1 /*
2  * acpi-cpufreq.c - ACPI Processor P-States Driver
3  *
4  *  Copyright (C) 2001, 2002 Andy Grover <andrew.grover@intel.com>
5  *  Copyright (C) 2001, 2002 Paul Diefenbaugh <paul.s.diefenbaugh@intel.com>
6  *  Copyright (C) 2002 - 2004 Dominik Brodowski <linux@brodo.de>
7  *  Copyright (C) 2006       Denis Sadykov <denis.m.sadykov@intel.com>
8  *
9  * ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
10  *
11  *  This program is free software; you can redistribute it and/or modify
12  *  it under the terms of the GNU General Public License as published by
13  *  the Free Software Foundation; either version 2 of the License, or (at
14  *  your option) any later version.
15  *
16  *  This program is distributed in the hope that it will be useful, but
17  *  WITHOUT ANY WARRANTY; without even the implied warranty of
18  *  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
19  *  General Public License for more details.
20  *
21  *  You should have received a copy of the GNU General Public License along
22  *  with this program; if not, write to the Free Software Foundation, Inc.,
23  *  59 Temple Place, Suite 330, Boston, MA 02111-1307 USA.
24  *
25  * ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
26  */
27 
28 #include <linux/kernel.h>
29 #include <linux/module.h>
30 #include <linux/init.h>
31 #include <linux/smp.h>
32 #include <linux/sched.h>
33 #include <linux/cpufreq.h>
34 #include <linux/compiler.h>
35 #include <linux/dmi.h>
36 #include <linux/slab.h>
37 
38 #include <linux/acpi.h>
39 #include <linux/io.h>
40 #include <linux/delay.h>
41 #include <linux/uaccess.h>
42 
43 #include <acpi/processor.h>
44 
45 #include <asm/msr.h>
46 #include <asm/processor.h>
47 #include <asm/cpufeature.h>
48 
49 MODULE_AUTHOR("Paul Diefenbaugh, Dominik Brodowski");
50 MODULE_DESCRIPTION("ACPI Processor P-States Driver");
51 MODULE_LICENSE("GPL");
52 
53 #define PFX "acpi-cpufreq: "
54 
55 enum {
56 	UNDEFINED_CAPABLE = 0,
57 	SYSTEM_INTEL_MSR_CAPABLE,
58 	SYSTEM_AMD_MSR_CAPABLE,
59 	SYSTEM_IO_CAPABLE,
60 };
61 
62 #define INTEL_MSR_RANGE		(0xffff)
63 #define AMD_MSR_RANGE		(0x7)
64 
65 #define MSR_K7_HWCR_CPB_DIS	(1ULL << 25)
66 
67 struct acpi_cpufreq_data {
68 	struct cpufreq_frequency_table *freq_table;
69 	unsigned int resume;
70 	unsigned int cpu_feature;
71 	unsigned int acpi_perf_cpu;
72 	cpumask_var_t freqdomain_cpus;
73 };
74 
75 /* acpi_perf_data is a pointer to percpu data. */
76 static struct acpi_processor_performance __percpu *acpi_perf_data;
77 
78 static inline struct acpi_processor_performance *to_perf_data(struct acpi_cpufreq_data *data)
79 {
80 	return per_cpu_ptr(acpi_perf_data, data->acpi_perf_cpu);
81 }
82 
83 static struct cpufreq_driver acpi_cpufreq_driver;
84 
85 static unsigned int acpi_pstate_strict;
86 static struct msr __percpu *msrs;
87 
88 static bool boost_state(unsigned int cpu)
89 {
90 	u32 lo, hi;
91 	u64 msr;
92 
93 	switch (boot_cpu_data.x86_vendor) {
94 	case X86_VENDOR_INTEL:
95 		rdmsr_on_cpu(cpu, MSR_IA32_MISC_ENABLE, &lo, &hi);
96 		msr = lo | ((u64)hi << 32);
97 		return !(msr & MSR_IA32_MISC_ENABLE_TURBO_DISABLE);
98 	case X86_VENDOR_AMD:
99 		rdmsr_on_cpu(cpu, MSR_K7_HWCR, &lo, &hi);
100 		msr = lo | ((u64)hi << 32);
101 		return !(msr & MSR_K7_HWCR_CPB_DIS);
102 	}
103 	return false;
104 }
105 
106 static void boost_set_msrs(bool enable, const struct cpumask *cpumask)
107 {
108 	u32 cpu;
109 	u32 msr_addr;
110 	u64 msr_mask;
111 
112 	switch (boot_cpu_data.x86_vendor) {
