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/dmi.h>
20#include <linux/irq_work.h>
21#include <linux/kthread.h>
22#include <linux/time.h>
23#include <linux/vmalloc.h>
24#include <uapi/linux/sched/types.h>
25
26#include <asm/unaligned.h>
27
28#include <acpi/cppc_acpi.h>
29
30/* Minimum struct length needed for the DMI processor entry we want */
31#define DMI_ENTRY_PROCESSOR_MIN_LENGTH	48
32
33/* Offset in the DMI processor structure for the max frequency */
34#define DMI_PROCESSOR_MAX_SPEED		0x14
35
36/*
37 * This list contains information parsed from per CPU ACPI _CPC and _PSD
38 * structures: e.g. the highest and lowest supported performance, capabilities,
39 * desired performance, level requested etc. Depending on the share_type, not
40 * all CPUs will have an entry in the list.
41 */
42static LIST_HEAD(cpu_data_list);
43
44static bool boost_supported;
45
46struct cppc_workaround_oem_info {
47	char oem_id[ACPI_OEM_ID_SIZE + 1];
48	char oem_table_id[ACPI_OEM_TABLE_ID_SIZE + 1];
49	u32 oem_revision;
50};
51
52static struct cppc_workaround_oem_info wa_info[] = {
53	{
54		.oem_id		= "HISI  ",
55		.oem_table_id	= "HIP07   ",
56		.oem_revision	= 0,
57	}, {
58		.oem_id		= "HISI  ",
59		.oem_table_id	= "HIP08   ",
60		.oem_revision	= 0,
61	}
62};
63
64static struct cpufreq_driver cppc_cpufreq_driver;
65
66static enum {
67	FIE_UNSET = -1,
68	FIE_ENABLED,
69	FIE_DISABLED
70} fie_disabled = FIE_UNSET;
71
72#ifdef CONFIG_ACPI_CPPC_CPUFREQ_FIE
73module_param(fie_disabled, int, 0444);
74MODULE_PARM_DESC(fie_disabled, "Disable Frequency Invariance Engine (FIE)");
75
76/* Frequency invariance support */
77struct cppc_freq_invariance {
78	int cpu;
79	struct irq_work irq_work;
80	struct kthread_work work;
81	struct cppc_perf_fb_ctrs prev_perf_fb_ctrs;
82	struct cppc_cpudata *cpu_data;
83};
84
85static DEFINE_PER_CPU(struct cppc_freq_invariance, cppc_freq_inv);
86static struct kthread_worker *kworker_fie;
87
88static unsigned int hisi_cppc_cpufreq_get_rate(unsigned int cpu);
89static int cppc_perf_from_fbctrs(struct cppc_cpudata *cpu_data,
90				 struct cppc_perf_fb_ctrs *fb_ctrs_t0,
91				 struct cppc_perf_fb_ctrs *fb_ctrs_t1);
92
93/**
94 * cppc_scale_freq_workfn - CPPC arch_freq_scale updater for frequency invariance
95 * @work: The work item.
96 *
97 * The CPPC driver register itself with the topology core to provide its own
98 * implementation (cppc_scale_freq_tick()) of topology_scale_freq_tick() which
99 * gets called by the scheduler on every tick.
100 *
101 * Note that the arch specific counters have higher priority than CPPC counters,
102 * if available, though the CPPC driver doesn't need to have any special
103 * handling for that.
104 *
105 * On an invocation of cppc_scale_freq_tick(), we schedule an irq work (since we
106 * reach here from hard-irq context), which then schedules a normal work item
107 * and cppc_scale_freq_workfn() updates the per_cpu arch_freq_scale variable
108 * based on the counter updates since the last tick.
109 */
110static void cppc_scale_freq_workfn(struct kthread_work *work)
111{
112	struct cppc_freq_invariance *cppc_fi;
113	struct cppc_perf_fb_ctrs fb_ctrs = {0};
114	struct cppc_cpudata *cpu_data;
115	unsigned long local_freq_scale;
116	u64 perf;
117
118	cppc_fi = container_of(work, struct cppc_freq_invariance, work);
119	cpu_data = cppc_fi->cpu_data;
120
121	if (cppc_get_perf_ctrs(cppc_fi->cpu, &fb_ctrs)) {
122		pr_warn("%s: failed to read perf counters\n", __func__);
123		return;
124	}
125
126	perf = cppc_perf_from_fbctrs(cpu_data, &cppc_fi->prev_perf_fb_ctrs,
127				     &fb_ctrs);
128	cppc_fi->prev_perf_fb_ctrs = fb_ctrs;
129
130	perf <<= SCHED_CAPACITY_SHIFT;
131	local_freq_scale = div64_u64(perf, cpu_data->perf_caps.highest_perf);
132
133	/* This can happen due to counter's overflow */
134	if (unlikely(local_freq_scale > 1024))
135		local_freq_scale = 1024;
136
137	per_cpu(arch_freq_scale, cppc_fi->cpu) = local_freq_scale;
138}
139
140static void cppc_irq_work(struct irq_work *irq_work)
141{
142	struct cppc_freq_invariance *cppc_fi;
143
144	cppc_fi = container_of(irq_work, struct cppc_freq_invariance, irq_work);
145	kthread_queue_work(kworker_fie, &cppc_fi->work);
146}
147
148static void cppc_scale_freq_tick(void)
149{
150	struct cppc_freq_invariance *cppc_fi = &per_cpu(cppc_freq_inv, smp_processor_id());
151
152	/*
153	 * cppc_get_perf_ctrs() can potentially sleep, call that from the right
154	 * context.
