[PATCH] clk: Fix race conditions between clk_set_parent() and clk_enable()

[skannan@codeaurora.org (Saravana Kannan)]

On 05/15/2012 01:00 PM, Sascha Hauer wrote:
> On Tue, May 15, 2012 at 12:51:06PM -0700, Saravana Kannan wrote:
>>>>   	ret = clk->ops->set_parent(clk->hw, i);
>>>
>>> You call ->set_parent while holding a spinlock. This won't work with i2c
>>> clocks.
>>
>> I did account for that. I explained it in the commit text. Please
>> let me know if any part of that is not clear or is not correct.
>>
>
> I missed this part in the commit log. I have no idea whether we can live
> with this limitation though.
>
> Sascha
>

It's not really an artificial limitation of the patch. This has to be 
enforced if the clock is to be managed correctly while allowing 
.set_parent to NOT be atomic.

There is no way to guarantee that the enable/disable is properly 
propagated to the parent clock if we can't guarantee mutual exclusion 
between changing parents and calling enable/disable.

Since we can't do mutual exclusion be using spinlock (since .set_parent 
is NOT atomic for these clocks), then only other way of ensuring mutual 
exclusion is to force an unprepare and then mutually exclude a prepare 
while changing the parent. This by association (can't enable unprepared 
clock) mutually excludes the changing of parent and calling enable/disable.

Thanks,
Saravana

-- 
Sent by an employee of the Qualcomm Innovation Center, Inc.
The Qualcomm Innovation Center, Inc. is a member of the Code Aurora Forum.