With Android UI and benchmarks the latency of cpufreq response to
certain scheduling events can become very critical. Currently, callbacks
into cpufreq governors are only made from the scheduler if the target
CPU of the event is the same as the current CPU. This means there are
certain situations where a target CPU may not run the cpufreq governor
for some time.

One testcase to show this behavior is where a task starts running on
CPU0, then a new task is also spawned on CPU0 by a task on CPU1. If the
system is configured such that the new tasks should receive maximum
demand initially, this should result in CPU0 increasing frequency
immediately. But because of the above mentioned limitation though, this
does not occur.

This patch updates the scheduler core to call the cpufreq callbacks for
remote CPUs as well.

The schedutil, ondemand and conservative governors are updated to
process cpufreq utilization update hooks called for remote CPUs where
the remote CPU is managed by the cpufreq policy of the local CPU.

The intel_pstate driver is updated to always reject remote callbacks.

This is tested with couple of usecases (Android: hackbench, recentfling,
galleryfling, vellamo, Ubuntu: hackbench) on ARM hikey board (64 bit
octa-core, single policy). Only galleryfling showed minor improvements,
while others didn't had much deviation.

The reason being that this patch only targets a corner case, where
following are required to be true to improve performance and that
doesn't happen too often with these tests:

- Task is migrated to another CPU.
- The task has high demand, and should take the target CPU to higher
  OPPs.
- And the target CPU doesn't call into the cpufreq governor until the
  next tick.

Based on initial work from Steve Muckle.
Signed-off-by: NViresh Kumar <viresh.kumar@linaro.org>
Acked-by: NSaravana Kannan <skannan@codeaurora.org>
Acked-by: NPeter Zijlstra (Intel) <peterz@infradead.org>
Signed-off-by: NRafael J. Wysocki <rafael.j.wysocki@intel.com>

Make iowait_boost and iowait_boost_max as unsigned int since its unit
is kHz and this is consistent with struct cpufreq_policy.  Also change
the local variables in sugov_iowait_boost() to match this.
Signed-off-by: NJoel Fernandes <joelaf@google.com>
Signed-off-by: NRafael J. Wysocki <rafael.j.wysocki@intel.com>

Currently the iowait_boost feature in schedutil makes the frequency
go to max on iowait wakeups.  This feature was added to handle a case
that Peter described where the throughput of operations involving
continuous I/O requests [1] is reduced due to running at a lower
frequency, however the lower throughput itself causes utilization to
be low and hence causing frequency to be low hence its "stuck".

Instead of going to max, its also possible to achieve the same effect
by ramping up to max if there are repeated in_iowait wakeups
happening. This patch is an attempt to do that. We start from a lower
frequency (policy->min) and double the boost for every consecutive
iowait update until we reach the maximum iowait boost frequency
(iowait_boost_max).

I ran a synthetic test (continuous O_DIRECT writes in a loop) on an
x86 machine with intel_pstate in passive mode using schedutil.  In
this test the iowait_boost value ramped from 800MHz to 4GHz in 60ms.
The patch achieves the desired improved throughput as the existing
behavior.

[1] https://patchwork.kernel.org/patch/9735885/Suggested-by: NPeter Zijlstra <peterz@infradead.org>
Suggested-by: NViresh Kumar <viresh.kumar@linaro.org>
Signed-off-by: NJoel Fernandes <joelaf@google.com>
Acked-by: NViresh Kumar <viresh.kumar@linaro.org>
Signed-off-by: NRafael J. Wysocki <rafael.j.wysocki@intel.com>

With a shared policy in place, when one of the CPUs in the policy is
hotplugged out and then brought back online, sugov_stop() and
sugov_start() are called in order.

sugov_stop() removes utilization hooks for each CPU in the policy and
does nothing else in the for_each_cpu() loop. sugov_start() on the
other hand iterates through the CPUs in the policy and re-initializes
the per-cpu structure _and_ adds the utilization hook.  This implies
that the scheduler is allowed to invoke a CPU's utilization update
hook when the rest of the per-cpu structures have yet to be
re-inited.

