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 <asm/unaligned.h>
26 
27 #include <acpi/cppc_acpi.h>
28 
29 /*
30  * This list contains information parsed from per CPU ACPI _CPC and _PSD
31  * structures: e.g. the highest and lowest supported performance, capabilities,
32  * desired performance, level requested etc. Depending on the share_type, not
33  * all CPUs will have an entry in the list.
34  */
35 static LIST_HEAD(cpu_data_list);
36 
37 static bool boost_supported;
38 
39 struct cppc_workaround_oem_info {
40 	char oem_id[ACPI_OEM_ID_SIZE + 1];
41 	char oem_table_id[ACPI_OEM_TABLE_ID_SIZE + 1];
42 	u32 oem_revision;
43 };
44 
45 static struct cppc_workaround_oem_info wa_info[] = {
46 	{
47 		.oem_id		= "HISI  ",
48 		.oem_table_id	= "HIP07   ",
49 		.oem_revision	= 0,
50 	}, {
51 		.oem_id		= "HISI  ",
52 		.oem_table_id	= "HIP08   ",
53 		.oem_revision	= 0,
54 	}
55 };
56 
57 static struct cpufreq_driver cppc_cpufreq_driver;
58 
59 static enum {
60 	FIE_UNSET = -1,
61 	FIE_ENABLED,
62 	FIE_DISABLED
63 } fie_disabled = FIE_UNSET;
64 
65 #ifdef CONFIG_ACPI_CPPC_CPUFREQ_FIE
66 module_param(fie_disabled, int, 0444);
67 MODULE_PARM_DESC(fie_disabled, "Disable Frequency Invariance Engine (FIE)");
68 
69 /* Frequency invariance support */
70 struct cppc_freq_invariance {
71 	int cpu;
72 	struct irq_work irq_work;
73 	struct kthread_work work;
74 	struct cppc_perf_fb_ctrs prev_perf_fb_ctrs;
75 	struct cppc_cpudata *cpu_data;
76 };
77 
78 static DEFINE_PER_CPU(struct cppc_freq_invariance, cppc_freq_inv);
79 static struct kthread_worker *kworker_fie;
80 
81 static unsigned int hisi_cppc_cpufreq_get_rate(unsigned int cpu);
82 static int cppc_perf_from_fbctrs(struct cppc_cpudata *cpu_data,
83 				 struct cppc_perf_fb_ctrs *fb_ctrs_t0,
84 				 struct cppc_perf_fb_ctrs *fb_ctrs_t1);
85 
86 /**
87  * cppc_scale_freq_workfn - CPPC arch_freq_scale updater for frequency invariance
88  * @work: The work item.
89  *
90  * The CPPC driver register itself with the topology core to provide its own
91  * implementation (cppc_scale_freq_tick()) of topology_scale_freq_tick() which
92  * gets called by the scheduler on every tick.
93  *
94  * Note that the arch specific counters have higher priority than CPPC counters,
95  * if available, though the CPPC driver doesn't need to have any special
96  * handling for that.
97  *
98  * On an invocation of cppc_scale_freq_tick(), we schedule an irq work (since we
99  * reach here from hard-irq context), which then schedules a normal work item
100  * and cppc_scale_freq_workfn() updates the per_cpu arch_freq_scale variable
101  * based on the counter updates since the last tick.
102  */
103 static void cppc_scale_freq_workfn(struct kthread_work *work)
104 {
105 	struct cppc_freq_invariance *cppc_fi;
106 	struct cppc_perf_fb_ctrs fb_ctrs = {0};
107 	struct cppc_cpudata *cpu_data;
108 	unsigned long local_freq_scale;
109 	u64 perf;
110 
111 	cppc_fi = container_of(work, struct cppc_freq_invariance, work);
112 	cpu_data = cppc_fi->cpu_data;
113 
114 	if (cppc_get_perf_ctrs(cppc_fi->cpu, &fb_ctrs)) {
115 		pr_warn("%s: failed to read perf counters\n", __func__);
116 		return;
117 	}
118 
119 	perf = cppc_perf_from_fbctrs(cpu_data, &cppc_fi->prev_perf_fb_ctrs,
120 				     &fb_ctrs);
121 	cppc_fi->prev_perf_fb_ctrs = fb_ctrs;
122 
123 	perf <<= SCHED_CAPACITY_SHIFT;
124 	local_freq_scale = div64_u64(perf, cpu_data->perf_caps.highest_perf);
125 
126 	/* This can happen due to counter's overflow */
127 	if (unlikely(local_freq_scale > 1024))
128 		local_freq_scale = 1024;
129 
130 	per_cpu(arch_freq_scale, cppc_fi->cpu) = local_freq_scale;
131 }
132 
133 static void cppc_irq_work(struct irq_work *irq_work)
134 {
135 	struct cppc_freq_invariance *cppc_fi;
136 
137 	cppc_fi = container_of(irq_work, struct cppc_freq_invariance, irq_work);
138 	kthread_queue_work(kworker_fie, &cppc_fi->work);
139 }
140 
141 static void cppc_scale_freq_tick(void)
142 {
143 	struct cppc_freq_invariance *cppc_fi = &per_cpu(cppc_freq_inv, smp_processor_id());
144 
145 	/*
146 	 * cppc_get_perf_ctrs() can potentially sleep, call that from the right
147 	 * context.
