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 
23 #include <linux/acpi.h>
24 #include <linux/io.h>
25 #include <linux/delay.h>
26 #include <linux/uaccess.h>
27 
28 #include <acpi/processor.h>
29 #include <acpi/cppc_acpi.h>
30 
31 #include <asm/msr.h>
32 #include <asm/processor.h>
33 #include <asm/cpufeature.h>
34 #include <asm/cpu_device_id.h>
35 
36 MODULE_AUTHOR("Paul Diefenbaugh, Dominik Brodowski");
37 MODULE_DESCRIPTION("ACPI Processor P-States Driver");
38 MODULE_LICENSE("GPL");
39 
40 enum {
41 	UNDEFINED_CAPABLE = 0,
42 	SYSTEM_INTEL_MSR_CAPABLE,
43 	SYSTEM_AMD_MSR_CAPABLE,
44 	SYSTEM_IO_CAPABLE,
45 };
46 
47 #define INTEL_MSR_RANGE		(0xffff)
48 #define AMD_MSR_RANGE		(0x7)
49 #define HYGON_MSR_RANGE		(0x7)
50 
51 #define MSR_K7_HWCR_CPB_DIS	(1ULL << 25)
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 		rdmsr_on_cpu(cpu, MSR_IA32_MISC_ENABLE, &lo, &hi);
82 		msr = lo | ((u64)hi << 32);
83 		return !(msr & MSR_IA32_MISC_ENABLE_TURBO_DISABLE);
84 	case X86_VENDOR_HYGON:
85 	case X86_VENDOR_AMD:
86 		rdmsr_on_cpu(cpu, MSR_K7_HWCR, &lo, &hi);
87 		msr = lo | ((u64)hi << 32);
88 		return !(msr & MSR_K7_HWCR_CPB_DIS);
89 	}
90 	return false;
91 }
92 
93 static int boost_set_msr(bool enable)
94 {
95 	u32 msr_addr;
96 	u64 msr_mask, val;
97 
98 	switch (boot_cpu_data.x86_vendor) {
99 	case X86_VENDOR_INTEL:
100 		msr_addr = MSR_IA32_MISC_ENABLE;
101 		msr_mask = MSR_IA32_MISC_ENABLE_TURBO_DISABLE;
102 		break;
103 	case X86_VENDOR_HYGON:
104 	case X86_VENDOR_AMD:
105 		msr_addr = MSR_K7_HWCR;
106 		msr_mask = MSR_K7_HWCR_CPB_DIS;
107 		break;
108 	default:
109 		return -EINVAL;
110 	}
111 
112 	rdmsrl(msr_addr, val);
113 
114 	if (enable)
115 		val &= ~msr_mask;
116 	else
117 		val |= msr_mask;
118 
119 	wrmsrl(msr_addr, val);
120 	return 0;
121 }
122 
123 static void boost_set_msr_each(void *p_en)
124 {
125 	bool enable = (bool) p_en;
126 
127 	boost_set_msr(enable);
128 }
129 
130 static int set_boost(struct cpufreq_policy *policy, int val)
131 {
132 	on_each_cpu_mask(policy->cpus, boost_set_msr_each,
133 			 (void *)(long)val, 1);
134 	pr_debug("CPU %*pbl: Core Boosting %sabled.\n",
135 		 cpumask_pr_args(policy->cpus), val ? "en" : "dis");
136 
137 	return 0;
138 }
139 
140 static ssize_t show_freqdomain_cpus(struct cpufreq_policy *policy, char *buf)
141 {
142 	struct acpi_cpufreq_data *data = policy->driver_data;
143 
144 	if (unlikely(!data))
145 		return -ENODEV;
146 
147 	return cpufreq_show_cpus(data->freqdomain_cpus, buf);
148 }
149 
150 cpufreq_freq_attr_ro(freqdomain_cpus);
151 
152 #ifdef CONFIG_X86_ACPI_CPUFREQ_CPB
153 static ssize_t store_cpb(struct cpufreq_policy *policy, const char *buf,
154 			 size_t count)
155 {
156 	int ret;
157 	unsigned int val = 0;
158 
159 	if (!acpi_cpufreq_driver.set_boost)
160 		return -EINVAL;
161 
162 	ret = kstrtouint(buf, 10, &val);
163 	if (ret || val > 1)
164 		return -EINVAL;
165 
166 	get_online_cpus();
167 	set_boost(policy, val);
168 	put_online_cpus();
169 
170 	return count;
171 }
172 
173 static ssize_t show_cpb(struct cpufreq_policy *policy, char *buf)
174 {
175 	return sprintf(buf, "%u\n", acpi_cpufreq_driver.boost_enabled);
176 }
177 
178 cpufreq_freq_attr_rw(cpb);
179 #endif
180 
181 static int check_est_cpu(unsigned int cpuid)
182 {
183 	struct cpuinfo_x86 *cpu = &cpu_data(cpuid);
184 
185 	return cpu_has(cpu, X86_FEATURE_EST);
186 }
187 
188 static int check_amd_hwpstate_cpu(unsigned int cpuid)
189 {
190 	struct cpuinfo_x86 *cpu = &cpu_data(cpuid);
191 
192 	return cpu_has(cpu, X86_FEATURE_HW_PSTATE);
193 }
194 
195 static unsigned extract_io(struct cpufreq_policy *policy, u32 value)
196 {