Apart from some strange values in tracepoints this doesn't cause a
problem, but if we do end up accessing a pointer from the per-cpu
sugov_cpu structure somewhere in the sugov_update_shared() path,
we will likely see crashes since the memset for another CPU in the
policy is free to race with sugov_update_shared from the CPU that is
ready to go.  So let's fix this now to first init all per-cpu
structures, and then add the per-cpu utilization update hooks all at
once.
Signed-off-by: NVikram Mulukutla <markivx@codeaurora.org>
Acked-by: NViresh Kumar <viresh.kumar@linaro.org>
Signed-off-by: NRafael J. Wysocki <rafael.j.wysocki@intel.com>

Revert commit 39b64aa1 (cpufreq: schedutil: Reduce frequencies
slower) that introduced unintentional changes in behavior leading
to adverse effects on some systems.
Reported-by: NViresh Kumar <viresh.kumar@linaro.org>
Signed-off-by: NRafael J. Wysocki <rafael.j.wysocki@intel.com>

Currently, sugov_next_freq_shared() uses last_freq_update_time as a
reference to decide when to start considering CPU contributions as
stale.

However, since last_freq_update_time is set by the last CPU that issued
a frequency transition, this might cause problems in certain cases. In
practice, the detection of stale utilization values fails whenever the
CPU with such values was the last to update the policy. For example (and
please note again that the SCHED_CPUFREQ_RT flag is not the problem
here, but only the detection of after how much time that flag has to be
considered stale), suppose a policy with 2 CPUs:

               CPU0                |               CPU1
                                   |
                                   |     RT task scheduled
                                   |     SCHED_CPUFREQ_RT is set
                                   |     CPU1->last_update = now
                                   |     freq transition to max
                                   |     last_freq_update_time = now
                                   |

                        more than TICK_NSEC nsecs

                                   |
     a small CFS wakes up          |
     CPU0->last_update = now1      |
     delta_ns(CPU0) < TICK_NSEC*   |
     CPU0's util is considered     |
     delta_ns(CPU1) =              |
      last_freq_update_time -      |
      CPU1->last_update = 0        |
      < TICK_NSEC                  |
     CPU1 is still considered      |
     CPU1->SCHED_CPUFREQ_RT is set |
     we stay at max (until CPU1    |
     exits from idle)              |

* delta_ns is actually negative as now1 > last_freq_update_time

While last_freq_update_time is a sensible reference for rate limiting,
it doesn't seem to be useful for working around stale CPU states.

Fix the problem by always considering now (time) as the reference for
deciding when CPUs have stale contributions.
Signed-off-by: NJuri Lelli <juri.lelli@arm.com>
Acked-by: NVincent Guittot <vincent.guittot@linaro.org>
Acked-by: NViresh Kumar <viresh.kumar@linaro.org>
Signed-off-by: NRafael J. Wysocki <rafael.j.wysocki@intel.com>

Make the schedutil governor take the initial (default) value of the
rate_limit_us sysfs attribute from the (new) transition_delay_us
policy parameter (to be set by the scaling driver).

That will allow scaling drivers to make schedutil use smaller default
values of rate_limit_us and reduce the default average time interval
between consecutive frequency changes.

Make intel_pstate set transition_delay_us to 500.
Signed-off-by: NRafael J. Wysocki <rafael.j.wysocki@intel.com>
Acked-by: NViresh Kumar <viresh.kumar@linaro.org>

The schedutil governor reduces frequencies too fast in some
situations which cases undesirable performance drops to
appear.

To address that issue, make schedutil reduce the frequency slower by
setting it to the average of the value chosen during the previous
iteration of governor computations and the new one coming from its
frequency selection formula.

Link: https://bugzilla.kernel.org/show_bug.cgi?id=194963Reported-by: NJohn <john.ettedgui@gmail.com>
Signed-off-by: NRafael J. Wysocki <rafael.j.wysocki@intel.com>
Acked-by: NViresh Kumar <viresh.kumar@linaro.org>

sugov_update_commit() calls trace_cpu_frequency() to record the
current CPU frequency if it has not changed in the fast switch case
to prevent utilities from getting confused (they may report that the
CPU is idle if the frequency has not been recorded for too long, for
example).

