xref: /linux/Documentation/admin-guide/pm/intel_pstate.rst (revision 7fc2cd2e4b398c57c9cf961cfea05eadbf34c05c)
1.. SPDX-License-Identifier: GPL-2.0
2.. include:: <isonum.txt>
3
4===============================================
5``intel_pstate`` CPU Performance Scaling Driver
6===============================================
7
8:Copyright: |copy| 2017 Intel Corporation
9
10:Author: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
11
12
13General Information
14===================
15
16``intel_pstate`` is a part of the
17:doc:`CPU performance scaling subsystem <cpufreq>` in the Linux kernel
18(``CPUFreq``).  It is a scaling driver for the Sandy Bridge and later
19generations of Intel processors.  Note, however, that some of those processors
20may not be supported.  [To understand ``intel_pstate`` it is necessary to know
21how ``CPUFreq`` works in general, so this is the time to read
22Documentation/admin-guide/pm/cpufreq.rst if you have not done that yet.]
23
24For the processors supported by ``intel_pstate``, the P-state concept is broader
25than just an operating frequency or an operating performance point (see the
26LinuxCon Europe 2015 presentation by Kristen Accardi [1]_ for more
27information about that).  For this reason, the representation of P-states used
28by ``intel_pstate`` internally follows the hardware specification (for details
29refer to Intel Software Developer’s Manual [2]_).  However, the ``CPUFreq`` core
30uses frequencies for identifying operating performance points of CPUs and
31frequencies are involved in the user space interface exposed by it, so
32``intel_pstate`` maps its internal representation of P-states to frequencies too
33(fortunately, that mapping is unambiguous).  At the same time, it would not be
34practical for ``intel_pstate`` to supply the ``CPUFreq`` core with a table of
35available frequencies due to the possible size of it, so the driver does not do
36that.  Some functionality of the core is limited by that.
37
38Since the hardware P-state selection interface used by ``intel_pstate`` is
39available at the logical CPU level, the driver always works with individual
40CPUs.  Consequently, if ``intel_pstate`` is in use, every ``CPUFreq`` policy
41object corresponds to one logical CPU and ``CPUFreq`` policies are effectively
42equivalent to CPUs.  In particular, this means that they become "inactive" every
43time the corresponding CPU is taken offline and need to be re-initialized when
44it goes back online.
45
46``intel_pstate`` is not modular, so it cannot be unloaded, which means that the
47only way to pass early-configuration-time parameters to it is via the kernel
48command line.  However, its configuration can be adjusted via ``sysfs`` to a
49great extent.  In some configurations it even is possible to unregister it via
50``sysfs`` which allows another ``CPUFreq`` scaling driver to be loaded and
51registered (see :ref:`below <status_attr>`).
52
53.. _operation_modes:
54
55Operation Modes
56===============
57
58``intel_pstate`` can operate in two different modes, active or passive.  In the
59active mode, it uses its own internal performance scaling governor algorithm or
60allows the hardware to do performance scaling by itself, while in the passive
61mode it responds to requests made by a generic ``CPUFreq`` governor implementing
62a certain performance scaling algorithm.  Which of them will be in effect
63depends on what kernel command line options are used and on the capabilities of
64the processor.
65
66.. _active_mode:
67
68Active Mode
69-----------
70
71This is the default operation mode of ``intel_pstate`` for processors with
72hardware-managed P-states (HWP) support.  If it works in this mode, the
73``scaling_driver`` policy attribute in ``sysfs`` for all ``CPUFreq`` policies
74contains the string "intel_pstate".
75
76In this mode the driver bypasses the scaling governors layer of ``CPUFreq`` and
77provides its own scaling algorithms for P-state selection.  Those algorithms
78can be applied to ``CPUFreq`` policies in the same way as generic scaling
79governors (that is, through the ``scaling_governor`` policy attribute in
80``sysfs``).  [Note that different P-state selection algorithms may be chosen for
81different policies, but that is not recommended.]
82
83They are not generic scaling governors, but their names are the same as the
84names of some of those governors.  Moreover, confusingly enough, they generally
85do not work in the same way as the generic governors they share the names with.
86For example, the ``powersave`` P-state selection algorithm provided by
87``intel_pstate`` is not a counterpart of the generic ``powersave`` governor
88(roughly, it corresponds to the ``schedutil`` and ``ondemand`` governors).
89
90There are two P-state selection algorithms provided by ``intel_pstate`` in the
91active mode: ``powersave`` and ``performance``.  The way they both operate
92depends on whether or not the hardware-managed P-states (HWP) feature has been
93enabled in the processor and possibly on the processor model.
94
95Which of the P-state selection algorithms is used by default depends on the
96:c:macro:`CONFIG_CPU_FREQ_DEFAULT_GOV_PERFORMANCE` kernel configuration option.
97Namely, if that option is set, the ``performance`` algorithm will be used by
98default, and the other one will be used by default if it is not set.
99
100.. _active_mode_hwp:
101
102Active Mode With HWP
103~~~~~~~~~~~~~~~~~~~~
104
105If the processor supports the HWP feature, it will be enabled during the
106processor initialization and cannot be disabled after that.  It is possible
107to avoid enabling it by passing the ``intel_pstate=no_hwp`` argument to the
108kernel in the command line.