197 	struct acpi_cpufreq_data *data = policy->driver_data;
198 	struct acpi_processor_performance *perf;
199 	int i;
200 
201 	perf = to_perf_data(data);
202 
203 	for (i = 0; i < perf->state_count; i++) {
204 		if (value == perf->states[i].status)
205 			return policy->freq_table[i].frequency;
206 	}
207 	return 0;
208 }
209 
210 static unsigned extract_msr(struct cpufreq_policy *policy, u32 msr)
211 {
212 	struct acpi_cpufreq_data *data = policy->driver_data;
213 	struct cpufreq_frequency_table *pos;
214 	struct acpi_processor_performance *perf;
215 
216 	if (boot_cpu_data.x86_vendor == X86_VENDOR_AMD)
217 		msr &= AMD_MSR_RANGE;
218 	else if (boot_cpu_data.x86_vendor == X86_VENDOR_HYGON)
219 		msr &= HYGON_MSR_RANGE;
220 	else
221 		msr &= INTEL_MSR_RANGE;
222 
223 	perf = to_perf_data(data);
224 
225 	cpufreq_for_each_entry(pos, policy->freq_table)
226 		if (msr == perf->states[pos->driver_data].status)
227 			return pos->frequency;
228 	return policy->freq_table[0].frequency;
229 }
230 
231 static unsigned extract_freq(struct cpufreq_policy *policy, u32 val)
232 {
233 	struct acpi_cpufreq_data *data = policy->driver_data;
234 
235 	switch (data->cpu_feature) {
236 	case SYSTEM_INTEL_MSR_CAPABLE:
237 	case SYSTEM_AMD_MSR_CAPABLE:
238 		return extract_msr(policy, val);
239 	case SYSTEM_IO_CAPABLE:
240 		return extract_io(policy, val);
241 	default:
242 		return 0;
243 	}
244 }
245 
246 static u32 cpu_freq_read_intel(struct acpi_pct_register *not_used)
247 {
248 	u32 val, dummy __always_unused;
249 
250 	rdmsr(MSR_IA32_PERF_CTL, val, dummy);
251 	return val;
252 }
253 
254 static void cpu_freq_write_intel(struct acpi_pct_register *not_used, u32 val)
255 {
256 	u32 lo, hi;
257 
258 	rdmsr(MSR_IA32_PERF_CTL, lo, hi);
259 	lo = (lo & ~INTEL_MSR_RANGE) | (val & INTEL_MSR_RANGE);
260 	wrmsr(MSR_IA32_PERF_CTL, lo, hi);
261 }
262 
263 static u32 cpu_freq_read_amd(struct acpi_pct_register *not_used)
264 {
265 	u32 val, dummy __always_unused;
266 
267 	rdmsr(MSR_AMD_PERF_CTL, val, dummy);
268 	return val;
269 }
270 
271 static void cpu_freq_write_amd(struct acpi_pct_register *not_used, u32 val)
272 {
273 	wrmsr(MSR_AMD_PERF_CTL, val, 0);
274 }
275 
276 static u32 cpu_freq_read_io(struct acpi_pct_register *reg)
277 {
278 	u32 val;
279 
280 	acpi_os_read_port(reg->address, &val, reg->bit_width);
281 	return val;
282 }
283 
284 static void cpu_freq_write_io(struct acpi_pct_register *reg, u32 val)
285 {
286 	acpi_os_write_port(reg->address, val, reg->bit_width);
287 }
288 
289 struct drv_cmd {
290 	struct acpi_pct_register *reg;
291 	u32 val;
292 	union {
293 		void (*write)(struct acpi_pct_register *reg, u32 val);
294 		u32 (*read)(struct acpi_pct_register *reg);
295 	} func;
296 };
297 
298 /* Called via smp_call_function_single(), on the target CPU */
299 static void do_drv_read(void *_cmd)
300 {
301 	struct drv_cmd *cmd = _cmd;
302 
303 	cmd->val = cmd->func.read(cmd->reg);
304 }
305 
306 static u32 drv_read(struct acpi_cpufreq_data *data, const struct cpumask *mask)
307 {
308 	struct acpi_processor_performance *perf = to_perf_data(data);
309 	struct drv_cmd cmd = {
310 		.reg = &perf->control_register,
311 		.func.read = data->cpu_freq_read,
312 	};
313 	int err;
314 
315 	err = smp_call_function_any(mask, do_drv_read, &cmd, 1);
316 	WARN_ON_ONCE(err);	/* smp_call_function_any() was buggy? */
317 	return cmd.val;
318 }
319 
320 /* Called via smp_call_function_many(), on the target CPUs */
321 static void do_drv_write(void *_cmd)
322 {
323 	struct drv_cmd *cmd = _cmd;
324 
325 	cmd->func.write(cmd->reg, cmd->val);
326 }
327 
328 static 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 	int this_cpu;
