xref: /linux/drivers/cpufreq/cppc_cpufreq.c (revision bba2c3615bd6cfee7456d1130f2e6b01b3f4e9ba)
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(u64 reference_perf,
54 				 struct cppc_perf_fb_ctrs *fb_ctrs_t0,
55 				 struct cppc_perf_fb_ctrs *fb_ctrs_t1);
56 
57 /**
58  * __cppc_scale_freq_tick - CPPC arch_freq_scale updater for frequency invariance
59  * @cppc_fi: per-cpu CPPC FIE data.
60  *
61  * The CPPC driver registers itself with the topology core to provide its own
62  * implementation (cppc_scale_freq_tick()) of topology_scale_freq_tick() which
63  * gets called by the scheduler on every tick.
64  *
65  * Note that the arch specific counters have higher priority than CPPC counters,
66  * if available, though the CPPC driver doesn't need to have any special
67  * handling for that.
68  */
69 static void __cppc_scale_freq_tick(struct cppc_freq_invariance *cppc_fi)
70 {
71 	struct cppc_perf_fb_ctrs fb_ctrs = {0};
72 	struct cppc_cpudata *cpu_data;
73 	unsigned long local_freq_scale;
74 	u64 perf, ref_perf;
75 
76 	cpu_data = cppc_fi->cpu_data;
77 
78 	if (cppc_get_perf_ctrs(cppc_fi->cpu, &fb_ctrs)) {
79 		pr_warn("%s: failed to read perf counters\n", __func__);
80 		return;
81 	}
82 
83 	ref_perf = cpu_data->perf_caps.reference_perf;
84 	perf = cppc_perf_from_fbctrs(ref_perf,
85 				     &cppc_fi->prev_perf_fb_ctrs, &fb_ctrs);
86 	if (!perf)
87 		return;
88 
89 	cppc_fi->prev_perf_fb_ctrs = fb_ctrs;
90 
91 	perf <<= SCHED_CAPACITY_SHIFT;
92 	local_freq_scale = div64_u64(perf, cpu_data->perf_caps.highest_perf);
93 
94 	/* This can happen due to counter's overflow */
95 	if (unlikely(local_freq_scale > 1024))
96 		local_freq_scale = 1024;
97 
98 	per_cpu(arch_freq_scale, cppc_fi->cpu) = local_freq_scale;
99 }
100 
101 static void cppc_scale_freq_tick(void)
102 {
103 	__cppc_scale_freq_tick(&per_cpu(cppc_freq_inv, smp_processor_id()));
104 }
105 
106 static struct scale_freq_data cppc_sftd = {
107 	.source = SCALE_FREQ_SOURCE_CPPC,
108 	.set_freq_scale = cppc_scale_freq_tick,
109 };
110 
111 static void cppc_scale_freq_workfn(struct kthread_work *work)
112 {
113 	struct cppc_freq_invariance *cppc_fi;
114 
115 	cppc_fi = container_of(work, struct cppc_freq_invariance, work);
116 	__cppc_scale_freq_tick(cppc_fi);
117 }
118 
119 static void cppc_irq_work(struct irq_work *irq_work)
120 {
121 	struct cppc_freq_invariance *cppc_fi;
122 
123 	cppc_fi = container_of(irq_work, struct cppc_freq_invariance, irq_work);
124 	kthread_queue_work(kworker_fie, &cppc_fi->work);
125 }
126 
127 /*
128  * Reading perf counters may sleep if the CPC regs are in PCC.  Thus, we
129  * schedule an irq work in scale_freq_tick (since we reach here from hard-irq
130  * context), which then schedules a normal work item cppc_scale_freq_workfn()
131  * that updates the per_cpu arch_freq_scale variable based on the counter
132  * updates since the last tick.
133  */
134 static void cppc_scale_freq_tick_pcc(void)
135 {
136 	struct cppc_freq_invariance *cppc_fi = &per_cpu(cppc_freq_inv, smp_processor_id());
137 
138 	/*
139 	 * cppc_get_perf_ctrs() can potentially sleep, call that from the right
140 	 * context.
141 	 */
142 	irq_work_queue(&cppc_fi->irq_work);
143 }
144 
145 static struct scale_freq_data cppc_sftd_pcc = {
146 	.source = SCALE_FREQ_SOURCE_CPPC,
147 	.set_freq_scale = cppc_scale_freq_tick_pcc,
148 };
149 
150 static void cppc_cpufreq_cpu_fie_init(struct cpufreq_policy *policy)
151 {
152 	struct scale_freq_data *sftd = &cppc_sftd;
153 	struct cppc_freq_invariance *cppc_fi;
154 	int cpu, ret;
155 
156 	if (fie_disabled)
157 		return;
158 
159 	for_each_cpu(cpu, policy->cpus) {
160 		cppc_fi = &per_cpu(cppc_freq_inv, cpu);
161 		cppc_fi->cpu = cpu;
162 		cppc_fi->cpu_data = policy->driver_data;
163 		if (cppc_perf_ctrs_in_pcc_cpu(cpu)) {
164 			kthread_init_work(&cppc_fi->work, cppc_scale_freq_workfn);
165 			init_irq_work(&cppc_fi->irq_work, cppc_irq_work);
166 			sftd = &cppc_sftd_pcc;
167 		}
168 
169 		ret = cppc_get_perf_ctrs(cpu, &cppc_fi->prev_perf_fb_ctrs);
170 
171 		/*
172 		 * Don't abort as the CPU was offline while the driver was
173 		 * getting registered.
