Accept a string of delays and speeds at which to apply the delay before
raising each step above hispeed. For example, "80000 1300000:200000
1500000:40000" means that the delay at or above 1GHz, until 1.3GHz is 80 msecs,
the delay until 1.5GHz is 200 msecs and the delay at or above 1.5GHz is 40
msecs when hispeed_freq is 1GHz.
[toddpoynor@google.com: add documentation]
Change-Id: Ifeebede8b1acbdd0a53e5c6916bccbf764dc854f
Signed-off-by: Minsung Kim <ms925.kim@samsung.com>
There is race condition when both two cpu do CPUFREQ_GOV_STOP and one cpu
do CPUFREQ_GOV_START soon. The sysfs_remove_group is not done yet on one
cpu, but sysfs_create_group is called on another cpu, which cause governor
start failed and then kernel panic in timer callback because the policy and
cpu mask are all kfree in cpufreq driver.
Replace atomic with mutex to lock the whole START/STOP sequence.
Change-Id: I3762b3d44315ae021b8275aca84f5ea9147cc540
Signed-off-by: Lianwei Wang <a22439@motorola.com>
Need to use irqsave/restore spinlock calls to avoid a deadlock in calls
from the timer.
Change-Id: I15b6b590045ba1447e34ca7b5ff342723e53a605
Signed-off-by: Todd Poynor <toddpoynor@google.com>
If multiple governors are in use then avoid processing frequency transition
notifications for CPUs on which the interactive governor is not enabled.
Change-Id: Ibd75255b921d887501a64774a8c4f62302f2d4e4
Reported-by: Francisco Franco <francisco.franco@cloudcar.com>
Signed-off-by: Todd Poynor <toddpoynor@google.com>
Update default go_hispeed_load from 85% to 99%. Recent changes to the
governor now use a default target_load of 90%. go_hispeed_load should
not be lower than the target load for hispeed_freq, which could lead
to oscillating speed decisions. Other recent changes reduce the need
to dampen speed jumps on load spikes, while input event boosts from
userspace are the preferred method for anticipating load spikes with
UI impacts.
General update to the documentation to reflect recent changes.
Change-Id: I1b92f3091f42c04b10503cd1169a943b5dfd6faf
Signed-off-by: Todd Poynor <toddpoynor@google.com>
Starting the governor, or restarting on a hotplugged-in CPU, can race
with the timer start in idle, triggering a BUG on timer already pending.
Start the timer before setting the enable flag, and use enable_sem to
protect the sequence (and ensure correct order of the update to the
enable flag). Delete any existing timer for safety.
Change-Id: Ife77cf9fe099e8fd8543224cbf148c6722c2ffb0
Reported-by: Francisco Franco <francisco.franco@cloudcar.com>
Signed-off-by: Todd Poynor <toddpoynor@google.com>
When stopping the governor, del_timer_sync() can race against an
invocation of the idle notifier callback, which has the potential
to reactivate the timer.
To fix this issue, a read-write semaphore is used. Multiple readers are
allowed as long as pcpu->governor_enabled is true. However it can be
moved to false only after taking a write semaphore which would wait for
any on-going asynchronous activities to complete and prevent any more of
those activities to be initiated.
[toddpoynor@google.com: cosmetic and commit text changes]
Change-Id: Ib51165a735d73dcf964a06754c48bdc1913e13d0
Signed-off-by: Nicolas Pitre <nicolas.pitre@linaro.org>
35a84de cpufreq: interactive: apply above_hispeed_delay to each step above hispeed
caused the speed choice logic to osciallate between boosting and not boosting.
Add back code to ensure speed does not drop below boost frequency while
boosting.
Change-Id: Id420068480fcc7f5c4989ff523e2a8d22e2f4db2
Signed-off-by: Todd Poynor <toddpoynor@google.com>
Always use deferrable timer for load sampling.
Set a non-deferrable timer to an additional slack time to allow prior to
waking up from idle to drop speed when not at minimum speed. Slack value
-1 avoids wakeups to drop speed. Default is 80ms.
