2017-07-11 10:21+0200, Christian Borntraeger:
> On 07/10/2017 11:23 PM, Gleb Fotengauer-Malinovskiy wrote:
> > On Mon, Jul 10, 2017 at 08:43:12PM +0200, Christian Borntraeger wrote:
> >> On 07/10/2017 04:44 PM, Gleb Fotengauer-Malinovskiy wrote:
> >>> This ioctl actually writes to parameter too.
> >>
> >> Maybe rephrase that to:
> >> The kernel does not only read struct kvm_s390_cmma_log for KVM_S390_GET_CMMA_BITS,
> >> it also writes back a return value making this _IOWR instead of _IOW.
> >
> > Ok, see v2.
> >
> >>> Fixes: 4036e387 ("KVM: s390: ioctls to get and set guest storage attributes")
> >>> Signed-off-by: Gleb Fotengauer-Malinovskiy <glebfm@xxxxxxxxxxxx>
> >> Acked-by: Christian Borntraeger <borntraeger@xxxxxxxxxx>
> >>
> >>
> >> Out of curiosity, how did you notice that?
> >
> > I regenerated strace's ioctl lists. It was obvious from the diff that
> > *GET* and *SET* could not be both _IOC_WRITE.
> >
>
> In fact we do have multiple GET/SET ioctls in KVM, where both provide a control
> block that is _IOC_WRITE only. That control block then has an address that will
> be read/written to depending on get/set.
> E.g. look at
> #define KVM_SET_DEVICE_ATTR _IOW(KVMIO, 0xe1, struct kvm_device_attr)
> #define KVM_GET_DEVICE_ATTR _IOW(KVMIO, 0xe2, struct kvm_device_attr)
>
> but as far as I understand, the direction hints only qualify the referenced
> struct and not the side effects. So for KVM_*_DEVICE_ATTR it is correct to have
> IOW for both cases.
>
> But for GET_CMMA we do indeed write back data.
>
> Paolo, Radim,
>
> if we want to fix the direction hint, it would be good to merge this in as soon
> as possible. The new interface was added during this merge window.
Having correct hints would allow us to have one common
copy_from_user/copy_to_user and I think it's not too late to rename it
with the real behavior. Applied for the second merge-window pull
request,
thanks.
