xref: /linux/drivers/cpufreq/acpi-cpufreq.c (revision 2fe3c78a2c26dd5ee811024a1b7d6cfb4d654319)
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 
38 MODULE_AUTHOR("Paul Diefenbaugh, Dominik Brodowski");
39 MODULE_DESCRIPTION("ACPI Processor P-States Driver");
40 MODULE_LICENSE("GPL");
41 
42 enum {
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 
53 struct 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. */
63 static struct acpi_processor_performance __percpu *acpi_perf_data;
64 
65 static 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 
70 static struct cpufreq_driver acpi_cpufreq_driver;
71 
72 static unsigned int acpi_pstate_strict;
73 
74 static bool boost_state(unsigned int cpu)
75 {
76 	u32 lo, hi;
77 	u64 msr;
78 
79 	switch (boot_cpu_data.x86_vendor) {
80 	case X86_VENDOR_INTEL:
81 	case X86_VENDOR_CENTAUR:
82 	case X86_VENDOR_ZHAOXIN:
83 		rdmsr_on_cpu(cpu, MSR_IA32_MISC_ENABLE, &lo, &hi);
84 		msr = lo | ((u64)hi << 32);
85 		return !(msr & MSR_IA32_MISC_ENABLE_TURBO_DISABLE);
86 	case X86_VENDOR_HYGON:
87 	case X86_VENDOR_AMD:
88 		rdmsr_on_cpu(cpu, MSR_K7_HWCR, &lo, &hi);
89 		msr = lo | ((u64)hi << 32);
90 		return !(msr & MSR_K7_HWCR_CPB_DIS);
91 	}
92 	return false;
93 }
94 
95 static int boost_set_msr(bool enable)
96 {
97 	u32 msr_addr;
98 	u64 msr_mask, val;
99 
100 	switch (boot_cpu_data.x86_vendor) {
101 	case X86_VENDOR_INTEL:
102 	case X86_VENDOR_CENTAUR:
103 	case X86_VENDOR_ZHAOXIN:
104 		msr_addr = MSR_IA32_MISC_ENABLE;
105 		msr_mask = MSR_IA32_MISC_ENABLE_TURBO_DISABLE;
106 		break;
107 	case X86_VENDOR_HYGON:
108 	case X86_VENDOR_AMD:
109 		msr_addr = MSR_K7_HWCR;
110 		msr_mask = MSR_K7_HWCR_CPB_DIS;
111 		break;
112 	default:
113 		return -EINVAL;
114 	}
115 
116 	rdmsrl(msr_addr, val);
117 
118 	if (enable)
119 		val &= ~msr_mask;
120 	else
121 		val |= msr_mask;
122 
123 	wrmsrl(msr_addr, val);
124 	return 0;
125 }
126 
127 static void boost_set_msr_each(void *p_en)
128 {
129 	bool enable = (bool) p_en;
130 
131 	boost_set_msr(enable);
132 }
133 
134 static int set_boost(struct cpufreq_policy *policy, int val)
135 {
136 	on_each_cpu_mask(policy->cpus, boost_set_msr_each,
137 			 (void *)(long)val, 1);
138 	pr_debug("CPU %*pbl: Core Boosting %s.\n",
139 		 cpumask_pr_args(policy->cpus), str_enabled_disabled(val));
140 
141 	return 0;
142 }
143 
144 static ssize_t show_freqdomain_cpus(struct cpufreq_policy *policy, char *buf)
145 {
146 	struct acpi_cpufreq_data *data = policy->driver_data;
147 
148 	if (unlikely(!data))
149 		return -ENODEV;
150 
151 	return cpufreq_show_cpus(data->freqdomain_cpus, buf);
152 }
153 
154 cpufreq_freq_attr_ro(freqdomain_cpus);
155 
156 #ifdef CONFIG_X86_ACPI_CPUFREQ_CPB
157 static ssize_t store_cpb(struct cpufreq_policy *policy, const char *buf,
158 			 size_t count)
159 {
160 	int ret;
161 	unsigned int val = 0;
162 
163 	if (!acpi_cpufreq_driver.set_boost)
164 		return -EINVAL;
165 
166 	ret = kstrtouint(buf, 10, &val);
167 	if (ret || val > 1)
168 		return -EINVAL;
169 
170 	cpus_read_lock();
171 	set_boost(policy, val);
172 	cpus_read_unlock();
173 
174 	return count;
175 }
176 
177 static ssize_t show_cpb(struct cpufreq_policy *policy, char *buf)
178 {
179 	return sprintf(buf, "%u\n", acpi_cpufreq_driver.boost_enabled);
180 }
181 
182 cpufreq_freq_attr_rw(cpb);
183 #endif
184 
185 static int check_est_cpu(unsigned int cpuid)
186 {
187 	struct cpuinfo_x86 *cpu = &cpu_data(cpuid);
188 
189 	return cpu_has(cpu, X86_FEATURE_EST);
190 }
191 
