On Fri, Jan 21, 2011 at 5:35 PM, Saravana Kannan [off-list ref] wrote:

On 01/21/2011 02:28 PM, Colin Cross wrote:

quoted

On Fri, Jan 21, 2011 at 2:02 PM, Russell King - ARM Linux
[off-list ref]  wrote:

quoted

On Fri, Jan 21, 2011 at 04:53:44PM -0500, Nicolas Pitre wrote:

quoted

So I think that the API must be augmented with more methods, such as:

clk_slow_enable():
  - may sleep
  - may be a no-op if the clk_fast_enable() is supported

clk_fast_enable():
  - may not sleep, used in atomic context
  - may be a no-op if controlling the clock takes time, in which case
    clk_slow_enable() must have set the clock up entirely

... and similar for clk_slow_disable() and clk_fast_disable().

Isn't this along the same lines as my clk_prepare() vs clk_enable()
suggestion?

I suggested that clk_prepare() be callable only from non-atomic contexts,
and do whatever's required to ensure that the clock is available.  That
may end up enabling the clock as a result.

clk_enable() callable from atomic contexts, and turns the clock on if
the hardware supports such an operation.

So, if you have something like:

Xtal--->PLL--->Routing/Masking--->Device

clk = clk_get() returns the clock for the device.

clk_prepare(clk) would walk up the clock tree, selecting the routing and
preparing each clock.  Clocks prior to _and_ including the PLL would need
to be enabled.

clk_enable(clk) would walk up the tree if the clock isn't already
enabled,
calling clk_enable() on the parent clock.  As we require prepared clocks
to already be enabled, this automatically stops at the PLL.

To encourage correct usage, we just need to make sure that clk_prepare()
has a might_sleep() thing, and clk_enable() throws a fit if it's used
on a clk without prepare being used first.  The second point is not easy
to do in a foolproof manner though, but doing _something_ is better than
nothing.

I like this proposal, and I prefer the clk_prepare naming over
clk_slow_enable - too many people would call clk_slow_enable instead
of, and not as well as, clk_fast_enable.

On Tegra, I currently use the ugly conditional mutex or spinlock
method to deal with voltage scaling based on clock frequency.

Colin,

MSM is in a similar situation, so thought I should bring this up to you
attention -- do you have no use case for changing the rate in atomic
context? If you do, the clk_prepare/unprepare() approach won't work.

Do you have no such requirement?

I can't think of any case that requires clk_set_rate in atomic
context.  We usually only scale frequencies to save power, so we can
always scale up early in sleepable context for a minor power penalty.