338 
339 	this_cpu = get_cpu();
340 	if (cpumask_test_cpu(this_cpu, mask))
341 		do_drv_write(&cmd);
342 
343 	smp_call_function_many(mask, do_drv_write, &cmd, 1);
344 	put_cpu();
345 }
346 
347 static u32 get_cur_val(const struct cpumask *mask, struct acpi_cpufreq_data *data)
348 {
349 	u32 val;
350 
351 	if (unlikely(cpumask_empty(mask)))
352 		return 0;
353 
354 	val = drv_read(data, mask);
355 
356 	pr_debug("%s = %u\n", __func__, val);
357 
358 	return val;
359 }
360 
361 static unsigned int get_cur_freq_on_cpu(unsigned int cpu)
362 {
363 	struct acpi_cpufreq_data *data;
364 	struct cpufreq_policy *policy;
365 	unsigned int freq;
366 	unsigned int cached_freq;
367 
368 	pr_debug("%s (%d)\n", __func__, cpu);
369 
370 	policy = cpufreq_cpu_get_raw(cpu);
371 	if (unlikely(!policy))
372 		return 0;
373 
374 	data = policy->driver_data;
375 	if (unlikely(!data || !policy->freq_table))
376 		return 0;
377 
378 	cached_freq = policy->freq_table[to_perf_data(data)->state].frequency;
379 	freq = extract_freq(policy, get_cur_val(cpumask_of(cpu), data));
380 	if (freq != cached_freq) {
381 		/*
382 		 * The dreaded BIOS frequency change behind our back.
383 		 * Force set the frequency on next target call.
384 		 */
385 		data->resume = 1;
386 	}
387 
388 	pr_debug("cur freq = %u\n", freq);
389 
390 	return freq;
391 }
392 
393 static unsigned int check_freqs(struct cpufreq_policy *policy,
394 				const struct cpumask *mask, unsigned int freq)
395 {
396 	struct acpi_cpufreq_data *data = policy->driver_data;
397 	unsigned int cur_freq;
398 	unsigned int i;
399 
400 	for (i = 0; i < 100; i++) {
401 		cur_freq = extract_freq(policy, get_cur_val(mask, data));
402 		if (cur_freq == freq)
403 			return 1;
404 		udelay(10);
405 	}
406 	return 0;
407 }
408 
409 static int acpi_cpufreq_target(struct cpufreq_policy *policy,
410 			       unsigned int index)
411 {
412 	struct acpi_cpufreq_data *data = policy->driver_data;
413 	struct acpi_processor_performance *perf;
414 	const struct cpumask *mask;
415 	unsigned int next_perf_state = 0; /* Index into perf table */
416 	int result = 0;
417 
418 	if (unlikely(!data)) {
419 		return -ENODEV;
420 	}
421 
422 	perf = to_perf_data(data);
423 	next_perf_state = policy->freq_table[index].driver_data;
424 	if (perf->state == next_perf_state) {
425 		if (unlikely(data->resume)) {
426 			pr_debug("Called after resume, resetting to P%d\n",
427 				next_perf_state);
428 			data->resume = 0;
429 		} else {
430 			pr_debug("Already at target state (P%d)\n",
431 				next_perf_state);
432 			return 0;
433 		}
434 	}
435 
436 	/*
437 	 * The core won't allow CPUs to go away until the governor has been
438 	 * stopped, so we can rely on the stability of policy->cpus.
439 	 */
440 	mask = policy->shared_type == CPUFREQ_SHARED_TYPE_ANY ?
441 		cpumask_of(policy->cpu) : policy->cpus;
442 
443 	drv_write(data, mask, perf->states[next_perf_state].control);
444 
445 	if (acpi_pstate_strict) {
446 		if (!check_freqs(policy, mask,
447 				 policy->freq_table[index].frequency)) {
448 			pr_debug("%s (%d)\n", __func__, policy->cpu);
449 			result = -EAGAIN;
450 		}
451 	}
452 
453 	if (!result)
454 		perf->state = next_perf_state;
455 
456 	return result;
457 }
458 
459 static unsigned int acpi_cpufreq_fast_switch(struct cpufreq_policy *policy,
460 					     unsigned int target_freq)
461 {
462 	struct acpi_cpufreq_data *data = policy->driver_data;
463 	struct acpi_processor_performance *perf;
464 	struct cpufreq_frequency_table *entry;
465 	unsigned int next_perf_state, next_freq, index;
466 
467 	/*
468 	 * Find the closest frequency above target_freq.