Remove the governidle module param and its timer management in idle. For
platforms on which holding speed above mimum in idle costs power, use the
new timer slack to select how long to wait before waking up to drop speed.
Change-Id: I270b3980667e2c70a68e5bff534124b4411dbad5
Signed-off-by: Todd Poynor <toddpoynor@google.com>
Sysfs attribute boostpulse_duration specifies the duration of boosting CPU
speed in response to bootpulse events. Duration is specified in usecs,
default 80ms.
Change-Id: Ifd41625574891a44f1787a4e85d1e7b4f2afb52b
Signed-off-by: Todd Poynor <toddpoynor@google.com>
Add notifier for speed transitions. Keep a count of CPU active
microseconds times current frequency, converted to a percentage relative
to the current frequency when load is evaluated.
Change-Id: I5c27adb11081c50490219784ca57cc46e97fc28c
Signed-off-by: Todd Poynor <toddpoynor@google.com>
The longer-term load since last speed change isn't terribly useful,
may delay recognition of dropping load, and would need forthcoming
changes to adjust load for changing CPU speeds. Drop it.
Change-Id: Ic3cbb0542cc3484617031787e03ed9bdd632dec1
Signed-off-by: Todd Poynor <toddpoynor@google.com>
Accept a string of target loads and speeds at which to apply the
target loads, per the documentation update in this patch. For example,
"85 1000000:90 1700000:99" targets CPU load 85% below speed 1GHz, 90%
at or above 1GHz, until 1.7GHz and above, at which load 99% is targeted.
Attempt to avoid oscillations by evaluating the current speed
weighted by current load against each new choice of speed, choosing a
higher speed if the current load requires a higher speed.
Change-Id: Ie3300206047c84eca5a26b0b63ea512e5207550e
Signed-off-by: Todd Poynor <toddpoynor@google.com>
Apply above_hispeed_delay whenever increasing speed to a new speed above
hispeed (not just the first step above hispeed).
Change-Id: Ibb7add7db47f2a4306a9458c4e1ebabb60698636
Signed-off-by: Todd Poynor <toddpoynor@google.com>
Add a target_load attribute that specifies how aggressively the governor is
to adjust speed to meet the observed load. New target speed is calculated
as the current actual speed (may be higher than target speed on SMP) times
the CPU load (as a fraction) divided by target load (fraction).
cpufreq_frequency_table_target() call use CPUFREQ_RELATION_L to set
the next higher speed rather than next lower speed.
Change-Id: If432451da82f5fed12e15c9421d7d27792376150
Signed-off-by: Todd Poynor <toddpoynor@google.com>
Tracing adds actual speed since this is expected to be key to the
choice of target speed.
Change-Id: Iec936102d0010c4e9dfa143c38a9fd0d551189c3
Signed-off-by: Todd Poynor <toddpoynor@google.com>
The deferrable timer list isn't checked on all idle exits, such as when
hi-res timers expire or ISRs schedule workers. If the idle loop is
exited and it's past time to run the governor load polling timer,
run it immediately. This ensures we handle load spikes caused by actvity
that does not run the normal timer list.
Rename the field that timestamps the "time_in_idle" value to be more
accurate.
Change-Id: Ied590ecbefc83c9a9ec5eb9e31903557f6fa1614
Signed-off-by: Todd Poynor <toddpoynor@google.com>
Avoid wakeups only to handle the governor timer when the system is otherwise
idle.
For platforms where the power cost of remaining in idle at higher CPU
speed may outweigh the cost of a governor wakeup from idle to lower the speed,
set parameter cpufreq_interactive.governidle=1.
Change-Id: Id6c43eb35caecf9b0574fcdd5b769711bc7e6de6
Signed-off-by: LianWei WANG <a22439@motorola.com>
Signed-off-by: Todd Poynor <toddpoynor@google.com>
Helps avoid waking up other CPUs to react to activity on the local CPU.
