Linux Virtualization - Jan 2015 - [PATCH 2/2] virtio: don't free memory until the underlying struct device has been released

Device VQs were getting freed twice: once in every devices removal functions,
and then again in virtio_pci_legacy_remove().

Signed-off-by: Sasha Levin <sasha.levin at oracle.com>
---
 drivers/virtio/virtio_pci_legacy.c |    1 -
 1 file changed, 1 deletion(-)

diff --git a/drivers/virtio/virtio_pci_legacy.c
b/drivers/virtio/virtio_pci_legacy.c
index 6c76f0f..913ca23 100644
--- a/drivers/virtio/virtio_pci_legacy.c
+++ b/drivers/virtio/virtio_pci_legacy.c
@@ -298,7 +298,6 @@ void virtio_pci_legacy_remove(struct pci_dev *pci_dev)
 
 	unregister_virtio_device(&vp_dev->vdev);
 
-	vp_del_vqs(&vp_dev->vdev);
 	pci_iounmap(pci_dev, vp_dev->ioaddr);
 	pci_release_regions(pci_dev);
 	pci_disable_device(pci_dev);
-- 
1.7.10.4

When releasing a virtio device, We can't free a struct virtio_device until
the
underlying struct device has been released, which might not happen immediately
on device_unregister() even if that was the device's last reference.

Instead, free the memory only once we know the device is gone in the release
callback.

Signed-off-by: Sasha Levin <sasha.levin at oracle.com>
---
 drivers/virtio/virtio_pci_common.c |    9 ++++-----
 drivers/virtio/virtio_pci_legacy.c |    1 -
 2 files changed, 4 insertions(+), 6 deletions(-)

diff --git a/drivers/virtio/virtio_pci_common.c
b/drivers/virtio/virtio_pci_common.c
index 59d3685..caa483d 100644
--- a/drivers/virtio/virtio_pci_common.c
+++ b/drivers/virtio/virtio_pci_common.c
@@ -423,11 +423,10 @@ int vp_set_vq_affinity(struct virtqueue *vq, int cpu)
 
 void virtio_pci_release_dev(struct device *_d)
 {
-	/*
-	 * No need for a release method as we allocate/free
-	 * all devices together with the pci devices.
-	 * Provide an empty one to avoid getting a warning from core.
-	 */
+	struct virtio_device *vdev = dev_to_virtio(_d);
+	struct virtio_pci_device *vp_dev = to_vp_device(vdev);
+
+	kfree(vp_dev);
 }
 
 #ifdef CONFIG_PM_SLEEP
diff --git a/drivers/virtio/virtio_pci_legacy.c
b/drivers/virtio/virtio_pci_legacy.c
index 913ca23..15e6e6d 100644
--- a/drivers/virtio/virtio_pci_legacy.c
+++ b/drivers/virtio/virtio_pci_legacy.c
@@ -301,5 +301,4 @@ void virtio_pci_legacy_remove(struct pci_dev *pci_dev)
 	pci_iounmap(pci_dev, vp_dev->ioaddr);
 	pci_release_regions(pci_dev);
 	pci_disable_device(pci_dev);
-	kfree(vp_dev);
 }
-- 
1.7.10.4

On Fri, Jan 02, 2015 at 02:47:39PM -0500, Sasha Levin
wrote:
> Device VQs were getting freed twice: once in every devices removal
functions,
> and then again in virtio_pci_legacy_remove().
> 
> Signed-off-by: Sasha Levin <sasha.levin at oracle.com>
I see.  This used to be harmless because it used to be idempotent,
and safe to call any number of times.

I think your patch is right for 3.20:

Acked-by: Michael S. Tsirkin <mst at redhat.com>

but for 3.19 I'd prefer a more conservative approach of making
del_vqs idempotent again.

I'll post a patch like this shortly.

> ---
>  drivers/virtio/virtio_pci_legacy.c |    1 -
>  1 file changed, 1 deletion(-)
> 
> diff --git a/drivers/virtio/virtio_pci_legacy.c
b/drivers/virtio/virtio_pci_legacy.c
> index 6c76f0f..913ca23 100644
> --- a/drivers/virtio/virtio_pci_legacy.c
> +++ b/drivers/virtio/virtio_pci_legacy.c
> @@ -298,7 +298,6 @@ void virtio_pci_legacy_remove(struct pci_dev *pci_dev)
>  
>  	unregister_virtio_device(&vp_dev->vdev);
>  
> -	vp_del_vqs(&vp_dev->vdev);
>  	pci_iounmap(pci_dev, vp_dev->ioaddr);
>  	pci_release_regions(pci_dev);
>  	pci_disable_device(pci_dev);
> -- 
> 1.7.10.4

On Fri, Jan 02, 2015 at 02:47:40PM -0500, Sasha Levin
wrote:
> When releasing a virtio device, We can't free a struct virtio_device
until the
> underlying struct device has been released, which might not happen
immediately
> on device_unregister() even if that was the device's last reference.
> 
> Instead, free the memory only once we know the device is gone in the
release
> callback.
> 
> Signed-off-by: Sasha Levin <sasha.levin at oracle.com>

Isn't this an old bug: do we need to copy stable on a fix?

What is the behaviour without this patch?
Is there a way to make this cause a crash?
> ---
>  drivers/virtio/virtio_pci_common.c |    9 ++++-----
>  drivers/virtio/virtio_pci_legacy.c |    1 -
>  2 files changed, 4 insertions(+), 6 deletions(-)
> 
> diff --git a/drivers/virtio/virtio_pci_common.c
b/drivers/virtio/virtio_pci_common.c
> index 59d3685..caa483d 100644
> --- a/drivers/virtio/virtio_pci_common.c
> +++ b/drivers/virtio/virtio_pci_common.c
> @@ -423,11 +423,10 @@ int vp_set_vq_affinity(struct virtqueue *vq, int cpu)
>  
>  void virtio_pci_release_dev(struct device *_d)
>  {
> -	/*
> -	 * No need for a release method as we allocate/free
> -	 * all devices together with the pci devices.
> -	 * Provide an empty one to avoid getting a warning from core.
> -	 */
> +	struct virtio_device *vdev = dev_to_virtio(_d);
> +	struct virtio_pci_device *vp_dev = to_vp_device(vdev);
> +
> +	kfree(vp_dev);
>  }
>  
>  #ifdef CONFIG_PM_SLEEP
> diff --git a/drivers/virtio/virtio_pci_legacy.c
b/drivers/virtio/virtio_pci_legacy.c
> index 913ca23..15e6e6d 100644
> --- a/drivers/virtio/virtio_pci_legacy.c
> +++ b/drivers/virtio/virtio_pci_legacy.c
> @@ -301,5 +301,4 @@ void virtio_pci_legacy_remove(struct pci_dev *pci_dev)
>  	pci_iounmap(pci_dev, vp_dev->ioaddr);
>  	pci_release_regions(pci_dev);
>  	pci_disable_device(pci_dev);
> -	kfree(vp_dev);
>  }

It seems inelegant to free a structure allocated in another file:
I think we should move this function to virtio_pci_legacy.
Will send a patch in a minute.




> -- 
> 1.7.10.4