1 /*
2  * cpufreq driver for Enhanced SpeedStep, as found in Intel's Pentium
3  * M (part of the Centrino chipset).
4  *
5  * Since the original Pentium M, most new Intel CPUs support Enhanced
6  * SpeedStep.
7  *
8  * Despite the "SpeedStep" in the name, this is almost entirely unlike
9  * traditional SpeedStep.
10  *
11  * Modelled on speedstep.c
12  *
13  * Copyright (C) 2003 Jeremy Fitzhardinge <jeremy@goop.org>
14  */
15 
16 #include <linux/kernel.h>
17 #include <linux/module.h>
18 #include <linux/init.h>
19 #include <linux/cpufreq.h>
20 #include <linux/sched.h>	/* current */
21 #include <linux/delay.h>
22 #include <linux/compiler.h>
23 #include <linux/gfp.h>
24 
25 #include <asm/msr.h>
26 #include <asm/processor.h>
27 #include <asm/cpufeature.h>
28 #include <asm/cpu_device_id.h>
29 
30 #define PFX		"speedstep-centrino: "
31 #define MAINTAINER	"cpufreq@vger.kernel.org"
32 
33 #define INTEL_MSR_RANGE	(0xffff)
34 
35 struct cpu_id
36 {
37 	__u8	x86;            /* CPU family */
38 	__u8	x86_model;	/* model */
39 	__u8	x86_mask;	/* stepping */
40 };
41 
42 enum {
43 	CPU_BANIAS,
44 	CPU_DOTHAN_A1,
45 	CPU_DOTHAN_A2,
46 	CPU_DOTHAN_B0,
47 	CPU_MP4HT_D0,
48 	CPU_MP4HT_E0,
49 };
50 
51 static const struct cpu_id cpu_ids[] = {
52 	[CPU_BANIAS]	= { 6,  9, 5 },
53 	[CPU_DOTHAN_A1]	= { 6, 13, 1 },
54 	[CPU_DOTHAN_A2]	= { 6, 13, 2 },
55 	[CPU_DOTHAN_B0]	= { 6, 13, 6 },
56 	[CPU_MP4HT_D0]	= {15,  3, 4 },
57 	[CPU_MP4HT_E0]	= {15,  4, 1 },
58 };
59 #define N_IDS	ARRAY_SIZE(cpu_ids)
60 
61 struct cpu_model
62 {
63 	const struct cpu_id *cpu_id;
64 	const char	*model_name;
65 	unsigned	max_freq; /* max clock in kHz */
66 
67 	struct cpufreq_frequency_table *op_points; /* clock/voltage pairs */
68 };
69 static int centrino_verify_cpu_id(const struct cpuinfo_x86 *c,
70 				  const struct cpu_id *x);
71 
72 /* Operating points for current CPU */
73 static DEFINE_PER_CPU(struct cpu_model *, centrino_model);
74 static DEFINE_PER_CPU(const struct cpu_id *, centrino_cpu);
75 
76 static struct cpufreq_driver centrino_driver;
77 
78 #ifdef CONFIG_X86_SPEEDSTEP_CENTRINO_TABLE
79 
80 /* Computes the correct form for IA32_PERF_CTL MSR for a particular
81    frequency/voltage operating point; frequency in MHz, volts in mV.
82    This is stored as "driver_data" in the structure. */
83 #define OP(mhz, mv)							\
84 	{								\
85 		.frequency = (mhz) * 1000,				\
86 		.driver_data = (((mhz)/100) << 8) | ((mv - 700) / 16)		\
87 	}
88 
89 /*
90  * These voltage tables were derived from the Intel Pentium M
91  * datasheet, document 25261202.pdf, Table 5.  I have verified they
92  * are consistent with my IBM ThinkPad X31, which has a 1.3GHz Pentium
93  * M.
