Back to home page

OSCL-LXR
 
 

     

0001 // SPDX-License-Identifier: GPL-2.0-or-later
0002 /*
0003  * acpi-cpufreq.c - ACPI Processor P-States Driver
0004  *
0005  *  Copyright (C) 2001, 2002 Andy Grover <andrew.grover@intel.com>
0006  *  Copyright (C) 2001, 2002 Paul Diefenbaugh <paul.s.diefenbaugh@intel.com>
0007  *  Copyright (C) 2002 - 2004 Dominik Brodowski <linux@brodo.de>
0008  *  Copyright (C) 2006       Denis Sadykov <denis.m.sadykov@intel.com>
0009  */
0010 
0011 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
0012 
0013 #include <linux/kernel.h>
0014 #include <linux/module.h>
0015 #include <linux/init.h>
0016 #include <linux/smp.h>
0017 #include <linux/sched.h>
0018 #include <linux/cpufreq.h>
0019 #include <linux/compiler.h>
0020 #include <linux/dmi.h>
0021 #include <linux/slab.h>
0022 
0023 #include <linux/acpi.h>
0024 #include <linux/io.h>
0025 #include <linux/delay.h>
0026 #include <linux/uaccess.h>
0027 
0028 #include <acpi/processor.h>
0029 #include <acpi/cppc_acpi.h>
0030 
0031 #include <asm/msr.h>
0032 #include <asm/processor.h>
0033 #include <asm/cpufeature.h>
0034 #include <asm/cpu_device_id.h>
0035 
0036 MODULE_AUTHOR("Paul Diefenbaugh, Dominik Brodowski");
0037 MODULE_DESCRIPTION("ACPI Processor P-States Driver");
0038 MODULE_LICENSE("GPL");
0039 
0040 enum {
0041     UNDEFINED_CAPABLE = 0,
0042     SYSTEM_INTEL_MSR_CAPABLE,
0043     SYSTEM_AMD_MSR_CAPABLE,
0044     SYSTEM_IO_CAPABLE,
0045 };
0046 
0047 #define INTEL_MSR_RANGE     (0xffff)
0048 #define AMD_MSR_RANGE       (0x7)
0049 #define HYGON_MSR_RANGE     (0x7)
0050 
0051 #define MSR_K7_HWCR_CPB_DIS (1ULL << 25)
0052 
0053 struct acpi_cpufreq_data {
0054     unsigned int resume;
0055     unsigned int cpu_feature;
0056     unsigned int acpi_perf_cpu;
0057     cpumask_var_t freqdomain_cpus;
0058     void (*cpu_freq_write)(struct acpi_pct_register *reg, u32 val);
0059     u32 (*cpu_freq_read)(struct acpi_pct_register *reg);
0060 };
0061 
0062 /* acpi_perf_data is a pointer to percpu data. */
0063 static struct acpi_processor_performance __percpu *acpi_perf_data;
0064 
0065 static inline struct acpi_processor_performance *to_perf_data(struct acpi_cpufreq_data *data)
0066 {
0067     return per_cpu_ptr(acpi_perf_data, data->acpi_perf_cpu);
0068 }
0069 
0070 static struct cpufreq_driver acpi_cpufreq_driver;
0071 
0072 static unsigned int acpi_pstate_strict;
0073 
0074 static bool boost_state(unsigned int cpu)
0075 {
0076     u32 lo, hi;
0077     u64 msr;
0078 
0079     switch (boot_cpu_data.x86_vendor) {
0080     case X86_VENDOR_INTEL:
0081     case X86_VENDOR_CENTAUR:
0082     case X86_VENDOR_ZHAOXIN:
0083         rdmsr_on_cpu(cpu, MSR_IA32_MISC_ENABLE, &lo, &hi);
0084         msr = lo | ((u64)hi << 32);
0085         return !(msr & MSR_IA32_MISC_ENABLE_TURBO_DISABLE);
0086     case X86_VENDOR_HYGON:
0087     case X86_VENDOR_AMD:
0088         rdmsr_on_cpu(cpu, MSR_K7_HWCR, &lo, &hi);
0089         msr = lo | ((u64)hi << 32);
0090         return !(msr & MSR_K7_HWCR_CPB_DIS);
0091     }
0092     return false;
0093 }
0094 
0095 static int boost_set_msr(bool enable)
0096 {
0097     u32 msr_addr;
0098     u64 msr_mask, val;
0099 
0100     switch (boot_cpu_data.x86_vendor) {
0101     case X86_VENDOR_INTEL:
0102     case X86_VENDOR_CENTAUR:
0103     case X86_VENDOR_ZHAOXIN:
0104         msr_addr = MSR_IA32_MISC_ENABLE;
0105         msr_mask = MSR_IA32_MISC_ENABLE_TURBO_DISABLE;
0106         break;
0107     case X86_VENDOR_HYGON:
0108     case X86_VENDOR_AMD:
0109         msr_addr = MSR_K7_HWCR;
0110         msr_mask = MSR_K7_HWCR_CPB_DIS;
0111         break;
0112     default:
0113         return -EINVAL;
0114     }
0115 
0116     rdmsrl(msr_addr, val);
0117 
0118     if (enable)
0119         val &= ~msr_mask;
0120     else
0121         val |= msr_mask;
0122 
0123     wrmsrl(msr_addr, val);
0124     return 0;
0125 }
0126 
