高通Thermal Overview之thermal-engine

网络上的参考有限，仅有的一个thermal-engine的一篇博客也不是最新的，参考的这篇博客再加上最新的msm8953平台的thermal-engine，对thermal-engine的配置、算法、加载进行介绍。

代码位置vendor/qcom/proprietary/thermal-engine/

默认配置文件thermal-engine.conf，里面目前用到的thermal控制算法monitor、ss、virtual。

1.base format:

Strings

说明

{debug}

可选，若有此选项表示使能调试信息

sampling

默认采样率，以ms为单位

[<Algorithminstance label>]

算法实例标签

algo_type

算法类型，必须是算法实例的第一个字段

disable

可选，用于默认情况下禁止该算法实例起作用

2.算法类型

inc/thermal_config.henumalgo_type {UNKNOWN_ALGO_TYPE= -1,MONITOR_ALGO_TYPE= 0,PID_ALGO_TYPE,SS_ALGO_TYPE,TB_ALGO_TYPE,EQUILIBRIUM_TYPE,VIRTUAL_SENSOR_TYPE,BWLM_ALGO_TYPE,ALGO_IDX_MAX,};

■monitor

如果温度达到thresholds，触发响应的动作(限制)，当温度低于thresholds_clr，结束动作(限制)

Strings

说明

algo_type

monitor

sensor

传感器名称

sampling

采样率ms

descending

可选，默认门限是升序，有此字段后门限顺序为降序

thresholds

门限值 mCor mA

thresholds_clr

清除门限值

actions

达到门限时的动作，多个动作时用'+'连接

action_info

动作额外信息，多个额外信息间用'+'连接

■ pid

ProportionalIntegralDerivative，比例-积分-微分控制，当温度超过最大允许值时，CPU需要降頻，并且使用pid算法将温度控制在set_point，该控制器输出的结果可以被转换成最大允许CPU频率。

pid简易的解释，操作人员控制电加热炉的炉温，炉温会设定一个set_point(需要维持的温度)，并且显示当前的温度，操作人员根据观察后，通过旋转电位器的转角，来控制电流的大小，进一步来控制温度。增加或者减小电流的比例，与P值相关；这个比例的误差，与I值相关；变化速率和未来趋势与D值相关。

Strings

说明

algo_type

pid

sampling

采样率ms

sensor

传感器名称

device

被PID算法调整的设备

set_point

PID算法调整的目标值mCor mA

set_point_clr

PID算法停止调整的值

p_const

PID算法中的P常量

i_const

PID算法中的I常量

d_const

PID算法中的D常量

i_samples

积分组件报错的积分样本数

dev_units_per_calc

Unitsof device adjustment per PID calculation outcome

每一个PID算法输出调整设备单元值

freq_scale

Frequencyscaling factor for DUAL PID

■ ss

DTM策略，即DynamicThermalManagement，动态控制算法有两个模式，DTM(用于频率控制)和PID(用于温度保持)。PID上文已经描述，DTM是基于温度setpoint，提升或这降低CPU的最大允许频率。每一个最大允许频率的调解是基于DCVSfrequency table中的可用频率steps。温度高于setpoint时，最大允许频率会按step逐渐下降，温度高于setpoint时，最大允许频率会按step逐渐上升。该策略可以用于控制GPU频率。

Strings

说明

algo_type

ss

sampling

采样率ms

sensor

传感器名称

device

被PID算法调整的设备

set_point

PID算法调整的目标值mCor mA

set_point_clr

PID算法停止调整的值

time_constant

Multiplierof sampling period for holding off adjustments

whencurrent and last error sample are equivalent

device_max_limit

可选，fieldto specify device performance mitgation

Ifit is defined, this instance mitigates device up todevice_max_limit value.

Itexpects value in KHz for cpu device and Hz for gpu device

device_perf_floor

可选，fieldto specify device performance mitgation floor.

Ifit is defined, this instance stops mitigating a device to a levelwith a corresponding perf_lvl at or above device_perf_floor.