109
110If the HWP feature has been enabled, ``intel_pstate`` relies on the processor to
111select P-states by itself, but still it can give hints to the processor's
112internal P-state selection logic.  What those hints are depends on which P-state
113selection algorithm has been applied to the given policy (or to the CPU it
114corresponds to).
115
116Even though the P-state selection is carried out by the processor automatically,
117``intel_pstate`` registers utilization update callbacks with the CPU scheduler
118in this mode.  However, they are not used for running a P-state selection
119algorithm, but for periodic updates of the current CPU frequency information to
120be made available from the ``scaling_cur_freq`` policy attribute in ``sysfs``.
121
122HWP + ``performance``
123.....................
124
125In this configuration ``intel_pstate`` will write 0 to the processor's
126Energy-Performance Preference (EPP) knob (if supported) or its
127Energy-Performance Bias (EPB) knob (otherwise), which means that the processor's
128internal P-state selection logic is expected to focus entirely on performance.
129
130This will override the EPP/EPB setting coming from the ``sysfs`` interface
131(see :ref:`energy_performance_hints` below).  Moreover, any attempts to change
132the EPP/EPB to a value different from 0 ("performance") via ``sysfs`` in this
133configuration will be rejected.
134
135Also, in this configuration the range of P-states available to the processor's
136internal P-state selection logic is always restricted to the upper boundary
137(that is, the maximum P-state that the driver is allowed to use).
138
139HWP + ``powersave``
140...................
141
142In this configuration ``intel_pstate`` will set the processor's
143Energy-Performance Preference (EPP) knob (if supported) or its
144Energy-Performance Bias (EPB) knob (otherwise) to whatever value it was
145previously set to via ``sysfs`` (or whatever default value it was
146set to by the platform firmware).  This usually causes the processor's
147internal P-state selection logic to be less performance-focused.
148
149Active Mode Without HWP
150~~~~~~~~~~~~~~~~~~~~~~~
151
152This operation mode is optional for processors that do not support the HWP
153feature or when the ``intel_pstate=no_hwp`` argument is passed to the kernel in
154the command line.  The active mode is used in those cases if the
155``intel_pstate=active`` argument is passed to the kernel in the command line.
156In this mode ``intel_pstate`` may refuse to work with processors that are not
157recognized by it.  [Note that ``intel_pstate`` will never refuse to work with
158any processor with the HWP feature enabled.]
159
160In this mode ``intel_pstate`` registers utilization update callbacks with the
161CPU scheduler in order to run a P-state selection algorithm, either
162``powersave`` or ``performance``, depending on the ``scaling_governor`` policy
163setting in ``sysfs``.  The current CPU frequency information to be made
164available from the ``scaling_cur_freq`` policy attribute in ``sysfs`` is
165periodically updated by those utilization update callbacks too.
166
167``performance``
168...............
169
170Without HWP, this P-state selection algorithm is always the same regardless of
171the processor model and platform configuration.
172
173It selects the maximum P-state it is allowed to use, subject to limits set via
174``sysfs``, every time the driver configuration for the given CPU is updated
175(e.g. via ``sysfs``).
176
177This is the default P-state selection algorithm if the
178:c:macro:`CONFIG_CPU_FREQ_DEFAULT_GOV_PERFORMANCE` kernel configuration option
179is set.
180
181``powersave``
182.............
183
184Without HWP, this P-state selection algorithm is similar to the algorithm
185implemented by the generic ``schedutil`` scaling governor except that the
186utilization metric used by it is based on numbers coming from feedback
187registers of the CPU.  It generally selects P-states proportional to the
188current CPU utilization.
189
190This algorithm is run by the driver's utilization update callback for the
191given CPU when it is invoked by the CPU scheduler, but not more often than
192every 10 ms.  Like in the ``performance`` case, the hardware configuration
193is not touched if the new P-state turns out to be the same as the current
194one.
195
196This is the default P-state selection algorithm if the
197:c:macro:`CONFIG_CPU_FREQ_DEFAULT_GOV_PERFORMANCE` kernel configuration option
198is not set.
199
200.. _passive_mode:
201
202Passive Mode
203------------
204
205This is the default operation mode of ``intel_pstate`` for processors without
206hardware-managed P-states (HWP) support.  It is always used if the
207``intel_pstate=passive`` argument is passed to the kernel in the command line
208regardless of whether or not the given processor supports HWP.  [Note that the
209``intel_pstate=no_hwp`` setting causes the driver to start in the passive mode
210if it is not combined with ``intel_pstate=active``.]  Like in the active mode
211without HWP support, in this mode ``intel_pstate`` may refuse to work with
212processors that are not recognized by it if HWP is prevented from being enabled
213through the kernel command line.
214
215If the driver works in this mode, the ``scaling_driver`` policy attribute in
216``sysfs`` for all ``CPUFreq`` policies contains the string "intel_cpufreq".
217Then, the driver behaves like a regular ``CPUFreq`` scaling driver.  That is,
218it is invoked by generic scaling governors when necessary to talk to the
219hardware in order to change the P-state of a CPU (in particular, the
220``schedutil`` governor can invoke it directly from scheduler context).
