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 #include "mperf.h"
49 
50 MODULE_AUTHOR("Paul Diefenbaugh, Dominik Brodowski");
51 MODULE_DESCRIPTION("ACPI Processor P-States Driver");
52 MODULE_LICENSE("GPL");
53 
54 #define PFX "acpi-cpufreq: "
55 
56 enum {
57 	UNDEFINED_CAPABLE = 0,
58 	SYSTEM_INTEL_MSR_CAPABLE,
59 	SYSTEM_AMD_MSR_CAPABLE,
60 	SYSTEM_IO_CAPABLE,
61 };
62 
63 #define INTEL_MSR_RANGE		(0xffff)
64 #define AMD_MSR_RANGE		(0x7)
65 
66 #define MSR_K7_HWCR_CPB_DIS	(1ULL << 25)
67 
68 struct acpi_cpufreq_data {
69 	struct acpi_processor_performance *acpi_data;
70 	struct cpufreq_frequency_table *freq_table;
71 	unsigned int resume;
72 	unsigned int cpu_feature;
73 };
74 
75 static DEFINE_PER_CPU(struct acpi_cpufreq_data *, acfreq_data);
76 
77 /* acpi_perf_data is a pointer to percpu data. */
78 static struct acpi_processor_performance __percpu *acpi_perf_data;
79 
80 static struct cpufreq_driver acpi_cpufreq_driver;
81 
82 static unsigned int acpi_pstate_strict;
83 static bool boost_enabled, boost_supported;
84 static struct msr __percpu *msrs;
85 
86 static bool boost_state(unsigned int cpu)
87 {
88 	u32 lo, hi;
89 	u64 msr;
90 
91 	switch (boot_cpu_data.x86_vendor) {
92 	case X86_VENDOR_INTEL:
93 		rdmsr_on_cpu(cpu, MSR_IA32_MISC_ENABLE, &lo, &hi);
94 		msr = lo | ((u64)hi << 32);
95 		return !(msr & MSR_IA32_MISC_ENABLE_TURBO_DISABLE);
96 	case X86_VENDOR_AMD:
97 		rdmsr_on_cpu(cpu, MSR_K7_HWCR, &lo, &hi);
98 		msr = lo | ((u64)hi << 32);
99 		return !(msr & MSR_K7_HWCR_CPB_DIS);
100 	}
101 	return false;
102 }
103 
104 static void boost_set_msrs(bool enable, const struct cpumask *cpumask)
105 {
106 	u32 cpu;
107 	u32 msr_addr;
108 	u64 msr_mask;
109 
110 	switch (boot_cpu_data.x86_vendor) {
111 	case X86_VENDOR_INTEL:
112 		msr_addr = MSR_IA32_MISC_ENABLE;
113 		msr_mask = MSR_IA32_MISC_ENABLE_TURBO_DISABLE;
114 		break;
115 	case X86_VENDOR_AMD:
116 		msr_addr = MSR_K7_HWCR;
117 		msr_mask = MSR_K7_HWCR_CPB_DIS;
118 		break;
119 	default:
120 		return;
121 	}
122 
123 	rdmsr_on_cpus(cpumask, msr_addr, msrs);
124 
125 	for_each_cpu(cpu, cpumask) {
126 		struct msr *reg = per_cpu_ptr(msrs, cpu);
127 		if (enable)
128 			reg->q &= ~msr_mask;
129 		else
130 			reg->q |= msr_mask;
131 	}
132 
133 	wrmsr_on_cpus(cpumask, msr_addr, msrs);
134 }
135 
136 static ssize_t _store_boost(const char *buf, size_t count)
137 {
138 	int ret;
139 	unsigned long val = 0;
140 
141 	if (!boost_supported)
142 		return -EINVAL;
143 
144 	ret = kstrtoul(buf, 10, &val);
145 	if (ret || (val > 1))
146 		return -EINVAL;
147 
148 	if ((val && boost_enabled) || (!val && !boost_enabled))
149 		return count;
150 
151 	get_online_cpus();
152 
153 	boost_set_msrs(val, cpu_online_mask);
154 
155 	put_online_cpus();
156 
157 	boost_enabled = val;
158 	pr_debug("Core Boosting %sabled.\n", val ? "en" : "dis");
159 
160 	return count;
161 }
162 
163 static ssize_t store_global_boost(struct kobject *kobj, struct attribute *attr,
164 				  const char *buf, size_t count)
165 {
166 	return _store_boost(buf, count);
167 }
168 
169 static ssize_t show_global_boost(struct kobject *kobj,
170 				 struct attribute *attr, char *buf)
171 {
172 	return sprintf(buf, "%u\n", boost_enabled);
173 }
174 
175 static struct global_attr global_boost = __ATTR(boost, 0644,
176 						show_global_boost,
177 						store_global_boost);
178 
179 #ifdef CONFIG_X86_ACPI_CPUFREQ_CPB
