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(max_cap * perf_caps->lowest_perf, perf_caps->highest_perf);
491 	if ((min_cap == 0) || (max_cap < min_cap))
492 		return 0;
493 	return 1 + max_cap / CPPC_EM_CAP_STEP - min_cap / CPPC_EM_CAP_STEP;
494 }
495 
496 /*
497  * The cost is defined as:
498  *   cost = power * max_frequency / frequency
499  */
500 static inline unsigned long compute_cost(int cpu, int step)
501 {
502 	return CPPC_EM_COST_GAP * per_cpu(efficiency_class, cpu) +
503 			step * CPPC_EM_COST_STEP;
504 }
505 
506 static int cppc_get_cpu_power(struct device *cpu_dev,
507 		unsigned long *power, unsigned long *KHz)
508 {
509 	unsigned long perf_step, perf_prev, perf, perf_check;
510 	unsigned int min_step, max_step, step, step_check;
511 	unsigned long prev_freq = *KHz;
512 	unsigned int min_cap, max_cap;
513 	struct cpufreq_policy *policy;
514 
515 	struct cppc_perf_caps *perf_caps;
516 	struct cppc_cpudata *cpu_data;
517 
518 	policy = cpufreq_cpu_get_raw(cpu_dev->id);
519 	cpu_data = policy->driver_data;
520 	perf_caps = &cpu_data->perf_caps;
521 	max_cap = arch_scale_cpu_capacity(cpu_dev->id);
522 	min_cap = div_u64(max_cap * perf_caps->lowest_perf,
523 			perf_caps->highest_perf);
524 
525 	perf_step = CPPC_EM_CAP_STEP * perf_caps->highest_perf / max_cap;
526 	min_step = min_cap / CPPC_EM_CAP_STEP;
527 	max_step = max_cap / CPPC_EM_CAP_STEP;
528 
529 	perf_prev = cppc_cpufreq_khz_to_perf(cpu_data, *KHz);
530 	step = perf_prev / perf_step;
531 
532 	if (step > max_step)
533 		return -EINVAL;
534 
535 	if (min_step == max_step) {
536 		step = max_step;
537 		perf = perf_caps->highest_perf;
538 	} else if (step < min_step) {
539 		step = min_step;
540 		perf = perf_caps->lowest_perf;
541 	} else {
542 		step++;
543 		if (step == max_step)
544 			perf = perf_caps->highest_perf;
545 		else
546 			perf = step * perf_step;
547 	}
548 
549 	*KHz = cppc_cpufreq_perf_to_khz(cpu_data, perf);
550 	perf_check = cppc_cpufreq_khz_to_perf(cpu_data, *KHz);
551 	step_check = perf_check / perf_step;
552 
553 	/*
554 	 * To avoid bad integer approximation, check that new frequency value
555 	 * increased and that the new frequency will be converted to the
556 	 * desired step value.
557 	 */
558 	while ((*KHz == prev_freq) || (step_check != step)) {
559 		perf++;
560 		*KHz = cppc_cpufreq_perf_to_khz(cpu_data, perf);
561 		perf_check = cppc_cpufreq_khz_to_perf(cpu_data, *KHz);
562 		step_check = perf_check / perf_step;
563 	}
564 
565 	/*
566 	 * With an artificial EM, only the cost value is used. Still the power
567 	 * is populated such as 0 < power < EM_MAX_POWER. This allows to add
568 	 * more sense to the artificial performance states.
