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 		return;
254 
255 	ret = sched_setattr_nocheck(kworker_fie->task, &attr);
256 	if (ret) {
257 		pr_warn("%s: failed to set SCHED_DEADLINE: %d\n", __func__,
258 			ret);
259 		kthread_destroy_worker(kworker_fie);
260 		return;
261 	}
262 }
263 
264 static void cppc_freq_invariance_exit(void)
265 {
266 	if (fie_disabled)
267 		return;
268 
269 	kthread_destroy_worker(kworker_fie);
270 	kworker_fie = NULL;
271 }
272 
273 #else
274 static inline void cppc_cpufreq_cpu_fie_init(struct cpufreq_policy *policy)
275 {
276 }
277 
278 static inline void cppc_cpufreq_cpu_fie_exit(struct cpufreq_policy *policy)
279 {
280 }
281 
282 static inline void cppc_freq_invariance_init(void)
283 {
284 }
285 
286 static inline void cppc_freq_invariance_exit(void)
287 {
288 }
289 #endif /* CONFIG_ACPI_CPPC_CPUFREQ_FIE */
290 
291 /* Callback function used to retrieve the max frequency from DMI */
292 static void cppc_find_dmi_mhz(const struct dmi_header *dm, void *private)
293 {
294 	const u8 *dmi_data = (const u8 *)dm;
295 	u16 *mhz = (u16 *)private;
296 
297 	if (dm->type == DMI_ENTRY_PROCESSOR &&
298 	    dm->length >= DMI_ENTRY_PROCESSOR_MIN_LENGTH) {
299 		u16 val = (u16)get_unaligned((const u16 *)
300 				(dmi_data + DMI_PROCESSOR_MAX_SPEED));
301 		*mhz = val > *mhz ? val : *mhz;
302 	}
303 }
304 
305 /* Look up the max frequency in DMI */
306 static u64 cppc_get_dmi_max_khz(void)
307 {
308 	u16 mhz = 0;
309 
310 	dmi_walk(cppc_find_dmi_mhz, &mhz);
311 
312 	/*
313 	 * Real stupid fallback value, just in case there is no
314 	 * actual value set.
315 	 */
316 	mhz = mhz ? mhz : 1;
317 
318 	return (1000 * mhz);
319 }
320 
321 /*
322  * If CPPC lowest_freq and nominal_freq registers are exposed then we can
323  * use them to convert perf to freq and vice versa. The conversion is
324  * extrapolated as an affine function passing by the 2 points:
325  *  - (Low perf, Low freq)
326  *  - (Nominal perf, Nominal perf)
327  */
328 static unsigned int cppc_cpufreq_perf_to_khz(struct cppc_cpudata *cpu_data,
329 					     unsigned int perf)
330 {
331 	struct cppc_perf_caps *caps = &cpu_data->perf_caps;
332 	s64 retval, offset = 0;
333 	static u64 max_khz;
334 	u64 mul, div;
335 
336 	if (caps->lowest_freq && caps->nominal_freq) {
337 		mul = caps->nominal_freq - caps->lowest_freq;
338 		div = caps->nominal_perf - caps->lowest_perf;
339 		offset = caps->nominal_freq - div64_u64(caps->nominal_perf * mul, div);
340 	} else {
341 		if (!max_khz)
342 			max_khz = cppc_get_dmi_max_khz();
343 		mul = max_khz;
344 		div = caps->highest_perf;
345 	}
346 
347 	retval = offset + div64_u64(perf * mul, div);
348 	if (retval >= 0)
349 		return retval;
350 	return 0;
351 }
352 
353 static unsigned int cppc_cpufreq_khz_to_perf(struct cppc_cpudata *cpu_data,
354 					     unsigned int freq)
355 {
356 	struct cppc_perf_caps *caps = &cpu_data->perf_caps;
357 	s64 retval, offset = 0;
358 	static u64 max_khz;
359 	u64  mul, div;
360 
361 	if (caps->lowest_freq && caps->nominal_freq) {
362 		mul = caps->nominal_perf - caps->lowest_perf;
363 		div = caps->nominal_freq - caps->lowest_freq;
364 		offset = caps->nominal_perf - div64_u64(caps->nominal_freq * mul, div);
365 	} else {
366 		if (!max_khz)
367 			max_khz = cppc_get_dmi_max_khz();