148 	 */
149 	irq_work_queue(&cppc_fi->irq_work);
150 }
151 
152 static struct scale_freq_data cppc_sftd = {
153 	.source = SCALE_FREQ_SOURCE_CPPC,
154 	.set_freq_scale = cppc_scale_freq_tick,
155 };
156 
157 static void cppc_cpufreq_cpu_fie_init(struct cpufreq_policy *policy)
158 {
159 	struct cppc_freq_invariance *cppc_fi;
160 	int cpu, ret;
161 
162 	if (fie_disabled)
163 		return;
164 
165 	for_each_cpu(cpu, policy->cpus) {
166 		cppc_fi = &per_cpu(cppc_freq_inv, cpu);
167 		cppc_fi->cpu = cpu;
168 		cppc_fi->cpu_data = policy->driver_data;
169 		kthread_init_work(&cppc_fi->work, cppc_scale_freq_workfn);
170 		init_irq_work(&cppc_fi->irq_work, cppc_irq_work);
171 
172 		ret = cppc_get_perf_ctrs(cpu, &cppc_fi->prev_perf_fb_ctrs);
173 		if (ret) {
174 			pr_warn("%s: failed to read perf counters for cpu:%d: %d\n",
175 				__func__, cpu, ret);
176 
177 			/*
178 			 * Don't abort if the CPU was offline while the driver
179 			 * was getting registered.
180 			 */
181 			if (cpu_online(cpu))
182 				return;
183 		}
184 	}
185 
186 	/* Register for freq-invariance */
187 	topology_set_scale_freq_source(&cppc_sftd, policy->cpus);
188 }
189 
190 /*
191  * We free all the resources on policy's removal and not on CPU removal as the
192  * irq-work are per-cpu and the hotplug core takes care of flushing the pending
193  * irq-works (hint: smpcfd_dying_cpu()) on CPU hotplug. Even if the kthread-work
194  * fires on another CPU after the concerned CPU is removed, it won't harm.
195  *
196  * We just need to make sure to remove them all on policy->exit().
197  */
198 static void cppc_cpufreq_cpu_fie_exit(struct cpufreq_policy *policy)
199 {
200 	struct cppc_freq_invariance *cppc_fi;
201 	int cpu;
202 
203 	if (fie_disabled)
204 		return;
205 
206 	/* policy->cpus will be empty here, use related_cpus instead */
207 	topology_clear_scale_freq_source(SCALE_FREQ_SOURCE_CPPC, policy->related_cpus);
208 
209 	for_each_cpu(cpu, policy->related_cpus) {
210 		cppc_fi = &per_cpu(cppc_freq_inv, cpu);
211 		irq_work_sync(&cppc_fi->irq_work);
212 		kthread_cancel_work_sync(&cppc_fi->work);
213 	}
214 }
215 
216 static void __init cppc_freq_invariance_init(void)
217 {
218 	struct sched_attr attr = {
219 		.size		= sizeof(struct sched_attr),
220 		.sched_policy	= SCHED_DEADLINE,
221 		.sched_nice	= 0,
222 		.sched_priority	= 0,
223 		/*
224 		 * Fake (unused) bandwidth; workaround to "fix"
225 		 * priority inheritance.