569 	 */
570 	*power = compute_cost(cpu_dev->id, step);
571 
572 	return 0;
573 }
574 
575 static int cppc_get_cpu_cost(struct device *cpu_dev, unsigned long KHz,
576 		unsigned long *cost)
577 {
578 	unsigned long perf_step, perf_prev;
579 	struct cppc_perf_caps *perf_caps;
580 	struct cpufreq_policy *policy;
581 	struct cppc_cpudata *cpu_data;
582 	unsigned int max_cap;
583 	int step;
584 
585 	policy = cpufreq_cpu_get_raw(cpu_dev->id);
586 	cpu_data = policy->driver_data;
587 	perf_caps = &cpu_data->perf_caps;
588 	max_cap = arch_scale_cpu_capacity(cpu_dev->id);
589 
590 	perf_prev = cppc_cpufreq_khz_to_perf(cpu_data, KHz);
591 	perf_step = CPPC_EM_CAP_STEP * perf_caps->highest_perf / max_cap;
592 	step = perf_prev / perf_step;
593 
594 	*cost = compute_cost(cpu_dev->id, step);
595 
596 	return 0;
597 }
598 
599 static int populate_efficiency_class(void)
600 {
601 	struct acpi_madt_generic_interrupt *gicc;
602 	DECLARE_BITMAP(used_classes, 256) = {};
603 	int class, cpu, index;
604 
605 	for_each_possible_cpu(cpu) {
606 		gicc = acpi_cpu_get_madt_gicc(cpu);
607 		class = gicc->efficiency_class;
608 		bitmap_set(used_classes, class, 1);
609 	}
610 
611 	if (bitmap_weight(used_classes, 256) <= 1) {
612 		pr_debug("Efficiency classes are all equal (=%d). "
613 			"No EM registered", class);
614 		return -EINVAL;
615 	}
616 
617 	/*
618 	 * Squeeze efficiency class values on [0:#efficiency_class-1].
619 	 * Values are per spec in [0:255].
620 	 */
621 	index = 0;
622 	for_each_set_bit(class, used_classes, 256) {
623 		for_each_possible_cpu(cpu) {
624 			gicc = acpi_cpu_get_madt_gicc(cpu);
625 			if (gicc->efficiency_class == class)
626 				per_cpu(efficiency_class, cpu) = index;
627 		}
628 		index++;
629 	}
630 	cppc_cpufreq_driver.register_em = cppc_cpufreq_register_em;
631 
632 	return 0;
633 }
634 
635 static void cppc_cpufreq_register_em(struct cpufreq_policy *policy)
636 {
637 	struct cppc_cpudata *cpu_data;
638 	struct em_data_callback em_cb =
639 		EM_ADV_DATA_CB(cppc_get_cpu_power, cppc_get_cpu_cost);
640 
641 	cpu_data = policy->driver_data;
642 	em_dev_register_perf_domain(get_cpu_device(policy->cpu),
643 			get_perf_level_count(policy), &em_cb,
644 			cpu_data->shared_cpu_map, 0);
645 }
646 
647 #else
648 static int populate_efficiency_class(void)
649 {
650 	return 0;
651 }
652 #endif
653 
654 static struct cppc_cpudata *cppc_cpufreq_get_cpu_data(unsigned int cpu)
655 {
656 	struct cppc_cpudata *cpu_data;
657 	int ret;
658 
659 	cpu_data = kzalloc(sizeof(struct cppc_cpudata), GFP_KERNEL);
660 	if (!cpu_data)
661 		goto out;
662 
663 	if (!zalloc_cpumask_var(&cpu_data->shared_cpu_map, GFP_KERNEL))
664 		goto free_cpu;
665 
666 	ret = acpi_get_psd_map(cpu, cpu_data);
667 	if (ret) {
668 		pr_debug("Err parsing CPU%d PSD data: ret:%d\n", cpu, ret);
669 		goto free_mask;
670 	}
671 
672 	ret = cppc_get_perf_caps(cpu, &cpu_data->perf_caps);
673 	if (ret) {
674 		pr_debug("Err reading CPU%d perf caps: ret:%d\n", cpu, ret);
675 		goto free_mask;
676 	}
677 
678 	/* Convert the lowest and nominal freq from MHz to KHz */
679 	cpu_data->perf_caps.lowest_freq *= 1000;
680 	cpu_data->perf_caps.nominal_freq *= 1000;
681 
682 	list_add(&cpu_data->node, &cpu_data_list);
683 
684 	return cpu_data;
685 
686 free_mask:
687 	free_cpumask_var(cpu_data->shared_cpu_map);
688 free_cpu:
689 	kfree(cpu_data);
690 out:
691 	return NULL;
692 }
693 
694 static void cppc_cpufreq_put_cpu_data(struct cpufreq_policy *policy)
695 {
696 	struct cppc_cpudata *cpu_data = policy->driver_data;