155	 */
156	irq_work_queue(&cppc_fi->irq_work);
157}
158
159static struct scale_freq_data cppc_sftd = {
160	.source = SCALE_FREQ_SOURCE_CPPC,
161	.set_freq_scale = cppc_scale_freq_tick,
162};
163
164static void cppc_cpufreq_cpu_fie_init(struct cpufreq_policy *policy)
165{
166	struct cppc_freq_invariance *cppc_fi;
167	int cpu, ret;
168
169	if (fie_disabled)
170		return;
171
172	for_each_cpu(cpu, policy->cpus) {
173		cppc_fi = &per_cpu(cppc_freq_inv, cpu);
174		cppc_fi->cpu = cpu;
175		cppc_fi->cpu_data = policy->driver_data;
176		kthread_init_work(&cppc_fi->work, cppc_scale_freq_workfn);
177		init_irq_work(&cppc_fi->irq_work, cppc_irq_work);
178
179		ret = cppc_get_perf_ctrs(cpu, &cppc_fi->prev_perf_fb_ctrs);
180		if (ret) {
181			pr_warn("%s: failed to read perf counters for cpu:%d: %d\n",
182				__func__, cpu, ret);
183
184			/*
185			 * Don't abort if the CPU was offline while the driver
186			 * was getting registered.
187			 */
188			if (cpu_online(cpu))
189				return;
190		}
191	}
192
193	/* Register for freq-invariance */
194	topology_set_scale_freq_source(&cppc_sftd, policy->cpus);
195}
196
197/*
198 * We free all the resources on policy's removal and not on CPU removal as the
199 * irq-work are per-cpu and the hotplug core takes care of flushing the pending
200 * irq-works (hint: smpcfd_dying_cpu()) on CPU hotplug. Even if the kthread-work
201 * fires on another CPU after the concerned CPU is removed, it won't harm.
202 *
203 * We just need to make sure to remove them all on policy->exit().
204 */
205static void cppc_cpufreq_cpu_fie_exit(struct cpufreq_policy *policy)
206{
207	struct cppc_freq_invariance *cppc_fi;
208	int cpu;
209
210	if (fie_disabled)
211		return;
212
213	/* policy->cpus will be empty here, use related_cpus instead */
214	topology_clear_scale_freq_source(SCALE_FREQ_SOURCE_CPPC, policy->related_cpus);
215
216	for_each_cpu(cpu, policy->related_cpus) {
217		cppc_fi = &per_cpu(cppc_freq_inv, cpu);
218		irq_work_sync(&cppc_fi->irq_work);
219		kthread_cancel_work_sync(&cppc_fi->work);
220	}
221}
222
223static void __init cppc_freq_invariance_init(void)
224{
225	struct sched_attr attr = {
226		.size		= sizeof(struct sched_attr),
227		.sched_policy	= SCHED_DEADLINE,
228		.sched_nice	= 0,
229		.sched_priority	= 0,
230		/*
231		 * Fake (unused) bandwidth; workaround to "fix"
232		 * priority inheritance.
