Documentation/admin-guide/pm/amd-pstate.rst

   1 .. SPDX-License-Identifier: GPL-2.0
   2 .. include:: <isonum.txt>
   3
   4 ===============================================
   5 ``amd-pstate`` CPU Performance Scaling Driver
   6 ===============================================
   7
   8 :Copyright: |copy| 2021 Advanced Micro Devices, Inc.
   9
  10 :Author: Huang Rui <ray.huang@amd.com>
  11
  12
  13 Introduction
  14 ===================
  15
  16 ``amd-pstate`` is the AMD CPU performance scaling driver that introduces a
  17 new CPU frequency control mechanism on modern AMD APU and CPU series in
  18 Linux kernel. The new mechanism is based on Collaborative Processor
  19 Performance Control (CPPC) which provides finer grain frequency management
  20 than legacy ACPI hardware P-States. Current AMD CPU/APU platforms are using
  21 the ACPI P-states driver to manage CPU frequency and clocks with switching
  22 only in 3 P-states. CPPC replaces the ACPI P-states controls and allows a
  23 flexible, low-latency interface for the Linux kernel to directly
  24 communicate the performance hints to hardware.
  25
  26 ``amd-pstate`` leverages the Linux kernel governors such as ``schedutil``,
  27 ``ondemand``, etc. to manage the performance hints which are provided by
  28 CPPC hardware functionality that internally follows the hardware
  29 specification (for details refer to AMD64 Architecture Programmer's Manual
  30 Volume 2: System Programming [1]_). Currently, ``amd-pstate`` supports basic
  31 frequency control function according to kernel governors on some of the
  32 Zen2 and Zen3 processors, and we will implement more AMD specific functions
  33 in future after we verify them on the hardware and SBIOS.
  34
  35
  36 AMD CPPC Overview
  37 =======================
  38
  39 Collaborative Processor Performance Control (CPPC) interface enumerates a
  40 continuous, abstract, and unit-less performance value in a scale that is
  41 not tied to a specific performance state / frequency. This is an ACPI
  42 standard [2]_ which software can specify application performance goals and
  43 hints as a relative target to the infrastructure limits. AMD processors
  44 provide the low latency register model (MSR) instead of an AML code
  45 interpreter for performance adjustments. ``amd-pstate`` will initialize a
  46 ``struct cpufreq_driver`` instance, ``amd_pstate_driver``, with the callbacks
  47 to manage each performance update behavior. ::
  48
  49  Highest Perf ------>+-----------------------+                         +-----------------------+
  50                      |                       |                         |                       |
  51                      |                       |                         |                       |
  52                      |                       |          Max Perf  ---->|                       |
  53                      |                       |                         |                       |
  54                      |                       |                         |                       |
  55  Nominal Perf ------>+-----------------------+                         +-----------------------+
  56                      |                       |                         |                       |
  57                      |                       |                         |                       |
  58                      |                       |                         |                       |
  59                      |                       |                         |                       |
  60                      |                       |                         |                       |
  61                      |                       |                         |                       |
  62                      |                       |      Desired Perf  ---->|                       |
  63                      |                       |                         |                       |
  64                      |                       |                         |                       |
  65                      |                       |                         |                       |
  66                      |                       |                         |                       |
  67                      |                       |                         |                       |
  68                      |                       |                         |                       |
  69                      |                       |                         |                       |
  70                      |                       |                         |                       |
  71                      |                       |                         |                       |
  72   Lowest non-        |                       |                         |                       |
  73   linear perf ------>+-----------------------+                         +-----------------------+
  74                      |                       |                         |                       |
  75                      |                       |       Lowest perf  ---->|                       |
  76                      |                       |                         |                       |
  77   Lowest perf ------>+-----------------------+                         +-----------------------+
  78                      |                       |                         |                       |
  79                      |                       |                         |                       |
  80                      |                       |                         |                       |
  81           0   ------>+-----------------------+                         +-----------------------+
  82
  83                                      AMD P-States Performance Scale
  84
  85
  86 .. _perf_cap:
  87
  88 AMD CPPC Performance Capability
  89 --------------------------------
  90
  91 Highest Performance (RO)
  92 .........................
  93
  94 This is the absolute maximum performance an individual processor may reach,
  95 assuming ideal conditions. This performance level may not be sustainable
  96 for long durations and may only be achievable if other platform components
  97 are in a specific state; for example, it may require other processors to be in
  98 an idle state. This would be equivalent to the highest frequencies
  99 supported by the processor.
 100
 101 Nominal (Guaranteed) Performance (RO)
 102 ......................................
 103
 104 This is the maximum sustained performance level of the processor, assuming
 105 ideal operating conditions. In the absence of an external constraint (power,
 106 thermal, etc.), this is the performance level the processor is expected to
 107 be able to maintain continuously. All cores/processors are expected to be
 108 able to sustain their nominal performance state simultaneously.
 109
 110 Lowest non-linear Performance (RO)
 111 ...................................
 112
 113 This is the lowest performance level at which nonlinear power savings are
 114 achieved, for example, due to the combined effects of voltage and frequency
 115 scaling. Above this threshold, lower performance levels should be generally
 116 more energy efficient than higher performance levels. This register
 117 effectively conveys the most efficient performance level to ``amd-pstate``.