469 	 */
470 	if (policy->cached_target_freq == target_freq)
471 		index = policy->cached_resolved_idx;
472 	else
473 		index = cpufreq_table_find_index_dl(policy, target_freq);
474 
475 	entry = &policy->freq_table[index];
476 	next_freq = entry->frequency;
477 	next_perf_state = entry->driver_data;
478 
479 	perf = to_perf_data(data);
480 	if (perf->state == next_perf_state) {
481 		if (unlikely(data->resume))
482 			data->resume = 0;
483 		else
484 			return next_freq;
485 	}
486 
487 	data->cpu_freq_write(&perf->control_register,
488 			     perf->states[next_perf_state].control);
489 	perf->state = next_perf_state;
490 	return next_freq;
491 }
492 
493 static unsigned long
494 acpi_cpufreq_guess_freq(struct acpi_cpufreq_data *data, unsigned int cpu)
495 {
496 	struct acpi_processor_performance *perf;
497 
498 	perf = to_perf_data(data);
499 	if (cpu_khz) {
500 		/* search the closest match to cpu_khz */
501 		unsigned int i;
502 		unsigned long freq;
503 		unsigned long freqn = perf->states[0].core_frequency * 1000;
504 
505 		for (i = 0; i < (perf->state_count-1); i++) {
506 			freq = freqn;
507 			freqn = perf->states[i+1].core_frequency * 1000;
508 			if ((2 * cpu_khz) > (freqn + freq)) {
509 				perf->state = i;
510 				return freq;
511 			}
512 		}
513 		perf->state = perf->state_count-1;
514 		return freqn;
515 	} else {
516 		/* assume CPU is at P0... */
517 		perf->state = 0;
518 		return perf->states[0].core_frequency * 1000;
519 	}
520 }
521 
522 static void free_acpi_perf_data(void)
523 {
524 	unsigned int i;
525 
526 	/* Freeing a NULL pointer is OK, and alloc_percpu zeroes. */
527 	for_each_possible_cpu(i)
528 		free_cpumask_var(per_cpu_ptr(acpi_perf_data, i)
529 				 ->shared_cpu_map);
530 	free_percpu(acpi_perf_data);
531 }
532 
533 static int cpufreq_boost_online(unsigned int cpu)
534 {
535 	/*
536 	 * On the CPU_UP path we simply keep the boost-disable flag
537 	 * in sync with the current global state.
538 	 */
539 	return boost_set_msr(acpi_cpufreq_driver.boost_enabled);
540 }
541 
542 static int cpufreq_boost_down_prep(unsigned int cpu)
543 {
544 	/*
545 	 * Clear the boost-disable bit on the CPU_DOWN path so that
546 	 * this cpu cannot block the remaining ones from boosting.
547 	 */
548 	return boost_set_msr(1);
549 }
550 
551 /*
552  * acpi_cpufreq_early_init - initialize ACPI P-States library
553  *
554  * Initialize the ACPI P-States library (drivers/acpi/processor_perflib.c)
555  * in order to determine correct frequency and voltage pairings. We can
556  * do _PDC and _PSD and find out the processor dependency for the
557  * actual init that will happen later...
558  */
559 static int __init acpi_cpufreq_early_init(void)
560 {
561 	unsigned int i;
562 	pr_debug("%s\n", __func__);
563 
564 	acpi_perf_data = alloc_percpu(struct acpi_processor_performance);
565 	if (!acpi_perf_data) {
566 		pr_debug("Memory allocation error for acpi_perf_data.\n");
567 		return -ENOMEM;
568 	}
569 	for_each_possible_cpu(i) {
570 		if (!zalloc_cpumask_var_node(
571 			&per_cpu_ptr(acpi_perf_data, i)->shared_cpu_map,
572 			GFP_KERNEL, cpu_to_node(i))) {
573 
574 			/* Freeing a NULL pointer is OK: alloc_percpu zeroes. */
575 			free_acpi_perf_data();
576 			return -ENOMEM;
577 		}
578 	}
579 
580 	/* Do initialization in ACPI core */
581 	acpi_processor_preregister_performance(acpi_perf_data);
582 	return 0;
583 }
584 
585 #ifdef CONFIG_SMP
586 /*
587  * Some BIOSes do SW_ANY coordination internally, either set it up in hw
588  * or do it in BIOS firmware and won't inform about it to OS. If not
589  * detected, this has a side effect of making CPU run at a different speed
590  * than OS intended it to run at. Detect it and handle it cleanly.