368 		mul = caps->highest_perf;
369 		div = max_khz;
370 	}
371 
372 	retval = offset + div64_u64(freq * mul, div);
373 	if (retval >= 0)
374 		return retval;
375 	return 0;
376 }
377 
378 static int cppc_cpufreq_set_target(struct cpufreq_policy *policy,
379 				   unsigned int target_freq,
380 				   unsigned int relation)
381 
382 {
383 	struct cppc_cpudata *cpu_data = policy->driver_data;
384 	unsigned int cpu = policy->cpu;
385 	struct cpufreq_freqs freqs;
386 	u32 desired_perf;
387 	int ret = 0;
388 
389 	desired_perf = cppc_cpufreq_khz_to_perf(cpu_data, target_freq);
390 	/* Return if it is exactly the same perf */
391 	if (desired_perf == cpu_data->perf_ctrls.desired_perf)
392 		return ret;
393 
394 	cpu_data->perf_ctrls.desired_perf = desired_perf;
395 	freqs.old = policy->cur;
396 	freqs.new = target_freq;
397 
398 	cpufreq_freq_transition_begin(policy, &freqs);
399 	ret = cppc_set_perf(cpu, &cpu_data->perf_ctrls);
400 	cpufreq_freq_transition_end(policy, &freqs, ret != 0);
401 
402 	if (ret)
403 		pr_debug("Failed to set target on CPU:%d. ret:%d\n",
404 			 cpu, ret);
405 
406 	return ret;
407 }
408 
409 static unsigned int cppc_cpufreq_fast_switch(struct cpufreq_policy *policy,
410 					      unsigned int target_freq)
411 {
412 	struct cppc_cpudata *cpu_data = policy->driver_data;
413 	unsigned int cpu = policy->cpu;
414 	u32 desired_perf;
415 	int ret;
416 
417 	desired_perf = cppc_cpufreq_khz_to_perf(cpu_data, target_freq);
418 	cpu_data->perf_ctrls.desired_perf = desired_perf;
419 	ret = cppc_set_perf(cpu, &cpu_data->perf_ctrls);
420 
421 	if (ret) {
422 		pr_debug("Failed to set target on CPU:%d. ret:%d\n",
423 			 cpu, ret);
424 		return 0;
425 	}
426 
427 	return target_freq;
428 }
429 
430 static int cppc_verify_policy(struct cpufreq_policy_data *policy)
431 {
432 	cpufreq_verify_within_cpu_limits(policy);
433 	return 0;
434 }
435 
436 /*
437  * The PCC subspace describes the rate at which platform can accept commands
438  * on the shared PCC channel (including READs which do not count towards freq
439  * transition requests), so ideally we need to use the PCC values as a fallback
440  * if we don't have a platform specific transition_delay_us
441  */
442 #ifdef CONFIG_ARM64
443 #include <asm/cputype.h>
444 
445 static unsigned int cppc_cpufreq_get_transition_delay_us(unsigned int cpu)
446 {
447 	unsigned long implementor = read_cpuid_implementor();
448 	unsigned long part_num = read_cpuid_part_number();
449 
450 	switch (implementor) {
451 	case ARM_CPU_IMP_QCOM:
452 		switch (part_num) {
453 		case QCOM_CPU_PART_FALKOR_V1:
454 		case QCOM_CPU_PART_FALKOR:
455 			return 10000;
456 		}
457 	}
458 	return cppc_get_transition_latency(cpu) / NSEC_PER_USEC;
459 }
460 #else
461 static unsigned int cppc_cpufreq_get_transition_delay_us(unsigned int cpu)
462 {
463 	return cppc_get_transition_latency(cpu) / NSEC_PER_USEC;
464 }
465 #endif
466 
467 #if defined(CONFIG_ARM64) && defined(CONFIG_ENERGY_MODEL)
468 
469 static DEFINE_PER_CPU(unsigned int, efficiency_class);
470 static void cppc_cpufreq_register_em(struct cpufreq_policy *policy);
471 
472 /* Create an artificial performance state every CPPC_EM_CAP_STEP capacity unit. */
473 #define CPPC_EM_CAP_STEP	(20)
474 /* Increase the cost value by CPPC_EM_COST_STEP every performance state. */
