This change uses SCHED_CPUFREQ_IOWAIT flag to boost HWP performance.
Since SCHED_CPUFREQ_IOWAIT flag is set frequently, we don't start
boosting steps unless we see two consecutive flags in two ticks. This
avoids boosting due to IO because of regular system activities.
To avoid synchronization issues, the actual processing of the flag is
done on the local CPU callback.
Reported-by: Mel Gorman <mgorman@techsingularity.net>
Tested-by: Giovanni Gherdovich <ggherdovich@suse.cz>
Signed-off-by: Srinivas Pandruvada <srinivas.pandruvada@linux.intel.com>
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
Added two utility functions to HWP boost up gradually and boost down to
the default cached HWP request values.
Boost up:
Boost up updates HWP request minimum value in steps. This minimum value
can reach upto at HWP request maximum values depends on how frequently,
this boost up function is called. At max, boost up will take three steps
to reach the maximum, depending on the current HWP request levels and HWP
capabilities. For example, if the current settings are:
If P0 (Turbo max) = P1 (Guaranteed max) = min
No boost at all.
If P0 (Turbo max) > P1 (Guaranteed max) = min
Should result in one level boost only for P0.
If P0 (Turbo max) = P1 (Guaranteed max) > min
Should result in two level boost:
(min + p1)/2 and P1.
If P0 (Turbo max) > P1 (Guaranteed max) > min
Should result in three level boost:
(min + p1)/2, P1 and P0.
We don't set any level between P0 and P1 as there is no guarantee that
they will be honored.
Boost down:
After the system is idle for hold time of 3ms, the HWP request is reset
to the default value from HWP init or user modified one via sysfs.
Caching of HWP Request and Capabilities
Store the HWP request value last set using MSR_HWP_REQUEST and read
MSR_HWP_CAPABILITIES. This avoid reading of MSRs in the boost utility
functions.
These boost utility functions calculated limits are based on the latest
HWP request value, which can be modified by setpolicy() callback. So if
user space modifies the minimum perf value, that will be accounted for
every time the boost up is called. There will be case when there can be
contention with the user modified minimum perf, in that case user value
will gain precedence. For example just before HWP_REQUEST MSR is updated
from setpolicy() callback, the boost up function is called via scheduler
tick callback. Here the cached MSR value is already the latest and limits
are updated based on the latest user limits, but on return the MSR write
callback called from setpolicy() callback will update the HWP_REQUEST
value. This will be used till next time the boost up function is called.
In addition add a variable to control HWP dynamic boosting. When HWP
dynamic boost is active then set the HWP specific update util hook. The
contents in the utility hooks will be filled in the subsequent patches.
Reported-by: Mel Gorman <mgorman@techsingularity.net>
Tested-by: Giovanni Gherdovich <ggherdovich@suse.cz>
Signed-off-by: Srinivas Pandruvada <srinivas.pandruvada@linux.intel.com>
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
Allow use of the trace_pstate_sample trace function
when the intel_pstate driver is in passive mode.
Since the core_busy and scaled_busy fields are not
used, and it might be desirable to know which path
through the driver was used, either intel_cpufreq_target
or intel_cpufreq_fast_switch, re-task the core_busy
field as a flag indicator.
The user can then use the intel_pstate_tracer.py utility
to summarize and plot the trace.
Note: The core_busy feild still goes by that name
in include/trace/events/power.h and within the
intel_pstate_tracer.py script and csv file headers,
but it is graphed as "performance", and called
core_avg_perf now in the intel_pstate driver.
Sometimes, in passive mode, the driver is not called for
many tens or even hundreds of seconds. The user
needs to understand, and not be confused by, this limitation.
Signed-off-by: Doug Smythies <dsmythies@telus.net>
Signed-off-by: Srinivas Pandruvada <srinivas.pandruvada@linux.intel.com>
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
The intel_pstate driver doesn't use debugfs any more, so drop
linux/debugfs.h from the list of included headers in it.
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
Acked-by: Viresh Kumar <viresh.kumar@linaro.org>
When maxcpus=1 is in the kernel command line, the BP is responsible
for re-enabling the HWP - because currently only the APs invoke
intel_pstate_hwp_enable() during their online process - which might
put the system into unstable state after resume.
Fix this by enabling the HWP explicitly on BP during resume.
