xref: /linux/drivers/cpufreq/cppc_cpufreq.c (revision 32daa5d7899e03433429bedf9e20d7963179703a)
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/kernel.h>
14 #include <linux/module.h>
15 #include <linux/delay.h>
16 #include <linux/cpu.h>
17 #include <linux/cpufreq.h>
18 #include <linux/dmi.h>
19 #include <linux/time.h>
20 #include <linux/vmalloc.h>
21 
22 #include <asm/unaligned.h>
23 
24 #include <acpi/cppc_acpi.h>
25 
26 /* Minimum struct length needed for the DMI processor entry we want */
27 #define DMI_ENTRY_PROCESSOR_MIN_LENGTH	48
28 
29 /* Offset in the DMI processor structure for the max frequency */
30 #define DMI_PROCESSOR_MAX_SPEED		0x14
31 
32 /*
33  * This list contains information parsed from per CPU ACPI _CPC and _PSD
34  * structures: e.g. the highest and lowest supported performance, capabilities,
35  * desired performance, level requested etc. Depending on the share_type, not
36  * all CPUs will have an entry in the list.
37  */
38 static LIST_HEAD(cpu_data_list);
39 
40 static bool boost_supported;
41 
42 struct cppc_workaround_oem_info {
43 	char oem_id[ACPI_OEM_ID_SIZE + 1];
44 	char oem_table_id[ACPI_OEM_TABLE_ID_SIZE + 1];
45 	u32 oem_revision;
46 };
47 
48 static struct cppc_workaround_oem_info wa_info[] = {
49 	{
50 		.oem_id		= "HISI  ",
51 		.oem_table_id	= "HIP07   ",
52 		.oem_revision	= 0,
53 	}, {
54 		.oem_id		= "HISI  ",
55 		.oem_table_id	= "HIP08   ",
56 		.oem_revision	= 0,
57 	}
58 };
59 
60 /* Callback function used to retrieve the max frequency from DMI */
61 static void cppc_find_dmi_mhz(const struct dmi_header *dm, void *private)
62 {
63 	const u8 *dmi_data = (const u8 *)dm;
64 	u16 *mhz = (u16 *)private;
65 
66 	if (dm->type == DMI_ENTRY_PROCESSOR &&
67 	    dm->length >= DMI_ENTRY_PROCESSOR_MIN_LENGTH) {
68 		u16 val = (u16)get_unaligned((const u16 *)
69 				(dmi_data + DMI_PROCESSOR_MAX_SPEED));
70 		*mhz = val > *mhz ? val : *mhz;
71 	}
72 }
73 
74 /* Look up the max frequency in DMI */
75 static u64 cppc_get_dmi_max_khz(void)
76 {
77 	u16 mhz = 0;
78 
79 	dmi_walk(cppc_find_dmi_mhz, &mhz);
80 
81 	/*
82 	 * Real stupid fallback value, just in case there is no
83 	 * actual value set.
