acpi-cpufreq.c

   1// SPDX-License-Identifier: GPL-2.0-or-later
   2/*
   3 * acpi-cpufreq.c - ACPI Processor P-States Driver
   4 *
   5 *  Copyright (C) 2001, 2002 Andy Grover <andrew.grover@intel.com>
   6 *  Copyright (C) 2001, 2002 Paul Diefenbaugh <paul.s.diefenbaugh@intel.com>
   7 *  Copyright (C) 2002 - 2004 Dominik Brodowski <linux@brodo.de>
   8 *  Copyright (C) 2006       Denis Sadykov <denis.m.sadykov@intel.com>
   9 */
  10
  11#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  12
  13#include <linux/kernel.h>
  14#include <linux/module.h>
  15#include <linux/init.h>
  16#include <linux/smp.h>
  17#include <linux/sched.h>
  18#include <linux/cpufreq.h>
  19#include <linux/compiler.h>
  20#include <linux/dmi.h>
  21#include <linux/slab.h>
  22#include <linux/string_helpers.h>
  23#include <linux/platform_device.h>
  24
  25#include <linux/acpi.h>
  26#include <linux/io.h>
  27#include <linux/delay.h>
  28#include <linux/uaccess.h>
  29
  30#include <acpi/processor.h>
  31#include <acpi/cppc_acpi.h>
  32
  33#include <asm/msr.h>
  34#include <asm/processor.h>
  35#include <asm/cpufeature.h>
  36#include <asm/cpu_device_id.h>
  37
  38MODULE_AUTHOR("Paul Diefenbaugh, Dominik Brodowski");
  39MODULE_DESCRIPTION("ACPI Processor P-States Driver");
  40MODULE_LICENSE("GPL");
  41
  42enum {
  43	UNDEFINED_CAPABLE = 0,
  44	SYSTEM_INTEL_MSR_CAPABLE,
  45	SYSTEM_AMD_MSR_CAPABLE,
  46	SYSTEM_IO_CAPABLE,
  47};
  48
  49#define INTEL_MSR_RANGE		(0xffff)
  50#define AMD_MSR_RANGE		(0x7)
  51#define HYGON_MSR_RANGE		(0x7)
  52
  53struct acpi_cpufreq_data {
  54	unsigned int resume;
  55	unsigned int cpu_feature;
  56	unsigned int acpi_perf_cpu;
  57	cpumask_var_t freqdomain_cpus;
  58	void (*cpu_freq_write)(struct acpi_pct_register *reg, u32 val);
  59	u32 (*cpu_freq_read)(struct acpi_pct_register *reg);
  60};
  61
  62/* acpi_perf_data is a pointer to percpu data. */
  63static struct acpi_processor_performance __percpu *acpi_perf_data;
  64
  65static inline struct acpi_processor_performance *to_perf_data(struct acpi_cpufreq_data *data)
  66{
  67	return per_cpu_ptr(acpi_perf_data, data->acpi_perf_cpu);
  68}
  69
  70static struct cpufreq_driver acpi_cpufreq_driver;
  71
  72static unsigned int acpi_pstate_strict;
  73
  74static bool boost_state(unsigned int cpu)
  75{
  76	u64 msr;
  77
  78	switch (boot_cpu_data.x86_vendor) {
  79	case X86_VENDOR_INTEL:
  80	case X86_VENDOR_CENTAUR:
  81	case X86_VENDOR_ZHAOXIN:
  82		rdmsrq_on_cpu(cpu, MSR_IA32_MISC_ENABLE, &msr);
  83		return !(msr & MSR_IA32_MISC_ENABLE_TURBO_DISABLE);
  84	case X86_VENDOR_HYGON:
  85	case X86_VENDOR_AMD:
  86		rdmsrq_on_cpu(cpu, MSR_K7_HWCR, &msr);
  87		return !(msr & MSR_K7_HWCR_CPB_DIS);
  88	}
  89	return false;
  90}
  91
  92static int boost_set_msr(bool enable)
  93{
  94	u32 msr_addr;
  95	u64 msr_mask, val;
  96
  97	switch (boot_cpu_data.x86_vendor) {
  98	case X86_VENDOR_INTEL:
  99	case X86_VENDOR_CENTAUR:
 100	case X86_VENDOR_ZHAOXIN:
 101		msr_addr = MSR_IA32_MISC_ENABLE;
 102		msr_mask = MSR_IA32_MISC_ENABLE_TURBO_DISABLE;
 103		break;
 104	case X86_VENDOR_HYGON:
 105	case X86_VENDOR_AMD:
 106		msr_addr = MSR_K7_HWCR;
 107		msr_mask = MSR_K7_HWCR_CPB_DIS;
 108		break;
 109	default:
 110		return -EINVAL;
 111	}
 112
 113	rdmsrq(msr_addr, val);
 114
 115	if (enable)
 116		val &= ~msr_mask;
 117	else
 118		val |= msr_mask;
 119
 120	wrmsrq(msr_addr, val);
 121	return 0;
 122}
 123
 124static void boost_set_msr_each(void *p_en)
 125{
 126	bool enable = (bool) p_en;
 127
 128	boost_set_msr(enable);
 129}
 130
 131static int set_boost(struct cpufreq_policy *policy, int val)
 132{
 133	on_each_cpu_mask(policy->cpus, boost_set_msr_each,
 134			 (void *)(long)val, 1);
 135	pr_debug("CPU %*pbl: Core Boosting %s.\n",
 136		 cpumask_pr_args(policy->cpus), str_enabled_disabled(val));
 137
 138	return 0;
 139}
 140
