Re: [PATCH] vfio-pci/zdev: require KVM to be built-in

[Matthew Rosato <mjrosato@xxxxxxxxxxxxx>]

On 8/16/22 3:55 AM, Pierre Morel wrote:
> 
> 
> On 8/16/22 08:04, Randy Dunlap wrote:
>> Hi--
>>
>> On 8/15/22 02:43, Pierre Morel wrote:
>>> Thank you Randy for this good catch.
>>> However forcing KVM to be include statically in the kernel when using VFIO_PCI extensions is not a good solution for us I think.
>>>
>>> I suggest we better do something like:
>>>
>>> ----
>>>
>>> diff --git a/arch/s390/include/asm/kvm_host.h b/arch/s390/include/asm/kvm_host.h
>>> index 6287a843e8bc..1733339cc4eb 100644
>>> --- a/arch/s390/include/asm/kvm_host.h
>>> +++ b/arch/s390/include/asm/kvm_host.h
>>> @@ -1038,7 +1038,7 @@ static inline void kvm_arch_vcpu_unblocking(struct kvm_vcpu *vcpu) {}
>>>   #define __KVM_HAVE_ARCH_VM_FREE
>>>   void kvm_arch_free_vm(struct kvm *kvm);
>>>
>>> -#ifdef CONFIG_VFIO_PCI_ZDEV_KVM
>>> +#if defined(CONFIG_VFIO_PCI_ZDEV_KVM) || defined(CONFIG_VFIO_PCI_ZDEV_KVM_MODULE)
>>
>> This all looks good except for the line above.
>> It should be:
>>
>> #if IS_ENABLED(CONFIG_VFIO_PCI_ZDEV_KVM)
>>
>> Thanks.
> 
> Yes, better, thanks.
> How do we do? Should I repost it with reported-by you or do you want to post it?
> 
> Pierre

Thanks for looking into this while I was away.

I think the issue is not just CONFIG_KVM=m && CONFIG_VFIO_PCI_ZDEV_KVM=y -- it also requires CONFIG_VFIO_PCI=y && CONFIG_VFIO_PCI_CORE=y.  This combination results in building in vfio_pci (which tries to link the calls to kvm_s390_pci_register_kvm and kvm_s390_pci_unregister_kvm which is not built in).

However... this tristate + IS_ENABLED(CONFIG_VFIO_PCI_ZDEV_KVM) check in kvm_host.h will not solve the issue.  Rather, due to the #ifdef CONFIG_VFIO_PCI_ZDEV_KVM in include/linux/vfio_pci_core.h, this change will just cause us to never call kvm_s390_pci_register_kvm or kvm_s390_pci_unregister_kvm when CONFIG_VFIO_PCI_ZDEV_KVM=m, effectively treating CONFIG_VFIO_PCI_ZDEV_KVM=m as a 'n' and we don't get the zdev kvm extensions, which I don't think was the intent.

I'm still thinking & am open to other ideas, but one solution to avoiding building in KVM would be to go back to using symbol_get for these 2 interfaces (kvm_s390_pci_register_kvm and kvm_s390_pci_unregister_kvm) as was done in a prior version of this series like virt/kvm/vfio.c does and otherwise leave CONFIG_VFIO_PCI_ZDEV_KVM as-is. 

diff --git a/drivers/vfio/pci/vfio_pci_zdev.c b/drivers/vfio/pci/vfio_pci_zdev.c
index e163aa9f6144..09c2758134c7 100644
--- a/drivers/vfio/pci/vfio_pci_zdev.c
+++ b/drivers/vfio/pci/vfio_pci_zdev.c
@@ -144,6 +144,8 @@ int vfio_pci_info_zdev_add_caps(struct vfio_pci_core_device *vdev,
 int vfio_pci_zdev_open_device(struct vfio_pci_core_device *vdev)
 {
        struct zpci_dev *zdev = to_zpci(vdev->pdev);
+       int (*fn)(struct zpci_dev *zdev, struct kvm *kvm);
+       int rc;
 
        if (!zdev)
                return -ENODEV;
@@ -151,15 +153,30 @@ int vfio_pci_zdev_open_device(struct vfio_pci_core_device *vdev)
        if (!vdev->vdev.kvm)
                return 0;
 
-       return kvm_s390_pci_register_kvm(zdev, vdev->vdev.kvm);
+       fn = symbol_get(kvm_s390_pci_register_kvm);
+       if (!fn)
+               return -EPERM;
+
+       rc = fn(zdev, vdev->vdev.kvm);
+
+       symbol_put(kvm_s390_pci_register_kvm);
+
+       return rc;
 }
 
 void vfio_pci_zdev_close_device(struct vfio_pci_core_device *vdev)
 {
        struct zpci_dev *zdev = to_zpci(vdev->pdev);
+       void (*fn)(struct zpci_dev *zdev);
 
        if (!zdev || !vdev->vdev.kvm)
                return;
 
-       kvm_s390_pci_unregister_kvm(zdev);
+       fn = symbol_get(kvm_s390_pci_unregister_kvm);
+       if (!fn)
+               return;
+
+       fn(zdev);
+
+       symbol_put(kvm_s390_pci_unregister_kvm);
 }