Previously we were using count-weight of the T124 for T30 in order to
get EMC clock rate that was reasonable for T30. In fact the count-weight
should be x2 times smaller on T30, but then devfreq was producing a bit
too low EMC clock rate for ISO memory clients, like display controller
for example.
Now both Tegra ACTMON and Tegra DRM display drivers support interconnect
framework and display driver tells to ICC what a minimum memory bandwidth
is needed, preventing FIFO underflows. Thus, now we can use a proper
count-weight value for Tegra30 and MC_ALL device config needs a bit more
aggressive boosting.
Add a separate ACTMON driver configuration that is specific to Tegra30.
Tested-by: Peter Geis <pgwipeout@gmail.com>
Tested-by: Nicolas Chauvet <kwizart@gmail.com>
Acked-by: Chanwoo Choi <cw00.choi@samsung.com>
Signed-off-by: Dmitry Osipenko <digetx@gmail.com>
Signed-off-by: Chanwoo Choi <cw00.choi@samsung.com>
This patch moves ACTMON driver away from generating OPP table by itself,
transitioning it to use the table which comes from device-tree. This
change breaks compatibility with older device-trees and brings support
for the interconnect framework to the driver. This is a mandatory change
which needs to be done in order to implement interconnect-based memory
DVFS, i.e. device-trees need to be updated. Now ACTMON issues a memory
bandwidth requests using dev_pm_opp_set_bw() instead of driving EMC clock
rate directly.
Tested-by: Peter Geis <pgwipeout@gmail.com>
Tested-by: Nicolas Chauvet <kwizart@gmail.com>
Acked-by: Chanwoo Choi <cw00.choi@samsung.com>
Signed-off-by: Dmitry Osipenko <digetx@gmail.com>
Signed-off-by: Chanwoo Choi <cw00.choi@samsung.com>
Tegra EMC driver was turned into a regular kernel driver, meaning that it
could be compiled as a loadable kernel module now. Hence EMC clock isn't
guaranteed to be available and clk_get("emc") may return -EPROBE_DEFER.
Let's silence the deferred probe error.
Acked-by: Chanwoo Choi <cw00.choi@samsung.com>
Signed-off-by: Dmitry Osipenko <digetx@gmail.com>
Signed-off-by: Chanwoo Choi <cw00.choi@samsung.com>
DEVFREQ supports the default governors like performance, simple_ondemand and
also allows the devfreq driver to add their own governor like tegra30-devfreq.c
according to their requirement. In result, some sysfs attributes are useful
or not useful. Prior to that the user can access all sysfs attributes
regardless of the available attributes.
So, clarify the access permission of sysfs attributes according to governor.
When adding the devfreq governor, can specify the available attribute
information by using DEVFREQ_GOV_ATTR_* constant variable. The user can
read or write the sysfs attributes in accordance to the specified attributes.
When adding the governor, can add the following attributes
according to the governor feature.
[Definition for speific sysfs attributes]
- DEVFREQ_GOV_ATTR_POLLING_INTERVAL to update polling interval for timer.
: /sys/class/devfreq/[devfreq dev name]/polling_interval
- DEVFREQ_GOV_ATTR_TIMER to change the type of timer on either deferrable
or dealyed timer.
: /sys/class/devfreq/[devfreq dev name]/timer
And all devfreq governors have to support the following common attributes.
The common attributes are added to devfreq class by default.
- governor
- available_governors
- available_frequencies
- cur_freq
- target_freq
- min_freq
- max_freq
- trans_stat
[Table of governor attribute flags for devfreq governors]
------------------------------------------------------------------------------
| simple | perfor | power | user | passive | tegra30
| ondemand | mance | save | space| |
------------------------------------------------------------------------------
governor | O | O | O | O | O | O
available_governors | O | O | O | O | O | O
available_frequencies | O | O | O | O | O | O
cur_freq | O | O | O | O | O | O
target_freq | O | O | O | O | O | O
min_freq | O | O | O | O | O | O
max_freq | O | O | O | O | O | O
trans_stat | O | O | O | O | O | O
--------------------------------------------------------
polling_interval | O | X | X | X | X | O
timer | O | X | X | X | X | X
------------------------------------------------------------------------------
Reviewed-by: Dmitry Osipenko <digetx@gmail.com>
Tested-by: Dmitry Osipenko <digetx@gmail.com>
Signed-off-by: Chanwoo Choi <cw00.choi@samsung.com>
The devfreq governor is able to have the specific flag as follows
in order to implement the specific feature. For example, devfreq allows
user to change the governors on runtime via sysfs interface.
