xref: /linux/drivers/cpufreq/cppc_cpufreq.c (revision 7fc2cd2e4b398c57c9cf961cfea05eadbf34c05c)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * CPPC (Collaborative Processor Performance Control) driver for
4  * interfacing with the CPUfreq layer and governors. See
5  * cppc_acpi.c for CPPC specific methods.
6  *
7  * (C) Copyright 2014, 2015 Linaro Ltd.
8  * Author: Ashwin Chaugule <ashwin.chaugule@linaro.org>
9  */
10 
11 #define pr_fmt(fmt)	"CPPC Cpufreq:"	fmt
12 
13 #include <linux/arch_topology.h>
14 #include <linux/kernel.h>
15 #include <linux/module.h>
16 #include <linux/delay.h>
17 #include <linux/cpu.h>
18 #include <linux/cpufreq.h>
19 #include <linux/irq_work.h>
20 #include <linux/kthread.h>
21 #include <linux/time.h>
22 #include <linux/vmalloc.h>
23 #include <uapi/linux/sched/types.h>
24 
25 #include <linux/unaligned.h>
26 
27 #include <acpi/cppc_acpi.h>
28 
29 static struct cpufreq_driver cppc_cpufreq_driver;
30 
31 #ifdef CONFIG_ACPI_CPPC_CPUFREQ_FIE
32 static enum {
33 	FIE_UNSET = -1,
34 	FIE_ENABLED,
35 	FIE_DISABLED
36 } fie_disabled = FIE_UNSET;
37 
38 module_param(fie_disabled, int, 0444);
39 MODULE_PARM_DESC(fie_disabled, "Disable Frequency Invariance Engine (FIE)");
40 
41 /* Frequency invariance support */
42 struct cppc_freq_invariance {
43 	int cpu;
44 	struct irq_work irq_work;
45 	struct kthread_work work;
46 	struct cppc_perf_fb_ctrs prev_perf_fb_ctrs;
47 	struct cppc_cpudata *cpu_data;
48 };
49 
50 static DEFINE_PER_CPU(struct cppc_freq_invariance, cppc_freq_inv);
51 static struct kthread_worker *kworker_fie;
52 
53 static int cppc_perf_from_fbctrs(struct cppc_perf_fb_ctrs *fb_ctrs_t0,
54 				 struct cppc_perf_fb_ctrs *fb_ctrs_t1);
55 
56 /**
57  * cppc_scale_freq_workfn - CPPC arch_freq_scale updater for frequency invariance
58  * @work: The work item.
59  *
60  * The CPPC driver register itself with the topology core to provide its own
61  * implementation (cppc_scale_freq_tick()) of topology_scale_freq_tick() which
62  * gets called by the scheduler on every tick.
63  *
64  * Note that the arch specific counters have higher priority than CPPC counters,
65  * if available, though the CPPC driver doesn't need to have any special
66  * handling for that.
67  *
68  * On an invocation of cppc_scale_freq_tick(), we schedule an irq work (since we
69  * reach here from hard-irq context), which then schedules a normal work item
70  * and cppc_scale_freq_workfn() updates the per_cpu arch_freq_scale variable
71  * based on the counter updates since the last tick.
72  */
73 static void cppc_scale_freq_workfn(struct kthread_work *work)
74 {
75 	struct cppc_freq_invariance *cppc_fi;
76 	struct cppc_perf_fb_ctrs fb_ctrs = {0};
77 	struct cppc_cpudata *cpu_data;
78 	unsigned long local_freq_scale;
79 	u64 perf;
80 
81 	cppc_fi = container_of(work, struct cppc_freq_invariance, work);
82 	cpu_data = cppc_fi->cpu_data;
83 
84 	if (cppc_get_perf_ctrs(cppc_fi->cpu, &fb_ctrs)) {
85 		pr_warn("%s: failed to read perf counters\n", __func__);
86 		return;
87 	}
88 
89 	perf = cppc_perf_from_fbctrs(&cppc_fi->prev_perf_fb_ctrs, &fb_ctrs);
90 	if (!perf)
91 		return;
92 
93 	cppc_fi->prev_perf_fb_ctrs = fb_ctrs;
94 
95 	perf <<= SCHED_CAPACITY_SHIFT;
96 	local_freq_scale = div64_u64(perf, cpu_data->perf_caps.highest_perf);
97 
98 	/* This can happen due to counter's overflow */
99 	if (unlikely(local_freq_scale > 1024))
100 		local_freq_scale = 1024;
101 
102 	per_cpu(arch_freq_scale, cppc_fi->cpu) = local_freq_scale;
103 }
104 
105 static void cppc_irq_work(struct irq_work *irq_work)
106 {
107 	struct cppc_freq_invariance *cppc_fi;
108 
109 	cppc_fi = container_of(irq_work, struct cppc_freq_invariance, irq_work);
110 	kthread_queue_work(kworker_fie, &cppc_fi->work);
111 }
112 
113 static void cppc_scale_freq_tick(void)
114 {
115 	struct cppc_freq_invariance *cppc_fi = &per_cpu(cppc_freq_inv, smp_processor_id());
116 
117 	/*
118 	 * cppc_get_perf_ctrs() can potentially sleep, call that from the right
119 	 * context.