 118
 119 Lowest Performance (RO)
 120 ........................
 121
 122 This is the absolute lowest performance level of the processor. Selecting a
 123 performance level lower than the lowest nonlinear performance level may
 124 cause an efficiency penalty but should reduce the instantaneous power
 125 consumption of the processor.
 126
 127 AMD CPPC Performance Control
 128 ------------------------------
 129
 130 ``amd-pstate`` passes performance goals through these registers. The
 131 register drives the behavior of the desired performance target.
 132
 133 Minimum requested performance (RW)
 134 ...................................
 135
 136 ``amd-pstate`` specifies the minimum allowed performance level.
 137
 138 Maximum requested performance (RW)
 139 ...................................
 140
 141 ``amd-pstate`` specifies a limit the maximum performance that is expected
 142 to be supplied by the hardware.
 143
 144 Desired performance target (RW)
 145 ...................................
 146
 147 ``amd-pstate`` specifies a desired target in the CPPC performance scale as
 148 a relative number. This can be expressed as percentage of nominal
 149 performance (infrastructure max). Below the nominal sustained performance
 150 level, desired performance expresses the average performance level of the
 151 processor subject to hardware. Above the nominal performance level,
 152 the processor must provide at least nominal performance requested and go higher
 153 if current operating conditions allow.
 154
 155 Energy Performance Preference (EPP) (RW)
 156 .........................................
 157
 158 This attribute provides a hint to the hardware if software wants to bias
 159 toward performance (0x0) or energy efficiency (0xff).
 160
 161
 162 Key Governors Support
 163 =======================
 164
 165 ``amd-pstate`` can be used with all the (generic) scaling governors listed
 166 by the ``scaling_available_governors`` policy attribute in ``sysfs``. Then,
 167 it is responsible for the configuration of policy objects corresponding to
 168 CPUs and provides the ``CPUFreq`` core (and the scaling governors attached
 169 to the policy objects) with accurate information on the maximum and minimum
 170 operating frequencies supported by the hardware. Users can check the
 171 ``scaling_cur_freq`` information comes from the ``CPUFreq`` core.
 172
 173 ``amd-pstate`` mainly supports ``schedutil`` and ``ondemand`` for dynamic
 174 frequency control. It is to fine tune the processor configuration on
 175 ``amd-pstate`` to the ``schedutil`` with CPU CFS scheduler. ``amd-pstate``
 176 registers the adjust_perf callback to implement performance update behavior
 177 similar to CPPC. It is initialized by ``sugov_start`` and then populates the
 178 CPU's update_util_data pointer to assign ``sugov_update_single_perf`` as the
 179 utilization update callback function in the CPU scheduler. The CPU scheduler
 180 will call ``cpufreq_update_util`` and assigns the target performance according
 181 to the ``struct sugov_cpu`` that the utilization update belongs to.
 182 Then, ``amd-pstate`` updates the desired performance according to the CPU
 183 scheduler assigned.
 184
 185 .. _processor_support:
 186
 187 Processor Support
 188 =======================
 189
 190 The ``amd-pstate`` initialization will fail if the ``_CPC`` entry in the ACPI
 191 SBIOS does not exist in the detected processor. It uses ``acpi_cpc_valid``
 192 to check the existence of ``_CPC``. All Zen based processors support the legacy
 193 ACPI hardware P-States function, so when ``amd-pstate`` fails initialization,
 194 the kernel will fall back to initialize the ``acpi-cpufreq`` driver.
 195
 196 There are two types of hardware implementations for ``amd-pstate``: one is
 197 `Full MSR Support <perf_cap_>`_ and another is `Shared Memory Support
 198 <perf_cap_>`_. It can use the :c:macro:`X86_FEATURE_CPPC` feature flag to
 199 indicate the different types. (For details, refer to the Processor Programming
 200 Reference (PPR) for AMD Family 19h Model 51h, Revision A1 Processors [3]_.)
 201 ``amd-pstate`` is to register different ``static_call`` instances for different
 202 hardware implementations.
 203
 204 Currently, some of the Zen2 and Zen3 processors support ``amd-pstate``. In the
 205 future, it will be supported on more and more AMD processors.
 206
 207 Full MSR Support
 208 -----------------
 209
 210 Some new Zen3 processors such as Cezanne provide the MSR registers directly
 211 while the :c:macro:`X86_FEATURE_CPPC` CPU feature flag is set.
 212 ``amd-pstate`` can handle the MSR register to implement the fast switch
 213 function in ``CPUFreq`` that can reduce the latency of frequency control in
 214 interrupt context. The functions with a ``pstate_xxx`` prefix represent the
 215 operations on MSR registers.
 216
 217 Shared Memory Support
 218 ----------------------
 219
 220 If the :c:macro:`X86_FEATURE_CPPC` CPU feature flag is not set, the
 221 processor supports the shared memory solution. In this case, ``amd-pstate``
 222 uses the ``cppc_acpi`` helper methods to implement the callback functions
 223 that are defined on ``static_call``. The functions with the ``cppc_xxx`` prefix
 224 represent the operations of ACPI CPPC helpers for the shared memory solution.