94  */
95 
96 /* Ultra Low Voltage Intel Pentium M processor 900MHz (Banias) */
97 static struct cpufreq_frequency_table banias_900[] =
98 {
99 	OP(600,  844),
100 	OP(800,  988),
101 	OP(900, 1004),
102 	{ .frequency = CPUFREQ_TABLE_END }
103 };
104 
105 /* Ultra Low Voltage Intel Pentium M processor 1000MHz (Banias) */
106 static struct cpufreq_frequency_table banias_1000[] =
107 {
108 	OP(600,   844),
109 	OP(800,   972),
110 	OP(900,   988),
111 	OP(1000, 1004),
112 	{ .frequency = CPUFREQ_TABLE_END }
113 };
114 
115 /* Low Voltage Intel Pentium M processor 1.10GHz (Banias) */
116 static struct cpufreq_frequency_table banias_1100[] =
117 {
118 	OP( 600,  956),
119 	OP( 800, 1020),
120 	OP( 900, 1100),
121 	OP(1000, 1164),
122 	OP(1100, 1180),
123 	{ .frequency = CPUFREQ_TABLE_END }
124 };
125 
126 
127 /* Low Voltage Intel Pentium M processor 1.20GHz (Banias) */
128 static struct cpufreq_frequency_table banias_1200[] =
129 {
130 	OP( 600,  956),
131 	OP( 800, 1004),
132 	OP( 900, 1020),
133 	OP(1000, 1100),
134 	OP(1100, 1164),
135 	OP(1200, 1180),
136 	{ .frequency = CPUFREQ_TABLE_END }
137 };
138 
139 /* Intel Pentium M processor 1.30GHz (Banias) */
140 static struct cpufreq_frequency_table banias_1300[] =
141 {
142 	OP( 600,  956),
143 	OP( 800, 1260),
144 	OP(1000, 1292),
145 	OP(1200, 1356),
146 	OP(1300, 1388),
147 	{ .frequency = CPUFREQ_TABLE_END }
148 };
149 
150 /* Intel Pentium M processor 1.40GHz (Banias) */
151 static struct cpufreq_frequency_table banias_1400[] =
152 {
153 	OP( 600,  956),
154 	OP( 800, 1180),
155 	OP(1000, 1308),
156 	OP(1200, 1436),
157 	OP(1400, 1484),
158 	{ .frequency = CPUFREQ_TABLE_END }
159 };
160 
161 /* Intel Pentium M processor 1.50GHz (Banias) */
162 static struct cpufreq_frequency_table banias_1500[] =
163 {
164 	OP( 600,  956),
165 	OP( 800, 1116),
166 	OP(1000, 1228),
167 	OP(1200, 1356),
168 	OP(1400, 1452),
169 	OP(1500, 1484),
170 	{ .frequency = CPUFREQ_TABLE_END }
171 };
172 
173 /* Intel Pentium M processor 1.60GHz (Banias) */
174 static struct cpufreq_frequency_table banias_1600[] =
175 {
176 	OP( 600,  956),
177 	OP( 800, 1036),
178 	OP(1000, 1164),
179 	OP(1200, 1276),
180 	OP(1400, 1420),
181 	OP(1600, 1484),
182 	{ .frequency = CPUFREQ_TABLE_END }
183 };
184 
185 /* Intel Pentium M processor 1.70GHz (Banias) */
186 static struct cpufreq_frequency_table banias_1700[] =
187 {
188 	OP( 600,  956),
189 	OP( 800, 1004),
190 	OP(1000, 1116),
191 	OP(1200, 1228),
192 	OP(1400, 1308),
193 	OP(1700, 1484),
194 	{ .frequency = CPUFREQ_TABLE_END }
195 };
196 #undef OP
197 
198 #define _BANIAS(cpuid, max, name)	\
199 {	.cpu_id		= cpuid,	\
200 	.model_name	= "Intel(R) Pentium(R) M processor " name "MHz", \
201 	.max_freq	= (max)*1000,	\
202 	.op_points	= banias_##max,	\
203 }
204 #define BANIAS(max)	_BANIAS(&cpu_ids[CPU_BANIAS], max, #max)
205 
206 /* CPU models, their operating frequency range, and freq/voltage
207    operating points */
208 static struct cpu_model models[] =
209 {
210 	_BANIAS(&cpu_ids[CPU_BANIAS], 900, " 900"),
211 	BANIAS(1000),
212 	BANIAS(1100),
213 	BANIAS(1200),
214 	BANIAS(1300),
215 	BANIAS(1400),
216 	BANIAS(1500),
217 	BANIAS(1600),
218 	BANIAS(1700),
219 
220 	/* NULL model_name is a wildcard */
221 	{ &cpu_ids[CPU_DOTHAN_A1], NULL, 0, NULL },
222 	{ &cpu_ids[CPU_DOTHAN_A2], NULL, 0, NULL },
223 	{ &cpu_ids[CPU_DOTHAN_B0], NULL, 0, NULL },