0127 static void boost_set_msr_each(void *p_en)
0128 {
0129     bool enable = (bool) p_en;
0130 
0131     boost_set_msr(enable);
0132 }
0133 
0134 static int set_boost(struct cpufreq_policy *policy, int val)
0135 {
0136     on_each_cpu_mask(policy->cpus, boost_set_msr_each,
0137              (void *)(long)val, 1);
0138     pr_debug("CPU %*pbl: Core Boosting %sabled.\n",
0139          cpumask_pr_args(policy->cpus), val ? "en" : "dis");
0140 
0141     return 0;
0142 }
0143 
0144 static ssize_t show_freqdomain_cpus(struct cpufreq_policy *policy, char *buf)
0145 {
0146     struct acpi_cpufreq_data *data = policy->driver_data;
0147 
0148     if (unlikely(!data))
0149         return -ENODEV;
0150 
0151     return cpufreq_show_cpus(data->freqdomain_cpus, buf);
0152 }
0153 
0154 cpufreq_freq_attr_ro(freqdomain_cpus);
0155 
0156 #ifdef CONFIG_X86_ACPI_CPUFREQ_CPB
0157 static ssize_t store_cpb(struct cpufreq_policy *policy, const char *buf,
0158              size_t count)
0159 {
0160     int ret;
0161     unsigned int val = 0;
0162 
0163     if (!acpi_cpufreq_driver.set_boost)
0164         return -EINVAL;
0165 
0166     ret = kstrtouint(buf, 10, &val);
0167     if (ret || val > 1)
0168         return -EINVAL;
0169 
0170     cpus_read_lock();
0171     set_boost(policy, val);
0172     cpus_read_unlock();
0173 
0174     return count;
0175 }
0176 
0177 static ssize_t show_cpb(struct cpufreq_policy *policy, char *buf)
0178 {
0179     return sprintf(buf, "%u\n", acpi_cpufreq_driver.boost_enabled);
0180 }
0181 
0182 cpufreq_freq_attr_rw(cpb);
0183 #endif
0184 
0185 static int check_est_cpu(unsigned int cpuid)
0186 {
0187     struct cpuinfo_x86 *cpu = &cpu_data(cpuid);
0188 
0189     return cpu_has(cpu, X86_FEATURE_EST);
0190 }
0191 
0192 static int check_amd_hwpstate_cpu(unsigned int cpuid)
0193 {
0194     struct cpuinfo_x86 *cpu = &cpu_data(cpuid);
0195 
0196     return cpu_has(cpu, X86_FEATURE_HW_PSTATE);
0197 }
0198 
0199 static unsigned extract_io(struct cpufreq_policy *policy, u32 value)
0200 {
0201     struct acpi_cpufreq_data *data = policy->driver_data;
0202     struct acpi_processor_performance *perf;
0203     int i;
0204 
0205     perf = to_perf_data(data);
0206 
0207     for (i = 0; i < perf->state_count; i++) {
0208         if (value == perf->states[i].status)
0209             return policy->freq_table[i].frequency;
0210     }
0211     return 0;
0212 }
0213 
0214 static unsigned extract_msr(struct cpufreq_policy *policy, u32 msr)
0215 {
0216     struct acpi_cpufreq_data *data = policy->driver_data;
0217     struct cpufreq_frequency_table *pos;
0218     struct acpi_processor_performance *perf;
0219 
0220     if (boot_cpu_data.x86_vendor == X86_VENDOR_AMD)
0221         msr &= AMD_MSR_RANGE;
0222     else if (boot_cpu_data.x86_vendor == X86_VENDOR_HYGON)
0223         msr &= HYGON_MSR_RANGE;
0224     else
0225         msr &= INTEL_MSR_RANGE;
0226 
0227     perf = to_perf_data(data);
0228 
0229     cpufreq_for_each_entry(pos, policy->freq_table)
0230         if (msr == perf->states[pos->driver_data].status)
0231             return pos->frequency;
0232     return policy->freq_table[0].frequency;
0233 }
0234 
0235 static unsigned extract_freq(struct cpufreq_policy *policy, u32 val)
0236 {
0237     struct acpi_cpufreq_data *data = policy->driver_data;
0238 
0239     switch (data->cpu_feature) {
0240     case SYSTEM_INTEL_MSR_CAPABLE:
0241     case SYSTEM_AMD_MSR_CAPABLE:
0242         return extract_msr(policy, val);
0243     case SYSTEM_IO_CAPABLE:
0244         return extract_io(policy, val);
0245     default:
0246         return 0;
0247     }
0248 }
0249 
0250 static u32 cpu_freq_read_intel(struct acpi_pct_register *not_used)
0251 {
0252     u32 val, dummy __always_unused;
0253 
0254     rdmsr(MSR_IA32_PERF_CTL, val, dummy);
0255     return val;
0256 }
0257 
0258 static void cpu_freq_write_intel(struct acpi_pct_register *not_used, u32 val)
