Re: [PATCH] opp: introduce library for device-specific OPPs

[Kevin Hilman <khilman@deeprootsystems.com>]

"Rafael J. Wysocki" <rjw@sisk.pl> writes:

> On Monday, September 20, 2010, Kevin Hilman wrote:
>> "Rafael J. Wysocki" <rjw@sisk.pl> writes:
>> 
>> [...]
>> 
>> >> >> In terms of the lifetime rules on the nodes in the list:
>> >> >> The list is expected to be maintained once created, entries are expected 
>> >> >> to be added optimally and not expected to be destroyed, the choice of 
>> >> >> list implementation was for reducing the complexity of the code itself 
>> >> >> and not yet meant as a mechanism to dynamically add and delete nodes on 
>> >> >> the fly.. Essentially, it was intended for the SOC framework to ensure 
>> >> >> it plugs in the OPP entries optimally and not create a humongous list of 
>> >> >> all possible OPPs for all families of the vendor SOCs - even though it 
>> >> >> is possible to use the OPP layer so - it just wont be smart to do so 
>> >> >> considering list scan latencies on hot paths such as cpufreq transitions 
>> >> >> or idle transitions.
>> >> > 
>> >> > If the list nodes are not supposed to be added and removed dynamically,
>> >> > it probably would make sense to create data structures containing
>> >> > the "available" OPPs only, once they are known, and simply free the object
>> >> > representing the other ones.
>> >> I covered the usage in my reply here: 
>> >> http://marc.info/?l=linux-arm-kernel&m=128476570300466&w=2
>> >> but to repeat, the list is dynamic during initialization but remains 
>> >> static after initialization based on SOC framework implementation - this 
>> >> is best implemented with a list (we had started with an original array 
>> >> implementation which evolved to the current list implementation 
>> >> http://marc.info/?l=linux-omap&m=125912217718770&w=2)
>> >
>> > Well, my point is, since the _final_ set of OPPs doesn't really
>> > change, there's no need to use a list for storing it in principle.
>> >
>> > Your current algorithm seems to be:
>> > (1) Create a list of all _possible_ OPPs.
>> > (2) Mark the ones that can actually be used on the given hardware as
>> >     "available".
>> > (3) Whenever we need to find an OPP to use, browse the entire list.
>> > This isn't optimal, because the OPPs that are not marked as "available" in (2)
>> > will never be used, although they _will_ be inspected while browsing the list.
>> 
>> A little clarificaion about "will never be used" below...
>> 
>> > So, I think a better algorithm would be:
>> > (1) Create a list of all possible OPPs.
>> > (2) Drop the nonavailable OPPs from the list.
>> > (3) Whenever we need to find an OPP to use, browse the entire list.
>> >
>> > But then, it may be better to simply move the list we get in (2) into an
>> > array, because the browsing is going to require fewer memory accesses in
>> > that case (also, an array would use less memory than the list).  So, perhaps,
>> > it's better to change the algorithm even further:
>> > (1) Create a list of all possible OPPs.
>> > (2) Drop the nonavailable OPPs from the list.
>> > (3) Move the list we got in (2) into a sorted array.
>> > (4) Whenever we need to find an OPP to use, browse the array
>> >      (perhaps using binary search).
>> 
>> Just a little clarification on "available."  The intended use of this flag
>> was not just a one-time "available on hardware X."  It was also intended
>> to be able to add/remove availbale OPPs dynamically at run-time.
>> 
>> More specifically, it's intended for use to *temporarily* remove an OPP
>> from being selected.  The production usage of this would primarily for
>> thermal considerations (e.g. don't use OPPx until the temperature drops)
>> 
>> However, for PM development & debug, we also use this to temporarily
>> take a class of OPPs out of the running for test/debug purposes
>> (e.g. driver X runs great at OPPx and OPPy, but not OPPz.)  So the
>> ability to temporarily be selective about OPPs at runtime for
>> debug/development is extremely useful.
>> 
>> So, to summarize, "most of the time", all the OPPs that were added (via
>> opp_add()) will be "available".  Ones that are !availble will likely
>> only be so temporarily, so I'm not sure that the overhead of keeping a
>> separate structure for the available and !available OPPs is worth it.
>> Especially, since OPP changes are relatively infrequent.
>
> Well, the Nishanth's description doesn't match this, so thanks for the
> clarification.

Agreed, we need to update the doc file to reflect this.

> In that case you might consider using a red-black tree for storing the
> "available" OPPs, so that you can add-remove them dynamically, but
> you can avoid a linear search through the entire list every time you need to
> find and available OPP.  Since we have standard helpers for handling rbtrees,
> that shouldn't be a big deal.

That's a possibility, but do you think rbtrees are worth it for a
relatively small number of OPPs for any given device?  We're using this
to track a list of OPPs for any struct device, so there may be multiple
independent OPP lists, but each would have a small number of OPPs.

For example, on OMAP, while the CPU might have a larger number of OPPs
(5-6), most devices will have a small number of OPPs (1-3.)  I gather
this is similar for many of the embedded SoCs available today.

Considering such a small number of OPPs, is the extra complexity of an
rbtree worth it?

Kevin