224 	{ &cpu_ids[CPU_MP4HT_D0], NULL, 0, NULL },
225 	{ &cpu_ids[CPU_MP4HT_E0], NULL, 0, NULL },
226 
227 	{ NULL, }
228 };
229 #undef _BANIAS
230 #undef BANIAS
231 
232 static int centrino_cpu_init_table(struct cpufreq_policy *policy)
233 {
234 	struct cpuinfo_x86 *cpu = &cpu_data(policy->cpu);
235 	struct cpu_model *model;
236 
237 	for(model = models; model->cpu_id != NULL; model++)
238 		if (centrino_verify_cpu_id(cpu, model->cpu_id) &&
239 		    (model->model_name == NULL ||
240 		     strcmp(cpu->x86_model_id, model->model_name) == 0))
241 			break;
242 
243 	if (model->cpu_id == NULL) {
244 		/* No match at all */
245 		pr_debug("no support for CPU model \"%s\": "
246 		       "send /proc/cpuinfo to " MAINTAINER "\n",
247 		       cpu->x86_model_id);
248 		return -ENOENT;
249 	}
250 
251 	if (model->op_points == NULL) {
252 		/* Matched a non-match */
253 		pr_debug("no table support for CPU model \"%s\"\n",
254 		       cpu->x86_model_id);
255 		pr_debug("try using the acpi-cpufreq driver\n");
256 		return -ENOENT;
257 	}
258 
259 	per_cpu(centrino_model, policy->cpu) = model;
260 
261 	pr_debug("found \"%s\": max frequency: %dkHz\n",
262 	       model->model_name, model->max_freq);
263 
264 	return 0;
265 }
266 
267 #else
268 static inline int centrino_cpu_init_table(struct cpufreq_policy *policy)
269 {
270 	return -ENODEV;
271 }
272 #endif /* CONFIG_X86_SPEEDSTEP_CENTRINO_TABLE */
273 
274 static int centrino_verify_cpu_id(const struct cpuinfo_x86 *c,
275 				  const struct cpu_id *x)
276 {
277 	if ((c->x86 == x->x86) &&
278 	    (c->x86_model == x->x86_model) &&
279 	    (c->x86_mask == x->x86_mask))
280 		return 1;
281 	return 0;
282 }
283 
284 /* To be called only after centrino_model is initialized */
285 static unsigned extract_clock(unsigned msr, unsigned int cpu, int failsafe)
286 {
287 	int i;
288 
289 	/*
290 	 * Extract clock in kHz from PERF_CTL value
291 	 * for centrino, as some DSDTs are buggy.
292 	 * Ideally, this can be done using the acpi_data structure.
293 	 */
294 	if ((per_cpu(centrino_cpu, cpu) == &cpu_ids[CPU_BANIAS]) ||
295 	    (per_cpu(centrino_cpu, cpu) == &cpu_ids[CPU_DOTHAN_A1]) ||
296 	    (per_cpu(centrino_cpu, cpu) == &cpu_ids[CPU_DOTHAN_B0])) {
297 		msr = (msr >> 8) & 0xff;
298 		return msr * 100000;
299 	}
300 
301 	if ((!per_cpu(centrino_model, cpu)) ||
302 	    (!per_cpu(centrino_model, cpu)->op_points))
303 		return 0;
304 
305 	msr &= 0xffff;
306 	for (i = 0;
307 		per_cpu(centrino_model, cpu)->op_points[i].frequency
308 							!= CPUFREQ_TABLE_END;
309 	     i++) {
310 		if (msr == per_cpu(centrino_model, cpu)->op_points[i].driver_data)
311 			return per_cpu(centrino_model, cpu)->
312 							op_points[i].frequency;
313 	}
314 	if (failsafe)
315 		return per_cpu(centrino_model, cpu)->op_points[i-1].frequency;
316 	else
317 		return 0;
318 }
319 
320 /* Return the current CPU frequency in kHz */
321 static unsigned int get_cur_freq(unsigned int cpu)
322 {
323 	unsigned l, h;
324 	unsigned clock_freq;
325 
326 	rdmsr_on_cpu(cpu, MSR_IA32_PERF_STATUS, &l, &h);
327 	clock_freq = extract_clock(l, cpu, 0);
328 
329 	if (unlikely(clock_freq == 0)) {
330 		/*
331 		 * On some CPUs, we can see transient MSR values (which are
332 		 * not present in _PSS), while CPU is doing some automatic
333 		 * P-state transition (like TM2). Get the last freq set
334 		 * in PERF_CTL.