192 static int check_amd_hwpstate_cpu(unsigned int cpuid)
193 {
194 	struct cpuinfo_x86 *cpu = &cpu_data(cpuid);
195 
196 	return cpu_has(cpu, X86_FEATURE_HW_PSTATE);
197 }
198 
199 static unsigned extract_io(struct cpufreq_policy *policy, u32 value)
200 {
201 	struct acpi_cpufreq_data *data = policy->driver_data;
202 	struct acpi_processor_performance *perf;
203 	int i;
204 
205 	perf = to_perf_data(data);
206 
207 	for (i = 0; i < perf->state_count; i++) {
208 		if (value == perf->states[i].status)
209 			return policy->freq_table[i].frequency;
210 	}
211 	return 0;
212 }
213 
214 static unsigned extract_msr(struct cpufreq_policy *policy, u32 msr)
215 {
216 	struct acpi_cpufreq_data *data = policy->driver_data;
217 	struct cpufreq_frequency_table *pos;
218 	struct acpi_processor_performance *perf;
219 
220 	if (boot_cpu_data.x86_vendor == X86_VENDOR_AMD)
221 		msr &= AMD_MSR_RANGE;
222 	else if (boot_cpu_data.x86_vendor == X86_VENDOR_HYGON)
223 		msr &= HYGON_MSR_RANGE;
224 	else
225 		msr &= INTEL_MSR_RANGE;
226 
227 	perf = to_perf_data(data);
228 
229 	cpufreq_for_each_entry(pos, policy->freq_table)
230 		if (msr == perf->states[pos->driver_data].status)
231 			return pos->frequency;
232 	return policy->freq_table[0].frequency;
233 }
234 
235 static unsigned extract_freq(struct cpufreq_policy *policy, u32 val)
236 {
237 	struct acpi_cpufreq_data *data = policy->driver_data;
238 
239 	switch (data->cpu_feature) {
240 	case SYSTEM_INTEL_MSR_CAPABLE:
241 	case SYSTEM_AMD_MSR_CAPABLE:
242 		return extract_msr(policy, val);
243 	case SYSTEM_IO_CAPABLE:
244 		return extract_io(policy, val);
245 	default:
246 		return 0;
247 	}
248 }
249 
250 static u32 cpu_freq_read_intel(struct acpi_pct_register *not_used)
251 {
252 	u32 val, dummy __always_unused;
253 
254 	rdmsr(MSR_IA32_PERF_CTL, val, dummy);
255 	return val;
256 }
257 
258 static void cpu_freq_write_intel(struct acpi_pct_register *not_used, u32 val)
259 {
260 	u32 lo, hi;
261 
262 	rdmsr(MSR_IA32_PERF_CTL, lo, hi);
263 	lo = (lo & ~INTEL_MSR_RANGE) | (val & INTEL_MSR_RANGE);
264 	wrmsr(MSR_IA32_PERF_CTL, lo, hi);
265 }
266 
267 static u32 cpu_freq_read_amd(struct acpi_pct_register *not_used)
268 {
269 	u32 val, dummy __always_unused;
270 
271 	rdmsr(MSR_AMD_PERF_CTL, val, dummy);
272 	return val;
273 }
274 
275 static void cpu_freq_write_amd(struct acpi_pct_register *not_used, u32 val)
276 {
277 	wrmsr(MSR_AMD_PERF_CTL, val, 0);
278 }
279 
280 static u32 cpu_freq_read_io(struct acpi_pct_register *reg)
281 {
282 	u32 val;
283 
284 	acpi_os_read_port(reg->address, &val, reg->bit_width);
285 	return val;
286 }
287 
288 static void cpu_freq_write_io(struct acpi_pct_register *reg, u32 val)
289 {
290 	acpi_os_write_port(reg->address, val, reg->bit_width);
291 }
292 
293 struct drv_cmd {
294 	struct acpi_pct_register *reg;
295 	u32 val;
296 	union {
297 		void (*write)(struct acpi_pct_register *reg, u32 val);
298 		u32 (*read)(struct acpi_pct_register *reg);
299 	} func;
300 };
301 
302 /* Called via smp_call_function_single(), on the target CPU */
303 static void do_drv_read(void *_cmd)
304 {
305 	struct drv_cmd *cmd = _cmd;
306 
307 	cmd->val = cmd->func.read(cmd->reg);
308 }
309 
310 static u32 drv_read(struct acpi_cpufreq_data *data, const struct cpumask *mask)
311 {
312 	struct acpi_processor_performance *perf = to_perf_data(data);
313 	struct drv_cmd cmd = {
314 		.reg = &perf->control_register,
315 		.func.read = data->cpu_freq_read,
316 	};
317 	int err;
318 
