Re: [kvm-unit-tests PATCH v8 09/12] s390x: Library resources for CSS tests

[Thomas Huth <thuth@redhat.com>]

On 09/06/2020 17.01, Pierre Morel wrote:
> 
> 
> On 2020-06-09 09:09, Thomas Huth wrote:
>> On 08/06/2020 10.12, Pierre Morel wrote:
>>> Provide some definitions and library routines that can be used by
> 
> ...snip...
> 
>>> +static inline int ssch(unsigned long schid, struct orb *addr)
>>> +{
>>> +    register long long reg1 asm("1") = schid;
>>> +    int cc;
>>> +
>>> +    asm volatile(
>>> +        "    ssch    0(%2)\n"
>>> +        "    ipm    %0\n"
>>> +        "    srl    %0,28\n"
>>> +        : "=d" (cc)
>>> +        : "d" (reg1), "a" (addr), "m" (*addr)
>>
>> Hmm... What's the "m" (*addr) here good for? %3 is not used in the
>> assembly code?
> 
> addr is %2
> "m" (*addr) means memory pointed by addr is read
> 
>>
>>> +        : "cc", "memory");
>>
>> Why "memory" ? Can this instruction also change the orb?
> 
> The orb not but this instruction modifies memory as follow:
> orb -> ccw -> data
> 
> The CCW can be a READ or a WRITE instruction and the data my be anywhere
> in memory (<2G)
> 
> A compiler memory barrier is need to avoid write instructions started
> before the SSCH instruction to occur after for a write
> and memory read made after the instruction to be executed before for a
> read.

Ok, makes sense now, thanks!

>>> +static inline int msch(unsigned long schid, struct schib *addr)
>>> +{
>>> +    register unsigned long reg1 asm ("1") = schid;
>>> +    int cc;
>>> +
>>> +    asm volatile(
>>> +        "    msch    0(%3)\n"
>>> +        "    ipm    %0\n"
>>> +        "    srl    %0,28"
>>> +        : "=d" (cc), "=m" (*addr)
>>> +        : "d" (reg1), "a" (addr)
>>
>> I'm not an expert with these IO instructions, but this looks wrong to me
>> ... Is MSCH reading or writing the SCHIB data?
> 
> MSCH is reading the SCHIB data in memory.

So if it is reading, you don't need the  "=m" (*addr) in the output
list, do you? You should rather use "m" (*addr) in the input list instead?

 Thomas