Reported-by: Doug Smythies <dsmythies@telus.net>
Suggested-by: Srinivas Pandruvada <srinivas.pandruvada@linux.intel.com>
Signed-off-by: Yu Chen <yu.c.chen@intel.com>
[ rjw: Subject/changelog, minor modifications ]
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
Currently intel_pstate can function only in HWP mode on Skylake servers.
When HWP feature is not enabled on the processor then acpi-cpufreq is
driver is used.
Based on the power and performance tests using intel_pstate scaling
algorithm the results are comparable. But intel_pstate brings in
additional features:
- Display of turbo frequency range, which many users like to see
- Place limits in the turbo frequency range when platform allows
Since these tests are done only using non PID algorithm introduced in
kernel version 4.14, this patch is not a backport candidate. So each user
has to carefully weigh the benefits before he backports.
Signed-off-by: Srinivas Pandruvada <srinivas.pandruvada@linux.intel.com>
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
Since core_funcs and bxt_funcs have same set of callbacks, replace
bxt_funcs with core_funcs.
Signed-off-by: Srinivas Pandruvada <srinivas.pandruvada@linux.intel.com>
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
- Update the ACPICA code in the kernel to upstream revision 20170728
including:
* Alias operator handling update (Bob Moore).
* Deferred resolution of reference package elements (Bob Moore).
* Support for the _DMA method in walk resources (Bob Moore).
* Tables handling update and support for deferred table
verification (Lv Zheng).
* Update of SMMU models for IORT (Robin Murphy).
* Compiler and disassembler updates (Alex James, Erik Schmauss,
Ganapatrao Kulkarni, James Morse).
* Tools updates (Erik Schmauss, Lv Zheng).
* Assorted minor fixes and cleanups (Bob Moore, Kees Cook,
Lv Zheng, Shao Ming).
- Rework the initialization of non-wakeup GPEs with method handlers
in order to address a boot crash on some systems with Thunderbolt
devices connected at boot time where we miss an early hotplug
event due to a delay in GPE enabling (Rafael Wysocki).
- Rework the handling of PCI bridges when setting up ACPI-based
device wakeup in order to avoid disabling wakeup for bridges
prematurely (Rafael Wysocki).
- Consolidate Apple DMI checks throughout the tree, add support for
Apple device properties to the device properties framework and
use these properties for the handling of I2C and SPI devices on
Apple systems (Lukas Wunner).
- Add support for _DMA to the ACPI-based device properties lookup
code and make it possible to use the information from there to
configure DMA regions on ARM64 systems (Lorenzo Pieralisi).
- Fix several issues in the APEI code, add support for exporting
the BERT error region over sysfs and update APEI MAINTAINERS
entry with reviewers information (Borislav Petkov, Dongjiu Geng,
Loc Ho, Punit Agrawal, Tony Luck, Yazen Ghannam).
- Fix a potential initialization ordering issue in the ACPI EC
driver and clean it up somewhat (Lv Zheng).
- Update the ACPI SPCR driver to extend the existing XGENE 8250
workaround in it to a new platform (m400) and to work around
an Xgene UART clock issue (Graeme Gregory).
- Add a new utility function to the ACPI core to support using
ACPI OEM ID / OEM Table ID / Revision for system identification
in blacklisting or similar and switch over the existing code
already using this information to this new interface (Toshi Kani).
- Fix an xpower PMIC issue related to GPADC reads that always return
0 without extra pin manipulations (Hans de Goede).
- Add statements to print debug messages in a couple of places in
the ACPI core for easier diagnostics (Rafael Wysocki).
- Clean up the ACPI processor driver slightly (Colin Ian King,
Hanjun Guo).
- Clean up the ACPI x86 boot code somewhat (Andy Shevchenko).
- Add a quirk for Dell OptiPlex 9020M to the ACPI backlight
driver (Alex Hung).
- Assorted fixes, cleanups and updates related to ACPI (Amitoj Kaur
Chawla, Bhumika Goyal, Frank Rowand, Jean Delvare, Punit Agrawal,
Ronald Tschalär, Sumeet Pawnikar).
-----BEGIN PGP SIGNATURE-----
Version: GnuPG v2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=6Ke9
-----END PGP SIGNATURE-----
Merge tag 'acpi-4.14-rc1' of git://git.kernel.org/pub/scm/linux/kernel/git/rafael/linux-pm
Pull ACPI updates from Rafael Wysocki:
"These include a usual ACPICA code update (this time to upstream
revision 20170728), a fix for a boot crash on some systems with
Thunderbolt devices connected at boot time, a rework of the handling
of PCI bridges when setting up device wakeup, new support for Apple
device properties, support for DMA configurations reported via ACPI on
ARM64, APEI-related updates, ACPI EC driver updates and assorted minor
modifications in several places.