174 		 */
175 		if (ret && cpu_online(cpu)) {
176 			pr_debug("%s: failed to read perf counters for cpu:%d: %d\n",
177 				__func__, cpu, ret);
178 			return;
179 		}
180 	}
181 
182 	/* Register for freq-invariance */
183 	topology_set_scale_freq_source(sftd, policy->cpus);
184 }
185 
186 /*
187  * We free all the resources on policy's removal and not on CPU removal as the
188  * irq-work are per-cpu and the hotplug core takes care of flushing the pending
189  * irq-works (hint: smpcfd_dying_cpu()) on CPU hotplug. Even if the kthread-work
190  * fires on another CPU after the concerned CPU is removed, it won't harm.
191  *
192  * We just need to make sure to remove them all on policy->exit().
193  */
194 static void cppc_cpufreq_cpu_fie_exit(struct cpufreq_policy *policy)
195 {
196 	struct cppc_freq_invariance *cppc_fi;
197 	int cpu;
198 
199 	if (fie_disabled)
200 		return;
201 
202 	/* policy->cpus will be empty here, use related_cpus instead */
203 	topology_clear_scale_freq_source(SCALE_FREQ_SOURCE_CPPC, policy->related_cpus);
204 
205 	for_each_cpu(cpu, policy->related_cpus) {
206 		if (!cppc_perf_ctrs_in_pcc_cpu(cpu))
207 			continue;
208 		cppc_fi = &per_cpu(cppc_freq_inv, cpu);
209 		irq_work_sync(&cppc_fi->irq_work);
210 		kthread_cancel_work_sync(&cppc_fi->work);
211 	}
212 }
213 
214 static void cppc_fie_kworker_init(void)
215 {
216 	struct sched_attr attr = {
217 		.size		= sizeof(struct sched_attr),
218 		.sched_policy	= SCHED_DEADLINE,
219 		.sched_nice	= 0,
220 		.sched_priority	= 0,
221 		/*
222 		 * Fake (unused) bandwidth; workaround to "fix"
223 		 * priority inheritance.
224 		 */
225 		.sched_runtime	= NSEC_PER_MSEC,
226 		.sched_deadline = 10 * NSEC_PER_MSEC,
227 		.sched_period	= 10 * NSEC_PER_MSEC,
228 	};
229 	int ret;
230 
231 	kworker_fie = kthread_run_worker(0, "cppc_fie");
232 	if (IS_ERR(kworker_fie)) {
233 		pr_warn("%s: failed to create kworker_fie: %ld\n", __func__,
234 			PTR_ERR(kworker_fie));
235 		fie_disabled = FIE_DISABLED;
236 		kworker_fie = NULL;
237 		return;
238 	}
239 
240 	ret = sched_setattr_nocheck(kworker_fie->task, &attr);
241 	if (ret) {
242 		pr_warn("%s: failed to set SCHED_DEADLINE: %d\n", __func__,
243 			ret);
244 		kthread_destroy_worker(kworker_fie);
245 		fie_disabled = FIE_DISABLED;
246 		kworker_fie = NULL;
247 	}
248 }
249 
250 static void __init cppc_freq_invariance_init(void)
251 {
252 	bool perf_ctrs_in_pcc = cppc_perf_ctrs_in_pcc();
253 
254 	if (fie_disabled == FIE_UNSET) {
255 		if (perf_ctrs_in_pcc) {
256 			pr_info("FIE not enabled on systems with registers in PCC\n");
257 			fie_disabled = FIE_DISABLED;
258 		} else {
259 			fie_disabled = FIE_ENABLED;
260 		}
261 	}
262 
263 	if (fie_disabled || !perf_ctrs_in_pcc)
264 		return;
265 
266 	cppc_fie_kworker_init();
267 }
268 
269 static void cppc_freq_invariance_exit(void)
270 {
271 	if (kworker_fie)
272 		kthread_destroy_worker(kworker_fie);
273 }
274 
275 #else
276 static inline void cppc_cpufreq_cpu_fie_init(struct cpufreq_policy *policy)
277 {
278 }
279 
280 static inline void cppc_cpufreq_cpu_fie_exit(struct cpufreq_policy *policy)
281 {
282 }
283 
284 static inline void cppc_freq_invariance_init(void)
285 {
286 }
287 
288 static inline void cppc_freq_invariance_exit(void)
289 {
290 }
291 #endif /* CONFIG_ACPI_CPPC_CPUFREQ_FIE */
292 
293 static void cppc_cpufreq_update_perf_limits(struct cppc_cpudata *cpu_data,
294 					    struct cpufreq_policy *policy)
295 {
296 	struct cppc_perf_caps *caps = &cpu_data->perf_caps;
297 	u32 min_perf, max_perf;
298 
299 	min_perf = cppc_khz_to_perf(caps, policy->min);
300 	max_perf = cppc_khz_to_perf(caps, policy->max);
301 
302 	cpu_data->perf_ctrls.min_perf =
303 		clamp_t(u32, min_perf, caps->lowest_perf, caps->highest_perf);
304 	cpu_data->perf_ctrls.max_perf =
305 		clamp_t(u32, max_perf, caps->lowest_perf, caps->highest_perf);
306 }
307 
308 static int cppc_cpufreq_set_target(struct cpufreq_policy *policy,
309 				   unsigned int target_freq,
310 				   unsigned int relation)
311 {
312 	struct cppc_cpudata *cpu_data = policy->driver_data;
313 	unsigned int cpu = policy->cpu;