120 	 */
121 	irq_work_queue(&cppc_fi->irq_work);
122 }
123 
124 static struct scale_freq_data cppc_sftd = {
125 	.source = SCALE_FREQ_SOURCE_CPPC,
126 	.set_freq_scale = cppc_scale_freq_tick,
127 };
128 
129 static void cppc_cpufreq_cpu_fie_init(struct cpufreq_policy *policy)
130 {
131 	struct cppc_freq_invariance *cppc_fi;
132 	int cpu, ret;
133 
134 	if (fie_disabled)
135 		return;
136 
137 	for_each_cpu(cpu, policy->cpus) {
138 		cppc_fi = &per_cpu(cppc_freq_inv, cpu);
139 		cppc_fi->cpu = cpu;
140 		cppc_fi->cpu_data = policy->driver_data;
141 		kthread_init_work(&cppc_fi->work, cppc_scale_freq_workfn);
142 		init_irq_work(&cppc_fi->irq_work, cppc_irq_work);
143 
144 		ret = cppc_get_perf_ctrs(cpu, &cppc_fi->prev_perf_fb_ctrs);
145 
146 		/*
147 		 * Don't abort as the CPU was offline while the driver was
148 		 * getting registered.
149 		 */
150 		if (ret && cpu_online(cpu)) {
151 			pr_debug("%s: failed to read perf counters for cpu:%d: %d\n",
152 				__func__, cpu, ret);
153 			return;
154 		}
155 	}
156 
157 	/* Register for freq-invariance */
158 	topology_set_scale_freq_source(&cppc_sftd, policy->cpus);
159 }
160 
161 /*
162  * We free all the resources on policy's removal and not on CPU removal as the
163  * irq-work are per-cpu and the hotplug core takes care of flushing the pending
164  * irq-works (hint: smpcfd_dying_cpu()) on CPU hotplug. Even if the kthread-work
165  * fires on another CPU after the concerned CPU is removed, it won't harm.
166  *
167  * We just need to make sure to remove them all on policy->exit().
168  */
169 static void cppc_cpufreq_cpu_fie_exit(struct cpufreq_policy *policy)
170 {
171 	struct cppc_freq_invariance *cppc_fi;
172 	int cpu;
173 
174 	if (fie_disabled)
175 		return;
176 
177 	/* policy->cpus will be empty here, use related_cpus instead */
178 	topology_clear_scale_freq_source(SCALE_FREQ_SOURCE_CPPC, policy->related_cpus);
179 
180 	for_each_cpu(cpu, policy->related_cpus) {
181 		cppc_fi = &per_cpu(cppc_freq_inv, cpu);
182 		irq_work_sync(&cppc_fi->irq_work);
183 		kthread_cancel_work_sync(&cppc_fi->work);
184 	}
185 }
186 
187 static void __init cppc_freq_invariance_init(void)
188 {
189 	struct sched_attr attr = {
190 		.size		= sizeof(struct sched_attr),
191 		.sched_policy	= SCHED_DEADLINE,
192 		.sched_nice	= 0,
193 		.sched_priority	= 0,
194 		/*
195 		 * Fake (unused) bandwidth; workaround to "fix"
196 		 * priority inheritance.