226 		 */
227 		.sched_runtime	= 1000000,
228 		.sched_deadline = 10000000,
229 		.sched_period	= 10000000,
230 	};
231 	int ret;
232 
233 	if (fie_disabled != FIE_ENABLED && fie_disabled != FIE_DISABLED) {
234 		fie_disabled = FIE_ENABLED;
235 		if (cppc_perf_ctrs_in_pcc()) {
236 			pr_info("FIE not enabled on systems with registers in PCC\n");
237 			fie_disabled = FIE_DISABLED;
238 		}
239 	}
240 
241 	if (fie_disabled)
242 		return;
243 
244 	kworker_fie = kthread_create_worker(0, "cppc_fie");
245 	if (IS_ERR(kworker_fie)) {
246 		pr_warn("%s: failed to create kworker_fie: %ld\n", __func__,
247 			PTR_ERR(kworker_fie));
248 		fie_disabled = FIE_DISABLED;
249 		return;
250 	}
251 
252 	ret = sched_setattr_nocheck(kworker_fie->task, &attr);
253 	if (ret) {
254 		pr_warn("%s: failed to set SCHED_DEADLINE: %d\n", __func__,
255 			ret);
256 		kthread_destroy_worker(kworker_fie);
257 		fie_disabled = FIE_DISABLED;
258 	}
259 }
260 
261 static void cppc_freq_invariance_exit(void)
262 {
263 	if (fie_disabled)
264 		return;
265 
266 	kthread_destroy_worker(kworker_fie);
267 }
268 
269 #else
270 static inline void cppc_cpufreq_cpu_fie_init(struct cpufreq_policy *policy)
271 {
272 }
273 
274 static inline void cppc_cpufreq_cpu_fie_exit(struct cpufreq_policy *policy)
275 {
276 }
277 
278 static inline void cppc_freq_invariance_init(void)
279 {
280 }
281 
282 static inline void cppc_freq_invariance_exit(void)
283 {
284 }
285 #endif /* CONFIG_ACPI_CPPC_CPUFREQ_FIE */
286 
287 static int cppc_cpufreq_set_target(struct cpufreq_policy *policy,
288 				   unsigned int target_freq,
289 				   unsigned int relation)
290 {
291 	struct cppc_cpudata *cpu_data = policy->driver_data;
292 	unsigned int cpu = policy->cpu;
293 	struct cpufreq_freqs freqs;
294 	u32 desired_perf;
295 	int ret = 0;
296 
297 	desired_perf = cppc_khz_to_perf(&cpu_data->perf_caps, target_freq);
298 	/* Return if it is exactly the same perf */
299 	if (desired_perf == cpu_data->perf_ctrls.desired_perf)
300 		return ret;
301 
302 	cpu_data->perf_ctrls.desired_perf = desired_perf;
303 	freqs.old = policy->cur;
304 	freqs.new = target_freq;
305 
306 	cpufreq_freq_transition_begin(policy, &freqs);
307 	ret = cppc_set_perf(cpu, &cpu_data->perf_ctrls);
308 	cpufreq_freq_transition_end(policy, &freqs, ret != 0);
309 
310 	if (ret)
311 		pr_debug("Failed to set target on CPU:%d. ret:%d\n",
312 			 cpu, ret);
313 
314 	return ret;
315 }
316 
317 static unsigned int cppc_cpufreq_fast_switch(struct cpufreq_policy *policy,
318 					      unsigned int target_freq)
319 {
320 	struct cppc_cpudata *cpu_data = policy->driver_data;
321 	unsigned int cpu = policy->cpu;
322 	u32 desired_perf;
323 	int ret;
324 
325 	desired_perf = cppc_khz_to_perf(&cpu_data->perf_caps, target_freq);
326 	cpu_data->perf_ctrls.desired_perf = desired_perf;
327 	ret = cppc_set_perf(cpu, &cpu_data->perf_ctrls);
328 
329 	if (ret) {
330 		pr_debug("Failed to set target on CPU:%d. ret:%d\n",
331 			 cpu, ret);
332 		return 0;
333 	}
334 
335 	return target_freq;
336 }
337 
338 static int cppc_verify_policy(struct cpufreq_policy_data *policy)
339 {
340 	cpufreq_verify_within_cpu_limits(policy);
341 	return 0;
342 }
343 
344 /*
345  * The PCC subspace describes the rate at which platform can accept commands
346  * on the shared PCC channel (including READs which do not count towards freq
347  * transition requests), so ideally we need to use the PCC values as a fallback
348  * if we don't have a platform specific transition_delay_us
349  */
350 #ifdef CONFIG_ARM64
351 #include <asm/cputype.h>
352 
353 static unsigned int cppc_cpufreq_get_transition_delay_us(unsigned int cpu)
354 {
355 	unsigned long implementor = read_cpuid_implementor();