■ tb

8998最新支持的算法TokenBucket算法。频率输出通过反馈进行控制，类似PID算法概念。当温度高于set_point的时候，频率降幅相对较大，然后通过频率限制，进行温控维持。

Strings

说明

algo_type

ss

sampling

采样率ms

sensor

传感器名称

device

被PID算法调整的设备

set_point

PID算法调整的目标值mCor mA

set_point_clr

PID算法停止调整的值

time_constant

Multiplierof sampling period for holding off adjustments

whencurrent and last error sample are equivalent

up_loop_gain

tokenbucket reward calculations的增益乘法器

down_loop_gain

tokenbucket penalty calculations的增益乘法器

auto_penalty

高于set_point时，增加的额外penalty。

Shouldbe in units of temperature/temp_scale_factor

auto_reward

低于set_point时，增加的额外reward。

Shouldbe in units of temperature/temp_scale_factor

temp_scale_factor

温度级别划分因子。Dividerfor temperature to get it to units of Degree Celsius from thevalue returned by reading the sensor.

freq_scale_factor

频率级别划分因子。Multiplierfor device frequency readings. Default device frequency unit isKHz, so set to 1 if device frequency is read in KHz, 1000 for Hz.This is necessary because GPU freq is often in Hz.

quadratic_reward

Boolean(zero false, nonzero true) to use a quadratic reward system in thecontrol loop

quadratic_penalty

Boolean(zero false, nonzero true) to use a quadratic penalty system inthe control loop

use_timeout

Boolean(zero false, nonzero true) to use a timeout when exiting theactive control zone for this client. This helps in smoothinghandoff between LMH and thermal sensors

timeout

如果use_timeout==true，当小于等于set_point时，该算法结束之前将等待'timeout'number of interrupts，feedingreward into the system of value 'set_point' - 'set_point_clr'.

unified_rail

If'unified_rail' is true for any configuration instance of thealgorithm, the unified rail logic of token bucket is applied

■virtual

可以利用两个以上的Tsensor，通过加权，组成一个虚拟的sensor。8998视频时间15:20

Strings

说明

algo_type

virtual

trip_sensor

传感器名称

set_point

当高于此温度时，虚拟传感器开始polling模式

set_point_clr

当高于此温度时，虚拟传感器停止polling模式

sensors

用于计算温度总和的传感器数组

weights

权重值数组

math

默认值0，使用weigths，可选的，1for minimum of sensors, 2 for max of sensors

sampling

采样率ms

3.配置文件字段取值解释

●'thresholds'/'thresholds_clr'/'actions'/'action_info'，最多接受8个空格分开的values

●'actions'field

actions

说明

none

-ACTION - Do nothing

-ACTION_INFO - ignored

report

-ACTION - Report threshold crossing to UI

-ACTION_INFO - ignored

备注：

通过抽象本地socket"THERMALD_UI"，以每行独立的string格式，发送thresholdcrossing information。清除或者触发alevel，或者其他action达到这个level将会导致消息的发送。参数按照以下顺序发送

*sensorname - Name of sensor reporting

*temperature - Current temperature

*current_threshold_level - current threshold level triggered orcleared

*is_trigger - "true" on level trigger, "false"on level clearing

cpu

-ACTION - CPU调整cpu频率

-ACTION_INFO - Max CPU frequency in KHz

cpuN

-ACTION - CPU frequency scaling where N is the specific CPU core[0..MAX CORES]

-ACTION_INFO - Max CPU frequency in KHz

clusterN

-ACTION - CLUSTER frequency scaling where N is the specific CLUSTERID

-ACTION_INFO - Max CLUSTER frequency in KHz

hotplug_N

-ACTION –拔掉第N个cpu

-ACTION_INFO - 0 for online a core, or 1 to offline it.