591  */
592 static int bios_with_sw_any_bug;
593 
594 static int sw_any_bug_found(const struct dmi_system_id *d)
595 {
596 	bios_with_sw_any_bug = 1;
597 	return 0;
598 }
599 
600 static const struct dmi_system_id sw_any_bug_dmi_table[] = {
601 	{
602 		.callback = sw_any_bug_found,
603 		.ident = "Supermicro Server X6DLP",
604 		.matches = {
605 			DMI_MATCH(DMI_SYS_VENDOR, "Supermicro"),
606 			DMI_MATCH(DMI_BIOS_VERSION, "080010"),
607 			DMI_MATCH(DMI_PRODUCT_NAME, "X6DLP"),
608 		},
609 	},
610 	{ }
611 };
612 
613 static int acpi_cpufreq_blacklist(struct cpuinfo_x86 *c)
614 {
615 	/* Intel Xeon Processor 7100 Series Specification Update
616 	 * https://www.intel.com/Assets/PDF/specupdate/314554.pdf
617 	 * AL30: A Machine Check Exception (MCE) Occurring during an
618 	 * Enhanced Intel SpeedStep Technology Ratio Change May Cause
619 	 * Both Processor Cores to Lock Up. */
620 	if (c->x86_vendor == X86_VENDOR_INTEL) {
621 		if ((c->x86 == 15) &&
622 		    (c->x86_model == 6) &&
623 		    (c->x86_stepping == 8)) {
624 			pr_info("Intel(R) Xeon(R) 7100 Errata AL30, processors may lock up on frequency changes: disabling acpi-cpufreq\n");
625 			return -ENODEV;
626 		    }
627 		}
628 	return 0;
629 }
630 #endif
631 
632 #ifdef CONFIG_ACPI_CPPC_LIB
633 static u64 get_max_boost_ratio(unsigned int cpu)
634 {
635 	struct cppc_perf_caps perf_caps;
636 	u64 highest_perf, nominal_perf;
637 	int ret;
638 
639 	if (acpi_pstate_strict)
640 		return 0;
641 
642 	ret = cppc_get_perf_caps(cpu, &perf_caps);
643 	if (ret) {
644 		pr_debug("CPU%d: Unable to get performance capabilities (%d)\n",
645 			 cpu, ret);
646 		return 0;
647 	}
648 
649 	highest_perf = perf_caps.highest_perf;
650 	nominal_perf = perf_caps.nominal_perf;
651 
652 	if (!highest_perf || !nominal_perf) {
653 		pr_debug("CPU%d: highest or nominal performance missing\n", cpu);
654 		return 0;
655 	}
656 
657 	if (highest_perf < nominal_perf) {
658 		pr_debug("CPU%d: nominal performance above highest\n", cpu);
659 		return 0;
660 	}
661 
662 	return div_u64(highest_perf << SCHED_CAPACITY_SHIFT, nominal_perf);
663 }
664 #else
665 static inline u64 get_max_boost_ratio(unsigned int cpu) { return 0; }
666 #endif
667 
668 static int acpi_cpufreq_cpu_init(struct cpufreq_policy *policy)
669 {
670 	struct cpufreq_frequency_table *freq_table;
671 	struct acpi_processor_performance *perf;
672 	struct acpi_cpufreq_data *data;
673 	unsigned int cpu = policy->cpu;
674 	struct cpuinfo_x86 *c = &cpu_data(cpu);
675 	unsigned int valid_states = 0;
676 	unsigned int result = 0;
677 	u64 max_boost_ratio;
678 	unsigned int i;
679 #ifdef CONFIG_SMP
680 	static int blacklisted;
681 #endif
682 
683 	pr_debug("%s\n", __func__);
684 
685 #ifdef CONFIG_SMP
686 	if (blacklisted)
687 		return blacklisted;
688 	blacklisted = acpi_cpufreq_blacklist(c);
689 	if (blacklisted)
690 		return blacklisted;
691 #endif
692 
693 	data = kzalloc(sizeof(*data), GFP_KERNEL);
694 	if (!data)
695 		return -ENOMEM;
696 
697 	if (!zalloc_cpumask_var(&data->freqdomain_cpus, GFP_KERNEL)) {
698 		result = -ENOMEM;
699 		goto err_free;
700 	}
701 
702 	perf = per_cpu_ptr(acpi_perf_data, cpu);
703 	data->acpi_perf_cpu = cpu;
704 	policy->driver_data = data;
705 
706 	if (cpu_has(c, X86_FEATURE_CONSTANT_TSC))
707 		acpi_cpufreq_driver.flags |= CPUFREQ_CONST_LOOPS;
708 
709 	result = acpi_processor_register_performance(perf, cpu);
710 	if (result)
711 		goto err_free_mask;
712 
713 	policy->shared_type = perf->shared_type;
714 
715 	/*
716 	 * Will let policy->cpus know about dependency only when software
717 	 * coordination is required.
