1.. SPDX-License-Identifier: GPL-2.0 2 3=============================================== 4How to Implement a new CPUFreq Processor Driver 5=============================================== 6 7Authors: 8 9 10 - Dominik Brodowski <linux@brodo.de> 11 - Rafael J. Wysocki <rafael.j.wysocki@intel.com> 12 - Viresh Kumar <viresh.kumar@linaro.org> 13 14.. Contents 15 16 1. What To Do? 17 1.1 Initialization 18 1.2 Per-CPU Initialization 19 1.3 verify 20 1.4 target/target_index or setpolicy? 21 1.5 target/target_index 22 1.6 setpolicy 23 1.7 get_intermediate and target_intermediate 24 2. Frequency Table Helpers 25 26 27 281. What To Do? 29============== 30 31So, you just got a brand-new CPU / chipset with datasheets and want to 32add cpufreq support for this CPU / chipset? Great. Here are some hints 33on what is necessary: 34 35 361.1 Initialization 37------------------ 38 39First of all, in an __initcall level 7 (module_init()) or later 40function check whether this kernel runs on the right CPU and the right 41chipset. If so, register a struct cpufreq_driver with the CPUfreq core 42using cpufreq_register_driver() 43 44What shall this struct cpufreq_driver contain? 45 46 .name - The name of this driver. 47 48 .init - A pointer to the per-policy initialization function. 49 50 .verify - A pointer to a "verification" function. 51 52 .setpolicy _or_ .fast_switch _or_ .target _or_ .target_index - See 53 below on the differences. 54 55And optionally 56 57 .flags - Hints for the cpufreq core. 58 59 .driver_data - cpufreq driver specific data. 60 61 .resolve_freq - Returns the most appropriate frequency for a target 62 frequency. Doesn't change the frequency though. 63 64 .get_intermediate and target_intermediate - Used to switch to stable 65 frequency while changing CPU frequency. 66 67 .get - Returns current frequency of the CPU. 68 69 .bios_limit - Returns HW/BIOS max frequency limitations for the CPU. 70 71 .exit - A pointer to a per-policy cleanup function called during 72 CPU_POST_DEAD phase of cpu hotplug process. 73 74 .stop_cpu - A pointer to a per-policy stop function called during 75 CPU_DOWN_PREPARE phase of cpu hotplug process. 76 77 .suspend - A pointer to a per-policy suspend function which is called 78 with interrupts disabled and _after_ the governor is stopped for the 79 policy. 80 81 .resume - A pointer to a per-policy resume function which is called 82 with interrupts disabled and _before_ the governor is started again. 83 84 .ready - A pointer to a per-policy ready function which is called after 85 the policy is fully initialized. 86 87 .attr - A pointer to a NULL-terminated list of "struct freq_attr" which 88 allow to export values to sysfs. 89 90 .boost_enabled - If set, boost frequencies are enabled. 91 92 .set_boost - A pointer to a per-policy function to enable/disable boost 93 frequencies. 94 95 961.2 Per-CPU Initialization 97-------------------------- 98 99Whenever a new CPU is registered with the device model, or after the 100cpufreq driver registers itself, the per-policy initialization function 101cpufreq_driver.init is called if no cpufreq policy existed for the CPU. 102Note that the .init() and .exit() routines are called only once for the 103policy and not for each CPU managed by the policy. It takes a ``struct 104cpufreq_policy *policy`` as argument. What to do now? 105 106If necessary, activate the CPUfreq support on your CPU. 107 108Then, the driver must fill in the following values: 109 110+-----------------------------------+--------------------------------------+ 111|policy->cpuinfo.min_freq _and_ | | 112|policy->cpuinfo.max_freq | the minimum and maximum frequency | 113| | (in kHz) which is supported by | 114| | this CPU | 115+-----------------------------------+--------------------------------------+ 116|policy->cpuinfo.transition_latency | the time it takes on this CPU to | 117| | switch between two frequencies in | 118| | nanoseconds (if appropriate, else | 119| | specify CPUFREQ_ETERNAL) | 120+-----------------------------------+--------------------------------------+ 121|policy->cur | The current operating frequency of | 122| | this CPU (if appropriate) | 123+-----------------------------------+--------------------------------------+ 124|policy->min, | | 125|policy->max, | | 126|policy->policy and, if necessary, | | 127|policy->governor | must contain the "default policy" for| 128| | this CPU. A few moments later, | 129| | cpufreq_driver.verify and either | 130| | cpufreq_driver.setpolicy or | 131| | cpufreq_driver.target/target_index is| 132| | called with these values. | 133+-----------------------------------+--------------------------------------+ 134|policy->cpus | Update this with the masks of the | 135| | (online + offline) CPUs that do DVFS | 136| | along with this CPU (i.e. that share| 137| | clock/voltage rails with it). | 138+-----------------------------------+--------------------------------------+ 139 140For setting some of these values (cpuinfo.min[max]_freq, policy->min[max]), the 141frequency table helpers might be helpful. See the section 2 for more information 142on them. 143 144 1451.3 verify 146---------- 147 148When the user decides a new policy (consisting of 149"policy,governor,min,max") shall be set, this policy must be validated 150so that incompatible values can be corrected. For verifying these 151values cpufreq_verify_within_limits(``struct cpufreq_policy *policy``, 152``unsigned int min_freq``, ``unsigned int max_freq``) function might be helpful. 153See section 2 for details on frequency table helpers. 154 155You need to make sure that at least one valid frequency (or operating 156range) is within policy->min and policy->max. If necessary, increase 157policy->max first, and only if this is no solution, decrease policy->min. 158 159 1601.4 target or target_index or setpolicy or fast_switch? 