[PATCH] virtio: decrement dev_index when device is unregistered

[PATCH] virtio: decrement dev_index when device is unregistered

[Takuma Umeya]

When virtio device is removed, dev_index does not get decremented. 
The next device hotplug event results in consuming the next pci to 
the one that is suppose to be available. 

Signed-off-by: Takuma Umeya <tumeya@redhat.com>

diff --git a/drivers/virtio/virtio.c b/drivers/virtio/virtio.c
index efb35aa..67fe71d 100644
--- a/drivers/virtio/virtio.c
+++ b/drivers/virtio/virtio.c
@@ -216,6 +216,7 @@ EXPORT_SYMBOL_GPL(register_virtio_device);
 void unregister_virtio_device(struct virtio_device *dev)
 {
        device_unregister(&dev->dev);
+       dev_index--;
 }
 EXPORT_SYMBOL_GPL(unregister_virtio_device);
 

-- 
Umeya, Takuma
Technical Account Manager
Red Hat GSS APAC
+81.3.5798.8584 (direct)
tumeya@redhat.com

Re: [PATCH] virtio: decrement dev_index when device is unregistered

[Stefan Hajnoczi]

On Tue, Apr 5, 2011 at 5:49 AM, Takuma Umeya <tumeya@redhat.com> wrote:
> When virtio device is removed, dev_index does not get decremented.
> The next device hotplug event results in consuming the next pci to
> the one that is suppose to be available.
>
> Signed-off-by: Takuma Umeya <tumeya@redhat.com>
>
> diff --git a/drivers/virtio/virtio.c b/drivers/virtio/virtio.c
> index efb35aa..67fe71d 100644
> --- a/drivers/virtio/virtio.c
> +++ b/drivers/virtio/virtio.c
> @@ -216,6 +216,7 @@ EXPORT_SYMBOL_GPL(register_virtio_device);
>  void unregister_virtio_device(struct virtio_device *dev)
>  {
>        device_unregister(&dev->dev);
> +       dev_index--;

I don't think there is any guarantee that virtio devices are
added/removed in first-in-last-out order.

That means I could add a virtio-net device (index 0) followed by a
virtio-blk device (index 1).  Now I remove the virtio-net device
(index 0) which causes me to decrement dev_index and hand index 1 out
again to the next device.  This leaves us with virtio-blk (index 1)
and the new device with index 1, which is not unique.

Perhaps I missed a constraint which prevents this from occurring?

Stefan

Re: [PATCH] virtio: decrement dev_index when device is unregistered

[Takuma Umeya]

----- Original Message -----
> On Tue, Apr 5, 2011 at 5:49 AM, Takuma Umeya <tumeya@redhat.com>
> wrote:
> > When virtio device is removed, dev_index does not get decremented.
> > The next device hotplug event results in consuming the next pci to
> > the one that is suppose to be available.
> >
> > Signed-off-by: Takuma Umeya <tumeya@redhat.com>
> >
> > diff --git a/drivers/virtio/virtio.c b/drivers/virtio/virtio.c
> > index efb35aa..67fe71d 100644
> > --- a/drivers/virtio/virtio.c
> > +++ b/drivers/virtio/virtio.c
> > @@ -216,6 +216,7 @@ EXPORT_SYMBOL_GPL(register_virtio_device);
> >  void unregister_virtio_device(struct virtio_device *dev)
> >  {
> >        device_unregister(&dev->dev);
> > + dev_index--;
> 
> I don't think there is any guarantee that virtio devices are
> added/removed in first-in-last-out order.
> 
> That means I could add a virtio-net device (index 0) followed by a
> virtio-blk device (index 1). Now I remove the virtio-net device
> (index 0) which causes me to decrement dev_index and hand index 1 out
> again to the next device. This leaves us with virtio-blk (index 1)
> and the new device with index 1, which is not unique.
> 
> Perhaps I missed a constraint which prevents this from occurring?
I believe the address is assigned up to 1f so using u32 value 
to track use/free. This should make the code immune to the scenario. 
Would this be adequate? 