475 #define CPPC_EM_COST_STEP	(1)
476 /* Add a cost gap correspnding to the energy of 4 CPUs. */
477 #define CPPC_EM_COST_GAP	(4 * SCHED_CAPACITY_SCALE * CPPC_EM_COST_STEP \
478 				/ CPPC_EM_CAP_STEP)
479 
480 static unsigned int get_perf_level_count(struct cpufreq_policy *policy)
481 {
482 	struct cppc_perf_caps *perf_caps;
483 	unsigned int min_cap, max_cap;
484 	struct cppc_cpudata *cpu_data;
485 	int cpu = policy->cpu;
486 
487 	cpu_data = policy->driver_data;
488 	perf_caps = &cpu_data->perf_caps;
489 	max_cap = arch_scale_cpu_capacity(cpu);
490 	min_cap = div_u64((u64)max_cap * perf_caps->lowest_perf,
491 			  perf_caps->highest_perf);
492 	if ((min_cap == 0) || (max_cap < min_cap))
493 		return 0;
494 	return 1 + max_cap / CPPC_EM_CAP_STEP - min_cap / CPPC_EM_CAP_STEP;
495 }
496 
497 /*
498  * The cost is defined as:
499  *   cost = power * max_frequency / frequency
500  */
501 static inline unsigned long compute_cost(int cpu, int step)
502 {
503 	return CPPC_EM_COST_GAP * per_cpu(efficiency_class, cpu) +
504 			step * CPPC_EM_COST_STEP;
505 }
506 
507 static int cppc_get_cpu_power(struct device *cpu_dev,
508 		unsigned long *power, unsigned long *KHz)
509 {
510 	unsigned long perf_step, perf_prev, perf, perf_check;
511 	unsigned int min_step, max_step, step, step_check;
512 	unsigned long prev_freq = *KHz;
513 	unsigned int min_cap, max_cap;
514 	struct cpufreq_policy *policy;
515 
516 	struct cppc_perf_caps *perf_caps;
517 	struct cppc_cpudata *cpu_data;
518 
519 	policy = cpufreq_cpu_get_raw(cpu_dev->id);
520 	cpu_data = policy->driver_data;
521 	perf_caps = &cpu_data->perf_caps;
522 	max_cap = arch_scale_cpu_capacity(cpu_dev->id);
523 	min_cap = div_u64((u64)max_cap * perf_caps->lowest_perf,
524 			  perf_caps->highest_perf);
525 	perf_step = div_u64((u64)CPPC_EM_CAP_STEP * perf_caps->highest_perf,
526 			    max_cap);
527 	min_step = min_cap / CPPC_EM_CAP_STEP;
528 	max_step = max_cap / CPPC_EM_CAP_STEP;
529 
530 	perf_prev = cppc_cpufreq_khz_to_perf(cpu_data, *KHz);
531 	step = perf_prev / perf_step;
532 
533 	if (step > max_step)
534 		return -EINVAL;
535 
536 	if (min_step == max_step) {
537 		step = max_step;
538 		perf = perf_caps->highest_perf;
539 	} else if (step < min_step) {
540 		step = min_step;
541 		perf = perf_caps->lowest_perf;
542 	} else {
543 		step++;
544 		if (step == max_step)
545 			perf = perf_caps->highest_perf;
546 		else
547 			perf = step * perf_step;
548 	}
549 
550 	*KHz = cppc_cpufreq_perf_to_khz(cpu_data, perf);
551 	perf_check = cppc_cpufreq_khz_to_perf(cpu_data, *KHz);
552 	step_check = perf_check / perf_step;
553 
554 	/*
555 	 * To avoid bad integer approximation, check that new frequency value
556 	 * increased and that the new frequency will be converted to the
557 	 * desired step value.
558 	 */
559 	while ((*KHz == prev_freq) || (step_check != step)) {
560 		perf++;
561 		*KHz = cppc_cpufreq_perf_to_khz(cpu_data, perf);
562 		perf_check = cppc_cpufreq_khz_to_perf(cpu_data, *KHz);
563 		step_check = perf_check / perf_step;
564 	}
565 
566 	/*
567 	 * With an artificial EM, only the cost value is used. Still the power
568 	 * is populated such as 0 < power < EM_MAX_POWER. This allows to add
569 	 * more sense to the artificial performance states.