197 		 */
198 		.sched_runtime	= NSEC_PER_MSEC,
199 		.sched_deadline = 10 * NSEC_PER_MSEC,
200 		.sched_period	= 10 * NSEC_PER_MSEC,
201 	};
202 	int ret;
203 
204 	if (fie_disabled != FIE_ENABLED && fie_disabled != FIE_DISABLED) {
205 		fie_disabled = FIE_ENABLED;
206 		if (cppc_perf_ctrs_in_pcc()) {
207 			pr_info("FIE not enabled on systems with registers in PCC\n");
208 			fie_disabled = FIE_DISABLED;
209 		}
210 	}
211 
212 	if (fie_disabled)
213 		return;
214 
215 	kworker_fie = kthread_run_worker(0, "cppc_fie");
216 	if (IS_ERR(kworker_fie)) {
217 		pr_warn("%s: failed to create kworker_fie: %ld\n", __func__,
218 			PTR_ERR(kworker_fie));
219 		fie_disabled = FIE_DISABLED;
220 		return;
221 	}
222 
223 	ret = sched_setattr_nocheck(kworker_fie->task, &attr);
224 	if (ret) {
225 		pr_warn("%s: failed to set SCHED_DEADLINE: %d\n", __func__,
226 			ret);
227 		kthread_destroy_worker(kworker_fie);
228 		fie_disabled = FIE_DISABLED;
229 	}
230 }
231 
232 static void cppc_freq_invariance_exit(void)
233 {
234 	if (fie_disabled)
235 		return;
236 
237 	kthread_destroy_worker(kworker_fie);
238 }
239 
240 #else
241 static inline void cppc_cpufreq_cpu_fie_init(struct cpufreq_policy *policy)
242 {
243 }
244 
245 static inline void cppc_cpufreq_cpu_fie_exit(struct cpufreq_policy *policy)
246 {
247 }
248 
249 static inline void cppc_freq_invariance_init(void)
250 {
251 }
252 
253 static inline void cppc_freq_invariance_exit(void)
254 {
255 }
256 #endif /* CONFIG_ACPI_CPPC_CPUFREQ_FIE */
257 
258 static int cppc_cpufreq_set_target(struct cpufreq_policy *policy,
259 				   unsigned int target_freq,
260 				   unsigned int relation)
261 {
262 	struct cppc_cpudata *cpu_data = policy->driver_data;
263 	unsigned int cpu = policy->cpu;
264 	struct cpufreq_freqs freqs;
265 	int ret = 0;
266 
267 	cpu_data->perf_ctrls.desired_perf =
268 			cppc_khz_to_perf(&cpu_data->perf_caps, target_freq);
269 	freqs.old = policy->cur;
270 	freqs.new = target_freq;
271 
272 	cpufreq_freq_transition_begin(policy, &freqs);
273 	ret = cppc_set_perf(cpu, &cpu_data->perf_ctrls);
274 	cpufreq_freq_transition_end(policy, &freqs, ret != 0);
275 
276 	if (ret)
277 		pr_debug("Failed to set target on CPU:%d. ret:%d\n",
278 			 cpu, ret);
279 
280 	return ret;
281 }
282 
283 static unsigned int cppc_cpufreq_fast_switch(struct cpufreq_policy *policy,
284 					      unsigned int target_freq)
285 {
286 	struct cppc_cpudata *cpu_data = policy->driver_data;
287 	unsigned int cpu = policy->cpu;
288 	u32 desired_perf;
289 	int ret;
290 
291 	desired_perf = cppc_khz_to_perf(&cpu_data->perf_caps, target_freq);
292 	cpu_data->perf_ctrls.desired_perf = desired_perf;
293 	ret = cppc_set_perf(cpu, &cpu_data->perf_ctrls);
294 
295 	if (ret) {
296 		pr_debug("Failed to set target on CPU:%d. ret:%d\n",
297 			 cpu, ret);
298 		return 0;
299 	}
300 
301 	return target_freq;
302 }
303 
304 static int cppc_verify_policy(struct cpufreq_policy_data *policy)
305 {
306 	cpufreq_verify_within_cpu_limits(policy);
307 	return 0;
308 }
309 
310 static unsigned int __cppc_cpufreq_get_transition_delay_us(unsigned int cpu)
311 {
312 	int transition_latency_ns = cppc_get_transition_latency(cpu);
313 
314 	if (transition_latency_ns < 0)
315 		return CPUFREQ_DEFAULT_TRANSITION_LATENCY_NS / NSEC_PER_USEC;
316 
317 	return transition_latency_ns / NSEC_PER_USEC;
318 }
319 
320 /*
321  * The PCC subspace describes the rate at which platform can accept commands
322  * on the shared PCC channel (including READs which do not count towards freq
323  * transition requests), so ideally we need to use the PCC values as a fallback
324  * if we don't have a platform specific transition_delay_us
325  */
326 #ifdef CONFIG_ARM64
327 #include <asm/cputype.h>
328 
329 static unsigned int cppc_cpufreq_get_transition_delay_us(unsigned int cpu)
330 {
331 	unsigned long implementor = read_cpuid_implementor();
332 	unsigned long part_num = read_cpuid_part_number();
333 
334 	switch (implementor) {