718 	 */
719 	if (policy->shared_type == CPUFREQ_SHARED_TYPE_ALL ||
720 	    policy->shared_type == CPUFREQ_SHARED_TYPE_ANY) {
721 		cpumask_copy(policy->cpus, perf->shared_cpu_map);
722 	}
723 	cpumask_copy(data->freqdomain_cpus, perf->shared_cpu_map);
724 
725 #ifdef CONFIG_SMP
726 	dmi_check_system(sw_any_bug_dmi_table);
727 	if (bios_with_sw_any_bug && !policy_is_shared(policy)) {
728 		policy->shared_type = CPUFREQ_SHARED_TYPE_ALL;
729 		cpumask_copy(policy->cpus, topology_core_cpumask(cpu));
730 	}
731 
732 	if (check_amd_hwpstate_cpu(cpu) && boot_cpu_data.x86 < 0x19 &&
733 	    !acpi_pstate_strict) {
734 		cpumask_clear(policy->cpus);
735 		cpumask_set_cpu(cpu, policy->cpus);
736 		cpumask_copy(data->freqdomain_cpus,
737 			     topology_sibling_cpumask(cpu));
738 		policy->shared_type = CPUFREQ_SHARED_TYPE_HW;
739 		pr_info_once("overriding BIOS provided _PSD data\n");
740 	}
741 #endif
742 
743 	/* capability check */
744 	if (perf->state_count <= 1) {
745 		pr_debug("No P-States\n");
746 		result = -ENODEV;
747 		goto err_unreg;
748 	}
749 
750 	if (perf->control_register.space_id != perf->status_register.space_id) {
751 		result = -ENODEV;
752 		goto err_unreg;
753 	}
754 
755 	switch (perf->control_register.space_id) {
756 	case ACPI_ADR_SPACE_SYSTEM_IO:
757 		if (boot_cpu_data.x86_vendor == X86_VENDOR_AMD &&
758 		    boot_cpu_data.x86 == 0xf) {
759 			pr_debug("AMD K8 systems must use native drivers.\n");
760 			result = -ENODEV;
761 			goto err_unreg;
762 		}
763 		pr_debug("SYSTEM IO addr space\n");
764 		data->cpu_feature = SYSTEM_IO_CAPABLE;
765 		data->cpu_freq_read = cpu_freq_read_io;
766 		data->cpu_freq_write = cpu_freq_write_io;
767 		break;
768 	case ACPI_ADR_SPACE_FIXED_HARDWARE:
769 		pr_debug("HARDWARE addr space\n");
770 		if (check_est_cpu(cpu)) {
771 			data->cpu_feature = SYSTEM_INTEL_MSR_CAPABLE;
772 			data->cpu_freq_read = cpu_freq_read_intel;
773 			data->cpu_freq_write = cpu_freq_write_intel;
774 			break;
775 		}
776 		if (check_amd_hwpstate_cpu(cpu)) {
777 			data->cpu_feature = SYSTEM_AMD_MSR_CAPABLE;
778 			data->cpu_freq_read = cpu_freq_read_amd;
779 			data->cpu_freq_write = cpu_freq_write_amd;
780 			break;
781 		}
782 		result = -ENODEV;
783 		goto err_unreg;
784 	default:
785 		pr_debug("Unknown addr space %d\n",
786 			(u32) (perf->control_register.space_id));
787 		result = -ENODEV;
788 		goto err_unreg;
789 	}
790 
791 	freq_table = kcalloc(perf->state_count + 1, sizeof(*freq_table),
792 			     GFP_KERNEL);
793 	if (!freq_table) {
794 		result = -ENOMEM;
795 		goto err_unreg;
796 	}
797 
798 	/* detect transition latency */
799 	policy->cpuinfo.transition_latency = 0;
800 	for (i = 0; i < perf->state_count; i++) {
801 		if ((perf->states[i].transition_latency * 1000) >
802 		    policy->cpuinfo.transition_latency)
803 			policy->cpuinfo.transition_latency =
804 			    perf->states[i].transition_latency * 1000;
805 	}
806 
807 	/* Check for high latency (>20uS) from buggy BIOSes, like on T42 */
808 	if (perf->control_register.space_id == ACPI_ADR_SPACE_FIXED_HARDWARE &&
809 	    policy->cpuinfo.transition_latency > 20 * 1000) {
810 		policy->cpuinfo.transition_latency = 20 * 1000;
811 		pr_info_once("P-state transition latency capped at 20 uS\n");
812 	}
813 
814 	/* table init */
815 	for (i = 0; i < perf->state_count; i++) {
816 		if (i > 0 && perf->states[i].core_frequency >=
817 		    freq_table[valid_states-1].frequency / 1000)
818 			continue;
819 
820 		freq_table[valid_states].driver_data = i;
821 		freq_table[valid_states].frequency =
822 		    perf->states[i].core_frequency * 1000;
823 		valid_states++;
824 	}
825 	freq_table[valid_states].frequency = CPUFREQ_TABLE_END;
826 
827 	max_boost_ratio = get_max_boost_ratio(cpu);
828 	if (max_boost_ratio) {
829 		unsigned int freq = freq_table[0].frequency;
830 
831 		/*
832 		 * Because the loop above sorts the freq_table entries in the
833 		 * descending order, freq is the maximum frequency in the table.