However, that may cause the tracepoint to be triggered quite often
for no real reason (if the frequency doesn't change, we will not
modify the last update time stamp and governor computations may
run again shortly when that happens), so don't do that (arguably, it
is done to work around a utilities bug anyway).

That allows code duplication in sugov_update_commit() to be reduced
somewhat too.
Signed-off-by: NRafael J. Wysocki <rafael.j.wysocki@intel.com>
Acked-by: NViresh Kumar <viresh.kumar@linaro.org>

The way the schedutil governor uses the PELT metric causes it to
underestimate the CPU utilization in some cases.

That can be easily demonstrated by running kernel compilation on
a Sandy Bridge Intel processor, running turbostat in parallel with
it and looking at the values written to the MSR_IA32_PERF_CTL
register.  Namely, the expected result would be that when all CPUs
were 100% busy, all of them would be requested to run in the maximum
P-state, but observation shows that this clearly isn't the case.
The CPUs run in the maximum P-state for a while and then are
requested to run slower and go back to the maximum P-state after
a while again.  That causes the actual frequency of the processor to
visibly oscillate below the sustainable maximum in a jittery fashion
which clearly is not desirable.

That has been attributed to CPU utilization metric updates on task
migration that cause the total utilization value for the CPU to be
reduced by the utilization of the migrated task.  If that happens,
the schedutil governor may see a CPU utilization reduction and will
attempt to reduce the CPU frequency accordingly right away.  That
may be premature, though, for example if the system is generally
busy and there are other runnable tasks waiting to be run on that
CPU already.

This is unlikely to be an issue on systems where cpufreq policies are
shared between multiple CPUs, because in those cases the policy
utilization is computed as the maximum of the CPU utilization values
over the whole policy and if that turns out to be low, reducing the
frequency for the policy most likely is a good idea anyway.  On
systems with one CPU per policy, however, it may affect performance
adversely and even lead to increased energy consumption in some cases.

On those systems it may be addressed by taking another utilization
metric into consideration, like whether or not the CPU whose
frequency is about to be reduced has been idle recently, because if
that's not the case, the CPU is likely to be busy in the near future
and its frequency should not be reduced.

To that end, use the counter of idle calls in the timekeeping code.
Namely, make the schedutil governor look at that counter for the
current CPU every time before its frequency is about to be reduced.
If the counter has not changed since the previous iteration of the
governor computations for that CPU, the CPU has been busy for all
that time and its frequency should not be decreased, so if the new
frequency would be lower than the one set previously, the governor
will skip the frequency update.
Signed-off-by: NRafael J. Wysocki <rafael.j.wysocki@intel.com>
Acked-by: NPeter Zijlstra (Intel) <peterz@infradead.org>
Acked-by: NViresh Kumar <viresh.kumar@linaro.org>
Reviewed-by: NJoel Fernandes <joelaf@google.com>

sugov_start() only initializes struct sugov_cpu per-CPU structures
for shared policies, but it should do that for single-CPU policies too.

That in particular makes the IO-wait boost mechanism work in the
cases when cpufreq policies correspond to individual CPUs.

Fixes: 21ca6d2c (cpufreq: schedutil: Add iowait boosting)
Signed-off-by: NRafael J. Wysocki <rafael.j.wysocki@intel.com>
Acked-by: NViresh Kumar <viresh.kumar@linaro.org>
Cc: 4.9+ <stable@vger.kernel.org> # 4.9+

The loop in sugov_next_freq_shared() contains an if block to skip the
loop for the current CPU. This turns out to be an unnecessary
conditional in the scheduler's hot-path for every CPU in the policy.

It would be better to drop the conditional and make the loop treat all
the CPUs in the same way. That would eliminate the need of calling
sugov_iowait_boost() at the top of the routine.

To keep the code optimized to return early if the current CPU has RT/DL
flags set, move the flags check to sugov_update_shared() instead in
order to avoid the function call entirely.
Signed-off-by: NViresh Kumar <viresh.kumar@linaro.org>
Signed-off-by: NRafael J. Wysocki <rafael.j.wysocki@intel.com>

The rate_limit_us tunable is intended to reduce the possible overhead
from running the schedutil governor.  However, that overhead can be
divided into two separate parts: the governor computations and the
invocation of the scaling driver to set the CPU frequency.  The latter
is where the real overhead comes from.  The former is much less
expensive in terms of execution time and running it every time the
governor callback is invoked by the scheduler, after rate_limit_us
interval has passed since the last frequency update, would not be a
problem.

