Hi,

2016-01-13 12:18 GMT+01:00 Maxime Ripard [off-list ref]:

Sorry for the reviewing delay.

No problem at all, thanks for the review!

On Mon, Dec 28, 2015 at 06:31:32PM +0100, Marcus Weseloh wrote:

quoted

This patch fixes some problems in the mod0 clock calculation. It has
the potential to break stuff, as the issues explained below had the
effect that clk_set_rate would always return successfully, sometimes
setting a frequency that is higher than the requested value.

That's actually the expected behaviour of clk_set_rate.

clk_set_rate is supposed to adjust the given clock rate to something
that the clock drivers seems fit. It should only return an error in a
case where you can't change the rate at all (because you didn't pass a
valid struct clk pointer, because changing the rate would violate some
clock flags, etc.). Otherwise, clk_set_rate should succeed.

By returning an error code the clock is higher than the one passed,
you violate that expectation, especially since that is relative to the
clock you passed.

It makes sense in your case to never exceed the given rate, it might
not for a different clock in the tree, or even for a different
instance of the same clock. For example, you could very well have
another case in your system where you should not have rates set that
are below the one given because that would prevent the consumer
device to be usable.

This is why the adjustment is left to the clock driver, and is not
enforced by the framework itself, simply because the framework has no
idea how you want to round your clock rate on that particular clock in
your system.

I understand now, thanks a lot for the good explanation! So my
thinking is wrong for the general case of the clock framework itself,
and that actually makes a lot of sense.

But the clk_factors_determine_rate function in
drivers/clk/sunxi/clk-factors.c works on the assumption that the
returned rate must be less or equal to the requested rate. At least
that is what the code in that function tries to do. That the mod0
factor calculation doesn't check the m and div variables for overflow
undermines the intended behaviour, as it "lies" about the frequencies
that the hardware can support. And for very low frequencies below
80kHz, clk_set_rate does currently return -EINVAL. There are even
cases when it results in a division by zero error, for example if you
request a rate of 94kHz from a 24Mhz parent (24000000 / 94000 =
255,32, rounded up to 256 = 0 on the u8 variable).

Now I'm unsure what to do here... If the clock driver should only
return an error in real error cases and not when the requested
frequency isn't reachable, then clk_factors_determine_rate needs to be
changed as well?

quoted

Code that "accidentally worked" because of this might fail after
applying this patch.

The problems in detail:

1. If a very low frequency is requested from a high parent clock, the
divisors "div" and "calcm" might be > 255. This patch changes the type
of both variables to unsigned int, because the silent cast to u8 will
result in invalid frequencies and register values.

2. The width of the "m" divisor in the clock control registers is only
4 bit, but that limitation is not checked when calculating the divisor
and the resulting frequency. This patch adds a check that m never
exceeds the field width.

3. During a call to clk_set_rate, the sun4i_a10_get_mod0_factors
function is called multiple times: first to find the best parent and
frequency, then again to calculate the p and m divisors, passing the
frequencies returned by the previous call(s). In certain cases
those chained calls do not result in the best frequency choice.

You know the drill by now :)

You're fixing three different issues, please send three different
patches.

Yes, thanks for the reminder :-) I thought about it just as I hit the
"send" button, but it was too late then.

quoted

An example:
parent_rate = 24Mhz, freq = 1.4Mhz results in p=1, m=9, freq=1333333,3333
(which gets rounded down to 1333333).
Calling the function again with parent_rate = 24Mhz and freq = 1333333
results in p=1, m=10, freq=1200000.

Rounding up the returned frequency removes this problem.

Signed-off-by: Marcus Weseloh <redacted>
---
 drivers/clk/sunxi/clk-mod0.c | 7 +++++--
 1 file changed, 5 insertions(+), 2 deletions(-)

diff --git a/drivers/clk/sunxi/clk-mod0.c b/drivers/clk/sunxi/clk-mod0.c
index d167e1e..d03f099 100644
--- a/drivers/clk/sunxi/clk-mod0.c
+++ b/drivers/clk/sunxi/clk-mod0.c

@@ -31,7 +31,8 @@
 static void sun4i_a10_get_mod0_factors(u32 *freq, u32 parent_rate,
                                     u8 *n, u8 *k, u8 *m, u8 *p)
 {
-     u8 div, calcm, calcp;
+     unsigned int div, calcm;
+     u8 calcp;

      /* These clocks can only divide, so we will never be able to achieve
       * frequencies higher than the parent frequency */

@@ -50,8 +51,10 @@ static void sun4i_a10_get_mod0_factors(u32 *freq, u32 parent_rate,
              calcp = 3;

      calcm = DIV_ROUND_UP(div, 1 << calcp);
+     if (calcm > 16)
+             calcm = 16;

-     *freq = (parent_rate >> calcp) / calcm;
+     *freq = DIV_ROUND_UP(parent_rate >> calcp, calcm);

While the two above seems harmless, this one concerns me a bit. Did
you test the various mod0 clock users and made sure that they were
still working as they used to?

No, I didn't do a thorough test, only booted the board with some mod0
users (mmc, spi, ss) and watched them request their frequencies
successfully. But this is an edge case, only affecting certain "weird"
frequencies. And the only effect is that the chosen frequency is not
the optimal one. So maybe I should drop it because it looks too
disruptive for too little gain?

Cheers,

   Marcus