180 static ssize_t store_cpb(struct cpufreq_policy *policy, const char *buf,
181 			 size_t count)
182 {
183 	return _store_boost(buf, count);
184 }
185 
186 static ssize_t show_cpb(struct cpufreq_policy *policy, char *buf)
187 {
188 	return sprintf(buf, "%u\n", boost_enabled);
189 }
190 
191 static struct freq_attr cpb = __ATTR(cpb, 0644, show_cpb, store_cpb);
192 #endif
193 
194 static int check_est_cpu(unsigned int cpuid)
195 {
196 	struct cpuinfo_x86 *cpu = &cpu_data(cpuid);
197 
198 	return cpu_has(cpu, X86_FEATURE_EST);
199 }
200 
201 static int check_amd_hwpstate_cpu(unsigned int cpuid)
202 {
203 	struct cpuinfo_x86 *cpu = &cpu_data(cpuid);
204 
205 	return cpu_has(cpu, X86_FEATURE_HW_PSTATE);
206 }
207 
208 static unsigned extract_io(u32 value, struct acpi_cpufreq_data *data)
209 {
210 	struct acpi_processor_performance *perf;
211 	int i;
212 
213 	perf = data->acpi_data;
214 
215 	for (i = 0; i < perf->state_count; i++) {
216 		if (value == perf->states[i].status)
217 			return data->freq_table[i].frequency;
218 	}
219 	return 0;
220 }
221 
222 static unsigned extract_msr(u32 msr, struct acpi_cpufreq_data *data)
223 {
224 	int i;
225 	struct acpi_processor_performance *perf;
226 
227 	if (boot_cpu_data.x86_vendor == X86_VENDOR_AMD)
228 		msr &= AMD_MSR_RANGE;
229 	else
230 		msr &= INTEL_MSR_RANGE;
231 
232 	perf = data->acpi_data;
233 
234 	for (i = 0; data->freq_table[i].frequency != CPUFREQ_TABLE_END; i++) {
235 		if (msr == perf->states[data->freq_table[i].index].status)
236 			return data->freq_table[i].frequency;
237 	}
238 	return data->freq_table[0].frequency;
239 }
240 
241 static unsigned extract_freq(u32 val, struct acpi_cpufreq_data *data)
242 {
243 	switch (data->cpu_feature) {
244 	case SYSTEM_INTEL_MSR_CAPABLE:
245 	case SYSTEM_AMD_MSR_CAPABLE:
246 		return extract_msr(val, data);
247 	case SYSTEM_IO_CAPABLE:
248 		return extract_io(val, data);
249 	default:
250 		return 0;
251 	}
252 }
253 
254 struct msr_addr {
255 	u32 reg;
256 };
257 
258 struct io_addr {
259 	u16 port;
260 	u8 bit_width;
261 };
262 
263 struct drv_cmd {
264 	unsigned int type;
265 	const struct cpumask *mask;
266 	union {
267 		struct msr_addr msr;
268 		struct io_addr io;
269 	} addr;
270 	u32 val;
271 };
272 
273 /* Called via smp_call_function_single(), on the target CPU */
274 static void do_drv_read(void *_cmd)
275 {
276 	struct drv_cmd *cmd = _cmd;
277 	u32 h;
278 
279 	switch (cmd->type) {
280 	case SYSTEM_INTEL_MSR_CAPABLE:
281 	case SYSTEM_AMD_MSR_CAPABLE:
282 		rdmsr(cmd->addr.msr.reg, cmd->val, h);
283 		break;
284 	case SYSTEM_IO_CAPABLE:
285 		acpi_os_read_port((acpi_io_address)cmd->addr.io.port,
286 				&cmd->val,
287 				(u32)cmd->addr.io.bit_width);
288 		break;
289 	default:
290 		break;
291 	}
292 }
293 
294 /* Called via smp_call_function_many(), on the target CPUs */
295 static void do_drv_write(void *_cmd)
296 {
297 	struct drv_cmd *cmd = _cmd;
298 	u32 lo, hi;
299 
300 	switch (cmd->type) {
301 	case SYSTEM_INTEL_MSR_CAPABLE:
302 		rdmsr(cmd->addr.msr.reg, lo, hi);
303 		lo = (lo & ~INTEL_MSR_RANGE) | (cmd->val & INTEL_MSR_RANGE);
304 		wrmsr(cmd->addr.msr.reg, lo, hi);
305 		break;
306 	case SYSTEM_AMD_MSR_CAPABLE:
307 		wrmsr(cmd->addr.msr.reg, cmd->val, 0);
308 		break;
309 	case SYSTEM_IO_CAPABLE:
310 		acpi_os_write_port((acpi_io_address)cmd->addr.io.port,
311 				cmd->val,
312 				(u32)cmd->addr.io.bit_width);
313 		break;
314 	default:
315 		break;
316 	}
317 }