356 	unsigned long part_num = read_cpuid_part_number();
357 
358 	switch (implementor) {
359 	case ARM_CPU_IMP_QCOM:
360 		switch (part_num) {
361 		case QCOM_CPU_PART_FALKOR_V1:
362 		case QCOM_CPU_PART_FALKOR:
363 			return 10000;
364 		}
365 	}
366 	return cppc_get_transition_latency(cpu) / NSEC_PER_USEC;
367 }
368 #else
369 static unsigned int cppc_cpufreq_get_transition_delay_us(unsigned int cpu)
370 {
371 	return cppc_get_transition_latency(cpu) / NSEC_PER_USEC;
372 }
373 #endif
374 
375 #if defined(CONFIG_ARM64) && defined(CONFIG_ENERGY_MODEL)
376 
377 static DEFINE_PER_CPU(unsigned int, efficiency_class);
378 static void cppc_cpufreq_register_em(struct cpufreq_policy *policy);
379 
380 /* Create an artificial performance state every CPPC_EM_CAP_STEP capacity unit. */
381 #define CPPC_EM_CAP_STEP	(20)
382 /* Increase the cost value by CPPC_EM_COST_STEP every performance state. */
383 #define CPPC_EM_COST_STEP	(1)
384 /* Add a cost gap correspnding to the energy of 4 CPUs. */
385 #define CPPC_EM_COST_GAP	(4 * SCHED_CAPACITY_SCALE * CPPC_EM_COST_STEP \
386 				/ CPPC_EM_CAP_STEP)
387 
388 static unsigned int get_perf_level_count(struct cpufreq_policy *policy)
389 {
390 	struct cppc_perf_caps *perf_caps;
391 	unsigned int min_cap, max_cap;
392 	struct cppc_cpudata *cpu_data;
393 	int cpu = policy->cpu;
394 
395 	cpu_data = policy->driver_data;
396 	perf_caps = &cpu_data->perf_caps;
397 	max_cap = arch_scale_cpu_capacity(cpu);
398 	min_cap = div_u64((u64)max_cap * perf_caps->lowest_perf,
399 			  perf_caps->highest_perf);
400 	if ((min_cap == 0) || (max_cap < min_cap))
401 		return 0;
402 	return 1 + max_cap / CPPC_EM_CAP_STEP - min_cap / CPPC_EM_CAP_STEP;
403 }
404 
405 /*
406  * The cost is defined as:
407  *   cost = power * max_frequency / frequency
408  */
409 static inline unsigned long compute_cost(int cpu, int step)
410 {
411 	return CPPC_EM_COST_GAP * per_cpu(efficiency_class, cpu) +
412 			step * CPPC_EM_COST_STEP;
413 }
414 
415 static int cppc_get_cpu_power(struct device *cpu_dev,
416 		unsigned long *power, unsigned long *KHz)
417 {
418 	unsigned long perf_step, perf_prev, perf, perf_check;
419 	unsigned int min_step, max_step, step, step_check;
420 	unsigned long prev_freq = *KHz;
421 	unsigned int min_cap, max_cap;
422 	struct cpufreq_policy *policy;
423 
424 	struct cppc_perf_caps *perf_caps;
425 	struct cppc_cpudata *cpu_data;
426 
427 	policy = cpufreq_cpu_get_raw(cpu_dev->id);
428 	cpu_data = policy->driver_data;
429 	perf_caps = &cpu_data->perf_caps;
430 	max_cap = arch_scale_cpu_capacity(cpu_dev->id);
431 	min_cap = div_u64((u64)max_cap * perf_caps->lowest_perf,
432 			  perf_caps->highest_perf);
433 	perf_step = div_u64((u64)CPPC_EM_CAP_STEP * perf_caps->highest_perf,
434 			    max_cap);
435 	min_step = min_cap / CPPC_EM_CAP_STEP;
436 	max_step = max_cap / CPPC_EM_CAP_STEP;
437 
438 	perf_prev = cppc_khz_to_perf(perf_caps, *KHz);
439 	step = perf_prev / perf_step;
440 
441 	if (step > max_step)
442 		return -EINVAL;
443 
444 	if (min_step == max_step) {
445 		step = max_step;
446 		perf = perf_caps->highest_perf;
447 	} else if (step < min_step) {
448 		step = min_step;
449 		perf = perf_caps->lowest_perf;
450 	} else {
451 		step++;
452 		if (step == max_step)
453 			perf = perf_caps->highest_perf;
454 		else
455 			perf = step * perf_step;
456 	}
457 
458 	*KHz = cppc_perf_to_khz(perf_caps, perf);
459 	perf_check = cppc_khz_to_perf(perf_caps, *KHz);
460 	step_check = perf_check / perf_step;
461 
462 	/*
463 	 * To avoid bad integer approximation, check that new frequency value
464 	 * increased and that the new frequency will be converted to the
465 	 * desired step value.