834 		 * Assume that it corresponds to the CPPC nominal frequency and
835 		 * use it to set cpuinfo.max_freq.
836 		 */
837 		policy->cpuinfo.max_freq = freq * max_boost_ratio >> SCHED_CAPACITY_SHIFT;
838 	} else {
839 		/*
840 		 * If the maximum "boost" frequency is unknown, ask the arch
841 		 * scale-invariance code to use the "nominal" performance for
842 		 * CPU utilization scaling so as to prevent the schedutil
843 		 * governor from selecting inadequate CPU frequencies.
844 		 */
845 		arch_set_max_freq_ratio(true);
846 	}
847 
848 	policy->freq_table = freq_table;
849 	perf->state = 0;
850 
851 	switch (perf->control_register.space_id) {
852 	case ACPI_ADR_SPACE_SYSTEM_IO:
853 		/*
854 		 * The core will not set policy->cur, because
855 		 * cpufreq_driver->get is NULL, so we need to set it here.
856 		 * However, we have to guess it, because the current speed is
857 		 * unknown and not detectable via IO ports.
858 		 */
859 		policy->cur = acpi_cpufreq_guess_freq(data, policy->cpu);
860 		break;
861 	case ACPI_ADR_SPACE_FIXED_HARDWARE:
862 		acpi_cpufreq_driver.get = get_cur_freq_on_cpu;
863 		break;
864 	default:
865 		break;
866 	}
867 
868 	/* notify BIOS that we exist */
869 	acpi_processor_notify_smm(THIS_MODULE);
870 
871 	pr_debug("CPU%u - ACPI performance management activated.\n", cpu);
872 	for (i = 0; i < perf->state_count; i++)
873 		pr_debug("     %cP%d: %d MHz, %d mW, %d uS\n",
874 			(i == perf->state ? '*' : ' '), i,
875 			(u32) perf->states[i].core_frequency,
876 			(u32) perf->states[i].power,
877 			(u32) perf->states[i].transition_latency);
878 
879 	/*
880 	 * the first call to ->target() should result in us actually
881 	 * writing something to the appropriate registers.
882 	 */
883 	data->resume = 1;
884 
885 	policy->fast_switch_possible = !acpi_pstate_strict &&
886 		!(policy_is_shared(policy) && policy->shared_type != CPUFREQ_SHARED_TYPE_ANY);
887 
888 	return result;
889 
890 err_unreg:
891 	acpi_processor_unregister_performance(cpu);
892 err_free_mask:
893 	free_cpumask_var(data->freqdomain_cpus);
894 err_free:
895 	kfree(data);
896 	policy->driver_data = NULL;
897 
898 	return result;
899 }
900 
901 static int acpi_cpufreq_cpu_exit(struct cpufreq_policy *policy)
902 {
903 	struct acpi_cpufreq_data *data = policy->driver_data;
904 
905 	pr_debug("%s\n", __func__);
906 
907 	policy->fast_switch_possible = false;
908 	policy->driver_data = NULL;
909 	acpi_processor_unregister_performance(data->acpi_perf_cpu);
910 	free_cpumask_var(data->freqdomain_cpus);
911 	kfree(policy->freq_table);
912 	kfree(data);
913 
914 	return 0;
915 }
916 
917 static void acpi_cpufreq_cpu_ready(struct cpufreq_policy *policy)
918 {
919 	struct acpi_processor_performance *perf = per_cpu_ptr(acpi_perf_data,
920 							      policy->cpu);
921 	unsigned int freq = policy->freq_table[0].frequency;
922 
923 	if (perf->states[0].core_frequency * 1000 != freq)
924 		pr_warn(FW_WARN "P-state 0 is not max freq\n");
925 }
926 
927 static int acpi_cpufreq_resume(struct cpufreq_policy *policy)
928 {
929 	struct acpi_cpufreq_data *data = policy->driver_data;
930 
931 	pr_debug("%s\n", __func__);
932 
933 	data->resume = 1;
934 
935 	return 0;
936 }
937 
938 static struct freq_attr *acpi_cpufreq_attr[] = {
939 	&cpufreq_freq_attr_scaling_available_freqs,
940 	&freqdomain_cpus,
941 #ifdef CONFIG_X86_ACPI_CPUFREQ_CPB
942 	&cpb,
943 #endif
944 	NULL,
945 };
946 
947 static struct cpufreq_driver acpi_cpufreq_driver = {
948 	.verify		= cpufreq_generic_frequency_table_verify,
949 	.target_index	= acpi_cpufreq_target,
950 	.fast_switch	= acpi_cpufreq_fast_switch,
951 	.bios_limit	= acpi_processor_get_bios_limit,
952 	.init		= acpi_cpufreq_cpu_init,
953 	.exit		= acpi_cpufreq_cpu_exit,
954 	.ready		= acpi_cpufreq_cpu_ready,
955 	.resume		= acpi_cpufreq_resume,
956 	.name		= "acpi-cpufreq",
957 	.attr		= acpi_cpufreq_attr,
958 };
959 
960 static enum cpuhp_state acpi_cpufreq_online;
961 
962 static void __init acpi_cpufreq_boost_init(void)
963 {
964 	int ret;
965 
966 	if (!(boot_cpu_has(X86_FEATURE_CPB) || boot_cpu_has(X86_FEATURE_IDA))) {
967 		pr_debug("Boost capabilities not present in the processor\n");
968 		return;
969 	}
970 
971 	acpi_cpufreq_driver.set_boost = set_boost;
972 	acpi_cpufreq_driver.boost_enabled = boost_state(0);
973 
974 	/*
975 	 * This calls the online callback on all online cpu and forces all
976 	 * MSRs to the same value.