319 	err = smp_call_function_any(mask, do_drv_read, &cmd, 1);
320 	WARN_ON_ONCE(err);	/* smp_call_function_any() was buggy? */
321 	return cmd.val;
322 }
323 
324 /* Called via smp_call_function_many(), on the target CPUs */
325 static void do_drv_write(void *_cmd)
326 {
327 	struct drv_cmd *cmd = _cmd;
328 
329 	cmd->func.write(cmd->reg, cmd->val);
330 }
331 
332 static void drv_write(struct acpi_cpufreq_data *data,
333 		      const struct cpumask *mask, u32 val)
334 {
335 	struct acpi_processor_performance *perf = to_perf_data(data);
336 	struct drv_cmd cmd = {
337 		.reg = &perf->control_register,
338 		.val = val,
339 		.func.write = data->cpu_freq_write,
340 	};
341 	int this_cpu;
342 
343 	this_cpu = get_cpu();
344 	if (cpumask_test_cpu(this_cpu, mask))
345 		do_drv_write(&cmd);
346 
347 	smp_call_function_many(mask, do_drv_write, &cmd, 1);
348 	put_cpu();
349 }
350 
351 static u32 get_cur_val(const struct cpumask *mask, struct acpi_cpufreq_data *data)
352 {
353 	u32 val;
354 
355 	if (unlikely(cpumask_empty(mask)))
356 		return 0;
357 
358 	val = drv_read(data, mask);
359 
360 	pr_debug("%s = %u\n", __func__, val);
361 
362 	return val;
363 }
364 
365 static unsigned int get_cur_freq_on_cpu(unsigned int cpu)
366 {
367 	struct acpi_cpufreq_data *data;
368 	struct cpufreq_policy *policy;
369 	unsigned int freq;
370 	unsigned int cached_freq;
371 
372 	pr_debug("%s (%d)\n", __func__, cpu);
373 
374 	policy = cpufreq_cpu_get_raw(cpu);
375 	if (unlikely(!policy))
376 		return 0;
377 
378 	data = policy->driver_data;
379 	if (unlikely(!data || !policy->freq_table))
380 		return 0;
381 
382 	cached_freq = policy->freq_table[to_perf_data(data)->state].frequency;
383 	freq = extract_freq(policy, get_cur_val(cpumask_of(cpu), data));
384 	if (freq != cached_freq) {
385 		/*
386 		 * The dreaded BIOS frequency change behind our back.
387 		 * Force set the frequency on next target call.
388 		 */
389 		data->resume = 1;
390 	}
391 
392 	pr_debug("cur freq = %u\n", freq);
393 
394 	return freq;
395 }
396 
397 static unsigned int check_freqs(struct cpufreq_policy *policy,
398 				const struct cpumask *mask, unsigned int freq)
399 {
400 	struct acpi_cpufreq_data *data = policy->driver_data;
401 	unsigned int cur_freq;
402 	unsigned int i;
403 
404 	for (i = 0; i < 100; i++) {
405 		cur_freq = extract_freq(policy, get_cur_val(mask, data));
406 		if (cur_freq == freq)
407 			return 1;
408 		udelay(10);
409 	}
410 	return 0;
411 }
412 
413 static int acpi_cpufreq_target(struct cpufreq_policy *policy,
414 			       unsigned int index)
415 {
416 	struct acpi_cpufreq_data *data = policy->driver_data;
417 	struct acpi_processor_performance *perf;
418 	const struct cpumask *mask;
419 	unsigned int next_perf_state = 0; /* Index into perf table */
420 	int result = 0;
421 
422 	if (unlikely(!data)) {
423 		return -ENODEV;
424 	}
425 
426 	perf = to_perf_data(data);
427 	next_perf_state = policy->freq_table[index].driver_data;
428 	if (perf->state == next_perf_state) {
429 		if (unlikely(data->resume)) {
430 			pr_debug("Called after resume, resetting to P%d\n",
431 				next_perf_state);
432 			data->resume = 0;
433 		} else {
434 			pr_debug("Already at target state (P%d)\n",
435 				next_perf_state);
436 			return 0;
437 		}
438 	}
439 
440 	/*
441 	 * The core won't allow CPUs to go away until the governor has been
442 	 * stopped, so we can rely on the stability of policy->cpus.
443 	 */
444 	mask = policy->shared_type == CPUFREQ_SHARED_TYPE_ANY ?