466 	 */
467 	while ((*KHz == prev_freq) || (step_check != step)) {
468 		perf++;
469 		*KHz = cppc_perf_to_khz(perf_caps, perf);
470 		perf_check = cppc_khz_to_perf(perf_caps, *KHz);
471 		step_check = perf_check / perf_step;
472 	}
473 
474 	/*
475 	 * With an artificial EM, only the cost value is used. Still the power
476 	 * is populated such as 0 < power < EM_MAX_POWER. This allows to add
477 	 * more sense to the artificial performance states.
478 	 */
479 	*power = compute_cost(cpu_dev->id, step);
480 
481 	return 0;
482 }
483 
484 static int cppc_get_cpu_cost(struct device *cpu_dev, unsigned long KHz,
485 		unsigned long *cost)
486 {
487 	unsigned long perf_step, perf_prev;
488 	struct cppc_perf_caps *perf_caps;
489 	struct cpufreq_policy *policy;
490 	struct cppc_cpudata *cpu_data;
491 	unsigned int max_cap;
492 	int step;
493 
494 	policy = cpufreq_cpu_get_raw(cpu_dev->id);
495 	cpu_data = policy->driver_data;
496 	perf_caps = &cpu_data->perf_caps;
497 	max_cap = arch_scale_cpu_capacity(cpu_dev->id);
498 
499 	perf_prev = cppc_khz_to_perf(perf_caps, KHz);
500 	perf_step = CPPC_EM_CAP_STEP * perf_caps->highest_perf / max_cap;
501 	step = perf_prev / perf_step;
502 
503 	*cost = compute_cost(cpu_dev->id, step);
504 
505 	return 0;
506 }
507 
508 static int populate_efficiency_class(void)
509 {
510 	struct acpi_madt_generic_interrupt *gicc;
511 	DECLARE_BITMAP(used_classes, 256) = {};
512 	int class, cpu, index;
513 
514 	for_each_possible_cpu(cpu) {
515 		gicc = acpi_cpu_get_madt_gicc(cpu);
516 		class = gicc->efficiency_class;
517 		bitmap_set(used_classes, class, 1);
518 	}
519 
520 	if (bitmap_weight(used_classes, 256) <= 1) {
521 		pr_debug("Efficiency classes are all equal (=%d). "
522 			"No EM registered", class);
523 		return -EINVAL;
524 	}
525 
526 	/*
527 	 * Squeeze efficiency class values on [0:#efficiency_class-1].
528 	 * Values are per spec in [0:255].
529 	 */
530 	index = 0;
531 	for_each_set_bit(class, used_classes, 256) {
532 		for_each_possible_cpu(cpu) {
533 			gicc = acpi_cpu_get_madt_gicc(cpu);
534 			if (gicc->efficiency_class == class)
535 				per_cpu(efficiency_class, cpu) = index;
536 		}
537 		index++;
538 	}
539 	cppc_cpufreq_driver.register_em = cppc_cpufreq_register_em;
540 
541 	return 0;
542 }
543 
544 static void cppc_cpufreq_register_em(struct cpufreq_policy *policy)
545 {
546 	struct cppc_cpudata *cpu_data;
547 	struct em_data_callback em_cb =
548 		EM_ADV_DATA_CB(cppc_get_cpu_power, cppc_get_cpu_cost);
549 
550 	cpu_data = policy->driver_data;
551 	em_dev_register_perf_domain(get_cpu_device(policy->cpu),
552 			get_perf_level_count(policy), &em_cb,
553 			cpu_data->shared_cpu_map, 0);
554 }
555 
556 #else
557 static int populate_efficiency_class(void)
558 {
559 	return 0;
560 }
561 #endif
562 
563 static struct cppc_cpudata *cppc_cpufreq_get_cpu_data(unsigned int cpu)
564 {
565 	struct cppc_cpudata *cpu_data;
566 	int ret;
567 
568 	cpu_data = kzalloc(sizeof(struct cppc_cpudata), GFP_KERNEL);
569 	if (!cpu_data)