Specifics:
- Update the ACPICA code in the kernel to upstream revision 20170728
including:
* Alias operator handling update (Bob Moore).
* Deferred resolution of reference package elements (Bob Moore).
* Support for the _DMA method in walk resources (Bob Moore).
* Tables handling update and support for deferred table
verification (Lv Zheng).
* Update of SMMU models for IORT (Robin Murphy).
* Compiler and disassembler updates (Alex James, Erik Schmauss,
Ganapatrao Kulkarni, James Morse).
* Tools updates (Erik Schmauss, Lv Zheng).
* Assorted minor fixes and cleanups (Bob Moore, Kees Cook, Lv
Zheng, Shao Ming).
- Rework the initialization of non-wakeup GPEs with method handlers
in order to address a boot crash on some systems with Thunderbolt
devices connected at boot time where we miss an early hotplug event
due to a delay in GPE enabling (Rafael Wysocki).
- Rework the handling of PCI bridges when setting up ACPI-based
device wakeup in order to avoid disabling wakeup for bridges
prematurely (Rafael Wysocki).
- Consolidate Apple DMI checks throughout the tree, add support for
Apple device properties to the device properties framework and use
these properties for the handling of I2C and SPI devices on Apple
systems (Lukas Wunner).
- Add support for _DMA to the ACPI-based device properties lookup
code and make it possible to use the information from there to
configure DMA regions on ARM64 systems (Lorenzo Pieralisi).
- Fix several issues in the APEI code, add support for exporting the
BERT error region over sysfs and update APEI MAINTAINERS entry with
reviewers information (Borislav Petkov, Dongjiu Geng, Loc Ho, Punit
Agrawal, Tony Luck, Yazen Ghannam).
- Fix a potential initialization ordering issue in the ACPI EC driver
and clean it up somewhat (Lv Zheng).
- Update the ACPI SPCR driver to extend the existing XGENE 8250
workaround in it to a new platform (m400) and to work around an
Xgene UART clock issue (Graeme Gregory).
- Add a new utility function to the ACPI core to support using ACPI
OEM ID / OEM Table ID / Revision for system identification in
blacklisting or similar and switch over the existing code already
using this information to this new interface (Toshi Kani).
- Fix an xpower PMIC issue related to GPADC reads that always return
0 without extra pin manipulations (Hans de Goede).
- Add statements to print debug messages in a couple of places in the
ACPI core for easier diagnostics (Rafael Wysocki).
- Clean up the ACPI processor driver slightly (Colin Ian King, Hanjun
Guo).
- Clean up the ACPI x86 boot code somewhat (Andy Shevchenko).
- Add a quirk for Dell OptiPlex 9020M to the ACPI backlight driver
(Alex Hung).
- Assorted fixes, cleanups and updates related to ACPI (Amitoj Kaur
Chawla, Bhumika Goyal, Frank Rowand, Jean Delvare, Punit Agrawal,
Ronald Tschalär, Sumeet Pawnikar)"
* tag 'acpi-4.14-rc1' of git://git.kernel.org/pub/scm/linux/kernel/git/rafael/linux-pm: (75 commits)
ACPI / APEI: Suppress message if HEST not present
intel_pstate: convert to use acpi_match_platform_list()
ACPI / blacklist: add acpi_match_platform_list()
ACPI, APEI, EINJ: Subtract any matching Register Region from Trigger resources
ACPI: make device_attribute const
ACPI / sysfs: Extend ACPI sysfs to provide access to boot error region
ACPI: APEI: fix the wrong iteration of generic error status block
ACPI / processor: make function acpi_processor_check_duplicates() static
ACPI / EC: Clean up EC GPE mask flag
ACPI: EC: Fix possible issues related to EC initialization order
ACPI / PM: Add debug statements to acpi_pm_notify_handler()
ACPI: Add debug statements to acpi_global_event_handler()
ACPI / scan: Enable GPEs before scanning the namespace
ACPICA: Make it possible to enable runtime GPEs earlier
ACPICA: Dispatch active GPEs at init time
ACPI: SPCR: work around clock issue on xgene UART
ACPI: SPCR: extend XGENE 8250 workaround to m400
ACPI / LPSS: Don't abort ACPI scan on missing mem resource
mailbox: pcc: Drop uninformative output during boot
ACPI/IORT: Add IORT named component memory address limits
...
