]> git.karo-electronics.de Git - linux-beck.git/commitdiff
cpufreq: intel_pstate: Generic governors support
authorRafael J. Wysocki <rafael.j.wysocki@intel.com>
Thu, 17 Nov 2016 22:34:17 +0000 (23:34 +0100)
committerRafael J. Wysocki <rafael.j.wysocki@intel.com>
Mon, 21 Nov 2016 13:32:32 +0000 (14:32 +0100)
There may be reasons to use generic cpufreq governors (eg. schedutil)
on Intel platforms instead of the intel_pstate driver's internal
governor.  However, that currently can only be done by disabling
intel_pstate altogether and using the acpi-cpufreq driver instead
of it, which is subject to limitations.

First of all, acpi-cpufreq only works on systems where the _PSS
object is present in the ACPI tables for all logical CPUs.  Second,
on those systems acpi-cpufreq will only use frequencies listed by
_PSS which may be suboptimal.  In particular, by convention, the
whole turbo range is represented in _PSS as a single P-state and
the frequency assigned to it is greater by 1 MHz than the greatest
non-turbo frequency listed by _PSS.  That may confuse governors to
use turbo frequencies less frequently which may lead to suboptimal
performance.

For this reason, make it possible to use the intel_pstate driver
with generic cpufreq governors as a "normal" cpufreq driver.  That
mode is enforced by adding intel_pstate=passive to the kernel
command line and cannot be disabled at run time.  In that mode,
intel_pstate provides a cpufreq driver interface including
the ->target() and ->fast_switch() callbacks and is listed in
scaling_driver as "intel_cpufreq".

Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
Tested-by: Doug Smythies <dsmythies@telus.net>
Documentation/kernel-parameters.txt
drivers/cpufreq/intel_pstate.c

index fa2336ec1f56fd34de75682de3302c74e8dbc343..dfdd38ec149b76d4ae913260fa9f53b97381a1d6 100644 (file)
@@ -1760,6 +1760,12 @@ bytes respectively. Such letter suffixes can also be entirely omitted.
                       disable
                         Do not enable intel_pstate as the default
                         scaling driver for the supported processors
+                      passive
+                        Use intel_pstate as a scaling driver, but configure it
+                        to work with generic cpufreq governors (instead of
+                        enabling its internal governor).  This mode cannot be
+                        used along with the hardware-managed P-states (HWP)
+                        feature.
                       force
                         Enable intel_pstate on systems that prohibit it by default
                         in favor of acpi-cpufreq. Forcing the intel_pstate driver
index ec1664bf6ef09f56f50ffce348e0e7bed042e685..0d82bf320838c5188fd67d99b97bb462f9f721a1 100644 (file)
@@ -37,6 +37,8 @@
 #include <asm/cpufeature.h>
 #include <asm/intel-family.h>
 
+#define INTEL_CPUFREQ_TRANSITION_LATENCY       20000
+
 #define ATOM_RATIOS            0x66a
 #define ATOM_VIDS              0x66b
 #define ATOM_TURBO_RATIOS      0x66c
@@ -122,6 +124,8 @@ struct sample {
  * @scaling:           Scaling factor to  convert frequency to cpufreq
  *                     frequency units
  * @turbo_pstate:      Max Turbo P state possible for this platform
+ * @max_freq:          @max_pstate frequency in cpufreq units
+ * @turbo_freq:                @turbo_pstate frequency in cpufreq units
  *
  * Stores the per cpu model P state limits and current P state.
  */
@@ -132,6 +136,8 @@ struct pstate_data {
        int     max_pstate_physical;
        int     scaling;
        int     turbo_pstate;
+       unsigned int max_freq;
+       unsigned int turbo_freq;
 };
 
 /**
@@ -470,7 +476,7 @@ static void intel_pstate_init_acpi_perf_limits(struct cpufreq_policy *policy)
 {
 }
 