 225
 226
 227 AMD P-States and ACPI hardware P-States always can be supported in one
 228 processor. But AMD P-States has the higher priority and if it is enabled
 229 with :c:macro:`MSR_AMD_CPPC_ENABLE` or ``cppc_set_enable``, it will respond
 230 to the request from AMD P-States.
 231
 232
 233 User Space Interface in ``sysfs`` - Per-policy control
 234 ======================================================
 235
 236 ``amd-pstate`` exposes several global attributes (files) in ``sysfs`` to
 237 control its functionality at the system level. They are located in the
 238 ``/sys/devices/system/cpu/cpufreq/policyX/`` directory and affect all CPUs. ::
 239
 240  root@hr-test1:/home/ray# ls /sys/devices/system/cpu/cpufreq/policy0/*amd*
 241  /sys/devices/system/cpu/cpufreq/policy0/amd_pstate_highest_perf
 242  /sys/devices/system/cpu/cpufreq/policy0/amd_pstate_lowest_nonlinear_freq
 243  /sys/devices/system/cpu/cpufreq/policy0/amd_pstate_max_freq
 244
 245
 246 ``amd_pstate_highest_perf / amd_pstate_max_freq``
 247
 248 Maximum CPPC performance and CPU frequency that the driver is allowed to
 249 set, in percent of the maximum supported CPPC performance level (the highest
 250 performance supported in `AMD CPPC Performance Capability <perf_cap_>`_).
 251 In some ASICs, the highest CPPC performance is not the one in the ``_CPC``
 252 table, so we need to expose it to sysfs. If boost is not active, but
 253 still supported, this maximum frequency will be larger than the one in
 254 ``cpuinfo``.
 255 This attribute is read-only.
 256
 257 ``amd_pstate_lowest_nonlinear_freq``
 258
 259 The lowest non-linear CPPC CPU frequency that the driver is allowed to set,
 260 in percent of the maximum supported CPPC performance level. (Please see the
 261 lowest non-linear performance in `AMD CPPC Performance Capability
 262 <perf_cap_>`_.)
 263 This attribute is read-only.
 264
 265 ``energy_performance_available_preferences``
 266
 267 A list of all the supported EPP preferences that could be used for
 268 ``energy_performance_preference`` on this system.
 269 These profiles represent different hints that are provided
 270 to the low-level firmware about the user's desired energy vs efficiency
 271 tradeoff.  ``default`` represents the epp value is set by platform
 272 firmware. This attribute is read-only.
 273
 274 ``energy_performance_preference``
 275
 276 The current energy performance preference can be read from this attribute.
 277 and user can change current preference according to energy or performance needs
 278 Please get all support profiles list from
 279 ``energy_performance_available_preferences`` attribute, all the profiles are
 280 integer values defined between 0 to 255 when EPP feature is enabled by platform
 281 firmware, if EPP feature is disabled, driver will ignore the written value
 282 This attribute is read-write.
 283
 284 Other performance and frequency values can be read back from
 285 ``/sys/devices/system/cpu/cpuX/acpi_cppc/``, see :ref:`cppc_sysfs`.
 286
 287
 288 ``amd-pstate`` vs ``acpi-cpufreq``
 289 ======================================
 290
 291 On the majority of AMD platforms supported by ``acpi-cpufreq``, the ACPI tables
 292 provided by the platform firmware are used for CPU performance scaling, but
 293 only provide 3 P-states on AMD processors.
 294 However, on modern AMD APU and CPU series, hardware provides the Collaborative
 295 Processor Performance Control according to the ACPI protocol and customizes this
 296 for AMD platforms. That is, fine-grained and continuous frequency ranges
 297 instead of the legacy hardware P-states. ``amd-pstate`` is the kernel
 298 module which supports the new AMD P-States mechanism on most of the future AMD
 299 platforms. The AMD P-States mechanism is the more performance and energy
 300 efficiency frequency management method on AMD processors.
 301
 302
 303 AMD Pstate Driver Operation Modes
 304 =================================
 305
 306 ``amd_pstate`` CPPC has two operation modes: CPPC Autonomous(active) mode and
 307 CPPC non-autonomous(passive) mode.
 308 active mode and passive mode can be chosen by different kernel parameters.
 309 When in Autonomous mode, CPPC ignores requests done in the Desired Performance
 310 Target register and takes into account only the values set to the Minimum requested
 311 performance, Maximum requested performance, and Energy Performance Preference
 312 registers. When Autonomous is disabled, it only considers the Desired Performance Target.
 313
 314 Active Mode
 315 ------------
 316
 317 ``amd_pstate=active``
 318
 319 This is the low-level firmware control mode which is implemented by ``amd_pstate_epp``
 320 driver with ``amd_pstate=active`` passed to the kernel in the command line.
 321 In this mode, ``amd_pstate_epp`` driver provides a hint to the hardware if software
 322 wants to bias toward performance (0x0) or energy efficiency (0xff) to the CPPC firmware.
 323 then CPPC power algorithm will calculate the runtime workload and adjust the realtime
 324 cores frequency according to the power supply and thermal, core voltage and some other
 325 hardware conditions.