318 
319 static void drv_read(struct drv_cmd *cmd)
320 {
321 	int err;
322 	cmd->val = 0;
323 
324 	err = smp_call_function_any(cmd->mask, do_drv_read, cmd, 1);
325 	WARN_ON_ONCE(err);	/* smp_call_function_any() was buggy? */
326 }
327 
328 static void drv_write(struct drv_cmd *cmd)
329 {
330 	int this_cpu;
331 
332 	this_cpu = get_cpu();
333 	if (cpumask_test_cpu(this_cpu, cmd->mask))
334 		do_drv_write(cmd);
335 	smp_call_function_many(cmd->mask, do_drv_write, cmd, 1);
336 	put_cpu();
337 }
338 
339 static u32 get_cur_val(const struct cpumask *mask)
340 {
341 	struct acpi_processor_performance *perf;
342 	struct drv_cmd cmd;
343 
344 	if (unlikely(cpumask_empty(mask)))
345 		return 0;
346 
347 	switch (per_cpu(acfreq_data, cpumask_first(mask))->cpu_feature) {
348 	case SYSTEM_INTEL_MSR_CAPABLE:
349 		cmd.type = SYSTEM_INTEL_MSR_CAPABLE;
350 		cmd.addr.msr.reg = MSR_IA32_PERF_STATUS;
351 		break;
352 	case SYSTEM_AMD_MSR_CAPABLE:
353 		cmd.type = SYSTEM_AMD_MSR_CAPABLE;
354 		cmd.addr.msr.reg = MSR_AMD_PERF_STATUS;
355 		break;
356 	case SYSTEM_IO_CAPABLE:
357 		cmd.type = SYSTEM_IO_CAPABLE;
358 		perf = per_cpu(acfreq_data, cpumask_first(mask))->acpi_data;
359 		cmd.addr.io.port = perf->control_register.address;
360 		cmd.addr.io.bit_width = perf->control_register.bit_width;
361 		break;
362 	default:
363 		return 0;
364 	}
365 
366 	cmd.mask = mask;
367 	drv_read(&cmd);
368 
369 	pr_debug("get_cur_val = %u\n", cmd.val);
370 
371 	return cmd.val;
372 }
373 
374 static unsigned int get_cur_freq_on_cpu(unsigned int cpu)
375 {
376 	struct acpi_cpufreq_data *data = per_cpu(acfreq_data, cpu);
377 	unsigned int freq;
378 	unsigned int cached_freq;
379 
380 	pr_debug("get_cur_freq_on_cpu (%d)\n", cpu);
381 
382 	if (unlikely(data == NULL ||
383 		     data->acpi_data == NULL || data->freq_table == NULL)) {
384 		return 0;
385 	}
386 
387 	cached_freq = data->freq_table[data->acpi_data->state].frequency;
388 	freq = extract_freq(get_cur_val(cpumask_of(cpu)), 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);
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 target_freq, unsigned int relation)
419 {
420 	struct acpi_cpufreq_data *data = per_cpu(acfreq_data, policy->cpu);
421 	struct acpi_processor_performance *perf;
422 	struct cpufreq_freqs freqs;
423 	struct drv_cmd cmd;
424 	unsigned int next_state = 0; /* Index into freq_table */
425 	unsigned int next_perf_state = 0; /* Index into perf table */
426 	unsigned int i;
427 	int result = 0;
428 
429 	pr_debug("acpi_cpufreq_target %d (%d)\n", target_freq, policy->cpu);
430 
431 	if (unlikely(data == NULL ||
432 	     data->acpi_data == NULL || data->freq_table == NULL)) {
433 		return -ENODEV;
434 	}
435 
436 	perf = data->acpi_data;
437 	result = cpufreq_frequency_table_target(policy,
438 						data->freq_table,
439 						target_freq,
440 						relation, &next_state);
441 	if (unlikely(result)) {
442 		result = -ENODEV;
443 		goto out;
444 	}
445 
446 	next_perf_state = data->freq_table[next_state].index;
447 	if (perf->state == next_perf_state) {
448 		if (unlikely(data->resume)) {
449 			pr_debug("Called after resume, resetting to P%d\n",
450 				next_perf_state);
451 			data->resume = 0;
452 		} else {
453 			pr_debug("Already at target state (P%d)\n",
454 				next_perf_state);
455 			goto out;
456 		}
457 	}
458 
459 	switch (data->cpu_feature) {
460 	case SYSTEM_INTEL_MSR_CAPABLE:
461 		cmd.type = SYSTEM_INTEL_MSR_CAPABLE;