 141static ssize_t show_freqdomain_cpus(struct cpufreq_policy *policy, char *buf)
 142{
 143	struct acpi_cpufreq_data *data = policy->driver_data;
 144
 145	if (unlikely(!data))
 146		return -ENODEV;
 147
 148	return cpufreq_show_cpus(data->freqdomain_cpus, buf);
 149}
 150
 151cpufreq_freq_attr_ro(freqdomain_cpus);
 152
 153#ifdef CONFIG_X86_ACPI_CPUFREQ_CPB
 154static ssize_t store_cpb(struct cpufreq_policy *policy, const char *buf,
 155			 size_t count)
 156{
 157	int ret;
 158	unsigned int val = 0;
 159
 160	if (!acpi_cpufreq_driver.set_boost)
 161		return -EINVAL;
 162
 163	ret = kstrtouint(buf, 10, &val);
 164	if (ret || val > 1)
 165		return -EINVAL;
 166
 167	cpus_read_lock();
 168	set_boost(policy, val);
 169	cpus_read_unlock();
 170
 171	return count;
 172}
 173
 174static ssize_t show_cpb(struct cpufreq_policy *policy, char *buf)
 175{
 176	return sprintf(buf, "%u\n", acpi_cpufreq_driver.boost_enabled);
 177}
 178
 179cpufreq_freq_attr_rw(cpb);
 180#endif
 181
 182static int check_est_cpu(unsigned int cpuid)
 183{
 184	struct cpuinfo_x86 *cpu = &cpu_data(cpuid);
 185
 186	return cpu_has(cpu, X86_FEATURE_EST);
 187}
 188
 189static int check_amd_hwpstate_cpu(unsigned int cpuid)
 190{
 191	struct cpuinfo_x86 *cpu = &cpu_data(cpuid);
 192
 193	return cpu_has(cpu, X86_FEATURE_HW_PSTATE);
 194}
 195
 196static unsigned extract_io(struct cpufreq_policy *policy, u32 value)
 197{
 198	struct acpi_cpufreq_data *data = policy->driver_data;
 199	struct acpi_processor_performance *perf;
 200	int i;
 201
 202	perf = to_perf_data(data);
 203
 204	for (i = 0; i < perf->state_count; i++) {
 205		if (value == perf->states[i].status)
 206			return policy->freq_table[i].frequency;
 207	}
 208	return 0;
 209}
 210
 211static unsigned extract_msr(struct cpufreq_policy *policy, u32 msr)
 212{
 213	struct acpi_cpufreq_data *data = policy->driver_data;
 214	struct cpufreq_frequency_table *pos;
 215	struct acpi_processor_performance *perf;
 216
 217	if (boot_cpu_data.x86_vendor == X86_VENDOR_AMD)
 218		msr &= AMD_MSR_RANGE;
 219	else if (boot_cpu_data.x86_vendor == X86_VENDOR_HYGON)
 220		msr &= HYGON_MSR_RANGE;
 221	else
 222		msr &= INTEL_MSR_RANGE;
 223
 224	perf = to_perf_data(data);
 225
 226	cpufreq_for_each_entry(pos, policy->freq_table)
 227		if (msr == perf->states[pos->driver_data].status)
 228			return pos->frequency;
 229	return policy->freq_table[0].frequency;
 230}
 231
 232static unsigned extract_freq(struct cpufreq_policy *policy, u32 val)
 233{
 234	struct acpi_cpufreq_data *data = policy->driver_data;
 235
 236	switch (data->cpu_feature) {
 237	case SYSTEM_INTEL_MSR_CAPABLE:
 238	case SYSTEM_AMD_MSR_CAPABLE:
 239		return extract_msr(policy, val);
 240	case SYSTEM_IO_CAPABLE:
 241		return extract_io(policy, val);
 242	default:
 243		return 0;
 244	}
 245}
 246
 247static u32 cpu_freq_read_intel(struct acpi_pct_register *not_used)
 248{
 249	u32 val, dummy __always_unused;
 250
 251	rdmsr(MSR_IA32_PERF_CTL, val, dummy);
 252	return val;
 253}
 254
 255static void cpu_freq_write_intel(struct acpi_pct_register *not_used, u32 val)
 256{
 257	u32 lo, hi;
 258
 259	rdmsr(MSR_IA32_PERF_CTL, lo, hi);
 260	lo = (lo & ~INTEL_MSR_RANGE) | (val & INTEL_MSR_RANGE);
 261	wrmsr(MSR_IA32_PERF_CTL, lo, hi);
 262}
 263
 264static u32 cpu_freq_read_amd(struct acpi_pct_register *not_used)
 265{
 266	u32 val, dummy __always_unused;
 267
 268	rdmsr(MSR_AMD_PERF_CTL, val, dummy);
 269	return val;
 270}
 271
 272static void cpu_freq_write_amd(struct acpi_pct_register *not_used, u32 val)
 273{
 274	wrmsr(MSR_AMD_PERF_CTL, val, 0);
 275}
 276
 277static u32 cpu_freq_read_io(struct acpi_pct_register *reg)
 278{
 279	u32 val;
 280
 281	acpi_os_read_port(reg->address, &val, reg->bit_width);
 282	return val;
 283}
 284
 285static void cpu_freq_write_io(struct acpi_pct_register *reg, u32 val)
 286{
 287	acpi_os_write_port(reg->address, val, reg->bit_width);
 288}
 289
 290struct drv_cmd {