diff --git a/drivers/virtio/virtio.c b/drivers/virtio/virtio.c
index efb35aa..0c73507 100644
--- a/drivers/virtio/virtio.c
+++ b/drivers/virtio/virtio.c
@@ -3,7 +3,7 @@
 #include <linux/virtio_config.h>
 
 /* Unique numbering for virtio devices. */
-static unsigned int dev_index;
+static u32 dev_index;
 
 static ssize_t device_show(struct device *_d,
                           struct device_attribute *attr, char *buf)
@@ -187,12 +187,23 @@ EXPORT_SYMBOL_GPL(unregister_virtio_driver);
 
 int register_virtio_device(struct virtio_device *dev)
 {
-       int err;
+       int err, count;
+       u32 testbit;
+
+       count = 0; 
 
        dev->dev.bus = &virtio_bus;
 
        /* Assign a unique device index and hence name. */
-       dev->index = dev_index++;
+       while (count < 32){
+            testbit = 1UL << count;
+            if(!(dev_index & testbit)){
+                dev->index = count;
+                dev_index |= testbit;
+                break;
+            }
+            count++;
+        }
        dev_set_name(&dev->dev, "virtio%u", dev->index);
 
        /* We always start by resetting the device, in case a previous
@@ -215,7 +226,13 @@ EXPORT_SYMBOL_GPL(register_virtio_device);
 
 void unregister_virtio_device(struct virtio_device *dev)
 {
+       u32 removebit;
+
+        removebit = 1UL << (dev->index);
+
        device_unregister(&dev->dev);
+
+       dev_index ^= removebit;
 }
 EXPORT_SYMBOL_GPL(unregister_virtio_device);
 
> 
> Stefan

Re: [PATCH] virtio: decrement dev_index when device is unregistered

[Stefan Hajnoczi]

On Mon, Apr 11, 2011 at 10:11 AM, Takuma Umeya <tumeya@redhat.com> wrote:
> ----- Original Message -----
>> On Tue, Apr 5, 2011 at 5:49 AM, Takuma Umeya <tumeya@redhat.com>
>> wrote:
>> > When virtio device is removed, dev_index does not get decremented.
>> > The next device hotplug event results in consuming the next pci to
>> > the one that is suppose to be available.
>> >
>> > Signed-off-by: Takuma Umeya <tumeya@redhat.com>
>> >
>> > diff --git a/drivers/virtio/virtio.c b/drivers/virtio/virtio.c
>> > index efb35aa..67fe71d 100644
>> > --- a/drivers/virtio/virtio.c
>> > +++ b/drivers/virtio/virtio.c
>> > @@ -216,6 +216,7 @@ EXPORT_SYMBOL_GPL(register_virtio_device);
>> >  void unregister_virtio_device(struct virtio_device *dev)
>> >  {
>> >        device_unregister(&dev->dev);
>> > + dev_index--;
>>
>> I don't think there is any guarantee that virtio devices are
>> added/removed in first-in-last-out order.
>>
>> That means I could add a virtio-net device (index 0) followed by a
>> virtio-blk device (index 1). Now I remove the virtio-net device
>> (index 0) which causes me to decrement dev_index and hand index 1 out
>> again to the next device. This leaves us with virtio-blk (index 1)
>> and the new device with index 1, which is not unique.
>>
>> Perhaps I missed a constraint which prevents this from occurring?
> I believe the address is assigned up to 1f so using u32 value
> to track use/free. This should make the code immune to the scenario.
> Would this be adequate?

This issue was also brought up by Jens Axboe on your other patch for
virtio block devices.  He suggested using idr.  I think that would be
a nicer solution than a u32 bitfield.

I'm not sure where you got 0x1f from but that seems like an artifical
limitation.  Nothing should stop us from having more virtio devices.

Stefan