462 		cmd.addr.msr.reg = MSR_IA32_PERF_CTL;
463 		cmd.val = (u32) perf->states[next_perf_state].control;
464 		break;
465 	case SYSTEM_AMD_MSR_CAPABLE:
466 		cmd.type = SYSTEM_AMD_MSR_CAPABLE;
467 		cmd.addr.msr.reg = MSR_AMD_PERF_CTL;
468 		cmd.val = (u32) perf->states[next_perf_state].control;
469 		break;
470 	case SYSTEM_IO_CAPABLE:
471 		cmd.type = SYSTEM_IO_CAPABLE;
472 		cmd.addr.io.port = perf->control_register.address;
473 		cmd.addr.io.bit_width = perf->control_register.bit_width;
474 		cmd.val = (u32) perf->states[next_perf_state].control;
475 		break;
476 	default:
477 		result = -ENODEV;
478 		goto out;
479 	}
480 
481 	/* cpufreq holds the hotplug lock, so we are safe from here on */
482 	if (policy->shared_type != CPUFREQ_SHARED_TYPE_ANY)
483 		cmd.mask = policy->cpus;
484 	else
485 		cmd.mask = cpumask_of(policy->cpu);
486 
487 	freqs.old = perf->states[perf->state].core_frequency * 1000;
488 	freqs.new = data->freq_table[next_state].frequency;
489 	for_each_cpu(i, policy->cpus) {
490 		freqs.cpu = i;
491 		cpufreq_notify_transition(&freqs, CPUFREQ_PRECHANGE);
492 	}
493 
494 	drv_write(&cmd);
495 
496 	if (acpi_pstate_strict) {
497 		if (!check_freqs(cmd.mask, freqs.new, data)) {
498 			pr_debug("acpi_cpufreq_target failed (%d)\n",
499 				policy->cpu);
500 			result = -EAGAIN;
501 			goto out;
502 		}
503 	}
504 
505 	for_each_cpu(i, policy->cpus) {
506 		freqs.cpu = i;
507 		cpufreq_notify_transition(&freqs, CPUFREQ_POSTCHANGE);
508 	}
509 	perf->state = next_perf_state;
510 
511 out:
512 	return result;
513 }
514 
515 static int acpi_cpufreq_verify(struct cpufreq_policy *policy)
516 {
517 	struct acpi_cpufreq_data *data = per_cpu(acfreq_data, policy->cpu);
518 
519 	pr_debug("acpi_cpufreq_verify\n");
520 
521 	return cpufreq_frequency_table_verify(policy, data->freq_table);
522 }
523 
524 static unsigned long
525 acpi_cpufreq_guess_freq(struct acpi_cpufreq_data *data, unsigned int cpu)
526 {
527 	struct acpi_processor_performance *perf = data->acpi_data;
528 
529 	if (cpu_khz) {
530 		/* search the closest match to cpu_khz */
531 		unsigned int i;
532 		unsigned long freq;
533 		unsigned long freqn = perf->states[0].core_frequency * 1000;
534 
535 		for (i = 0; i < (perf->state_count-1); i++) {
536 			freq = freqn;
537 			freqn = perf->states[i+1].core_frequency * 1000;
538 			if ((2 * cpu_khz) > (freqn + freq)) {
539 				perf->state = i;
540 				return freq;
541 			}
542 		}
543 		perf->state = perf->state_count-1;
544 		return freqn;
545 	} else {
546 		/* assume CPU is at P0... */
547 		perf->state = 0;
548 		return perf->states[0].core_frequency * 1000;
549 	}
550 }
551 
552 static void free_acpi_perf_data(void)
553 {
554 	unsigned int i;
555 
556 	/* Freeing a NULL pointer is OK, and alloc_percpu zeroes. */
557 	for_each_possible_cpu(i)
558 		free_cpumask_var(per_cpu_ptr(acpi_perf_data, i)
559 				 ->shared_cpu_map);
560 	free_percpu(acpi_perf_data);
561 }
562 
563 static int boost_notify(struct notifier_block *nb, unsigned long action,
564 		      void *hcpu)
565 {
566 	unsigned cpu = (long)hcpu;
567 	const struct cpumask *cpumask;
568 
569 	cpumask = get_cpu_mask(cpu);
570 
571 	/*
572 	 * Clear the boost-disable bit on the CPU_DOWN path so that
573 	 * this cpu cannot block the remaining ones from boosting. On
574 	 * the CPU_UP path we simply keep the boost-disable flag in
575 	 * sync with the current global state.