Change-Id: Ife272aaa7916894a437705d44521b1a1693fbe8e
Signed-off-by: Todd Poynor <toddpoynor@google.com>
When load is below go_hispeed_load, apply the percentage of CPU load to
a max frequency of hispeed_freq instead of the max speed. This avoids
jumping too quickly to hispeed_freq when it is a relatively low
percentage of max speed. This also allows go_hispeed_load to be set to
a high percentage relative to hispeed_freq (as a percentage of max speed,
again useful when hispeed_freq is a low fraction of max speed), to cap
larger loads at hispeed_freq. For example, a load of 60% will typically
move to 60% of hispeed_freq, not 60% of max speed. This causes the
governor to apply two different speed caps, depending on whether load is
below or above go_hispeed_load.
Also fix the type of hispeed_freq, which was u64, to match other
speed data types (and avoid overhead and allow division).
Change-Id: Ie2d0668be161c074aaad77db2037505431457b3a
Signed-off-by: Todd Poynor <toddpoynor@google.com>
First bump speed up to hispeed_freq whenever the current speed is below
hispeed_freq, instead of only when the current speed is the minimum speed.
The previous code made it too difficult to use hispeed_freq as a common
intermediate speed on systems that frequently run at speeds between
minimum and hispeed_freq.
Change-Id: I04ec30bafabf5741e267ff289209b8c2d846824b
Signed-off-by: Todd Poynor <toddpoynor@google.com>
Not useful to have a separate, non-realtime workqueue for speed down
events, avoid priority inversion for speed up events.
Change-Id: Iddcd05545245c847aa1bbe0b8790092914c813d2
Signed-off-by: Todd Poynor <toddpoynor@google.com>
Fix a problem where the hung task mechanism was deeming the interactive
clock boost thread as hung. This was because the thread is created at
module init but never run/woken up until needed. If the governor is not
being used this can be forever. To workaround this explicitly wake up
the thread once all the necessary data structures are initialized. The
latter required some minor code shuffle.
Signed-off-by: Sam Leffler <sleffler@chromium.org>
Change-Id: Ie2c058dd75dcb6460ea10e7ac997e46baf66b1fe
Register an idle notifier only when the governor is active. Also
short-circuit work of idle end if the governor is not enabled.
Signed-off-by: Sam Leffler <sleffler@chromium.org>
Change-Id: I4cae36dd2e7389540d337d74745ffbaa0131870f
Allow speed to drop to flooor frequency but not below, don't pin
to speed at last boost.
Change-Id: I0147c2b7a2e61ba16820605af6baaf09570be787
Signed-off-by: Todd Poynor <toddpoynor@google.com>
Based on previous patches by Tero Kristo <tero.kristo@nokia.com>,
Brian Steuer <bsteuer@codeaurora.org>,
David Ng <dave@codeaurora.org>,
Antti P Miettinen <amiettinen@nvidia.com>, and
Thomas Renninger <trenn@suse.de>
Change-Id: Ic55fedcf6f9310f43a7022fb88e23b0392122769
Signed-off-by: Todd Poynor <toddpoynor@google.com>
Allow speed drop after min_sample_time elapses from last time
the current speed was last re-validated as appropriate for
current load / input boost.
Allow speed bump after min_sample_time (or above_hispeed_delay)
elapses from the time the current speed was originally set.
Change-Id: Ic25687a7a53d25e6544c30c47d7ab6f27a47bee8
Signed-off-by: Todd Poynor <toddpoynor@google.com>
For systems that set a common speed for all CPUs, checking current
speed here could bypass the intermediate hispeed bump decision for
this CPU when another CPU was already at hispeed. This could
result in an overly high setting (for all CPUs) in situations
where all CPUs were about to drop to load levels that map to
hispeed or below.
Change-Id: I186f23dcfc5e2b6336cab8b0327f0c8a9a4482bc
Signed-off-by: Todd Poynor <toddpoynor@google.com>
Apply min_sample_time to the last time the current target speed
was originally requested or re-validated as appropriate for the
current load, not to the time since the current speed was
originally set. Avoids periodic dips in speed during bursty
loads.