 326
 327 Passive Mode
 328 ------------
 329
 330 ``amd_pstate=passive``
 331
 332 It will be enabled if the ``amd_pstate=passive`` is passed to the kernel in the command line.
 333 In this mode, ``amd_pstate`` driver software specifies a desired QoS target in the CPPC
 334 performance scale as a relative number. This can be expressed as percentage of nominal
 335 performance (infrastructure max). Below the nominal sustained performance level,
 336 desired performance expresses the average performance level of the processor subject
 337 to the Performance Reduction Tolerance register. Above the nominal performance level,
 338 processor must provide at least nominal performance requested and go higher if current
 339 operating conditions allow.
 340
 341
 342 User Space Interface in ``sysfs`` - General
 343 ===========================================
 344
 345 Global Attributes
 346 -----------------
 347
 348 ``amd-pstate`` exposes several global attributes (files) in ``sysfs`` to
 349 control its functionality at the system level.  They are located in the
 350 ``/sys/devices/system/cpu/amd-pstate/`` directory and affect all CPUs.
 351
 352 ``status``
 353         Operation mode of the driver: "active", "passive" or "disable".
 354
 355         "active"
 356                 The driver is functional and in the ``active mode``
 357
 358         "passive"
 359                 The driver is functional and in the ``passive mode``
 360
 361         "disable"
 362                 The driver is unregistered and not functional now.
 363
 364         This attribute can be written to in order to change the driver's
 365         operation mode or to unregister it.  The string written to it must be
 366         one of the possible values of it and, if successful, writing one of
 367         these values to the sysfs file will cause the driver to switch over
 368         to the operation mode represented by that string - or to be
 369         unregistered in the "disable" case.
 370
 371 ``cpupower`` tool support for ``amd-pstate``
 372 ===============================================
 373
 374 ``amd-pstate`` is supported by the ``cpupower`` tool, which can be used to dump
 375 frequency information. Development is in progress to support more and more
 376 operations for the new ``amd-pstate`` module with this tool. ::
 377
 378  root@hr-test1:/home/ray# cpupower frequency-info
 379  analyzing CPU 0:
 380    driver: amd-pstate
 381    CPUs which run at the same hardware frequency: 0
 382    CPUs which need to have their frequency coordinated by software: 0
 383    maximum transition latency: 131 us
 384    hardware limits: 400 MHz - 4.68 GHz
 385    available cpufreq governors: ondemand conservative powersave userspace performance schedutil
 386    current policy: frequency should be within 400 MHz and 4.68 GHz.
 387                    The governor "schedutil" may decide which speed to use
 388                    within this range.
 389    current CPU frequency: Unable to call hardware
 390    current CPU frequency: 4.02 GHz (asserted by call to kernel)
 391    boost state support:
 392      Supported: yes
 393      Active: yes
 394      AMD PSTATE Highest Performance: 166. Maximum Frequency: 4.68 GHz.
 395      AMD PSTATE Nominal Performance: 117. Nominal Frequency: 3.30 GHz.
 396      AMD PSTATE Lowest Non-linear Performance: 39. Lowest Non-linear Frequency: 1.10 GHz.
 397      AMD PSTATE Lowest Performance: 15. Lowest Frequency: 400 MHz.
 398
 399
 400 Diagnostics and Tuning
 401 =======================
 402
 403 Trace Events
 404 --------------
 405
 406 There are two static trace events that can be used for ``amd-pstate``
 407 diagnostics. One of them is the ``cpu_frequency`` trace event generally used
 408 by ``CPUFreq``, and the other one is the ``amd_pstate_perf`` trace event
 409 specific to ``amd-pstate``.  The following sequence of shell commands can
 410 be used to enable them and see their output (if the kernel is
 411 configured to support event tracing). ::
 412
 413  root@hr-test1:/home/ray# cd /sys/kernel/tracing/
 414  root@hr-test1:/sys/kernel/tracing# echo 1 > events/amd_cpu/enable
 415  root@hr-test1:/sys/kernel/tracing# cat trace
 416  # tracer: nop
 417  #
 418  # entries-in-buffer/entries-written: 47827/42233061   #P:2
 419  #
 420  #                                _-----=> irqs-off
 421  #                               / _----=> need-resched
 422  #                              | / _---=> hardirq/softirq
 423  #                              || / _--=> preempt-depth
 424  #                              ||| /     delay
 425  #           TASK-PID     CPU#  ||||   TIMESTAMP  FUNCTION
 426  #              | |         |   ||||      |         |
 427           <idle>-0       [015] dN...  4995.979886: amd_pstate_perf: amd_min_perf=85 amd_des_perf=85 amd_max_perf=166 cpu_id=15 changed=false fast_switch=true
 428           <idle>-0       [007] d.h..  4995.979893: amd_pstate_perf: amd_min_perf=85 amd_des_perf=85 amd_max_perf=166 cpu_id=7 changed=false fast_switch=true
 429              cat-2161    [000] d....  4995.980841: amd_pstate_perf: amd_min_perf=85 amd_des_perf=85 amd_max_perf=166 cpu_id=0 changed=false fast_switch=true
 430             sshd-2125    [004] d.s..  4995.980968: amd_pstate_perf: amd_min_perf=85 amd_des_perf=85 amd_max_perf=166 cpu_id=4 changed=false fast_switch=true
 431           <idle>-0       [007] d.s..  4995.980968: amd_pstate_perf: amd_min_perf=85 amd_des_perf=85 amd_max_perf=166 cpu_id=7 changed=false fast_switch=true
 432           <idle>-0       [003] d.s..  4995.980971: amd_pstate_perf: amd_min_perf=85 amd_des_perf=85 amd_max_perf=166 cpu_id=3 changed=false fast_switch=true
 433           <idle>-0       [011] d.s..  4995.980996: amd_pstate_perf: amd_min_perf=85 amd_des_perf=85 amd_max_perf=166 cpu_id=11 changed=false fast_switch=true
 434
 435 The ``cpu_frequency`` trace event will be triggered either by the ``schedutil`` scaling
 436 governor (for the policies it is attached to), or by the ``CPUFreq`` core (for the
 437 policies with other scaling governors).