314 	struct cpufreq_freqs freqs;
315 	int ret = 0;
316 
317 	cpu_data->perf_ctrls.desired_perf =
318 			cppc_khz_to_perf(&cpu_data->perf_caps, target_freq);
319 	cppc_cpufreq_update_perf_limits(cpu_data, policy);
320 
321 	freqs.old = policy->cur;
322 	freqs.new = target_freq;
323 
324 	cpufreq_freq_transition_begin(policy, &freqs);
325 	ret = cppc_set_perf(cpu, &cpu_data->perf_ctrls);
326 	cpufreq_freq_transition_end(policy, &freqs, ret != 0);
327 
328 	if (ret)
329 		pr_debug("Failed to set target on CPU:%d. ret:%d\n",
330 			 cpu, ret);
331 
332 	return ret;
333 }
334 
335 static unsigned int cppc_cpufreq_fast_switch(struct cpufreq_policy *policy,
336 					      unsigned int target_freq)
337 {
338 	struct cppc_cpudata *cpu_data = policy->driver_data;
339 	unsigned int cpu = policy->cpu;
340 	u32 desired_perf;
341 	int ret;
342 
343 	desired_perf = cppc_khz_to_perf(&cpu_data->perf_caps, target_freq);
344 	cpu_data->perf_ctrls.desired_perf = desired_perf;
345 	cppc_cpufreq_update_perf_limits(cpu_data, policy);
346 
347 	ret = cppc_set_perf(cpu, &cpu_data->perf_ctrls);
348 	if (ret) {
349 		pr_debug("Failed to set target on CPU:%d. ret:%d\n",
350 			 cpu, ret);
351 		return 0;
352 	}
353 
354 	return target_freq;
355 }
356 
357 static int cppc_verify_policy(struct cpufreq_policy_data *policy)
358 {
359 	cpufreq_verify_within_cpu_limits(policy);
360 	return 0;
361 }
362 
363 static unsigned int __cppc_cpufreq_get_transition_delay_us(unsigned int cpu)
364 {
365 	int transition_latency_ns = cppc_get_transition_latency(cpu);
366 
367 	if (transition_latency_ns < 0)
368 		return CPUFREQ_DEFAULT_TRANSITION_LATENCY_NS / NSEC_PER_USEC;
369 
370 	return transition_latency_ns / NSEC_PER_USEC;
371 }
372 
373 /*
374  * The PCC subspace describes the rate at which platform can accept commands
375  * on the shared PCC channel (including READs which do not count towards freq
376  * transition requests), so ideally we need to use the PCC values as a fallback
377  * if we don't have a platform specific transition_delay_us
378  */
379 #ifdef CONFIG_ARM64
380 #include <asm/cputype.h>
381 
382 static unsigned int cppc_cpufreq_get_transition_delay_us(unsigned int cpu)
383 {
384 	unsigned long implementor = read_cpuid_implementor();
385 	unsigned long part_num = read_cpuid_part_number();
386 
387 	switch (implementor) {
388 	case ARM_CPU_IMP_QCOM:
389 		switch (part_num) {
390 		case QCOM_CPU_PART_FALKOR_V1:
391 		case QCOM_CPU_PART_FALKOR:
392 			return 10000;
393 		}
394 	}
395 	return __cppc_cpufreq_get_transition_delay_us(cpu);
396 }
397 #else
398 static unsigned int cppc_cpufreq_get_transition_delay_us(unsigned int cpu)
399 {
400 	return __cppc_cpufreq_get_transition_delay_us(cpu);
401 }
402 #endif
403 
404 #if defined(CONFIG_ARM64) && defined(CONFIG_ENERGY_MODEL)
405 
406 static DEFINE_PER_CPU(unsigned int, efficiency_class);
407 
408 /* Create an artificial performance state every CPPC_EM_CAP_STEP capacity unit. */
409 #define CPPC_EM_CAP_STEP	(20)
410 /* Increase the cost value by CPPC_EM_COST_STEP every performance state. */
411 #define CPPC_EM_COST_STEP	(1)
412 /* Add a cost gap correspnding to the energy of 4 CPUs. */
413 #define CPPC_EM_COST_GAP	(4 * SCHED_CAPACITY_SCALE * CPPC_EM_COST_STEP \
414 				/ CPPC_EM_CAP_STEP)
415 
416 static unsigned int get_perf_level_count(struct cpufreq_policy *policy)
417 {
418 	struct cppc_perf_caps *perf_caps;
419 	unsigned int min_cap, max_cap;
420 	struct cppc_cpudata *cpu_data;
421 	int cpu = policy->cpu;
422 
423 	cpu_data = policy->driver_data;
424 	perf_caps = &cpu_data->perf_caps;
425 	max_cap = arch_scale_cpu_capacity(cpu);
426 	min_cap = div_u64((u64)max_cap * perf_caps->lowest_perf,
427 			  perf_caps->highest_perf);
428 	if ((min_cap == 0) || (max_cap < min_cap))
429 		return 0;
430 	return 1 + max_cap / CPPC_EM_CAP_STEP - min_cap / CPPC_EM_CAP_STEP;
431 }
432 
433 /*
434  * The cost is defined as:
435  *   cost = power * max_frequency / frequency
436  */
437 static inline unsigned long compute_cost(int cpu, int step)
438 {
439 	return CPPC_EM_COST_GAP * per_cpu(efficiency_class, cpu) +