 291	struct acpi_pct_register *reg;
 292	u32 val;
 293	union {
 294		void (*write)(struct acpi_pct_register *reg, u32 val);
 295		u32 (*read)(struct acpi_pct_register *reg);
 296	} func;
 297};
 298
 299/* Called via smp_call_function_single(), on the target CPU */
 300static void do_drv_read(void *_cmd)
 301{
 302	struct drv_cmd *cmd = _cmd;
 303
 304	cmd->val = cmd->func.read(cmd->reg);
 305}
 306
 307static u32 drv_read(struct acpi_cpufreq_data *data, const struct cpumask *mask)
 308{
 309	struct acpi_processor_performance *perf = to_perf_data(data);
 310	struct drv_cmd cmd = {
 311		.reg = &perf->control_register,
 312		.func.read = data->cpu_freq_read,
 313	};
 314	int err;
 315
 316	err = smp_call_function_any(mask, do_drv_read, &cmd, 1);
 317	WARN_ON_ONCE(err);	/* smp_call_function_any() was buggy? */
 318	return cmd.val;
 319}
 320
 321static void do_drv_write(void *_cmd)
 322{
 323	struct drv_cmd *cmd = _cmd;
 324
 325	cmd->func.write(cmd->reg, cmd->val);
 326}
 327
 328static void drv_write(struct acpi_cpufreq_data *data,
 329		      const struct cpumask *mask, u32 val)
 330{
 331	struct acpi_processor_performance *perf = to_perf_data(data);
 332	struct drv_cmd cmd = {
 333		.reg = &perf->control_register,
 334		.val = val,
 335		.func.write = data->cpu_freq_write,
 336	};
 337
 338	on_each_cpu_mask(mask, do_drv_write, &cmd, true);
 339}
 340
 341static u32 get_cur_val(const struct cpumask *mask, struct acpi_cpufreq_data *data)
 342{
 343	u32 val;
 344
 345	if (unlikely(cpumask_empty(mask)))
 346		return 0;
 347
 348	val = drv_read(data, mask);
 349
 350	pr_debug("%s = %u\n", __func__, val);
 351
 352	return val;
 353}
 354
 355static unsigned int get_cur_freq_on_cpu(unsigned int cpu)
 356{
 357	struct acpi_cpufreq_data *data;
 358	struct cpufreq_policy *policy;
 359	unsigned int freq;
 360	unsigned int cached_freq;
 361
 362	pr_debug("%s (%d)\n", __func__, cpu);
 363
 364	policy = cpufreq_cpu_get_raw(cpu);
 365	if (unlikely(!policy))
 366		return 0;
 367
 368	data = policy->driver_data;
 369	if (unlikely(!data || !policy->freq_table))
 370		return 0;
 371
 372	cached_freq = policy->freq_table[to_perf_data(data)->state].frequency;
 373	freq = extract_freq(policy, get_cur_val(cpumask_of(cpu), data));
 374	if (freq != cached_freq) {
 375		/*
 376		 * The dreaded BIOS frequency change behind our back.
 377		 * Force set the frequency on next target call.
 378		 */
 379		data->resume = 1;
 380	}
 381
 382	pr_debug("cur freq = %u\n", freq);
 383
 384	return freq;
 385}
 386
 387static unsigned int check_freqs(struct cpufreq_policy *policy,
 388				const struct cpumask *mask, unsigned int freq)
 389{
 390	struct acpi_cpufreq_data *data = policy->driver_data;
 391	unsigned int cur_freq;
 392	unsigned int i;
 393
 394	for (i = 0; i < 100; i++) {
 395		cur_freq = extract_freq(policy, get_cur_val(mask, data));
 396		if (cur_freq == freq)
 397			return 1;
 398		usleep_range(10, 15);
 399	}
 400	return 0;
 401}
 402
 403static int acpi_cpufreq_target(struct cpufreq_policy *policy,
 404			       unsigned int index)
 405{
 406	struct acpi_cpufreq_data *data = policy->driver_data;
 407	struct acpi_processor_performance *perf;
 408	const struct cpumask *mask;
 409	unsigned int next_perf_state = 0; /* Index into perf table */
 410	int result = 0;
 411
 412	if (unlikely(!data)) {
 413		return -ENODEV;
 414	}
 415
 416	perf = to_perf_data(data);
 417	next_perf_state = policy->freq_table[index].driver_data;
 418	if (perf->state == next_perf_state) {
 419		if (unlikely(data->resume)) {
 420			pr_debug("Called after resume, resetting to P%d\n",
 421				next_perf_state);
 422			data->resume = 0;
 423		} else {
 424			pr_debug("Already at target state (P%d)\n",
 425				next_perf_state);
 426			return 0;
 427		}
 428	}
 429
 430	/*
 431	 * The core won't allow CPUs to go away until the governor has been
 432	 * stopped, so we can rely on the stability of policy->cpus.