But, if devfreq device uses 'passive' governor, don't allow user to change
the governor. For this case, define the DEVFREQ_GOV_FLAG_IMMUTABLE
and set it to flag of passive governor.
[Definition for governor flag]
- DEVFREQ_GOV_FLAG_IMMUTABLE
: If immutable flag is set, governor is never changeable to other governors.
- DEVFREQ_GOV_FLAG_IRQ_DRIVEN
: Devfreq core won't schedule polling work for this governor if value is set.
[Table of governor flag for devfreq governors]
------------------------------------------------------------------------------
| simple | perfor | power | user | passive | tegra30
| ondemand | mance | save | space| |
------------------------------------------------------------------------------
immutable | X | X | X | X | O | O
interrupt_driven | X(polling)| X | X | X | X | O (irq)
------------------------------------------------------------------------------
Reviewed-by: Dmitry Osipenko <digetx@gmail.com>
Tested-by: Dmitry Osipenko <digetx@gmail.com>
Signed-off-by: Chanwoo Choi <cw00.choi@samsung.com>
It's safe to enable the ACTMON clock at any time during driver probing,
even if we don't know the state of hardware, because it's used only for
collecting and processing stats, and interrupt is kept disabled. This
allows us to slightly improve code which performs initial hardware
resetting by making use of a single reset_control_reset() instead of
assert/deassert pair. Secondly, a potential error of the reset-control
API is handled nicely now.
Signed-off-by: Dmitry Osipenko <digetx@gmail.com>
Signed-off-by: Chanwoo Choi <cw00.choi@samsung.com>
The function “platform_get_irq” can log an error already.
Thus omit a redundant message for the exception handling in the
calling function.
This issue was detected by using the Coccinelle software.
Signed-off-by: Markus Elfring <elfring@users.sourceforge.net>
Reviewed-by: Dmitry Osipenko <digetx@gmail.com>
Signed-off-by: Chanwoo Choi <cw00.choi@samsung.com>
We're taking into account both HW memory-accesses + CPU activity based on
current CPU's frequency. For memory-accesses there is a kind of hysteresis
in a form of "boosting" which is managed by the tegra30-devfreq driver.
If current HW memory activity is higher than activity judged based of the
CPU's frequency, then there is no need to schedule cpufreq_update_work
because the result of the work will be a NO-OP. And thus,
tegra_actmon_cpufreq_contribution() should return 0, meaning that at the
moment CPU frequency doesn't contribute anything to the final decision
about required memory clock rate.
Signed-off-by: Dmitry Osipenko <digetx@gmail.com>
Signed-off-by: Chanwoo Choi <cw00.choi@samsung.com>
DEVFREQ_GOV_INTERVAL event indicates that update the interval
for polling mode of devfreq device. But, this event name doesn't
specify exactly what to do.
Change DEVFREQ_GOV_INTERVAL event name to DEVFREQ_GOV_UPDATE_INTERVAL
which specifies what to do by event name.
And modify the function name to DEVFREQ_GOV_UPDATE_INTERVAL
with 'devfreq_' prefix + verb + object as following:
- devfreq_interval_update -> devfreq_updatee_interval
Reviewed-by: Lukasz Luba <lukasz.luba@arm.com>
Signed-off-by: Chanwoo Choi <cw00.choi@samsung.com>
MCCPU boosts up very aggressively by 800% and boosts down very mildly by
10%. This doesn't work well when system is idling because the very slow
de-boosting results in lots of consecutive-down interrupts, in result
memory stays clocked high and CPU doesn't enter deepest idling state
instead of keeping memory at lowest freq and having CPU cluster turned
off. A more faster de-boosting fixes the case of idling system and doesn't
affect the case of an active system.