113 	case X86_VENDOR_INTEL:
114 		msr_addr = MSR_IA32_MISC_ENABLE;
115 		msr_mask = MSR_IA32_MISC_ENABLE_TURBO_DISABLE;
116 		break;
117 	case X86_VENDOR_AMD:
118 		msr_addr = MSR_K7_HWCR;
119 		msr_mask = MSR_K7_HWCR_CPB_DIS;
120 		break;
121 	default:
122 		return;
123 	}
124 
125 	rdmsr_on_cpus(cpumask, msr_addr, msrs);
126 
127 	for_each_cpu(cpu, cpumask) {
128 		struct msr *reg = per_cpu_ptr(msrs, cpu);
129 		if (enable)
130 			reg->q &= ~msr_mask;
131 		else
132 			reg->q |= msr_mask;
133 	}
134 
135 	wrmsr_on_cpus(cpumask, msr_addr, msrs);
136 }
137 
138 static int _store_boost(int val)
139 {
140 	get_online_cpus();
141 	boost_set_msrs(val, cpu_online_mask);
142 	put_online_cpus();
143 	pr_debug("Core Boosting %sabled.\n", val ? "en" : "dis");
144 
145 	return 0;
146 }
147 
148 static ssize_t show_freqdomain_cpus(struct cpufreq_policy *policy, char *buf)
149 {
150 	struct acpi_cpufreq_data *data = policy->driver_data;
151 
152 	return cpufreq_show_cpus(data->freqdomain_cpus, buf);
153 }
154 
155 cpufreq_freq_attr_ro(freqdomain_cpus);
156 
157 #ifdef CONFIG_X86_ACPI_CPUFREQ_CPB
158 static ssize_t store_boost(const char *buf, size_t count)
159 {
160 	int ret;
161 	unsigned long val = 0;
162 
163 	if (!acpi_cpufreq_driver.boost_supported)
164 		return -EINVAL;
165 
166 	ret = kstrtoul(buf, 10, &val);
167 	if (ret || (val > 1))
168 		return -EINVAL;
169 
170 	_store_boost((int) val);
171 
172 	return count;
173 }
174 
175 static ssize_t store_cpb(struct cpufreq_policy *policy, const char *buf,
176 			 size_t count)
177 {
178 	return store_boost(buf, count);
179 }
180 
181 static ssize_t show_cpb(struct cpufreq_policy *policy, char *buf)
182 {
183 	return sprintf(buf, "%u\n", acpi_cpufreq_driver.boost_enabled);
184 }
185 
186 cpufreq_freq_attr_rw(cpb);
187 #endif
188 
189 static int check_est_cpu(unsigned int cpuid)
190 {
191 	struct cpuinfo_x86 *cpu = &cpu_data(cpuid);
192 
193 	return cpu_has(cpu, X86_FEATURE_EST);
194 }
195 
196 static int check_amd_hwpstate_cpu(unsigned int cpuid)
197 {
198 	struct cpuinfo_x86 *cpu = &cpu_data(cpuid);
199 
200 	return cpu_has(cpu, X86_FEATURE_HW_PSTATE);
201 }
202 
203 static unsigned extract_io(u32 value, struct acpi_cpufreq_data *data)
204 {
205 	struct acpi_processor_performance *perf;
206 	int i;
207 
208 	perf = to_perf_data(data);
209 
210 	for (i = 0; i < perf->state_count; i++) {
211 		if (value == perf->states[i].status)
212 			return data->freq_table[i].frequency;
213 	}
214 	return 0;
215 }
216 
217 static unsigned extract_msr(u32 msr, struct acpi_cpufreq_data *data)
218 {
219 	struct cpufreq_frequency_table *pos;
220 	struct acpi_processor_performance *perf;
221 
222 	if (boot_cpu_data.x86_vendor == X86_VENDOR_AMD)
223 		msr &= AMD_MSR_RANGE;
224 	else
225 		msr &= INTEL_MSR_RANGE;
226 
227 	perf = to_perf_data(data);
228 
229 	cpufreq_for_each_entry(pos, data->freq_table)
230 		if (msr == perf->states[pos->driver_data].status)
231 			return pos->frequency;
232 	return data->freq_table[0].frequency;
233 }
234 
235 static unsigned extract_freq(u32 val, struct acpi_cpufreq_data *data)
236 {
237 	switch (data->cpu_feature) {
238 	case SYSTEM_INTEL_MSR_CAPABLE:
239 	case SYSTEM_AMD_MSR_CAPABLE:
240 		return extract_msr(val, data);
241 	case SYSTEM_IO_CAPABLE:
242 		return extract_io(val, data);
243 	default:
244 		return 0;
245 	}
246 }
247 
248 struct msr_addr {
249 	u32 reg;
250 };
251 
252 struct io_addr {
253 	u16 port;
254 	u8 bit_width;
255 };
256 
257 struct drv_cmd {
258 	unsigned int type;