570 		goto out;
571 
572 	if (!zalloc_cpumask_var(&cpu_data->shared_cpu_map, GFP_KERNEL))
573 		goto free_cpu;
574 
575 	ret = acpi_get_psd_map(cpu, cpu_data);
576 	if (ret) {
577 		pr_debug("Err parsing CPU%d PSD data: ret:%d\n", cpu, ret);
578 		goto free_mask;
579 	}
580 
581 	ret = cppc_get_perf_caps(cpu, &cpu_data->perf_caps);
582 	if (ret) {
583 		pr_debug("Err reading CPU%d perf caps: ret:%d\n", cpu, ret);
584 		goto free_mask;
585 	}
586 
587 	list_add(&cpu_data->node, &cpu_data_list);
588 
589 	return cpu_data;
590 
591 free_mask:
592 	free_cpumask_var(cpu_data->shared_cpu_map);
593 free_cpu:
594 	kfree(cpu_data);
595 out:
596 	return NULL;
597 }
598 
599 static void cppc_cpufreq_put_cpu_data(struct cpufreq_policy *policy)
600 {
601 	struct cppc_cpudata *cpu_data = policy->driver_data;
602 
603 	list_del(&cpu_data->node);
604 	free_cpumask_var(cpu_data->shared_cpu_map);
605 	kfree(cpu_data);
606 	policy->driver_data = NULL;
607 }
608 
609 static int cppc_cpufreq_cpu_init(struct cpufreq_policy *policy)
610 {
611 	unsigned int cpu = policy->cpu;
612 	struct cppc_cpudata *cpu_data;
613 	struct cppc_perf_caps *caps;
614 	int ret;
615 
616 	cpu_data = cppc_cpufreq_get_cpu_data(cpu);
617 	if (!cpu_data) {
618 		pr_err("Error in acquiring _CPC/_PSD data for CPU%d.\n", cpu);
619 		return -ENODEV;
620 	}
621 	caps = &cpu_data->perf_caps;
622 	policy->driver_data = cpu_data;
623 
624 	/*
625 	 * Set min to lowest nonlinear perf to avoid any efficiency penalty (see
626 	 * Section 8.4.7.1.1.5 of ACPI 6.1 spec)
627 	 */
628 	policy->min = cppc_perf_to_khz(caps, caps->lowest_nonlinear_perf);
629 	policy->max = cppc_perf_to_khz(caps, caps->nominal_perf);
630 
631 	/*
632 	 * Set cpuinfo.min_freq to Lowest to make the full range of performance
633 	 * available if userspace wants to use any perf between lowest & lowest
634 	 * nonlinear perf
635 	 */
636 	policy->cpuinfo.min_freq = cppc_perf_to_khz(caps, caps->lowest_perf);
637 	policy->cpuinfo.max_freq = cppc_perf_to_khz(caps, caps->nominal_perf);
638 
639 	policy->transition_delay_us = cppc_cpufreq_get_transition_delay_us(cpu);
640 	policy->shared_type = cpu_data->shared_type;
641 
642 	switch (policy->shared_type) {
643 	case CPUFREQ_SHARED_TYPE_HW:
644 	case CPUFREQ_SHARED_TYPE_NONE:
645 		/* Nothing to be done - we'll have a policy for each CPU */
646 		break;
647 	case CPUFREQ_SHARED_TYPE_ANY:
648 		/*
649 		 * All CPUs in the domain will share a policy and all cpufreq
650 		 * operations will use a single cppc_cpudata structure stored
651 		 * in policy->driver_data.
652 		 */
653 		cpumask_copy(policy->cpus, cpu_data->shared_cpu_map);
654 		break;
655 	default:
656 		pr_debug("Unsupported CPU co-ord type: %d\n",
657 			 policy->shared_type);
658 		ret = -EFAULT;
659 		goto out;
660 	}
661 
662 	policy->fast_switch_possible = cppc_allow_fast_switch();
663 	policy->dvfs_possible_from_any_cpu = true;
664 
665 	/*
666 	 * If 'highest_perf' is greater than 'nominal_perf', we assume CPU Boost
667 	 * is supported.