440 			step * CPPC_EM_COST_STEP;
441 }
442 
443 static int cppc_get_cpu_power(struct device *cpu_dev,
444 		unsigned long *power, unsigned long *KHz)
445 {
446 	unsigned long perf_step, perf_prev, perf, perf_check;
447 	unsigned int min_step, max_step, step, step_check;
448 	unsigned long prev_freq = *KHz;
449 	unsigned int min_cap, max_cap;
450 	struct cpufreq_policy *policy;
451 
452 	struct cppc_perf_caps *perf_caps;
453 	struct cppc_cpudata *cpu_data;
454 
455 	policy = cpufreq_cpu_get_raw(cpu_dev->id);
456 	if (!policy)
457 		return -EINVAL;
458 
459 	cpu_data = policy->driver_data;
460 	perf_caps = &cpu_data->perf_caps;
461 	max_cap = arch_scale_cpu_capacity(cpu_dev->id);
462 	min_cap = div_u64((u64)max_cap * perf_caps->lowest_perf,
463 			  perf_caps->highest_perf);
464 	perf_step = div_u64((u64)CPPC_EM_CAP_STEP * perf_caps->highest_perf,
465 			    max_cap);
466 	min_step = min_cap / CPPC_EM_CAP_STEP;
467 	max_step = max_cap / CPPC_EM_CAP_STEP;
468 
469 	perf_prev = cppc_khz_to_perf(perf_caps, *KHz);
470 	step = perf_prev / perf_step;
471 
472 	if (step > max_step)
473 		return -EINVAL;
474 
475 	if (min_step == max_step) {
476 		step = max_step;
477 		perf = perf_caps->highest_perf;
478 	} else if (step < min_step) {
479 		step = min_step;
480 		perf = perf_caps->lowest_perf;
481 	} else {
482 		step++;
483 		if (step == max_step)
484 			perf = perf_caps->highest_perf;
485 		else
486 			perf = step * perf_step;
487 	}
488 
489 	*KHz = cppc_perf_to_khz(perf_caps, perf);
490 	perf_check = cppc_khz_to_perf(perf_caps, *KHz);
491 	step_check = perf_check / perf_step;
492 
493 	/*
494 	 * To avoid bad integer approximation, check that new frequency value
495 	 * increased and that the new frequency will be converted to the
496 	 * desired step value.
497 	 */
498 	while ((*KHz == prev_freq) || (step_check != step)) {
499 		perf++;
500 		*KHz = cppc_perf_to_khz(perf_caps, perf);
501 		perf_check = cppc_khz_to_perf(perf_caps, *KHz);
502 		step_check = perf_check / perf_step;
503 	}
504 
505 	/*
506 	 * With an artificial EM, only the cost value is used. Still the power
507 	 * is populated such as 0 < power < EM_MAX_POWER. This allows to add
508 	 * more sense to the artificial performance states.
509 	 */
510 	*power = compute_cost(cpu_dev->id, step);
511 
512 	return 0;
513 }
514 
515 static int cppc_get_cpu_cost(struct device *cpu_dev, unsigned long KHz,
516 		unsigned long *cost)
517 {
518 	unsigned long perf_step, perf_prev;
519 	struct cppc_perf_caps *perf_caps;
520 	struct cpufreq_policy *policy;
521 	struct cppc_cpudata *cpu_data;
522 	unsigned int max_cap;
523 	int step;
524 
525 	policy = cpufreq_cpu_get_raw(cpu_dev->id);
526 	if (!policy)
527 		return -EINVAL;
528 
529 	cpu_data = policy->driver_data;
530 	perf_caps = &cpu_data->perf_caps;
531 	max_cap = arch_scale_cpu_capacity(cpu_dev->id);
532 
533 	perf_prev = cppc_khz_to_perf(perf_caps, KHz);
534 	perf_step = CPPC_EM_CAP_STEP * perf_caps->highest_perf / max_cap;
535 	step = perf_prev / perf_step;
536 
537 	*cost = compute_cost(cpu_dev->id, step);
538 
539 	return 0;
540 }
541 
542 static void cppc_cpufreq_register_em(struct cpufreq_policy *policy)
543 {
544 	struct cppc_cpudata *cpu_data;
545 	struct em_data_callback em_cb =
546 		EM_ADV_DATA_CB(cppc_get_cpu_power, cppc_get_cpu_cost);
547 
548 	cpu_data = policy->driver_data;
549 	em_dev_register_perf_domain(get_cpu_device(policy->cpu),
550 			get_perf_level_count(policy), &em_cb,
551 			cpu_data->shared_cpu_map, 0);
552 }
553 
554 static void populate_efficiency_class(void)
555 {
556 	struct acpi_madt_generic_interrupt *gicc;
557 	DECLARE_BITMAP(used_classes, 256) = {};
558 	int class, cpu, index;
559 
560 	for_each_possible_cpu(cpu) {
561 		gicc = acpi_cpu_get_madt_gicc(cpu);
562 		class = gicc->efficiency_class;
563 		bitmap_set(used_classes, class, 1);
564 	}
565 
566 	if (bitmap_weight(used_classes, 256) <= 1) {
567 		pr_debug("Efficiency classes are all equal (=%d). "
568 			"No EM registered", class);
569 		return;
570 	}
571 
572 	/*
573 	 * Squeeze efficiency class values on [0:#efficiency_class-1].