445 		cpumask_of(policy->cpu) : policy->cpus;
446 
447 	drv_write(data, mask, perf->states[next_perf_state].control);
448 
449 	if (acpi_pstate_strict) {
450 		if (!check_freqs(policy, mask,
451 				 policy->freq_table[index].frequency)) {
452 			pr_debug("%s (%d)\n", __func__, policy->cpu);
453 			result = -EAGAIN;
454 		}
455 	}
456 
457 	if (!result)
458 		perf->state = next_perf_state;
459 
460 	return result;
461 }
462 
463 static unsigned int acpi_cpufreq_fast_switch(struct cpufreq_policy *policy,
464 					     unsigned int target_freq)
465 {
466 	struct acpi_cpufreq_data *data = policy->driver_data;
467 	struct acpi_processor_performance *perf;
468 	struct cpufreq_frequency_table *entry;
469 	unsigned int next_perf_state, next_freq, index;
470 
471 	/*
472 	 * Find the closest frequency above target_freq.
473 	 */
474 	if (policy->cached_target_freq == target_freq)
475 		index = policy->cached_resolved_idx;
476 	else
477 		index = cpufreq_table_find_index_dl(policy, target_freq,
478 						    false);
479 
480 	entry = &policy->freq_table[index];
481 	next_freq = entry->frequency;
482 	next_perf_state = entry->driver_data;
483 
484 	perf = to_perf_data(data);
485 	if (perf->state == next_perf_state) {
486 		if (unlikely(data->resume))
487 			data->resume = 0;
488 		else
489 			return next_freq;
490 	}
491 
492 	data->cpu_freq_write(&perf->control_register,
493 			     perf->states[next_perf_state].control);
494 	perf->state = next_perf_state;
495 	return next_freq;
496 }
497 
498 static unsigned long
499 acpi_cpufreq_guess_freq(struct acpi_cpufreq_data *data, unsigned int cpu)
500 {
501 	struct acpi_processor_performance *perf;
502 
503 	perf = to_perf_data(data);
504 	if (cpu_khz) {
505 		/* search the closest match to cpu_khz */
506 		unsigned int i;
507 		unsigned long freq;
508 		unsigned long freqn = perf->states[0].core_frequency * 1000;
509 
510 		for (i = 0; i < (perf->state_count-1); i++) {
511 			freq = freqn;
512 			freqn = perf->states[i+1].core_frequency * 1000;
513 			if ((2 * cpu_khz) > (freqn + freq)) {
514 				perf->state = i;
515 				return freq;
516 			}
517 		}
518 		perf->state = perf->state_count-1;
519 		return freqn;
520 	} else {
521 		/* assume CPU is at P0... */
522 		perf->state = 0;
523 		return perf->states[0].core_frequency * 1000;
524 	}
525 }
526 
527 static void free_acpi_perf_data(void)
528 {
529 	unsigned int i;
530 
531 	/* Freeing a NULL pointer is OK, and alloc_percpu zeroes. */
532 	for_each_possible_cpu(i)
533 		free_cpumask_var(per_cpu_ptr(acpi_perf_data, i)
534 				 ->shared_cpu_map);
535 	free_percpu(acpi_perf_data);
536 }
537 
538 static int cpufreq_boost_down_prep(unsigned int cpu)
539 {
540 	/*
541 	 * Clear the boost-disable bit on the CPU_DOWN path so that
542 	 * this cpu cannot block the remaining ones from boosting.
543 	 */
544 	return boost_set_msr(1);
545 }
546 
547 /*
548  * acpi_cpufreq_early_init - initialize ACPI P-States library
549  *
550  * Initialize the ACPI P-States library (drivers/acpi/processor_perflib.c)
551  * in order to determine correct frequency and voltage pairings. We can
552  * do _PDC and _PSD and find out the processor dependency for the
553  * actual init that will happen later...
554  */
555 static int __init acpi_cpufreq_early_init(void)
556 {
557 	unsigned int i;
558 	pr_debug("%s\n", __func__);
559 
560 	acpi_perf_data = alloc_percpu(struct acpi_processor_performance);
561 	if (!acpi_perf_data) {
562 		pr_debug("Memory allocation error for acpi_perf_data.\n");
563 		return -ENOMEM;
564 	}
565 	for_each_possible_cpu(i) {
566 		if (!zalloc_cpumask_var_node(
567 			&per_cpu_ptr(acpi_perf_data, i)->shared_cpu_map,
568 			GFP_KERNEL, cpu_to_node(i))) {
569 
570 			/* Freeing a NULL pointer is OK: alloc_percpu zeroes. */
571 			free_acpi_perf_data();
572 			return -ENOMEM;
573 		}
574 	}
575 
576 	/* Do initialization in ACPI core */
577 	acpi_processor_preregister_performance(acpi_perf_data);
578 	return 0;
579 }
580 
581 #ifdef CONFIG_SMP
582 /*
583  * Some BIOSes do SW_ANY coordination internally, either set it up in hw
584  * or do it in BIOS firmware and won't inform about it to OS. If not
585  * detected, this has a side effect of making CPU run at a different speed
586  * than OS intended it to run at. Detect it and handle it cleanly.
