virtio_pci: Question about virtio_pci kernel module refcnt

virtio_pci: Question about virtio_pci kernel module refcnt

[Zhang Tianci]

Hi,
I noticed that the reference count of the virtio_pci kernel module
in the VM is always 0, even though there are at least two
virtio block devices and one virtiofs device in the VM.

Using rmmod virtio_pci can unload the virtio_pci module,
but this renders the virtio block devices unusable.

I wonder if this result is expected? Why don't the virtio block devices
and virtiofs device hold a reference count to the virtio_pci kernel module?

Re: virtio_pci: Question about virtio_pci kernel module refcnt

[Michael S. Tsirkin]

On Wed, Oct 22, 2025 at 07:21:45PM +0800, Zhang Tianci wrote:
> Hi,
> I noticed that the reference count of the virtio_pci kernel module
> in the VM is always 0, even though there are at least two
> virtio block devices and one virtiofs device in the VM.
> 
> Using rmmod virtio_pci can unload the virtio_pci module,
> but this renders the virtio block devices unusable.

it removes them. you can put it back in and you will
get devices again.

> I wonder if this result is expected? Why don't the virtio block devices
> and virtiofs device hold a reference count to the virtio_pci kernel module?

because they don't have to.

-- 
MST

Re: [External] Re: virtio_pci: Question about virtio_pci kernel module refcnt

[Zhang Tianci]

Hi, thank you for your reply, but I still have a few questions:

On Wed, Oct 22, 2025 at 7:31 PM Michael S. Tsirkin <mst@redhat.com> wrote:
>
> On Wed, Oct 22, 2025 at 07:21:45PM +0800, Zhang Tianci wrote:
> > Hi,
> > I noticed that the reference count of the virtio_pci kernel module
> > in the VM is always 0, even though there are at least two
> > virtio block devices and one virtiofs device in the VM.
> >
> > Using rmmod virtio_pci can unload the virtio_pci module,
> > but this renders the virtio block devices unusable.
>
> it removes them. you can put it back in and you will
> get devices again.

But I can't do it anymore because the rootfs is no longer available,
or rather, the OS is no longer available.

>
> > I wonder if this result is expected? Why don't the virtio block devices
> > and virtiofs device hold a reference count to the virtio_pci kernel module?
>
> because they don't have to.

Is it due to design reasons or implementation reasons?
Is it because the semantics of kernel module refcnt does not include such
logical dependency?

Thanks,
Tianci

Re: [External] Re: virtio_pci: Question about virtio_pci kernel module refcnt

[Jason Wang]

On Thu, Oct 23, 2025 at 10:34 AM Zhang Tianci
<zhangtianci.1997@bytedance.com> wrote:
>
> Hi, thank you for your reply, but I still have a few questions:
>
> On Wed, Oct 22, 2025 at 7:31 PM Michael S. Tsirkin <mst@redhat.com> wrote:
> >
> > On Wed, Oct 22, 2025 at 07:21:45PM +0800, Zhang Tianci wrote:
> > > Hi,
> > > I noticed that the reference count of the virtio_pci kernel module
> > > in the VM is always 0, even though there are at least two
> > > virtio block devices and one virtiofs device in the VM.
> > >
> > > Using rmmod virtio_pci can unload the virtio_pci module,
> > > but this renders the virtio block devices unusable.
> >
> > it removes them. you can put it back in and you will
> > get devices again.
>
> But I can't do it anymore because the rootfs is no longer available,
> or rather, the OS is no longer available.
>
> >
> > > I wonder if this result is expected? Why don't the virtio block devices
> > > and virtiofs device hold a reference count to the virtio_pci kernel module?
> >
> > because they don't have to.
>
> Is it due to design reasons or implementation reasons?
> Is it because the semantics of kernel module refcnt does not include such
> logical dependency?

Note that virito has a bus so it follows the device/driver model so if
I was not wrong the refcnt was handled by the driver core.

Thanks

>
> Thanks,
> Tianci
>

Re: [External] Re: virtio_pci: Question about virtio_pci kernel module refcnt

[Zhang Tianci]

Hi, thanks for your reply

On Fri, Oct 24, 2025 at 9:53 AM Jason Wang <jasowang@redhat.com> wrote:
>
> On Thu, Oct 23, 2025 at 10:34 AM Zhang Tianci
> <zhangtianci.1997@bytedance.com> wrote:
> >
> > Hi, thank you for your reply, but I still have a few questions:
> >
> > On Wed, Oct 22, 2025 at 7:31 PM Michael S. Tsirkin <mst@redhat.com> wrote:
> > >
> > > On Wed, Oct 22, 2025 at 07:21:45PM +0800, Zhang Tianci wrote:
> > > > Hi,
> > > > I noticed that the reference count of the virtio_pci kernel module
> > > > in the VM is always 0, even though there are at least two
> > > > virtio block devices and one virtiofs device in the VM.
> > > >
> > > > Using rmmod virtio_pci can unload the virtio_pci module,
> > > > but this renders the virtio block devices unusable.
> > >
> > > it removes them. you can put it back in and you will
> > > get devices again.
> >
> > But I can't do it anymore because the rootfs is no longer available,
> > or rather, the OS is no longer available.
> >
> > >
> > > > I wonder if this result is expected? Why don't the virtio block devices
> > > > and virtiofs device hold a reference count to the virtio_pci kernel module?
> > >
> > > because they don't have to.
> >
> > Is it due to design reasons or implementation reasons?
> > Is it because the semantics of kernel module refcnt does not include such
> > logical dependency?
>
> Note that virito has a bus so it follows the device/driver model so if
> I was not wrong the refcnt was handled by the driver core.

