cpufreq: powernv: Add sysfs attributes to show throttle stats

Create sysfs attributes to export throttle information in
/sys/devices/system/cpu/cpuX/cpufreq/throttle_stats directory. The
newly added sysfs files are as follows:

 1)/sys/devices/system/cpu/cpuX/cpufreq/throttle_stats/turbo_stat
 2)/sys/devices/system/cpu/cpuX/cpufreq/throttle_stats/sub-turbo_stat
 3)/sys/devices/system/cpu/cpuX/cpufreq/throttle_stats/unthrottle
 4)/sys/devices/system/cpu/cpuX/cpufreq/throttle_stats/powercap
 5)/sys/devices/system/cpu/cpuX/cpufreq/throttle_stats/overtemp
 6)/sys/devices/system/cpu/cpuX/cpufreq/throttle_stats/supply_fault
 7)/sys/devices/system/cpu/cpuX/cpufreq/throttle_stats/overcurrent
 8)/sys/devices/system/cpu/cpuX/cpufreq/throttle_stats/occ_reset

Detailed explanation of each attribute is added to
Documentation/ABI/testing/sysfs-devices-system-cpu

Signed-off-by: Shilpasri G Bhat <shilpa.bhat@linux.vnet.ibm.com>
Acked-by: Viresh Kumar <viresh.kumar@linaro.org>
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>

cpufreq: acpi-cpufreq: make Intel/AMD MSR access, io port access static

These frequency register read/write operations' implementations for the
given processor (Intel/AMD MSR access or I/O port access) are only used
internally in acpi-cpufreq, so make them static.

Signed-off-by: Jisheng Zhang <jszhang@marvell.com>
Acked-by: Viresh Kumar <viresh.kumar@linaro.org>
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>

cpufreq: powernv: Define per_cpu chip pointer to optimize hot-path

Commit 96c4726f01cd "cpufreq: powernv: Remove cpu_to_chip_id() from
hot-path" introduced a 'core_to_chip_map' array to cache the chip-ids
of all cores.

Replace this with a per-CPU variable that stores the pointer to the
chip-array. This removes the linear lookup and provides a neater and
simpler solution.

Signed-off-by: Michael Neuling <mikey@neuling.org>
Signed-off-by: Shilpasri G Bhat <shilpa.bhat@linux.vnet.ibm.com>
Acked-by: Viresh Kumar <viresh.kumar@linaro.org>
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>

cpufreq: acpi-cpufreq: Clean up hot plug notifier callback

This driver has two issues.  First, it tries to fiddle with the hot
plugged CPU's MSR on the UP_PREPARE event, at a time when the CPU is
not yet online.  Second, the driver sets the "boost-disable" bit for a
CPU when going down, but does not clear the bit again if the CPU comes
up again due to DOWN_FAILED.

This patch fixes the issues by changing the driver to react to the
ONLINE/DOWN_FAILED events instead of UP_PREPARE.  As an added benefit,
the driver also becomes symmetric with respect to the hot plug
mechanism.

Signed-off-by: Richard Cochran <rcochran@linutronix.de>
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>

intel_pstate: Do not call wrmsrl_on_cpu() with disabled interrupts

After commit a4675fbc4a7a (cpufreq: intel_pstate: Replace timers with
utilization update callbacks) wrmsrl_on_cpu() cannot be called in the
intel_pstate_adjust_busy_pstate() path as that is executed with
disabled interrupts.  However, atom_set_pstate() called from there
via intel_pstate_set_pstate() uses wrmsrl_on_cpu() to update the
IA32_PERF_CTL MSR which triggers the WARN_ON_ONCE() in
smp_call_function_single().

The reason why wrmsrl_on_cpu() is used by atom_set_pstate() is
because intel_pstate_set_pstate() calling it is also invoked during
the initialization and cleanup of the driver and in those cases it is
not guaranteed to be run on the CPU that is being updated.  However,
in the case when intel_pstate_set_pstate() is called by
intel_pstate_adjust_busy_pstate(), wrmsrl() can be used to update
the register safely.  Moreover, intel_pstate_set_pstate() already
contains code that only is executed if the function is called by
intel_pstate_adjust_busy_pstate() and there is a special argument
passed to it because of that.

To fix the problem at hand, rearrange the code taking the above
observations into account.

First, replace the ->set() callback in struct pstate_funcs with a
->get_val() one that will return the value to be written to the
IA32_PERF_CTL MSR without updating the register.

Second, split intel_pstate_set_pstate() into two functions,
intel_pstate_update_pstate() to be called by
intel_pstate_adjust_busy_pstate() that will contain all of the
intel_pstate_set_pstate() code which only needs to be executed in
that case and will use wrmsrl() to update the MSR (after obtaining
the value to write to it from the ->get_val() callback), and
intel_pstate_set_min_pstate() to be invoked during the
initialization and cleanup that will set the P-state to the
minimum one and will update the MSR using wrmsrl_on_cpu().

Finally, move the code shared between intel_pstate_update_pstate()
and intel_pstate_set_min_pstate() to a new static inline function
intel_pstate_record_pstate() and make them both call it.

Of course, that unifies the handling of the IA32_PERF_CTL MSR writes
between Atom and Core.

Fixes: a4675fbc4a7a (cpufreq: intel_pstate: Replace timers with utilization update callbacks)
Reported-and-tested-by: Josh Boyer <jwboyer@fedoraproject.org>
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>

cpufreq: Make cpufreq_quick_get() safe to call

The function, cpufreq_quick_get, accesses the global 'cpufreq_driver' and
its fields without taking the associated lock, cpufreq_driver_lock.

Without the locking, nothing guarantees that 'cpufreq_driver' remains
consistent during the call.  This patch fixes the issue by taking the lock
before accessing the data structure.

Signed-off-by: Richard Cochran <rcochran@linutronix.de>
Acked-by: Viresh Kumar <viresh.kumar@linaro.org>
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>

intel_pstate: Do not skip samples partially

If the current value of MPERF or the current value of TSC is the
same as the previous one, respectively, intel_pstate_sample() bails
out early and skips the sample.

However, intel_pstate_adjust_busy_pstate() is still called in that
case which is not correct, so modify intel_pstate_sample() to
return a bool value indicating whether or not the sample has been
taken and use it to decide whether or not to call
intel_pstate_adjust_busy_pstate().

While at it, remove redundant parentheses from the MPERF/TSC
check in intel_pstate_sample().

Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
Acked-by: Srinivas Pandruvada <srinivas.pandruvada@linux.intel.com>

intel_pstate: Remove freq calculation from intel_pstate_calc_busy()

Use a helper function to compute the average pstate and call it only
where it is needed (only when tracing or in intel_pstate_get).

Signed-off-by: Philippe Longepe <philippe.longepe@linux.intel.com>
Acked-by: Srinivas Pandruvada <srinivas.pandruvada@linux.intel.com>
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>

intel_pstate: Move intel_pstate_calc_busy() into get_target_pstate_use_performance()

The cpu_load algorithm doesn't need to invoke intel_pstate_calc_busy(),
so move that call from intel_pstate_sample() to
get_target_pstate_use_performance().