587  */
588 static int bios_with_sw_any_bug;
589 
590 static int sw_any_bug_found(const struct dmi_system_id *d)
591 {
592 	bios_with_sw_any_bug = 1;
593 	return 0;
594 }
595 
596 static const struct dmi_system_id sw_any_bug_dmi_table[] = {
597 	{
598 		.callback = sw_any_bug_found,
599 		.ident = "Supermicro Server X6DLP",
600 		.matches = {
601 			DMI_MATCH(DMI_SYS_VENDOR, "Supermicro"),
602 			DMI_MATCH(DMI_BIOS_VERSION, "080010"),
603 			DMI_MATCH(DMI_PRODUCT_NAME, "X6DLP"),
604 		},
605 	},
606 	{ }
607 };
608 
609 static int acpi_cpufreq_blacklist(struct cpuinfo_x86 *c)
610 {
611 	/* Intel Xeon Processor 7100 Series Specification Update
612 	 * https://www.intel.com/Assets/PDF/specupdate/314554.pdf
613 	 * AL30: A Machine Check Exception (MCE) Occurring during an
614 	 * Enhanced Intel SpeedStep Technology Ratio Change May Cause
615 	 * Both Processor Cores to Lock Up. */
616 	if (c->x86_vendor == X86_VENDOR_INTEL) {
617 		if ((c->x86 == 15) &&
618 		    (c->x86_model == 6) &&
619 		    (c->x86_stepping == 8)) {
620 			pr_info("Intel(R) Xeon(R) 7100 Errata AL30, processors may lock up on frequency changes: disabling acpi-cpufreq\n");
621 			return -ENODEV;
622 		    }
623 		}
624 	return 0;
625 }
626 #endif
627 
628 #ifdef CONFIG_ACPI_CPPC_LIB
629 static u64 get_max_boost_ratio(unsigned int cpu)
630 {
631 	struct cppc_perf_caps perf_caps;
632 	u64 highest_perf, nominal_perf;
633 	int ret;
634 
635 	if (acpi_pstate_strict)
636 		return 0;
637 
638 	ret = cppc_get_perf_caps(cpu, &perf_caps);
639 	if (ret) {
640 		pr_debug("CPU%d: Unable to get performance capabilities (%d)\n",
641 			 cpu, ret);
642 		return 0;
643 	}
644 
645 	if (boot_cpu_data.x86_vendor == X86_VENDOR_AMD) {
646 		ret = amd_get_boost_ratio_numerator(cpu, &highest_perf);
647 		if (ret) {
648 			pr_debug("CPU%d: Unable to get boost ratio numerator (%d)\n",
649 				 cpu, ret);
650 			return 0;
651 		}
652 	} else {
653 		highest_perf = perf_caps.highest_perf;
654 	}
655 
656 	nominal_perf = perf_caps.nominal_perf;
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 #else
671 static inline u64 get_max_boost_ratio(unsigned int cpu) { return 0; }
672 #endif
673 
674 static int acpi_cpufreq_cpu_init(struct cpufreq_policy *policy)
675 {
676 	struct cpufreq_frequency_table *freq_table;
677 	struct acpi_processor_performance *perf;
678 	struct acpi_cpufreq_data *data;
679 	unsigned int cpu = policy->cpu;
680 	struct cpuinfo_x86 *c = &cpu_data(cpu);
681 	unsigned int valid_states = 0;
682 	unsigned int result = 0;
683 	u64 max_boost_ratio;
684 	unsigned int i;
685 #ifdef CONFIG_SMP
686 	static int blacklisted;
687 #endif
688 
689 	pr_debug("%s\n", __func__);
690 
691 #ifdef CONFIG_SMP
692 	if (blacklisted)
693 		return blacklisted;
694 	blacklisted = acpi_cpufreq_blacklist(c);
695 	if (blacklisted)
696 		return blacklisted;
697 #endif
698 
699 	data = kzalloc(sizeof(*data), GFP_KERNEL);
700 	if (!data)
701 		return -ENOMEM;
702 
703 	if (!zalloc_cpumask_var(&data->freqdomain_cpus, GFP_KERNEL)) {
704 		result = -ENOMEM;
705 		goto err_free;
706 	}
707 
708 	perf = per_cpu_ptr(acpi_perf_data, cpu);
709 	data->acpi_perf_cpu = cpu;
710 	policy->driver_data = data;
711 
712 	if (cpu_has(c, X86_FEATURE_CONSTANT_TSC))
713 		acpi_cpufreq_driver.flags |= CPUFREQ_CONST_LOOPS;
714 
715 	result = acpi_processor_register_performance(perf, cpu);
716 	if (result)
717 		goto err_free_mask;
718 
719 	policy->shared_type = perf->shared_type;
720 
721 	/*
722 	 * Will let policy->cpus know about dependency only when software
723 	 * coordination is required.