For this reason, redefine the rate_limit_us tunable so that it means the
minimum time that has to pass between two consecutive invocations of the
scaling driver by the schedutil governor (to set the CPU frequency).
Signed-off-by: NViresh Kumar <viresh.kumar@linaro.org>
Signed-off-by: NRafael J. Wysocki <rafael.j.wysocki@intel.com>

get_next_freq() uses sg_cpu only to get sg_policy, which the callers of
get_next_freq() already have. Pass sg_policy instead of sg_cpu to
get_next_freq(), to make it more efficient.
Signed-off-by: NViresh Kumar <viresh.kumar@linaro.org>
Signed-off-by: NRafael J. Wysocki <rafael.j.wysocki@intel.com>

cached_raw_freq applies to the entire cpufreq policy and not individual
CPUs. Apart from wasting per-cpu memory, it is actually wrong to keep it
in struct sugov_cpu as we may end up comparing next_freq with a stale
cached_raw_freq of a random CPU.

Move cached_raw_freq to struct sugov_policy.

Fixes: 5cbea469 (cpufreq: schedutil: map raw required frequency to driver frequency)
Signed-off-by: NViresh Kumar <viresh.kumar@linaro.org>
Signed-off-by: NRafael J. Wysocki <rafael.j.wysocki@intel.com>

We are going to move scheduler ABI details to <uapi/linux/sched/types.h>,
which will be used from a number of .c files.

Create empty placeholder header that maps to <linux/types.h>.

Include the new header in the files that are going to need it.
Acked-by: NLinus Torvalds <torvalds@linux-foundation.org>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: linux-kernel@vger.kernel.org
Signed-off-by: NIngo Molnar <mingo@kernel.org>

This patch rectifies a comment present in sugov_irq_work() function to
follow proper grammar.
Suggested-by: NIngo Molnar <mingo@kernel.org>
Signed-off-by: NViresh Kumar <viresh.kumar@linaro.org>
Acked-by: NIngo Molnar <mingo@kernel.org>
Signed-off-by: NRafael J. Wysocki <rafael.j.wysocki@intel.com>

Execute the irq-work specific initialization/exit code only when the
fast path isn't available.
Signed-off-by: NViresh Kumar <viresh.kumar@linaro.org>
Signed-off-by: NRafael J. Wysocki <rafael.j.wysocki@intel.com>

If slow path frequency changes are conducted in a SCHED_OTHER context
then they may be delayed for some amount of time, including
indefinitely, when real time or deadline activity is taking place.

Move the slow path to a real time kernel thread. In the future the
thread should be made SCHED_DEADLINE. The RT priority is arbitrarily set
to 50 for now.

Hackbench results on ARM Exynos, dual core A15 platform for 10
iterations:

$ hackbench -s 100 -l 100 -g 10 -f 20

Before			After
---------------------------------
1.808			1.603
1.847			1.251
2.229			1.590
1.952			1.600
1.947			1.257
1.925			1.627
2.694			1.620
1.258			1.621
1.919			1.632
1.250			1.240

Average:

1.8829			1.5041

Based on initial work by Steve Muckle.
Signed-off-by: NSteve Muckle <smuckle.linux@gmail.com>
Signed-off-by: NViresh Kumar <viresh.kumar@linaro.org>
Acked-by: NPeter Zijlstra (Intel) <peterz@infradead.org>
Signed-off-by: NRafael J. Wysocki <rafael.j.wysocki@intel.com>

The fast_switch_enabled flag will be used by both sugov_policy_alloc()
and sugov_policy_free() with a later patch.

Prepare for that by moving the calls to enable and disable it to the
beginning of sugov_init() and end of sugov_exit().
Signed-off-by: NViresh Kumar <viresh.kumar@linaro.org>
Signed-off-by: NRafael J. Wysocki <rafael.j.wysocki@intel.com>

Switch to the more common practice of writing labels.
Suggested-by: NPeter Zijlstra <peterz@infradead.org>
Signed-off-by: NViresh Kumar <viresh.kumar@linaro.org>
Signed-off-by: NRafael J. Wysocki <rafael.j.wysocki@intel.com>

Modify the schedutil cpufreq governor to boost the CPU
frequency if the SCHED_CPUFREQ_IOWAIT flag is passed to
it via cpufreq_update_util().