161------------------------------------------------------- 162 163Most cpufreq drivers or even most cpu frequency scaling algorithms 164only allow the CPU frequency to be set to predefined fixed values. For 165these, you use the ->target(), ->target_index() or ->fast_switch() 166callbacks. 167 168Some cpufreq capable processors switch the frequency between certain 169limits on their own. These shall use the ->setpolicy() callback. 170 171 1721.5. target/target_index 173------------------------ 174 175The target_index call has two arguments: ``struct cpufreq_policy *policy``, 176and ``unsigned int`` index (into the exposed frequency table). 177 178The CPUfreq driver must set the new frequency when called here. The 179actual frequency must be determined by freq_table[index].frequency. 180 181It should always restore to earlier frequency (i.e. policy->restore_freq) in 182case of errors, even if we switched to intermediate frequency earlier. 183 184Deprecated 185---------- 186The target call has three arguments: ``struct cpufreq_policy *policy``, 187unsigned int target_frequency, unsigned int relation. 188 189The CPUfreq driver must set the new frequency when called here. The 190actual frequency must be determined using the following rules: 191 192- keep close to "target_freq" 193- policy->min <= new_freq <= policy->max (THIS MUST BE VALID!!!) 194- if relation==CPUFREQ_REL_L, try to select a new_freq higher than or equal 195 target_freq. ("L for lowest, but no lower than") 196- if relation==CPUFREQ_REL_H, try to select a new_freq lower than or equal 197 target_freq. ("H for highest, but no higher than") 198 199Here again the frequency table helper might assist you - see section 2 200for details. 201 2021.6. fast_switch 203---------------- 204 205This function is used for frequency switching from scheduler's context. 206Not all drivers are expected to implement it, as sleeping from within 207this callback isn't allowed. This callback must be highly optimized to 208do switching as fast as possible. 209 210This function has two arguments: ``struct cpufreq_policy *policy`` and 211``unsigned int target_frequency``. 212 213 2141.7 setpolicy 215------------- 216 217The setpolicy call only takes a ``struct cpufreq_policy *policy`` as 218argument. You need to set the lower limit of the in-processor or 219in-chipset dynamic frequency switching to policy->min, the upper limit 220to policy->max, and -if supported- select a performance-oriented 221setting when policy->policy is CPUFREQ_POLICY_PERFORMANCE, and a 222powersaving-oriented setting when CPUFREQ_POLICY_POWERSAVE. Also check 223the reference implementation in drivers/cpufreq/longrun.c 224 2251.8 get_intermediate and target_intermediate 226-------------------------------------------- 227 228Only for drivers with target_index() and CPUFREQ_ASYNC_NOTIFICATION unset. 229 230get_intermediate should return a stable intermediate frequency platform wants to 231switch to, and target_intermediate() should set CPU to that frequency, before 232jumping to the frequency corresponding to 'index'. Core will take care of 233sending notifications and driver doesn't have to handle them in 234target_intermediate() or target_index(). 235 236Drivers can return '0' from get_intermediate() in case they don't wish to switch 237to intermediate frequency for some target frequency. In that case core will 238directly call ->target_index(). 239 240NOTE: ->target_index() should restore to policy->restore_freq in case of 241failures as core would send notifications for that. 242 243 2442. Frequency Table Helpers 245========================== 246 247As most cpufreq processors only allow for being set to a few specific 248frequencies, a "frequency table" with some functions might assist in 249some work of the processor driver. Such a "frequency table" consists of 250an array of struct cpufreq_frequency_table entries, with driver specific 251values in "driver_data", the corresponding frequency in "frequency" and 252flags set. At the end of the table, you need to add a 253cpufreq_frequency_table entry with frequency set to CPUFREQ_TABLE_END. 254And if you want to skip one entry in the table, set the frequency to 255CPUFREQ_ENTRY_INVALID. The entries don't need to be in sorted in any 256particular order, but if they are cpufreq core will do DVFS a bit 257quickly for them as search for best match is faster. 258 259The cpufreq table is verified automatically by the core if the policy contains a 260valid pointer in its policy->freq_table field. 261 262cpufreq_frequency_table_verify() assures that at least one valid 263frequency is within policy->min and policy->max, and all other criteria 264are met. This is helpful for the ->verify call. 265 266cpufreq_frequency_table_target() is the corresponding frequency table 267helper for the ->target stage. Just pass the values to this function, 268and this function returns the of the frequency table entry which 269contains the frequency the CPU shall be set to. 270 271The following macros can be used as iterators over cpufreq_frequency_table: 272 273cpufreq_for_each_entry(pos, table) - iterates over all entries of frequency 274table. 275 276cpufreq_for_each_valid_entry(pos, table) - iterates over all entries, 277excluding CPUFREQ_ENTRY_INVALID frequencies. 278Use arguments "pos" - a ``cpufreq_frequency_table *`` as a loop cursor and 279"table" - the ``cpufreq_frequency_table *`` you want to iterate over. 280 281For example:: 282 283 struct cpufreq_frequency_table *pos, *driver_freq_table; 284 285 cpufreq_for_each_entry(pos, driver_freq_table) { 286 /* Do something with pos */ 287 pos->frequency = ... 288 } 289 290If you need to work with the position of pos within driver_freq_table, 291do not subtract the pointers, as it is quite costly. Instead, use the 292macros cpufreq_for_each_entry_idx() and cpufreq_for_each_valid_entry_idx(). 293