335 	case ARM_CPU_IMP_QCOM:
336 		switch (part_num) {
337 		case QCOM_CPU_PART_FALKOR_V1:
338 		case QCOM_CPU_PART_FALKOR:
339 			return 10000;
340 		}
341 	}
342 	return __cppc_cpufreq_get_transition_delay_us(cpu);
343 }
344 #else
345 static unsigned int cppc_cpufreq_get_transition_delay_us(unsigned int cpu)
346 {
347 	return __cppc_cpufreq_get_transition_delay_us(cpu);
348 }
349 #endif
350 
351 #if defined(CONFIG_ARM64) && defined(CONFIG_ENERGY_MODEL)
352 
353 static DEFINE_PER_CPU(unsigned int, efficiency_class);
354 
355 /* Create an artificial performance state every CPPC_EM_CAP_STEP capacity unit. */
356 #define CPPC_EM_CAP_STEP	(20)
357 /* Increase the cost value by CPPC_EM_COST_STEP every performance state. */
358 #define CPPC_EM_COST_STEP	(1)
359 /* Add a cost gap correspnding to the energy of 4 CPUs. */
360 #define CPPC_EM_COST_GAP	(4 * SCHED_CAPACITY_SCALE * CPPC_EM_COST_STEP \
361 				/ CPPC_EM_CAP_STEP)
362 
363 static unsigned int get_perf_level_count(struct cpufreq_policy *policy)
364 {
365 	struct cppc_perf_caps *perf_caps;
366 	unsigned int min_cap, max_cap;
367 	struct cppc_cpudata *cpu_data;
368 	int cpu = policy->cpu;
369 
370 	cpu_data = policy->driver_data;
371 	perf_caps = &cpu_data->perf_caps;
372 	max_cap = arch_scale_cpu_capacity(cpu);
373 	min_cap = div_u64((u64)max_cap * perf_caps->lowest_perf,
374 			  perf_caps->highest_perf);
375 	if ((min_cap == 0) || (max_cap < min_cap))
376 		return 0;
377 	return 1 + max_cap / CPPC_EM_CAP_STEP - min_cap / CPPC_EM_CAP_STEP;
378 }
379 
380 /*
381  * The cost is defined as:
382  *   cost = power * max_frequency / frequency
383  */
384 static inline unsigned long compute_cost(int cpu, int step)
385 {
386 	return CPPC_EM_COST_GAP * per_cpu(efficiency_class, cpu) +
387 			step * CPPC_EM_COST_STEP;
388 }
389 
390 static int cppc_get_cpu_power(struct device *cpu_dev,
391 		unsigned long *power, unsigned long *KHz)
392 {
393 	unsigned long perf_step, perf_prev, perf, perf_check;
394 	unsigned int min_step, max_step, step, step_check;
395 	unsigned long prev_freq = *KHz;
396 	unsigned int min_cap, max_cap;
397 	struct cpufreq_policy *policy;
398 
399 	struct cppc_perf_caps *perf_caps;
400 	struct cppc_cpudata *cpu_data;
401 
402 	policy = cpufreq_cpu_get_raw(cpu_dev->id);
403 	if (!policy)
404 		return -EINVAL;
405 
406 	cpu_data = policy->driver_data;
407 	perf_caps = &cpu_data->perf_caps;
408 	max_cap = arch_scale_cpu_capacity(cpu_dev->id);
409 	min_cap = div_u64((u64)max_cap * perf_caps->lowest_perf,
410 			  perf_caps->highest_perf);
411 	perf_step = div_u64((u64)CPPC_EM_CAP_STEP * perf_caps->highest_perf,
412 			    max_cap);
413 	min_step = min_cap / CPPC_EM_CAP_STEP;
414 	max_step = max_cap / CPPC_EM_CAP_STEP;
415 
416 	perf_prev = cppc_khz_to_perf(perf_caps, *KHz);
417 	step = perf_prev / perf_step;
418 
419 	if (step > max_step)
420 		return -EINVAL;
421 
422 	if (min_step == max_step) {
423 		step = max_step;
424 		perf = perf_caps->highest_perf;
425 	} else if (step < min_step) {
426 		step = min_step;
427 		perf = perf_caps->lowest_perf;
428 	} else {
429 		step++;
430 		if (step == max_step)
431 			perf = perf_caps->highest_perf;
432 		else
433 			perf = step * perf_step;
434 	}
435 
436 	*KHz = cppc_perf_to_khz(perf_caps, perf);
437 	perf_check = cppc_khz_to_perf(perf_caps, *KHz);
438 	step_check = perf_check / perf_step;
439 
440 	/*
441 	 * To avoid bad integer approximation, check that new frequency value
442 	 * increased and that the new frequency will be converted to the
443 	 * desired step value.
444 	 */
445 	while ((*KHz == prev_freq) || (step_check != step)) {
446 		perf++;
447 		*KHz = cppc_perf_to_khz(perf_caps, perf);
448 		perf_check = cppc_khz_to_perf(perf_caps, *KHz);
449 		step_check = perf_check / perf_step;
450 	}
451 
452 	/*
453 	 * With an artificial EM, only the cost value is used. Still the power
454 	 * is populated such as 0 < power < EM_MAX_POWER. This allows to add
455 	 * more sense to the artificial performance states.