* intel_pstate:
cpufreq: intel_pstate: Shorten a couple of long names
cpufreq: intel_pstate: Simplify intel_pstate_adjust_pstate()
cpufreq: intel_pstate: Improve IO performance with per-core P-states
cpufreq: intel_pstate: Drop INTEL_PSTATE_HWP_SAMPLING_INTERVAL
cpufreq: intel_pstate: Drop ->update_util from pstate_funcs
cpufreq: intel_pstate: Do not use PID-based P-state selection
* pm-cpufreq-sched:
cpufreq: schedutil: Always process remote callback with slow switching
cpufreq: schedutil: Don't restrict kthread to related_cpus unnecessarily
cpufreq: Return 0 from ->fast_switch() on errors
cpufreq: Simplify cpufreq_can_do_remote_dvfs()
cpufreq: Process remote callbacks from any CPU if the platform permits
sched: cpufreq: Allow remote cpufreq callbacks
cpufreq: schedutil: Use unsigned int for iowait boost
cpufreq: schedutil: Make iowait boost more energy efficient
* pm-cpufreq: (33 commits)
cpufreq: imx6q: Fix imx6sx low frequency support
cpufreq: speedstep-lib: make several arrays static, makes code smaller
cpufreq: ti: Fix 'of_node_put' being called twice in error handling path
cpufreq: dt-platdev: Drop few entries from whitelist
cpufreq: dt-platdev: Automatically create cpufreq device with OPP v2
ARM: ux500: don't select CPUFREQ_DT
cpufreq: Convert to using %pOF instead of full_name
cpufreq: Cap the default transition delay value to 10 ms
cpufreq: dbx500: Delete obsolete driver
mfd: db8500-prcmu: Get rid of cpufreq dependency
cpufreq: enable the DT cpufreq driver on the Ux500
cpufreq: Loongson2: constify platform_device_id
cpufreq: dt: Add r8a7796 support to to use generic cpufreq driver
cpufreq: remove setting of policy->cpu in policy->cpus during init
cpufreq: mediatek: add support of cpufreq to MT7622 SoC
cpufreq: mediatek: add cleanups with the more generic naming
cpufreq: rcar: Add support for R8A7795 SoC
cpufreq: dt: Add rk3328 compatible to use generic cpufreq driver
cpufreq: s5pv210: add missing of_node_put()
cpufreq: Allow dynamic switching with CPUFREQ_ETERNAL latency
...
Convert to use acpi_match_platform_list() for the platform check.
There is no change in functionality.
Signed-off-by: Toshi Kani <toshi.kani@hpe.com>
Acked-by: Srinivas Pandruvada <srinivas.pandruvada@linux.intel.com>
Reviewed-by: Borislav Petkov <bp@suse.de>
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
policy->cpu is copied into policy->cpus in cpufreq_online() before
calling into cpufreq_driver->init(). So there's no need to set the
same in the individual driver init() functions again.
This patch removes the redundant setting of policy->cpu in policy->cpus
in intel_pstate and cppc drivers.
Reported-by: Viresh Kumar <viresh.kumar@linaro.org>
Signed-off-by: Sudeep Holla <sudeep.holla@arm.com>
Acked-by: Viresh Kumar <viresh.kumar@linaro.org>
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
The intel_pstate CPU frequency scaling driver has always
calculated CPU frequency incorrectly. Recent changes have
eliminted most of the issues, however the frequency reported
in the trace buffer, if used, is incorrect.
It remains desireable that cpu->pstate.scaling still be a nice
round number for things such as when setting max and min frequencies.
So the proposal is to just fix the reported frequency in the trace data.
Fixes what remains of [1].
Link: https://bugzilla.kernel.org/show_bug.cgi?id=96521 # [1]
Signed-off-by: Doug Smythies <dsmythies@telus.net>
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
The names of the INTEL_PSTATE_DEFAULT_SAMPLING_INTERVAL symbol and
the get_target_pstate_use_cpu_load() function don't need to be so
long any more, so make them shorter.
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
Since there is only one P-state selection routine in intel_pstate
now, make intel_pstate_adjust_pstate() call it directly and drop
the target_pstate argument from that function.