335 		 */
336 		rdmsr_on_cpu(cpu, MSR_IA32_PERF_CTL, &l, &h);
337 		clock_freq = extract_clock(l, cpu, 1);
338 	}
339 	return clock_freq;
340 }
341 
342 
343 static int centrino_cpu_init(struct cpufreq_policy *policy)
344 {
345 	struct cpuinfo_x86 *cpu = &cpu_data(policy->cpu);
346 	unsigned freq;
347 	unsigned l, h;
348 	int ret;
349 	int i;
350 
351 	/* Only Intel makes Enhanced Speedstep-capable CPUs */
352 	if (cpu->x86_vendor != X86_VENDOR_INTEL ||
353 	    !cpu_has(cpu, X86_FEATURE_EST))
354 		return -ENODEV;
355 
356 	if (cpu_has(cpu, X86_FEATURE_CONSTANT_TSC))
357 		centrino_driver.flags |= CPUFREQ_CONST_LOOPS;
358 
359 	if (policy->cpu != 0)
360 		return -ENODEV;
361 
362 	for (i = 0; i < N_IDS; i++)
363 		if (centrino_verify_cpu_id(cpu, &cpu_ids[i]))
364 			break;
365 
366 	if (i != N_IDS)
367 		per_cpu(centrino_cpu, policy->cpu) = &cpu_ids[i];
368 
369 	if (!per_cpu(centrino_cpu, policy->cpu)) {
370 		pr_debug("found unsupported CPU with "
371 		"Enhanced SpeedStep: send /proc/cpuinfo to "
372 		MAINTAINER "\n");
373 		return -ENODEV;
374 	}
375 
376 	if (centrino_cpu_init_table(policy)) {
377 		return -ENODEV;
378 	}
379 
380 	/* Check to see if Enhanced SpeedStep is enabled, and try to
381 	   enable it if not. */
382 	rdmsr(MSR_IA32_MISC_ENABLE, l, h);
383 
384 	if (!(l & MSR_IA32_MISC_ENABLE_ENHANCED_SPEEDSTEP)) {
385 		l |= MSR_IA32_MISC_ENABLE_ENHANCED_SPEEDSTEP;
386 		pr_debug("trying to enable Enhanced SpeedStep (%x)\n", l);
387 		wrmsr(MSR_IA32_MISC_ENABLE, l, h);
388 
389 		/* check to see if it stuck */
390 		rdmsr(MSR_IA32_MISC_ENABLE, l, h);
391 		if (!(l & MSR_IA32_MISC_ENABLE_ENHANCED_SPEEDSTEP)) {
392 			printk(KERN_INFO PFX
393 				"couldn't enable Enhanced SpeedStep\n");
394 			return -ENODEV;
395 		}
396 	}
397 
398 	freq = get_cur_freq(policy->cpu);
399 	policy->cpuinfo.transition_latency = 10000;
400 						/* 10uS transition latency */
401 	policy->cur = freq;
402 
403 	pr_debug("centrino_cpu_init: cur=%dkHz\n", policy->cur);
404 
405 	ret = cpufreq_frequency_table_cpuinfo(policy,
406 		per_cpu(centrino_model, policy->cpu)->op_points);
407 	if (ret)
408 		return (ret);
409 
410 	cpufreq_frequency_table_get_attr(
411 		per_cpu(centrino_model, policy->cpu)->op_points, policy->cpu);
412 
413 	return 0;
414 }
415 
416 static int centrino_cpu_exit(struct cpufreq_policy *policy)
417 {
418 	unsigned int cpu = policy->cpu;
419 
420 	if (!per_cpu(centrino_model, cpu))
421 		return -ENODEV;
422 
423 	cpufreq_frequency_table_put_attr(cpu);
424 
425 	per_cpu(centrino_model, cpu) = NULL;
426 
427 	return 0;
428 }
429 
430 /**
431  * centrino_verify - verifies a new CPUFreq policy
432  * @policy: new policy
433  *
434  * Limit must be within this model's frequency range at least one
435  * border included.
