Re: [PATCH 6/7] KVM: arm64: Change return type of kvm_vm_ioctl_mte_copy_tags() to "int"

[Thomas Huth]

On 08/02/2023 12.51, Steven Price wrote:
> On 08/02/2023 08:49, Cornelia Huck wrote:
> > On Wed, Feb 08 2023, Gavin Shan <gshan@xxxxxxxxxx> wrote:
> >
> > > On 2/7/23 9:09 PM, Thomas Huth wrote:
> > > > Oh, drat, I thought I had checked all return statements ... this must have fallen through the cracks, sorry!
> > > >
> > > > Anyway, this is already a problem now: The function is called from kvm_arch_vm_ioctl() (which still returns a long), which in turn is called from kvm_vm_ioctl() in virt/kvm/kvm_main.c. And that functions stores the return value in an "int r" variable. So the upper bits are already lost there.
>
> Sorry about that, I was caught out by kvm_arch_vm_ioctl() returning long...

That's why I'm trying to fix that return type mess with my series, to avoid 
such problems in the future :-)

> > > > Also, how is this supposed to work from user space? The normal "ioctl()" libc function just returns an "int" ? Is this ioctl already used in a userspace application somewhere? ... at least in QEMU, I didn't spot it yet...
> >
> > We will need it in QEMU to implement migration with MTE (the current
> > proposal simply adds a migration blocker when MTE is enabled, as there
> > are various other things that need to be figured out for this to work.)
> > But maybe other VMMs already use it (and have been lucky because they
> > always dealt with shorter lengths?)
> >
> > > The ioctl command KVM_ARM_MTE_COPY_TAGS was merged recently and not used
> > > by QEMU yet. I think struct kvm_arm_copy_mte_tags::length needs to be
> > > '__u32' instead of '__u64' in order to standardize the return value.
> > > Something like below. Documentation/virt/kvm/api.rst::section-4.130
> > > needs update accordingly.
> > >
> > >      struct kvm_arm_copy_mte_tags {
> > >           __u64 guest_ipa;
> > >           __u32 pad;
> > >           __u32 length;
> > >           void __user *addr;
> > >           __u64 flags;
> > >           __u64 reserved[2];
> > >     };
> >
> > Can we do this in a more compatible way, as we are dealing with an API?
> > Like returning -EINVAL if length is too big?
>
> I agree the simplest fix for the problem is simply to reject any
> lengths>INT_MAX:
>
> diff --git a/arch/arm64/kvm/guest.c b/arch/arm64/kvm/guest.c
> index cf4c495a4321..94aed7ce85c4 100644
> --- a/arch/arm64/kvm/guest.c
> +++ b/arch/arm64/kvm/guest.c
> @@ -1032,6 +1032,13 @@ long kvm_vm_ioctl_mte_copy_tags(struct kvm *kvm,
>   	if (copy_tags->flags & ~KVM_ARM_TAGS_FROM_GUEST)
>   		return -EINVAL;
>
> +	/*
> +	 * ioctl returns int, so lengths above INT_MAX cannot be
> +	 * represented in the return value
> +	 */
> +	if (length > INT_MAX)
> +		return -EINVAL;
> +
>   	if (length & ~PAGE_MASK || guest_ipa & ~PAGE_MASK)
>   		return -EINVAL;
>
> This could also be fixed in a useable way by including a new flag which
> returns the length in an output field of the ioctl structure. I'm
> guessing a 2GB limit would be annoying to work around.

I agree that checking for length > INT_MAX is likely the best thing to do 
here right now. I'll add that in v2 of my series.

But actually, this might even be a good thing from another point of view (so 
I'm not sure whether your idea with the flag should really be pursued): The 
code here takes a mutex and then runs a while loop that depends on the 
length - which could cause the lock to be held for a rather long time if 
length is a 64-bit value. Forcing the user space to limit the length here 
could help to avoid taking the lock for too long.

 Thomas