 438
 439
 440 Tracer Tool
 441 -------------
 442
 443 ``amd_pstate_tracer.py`` can record and parse ``amd-pstate`` trace log, then
 444 generate performance plots. This utility can be used to debug and tune the
 445 performance of ``amd-pstate`` driver. The tracer tool needs to import intel
 446 pstate tracer.
 447
 448 Tracer tool located in ``linux/tools/power/x86/amd_pstate_tracer``. It can be
 449 used in two ways. If trace file is available, then directly parse the file
 450 with command ::
 451
 452  ./amd_pstate_trace.py [-c cpus] -t <trace_file> -n <test_name>
 453
 454 Or generate trace file with root privilege, then parse and plot with command ::
 455
 456  sudo ./amd_pstate_trace.py [-c cpus] -n <test_name> -i <interval> [-m kbytes]
 457
 458 The test result can be found in ``results/test_name``. Following is the example
 459 about part of the output. ::
 460
 461  common_cpu  common_secs  common_usecs  min_perf  des_perf  max_perf  freq    mperf   apef    tsc       load   duration_ms  sample_num  elapsed_time  common_comm
 462  CPU_005     712          116384        39        49        166       0.7565  9645075 2214891 38431470  25.1   11.646       469         2.496         kworker/5:0-40
 463  CPU_006     712          116408        39        49        166       0.6769  8950227 1839034 37192089  24.06  11.272       470         2.496         kworker/6:0-1264
 464
 465 Unit Tests for amd-pstate
 466 -------------------------
 467
 468 ``amd-pstate-ut`` is a test module for testing the ``amd-pstate`` driver.
 469
 470  * It can help all users to verify their processor support (SBIOS/Firmware or Hardware).
 471
 472  * Kernel can have a basic function test to avoid the kernel regression during the update.
 473
 474  * We can introduce more functional or performance tests to align the result together, it will benefit power and performance scale optimization.
 475
 476 1. Test case descriptions
 477
 478     1). Basic tests
 479
 480         Test prerequisite and basic functions for the ``amd-pstate`` driver.
 481
 482         +---------+--------------------------------+------------------------------------------------------------------------------------+
 483         | Index   | Functions                      | Description                                                                        |
 484         +=========+================================+====================================================================================+
 485         | 1       | amd_pstate_ut_acpi_cpc_valid   || Check whether the _CPC object is present in SBIOS.                                |
 486         |         |                                ||                                                                                   |
 487         |         |                                || The detail refer to `Processor Support <processor_support_>`_.                    |
 488         +---------+--------------------------------+------------------------------------------------------------------------------------+
 489         | 2       | amd_pstate_ut_check_enabled    || Check whether AMD P-State is enabled.                                             |
 490         |         |                                ||                                                                                   |
 491         |         |                                || AMD P-States and ACPI hardware P-States always can be supported in one processor. |
 492         |         |                                | But AMD P-States has the higher priority and if it is enabled with                 |
 493         |         |                                | :c:macro:`MSR_AMD_CPPC_ENABLE` or ``cppc_set_enable``, it will respond to the      |
 494         |         |                                | request from AMD P-States.                                                         |
 495         +---------+--------------------------------+------------------------------------------------------------------------------------+
 496         | 3       | amd_pstate_ut_check_perf       || Check if the each performance values are reasonable.                              |
 497         |         |                                || highest_perf >= nominal_perf > lowest_nonlinear_perf > lowest_perf > 0.           |
 498         +---------+--------------------------------+------------------------------------------------------------------------------------+
 499         | 4       | amd_pstate_ut_check_freq       || Check if the each frequency values and max freq when set support boost mode       |
 500         |         |                                | are reasonable.                                                                    |
 501         |         |                                || max_freq >= nominal_freq > lowest_nonlinear_freq > min_freq > 0                   |
 502         |         |                                || If boost is not active but supported, this maximum frequency will be larger than  |
 503         |         |                                | the one in ``cpuinfo``.                                                            |
 504         +---------+--------------------------------+------------------------------------------------------------------------------------+