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
In the current implementation, the response latency between seeing
SCHED_CPUFREQ_IOWAIT set and the actual P-state adjustment can be up
to 10ms. It can be reduced by bumping up the P-state to the max at
the time SCHED_CPUFREQ_IOWAIT is passed to intel_pstate_update_util().
With this change, the IO performance improves significantly.
For a simple "grep -r . linux" (Here linux is the kernel source
folder) with caches dropped every time on a Broadwell Xeon workstation
with per-core P-states, the user and system time is shorter by as much
as 30% - 40%.
The same performance difference was not observed on clients that don't
support per-core P-state.
Signed-off-by: Srinivas Pandruvada <srinivas.pandruvada@linux.intel.com>
[ rjw: Changelog ]
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
With Android UI and benchmarks the latency of cpufreq response to
certain scheduling events can become very critical. Currently, callbacks
into cpufreq governors are only made from the scheduler if the target
CPU of the event is the same as the current CPU. This means there are
certain situations where a target CPU may not run the cpufreq governor
for some time.
One testcase to show this behavior is where a task starts running on
CPU0, then a new task is also spawned on CPU0 by a task on CPU1. If the
system is configured such that the new tasks should receive maximum
demand initially, this should result in CPU0 increasing frequency
immediately. But because of the above mentioned limitation though, this
does not occur.
This patch updates the scheduler core to call the cpufreq callbacks for
remote CPUs as well.
The schedutil, ondemand and conservative governors are updated to
process cpufreq utilization update hooks called for remote CPUs where
the remote CPU is managed by the cpufreq policy of the local CPU.
The intel_pstate driver is updated to always reject remote callbacks.
This is tested with couple of usecases (Android: hackbench, recentfling,
galleryfling, vellamo, Ubuntu: hackbench) on ARM hikey board (64 bit
octa-core, single policy). Only galleryfling showed minor improvements,
while others didn't had much deviation.
The reason being that this patch only targets a corner case, where
following are required to be true to improve performance and that
doesn't happen too often with these tests:
- Task is migrated to another CPU.
- The task has high demand, and should take the target CPU to higher
OPPs.
- And the target CPU doesn't call into the cpufreq governor until the
next tick.
Based on initial work from Steve Muckle.
Signed-off-by: Viresh Kumar <viresh.kumar@linaro.org>
Acked-by: Saravana Kannan <skannan@codeaurora.org>
Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
After commit 62611cb912 (intel_pstate: delete scheduler hook in HWP
mode) the INTEL_PSTATE_HWP_SAMPLING_INTERVAL is not used anywhere in
the code, so drop it.
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
The ->get callback in the intel_pstate structure was mostly there
for the scaling_cur_freq sysfs attribute to work, but after commit
f8475cef90 (x86: use common aperfmperf_khz_on_cpu() to calculate
KHz using APERF/MPERF) that attribute uses arch_freq_get_on_cpu()
provided by the x86 arch code on all processors supported by
intel_pstate, so it doesn't need the ->get callback from the
driver any more.
Moreover, the very presence of the ->get callback in the intel_pstate
structure causes the cpuinfo_cur_freq attribute to be present when
intel_pstate operates in the active mode, which is bogus, because
the role of that attribute is to return the current CPU frequency
as seen by the hardware. For intel_pstate, though, this is just an
average frequency and not really current, but computed for the
previous sampling interval (the actual current frequency may be
way different at the point this value is obtained by reading from
cpuinfo_cur_freq), and after commit 82b4e03e01 (intel_pstate: skip
scheduler hook when in "performance" mode) the value in
cpuinfo_cur_freq may be stale or just 0, depending on the driver's
operation mode. In fact, however, on the hardware supported by
intel_pstate there is no way to read the current CPU frequency
from it, so the cpuinfo_cur_freq attribute should not be present
at all when this driver is in use.
For this reason, drop intel_pstate_get() and clear the ->get
callback pointer pointing to it, so that the cpuinfo_cur_freq is
not present for intel_pstate in the active mode any more.
Fixes: 82b4e03e01 (intel_pstate: skip scheduler hook when in "performance" mode)
Reported-by: Huaisheng Ye <yehs1@lenovo.com>
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
Acked-by: Viresh Kumar <viresh.kumar@linaro.org>
All systems use the same P-state selection "powersave" algorithm
in the active mode if HWP is not used, so there's no need to provide
a pointer for it in struct pstate_funcs any more.