259 	const struct cpumask *mask;
260 	union {
261 		struct msr_addr msr;
262 		struct io_addr io;
263 	} addr;
264 	u32 val;
265 };
266 
267 /* Called via smp_call_function_single(), on the target CPU */
268 static void do_drv_read(void *_cmd)
269 {
270 	struct drv_cmd *cmd = _cmd;
271 	u32 h;
272 
273 	switch (cmd->type) {
274 	case SYSTEM_INTEL_MSR_CAPABLE:
275 	case SYSTEM_AMD_MSR_CAPABLE:
276 		rdmsr(cmd->addr.msr.reg, cmd->val, h);
277 		break;
278 	case SYSTEM_IO_CAPABLE:
279 		acpi_os_read_port((acpi_io_address)cmd->addr.io.port,
280 				&cmd->val,
281 				(u32)cmd->addr.io.bit_width);
282 		break;
283 	default:
284 		break;
285 	}
286 }
287 
288 /* Called via smp_call_function_many(), on the target CPUs */
289 static void do_drv_write(void *_cmd)
290 {
291 	struct drv_cmd *cmd = _cmd;
292 	u32 lo, hi;
293 
294 	switch (cmd->type) {
295 	case SYSTEM_INTEL_MSR_CAPABLE:
296 		rdmsr(cmd->addr.msr.reg, lo, hi);
297 		lo = (lo & ~INTEL_MSR_RANGE) | (cmd->val & INTEL_MSR_RANGE);
298 		wrmsr(cmd->addr.msr.reg, lo, hi);
299 		break;
300 	case SYSTEM_AMD_MSR_CAPABLE:
301 		wrmsr(cmd->addr.msr.reg, cmd->val, 0);
302 		break;
303 	case SYSTEM_IO_CAPABLE:
304 		acpi_os_write_port((acpi_io_address)cmd->addr.io.port,
305 				cmd->val,
306 				(u32)cmd->addr.io.bit_width);
307 		break;
308 	default:
309 		break;
310 	}
311 }
312 
313 static void drv_read(struct drv_cmd *cmd)
314 {
315 	int err;
316 	cmd->val = 0;
317 
318 	err = smp_call_function_any(cmd->mask, do_drv_read, cmd, 1);
319 	WARN_ON_ONCE(err);	/* smp_call_function_any() was buggy? */
320 }
321 
322 static void drv_write(struct drv_cmd *cmd)
323 {
324 	int this_cpu;
325 
326 	this_cpu = get_cpu();
327 	if (cpumask_test_cpu(this_cpu, cmd->mask))
328 		do_drv_write(cmd);
329 	smp_call_function_many(cmd->mask, do_drv_write, cmd, 1);
330 	put_cpu();
331 }
332 
333 static u32
334 get_cur_val(const struct cpumask *mask, struct acpi_cpufreq_data *data)
335 {
336 	struct acpi_processor_performance *perf;
337 	struct drv_cmd cmd;
338 
339 	if (unlikely(cpumask_empty(mask)))
340 		return 0;
341 
342 	switch (data->cpu_feature) {
343 	case SYSTEM_INTEL_MSR_CAPABLE:
344 		cmd.type = SYSTEM_INTEL_MSR_CAPABLE;
345 		cmd.addr.msr.reg = MSR_IA32_PERF_CTL;
346 		break;
347 	case SYSTEM_AMD_MSR_CAPABLE:
348 		cmd.type = SYSTEM_AMD_MSR_CAPABLE;
349 		cmd.addr.msr.reg = MSR_AMD_PERF_CTL;
350 		break;
351 	case SYSTEM_IO_CAPABLE:
352 		cmd.type = SYSTEM_IO_CAPABLE;
353 		perf = to_perf_data(data);
354 		cmd.addr.io.port = perf->control_register.address;
355 		cmd.addr.io.bit_width = perf->control_register.bit_width;
356 		break;
357 	default:
358 		return 0;
359 	}
360 
361 	cmd.mask = mask;
362 	drv_read(&cmd);
363 
364 	pr_debug("get_cur_val = %u\n", cmd.val);
365 
366 	return cmd.val;
367 }
368 
369 static unsigned int get_cur_freq_on_cpu(unsigned int cpu)
370 {
371 	struct acpi_cpufreq_data *data;
372 	struct cpufreq_policy *policy;
373 	unsigned int freq;
374 	unsigned int cached_freq;
375 
376 	pr_debug("get_cur_freq_on_cpu (%d)\n", cpu);
377 
378 	policy = cpufreq_cpu_get(cpu);
379 	if (unlikely(!policy))
380 		return 0;
381 
382 	data = policy->driver_data;
383 	cpufreq_cpu_put(policy);
384 	if (unlikely(!data || !data->freq_table))
385 		return 0;
386 
387 	cached_freq = data->freq_table[to_perf_data(data)->state].frequency;
388 	freq = extract_freq(get_cur_val(cpumask_of(cpu), data), data);
389 	if (freq != cached_freq) {
390 		/*
391 		 * The dreaded BIOS frequency change behind our back.