697 
698 	list_del(&cpu_data->node);
699 	free_cpumask_var(cpu_data->shared_cpu_map);
700 	kfree(cpu_data);
701 	policy->driver_data = NULL;
702 }
703 
704 static int cppc_cpufreq_cpu_init(struct cpufreq_policy *policy)
705 {
706 	unsigned int cpu = policy->cpu;
707 	struct cppc_cpudata *cpu_data;
708 	struct cppc_perf_caps *caps;
709 	int ret;
710 
711 	cpu_data = cppc_cpufreq_get_cpu_data(cpu);
712 	if (!cpu_data) {
713 		pr_err("Error in acquiring _CPC/_PSD data for CPU%d.\n", cpu);
714 		return -ENODEV;
715 	}
716 	caps = &cpu_data->perf_caps;
717 	policy->driver_data = cpu_data;
718 
719 	/*
720 	 * Set min to lowest nonlinear perf to avoid any efficiency penalty (see
721 	 * Section 8.4.7.1.1.5 of ACPI 6.1 spec)
722 	 */
723 	policy->min = cppc_cpufreq_perf_to_khz(cpu_data,
724 					       caps->lowest_nonlinear_perf);
725 	policy->max = cppc_cpufreq_perf_to_khz(cpu_data,
726 					       caps->nominal_perf);
727 
728 	/*
729 	 * Set cpuinfo.min_freq to Lowest to make the full range of performance
730 	 * available if userspace wants to use any perf between lowest & lowest
731 	 * nonlinear perf
732 	 */
733 	policy->cpuinfo.min_freq = cppc_cpufreq_perf_to_khz(cpu_data,
734 							    caps->lowest_perf);
735 	policy->cpuinfo.max_freq = cppc_cpufreq_perf_to_khz(cpu_data,
736 							    caps->nominal_perf);
737 
738 	policy->transition_delay_us = cppc_cpufreq_get_transition_delay_us(cpu);
739 	policy->shared_type = cpu_data->shared_type;
740 
741 	switch (policy->shared_type) {
742 	case CPUFREQ_SHARED_TYPE_HW:
743 	case CPUFREQ_SHARED_TYPE_NONE:
744 		/* Nothing to be done - we'll have a policy for each CPU */
745 		break;
746 	case CPUFREQ_SHARED_TYPE_ANY:
747 		/*
748 		 * All CPUs in the domain will share a policy and all cpufreq
749 		 * operations will use a single cppc_cpudata structure stored
750 		 * in policy->driver_data.
751 		 */
752 		cpumask_copy(policy->cpus, cpu_data->shared_cpu_map);
753 		break;
754 	default:
755 		pr_debug("Unsupported CPU co-ord type: %d\n",
756 			 policy->shared_type);
757 		ret = -EFAULT;
758 		goto out;
759 	}
760 
761 	policy->fast_switch_possible = cppc_allow_fast_switch();
762 	policy->dvfs_possible_from_any_cpu = true;
763 
764 	/*
765 	 * If 'highest_perf' is greater than 'nominal_perf', we assume CPU Boost
766 	 * is supported.
767 	 */
768 	if (caps->highest_perf > caps->nominal_perf)
769 		boost_supported = true;
770 
771 	/* Set policy->cur to max now. The governors will adjust later. */
772 	policy->cur = cppc_cpufreq_perf_to_khz(cpu_data, caps->highest_perf);
773 	cpu_data->perf_ctrls.desired_perf =  caps->highest_perf;
774 
775 	ret = cppc_set_perf(cpu, &cpu_data->perf_ctrls);
776 	if (ret) {
777 		pr_debug("Err setting perf value:%d on CPU:%d. ret:%d\n",
778 			 caps->highest_perf, cpu, ret);
779 		goto out;
780 	}
781 
782 	cppc_cpufreq_cpu_fie_init(policy);
783 	return 0;
784 
785 out:
786 	cppc_cpufreq_put_cpu_data(policy);
787 	return ret;
788 }
789 
790 static int cppc_cpufreq_cpu_exit(struct cpufreq_policy *policy)
791 {
792 	struct cppc_cpudata *cpu_data = policy->driver_data;
793 	struct cppc_perf_caps *caps = &cpu_data->perf_caps;
794 	unsigned int cpu = policy->cpu;
795 	int ret;
796 
797 	cppc_cpufreq_cpu_fie_exit(policy);
798 
799 	cpu_data->perf_ctrls.desired_perf = caps->lowest_perf;
800 
801 	ret = cppc_set_perf(cpu, &cpu_data->perf_ctrls);
802 	if (ret)
803 		pr_debug("Err setting perf value:%d on CPU:%d. ret:%d\n",
804 			 caps->lowest_perf, cpu, ret);
805 
