by Matthew Rosato

[permalink] [raw]

Subject: Re: [PATCH] KVM: s390: pci: Hook to access KVM lowlevel from VFIO

On 8/18/22 10:20 AM, Niklas Schnelle wrote:
> On Thu, 2022-08-18 at 09:33 -0400, Matthew Rosato wrote:
>> On 8/18/22 6:23 AM, Pierre Morel wrote:
>>> We have a cross dependency between KVM and VFIO.
>>
>> maybe add something like 'when using s390 vfio_pci_zdev extensions for PCI passthrough'
>>
>>> To be able to keep both subsystem modular we add a registering
>>> hook inside the S390 core code.
>>>
>>> This fixes a build problem when VFIO is built-in and KVM is built
>>> as a module or excluded.
>>
>> s/or excluded//
>>
>> There's no problem when KVM is excluded, that forces CONFIG_VFIO_PCI_ZDEV_KVM=n because of the 'depends on S390 && KVM'.
>>
>>> Reported-by: Randy Dunlap <[email protected]>
>>> Reported-by: kernel test robot <[email protected]>
>>> Signed-off-by: Pierre Morel <[email protected]>
>>> Fixes: 09340b2fca007 ("KVM: s390: pci: add routines to start/stop inter..")
>>> Cc: <[email protected]>
>>> ---
>>> arch/s390/include/asm/kvm_host.h | 17 ++++++-----------
>>> arch/s390/kvm/pci.c | 10 ++++++----
>>> arch/s390/pci/Makefile | 2 ++
>>> arch/s390/pci/pci_kvm_hook.c | 11 +++++++++++
>>> drivers/vfio/pci/vfio_pci_zdev.c | 8 ++++++--
>>> 5 files changed, 31 insertions(+), 17 deletions(-)
>>> create mode 100644 arch/s390/pci/pci_kvm_hook.c
>>>
>>> diff --git a/arch/s390/include/asm/kvm_host.h b/arch/s390/include/asm/kvm_host.h
>>> index f39092e0ceaa..8312ed9d1937 100644
>>> --- a/arch/s390/include/asm/kvm_host.h
>>> +++ b/arch/s390/include/asm/kvm_host.h
>
> I added Janosch as second S390 KVM maintainer in case he wants to chime
> in.
>
>>> @@ -1038,16 +1038,11 @@ 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
>>> -int kvm_s390_pci_register_kvm(struct zpci_dev *zdev, struct kvm *kvm);
>>> -void kvm_s390_pci_unregister_kvm(struct zpci_dev *zdev);
>>> -#else
>>> -static inline int kvm_s390_pci_register_kvm(struct zpci_dev *dev,
>>> - struct kvm *kvm)
>>> -{
>>> - return -EPERM;
>>> -}
>>> -static inline void kvm_s390_pci_unregister_kvm(struct zpci_dev *dev) {}
>>> -#endif
>>> +struct kvm_register_hook {
>>
>> Nit: zpci_kvm_register_hook ? Just to make it clear it's for zpci.
>
> Hmm, I guess one could re-use the same struct for another such KVM
> dependency but I lean towards the same thinking as Matt, for now this
> is for zpci so stay specific we can always generalize later.

Yes, let's keep this zpci-specific.

>
> Nit: For me hook and register together sound a bit redudant, maybe
> "zpci_kvm_register"? Also question for Matt as a native speaker, should
> it rather be "registration" when used as a noun here?
>

Maybe just drop the 'register'. If there is a need for a 3rd function later, for example, it might not be related to registration.

e.g. struct kvm_zpci_hook {
...
};

extern struct kvm_zpci_hook zpci_kvm;

>
>>
>>> + int (*kvm_register)(void *opaque, struct kvm *kvm);
>>> + void (*kvm_unregister)(void *opaque);
>
> I do wonder if this needs to be opague "struct zpci_dev" should be
> defined even if CONFIG_PCI is unset.
>
>
>>> +};
>>> +
>>> +extern struct kvm_register_hook kvm_pci_hook;
>>
>> Nit: kvm_zpci_hook ?
>
> Analogous to zpci_kvm_regist(er|ration) I would call the variable
> simply zpci_kvm i.e. the type is a registration and the variable is the
> instance of it that links zpci and kvm.
>

Yeah, see above.

