Re: [PATCH v4 5/7] s390: ap: implement PAPQ AQIC interception in kernel

[Pierre Morel <pmorel@linux.ibm.com>]

On 28/02/2019 21:20, Christian Borntraeger wrote:
> 
> 
> On 22.02.2019 16:29, Pierre Morel wrote:
>> We register the AP PQAP instruction hook during the open
>> of the mediated device. And unregister it on release.
>>
>> In the AP PQAP instruction hook, if we receive a demand to
>> enable IRQs,
>> - we retrieve the vfio_ap_queue based on the APQN we receive
>>    in REG1,
>> - we retrieve the page of the guest address, (NIB), from
>>    register REG2
>> - we the mediated device to use the VFIO pinning infratrsucture
>>    to pin the page of the guest address,
>> - we retrieve the pointer to KVM to register the guest ISC
>>    and retrieve the host ISC
>> - finaly we activate GISA
>>
>> If we receive a demand to disable IRQs,
>> - we deactivate GISA
>> - unregister from the GIB
>> - unping the NIB
>>
>> Signed-off-by: Pierre Morel <pmorel@linux.ibm.com>
>> ---
>>   arch/s390/include/asm/kvm_host.h      |   1 +
>>   drivers/s390/crypto/ap_bus.h          |   1 +
>>   drivers/s390/crypto/vfio_ap_ops.c     | 199 +++++++++++++++++++++++++++++++++-
>>   drivers/s390/crypto/vfio_ap_private.h |   1 +
>>   4 files changed, 199 insertions(+), 3 deletions(-)
>>
>> diff --git a/arch/s390/include/asm/kvm_host.h b/arch/s390/include/asm/kvm_host.h
>> index 49cc8b0..5f3bb8c 100644
>> --- a/arch/s390/include/asm/kvm_host.h
>> +++ b/arch/s390/include/asm/kvm_host.h
>> @@ -720,6 +720,7 @@ struct kvm_s390_cpu_model {
>>   struct kvm_s390_crypto {
>>   	struct kvm_s390_crypto_cb *crycb;
>>   	int (*pqap_hook)(struct kvm_vcpu *vcpu);
>> +	void *vfio_private;
>>   	__u32 crycbd;
>>   	__u8 aes_kw;
>>   	__u8 dea_kw;
>> diff --git a/drivers/s390/crypto/ap_bus.h b/drivers/s390/crypto/ap_bus.h
>> index bfc66e4..323f2aa 100644
>> --- a/drivers/s390/crypto/ap_bus.h
>> +++ b/drivers/s390/crypto/ap_bus.h
>> @@ -43,6 +43,7 @@ static inline int ap_test_bit(unsigned int *ptr, unsigned int nr)
>>   #define AP_RESPONSE_BUSY		0x05
>>   #define AP_RESPONSE_INVALID_ADDRESS	0x06
>>   #define AP_RESPONSE_OTHERWISE_CHANGED	0x07
>> +#define AP_RESPONSE_INVALID_GISA	0x08
>>   #define AP_RESPONSE_Q_FULL		0x10
>>   #define AP_RESPONSE_NO_PENDING_REPLY	0x10
>>   #define AP_RESPONSE_INDEX_TOO_BIG	0x11
>> diff --git a/drivers/s390/crypto/vfio_ap_ops.c b/drivers/s390/crypto/vfio_ap_ops.c
>> index 1b5130a..0196065 100644
>> --- a/drivers/s390/crypto/vfio_ap_ops.c
>> +++ b/drivers/s390/crypto/vfio_ap_ops.c
>> @@ -43,7 +43,7 @@ struct vfio_ap_queue *vfio_ap_get_queue(int apqn, struct list_head *l)
>>   	return NULL;
>>   }
>>   
>> -static int vfio_ap_mdev_reset_queue(struct vfio_ap_queue *q)
>> +int vfio_ap_mdev_reset_queue(struct vfio_ap_queue *q)
>>   {
>>   	struct ap_queue_status status;
>>   	int retry = 20;
>> @@ -75,6 +75,27 @@ static int vfio_ap_mdev_reset_queue(struct vfio_ap_queue *q)
>>   	return -EBUSY;
>>   }
>>   
>> +/**
>> + * vfio_ap_free_irq:
>> + * @q: The vfio_ap_queue
>> + *
>> + * Unpin the guest NIB
>> + * Unregister the ISC from the GIB alert
>> + * Clear the vfio_ap_queue intern fields
>> + */
>> +static void vfio_ap_free_irq(struct vfio_ap_queue *q)
>> +{
>> +	if (!q)
>> +		return;
>> +	if (q->g_pfn)
>> +		vfio_unpin_pages(mdev_dev(q->matrix_mdev->mdev), &q->g_pfn, 1);
>> +	if (q->isc)
>> +		kvm_s390_gisc_unregister(q->matrix_mdev->kvm, q->isc);
>> +	q->nib = 0;
>> +	q->isc = 0;
>> +	q->g_pfn = 0;
>> +}
> 
> Pierre, unless there is some magic, I think we need to free the gisa stuff before kvm exit.
> 
> Imagine a malicious userspace that setups everything fine, but then closes all kvm file
> descriptors but not the vfio file descriptor. This might result in random access to the
> memory that contained the gisa potentially resulting in random memory overwrites.
> 
> the problem is that kvm_destroy_vm calls kvm_arch_destroy_vm(kvm) before it calls
> kvm_destroy_devices(kvm); So we already free the gisa before we do the unregister call.
> 
> What about calling kvm_get_kvm/put from some of the callbacks in the right places.
> 
> Debugging random memory overwrites is a PITA, so we either should document why I cannot
> happen (even with malicious userspace) or simply fix the refcounting.
> 

OK, understood.
I think we can do something simple by using kvm_get/kvm_put, as you 
suggested, in the vfio KVM notifier to ensure the order of the calls and 
also unregister the GISA at this moment.
I will investigate in this direction.

Thanks
Pierre




-- 
Pierre Morel
Linux/KVM/QEMU in Böblingen - Germany