806 	cppc_cpufreq_put_cpu_data(policy);
807 	return 0;
808 }
809 
810 static inline u64 get_delta(u64 t1, u64 t0)
811 {
812 	if (t1 > t0 || t0 > ~(u32)0)
813 		return t1 - t0;
814 
815 	return (u32)t1 - (u32)t0;
816 }
817 
818 static int cppc_perf_from_fbctrs(struct cppc_cpudata *cpu_data,
819 				 struct cppc_perf_fb_ctrs *fb_ctrs_t0,
820 				 struct cppc_perf_fb_ctrs *fb_ctrs_t1)
821 {
822 	u64 delta_reference, delta_delivered;
823 	u64 reference_perf;
824 
825 	reference_perf = fb_ctrs_t0->reference_perf;
826 
827 	delta_reference = get_delta(fb_ctrs_t1->reference,
828 				    fb_ctrs_t0->reference);
829 	delta_delivered = get_delta(fb_ctrs_t1->delivered,
830 				    fb_ctrs_t0->delivered);
831 
832 	/* Check to avoid divide-by zero and invalid delivered_perf */
833 	if (!delta_reference || !delta_delivered)
834 		return cpu_data->perf_ctrls.desired_perf;
835 
836 	return (reference_perf * delta_delivered) / delta_reference;
837 }
838 
839 static unsigned int cppc_cpufreq_get_rate(unsigned int cpu)
840 {
841 	struct cppc_perf_fb_ctrs fb_ctrs_t0 = {0}, fb_ctrs_t1 = {0};
842 	struct cpufreq_policy *policy = cpufreq_cpu_get(cpu);
843 	struct cppc_cpudata *cpu_data = policy->driver_data;
844 	u64 delivered_perf;
845 	int ret;
846 
847 	cpufreq_cpu_put(policy);
848 
849 	ret = cppc_get_perf_ctrs(cpu, &fb_ctrs_t0);
850 	if (ret)
851 		return ret;
852 
853 	udelay(2); /* 2usec delay between sampling */
854 
855 	ret = cppc_get_perf_ctrs(cpu, &fb_ctrs_t1);
856 	if (ret)
857 		return ret;
858 
859 	delivered_perf = cppc_perf_from_fbctrs(cpu_data, &fb_ctrs_t0,
860 					       &fb_ctrs_t1);
861 
862 	return cppc_cpufreq_perf_to_khz(cpu_data, delivered_perf);
863 }
864 
865 static int cppc_cpufreq_set_boost(struct cpufreq_policy *policy, int state)
866 {
867 	struct cppc_cpudata *cpu_data = policy->driver_data;
868 	struct cppc_perf_caps *caps = &cpu_data->perf_caps;
869 	int ret;
870 
871 	if (!boost_supported) {
872 		pr_err("BOOST not supported by CPU or firmware\n");
873 		return -EINVAL;
874 	}
875 
876 	if (state)
877 		policy->max = cppc_cpufreq_perf_to_khz(cpu_data,
878 						       caps->highest_perf);
879 	else
880 		policy->max = cppc_cpufreq_perf_to_khz(cpu_data,
881 						       caps->nominal_perf);
882 	policy->cpuinfo.max_freq = policy->max;
883 
884 	ret = freq_qos_update_request(policy->max_freq_req, policy->max);
885 	if (ret < 0)
886 		return ret;
887 
888 	return 0;
889 }
890 
891 static ssize_t show_freqdomain_cpus(struct cpufreq_policy *policy, char *buf)
892 {
893 	struct cppc_cpudata *cpu_data = policy->driver_data;
894 
895 	return cpufreq_show_cpus(cpu_data->shared_cpu_map, buf);
896 }
897 cpufreq_freq_attr_ro(freqdomain_cpus);
898 
899 static struct freq_attr *cppc_cpufreq_attr[] = {
900 	&freqdomain_cpus,
901 	NULL,
902 };
903 
904 static struct cpufreq_driver cppc_cpufreq_driver = {
905 	.flags = CPUFREQ_CONST_LOOPS,
906 	.verify = cppc_verify_policy,
907 	.target = cppc_cpufreq_set_target,
908 	.get = cppc_cpufreq_get_rate,
909 	.fast_switch = cppc_cpufreq_fast_switch,
910 	.init = cppc_cpufreq_cpu_init,
911 	.exit = cppc_cpufreq_cpu_exit,
912 	.set_boost = cppc_cpufreq_set_boost,
913 	.attr = cppc_cpufreq_attr,
914 	.name = "cppc_cpufreq",
915 };
916 
917 /*
918  * HISI platform does not support delivered performance counter and
919  * reference performance counter. It can calculate the performance using the
920  * platform specific mechanism. We reuse the desired performance register to
921  * store the real performance calculated by the platform.