>>
>>>
>>> #endif
>>> diff --git a/arch/s390/kvm/pci.c b/arch/s390/kvm/pci.c
>>> index 4946fb7757d6..e173fce64c4f 100644
>>> --- a/arch/s390/kvm/pci.c
>>> +++ b/arch/s390/kvm/pci.c
>>> @@ -431,8 +431,9 @@ static void kvm_s390_pci_dev_release(struct zpci_dev *zdev)
>>> * available, enable them and let userspace indicate whether or not they will
>>> * be used (specify SHM bit to disable).
>>> */
>>> -int kvm_s390_pci_register_kvm(struct zpci_dev *zdev, struct kvm *kvm)
>>> +static int kvm_s390_pci_register_kvm(void *opaque, struct kvm *kvm)
>>> {
>>> + struct zpci_dev *zdev = opaque;
>>> int rc;
>>>
>>> if (!zdev)
>>> @@ -510,10 +511,10 @@ int kvm_s390_pci_register_kvm(struct zpci_dev *zdev, struct kvm *kvm)
>>> kvm_put_kvm(kvm);
>>> return rc;
>>> }
>>> -EXPORT_SYMBOL_GPL(kvm_s390_pci_register_kvm);
>>>
>>> -void kvm_s390_pci_unregister_kvm(struct zpci_dev *zdev)
>>> +static void kvm_s390_pci_unregister_kvm(void *opaque)
>>> {
>>> + struct zpci_dev *zdev = opaque;
>>> struct kvm *kvm;
>>>
>>> if (!zdev)
>>> @@ -566,7 +567,6 @@ void kvm_s390_pci_unregister_kvm(struct zpci_dev *zdev)
>>>
>>> kvm_put_kvm(kvm);
>>> }
>>> -EXPORT_SYMBOL_GPL(kvm_s390_pci_unregister_kvm);
>>>
>>> void kvm_s390_pci_init_list(struct kvm *kvm)
>>> {
>>> @@ -678,6 +678,8 @@ int kvm_s390_pci_init(void)
>>>
>>> spin_lock_init(&aift->gait_lock);
>>> mutex_init(&aift->aift_lock);
>>> + kvm_pci_hook.kvm_register = kvm_s390_pci_register_kvm;
>>> + kvm_pci_hook.kvm_unregister = kvm_s390_pci_unregister_kvm;
>>>
>>> return 0;
>>> }
>>> diff --git a/arch/s390/pci/Makefile b/arch/s390/pci/Makefile
>>> index bf557a1b789c..c02dbfb415d9 100644
>>> --- a/arch/s390/pci/Makefile
>>> +++ b/arch/s390/pci/Makefile
>>> @@ -7,3 +7,5 @@ obj-$(CONFIG_PCI) += pci.o pci_irq.o pci_dma.o pci_clp.o pci_sysfs.o \
>>> pci_event.o pci_debug.o pci_insn.o pci_mmio.o \
>>> pci_bus.o
>>> obj-$(CONFIG_PCI_IOV) += pci_iov.o
>>> +
>>> +obj-y += pci_kvm_hook.o
>>
>> I guess it doesn't harm anything to add this unconditionally, but I think it would also be OK to just include this in the CONFIG_PCI list - vfio_pci_zdev and arch/s390/kvm/pci all rely on CONFIG_PCI via CONFIG_VFIO_PCI_ZDEV_KVM which implies PCI via VFIO_PCI.
>>
>>> diff --git a/arch/s390/pci/pci_kvm_hook.c b/arch/s390/pci/pci_kvm_hook.c
>>> new file mode 100644
>>> index 000000000000..9d8799b72dbf
>>> --- /dev/null
>>> +++ b/arch/s390/pci/pci_kvm_hook.c
>>> @@ -0,0 +1,11 @@
>>> +// SPDX-License-Identifier: GPL-2.0-only
>>> +/*
>>> + * VFIO ZPCI devices support
>>> + *
>>> + * Copyright (C) IBM Corp. 2022. All rights reserved.
>>> + * Author(s): Pierre Morel <[email protected]>
>>> + */
>>> +#include <linux/kvm_host.h>
>>> +
>>> +struct kvm_register_hook kvm_pci_hook;
>>> +EXPORT_SYMBOL_GPL(kvm_pci_hook);
>>
>> Following the comments above, zpci_kvm_register_hook, kvm_zpci_hook ?
>>
>> I'm not sure if this really needs to be in a separate file or if it could just go into arch/s390/pci.c with the zpci_aipb -- If going the route of a separate file, up to Niklas whether he wants this under the S390 PCI maintainership or added to the list for s390 vfio-pci like arch/kvm/pci* and vfio_pci_zdev.
>
> I'm fine with a separate file, pci.c is long enough as it is. I also
> don't have a problem with having it maintained as part of S390 PCI but
> logically I think it does fall more under arch/kvm/pci* so one could
> argue it should be added in the MAINTAINERS file in that section.
> If you change the struct name as I proposed above I would probably go
> with "pci_kvm_register.c"

OK, no problem with me for a separate file then, or maintaining said file. But I guess not pci_kvm_register.c per my comments above

>
>>
>>> diff --git a/drivers/vfio/pci/vfio_pci_zdev.c b/drivers/vfio/pci/vfio_pci_zdev.c
>>> index e163aa9f6144..3b7a707e2fe5 100644
>>> --- a/drivers/vfio/pci/vfio_pci_zdev.c
>>> +++ b/drivers/vfio/pci/vfio_pci_zdev.c
>>> @@ -151,7 +151,10 @@ 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);
>>> + if (kvm_pci_hook.kvm_register)
>>> + return kvm_pci_hook.kvm_register(zdev, vdev->vdev.kvm);
>>> +
>>> + return -ENOENT;
>>> }
>>>
>>> void vfio_pci_zdev_close_device(struct vfio_pci_core_device *vdev)
>>> @@ -161,5 +164,6 @@ void vfio_pci_zdev_close_device(struct vfio_pci_core_device *vdev)
>>> if (!zdev || !vdev->vdev.kvm)
>>> return;
>>>
>>> - kvm_s390_pci_unregister_kvm(zdev);
>>> + if (kvm_pci_hook.kvm_unregister)
>>> + return kvm_pci_hook.kvm_unregister(zdev);
>>
>> No need for the return here, this is a void function calling a void function.
>>
>>
>> Overall, this looks good to me and survives a series of compile and device passthrough tests on my end, just a matter of a few of these minor comments above. Thanks for tackling this Pierre!
>
> Yes I agree, overall this looks good to me though I'm admittedly not
> very knowledgable about how to best handle module dependencies like
> this. It does look cleaner than the symbol_get() alternative we
> discussed.
>
>