Signed-off-by: Philippe Longepe <philippe.longepe@linux.intel.com>
Acked-by: Srinivas Pandruvada <srinivas.pandruvada@linux.intel.com>
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>

intel_pstate: Optimize calculation for max/min_perf_adj

mul_fp(int_tofp(A), B) expands to:
((A << FRAC_BITS) * B) >> FRAC_BITS, so the same result can be obtained
via simple multiplication A * B.  Apply this observation to
max_perf * limits->max_perf and max_perf * limits->min_perf in
intel_pstate_get_min_max()."

Signed-off-by: Philippe Longepe <philippe.longepe@linux.intel.com>
Acked-by: Srinivas Pandruvada <srinivas.pandruvada@linux.intel.com>
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>

intel_pstate: Remove extra conversions in pid calculation

pid->setpoint and pid->deadband can be initialized in fixed point, so we
can avoid the int_tofp in pid_calc.

Signed-off-by: Philippe Longepe <philippe.longepe@linux.intel.com>
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>

Merge branch 'pm-cpufreq-governor' into pm-cpufreq

cpufreq: Move scheduler-related code to the sched directory

Create cpufreq.c under kernel/sched/ and move the cpufreq code
related to the scheduler to that file and to sched.h.

Redefine cpufreq_update_util() as a static inline function to avoid
function calls at its call sites in the scheduler code (as suggested
by Peter Zijlstra).

Also move the definition of struct update_util_data and declaration
of cpufreq_set_update_util_data() from include/linux/cpufreq.h to
include/linux/sched.h.

Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org>

Revert "cpufreq: postfix policy directory with the first CPU in related_cpus"

Revert commit 3510fac45492 (cpufreq: postfix policy directory with the
first CPU in related_cpus).

Earlier, the policy->kobj was added to the kobject core, before ->init()
callback was called for the cpufreq drivers. Which allowed those drivers
to add or remove, driver dependent, sysfs files/directories to the same
kobj from their ->init() and ->exit() callbacks.

That isn't possible anymore after commit 3510fac45492.

Now, there is no other clean alternative that people can adopt.

Its better to revert the earlier commit to allow cpufreq drivers to
create/remove sysfs files from ->init() and ->exit() callbacks.

Signed-off-by: Viresh Kumar <viresh.kumar@linaro.org>
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>

cpufreq: Reduce cpufreq_update_util() overhead a bit

Use the observation that cpufreq_update_util() is only called
by the scheduler with rq->lock held, so the callers of
cpufreq_set_update_util_data() can use synchronize_sched()
instead of synchronize_rcu() to wait for cpufreq_update_util()
to complete.  Moreover, if they are updated to do that,
rcu_read_(un)lock() calls in cpufreq_update_util() might be
replaced with rcu_read_(un)lock_sched(), respectively, but
those aren't really necessary, because the scheduler calls
that function from RCU-sched read-side critical sections
already.

In addition to that, if cpufreq_set_update_util_data() checks
the func field in the struct update_util_data before setting
the per-CPU pointer to it, the data->func check may be dropped
from cpufreq_update_util() as well.

Make the above changes to reduce the overhead from
cpufreq_update_util() in the scheduler paths invoking it
and to make the cleanup after removing its callbacks less
heavy-weight somewhat.

Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
Acked-by: Viresh Kumar <viresh.kumar@linaro.org>
Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org>

cpufreq: Select IRQ_WORK if CPU_FREQ_GOV_COMMON is set

Commit 0eb463be3436 (cpufreq: governor: Replace timers with utilization
update callbacks) made CPU_FREQ select IRQ_WORK, but that's not
necessary, as it is sufficient for IRQ_WORK to be selected by
CPU_FREQ_GOV_COMMON, so modify the cpufreq Kconfig to that effect.

Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
Acked-by: Viresh Kumar <viresh.kumar@linaro.org>

cpufreq: Remove 'policy->governor_enabled'

The entire sequence of events (like INIT/START or STOP/EXIT) for which
cpufreq_governor() is called, is guaranteed to be protected by
policy->rwsem now.

The additional checks that were added earlier (as we were forced to drop
policy->rwsem before calling cpufreq_governor() for EXIT event), aren't
required anymore.

Over that, they weren't sufficient really. They just take care of
START/STOP events, but not INIT/EXIT and the state machine was never
maintained properly by them.

Kill the unnecessary checks and policy->governor_enabled field.

Signed-off-by: Viresh Kumar <viresh.kumar@linaro.org>
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>

cpufreq: Rename __cpufreq_governor() to cpufreq_governor()

The __ at the beginning of the routine aren't really necessary at all.
Rename it to cpufreq_governor() instead.

Signed-off-by: Viresh Kumar <viresh.kumar@linaro.org>
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>

cpufreq: Relocate handle_update() to kill its declaration

handle_update() is declared at the top of the file as its user appear
before its definition. Relocate the routine to get rid of this.

Signed-off-by: Viresh Kumar <viresh.kumar@linaro.org>
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>

cpufreq: governor: Drop unnecessary checks from show() and store()

The show() and store() routines in the cpufreq-governor core don't need
to check if the struct governor_attr they want to use really provides
the callbacks they need as expected (if that's not the case, it means a
bug in the code anyway), so change them to avoid doing that.

Also change the error value to -EBUSY, if the governor is getting
removed and we aren't allowed to store any more changes.

Signed-off-by: Viresh Kumar <viresh.kumar@linaro.org>
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>

cpufreq: governor: Fix race in dbs_update_util_handler()

There is a scenario that may lead to undesired results in
dbs_update_util_handler().  Namely, if two CPUs sharing a policy
enter the funtion at the same time, pass the sample delay check
and then one of them is stalled until dbs_work_handler() (queued
up by the other CPU) clears the work counter, it may update the
work counter and queue up another work item prematurely.

To prevent that from happening, use the observation that the CPU
queuing up a work item in dbs_update_util_handler() updates the
last sample time.  This means that if another CPU was stalling after
passing the sample delay check and now successfully updated the work
counter as a result of the race described above, it will see the new
value of the last sample time which is different from what it used in
the sample delay check before.  If that happens, the sample delay
check passed previously is not valid any more, so the CPU should not
continue.

Fixes: f17cbb53783c (cpufreq: governor: Avoid atomic operations in hot paths)
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
Acked-by: Viresh Kumar <viresh.kumar@linaro.org>

cpufreq: governor: Make gov_set_update_util() static

The gov_set_update_util() routine is only used internally by the
common governor code and it doesn't need to be exported, so make
it static.

No functional changes.

Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
Acked-by: Viresh Kumar <viresh.kumar@linaro.org>

cpufreq: governor: Narrow down the dbs_data_mutex coverage

Since cpufreq_governor_dbs() is now always called with policy->rwsem
held, it cannot be executed twice in parallel for the same policy.
Thus it is not necessary to hold dbs_data_mutex around the invocations
of cpufreq_governor_start/stop/limits() from it as those functions
never modify any data that can be shared between different policies.

However, cpufreq_governor_dbs() may be executed twice in parallal
for different policies using the same gov->gdbs_data object and
dbs_data_mutex is still necessary to protect that object against
concurrent updates.