392 		 * Force set the frequency on next target call.
393 		 */
394 		data->resume = 1;
395 	}
396 
397 	pr_debug("cur freq = %u\n", freq);
398 
399 	return freq;
400 }
401 
402 static unsigned int check_freqs(const struct cpumask *mask, unsigned int freq,
403 				struct acpi_cpufreq_data *data)
404 {
405 	unsigned int cur_freq;
406 	unsigned int i;
407 
408 	for (i = 0; i < 100; i++) {
409 		cur_freq = extract_freq(get_cur_val(mask, data), data);
410 		if (cur_freq == freq)
411 			return 1;
412 		udelay(10);
413 	}
414 	return 0;
415 }
416 
417 static int acpi_cpufreq_target(struct cpufreq_policy *policy,
418 			       unsigned int index)
419 {
420 	struct acpi_cpufreq_data *data = policy->driver_data;
421 	struct acpi_processor_performance *perf;
422 	struct drv_cmd cmd;
423 	unsigned int next_perf_state = 0; /* Index into perf table */
424 	int result = 0;
425 
426 	if (unlikely(data == NULL || data->freq_table == NULL)) {
427 		return -ENODEV;
428 	}
429 
430 	perf = to_perf_data(data);
431 	next_perf_state = data->freq_table[index].driver_data;
432 	if (perf->state == next_perf_state) {
433 		if (unlikely(data->resume)) {
434 			pr_debug("Called after resume, resetting to P%d\n",
435 				next_perf_state);
436 			data->resume = 0;
437 		} else {
438 			pr_debug("Already at target state (P%d)\n",
439 				next_perf_state);
440 			goto out;
441 		}
442 	}
443 
444 	switch (data->cpu_feature) {
445 	case SYSTEM_INTEL_MSR_CAPABLE:
446 		cmd.type = SYSTEM_INTEL_MSR_CAPABLE;
447 		cmd.addr.msr.reg = MSR_IA32_PERF_CTL;
448 		cmd.val = (u32) perf->states[next_perf_state].control;
449 		break;
450 	case SYSTEM_AMD_MSR_CAPABLE:
451 		cmd.type = SYSTEM_AMD_MSR_CAPABLE;
452 		cmd.addr.msr.reg = MSR_AMD_PERF_CTL;
453 		cmd.val = (u32) perf->states[next_perf_state].control;
454 		break;
455 	case SYSTEM_IO_CAPABLE:
456 		cmd.type = SYSTEM_IO_CAPABLE;
457 		cmd.addr.io.port = perf->control_register.address;
458 		cmd.addr.io.bit_width = perf->control_register.bit_width;
459 		cmd.val = (u32) perf->states[next_perf_state].control;
460 		break;
461 	default:
462 		result = -ENODEV;
463 		goto out;
464 	}
465 
466 	/* cpufreq holds the hotplug lock, so we are safe from here on */
467 	if (policy->shared_type != CPUFREQ_SHARED_TYPE_ANY)
468 		cmd.mask = policy->cpus;
469 	else
470 		cmd.mask = cpumask_of(policy->cpu);
471 
472 	drv_write(&cmd);
473 
474 	if (acpi_pstate_strict) {
475 		if (!check_freqs(cmd.mask, data->freq_table[index].frequency,
476 					data)) {
477 			pr_debug("acpi_cpufreq_target failed (%d)\n",
478 				policy->cpu);
479 			result = -EAGAIN;
480 		}
481 	}
482 
483 	if (!result)
484 		perf->state = next_perf_state;
485 
486 out:
487 	return result;
488 }
489 
490 static unsigned long
491 acpi_cpufreq_guess_freq(struct acpi_cpufreq_data *data, unsigned int cpu)
492 {
493 	struct acpi_processor_performance *perf;
494 
495 	perf = to_perf_data(data);
496 	if (cpu_khz) {
497 		/* search the closest match to cpu_khz */
498 		unsigned int i;
499 		unsigned long freq;
500 		unsigned long freqn = perf->states[0].core_frequency * 1000;
501 
502 		for (i = 0; i < (perf->state_count-1); i++) {
503 			freq = freqn;
504 			freqn = perf->states[i+1].core_frequency * 1000;
505 			if ((2 * cpu_khz) > (freqn + freq)) {
506 				perf->state = i;
507 				return freq;
508 			}
509 		}
510 		perf->state = perf->state_count-1;
511 		return freqn;
512 	} else {
513 		/* assume CPU is at P0... */
514 		perf->state = 0;
515 		return perf->states[0].core_frequency * 1000;
516 	}
517 }
518 
519 static void free_acpi_perf_data(void)
520 {
521 	unsigned int i;
522 
523 	/* Freeing a NULL pointer is OK, and alloc_percpu zeroes. */
524 	for_each_possible_cpu(i)
525 		free_cpumask_var(per_cpu_ptr(acpi_perf_data, i)
526 				 ->shared_cpu_map);
527 	free_percpu(acpi_perf_data);
528 }
529 
530 static int boost_notify(struct notifier_block *nb, unsigned long action,
531 		      void *hcpu)
532 {
533 	unsigned cpu = (long)hcpu;
534 	const struct cpumask *cpumask;
535 
536 	cpumask = get_cpu_mask(cpu);
537 
538 	/*
539 	 * Clear the boost-disable bit on the CPU_DOWN path so that
540 	 * this cpu cannot block the remaining ones from boosting. On
541 	 * the CPU_UP path we simply keep the boost-disable flag in
542 	 * sync with the current global state.