84 	 */
85 	mhz = mhz ? mhz : 1;
86 
87 	return (1000 * mhz);
88 }
89 
90 /*
91  * If CPPC lowest_freq and nominal_freq registers are exposed then we can
92  * use them to convert perf to freq and vice versa
93  *
94  * If the perf/freq point lies between Nominal and Lowest, we can treat
95  * (Low perf, Low freq) and (Nom Perf, Nom freq) as 2D co-ordinates of a line
96  * and extrapolate the rest
97  * For perf/freq > Nominal, we use the ratio perf:freq at Nominal for conversion
98  */
99 static unsigned int cppc_cpufreq_perf_to_khz(struct cppc_cpudata *cpu_data,
100 					     unsigned int perf)
101 {
102 	struct cppc_perf_caps *caps = &cpu_data->perf_caps;
103 	static u64 max_khz;
104 	u64 mul, div;
105 
106 	if (caps->lowest_freq && caps->nominal_freq) {
107 		if (perf >= caps->nominal_perf) {
108 			mul = caps->nominal_freq;
109 			div = caps->nominal_perf;
110 		} else {
111 			mul = caps->nominal_freq - caps->lowest_freq;
112 			div = caps->nominal_perf - caps->lowest_perf;
113 		}
114 	} else {
115 		if (!max_khz)
116 			max_khz = cppc_get_dmi_max_khz();
117 		mul = max_khz;
118 		div = caps->highest_perf;
119 	}
120 	return (u64)perf * mul / div;
121 }
122 
123 static unsigned int cppc_cpufreq_khz_to_perf(struct cppc_cpudata *cpu_data,
124 					     unsigned int freq)
125 {
126 	struct cppc_perf_caps *caps = &cpu_data->perf_caps;
127 	static u64 max_khz;
128 	u64  mul, div;
129 
130 	if (caps->lowest_freq && caps->nominal_freq) {
131 		if (freq >= caps->nominal_freq) {
132 			mul = caps->nominal_perf;
133 			div = caps->nominal_freq;
134 		} else {
135 			mul = caps->lowest_perf;
136 			div = caps->lowest_freq;
137 		}
138 	} else {
139 		if (!max_khz)
140 			max_khz = cppc_get_dmi_max_khz();
141 		mul = caps->highest_perf;
142 		div = max_khz;
143 	}
144 
145 	return (u64)freq * mul / div;
146 }
147 
148 static int cppc_cpufreq_set_target(struct cpufreq_policy *policy,
149 				   unsigned int target_freq,
150 				   unsigned int relation)
151 
152 {
153 	struct cppc_cpudata *cpu_data = policy->driver_data;
154 	unsigned int cpu = policy->cpu;
155 	struct cpufreq_freqs freqs;
156 	u32 desired_perf;
157 	int ret = 0;
158 
159 	desired_perf = cppc_cpufreq_khz_to_perf(cpu_data, target_freq);
160 	/* Return if it is exactly the same perf */
161 	if (desired_perf == cpu_data->perf_ctrls.desired_perf)
162 		return ret;
163 
164 	cpu_data->perf_ctrls.desired_perf = desired_perf;
165 	freqs.old = policy->cur;
166 	freqs.new = target_freq;
167 
168 	cpufreq_freq_transition_begin(policy, &freqs);
169 	ret = cppc_set_perf(cpu, &cpu_data->perf_ctrls);
170 	cpufreq_freq_transition_end(policy, &freqs, ret != 0);
171 
172 	if (ret)
173 		pr_debug("Failed to set target on CPU:%d. ret:%d\n",
174 			 cpu, ret);
175 
176 	return ret;
177 }
178 
179 static int cppc_verify_policy(struct cpufreq_policy_data *policy)
180 {
181 	cpufreq_verify_within_cpu_limits(policy);
182 	return 0;
183 }
184 
185 static void cppc_cpufreq_stop_cpu(struct cpufreq_policy *policy)
186 {
187 	struct cppc_cpudata *cpu_data = policy->driver_data;
188 	struct cppc_perf_caps *caps = &cpu_data->perf_caps;
189 	unsigned int cpu = policy->cpu;
190 	int ret;
191 
192 	cpu_data->perf_ctrls.desired_perf = caps->lowest_perf;
193 
194 	ret = cppc_set_perf(cpu, &cpu_data->perf_ctrls);
195 	if (ret)
196 		pr_debug("Err setting perf value:%d on CPU:%d. ret:%d\n",
197 			 caps->lowest_perf, cpu, ret);
198 
199 	/* Remove CPU node from list and free driver data for policy */