922  */
923 static unsigned int hisi_cppc_cpufreq_get_rate(unsigned int cpu)
924 {
925 	struct cpufreq_policy *policy = cpufreq_cpu_get(cpu);
926 	struct cppc_cpudata *cpu_data = policy->driver_data;
927 	u64 desired_perf;
928 	int ret;
929 
930 	cpufreq_cpu_put(policy);
931 
932 	ret = cppc_get_desired_perf(cpu, &desired_perf);
933 	if (ret < 0)
934 		return -EIO;
935 
936 	return cppc_cpufreq_perf_to_khz(cpu_data, desired_perf);
937 }
938 
939 static void cppc_check_hisi_workaround(void)
940 {
941 	struct acpi_table_header *tbl;
942 	acpi_status status = AE_OK;
943 	int i;
944 
945 	status = acpi_get_table(ACPI_SIG_PCCT, 0, &tbl);
946 	if (ACPI_FAILURE(status) || !tbl)
947 		return;
948 
949 	for (i = 0; i < ARRAY_SIZE(wa_info); i++) {
950 		if (!memcmp(wa_info[i].oem_id, tbl->oem_id, ACPI_OEM_ID_SIZE) &&
951 		    !memcmp(wa_info[i].oem_table_id, tbl->oem_table_id, ACPI_OEM_TABLE_ID_SIZE) &&
952 		    wa_info[i].oem_revision == tbl->oem_revision) {
953 			/* Overwrite the get() callback */
954 			cppc_cpufreq_driver.get = hisi_cppc_cpufreq_get_rate;
955 			fie_disabled = FIE_DISABLED;
956 			break;
957 		}
958 	}
959 
960 	acpi_put_table(tbl);
961 }
962 
963 static int __init cppc_cpufreq_init(void)
964 {
965 	int ret;
966 
967 	if (!acpi_cpc_valid())
968 		return -ENODEV;
969 
970 	cppc_check_hisi_workaround();
971 	cppc_freq_invariance_init();
972 	populate_efficiency_class();
973 
974 	ret = cpufreq_register_driver(&cppc_cpufreq_driver);
975 	if (ret)
976 		cppc_freq_invariance_exit();
977 
978 	return ret;
979 }
980 
981 static inline void free_cpu_data(void)
982 {
983 	struct cppc_cpudata *iter, *tmp;
984 
985 	list_for_each_entry_safe(iter, tmp, &cpu_data_list, node) {
986 		free_cpumask_var(iter->shared_cpu_map);
987 		list_del(&iter->node);
988 		kfree(iter);
989 	}
990 
991 }
992 
993 static void __exit cppc_cpufreq_exit(void)
994 {
995 	cpufreq_unregister_driver(&cppc_cpufreq_driver);
996 	cppc_freq_invariance_exit();
997 
998 	free_cpu_data();
999 }
1000 
1001 module_exit(cppc_cpufreq_exit);
1002 MODULE_AUTHOR("Ashwin Chaugule");
1003 MODULE_DESCRIPTION("CPUFreq driver based on the ACPI CPPC v5.0+ spec");
1004 MODULE_LICENSE("GPL");
1005 
1006 late_initcall(cppc_cpufreq_init);
1007 
1008 static const struct acpi_device_id cppc_acpi_ids[] __used = {
1009 	{ACPI_PROCESSOR_DEVICE_HID, },
1010 	{}
1011 };
1012 
1013 MODULE_DEVICE_TABLE(acpi, cppc_acpi_ids);
1014