0259 {
0260     u32 lo, hi;
0261 
0262     rdmsr(MSR_IA32_PERF_CTL, lo, hi);
0263     lo = (lo & ~INTEL_MSR_RANGE) | (val & INTEL_MSR_RANGE);
0264     wrmsr(MSR_IA32_PERF_CTL, lo, hi);
0265 }
0266 
0267 static u32 cpu_freq_read_amd(struct acpi_pct_register *not_used)
0268 {
0269     u32 val, dummy __always_unused;
0270 
0271     rdmsr(MSR_AMD_PERF_CTL, val, dummy);
0272     return val;
0273 }
0274 
0275 static void cpu_freq_write_amd(struct acpi_pct_register *not_used, u32 val)
0276 {
0277     wrmsr(MSR_AMD_PERF_CTL, val, 0);
0278 }
0279 
0280 static u32 cpu_freq_read_io(struct acpi_pct_register *reg)
0281 {
0282     u32 val;
0283 
0284     acpi_os_read_port(reg->address, &val, reg->bit_width);
0285     return val;
0286 }
0287 
0288 static void cpu_freq_write_io(struct acpi_pct_register *reg, u32 val)
0289 {
0290     acpi_os_write_port(reg->address, val, reg->bit_width);
0291 }
0292 
0293 struct drv_cmd {
0294     struct acpi_pct_register *reg;
0295     u32 val;
0296     union {
0297         void (*write)(struct acpi_pct_register *reg, u32 val);
0298         u32 (*read)(struct acpi_pct_register *reg);
0299     } func;
0300 };
0301 
0302 /* Called via smp_call_function_single(), on the target CPU */
0303 static void do_drv_read(void *_cmd)
0304 {
0305     struct drv_cmd *cmd = _cmd;
0306 
0307     cmd->val = cmd->func.read(cmd->reg);
0308 }
0309 
0310 static u32 drv_read(struct acpi_cpufreq_data *data, const struct cpumask *mask)
0311 {
0312     struct acpi_processor_performance *perf = to_perf_data(data);
0313     struct drv_cmd cmd = {
0314         .reg = &perf->control_register,
0315         .func.read = data->cpu_freq_read,
0316     };
0317     int err;
0318 
0319     err = smp_call_function_any(mask, do_drv_read, &cmd, 1);
0320     WARN_ON_ONCE(err);  /* smp_call_function_any() was buggy? */
0321     return cmd.val;
0322 }
0323 
0324 /* Called via smp_call_function_many(), on the target CPUs */
0325 static void do_drv_write(void *_cmd)
0326 {
0327     struct drv_cmd *cmd = _cmd;
0328 
0329     cmd->func.write(cmd->reg, cmd->val);
0330 }
0331 
0332 static void drv_write(struct acpi_cpufreq_data *data,
0333               const struct cpumask *mask, u32 val)
0334 {
0335     struct acpi_processor_performance *perf = to_perf_data(data);
0336     struct drv_cmd cmd = {
0337         .reg = &perf->control_register,
0338         .val = val,
0339         .func.write = data->cpu_freq_write,
0340     };
0341     int this_cpu;
0342 
0343     this_cpu = get_cpu();
0344     if (cpumask_test_cpu(this_cpu, mask))
0345         do_drv_write(&cmd);
0346 
0347     smp_call_function_many(mask, do_drv_write, &cmd, 1);
0348     put_cpu();
0349 }
0350 
0351 static u32 get_cur_val(const struct cpumask *mask, struct acpi_cpufreq_data *data)
0352 {
0353     u32 val;
0354 
0355     if (unlikely(cpumask_empty(mask)))
0356         return 0;
0357 
0358     val = drv_read(data, mask);
0359 
0360     pr_debug("%s = %u\n", __func__, val);
0361 
0362     return val;
0363 }
0364 
0365 static unsigned int get_cur_freq_on_cpu(unsigned int cpu)
0366 {
0367     struct acpi_cpufreq_data *data;
0368     struct cpufreq_policy *policy;
0369     unsigned int freq;
0370     unsigned int cached_freq;
0371 
0372     pr_debug("%s (%d)\n", __func__, cpu);
0373 
0374     policy = cpufreq_cpu_get_raw(cpu);
0375     if (unlikely(!policy))
0376         return 0;
0377 
0378     data = policy->driver_data;
0379     if (unlikely(!data || !policy->freq_table))
0380         return 0;
0381 
0382     cached_freq = policy->freq_table[to_perf_data(data)->state].frequency;
0383     freq = extract_freq(policy, get_cur_val(cpumask_of(cpu), data));
0384     if (freq != cached_freq) {
0385         /*
0386          * The dreaded BIOS frequency change behind our back.
0387          * Force set the frequency on next target call.