For this reason, narrow down the dbs_data_mutex locking to
cpufreq_governor_init/exit() where it is needed and rename the
mutex to gov_dbs_data_mutex to reflect its purpose.

Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
Acked-by: Viresh Kumar <viresh.kumar@linaro.org>

cpufreq: governor: Make dbs_data_mutex static

That mutex is only used by cpufreq_governor_dbs() and it doesn't
need to be exported to modules, so make it static and drop the
export incantation.

No functional changes.

Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
Acked-by: Viresh Kumar <viresh.kumar@linaro.org>

cpufreq: governor: Relocate definitions of tuners structures

Move the definitions of struct od_dbs_tuners and struct cs_dbs_tuners
from the common governor header to the ondemand and conservative
governor code, respectively, as they don't need to be in the common
header any more.

No functional changes.

Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
Acked-by: Viresh Kumar <viresh.kumar@linaro.org>

cpufreq: governor: Move per-CPU data to the common code

After previous changes there is only one piece of code in the
ondemand governor making references to per-CPU data structures,
but it can be easily modified to avoid doing that, so modify it
accordingly and move the definition of per-CPU data used by the
ondemand and conservative governors to the common code.  Next,
change that code to access the per-CPU data structures directly
rather than via a governor callback.

This causes the ->get_cpu_cdbs governor callback to become
unnecessary, so drop it along with the macro and function
definitions related to it.

Finally, drop the definitions of struct od_cpu_dbs_info_s and
struct cs_cpu_dbs_info_s that aren't necessary any more.

Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
Acked-by: Viresh Kumar <viresh.kumar@linaro.org>

cpufreq: governor: Make governor private data per-policy

Some fields in struct od_cpu_dbs_info_s and struct cs_cpu_dbs_info_s
are only used for a limited set of CPUs.  Namely, if a policy is
shared between multiple CPUs, those fields will only be used for one
of them (policy->cpu).  This means that they really are per-policy
rather than per-CPU and holding room for them in per-CPU data
structures is generally wasteful.  Also moving those fields into
per-policy data structures will allow some significant simplifications
to be made going forward.

For this reason, introduce struct cs_policy_dbs_info and
struct od_policy_dbs_info to hold those fields.  Define each of the
new structures as an extension of struct policy_dbs_info (such that
struct policy_dbs_info is embedded in each of them) and introduce
new ->alloc and ->free governor callbacks to allocate and free
those structures, respectively, such that ->alloc() will return
a pointer to the struct policy_dbs_info embedded in the allocated
data structure and ->free() will take that pointer as its argument.

With that, modify the code accessing the data fields in question
in per-CPU data objects to look for them in the new structures
via the struct policy_dbs_info pointer available to it and drop
them from struct od_cpu_dbs_info_s and struct cs_cpu_dbs_info_s.

Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
Acked-by: Viresh Kumar <viresh.kumar@linaro.org>

cpufreq: ondemand: Rework the handling of powersave bias updates

The ondemand_powersave_bias_init() function used for resetting data
fields related to the powersave bias tunable of the ondemand governor
works by walking all of the online CPUs in the system and updating the
od_cpu_dbs_info_s structures for all of them.

However, if governor tunables are per policy, the update should not
touch the CPUs that are not associated with the given dbs_data.

Moreover, since the data fields in question are only ever used for
policy->cpu in each policy governed by ondemand, the update can be
limited to those specific CPUs.

Rework the code to take the above observations into account.

Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
Acked-by: Viresh Kumar <viresh.kumar@linaro.org>

cpufreq: governor: Fix CPU load information updates via ->store

The ->store() callbacks of some tunable sysfs attributes of the
ondemand and conservative governors trigger immediate updates of
the CPU load information for all CPUs "governed" by the given
dbs_data by walking the cpu_dbs_info structures for all online
CPUs in the system and updating them.

This is questionable for two reasons.  First, it may lead to a lot of
extra overhead on a system with many CPUs if the given dbs_data is
only associated with a few of them.  Second, if governor tunables are
per-policy, the CPUs associated with the other sets of governor
tunables should not be updated.

To address this issue, use the observation that in all of the places
in question the update operation may be carried out in the same way
(because all of the tunables involved are now located in struct
dbs_data and readily available to the common code) and make the
code in those places invoke the same (new) helper function that
will carry out the update correctly.

That new function always checks the ignore_nice_load tunable value
and updates the CPUs' prev_cpu_nice data fields if that's set, which
wasn't done by the original code in store_io_is_busy(), but it
should have been done in there too.

Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
Acked-by: Viresh Kumar <viresh.kumar@linaro.org>

cpufreq: ondemand: Drop one more callback from struct od_ops

The ->powersave_bias_init_cpu callback in struct od_ops is only used
in one place and that invocation may be replaced with a direct call
to the function pointed to by that callback, so change the code
accordingly and drop the callback.

Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
Acked-by: Viresh Kumar <viresh.kumar@linaro.org>

cpufreq: governor: Drop unused governor callback and data fields

After some previous changes, the ->get_cpu_dbs_info_s governor
callback and the "governor" field in struct dbs_governor (whose
value represents the governor type) are not used any more, so
drop them.

Also drop the unused gov_ops field from struct dbs_governor.

No functional changes.

Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
Acked-by: Viresh Kumar <viresh.kumar@linaro.org>

cpufreq: governor: Add a ->start callback for governors

To avoid having to check the governor type explicitly in the common
code in order to initialize data structures specific to the governor
type properly, add a ->start callback to struct dbs_governor and
use it to initialize those data structures for the ondemand and
conservative governors.

Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
Acked-by: Viresh Kumar <viresh.kumar@linaro.org>

cpufreq: governor: Move io_is_busy to struct dbs_data

The io_is_busy governor tunable is only used by the ondemand governor
and is located in the ondemand-specific data structure, but it is
looked at by the common governor code that has to do ugly things to
get to that value, so move it to struct dbs_data and modify ondemand
accordingly.

Since the conservative governor never touches that field, it will
be always 0 for that governor and it won't have any effect on the
results of computations in that case.

Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
Acked-by: Viresh Kumar <viresh.kumar@linaro.org>

cpufreq: governor: Close dbs_data update race condition

It is possible for a dbs_data object to be updated after its
usage counter has become 0.  That may happen if governor_store()
runs (via a govenor tunable sysfs attribute write) in parallel
with cpufreq_governor_exit() called for the last cpufreq policy
associated with the dbs_data in question.  In that case, if
governor_store() acquires dbs_data->mutex right after
cpufreq_governor_exit() has released it, the ->store() callback
invoked by it may operate on dbs_data with no users.  Although
sysfs will cause the kobject_put() in cpufreq_governor_exit() to
block until governor_store() has returned, that situation may
lead to some unexpected results, depending on the implementation
of the ->store callback, and therefore it should be avoided.

To that end, modify governor_store() to check the dbs_data's
usage count before invoking the ->store() callback and return
an error if it is 0 at that point.

Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
Acked-by: Viresh Kumar <viresh.kumar@linaro.org>

cpufreq: ondemand: Drop unused callback from struct od_ops

The ->freq_increase callback in struct od_ops is never invoked,
so drop it.

No functional changes.

Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
Acked-by: Viresh Kumar <viresh.kumar@linaro.org>

cpufreq: ondemand: Simplify od_update() slightly

Drop some lines of code from od_update() by arranging the statements
in there in a more logical way.