221
222While in this mode, ``intel_pstate`` can be used with all of the (generic)
223scaling governors listed by the ``scaling_available_governors`` policy attribute
224in ``sysfs`` (and the P-state selection algorithms described above are not
225used).  Then, it is responsible for the configuration of policy objects
226corresponding to CPUs and provides the ``CPUFreq`` core (and the scaling
227governors attached to the policy objects) with accurate information on the
228maximum and minimum operating frequencies supported by the hardware (including
229the so-called "turbo" frequency ranges).  In other words, in the passive mode
230the entire range of available P-states is exposed by ``intel_pstate`` to the
231``CPUFreq`` core.  However, in this mode the driver does not register
232utilization update callbacks with the CPU scheduler and the ``scaling_cur_freq``
233information comes from the ``CPUFreq`` core (and is the last frequency selected
234by the current scaling governor for the given policy).
235
236
237.. _turbo:
238
239Turbo P-states Support
240======================
241
242In the majority of cases, the entire range of P-states available to
243``intel_pstate`` can be divided into two sub-ranges that correspond to
244different types of processor behavior, above and below a boundary that
245will be referred to as the "turbo threshold" in what follows.
246
247The P-states above the turbo threshold are referred to as "turbo P-states" and
248the whole sub-range of P-states they belong to is referred to as the "turbo
249range".  These names are related to the Turbo Boost technology allowing a
250multicore processor to opportunistically increase the P-state of one or more
251cores if there is enough power to do that and if that is not going to cause the
252thermal envelope of the processor package to be exceeded.
253
254Specifically, if software sets the P-state of a CPU core within the turbo range
255(that is, above the turbo threshold), the processor is permitted to take over
256performance scaling control for that core and put it into turbo P-states of its
257choice going forward.  However, that permission is interpreted differently by
258different processor generations.  Namely, the Sandy Bridge generation of
259processors will never use any P-states above the last one set by software for
260the given core, even if it is within the turbo range, whereas all of the later
261processor generations will take it as a license to use any P-states from the
262turbo range, even above the one set by software.  In other words, on those
263processors setting any P-state from the turbo range will enable the processor
264to put the given core into all turbo P-states up to and including the maximum
265supported one as it sees fit.
266
267One important property of turbo P-states is that they are not sustainable.  More
268precisely, there is no guarantee that any CPUs will be able to stay in any of
269those states indefinitely, because the power distribution within the processor
270package may change over time  or the thermal envelope it was designed for might
271be exceeded if a turbo P-state was used for too long.
272
273In turn, the P-states below the turbo threshold generally are sustainable.  In
274fact, if one of them is set by software, the processor is not expected to change
275it to a lower one unless in a thermal stress or a power limit violation
276situation (a higher P-state may still be used if it is set for another CPU in
277the same package at the same time, for example).
278
279Some processors allow multiple cores to be in turbo P-states at the same time,
280but the maximum P-state that can be set for them generally depends on the number
281of cores running concurrently.  The maximum turbo P-state that can be set for 3
282cores at the same time usually is lower than the analogous maximum P-state for
2832 cores, which in turn usually is lower than the maximum turbo P-state that can
284be set for 1 core.  The one-core maximum turbo P-state is thus the maximum
285supported one overall.
286
287The maximum supported turbo P-state, the turbo threshold (the maximum supported
288non-turbo P-state) and the minimum supported P-state are specific to the
289processor model and can be determined by reading the processor's model-specific
290registers (MSRs).  Moreover, some processors support the Configurable TDP
291(Thermal Design Power) feature and, when that feature is enabled, the turbo
292threshold effectively becomes a configurable value that can be set by the
293platform firmware.
294
295Unlike ``_PSS`` objects in the ACPI tables, ``intel_pstate`` always exposes
296the entire range of available P-states, including the whole turbo range, to the
297``CPUFreq`` core and (in the passive mode) to generic scaling governors.  This
298generally causes turbo P-states to be set more often when ``intel_pstate`` is
299used relative to ACPI-based CPU performance scaling (see
300:ref:`below <acpi-cpufreq>` for more information).
301
302Moreover, since ``intel_pstate`` always knows what the real turbo threshold is
303(even if the Configurable TDP feature is enabled in the processor), its
304``no_turbo`` attribute in ``sysfs`` (described :ref:`below <no_turbo_attr>`) should
305work as expected in all cases (that is, if set to disable turbo P-states, it
306always should prevent ``intel_pstate`` from using them).
307
308
309Processor Support
310=================
311
312To handle a given processor ``intel_pstate`` requires a number of different
313pieces of information on it to be known, including:
314
315 * The minimum supported P-state.
316
317 * The maximum supported :ref:`non-turbo P-state <turbo>`.
318
319 * Whether or not turbo P-states are supported at all.
320
321 * The maximum supported :ref:`one-core turbo P-state <turbo>` (if turbo
322   P-states are supported).
323
324 * The scaling formula to translate the driver's internal representation
325   of P-states into frequencies and the other way around.