574 	 * Values are per spec in [0:255].
575 	 */
576 	index = 0;
577 	for_each_set_bit(class, used_classes, 256) {
578 		for_each_possible_cpu(cpu) {
579 			gicc = acpi_cpu_get_madt_gicc(cpu);
580 			if (gicc->efficiency_class == class)
581 				per_cpu(efficiency_class, cpu) = index;
582 		}
583 		index++;
584 	}
585 	cppc_cpufreq_driver.register_em = cppc_cpufreq_register_em;
586 }
587 
588 #else
589 static void populate_efficiency_class(void)
590 {
591 }
592 #endif
593 
594 static struct cppc_cpudata *cppc_cpufreq_get_cpu_data(unsigned int cpu)
595 {
596 	struct cppc_cpudata *cpu_data;
597 	int ret;
598 
599 	cpu_data = kzalloc_obj(struct cppc_cpudata);
600 	if (!cpu_data)
601 		goto out;
602 
603 	if (!zalloc_cpumask_var(&cpu_data->shared_cpu_map, GFP_KERNEL))
604 		goto free_cpu;
605 
606 	ret = acpi_get_psd_map(cpu, cpu_data);
607 	if (ret) {
608 		pr_debug("Err parsing CPU%d PSD data: ret:%d\n", cpu, ret);
609 		goto free_mask;
610 	}
611 
612 	ret = cppc_get_perf_caps(cpu, &cpu_data->perf_caps);
613 	if (ret) {
614 		pr_debug("Err reading CPU%d perf caps: ret:%d\n", cpu, ret);
615 		goto free_mask;
616 	}
617 
618 	ret = cppc_get_perf(cpu, &cpu_data->perf_ctrls);
619 	if (ret) {
620 		pr_debug("Err reading CPU%d perf ctrls: ret:%d\n", cpu, ret);
621 		goto free_mask;
622 	}
623 
624 	return cpu_data;
625 
626 free_mask:
627 	free_cpumask_var(cpu_data->shared_cpu_map);
628 free_cpu:
629 	kfree(cpu_data);
630 out:
631 	return NULL;
632 }
633 
634 static void cppc_cpufreq_put_cpu_data(struct cpufreq_policy *policy)
635 {
636 	struct cppc_cpudata *cpu_data = policy->driver_data;
637 
638 	free_cpumask_var(cpu_data->shared_cpu_map);
639 	kfree(cpu_data);
640 	policy->driver_data = NULL;
641 }
642 
643 static int cppc_cpufreq_cpu_init(struct cpufreq_policy *policy)
644 {
645 	unsigned int cpu = policy->cpu;
646 	struct cppc_cpudata *cpu_data;
647 	struct cppc_perf_caps *caps;
648 	int ret;
649 
650 	cpu_data = cppc_cpufreq_get_cpu_data(cpu);
651 	if (!cpu_data) {
652 		pr_err("Error in acquiring _CPC/_PSD data for CPU%d.\n", cpu);
653 		return -ENODEV;
654 	}
655 	caps = &cpu_data->perf_caps;
656 	policy->driver_data = cpu_data;
657 
658 	/*
659 	 * Set min to lowest nonlinear perf to avoid any efficiency penalty (see
660 	 * Section 8.4.7.1.1.5 of ACPI 6.1 spec)
661 	 */
662 	policy->min = cppc_perf_to_khz(caps, caps->lowest_nonlinear_perf);
663 
664 	/*
665 	 * Set cpuinfo.min_freq to Lowest to make the full range of performance
666 	 * available if userspace wants to use any perf between lowest & lowest
667 	 * nonlinear perf
668 	 */
669 	policy->cpuinfo.min_freq = cppc_perf_to_khz(caps, caps->lowest_perf);
670 	policy->cpuinfo.max_freq = cppc_perf_to_khz(caps, policy->boost_enabled ?
671 						    caps->highest_perf : caps->nominal_perf);
672 
673 	policy->transition_delay_us = cppc_cpufreq_get_transition_delay_us(cpu);
674 	policy->shared_type = cpu_data->shared_type;
675 
676 	switch (policy->shared_type) {
677 	case CPUFREQ_SHARED_TYPE_HW:
678 	case CPUFREQ_SHARED_TYPE_NONE:
679 		/* Nothing to be done - we'll have a policy for each CPU */
680 		break;
681 	case CPUFREQ_SHARED_TYPE_ANY:
682 		/*
683 		 * All CPUs in the domain will share a policy and all cpufreq
684 		 * operations will use a single cppc_cpudata structure stored
685 		 * in policy->driver_data.