Drop ->update_util from struct pstate_funcs and make
intel_pstate_set_update_util_hook() use intel_pstate_update_util()
directly.
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
All systems with a defined ACPI preferred profile that are not
"servers" have been using the load-based P-state selection algorithm
in intel_pstate since 4.12-rc1 (mobile systems and laptops have been
using it since 4.10-rc1) and no problems with it have been reported
to date. In particular, no regressions with respect to the PID-based
P-state selection have been reported. Also testing indicates that
the P-state selection algorithm based on CPU load is generally on par
with the PID-based algorithm performance-wise, and for some workloads
it turns out to be better than the other one, while being more
straightforward and easier to understand at the same time.
Moreover, the PID-based P-state selection algorithm in intel_pstate
is known to be unstable in some situation and generally problematic,
the issues with it are hard to address and it has become a
significant maintenance burden.
For these reasons, make intel_pstate use the "powersave" P-state
selection algorithm based on CPU load in the active mode on all
systems and drop the PID-based P-state selection code along with
all things related to it from the driver. Also update the
documentation accordingly.
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
The transition_latency field isn't used for drivers with ->setpolicy()
callback present and there is no point setting it from the drivers.
Signed-off-by: Viresh Kumar <viresh.kumar@linaro.org>
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
The busy percent calculated for the Knights Landing (KNL) platform
is 1024 times smaller than the correct busy value. This causes
performance to get stuck at the lowest ratio.
The scaling algorithm used for KNL is performance-based, but it still
looks at the CPU load to set the scaled busy factor to 0 when the
load is less than 1 percent. In this case, since the computed load
is 1024x smaller than it should be, the scaled busy factor will
always be 0, irrespective of CPU business.
This needs a fix similar to the turbostat one in commit b2b34dfe4d
(tools/power turbostat: KNL workaround for %Busy and Avg_MHz).
For this reason, add one more callback to processor-specific
callbacks to specify an MPERF multiplier represented by a number of
bit positions to shift the value of that register to the left to
copmensate for its rate difference with respect to the TSC. This
shift value is used during CPU busy calculations.
Fixes: ffb810563c (intel_pstate: Avoid getting stuck in high P-states when idle)
Reported-and-tested-by: Artem Bityutskiy <artem.bityutskiy@linux.intel.com>
Signed-off-by: Srinivas Pandruvada <srinivas.pandruvada@linux.intel.com>
Cc: 4.6+ <stable@vger.kernel.org> # 4.6+
[ rjw: Changelog ]
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
When the minimum performance limit percentage is set to the power-up
default, it is possible that minimum performance ratio is off by one.
In the set_policy() callback the minimum ratio is calculated by
applying global.min_perf_pct to turbo_ratio and rounding up, but the
power-up default global.min_perf_pct is already rounded up to the
next percent in min_perf_pct_min(). That results in two round up
operations, so for the default min_perf_pct one of them is not
required.
It is better to remove rounding up in min_perf_pct_min() as this
matches the displayed min_perf_pct prior to commit c5a2ee7dde
(cpufreq: intel_pstate: Active mode P-state limits rework) in 4.12.
For example on a platform with max turbo ratio of 37 and minimum
ratio of 10, the min_perf_pct resulted in 28 with the above commit.
Before this commit it was 27 and it will be the same after this
change.
Fixes: 1a4fe38add (cpufreq: intel_pstate: Remove max/min fractions to limit performance)
Reported-by: Artem Bityutskiy <artem.bityutskiy@linux.intel.com>
Signed-off-by: Srinivas Pandruvada <srinivas.pandruvada@linux.intel.com>
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
attribute_groups are not supposed to change at runtime. All functions
working with attribute_groups provided by <linux/sysfs.h> work with const
attribute_group. So mark the non-const structs as const.