543 	 */
544 
545 	switch (action) {
546 	case CPU_UP_PREPARE:
547 	case CPU_UP_PREPARE_FROZEN:
548 		boost_set_msrs(acpi_cpufreq_driver.boost_enabled, cpumask);
549 		break;
550 
551 	case CPU_DOWN_PREPARE:
552 	case CPU_DOWN_PREPARE_FROZEN:
553 		boost_set_msrs(1, cpumask);
554 		break;
555 
556 	default:
557 		break;
558 	}
559 
560 	return NOTIFY_OK;
561 }
562 
563 
564 static struct notifier_block boost_nb = {
565 	.notifier_call          = boost_notify,
566 };
567 
568 /*
569  * acpi_cpufreq_early_init - initialize ACPI P-States library
570  *
571  * Initialize the ACPI P-States library (drivers/acpi/processor_perflib.c)
572  * in order to determine correct frequency and voltage pairings. We can
573  * do _PDC and _PSD and find out the processor dependency for the
574  * actual init that will happen later...
575  */
576 static int __init acpi_cpufreq_early_init(void)
577 {
578 	unsigned int i;
579 	pr_debug("acpi_cpufreq_early_init\n");
580 
581 	acpi_perf_data = alloc_percpu(struct acpi_processor_performance);
582 	if (!acpi_perf_data) {
583 		pr_debug("Memory allocation error for acpi_perf_data.\n");
584 		return -ENOMEM;
585 	}
586 	for_each_possible_cpu(i) {
587 		if (!zalloc_cpumask_var_node(
588 			&per_cpu_ptr(acpi_perf_data, i)->shared_cpu_map,
589 			GFP_KERNEL, cpu_to_node(i))) {
590 
591 			/* Freeing a NULL pointer is OK: alloc_percpu zeroes. */
592 			free_acpi_perf_data();
593 			return -ENOMEM;
594 		}
595 	}
596 
597 	/* Do initialization in ACPI core */
598 	acpi_processor_preregister_performance(acpi_perf_data);
599 	return 0;
600 }
601 
602 #ifdef CONFIG_SMP
603 /*
604  * Some BIOSes do SW_ANY coordination internally, either set it up in hw
605  * or do it in BIOS firmware and won't inform about it to OS. If not
606  * detected, this has a side effect of making CPU run at a different speed
607  * than OS intended it to run at. Detect it and handle it cleanly.
608  */
609 static int bios_with_sw_any_bug;
610 
611 static int sw_any_bug_found(const struct dmi_system_id *d)
612 {
613 	bios_with_sw_any_bug = 1;
614 	return 0;
615 }
616 
617 static const struct dmi_system_id sw_any_bug_dmi_table[] = {
618 	{
619 		.callback = sw_any_bug_found,
620 		.ident = "Supermicro Server X6DLP",
621 		.matches = {
622 			DMI_MATCH(DMI_SYS_VENDOR, "Supermicro"),
623 			DMI_MATCH(DMI_BIOS_VERSION, "080010"),
624 			DMI_MATCH(DMI_PRODUCT_NAME, "X6DLP"),
625 		},
626 	},
627 	{ }
628 };
629 
630 static int acpi_cpufreq_blacklist(struct cpuinfo_x86 *c)
631 {
632 	/* Intel Xeon Processor 7100 Series Specification Update
633 	 * http://www.intel.com/Assets/PDF/specupdate/314554.pdf
634 	 * AL30: A Machine Check Exception (MCE) Occurring during an
635 	 * Enhanced Intel SpeedStep Technology Ratio Change May Cause
636 	 * Both Processor Cores to Lock Up. */
637 	if (c->x86_vendor == X86_VENDOR_INTEL) {
638 		if ((c->x86 == 15) &&
639 		    (c->x86_model == 6) &&
640 		    (c->x86_mask == 8)) {
641 			printk(KERN_INFO "acpi-cpufreq: Intel(R) "
642 			    "Xeon(R) 7100 Errata AL30, processors may "
643 			    "lock up on frequency changes: disabling "
644 			    "acpi-cpufreq.\n");
645 			return -ENODEV;
646 		    }
647 		}
648 	return 0;
649 }
650 #endif
651 
652 static int acpi_cpufreq_cpu_init(struct cpufreq_policy *policy)
653 {
654 	unsigned int i;
655 	unsigned int valid_states = 0;
656 	unsigned int cpu = policy->cpu;
657 	struct acpi_cpufreq_data *data;
658 	unsigned int result = 0;
659 	struct cpuinfo_x86 *c = &cpu_data(policy->cpu);