 433	 */
 434	mask = policy->shared_type == CPUFREQ_SHARED_TYPE_ANY ?
 435		cpumask_of(policy->cpu) : policy->cpus;
 436
 437	drv_write(data, mask, perf->states[next_perf_state].control);
 438
 439	if (acpi_pstate_strict) {
 440		if (!check_freqs(policy, mask,
 441				 policy->freq_table[index].frequency)) {
 442			pr_debug("%s (%d)\n", __func__, policy->cpu);
 443			result = -EAGAIN;
 444		}
 445	}
 446
 447	if (!result)
 448		perf->state = next_perf_state;
 449
 450	return result;
 451}
 452
 453static unsigned int acpi_cpufreq_fast_switch(struct cpufreq_policy *policy,
 454					     unsigned int target_freq)
 455{
 456	struct acpi_cpufreq_data *data = policy->driver_data;
 457	struct acpi_processor_performance *perf;
 458	struct cpufreq_frequency_table *entry;
 459	unsigned int next_perf_state, next_freq, index;
 460
 461	/*
 462	 * Find the closest frequency above target_freq.
 463	 */
 464	if (policy->cached_target_freq == target_freq)
 465		index = policy->cached_resolved_idx;
 466	else
 467		index = cpufreq_table_find_index_dl(policy, target_freq,
 468						    false);
 469
 470	entry = &policy->freq_table[index];
 471	next_freq = entry->frequency;
 472	next_perf_state = entry->driver_data;
 473
 474	perf = to_perf_data(data);
 475	if (perf->state == next_perf_state) {
 476		if (unlikely(data->resume))
 477			data->resume = 0;
 478		else
 479			return next_freq;
 480	}
 481
 482	data->cpu_freq_write(&perf->control_register,
 483			     perf->states[next_perf_state].control);
 484	perf->state = next_perf_state;
 485	return next_freq;
 486}
 487
 488static unsigned long
 489acpi_cpufreq_guess_freq(struct acpi_cpufreq_data *data, unsigned int cpu)
 490{
 491	struct acpi_processor_performance *perf;
 492
 493	perf = to_perf_data(data);
 494	if (cpu_khz) {
 495		/* search the closest match to cpu_khz */
 496		unsigned int i;
 497		unsigned long freq;
 498		unsigned long freqn = perf->states[0].core_frequency * 1000;
 499
 500		for (i = 0; i < (perf->state_count-1); i++) {
 501			freq = freqn;
 502			freqn = perf->states[i+1].core_frequency * 1000;
 503			if ((2 * cpu_khz) > (freqn + freq)) {
 504				perf->state = i;
 505				return freq;
 506			}
 507		}
 508		perf->state = perf->state_count-1;
 509		return freqn;
 510	} else {
 511		/* assume CPU is at P0... */
 512		perf->state = 0;
 513		return perf->states[0].core_frequency * 1000;
 514	}
 515}
 516
 517static void free_acpi_perf_data(void)
 518{
 519	unsigned int i;
 520
 521	/* Freeing a NULL pointer is OK, and alloc_percpu zeroes. */
 522	for_each_possible_cpu(i)
 523		free_cpumask_var(per_cpu_ptr(acpi_perf_data, i)
 524				 ->shared_cpu_map);
 525	free_percpu(acpi_perf_data);
 526}
 527
 528static int cpufreq_boost_down_prep(unsigned int cpu)
 529{
 530	/*
 531	 * Clear the boost-disable bit on the CPU_DOWN path so that
 532	 * this cpu cannot block the remaining ones from boosting.
 533	 */
 534	return boost_set_msr(1);
 535}
 536
 537/*
 538 * acpi_cpufreq_early_init - initialize ACPI P-States library
 539 *
 540 * Initialize the ACPI P-States library (drivers/acpi/processor_perflib.c)
 541 * in order to determine correct frequency and voltage pairings. We can
 542 * do _PDC and _PSD and find out the processor dependency for the
 543 * actual init that will happen later...