 505
 506     2). Tbench test
 507
 508         Test and monitor the cpu changes when running tbench benchmark under the specified governor.
 509         These changes include desire performance, frequency, load, performance, energy etc.
 510         The specified governor is ondemand or schedutil.
 511         Tbench can also be tested on the ``acpi-cpufreq`` kernel driver for comparison.
 512
 513     3). Gitsource test
 514
 515         Test and monitor the cpu changes when running gitsource benchmark under the specified governor.
 516         These changes include desire performance, frequency, load, time, energy etc.
 517         The specified governor is ondemand or schedutil.
 518         Gitsource can also be tested on the ``acpi-cpufreq`` kernel driver for comparison.
 519
 520 #. How to execute the tests
 521
 522    We use test module in the kselftest frameworks to implement it.
 523    We create ``amd-pstate-ut`` module and tie it into kselftest.(for
 524    details refer to Linux Kernel Selftests [4]_).
 525
 526     1). Build
 527
 528         + open the :c:macro:`CONFIG_X86_AMD_PSTATE` configuration option.
 529         + set the :c:macro:`CONFIG_X86_AMD_PSTATE_UT` configuration option to M.
 530         + make project
 531         + make selftest ::
 532
 533             $ cd linux
 534             $ make -C tools/testing/selftests
 535
 536         + make perf ::
 537
 538             $ cd tools/perf/
 539             $ make
 540
 541
 542     2). Installation & Steps ::
 543
 544         $ make -C tools/testing/selftests install INSTALL_PATH=~/kselftest
 545         $ cp tools/perf/perf /usr/bin/perf
 546         $ sudo ./kselftest/run_kselftest.sh -c amd-pstate
 547
 548     3). Specified test case ::
 549
 550         $ cd ~/kselftest/amd-pstate
 551         $ sudo ./run.sh -t basic
 552         $ sudo ./run.sh -t tbench
 553         $ sudo ./run.sh -t tbench -m acpi-cpufreq
 554         $ sudo ./run.sh -t gitsource
 555         $ sudo ./run.sh -t gitsource -m acpi-cpufreq
 556         $ ./run.sh --help
 557         ./run.sh: illegal option -- -
 558         Usage: ./run.sh [OPTION...]
 559                 [-h <help>]
 560                 [-o <output-file-for-dump>]
 561                 [-c <all: All testing,
 562                      basic: Basic testing,
 563                      tbench: Tbench testing,
 564                      gitsource: Gitsource testing.>]
 565                 [-t <tbench time limit>]
 566                 [-p <tbench process number>]
 567                 [-l <loop times for tbench>]
 568                 [-i <amd tracer interval>]
 569                 [-m <comparative test: acpi-cpufreq>]
 570
 571
 572     4). Results
 573
 574         + basic
 575
 576          When you finish test, you will get the following log info ::
 577
 578           $ dmesg | grep "amd_pstate_ut" | tee log.txt
 579           [12977.570663] amd_pstate_ut: 1    amd_pstate_ut_acpi_cpc_valid  success!
 580           [12977.570673] amd_pstate_ut: 2    amd_pstate_ut_check_enabled   success!
 581           [12977.571207] amd_pstate_ut: 3    amd_pstate_ut_check_perf      success!
 582           [12977.571212] amd_pstate_ut: 4    amd_pstate_ut_check_freq      success!
 583
 584         + tbench
 585
 586          When you finish test, you will get selftest.tbench.csv and png images.
 587          The selftest.tbench.csv file contains the raw data and the drop of the comparative test.
 588          The png images shows the performance, energy and performan per watt of each test.
 589          Open selftest.tbench.csv :
 590
 591          +-------------------------------------------------+--------------+----------+---------+----------+-------------+---------+----------------------+
 592          + Governor                                        | Round        | Des-perf | Freq    | Load     | Performance | Energy  | Performance Per Watt |
 593          +-------------------------------------------------+--------------+----------+---------+----------+-------------+---------+----------------------+
 594          + Unit                                            |              |          | GHz     |          | MB/s        | J       | MB/J                 |
 595          +=================================================+==============+==========+=========+==========+=============+=========+======================+
 596          + amd-pstate-ondemand                             | 1            |          |         |          | 2504.05     | 1563.67 | 158.5378             |
 597          +-------------------------------------------------+--------------+----------+---------+----------+-------------+---------+----------------------+
 598          + amd-pstate-ondemand                             | 2            |          |         |          | 2243.64     | 1430.32 | 155.2941             |
 599          +-------------------------------------------------+--------------+----------+---------+----------+-------------+---------+----------------------+
 600          + amd-pstate-ondemand                             | 3            |          |         |          | 2183.88     | 1401.32 | 154.2860             |
 601          +-------------------------------------------------+--------------+----------+---------+----------+-------------+---------+----------------------+
