On Sat, Jul 10, 2010 at 01:55:26AM +0200, Linus Walleij wrote:

This looks mostly good to me (comments below) but will have as
a side effect that all clock frameworks on all platforms using PrimeCells
will need to be initialized *before* the PrimeCell devices are added.

No it doesn't - it needs the clock API to be initialized before the
first driver is probed.  As drivers are registered during the initcalls,
later than postcore level

quoted

- ? ? ? id = amba_lookup(pcdrv->id_table, pcdev);
+ ? ? ? ? ? ? ? ret = pcdrv->probe(pcdev, id);
+ ? ? ? ? ? ? ? if (ret == 0)
+ ? ? ? ? ? ? ? ? ? ? ? break;

So this means you always turn on the block clock at
probe, and then leave it on if the probe was successful.

Correct, which is what drivers currently expect.  This is also why
two additional amba_* callbacks are supplied - if it is appropriate
for a driver to disable their bus clock after being probed, they can
do that but they have to participate in that.

Basically, they have to call amba_bus_clk_disable() on successful
probe, and then amba_bus_clk_enable() each time that they want to
access their registers, followed by an amba_bus_clk_disable() when
they've finished.  Finally, they must call amba_bus_clk_enable()
in their remove method and avoid amba_bus_clk_disable() before
returning.

For drivers that need to conserve power, also the block
clock has to be gated when the device is not in use,
sometimes since that affects some hardware that keep
track of the usecount of the clock tree for a larger region of
silicon, so the power savings can be drastic.

You can't have the core code doing that.  If you unconditionally turn
the bus clock off after probe, what happens when a driver receives an
interrupt and tries to access its registers?

Hint: the core code can't know that the driver has registered an IRQ
handler.