326
327Generally, ways to obtain that information are specific to the processor model
328or family.  Although it often is possible to obtain all of it from the processor
329itself (using model-specific registers), there are cases in which hardware
330manuals need to be consulted to get to it too.
331
332For this reason, there is a list of supported processors in ``intel_pstate`` and
333the driver initialization will fail if the detected processor is not in that
334list, unless it supports the HWP feature.  [The interface to obtain all of the
335information listed above is the same for all of the processors supporting the
336HWP feature, which is why ``intel_pstate`` works with all of them.]
337
338
339Support for Hybrid Processors
340=============================
341
342Some processors supported by ``intel_pstate`` contain two or more types of CPU
343cores differing by the maximum turbo P-state, performance vs power characteristics,
344cache sizes, and possibly other properties.  They are commonly referred to as
345hybrid processors.  To support them, ``intel_pstate`` requires HWP to be enabled
346and it assumes the HWP performance units to be the same for all CPUs in the
347system, so a given HWP performance level always represents approximately the
348same physical performance regardless of the core (CPU) type.
349
350Hybrid Processors with SMT
351--------------------------
352
353On systems where SMT (Simultaneous Multithreading), also referred to as
354HyperThreading (HT) in the context of Intel processors, is enabled on at least
355one core, ``intel_pstate`` assigns performance-based priorities to CPUs.  Namely,
356the priority of a given CPU reflects its highest HWP performance level which
357causes the CPU scheduler to generally prefer more performant CPUs, so the less
358performant CPUs are used when the other ones are fully loaded.  However, SMT
359siblings (that is, logical CPUs sharing one physical core) are treated in a
360special way such that if one of them is in use, the effective priority of the
361other ones is lowered below the priorities of the CPUs located in the other
362physical cores.
363
364This approach maximizes performance in the majority of cases, but unfortunately
365it also leads to excessive energy usage in some important scenarios, like video
366playback, which is not generally desirable.  While there is no other viable
367choice with SMT enabled because the effective capacity and utilization of SMT
368siblings are hard to determine, hybrid processors without SMT can be handled in
369more energy-efficient ways.
370
371.. _CAS:
372
373Capacity-Aware Scheduling Support
374---------------------------------
375
376The capacity-aware scheduling (CAS) support in the CPU scheduler is enabled by
377``intel_pstate`` by default on hybrid processors without SMT.  CAS generally
378causes the scheduler to put tasks on a CPU so long as there is a sufficient
379amount of spare capacity on it, and if the utilization of a given task is too
380high for it, the task will need to go somewhere else.
381
382Since CAS takes CPU capacities into account, it does not require CPU
383prioritization and it allows tasks to be distributed more symmetrically among
384the more performant and less performant CPUs.  Once placed on a CPU with enough
385capacity to accommodate it, a task may just continue to run there regardless of
386whether or not the other CPUs are fully loaded, so on average CAS reduces the
387utilization of the more performant CPUs which causes the energy usage to be more
388balanced because the more performant CPUs are generally less energy-efficient
389than the less performant ones.
390
391In order to use CAS, the scheduler needs to know the capacity of each CPU in
392the system and it needs to be able to compute scale-invariant utilization of
393CPUs, so ``intel_pstate`` provides it with the requisite information.
394
395First of all, the capacity of each CPU is represented by the ratio of its highest
396HWP performance level, multiplied by 1024, to the highest HWP performance level
397of the most performant CPU in the system, which works because the HWP performance
398units are the same for all CPUs.  Second, the frequency-invariance computations,
399carried out by the scheduler to always express CPU utilization in the same units
400regardless of the frequency it is currently running at, are adjusted to take the
401CPU capacity into account.  All of this happens when ``intel_pstate`` has
402registered itself with the ``CPUFreq`` core and it has figured out that it is
403running on a hybrid processor without SMT.
404
405Energy-Aware Scheduling Support
406-------------------------------
407
408If ``CONFIG_ENERGY_MODEL`` has been set during kernel configuration and
409``intel_pstate`` runs on a hybrid processor without SMT, in addition to enabling
410:ref:`CAS` it registers an Energy Model for the processor.  This allows the
411Energy-Aware Scheduling (EAS) support to be enabled in the CPU scheduler if
412``schedutil`` is used as the  ``CPUFreq`` governor which requires ``intel_pstate``
413to operate in the :ref:`passive mode <passive_mode>`.
414
415The Energy Model registered by ``intel_pstate`` is artificial (that is, it is
416based on abstract cost values and it does not include any real power numbers)
417and it is relatively simple to avoid unnecessary computations in the scheduler.
418There is a performance domain in it for every CPU in the system and the cost
419values for these performance domains have been chosen so that running a task on
420a less performant (small) CPU appears to be always cheaper than running that
421task on a more performant (big) CPU.  However, for two CPUs of the same type,
422the cost difference depends on their current utilization, and the CPU whose
423current utilization is higher generally appears to be a more expensive
424destination for a given task.  This helps to balance the load among CPUs of the
425same type.
426
427Since EAS works on top of CAS, high-utilization tasks are always migrated to
428CPUs with enough capacity to accommodate them, but thanks to EAS, low-utilization
429tasks tend to be placed on the CPUs that look less expensive to the scheduler.