lcd

-ACTION - LCD brightness throttling，限制lcd的亮度

-ACTION_INFO - 0-255 value for max LCD brightness

modem

-ACTION - Request throttling of modem functionality，请求限制modem功能

-ACTION_INFO - 0-3 throttling level for modem mitigation

fusion

-ACTION - Request throttling of fusion modem functionality

-ACTION_INFO - 0-3 throttling level for fusion modem mitigation

battery

-ACTION - Battery charging current throttling

-ACTION_INFO - 0-3 throttling level for battery charging current

gpu

-ACTION - GPU frequency scaling

-ACTION_INFO - Max GPU frequency in Hz

wlan

-ACTION - WLAN throttling

-ACTION_INFO - 0-4 throttling level for WLAN mitigation

shutdown

-ACTION - Shutdown target

-ACTION_INFO - Shutdown delay in ms

vdd_

restriction

-ACTION - Request voltage restriction of all vdd rails on SoC

-ACTION_INFO - 1 for request for vdd_restriction,

0for release vdd_restriction

camera

-ACTION - camera fps throttling and camera shutdown mitigation

-ACTION_INFO - 0-3 throttling level for camera fps mitigation,

10level for camera app shutdown

camcorder

-ACTION - camcorder fps throttling and camcorder shutdownmitigation

-ACTION_INFO - 0-3 throttling level for camcorder fps mitigation,

10level for camcorder app shutdown

mdp

-ACTION - Request throttling of MDP CX voting

-ACTION_INFO - 0-3 throttling level for MDP mitigation

venus

-ACTION - Request throttling of VENUS CX voting

-ACTION_INFO - 0-3 throttling level for VENUS mitigation

modem_cx

-ACTION - Request throttling of modem CX voting

-ACTION_INFO - 0-3 throttling level for MODEM CX mitigation

●'device'field

device

说明

cpu

-DEVICE - Dynamic CPU frequency scaling

cpuN

-DEVICE - Dynamic CPU frequency scaling where N is the specific CPUcore [0..MAX CORES]

clusterN

-DEVICE - Dynamic CLUSTER frequency scaling where N is the specificCLUSTER ID

gpu

-DEVICE - Dynamic GPU frequency scaling

4.配置文件示例

例1:

sampling 1000[PMIC_THERM_MON]algo_type monitorsensor PMIC_THERMsampling 5000thresholds 40200 45000 50000thresholds_clr 38000 43000 48000actions cpu+report cpu cpuaction_info 1188000+0 368640 245760

描述：

1)默认采样率为1s；sensorPMIC_THERM设置的采样率是5s，覆盖了默认的；

2)当温度升到40.2度以上时，触发门限1，调节CPU最大允许频率为1188000KHz，在本例中，由于此频率是最大值，因此实际无动作；同时上报此消息，action_info值0被忽略；

3)上升到45度以上时，触发门限2，下降到43度以下时，清除门限2，调节CPU最大允许频率为368640 KHz；

4)当门限2在低于43度清除时，调节CPU的最大允许频率回到1188000KHz上；

5)当门限1在低于38度清除时，产生一个清楚门限1的report，allmitigation被reset。

例2:

debugsampling 2000[PMIC_THERM_MONITOR]algo_type monitorsensor PMIC_THERMsampling 5000thresholds 40200 45000 50000thresholds_clr 38000 43000 48000actions cpu+report cpu report+shutdownaction_info 768000+0 368640 0+6000

描述：

1)使能debuglogging输出；

2)默认采样率2s，sensorPMIC_THERM的采样率设置为5s；

3)上升到大于40.2度触发门限1，下降到低于38度清除门限1，当门限1触发时，最大允许CPU的频

率是768000KHz，并且产生一个report(action_infovalue 0 is ignored)；

4)上升到大于45度触发门限2，下降到低于43度清除门限2，当门限2触发时，最大允许CPU的频率

是368640KHz；

5)上升到大于50度触发门限3，下降到低于48度清除门限3，当门限3触发时，产生一个report并且

6s后关机。

例3:

debugsampling 2000[bcl_monitor]algo_type monitordescendingsensor bclsampling 1000thresholds 100 0thresholds_clr 500 100actions report reportaction_info 0 0

描述：

1)使能信息调试输出；

2)默认采样率是2s，batterycurrent limit 'bcl'的采样率设置为1s

3)当ibat上升到(imax- 100mA)时，触发门限1，当下降到(imax– 500mA)时，清除门限1，When

triggered,generate a report (action_info value 0 is ignored).

4)当ibat上升到imax时，触发门限2，当下降到(imax– 100mA)时，清除门限2，When

triggered,generate a report (action_info value 0 is ignored).

例4:

debug[TEST_PID]algo_type pidsensor tsens_tz_sensor0device cpusampling 1000set_point 85000set_point_clr 65000p_const 1.0i_const 1.0d_const 1.0i_samples 10dev_units_per_calc10000

描述：

1)使能debuglogging

2)PID算法实例标签为TEST_PID

3)使用tsens_tz_sensor0传感器

4)被调节的设备是CPU

5)采样时间1s

6)set_point值是PID调节算法的门限值和PID算法的设定值；

7)set_point_clr是停止PID调节算法的门限值；

8)p_const, i_const, d_const是PID等式中的p,i,d常量；

9)dev_units_per_calc 10000kHz(kHz unit because this is CPU device) ismultiplied

withPID calculation outcome to determine adjustment on the cpu device.