No functional changes.

Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
Acked-by: Viresh Kumar <viresh.kumar@linaro.org>

cpufreq: governor: Use microseconds in sample delay computations

Do not convert microseconds to jiffies and the other way around
in governor computations related to the sampling rate and sample
delay and drop delay_for_sampling_rate() which isn't of any use
then.

Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
Acked-by: Viresh Kumar <viresh.kumar@linaro.org>

cpufreq: ondemand: Simplify conditionals in od_dbs_timer()

Reduce the indentation level in the conditionals in od_dbs_timer()
and drop the delay variable from it.

No functional changes.

Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
Reviewed-by: Viresh Kumar <viresh.kumar@linaro.org>

cpufreq: governor: Move rate_mult to struct policy_dbs

The rate_mult field in struct od_cpu_dbs_info_s is used by the code
shared with the conservative governor and to access it that code
has to do an ugly governor type check.  However, first of all it
is ever only used for policy->cpu, so it is per-policy rather than
per-CPU and second, it is initialized to 1 by cpufreq_governor_start(),
so if the conservative governor never modifies it, it will have no
effect on the results of any computations.

For these reasons, move rate_mult to struct policy_dbs_info (as a
common field).

Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
Acked-by: Viresh Kumar <viresh.kumar@linaro.org>

cpufreq: governor: Reset sample delay in store_sampling_rate()

If store_sampling_rate() updates the sample delay when the ondemand
governor is in the middle of its high/low dance (OD_SUB_SAMPLE sample
type is set), the governor will still do the bottom half of the
previous sample which may take too much time.

To prevent that from happening, change store_sampling_rate() to always
reset the sample delay to 0 which also is consistent with the new
behavior of cpufreq_governor_limits().

Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
Acked-by: Viresh Kumar <viresh.kumar@linaro.org>

cpufreq: governor: Get rid of the ->gov_check_cpu callback

The way the ->gov_check_cpu governor callback is used by the ondemand
and conservative governors is not really straightforward.  Namely, the
governor calls dbs_check_cpu() that updates the load information for
the policy and the invokes ->gov_check_cpu() for the governor.

To get rid of that entanglement, notice that cpufreq_governor_limits()
doesn't need to call dbs_check_cpu() directly.  Instead, it can simply
reset the sample delay to 0 which will cause a sample to be taken
immediately.  The result of that is practically equivalent to calling
dbs_check_cpu() except that it will trigger a full update of governor
internal state and not just the ->gov_check_cpu() part.

Following that observation, make cpufreq_governor_limits() reset
the sample delay and turn dbs_check_cpu() into a function that will
simply evaluate the load and return the result called dbs_update().

That function can now be called by governors from the routines that
previously were pointed to by ->gov_check_cpu and those routines
can be called directly by each governor instead of dbs_check_cpu().
This way ->gov_check_cpu becomes unnecessary, so drop it.

Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
Acked-by: Viresh Kumar <viresh.kumar@linaro.org>

cpufreq: governor: Clean up load-related computations

Clean up some load-related computations in dbs_check_cpu() and
cpufreq_governor_start() to get rid of unnecessary operations and
type casts and make the code easier to read.

No functional changes.

Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
Acked-by: Viresh Kumar <viresh.kumar@linaro.org>

cpufreq: governor: Fix nice contribution computation in dbs_check_cpu()

The contribution of the CPU nice time to the idle time in dbs_check_cpu()
is computed in a bogus way, as the code may subtract current and previous
nice values for different CPUs.

That doesn't matter for cases when cpufreq policies are not shared,
but may lead to problems otherwise.

Fix the computation and simplify it to avoid taking unnecessary steps.

Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
Acked-by: Viresh Kumar <viresh.kumar@linaro.org>

cpufreq: governor: Avoid atomic operations in hot paths

Rework the handling of work items by dbs_update_util_handler() and
dbs_work_handler() so the former (which is executed in scheduler
paths) only uses atomic operations when absolutely necessary.  That
is, when the policy is shared and dbs_update_util_handler() has
already decided that this is the time to queue up a work item.

In particular, this avoids the atomic ops entirely on platforms where
policy objects are never shared.

Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
Acked-by: Viresh Kumar <viresh.kumar@linaro.org>

cpufreq: governor: Simplify gov_cancel_work() slightly

The atomic work counter incrementation in gov_cancel_work() is not
necessary any more, because work items won't be queued up after
gov_clear_update_util() anyway, so drop it along with the comment
about how it may be missed by the gov_clear_update_util().

Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
Acked-by: Viresh Kumar <viresh.kumar@linaro.org>

cpufreq: governor: Avoid irq_work_queue_on() crash on non-SMP ARM

As it turns out, irq_work_queue_on() will crash if invoked on
non-SMP ARM platforms, but in fact it is not necessary to use that
function in the cpufreq governor code (as it doesn't matter to that
code which CPU will handle the irq_work), so change it to always use
irq_work_queue().

Fixes: 8fb47ff100af (cpufreq: governor: Replace timers with utilization update callbacks)
Reported-and-tested-by: Guenter Roeck <linux@roeck-us.net>
Reported-and-tested-by: Tony Lindgren <tony@atomide.com>
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>

cpufreq: ondemand: Rearrange od_dbs_timer() to avoid updating delay

Avoid extra checks in od_dbs_timer() by rearranging updates to the
local delay variable in it.

Signed-off-by: Viresh Kumar <viresh.kumar@linaro.org>
[ rjw: Changelog ]
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>

cpufreq: conservative: Update sample_delay_ns immediately

The ondemand governor already updates sample_delay_ns immediately on
updates to the sampling rate, but conservative doesn't do that.

It was left out earlier as the code was really too complex to get
that done easily.  Things are sorted out very well now, however, and
the conservative governor can be modified to follow ondemand in that
respect.

Moreover, since the code needed to implement that in the
conservative governor would be identical to the corresponding
ondemand governor's code, make that code common and change both
governors to use it.

Signed-off-by: Viresh Kumar <viresh.kumar@linaro.org>
Tested-by: Juri Lelli <juri.lelli@arm.com>
Tested-by: Shilpasri G Bhat <shilpa.bhat@linux.vnet.ibm.com>
[ rjw: Changelog ]
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>

cpufreq: governor: No need to manage state machine now

The cpufreq core now guarantees that policy->rwsem won't be dropped
while running the ->governor callback for the CPUFREQ_GOV_POLICY_EXIT
event and will be held acquired until the complete sequence of governor
state changes has finished.

This allows governor state machine checks to be dropped from multiple
functions in cpufreq_governor.c.

This also means that policy_dbs->policy can be initialized upfront, so
the entire initialization of struct policy_dbs can be carried out in
one place.

Signed-off-by: Viresh Kumar <viresh.kumar@linaro.org>
Tested-by: Juri Lelli <juri.lelli@arm.com>
Tested-by: Shilpasri G Bhat <shilpa.bhat@linux.vnet.ibm.com>
[ rjw: Changelog ]
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>

cpufreq: Remove cpufreq_governor_lock

We used to drop policy->rwsem just before calling __cpufreq_governor()
in some cases earlier and so it was possible that __cpufreq_governor()
ran concurrently via separate threads for the same policy.