430Effectively, this causes the less performant and less loaded CPUs to be
431preferred as long as they have enough spare capacity to run the given task
432which generally leads to reduced energy usage.
433
434The Energy Model created by ``intel_pstate`` can be inspected by looking at
435the ``energy_model`` directory in ``debugfs`` (typlically mounted on
436``/sys/kernel/debug/``).
437
438
439User Space Interface in ``sysfs``
440=================================
441
442.. _global_attributes:
443
444Global Attributes
445-----------------
446
447``intel_pstate`` exposes several global attributes (files) in ``sysfs`` to
448control its functionality at the system level.  They are located in the
449``/sys/devices/system/cpu/intel_pstate/`` directory and affect all CPUs.
450
451Some of them are not present if the ``intel_pstate=per_cpu_perf_limits``
452argument is passed to the kernel in the command line.
453
454``max_perf_pct``
455	Maximum P-state the driver is allowed to set in percent of the
456	maximum supported performance level (the highest supported :ref:`turbo
457	P-state <turbo>`).
458
459	This attribute will not be exposed if the
460	``intel_pstate=per_cpu_perf_limits`` argument is present in the kernel
461	command line.
462
463``min_perf_pct``
464	Minimum P-state the driver is allowed to set in percent of the
465	maximum supported performance level (the highest supported :ref:`turbo
466	P-state <turbo>`).
467
468	This attribute will not be exposed if the
469	``intel_pstate=per_cpu_perf_limits`` argument is present in the kernel
470	command line.
471
472``num_pstates``
473	Number of P-states supported by the processor (between 0 and 255
474	inclusive) including both turbo and non-turbo P-states (see
475	:ref:`turbo`).
476
477	This attribute is present only if the value exposed by it is the same
478	for all of the CPUs in the system.
479
480	The value of this attribute is not affected by the ``no_turbo``
481	setting described :ref:`below <no_turbo_attr>`.
482
483	This attribute is read-only.
484
485``turbo_pct``
486	Ratio of the :ref:`turbo range <turbo>` size to the size of the entire
487	range of supported P-states, in percent.
488
489	This attribute is present only if the value exposed by it is the same
490	for all of the CPUs in the system.
491
492	This attribute is read-only.
493
494.. _no_turbo_attr:
495
496``no_turbo``
497	If set (equal to 1), the driver is not allowed to set any turbo P-states
498	(see :ref:`turbo`).  If unset (equal to 0, which is the
499	default), turbo P-states can be set by the driver.
500	[Note that ``intel_pstate`` does not support the general ``boost``
501	attribute (supported by some other scaling drivers) which is replaced
502	by this one.]
503
504	This attribute does not affect the maximum supported frequency value
505	supplied to the ``CPUFreq`` core and exposed via the policy interface,
506	but it affects the maximum possible value of per-policy P-state	limits
507	(see :ref:`policy_attributes_interpretation` below for details).
508
509``hwp_dynamic_boost``
510	This attribute is only present if ``intel_pstate`` works in the
511	:ref:`active mode with the HWP feature enabled <active_mode_hwp>` in
512	the processor.  If set (equal to 1), it causes the minimum P-state limit
513	to be increased dynamically for a short time whenever a task previously
514	waiting on I/O is selected to run on a given logical CPU (the purpose
515	of this mechanism is to improve performance).
516
517	This setting has no effect on logical CPUs whose minimum P-state limit
518	is directly set to the highest non-turbo P-state or above it.
519
520.. _status_attr:
521
522``status``
523	Operation mode of the driver: "active", "passive" or "off".
524
525	"active"
526		The driver is functional and in the :ref:`active mode
527		<active_mode>`.
528
529	"passive"
530		The driver is functional and in the :ref:`passive mode
531		<passive_mode>`.
532
533	"off"
534		The driver is not functional (it is not registered as a scaling
535		driver with the ``CPUFreq`` core).
536
537	This attribute can be written to in order to change the driver's
538	operation mode or to unregister it.  The string written to it must be
539	one of the possible values of it and, if successful, the write will
540	cause the driver to switch over to the operation mode represented by
541	that string - or to be unregistered in the "off" case.  [Actually,
542	switching over from the active mode to the passive mode or the other
543	way around causes the driver to be unregistered and registered again
544	with a different set of callbacks, so all of its settings (the global
545	as well as the per-policy ones) are then reset to their default
546	values, possibly depending on the target operation mode.]
547
548``energy_efficiency``
549	This attribute is only present on platforms with CPUs matching the Kaby
550	Lake or Coffee Lake desktop CPU model. By default, energy-efficiency
551	optimizations are disabled on these CPU models if HWP is enabled.
552	Enabling energy-efficiency optimizations may limit maximum operating
553	frequency with or without the HWP feature.  With HWP enabled, the
554	optimizations are done only in the turbo frequency range.  Without it,
555	they are done in the entire available frequency range.  Setting this
556	attribute to "1" enables the energy-efficiency optimizations and setting
557	to "0" disables them.