-static void intel_pstate_exit_perf_limits(struct cpufreq_policy *policy)
+static inline int intel_pstate_exit_perf_limits(struct cpufreq_policy *policy)
 {
 }
 #endif
@@ -1225,6 +1231,8 @@ static void intel_pstate_get_cpu_pstates(struct cpudata *cpu)
        cpu->pstate.max_pstate_physical = pstate_funcs.get_max_physical();
        cpu->pstate.turbo_pstate = pstate_funcs.get_turbo();
        cpu->pstate.scaling = pstate_funcs.get_scaling();
+       cpu->pstate.max_freq = cpu->pstate.max_pstate * cpu->pstate.scaling;
+       cpu->pstate.turbo_freq = cpu->pstate.turbo_pstate * cpu->pstate.scaling;
 
        if (pstate_funcs.get_vid)
                pstate_funcs.get_vid(cpu);
@@ -1363,15 +1371,19 @@ static inline int32_t get_target_pstate_use_performance(struct cpudata *cpu)
        return cpu->pstate.current_pstate - pid_calc(&cpu->pid, perf_scaled);
 }
 
-static inline void intel_pstate_update_pstate(struct cpudata *cpu, int pstate)
+static int intel_pstate_prepare_request(struct cpudata *cpu, int pstate)
 {
        int max_perf, min_perf;
 
-       update_turbo_state();
-
        intel_pstate_get_min_max(cpu, &min_perf, &max_perf);
        pstate = clamp_t(int, pstate, min_perf, max_perf);
        trace_cpu_frequency(pstate * cpu->pstate.scaling, cpu->cpu);
+       return pstate;
+}
+
+static void intel_pstate_update_pstate(struct cpudata *cpu, int pstate)
+{
+       pstate = intel_pstate_prepare_request(cpu, pstate);
        if (pstate == cpu->pstate.current_pstate)
                return;
 
@@ -1389,6 +1401,8 @@ static inline void intel_pstate_adjust_busy_pstate(struct cpudata *cpu)
        target_pstate = cpu->policy == CPUFREQ_POLICY_PERFORMANCE ?
                cpu->pstate.turbo_pstate : pstate_funcs.get_target_pstate(cpu);
 
+       update_turbo_state();
+
        intel_pstate_update_pstate(cpu, target_pstate);
 
        sample = &cpu->sample;
@@ -1670,22 +1684,30 @@ static int intel_pstate_verify_policy(struct cpufreq_policy *policy)
        return 0;
 }
 
+static void intel_cpufreq_stop_cpu(struct cpufreq_policy *policy)
+{
+       intel_pstate_set_min_pstate(all_cpu_data[policy->cpu]);
+}
+
 static void intel_pstate_stop_cpu(struct cpufreq_policy *policy)
 {
-       int cpu_num = policy->cpu;
-       struct cpudata *cpu = all_cpu_data[cpu_num];
+       pr_debug("CPU %d exiting\n", policy->cpu);
 
-       pr_debug("CPU %d exiting\n", cpu_num);
+       intel_pstate_clear_update_util_hook(policy->cpu);
+       if (!hwp_active)
+               intel_cpufreq_stop_cpu(policy);
+}
 
-       intel_pstate_clear_update_util_hook(cpu_num);
+static int intel_pstate_cpu_exit(struct cpufreq_policy *policy)
+{
+       intel_pstate_exit_perf_limits(policy);
 
-       if (hwp_active)
-               return;
+       policy->fast_switch_possible = false;
 
-       intel_pstate_set_min_pstate(cpu);
+       return 0;
 }
 
-static int intel_pstate_cpu_init(struct cpufreq_policy *policy)
+static int __intel_pstate_cpu_init(struct cpufreq_policy *policy)
 {
        struct cpudata *cpu;
        int rc;
@@ -1696,11 +1718,6 @@ static int intel_pstate_cpu_init(struct cpufreq_policy *policy)
 
        cpu = all_cpu_data[policy->cpu];
 