0388          */
0389         data->resume = 1;
0390     }
0391 
0392     pr_debug("cur freq = %u\n", freq);
0393 
0394     return freq;
0395 }
0396 
0397 static unsigned int check_freqs(struct cpufreq_policy *policy,
0398                 const struct cpumask *mask, unsigned int freq)
0399 {
0400     struct acpi_cpufreq_data *data = policy->driver_data;
0401     unsigned int cur_freq;
0402     unsigned int i;
0403 
0404     for (i = 0; i < 100; i++) {
0405         cur_freq = extract_freq(policy, get_cur_val(mask, data));
0406         if (cur_freq == freq)
0407             return 1;
0408         udelay(10);
0409     }
0410     return 0;
0411 }
0412 
0413 static int acpi_cpufreq_target(struct cpufreq_policy *policy,
0414                    unsigned int index)
0415 {
0416     struct acpi_cpufreq_data *data = policy->driver_data;
0417     struct acpi_processor_performance *perf;
0418     const struct cpumask *mask;
0419     unsigned int next_perf_state = 0; /* Index into perf table */
0420     int result = 0;
0421 
0422     if (unlikely(!data)) {
0423         return -ENODEV;
0424     }
0425 
0426     perf = to_perf_data(data);
0427     next_perf_state = policy->freq_table[index].driver_data;
0428     if (perf->state == next_perf_state) {
0429         if (unlikely(data->resume)) {
0430             pr_debug("Called after resume, resetting to P%d\n",
0431                 next_perf_state);
0432             data->resume = 0;
0433         } else {
0434             pr_debug("Already at target state (P%d)\n",
0435                 next_perf_state);
0436             return 0;
0437         }
0438     }
0439 
0440     /*
0441      * The core won't allow CPUs to go away until the governor has been
0442      * stopped, so we can rely on the stability of policy->cpus.
0443      */
0444     mask = policy->shared_type == CPUFREQ_SHARED_TYPE_ANY ?
0445         cpumask_of(policy->cpu) : policy->cpus;
0446 
0447     drv_write(data, mask, perf->states[next_perf_state].control);
0448 
0449     if (acpi_pstate_strict) {
0450         if (!check_freqs(policy, mask,
0451                  policy->freq_table[index].frequency)) {
0452             pr_debug("%s (%d)\n", __func__, policy->cpu);
0453             result = -EAGAIN;
0454         }
0455     }
0456 
0457     if (!result)
0458         perf->state = next_perf_state;
0459 
0460     return result;
0461 }
0462 
0463 static unsigned int acpi_cpufreq_fast_switch(struct cpufreq_policy *policy,
0464                          unsigned int target_freq)
0465 {
0466     struct acpi_cpufreq_data *data = policy->driver_data;
0467     struct acpi_processor_performance *perf;
0468     struct cpufreq_frequency_table *entry;
0469     unsigned int next_perf_state, next_freq, index;
0470 
0471     /*
0472      * Find the closest frequency above target_freq.
0473      */
0474     if (policy->cached_target_freq == target_freq)
0475         index = policy->cached_resolved_idx;
0476     else
0477         index = cpufreq_table_find_index_dl(policy, target_freq,
0478                             false);
0479 
0480     entry = &policy->freq_table[index];
0481     next_freq = entry->frequency;
0482     next_perf_state = entry->driver_data;
0483 
0484     perf = to_perf_data(data);
0485     if (perf->state == next_perf_state) {
0486         if (unlikely(data->resume))
0487             data->resume = 0;
0488         else
0489             return next_freq;
0490     }
0491 
0492     data->cpu_freq_write(&perf->control_register,
0493                  perf->states[next_perf_state].control);
0494     perf->state = next_perf_state;
0495     return next_freq;
0496 }
0497 
0498 static unsigned long
0499 acpi_cpufreq_guess_freq(struct acpi_cpufreq_data *data, unsigned int cpu)
0500 {
0501     struct acpi_processor_performance *perf;
0502 
0503     perf = to_perf_data(data);
0504     if (cpu_khz) {
0505         /* search the closest match to cpu_khz */
0506         unsigned int i;
0507         unsigned long freq;
0508         unsigned long freqn = perf->states[0].core_frequency * 1000;
0509 
0510         for (i = 0; i < (perf->state_count-1); i++) {
0511             freq = freqn;
0512             freqn = perf->states[i+1].core_frequency * 1000;
0513             if ((2 * cpu_khz) > (freqn + freq)) {
0514                 perf->state = i;
0515                 return freq;
0516             }
0517         }
0518         perf->state = perf->state_count-1;
0519         return freqn;
0520     } else {
0521         /* assume CPU is at P0... */
0522         perf->state = 0;
0523         return perf->states[0].core_frequency * 1000;
0524     }
0525 }
0526 
0527 static void free_acpi_perf_data(void)
0528 {
0529     unsigned int i;
0530 
0531     /* Freeing a NULL pointer is OK, and alloc_percpu zeroes. */
0532     for_each_possible_cpu(i)
0533         free_cpumask_var(per_cpu_ptr(acpi_perf_data, i)
0534                  ->shared_cpu_map);
0535     free_percpu(acpi_perf_data);
0536 }
0537 
0538 static int cpufreq_boost_online(unsigned int cpu)
0539 {
0540     /*
0541      * On the CPU_UP path we simply keep the boost-disable flag
0542      * in sync with the current global state.
0543      */
0544     return boost_set_msr(acpi_cpufreq_driver.boost_enabled);
0545 }
0546 
0547 static int cpufreq_boost_down_prep(unsigned int cpu)
0548 {
0549     /*
0550      * Clear the boost-disable bit on the CPU_DOWN path so that
0551      * this cpu cannot block the remaining ones from boosting.
0552      */
0553     return boost_set_msr(1);
0554 }
0555 
0556 /*
0557  * acpi_cpufreq_early_init - initialize ACPI P-States library
0558  *
0559  * Initialize the ACPI P-States library (drivers/acpi/processor_perflib.c)
0560  * in order to determine correct frequency and voltage pairings. We can
0561  * do _PDC and _PSD and find out the processor dependency for the
0562  * actual init that will happen later...