Change-Id: I250bda657985de60373f9897cc41f480664d51a1
Signed-off-by: Todd Poynor <toddpoynor@google.com>
If load is above go_hispeed_load, always go to at least hispeed_freq,
even when reducing speed from a higher speed, not just when jumping
up from minimum speed. Avoids running at a lower than intended
speed after a burst of even higher load.
Change-Id: I5b9d2a15ba25ce609b21bac7c724265cf6838dee
Signed-off-by: Todd Poynor <toddpoynor@google.com>
Evaluate spikes in load (below go_hispeed_load) against the maximum
speed supported by the device, not the current speed (which tends to
make it too difficult to raise speed to intermediate levels until
very busy).
Change-Id: Ib937006abf8bedb60891a739acd733e89b732ae0
Signed-off-by: Todd Poynor <toddpoynor@google.com>
This governor is designed for latency-sensitive workloads, such as
interactive user interfaces. The interactive governor aims to be
significantly more responsive to ramp CPU quickly up when CPU-intensive
activity begins.
Existing governors sample CPU load at a particular rate, typically
every X ms. This can lead to under-powering UI threads for the period of
time during which the user begins interacting with a previously-idle system
until the next sample period happens.
The 'interactive' governor uses a different approach. Instead of sampling
the CPU at a specified rate, the governor will check whether to scale the
CPU frequency up soon after coming out of idle. When the CPU comes out of
idle, a timer is configured to fire within 1-2 ticks. If the CPU is very
busy from exiting idle to when the timer fires then we assume the CPU is
underpowered and ramp to MAX speed.
If the CPU was not sufficiently busy to immediately ramp to MAX speed, then
the governor evaluates the CPU load since the last speed adjustment,
choosing the highest value between that longer-term load or the short-term
load since idle exit to determine the CPU speed to ramp to.
A realtime thread is used for scaling up, giving the remaining tasks the
CPU performance benefit, unlike existing governors which are more likely to
schedule rampup work to occur after your performance starved tasks have
completed.
The tuneables for this governor are:
/sys/devices/system/cpu/cpufreq/interactive/min_sample_time:
The minimum amount of time to spend at the current frequency before
ramping down. This is to ensure that the governor has seen enough
historic CPU load data to determine the appropriate workload.
Default is 80000 uS.
/sys/devices/system/cpu/cpufreq/interactive/go_maxspeed_load
The CPU load at which to ramp to max speed. Default is 85.
Change-Id: Ib2b362607c62f7c56d35f44a9ef3280f98c17585
Signed-off-by: Mike Chan <mike@android.com>
Signed-off-by: Todd Poynor <toddpoynor@google.com>
Bug: 3152864
Move the x86_64 idle notifiers originally by Andi Kleen and Venkatesh
Pallipadi to generic.
Change-Id: Idf29cda15be151f494ff245933c12462643388d5
Acked-by: Nicolas Pitre <nicolas.pitre@linaro.org>
Signed-off-by: Todd Poynor <toddpoynor@google.com>
get_loadavg doesn't work as intended. According to the comments, it
should be returning an average over a few seconds, but it is actually
reading the instantaneous load. It is almost always returning 0, but
can sometimes, depending on workload, spike very high into the hundreds
even when the average cpu load is under 10%. Disable it for now.
Change-Id: I63ed100af1cf9463549939b8113ed83676db5f86
Signed-off-by: Colin Cross <ccross@android.com>
Decare war on uninterruptible sleep. Add a tracepoint which
walks the kernel stack and dumps the first non-scheduler function
called before the scheduler is invoked.
Change-Id: I19e965d5206329360a92cbfe2afcc8c30f65c229
Signed-off-by: Riley Andrews <riandrews@google.com>
Dumps registers and stacktrace into console-ramoops when called
from a watchdog fiq.
Change-Id: Ib6fab5a52f670db18e64214d5e4890e8292a749c
Signed-off-by: Arve Hjønnevåg <arve@android.com>