456 	 */
457 	*power = compute_cost(cpu_dev->id, step);
458 
459 	return 0;
460 }
461 
462 static int cppc_get_cpu_cost(struct device *cpu_dev, unsigned long KHz,
463 		unsigned long *cost)
464 {
465 	unsigned long perf_step, perf_prev;
466 	struct cppc_perf_caps *perf_caps;
467 	struct cpufreq_policy *policy;
468 	struct cppc_cpudata *cpu_data;
469 	unsigned int max_cap;
470 	int step;
471 
472 	policy = cpufreq_cpu_get_raw(cpu_dev->id);
473 	if (!policy)
474 		return -EINVAL;
475 
476 	cpu_data = policy->driver_data;
477 	perf_caps = &cpu_data->perf_caps;
478 	max_cap = arch_scale_cpu_capacity(cpu_dev->id);
479 
480 	perf_prev = cppc_khz_to_perf(perf_caps, KHz);
481 	perf_step = CPPC_EM_CAP_STEP * perf_caps->highest_perf / max_cap;
482 	step = perf_prev / perf_step;
483 
484 	*cost = compute_cost(cpu_dev->id, step);
485 
486 	return 0;
487 }
488 
489 static void cppc_cpufreq_register_em(struct cpufreq_policy *policy)
490 {
491 	struct cppc_cpudata *cpu_data;
492 	struct em_data_callback em_cb =
493 		EM_ADV_DATA_CB(cppc_get_cpu_power, cppc_get_cpu_cost);
494 
495 	cpu_data = policy->driver_data;
496 	em_dev_register_perf_domain(get_cpu_device(policy->cpu),
497 			get_perf_level_count(policy), &em_cb,
498 			cpu_data->shared_cpu_map, 0);
499 }
500 
501 static void populate_efficiency_class(void)
502 {
503 	struct acpi_madt_generic_interrupt *gicc;
504 	DECLARE_BITMAP(used_classes, 256) = {};
505 	int class, cpu, index;
506 
507 	for_each_possible_cpu(cpu) {
508 		gicc = acpi_cpu_get_madt_gicc(cpu);
509 		class = gicc->efficiency_class;
510 		bitmap_set(used_classes, class, 1);
511 	}
512 
513 	if (bitmap_weight(used_classes, 256) <= 1) {
514 		pr_debug("Efficiency classes are all equal (=%d). "
515 			"No EM registered", class);
516 		return;
517 	}
518 
519 	/*
520 	 * Squeeze efficiency class values on [0:#efficiency_class-1].
521 	 * Values are per spec in [0:255].
522 	 */
523 	index = 0;
524 	for_each_set_bit(class, used_classes, 256) {
525 		for_each_possible_cpu(cpu) {
526 			gicc = acpi_cpu_get_madt_gicc(cpu);
527 			if (gicc->efficiency_class == class)
528 				per_cpu(efficiency_class, cpu) = index;
529 		}
530 		index++;
531 	}
532 	cppc_cpufreq_driver.register_em = cppc_cpufreq_register_em;
533 }
534 
535 #else
536 static void populate_efficiency_class(void)
537 {
538 }
539 #endif
540 
541 static struct cppc_cpudata *cppc_cpufreq_get_cpu_data(unsigned int cpu)
542 {
543 	struct cppc_cpudata *cpu_data;
544 	int ret;
545 
546 	cpu_data = kzalloc(sizeof(struct cppc_cpudata), GFP_KERNEL);
547 	if (!cpu_data)
548 		goto out;
549 
550 	if (!zalloc_cpumask_var(&cpu_data->shared_cpu_map, GFP_KERNEL))
551 		goto free_cpu;
552 
553 	ret = acpi_get_psd_map(cpu, cpu_data);
554 	if (ret) {
555 		pr_debug("Err parsing CPU%d PSD data: ret:%d\n", cpu, ret);
556 		goto free_mask;
557 	}
558 
559 	ret = cppc_get_perf_caps(cpu, &cpu_data->perf_caps);
560 	if (ret) {
561 		pr_debug("Err reading CPU%d perf caps: ret:%d\n", cpu, ret);
562 		goto free_mask;
563 	}
564 
565 	return cpu_data;
566 
567 free_mask:
568 	free_cpumask_var(cpu_data->shared_cpu_map);
569 free_cpu:
570 	kfree(cpu_data);
571 out:
572 	return NULL;
573 }
574 
575 static void cppc_cpufreq_put_cpu_data(struct cpufreq_policy *policy)
576 {
577 	struct cppc_cpudata *cpu_data = policy->driver_data;
578 
579 	free_cpumask_var(cpu_data->shared_cpu_map);
580 	kfree(cpu_data);
581 	policy->driver_data = NULL;
582 }
583 
584 static int cppc_cpufreq_cpu_init(struct cpufreq_policy *policy)
585 {
586 	unsigned int cpu = policy->cpu;
587 	struct cppc_cpudata *cpu_data;
588 	struct cppc_perf_caps *caps;
589 	int ret;
590 
591 	cpu_data = cppc_cpufreq_get_cpu_data(cpu);
592 	if (!cpu_data) {
593 		pr_err("Error in acquiring _CPC/_PSD data for CPU%d.\n", cpu);
594 		return -ENODEV;
595 	}
596 	caps = &cpu_data->perf_caps;
597 	policy->driver_data = cpu_data;
598 
599 	/*
600 	 * Set min to lowest nonlinear perf to avoid any efficiency penalty (see
601 	 * Section 8.4.7.1.1.5 of ACPI 6.1 spec)
602 	 */
603 	policy->min = cppc_perf_to_khz(caps, caps->lowest_nonlinear_perf);
604 	policy->max = cppc_perf_to_khz(caps, policy->boost_enabled ?