0563  */
0564 static int __init acpi_cpufreq_early_init(void)
0565 {
0566     unsigned int i;
0567     pr_debug("%s\n", __func__);
0568 
0569     acpi_perf_data = alloc_percpu(struct acpi_processor_performance);
0570     if (!acpi_perf_data) {
0571         pr_debug("Memory allocation error for acpi_perf_data.\n");
0572         return -ENOMEM;
0573     }
0574     for_each_possible_cpu(i) {
0575         if (!zalloc_cpumask_var_node(
0576             &per_cpu_ptr(acpi_perf_data, i)->shared_cpu_map,
0577             GFP_KERNEL, cpu_to_node(i))) {
0578 
0579             /* Freeing a NULL pointer is OK: alloc_percpu zeroes. */
0580             free_acpi_perf_data();
0581             return -ENOMEM;
0582         }
0583     }
0584 
0585     /* Do initialization in ACPI core */
0586     acpi_processor_preregister_performance(acpi_perf_data);
0587     return 0;
0588 }
0589 
0590 #ifdef CONFIG_SMP
0591 /*
0592  * Some BIOSes do SW_ANY coordination internally, either set it up in hw
0593  * or do it in BIOS firmware and won't inform about it to OS. If not
0594  * detected, this has a side effect of making CPU run at a different speed
0595  * than OS intended it to run at. Detect it and handle it cleanly.
0596  */
0597 static int bios_with_sw_any_bug;
0598 
0599 static int sw_any_bug_found(const struct dmi_system_id *d)
0600 {
0601     bios_with_sw_any_bug = 1;
0602     return 0;
0603 }
0604 
0605 static const struct dmi_system_id sw_any_bug_dmi_table[] = {
0606     {
0607         .callback = sw_any_bug_found,
0608         .ident = "Supermicro Server X6DLP",
0609         .matches = {
0610             DMI_MATCH(DMI_SYS_VENDOR, "Supermicro"),
0611             DMI_MATCH(DMI_BIOS_VERSION, "080010"),
0612             DMI_MATCH(DMI_PRODUCT_NAME, "X6DLP"),
0613         },
0614     },
0615     { }
0616 };
0617 
0618 static int acpi_cpufreq_blacklist(struct cpuinfo_x86 *c)
0619 {
0620     /* Intel Xeon Processor 7100 Series Specification Update
0621      * https://www.intel.com/Assets/PDF/specupdate/314554.pdf
0622      * AL30: A Machine Check Exception (MCE) Occurring during an
0623      * Enhanced Intel SpeedStep Technology Ratio Change May Cause
0624      * Both Processor Cores to Lock Up. */
0625     if (c->x86_vendor == X86_VENDOR_INTEL) {
0626         if ((c->x86 == 15) &&
0627             (c->x86_model == 6) &&
0628             (c->x86_stepping == 8)) {
0629             pr_info("Intel(R) Xeon(R) 7100 Errata AL30, processors may lock up on frequency changes: disabling acpi-cpufreq\n");
0630             return -ENODEV;
0631             }
0632         }
0633     return 0;
0634 }
0635 #endif
0636 
0637 #ifdef CONFIG_ACPI_CPPC_LIB
0638 static u64 get_max_boost_ratio(unsigned int cpu)
0639 {
0640     struct cppc_perf_caps perf_caps;
0641     u64 highest_perf, nominal_perf;
0642     int ret;
0643 
0644     if (acpi_pstate_strict)
0645         return 0;
0646 
0647     ret = cppc_get_perf_caps(cpu, &perf_caps);
0648     if (ret) {
0649         pr_debug("CPU%d: Unable to get performance capabilities (%d)\n",
0650              cpu, ret);
0651         return 0;
0652     }
0653 
0654     if (boot_cpu_data.x86_vendor == X86_VENDOR_AMD)
0655         highest_perf = amd_get_highest_perf();
0656     else
0657         highest_perf = perf_caps.highest_perf;
0658 
0659     nominal_perf = perf_caps.nominal_perf;
0660 
0661     if (!highest_perf || !nominal_perf) {
0662         pr_debug("CPU%d: highest or nominal performance missing\n", cpu);
0663         return 0;
0664     }
0665 
0666     if (highest_perf < nominal_perf) {
0667         pr_debug("CPU%d: nominal performance above highest\n", cpu);
0668         return 0;
0669     }
0670 
0671     return div_u64(highest_perf << SCHED_CAPACITY_SHIFT, nominal_perf);
0672 }
0673 #else
0674 static inline u64 get_max_boost_ratio(unsigned int cpu) { return 0; }
0675 #endif
0676 
0677 static int acpi_cpufreq_cpu_init(struct cpufreq_policy *policy)
0678 {
0679     struct cpufreq_frequency_table *freq_table;
0680     struct acpi_processor_performance *perf;
0681     struct acpi_cpufreq_data *data;
0682     unsigned int cpu = policy->cpu;
0683     struct cpuinfo_x86 *c = &cpu_data(cpu);
0684     unsigned int valid_states = 0;
0685     unsigned int result = 0;
0686     u64 max_boost_ratio;
0687     unsigned int i;
0688 #ifdef CONFIG_SMP
0689     static int blacklisted;
0690 #endif
0691 
0692     pr_debug("%s\n", __func__);
0693 
0694 #ifdef CONFIG_SMP
0695     if (blacklisted)
0696         return blacklisted;
0697     blacklisted = acpi_cpufreq_blacklist(c);
0698     if (blacklisted)
0699         return blacklisted;
0700 #endif
0701 
0702     data = kzalloc(sizeof(*data), GFP_KERNEL);
0703     if (!data)
0704         return -ENOMEM;
0705 
0706     if (!zalloc_cpumask_var(&data->freqdomain_cpus, GFP_KERNEL)) {
0707         result = -ENOMEM;
0708         goto err_free;
0709     }
0710 
0711     perf = per_cpu_ptr(acpi_perf_data, cpu);
0712     data->acpi_perf_cpu = cpu;
0713     policy->driver_data = data;
0714 
0715     if (cpu_has(c, X86_FEATURE_CONSTANT_TSC))
0716         acpi_cpufreq_driver.flags |= CPUFREQ_CONST_LOOPS;
0717 
0718     result = acpi_processor_register_performance(perf, cpu);
0719     if (result)
0720         goto err_free_mask;
0721 
0722     policy->shared_type = perf->shared_type;
0723 
0724     /*
0725      * Will let policy->cpus know about dependency only when software
0726      * coordination is required.
