[PATCH RFC 0/6] ARM: OMAP2+: AM43x/AM335x prcm reset driver

[Afzal Mohammed]

Hi Philipp,

On Thursday 05 September 2013 03:37 PM, Philipp Zabel wrote:
> Am Montag, den 02.09.2013, 19:41 +0530 schrieb Afzal Mohammed:

>> Two new reset API's are provided to check whether reset is ready and
>> to clear reset. This would be required in case IP needs to mix reset
>> handling procedure with power/clock managment procedure to achieve
>> proper reset and these procedures are sometimes IP specific that would
>> make it difficult to handle reset fully in pm_runtime platform support.

>> client IP handling s/w (DT based) should do as follows:

>> 2. In driver, that require reset to be deasserted,
>>  (this is the sequence required for gfx on AM43x/AM335x, this would
>>   depend on requirements of the IP)
>>
>> 	mydriver_probe(struct platform device *pdev)
>> 	{
>> 		:
>> 		:
>> 		reset_control_get(&pdev->dev, NULL);
>> 		reset_control_clear_reset();
>> 		reset_control_deassert();
>> 		pm_runtime_get_sync();
>> 		if (reset_control_is_reset() != true)
>> 			goto err;
>> 		reset_control_put();
>> 		:
>> 		:
>> 	}

> if possible, I'd like to move this logic into the reset controller
> driver. Can this be reordered to enable power before deasserting the
> reset line (assuming it is initially asserted)? In this case, I'd
> suggest to just call device_reset:
> 
> 	pm_runtime_get_sync(&pdev->dev);
> 	ret = device_reset(&pdev->dev);
> 	if (ret)
> 		goto err;
> 
> The ops.reset callback in the prcm driver then can handle clearing
> the reset status bit, deasserting the reset control bit, and waiting for
> the reset status bit to be set (or timing out with an error).

I too would have loved to have it in such a clean way and was initially
proceeding in that direction, but there is an issue specific to OMAP
family SoC's, which was required to be taken care of (even though
present use case for AM335x/AM43x would work [as it does not have
hardware supervised clockdomain mode] with a small change in platform
level power management support code - a generic one shared with other
OMAP family SoC's)

For a module to be reset, clock domain to which the module clock belongs
should be set to software supervised mode. During "pm_runtime_get_sync",
in OMAP platform level handling code, it first put clockdomain into
software supervised mode, enables clock to module, and once module is
ready, module will be put to hardware supervised mode. In hardware
supervised mode, reset may not happen.

So if device_reset() is done after pm_runtime_get_sync(), reset may not
happen as by the time device_reset() is called, clockdomain would be in
hardware supervised mode. But in other case, as reset is already
deasserted when pm_runtime_get_sync() is called, module would be reset
as it first puts to software supervised mode.

And device reset would happen only upon enabling clock to module (if
reset was deasserted) by pm_runtime_get_sync(), so reset status has to
be checked after pm call, preventing us having device_reset() before
pm_runtime_get_sync(), or else in that case we have to sacrifice on
reset status checking (which may not be reliable)

Another alternative would have been to integrate this reset handling in
low level power management code thus hiding all reset handling from
driver, but as far as I know the reset sequence to be done is sometimes
IP specific, preventing it.

Regards
Afzal