576 	 */
577 
578 	switch (action) {
579 	case CPU_UP_PREPARE:
580 	case CPU_UP_PREPARE_FROZEN:
581 		boost_set_msrs(boost_enabled, cpumask);
582 		break;
583 
584 	case CPU_DOWN_PREPARE:
585 	case CPU_DOWN_PREPARE_FROZEN:
586 		boost_set_msrs(1, cpumask);
587 		break;
588 
589 	default:
590 		break;
591 	}
592 
593 	return NOTIFY_OK;
594 }
595 
596 
597 static struct notifier_block boost_nb = {
598 	.notifier_call          = boost_notify,
599 };
600 
601 /*
602  * acpi_cpufreq_early_init - initialize ACPI P-States library
603  *
604  * Initialize the ACPI P-States library (drivers/acpi/processor_perflib.c)
605  * in order to determine correct frequency and voltage pairings. We can
606  * do _PDC and _PSD and find out the processor dependency for the
607  * actual init that will happen later...
608  */
609 static int __init acpi_cpufreq_early_init(void)
610 {
611 	unsigned int i;
612 	pr_debug("acpi_cpufreq_early_init\n");
613 
614 	acpi_perf_data = alloc_percpu(struct acpi_processor_performance);
615 	if (!acpi_perf_data) {
616 		pr_debug("Memory allocation error for acpi_perf_data.\n");
617 		return -ENOMEM;
618 	}
619 	for_each_possible_cpu(i) {
620 		if (!zalloc_cpumask_var_node(
621 			&per_cpu_ptr(acpi_perf_data, i)->shared_cpu_map,
622 			GFP_KERNEL, cpu_to_node(i))) {
623 
624 			/* Freeing a NULL pointer is OK: alloc_percpu zeroes. */
625 			free_acpi_perf_data();
626 			return -ENOMEM;
627 		}
628 	}
629 
630 	/* Do initialization in ACPI core */
631 	acpi_processor_preregister_performance(acpi_perf_data);
632 	return 0;
633 }
634 
635 #ifdef CONFIG_SMP
636 /*
637  * Some BIOSes do SW_ANY coordination internally, either set it up in hw
638  * or do it in BIOS firmware and won't inform about it to OS. If not
639  * detected, this has a side effect of making CPU run at a different speed
640  * than OS intended it to run at. Detect it and handle it cleanly.