0727      */
0728     if (policy->shared_type == CPUFREQ_SHARED_TYPE_ALL ||
0729         policy->shared_type == CPUFREQ_SHARED_TYPE_ANY) {
0730         cpumask_copy(policy->cpus, perf->shared_cpu_map);
0731     }
0732     cpumask_copy(data->freqdomain_cpus, perf->shared_cpu_map);
0733 
0734 #ifdef CONFIG_SMP
0735     dmi_check_system(sw_any_bug_dmi_table);
0736     if (bios_with_sw_any_bug && !policy_is_shared(policy)) {
0737         policy->shared_type = CPUFREQ_SHARED_TYPE_ALL;
0738         cpumask_copy(policy->cpus, topology_core_cpumask(cpu));
0739     }
0740 
0741     if (check_amd_hwpstate_cpu(cpu) && boot_cpu_data.x86 < 0x19 &&
0742         !acpi_pstate_strict) {
0743         cpumask_clear(policy->cpus);
0744         cpumask_set_cpu(cpu, policy->cpus);
0745         cpumask_copy(data->freqdomain_cpus,
0746                  topology_sibling_cpumask(cpu));
0747         policy->shared_type = CPUFREQ_SHARED_TYPE_HW;
0748         pr_info_once("overriding BIOS provided _PSD data\n");
0749     }
0750 #endif
0751 
0752     /* capability check */
0753     if (perf->state_count <= 1) {
0754         pr_debug("No P-States\n");
0755         result = -ENODEV;
0756         goto err_unreg;
0757     }
0758 
0759     if (perf->control_register.space_id != perf->status_register.space_id) {
0760         result = -ENODEV;
0761         goto err_unreg;
0762     }
0763 
0764     switch (perf->control_register.space_id) {
0765     case ACPI_ADR_SPACE_SYSTEM_IO:
0766         if (boot_cpu_data.x86_vendor == X86_VENDOR_AMD &&
0767             boot_cpu_data.x86 == 0xf) {
0768             pr_debug("AMD K8 systems must use native drivers.\n");
0769             result = -ENODEV;
0770             goto err_unreg;
0771         }
0772         pr_debug("SYSTEM IO addr space\n");
0773         data->cpu_feature = SYSTEM_IO_CAPABLE;
0774         data->cpu_freq_read = cpu_freq_read_io;
0775         data->cpu_freq_write = cpu_freq_write_io;
0776         break;
0777     case ACPI_ADR_SPACE_FIXED_HARDWARE:
0778         pr_debug("HARDWARE addr space\n");
0779         if (check_est_cpu(cpu)) {
0780             data->cpu_feature = SYSTEM_INTEL_MSR_CAPABLE;
0781             data->cpu_freq_read = cpu_freq_read_intel;
0782             data->cpu_freq_write = cpu_freq_write_intel;
0783             break;
0784         }
0785         if (check_amd_hwpstate_cpu(cpu)) {
0786             data->cpu_feature = SYSTEM_AMD_MSR_CAPABLE;
0787             data->cpu_freq_read = cpu_freq_read_amd;
0788             data->cpu_freq_write = cpu_freq_write_amd;
0789             break;
0790         }
0791         result = -ENODEV;
0792         goto err_unreg;
0793     default:
0794         pr_debug("Unknown addr space %d\n",
0795             (u32) (perf->control_register.space_id));
0796         result = -ENODEV;
0797         goto err_unreg;
0798     }
0799 
0800     freq_table = kcalloc(perf->state_count + 1, sizeof(*freq_table),
0801                  GFP_KERNEL);
0802     if (!freq_table) {
0803         result = -ENOMEM;
0804         goto err_unreg;
0805     }
0806 
0807     /* detect transition latency */
0808     policy->cpuinfo.transition_latency = 0;
0809     for (i = 0; i < perf->state_count; i++) {
0810         if ((perf->states[i].transition_latency * 1000) >
0811             policy->cpuinfo.transition_latency)
0812             policy->cpuinfo.transition_latency =
0813                 perf->states[i].transition_latency * 1000;
0814     }
0815 
0816     /* Check for high latency (>20uS) from buggy BIOSes, like on T42 */
0817     if (perf->control_register.space_id == ACPI_ADR_SPACE_FIXED_HARDWARE &&
0818         policy->cpuinfo.transition_latency > 20 * 1000) {
0819         policy->cpuinfo.transition_latency = 20 * 1000;
0820         pr_info_once("P-state transition latency capped at 20 uS\n");
0821     }
0822 
0823     /* table init */
0824     for (i = 0; i < perf->state_count; i++) {
0825         if (i > 0 && perf->states[i].core_frequency >=
0826             freq_table[valid_states-1].frequency / 1000)
0827             continue;
0828 
0829         freq_table[valid_states].driver_data = i;
0830         freq_table[valid_states].frequency =
0831             perf->states[i].core_frequency * 1000;
0832         valid_states++;
0833     }
0834     freq_table[valid_states].frequency = CPUFREQ_TABLE_END;
0835 
0836     max_boost_ratio = get_max_boost_ratio(cpu);
0837     if (max_boost_ratio) {
0838         unsigned int freq = freq_table[0].frequency;
0839 
0840         /*
0841          * Because the loop above sorts the freq_table entries in the
0842          * descending order, freq is the maximum frequency in the table.