570 	 */
571 	*power = compute_cost(cpu_dev->id, step);
572 
573 	return 0;
574 }
575 
576 static int cppc_get_cpu_cost(struct device *cpu_dev, unsigned long KHz,
577 		unsigned long *cost)
578 {
579 	unsigned long perf_step, perf_prev;
580 	struct cppc_perf_caps *perf_caps;
581 	struct cpufreq_policy *policy;
582 	struct cppc_cpudata *cpu_data;
583 	unsigned int max_cap;
584 	int step;
585 
586 	policy = cpufreq_cpu_get_raw(cpu_dev->id);
587 	cpu_data = policy->driver_data;
588 	perf_caps = &cpu_data->perf_caps;
589 	max_cap = arch_scale_cpu_capacity(cpu_dev->id);
590 
591 	perf_prev = cppc_cpufreq_khz_to_perf(cpu_data, KHz);
592 	perf_step = CPPC_EM_CAP_STEP * perf_caps->highest_perf / max_cap;
593 	step = perf_prev / perf_step;
594 
595 	*cost = compute_cost(cpu_dev->id, step);
596 
597 	return 0;
598 }
599 
600 static int populate_efficiency_class(void)
601 {
602 	struct acpi_madt_generic_interrupt *gicc;
603 	DECLARE_BITMAP(used_classes, 256) = {};
604 	int class, cpu, index;
605 
606 	for_each_possible_cpu(cpu) {
607 		gicc = acpi_cpu_get_madt_gicc(cpu);
608 		class = gicc->efficiency_class;
609 		bitmap_set(used_classes, class, 1);
610 	}
611 
612 	if (bitmap_weight(used_classes, 256) <= 1) {
613 		pr_debug("Efficiency classes are all equal (=%d). "
614 			"No EM registered", class);
615 		return -EINVAL;
616 	}
617 
618 	/*
619 	 * Squeeze efficiency class values on [0:#efficiency_class-1].
620 	 * Values are per spec in [0:255].
621 	 */
622 	index = 0;
623 	for_each_set_bit(class, used_classes, 256) {
624 		for_each_possible_cpu(cpu) {
625 			gicc = acpi_cpu_get_madt_gicc(cpu);
626 			if (gicc->efficiency_class == class)
627 				per_cpu(efficiency_class, cpu) = index;
628 		}
629 		index++;
630 	}
631 	cppc_cpufreq_driver.register_em = cppc_cpufreq_register_em;
632 
633 	return 0;
634 }
635 
636 static void cppc_cpufreq_register_em(struct cpufreq_policy *policy)
637 {
638 	struct cppc_cpudata *cpu_data;
639 	struct em_data_callback em_cb =
640 		EM_ADV_DATA_CB(cppc_get_cpu_power, cppc_get_cpu_cost);
641 
642 	cpu_data = policy->driver_data;
643 	em_dev_register_perf_domain(get_cpu_device(policy->cpu),
644 			get_perf_level_count(policy), &em_cb,
645 			cpu_data->shared_cpu_map, 0);
646 }
647 
648 #else
649 static int populate_efficiency_class(void)
650 {
651 	return 0;
652 }
653 #endif
654 
655 static struct cppc_cpudata *cppc_cpufreq_get_cpu_data(unsigned int cpu)
656 {
657 	struct cppc_cpudata *cpu_data;
658 	int ret;
659 
660 	cpu_data = kzalloc(sizeof(struct cppc_cpudata), GFP_KERNEL);
661 	if (!cpu_data)
662 		goto out;
663 
664 	if (!zalloc_cpumask_var(&cpu_data->shared_cpu_map, GFP_KERNEL))
665 		goto free_cpu;
666 
667 	ret = acpi_get_psd_map(cpu, cpu_data);
668 	if (ret) {
669 		pr_debug("Err parsing CPU%d PSD data: ret:%d\n", cpu, ret);
670 		goto free_mask;
671 	}
672 
673 	ret = cppc_get_perf_caps(cpu, &cpu_data->perf_caps);
674 	if (ret) {