233		 */
234		.sched_runtime	= 1000000,
235		.sched_deadline = 10000000,
236		.sched_period	= 10000000,
237	};
238	int ret;
239
240	if (fie_disabled != FIE_ENABLED && fie_disabled != FIE_DISABLED) {
241		fie_disabled = FIE_ENABLED;
242		if (cppc_perf_ctrs_in_pcc()) {
243			pr_info("FIE not enabled on systems with registers in PCC\n");
244			fie_disabled = FIE_DISABLED;
245		}
246	}
247
248	if (fie_disabled)
249		return;
250
251	kworker_fie = kthread_create_worker(0, "cppc_fie");
252	if (IS_ERR(kworker_fie)) {
253		pr_warn("%s: failed to create kworker_fie: %ld\n", __func__,
254			PTR_ERR(kworker_fie));
255		fie_disabled = FIE_DISABLED;
256		return;
257	}
258
259	ret = sched_setattr_nocheck(kworker_fie->task, &attr);
260	if (ret) {
261		pr_warn("%s: failed to set SCHED_DEADLINE: %d\n", __func__,
262			ret);
263		kthread_destroy_worker(kworker_fie);
264		fie_disabled = FIE_DISABLED;
265	}
266}
267
268static void cppc_freq_invariance_exit(void)
269{
270	if (fie_disabled)
271		return;
272
273	kthread_destroy_worker(kworker_fie);
274}
275
276#else
277static inline void cppc_cpufreq_cpu_fie_init(struct cpufreq_policy *policy)
278{
279}
280
281static inline void cppc_cpufreq_cpu_fie_exit(struct cpufreq_policy *policy)
282{
283}
284
285static inline void cppc_freq_invariance_init(void)
286{
287}
288
289static inline void cppc_freq_invariance_exit(void)
290{
291}
292#endif /* CONFIG_ACPI_CPPC_CPUFREQ_FIE */
293
294/* Callback function used to retrieve the max frequency from DMI */
295static void cppc_find_dmi_mhz(const struct dmi_header *dm, void *private)
296{
297	const u8 *dmi_data = (const u8 *)dm;
298	u16 *mhz = (u16 *)private;
299
300	if (dm->type == DMI_ENTRY_PROCESSOR &&
301	    dm->length >= DMI_ENTRY_PROCESSOR_MIN_LENGTH) {
302		u16 val = (u16)get_unaligned((const u16 *)
303				(dmi_data + DMI_PROCESSOR_MAX_SPEED));
304		*mhz = val > *mhz ? val : *mhz;
305	}
306}
307
308/* Look up the max frequency in DMI */
309static u64 cppc_get_dmi_max_khz(void)
310{
311	u16 mhz = 0;
312
313	dmi_walk(cppc_find_dmi_mhz, &mhz);
314
315	/*
316	 * Real stupid fallback value, just in case there is no
317	 * actual value set.
318	 */
319	mhz = mhz ? mhz : 1;
320
321	return (1000 * mhz);
322}
323
324/*
325 * If CPPC lowest_freq and nominal_freq registers are exposed then we can
326 * use them to convert perf to freq and vice versa. The conversion is
327 * extrapolated as an affine function passing by the 2 points:
328 *  - (Low perf, Low freq)
329 *  - (Nominal perf, Nominal perf)
330 */
331static unsigned int cppc_cpufreq_perf_to_khz(struct cppc_cpudata *cpu_data,
332					     unsigned int perf)
333{
334	struct cppc_perf_caps *caps = &cpu_data->perf_caps;
335	s64 retval, offset = 0;
336	static u64 max_khz;
337	u64 mul, div;
338
339	if (caps->lowest_freq && caps->nominal_freq) {
340		mul = caps->nominal_freq - caps->lowest_freq;
341		div = caps->nominal_perf - caps->lowest_perf;
342		offset = caps->nominal_freq - div64_u64(caps->nominal_perf * mul, div);
343	} else {
344		if (!max_khz)
345			max_khz = cppc_get_dmi_max_khz();
346		mul = max_khz;
347		div = caps->highest_perf;
348	}
349
350	retval = offset + div64_u64(perf * mul, div);
351	if (retval >= 0)
352		return retval;
353	return 0;
354}
355
356static unsigned int cppc_cpufreq_khz_to_perf(struct cppc_cpudata *cpu_data,
357					     unsigned int freq)
358{
359	struct cppc_perf_caps *caps = &cpu_data->perf_caps;
360	s64 retval, offset = 0;
361	static u64 max_khz;
362	u64  mul, div;
363
364	if (caps->lowest_freq && caps->nominal_freq) {
365		mul = caps->nominal_perf - caps->lowest_perf;
366		div = caps->nominal_freq - caps->lowest_freq;
367		offset = caps->nominal_perf - div64_u64(caps->nominal_freq * mul, div);
368	} else {
369		if (!max_khz)
370			max_khz = cppc_get_dmi_max_khz();
371		mul = caps->highest_perf;
372		div = max_khz;
373	}
374