558
559.. _policy_attributes_interpretation:
560
561Interpretation of Policy Attributes
562-----------------------------------
563
564The interpretation of some ``CPUFreq`` policy attributes described in
565Documentation/admin-guide/pm/cpufreq.rst is special with ``intel_pstate``
566as the current scaling driver and it generally depends on the driver's
567:ref:`operation mode <operation_modes>`.
568
569First of all, the values of the ``cpuinfo_max_freq``, ``cpuinfo_min_freq`` and
570``scaling_cur_freq`` attributes are produced by applying a processor-specific
571multiplier to the internal P-state representation used by ``intel_pstate``.
572Also, the values of the ``scaling_max_freq`` and ``scaling_min_freq``
573attributes are capped by the frequency corresponding to the maximum P-state that
574the driver is allowed to set.
575
576If the ``no_turbo`` :ref:`global attribute <no_turbo_attr>` is set, the driver
577is not allowed to use turbo P-states, so the maximum value of
578``scaling_max_freq`` and ``scaling_min_freq`` is limited to the maximum
579non-turbo P-state frequency.
580Accordingly, setting ``no_turbo`` causes ``scaling_max_freq`` and
581``scaling_min_freq`` to go down to that value if they were above it before.
582However, the old values of ``scaling_max_freq`` and ``scaling_min_freq`` will be
583restored after unsetting ``no_turbo``, unless these attributes have been written
584to after ``no_turbo`` was set.
585
586If ``no_turbo`` is not set, the maximum possible value of ``scaling_max_freq``
587and ``scaling_min_freq`` corresponds to the maximum supported turbo P-state,
588which also is the value of ``cpuinfo_max_freq`` in either case.
589
590Next, the following policy attributes have special meaning if
591``intel_pstate`` works in the :ref:`active mode <active_mode>`:
592
593``scaling_available_governors``
594	List of P-state selection algorithms provided by ``intel_pstate``.
595
596``scaling_governor``
597	P-state selection algorithm provided by ``intel_pstate`` currently in
598	use with the given policy.
599
600``scaling_cur_freq``
601	Frequency of the average P-state of the CPU represented by the given
602	policy for the time interval between the last two invocations of the
603	driver's utilization update callback by the CPU scheduler for that CPU.
604
605One more policy attribute is present if the HWP feature is enabled in the
606processor:
607
608``base_frequency``
609	Shows the base frequency of the CPU. Any frequency above this will be
610	in the turbo frequency range.
611
612The meaning of these attributes in the :ref:`passive mode <passive_mode>` is the
613same as for other scaling drivers.
614
615Additionally, the value of the ``scaling_driver`` attribute for ``intel_pstate``
616depends on the operation mode of the driver.  Namely, it is either
617"intel_pstate" (in the :ref:`active mode <active_mode>`) or "intel_cpufreq"
618(in the :ref:`passive mode <passive_mode>`).
619
620.. _pstate_limits_coordination:
621
622Coordination of P-State Limits
623------------------------------
624
625``intel_pstate`` allows P-state limits to be set in two ways: with the help of
626the ``max_perf_pct`` and ``min_perf_pct`` :ref:`global attributes
627<global_attributes>` or via the ``scaling_max_freq`` and ``scaling_min_freq``
628``CPUFreq`` policy attributes.  The coordination between those limits is based
629on the following rules, regardless of the current operation mode of the driver:
630
631 1. All CPUs are affected by the global limits (that is, none of them can be
632    requested to run faster than the global maximum and none of them can be
633    requested to run slower than the global minimum).
634
635 2. Each individual CPU is affected by its own per-policy limits (that is, it
636    cannot be requested to run faster than its own per-policy maximum and it
637    cannot be requested to run slower than its own per-policy minimum). The
638    effective performance depends on whether the platform supports per core
639    P-states, hyper-threading is enabled and on current performance requests
640    from other CPUs. When platform doesn't support per core P-states, the
641    effective performance can be more than the policy limits set on a CPU, if
642    other CPUs are requesting higher performance at that moment. Even with per
643    core P-states support, when hyper-threading is enabled, if the sibling CPU
644    is requesting higher performance, the other siblings will get higher
645    performance than their policy limits.
646
647 3. The global and per-policy limits can be set independently.
648
649In the :ref:`active mode with the HWP feature enabled <active_mode_hwp>`, the
650resulting effective values are written into hardware registers whenever the
651limits change in order to request its internal P-state selection logic to always
652set P-states within these limits.  Otherwise, the limits are taken into account
653by scaling governors (in the :ref:`passive mode <passive_mode>`) and by the
654driver every time before setting a new P-state for a CPU.
655
656Additionally, if the ``intel_pstate=per_cpu_perf_limits`` command line argument
657is passed to the kernel, ``max_perf_pct`` and ``min_perf_pct`` are not exposed
658at all and the only way to set the limits is by using the policy attributes.
659
660.. _energy_performance_hints:
661
662Energy vs Performance Hints
663---------------------------
664
665If the hardware-managed P-states (HWP) is enabled in the processor, additional
666attributes, intended to allow user space to help ``intel_pstate`` to adjust the
667processor's internal P-state selection logic by focusing it on performance or on
668energy-efficiency, or somewhere between the two extremes, are present in every
669``CPUFreq`` policy directory in ``sysfs``.  They are :
670
671``energy_performance_preference``
672	Current value of the energy vs performance hint for the given policy
673	(or the CPU represented by it).
