Re: [PATCH v2 05/11] KVM: s390: Add optional storage key checking to MEMOP IOCTL

[Christian Borntraeger <borntraeger@linux.ibm.com>]

Am 09.02.22 um 11:39 schrieb Janis Schoetterl-Glausch:
> On 2/9/22 11:08, Christian Borntraeger wrote:
>>
>>
>> Am 09.02.22 um 11:01 schrieb Janis Schoetterl-Glausch:
>>> On Wed, 2022-02-09 at 10:08 +0100, Christian Borntraeger wrote:
>>>>
>>>> Am 09.02.22 um 09:49 schrieb Janis Schoetterl-Glausch:
>>>>> On 2/9/22 08:34, Christian Borntraeger wrote:
>>>>>> Am 07.02.22 um 17:59 schrieb Janis Schoetterl-Glausch:
>>>>>>> User space needs a mechanism to perform key checked accesses when
>>>>>>> emulating instructions.
>>>>>>>
>>>>>>> The key can be passed as an additional argument.
>>>>>>> Having an additional argument is flexible, as user space can
>>>>>>> pass the guest PSW's key, in order to make an access the same way the
>>>>>>> CPU would, or pass another key if necessary.
>>>>>>>
>>>>>>> Signed-off-by: Janis Schoetterl-Glausch <scgl@linux.ibm.com>
>>>>>>> Acked-by: Janosch Frank <frankja@linux.ibm.com>
>>>>>>> Reviewed-by: Claudio Imbrenda <imbrenda@linux.ibm.com>
>>>>>>> ---
>>>>>>>      arch/s390/kvm/kvm-s390.c | 49 +++++++++++++++++++++++++++++++---------
>>>>>>>      include/uapi/linux/kvm.h |  8 +++++--
>>>>>>>      2 files changed, 44 insertions(+), 13 deletions(-)
>>>>>>>
>>>>>>> diff --git a/arch/s390/kvm/kvm-s390.c b/arch/s390/kvm/kvm-s390.c
>>>>>>> index cf347e1a4f17..71e61fb3f0d9 100644
>>>>>>> --- a/arch/s390/kvm/kvm-s390.c
>>>>>>> +++ b/arch/s390/kvm/kvm-s390.c
>>>>>>> @@ -32,6 +32,7 @@
>>>>>>>      #include <linux/sched/signal.h>
>>>>>>>      #include <linux/string.h>
>>>>>>>      #include <linux/pgtable.h>
>>>>>>> +#include <linux/bitfield.h>
>>>>>>>        #include <asm/asm-offsets.h>
>>>>>>>      #include <asm/lowcore.h>
>>>>>>> @@ -2359,6 +2360,11 @@ static int kvm_s390_handle_pv(struct kvm *kvm, struct kvm_pv_cmd *cmd)
>>>>>>>          return r;
>>>>>>>      }
>>>>>>>      +static bool access_key_invalid(u8 access_key)
>>>>>>> +{
>>>>>>> +    return access_key > 0xf;
>>>>>>> +}
>>>>>>> +
>>>>>>>      long kvm_arch_vm_ioctl(struct file *filp,
>>>>>>>                     unsigned int ioctl, unsigned long arg)
>>>>>>>      {
>>>>>>> @@ -4687,34 +4693,54 @@ static long kvm_s390_guest_mem_op(struct kvm_vcpu *vcpu,
>>>>>>>                        struct kvm_s390_mem_op *mop)
>>>>>>>      {
>>>>>>>          void __user *uaddr = (void __user *)mop->buf;
>>>>>>> +    u8 access_key = 0, ar = 0;
>>>>>>>          void *tmpbuf = NULL;
>>>>>>> +    bool check_reserved;
>>>>>>>          int r = 0;
>>>>>>>          const u64 supported_flags = KVM_S390_MEMOP_F_INJECT_EXCEPTION
>>>>>>> -                    | KVM_S390_MEMOP_F_CHECK_ONLY;
>>>>>>> +                    | KVM_S390_MEMOP_F_CHECK_ONLY
>>>>>>> +                    | KVM_S390_MEMOP_F_SKEY_PROTECTION;
>>>>>>>      -    if (mop->flags & ~supported_flags || mop->ar >= NUM_ACRS || !mop->size)
>>>>>>> +    if (mop->flags & ~supported_flags || !mop->size)
>>>>>>>              return -EINVAL;
>>>>>>> -
>>>>>>>          if (mop->size > MEM_OP_MAX_SIZE)
>>>>>>>              return -E2BIG;
>>>>>>> -
>>>>>>>          if (kvm_s390_pv_cpu_is_protected(vcpu))
>>>>>>>              return -EINVAL;
>>>>>>> -
>>>>>>>          if (!(mop->flags & KVM_S390_MEMOP_F_CHECK_ONLY)) {
>>>>>>>              tmpbuf = vmalloc(mop->size);
>>>>>>>              if (!tmpbuf)
>>>>>>>                  return -ENOMEM;
>>>>>>>          }
>>>>>>> +    ar = mop->ar;
>>>>>>> +    mop->ar = 0;
>>>>>>
>>>>>> Why this assignment to 0?
>>>>>
>>>>> It's so the check of reserved below works like that, they're all part of the anonymous union.
>>>>
>>>> Ah, I see. This is ugly :-)
>>>
>>> Yes :)
>>>>
>>>>>>> +    if (ar >= NUM_ACRS)
>>>>>>> +        return -EINVAL;
>>>>>>> +    if (mop->flags & KVM_S390_MEMOP_F_SKEY_PROTECTION) {
>>>>>>> +        access_key = mop->key;
>>>>>>> +        mop->key = 0;
>>>>>>
>>>>>> and this? I think we can leave mop unchanged.
>>>>>>
>>>>>> In fact, why do we add the ar and access_key variable?
>>>>>> This breaks the check from above (if (mop->flags & ~supported_flags || mop->ar >= NUM_ACRS || !mop->size))  into two checks
>>>>>> and it will create a memleak for tmpbuf.
>>>>>
>>>>> I can move the allocation down, goto out or get rid of the reserved check and keep everything as before.
>>>>> First is simpler, but second makes handling that case more explicit and might help in the future.
>>>>
>>>> Maybe add a reserved_02 field in the anon struct and check this for being zero and get rid of the local variables?
>>>
>>> I think that would require us adding new fields in the struct by putting them in a union with reserved_02 and so on,
>>> which could get rather messy.
>>
>> I think it is fine to rename reserved_02. Maybe rename that to dont_use_02 ?
> 
> I don't know what kind of stability guarantees we give here, since it can only happen when recompiling with
> a new header. dont_use is a lot better than reserved here, after all we tell user space to set
> reserved bytes to 0, using reserved_02 to do that would be quite handy and therefore likely.
> 
> The question is also what semantic we want for the check.
> The way it works right now, user space also needs to set unused fields to 0, e.g. key if the flag is not set.
> At least this is the case for the vm memop, the vcpu memop cannot do that because of backward compatibility.

As an alternative just remove the check for reserved == 0 and do that later on as an add-on patch?