641  */
642 static int bios_with_sw_any_bug;
643 
644 static int sw_any_bug_found(const struct dmi_system_id *d)
645 {
646 	bios_with_sw_any_bug = 1;
647 	return 0;
648 }
649 
650 static const struct dmi_system_id sw_any_bug_dmi_table[] = {
651 	{
652 		.callback = sw_any_bug_found,
653 		.ident = "Supermicro Server X6DLP",
654 		.matches = {
655 			DMI_MATCH(DMI_SYS_VENDOR, "Supermicro"),
656 			DMI_MATCH(DMI_BIOS_VERSION, "080010"),
657 			DMI_MATCH(DMI_PRODUCT_NAME, "X6DLP"),
658 		},
659 	},
660 	{ }
661 };
662 
663 static int acpi_cpufreq_blacklist(struct cpuinfo_x86 *c)
664 {
665 	/* Intel Xeon Processor 7100 Series Specification Update
666 	 * http://www.intel.com/Assets/PDF/specupdate/314554.pdf
667 	 * AL30: A Machine Check Exception (MCE) Occurring during an
668 	 * Enhanced Intel SpeedStep Technology Ratio Change May Cause
669 	 * Both Processor Cores to Lock Up. */
670 	if (c->x86_vendor == X86_VENDOR_INTEL) {
671 		if ((c->x86 == 15) &&
672 		    (c->x86_model == 6) &&
673 		    (c->x86_mask == 8)) {
674 			printk(KERN_INFO "acpi-cpufreq: Intel(R) "
675 			    "Xeon(R) 7100 Errata AL30, processors may "
676 			    "lock up on frequency changes: disabling "
677 			    "acpi-cpufreq.\n");
678 			return -ENODEV;
679 		    }
680 		}
681 	return 0;
682 }
683 #endif
684 
685 static int acpi_cpufreq_cpu_init(struct cpufreq_policy *policy)
686 {
687 	unsigned int i;
688 	unsigned int valid_states = 0;
689 	unsigned int cpu = policy->cpu;
690 	struct acpi_cpufreq_data *data;
691 	unsigned int result = 0;
692 	struct cpuinfo_x86 *c = &cpu_data(policy->cpu);
693 	struct acpi_processor_performance *perf;
694 #ifdef CONFIG_SMP
695 	static int blacklisted;
696 #endif
697 
698 	pr_debug("acpi_cpufreq_cpu_init\n");
699 
700 #ifdef CONFIG_SMP
701 	if (blacklisted)
702 		return blacklisted;
703 	blacklisted = acpi_cpufreq_blacklist(c);
704 	if (blacklisted)
705 		return blacklisted;
706 #endif
707 
708 	data = kzalloc(sizeof(struct acpi_cpufreq_data), GFP_KERNEL);
709 	if (!data)
710 		return -ENOMEM;
711 
712 	data->acpi_data = per_cpu_ptr(acpi_perf_data, cpu);
713 	per_cpu(acfreq_data, cpu) = data;
714 
715 	if (cpu_has(c, X86_FEATURE_CONSTANT_TSC))
716 		acpi_cpufreq_driver.flags |= CPUFREQ_CONST_LOOPS;
717 
718 	result = acpi_processor_register_performance(data->acpi_data, cpu);
719 	if (result)
720 		goto err_free;
721 
722 	perf = data->acpi_data;
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(policy->related_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 && cpumask_weight(policy->cpus) == 1) {
738 		policy->shared_type = CPUFREQ_SHARED_TYPE_ALL;
739 		cpumask_copy(policy->cpus, cpu_core_mask(cpu));
740 	}
741 
742 	if (check_amd_hwpstate_cpu(cpu) && !acpi_pstate_strict) {
743 		cpumask_clear(policy->cpus);
744 		cpumask_set_cpu(cpu, policy->cpus);
745 		cpumask_copy(policy->related_cpus, cpu_sibling_mask(cpu));
746 		policy->shared_type = CPUFREQ_SHARED_TYPE_HW;
747 		pr_info_once(PFX "overriding BIOS provided _PSD data\n");
748 	}
749 #endif
750 
751 	/* capability check */
752 	if (perf->state_count <= 1) {
753 		pr_debug("No P-States\n");
754 		result = -ENODEV;
755 		goto err_unreg;
756 	}
757 
758 	if (perf->control_register.space_id != perf->status_register.space_id) {
759 		result = -ENODEV;
760 		goto err_unreg;
761 	}
762 
763 	switch (perf->control_register.space_id) {
764 	case ACPI_ADR_SPACE_SYSTEM_IO:
765 		pr_debug("SYSTEM IO addr space\n");
766 		data->cpu_feature = SYSTEM_IO_CAPABLE;
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 			break;
773 		}
774 		if (check_amd_hwpstate_cpu(cpu)) {
775 			data->cpu_feature = SYSTEM_AMD_MSR_CAPABLE;
776 			break;
777 		}
778 		result = -ENODEV;
779 		goto err_unreg;
780 	default:
