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  */
42 static LIST_HEAD(cpu_data_list);
43 
44 static bool boost_supported;
45 
46 struct 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 
52 static 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 
64 static struct cpufreq_driver cppc_cpufreq_driver;
65 
66 static enum {
67 	FIE_UNSET = -1,
68 	FIE_ENABLED,
69 	FIE_DISABLED
70 } fie_disabled = FIE_UNSET;
71 
72 #ifdef CONFIG_ACPI_CPPC_CPUFREQ_FIE
73 module_param(fie_disabled, int, 0444);
74 MODULE_PARM_DESC(fie_disabled, "Disable Frequency Invariance Engine (FIE)");
75 
76 /* Frequency invariance support */
77 struct 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 
85 static DEFINE_PER_CPU(struct cppc_freq_invariance, cppc_freq_inv);
86 static struct kthread_worker *kworker_fie;
87 
88 static unsigned int hisi_cppc_cpufreq_get_rate(unsigned int cpu);
89 static 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  */
110 static 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 
140 static 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 
148 static 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 
159 static struct scale_freq_data cppc_sftd = {
160 	.source = SCALE_FREQ_SOURCE_CPPC,
161 	.set_freq_scale = cppc_scale_freq_tick,
162 };
163 
164 static 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  */
205 static 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 
223 static 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 
268 static void cppc_freq_invariance_exit(void)
269 {
270 	if (fie_disabled)
271 		return;
272 
273 	kthread_destroy_worker(kworker_fie);
274 }
275 
276 #else
277 static inline void cppc_cpufreq_cpu_fie_init(struct cpufreq_policy *policy)
278 {
279 }
280 
281 static inline void cppc_cpufreq_cpu_fie_exit(struct cpufreq_policy *policy)
282 {
283 }
284 
285 static inline void cppc_freq_invariance_init(void)
286 {
287 }
288 
289 static 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 */
295 static 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 */
309 static 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  */
331 static 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 
356 static 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 
381 static 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 
412 static 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 
433 static 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 
448 static 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
464 static 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 
472 static DEFINE_PER_CPU(unsigned int, efficiency_class);
473 static 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 
483 static 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  */
504 static 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 
510 static 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 
579 static 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 
603 static 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 
639 static 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
652 static int populate_efficiency_class(void)
653 {
654 	return 0;
655 }
656 #endif
657 
658 static 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 
690 free_mask:
691 	free_cpumask_var(cpu_data->shared_cpu_map);
692 free_cpu:
693 	kfree(cpu_data);
694 out:
695 	return NULL;
696 }
697 
698 static 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 
708 static 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 
789 out:
790 	cppc_cpufreq_put_cpu_data(policy);
791 	return ret;
792 }
793 
794 static 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 
814 static 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 
822 static 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 
843 static 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;
848 	u64 delivered_perf;
849 	int ret;
850 
851 	if (!policy)
852 		return -ENODEV;
853 
854 	cpu_data = policy->driver_data;
855 
856 	cpufreq_cpu_put(policy);
857 
858 	ret = cppc_get_perf_ctrs(cpu, &fb_ctrs_t0);
859 	if (ret)
860 		return 0;
861 
862 	udelay(2); /* 2usec delay between sampling */
863 
864 	ret = cppc_get_perf_ctrs(cpu, &fb_ctrs_t1);
865 	if (ret)
866 		return 0;
867 
868 	delivered_perf = cppc_perf_from_fbctrs(cpu_data, &fb_ctrs_t0,
869 					       &fb_ctrs_t1);
870 
871 	return cppc_cpufreq_perf_to_khz(cpu_data, delivered_perf);
872 }
873 
874 static int cppc_cpufreq_set_boost(struct cpufreq_policy *policy, int state)
875 {
876 	struct cppc_cpudata *cpu_data = policy->driver_data;
877 	struct cppc_perf_caps *caps = &cpu_data->perf_caps;
878 	int ret;
879 
880 	if (!boost_supported) {
881 		pr_err("BOOST not supported by CPU or firmware\n");
882 		return -EINVAL;
883 	}
884 
885 	if (state)
886 		policy->max = cppc_cpufreq_perf_to_khz(cpu_data,
887 						       caps->highest_perf);
888 	else
889 		policy->max = cppc_cpufreq_perf_to_khz(cpu_data,
890 						       caps->nominal_perf);
891 	policy->cpuinfo.max_freq = policy->max;
892 
893 	ret = freq_qos_update_request(policy->max_freq_req, policy->max);
894 	if (ret < 0)
895 		return ret;
896 
897 	return 0;
898 }
899 
900 static ssize_t show_freqdomain_cpus(struct cpufreq_policy *policy, char *buf)
901 {
902 	struct cppc_cpudata *cpu_data = policy->driver_data;
903 
904 	return cpufreq_show_cpus(cpu_data->shared_cpu_map, buf);
905 }
906 cpufreq_freq_attr_ro(freqdomain_cpus);
907 
908 static struct freq_attr *cppc_cpufreq_attr[] = {
909 	&freqdomain_cpus,
910 	NULL,
911 };
912 
913 static struct cpufreq_driver cppc_cpufreq_driver = {
914 	.flags = CPUFREQ_CONST_LOOPS,
915 	.verify = cppc_verify_policy,
916 	.target = cppc_cpufreq_set_target,
917 	.get = cppc_cpufreq_get_rate,
918 	.fast_switch = cppc_cpufreq_fast_switch,
919 	.init = cppc_cpufreq_cpu_init,
920 	.exit = cppc_cpufreq_cpu_exit,
921 	.set_boost = cppc_cpufreq_set_boost,
922 	.attr = cppc_cpufreq_attr,
923 	.name = "cppc_cpufreq",
924 };
925 
926 /*
927  * HISI platform does not support delivered performance counter and
928  * reference performance counter. It can calculate the performance using the
929  * platform specific mechanism. We reuse the desired performance register to
930  * store the real performance calculated by the platform.
