The device tree should describe the chips (or chip-like subblocks) in
the system, but it generally does not describe individual registers --
it should identify, rather than describe, a programming interface.

This has not been the case with the QorIQ clockgen nodes.  The
knowledge of what each bit setting of CLKCnCSR means is encoded in
three places (binding, pll node, and mux node), and the last also needs
to know which options are valid on a particular chip.  All three of
these locations are considered stable ABI, making it difficult to fix
mistakes (of which I have found several), much less refactor the
abstraction to be able to address problems, limitations, or new chips.

Under the current binding, a pll clock specifier of 2 means that the
PLL is divided by 4 -- and the driver implements this, unless there
happen to be four clock-output-names rather than 3, in which case it
interprets it as PLL divided by 3.  This does not appear in the binding
documentation at all.  That hack is now considered stable ABI.

The current device tree nodes contain errors, such as saying that
T1040 can set a core clock to PLL/4 when only PLL and PLL/2 are options.
The current binding also ignores some restrictions on clock selection,
such as p5020's requirement that if a core uses the "wrong" PLL, that
PLL must be clocked lower than the "correct" PLL and be at most 80% of
the rated CPU frequency.

Possibly because of the lack of the ability to express such nuance in
the binding, some valid options are omitted from the device trees, such
as the ability on p4080 to run cores 0-3 from PLL3 and cores 4-7 from
PLL1 (again, only if they are at most 80% of rated CPU frequency).
This omission, combined with excessive caution in the cpufreq driver
(addressed in a subsequent patch), means that currently on a 1500 MHz
p4080 with typical PLL configuration, cpufreq can lower the frequency
to 1200 MHz on half the CPUs and do nothing on the others.  With this
patchset, all CPUs can be lowered to 1200 MHz on a rev2 p4080, and on a
rev3 p4080 half can be lowered to 750 MHz and the other half to 600
MHz.

The current binding only deals with CPU clocks.  To describe FMan in
the device tree, we need to describe its clock.  Some chips have
additional muxes that work like the CPU muxes, but are not described in
the device tree.  Others require inspecting the Reset Control Word to
determine which PLL is used.  Rather than continue to extend this mess,
replace it.  Have the driver bind to the chip-specific clockgen
compatible, and keep the detailed description of quirky chip variations
in the driver, where it can be easily fixed, refactored, and extended.

Older device trees will continue to work (including a workaround for
old ls1021a device trees that are missing compatible and reg in the
clockgen node, which even the old binding required).  The pll/mux
details in old device trees will be ignored, but "clocks" properties
pointing at the old nodes will still work, and be directed at the
corresponding new clock.
Signed-off-by: Scott Wood <scottwood@freescale.com>
Acked-by: Stephen Boyd <sboyd@codeaurora.org>