Modify dbs_irq_work() to always schedule the process-context work
on the current CPU which also ran the dbs_update_util_handler()
that the irq_work being handled came from.
This causes the entire frequency update handling (involving the
"ondemand" or "conservative" governors) to be carried out by the
CPU whose frequency is to be updated and reduces the overall amount
of inter-CPU noise related to cpufreq.
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
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>
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>
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>
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>
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>
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>
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>
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>
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>
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>
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>
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>
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>
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>
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>
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>
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>
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>
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>
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>
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>
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>
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>
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>
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>
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>
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>
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>
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>
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>
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>
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>
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>
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>
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>
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>
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>
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>
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>
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>
It is reported that, with CONFIG_HZ_PERIODIC=y cpu stays at the
lowest frequency even if the usage goes to 100%, neither ondemand
nor conservative governor works, however performance and
userspace work as expected. If set with CONFIG_NO_HZ_FULL=y,
everything goes well.
This problem is caused by improper calculation of the idle_time
when the load is extremely high(near 100%). Firstly, cpufreq_governor
uses get_cpu_idle_time to get the total idle time for specific cpu, then:
1.If the system is configured with CONFIG_NO_HZ_FULL, the idle time is
returned by ktime_get, which is always increasing, it's OK.
2.However, if the system is configured with CONFIG_HZ_PERIODIC,
get_cpu_idle_time might not guarantee to be always increasing,
because it will leverage get_cpu_idle_time_jiffy to calculate the
idle_time, consider the following scenario:
At T1:
idle_tick_1 = total_tick_1 - user_tick_1
sample period(80ms)...
At T2: ( T2 = T1 + 80ms):
idle_tick_2 = total_tick_2 - user_tick_2
Currently the algorithm is using (idle_tick_2 - idle_tick_1) to
get the delta idle_time during the past sample period, however
it CAN NOT guarantee that idle_tick_2 >= idle_tick_1, especially
when cpu load is high.
(Yes, total_tick_2 >= total_tick_1, and user_tick_2 >= user_tick_1,
but how about idle_tick_2 and idle_tick_1? No guarantee.)
So governor might get a negative value of idle_time during the past
sample period, which might mislead the system that the idle time is
very big(converted to unsigned int), and the busy time is nearly zero,
which causes the governor to always choose the lowest cpufreq,
then cause this problem.
In theory there are two solutions:
1.The logic should not rely on the idle tick during every sample period,
but be based on the busy tick directly, as this is how 'top' is
implemented.
2.Or the logic must make sure that the idle_time is strictly increasing
during each sample period, then there would be no negative idle_time
anymore. This solution requires minimum modification to current code
and this patch uses method 2.
Link: https://bugzilla.kernel.org/show_bug.cgi?id=69821
Reported-by: Jan Fikar <j.fikar@gmail.com>
Signed-off-by: Chen Yu <yu.c.chen@intel.com>
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
It is possible to get rid of the timer_lock spinlock used by the
governor timer function for synchronization, but a couple of races
need to be avoided.
The first race is between multiple dbs_timer_handler() instances
that may be running in parallel with each other on different
CPUs. Namely, one of them has to queue up the work item, but it
cannot be queued up more than once. To achieve that,
atomic_inc_return() can be used on the skip_work field of
struct cpu_common_dbs_info.
The second race is between an already running dbs_timer_handler()
and gov_cancel_work(). In that case the dbs_timer_handler() might
not notice the skip_work incrementation in gov_cancel_work() and
it might queue up its work item after gov_cancel_work() had
returned (and that work item would corrupt skip_work going
forward). To prevent that from happening, gov_cancel_work()
can be made wait for the timer function to complete (on all CPUs)
right after skip_work has been incremented.
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
Acked-by: Viresh Kumar <viresh.kumar@linaro.org>
cpufreq governors evaluate load at sampling rate and based on that they
update frequency for a group of CPUs belonging to the same cpufreq
policy.
This is required to be done in a single thread for all policy->cpus, but
because we don't want to wakeup idle CPUs to do just that, we use
deferrable work for this. If we would have used a single delayed
deferrable work for the entire policy, there were chances that the CPU
required to run the handler can be in idle and we might end up not
changing the frequency for the entire group with load variations.
And so we were forced to keep per-cpu works, and only the one that
expires first need to do the real work and others are rescheduled for
next sampling time.
We have been using the more complex solution until now, where we used a
delayed deferrable work for this, which is a combination of a timer and
a work.
This could be made lightweight by keeping per-cpu deferred timers with a
single work item, which is scheduled by the first timer that expires.
This patch does just that and here are important changes:
- The timer handler will run in irq context and so we need to use a
spin_lock instead of the timer_mutex. And so a separate timer_lock is
created. This also makes the use of the mutex and lock quite clear, as
we know what exactly they are protecting.
- A new field 'skip_work' is added to track when the timer handlers can
queue a work. More comments present in code.
Suggested-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
Signed-off-by: Viresh Kumar <viresh.kumar@linaro.org>
Reviewed-by: Ashwin Chaugule <ashwin.chaugule@linaro.org>
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
timer_mutex is required to be initialized only while memory for 'shared'
is allocated and in a similar way it is required to be destroyed only
when memory for 'shared' is freed.
There is no need to do the same every time we start/stop the governor.
Move code to initialize/destroy timer_mutex to the relevant places.
Signed-off-by: Viresh Kumar <viresh.kumar@linaro.org>
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
Pass 'policy' as argument to ->gov_dbs_timer() instead of cdbs and
dbs_data.
Signed-off-by: Viresh Kumar <viresh.kumar@linaro.org>
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
gov_queue_work() acquires cpufreq_governor_lock to allow
cpufreq_governor_stop() to drain delayed work items possibly scheduled
on CPUs that share the policy with a CPU being taken offline.
However, the same goal may be achieved in a more straightforward way if
the policy pointer in the struct cpu_dbs_info matching the policy CPU is
reset upfront by cpufreq_governor_stop() under the timer_mutex belonging
to it and checked against NULL, under the same lock, at the beginning of
dbs_timer().
In that case every instance of dbs_timer() run for a struct cpu_dbs_info
sharing the policy pointer in question after cpufreq_governor_stop() has
started will notice that that pointer is NULL and bail out immediately
without queuing up any new work items. In turn, gov_cancel_work()
called by cpufreq_governor_stop() before destroying timer_mutex will
wait for all of the delayed work items currently running on the CPUs
sharing the policy to drop the mutex, so it may be destroyed safely.
Make cpufreq_governor_stop() and dbs_timer() work as described and
modify gov_queue_work() so it does not acquire cpufreq_governor_lock any
more.
Signed-off-by: Viresh Kumar <viresh.kumar@linaro.org>
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
Later patches will need to create policy specific directories in
/sys/devices/system/cpu/cpufreq/ directory and so the cpufreq directory
wouldn't be ever empty.
And so no fun creating/destroying it on need basis anymore. Create it
once on system boot.
Reviewed-by: Saravana Kannan <skannan@codeaurora.org>
Signed-off-by: Viresh Kumar <viresh.kumar@linaro.org>
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
Conservative governor has its own 'enable' field to check if
conservative governor is used for a CPU or not
This can be checked by policy->governor with 'cpufreq_gov_conservative'
and so this field can be dropped.
Because its not guaranteed that dbs_info->cdbs.shared will a valid
pointer for all CPUs (will be NULL for CPUs that don't use
ondemand/conservative governors), we can't use it anymore. Lets get
policy with cpufreq_cpu_get_raw() instead.
Signed-off-by: Viresh Kumar <viresh.kumar@linaro.org>
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
Rafael J. Wysocki