In order to guarantee valid state transitions for governors,
'governor_enabled' was required to be protected using some locking
and cpufreq_governor_lock was added for that.

But now __cpufreq_governor() is always called under policy->rwsem,
and 'governor_enabled' is protected against races even without
cpufreq_governor_lock.

Get rid of the extra lock now.

Signed-off-by: Viresh Kumar <viresh.kumar@linaro.org>
Tested-by: Juri Lelli <juri.lelli@arm.com>
Tested-by: Shilpasri G Bhat <shilpa.bhat@linux.vnet.ibm.com>
[ rjw : Changelog ]
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>

cpufreq: Call __cpufreq_governor() with policy->rwsem held

The cpufreq core code is not consistent with respect to invoking
__cpufreq_governor() under policy->rwsem.

Changing all code to always hold policy->rwsem around
__cpufreq_governor() invocations will allow us to remove
cpufreq_governor_lock that is used today because we can't
guarantee that __cpufreq_governor() isn't executed twice in
parallel for the same policy.

We should also ensure that policy->rwsem is held across governor
state changes.

For example, while adding a CPU to the policy in the CPU online path,
we need to stop the governor, change policy->cpus, start the governor
and then refresh its limits. The complete sequence must be guaranteed
to complete without interruptions by concurrent governor state
updates.  That can be achieved by holding policy->rwsem around those
sequences of operations.

Also note that after this patch cpufreq_driver->stop_cpu() and
->exit() will get called under policy->rwsem which wasn't the case
earlier. That shouldn't have any side effects, though.

Signed-off-by: Viresh Kumar <viresh.kumar@linaro.org>
Tested-by: Juri Lelli <juri.lelli@arm.com>
Tested-by: Shilpasri G Bhat <shilpa.bhat@linux.vnet.ibm.com>
[ rjw: Changelog ]
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>

cpufreq: Merge cpufreq_offline_prepare/finish routines

Commit 1aee40ac9c86 (cpufreq: Invoke __cpufreq_remove_dev_finish()
after releasing cpu_hotplug.lock) split the cpufreq's CPU offline
routine in two pieces, one of them to be run with CPU offline/online
locked and the other to be called later.  The reason for that split
was a possible deadlock scenario involving cpufreq sysfs attributes
and CPU offline.

However, the handling of CPU offline in cpufreq has changed since
then.  Policy sysfs attributes are never removed during CPU offline,
so there's no need to worry about accessing them during CPU offline,
because that can't lead to any deadlocks now.  Governor sysfs
attributes are still removed in __cpufreq_governor(_EXIT), but
there is a new kobject type for them now and its show/store
callbacks don't lock CPU offline/online (they don't need to do
that).

This means that the CPU offline code in cpufreq doesn't need to
be split any more, so combine cpufreq_offline_prepare() with
cpufreq_offline_finish().

Signed-off-by: Viresh Kumar <viresh.kumar@linaro.org>
[ rjw: Changelog ]
Tested-by: Juri Lelli <juri.lelli@arm.com>
Tested-by: Shilpasri G Bhat <shilpa.bhat@linux.vnet.ibm.com>
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>

cpufreq: governor: Create and traverse list of policy_dbs to avoid deadlock

The dbs_data_mutex lock is currently used in two places.  First,
cpufreq_governor_dbs() uses it to guarantee mutual exclusion between
invocations of governor operations from the core.  Second, it is used by
ondemand governor's update_sampling_rate() to ensure the stability of
data structures walked by it.

The second usage is quite problematic, because update_sampling_rate() is
called from a governor sysfs attribute's ->store callback and that leads
to a deadlock scenario involving cpufreq_governor_exit() which runs
under dbs_data_mutex.  Thus it is better to rework the code so
update_sampling_rate() doesn't need to acquire dbs_data_mutex.

To that end, rework update_sampling_rate() to walk a list of policy_dbs
objects supported by the dbs_data one it has been called for (instead of
walking cpu_dbs_info object for all CPUs).  The list manipulation is
protected with dbs_data->mutex which also is held around the execution
of update_sampling_rate(), it is not necessary to hold dbs_data_mutex in
that function any more.

Reported-by: Juri Lelli <juri.lelli@arm.com>
Reported-by: Shilpasri G Bhat <shilpa.bhat@linux.vnet.ibm.com>
Signed-off-by: Viresh Kumar <viresh.kumar@linaro.org>
[ rjw: Subject & changelog ]
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>

Revert "cpufreq: Drop rwsem lock around CPUFREQ_GOV_POLICY_EXIT"

Earlier, when the struct freq-attr was used to represent governor
attributes, the standard cpufreq show/store sysfs attribute callbacks
were applied to the governor tunable attributes and they always acquire
the policy->rwsem lock before carrying out the operation.  That could
have resulted in an ABBA deadlock if governor tunable attributes are
removed under policy->rwsem while one of them is being accessed
concurrently (if sysfs attributes removal wins the race, it will wait
for the access to complete with policy->rwsem held while the attribute
callback will block on policy->rwsem indefinitely).

We attempted to address this issue by dropping policy->rwsem around
governor tunable attributes removal (that is, around invocations of the
->governor callback with the event arg equal to CPUFREQ_GOV_POLICY_EXIT)
in cpufreq_set_policy(), but that opened up race conditions that had not
been possible with policy->rwsem held all the time.

The previous commit, "cpufreq: governor: New sysfs show/store callbacks
for governor tunables", fixed the original ABBA deadlock by adding new
governor specific show/store callbacks.

We don't have to drop rwsem around invocations of governor event
CPUFREQ_GOV_POLICY_EXIT anymore, and original fix can be reverted now.

Fixes: 955ef4833574 (cpufreq: Drop rwsem lock around CPUFREQ_GOV_POLICY_EXIT)
Signed-off-by: Viresh Kumar <viresh.kumar@linaro.org>
Reported-by: Juri Lelli <juri.lelli@arm.com>
Tested-by: Juri Lelli <juri.lelli@arm.com>
Tested-by: Shilpasri G Bhat <shilpa.bhat@linux.vnet.ibm.com>
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>

cpufreq: governor: Drop unused macros for creating governor tunable attributes

The previous commit introduced a new set of macros for creating sysfs
attributes that represent governor tunables and the old macros used for
this purpose are not needed any more, so drop them.

Signed-off-by: Viresh Kumar <viresh.kumar@linaro.org>
Tested-by: Juri Lelli <juri.lelli@arm.com>
Tested-by: Shilpasri G Bhat <shilpa.bhat@linux.vnet.ibm.com>
[ rjw: Subject & changelog ]
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>

cpufreq: governor: New sysfs show/store callbacks for governor tunables

The ondemand and conservative governors use the global-attr or freq-attr
structures to represent sysfs attributes corresponding to their tunables
(which of them is actually used depends on whether or not different
policy objects can use the same governor with different tunables at the
same time and, consequently, on where those attributes are located in
sysfs).