660 	struct acpi_processor_performance *perf;
661 #ifdef CONFIG_SMP
662 	static int blacklisted;
663 #endif
664 
665 	pr_debug("acpi_cpufreq_cpu_init\n");
666 
667 #ifdef CONFIG_SMP
668 	if (blacklisted)
669 		return blacklisted;
670 	blacklisted = acpi_cpufreq_blacklist(c);
671 	if (blacklisted)
672 		return blacklisted;
673 #endif
674 
675 	data = kzalloc(sizeof(*data), GFP_KERNEL);
676 	if (!data)
677 		return -ENOMEM;
678 
679 	if (!zalloc_cpumask_var(&data->freqdomain_cpus, GFP_KERNEL)) {
680 		result = -ENOMEM;
681 		goto err_free;
682 	}
683 
684 	perf = per_cpu_ptr(acpi_perf_data, cpu);
685 	data->acpi_perf_cpu = cpu;
686 	policy->driver_data = data;
687 
688 	if (cpu_has(c, X86_FEATURE_CONSTANT_TSC))
689 		acpi_cpufreq_driver.flags |= CPUFREQ_CONST_LOOPS;
690 
691 	result = acpi_processor_register_performance(perf, cpu);
692 	if (result)
693 		goto err_free_mask;
694 
695 	policy->shared_type = perf->shared_type;
696 
697 	/*
698 	 * Will let policy->cpus know about dependency only when software
699 	 * coordination is required.
700 	 */
701 	if (policy->shared_type == CPUFREQ_SHARED_TYPE_ALL ||
702 	    policy->shared_type == CPUFREQ_SHARED_TYPE_ANY) {
703 		cpumask_copy(policy->cpus, perf->shared_cpu_map);
704 	}
705 	cpumask_copy(data->freqdomain_cpus, perf->shared_cpu_map);
706 
707 #ifdef CONFIG_SMP
708 	dmi_check_system(sw_any_bug_dmi_table);
709 	if (bios_with_sw_any_bug && !policy_is_shared(policy)) {
710 		policy->shared_type = CPUFREQ_SHARED_TYPE_ALL;
711 		cpumask_copy(policy->cpus, topology_core_cpumask(cpu));
712 	}
713 
714 	if (check_amd_hwpstate_cpu(cpu) && !acpi_pstate_strict) {
715 		cpumask_clear(policy->cpus);
716 		cpumask_set_cpu(cpu, policy->cpus);
717 		cpumask_copy(data->freqdomain_cpus,
718 			     topology_sibling_cpumask(cpu));
719 		policy->shared_type = CPUFREQ_SHARED_TYPE_HW;
720 		pr_info_once(PFX "overriding BIOS provided _PSD data\n");
721 	}
722 #endif
723 
724 	/* capability check */
725 	if (perf->state_count <= 1) {
726 		pr_debug("No P-States\n");
727 		result = -ENODEV;
728 		goto err_unreg;
729 	}
730 
731 	if (perf->control_register.space_id != perf->status_register.space_id) {
732 		result = -ENODEV;
733 		goto err_unreg;
734 	}
735 
736 	switch (perf->control_register.space_id) {
737 	case ACPI_ADR_SPACE_SYSTEM_IO:
738 		if (boot_cpu_data.x86_vendor == X86_VENDOR_AMD &&
739 		    boot_cpu_data.x86 == 0xf) {
740 			pr_debug("AMD K8 systems must use native drivers.\n");
741 			result = -ENODEV;
742 			goto err_unreg;
743 		}
744 		pr_debug("SYSTEM IO addr space\n");
745 		data->cpu_feature = SYSTEM_IO_CAPABLE;
746 		break;
747 	case ACPI_ADR_SPACE_FIXED_HARDWARE:
748 		pr_debug("HARDWARE addr space\n");
749 		if (check_est_cpu(cpu)) {
750 			data->cpu_feature = SYSTEM_INTEL_MSR_CAPABLE;
751 			break;
752 		}
753 		if (check_amd_hwpstate_cpu(cpu)) {
754 			data->cpu_feature = SYSTEM_AMD_MSR_CAPABLE;
755 			break;
756 		}
757 		result = -ENODEV;
758 		goto err_unreg;
759 	default:
760 		pr_debug("Unknown addr space %d\n",
761 			(u32) (perf->control_register.space_id));
762 		result = -ENODEV;
763 		goto err_unreg;
764 	}
765 
766 	data->freq_table = kzalloc(sizeof(*data->freq_table) *
767 		    (perf->state_count+1), GFP_KERNEL);
768 	if (!data->freq_table) {
769 		result = -ENOMEM;
770 		goto err_unreg;
771 	}
772 
773 	/* detect transition latency */
774 	policy->cpuinfo.transition_latency = 0;
775 	for (i = 0; i < perf->state_count; i++) {
776 		if ((perf->states[i].transition_latency * 1000) >