605 						caps->highest_perf : caps->nominal_perf);
606 
607 	/*
608 	 * Set cpuinfo.min_freq to Lowest to make the full range of performance
609 	 * available if userspace wants to use any perf between lowest & lowest
610 	 * nonlinear perf
611 	 */
612 	policy->cpuinfo.min_freq = cppc_perf_to_khz(caps, caps->lowest_perf);
613 	policy->cpuinfo.max_freq = policy->max;
614 
615 	policy->transition_delay_us = cppc_cpufreq_get_transition_delay_us(cpu);
616 	policy->shared_type = cpu_data->shared_type;
617 
618 	switch (policy->shared_type) {
619 	case CPUFREQ_SHARED_TYPE_HW:
620 	case CPUFREQ_SHARED_TYPE_NONE:
621 		/* Nothing to be done - we'll have a policy for each CPU */
622 		break;
623 	case CPUFREQ_SHARED_TYPE_ANY:
624 		/*
625 		 * All CPUs in the domain will share a policy and all cpufreq
626 		 * operations will use a single cppc_cpudata structure stored
627 		 * in policy->driver_data.
628 		 */
629 		cpumask_copy(policy->cpus, cpu_data->shared_cpu_map);
630 		break;
631 	default:
632 		pr_debug("Unsupported CPU co-ord type: %d\n",
633 			 policy->shared_type);
634 		ret = -EFAULT;
635 		goto out;
636 	}
637 
638 	policy->fast_switch_possible = cppc_allow_fast_switch();
639 	policy->dvfs_possible_from_any_cpu = true;
640 
641 	/*
642 	 * If 'highest_perf' is greater than 'nominal_perf', we assume CPU Boost
643 	 * is supported.
644 	 */
645 	if (caps->highest_perf > caps->nominal_perf)
646 		policy->boost_supported = true;
647 
648 	/* Set policy->cur to max now. The governors will adjust later. */
649 	policy->cur = cppc_perf_to_khz(caps, caps->highest_perf);
650 	cpu_data->perf_ctrls.desired_perf =  caps->highest_perf;
651 
652 	ret = cppc_set_perf(cpu, &cpu_data->perf_ctrls);
653 	if (ret) {
654 		pr_debug("Err setting perf value:%d on CPU:%d. ret:%d\n",
655 			 caps->highest_perf, cpu, ret);
656 		goto out;
657 	}
658 
659 	cppc_cpufreq_cpu_fie_init(policy);
660 	return 0;
661 
662 out:
663 	cppc_cpufreq_put_cpu_data(policy);
664 	return ret;
665 }
666 
667 static void cppc_cpufreq_cpu_exit(struct cpufreq_policy *policy)
668 {
669 	struct cppc_cpudata *cpu_data = policy->driver_data;
670 	struct cppc_perf_caps *caps = &cpu_data->perf_caps;
671 	unsigned int cpu = policy->cpu;
672 	int ret;
673 
674 	cppc_cpufreq_cpu_fie_exit(policy);
675 
676 	cpu_data->perf_ctrls.desired_perf = caps->lowest_perf;
677 
678 	ret = cppc_set_perf(cpu, &cpu_data->perf_ctrls);
679 	if (ret)
680 		pr_debug("Err setting perf value:%d on CPU:%d. ret:%d\n",
681 			 caps->lowest_perf, cpu, ret);
682 
683 	cppc_cpufreq_put_cpu_data(policy);
684 }
685 
686 static inline u64 get_delta(u64 t1, u64 t0)
687 {
688 	if (t1 > t0 || t0 > ~(u32)0)
689 		return t1 - t0;
690 
691 	return (u32)t1 - (u32)t0;
692 }
693 
694 static int cppc_perf_from_fbctrs(struct cppc_perf_fb_ctrs *fb_ctrs_t0,
695 				 struct cppc_perf_fb_ctrs *fb_ctrs_t1)
696 {
697 	u64 delta_reference, delta_delivered;
698 	u64 reference_perf;
699 
700 	reference_perf = fb_ctrs_t0->reference_perf;
701 
702 	delta_reference = get_delta(fb_ctrs_t1->reference,
703 				    fb_ctrs_t0->reference);
704 	delta_delivered = get_delta(fb_ctrs_t1->delivered,
705 				    fb_ctrs_t0->delivered);
706 
707 	/*
708 	 * Avoid divide-by zero and unchanged feedback counters.