675 		pr_debug("Err reading CPU%d perf caps: ret:%d\n", cpu, ret);
676 		goto free_mask;
677 	}
678 
679 	/* Convert the lowest and nominal freq from MHz to KHz */
680 	cpu_data->perf_caps.lowest_freq *= 1000;
681 	cpu_data->perf_caps.nominal_freq *= 1000;
682 
683 	list_add(&cpu_data->node, &cpu_data_list);
684 
685 	return cpu_data;
686 
687 free_mask:
688 	free_cpumask_var(cpu_data->shared_cpu_map);
689 free_cpu:
690 	kfree(cpu_data);
691 out:
692 	return NULL;
693 }
694 
695 static void cppc_cpufreq_put_cpu_data(struct cpufreq_policy *policy)
696 {
697 	struct cppc_cpudata *cpu_data = policy->driver_data;
698 
699 	list_del(&cpu_data->node);
700 	free_cpumask_var(cpu_data->shared_cpu_map);
701 	kfree(cpu_data);
702 	policy->driver_data = NULL;
703 }
704 
705 static int cppc_cpufreq_cpu_init(struct cpufreq_policy *policy)
706 {
707 	unsigned int cpu = policy->cpu;
708 	struct cppc_cpudata *cpu_data;
709 	struct cppc_perf_caps *caps;
710 	int ret;
711 
712 	cpu_data = cppc_cpufreq_get_cpu_data(cpu);
713 	if (!cpu_data) {
714 		pr_err("Error in acquiring _CPC/_PSD data for CPU%d.\n", cpu);
715 		return -ENODEV;
716 	}
717 	caps = &cpu_data->perf_caps;
718 	policy->driver_data = cpu_data;
719 
720 	/*
721 	 * Set min to lowest nonlinear perf to avoid any efficiency penalty (see
722 	 * Section 8.4.7.1.1.5 of ACPI 6.1 spec)
723 	 */
724 	policy->min = cppc_cpufreq_perf_to_khz(cpu_data,
725 					       caps->lowest_nonlinear_perf);
726 	policy->max = cppc_cpufreq_perf_to_khz(cpu_data,
727 					       caps->nominal_perf);
728 
729 	/*
730 	 * Set cpuinfo.min_freq to Lowest to make the full range of performance
731 	 * available if userspace wants to use any perf between lowest & lowest
732 	 * nonlinear perf
733 	 */
734 	policy->cpuinfo.min_freq = cppc_cpufreq_perf_to_khz(cpu_data,
735 							    caps->lowest_perf);
736 	policy->cpuinfo.max_freq = cppc_cpufreq_perf_to_khz(cpu_data,
737 							    caps->nominal_perf);
738 
739 	policy->transition_delay_us = cppc_cpufreq_get_transition_delay_us(cpu);
740 	policy->shared_type = cpu_data->shared_type;
741 
742 	switch (policy->shared_type) {
743 	case CPUFREQ_SHARED_TYPE_HW:
744 	case CPUFREQ_SHARED_TYPE_NONE:
745 		/* Nothing to be done - we'll have a policy for each CPU */
746 		break;
747 	case CPUFREQ_SHARED_TYPE_ANY:
748 		/*
749 		 * All CPUs in the domain will share a policy and all cpufreq
750 		 * operations will use a single cppc_cpudata structure stored
751 		 * in policy->driver_data.
752 		 */
753 		cpumask_copy(policy->cpus, cpu_data->shared_cpu_map);
754 		break;
755 	default:
756 		pr_debug("Unsupported CPU co-ord type: %d\n",
757 			 policy->shared_type);
758 		ret = -EFAULT;
759 		goto out;
760 	}
761 
762 	policy->fast_switch_possible = cppc_allow_fast_switch();
763 	policy->dvfs_possible_from_any_cpu = true;
764 
765 	/*
766 	 * If 'highest_perf' is greater than 'nominal_perf', we assume CPU Boost
767 	 * is supported.