781 		pr_debug("Unknown addr space %d\n",
782 			(u32) (perf->control_register.space_id));
783 		result = -ENODEV;
784 		goto err_unreg;
785 	}
786 
787 	data->freq_table = kmalloc(sizeof(struct cpufreq_frequency_table) *
788 		    (perf->state_count+1), GFP_KERNEL);
789 	if (!data->freq_table) {
790 		result = -ENOMEM;
791 		goto err_unreg;
792 	}
793 
794 	/* detect transition latency */
795 	policy->cpuinfo.transition_latency = 0;
796 	for (i = 0; i < perf->state_count; i++) {
797 		if ((perf->states[i].transition_latency * 1000) >
798 		    policy->cpuinfo.transition_latency)
799 			policy->cpuinfo.transition_latency =
800 			    perf->states[i].transition_latency * 1000;
801 	}
802 
803 	/* Check for high latency (>20uS) from buggy BIOSes, like on T42 */
804 	if (perf->control_register.space_id == ACPI_ADR_SPACE_FIXED_HARDWARE &&
805 	    policy->cpuinfo.transition_latency > 20 * 1000) {
806 		policy->cpuinfo.transition_latency = 20 * 1000;
807 		printk_once(KERN_INFO
808 			    "P-state transition latency capped at 20 uS\n");
809 	}
810 
811 	/* table init */
812 	for (i = 0; i < perf->state_count; i++) {
813 		if (i > 0 && perf->states[i].core_frequency >=
814 		    data->freq_table[valid_states-1].frequency / 1000)
815 			continue;
816 
817 		data->freq_table[valid_states].index = i;
818 		data->freq_table[valid_states].frequency =
819 		    perf->states[i].core_frequency * 1000;
820 		valid_states++;
821 	}
822 	data->freq_table[valid_states].frequency = CPUFREQ_TABLE_END;
823 	perf->state = 0;
824 
825 	result = cpufreq_frequency_table_cpuinfo(policy, data->freq_table);
826 	if (result)
827 		goto err_freqfree;
828 
829 	if (perf->states[0].core_frequency * 1000 != policy->cpuinfo.max_freq)
830 		printk(KERN_WARNING FW_WARN "P-state 0 is not max freq\n");
831 
832 	switch (perf->control_register.space_id) {
833 	case ACPI_ADR_SPACE_SYSTEM_IO:
834 		/* Current speed is unknown and not detectable by IO port */
835 		policy->cur = acpi_cpufreq_guess_freq(data, policy->cpu);
836 		break;
837 	case ACPI_ADR_SPACE_FIXED_HARDWARE:
838 		acpi_cpufreq_driver.get = get_cur_freq_on_cpu;
839 		policy->cur = get_cur_freq_on_cpu(cpu);
840 		break;
841 	default:
842 		break;
843 	}
844 
845 	/* notify BIOS that we exist */
846 	acpi_processor_notify_smm(THIS_MODULE);
847 
848 	/* Check for APERF/MPERF support in hardware */
849 	if (boot_cpu_has(X86_FEATURE_APERFMPERF))
850 		acpi_cpufreq_driver.getavg = cpufreq_get_measured_perf;
851 
852 	pr_debug("CPU%u - ACPI performance management activated.\n", cpu);
853 	for (i = 0; i < perf->state_count; i++)
854 		pr_debug("     %cP%d: %d MHz, %d mW, %d uS\n",
855 			(i == perf->state ? '*' : ' '), i,
856 			(u32) perf->states[i].core_frequency,
857 			(u32) perf->states[i].power,
858 			(u32) perf->states[i].transition_latency);
859 
860 	cpufreq_frequency_table_get_attr(data->freq_table, policy->cpu);
861 
862 	/*
863 	 * the first call to ->target() should result in us actually
864 	 * writing something to the appropriate registers.
865 	 */
866 	data->resume = 1;
867 
868 	return result;
869 
870 err_freqfree:
871 	kfree(data->freq_table);
872 err_unreg:
873 	acpi_processor_unregister_performance(perf, cpu);
874 err_free:
875 	kfree(data);
876 	per_cpu(acfreq_data, cpu) = NULL;
877 
878 	return result;
879 }
880 
881 static int acpi_cpufreq_cpu_exit(struct cpufreq_policy *policy)
882 {
883 	struct acpi_cpufreq_data *data = per_cpu(acfreq_data, policy->cpu);
884 
885 	pr_debug("acpi_cpufreq_cpu_exit\n");
886 
887 	if (data) {
888 		cpufreq_frequency_table_put_attr(policy->cpu);
889 		per_cpu(acfreq_data, policy->cpu) = NULL;
890 		acpi_processor_unregister_performance(data->acpi_data,
891 						      policy->cpu);
892 		kfree(data->freq_table);