436  */
437 static int centrino_verify (struct cpufreq_policy *policy)
438 {
439 	return cpufreq_frequency_table_verify(policy,
440 			per_cpu(centrino_model, policy->cpu)->op_points);
441 }
442 
443 /**
444  * centrino_setpolicy - set a new CPUFreq policy
445  * @policy: new policy
446  * @target_freq: the target frequency
447  * @relation: how that frequency relates to achieved frequency
448  *	(CPUFREQ_RELATION_L or CPUFREQ_RELATION_H)
449  *
450  * Sets a new CPUFreq policy.
451  */
452 static int centrino_target (struct cpufreq_policy *policy,
453 			    unsigned int target_freq,
454 			    unsigned int relation)
455 {
456 	unsigned int    newstate = 0;
457 	unsigned int	msr, oldmsr = 0, h = 0, cpu = policy->cpu;
458 	struct cpufreq_freqs	freqs;
459 	int			retval = 0;
460 	unsigned int		j, first_cpu, tmp;
461 	cpumask_var_t covered_cpus;
462 
463 	if (unlikely(!zalloc_cpumask_var(&covered_cpus, GFP_KERNEL)))
464 		return -ENOMEM;
465 
466 	if (unlikely(per_cpu(centrino_model, cpu) == NULL)) {
467 		retval = -ENODEV;
468 		goto out;
469 	}
470 
471 	if (unlikely(cpufreq_frequency_table_target(policy,
472 			per_cpu(centrino_model, cpu)->op_points,
473 			target_freq,
474 			relation,
475 			&newstate))) {
476 		retval = -EINVAL;
477 		goto out;
478 	}
479 
480 	first_cpu = 1;
481 	for_each_cpu(j, policy->cpus) {
482 		int good_cpu;
483 
484 		/*
485 		 * Support for SMP systems.
486 		 * Make sure we are running on CPU that wants to change freq
487 		 */
488 		if (policy->shared_type == CPUFREQ_SHARED_TYPE_ANY)
489 			good_cpu = cpumask_any_and(policy->cpus,
490 						   cpu_online_mask);
491 		else
492 			good_cpu = j;
493 
494 		if (good_cpu >= nr_cpu_ids) {
495 			pr_debug("couldn't limit to CPUs in this domain\n");
496 			retval = -EAGAIN;
497 			if (first_cpu) {
498 				/* We haven't started the transition yet. */
499 				goto out;
500 			}
501 			break;
502 		}
503 
504 		msr = per_cpu(centrino_model, cpu)->op_points[newstate].driver_data;
505 
506 		if (first_cpu) {
507 			rdmsr_on_cpu(good_cpu, MSR_IA32_PERF_CTL, &oldmsr, &h);
508 			if (msr == (oldmsr & 0xffff)) {
509 				pr_debug("no change needed - msr was and needs "
510 					"to be %x\n", oldmsr);
511 				retval = 0;
512 				goto out;
513 			}
514 
515 			freqs.old = extract_clock(oldmsr, cpu, 0);
516 			freqs.new = extract_clock(msr, cpu, 0);
517 
518 			pr_debug("target=%dkHz old=%d new=%d msr=%04x\n",
519 				target_freq, freqs.old, freqs.new, msr);
520 
521 			cpufreq_notify_transition(policy, &freqs,
522 					CPUFREQ_PRECHANGE);
523 
524 			first_cpu = 0;
525 			/* all but 16 LSB are reserved, treat them with care */
526 			oldmsr &= ~0xffff;
527 			msr &= 0xffff;
528 			oldmsr |= msr;
529 		}
530 
531 		wrmsr_on_cpu(good_cpu, MSR_IA32_PERF_CTL, oldmsr, h);
532 		if (policy->shared_type == CPUFREQ_SHARED_TYPE_ANY)
533 			break;
534 
535 		cpumask_set_cpu(j, covered_cpus);
536 	}
537 
538 	cpufreq_notify_transition(policy, &freqs, CPUFREQ_POSTCHANGE);
539 
540 	if (unlikely(retval)) {
541 		/*
542 		 * We have failed halfway through the frequency change.
