Re: [RFC PATCH 00/19] cpufreq locking cleanups and documentation

[Juri Lelli]

Hi Saravana,

On 29/01/16 16:57, Saravana Kannan wrote:
> On 01/11/2016 09:35 AM, Juri Lelli wrote:
> >Hi all,
> >
> >In the context of the ongoing discussion about introducing a simple platform
> >energy model to guide scheduling decisions (Energy Aware Scheduling [1])
> >concerns have been expressed by Peter about the component in charge of driving
> >clock frequency selection (Steve recently posted an update of such component
> >[2]): https://lkml.org/lkml/2015/8/15/141.
> >
> >The problem is that, with this new approach, cpufreq core functions need to be
> >accessed from scheduler hot-paths and the overhead associated with the current
> >locking scheme might result to be unsustainable.
> >
> >Peter's proposed approach of using RCU logic to reduce locking overhead seems
> >reasonable, but things may not be so straightforward as originally thought. The
> >very first thing I actually realized when I started looking into this is that
> >it was hard for me to understand which locking mechanism was protecting which
> >data structure. As mostly a way to build a better understanding of the current
> >cpufreq locking scheme and also as preparatory work for implementing RCU logic,
> >I came up with this set of patches. In fact, at this stage, I would like each
> >patch to be considered as a question I'm asking rather than a proposed change,
> >thus the RFC tag for the series; with the intent of documenting current locking
> >scheme and modifying it a bit in order to make RCU logic implementation easier.
> >Actually, as you'll soon notice, I didn't really start from scratch. Mike
> >shared with me some patches he has been developing while looking at the same
> >problem. I've given Mike attribution for the patches that I took unchanged from
> >him, with thanks for sharing his findings with me.
> >
> >High level description of patches:
> >
> >  o [01-04] cleanup and move code around to make things (hopefully) cleaner
> >  o [05-14] insert lockdep assertions and fix uncovered erroneous situations
> >  o [15-18] remove overkill usage of locking mechanism
> >  o 19      adds documentation for the cleaned up locking scheme
> >
> >With Viresh' tests [3] on both arm TC2 and arm64 Juno boards I'm not seeing
> >anything bad happening. However, coverage is really small (as is my personal
> >confidence of not breaking things for other confs :-)).
> >
> >This set is based on top of linux-pm/linux-next as of today and it is also
> >available from here:
> >
> >  git://linux-arm.org/linux-jl.git upstream/cpufreq_cleanups
> >
> >Comments, concerns and rants are the primary goal of this posting; I'm thus
> >looking forward to them.
> >
> >Best,
> >
> >- Juri
> >
> >[1] https://lkml.org/lkml/2015/7/7/754
> >[2] https://lkml.org/lkml/2015/12/9/35
> >[3] https://git.linaro.org/people/viresh.kumar/cpufreq-tests.git
> >
> >Juri Lelli (16):
> >   cpufreq: kill for_each_policy
> >   cpufreq: bring data structures close to their locks
> >   cpufreq: assert locking when accessing cpufreq_policy_list
> >   cpufreq: always access cpufreq_policy_list while holding
> >     cpufreq_driver_lock
> >   cpufreq: assert locking when accessing cpufreq_governor_list
> >   cpufreq: fix warning for cpufreq_init_policy unlocked access to
> >     cpufreq_governor_list
> >   cpufreq: fix warning for show_scaling_available_governors unlocked
> >     access to cpufreq_governor_list
> >   cpufreq: assert policy->rwsem is held in cpufreq_set_policy
> >   cpufreq: assert policy->rwsem is held in __cpufreq_governor
> >   cpufreq: fix locking of policy->rwsem in cpufreq_init_policy
> >   cpufreq: fix locking of policy->rwsem in cpufreq_offline_prepare
> >   cpufreq: fix locking of policy->rwsem in cpufreq_offline_finish
> >   cpufreq: remove useless usage of cpufreq_governor_mutex in
> >     __cpufreq_governor
> >   cpufreq: hold policy->rwsem across CPUFREQ_GOV_POLICY_EXIT
> >   cpufreq: stop checking for cpufreq_driver being present in
> >     cpufreq_cpu_get
> >   cpufreq: documentation: document locking scheme
> >
> >Michael Turquette (3):
> >   cpufreq: do not expose cpufreq_governor_lock
> >   cpufreq: merge governor lock and mutex
> >   cpufreq: remove transition_lock
> >
> >  Documentation/cpu-freq/core.txt    |  44 +++++++++++++
> >  drivers/cpufreq/cpufreq.c          | 132 +++++++++++++++++++++++--------------
> >  drivers/cpufreq/cpufreq_governor.h |   2 -
> >  include/linux/cpufreq.h            |   5 --
> >  4 files changed, 125 insertions(+), 58 deletions(-)
> >
> 
> Juri,
> 
> I haven't looked at the cpufreq-tests, but I doubt they do hotplug
> testing where they remove all the CPUs of a policy (to trigger a
> policy exit).
> 

As already pointed out by Viresh, they do actually test the case when we
hotplug out all CPUs of a policy. I'm running those on b.L. arm and
arm64 targets (two policies).

> Can you please add that to your testing? I wouldn't be surprised if
> some of your clean ups would cause a dead lock. This clean up series
> is definitely appreciated, but I think the patch series might still
> be missing some patches that are needed to make things work without
> deadlocking.
> 

Right, the problem with governors sysfs attibutes was there. But Viresh
proposed a patch (different thread I'm afraid) to fix that.

Another thing that was still missing/wrong in this set is that governors
don't hold policy->rwsem when calling __cpufreq_driver_target(), so we
can't remove transition_lock. This seems to be a bit tricky to solve
thought, as we seems to be runnning into lot of ABBA deadlocks
(timer_mutex, etc.) if we do that.

> I'll try to do a deeper analysis/review/testing, but kinda hard
> pressed on time here.
> 

Thanks a lot for looking at this set. I should be able to post an
updated version fairly soon that should address review comments.

Best,

- Juri