686 		 */
687 		cpumask_copy(policy->cpus, cpu_data->shared_cpu_map);
688 		break;
689 	default:
690 		pr_debug("Unsupported CPU co-ord type: %d\n",
691 			 policy->shared_type);
692 		ret = -EFAULT;
693 		goto out;
694 	}
695 
696 	policy->fast_switch_possible = cppc_allow_fast_switch();
697 	policy->dvfs_possible_from_any_cpu = true;
698 
699 	/*
700 	 * If 'highest_perf' is greater than 'nominal_perf', we assume CPU Boost
701 	 * is supported.
702 	 */
703 	if (caps->highest_perf > caps->nominal_perf)
704 		policy->boost_supported = true;
705 
706 	/* Set policy->cur to max now. The governors will adjust later. */
707 	policy->cur = cppc_perf_to_khz(caps, caps->highest_perf);
708 	cpu_data->perf_ctrls.desired_perf =  caps->highest_perf;
709 
710 	ret = cppc_set_perf(cpu, &cpu_data->perf_ctrls);
711 	if (ret) {
712 		pr_debug("Err setting perf value:%d on CPU:%d. ret:%d\n",
713 			 caps->highest_perf, cpu, ret);
714 		goto out;
715 	}
716 
717 	cppc_cpufreq_cpu_fie_init(policy);
718 	return 0;
719 
720 out:
721 	cppc_cpufreq_put_cpu_data(policy);
722 	return ret;
723 }
724 
725 static void cppc_cpufreq_cpu_exit(struct cpufreq_policy *policy)
726 {
727 	struct cppc_cpudata *cpu_data = policy->driver_data;
728 	struct cppc_perf_caps *caps = &cpu_data->perf_caps;
729 	unsigned int cpu = policy->cpu;
730 	int ret;
731 
732 	cppc_cpufreq_cpu_fie_exit(policy);
733 
734 	cpu_data->perf_ctrls.desired_perf = caps->lowest_perf;
735 
736 	ret = cppc_set_perf(cpu, &cpu_data->perf_ctrls);
737 	if (ret)
738 		pr_debug("Err setting perf value:%d on CPU:%d. ret:%d\n",
739 			 caps->lowest_perf, cpu, ret);
740 
741 	cppc_cpufreq_put_cpu_data(policy);
742 }
743 
744 static inline u64 get_delta(u64 t1, u64 t0)
745 {
746 	if (t1 > t0 || t0 > ~(u32)0)
747 		return t1 - t0;
748 
749 	return (u32)t1 - (u32)t0;
750 }
751 
752 static int cppc_perf_from_fbctrs(u64 reference_perf,
753 				 struct cppc_perf_fb_ctrs *fb_ctrs_t0,
754 				 struct cppc_perf_fb_ctrs *fb_ctrs_t1)
755 {
756 	u64 delta_reference, delta_delivered;
757 
758 	delta_reference = get_delta(fb_ctrs_t1->reference,
759 				    fb_ctrs_t0->reference);
760 	delta_delivered = get_delta(fb_ctrs_t1->delivered,
761 				    fb_ctrs_t0->delivered);
762 
763 	/*
764 	 * Avoid divide-by zero and unchanged feedback counters.
765 	 * Leave it for callers to handle.
766 	 */
767 	if (!delta_reference || !delta_delivered)
768 		return 0;
769 
770 	return (reference_perf * delta_delivered) / delta_reference;
771 }
772 
773 static int cppc_get_perf_ctrs_sample(int cpu,
774 				     struct cppc_perf_fb_ctrs *fb_ctrs_t0,
775 				     struct cppc_perf_fb_ctrs *fb_ctrs_t1)
776 {
777 	int ret;
778 
779 	ret = cppc_get_perf_ctrs(cpu, fb_ctrs_t0);
780 	if (ret)
781 		return ret;
782 
783 	udelay(2); /* 2usec delay between sampling */
784 
785 	return cppc_get_perf_ctrs(cpu, fb_ctrs_t1);
786 }
787 
788 static unsigned int cppc_cpufreq_get_rate(unsigned int cpu)
789 {
790 	struct cpufreq_policy *policy __free(put_cpufreq_policy) = cpufreq_cpu_get(cpu);
791 	struct cppc_perf_fb_ctrs fb_ctrs_t0 = {0}, fb_ctrs_t1 = {0};
792 	struct cppc_cpudata *cpu_data;
793 	u64 delivered_perf, reference_perf;
794 	int ret;
795 
796 	if (!policy)
797 		return 0;
798 
799 	cpu_data = policy->driver_data;
800 
801 	ret = cppc_get_perf_ctrs_sample(cpu, &fb_ctrs_t0, &fb_ctrs_t1);
802 	if (ret) {
803 		if (ret == -EFAULT)
804 			/* Any of the associated CPPC regs is 0. */
805 			goto out_invalid_counters;
806 		else
807 			return 0;
808 	}
809 
810 	reference_perf = cpu_data->perf_caps.reference_perf;
811 	delivered_perf = cppc_perf_from_fbctrs(reference_perf,
812 					       &fb_ctrs_t0, &fb_ctrs_t1);
813 	if (!delivered_perf)
814 		goto out_invalid_counters;
815 
816 	return cppc_perf_to_khz(&cpu_data->perf_caps, delivered_perf);
817 
818 out_invalid_counters:
819 	/*
820 	 * Feedback counters could be unchanged or 0 when a cpu enters a
821 	 * low-power idle state, e.g. clock-gated or power-gated.