375	retval = offset + div64_u64(freq * mul, div);
376	if (retval >= 0)
377		return retval;
378	return 0;
379}
380
381static int cppc_cpufreq_set_target(struct cpufreq_policy *policy,
382				   unsigned int target_freq,
383				   unsigned int relation)
384
385{
386	struct cppc_cpudata *cpu_data = policy->driver_data;
387	unsigned int cpu = policy->cpu;
388	struct cpufreq_freqs freqs;
389	u32 desired_perf;
390	int ret = 0;
391
392	desired_perf = cppc_cpufreq_khz_to_perf(cpu_data, target_freq);
393	/* Return if it is exactly the same perf */
394	if (desired_perf == cpu_data->perf_ctrls.desired_perf)
395		return ret;
396
397	cpu_data->perf_ctrls.desired_perf = desired_perf;
398	freqs.old = policy->cur;
399	freqs.new = target_freq;
400
401	cpufreq_freq_transition_begin(policy, &freqs);
402	ret = cppc_set_perf(cpu, &cpu_data->perf_ctrls);
403	cpufreq_freq_transition_end(policy, &freqs, ret != 0);
404
405	if (ret)
406		pr_debug("Failed to set target on CPU:%d. ret:%d\n",
407			 cpu, ret);
408
409	return ret;
410}
411
412static unsigned int cppc_cpufreq_fast_switch(struct cpufreq_policy *policy,
413					      unsigned int target_freq)
414{
415	struct cppc_cpudata *cpu_data = policy->driver_data;
416	unsigned int cpu = policy->cpu;
417	u32 desired_perf;
418	int ret;
419
420	desired_perf = cppc_cpufreq_khz_to_perf(cpu_data, target_freq);
421	cpu_data->perf_ctrls.desired_perf = desired_perf;
422	ret = cppc_set_perf(cpu, &cpu_data->perf_ctrls);
423
424	if (ret) {
425		pr_debug("Failed to set target on CPU:%d. ret:%d\n",
426			 cpu, ret);
427		return 0;
428	}
429
430	return target_freq;
431}
432
433static int cppc_verify_policy(struct cpufreq_policy_data *policy)
434{
435	cpufreq_verify_within_cpu_limits(policy);
436	return 0;
437}
438
439/*
440 * The PCC subspace describes the rate at which platform can accept commands
441 * on the shared PCC channel (including READs which do not count towards freq
442 * transition requests), so ideally we need to use the PCC values as a fallback
443 * if we don't have a platform specific transition_delay_us
444 */
445#ifdef CONFIG_ARM64
446#include <asm/cputype.h>
447
448static unsigned int cppc_cpufreq_get_transition_delay_us(unsigned int cpu)
449{
450	unsigned long implementor = read_cpuid_implementor();
451	unsigned long part_num = read_cpuid_part_number();
452
453	switch (implementor) {
454	case ARM_CPU_IMP_QCOM:
455		switch (part_num) {
456		case QCOM_CPU_PART_FALKOR_V1:
457		case QCOM_CPU_PART_FALKOR:
458			return 10000;
459		}
460	}
461	return cppc_get_transition_latency(cpu) / NSEC_PER_USEC;
462}
463#else
464static unsigned int cppc_cpufreq_get_transition_delay_us(unsigned int cpu)
465{
466	return cppc_get_transition_latency(cpu) / NSEC_PER_USEC;
467}
468#endif
469
470#if defined(CONFIG_ARM64) && defined(CONFIG_ENERGY_MODEL)
471
472static DEFINE_PER_CPU(unsigned int, efficiency_class);
473static void cppc_cpufreq_register_em(struct cpufreq_policy *policy);
474
475/* Create an artificial performance state every CPPC_EM_CAP_STEP capacity unit. */
476#define CPPC_EM_CAP_STEP	(20)
477/* Increase the cost value by CPPC_EM_COST_STEP every performance state. */
478#define CPPC_EM_COST_STEP	(1)
479/* Add a cost gap correspnding to the energy of 4 CPUs. */
480#define CPPC_EM_COST_GAP	(4 * SCHED_CAPACITY_SCALE * CPPC_EM_COST_STEP \
481				/ CPPC_EM_CAP_STEP)
482
483static unsigned int get_perf_level_count(struct cpufreq_policy *policy)
484{
485	struct cppc_perf_caps *perf_caps;
486	unsigned int min_cap, max_cap;
487	struct cppc_cpudata *cpu_data;
488	int cpu = policy->cpu;
489
490	cpu_data = policy->driver_data;
491	perf_caps = &cpu_data->perf_caps;
492	max_cap = arch_scale_cpu_capacity(cpu);
493	min_cap = div_u64((u64)max_cap * perf_caps->lowest_perf,
494			  perf_caps->highest_perf);
495	if ((min_cap == 0) || (max_cap < min_cap))
496		return 0;