 544 */
 545static int __init acpi_cpufreq_early_init(void)
 546{
 547	unsigned int i;
 548	pr_debug("%s\n", __func__);
 549
 550	acpi_perf_data = alloc_percpu(struct acpi_processor_performance);
 551	if (!acpi_perf_data) {
 552		pr_debug("Memory allocation error for acpi_perf_data.\n");
 553		return -ENOMEM;
 554	}
 555	for_each_possible_cpu(i) {
 556		if (!zalloc_cpumask_var_node(
 557			&per_cpu_ptr(acpi_perf_data, i)->shared_cpu_map,
 558			GFP_KERNEL, cpu_to_node(i))) {
 559
 560			/* Freeing a NULL pointer is OK: alloc_percpu zeroes. */
 561			free_acpi_perf_data();
 562			return -ENOMEM;
 563		}
 564	}
 565
 566	/* Do initialization in ACPI core */
 567	acpi_processor_preregister_performance(acpi_perf_data);
 568	return 0;
 569}
 570
 571#ifdef CONFIG_SMP
 572/*
 573 * Some BIOSes do SW_ANY coordination internally, either set it up in hw
 574 * or do it in BIOS firmware and won't inform about it to OS. If not
 575 * detected, this has a side effect of making CPU run at a different speed
 576 * than OS intended it to run at. Detect it and handle it cleanly.
 577 */
 578static int bios_with_sw_any_bug;
 579
 580static int sw_any_bug_found(const struct dmi_system_id *d)
 581{
 582	bios_with_sw_any_bug = 1;
 583	return 0;
 584}
 585
 586static const struct dmi_system_id sw_any_bug_dmi_table[] = {
 587	{
 588		.callback = sw_any_bug_found,
 589		.ident = "Supermicro Server X6DLP",
 590		.matches = {
 591			DMI_MATCH(DMI_SYS_VENDOR, "Supermicro"),
 592			DMI_MATCH(DMI_BIOS_VERSION, "080010"),
 593			DMI_MATCH(DMI_PRODUCT_NAME, "X6DLP"),
 594		},
 595	},
 596	{ }
 597};
 598
 599static int acpi_cpufreq_blacklist(struct cpuinfo_x86 *c)
 600{
 601	/* Intel Xeon Processor 7100 Series Specification Update
 602	 * https://www.intel.com/Assets/PDF/specupdate/314554.pdf
 603	 * AL30: A Machine Check Exception (MCE) Occurring during an
 604	 * Enhanced Intel SpeedStep Technology Ratio Change May Cause
 605	 * Both Processor Cores to Lock Up. */
 606	if (c->x86_vendor == X86_VENDOR_INTEL) {
 607		if ((c->x86 == 15) &&
 608		    (c->x86_model == 6) &&
 609		    (c->x86_stepping == 8)) {
 610			pr_info("Intel(R) Xeon(R) 7100 Errata AL30, processors may lock up on frequency changes: disabling acpi-cpufreq\n");
 611			return -ENODEV;
 612		    }
 613		}
 614	return 0;
 615}
 616#endif
 617
 618#ifdef CONFIG_ACPI_CPPC_LIB
 619/*
 620 * get_max_boost_ratio: Computes the max_boost_ratio as the ratio
 621 * between the highest_perf and the nominal_perf.
 622 *
 623 * Returns the max_boost_ratio for @cpu. Returns the CPPC nominal
 624 * frequency via @nominal_freq if it is non-NULL pointer.
 625 */
 626static u64 get_max_boost_ratio(unsigned int cpu, u64 *nominal_freq)
 627{
 628	struct cppc_perf_caps perf_caps;
 629	u64 highest_perf, nominal_perf;
 630	int ret;
 631
 632	if (acpi_pstate_strict)
 633		return 0;
 634
 635	ret = cppc_get_perf_caps(cpu, &perf_caps);
 636	if (ret) {
 637		pr_debug("CPU%d: Unable to get performance capabilities (%d)\n",
 638			 cpu, ret);
 639		return 0;
 640	}
 641
 642	if (boot_cpu_data.x86_vendor == X86_VENDOR_AMD) {
 643		ret = amd_get_boost_ratio_numerator(cpu, &highest_perf);
 644		if (ret) {
 645			pr_debug("CPU%d: Unable to get boost ratio numerator (%d)\n",
 646				 cpu, ret);
 647			return 0;
 648		}
 649	} else {
 650		highest_perf = perf_caps.highest_perf;
 651	}
 652
 653	nominal_perf = perf_caps.nominal_perf;
 654
 655	if (nominal_freq)
 656		*nominal_freq = perf_caps.nominal_freq * 1000;
 657
 658	if (!highest_perf || !nominal_perf) {
 659		pr_debug("CPU%d: highest or nominal performance missing\n", cpu);
 660		return 0;
 661	}
 662
 663	if (highest_perf < nominal_perf) {
 664		pr_debug("CPU%d: nominal performance above highest\n", cpu);
 665		return 0;
 666	}
 667
 668	return div_u64(highest_perf << SCHED_CAPACITY_SHIFT, nominal_perf);
 669}
 670
 671#else
 672static inline u64 get_max_boost_ratio(unsigned int cpu, u64 *nominal_freq)
 673{
 674	return 0;
 675}
 676#endif
 677
 678static int acpi_cpufreq_cpu_init(struct cpufreq_policy *policy)
 679{
 680	struct cpufreq_frequency_table *freq_table;
 681	struct acpi_processor_performance *perf;
 682	struct acpi_cpufreq_data *data;
 683	unsigned int cpu = policy->cpu;
 684	struct cpuinfo_x86 *c = &cpu_data(cpu);
 685	u64 max_boost_ratio, nominal_freq = 0;
 686	unsigned int valid_states = 0;
 687	unsigned int result = 0;
 688	unsigned int i;
 689#ifdef CONFIG_SMP
 690	static int blacklisted;
 691#endif
 692
 693	pr_debug("%s\n", __func__);
 694
 695#ifdef CONFIG_SMP
 696	if (blacklisted)
 697		return blacklisted;
 698	blacklisted = acpi_cpufreq_blacklist(c);
 699	if (blacklisted)
 700		return blacklisted;
 701#endif
 702
 703	data = kzalloc(sizeof(*data), GFP_KERNEL);
 704	if (!data)
 705		return -ENOMEM;
 706
 707	if (!zalloc_cpumask_var(&data->freqdomain_cpus, GFP_KERNEL)) {
 708		result = -ENOMEM;
 709		goto err_free;
 710	}
 711
 712	perf = per_cpu_ptr(acpi_perf_data, cpu);
 713	data->acpi_perf_cpu = cpu;
 714	policy->driver_data = data;
 715
 716	if (cpu_has(c, X86_FEATURE_CONSTANT_TSC))
 717		acpi_cpufreq_driver.flags |= CPUFREQ_CONST_LOOPS;
 718
 719	result = acpi_processor_register_performance(perf, cpu);
 720	if (result)
 721		goto err_free_mask;
 722
 723	policy->shared_type = perf->shared_type;
 724
 725	/*
 726	 * Will let policy->cpus know about dependency only when software
 727	 * coordination is required.