File size before:
text data bss dec hex filename
15197 2552 40 17789 457d drivers/cpufreq/intel_pstate.o
File size After adding 'const':
text data bss dec hex filename
15261 2488 40 17789 457d drivers/cpufreq/intel_pstate.o
Signed-off-by: Arvind Yadav <arvind.yadav.cs@gmail.com>
Acked-by: Viresh Kumar <viresh.kumar@linaro.org>
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
* pm-cpufreq:
cpufreq / CPPC: Initialize policy->min to lowest nonlinear performance
cpufreq: sfi: make freq_table static
cpufreq: exynos5440: Fix inconsistent indenting
cpufreq: imx6q: imx6ull should use the same flow as imx6ul
cpufreq: dt: Add support for hi3660
* intel_pstate:
cpufreq: Update scaling_cur_freq documentation
cpufreq: intel_pstate: Clean up after performance governor changes
intel_pstate: skip scheduler hook when in "performance" mode
intel_pstate: delete scheduler hook in HWP mode
x86: use common aperfmperf_khz_on_cpu() to calculate KHz using APERF/MPERF
cpufreq: intel_pstate: Remove max/min fractions to limit performance
x86: do not use cpufreq_quick_get() for /proc/cpuinfo "cpu MHz"
* pm-cpuidle:
cpuidle: menu: allow state 0 to be disabled
intel_idle: Use more common logging style
x86/ACPI/cstate: Allow ACPI C1 FFH MWAIT use on AMD systems
ARM: cpuidle: Support asymmetric idle definition
* pm-tools:
cpupower: Add support for new AMD family 0x17
cpupower: Fix bug where return value was not used
tools/power turbostat: update version number
tools/power turbostat: decode MSR_IA32_MISC_ENABLE only on Intel
tools/power turbostat: stop migrating, unless '-m'
tools/power turbostat: if --debug, print sampling overhead
tools/power turbostat: hide SKL counters, when not requested
intel_pstate: use updated msr-index.h HWP.EPP values
tools/power x86_energy_perf_policy: support HWP.EPP
x86: msr-index.h: fix shifts to ULL results in HWP macros.
x86: msr-index.h: define HWP.EPP values
x86: msr-index.h: define EPB mid-points
After commit 82b4e03e01 (intel_pstate: skip scheduler hook when in
"performance" mode) get_target_pstate_use_performance() and
get_target_pstate_use_cpu_load() are never called if scaling_governor
is "performance", so drop the CPUFREQ_POLICY_PERFORMANCE checks from
them as they will never trigger anyway.
Moreover, the documentation needs to be updated to reflect the change
made by the above commit, so do that too.
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
Acked-by: Srinivas Pandruvada <srinivas.pandruvada@linux.intel.com>
When the governor is set to "performance", intel_pstate does not
need the scheduler hook for doing any calculations. Under these
conditions, its only purpose is to continue to maintain
cpufreq/scaling_cur_freq.
The cpufreq/scaling_cur_freq sysfs attribute is now provided by
shared x86 cpufreq code on modern x86 systems, including
all systems supported by the intel_pstate driver.
So in "performance" governor mode, the scheduler hook can be skipped.
This applies to both in Software and Hardware P-state control modes.
Suggested-by: Srinivas Pandruvada <srinivas.pandruvada@linux.intel.com>
Signed-off-by: Len Brown <len.brown@intel.com>
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
The cpufreq/scaling_cur_freq sysfs attribute is now provided by
shared x86 cpufreq code on modern x86 systems, including
all systems supported by the intel_pstate driver.
In HWP mode, maintaining that value was the sole purpose of
the scheduler hook, intel_pstate_update_util_hwp(),
so it can now be removed.
Signed-off-by: Len Brown <len.brown@intel.com>
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
In the current model the max/min perf limits are a fraction of current
user space limits to the allowed max_freq or 100% for global limits.
This results in wrong ratio limits calculation because of rounding
issues for some user space limits.
Initially we tried to solve this issue by issue by having more shift
bits to increase precision. Still there are isolated cases where we still
have error.
This can be avoided by using ratios all together. Since the way we get
cpuinfo.max_freq is by multiplying scaling factor to max ratio, we can
easily keep the max/min ratios in terms of ratios and not fractions.
For example:
if the max ratio = 36
cpuinfo.max_freq = 36 * 100000 = 3600000
Suppose user space sets a limit of 1200000, then we can calculate
max ratio limit as
= 36 * 1200000 / 3600000
= 12
This will be correct for any user limits.
The other advantage is that, we don't need to do any calculation in the
fast path as ratio limit is already calculated via set_policy() callback.
Signed-off-by: Srinivas Pandruvada <srinivas.pandruvada@linux.intel.com>
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
Commit c5a2ee7dde (cpufreq: intel_pstate: Active mode P-state
limits rework) incorrectly assumed that pstate.turbo_pstate would
always be nonzero for CPU0 in min_perf_pct_min() if
cpufreq_register_driver() had succeeded which may not be the case
in virtualized environments.