497	return 1 + max_cap / CPPC_EM_CAP_STEP - min_cap / CPPC_EM_CAP_STEP;
498}
499
500/*
501 * The cost is defined as:
502 *   cost = power * max_frequency / frequency
503 */
504static inline unsigned long compute_cost(int cpu, int step)
505{
506	return CPPC_EM_COST_GAP * per_cpu(efficiency_class, cpu) +
507			step * CPPC_EM_COST_STEP;
508}
509
510static int cppc_get_cpu_power(struct device *cpu_dev,
511		unsigned long *power, unsigned long *KHz)
512{
513	unsigned long perf_step, perf_prev, perf, perf_check;
514	unsigned int min_step, max_step, step, step_check;
515	unsigned long prev_freq = *KHz;
516	unsigned int min_cap, max_cap;
517	struct cpufreq_policy *policy;
518
519	struct cppc_perf_caps *perf_caps;
520	struct cppc_cpudata *cpu_data;
521
522	policy = cpufreq_cpu_get_raw(cpu_dev->id);
523	cpu_data = policy->driver_data;
524	perf_caps = &cpu_data->perf_caps;
525	max_cap = arch_scale_cpu_capacity(cpu_dev->id);
526	min_cap = div_u64((u64)max_cap * perf_caps->lowest_perf,
527			  perf_caps->highest_perf);
528	perf_step = div_u64((u64)CPPC_EM_CAP_STEP * perf_caps->highest_perf,
529			    max_cap);
530	min_step = min_cap / CPPC_EM_CAP_STEP;
531	max_step = max_cap / CPPC_EM_CAP_STEP;
532
533	perf_prev = cppc_cpufreq_khz_to_perf(cpu_data, *KHz);
534	step = perf_prev / perf_step;
535
536	if (step > max_step)
537		return -EINVAL;
538
539	if (min_step == max_step) {
540		step = max_step;
541		perf = perf_caps->highest_perf;
542	} else if (step < min_step) {
543		step = min_step;
544		perf = perf_caps->lowest_perf;
545	} else {
546		step++;
547		if (step == max_step)
548			perf = perf_caps->highest_perf;
549		else
550			perf = step * perf_step;
551	}
552
553	*KHz = cppc_cpufreq_perf_to_khz(cpu_data, perf);
554	perf_check = cppc_cpufreq_khz_to_perf(cpu_data, *KHz);
555	step_check = perf_check / perf_step;
556
557	/*
558	 * To avoid bad integer approximation, check that new frequency value
559	 * increased and that the new frequency will be converted to the
560	 * desired step value.
561	 */
562	while ((*KHz == prev_freq) || (step_check != step)) {
563		perf++;
564		*KHz = cppc_cpufreq_perf_to_khz(cpu_data, perf);
565		perf_check = cppc_cpufreq_khz_to_perf(cpu_data, *KHz);
566		step_check = perf_check / perf_step;
567	}
568
569	/*
570	 * With an artificial EM, only the cost value is used. Still the power
571	 * is populated such as 0 < power < EM_MAX_POWER. This allows to add
572	 * more sense to the artificial performance states.
573	 */
574	*power = compute_cost(cpu_dev->id, step);
575
576	return 0;
577}
578
579static int cppc_get_cpu_cost(struct device *cpu_dev, unsigned long KHz,
580		unsigned long *cost)
581{
582	unsigned long perf_step, perf_prev;
583	struct cppc_perf_caps *perf_caps;
584	struct cpufreq_policy *policy;
585	struct cppc_cpudata *cpu_data;
586	unsigned int max_cap;
587	int step;
588
589	policy = cpufreq_cpu_get_raw(cpu_dev->id);
590	cpu_data = policy->driver_data;
591	perf_caps = &cpu_data->perf_caps;
592	max_cap = arch_scale_cpu_capacity(cpu_dev->id);
593
594	perf_prev = cppc_cpufreq_khz_to_perf(cpu_data, KHz);
595	perf_step = CPPC_EM_CAP_STEP * perf_caps->highest_perf / max_cap;
596	step = perf_prev / perf_step;
597
598	*cost = compute_cost(cpu_dev->id, step);
599
600	return 0;
601}
602
603static int populate_efficiency_class(void)
604{
605	struct acpi_madt_generic_interrupt *gicc;
606	DECLARE_BITMAP(used_classes, 256) = {};
607	int class, cpu, index;
608
609	for_each_possible_cpu(cpu) {
610		gicc = acpi_cpu_get_madt_gicc(cpu);
611		class = gicc->efficiency_class;
612		bitmap_set(used_classes, class, 1);
613	}
614
615	if (bitmap_weight(used_classes, 256) <= 1) {
616		pr_debug("Efficiency classes are all equal (=%d). "
617			"No EM registered", class);
618		return -EINVAL;
619	}
620
621	/*
622	 * Squeeze efficiency class values on [0:#efficiency_class-1].
623	 * Values are per spec in [0:255].
