Re: [kvm-unit-tests PATCH v6 3/4] s390x: lib: add SPX and STPX instruction wrapper

[Claudio Imbrenda <imbrenda@linux.ibm.com>]

On Thu, 9 Jan 2020 18:05:42 +0100
Janosch Frank <frankja@linux.ibm.com> wrote:

> On 1/9/20 5:58 PM, Thomas Huth wrote:
> > On 09/01/2020 17.50, Claudio Imbrenda wrote:  
> >> On Thu, 9 Jan 2020 17:43:55 +0100
> >> Thomas Huth <thuth@redhat.com> wrote:
> >>  
> >>> On 09/01/2020 17.16, Claudio Imbrenda wrote:  
> >>>> Add a wrapper for the SET PREFIX and STORE PREFIX instructions,
> >>>> and use it instead of using inline assembly everywhere.
> >>>>
> >>>> Signed-off-by: Claudio Imbrenda <imbrenda@linux.ibm.com>
> >>>> ---
> >>>>  lib/s390x/asm/arch_def.h | 10 ++++++++++
> >>>>  s390x/intercept.c        | 33 +++++++++++++--------------------
> >>>>  2 files changed, 23 insertions(+), 20 deletions(-)
> >>>>
> >>>> diff --git a/lib/s390x/asm/arch_def.h b/lib/s390x/asm/arch_def.h
> >>>> index 1a5e3c6..465fe0f 100644
> >>>> --- a/lib/s390x/asm/arch_def.h
> >>>> +++ b/lib/s390x/asm/arch_def.h
> >>>> @@ -284,4 +284,14 @@ static inline int servc(uint32_t command,
> >>>> unsigned long sccb) return cc;
> >>>>  }
> >>>>  
> >>>> +static inline void spx(uint32_t *new_prefix)    
> >>>
> >>> Looking at this a second time ... why is new_prefix a pointer? A
> >>> normal value should be sufficient here, shouldn't it?  
> >>
> >> no. if you look at the code in the same patch, intercept.c at some
> >> points needs to pass "wrong" pointers to spx and stpx in order to
> >> test them, so this needs to be a pointer
> >>
> >> the instructions themselves expect pointers (base register +
> >> offset)  
> > 
> > Ah, you're right, that "Q" constraint always confuses me... I guess
> > you could do it without pointers when using the "r" constraint, but
> > it's likely better to do it the same way as stpx, so your patch
> > should be fine.  
> 
> Honestly, I'd rather have stpx return a u32 than passing a ptr.

that's what I had done initially, but it doesn't work, see above for
the reasons why we need a pointer

> That's how the kernel does it and is in-line with epswe/lpswe and
> sctlg/lctlg which are already in the library.

the kernel does not need to test wrong addresses.

I could have spx accept an int and stpx return an int, but then
intercept.c would still need some inline assembly for SPX and STPX

> Also, if possible names like set_prefix and store_prefix (or better
> get_prefix) prefix would make it much more readable.

this can be done, but that's not how all the other wrappers are

> >   
> >>>> +{
> >>>> +	asm volatile("spx %0" : : "Q" (*new_prefix) : "memory");
> >>>> +}
> >>>> +
> >>>> +static inline void stpx(uint32_t *current_prefix)
> >>>> +{
> >>>> +	asm volatile("stpx %0" : "=Q" (*current_prefix));
> >>>> +}
> >>>> +    
> > 
> > Reviewed-by: Thomas Huth <thuth@redhat.com>
> >   
> 
>