768 	 */
769 	if (caps->highest_perf > caps->nominal_perf)
770 		boost_supported = true;
771 
772 	/* Set policy->cur to max now. The governors will adjust later. */
773 	policy->cur = cppc_cpufreq_perf_to_khz(cpu_data, caps->highest_perf);
774 	cpu_data->perf_ctrls.desired_perf =  caps->highest_perf;
775 
776 	ret = cppc_set_perf(cpu, &cpu_data->perf_ctrls);
777 	if (ret) {
778 		pr_debug("Err setting perf value:%d on CPU:%d. ret:%d\n",
779 			 caps->highest_perf, cpu, ret);
780 		goto out;
781 	}
782 
783 	cppc_cpufreq_cpu_fie_init(policy);
784 	return 0;
785 
786 out:
787 	cppc_cpufreq_put_cpu_data(policy);
788 	return ret;
789 }
790 
791 static int cppc_cpufreq_cpu_exit(struct cpufreq_policy *policy)
792 {
793 	struct cppc_cpudata *cpu_data = policy->driver_data;
794 	struct cppc_perf_caps *caps = &cpu_data->perf_caps;
795 	unsigned int cpu = policy->cpu;
796 	int ret;
797 
798 	cppc_cpufreq_cpu_fie_exit(policy);
799 
800 	cpu_data->perf_ctrls.desired_perf = caps->lowest_perf;
801 
802 	ret = cppc_set_perf(cpu, &cpu_data->perf_ctrls);
803 	if (ret)
804 		pr_debug("Err setting perf value:%d on CPU:%d. ret:%d\n",
805 			 caps->lowest_perf, cpu, ret);
806 
807 	cppc_cpufreq_put_cpu_data(policy);
808 	return 0;
809 }
810 
811 static inline u64 get_delta(u64 t1, u64 t0)
812 {
813 	if (t1 > t0 || t0 > ~(u32)0)
814 		return t1 - t0;
815 
816 	return (u32)t1 - (u32)t0;
817 }
818 
819 static int cppc_perf_from_fbctrs(struct cppc_cpudata *cpu_data,
820 				 struct cppc_perf_fb_ctrs *fb_ctrs_t0,
821 				 struct cppc_perf_fb_ctrs *fb_ctrs_t1)
822 {
823 	u64 delta_reference, delta_delivered;
824 	u64 reference_perf;
825 
826 	reference_perf = fb_ctrs_t0->reference_perf;
827 
828 	delta_reference = get_delta(fb_ctrs_t1->reference,
829 				    fb_ctrs_t0->reference);
830 	delta_delivered = get_delta(fb_ctrs_t1->delivered,
831 				    fb_ctrs_t0->delivered);
832 
833 	/* Check to avoid divide-by zero and invalid delivered_perf */
834 	if (!delta_reference || !delta_delivered)
835 		return cpu_data->perf_ctrls.desired_perf;
836 
837 	return (reference_perf * delta_delivered) / delta_reference;
838 }
839 
840 static unsigned int cppc_cpufreq_get_rate(unsigned int cpu)
841 {
842 	struct cppc_perf_fb_ctrs fb_ctrs_t0 = {0}, fb_ctrs_t1 = {0};
843 	struct cpufreq_policy *policy = cpufreq_cpu_get(cpu);
844 	struct cppc_cpudata *cpu_data = policy->driver_data;
845 	u64 delivered_perf;
846 	int ret;
847 
848 	cpufreq_cpu_put(policy);
849 
850 	ret = cppc_get_perf_ctrs(cpu, &fb_ctrs_t0);
851 	if (ret)
852 		return ret;
853 
854 	udelay(2); /* 2usec delay between sampling */
855 
856 	ret = cppc_get_perf_ctrs(cpu, &fb_ctrs_t1);
857 	if (ret)
858 		return ret;
859 
860 	delivered_perf = cppc_perf_from_fbctrs(cpu_data, &fb_ctrs_t0,
861 					       &fb_ctrs_t1);
862 
863 	return cppc_cpufreq_perf_to_khz(cpu_data, delivered_perf);
864 }
865 
866 static int cppc_cpufreq_set_boost(struct cpufreq_policy *policy, int state)
867 {
868 	struct cppc_cpudata *cpu_data = policy->driver_data;
869 	struct cppc_perf_caps *caps = &cpu_data->perf_caps;
870 	int ret;
871 
872 	if (!boost_supported) {
873 		pr_err("BOOST not supported by CPU or firmware\n");
874 		return -EINVAL;
875 	}
876 
877 	if (state)
878 		policy->max = cppc_cpufreq_perf_to_khz(cpu_data,
879 						       caps->highest_perf);
880 	else
881 		policy->max = cppc_cpufreq_perf_to_khz(cpu_data,
882 						       caps->nominal_perf);
883 	policy->cpuinfo.max_freq = policy->max;
884 
885 	ret = freq_qos_update_request(policy->max_freq_req, policy->max);
886 	if (ret < 0)
887 		return ret;
888 
889 	return 0;
890 }
891 
892 static ssize_t show_freqdomain_cpus(struct cpufreq_policy *policy, char *buf)
893 {
894 	struct cppc_cpudata *cpu_data = policy->driver_data;
895 
896 	return cpufreq_show_cpus(cpu_data->shared_cpu_map, buf);
897 }
898 cpufreq_freq_attr_ro(freqdomain_cpus);
899 
900 static struct freq_attr *cppc_cpufreq_attr[] = {
901 	&freqdomain_cpus,
902 	NULL,
903 };
904 
905 static struct cpufreq_driver cppc_cpufreq_driver = {
906 	.flags = CPUFREQ_CONST_LOOPS,
907 	.verify = cppc_verify_policy,
908 	.target = cppc_cpufreq_set_target,
909 	.get = cppc_cpufreq_get_rate,
910 	.fast_switch = cppc_cpufreq_fast_switch,
911 	.init = cppc_cpufreq_cpu_init,
912 	.exit = cppc_cpufreq_cpu_exit,
913 	.set_boost = cppc_cpufreq_set_boost,
914 	.attr = cppc_cpufreq_attr,
915 	.name = "cppc_cpufreq",
916 };
917 
918 /*
919  * HISI platform does not support delivered performance counter and
920  * reference performance counter. It can calculate the performance using the
921  * platform specific mechanism. We reuse the desired performance register to
922  * store the real performance calculated by the platform.