 602          + amd-pstate-ondemand                             | Average      |          |         |          | 2310.52     | 1465.1  | 156.1268             |
 603          +-------------------------------------------------+--------------+----------+---------+----------+-------------+---------+----------------------+
 604          + amd-pstate-schedutil                            | 1            | 165.329  | 1.62257 | 99.798   | 2136.54     | 1395.26 | 151.5971             |
 605          +-------------------------------------------------+--------------+----------+---------+----------+-------------+---------+----------------------+
 606          + amd-pstate-schedutil                            | 2            | 166      | 1.49761 | 99.9993  | 2100.56     | 1380.5  | 150.6377             |
 607          +-------------------------------------------------+--------------+----------+---------+----------+-------------+---------+----------------------+
 608          + amd-pstate-schedutil                            | 3            | 166      | 1.47806 | 99.9993  | 2084.12     | 1375.76 | 149.9737             |
 609          +-------------------------------------------------+--------------+----------+---------+----------+-------------+---------+----------------------+
 610          + amd-pstate-schedutil                            | Average      | 165.776  | 1.53275 | 99.9322  | 2107.07     | 1383.84 | 150.7399             |
 611          +-------------------------------------------------+--------------+----------+---------+----------+-------------+---------+----------------------+
 612          + acpi-cpufreq-ondemand                           | 1            |          |         |          | 2529.9      | 1564.4  | 160.0997             |
 613          +-------------------------------------------------+--------------+----------+---------+----------+-------------+---------+----------------------+
 614          + acpi-cpufreq-ondemand                           | 2            |          |         |          | 2249.76     | 1432.97 | 155.4297             |
 615          +-------------------------------------------------+--------------+----------+---------+----------+-------------+---------+----------------------+
 616          + acpi-cpufreq-ondemand                           | 3            |          |         |          | 2181.46     | 1406.88 | 153.5060             |
 617          +-------------------------------------------------+--------------+----------+---------+----------+-------------+---------+----------------------+
 618          + acpi-cpufreq-ondemand                           | Average      |          |         |          | 2320.37     | 1468.08 | 156.4741             |
 619          +-------------------------------------------------+--------------+----------+---------+----------+-------------+---------+----------------------+
 620          + acpi-cpufreq-schedutil                          | 1            |          |         |          | 2137.64     | 1385.24 | 152.7723             |
 621          +-------------------------------------------------+--------------+----------+---------+----------+-------------+---------+----------------------+
 622          + acpi-cpufreq-schedutil                          | 2            |          |         |          | 2107.05     | 1372.23 | 152.0138             |
 623          +-------------------------------------------------+--------------+----------+---------+----------+-------------+---------+----------------------+
 624          + acpi-cpufreq-schedutil                          | 3            |          |         |          | 2085.86     | 1365.35 | 151.2433             |
 625          +-------------------------------------------------+--------------+----------+---------+----------+-------------+---------+----------------------+
 626          + acpi-cpufreq-schedutil                          | Average      |          |         |          | 2110.18     | 1374.27 | 152.0136             |
 627          +-------------------------------------------------+--------------+----------+---------+----------+-------------+---------+----------------------+
 628          + acpi-cpufreq-ondemand VS acpi-cpufreq-schedutil | Comprison(%) |          |         |          | -9.0584     | -6.3899 | -2.8506              |
 629          +-------------------------------------------------+--------------+----------+---------+----------+-------------+---------+----------------------+
 630          + amd-pstate-ondemand VS amd-pstate-schedutil     | Comprison(%) |          |         |          | 8.8053      | -5.5463 | -3.4503              |
 631          +-------------------------------------------------+--------------+----------+---------+----------+-------------+---------+----------------------+
 632          + acpi-cpufreq-ondemand VS amd-pstate-ondemand    | Comprison(%) |          |         |          | -0.4245     | -0.2029 | -0.2219              |
 633          +-------------------------------------------------+--------------+----------+---------+----------+-------------+---------+----------------------+
 634          + acpi-cpufreq-schedutil VS amd-pstate-schedutil  | Comprison(%) |          |         |          | -0.1473     | 0.6963  | -0.8378              |
 635          +-------------------------------------------------+--------------+----------+---------+----------+-------------+---------+----------------------+