977 	 */
978 	ret = cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "cpufreq/acpi:online",
979 				cpufreq_boost_online, cpufreq_boost_down_prep);
980 	if (ret < 0) {
981 		pr_err("acpi_cpufreq: failed to register hotplug callbacks\n");
982 		return;
983 	}
984 	acpi_cpufreq_online = ret;
985 }
986 
987 static void acpi_cpufreq_boost_exit(void)
988 {
989 	if (acpi_cpufreq_online > 0)
990 		cpuhp_remove_state_nocalls(acpi_cpufreq_online);
991 }
992 
993 static int __init acpi_cpufreq_init(void)
994 {
995 	int ret;
996 
997 	if (acpi_disabled)
998 		return -ENODEV;
999 
1000 	/* don't keep reloading if cpufreq_driver exists */
1001 	if (cpufreq_get_current_driver())
1002 		return -EEXIST;
1003 
1004 	pr_debug("%s\n", __func__);
1005 
1006 	ret = acpi_cpufreq_early_init();
1007 	if (ret)
1008 		return ret;
1009 
1010 #ifdef CONFIG_X86_ACPI_CPUFREQ_CPB
1011 	/* this is a sysfs file with a strange name and an even stranger
1012 	 * semantic - per CPU instantiation, but system global effect.
1013 	 * Lets enable it only on AMD CPUs for compatibility reasons and
1014 	 * only if configured. This is considered legacy code, which
1015 	 * will probably be removed at some point in the future.
1016 	 */
1017 	if (!check_amd_hwpstate_cpu(0)) {
1018 		struct freq_attr **attr;
1019 
1020 		pr_debug("CPB unsupported, do not expose it\n");
1021 
1022 		for (attr = acpi_cpufreq_attr; *attr; attr++)
1023 			if (*attr == &cpb) {
1024 				*attr = NULL;
1025 				break;
1026 			}
1027 	}
1028 #endif
1029 	acpi_cpufreq_boost_init();
1030 
1031 	ret = cpufreq_register_driver(&acpi_cpufreq_driver);
1032 	if (ret) {
1033 		free_acpi_perf_data();
1034 		acpi_cpufreq_boost_exit();
1035 	}
1036 	return ret;
1037 }
1038 
1039 static void __exit acpi_cpufreq_exit(void)
1040 {
1041 	pr_debug("%s\n", __func__);
1042 
1043 	acpi_cpufreq_boost_exit();
1044 
1045 	cpufreq_unregister_driver(&acpi_cpufreq_driver);
1046 
1047 	free_acpi_perf_data();
1048 }
1049 
1050 module_param(acpi_pstate_strict, uint, 0644);
1051 MODULE_PARM_DESC(acpi_pstate_strict,
1052 	"value 0 or non-zero. non-zero -> strict ACPI checks are "
1053 	"performed during frequency changes.");
1054 
1055 late_initcall(acpi_cpufreq_init);
1056 module_exit(acpi_cpufreq_exit);
1057 
1058 static const struct x86_cpu_id __maybe_unused acpi_cpufreq_ids[] = {
1059 	X86_MATCH_FEATURE(X86_FEATURE_ACPI, NULL),
1060 	X86_MATCH_FEATURE(X86_FEATURE_HW_PSTATE, NULL),
1061 	{}
1062 };
1063 MODULE_DEVICE_TABLE(x86cpu, acpi_cpufreq_ids);
1064 
1065 static const struct acpi_device_id __maybe_unused processor_device_ids[] = {
1066 	{ACPI_PROCESSOR_OBJECT_HID, },
1067 	{ACPI_PROCESSOR_DEVICE_HID, },
1068 	{},
1069 };
1070 MODULE_DEVICE_TABLE(acpi, processor_device_ids);
1071 
1072 MODULE_ALIAS("acpi");
1073