777 		    policy->cpuinfo.transition_latency)
778 			policy->cpuinfo.transition_latency =
779 			    perf->states[i].transition_latency * 1000;
780 	}
781 
782 	/* Check for high latency (>20uS) from buggy BIOSes, like on T42 */
783 	if (perf->control_register.space_id == ACPI_ADR_SPACE_FIXED_HARDWARE &&
784 	    policy->cpuinfo.transition_latency > 20 * 1000) {
785 		policy->cpuinfo.transition_latency = 20 * 1000;
786 		printk_once(KERN_INFO
787 			    "P-state transition latency capped at 20 uS\n");
788 	}
789 
790 	/* table init */
791 	for (i = 0; i < perf->state_count; i++) {
792 		if (i > 0 && perf->states[i].core_frequency >=
793 		    data->freq_table[valid_states-1].frequency / 1000)
794 			continue;
795 
796 		data->freq_table[valid_states].driver_data = i;
797 		data->freq_table[valid_states].frequency =
798 		    perf->states[i].core_frequency * 1000;
799 		valid_states++;
800 	}
801 	data->freq_table[valid_states].frequency = CPUFREQ_TABLE_END;
802 	perf->state = 0;
803 
804 	result = cpufreq_table_validate_and_show(policy, data->freq_table);
805 	if (result)
806 		goto err_freqfree;
807 
808 	if (perf->states[0].core_frequency * 1000 != policy->cpuinfo.max_freq)
809 		printk(KERN_WARNING FW_WARN "P-state 0 is not max freq\n");
810 
811 	switch (perf->control_register.space_id) {
812 	case ACPI_ADR_SPACE_SYSTEM_IO:
813 		/*
814 		 * The core will not set policy->cur, because
815 		 * cpufreq_driver->get is NULL, so we need to set it here.
816 		 * However, we have to guess it, because the current speed is
817 		 * unknown and not detectable via IO ports.
818 		 */
819 		policy->cur = acpi_cpufreq_guess_freq(data, policy->cpu);
820 		break;
821 	case ACPI_ADR_SPACE_FIXED_HARDWARE:
822 		acpi_cpufreq_driver.get = get_cur_freq_on_cpu;
823 		break;
824 	default:
825 		break;
826 	}
827 
828 	/* notify BIOS that we exist */
829 	acpi_processor_notify_smm(THIS_MODULE);
830 
831 	pr_debug("CPU%u - ACPI performance management activated.\n", cpu);
832 	for (i = 0; i < perf->state_count; i++)
833 		pr_debug("     %cP%d: %d MHz, %d mW, %d uS\n",
834 			(i == perf->state ? '*' : ' '), i,
835 			(u32) perf->states[i].core_frequency,
836 			(u32) perf->states[i].power,
837 			(u32) perf->states[i].transition_latency);
838 
839 	/*
840 	 * the first call to ->target() should result in us actually
841 	 * writing something to the appropriate registers.
842 	 */
843 	data->resume = 1;
844 
845 	return result;
846 
847 err_freqfree:
848 	kfree(data->freq_table);
849 err_unreg:
850 	acpi_processor_unregister_performance(cpu);
851 err_free_mask:
852 	free_cpumask_var(data->freqdomain_cpus);
853 err_free:
854 	kfree(data);
855 	policy->driver_data = NULL;
856 
857 	return result;
858 }
859 
860 static int acpi_cpufreq_cpu_exit(struct cpufreq_policy *policy)
861 {
862 	struct acpi_cpufreq_data *data = policy->driver_data;
863 
864 	pr_debug("acpi_cpufreq_cpu_exit\n");
865 
866 	if (data) {
867 		policy->driver_data = NULL;
868 		acpi_processor_unregister_performance(data->acpi_perf_cpu);
869 		free_cpumask_var(data->freqdomain_cpus);
870 		kfree(data->freq_table);
871 		kfree(data);
872 	}
873 
874 	return 0;
875 }
876 
877 static int acpi_cpufreq_resume(struct cpufreq_policy *policy)
878 {
879 	struct acpi_cpufreq_data *data = policy->driver_data;
880 
881 	pr_debug("acpi_cpufreq_resume\n");
882 
883 	data->resume = 1;
884 
885 	return 0;
886 }
887 
888 static struct freq_attr *acpi_cpufreq_attr[] = {
889 	&cpufreq_freq_attr_scaling_available_freqs,
890 	&freqdomain_cpus,