674
675	The hint can be changed by writing to this attribute.
676
677``energy_performance_available_preferences``
678	List of strings that can be written to the
679	``energy_performance_preference`` attribute.
680
681	They represent different energy vs performance hints and should be
682	self-explanatory, except that ``default`` represents whatever hint
683	value was set by the platform firmware.
684
685Strings written to the ``energy_performance_preference`` attribute are
686internally translated to integer values written to the processor's
687Energy-Performance Preference (EPP) knob (if supported) or its
688Energy-Performance Bias (EPB) knob. It is also possible to write a positive
689integer value between 0 to 255, if the EPP feature is present. If the EPP
690feature is not present, writing integer value to this attribute is not
691supported. In this case, user can use the
692"/sys/devices/system/cpu/cpu*/power/energy_perf_bias" interface.
693
694[Note that tasks may by migrated from one CPU to another by the scheduler's
695load-balancing algorithm and if different energy vs performance hints are
696set for those CPUs, that may lead to undesirable outcomes.  To avoid such
697issues it is better to set the same energy vs performance hint for all CPUs
698or to pin every task potentially sensitive to them to a specific CPU.]
699
700.. _acpi-cpufreq:
701
702``intel_pstate`` vs ``acpi-cpufreq``
703====================================
704
705On the majority of systems supported by ``intel_pstate``, the ACPI tables
706provided by the platform firmware contain ``_PSS`` objects returning information
707that can be used for CPU performance scaling (refer to the ACPI specification
708[3]_ for details on the ``_PSS`` objects and the format of the information
709returned by them).
710
711The information returned by the ACPI ``_PSS`` objects is used by the
712``acpi-cpufreq`` scaling driver.  On systems supported by ``intel_pstate``
713the ``acpi-cpufreq`` driver uses the same hardware CPU performance scaling
714interface, but the set of P-states it can use is limited by the ``_PSS``
715output.
716
717On those systems each ``_PSS`` object returns a list of P-states supported by
718the corresponding CPU which basically is a subset of the P-states range that can
719be used by ``intel_pstate`` on the same system, with one exception: the whole
720:ref:`turbo range <turbo>` is represented by one item in it (the topmost one).
721By convention, the frequency returned by ``_PSS`` for that item is greater by
7221 MHz than the frequency of the highest non-turbo P-state listed by it, but the
723corresponding P-state representation (following the hardware specification)
724returned for it matches the maximum supported turbo P-state (or is the
725special value 255 meaning essentially "go as high as you can get").
726
727The list of P-states returned by ``_PSS`` is reflected by the table of
728available frequencies supplied by ``acpi-cpufreq`` to the ``CPUFreq`` core and
729scaling governors and the minimum and maximum supported frequencies reported by
730it come from that list as well.  In particular, given the special representation
731of the turbo range described above, this means that the maximum supported
732frequency reported by ``acpi-cpufreq`` is higher by 1 MHz than the frequency
733of the highest supported non-turbo P-state listed by ``_PSS`` which, of course,
734affects decisions made by the scaling governors, except for ``powersave`` and
735``performance``.
736
737For example, if a given governor attempts to select a frequency proportional to
738estimated CPU load and maps the load of 100% to the maximum supported frequency
739(possibly multiplied by a constant), then it will tend to choose P-states below
740the turbo threshold if ``acpi-cpufreq`` is used as the scaling driver, because
741in that case the turbo range corresponds to a small fraction of the frequency
742band it can use (1 MHz vs 1 GHz or more).  In consequence, it will only go to
743the turbo range for the highest loads and the other loads above 50% that might
744benefit from running at turbo frequencies will be given non-turbo P-states
745instead.
746
747One more issue related to that may appear on systems supporting the
748:ref:`Configurable TDP feature <turbo>` allowing the platform firmware to set
749the turbo threshold.  Namely, if that is not coordinated with the lists of
750P-states returned by ``_PSS`` properly, there may be more than one item
751corresponding to a turbo P-state in those lists and there may be a problem with
752avoiding the turbo range (if desirable or necessary).  Usually, to avoid using
753turbo P-states overall, ``acpi-cpufreq`` simply avoids using the topmost state
754listed by ``_PSS``, but that is not sufficient when there are other turbo
755P-states in the list returned by it.
756
757Apart from the above, ``acpi-cpufreq`` works like ``intel_pstate`` in the
758:ref:`passive mode <passive_mode>`, except that the number of P-states it can
759set is limited to the ones listed by the ACPI ``_PSS`` objects.
760
761
762Kernel Command Line Options for ``intel_pstate``
763================================================
764
765Several kernel command line options can be used to pass early-configuration-time
766parameters to ``intel_pstate`` in order to enforce specific behavior of it.  All
767of them have to be prepended with the ``intel_pstate=`` prefix.
768
769``disable``
770	Do not register ``intel_pstate`` as the scaling driver even if the
771	processor is supported by it.
772
773``active``
774	Register ``intel_pstate`` in the :ref:`active mode <active_mode>` to
775        start with.