724 	 */
725 	if (policy->shared_type == CPUFREQ_SHARED_TYPE_ALL ||
726 	    policy->shared_type == CPUFREQ_SHARED_TYPE_ANY) {
727 		cpumask_copy(policy->cpus, perf->shared_cpu_map);
728 	}
729 	cpumask_copy(data->freqdomain_cpus, perf->shared_cpu_map);
730 
731 #ifdef CONFIG_SMP
732 	dmi_check_system(sw_any_bug_dmi_table);
733 	if (bios_with_sw_any_bug && !policy_is_shared(policy)) {
734 		policy->shared_type = CPUFREQ_SHARED_TYPE_ALL;
735 		cpumask_copy(policy->cpus, topology_core_cpumask(cpu));
736 	}
737 
738 	if (check_amd_hwpstate_cpu(cpu) && boot_cpu_data.x86 < 0x19 &&
739 	    !acpi_pstate_strict) {
740 		cpumask_clear(policy->cpus);
741 		cpumask_set_cpu(cpu, policy->cpus);
742 		cpumask_copy(data->freqdomain_cpus,
743 			     topology_sibling_cpumask(cpu));
744 		policy->shared_type = CPUFREQ_SHARED_TYPE_HW;
745 		pr_info_once("overriding BIOS provided _PSD data\n");
746 	}
747 #endif
748 
749 	/* capability check */
750 	if (perf->state_count <= 1) {
751 		pr_debug("No P-States\n");
752 		result = -ENODEV;
753 		goto err_unreg;
754 	}
755 
756 	if (perf->control_register.space_id != perf->status_register.space_id) {
757 		result = -ENODEV;
758 		goto err_unreg;
759 	}
760 
761 	switch (perf->control_register.space_id) {
762 	case ACPI_ADR_SPACE_SYSTEM_IO:
763 		if (boot_cpu_data.x86_vendor == X86_VENDOR_AMD &&
764 		    boot_cpu_data.x86 == 0xf) {
765 			pr_debug("AMD K8 systems must use native drivers.\n");
766 			result = -ENODEV;
767 			goto err_unreg;
768 		}
769 		pr_debug("SYSTEM IO addr space\n");
770 		data->cpu_feature = SYSTEM_IO_CAPABLE;
771 		data->cpu_freq_read = cpu_freq_read_io;
772 		data->cpu_freq_write = cpu_freq_write_io;
773 		break;
774 	case ACPI_ADR_SPACE_FIXED_HARDWARE:
775 		pr_debug("HARDWARE addr space\n");
776 		if (check_est_cpu(cpu)) {
777 			data->cpu_feature = SYSTEM_INTEL_MSR_CAPABLE;
778 			data->cpu_freq_read = cpu_freq_read_intel;
779 			data->cpu_freq_write = cpu_freq_write_intel;
780 			break;
781 		}
782 		if (check_amd_hwpstate_cpu(cpu)) {
783 			data->cpu_feature = SYSTEM_AMD_MSR_CAPABLE;
784 			data->cpu_freq_read = cpu_freq_read_amd;
785 			data->cpu_freq_write = cpu_freq_write_amd;
786 			break;
787 		}
788 		result = -ENODEV;
789 		goto err_unreg;
790 	default:
791 		pr_debug("Unknown addr space %d\n",
792 			(u32) (perf->control_register.space_id));
793 		result = -ENODEV;
794 		goto err_unreg;
795 	}
796 
797 	freq_table = kcalloc(perf->state_count + 1, sizeof(*freq_table),
798 			     GFP_KERNEL);
799 	if (!freq_table) {
800 		result = -ENOMEM;
801 		goto err_unreg;
802 	}
803 
804 	/* detect transition latency */
805 	policy->cpuinfo.transition_latency = 0;
806 	for (i = 0; i < perf->state_count; i++) {
807 		if ((perf->states[i].transition_latency * 1000) >
808 		    policy->cpuinfo.transition_latency)
809 			policy->cpuinfo.transition_latency =
810 			    perf->states[i].transition_latency * 1000;
811 	}
812 
813 	/* Check for high latency (>20uS) from buggy BIOSes, like on T42 */
814 	if (perf->control_register.space_id == ACPI_ADR_SPACE_FIXED_HARDWARE &&
815 	    policy->cpuinfo.transition_latency > 20 * 1000) {
816 		policy->cpuinfo.transition_latency = 20 * 1000;
817 		pr_info_once("P-state transition latency capped at 20 uS\n");
818 	}
819 
820 	/* table init */
821 	for (i = 0; i < perf->state_count; i++) {
822 		if (i > 0 && perf->states[i].core_frequency >=
823 		    freq_table[valid_states-1].frequency / 1000)
824 			continue;
825 
826 		freq_table[valid_states].driver_data = i;
827 		freq_table[valid_states].frequency =
828 		    perf->states[i].core_frequency * 1000;
829 		valid_states++;
830 	}
831 	freq_table[valid_states].frequency = CPUFREQ_TABLE_END;
832 
833 	max_boost_ratio = get_max_boost_ratio(cpu);
834 	if (max_boost_ratio) {
835 		unsigned int freq = freq_table[0].frequency;
836 
837 		/*
838 		 * Because the loop above sorts the freq_table entries in the
839 		 * descending order, freq is the maximum frequency in the table.