624	 */
625	index = 0;
626	for_each_set_bit(class, used_classes, 256) {
627		for_each_possible_cpu(cpu) {
628			gicc = acpi_cpu_get_madt_gicc(cpu);
629			if (gicc->efficiency_class == class)
630				per_cpu(efficiency_class, cpu) = index;
631		}
632		index++;
633	}
634	cppc_cpufreq_driver.register_em = cppc_cpufreq_register_em;
635
636	return 0;
637}
638
639static void cppc_cpufreq_register_em(struct cpufreq_policy *policy)
640{
641	struct cppc_cpudata *cpu_data;
642	struct em_data_callback em_cb =
643		EM_ADV_DATA_CB(cppc_get_cpu_power, cppc_get_cpu_cost);
644
645	cpu_data = policy->driver_data;
646	em_dev_register_perf_domain(get_cpu_device(policy->cpu),
647			get_perf_level_count(policy), &em_cb,
648			cpu_data->shared_cpu_map, 0);
649}
650
651#else
652static int populate_efficiency_class(void)
653{
654	return 0;
655}
656#endif
657
658static struct cppc_cpudata *cppc_cpufreq_get_cpu_data(unsigned int cpu)
659{
660	struct cppc_cpudata *cpu_data;
661	int ret;
662
663	cpu_data = kzalloc(sizeof(struct cppc_cpudata), GFP_KERNEL);
664	if (!cpu_data)
665		goto out;
666
667	if (!zalloc_cpumask_var(&cpu_data->shared_cpu_map, GFP_KERNEL))
668		goto free_cpu;
669
670	ret = acpi_get_psd_map(cpu, cpu_data);
671	if (ret) {
672		pr_debug("Err parsing CPU%d PSD data: ret:%d\n", cpu, ret);
673		goto free_mask;
674	}
675
676	ret = cppc_get_perf_caps(cpu, &cpu_data->perf_caps);
677	if (ret) {
678		pr_debug("Err reading CPU%d perf caps: ret:%d\n", cpu, ret);
679		goto free_mask;
680	}
681
682	/* Convert the lowest and nominal freq from MHz to KHz */
683	cpu_data->perf_caps.lowest_freq *= 1000;
684	cpu_data->perf_caps.nominal_freq *= 1000;
685
686	list_add(&cpu_data->node, &cpu_data_list);
687
688	return cpu_data;
689
690free_mask:
691	free_cpumask_var(cpu_data->shared_cpu_map);
692free_cpu:
693	kfree(cpu_data);
694out:
695	return NULL;
696}
697
698static void cppc_cpufreq_put_cpu_data(struct cpufreq_policy *policy)
699{
700	struct cppc_cpudata *cpu_data = policy->driver_data;
701
702	list_del(&cpu_data->node);
703	free_cpumask_var(cpu_data->shared_cpu_map);
704	kfree(cpu_data);
705	policy->driver_data = NULL;
706}
707
708static int cppc_cpufreq_cpu_init(struct cpufreq_policy *policy)
709{
710	unsigned int cpu = policy->cpu;
711	struct cppc_cpudata *cpu_data;
712	struct cppc_perf_caps *caps;
713	int ret;
714
715	cpu_data = cppc_cpufreq_get_cpu_data(cpu);
716	if (!cpu_data) {
717		pr_err("Error in acquiring _CPC/_PSD data for CPU%d.\n", cpu);
718		return -ENODEV;
719	}
720	caps = &cpu_data->perf_caps;
721	policy->driver_data = cpu_data;
722
723	/*
724	 * Set min to lowest nonlinear perf to avoid any efficiency penalty (see
725	 * Section 8.4.7.1.1.5 of ACPI 6.1 spec)
726	 */
727	policy->min = cppc_cpufreq_perf_to_khz(cpu_data,
728					       caps->lowest_nonlinear_perf);
729	policy->max = cppc_cpufreq_perf_to_khz(cpu_data,
730					       caps->nominal_perf);
731
732	/*
733	 * Set cpuinfo.min_freq to Lowest to make the full range of performance
734	 * available if userspace wants to use any perf between lowest & lowest
735	 * nonlinear perf
736	 */
737	policy->cpuinfo.min_freq = cppc_cpufreq_perf_to_khz(cpu_data,
738							    caps->lowest_perf);
739	policy->cpuinfo.max_freq = cppc_cpufreq_perf_to_khz(cpu_data,
740							    caps->nominal_perf);
741
742	policy->transition_delay_us = cppc_cpufreq_get_transition_delay_us(cpu);
743	policy->shared_type = cpu_data->shared_type;
744
745	switch (policy->shared_type) {
746	case CPUFREQ_SHARED_TYPE_HW:
747	case CPUFREQ_SHARED_TYPE_NONE:
748		/* Nothing to be done - we'll have a policy for each CPU */
749		break;
750	case CPUFREQ_SHARED_TYPE_ANY:
751		/*
752		 * All CPUs in the domain will share a policy and all cpufreq
753		 * operations will use a single cppc_cpudata structure stored
754		 * in policy->driver_data.
