Re: [kvm-unit-tests PATCH v4 4/6] s390x: Add library functions for exiting from snippet

[Janosch Frank <frankja@linux.ibm.com>]

On 10/21/24 10:21 AM, Nico Boehr wrote:
> Quoting Janosch Frank (2024-10-18 14:53:34)
>> On 10/18/24 12:56 PM, Nico Boehr wrote:
>>> Quoting Janosch Frank (2024-10-18 10:02:37)
>>> [...]
>>>>> +static inline uint64_t snippet_get_force_exit_value(struct vm *vm)
>>>>> +{
>>>>> +     struct kvm_s390_sie_block *sblk = vm->sblk;
>>>>> +
>>>>> +     assert(snippet_is_force_exit_value(vm));
>>>>> +
>>>>> +     return vm->save_area.guest.grs[sblk_ip_as_diag(sblk).r_1];
>>>>> +}
>>>>
>>>> The cpu address parameter for 9C is 16 bit.
>>>> While we could make it 64 bit for snippets I don't see a reason to do
>>>> so. The 16 bits are enough to indicate something to the host which can
>>>> then go and fetch memory for more data.
>>>
>>> Mh, how exactly would you "fetch memory"? That requires knowledge on where
>>> things are in guest memory which can be painful to figure out from the
>>> host.
>>>
>>> I've found it useful to be able to pass a pointer from guest to host. Maybe
>>> a diag500 is the better option? gr2 contains the cookie which is a 64-bit
>>> value - see Linux' Documentation/virt/kvm/s390/s390-diag.rst.
>>>
>>> P.S. Did I miss the part in the docs where the 16-bit restriction of 9c is
>>> documented or is it missing?
>>
>> For ASM snippets addresses 0x2000 to 0x4000 are a free area.
>> For C snippets that area is the stack.
>> The 16 bits should be good enough to point into that area.
> 
> Actually, it's currently just enough to point into the stack (snippet stack
> is 64K)... also requires additional fiddling in the host to figure out the
> complete address. Probably also in the guest, if you do it r15-relative
> (can't think of a different solution right now).

I think you looked into the wrong lds; the unit tests have 64k, the 
snippets don't.
The snippet stack is from 0x4000 to 0x2000 whereas the unit test stack 
is located behind the binary.

I have no recollection about why I placed it there, though the mail 
archive may or may not have answers.


> 
>> If the snippet requires a lot of memory then you can use constant
>> addresses which are way over the snippet binary or just store a 64 bit
>> address in a "free" lowcore location.
> 
> Would you prefer any of those over the diag500 solution?

I think the diag500 is a bit cleaner and more future proof.