 728	 */
 729	if (policy->shared_type == CPUFREQ_SHARED_TYPE_ALL ||
 730	    policy->shared_type == CPUFREQ_SHARED_TYPE_ANY) {
 731		cpumask_copy(policy->cpus, perf->shared_cpu_map);
 732	}
 733	cpumask_copy(data->freqdomain_cpus, perf->shared_cpu_map);
 734
 735#ifdef CONFIG_SMP
 736	dmi_check_system(sw_any_bug_dmi_table);
 737	if (bios_with_sw_any_bug && !policy_is_shared(policy)) {
 738		policy->shared_type = CPUFREQ_SHARED_TYPE_ALL;
 739		cpumask_copy(policy->cpus, topology_core_cpumask(cpu));
 740	}
 741
 742	if (check_amd_hwpstate_cpu(cpu) && boot_cpu_data.x86 < 0x19 &&
 743	    !acpi_pstate_strict) {
 744		cpumask_clear(policy->cpus);
 745		cpumask_set_cpu(cpu, policy->cpus);
 746		cpumask_copy(data->freqdomain_cpus,
 747			     topology_sibling_cpumask(cpu));
 748		policy->shared_type = CPUFREQ_SHARED_TYPE_HW;
 749		pr_info_once("overriding BIOS provided _PSD data\n");
 750	}
 751#endif
 752
 753	/* capability check */
 754	if (perf->state_count <= 1) {
 755		pr_debug("No P-States\n");
 756		result = -ENODEV;
 757		goto err_unreg;
 758	}
 759
 760	if (perf->control_register.space_id != perf->status_register.space_id) {
 761		result = -ENODEV;
 762		goto err_unreg;
 763	}
 764
 765	switch (perf->control_register.space_id) {
 766	case ACPI_ADR_SPACE_SYSTEM_IO:
 767		if (boot_cpu_data.x86_vendor == X86_VENDOR_AMD &&
 768		    boot_cpu_data.x86 == 0xf) {
 769			pr_debug("AMD K8 systems must use native drivers.\n");
 770			result = -ENODEV;
 771			goto err_unreg;
 772		}
 773		pr_debug("SYSTEM IO addr space\n");
 774		data->cpu_feature = SYSTEM_IO_CAPABLE;
 775		data->cpu_freq_read = cpu_freq_read_io;
 776		data->cpu_freq_write = cpu_freq_write_io;
 777		break;
 778	case ACPI_ADR_SPACE_FIXED_HARDWARE:
 779		pr_debug("HARDWARE addr space\n");
 780		if (check_est_cpu(cpu)) {
 781			data->cpu_feature = SYSTEM_INTEL_MSR_CAPABLE;
 782			data->cpu_freq_read = cpu_freq_read_intel;
 783			data->cpu_freq_write = cpu_freq_write_intel;
 784			break;
 785		}
 786		if (check_amd_hwpstate_cpu(cpu)) {
 787			data->cpu_feature = SYSTEM_AMD_MSR_CAPABLE;
 788			data->cpu_freq_read = cpu_freq_read_amd;
 789			data->cpu_freq_write = cpu_freq_write_amd;
 790			break;
 791		}
 792		result = -ENODEV;
 793		goto err_unreg;
 794	default:
 795		pr_debug("Unknown addr space %d\n",
 796			(u32) (perf->control_register.space_id));
 797		result = -ENODEV;
 798		goto err_unreg;
 799	}
 800
 801	freq_table = kcalloc(perf->state_count + 1, sizeof(*freq_table),
 802			     GFP_KERNEL);
 803	if (!freq_table) {
 804		result = -ENOMEM;
 805		goto err_unreg;
 806	}
 807
 808	/* detect transition latency */
 809	policy->cpuinfo.transition_latency = 0;
 810	for (i = 0; i < perf->state_count; i++) {
 811		if ((perf->states[i].transition_latency * 1000) >
 812		    policy->cpuinfo.transition_latency)
 813			policy->cpuinfo.transition_latency =
 814			    perf->states[i].transition_latency * 1000;
 815	}
 816
 817	/* Check for high latency (>20uS) from buggy BIOSes, like on T42 */
 818	if (perf->control_register.space_id == ACPI_ADR_SPACE_FIXED_HARDWARE &&
 819	    policy->cpuinfo.transition_latency > 20 * 1000) {
 820		policy->cpuinfo.transition_latency = 20 * 1000;