0843          * Assume that it corresponds to the CPPC nominal frequency and
0844          * use it to set cpuinfo.max_freq.
0845          */
0846         policy->cpuinfo.max_freq = freq * max_boost_ratio >> SCHED_CAPACITY_SHIFT;
0847     } else {
0848         /*
0849          * If the maximum "boost" frequency is unknown, ask the arch
0850          * scale-invariance code to use the "nominal" performance for
0851          * CPU utilization scaling so as to prevent the schedutil
0852          * governor from selecting inadequate CPU frequencies.
0853          */
0854         arch_set_max_freq_ratio(true);
0855     }
0856 
0857     policy->freq_table = freq_table;
0858     perf->state = 0;
0859 
0860     switch (perf->control_register.space_id) {
0861     case ACPI_ADR_SPACE_SYSTEM_IO:
0862         /*
0863          * The core will not set policy->cur, because
0864          * cpufreq_driver->get is NULL, so we need to set it here.
0865          * However, we have to guess it, because the current speed is
0866          * unknown and not detectable via IO ports.
0867          */
0868         policy->cur = acpi_cpufreq_guess_freq(data, policy->cpu);
0869         break;
0870     case ACPI_ADR_SPACE_FIXED_HARDWARE:
0871         acpi_cpufreq_driver.get = get_cur_freq_on_cpu;
0872         break;
0873     default:
0874         break;
0875     }
0876 
0877     /* notify BIOS that we exist */
0878     acpi_processor_notify_smm(THIS_MODULE);
0879 
0880     pr_debug("CPU%u - ACPI performance management activated.\n", cpu);
0881     for (i = 0; i < perf->state_count; i++)
0882         pr_debug("     %cP%d: %d MHz, %d mW, %d uS\n",
0883             (i == perf->state ? '*' : ' '), i,
0884             (u32) perf->states[i].core_frequency,
0885             (u32) perf->states[i].power,
0886             (u32) perf->states[i].transition_latency);
0887 
0888     /*
0889      * the first call to ->target() should result in us actually
0890      * writing something to the appropriate registers.
0891      */
0892     data->resume = 1;
0893 
0894     policy->fast_switch_possible = !acpi_pstate_strict &&
0895         !(policy_is_shared(policy) && policy->shared_type != CPUFREQ_SHARED_TYPE_ANY);
0896 
0897     if (perf->states[0].core_frequency * 1000 != freq_table[0].frequency)
0898         pr_warn(FW_WARN "P-state 0 is not max freq\n");
0899 
0900     return result;
0901 
0902 err_unreg:
0903     acpi_processor_unregister_performance(cpu);
0904 err_free_mask:
0905     free_cpumask_var(data->freqdomain_cpus);
0906 err_free:
0907     kfree(data);
0908     policy->driver_data = NULL;
0909 
0910     return result;
0911 }
0912 
0913 static int acpi_cpufreq_cpu_exit(struct cpufreq_policy *policy)
0914 {
0915     struct acpi_cpufreq_data *data = policy->driver_data;
0916 
0917     pr_debug("%s\n", __func__);
0918 
0919     policy->fast_switch_possible = false;
0920     policy->driver_data = NULL;
0921     acpi_processor_unregister_performance(data->acpi_perf_cpu);
0922     free_cpumask_var(data->freqdomain_cpus);
0923     kfree(policy->freq_table);
0924     kfree(data);
0925 
0926     return 0;
0927 }
0928 
0929 static int acpi_cpufreq_resume(struct cpufreq_policy *policy)
0930 {
0931     struct acpi_cpufreq_data *data = policy->driver_data;
0932 
0933     pr_debug("%s\n", __func__);
0934 
0935     data->resume = 1;
0936 
0937     return 0;
0938 }
0939 
0940 static struct freq_attr *acpi_cpufreq_attr[] = {
0941     &cpufreq_freq_attr_scaling_available_freqs,
0942     &freqdomain_cpus,
0943 #ifdef CONFIG_X86_ACPI_CPUFREQ_CPB
0944     &cpb,
0945 #endif
0946     NULL,
0947 };
0948 
0949 static struct cpufreq_driver acpi_cpufreq_driver = {
0950     .verify     = cpufreq_generic_frequency_table_verify,
0951     .target_index   = acpi_cpufreq_target,
0952     .fast_switch    = acpi_cpufreq_fast_switch,
0953     .bios_limit = acpi_processor_get_bios_limit,
0954     .init       = acpi_cpufreq_cpu_init,
0955     .exit       = acpi_cpufreq_cpu_exit,
0956     .resume     = acpi_cpufreq_resume,
0957     .name       = "acpi-cpufreq",
0958     .attr       = acpi_cpufreq_attr,
0959 };
0960 
0961 static enum cpuhp_state acpi_cpufreq_online;
0962 
0963 static void __init acpi_cpufreq_boost_init(void)
0964 {
0965     int ret;
0966 
0967     if (!(boot_cpu_has(X86_FEATURE_CPB) || boot_cpu_has(X86_FEATURE_IDA))) {
0968         pr_debug("Boost capabilities not present in the processor\n");
0969         return;
0970     }
0971 
0972     acpi_cpufreq_driver.set_boost = set_boost;
0973     acpi_cpufreq_driver.boost_enabled = boost_state(0);
0974 
0975     /*
0976      * This calls the online callback on all online cpu and forces all
0977      * MSRs to the same value.
