Re: [PATCH v3 2/9] Documentation: KVM: s390: Describe KVM_S390_MEMOP_F_CMPXCHG

From: Janis Schoetterl-Glausch
Date: Tue Nov 22 2022 - 08:11:07 EST


On Tue, 2022-11-22 at 08:47 +0100, Thomas Huth wrote:
> On 17/11/2022 23.17, Janis Schoetterl-Glausch wrote:
> > Describe the semantics of the new KVM_S390_MEMOP_F_CMPXCHG flag for
> > absolute vm write memops which allows user space to perform (storage key
> > checked) cmpxchg operations on guest memory.
> >
> > Signed-off-by: Janis Schoetterl-Glausch <scgl@xxxxxxxxxxxxx>
> > ---
> ...
> > Supported flags:
> > * ``KVM_S390_MEMOP_F_CHECK_ONLY``
> > * ``KVM_S390_MEMOP_F_SKEY_PROTECTION``
> > + * ``KVM_S390_MEMOP_F_CMPXCHG``
> > +
> > +The semantics of the flags common with logical acesses are as for logical
> > +accesses.
> > +
> > +For write accesses, the KVM_S390_MEMOP_F_CMPXCHG might be supported.
>
> I'd maybe merge this with the last sentence:
>
> For write accesses, the KVM_S390_MEMOP_F_CMPXCHG flag is supported if
> KVM_CAP_S390_MEM_OP_EXTENSION has bit 1 (i.e. bit with value 2) set.

Ok.
>
> ... and speaking of that, I wonder whether it's maybe a good idea to
> introduce some #defines for bit 1 / value 2, to avoid the confusion ?

Not sure, I don't feel it's too complicated. Where would you define it?
Next to the mem_op struct? KVM_S390_MEMOP_EXTENSION_CAP_CMPXCHG?
>
> > +In this case, instead of doing an unconditional write, the access occurs only
> > +if the target location contains the "size" byte long value pointed to by
> > +"old_p". This is performed as an atomic cmpxchg.
>
> I had to read the first sentence twice to understand it ... maybe it's
> easier to understand if you move the "size" part to the second sentence:
>
> In this case, instead of doing an unconditional write, the access occurs
> only if the target location contains value pointed to by "old_p". This is
> performed as an atomic cmpxchg with the length specified by the "size"
> parameter.
>
> ?

Ok.
>
> > "size" must be a power of two
> > +up to and including 16.
> > +The value at the target location is written to the location "old_p" points to.
>
> IMHO something like this would be better:
>
> The value at the target location is replaced with the value from the
> location that "old_p" points to.

I'm trying to say the opposite :).
I went with this:

If the exchange did not take place because the target value doesn't match the
old value, KVM_S390_MEMOP_R_NO_XCHG is returned.
In this case the value "old_addr" points to is replaced by the target value.
>
> > +If the exchange did not take place because the target value doesn't match the
> > +old value KVM_S390_MEMOP_R_NO_XCHG is returned.
> > +The KVM_S390_MEMOP_F_CMPXCHG flag is supported if KVM_CAP_S390_MEM_OP_EXTENSION
> > +has bit 1 (i.e. bit with value 2) set.
>
> Thomas
>
> PS: Please take my suggestions with a grain of salt ... I'm not a native
> speaker either.
>