Unfortunately, in the freq-attr case, the standard cpufreq show/store
sysfs attribute callbacks are applied to the governor tunable attributes
and they always acquire the policy->rwsem lock before carrying out the
operation.  That may lead to an ABBA deadlock if governor tunable
attributes are removed under policy->rwsem while one of them is being
accessed concurrently (if sysfs attributes removal wins the race, it
will wait for the access to complete with policy->rwsem held while the
attribute callback will block on policy->rwsem indefinitely).

We attempted to address this issue by dropping policy->rwsem around
governor tunable attributes removal (that is, around invocations of the
->governor callback with the event arg equal to CPUFREQ_GOV_POLICY_EXIT)
in cpufreq_set_policy(), but that opened up race conditions that had not
been possible with policy->rwsem held all the time.  Therefore
policy->rwsem cannot be dropped in cpufreq_set_policy() at any point,
but the deadlock situation described above must be avoided too.

To that end, use the observation that in principle governor tunables may
be represented by the same data type regardless of whether the governor
is system-wide or per-policy and introduce a new structure, struct
governor_attr, for representing them and new corresponding macros for
creating show/store sysfs callbacks for them.  Also make their parent
kobject use a new kobject type whose default show/store callbacks are
not related to the standard core cpufreq ones in any way (and they don't
acquire policy->rwsem in particular).

Signed-off-by: Viresh Kumar <viresh.kumar@linaro.org>
Tested-by: Juri Lelli <juri.lelli@arm.com>
Tested-by: Shilpasri G Bhat <shilpa.bhat@linux.vnet.ibm.com>
[ rjw: Subject & changelog + rebase ]
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>

cpufreq: governor: Move common tunables to 'struct dbs_data'

There are a few common tunables shared between the ondemand and
conservative governors.  Move them to struct dbs_data to simplify
code.

Signed-off-by: Viresh Kumar <viresh.kumar@linaro.org>
Tested-by: Juri Lelli <juri.lelli@arm.com>
Tested-by: Shilpasri G Bhat <shilpa.bhat@linux.vnet.ibm.com>
[ rjw: Changelog ]
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>

cpufreq: governor: Create generic macro for common tunables

Some tunables are present in governor-specific structures, whereas one
(min_sampling_rate) is located directly in struct dbs_data.

There is a special macro for creating its sysfs attribute and the
show/store callbacks, but since more tunables are going to be moved
to struct dbs_data, a new generic macro for such cases will be useful,
so add it and use it for min_sampling_rate.

Signed-off-by: Viresh Kumar <viresh.kumar@linaro.org>
Tested-by: Juri Lelli <juri.lelli@arm.com>
Tested-by: Shilpasri G Bhat <shilpa.bhat@linux.vnet.ibm.com>
[ rjw: Subject & changelog ]
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>

cpufreq: governor: Drop pointless goto from cpufreq_governor_init()

It is silly to jump around "return 0", so don't do that.

Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
Acked-by: Viresh Kumar <viresh.kumar@linaro.org>

cpufreq: governor: Rename skip_work to work_count

The skip_work field in struct policy_dbs_info technically is a
counter, so give it a new name to reflect that.

No functional changes.

Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
Acked-by: Viresh Kumar <viresh.kumar@linaro.org>

cpufreq: governor: Symmetrize cpu_dbs_info initialization and cleanup

Make the initialization of struct cpu_dbs_info objects in
alloc_policy_dbs_info() and the code that cleans them up in
free_policy_dbs_info() more symmetrical.  In particular,
set/clear the update_util.func field in those functions along
with the policy_dbs field.

Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
Acked-by: Viresh Kumar <viresh.kumar@linaro.org>

cpufreq: governor: Rearrange governor data structures

The struct policy_dbs_info objects representing per-policy governor
data are not accessible directly from the corresponding policy
objects.  To access them, one has to get a pointer to the
struct cpu_dbs_info of policy->cpu and use the policy_dbs field of
that which isn't really straightforward.

To address that rearrange the governor data structures so the
governor_data pointer in struct cpufreq_policy will point to
struct policy_dbs_info (instead of struct dbs_data) and that will
contain a pointer to struct dbs_data.

Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
Acked-by: Viresh Kumar <viresh.kumar@linaro.org>

cpufreq: governor: Simplify cpufreq_governor_limits()

Use the observation that cpufreq_governor_limits() doesn't have to
get to the policy object it wants to manipulate by walking the
reference chain cdbs->policy_dbs->policy, as the final pointer is
actually equal to its argument, and make it access the policy
object directy via its argument.

Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
Acked-by: Viresh Kumar <viresh.kumar@linaro.org>

cpufreq: governor: Drop cpu argument from dbs_check_cpu()

Since policy->cpu is always passed as the second argument to
dbs_check_cpu(), it is not really necessary to pass it, because
the function can obtain that value via its first argument just fine.

Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
Acked-by: Viresh Kumar <viresh.kumar@linaro.org>

cpufreq: governor: Rename cpu_common_dbs_info to policy_dbs_info

The struct cpu_common_dbs_info structure represents the per-policy
part of the governor data (for the ondemand and conservative
governors), but its name doesn't reflect its purpose.

Rename it to struct policy_dbs_info and rename variables related to
it accordingly.

No functional changes.

Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
Acked-by: Viresh Kumar <viresh.kumar@linaro.org>

cpufreq: governor: Drop the gov pointer from struct dbs_data

Since it is possible to obtain a pointer to struct dbs_governor
from a pointer to the struct governor embedded in it with the help
of container_of(), the additional gov pointer in struct dbs_data
isn't really necessary.

Drop that pointer and make the code using it reach the dbs_governor
object via policy->governor.

Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
Acked-by: Viresh Kumar <viresh.kumar@linaro.org>

cpufreq: governor: Rework cpufreq_governor_dbs()

Since it is possible to obtain a pointer to struct dbs_governor
from a pointer to the struct governor embedded in it via
container_of(), the second argument of cpufreq_governor_init()
is not necessary.  Accordingly, cpufreq_governor_dbs() doesn't
need its second argument either and the ->governor callbacks
for both the ondemand and conservative governors may be set
to cpufreq_governor_dbs() directly.  Make that happen.

Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
Acked-by: Saravana Kannan <skannan@codeaurora.org>
Acked-by: Viresh Kumar <viresh.kumar@linaro.org>

cpufreq: governor: Rename some data types and variables

The ondemand and conservative governors are represented by
struct common_dbs_data whose name doesn't reflect the purpose it
is used for, so rename it to struct dbs_governor and rename
variables of that type accordingly.

No functional changes.

Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
Acked-by: Viresh Kumar <viresh.kumar@linaro.org>

cpufreq: governor: Put governor structure into common_dbs_data

For the ondemand and conservative governors (generally, governors
that use the common code in cpufreq_governor.c), there are two static
data structures representing the governor, the struct governor
structure (the interface to the cpufreq core) and the struct
common_dbs_data one (the interface to the cpufreq_governor.c code).

There's no fundamental reason why those two structures have to be
separate.  Moreover, if the struct governor one is included into
struct common_dbs_data, it will be possible to reach the latter from
the policy via its policy->governor pointer, so it won't be necessary
to pass a separate pointer to it around.  For this reason, embed
struct governor in struct common_dbs_data.

Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
Acked-by: Saravana Kannan <skannan@codeaurora.org>
Acked-by: Viresh Kumar <viresh.kumar@linaro.org>

cpufreq: governor: Avoid passing dbs_data pointers around unnecessarily

Do not pass struct dbs_data pointers to the family of functions
implementing governor operations in cpufreq_governor.c as they can
take that pointer from policy->governor by themselves.

The cpufreq_governor_init() case is slightly more complicated, since
policy->governor may be NULL when it is invoked, but then it can reach
the pointer in question via its cdata argument just fine.

While at it, rework cpufreq_governor_dbs() to avoid a pointless
policy_governor check in the CPUFREQ_GOV_POLICY_INIT case.

Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
Acked-by: Viresh Kumar <viresh.kumar@linaro.org>

cpufreq: governor: Use common mutex for dbs_data protection

Every governor relying on the common code in cpufreq_governor.c
has to provide its own mutex in struct common_dbs_data.  However,
there actually is no need to have a separate mutex per governor
for this purpose, they may be using the same global mutex just
fine.  Accordingly, introduce a single common mutex for that and
drop the mutex field from struct common_dbs_data.

That at least will ensure that the mutex is always present and
initialized regardless of what the particular governors do.

Another benefit is that the common code does not need a pointer to
a governor-related structure to get to the mutex which sometimes
helps.

Finally, it makes the code generally easier to follow.

Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
Acked-by: Saravana Kannan <skannan@codeaurora.org>
Acked-by: Viresh Kumar <viresh.kumar@linaro.org>

cpufreq: governor: Replace timers with utilization update callbacks

Instead of using a per-CPU deferrable timer for queuing up governor
work items, register a utilization update callback that will be
invoked from the scheduler on utilization changes.

The sampling rate is still the same as what was used for the
deferrable timers and the added irq_work overhead should be offset by
the eliminated timers overhead, so in theory the functional impact of
this patch should not be significant.

Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
Acked-by: Viresh Kumar <viresh.kumar@linaro.org>
Tested-by: Gautham R. Shenoy <ego@linux.vnet.ibm.com>

cpufreq: intel_pstate: Replace timers with utilization update callbacks

Instead of using a per-CPU deferrable timer for utilization sampling
and P-states adjustments, register a utilization update callback that
will be invoked from the scheduler on utilization changes.

The sampling rate is still the same as what was used for the deferrable
timers, so the functional impact of this patch should not be significant.

Based on an earlier patch from Srinivas Pandruvada.

Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
Acked-by: Srinivas Pandruvada <srinivas.pandruvada@linux.intel.com>

cpufreq: Add mechanism for registering utilization update callbacks

Introduce a mechanism by which parts of the cpufreq subsystem
("setpolicy" drivers or the core) can register callbacks to be
executed from cpufreq_update_util() which is invoked by the
scheduler's update_load_avg() on CPU utilization changes.

This allows the "setpolicy" drivers to dispense with their timers
and do all of the computations they need and frequency/voltage
adjustments in the update_load_avg() code path, among other things.

The update_load_avg() changes were suggested by Peter Zijlstra.

Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
Acked-by: Viresh Kumar <viresh.kumar@linaro.org>
Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Acked-by: Ingo Molnar <mingo@kernel.org>

cpufreq: acpi-cpufreq: Make read and write operations more efficient

Setting a new CPU frequency and reading the current request value
in the ACPI cpufreq driver involves each at least two switch
instructions (there's more if the policy is shared).  One of
them is present in drv_read/write() that prepares a command
structure and the other happens in subsequent do_drv_read/write()
when that structure is interpreted.  However, all of those switches
may be avoided by using function pointers.

To that end, add two function pointers to struct acpi_cpufreq_data
to represent read and write operations on the frequency register
and set them up during policy intitialization to point to the pair
of routines suitable for the given processor (Intel/AMD MSR access
or I/O port access).  Then, use those pointers in do_drv_read/write()
and modify drv_read/write() to prepare the command structure for
them without any checks.

Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>

cpufreq: powernv: Fix bugs in powernv_cpufreq_{init/exit}

Unregister the notifiers if cpufreq_driver_register() fails in
powernv_cpufreq_init(). Re-arrange the unregistration and cleanup routines
in powernv_cpufreq_exit() to free all the resources after the driver
has unregistered.

Signed-off-by: Shilpasri G Bhat <shilpa.bhat@linux.vnet.ibm.com>
Acked-by: Viresh Kumar <viresh.kumar@linaro.org>
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>

cpufreq: intel_pstate: disable HWP notifications

Disable HWP Interrupt notification before enabling HWP. Since we don't
have HWP interrupt handling for possible performance interrupts, there
is not much use of enabling HWP interrupts.

Signed-off-by: Srinivas Pandruvada <srinivas.pandruvada@linux.intel.com>
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>

cpufreq: intel_pstate: Enable HWP by default

If the processor supports HWP, enable it by default without checking
for the cpu model. This will allow to enable HWP in all supported
processors without driver change.

Signed-off-by: Srinivas Pandruvada <srinivas.pandruvada@linux.intel.com>
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>

cpufreq: Simplify the cpufreq_for_each_valid_entry()

That macro uses an internal static inline function that is first
totally unnecessary and second hard to read, so simplify it and
get rid of that monster.

No functional changes.

Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
Acked-by: Viresh Kumar <viresh.kumar@linaro.org>

cpufreq: acpi-cpufreq: Drop pointless label from acpi_cpufreq_target()

The "out" label at the final return statement in acpi_cpufreq_target()
is totally pointless, so drop them and modify the code to return the
right values immediately instead of jumping to it.

No functional changes.

Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
Reviewed-by: Srinivas Pandruvada <srinivas.pandruvada@linux.intel.com>
Acked-by: Viresh Kumar <viresh.kumar@linaro.org>

cpufreq: Rearrange __cpufreq_driver_target()

Drop a pointless label at a return statement from
__cpufreq_driver_target() and rearrange that function
to reduce the indentation level.

No intentional functional changes.

Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
Acked-by: Viresh Kumar <viresh.kumar@linaro.org>

intel_pstate: Update frequencies of policy->cpus only from ->set_policy()

The intel-pstate driver is using intel_pstate_hwp_set() from two
separate paths, i.e. ->set_policy() callback and sysfs update path for
the files present in /sys/devices/system/cpu/intel_pstate/ directory.

While an update to the sysfs path applies to all the CPUs being managed
by the driver (which essentially means all the online CPUs), the update
via the ->set_policy() callback applies to a smaller group of CPUs
managed by the policy for which ->set_policy() is called.

And so, intel_pstate_hwp_set() should update frequencies of only the
CPUs that are part of policy->cpus mask, while it is called from
->set_policy() callback.

In order to do that, add a parameter (cpumask) to intel_pstate_hwp_set()
and apply the frequency changes only to the concerned CPUs.

For ->set_policy() path, we are only concerned about policy->cpus, and
so policy->rwsem lock taken by the core prior to calling ->set_policy()
is enough to take care of any races. The larger lock acquired by
get_online_cpus() is required only for the updates to sysfs files.

Add another routine, intel_pstate_hwp_set_online_cpus(), and call it
from the sysfs update paths.