709 	 * Leave it for callers to handle.
710 	 */
711 	if (!delta_reference || !delta_delivered)
712 		return 0;
713 
714 	return (reference_perf * delta_delivered) / delta_reference;
715 }
716 
717 static int cppc_get_perf_ctrs_sample(int cpu,
718 				     struct cppc_perf_fb_ctrs *fb_ctrs_t0,
719 				     struct cppc_perf_fb_ctrs *fb_ctrs_t1)
720 {
721 	int ret;
722 
723 	ret = cppc_get_perf_ctrs(cpu, fb_ctrs_t0);
724 	if (ret)
725 		return ret;
726 
727 	udelay(2); /* 2usec delay between sampling */
728 
729 	return cppc_get_perf_ctrs(cpu, fb_ctrs_t1);
730 }
731 
732 static unsigned int cppc_cpufreq_get_rate(unsigned int cpu)
733 {
734 	struct cpufreq_policy *policy __free(put_cpufreq_policy) = cpufreq_cpu_get(cpu);
735 	struct cppc_perf_fb_ctrs fb_ctrs_t0 = {0}, fb_ctrs_t1 = {0};
736 	struct cppc_cpudata *cpu_data;
737 	u64 delivered_perf;
738 	int ret;
739 
740 	if (!policy)
741 		return 0;
742 
743 	cpu_data = policy->driver_data;
744 
745 	ret = cppc_get_perf_ctrs_sample(cpu, &fb_ctrs_t0, &fb_ctrs_t1);
746 	if (ret) {
747 		if (ret == -EFAULT)
748 			/* Any of the associated CPPC regs is 0. */
749 			goto out_invalid_counters;
750 		else
751 			return 0;
752 	}
753 
754 	delivered_perf = cppc_perf_from_fbctrs(&fb_ctrs_t0, &fb_ctrs_t1);
755 	if (!delivered_perf)
756 		goto out_invalid_counters;
757 
758 	return cppc_perf_to_khz(&cpu_data->perf_caps, delivered_perf);
759 
760 out_invalid_counters:
761 	/*
762 	 * Feedback counters could be unchanged or 0 when a cpu enters a
763 	 * low-power idle state, e.g. clock-gated or power-gated.
764 	 * Use desired perf for reflecting frequency.  Get the latest register
765 	 * value first as some platforms may update the actual delivered perf
766 	 * there; if failed, resort to the cached desired perf.
767 	 */
768 	if (cppc_get_desired_perf(cpu, &delivered_perf))
769 		delivered_perf = cpu_data->perf_ctrls.desired_perf;
770 
771 	return cppc_perf_to_khz(&cpu_data->perf_caps, delivered_perf);
772 }
773 
774 static int cppc_cpufreq_set_boost(struct cpufreq_policy *policy, int state)
775 {
776 	struct cppc_cpudata *cpu_data = policy->driver_data;
777 	struct cppc_perf_caps *caps = &cpu_data->perf_caps;
778 	int ret;
779 
780 	if (state)
781 		policy->max = cppc_perf_to_khz(caps, caps->highest_perf);
782 	else
783 		policy->max = cppc_perf_to_khz(caps, caps->nominal_perf);
784 	policy->cpuinfo.max_freq = policy->max;
785 
786 	ret = freq_qos_update_request(policy->max_freq_req, policy->max);
787 	if (ret < 0)
788 		return ret;
789 
790 	return 0;
791 }
792 
793 static ssize_t show_freqdomain_cpus(struct cpufreq_policy *policy, char *buf)
794 {
795 	struct cppc_cpudata *cpu_data = policy->driver_data;
796 
797 	return cpufreq_show_cpus(cpu_data->shared_cpu_map, buf);
798 }
799 
800 static ssize_t show_auto_select(struct cpufreq_policy *policy, char *buf)
801 {
802 	bool val;
803 	int ret;
804 
805 	ret = cppc_get_auto_sel(policy->cpu, &val);
806 
807 	/* show "<unsupported>" when this register is not supported by cpc */
808 	if (ret == -EOPNOTSUPP)
809 		return sysfs_emit(buf, "<unsupported>\n");
810 
811 	if (ret)
812 		return ret;
813 
814 	return sysfs_emit(buf, "%d\n", val);
815 }
816 
817 static ssize_t store_auto_select(struct cpufreq_policy *policy,
818 				 const char *buf, size_t count)
819 {
820 	bool val;
821 	int ret;
822 
823 	ret = kstrtobool(buf, &val);
824 	if (ret)
825 		return ret;
826 