If that assumption doesn't hold, it leads to an early crash on boot
in intel_pstate_register_driver(), so add a sanity check to
min_perf_pct_min() to prevent the crash from happening.
Fixes: c5a2ee7dde (cpufreq: intel_pstate: Active mode P-state limits rework)
Reported-and-tested-by: Jongman Heo <jongman.heo@samsung.com>
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
intel_pstate exports sysfs attributes for setting and observing HWP.EPP.
These attributes use strings to describe 4 operating states, and
inside the driver, these strings are mapped to numerical register
values.
The authorative mapping between the strings and numerical HWP.EPP values
are now globally defined in msr-index.h, replacing the out-dated
mapping that were open-coded into intel_pstate.c
new old string
--- --- ------
0 0 performance
128 64 balance_performance
192 128 balance_power
255 192 power
Note that the HW and BIOS default value on most system is 128,
which intel_pstate will now call "balance_performance"
while it used to call it "balance_power".
Signed-off-by: Len Brown <len.brown@intel.com>
Make the schedutil governor take the initial (default) value of the
rate_limit_us sysfs attribute from the (new) transition_delay_us
policy parameter (to be set by the scaling driver).
That will allow scaling drivers to make schedutil use smaller default
values of rate_limit_us and reduce the default average time interval
between consecutive frequency changes.
Make intel_pstate set transition_delay_us to 500.
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
Acked-by: Viresh Kumar <viresh.kumar@linaro.org>
Use same parameters as INTEL_FAM6_ATOM_GOLDMONT to enable
Gemini Lake.
Signed-off-by: Box, David E <david.e.box@intel.com>
Signed-off-by: Srinivas Pandruvada <srinivas.pandruvada@linux.intel.com>
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
Some computations in intel_pstate_get_min_max() are not necessary
and one of its two callers doesn't even use the full result.
First off, the fixed-point value of cpu->max_perf represents a
non-negative number between 0 and 1 inclusive and cpu->min_perf
cannot be greater than cpu->max_perf. It is not necessary to check
those conditions every time the numbers in question are used.
Moreover, since intel_pstate_max_within_limits() only needs the
upper boundary, it doesn't make sense to compute the lower one in
there and returning min and max from intel_pstate_get_min_max()
via pointers doesn't look particularly nice.
For the above reasons, drop intel_pstate_get_min_max(), add a helper
to get the base P-state for min/max computations and carry out them
directly in the previous callers of intel_pstate_get_min_max().
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
intel_pstate_hwp_set() is the only function walking policy->cpus
in intel_pstate. The rest of the code simply assumes one CPU per
policy, including the initialization code.
Therefore it doesn't make sense for intel_pstate_hwp_set() to
walk policy->cpus as it is guaranteed to have only one bit set
for policy->cpu.
For this reason, rearrange intel_pstate_hwp_set() to take the CPU
number as the argument and drop the loop over policy->cpus from it.
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
Add a new function pid_in_use() to return the information on whether
or not the PID-based P-state selection algorithm is in use.
That allows a couple of complicated conditions in the code to be
reduced to simple checks against the new function's return value.
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
The cpu_defaults structure is redundant, because it only contains
one member of type struct pstate_funcs which can be used directly
instead of struct cpu_defaults.
For this reason, drop struct cpu_defaults, use struct pstate_funcs
directly instead of it where applicable and rename all of the
variables of that type accordingly.
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
Move the definitions of the cpu_defaults structures after the
definitions of utilization update callback routines to avoid
extra declarations of the latter.
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
Avoid using extra function pointers during P-state selection by
dropping the get_target_pstate member from struct pstate_funcs,
adding a new update_util callback to it (to be registered with
the CPU scheduler as the utilization update callback in the active
mode) and reworking the utilization update callback routines to
invoke specific P-state selection functions directly.
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
Notice that some overhead in the utilization update callbacks
registered by intel_pstate in the active mode can be avoided if
those callbacks are tailored to specific configurations of the
driver. For example, the utilization update callback for the HWP
enabled case only needs to update the average CPU performance
periodically whereas the utilization update callback for the
PID-based algorithm does not need to take IO-wait boosting into
account and so on.
With that in mind, define three utilization update callbacks for
three different use cases: HWP enabled, the CPU load "powersave"
P-state selection algorithm and the PID-based "powersave" P-state
selection algorithm and modify the driver initialization to
choose the callback matching its current configuration.
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