 636
 637         + gitsource
 638
 639          When you finish test, you will get selftest.gitsource.csv and png images.
 640          The selftest.gitsource.csv file contains the raw data and the drop of the comparative test.
 641          The png images shows the performance, energy and performan per watt of each test.
 642          Open selftest.gitsource.csv :
 643
 644          +-------------------------------------------------+--------------+----------+----------+----------+-------------+---------+----------------------+
 645          + Governor                                        | Round        | Des-perf | Freq     | Load     | Time        | Energy  | Performance Per Watt |
 646          +-------------------------------------------------+--------------+----------+----------+----------+-------------+---------+----------------------+
 647          + Unit                                            |              |          | GHz      |          | s           | J       | 1/J                  |
 648          +=================================================+==============+==========+==========+==========+=============+=========+======================+
 649          + amd-pstate-ondemand                             | 1            | 50.119   | 2.10509  | 23.3076  | 475.69      | 865.78  | 0.001155027          |
 650          +-------------------------------------------------+--------------+----------+----------+----------+-------------+---------+----------------------+
 651          + amd-pstate-ondemand                             | 2            | 94.8006  | 1.98771  | 56.6533  | 467.1       | 839.67  | 0.001190944          |
 652          +-------------------------------------------------+--------------+----------+----------+----------+-------------+---------+----------------------+
 653          + amd-pstate-ondemand                             | 3            | 76.6091  | 2.53251  | 43.7791  | 467.69      | 855.85  | 0.001168429          |
 654          +-------------------------------------------------+--------------+----------+----------+----------+-------------+---------+----------------------+
 655          + amd-pstate-ondemand                             | Average      | 73.8429  | 2.20844  | 41.2467  | 470.16      | 853.767 | 0.001171279          |
 656          +-------------------------------------------------+--------------+----------+----------+----------+-------------+---------+----------------------+
 657          + amd-pstate-schedutil                            | 1            | 165.919  | 1.62319  | 98.3868  | 464.17      | 866.8   | 0.001153668          |
 658          +-------------------------------------------------+--------------+----------+----------+----------+-------------+---------+----------------------+
 659          + amd-pstate-schedutil                            | 2            | 165.97   | 1.31309  | 99.5712  | 480.15      | 880.4   | 0.001135847          |
 660          +-------------------------------------------------+--------------+----------+----------+----------+-------------+---------+----------------------+
 661          + amd-pstate-schedutil                            | 3            | 165.973  | 1.28448  | 99.9252  | 481.79      | 867.02  | 0.001153375          |
 662          +-------------------------------------------------+--------------+----------+----------+----------+-------------+---------+----------------------+
 663          + amd-pstate-schedutil                            | Average      | 165.954  | 1.40692  | 99.2944  | 475.37      | 871.407 | 0.001147569          |
 664          +-------------------------------------------------+--------------+----------+----------+----------+-------------+---------+----------------------+
 665          + acpi-cpufreq-ondemand                           | 1            |          |          |          | 2379.62     | 742.96  | 0.001345967          |
 666          +-------------------------------------------------+--------------+----------+----------+----------+-------------+---------+----------------------+
 667          + acpi-cpufreq-ondemand                           | 2            |          |          |          | 441.74      | 817.49  | 0.001223256          |
 668          +-------------------------------------------------+--------------+----------+----------+----------+-------------+---------+----------------------+
 669          + acpi-cpufreq-ondemand                           | 3            |          |          |          | 455.48      | 820.01  | 0.001219497          |
 670          +-------------------------------------------------+--------------+----------+----------+----------+-------------+---------+----------------------+
 671          + acpi-cpufreq-ondemand                           | Average      |          |          |          | 425.613     | 793.487 | 0.001260260          |
 672          +-------------------------------------------------+--------------+----------+----------+----------+-------------+---------+----------------------+
 673          + acpi-cpufreq-schedutil                          | 1            |          |          |          | 459.69      | 838.54  | 0.001192548          |
 674          +-------------------------------------------------+--------------+----------+----------+----------+-------------+---------+----------------------+
 675          + acpi-cpufreq-schedutil                          | 2            |          |          |          | 466.55      | 830.89  | 0.001203528          |
 676          +-------------------------------------------------+--------------+----------+----------+----------+-------------+---------+----------------------+
 677          + acpi-cpufreq-schedutil                          | 3            |          |          |          | 470.38      | 837.32  | 0.001194286          |
 678          +-------------------------------------------------+--------------+----------+----------+----------+-------------+---------+----------------------+
 679          + acpi-cpufreq-schedutil                          | Average      |          |          |          | 465.54      | 835.583 | 0.001196769          |
 680          +-------------------------------------------------+--------------+----------+----------+----------+-------------+---------+----------------------+
 681          + acpi-cpufreq-ondemand VS acpi-cpufreq-schedutil | Comprison(%) |          |          |          | 9.3810      | 5.3051  | -5.0379              |
 682          +-------------------------------------------------+--------------+----------+----------+----------+-------------+---------+----------------------+
 683          + amd-pstate-ondemand VS amd-pstate-schedutil     | Comprison(%) | 124.7392 | -36.2934 | 140.7329 | 1.1081      | 2.0661  | -2.0242              |
 684          +-------------------------------------------------+--------------+----------+----------+----------+-------------+---------+----------------------+
 685          + acpi-cpufreq-ondemand VS amd-pstate-ondemand    | Comprison(%) |          |          |          | 10.4665     | 7.5968  | -7.0605              |
 686          +-------------------------------------------------+--------------+----------+----------+----------+-------------+---------+----------------------+
 687          + acpi-cpufreq-schedutil VS amd-pstate-schedutil  | Comprison(%) |          |          |          | 2.1115      | 4.2873  | -4.1110              |
 688          +-------------------------------------------------+--------------+----------+----------+----------+-------------+---------+----------------------+
 689
 690 Reference
 691 ===========
 692
 693 .. [1] AMD64 Architecture Programmer's Manual Volume 2: System Programming,
 694        https://www.amd.com/system/files/TechDocs/24593.pdf
 695
 696 .. [2] Advanced Configuration and Power Interface Specification,
 697        https://uefi.org/sites/default/files/resources/ACPI_Spec_6_4_Jan22.pdf
 698
 699 .. [3] Processor Programming Reference (PPR) for AMD Family 19h Model 51h, Revision A1 Processors
 700        https://www.amd.com/system/files/TechDocs/56569-A1-PUB.zip
 701
 702 .. [4] Linux Kernel Selftests,
 703        https://www.kernel.org/doc/html/latest/dev-tools/kselftest.html