822 	 * Use desired perf for reflecting frequency.  Get the latest register
823 	 * value first as some platforms may update the actual delivered perf
824 	 * there; if failed, resort to the cached desired perf.
825 	 */
826 	if (cppc_get_desired_perf(cpu, &delivered_perf))
827 		delivered_perf = cpu_data->perf_ctrls.desired_perf;
828 
829 	return cppc_perf_to_khz(&cpu_data->perf_caps, delivered_perf);
830 }
831 
832 static int cppc_cpufreq_set_boost(struct cpufreq_policy *policy, int state)
833 {
834 	struct cppc_cpudata *cpu_data = policy->driver_data;
835 	struct cppc_perf_caps *caps = &cpu_data->perf_caps;
836 
837 	if (state)
838 		policy->cpuinfo.max_freq = cppc_perf_to_khz(caps, caps->highest_perf);
839 	else
840 		policy->cpuinfo.max_freq = cppc_perf_to_khz(caps, caps->nominal_perf);
841 
842 	return 0;
843 }
844 
845 static ssize_t show_freqdomain_cpus(struct cpufreq_policy *policy, char *buf)
846 {
847 	struct cppc_cpudata *cpu_data = policy->driver_data;
848 
849 	return cpufreq_show_cpus(cpu_data->shared_cpu_map, buf);
850 }
851 
852 static ssize_t show_auto_select(struct cpufreq_policy *policy, char *buf)
853 {
854 	bool val;
855 	int ret;
856 
857 	ret = cppc_get_auto_sel(policy->cpu, &val);
858 
859 	/* show "<unsupported>" when this register is not supported by cpc */
860 	if (ret == -EOPNOTSUPP)
861 		return sysfs_emit(buf, "<unsupported>\n");
862 
863 	if (ret)
864 		return ret;
865 
866 	return sysfs_emit(buf, "%d\n", val);
867 }
868 
869 static ssize_t store_auto_select(struct cpufreq_policy *policy,
870 				 const char *buf, size_t count)
871 {
872 	struct cppc_cpudata *cpu_data = policy->driver_data;
873 	bool val;
874 	int ret;
875 
876 	ret = kstrtobool(buf, &val);
877 	if (ret)
878 		return ret;
879 
880 	ret = cppc_set_auto_sel(policy->cpu, val);
881 	if (ret)
882 		return ret;
883 
884 	cpu_data->perf_ctrls.auto_sel = val;
885 
886 	if (val) {
887 		u32 old_min_perf = cpu_data->perf_ctrls.min_perf;
888 		u32 old_max_perf = cpu_data->perf_ctrls.max_perf;
889 
890 		/*
891 		 * When enabling autonomous selection, program MIN_PERF and
892 		 * MAX_PERF from current policy limits so that the platform
893 		 * uses the correct performance bounds immediately.
894 		 */
895 		cppc_cpufreq_update_perf_limits(cpu_data, policy);
896 
897 		ret = cppc_set_perf(policy->cpu, &cpu_data->perf_ctrls);
898 		if (ret) {
899 			cpu_data->perf_ctrls.min_perf = old_min_perf;
900 			cpu_data->perf_ctrls.max_perf = old_max_perf;
901 			cppc_set_auto_sel(policy->cpu, false);
902 			cpu_data->perf_ctrls.auto_sel = false;
903 			return ret;
904 		}
905 	}
906 
907 	return count;
908 }
909 
910 static ssize_t cppc_cpufreq_sysfs_show_u64(unsigned int cpu,
911 					   int (*get_func)(int, u64 *),
912 					   char *buf)
913 {
914 	u64 val;
915 	int ret = get_func((int)cpu, &val);
916 
917 	if (ret == -EOPNOTSUPP)
918 		return sysfs_emit(buf, "<unsupported>\n");
919 
920 	if (ret)
921 		return ret;
922 
923 	return sysfs_emit(buf, "%llu\n", val);
924 }
925 
926 static ssize_t cppc_cpufreq_sysfs_store_u64(unsigned int cpu,
927 					    int (*set_func)(int, u64),
928 					    const char *buf, size_t count)
929 {
930 	u64 val;
931 	int ret;
932 
933 	ret = kstrtou64(buf, 0, &val);
934 	if (ret)
935 		return ret;
936 
937 	ret = set_func((int)cpu, val);
938 
939 	return ret ? ret : count;
940 }
941 
942 #define CPPC_CPUFREQ_ATTR_RW_U64(_name, _get_func, _set_func)		\
943 static ssize_t show_##_name(struct cpufreq_policy *policy, char *buf)	\
944 {									\
945 	return cppc_cpufreq_sysfs_show_u64(policy->cpu, _get_func, buf);\
946 }									\
947 static ssize_t store_##_name(struct cpufreq_policy *policy,		\