776
777``passive``
778	Register ``intel_pstate`` in the :ref:`passive mode <passive_mode>` to
779	start with.
780
781``force``
782	Register ``intel_pstate`` as the scaling driver instead of
783	``acpi-cpufreq`` even if the latter is preferred on the given system.
784
785	This may prevent some platform features (such as thermal controls and
786	power capping) that rely on the availability of ACPI P-states
787	information from functioning as expected, so it should be used with
788	caution.
789
790	This option does not work with processors that are not supported by
791	``intel_pstate`` and on platforms where the ``pcc-cpufreq`` scaling
792	driver is used instead of ``acpi-cpufreq``.
793
794``no_hwp``
795	Do not enable the hardware-managed P-states (HWP) feature even if it is
796	supported by the processor.
797
798``hwp_only``
799	Register ``intel_pstate`` as the scaling driver only if the
800	hardware-managed P-states (HWP) feature is supported by the processor.
801
802``support_acpi_ppc``
803	Take ACPI ``_PPC`` performance limits into account.
804
805	If the preferred power management profile in the FADT (Fixed ACPI
806	Description Table) is set to "Enterprise Server" or "Performance
807	Server", the ACPI ``_PPC`` limits are taken into account by default
808	and this option has no effect.
809
810``per_cpu_perf_limits``
811	Use per-logical-CPU P-State limits (see
812        :ref:`pstate_limits_coordination` for details).
813
814``no_cas``
815	Do not enable :ref:`capacity-aware scheduling <CAS>` which is enabled
816        by default on hybrid systems without SMT.
817
818Diagnostics and Tuning
819======================
820
821Trace Events
822------------
823
824There are two static trace events that can be used for ``intel_pstate``
825diagnostics.  One of them is the ``cpu_frequency`` trace event generally used
826by ``CPUFreq``, and the other one is the ``pstate_sample`` trace event specific
827to ``intel_pstate``.  Both of them are triggered by ``intel_pstate`` only if
828it works in the :ref:`active mode <active_mode>`.
829
830The following sequence of shell commands can be used to enable them and see
831their output (if the kernel is generally configured to support event tracing)::
832
833 # cd /sys/kernel/tracing/
834 # echo 1 > events/power/pstate_sample/enable
835 # echo 1 > events/power/cpu_frequency/enable
836 # cat trace
837 gnome-terminal--4510  [001] ..s.  1177.680733: pstate_sample: core_busy=107 scaled=94 from=26 to=26 mperf=1143818 aperf=1230607 tsc=29838618 freq=2474476
838 cat-5235  [002] ..s.  1177.681723: cpu_frequency: state=2900000 cpu_id=2
839
840If ``intel_pstate`` works in the :ref:`passive mode <passive_mode>`, the
841``cpu_frequency`` trace event will be triggered either by the ``schedutil``
842scaling governor (for the policies it is attached to), or by the ``CPUFreq``
843core (for the policies with other scaling governors).
844
845``ftrace``
846----------
847
848The ``ftrace`` interface can be used for low-level diagnostics of
849``intel_pstate``.  For example, to check how often the function to set a
850P-state is called, the ``ftrace`` filter can be set to
851:c:func:`intel_pstate_set_pstate`::
852
853 # cd /sys/kernel/tracing/
854 # cat available_filter_functions | grep -i pstate
855 intel_pstate_set_pstate
856 intel_pstate_cpu_init
857 ...
858 # echo intel_pstate_set_pstate > set_ftrace_filter
859 # echo function > current_tracer
860 # cat trace | head -15
861 # tracer: function
862 #
863 # entries-in-buffer/entries-written: 80/80   #P:4
864 #
865 #                              _-----=> irqs-off
866 #                             / _----=> need-resched
867 #                            | / _---=> hardirq/softirq
868 #                            || / _--=> preempt-depth
869 #                            ||| /     delay
870 #           TASK-PID   CPU#  ||||    TIMESTAMP  FUNCTION
871 #              | |       |   ||||       |         |
872             Xorg-3129  [000] ..s.  2537.644844: intel_pstate_set_pstate <-intel_pstate_timer_func
873  gnome-terminal--4510  [002] ..s.  2537.649844: intel_pstate_set_pstate <-intel_pstate_timer_func
874      gnome-shell-3409  [001] ..s.  2537.650850: intel_pstate_set_pstate <-intel_pstate_timer_func
875           <idle>-0     [000] ..s.  2537.654843: intel_pstate_set_pstate <-intel_pstate_timer_func
876
877
878References
879==========
880
881.. [1] Kristen Accardi, *Balancing Power and Performance in the Linux Kernel*,
882       https://events.static.linuxfound.org/sites/events/files/slides/LinuxConEurope_2015.pdf
883
884.. [2] *Intel® 64 and IA-32 Architectures Software Developer’s Manual Volume 3: System Programming Guide*,
885       https://www.intel.com/content/www/us/en/architecture-and-technology/64-ia-32-architectures-software-developer-system-programming-manual-325384.html
886
887.. [3] *Advanced Configuration and Power Interface Specification*,
888       https://uefi.org/sites/default/files/resources/ACPI_6_3_final_Jan30.pdf
889