If that happens, the frequency is set to the maximum during
the first update after receiving the SCHED_CPUFREQ_IOWAIT flag
and then the boost is reduced by half during each following update.
Signed-off-by: NRafael J. Wysocki <rafael.j.wysocki@intel.com>
Looks-good-to: Steve Muckle <smuckle@linaro.org>
Acked-by: NPeter Zijlstra (Intel) <peterz@infradead.org>

PELT does not consider SMT when scaling its utilization values via
arch_scale_cpu_capacity(). The value in rq->cpu_capacity_orig does
take SMT into consideration though and therefore may be smaller than
the utilization reported by PELT.

On an Intel i7-3630QM for example rq->cpu_capacity_orig is 589 but
util_avg scales up to 1024. This means that a 50% utilized CPU will show
up in schedutil as ~86% busy.

Fix this by using the same CPU scaling value in schedutil as that which
is used by PELT.
Signed-off-by: NSteve Muckle <smuckle@linaro.org>
Signed-off-by: NRafael J. Wysocki <rafael.j.wysocki@intel.com>

It is useful to know the reason why cpufreq_update_util() has just
been called and that can be passed as flags to cpufreq_update_util()
and to the ->func() callback in struct update_util_data.  However,
doing that in addition to passing the util and max arguments they
already take would be clumsy, so avoid it.

Instead, use the observation that the schedutil governor is part
of the scheduler proper, so it can access scheduler data directly.
This allows the util and max arguments of cpufreq_update_util()
and the ->func() callback in struct update_util_data to be replaced
with a flags one, but schedutil has to be modified to follow.

Thus make the schedutil governor obtain the CFS utilization
information from the scheduler and use the "RT" and "DL" flags
instead of the special utilization value of ULONG_MAX to track
updates from the RT and DL sched classes.  Make it non-modular
too to avoid having to export scheduler variables to modules at
large.

Next, update all of the other users of cpufreq_update_util()
and the ->func() callback in struct update_util_data accordingly.
Suggested-by: NPeter Zijlstra <peterz@infradead.org>
Signed-off-by: NRafael J. Wysocki <rafael.j.wysocki@intel.com>
Acked-by: NPeter Zijlstra (Intel) <peterz@infradead.org>
Acked-by: NViresh Kumar <viresh.kumar@linaro.org>

The slow-path frequency transition path is relatively expensive as it
requires waking up a thread to do work. Should support be added for
remote CPU cpufreq updates that is also expensive since it requires an
IPI. These activities should be avoided if they are not necessary.

To that end, calculate the actual driver-supported frequency required by
the new utilization value in schedutil by using the recently added
cpufreq_driver_resolve_freq API. If it is the same as the previously
requested driver frequency then there is no need to continue with the
update assuming the cpu frequency limits have not changed. This will
have additional benefits should the semantics of the rate limit be
changed to apply solely to frequency transitions rather than to
frequency calculations in schedutil.

The last raw required frequency is cached. This allows the driver
frequency lookup to be skipped in the event that the new raw required
frequency matches the last one, assuming a frequency update has not been
forced due to limits changing (indicated by a next_freq value of
UINT_MAX, see sugov_should_update_freq).
Signed-off-by: NSteve Muckle <smuckle@linaro.org>
Reviewed-by: NViresh Kumar <viresh.kumar@linaro.org>
Signed-off-by: NRafael J. Wysocki <rafael.j.wysocki@intel.com>

Create a new helper to avoid code duplication across governors.
Signed-off-by: NViresh Kumar <viresh.kumar@linaro.org>
Signed-off-by: NRafael J. Wysocki <rafael.j.wysocki@intel.com>

The design of the cpufreq governor API is not very straightforward,
as struct cpufreq_governor provides only one callback to be invoked
from different code paths for different purposes.  The purpose it is
invoked for is determined by its second "event" argument, causing it
to act as a "callback multiplexer" of sorts.