840 		 * Assume that it corresponds to the CPPC nominal frequency and
841 		 * use it to set cpuinfo.max_freq.
842 		 */
843 		policy->cpuinfo.max_freq = freq * max_boost_ratio >> SCHED_CAPACITY_SHIFT;
844 	} else {
845 		/*
846 		 * If the maximum "boost" frequency is unknown, ask the arch
847 		 * scale-invariance code to use the "nominal" performance for
848 		 * CPU utilization scaling so as to prevent the schedutil
849 		 * governor from selecting inadequate CPU frequencies.
850 		 */
851 		arch_set_max_freq_ratio(true);
852 	}
853 
854 	policy->freq_table = freq_table;
855 	perf->state = 0;
856 
857 	switch (perf->control_register.space_id) {
858 	case ACPI_ADR_SPACE_SYSTEM_IO:
859 		/*
860 		 * The core will not set policy->cur, because
861 		 * cpufreq_driver->get is NULL, so we need to set it here.
862 		 * However, we have to guess it, because the current speed is
863 		 * unknown and not detectable via IO ports.
864 		 */
865 		policy->cur = acpi_cpufreq_guess_freq(data, policy->cpu);
866 		break;
867 	case ACPI_ADR_SPACE_FIXED_HARDWARE:
868 		acpi_cpufreq_driver.get = get_cur_freq_on_cpu;
869 		break;
870 	default:
871 		break;
872 	}
873 
874 	/* notify BIOS that we exist */
875 	acpi_processor_notify_smm(THIS_MODULE);
876 
877 	pr_debug("CPU%u - ACPI performance management activated.\n", cpu);
878 	for (i = 0; i < perf->state_count; i++)
879 		pr_debug("     %cP%d: %d MHz, %d mW, %d uS\n",
880 			(i == perf->state ? '*' : ' '), i,
881 			(u32) perf->states[i].core_frequency,
882 			(u32) perf->states[i].power,
883 			(u32) perf->states[i].transition_latency);
884 
885 	/*
886 	 * the first call to ->target() should result in us actually
887 	 * writing something to the appropriate registers.
888 	 */
889 	data->resume = 1;
890 
891 	policy->fast_switch_possible = !acpi_pstate_strict &&
892 		!(policy_is_shared(policy) && policy->shared_type != CPUFREQ_SHARED_TYPE_ANY);
893 
894 	if (perf->states[0].core_frequency * 1000 != freq_table[0].frequency)
895 		pr_warn(FW_WARN "P-state 0 is not max freq\n");
896 
897 	if (acpi_cpufreq_driver.set_boost) {
898 		set_boost(policy, acpi_cpufreq_driver.boost_enabled);
899 		policy->boost_enabled = acpi_cpufreq_driver.boost_enabled;
900 	}
901 
902 	return result;
903 
904 err_unreg:
905 	acpi_processor_unregister_performance(cpu);
906 err_free_mask:
907 	free_cpumask_var(data->freqdomain_cpus);
908 err_free:
909 	kfree(data);
910 	policy->driver_data = NULL;
911 
912 	return result;
913 }
914 
915 static void acpi_cpufreq_cpu_exit(struct cpufreq_policy *policy)
916 {
917 	struct acpi_cpufreq_data *data = policy->driver_data;
918 
919 	pr_debug("%s\n", __func__);
920 
921 	cpufreq_boost_down_prep(policy->cpu);
922 	policy->fast_switch_possible = false;
923 	policy->driver_data = NULL;
924 	acpi_processor_unregister_performance(data->acpi_perf_cpu);
925 	free_cpumask_var(data->freqdomain_cpus);
926 	kfree(policy->freq_table);
927 	kfree(data);
928 }
929 
930 static int acpi_cpufreq_resume(struct cpufreq_policy *policy)
931 {
932 	struct acpi_cpufreq_data *data = policy->driver_data;
933 
934 	pr_debug("%s\n", __func__);
935 
936 	data->resume = 1;
937 
938 	return 0;
939 }
940 