827 	ret = cppc_set_auto_sel(policy->cpu, val);
828 	if (ret)
829 		return ret;
830 
831 	return count;
832 }
833 
834 static ssize_t show_auto_act_window(struct cpufreq_policy *policy, char *buf)
835 {
836 	u64 val;
837 	int ret;
838 
839 	ret = cppc_get_auto_act_window(policy->cpu, &val);
840 
841 	/* show "<unsupported>" when this register is not supported by cpc */
842 	if (ret == -EOPNOTSUPP)
843 		return sysfs_emit(buf, "<unsupported>\n");
844 
845 	if (ret)
846 		return ret;
847 
848 	return sysfs_emit(buf, "%llu\n", val);
849 }
850 
851 static ssize_t store_auto_act_window(struct cpufreq_policy *policy,
852 				     const char *buf, size_t count)
853 {
854 	u64 usec;
855 	int ret;
856 
857 	ret = kstrtou64(buf, 0, &usec);
858 	if (ret)
859 		return ret;
860 
861 	ret = cppc_set_auto_act_window(policy->cpu, usec);
862 	if (ret)
863 		return ret;
864 
865 	return count;
866 }
867 
868 static ssize_t show_energy_performance_preference_val(struct cpufreq_policy *policy, char *buf)
869 {
870 	u64 val;
871 	int ret;
872 
873 	ret = cppc_get_epp_perf(policy->cpu, &val);
874 
875 	/* show "<unsupported>" when this register is not supported by cpc */
876 	if (ret == -EOPNOTSUPP)
877 		return sysfs_emit(buf, "<unsupported>\n");
878 
879 	if (ret)
880 		return ret;
881 
882 	return sysfs_emit(buf, "%llu\n", val);
883 }
884 
885 static ssize_t store_energy_performance_preference_val(struct cpufreq_policy *policy,
886 						       const char *buf, size_t count)
887 {
888 	u64 val;
889 	int ret;
890 
891 	ret = kstrtou64(buf, 0, &val);
892 	if (ret)
893 		return ret;
894 
895 	ret = cppc_set_epp(policy->cpu, val);
896 	if (ret)
897 		return ret;
898 
899 	return count;
900 }
901 
902 cpufreq_freq_attr_ro(freqdomain_cpus);
903 cpufreq_freq_attr_rw(auto_select);
904 cpufreq_freq_attr_rw(auto_act_window);
905 cpufreq_freq_attr_rw(energy_performance_preference_val);
906 
907 static struct freq_attr *cppc_cpufreq_attr[] = {
908 	&freqdomain_cpus,
909 	&auto_select,
910 	&auto_act_window,
911 	&energy_performance_preference_val,
912 	NULL,
913 };
914 
915 static struct cpufreq_driver cppc_cpufreq_driver = {
916 	.flags = CPUFREQ_CONST_LOOPS | CPUFREQ_NEED_UPDATE_LIMITS,
917 	.verify = cppc_verify_policy,
918 	.target = cppc_cpufreq_set_target,
919 	.get = cppc_cpufreq_get_rate,
920 	.fast_switch = cppc_cpufreq_fast_switch,
921 	.init = cppc_cpufreq_cpu_init,
922 	.exit = cppc_cpufreq_cpu_exit,
923 	.set_boost = cppc_cpufreq_set_boost,
924 	.attr = cppc_cpufreq_attr,
925 	.name = "cppc_cpufreq",
926 };
927 
928 static int __init cppc_cpufreq_init(void)
929 {
930 	int ret;
931 
932 	if (!acpi_cpc_valid())
933 		return -ENODEV;
934 
935 	cppc_freq_invariance_init();
936 	populate_efficiency_class();
937 
938 	ret = cpufreq_register_driver(&cppc_cpufreq_driver);
939 	if (ret)
940 		cppc_freq_invariance_exit();
941 
942 	return ret;
943 }
944 
945 static void __exit cppc_cpufreq_exit(void)
946 {
947 	cpufreq_unregister_driver(&cppc_cpufreq_driver);
948 	cppc_freq_invariance_exit();
949 }
950 
951 module_exit(cppc_cpufreq_exit);
952 MODULE_AUTHOR("Ashwin Chaugule");
953 MODULE_DESCRIPTION("CPUFreq driver based on the ACPI CPPC v5.0+ spec");
954 MODULE_LICENSE("GPL");
955 
956 late_initcall(cppc_cpufreq_init);
957 
958 static const struct acpi_device_id cppc_acpi_ids[] __used = {
959 	{ACPI_PROCESSOR_DEVICE_HID, },
960 	{}
961 };
962 
963 MODULE_DEVICE_TABLE(acpi, cppc_acpi_ids);
964