948 			     const char *buf, size_t count)		\
949 {									\
950 	return cppc_cpufreq_sysfs_store_u64(policy->cpu, _set_func,	\
951 					    buf, count);		\
952 }
953 
954 CPPC_CPUFREQ_ATTR_RW_U64(auto_act_window, cppc_get_auto_act_window,
955 			 cppc_set_auto_act_window)
956 
957 static ssize_t
958 show_energy_performance_preference_val(struct cpufreq_policy *policy, char *buf)
959 {
960 	return cppc_cpufreq_sysfs_show_u64(policy->cpu, cppc_get_epp_perf, buf);
961 }
962 
963 static ssize_t
964 store_energy_performance_preference_val(struct cpufreq_policy *policy,
965 					const char *buf, size_t count)
966 {
967 	struct cppc_cpudata *cpu_data = policy->driver_data;
968 	u64 val;
969 	int ret;
970 
971 	ret = kstrtou64(buf, 0, &val);
972 	if (ret)
973 		return ret;
974 
975 	ret = cppc_set_epp(policy->cpu, val);
976 	if (ret)
977 		return ret;
978 
979 	cpu_data->perf_ctrls.energy_perf = val;
980 
981 	return count;
982 }
983 
984 static int cppc_get_perf_limited_filtered(int cpu, u64 *perf_limited)
985 {
986 	struct cpufreq_policy *policy;
987 	struct cppc_cpudata *cpu_data;
988 	int ret;
989 
990 	ret = cppc_get_perf_limited(cpu, perf_limited);
991 	if (ret)
992 		return ret;
993 
994 	policy = cpufreq_cpu_get_raw(cpu);
995 	if (!policy)
996 		return -EINVAL;
997 
998 	cpu_data = policy->driver_data;
999 
1000 	/*
1001 	 * Desired Excursion is ignored when autonomous selection is
1002 	 * enabled. Clear the bit to avoid exposing meaningless state
1003 	 * to userspace.
1004 	 */
1005 	if (cpu_data && cpu_data->perf_ctrls.auto_sel)
1006 		*perf_limited &= ~CPPC_PERF_LIMITED_DESIRED_EXCURSION;
1007 
1008 	return 0;
1009 }
1010 
1011 CPPC_CPUFREQ_ATTR_RW_U64(perf_limited, cppc_get_perf_limited_filtered,
1012 			 cppc_set_perf_limited)
1013 
1014 cpufreq_freq_attr_ro(freqdomain_cpus);
1015 cpufreq_freq_attr_rw(auto_select);
1016 cpufreq_freq_attr_rw(auto_act_window);
1017 cpufreq_freq_attr_rw(energy_performance_preference_val);
1018 cpufreq_freq_attr_rw(perf_limited);
1019 
1020 static struct freq_attr *cppc_cpufreq_attr[] = {
1021 	&freqdomain_cpus,
1022 	&auto_select,
1023 	&auto_act_window,
1024 	&energy_performance_preference_val,
1025 	&perf_limited,
1026 	NULL,
1027 };
1028 
1029 static struct cpufreq_driver cppc_cpufreq_driver = {
1030 	.flags = CPUFREQ_CONST_LOOPS | CPUFREQ_NEED_UPDATE_LIMITS,
1031 	.verify = cppc_verify_policy,
1032 	.target = cppc_cpufreq_set_target,
1033 	.get = cppc_cpufreq_get_rate,
1034 	.fast_switch = cppc_cpufreq_fast_switch,
1035 	.init = cppc_cpufreq_cpu_init,
1036 	.exit = cppc_cpufreq_cpu_exit,
1037 	.set_boost = cppc_cpufreq_set_boost,
1038 	.attr = cppc_cpufreq_attr,
1039 	.name = "cppc_cpufreq",
1040 };
1041 
1042 static int __init cppc_cpufreq_init(void)
1043 {
1044 	int ret;
1045 
1046 	if (!acpi_cpc_valid())
1047 		return -ENODEV;
1048 
1049 	cppc_freq_invariance_init();
1050 	populate_efficiency_class();
1051 
1052 	ret = cpufreq_register_driver(&cppc_cpufreq_driver);
1053 	if (ret)
1054 		cppc_freq_invariance_exit();
1055 
1056 	return ret;
1057 }
1058 
1059 static void __exit cppc_cpufreq_exit(void)
1060 {
1061 	cpufreq_unregister_driver(&cppc_cpufreq_driver);
1062 	cppc_freq_invariance_exit();
1063 }
1064 
1065 module_exit(cppc_cpufreq_exit);
1066 MODULE_AUTHOR("Ashwin Chaugule");
1067 MODULE_DESCRIPTION("CPUFreq driver based on the ACPI CPPC v5.0+ spec");
1068 MODULE_LICENSE("GPL");
1069 
1070 late_initcall(cppc_cpufreq_init);
1071 
1072 static const struct acpi_device_id cppc_acpi_ids[] __used = {
1073 	{ACPI_PROCESSOR_DEVICE_HID, },
1074 	{}
1075 };
1076 
1077 MODULE_DEVICE_TABLE(acpi, cppc_acpi_ids);
1078