931  */
932 static unsigned int hisi_cppc_cpufreq_get_rate(unsigned int cpu)
933 {
934 	struct cpufreq_policy *policy = cpufreq_cpu_get(cpu);
935 	struct cppc_cpudata *cpu_data;
936 	u64 desired_perf;
937 	int ret;
938 
939 	if (!policy)
940 		return -ENODEV;
941 
942 	cpu_data = policy->driver_data;
943 
944 	cpufreq_cpu_put(policy);
945 
946 	ret = cppc_get_desired_perf(cpu, &desired_perf);
947 	if (ret < 0)
948 		return -EIO;
949 
950 	return cppc_cpufreq_perf_to_khz(cpu_data, desired_perf);
951 }
952 
953 static void cppc_check_hisi_workaround(void)
954 {
955 	struct acpi_table_header *tbl;
956 	acpi_status status = AE_OK;
957 	int i;
958 
959 	status = acpi_get_table(ACPI_SIG_PCCT, 0, &tbl);
960 	if (ACPI_FAILURE(status) || !tbl)
961 		return;
962 
963 	for (i = 0; i < ARRAY_SIZE(wa_info); i++) {
964 		if (!memcmp(wa_info[i].oem_id, tbl->oem_id, ACPI_OEM_ID_SIZE) &&
965 		    !memcmp(wa_info[i].oem_table_id, tbl->oem_table_id, ACPI_OEM_TABLE_ID_SIZE) &&
966 		    wa_info[i].oem_revision == tbl->oem_revision) {
967 			/* Overwrite the get() callback */
968 			cppc_cpufreq_driver.get = hisi_cppc_cpufreq_get_rate;
969 			fie_disabled = FIE_DISABLED;
970 			break;
971 		}
972 	}
973 
974 	acpi_put_table(tbl);
975 }
976 
977 static int __init cppc_cpufreq_init(void)
978 {
979 	int ret;
980 
981 	if (!acpi_cpc_valid())
982 		return -ENODEV;
983 
984 	cppc_check_hisi_workaround();
985 	cppc_freq_invariance_init();
986 	populate_efficiency_class();
987 
988 	ret = cpufreq_register_driver(&cppc_cpufreq_driver);
989 	if (ret)
990 		cppc_freq_invariance_exit();
991 
992 	return ret;
993 }
994 
995 static inline void free_cpu_data(void)
996 {
997 	struct cppc_cpudata *iter, *tmp;
998 
999 	list_for_each_entry_safe(iter, tmp, &cpu_data_list, node) {
1000 		free_cpumask_var(iter->shared_cpu_map);
1001 		list_del(&iter->node);
1002 		kfree(iter);
1003 	}
1004 
1005 }
1006 
1007 static void __exit cppc_cpufreq_exit(void)
1008 {
1009 	cpufreq_unregister_driver(&cppc_cpufreq_driver);
1010 	cppc_freq_invariance_exit();
1011 
1012 	free_cpu_data();
1013 }
1014 
1015 module_exit(cppc_cpufreq_exit);
1016 MODULE_AUTHOR("Ashwin Chaugule");
1017 MODULE_DESCRIPTION("CPUFreq driver based on the ACPI CPPC v5.0+ spec");
1018 MODULE_LICENSE("GPL");
1019 
1020 late_initcall(cppc_cpufreq_init);
1021 
1022 static const struct acpi_device_id cppc_acpi_ids[] __used = {
1023 	{ACPI_PROCESSOR_DEVICE_HID, },
1024 	{}
1025 };
1026 
1027 MODULE_DEVICE_TABLE(acpi, cppc_acpi_ids);
1028