543 		 * We have sent callbacks to policy->cpus and
544 		 * MSRs have already been written on coverd_cpus.
545 		 * Best effort undo..
546 		 */
547 
548 		for_each_cpu(j, covered_cpus)
549 			wrmsr_on_cpu(j, MSR_IA32_PERF_CTL, oldmsr, h);
550 
551 		tmp = freqs.new;
552 		freqs.new = freqs.old;
553 		freqs.old = tmp;
554 		cpufreq_notify_transition(policy, &freqs, CPUFREQ_PRECHANGE);
555 		cpufreq_notify_transition(policy, &freqs, CPUFREQ_POSTCHANGE);
556 	}
557 	retval = 0;
558 
559 out:
560 	free_cpumask_var(covered_cpus);
561 	return retval;
562 }
563 
564 static struct freq_attr* centrino_attr[] = {
565 	&cpufreq_freq_attr_scaling_available_freqs,
566 	NULL,
567 };
568 
569 static struct cpufreq_driver centrino_driver = {
570 	.name		= "centrino", /* should be speedstep-centrino,
571 					 but there's a 16 char limit */
572 	.init		= centrino_cpu_init,
573 	.exit		= centrino_cpu_exit,
574 	.verify		= centrino_verify,
575 	.target		= centrino_target,
576 	.get		= get_cur_freq,
577 	.attr           = centrino_attr,
578 };
579 
580 /*
581  * This doesn't replace the detailed checks above because
582  * the generic CPU IDs don't have a way to match for steppings
583  * or ASCII model IDs.
584  */
585 static const struct x86_cpu_id centrino_ids[] = {
586 	{ X86_VENDOR_INTEL, 6, 9, X86_FEATURE_EST },
587 	{ X86_VENDOR_INTEL, 6, 13, X86_FEATURE_EST },
588 	{ X86_VENDOR_INTEL, 6, 13, X86_FEATURE_EST },
589 	{ X86_VENDOR_INTEL, 6, 13, X86_FEATURE_EST },
590 	{ X86_VENDOR_INTEL, 15, 3, X86_FEATURE_EST },
591 	{ X86_VENDOR_INTEL, 15, 4, X86_FEATURE_EST },
592 	{}
593 };
594 #if 0
595 /* Autoload or not? Do not for now. */
596 MODULE_DEVICE_TABLE(x86cpu, centrino_ids);
597 #endif
598 
599 /**
600  * centrino_init - initializes the Enhanced SpeedStep CPUFreq driver
601  *
602  * Initializes the Enhanced SpeedStep support. Returns -ENODEV on
603  * unsupported devices, -ENOENT if there's no voltage table for this
604  * particular CPU model, -EINVAL on problems during initiatization,
605  * and zero on success.
606  *
607  * This is quite picky.  Not only does the CPU have to advertise the
608  * "est" flag in the cpuid capability flags, we look for a specific
609  * CPU model and stepping, and we need to have the exact model name in
610  * our voltage tables.  That is, be paranoid about not releasing
611  * someone's valuable magic smoke.
612  */
613 static int __init centrino_init(void)
614 {
615 	if (!x86_match_cpu(centrino_ids))
616 		return -ENODEV;
617 	return cpufreq_register_driver(&centrino_driver);
618 }
619 
620 static void __exit centrino_exit(void)
621 {
622 	cpufreq_unregister_driver(&centrino_driver);
623 }
624 
625 MODULE_AUTHOR ("Jeremy Fitzhardinge <jeremy@goop.org>");
626 MODULE_DESCRIPTION ("Enhanced SpeedStep driver for Intel Pentium M processors.");
627 MODULE_LICENSE ("GPL");
628 
629 late_initcall(centrino_init);
630 module_exit(centrino_exit);
631