941 static struct freq_attr *acpi_cpufreq_attr[] = {
942 	&cpufreq_freq_attr_scaling_available_freqs,
943 	&freqdomain_cpus,
944 #ifdef CONFIG_X86_ACPI_CPUFREQ_CPB
945 	&cpb,
946 #endif
947 	NULL,
948 };
949 
950 static struct cpufreq_driver acpi_cpufreq_driver = {
951 	.verify		= cpufreq_generic_frequency_table_verify,
952 	.target_index	= acpi_cpufreq_target,
953 	.fast_switch	= acpi_cpufreq_fast_switch,
954 	.bios_limit	= acpi_processor_get_bios_limit,
955 	.init		= acpi_cpufreq_cpu_init,
956 	.exit		= acpi_cpufreq_cpu_exit,
957 	.resume		= acpi_cpufreq_resume,
958 	.name		= "acpi-cpufreq",
959 	.attr		= acpi_cpufreq_attr,
960 };
961 
962 static void __init acpi_cpufreq_boost_init(void)
963 {
964 	if (!(boot_cpu_has(X86_FEATURE_CPB) || boot_cpu_has(X86_FEATURE_IDA))) {
965 		pr_debug("Boost capabilities not present in the processor\n");
966 		return;
967 	}
968 
969 	acpi_cpufreq_driver.set_boost = set_boost;
970 	acpi_cpufreq_driver.boost_enabled = boost_state(0);
971 }
972 
973 static int __init acpi_cpufreq_probe(struct platform_device *pdev)
974 {
975 	int ret;
976 
977 	if (acpi_disabled)
978 		return -ENODEV;
979 
980 	/* don't keep reloading if cpufreq_driver exists */
981 	if (cpufreq_get_current_driver())
982 		return -ENODEV;
983 
984 	pr_debug("%s\n", __func__);
985 
986 	ret = acpi_cpufreq_early_init();
987 	if (ret)
988 		return ret;
989 
990 #ifdef CONFIG_X86_ACPI_CPUFREQ_CPB
991 	/* this is a sysfs file with a strange name and an even stranger
992 	 * semantic - per CPU instantiation, but system global effect.
993 	 * Lets enable it only on AMD CPUs for compatibility reasons and
994 	 * only if configured. This is considered legacy code, which
995 	 * will probably be removed at some point in the future.
996 	 */
997 	if (!check_amd_hwpstate_cpu(0)) {
998 		struct freq_attr **attr;
999 
1000 		pr_debug("CPB unsupported, do not expose it\n");
1001 
1002 		for (attr = acpi_cpufreq_attr; *attr; attr++)
1003 			if (*attr == &cpb) {
1004 				*attr = NULL;
1005 				break;
1006 			}
1007 	}
1008 #endif
1009 	acpi_cpufreq_boost_init();
1010 
1011 	ret = cpufreq_register_driver(&acpi_cpufreq_driver);
1012 	if (ret) {
1013 		free_acpi_perf_data();
1014 	}
1015 	return ret;
1016 }
1017 
1018 static void acpi_cpufreq_remove(struct platform_device *pdev)
1019 {
1020 	pr_debug("%s\n", __func__);
1021 
1022 	cpufreq_unregister_driver(&acpi_cpufreq_driver);
1023 
1024 	free_acpi_perf_data();
1025 }
1026 
1027 static struct platform_driver acpi_cpufreq_platdrv = {
1028 	.driver = {
1029 		.name	= "acpi-cpufreq",
1030 	},
1031 	.remove_new	= acpi_cpufreq_remove,
1032 };
1033 
1034 static int __init acpi_cpufreq_init(void)
1035 {
1036 	return platform_driver_probe(&acpi_cpufreq_platdrv, acpi_cpufreq_probe);
1037 }
1038 
1039 static void __exit acpi_cpufreq_exit(void)
1040 {
1041 	platform_driver_unregister(&acpi_cpufreq_platdrv);
1042 }
1043 
1044 module_param(acpi_pstate_strict, uint, 0644);
1045 MODULE_PARM_DESC(acpi_pstate_strict,
1046 	"value 0 or non-zero. non-zero -> strict ACPI checks are "
1047 	"performed during frequency changes.");
1048 
1049 late_initcall(acpi_cpufreq_init);
1050 module_exit(acpi_cpufreq_exit);
1051 
1052 MODULE_ALIAS("platform:acpi-cpufreq");
1053