668 	 */
669 	if (caps->highest_perf > caps->nominal_perf)
670 		boost_supported = true;
671 
672 	/* Set policy->cur to max now. The governors will adjust later. */
673 	policy->cur = cppc_perf_to_khz(caps, caps->highest_perf);
674 	cpu_data->perf_ctrls.desired_perf =  caps->highest_perf;
675 
676 	ret = cppc_set_perf(cpu, &cpu_data->perf_ctrls);
677 	if (ret) {
678 		pr_debug("Err setting perf value:%d on CPU:%d. ret:%d\n",
679 			 caps->highest_perf, cpu, ret);
680 		goto out;
681 	}
682 
683 	cppc_cpufreq_cpu_fie_init(policy);
684 	return 0;
685 
686 out:
687 	cppc_cpufreq_put_cpu_data(policy);
688 	return ret;
689 }
690 
691 static int cppc_cpufreq_cpu_exit(struct cpufreq_policy *policy)
692 {
693 	struct cppc_cpudata *cpu_data = policy->driver_data;
694 	struct cppc_perf_caps *caps = &cpu_data->perf_caps;
695 	unsigned int cpu = policy->cpu;
696 	int ret;
697 
698 	cppc_cpufreq_cpu_fie_exit(policy);
699 
700 	cpu_data->perf_ctrls.desired_perf = caps->lowest_perf;
701 
702 	ret = cppc_set_perf(cpu, &cpu_data->perf_ctrls);
703 	if (ret)
704 		pr_debug("Err setting perf value:%d on CPU:%d. ret:%d\n",
705 			 caps->lowest_perf, cpu, ret);
706 
707 	cppc_cpufreq_put_cpu_data(policy);
708 	return 0;
709 }
710 
711 static inline u64 get_delta(u64 t1, u64 t0)
712 {
713 	if (t1 > t0 || t0 > ~(u32)0)
714 		return t1 - t0;
715 
716 	return (u32)t1 - (u32)t0;
717 }
718 
719 static int cppc_perf_from_fbctrs(struct cppc_cpudata *cpu_data,
720 				 struct cppc_perf_fb_ctrs *fb_ctrs_t0,
721 				 struct cppc_perf_fb_ctrs *fb_ctrs_t1)
722 {
723 	u64 delta_reference, delta_delivered;
724 	u64 reference_perf;
725 
726 	reference_perf = fb_ctrs_t0->reference_perf;
727 
728 	delta_reference = get_delta(fb_ctrs_t1->reference,
729 				    fb_ctrs_t0->reference);
730 	delta_delivered = get_delta(fb_ctrs_t1->delivered,
731 				    fb_ctrs_t0->delivered);
732 
733 	/* Check to avoid divide-by zero and invalid delivered_perf */
734 	if (!delta_reference || !delta_delivered)
735 		return cpu_data->perf_ctrls.desired_perf;
736 
737 	return (reference_perf * delta_delivered) / delta_reference;
738 }
739 
740 static unsigned int cppc_cpufreq_get_rate(unsigned int cpu)
741 {
742 	struct cppc_perf_fb_ctrs fb_ctrs_t0 = {0}, fb_ctrs_t1 = {0};
743 	struct cpufreq_policy *policy = cpufreq_cpu_get(cpu);
744 	struct cppc_cpudata *cpu_data;
745 	u64 delivered_perf;
746 	int ret;
747 
748 	if (!policy)
749 		return -ENODEV;
750 
751 	cpu_data = policy->driver_data;
752 
753 	cpufreq_cpu_put(policy);
754 
755 	ret = cppc_get_perf_ctrs(cpu, &fb_ctrs_t0);
756 	if (ret)
757 		return 0;
758 
759 	udelay(2); /* 2usec delay between sampling */
760 
761 	ret = cppc_get_perf_ctrs(cpu, &fb_ctrs_t1);
762 	if (ret)
763 		return 0;
764 
765 	delivered_perf = cppc_perf_from_fbctrs(cpu_data, &fb_ctrs_t0,
766 					       &fb_ctrs_t1);
767 
768 	return cppc_perf_to_khz(&cpu_data->perf_caps, delivered_perf);
769 }
770 
771 static int cppc_cpufreq_set_boost(struct cpufreq_policy *policy, int state)
772 {
773 	struct cppc_cpudata *cpu_data = policy->driver_data;
774 	struct cppc_perf_caps *caps = &cpu_data->perf_caps;
775 	int ret;
776 
777 	if (!boost_supported) {
778 		pr_err("BOOST not supported by CPU or firmware\n");
779 		return -EINVAL;
780 	}
781 
782 	if (state)
783 		policy->max = cppc_perf_to_khz(caps, caps->highest_perf);
784 	else
785 		policy->max = cppc_perf_to_khz(caps, caps->nominal_perf);
786 	policy->cpuinfo.max_freq = policy->max;
787 
788 	ret = freq_qos_update_request(policy->max_freq_req, policy->max);
789 	if (ret < 0)
790 		return ret;
791 
792 	return 0;
793 }
794 
795 static ssize_t show_freqdomain_cpus(struct cpufreq_policy *policy, char *buf)
796 {
797 	struct cppc_cpudata *cpu_data = policy->driver_data;
798 
799 	return cpufreq_show_cpus(cpu_data->shared_cpu_map, buf);
800 }
801 cpufreq_freq_attr_ro(freqdomain_cpus);
802 
803 static struct freq_attr *cppc_cpufreq_attr[] = {
804 	&freqdomain_cpus,
805 	NULL,
806 };
807 
808 static struct cpufreq_driver cppc_cpufreq_driver = {
809 	.flags = CPUFREQ_CONST_LOOPS,
810 	.verify = cppc_verify_policy,
811 	.target = cppc_cpufreq_set_target,
812 	.get = cppc_cpufreq_get_rate,
813 	.fast_switch = cppc_cpufreq_fast_switch,
814 	.init = cppc_cpufreq_cpu_init,
815 	.exit = cppc_cpufreq_cpu_exit,
816 	.set_boost = cppc_cpufreq_set_boost,
817 	.attr = cppc_cpufreq_attr,
818 	.name = "cppc_cpufreq",
819 };
820 
821 /*
822  * HISI platform does not support delivered performance counter and
823  * reference performance counter. It can calculate the performance using the
824  * platform specific mechanism. We reuse the desired performance register to
825  * store the real performance calculated by the platform.