This also fixes a lockdep reported recently, where policy->rwsem and
get_online_cpus() could have been acquired in any order causing an ABBA
deadlock. The sequence of events leading to that was:

intel_pstate_init(...)
...cpufreq_online(...)
down_write(&policy->rwsem); // Locks policy->rwsem
...
cpufreq_init_policy(policy);
...intel_pstate_hwp_set();
get_online_cpus(); // Temporarily locks cpu_hotplug.lock
...
up_write(&policy->rwsem);

pm_suspend(...)
...disable_nonboot_cpus()
_cpu_down()
cpu_hotplug_begin(); // Locks cpu_hotplug.lock
__cpu_notify(CPU_DOWN_PREPARE, ...);
...cpufreq_offline_prepare();
down_write(&policy->rwsem); // Locks policy->rwsem

Reported-and-tested-by: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Signed-off-by: Viresh Kumar <viresh.kumar@linaro.org>
Reviewed-by: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Acked-by: Srinivas Pandruvada <srinivas.pandruvada@linux.intel.com>
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>

cpufreq: simplify for_each_suitable_policy() macro

Signed-off-by: Eric Biggers <ebiggers3@gmail.com>
Acked-by: Viresh Kumar <viresh.kumar@linaro.org>
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>

cpufreq: fix comment about return value of cpufreq_register_driver()

The comment has been incorrect since commit 4dea5806d332
("cpufreq: return EEXIST instead of EBUSY for second registering").

Signed-off-by: Eric Biggers <ebiggers3@gmail.com>
Acked-by: Viresh Kumar <viresh.kumar@linaro.org>
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>

Documentation: cpufreq: intel_pstate: fix typo

This just swaps a colon for a quote in the intel_pstate documentation.

Signed-off-by: Felipe Franciosi <felipe@nutanix.com>
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>

cpufreq: Drop unnecessary checks from show() and store()

The show() and store() routines in the cpufreq core don't need to
check if the struct freq_attr they want to use really provides the
callbacks they need as expected (if that's not the case, it means
a bug in the code anyway), so change them to avoid doing that.

Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
Acked-by: Viresh Kumar <viresh.kumar@linaro.org>

cpufreq: dt: No need to allocate resources anymore

OPP layer manages it now and cpufreq-dt driver doesn't need it. But, we
still need to check for availability of resources for deferred probing.

Signed-off-by: Viresh Kumar <viresh.kumar@linaro.org>
Reviewed-by: Stephen Boyd <sboyd@codeaurora.org>
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>

cpufreq: dt: No need to fetch voltage-tolerance

Its already done by core and we don't need to get it anymore.  And so,
we don't need to get of node in cpufreq_init() anymore, move that to
find_supply_name() instead.

Signed-off-by: Viresh Kumar <viresh.kumar@linaro.org>
Reviewed-by: Stephen Boyd <sboyd@codeaurora.org>
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>

cpufreq: dt: Use dev_pm_opp_set_rate() to switch frequency

OPP core supports frequency/voltage changes based on the target
frequency now, use that instead of open coding the same in cpufreq-dt
driver.

Signed-off-by: Viresh Kumar <viresh.kumar@linaro.org>
Reviewed-by: Stephen Boyd <sboyd@codeaurora.org>
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>

cpufreq: dt: Reuse dev_pm_opp_get_max_transition_latency()

OPP layer has all the information now to calculate transition latency
(clock_latency + voltage_latency). Lets reuse the OPP layer helper
dev_pm_opp_get_max_transition_latency() instead of open coding the same
in cpufreq-dt driver.

Signed-off-by: Viresh Kumar <viresh.kumar@linaro.org>
Reviewed-by: Stephen Boyd <sboyd@codeaurora.org>
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>

cpufreq: dt: Unsupported OPPs are already disabled

The core already have a valid regulator set for the device opp and the
unsupported OPPs are already disabled by the core. There is no need to
repeat that in the user drivers, get rid of it.

Signed-off-by: Viresh Kumar <viresh.kumar@linaro.org>
Reviewed-by: Stephen Boyd <sboyd@codeaurora.org>
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>

cpufreq: dt: Pass regulator name to the OPP core

OPP core can handle the regulators by itself, and but it needs to know
the name of the regulator to fetch. Add support for that.

Signed-off-by: Viresh Kumar <viresh.kumar@linaro.org>
Reviewed-by: Stephen Boyd <sboyd@codeaurora.org>
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>

cpufreq: dt: OPP layers handles clock-latency for V1 bindings as well

"clock-latency" is handled by OPP layer for all bindings and so there is
no need to make special calls for V1 bindings. Use
dev_pm_opp_get_max_clock_latency() for both the cases.

Signed-off-by: Viresh Kumar <viresh.kumar@linaro.org>
Reviewed-by: Stephen Boyd <sboyd@codeaurora.org>
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>

cpufreq: dt: Rename 'need_update' to 'opp_v1'

That's the real purpose of this field, i.e. to take special care of old
OPP V1 bindings. Lets name it accordingly, so that it can be used
elsewhere.

Signed-off-by: Viresh Kumar <viresh.kumar@linaro.org>
Reviewed-by: Stephen Boyd <sboyd@codeaurora.org>
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>

cpufreq: dt: Convert few pr_debug/err() calls to dev_dbg/err()

We have the device structure available now, lets use it for better print
messages.

Signed-off-by: Viresh Kumar <viresh.kumar@linaro.org>
Reviewed-by: Stephen Boyd <sboyd@codeaurora.org>
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>

Merge branch 'pm-opp' into pm-cpufreq

PM / OPP: Add dev_pm_opp_set_rate()

This adds a routine, dev_pm_opp_set_rate(), responsible for configuring
power-supply and clock source for an OPP.

The OPP is found by matching against the target_freq passed to the
routine. This shall replace similar code present in most of the OPP
users and help simplify them a lot.

Signed-off-by: Viresh Kumar <viresh.kumar@linaro.org>
Reviewed-by: Stephen Boyd <sboyd@codeaurora.org>
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>

PM / OPP: Manage device clk

OPP core has got almost everything now to manage device's OPP
transitions, the only thing left is device's clk. Get that as well.

Signed-off-by: Viresh Kumar <viresh.kumar@linaro.org>
Reviewed-by: Stephen Boyd <sboyd@codeaurora.org>
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>

PM / OPP: Parse clock-latency and voltage-tolerance for v1 bindings

V2 bindings have better support for clock-latency and voltage-tolerance
and doesn't need special care. To use callbacks, like
dev_pm_opp_get_max_{transition|volt}_latency(), irrespective of the
bindings, the core needs to know clock-latency/voltage-tolerance for the
earlier bindings.

This patch reads clock-latency/voltage-tolerance from the device node,
irrespective of the bindings (to keep it simple) and use them only for
V1 bindings.

Signed-off-by: Viresh Kumar <viresh.kumar@linaro.org>
Reviewed-by: Stephen Boyd <sboyd@codeaurora.org>
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>

PM / OPP: Introduce dev_pm_opp_get_max_transition_latency()

In few use cases (like: cpufreq), it is desired to get the maximum
latency for changing OPPs. Add support for that.

Signed-off-by: Viresh Kumar <viresh.kumar@linaro.org>
Reviewed-by: Stephen Boyd <sboyd@codeaurora.org>
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>