Let's look at an example:

$ lsmod | grep virtio
virtio_balloon                 20480  0
virtio_console                40960  1
virtiofs                            32768  1  <-  one virtiofs
fuse                                172032  2 virtiofs
virtio_net                         73728  0
net_failover                     20480  1 virtio_net
virtio_blk                         32768  2  <- two block devices
virtio_pci                         24576  0
virtio_pci_legacy_dev     16384  1 virtio_pci
virtio_pci_modern_dev   16384  1 virtio_pci
virtio                                16384  6
virtio_console,virtio_balloon,virtiofs,virtio_pci,virtio_blk,virtio_net
virtio_ring                         49152  6
virtio_console,virtio_balloon,virtiofs,virtio_pci,virtio_blk,virtio_net

There are two virtio_blk devices which are vda and vdb, and one virtiofs device.
We can find the block device and fs instance are holding virtio_blk
and virtiofs' refcnts.
But they don't hold virtio_pci(pci bus) or virtio(virtio bus) refcnts
even though they logically depend on
them.
So I might need to consult the driver core maintainers?(I have cc'd
them in this email.)

Thanks,
Tianci

Re: [External] Re: virtio_pci: Question about virtio_pci kernel module refcnt

[Jason Wang]

On Fri, Oct 24, 2025 at 10:34 AM Zhang Tianci
<zhangtianci.1997@bytedance.com> wrote:
>
> Hi, thanks for your reply
>
> On Fri, Oct 24, 2025 at 9:53 AM Jason Wang <jasowang@redhat.com> wrote:
> >
> > On Thu, Oct 23, 2025 at 10:34 AM Zhang Tianci
> > <zhangtianci.1997@bytedance.com> wrote:
> > >
> > > Hi, thank you for your reply, but I still have a few questions:
> > >
> > > On Wed, Oct 22, 2025 at 7:31 PM Michael S. Tsirkin <mst@redhat.com> wrote:
> > > >
> > > > On Wed, Oct 22, 2025 at 07:21:45PM +0800, Zhang Tianci wrote:
> > > > > Hi,
> > > > > I noticed that the reference count of the virtio_pci kernel module
> > > > > in the VM is always 0, even though there are at least two
> > > > > virtio block devices and one virtiofs device in the VM.
> > > > >
> > > > > Using rmmod virtio_pci can unload the virtio_pci module,
> > > > > but this renders the virtio block devices unusable.
> > > >
> > > > it removes them. you can put it back in and you will
> > > > get devices again.
> > >
> > > But I can't do it anymore because the rootfs is no longer available,
> > > or rather, the OS is no longer available.
> > >
> > > >
> > > > > I wonder if this result is expected? Why don't the virtio block devices
> > > > > and virtiofs device hold a reference count to the virtio_pci kernel module?
> > > >
> > > > because they don't have to.
> > >
> > > Is it due to design reasons or implementation reasons?
> > > Is it because the semantics of kernel module refcnt does not include such
> > > logical dependency?
> >
> > Note that virito has a bus so it follows the device/driver model so if
> > I was not wrong the refcnt was handled by the driver core.
>
> Let's look at an example:
>
> $ lsmod | grep virtio
> virtio_balloon                 20480  0
> virtio_console                40960  1
> virtiofs                            32768  1  <-  one virtiofs
> fuse                                172032  2 virtiofs
> virtio_net                         73728  0
> net_failover                     20480  1 virtio_net
> virtio_blk                         32768  2  <- two block devices
> virtio_pci                         24576  0
> virtio_pci_legacy_dev     16384  1 virtio_pci
> virtio_pci_modern_dev   16384  1 virtio_pci
> virtio                                16384  6
> virtio_console,virtio_balloon,virtiofs,virtio_pci,virtio_blk,virtio_net
> virtio_ring                         49152  6
> virtio_console,virtio_balloon,virtiofs,virtio_pci,virtio_blk,virtio_net
>
> There are two virtio_blk devices which are vda and vdb, and one virtiofs device.
> We can find the block device and fs instance are holding virtio_blk
> and virtiofs' refcnts.
> But they don't hold virtio_pci(pci bus) or virtio(virtio bus) refcnts
> even though they logically depend on
> them.

My understanding is the driver must be unregistered while the device
is being unregistered.

> So I might need to consult the driver core maintainers?(I have cc'd
> them in this email.)
>
> Thanks,
> Tianci
>

Thanks