例5:

debug[virtual-sensor-0]algo_type virtualtrip_sensor tsens_tz_sensor8set_point 35000set_point_clr 30000sensors tsens_tz_sensor1 tsens_tz_sensor5weights 40 60sampling 250[Test-PID]algo_type pidsensor virtual-sensor-0device cpu1sampling 250set_point 55000set_point_clr 50000p_const 1.25i_const 0.8d_const 0.5i_samples 10dev_units_per_calc5000

描述：

1)PID实例Test-PID基于virtual-sensor-0的结果；

2)virtual-sensor-0需要用户手动定义；

3)trip_sensor用来指示虚拟传感器何时开始进入polling模式(轮询)；

4)trip_sensor必须为常规传感器，不能为另外一个虚拟传感器；

5)set_point是tripsensor的门限值，当高于此门限值时，tripsensor将从中断模式进入轮询模式，轮询

频率由虚拟传感器的sampling字段设置

6)set_point_clr是tripsensor的门限值，当低于此门限值时，tripsensor将停止轮询模式然后等待下一个

门限事件；

7)sensors定义了常规传感器数组，这些传感器要参与权重温度的计算

8)weights给定了传感器数组的权重值；

9)虚拟传感器的set_point必须小于pid算法的set_point，以便当达到set_point时pid能收到通知，另外，

如果虚拟传感器未进入轮询模式，pid将不能获取到它的温度；

10)如果tirpsensor不支持从中断模式到查询模式的改变，第8条可以被忽略。此时虚拟传感器的采样率将应该和pid的采样率一致。

例6:

debug[bcm_monitor]algo_type monitorsensor bcmsampling 1000thresholds 70000 90000 //注意，单位为m%thresholds_clr 69000 89000actions cpu cpuaction_info 768000 384000

描述：

1)使能debuglogging

2)当采样电流达到imax的70%时，门限1触发，调整cpu最大允许频率为768000KHz；

3)当采样电流达到imax的90%时，门限2触发，调整cpu最大允许频率为384000KHz；

4)bcm仅支持2级门限调节；

5)有效的门限值取值为：40000,50000,60000,70000,80000,90000。

例7:

debug[TB-CPU4]algo_type tbsampling 10sensor tsens_tz_sensor13device cluster1set_point 85000set_point_clr 50000time_constant 1up_loop_gain 2down_loop_gain 3auto_penalty 1.0auto_reward 0.0temp_scale_factor 1000freq_scale_factor 1quadratic_reward 1quadratic_penalty 1use_timeout 0timeout 1

1)使能debuglogging；

2)我们配置CPU4被tokenbucket控制。由于算法是tb，sensor是tsens_tz_sensor13，device是cluster1，

这个instance将monitortsens_tz_sensor13并且控制cluster1的频率；

3)每10ms发生一次温度读取和频率mitigation

4)当温度上升到85度时，将触摸3)事件，温度下降到50度时clr

5)time_constant is currently unused but is reserved for futureuse.当前使用，为未来使用保留

6)up_loop_gain set to 2 means that for every degree thattsens_tz_sensor13 is below its set_point, the reward
willincrease by a factor of 2. --2倍

7)down_loop_gain set to 3 means that for every degree thattsens_tz_sensor13 overshoots its set_point, the

penalty willincrease by a factor of 3. --3倍

8)auto_penalty of 1.0 means that the number of degrees of overshootused to calculate penalty will always be

incrementedby 1.0. This means that at 85C, the temperature overshoot will be 1C,86C->2C, etc.

9)auto_reward of 0.0 means that the number of degrees of undershootused to calculate the reward will not be

augmented.This means that at 80C, the temperature undershoot will be 5C,81C->4C, etc.

10)temp_scale_factorof 1000 means that the temperature readings from tsens_tz_sensor13are received in mC

andnot C. This used to vary between tsens sensors and other types ofsensors in the thermal-engine such as

gensensors.

11)freq_scale_factorof 1 means that the frequency readings from device1 are recieved inkHz. This varies

betweendevices, for instance the GPU frequency readings on 8994 are receivedin Hz, and therefore the

freq_scale_factorfor the 8994 GPU is 1000.

12)quadratic_rewardof 1 means to square the amount of reward. Whereas reward wouldnormally be

(undershoot *up_loop_gain), if this value is nonzero then reward is calculated as(undershoot *

up_loop_gain)^2.

13)quadratic_penaltyof 1 means to square the amount of penalty. Whereas penaltywouldnormally be

(overshoot *down_loop_gain), if this value is nonzero then penalty is calculatedas (overshoot *

down_loop_gain)^2.

14)use_timeoutof 0 means that when the readings on tsens_tz_sensor13 reach theirset_point_clr, the algorithm

shouldimmediately stop polling and stop mitigating. If this value werenonzero, then the algorithm would

wait fortimeout number of sequential interrupts (i.e. timeout * sampling ms)in which the reading from

tsens_tz_sensor13was below set_point_clr before ceasing to poll and mitigate.