923  */
924 static unsigned int hisi_cppc_cpufreq_get_rate(unsigned int cpu)
925 {
926 	struct cpufreq_policy *policy = cpufreq_cpu_get(cpu);
927 	struct cppc_cpudata *cpu_data = policy->driver_data;
928 	u64 desired_perf;
929 	int ret;
930 
931 	cpufreq_cpu_put(policy);
932 
933 	ret = cppc_get_desired_perf(cpu, &desired_perf);
934 	if (ret < 0)
935 		return -EIO;
936 
937 	return cppc_cpufreq_perf_to_khz(cpu_data, desired_perf);
938 }
939 
940 static void cppc_check_hisi_workaround(void)
941 {
942 	struct acpi_table_header *tbl;
943 	acpi_status status = AE_OK;
944 	int i;
945 
946 	status = acpi_get_table(ACPI_SIG_PCCT, 0, &tbl);
947 	if (ACPI_FAILURE(status) || !tbl)
948 		return;
949 
950 	for (i = 0; i < ARRAY_SIZE(wa_info); i++) {
951 		if (!memcmp(wa_info[i].oem_id, tbl->oem_id, ACPI_OEM_ID_SIZE) &&
952 		    !memcmp(wa_info[i].oem_table_id, tbl->oem_table_id, ACPI_OEM_TABLE_ID_SIZE) &&
953 		    wa_info[i].oem_revision == tbl->oem_revision) {
954 			/* Overwrite the get() callback */
955 			cppc_cpufreq_driver.get = hisi_cppc_cpufreq_get_rate;
956 			fie_disabled = FIE_DISABLED;
957 			break;
958 		}
959 	}
960 
961 	acpi_put_table(tbl);
962 }
963 
964 static int __init cppc_cpufreq_init(void)
965 {
966 	int ret;
967 
968 	if (!acpi_cpc_valid())
969 		return -ENODEV;
970 
971 	cppc_check_hisi_workaround();
972 	cppc_freq_invariance_init();
973 	populate_efficiency_class();
974 
975 	ret = cpufreq_register_driver(&cppc_cpufreq_driver);
976 	if (ret)
977 		cppc_freq_invariance_exit();
978 
979 	return ret;
980 }
981 
982 static inline void free_cpu_data(void)
983 {
984 	struct cppc_cpudata *iter, *tmp;
985 
986 	list_for_each_entry_safe(iter, tmp, &cpu_data_list, node) {
987 		free_cpumask_var(iter->shared_cpu_map);
988 		list_del(&iter->node);
989 		kfree(iter);
990 	}
991 
992 }
993 
994 static void __exit cppc_cpufreq_exit(void)
995 {
996 	cpufreq_unregister_driver(&cppc_cpufreq_driver);
997 	cppc_freq_invariance_exit();
998 
999 	free_cpu_data();
1000 }
1001 
1002 module_exit(cppc_cpufreq_exit);
1003 MODULE_AUTHOR("Ashwin Chaugule");
1004 MODULE_DESCRIPTION("CPUFreq driver based on the ACPI CPPC v5.0+ spec");
1005 MODULE_LICENSE("GPL");
1006 
1007 late_initcall(cppc_cpufreq_init);
1008 
1009 static const struct acpi_device_id cppc_acpi_ids[] __used = {
1010 	{ACPI_PROCESSOR_DEVICE_HID, },
1011 	{}
1012 };
1013 
1014 MODULE_DEVICE_TABLE(acpi, cppc_acpi_ids);
1015