 821		pr_info_once("P-state transition latency capped at 20 uS\n");
 822	}
 823
 824	/* table init */
 825	for (i = 0; i < perf->state_count; i++) {
 826		if (i > 0 && perf->states[i].core_frequency >=
 827		    freq_table[valid_states-1].frequency / 1000)
 828			continue;
 829
 830		freq_table[valid_states].driver_data = i;
 831		freq_table[valid_states].frequency =
 832		    perf->states[i].core_frequency * 1000;
 833		valid_states++;
 834	}
 835	freq_table[valid_states].frequency = CPUFREQ_TABLE_END;
 836
 837	max_boost_ratio = get_max_boost_ratio(cpu, &nominal_freq);
 838	if (max_boost_ratio) {
 839		unsigned int freq = nominal_freq;
 840
 841		/*
 842		 * The loop above sorts the freq_table entries in the
 843		 * descending order. If ACPI CPPC has not advertised
 844		 * the nominal frequency (this is possible in CPPC
 845		 * revisions prior to 3), then use the first entry in
 846		 * the pstate table as a proxy for nominal frequency.
 847		 */
 848		if (!freq)
 849			freq = freq_table[0].frequency;
 850
 851		policy->cpuinfo.max_freq = freq * max_boost_ratio >> SCHED_CAPACITY_SHIFT;
 852	} else {
 853		/*
 854		 * If the maximum "boost" frequency is unknown, ask the arch
 855		 * scale-invariance code to use the "nominal" performance for
 856		 * CPU utilization scaling so as to prevent the schedutil
 857		 * governor from selecting inadequate CPU frequencies.
 858		 */
 859		arch_set_max_freq_ratio(true);
 860	}
 861
 862	policy->freq_table = freq_table;
 863	perf->state = 0;
 864
 865	switch (perf->control_register.space_id) {
 866	case ACPI_ADR_SPACE_SYSTEM_IO:
 867		/*
 868		 * The core will not set policy->cur, because
 869		 * cpufreq_driver->get is NULL, so we need to set it here.
 870		 * However, we have to guess it, because the current speed is
 871		 * unknown and not detectable via IO ports.
 872		 */
 873		policy->cur = acpi_cpufreq_guess_freq(data, policy->cpu);
 874		break;
 875	case ACPI_ADR_SPACE_FIXED_HARDWARE:
 876		acpi_cpufreq_driver.get = get_cur_freq_on_cpu;
 877		break;
 878	default:
 879		break;
 880	}
 881
 882	/* notify BIOS that we exist */
 883	acpi_processor_notify_smm(THIS_MODULE);
 884
 885	pr_debug("CPU%u - ACPI performance management activated.\n", cpu);
 886	for (i = 0; i < perf->state_count; i++)
 887		pr_debug("     %cP%d: %d MHz, %d mW, %d uS\n",
 888			(i == perf->state ? '*' : ' '), i,
 889			(u32) perf->states[i].core_frequency,
 890			(u32) perf->states[i].power,
 891			(u32) perf->states[i].transition_latency);
 892
 893	/*
 894	 * the first call to ->target() should result in us actually
 895	 * writing something to the appropriate registers.
 896	 */
 897	data->resume = 1;
 898
 899	policy->fast_switch_possible = !acpi_pstate_strict &&
 900		!(policy_is_shared(policy) && policy->shared_type != CPUFREQ_SHARED_TYPE_ANY);
 901
 902	if (perf->states[0].core_frequency * 1000 != freq_table[0].frequency)
 903		pr_warn(FW_WARN "P-state 0 is not max freq\n");
 904
 905	if (acpi_cpufreq_driver.set_boost) {
 906		if (policy->boost_supported) {
 907			/*
 908			 * The firmware may have altered boost state while the
 909			 * CPU was offline (for example during a suspend-resume
 910			 * cycle).