755		 */
756		cpumask_copy(policy->cpus, cpu_data->shared_cpu_map);
757		break;
758	default:
759		pr_debug("Unsupported CPU co-ord type: %d\n",
760			 policy->shared_type);
761		ret = -EFAULT;
762		goto out;
763	}
764
765	policy->fast_switch_possible = cppc_allow_fast_switch();
766	policy->dvfs_possible_from_any_cpu = true;
767
768	/*
769	 * If 'highest_perf' is greater than 'nominal_perf', we assume CPU Boost
770	 * is supported.
771	 */
772	if (caps->highest_perf > caps->nominal_perf)
773		boost_supported = true;
774
775	/* Set policy->cur to max now. The governors will adjust later. */
776	policy->cur = cppc_cpufreq_perf_to_khz(cpu_data, caps->highest_perf);
777	cpu_data->perf_ctrls.desired_perf =  caps->highest_perf;
778
779	ret = cppc_set_perf(cpu, &cpu_data->perf_ctrls);
780	if (ret) {
781		pr_debug("Err setting perf value:%d on CPU:%d. ret:%d\n",
782			 caps->highest_perf, cpu, ret);
783		goto out;
784	}
785
786	cppc_cpufreq_cpu_fie_init(policy);
787	return 0;
788
789out:
790	cppc_cpufreq_put_cpu_data(policy);
791	return ret;
792}
793
794static int cppc_cpufreq_cpu_exit(struct cpufreq_policy *policy)
795{
796	struct cppc_cpudata *cpu_data = policy->driver_data;
797	struct cppc_perf_caps *caps = &cpu_data->perf_caps;
798	unsigned int cpu = policy->cpu;
799	int ret;
800
801	cppc_cpufreq_cpu_fie_exit(policy);
802
803	cpu_data->perf_ctrls.desired_perf = caps->lowest_perf;
804
805	ret = cppc_set_perf(cpu, &cpu_data->perf_ctrls);
806	if (ret)
807		pr_debug("Err setting perf value:%d on CPU:%d. ret:%d\n",
808			 caps->lowest_perf, cpu, ret);
809
810	cppc_cpufreq_put_cpu_data(policy);
811	return 0;
812}
813
814static inline u64 get_delta(u64 t1, u64 t0)
815{
816	if (t1 > t0 || t0 > ~(u32)0)
817		return t1 - t0;
818
819	return (u32)t1 - (u32)t0;
820}
821
822static int cppc_perf_from_fbctrs(struct cppc_cpudata *cpu_data,
823				 struct cppc_perf_fb_ctrs *fb_ctrs_t0,
824				 struct cppc_perf_fb_ctrs *fb_ctrs_t1)
825{
826	u64 delta_reference, delta_delivered;
827	u64 reference_perf;
828
829	reference_perf = fb_ctrs_t0->reference_perf;
830
831	delta_reference = get_delta(fb_ctrs_t1->reference,
832				    fb_ctrs_t0->reference);
833	delta_delivered = get_delta(fb_ctrs_t1->delivered,
834				    fb_ctrs_t0->delivered);
835
836	/* Check to avoid divide-by zero and invalid delivered_perf */
837	if (!delta_reference || !delta_delivered)
838		return cpu_data->perf_ctrls.desired_perf;
839
840	return (reference_perf * delta_delivered) / delta_reference;
841}
842
843static unsigned int cppc_cpufreq_get_rate(unsigned int cpu)
844{
845	struct cppc_perf_fb_ctrs fb_ctrs_t0 = {0}, fb_ctrs_t1 = {0};
846	struct cpufreq_policy *policy = cpufreq_cpu_get(cpu);
847	struct cppc_cpudata *cpu_data = policy->driver_data;
848	u64 delivered_perf;
849	int ret;
850
851	cpufreq_cpu_put(policy);
852
853	ret = cppc_get_perf_ctrs(cpu, &fb_ctrs_t0);
854	if (ret)
855		return 0;
856
857	udelay(2); /* 2usec delay between sampling */
858
859	ret = cppc_get_perf_ctrs(cpu, &fb_ctrs_t1);
860	if (ret)
861		return 0;
862
863	delivered_perf = cppc_perf_from_fbctrs(cpu_data, &fb_ctrs_t0,
864					       &fb_ctrs_t1);
865
866	return cppc_cpufreq_perf_to_khz(cpu_data, delivered_perf);
867}
868
869static int cppc_cpufreq_set_boost(struct cpufreq_policy *policy, int state)
870{
871	struct cppc_cpudata *cpu_data = policy->driver_data;
872	struct cppc_perf_caps *caps = &cpu_data->perf_caps;
873	int ret;
874
875	if (!boost_supported) {
876		pr_err("BOOST not supported by CPU or firmware\n");
877		return -EINVAL;
878	}
879
880	if (state)
881		policy->max = cppc_cpufreq_perf_to_khz(cpu_data,
882						       caps->highest_perf);
883	else