891 #ifdef CONFIG_X86_ACPI_CPUFREQ_CPB
892 	&cpb,
893 #endif
894 	NULL,
895 };
896 
897 static struct cpufreq_driver acpi_cpufreq_driver = {
898 	.verify		= cpufreq_generic_frequency_table_verify,
899 	.target_index	= acpi_cpufreq_target,
900 	.bios_limit	= acpi_processor_get_bios_limit,
901 	.init		= acpi_cpufreq_cpu_init,
902 	.exit		= acpi_cpufreq_cpu_exit,
903 	.resume		= acpi_cpufreq_resume,
904 	.name		= "acpi-cpufreq",
905 	.attr		= acpi_cpufreq_attr,
906 	.set_boost      = _store_boost,
907 };
908 
909 static void __init acpi_cpufreq_boost_init(void)
910 {
911 	if (boot_cpu_has(X86_FEATURE_CPB) || boot_cpu_has(X86_FEATURE_IDA)) {
912 		msrs = msrs_alloc();
913 
914 		if (!msrs)
915 			return;
916 
917 		acpi_cpufreq_driver.boost_supported = true;
918 		acpi_cpufreq_driver.boost_enabled = boost_state(0);
919 
920 		cpu_notifier_register_begin();
921 
922 		/* Force all MSRs to the same value */
923 		boost_set_msrs(acpi_cpufreq_driver.boost_enabled,
924 			       cpu_online_mask);
925 
926 		__register_cpu_notifier(&boost_nb);
927 
928 		cpu_notifier_register_done();
929 	}
930 }
931 
932 static void acpi_cpufreq_boost_exit(void)
933 {
934 	if (msrs) {
935 		unregister_cpu_notifier(&boost_nb);
936 
937 		msrs_free(msrs);
938 		msrs = NULL;
939 	}
940 }
941 
942 static int __init acpi_cpufreq_init(void)
943 {
944 	int ret;
945 
946 	if (acpi_disabled)
947 		return -ENODEV;
948 
949 	/* don't keep reloading if cpufreq_driver exists */
950 	if (cpufreq_get_current_driver())
951 		return -EEXIST;
952 
953 	pr_debug("acpi_cpufreq_init\n");
954 
955 	ret = acpi_cpufreq_early_init();
956 	if (ret)
957 		return ret;
958 
959 #ifdef CONFIG_X86_ACPI_CPUFREQ_CPB
960 	/* this is a sysfs file with a strange name and an even stranger
961 	 * semantic - per CPU instantiation, but system global effect.
962 	 * Lets enable it only on AMD CPUs for compatibility reasons and
963 	 * only if configured. This is considered legacy code, which
964 	 * will probably be removed at some point in the future.
965 	 */
966 	if (!check_amd_hwpstate_cpu(0)) {
967 		struct freq_attr **attr;
968 
969 		pr_debug("CPB unsupported, do not expose it\n");
970 
971 		for (attr = acpi_cpufreq_attr; *attr; attr++)
972 			if (*attr == &cpb) {
973 				*attr = NULL;
974 				break;
975 			}
976 	}
977 #endif
978 	acpi_cpufreq_boost_init();
979 
980 	ret = cpufreq_register_driver(&acpi_cpufreq_driver);
981 	if (ret) {
982 		free_acpi_perf_data();
983 		acpi_cpufreq_boost_exit();
984 	}
985 	return ret;
986 }
987 
988 static void __exit acpi_cpufreq_exit(void)
989 {
990 	pr_debug("acpi_cpufreq_exit\n");
991 
992 	acpi_cpufreq_boost_exit();
993 
994 	cpufreq_unregister_driver(&acpi_cpufreq_driver);
995 
996 	free_acpi_perf_data();
997 }
998 
999 module_param(acpi_pstate_strict, uint, 0644);
1000 MODULE_PARM_DESC(acpi_pstate_strict,
1001 	"value 0 or non-zero. non-zero -> strict ACPI checks are "
1002 	"performed during frequency changes.");
1003 
1004 late_initcall(acpi_cpufreq_init);
1005 module_exit(acpi_cpufreq_exit);
1006 
1007 static const struct x86_cpu_id acpi_cpufreq_ids[] = {
1008 	X86_FEATURE_MATCH(X86_FEATURE_ACPI),
1009 	X86_FEATURE_MATCH(X86_FEATURE_HW_PSTATE),
1010 	{}
1011 };
1012 MODULE_DEVICE_TABLE(x86cpu, acpi_cpufreq_ids);
1013 
1014 static const struct acpi_device_id processor_device_ids[] = {
1015 	{ACPI_PROCESSOR_OBJECT_HID, },
1016 	{ACPI_PROCESSOR_DEVICE_HID, },
1017 	{},
1018 };
1019 MODULE_DEVICE_TABLE(acpi, processor_device_ids);
1020 
1021 MODULE_ALIAS("acpi");
1022