On 01/19/2018 11:16 AM, David Hildenbrand wrote:
> On 19.01.2018 11:11, Heiko Carstens wrote:
>> On Thu, Jan 18, 2018 at 09:45:03PM +0100, David Hildenbrand wrote:
>>> Actually, the problem is there are no atomic byte-based operations on
>>> s390x without Interlocked-Access-Facility 2.
>>>
>>> Even __sync_fetch_and_or(&gisa->ipm, value) falls back to a
>>> Compare-And-Swap loop. And Compare-And-Swap also operates at least on 32bit.
>>>
>>> So I assume there isn't too much we can do about it. As storage
>>> locations following the u8 are also written - but in an atomic matter,
>>> it should in general not matter.
>>>
>>> But can we avoid starting the bitmap at the beginning of the gisa?
>>>
>>> What about something like this:
>>>
>>> +void kvm_s390_gisa_set_ipm_gisc(struct kvm_s390_gisa *gisa, u8 gisc)
>>> +{
>>> + set_bit_inv(gisc, (unsigned long *) &gisa->ipm);
>>> +}
>>
>> set_bit_inv() may use a csg instruction which requires an 8 byte alignment
>> of the operand. What you propose would crash immediately.
>>
>> The code written by Michael is fine as-is.
>>
>
> That's unfortunate... and still looks hacky to me :) But if it works ...
I had looked at the same place, but the alignment thing makes this really
hard and I did not find a better solution.
