
> +
> +
> static void virtio_pci_remove(struct pci_dev *pci_dev)
> {
> struct virtio_pci_device *vp_dev = pci_get_drvdata(pci_dev);
> @@ -794,6 +802,7 @@ static struct pci_driver virtio_pci_driver = {
> .id_table = virtio_pci_id_table,
> .probe = virtio_pci_probe,
> .remove = virtio_pci_remove,
> + .shutdown = virtio_pci_shutdown,
> #ifdef CONFIG_PM_SLEEP
> .driver.pm = &virtio_pci_pm_ops,
> #endif
> --
> 1.9.3