893 		kfree(data);
894 	}
895 
896 	return 0;
897 }
898 
899 static int acpi_cpufreq_resume(struct cpufreq_policy *policy)
900 {
901 	struct acpi_cpufreq_data *data = per_cpu(acfreq_data, policy->cpu);
902 
903 	pr_debug("acpi_cpufreq_resume\n");
904 
905 	data->resume = 1;
906 
907 	return 0;
908 }
909 
910 static struct freq_attr *acpi_cpufreq_attr[] = {
911 	&cpufreq_freq_attr_scaling_available_freqs,
912 	NULL,	/* this is a placeholder for cpb, do not remove */
913 	NULL,
914 };
915 
916 static struct cpufreq_driver acpi_cpufreq_driver = {
917 	.verify		= acpi_cpufreq_verify,
918 	.target		= acpi_cpufreq_target,
919 	.bios_limit	= acpi_processor_get_bios_limit,
920 	.init		= acpi_cpufreq_cpu_init,
921 	.exit		= acpi_cpufreq_cpu_exit,
922 	.resume		= acpi_cpufreq_resume,
923 	.name		= "acpi-cpufreq",
924 	.owner		= THIS_MODULE,
925 	.attr		= acpi_cpufreq_attr,
926 };
927 
928 static void __init acpi_cpufreq_boost_init(void)
929 {
930 	if (boot_cpu_has(X86_FEATURE_CPB) || boot_cpu_has(X86_FEATURE_IDA)) {
931 		msrs = msrs_alloc();
932 
933 		if (!msrs)
934 			return;
935 
936 		boost_supported = true;
937 		boost_enabled = boost_state(0);
938 
939 		get_online_cpus();
940 
941 		/* Force all MSRs to the same value */
942 		boost_set_msrs(boost_enabled, cpu_online_mask);
943 
944 		register_cpu_notifier(&boost_nb);
945 
946 		put_online_cpus();
947 	} else
948 		global_boost.attr.mode = 0444;
949 
950 	/* We create the boost file in any case, though for systems without
951 	 * hardware support it will be read-only and hardwired to return 0.
952 	 */
953 	if (sysfs_create_file(cpufreq_global_kobject, &(global_boost.attr)))
954 		pr_warn(PFX "could not register global boost sysfs file\n");
955 	else
956 		pr_debug("registered global boost sysfs file\n");
957 }
958 
959 static void __exit acpi_cpufreq_boost_exit(void)
960 {
961 	sysfs_remove_file(cpufreq_global_kobject, &(global_boost.attr));
962 
963 	if (msrs) {
964 		unregister_cpu_notifier(&boost_nb);
965 
966 		msrs_free(msrs);
967 		msrs = NULL;
968 	}
969 }
970 
971 static int __init acpi_cpufreq_init(void)
972 {
973 	int ret;
974 
975 	if (acpi_disabled)
976 		return 0;
977 
978 	pr_debug("acpi_cpufreq_init\n");
979 
980 	ret = acpi_cpufreq_early_init();
981 	if (ret)
982 		return ret;
983 
984 #ifdef CONFIG_X86_ACPI_CPUFREQ_CPB
985 	/* this is a sysfs file with a strange name and an even stranger
986 	 * semantic - per CPU instantiation, but system global effect.
987 	 * Lets enable it only on AMD CPUs for compatibility reasons and
988 	 * only if configured. This is considered legacy code, which
989 	 * will probably be removed at some point in the future.
990 	 */
991 	if (check_amd_hwpstate_cpu(0)) {
992 		struct freq_attr **iter;
993 
994 		pr_debug("adding sysfs entry for cpb\n");
995 
996 		for (iter = acpi_cpufreq_attr; *iter != NULL; iter++)
997 			;
998 
999 		/* make sure there is a terminator behind it */
1000 		if (iter[1] == NULL)
1001 			*iter = &cpb;
1002 	}
1003 #endif
1004 
1005 	ret = cpufreq_register_driver(&acpi_cpufreq_driver);
1006 	if (ret)
1007 		free_acpi_perf_data();
1008 	else
1009 		acpi_cpufreq_boost_init();
1010 
1011 	return ret;
1012 }
1013 
1014 static void __exit acpi_cpufreq_exit(void)
1015 {
1016 	pr_debug("acpi_cpufreq_exit\n");
1017 
1018 	acpi_cpufreq_boost_exit();
1019 
1020 	cpufreq_unregister_driver(&acpi_cpufreq_driver);
1021 
1022 	free_acpi_perf_data();
1023 }
1024 
1025 module_param(acpi_pstate_strict, uint, 0644);
1026 MODULE_PARM_DESC(acpi_pstate_strict,
1027 	"value 0 or non-zero. non-zero -> strict ACPI checks are "
1028 	"performed during frequency changes.");
1029 
1030 late_initcall(acpi_cpufreq_init);
1031 module_exit(acpi_cpufreq_exit);
1032 
1033 MODULE_ALIAS("acpi");
1034