Reviewed-by: Chanwoo Choi <cw00.choi@samsung.com>
Signed-off-by: Dmitry Osipenko <digetx@gmail.com>
Signed-off-by: Chanwoo Choi <cw00.choi@samsung.com>
The ACTMON governor is interrupt-driven and currently hardware's polling
interval is fixed to 16ms in the driver. Devfreq supports variable polling
interval by the generic governors, let's re-use the generic interface for
changing of the polling interval. Now the polling interval can be changed
dynamically via /sys/class/devfreq/devfreq0/polling_interval.
Signed-off-by: Dmitry Osipenko <digetx@gmail.com>
Reviewed-by: Chanwoo Choi <cw00.choi@samsung.com>
Signed-off-by: Chanwoo Choi <cw00.choi@samsung.com>
The dependency threshold designates a memory activity level below which
CPU's frequency isn't accounted. Currently the threshold is given in
"memory cycle" units and that value depends on the polling interval which
is fixed to 12ms in the driver. Later on we'd want to add support for a
variable polling interval and thus the threshold value either needs to be
scaled in accordance to the polling interval or it needs to be represented
in a units that do not depend on the polling interval.
It is nicer to have threshold value being defined independently of the
polling interval, thus this patch converts the dependency threshold units
from "cycle" to "kHz". Having this change as a separate-preparatory patch
will make easier to follow further patches.
Signed-off-by: Dmitry Osipenko <digetx@gmail.com>
Reviewed-by: Chanwoo Choi <cw00.choi@samsung.com>
Signed-off-by: Chanwoo Choi <cw00.choi@samsung.com>
Consecutive interrupts should be disabled when boosting is completed.
Currently the disabling of "lower" interrupt happens only for MCCPU
monitor that uses dependency threshold, but even in a case of MCCPU the
interrupt isn't getting disabled if CPU's activity is above the threshold.
This results in a lot of dummy interrupt requests. The boosting feature is
used by both MCCPU and MCALL, boosting should be stopped once it reaches 0
for both of the monitors and regardless of the activity level.
The boosting stops to grow once the maximum limit is hit and thus the
"upper" interrupt needs to be disabled when the limit is reached.
Reviewed-by: Chanwoo Choi <cw00.choi@samsung.com>
Signed-off-by: Dmitry Osipenko <digetx@gmail.com>
Signed-off-by: Chanwoo Choi <cw00.choi@samsung.com>
Consecutive up/down interrupt-bit is set in the interrupt status register
only if that interrupt was previously enabled. Thus enabling the already
enabled interrupt doesn't do much for us.
Reviewed-by: Chanwoo Choi <cw00.choi@samsung.com>
Signed-off-by: Dmitry Osipenko <digetx@gmail.com>
Signed-off-by: Chanwoo Choi <cw00.choi@samsung.com>
It's not very correct to include mod_devicetable.h for the OF device
drivers and of_device.h should be included instead.
Reviewed-by: Chanwoo Choi <cw00.choi@samsung.com>
Tested-by: Peter Geis <pgwipeout@gmail.com>
Signed-off-by: Dmitry Osipenko <digetx@gmail.com>
Signed-off-by: Chanwoo Choi <cw00.choi@samsung.com>
The consecutive-down event tells that we should perform frequency
de-boosting, but boosting is in a reset state on start and hence the
event won't do anything useful for us and it will be just a dummy
interrupt request.
Reviewed-by: Chanwoo Choi <cw00.choi@samsung.com>
Tested-by: Peter Geis <pgwipeout@gmail.com>
Signed-off-by: Dmitry Osipenko <digetx@gmail.com>
Signed-off-by: Chanwoo Choi <cw00.choi@samsung.com>
Governor could be stopped while boosting is active. We have assumption
that everything is reset on governor's restart, including the boosting
value, which was missed.
Reviewed-by: Chanwoo Choi <cw00.choi@samsung.com>
Tested-by: Peter Geis <pgwipeout@gmail.com>
Signed-off-by: Dmitry Osipenko <digetx@gmail.com>
Signed-off-by: Chanwoo Choi <cw00.choi@samsung.com>
There is no point in receiving of the notifications while governor is
stopped, let's keep them disabled like we do for the CPU freq-change
notifications. This also fixes a potential use-after-free bug if
notification happens after device's removal.
