[PATCH] clk: debug clock tree

[Mike Turquette]

On Fri, Dec 14, 2012 at 2:43 AM, Peter De Schrijver
<pdeschrijver at nvidia.com> wrote:
> On Thu, Dec 13, 2012 at 07:01:31PM +0100, Stephen Warren wrote:
>> On 12/13/2012 09:27 AM, Mike Turquette wrote:
>> > On Wed, Dec 12, 2012 at 7:49 PM, Prashant Gaikwad <pgaikwad at nvidia.com> wrote:
>> >> Adds debug file "clock_tree" in /sys/kernel/debug/clk dir.
>> >> It helps to view all the clock registered in tree format.
>> >>
>> >
>> > Prashant,
>> >
>> > Thanks for submitting this.  We've been talking about having a single
>> > file for representing the tree for some time.
>> >
>> > Regarding the output format had you considered using a well known
>> > format which can be parsed using well known parsing libs?  This avoids
>> > needing a custom parser just for this one file.  JSON springs to mind
>> > as something lightweight and well-understood.
>>
>> One advantage of the format below is that it's very easily
>> human-readable, and it's not too hard to parse (although I guess you'd
>> have to parse the indent level to get parent/child relation, which would
>> suck a bit). Is there room to provide both? Otherwise, I guess the
>> kernel could include a script to convert from JSON/whatever into the
>> format below.
>
> We already have the clk directory for an easy to parse version. So I think
> this one should be focused on being human readable.
>

I'm really not sure the clk directory structure is such a good idea.
Hopping around directories and dumping a few files feels like a
strange way to get at the data.

Another key point is atomicity.  The prepare_lock mutex is only held
for each individual read of a file under the current clock directory
implementation.  So if you wish to read several different clock rates
there is a caveat that the rates could change between reads.  The
"summary" file that dumps all of the data improves on this since it
could hold the prepare_lock mutex across the operation.

This last point makes me realize that the current code above either
needs to hold that lock or at least use hlist_for_each_entry_safe
since clocks could be removed from the lists during the operation.  As
such it is a bit unsafe.

I'm OK letting the two implementations live along side each other but
I am also sure that the clock summary dump should be reasonably
machine readable.

Regards,
Mike

> Cheers,
>
> Peter.