-       if (limits->min_perf_pct == 100 && limits->max_perf_pct == 100)
-               policy->policy = CPUFREQ_POLICY_PERFORMANCE;
-       else
-               policy->policy = CPUFREQ_POLICY_POWERSAVE;
-
        /*
         * We need sane value in the cpu->perf_limits, so inherit from global
         * perf_limits limits, which are seeded with values based on the
@@ -1720,20 +1737,30 @@ static int intel_pstate_cpu_init(struct cpufreq_policy *policy)
        policy->cpuinfo.max_freq *= cpu->pstate.scaling;
 
        intel_pstate_init_acpi_perf_limits(policy);
-       policy->cpuinfo.transition_latency = CPUFREQ_ETERNAL;
        cpumask_set_cpu(policy->cpu, policy->cpus);
 
+       policy->fast_switch_possible = true;
+
        return 0;
 }
 
-static int intel_pstate_cpu_exit(struct cpufreq_policy *policy)
+static int intel_pstate_cpu_init(struct cpufreq_policy *policy)
 {
-       intel_pstate_exit_perf_limits(policy);
+       int ret = __intel_pstate_cpu_init(policy);
+
+       if (ret)
+               return ret;
+
+       policy->cpuinfo.transition_latency = CPUFREQ_ETERNAL;
+       if (limits->min_perf_pct == 100 && limits->max_perf_pct == 100)
+               policy->policy = CPUFREQ_POLICY_PERFORMANCE;
+       else
+               policy->policy = CPUFREQ_POLICY_POWERSAVE;
 
        return 0;
 }
 
-static struct cpufreq_driver intel_pstate_driver = {
+static struct cpufreq_driver intel_pstate = {
        .flags          = CPUFREQ_CONST_LOOPS,
        .verify         = intel_pstate_verify_policy,
        .setpolicy      = intel_pstate_set_policy,
@@ -1745,6 +1772,118 @@ static struct cpufreq_driver intel_pstate_driver = {
        .name           = "intel_pstate",
 };
 