5. 加载

关于sperakercall：80-N9649-1_D中的内容，ThermalCalibration Procedure for Speaker Coil Protection

speakercoil(线圈)的thermal校准的目标是准确的测量出来自温度传感器中的speakercoil的温度

可以通过三步实现：

1.转换resistanceto temperature –通过实际测量找出他们之间的线性关系b= T- mR

2.决定thermalmanagement algorithm中的offsetvalue – 查看logcat显示最小温度32度，实际万用表测量是27度，那么offset= 27 -32 = -5度

3.Program offset into device –这个是thermal-engine-o输出sperker-cal内容

[SPEAKER-CAL]sampling30000 30000 10 1800000sensorpm8953_tzsensorstsens_tz_sensor1 tsens_tz_sensor2 tsens_tz_sensor3 tsens_tz_sensor14tsens_tz_sensor15temp_range6000 10000 2000max_temp45000offset-4000

已经定义的宏，ANDROID

未定以的宏，ENABLE_OLD_PARSER

thermal.c

structthermal_setting_tthermal_settings;//包含一个setting_info链表，它包含关于thermal算法的基本数据intmain(int argc, char **argv){…if(!config_file) {if((config_file = get_target_default_thermal_config_file()))info("Usingtarget config file '%s'\n", config_file);else//这条路system/etc/thermal-engine.confinfo("Notarget config file, falling back to '%s'\n",CONFIG_FILE_DEFAULT);}if(output_conf) {//thermal-engine -o，dumpconfig file of active settingsdevices_manager_init();devices_init(minimum_mode);sensors_manager_init();...return 0;}…devices_manager_init();//啥也没做devices_init(minimum_mode);//通过读取sys节点信息，初始化gpufreq、cpufreq、clusterfreq、thermal_ioctl、qmi_communication、各种dev、vdd*、profile_switch、lcd、battery_mitigation等等target_algo_init();///*Vote to keep kernel mitigation enabled until init is done */kernel_dev =devices_manager_reg_clnt("kernel");//获取/sys/kernel/下的节点信息，并且组装成链表if(kernel_dev == NULL) {msg("%sFailed to create kernel device handle\n", __func__);}req.value =1;device_clnt_request(kernel_dev,&req);sensors_manager_init//啥也没做sensors_init(minimum_mode);//1.通过/sys/module/msm_thermal/sensor_info开始初始化sensors[]，2./sys/devices/virtual/thermal/thermal_zone*，继续初始化sensors[]，3.将bcl加入structsensors_mgr_sensor_info *sensor_listif(thermal_algo_framework_init() != THERMAL_ALGO_SUCCESS){//thermal_threads = NULL;info("%s:Error initializing thermal algorithm framework",__func__);return 1;}init_settings(&thermal_settings);//清0pid_init_data(&thermal_settings);//初始化定义好的pid数据，加入thermal_settings.setting_info里thermal_monitor_init_data(&thermal_settings);//monitorspeaker_cal_init_data(&thermal_settings);ss_init_data(&thermal_settings);tb_init_data(&thermal_settings);virtual_sensors_init_data(&thermal_settings);virtual_sensors_init(&thermal_settings,config_file);//初始化并且loadconfig_file数据到虚拟传感器链表load_config(&thermal_settings,config_file, LOAD_ALL_FLAG);//thermal_config_v2.cthermal_server_init();/*创建4个server套接字，并且使用select方法监听1.sockfd_server_send，通知已经注册client端thermalcurrent level2.sockfd_server_recv，获得client发送数据，并调用callback相应更新，在pid、montior、speaker、ss后面等初始化时加入了callback3.sockfd_server_log，将获得与客户端链接的新fd加入socket_fd[MAX_SOCKET_FD]4.sockfd_server_recv_passive，清缓存，清除select描述符集，清除thermal_send_fds[]，里面存储了每次新连接的clientfd*/pid_algo_init(&thermal_settings);//初始化pid，并且开启线程pid_algo_monitorthermal_monitor(&thermal_settings);ss_algo_init(&thermal_settings);speaker_cal_init(&thermal_settings);if(tb_algo_init(&thermal_settings) != THERMAL_ALGO_SUCCESS) {info("%s:Error initializing token bucket", __func__);return 1;}if(kernel_dev)device_clnt_cancel_request(kernel_dev);while (1)pause();...//释放资源}