Unfortunately, that leads to extra complexity in governors, some of
which implement the ->governor() callback as a switch statement
that simply checks the event argument and invokes a separate function
to handle that specific event.

That extra complexity can be eliminated by replacing the all-purpose
->governor() callback with a family of callbacks to carry out specific
governor operations: initialization and exit, start and stop and policy
limits updates.  That also turns out to reduce the code size too, so
do it.
Signed-off-by: NRafael J. Wysocki <rafael.j.wysocki@intel.com>
Acked-by: NViresh Kumar <viresh.kumar@linaro.org>

Prefix print messages with KBUILD_MODNAME, i.e 'cpufreq_schedutil: '.
This helps to keep similar formatting for all the print messages
particular to a file and identify those easily in kernel logs.

Its already done this way for rest of the governors.

Along with that, remove the (now) redundant bits from a print message.
Signed-off-by: NViresh Kumar <viresh.kumar@linaro.org>
Signed-off-by: NRafael J. Wysocki <rafael.j.wysocki@intel.com>

Due to differences in the cpufreq core's handling of runtime CPU
offline and nonboot CPUs disabling during system suspend-to-RAM,
fast frequency switching gets disabled after a suspend-to-RAM and
resume cycle on all of the nonboot CPUs.

To prevent that from happening, move the invocation of
cpufreq_disable_fast_switch() from cpufreq_exit_governor() to
sugov_exit(), as the schedutil governor is the only user of fast
frequency switching today anyway.

That simply prevents cpufreq_disable_fast_switch() from being called
without invoking the ->governor callback for the CPUFREQ_GOV_POLICY_EXIT
event (which happens during system suspend now).

Fixes: b7898fda (cpufreq: Support for fast frequency switching)
Signed-off-by: NRafael J. Wysocki <rafael.j.wysocki@intel.com>
Acked-by: NViresh Kumar <viresh.kumar@linaro.org>

Add a new cpufreq scaling governor, called "schedutil", that uses
scheduler-provided CPU utilization information as input for making
its decisions.

Doing that is possible after commit 34e2c555 (cpufreq: Add
mechanism for registering utilization update callbacks) that
introduced cpufreq_update_util() called by the scheduler on
utilization changes (from CFS) and RT/DL task status updates.
In particular, CPU frequency scaling decisions may be based on
the the utilization data passed to cpufreq_update_util() by CFS.

The new governor is relatively simple.

The frequency selection formula used by it depends on whether or not
the utilization is frequency-invariant.  In the frequency-invariant
case the new CPU frequency is given by

	next_freq = 1.25 * max_freq * util / max

where util and max are the last two arguments of cpufreq_update_util().
In turn, if util is not frequency-invariant, the maximum frequency in
the above formula is replaced with the current frequency of the CPU:

	next_freq = 1.25 * curr_freq * util / max

The coefficient 1.25 corresponds to the frequency tipping point at
(util / max) = 0.8.

All of the computations are carried out in the utilization update
handlers provided by the new governor.  One of those handlers is
used for cpufreq policies shared between multiple CPUs and the other
one is for policies with one CPU only (and therefore it doesn't need
to use any extra synchronization means).

The governor supports fast frequency switching if that is supported
by the cpufreq driver in use and possible for the given policy.
In the fast switching case, all operations of the governor take
place in its utilization update handlers.  If fast switching cannot
be used, the frequency switch operations are carried out with the
help of a work item which only calls __cpufreq_driver_target()
(under a mutex) to trigger a frequency update (to a value already
computed beforehand in one of the utilization update handlers).

Currently, the governor treats all of the RT and DL tasks as
"unknown utilization" and sets the frequency to the allowed
maximum when updated from the RT or DL sched classes.  That
heavy-handed approach should be replaced with something more
subtle and specifically targeted at RT and DL tasks.

The governor shares some tunables management code with the
"ondemand" and "conservative" governors and uses some common
definitions from cpufreq_governor.h, but apart from that it
is stand-alone.
Signed-off-by: NRafael J. Wysocki <rafael.j.wysocki@intel.com>
Acked-by: NViresh Kumar <viresh.kumar@linaro.org>
Acked-by: NPeter Zijlstra (Intel) <peterz@infradead.org>