[PATCH v3] powerpc/powernv: Use darn instr for random_seed on p9

[Michael Ellerman]

Matt Brown <matthew.brown.dev at gmail.com> writes:
> On Sat, Aug 5, 2017 at 3:06 AM, Tyrel Datwyler <tyreld at linux.vnet.ibm.com> wrote:
>> On 08/03/2017 06:12 PM, Matt Brown wrote:
>>> @@ -135,8 +152,9 @@ static __init int rng_create(struct device_node *dn)
>>>
>>>  static __init int rng_init(void)
>>>  {
>>> +     unsigned long darn_test;
>>>       struct device_node *dn;
>>> -     int rc;
>>> +     int rc, i;
>>>
>>>       for_each_compatible_node(dn, NULL, "ibm,power-rng") {
>>>               rc = rng_create(dn);
>>> @@ -150,6 +168,21 @@ static __init int rng_init(void)
>>>               of_platform_device_create(dn, NULL, NULL);
>>>       }
>>>
>>> +     if (cpu_has_feature(CPU_FTR_ARCH_300)) {
>>> +             for (i = 0; i < 10; i++) {
>>> +                     if (powernv_get_random_darn(&darn_test)) {
>>> +                             ppc_md.get_random_seed =
>>> +                                     powernv_get_random_darn;
>>> +                             break;
>>
>> If you return directly here you can avoid the (i == 9) conditional every iteration of the
>> loop by moving the pr_warn to just outside the loop.
>
> That's true, although it is very unlikely for the
> powernv_get_random_darn to fail. So in practice we should never reach
> the (i == 9) conditional.
> The loop is more of a backup in the rare case that it does fail.

All true, the runtime overhead is really not an issue. It's more that
testing for loop almost-termination in the loop is a bit gross :)

I think the best solution is to just pull it out into a separate
function, it also avoids adding sometimes unused variables to the outer
function, eg:

static int initialise_darn(void)
{
	unsigned long val;

	if (!cpu_has_feature(CPU_FTR_ARCH_300))
		return -ENODEV;

	for (i = 0; i < 10; i++) {
		if (powernv_get_random_darn(&val)) {
			ppc_md.get_random_seed = powernv_get_random_darn;
			return 0;
		}
	}

	return -EIO;
}

cheers