Reviewed-by: Chanwoo Choi <cw00.choi@samsung.com>
Tested-by: Peter Geis <pgwipeout@gmail.com>
Signed-off-by: Dmitry Osipenko <digetx@gmail.com>
Signed-off-by: Chanwoo Choi <cw00.choi@samsung.com>
The CPU's client need to take into account that CPUFreq may change
while memory activity not, staying high. Thus an appropriate frequency
notifier should be used in addition to the clk-notifier.
Reviewed-by: Chanwoo Choi <cw00.choi@samsung.com>
Tested-by: Peter Geis <pgwipeout@gmail.com>
Signed-off-by: Dmitry Osipenko <digetx@gmail.com>
Signed-off-by: Chanwoo Choi <cw00.choi@samsung.com>
Part of the code uses Hz units and the other kHz, let's switch to kHz
everywhere for consistency. A small benefit from this change (besides
code's cleanup) is that now powertop utility correctly displays devfreq's
stats, for some reason it expects them to be in kHz.
Tested-by: Peter Geis <pgwipeout@gmail.com>
Reviewed-by: Chanwoo Choi <cw00.choi@samsung.com>
Signed-off-by: Dmitry Osipenko <digetx@gmail.com>
Signed-off-by: Chanwoo Choi <cw00.choi@samsung.com>
There is another kHz-conversion bug in the code, resulting in integer
overflow. Although, this time the resulting value is 4294966296 and it's
close to ULONG_MAX, which is okay in this case.
Reviewed-by: Chanwoo Choi <cw00.choi@samsung.com>
Tested-by: Peter Geis <pgwipeout@gmail.com>
Signed-off-by: Dmitry Osipenko <digetx@gmail.com>
Signed-off-by: Chanwoo Choi <cw00.choi@samsung.com>
There is no need in a write-barrier now, given that interrupt masking is
handled by CPU's GIC now. Hence we know exactly that interrupt won't fire
after stopping the devfreq's governor. In other cases we don't care about
potential buffering of the writes to hardware and thus there is no need to
stall CPU.
Reviewed-by: Chanwoo Choi <cw00.choi@samsung.com>
Tested-by: Peter Geis <pgwipeout@gmail.com>
Signed-off-by: Dmitry Osipenko <digetx@gmail.com>
Signed-off-by: Chanwoo Choi <cw00.choi@samsung.com>
There is no real need to keep interrupt always-enabled, will be nicer
to keep it disabled while governor is inactive.
Suggested-by: Thierry Reding <thierry.reding@gmail.com>
Reviewed-by: Chanwoo Choi <cw00.choi@samsung.com>
Tested-by: Peter Geis <pgwipeout@gmail.com>
Signed-off-by: Dmitry Osipenko <digetx@gmail.com>
Signed-off-by: Chanwoo Choi <cw00.choi@samsung.com>
IRQ numbers are always positive, hence the corresponding variable should
be unsigned to keep types consistent. This is a minor change that cleans
up code a tad more.
Suggested-by: Thierry Reding <thierry.reding@gmail.com>
Acked-by: MyungJoo Ham <myungjoo.ham@samsung.com>
Reviewed-by: Chanwoo Choi <cw00.choi@samsung.com>
Tested-by: Peter Geis <pgwipeout@gmail.com>
Signed-off-by: Dmitry Osipenko <digetx@gmail.com>
Signed-off-by: Chanwoo Choi <cw00.choi@samsung.com>
In order to reflect that driver serves NVIDIA Tegra30 and later SoC
generations, let's rename the driver's source file to "tegra30-devfreq.c".
This will make driver files to look more consistent after addition of a
driver for Tegra20.
Reviewed-by: Chanwoo Choi <cw00.choi@samsung.com>
Signed-off-by: Dmitry Osipenko <digetx@gmail.com>
Signed-off-by: MyungJoo Ham <myungjoo.ham@samsung.com>
Dmitry Osipenko
Chanwoo Choi
Dan Carpenter
Markus Elfring