884		policy->max = cppc_cpufreq_perf_to_khz(cpu_data,
885						       caps->nominal_perf);
886	policy->cpuinfo.max_freq = policy->max;
887
888	ret = freq_qos_update_request(policy->max_freq_req, policy->max);
889	if (ret < 0)
890		return ret;
891
892	return 0;
893}
894
895static ssize_t show_freqdomain_cpus(struct cpufreq_policy *policy, char *buf)
896{
897	struct cppc_cpudata *cpu_data = policy->driver_data;
898
899	return cpufreq_show_cpus(cpu_data->shared_cpu_map, buf);
900}
901cpufreq_freq_attr_ro(freqdomain_cpus);
902
903static struct freq_attr *cppc_cpufreq_attr[] = {
904	&freqdomain_cpus,
905	NULL,
906};
907
908static struct cpufreq_driver cppc_cpufreq_driver = {
909	.flags = CPUFREQ_CONST_LOOPS,
910	.verify = cppc_verify_policy,
911	.target = cppc_cpufreq_set_target,
912	.get = cppc_cpufreq_get_rate,
913	.fast_switch = cppc_cpufreq_fast_switch,
914	.init = cppc_cpufreq_cpu_init,
915	.exit = cppc_cpufreq_cpu_exit,
916	.set_boost = cppc_cpufreq_set_boost,
917	.attr = cppc_cpufreq_attr,
918	.name = "cppc_cpufreq",
919};
920
921/*
922 * HISI platform does not support delivered performance counter and
923 * reference performance counter. It can calculate the performance using the
924 * platform specific mechanism. We reuse the desired performance register to
925 * store the real performance calculated by the platform.
926 */
927static unsigned int hisi_cppc_cpufreq_get_rate(unsigned int cpu)
928{
929	struct cpufreq_policy *policy = cpufreq_cpu_get(cpu);
930	struct cppc_cpudata *cpu_data = policy->driver_data;
931	u64 desired_perf;
932	int ret;
933
934	cpufreq_cpu_put(policy);
935
936	ret = cppc_get_desired_perf(cpu, &desired_perf);
937	if (ret < 0)
938		return -EIO;
939
940	return cppc_cpufreq_perf_to_khz(cpu_data, desired_perf);
941}
942
943static void cppc_check_hisi_workaround(void)
944{
945	struct acpi_table_header *tbl;
946	acpi_status status = AE_OK;
947	int i;
948
949	status = acpi_get_table(ACPI_SIG_PCCT, 0, &tbl);
950	if (ACPI_FAILURE(status) || !tbl)
951		return;
952
953	for (i = 0; i < ARRAY_SIZE(wa_info); i++) {
954		if (!memcmp(wa_info[i].oem_id, tbl->oem_id, ACPI_OEM_ID_SIZE) &&
955		    !memcmp(wa_info[i].oem_table_id, tbl->oem_table_id, ACPI_OEM_TABLE_ID_SIZE) &&
956		    wa_info[i].oem_revision == tbl->oem_revision) {
957			/* Overwrite the get() callback */
958			cppc_cpufreq_driver.get = hisi_cppc_cpufreq_get_rate;
959			fie_disabled = FIE_DISABLED;
960			break;
961		}
962	}
963
964	acpi_put_table(tbl);
965}
966
967static int __init cppc_cpufreq_init(void)
968{
969	int ret;
970
971	if (!acpi_cpc_valid())
972		return -ENODEV;
973
974	cppc_check_hisi_workaround();
975	cppc_freq_invariance_init();
976	populate_efficiency_class();
977
978	ret = cpufreq_register_driver(&cppc_cpufreq_driver);
979	if (ret)
980		cppc_freq_invariance_exit();
981
982	return ret;
983}
984
985static inline void free_cpu_data(void)
986{
987	struct cppc_cpudata *iter, *tmp;
988
989	list_for_each_entry_safe(iter, tmp, &cpu_data_list, node) {
990		free_cpumask_var(iter->shared_cpu_map);
991		list_del(&iter->node);
992		kfree(iter);
993	}
994
995}
996
997static void __exit cppc_cpufreq_exit(void)
998{
999	cpufreq_unregister_driver(&cppc_cpufreq_driver);
1000	cppc_freq_invariance_exit();
1001
1002	free_cpu_data();
1003}
1004
1005module_exit(cppc_cpufreq_exit);
1006MODULE_AUTHOR("Ashwin Chaugule");
1007MODULE_DESCRIPTION("CPUFreq driver based on the ACPI CPPC v5.0+ spec");
1008MODULE_LICENSE("GPL");
1009
1010late_initcall(cppc_cpufreq_init);
1011
1012static const struct acpi_device_id cppc_acpi_ids[] __used = {
1013	{ACPI_PROCESSOR_DEVICE_HID, },
1014	{}
1015};
1016
1017MODULE_DEVICE_TABLE(acpi, cppc_acpi_ids);
1018