+static int intel_cpufreq_verify_policy(struct cpufreq_policy *policy)
+{
+       struct cpudata *cpu = all_cpu_data[policy->cpu];
+       struct perf_limits *perf_limits = limits;
+
+       update_turbo_state();
+       policy->cpuinfo.max_freq = limits->turbo_disabled ?
+                       cpu->pstate.max_freq : cpu->pstate.turbo_freq;
+
+       cpufreq_verify_within_cpu_limits(policy);
+
+       if (per_cpu_limits)
+               perf_limits = cpu->perf_limits;
+
+       intel_pstate_update_perf_limits(policy, perf_limits);
+
+       return 0;
+}
+
+static unsigned int intel_cpufreq_turbo_update(struct cpudata *cpu,
+                                              struct cpufreq_policy *policy,
+                                              unsigned int target_freq)
+{
+       unsigned int max_freq;
+
+       update_turbo_state();
+
+       max_freq = limits->no_turbo || limits->turbo_disabled ?
+                       cpu->pstate.max_freq : cpu->pstate.turbo_freq;
+       policy->cpuinfo.max_freq = max_freq;
+       if (policy->max > max_freq)
+               policy->max = max_freq;
+
+       if (target_freq > max_freq)
+               target_freq = max_freq;
+
+       return target_freq;
+}
+
+static int intel_cpufreq_target(struct cpufreq_policy *policy,
+                               unsigned int target_freq,
+                               unsigned int relation)
+{
+       struct cpudata *cpu = all_cpu_data[policy->cpu];
+       struct cpufreq_freqs freqs;
+       int target_pstate;
+
+       freqs.old = policy->cur;
+       freqs.new = intel_cpufreq_turbo_update(cpu, policy, target_freq);
+
+       cpufreq_freq_transition_begin(policy, &freqs);
+       switch (relation) {
+       case CPUFREQ_RELATION_L:
+               target_pstate = DIV_ROUND_UP(freqs.new, cpu->pstate.scaling);
+               break;
+       case CPUFREQ_RELATION_H:
+               target_pstate = freqs.new / cpu->pstate.scaling;
+               break;
+       default:
+               target_pstate = DIV_ROUND_CLOSEST(freqs.new, cpu->pstate.scaling);
+               break;
+       }
+       target_pstate = intel_pstate_prepare_request(cpu, target_pstate);
+       if (target_pstate != cpu->pstate.current_pstate) {
+               cpu->pstate.current_pstate = target_pstate;
+               wrmsrl_on_cpu(policy->cpu, MSR_IA32_PERF_CTL,
+                             pstate_funcs.get_val(cpu, target_pstate));
+       }
+       cpufreq_freq_transition_end(policy, &freqs, false);
+
+       return 0;
+}
+
+static unsigned int intel_cpufreq_fast_switch(struct cpufreq_policy *policy,
+                                             unsigned int target_freq)
+{
+       struct cpudata *cpu = all_cpu_data[policy->cpu];
+       int target_pstate;
+
+       target_freq = intel_cpufreq_turbo_update(cpu, policy, target_freq);
+       target_pstate = DIV_ROUND_UP(target_freq, cpu->pstate.scaling);
+       intel_pstate_update_pstate(cpu, target_pstate);
+       return target_freq;
+}
+
+static int intel_cpufreq_cpu_init(struct cpufreq_policy *policy)
+{
+       int ret = __intel_pstate_cpu_init(policy);
+
+       if (ret)
+               return ret;
+
+       policy->cpuinfo.transition_latency = INTEL_CPUFREQ_TRANSITION_LATENCY;
+       /* This reflects the intel_pstate_get_cpu_pstates() setting. */
+       policy->cur = policy->cpuinfo.min_freq;
+
+       return 0;
+}
+
+static struct cpufreq_driver intel_cpufreq = {
+       .flags          = CPUFREQ_CONST_LOOPS,
+       .verify         = intel_cpufreq_verify_policy,
+       .target         = intel_cpufreq_target,
+       .fast_switch    = intel_cpufreq_fast_switch,
+       .init           = intel_cpufreq_cpu_init,
+       .exit           = intel_pstate_cpu_exit,
+       .stop_cpu       = intel_cpufreq_stop_cpu,
+       .name           = "intel_cpufreq",
+};
+
+static struct cpufreq_driver *intel_pstate_driver = &intel_pstate;
+
 static int no_load __initdata;
 static int no_hwp __initdata;
 static int hwp_only __initdata;
@@ -1976,7 +2115,7 @@ hwp_cpu_matched:
 
        intel_pstate_request_control_from_smm();
 
-       rc = cpufreq_register_driver(&intel_pstate_driver);
+       rc = cpufreq_register_driver(intel_pstate_driver);
        if (rc)
                goto out;
 
@@ -1991,7 +2130,9 @@ out:
        get_online_cpus();
        for_each_online_cpu(cpu) {
                if (all_cpu_data[cpu]) {
-                       intel_pstate_clear_update_util_hook(cpu);
+                       if (intel_pstate_driver == &intel_pstate)
+                               intel_pstate_clear_update_util_hook(cpu);
+
                        kfree(all_cpu_data[cpu]);
                }
        }
@@ -2007,8 +2148,13 @@ static int __init intel_pstate_setup(char *str)
        if (!str)
                return -EINVAL;
 
-       if (!strcmp(str, "disable"))
+       if (!strcmp(str, "disable")) {
                no_load = 1;
+       } else if (!strcmp(str, "passive")) {
+               pr_info("Passive mode enabled\n");
+               intel_pstate_driver = &intel_cpufreq;
+               no_hwp = 1;
+       }
        if (!strcmp(str, "no_hwp")) {
                pr_info("HWP disabled\n");
                no_hwp = 1;