0978      */
0979     ret = cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "cpufreq/acpi:online",
0980                 cpufreq_boost_online, cpufreq_boost_down_prep);
0981     if (ret < 0) {
0982         pr_err("acpi_cpufreq: failed to register hotplug callbacks\n");
0983         return;
0984     }
0985     acpi_cpufreq_online = ret;
0986 }
0987 
0988 static void acpi_cpufreq_boost_exit(void)
0989 {
0990     if (acpi_cpufreq_online > 0)
0991         cpuhp_remove_state_nocalls(acpi_cpufreq_online);
0992 }
0993 
0994 static int __init acpi_cpufreq_init(void)
0995 {
0996     int ret;
0997 
0998     if (acpi_disabled)
0999         return -ENODEV;
1000 
1001     /* don't keep reloading if cpufreq_driver exists */
1002     if (cpufreq_get_current_driver())
1003         return -EEXIST;
1004 
1005     pr_debug("%s\n", __func__);
1006 
1007     ret = acpi_cpufreq_early_init();
1008     if (ret)
1009         return ret;
1010 
1011 #ifdef CONFIG_X86_ACPI_CPUFREQ_CPB
1012     /* this is a sysfs file with a strange name and an even stranger
1013      * semantic - per CPU instantiation, but system global effect.
1014      * Lets enable it only on AMD CPUs for compatibility reasons and
1015      * only if configured. This is considered legacy code, which
1016      * will probably be removed at some point in the future.
1017      */
1018     if (!check_amd_hwpstate_cpu(0)) {
1019         struct freq_attr **attr;
1020 
1021         pr_debug("CPB unsupported, do not expose it\n");
1022 
1023         for (attr = acpi_cpufreq_attr; *attr; attr++)
1024             if (*attr == &cpb) {
1025                 *attr = NULL;
1026                 break;
1027             }
1028     }
1029 #endif
1030     acpi_cpufreq_boost_init();
1031 
1032     ret = cpufreq_register_driver(&acpi_cpufreq_driver);
1033     if (ret) {
1034         free_acpi_perf_data();
1035         acpi_cpufreq_boost_exit();
1036     }
1037     return ret;
1038 }
1039 
1040 static void __exit acpi_cpufreq_exit(void)
1041 {
1042     pr_debug("%s\n", __func__);
1043 
1044     acpi_cpufreq_boost_exit();
1045 
1046     cpufreq_unregister_driver(&acpi_cpufreq_driver);
1047 
1048     free_acpi_perf_data();
1049 }
1050 
1051 module_param(acpi_pstate_strict, uint, 0644);
1052 MODULE_PARM_DESC(acpi_pstate_strict,
1053     "value 0 or non-zero. non-zero -> strict ACPI checks are "
1054     "performed during frequency changes.");
1055 
1056 late_initcall(acpi_cpufreq_init);
1057 module_exit(acpi_cpufreq_exit);
1058 
1059 static const struct x86_cpu_id __maybe_unused acpi_cpufreq_ids[] = {
1060     X86_MATCH_FEATURE(X86_FEATURE_ACPI, NULL),
1061     X86_MATCH_FEATURE(X86_FEATURE_HW_PSTATE, NULL),
1062     {}
1063 };
1064 MODULE_DEVICE_TABLE(x86cpu, acpi_cpufreq_ids);
1065 
1066 static const struct acpi_device_id __maybe_unused processor_device_ids[] = {
1067     {ACPI_PROCESSOR_OBJECT_HID, },
1068     {ACPI_PROCESSOR_DEVICE_HID, },
1069     {},
1070 };
1071 MODULE_DEVICE_TABLE(acpi, processor_device_ids);
1072 
1073 MODULE_ALIAS("acpi");
[ Source navigation ] [ Diff markup ] [ Identifier search ] [ general search ]

This page was automatically generated by the 2.1.0 LXR engine.

The LXR team
Valid CSS 2.1! Valid HTML 4.01!