 911			 */
 912			if (policy->boost_enabled != boost_state(cpu))
 913				set_boost(policy, policy->boost_enabled);
 914		} else {
 915			policy->boost_supported = true;
 916		}
 917	}
 918
 919	return result;
 920
 921err_unreg:
 922	acpi_processor_unregister_performance(cpu);
 923err_free_mask:
 924	free_cpumask_var(data->freqdomain_cpus);
 925err_free:
 926	kfree(data);
 927	policy->driver_data = NULL;
 928
 929	return result;
 930}
 931
 932static void acpi_cpufreq_cpu_exit(struct cpufreq_policy *policy)
 933{
 934	struct acpi_cpufreq_data *data = policy->driver_data;
 935
 936	pr_debug("%s\n", __func__);
 937
 938	cpufreq_boost_down_prep(policy->cpu);
 939	policy->fast_switch_possible = false;
 940	policy->driver_data = NULL;
 941	acpi_processor_unregister_performance(data->acpi_perf_cpu);
 942	free_cpumask_var(data->freqdomain_cpus);
 943	kfree(policy->freq_table);
 944	kfree(data);
 945}
 946
 947static int acpi_cpufreq_resume(struct cpufreq_policy *policy)
 948{
 949	struct acpi_cpufreq_data *data = policy->driver_data;
 950
 951	pr_debug("%s\n", __func__);
 952
 953	data->resume = 1;
 954
 955	return 0;
 956}
 957
 958static struct freq_attr *acpi_cpufreq_attr[] = {
 959	&freqdomain_cpus,
 960#ifdef CONFIG_X86_ACPI_CPUFREQ_CPB
 961	&cpb,
 962#endif
 963	NULL,
 964};
 965
 966static struct cpufreq_driver acpi_cpufreq_driver = {
 967	.verify		= cpufreq_generic_frequency_table_verify,
 968	.target_index	= acpi_cpufreq_target,
 969	.fast_switch	= acpi_cpufreq_fast_switch,
 970	.bios_limit	= acpi_processor_get_bios_limit,
 971	.init		= acpi_cpufreq_cpu_init,
 972	.exit		= acpi_cpufreq_cpu_exit,
 973	.resume		= acpi_cpufreq_resume,
 974	.name		= "acpi-cpufreq",
 975	.attr		= acpi_cpufreq_attr,
 976};
 977
 978static void __init acpi_cpufreq_boost_init(void)
 979{
 980	if (!(boot_cpu_has(X86_FEATURE_CPB) || boot_cpu_has(X86_FEATURE_IDA))) {
 981		pr_debug("Boost capabilities not present in the processor\n");
 982		return;
 983	}
 984
 985	acpi_cpufreq_driver.set_boost = set_boost;
 986	acpi_cpufreq_driver.boost_enabled = boost_state(0);
 987}
 988
 989static int __init acpi_cpufreq_probe(struct platform_device *pdev)
 990{
 991	int ret;
 992
 993	if (acpi_disabled)
 994		return -ENODEV;
 995
 996	/* don't keep reloading if cpufreq_driver exists */
 997	if (cpufreq_get_current_driver())
 998		return -ENODEV;
 999
1000	pr_debug("%s\n", __func__);
1001
1002	ret = acpi_cpufreq_early_init();
1003	if (ret)
1004		return ret;
1005
1006#ifdef CONFIG_X86_ACPI_CPUFREQ_CPB
1007	/* this is a sysfs file with a strange name and an even stranger
1008	 * semantic - per CPU instantiation, but system global effect.
1009	 * Lets enable it only on AMD CPUs for compatibility reasons and
1010	 * only if configured. This is considered legacy code, which
1011	 * will probably be removed at some point in the future.
1012	 */
1013	if (!check_amd_hwpstate_cpu(0)) {
1014		struct freq_attr **attr;
1015
1016		pr_debug("CPB unsupported, do not expose it\n");
1017
1018		for (attr = acpi_cpufreq_attr; *attr; attr++)
1019			if (*attr == &cpb) {
1020				*attr = NULL;
1021				break;
1022			}
1023	}
1024#endif
1025	acpi_cpufreq_boost_init();
1026
1027	ret = cpufreq_register_driver(&acpi_cpufreq_driver);
1028	if (ret) {
1029		free_acpi_perf_data();
1030	}
1031	return ret;
1032}
1033
1034static void acpi_cpufreq_remove(struct platform_device *pdev)
1035{
1036	pr_debug("%s\n", __func__);
1037
1038	cpufreq_unregister_driver(&acpi_cpufreq_driver);
1039
1040	free_acpi_perf_data();
1041}
1042
1043static struct platform_driver acpi_cpufreq_platdrv = {
1044	.driver = {
1045		.name	= "acpi-cpufreq",
1046	},
1047	.remove = acpi_cpufreq_remove,
1048};
1049
1050static int __init acpi_cpufreq_init(void)
1051{
1052	return platform_driver_probe(&acpi_cpufreq_platdrv, acpi_cpufreq_probe);
1053}
1054
1055static void __exit acpi_cpufreq_exit(void)
1056{
1057	platform_driver_unregister(&acpi_cpufreq_platdrv);
1058}
1059
1060module_param(acpi_pstate_strict, uint, 0644);
1061MODULE_PARM_DESC(acpi_pstate_strict,
1062	"value 0 or non-zero. non-zero -> strict ACPI checks are "
1063	"performed during frequency changes.");
1064
1065late_initcall(acpi_cpufreq_init);
1066module_exit(acpi_cpufreq_exit);
1067
1068MODULE_ALIAS("platform:acpi-cpufreq");