On 29 December 2012 01:11, Todd Poynor toddpoynor@google.com wrote:
Thanks Viresh, I'm out of the office right now but will look at this more closely early next year.
Happy New Year!!
There's discussions starting among various folks about how to handle big.LITTLE and what role cpufreq governors vs. platform code play a part in that, so all this will probably be something a little longer-term as it shakes out. I've also recently added a target_loads attribute that can pair CPU frequencies with target loads (at speed 200M try for target load 85%, at speed 1.5G try for load 98%...), which may also need adjustment if the governors manage the little CPUs. If the little CPUs have CPU numbers exposed to userspace then maybe these attributes become per-CPU group instead of global?
Anyhow, lots to think about for big.LITTLE.
Currently, Linaro is looking for two solutions to big.LITTLE: - Make scheduler aware of big.LITTLE and so, cpufreq isn't required to do anything special (Long term solution, work in progress) - Hide big.LITTLE from scheduler and perform big-LITTLE switching using cpufreq. This is a short term solution and Linaro has released its first version already (Private to members) and this patch came out of the work on that release.
-- viresh