826  */
827 static unsigned int hisi_cppc_cpufreq_get_rate(unsigned int cpu)
828 {
829 	struct cpufreq_policy *policy = cpufreq_cpu_get(cpu);
830 	struct cppc_cpudata *cpu_data;
831 	u64 desired_perf;
832 	int ret;
833 
834 	if (!policy)
835 		return -ENODEV;
836 
837 	cpu_data = policy->driver_data;
838 
839 	cpufreq_cpu_put(policy);
840 
841 	ret = cppc_get_desired_perf(cpu, &desired_perf);
842 	if (ret < 0)
843 		return -EIO;
844 
845 	return cppc_perf_to_khz(&cpu_data->perf_caps, desired_perf);
846 }
847 
848 static void cppc_check_hisi_workaround(void)
849 {
850 	struct acpi_table_header *tbl;
851 	acpi_status status = AE_OK;
852 	int i;
853 
854 	status = acpi_get_table(ACPI_SIG_PCCT, 0, &tbl);
855 	if (ACPI_FAILURE(status) || !tbl)
856 		return;
857 
858 	for (i = 0; i < ARRAY_SIZE(wa_info); i++) {
859 		if (!memcmp(wa_info[i].oem_id, tbl->oem_id, ACPI_OEM_ID_SIZE) &&
860 		    !memcmp(wa_info[i].oem_table_id, tbl->oem_table_id, ACPI_OEM_TABLE_ID_SIZE) &&
861 		    wa_info[i].oem_revision == tbl->oem_revision) {
862 			/* Overwrite the get() callback */
863 			cppc_cpufreq_driver.get = hisi_cppc_cpufreq_get_rate;
864 			fie_disabled = FIE_DISABLED;
865 			break;
866 		}
867 	}
868 
869 	acpi_put_table(tbl);
870 }
871 
872 static int __init cppc_cpufreq_init(void)
873 {
874 	int ret;
875 
876 	if (!acpi_cpc_valid())
877 		return -ENODEV;
878 
879 	cppc_check_hisi_workaround();
880 	cppc_freq_invariance_init();
881 	populate_efficiency_class();
882 
883 	ret = cpufreq_register_driver(&cppc_cpufreq_driver);
884 	if (ret)
885 		cppc_freq_invariance_exit();
886 
887 	return ret;
888 }
889 
890 static inline void free_cpu_data(void)
891 {
892 	struct cppc_cpudata *iter, *tmp;
893 
894 	list_for_each_entry_safe(iter, tmp, &cpu_data_list, node) {
895 		free_cpumask_var(iter->shared_cpu_map);
896 		list_del(&iter->node);
897 		kfree(iter);
898 	}
899 
900 }
901 
902 static void __exit cppc_cpufreq_exit(void)
903 {
904 	cpufreq_unregister_driver(&cppc_cpufreq_driver);
905 	cppc_freq_invariance_exit();
906 
907 	free_cpu_data();
908 }
909 
910 module_exit(cppc_cpufreq_exit);
911 MODULE_AUTHOR("Ashwin Chaugule");
912 MODULE_DESCRIPTION("CPUFreq driver based on the ACPI CPPC v5.0+ spec");
913 MODULE_LICENSE("GPL");
914 
915 late_initcall(cppc_cpufreq_init);
916 
917 static const struct acpi_device_id cppc_acpi_ids[] __used = {
918 	{ACPI_PROCESSOR_DEVICE_HID, },
919 	{}
920 };
921 
922 MODULE_DEVICE_TABLE(acpi, cppc_acpi_ids);
923