Hi,

Dne ?etrtek, 18. januar 2018 ob 11:58:41 CET je Maxime Ripard napisal(a):

Hi,

On Wed, Jan 17, 2018 at 09:14:11PM +0100, Jernej Skrabec wrote:

quoted

This commit changes formula from this:

Freq = (parent_freq * N * K) / (M * P)

to this:

Freq = (parent_freq / M) * N * K / P

This improves situation when N is in the range 1-255. PLL parent clock
is almost always 24 MHz, which means that for N >= 180 original formula
overflows and result becomes useless. Situation can be improved if M is
used as predivider as it can be seen in the second formula. That way at
least M > 1 is considered, but it still leaves small gap for wrong result
when M = 1 and N >= 180.

Using M as predivider shouldn't cause any issue, because it is in range
1-4 at most, so there is no or only minimal rounding error.

Signed-off-by: Jernej Skrabec <redacted>

I'd really prefer to stick to the formula documented and that we've
used so far. NKMP clocks are most notably used for the CPU PLLs and
I've debugged way too many cpufreq bugs already :)

What about using long long types for the parent * n * k result?

Yes, using long long is the best possible solution and covers all cases 
whereas this patch does not.

I thought that do_div() would cause a lot of overhead, but I noticed that it's 
not big if both numbers fit in 32 bit, which in our case is true most of the 
time.

I will make a helper function for calculating rate, since using long long 
needs more than one line of code.

Best regards,
Jernej