200 	free_cpumask_var(cpu_data->shared_cpu_map);
201 	list_del(&cpu_data->node);
202 	kfree(policy->driver_data);
203 	policy->driver_data = NULL;
204 }
205 
206 /*
207  * The PCC subspace describes the rate at which platform can accept commands
208  * on the shared PCC channel (including READs which do not count towards freq
209  * transition requests), so ideally we need to use the PCC values as a fallback
210  * if we don't have a platform specific transition_delay_us
211  */
212 #ifdef CONFIG_ARM64
213 #include <asm/cputype.h>
214 
215 static unsigned int cppc_cpufreq_get_transition_delay_us(unsigned int cpu)
216 {
217 	unsigned long implementor = read_cpuid_implementor();
218 	unsigned long part_num = read_cpuid_part_number();
219 	unsigned int delay_us = 0;
220 
221 	switch (implementor) {
222 	case ARM_CPU_IMP_QCOM:
223 		switch (part_num) {
224 		case QCOM_CPU_PART_FALKOR_V1:
225 		case QCOM_CPU_PART_FALKOR:
226 			delay_us = 10000;
227 			break;
228 		default:
229 			delay_us = cppc_get_transition_latency(cpu) / NSEC_PER_USEC;
230 			break;
231 		}
232 		break;
233 	default:
234 		delay_us = cppc_get_transition_latency(cpu) / NSEC_PER_USEC;
235 		break;
236 	}
237 
238 	return delay_us;
239 }
240 
241 #else
242 
243 static unsigned int cppc_cpufreq_get_transition_delay_us(unsigned int cpu)
244 {
245 	return cppc_get_transition_latency(cpu) / NSEC_PER_USEC;
246 }
247 #endif
248 
249 
250 static struct cppc_cpudata *cppc_cpufreq_get_cpu_data(unsigned int cpu)
251 {
252 	struct cppc_cpudata *cpu_data;
253 	int ret;
254 
255 	cpu_data = kzalloc(sizeof(struct cppc_cpudata), GFP_KERNEL);
256 	if (!cpu_data)
257 		goto out;
258 
259 	if (!zalloc_cpumask_var(&cpu_data->shared_cpu_map, GFP_KERNEL))
260 		goto free_cpu;
261 
262 	ret = acpi_get_psd_map(cpu, cpu_data);
263 	if (ret) {
264 		pr_debug("Err parsing CPU%d PSD data: ret:%d\n", cpu, ret);
265 		goto free_mask;
266 	}
267 
268 	ret = cppc_get_perf_caps(cpu, &cpu_data->perf_caps);
269 	if (ret) {
270 		pr_debug("Err reading CPU%d perf caps: ret:%d\n", cpu, ret);
271 		goto free_mask;
272 	}
273 
274 	/* Convert the lowest and nominal freq from MHz to KHz */
275 	cpu_data->perf_caps.lowest_freq *= 1000;
276 	cpu_data->perf_caps.nominal_freq *= 1000;
277 
278 	list_add(&cpu_data->node, &cpu_data_list);
279 
280 	return cpu_data;
281 
282 free_mask:
283 	free_cpumask_var(cpu_data->shared_cpu_map);
284 free_cpu:
285 	kfree(cpu_data);
286 out:
287 	return NULL;
288 }
289 
290 static int cppc_cpufreq_cpu_init(struct cpufreq_policy *policy)
291 {
292 	unsigned int cpu = policy->cpu;
293 	struct cppc_cpudata *cpu_data;
294 	struct cppc_perf_caps *caps;
295 	int ret;
296 
297 	cpu_data = cppc_cpufreq_get_cpu_data(cpu);
298 	if (!cpu_data) {
299 		pr_err("Error in acquiring _CPC/_PSD data for CPU%d.\n", cpu);
300 		return -ENODEV;
301 	}
302 	caps = &cpu_data->perf_caps;
303 	policy->driver_data = cpu_data;
304 
305 	/*
306 	 * Set min to lowest nonlinear perf to avoid any efficiency penalty (see
307 	 * Section 8.4.7.1.1.5 of ACPI 6.1 spec)
308 	 */
309 	policy->min = cppc_cpufreq_perf_to_khz(cpu_data,
310 					       caps->lowest_nonlinear_perf);
311 	policy->max = cppc_cpufreq_perf_to_khz(cpu_data,
312 					       caps->nominal_perf);
313 
314 	/*
315 	 * Set cpuinfo.min_freq to Lowest to make the full range of performance
316 	 * available if userspace wants to use any perf between lowest & lowest
317 	 * nonlinear perf
318 	 */
319 	policy->cpuinfo.min_freq = cppc_cpufreq_perf_to_khz(cpu_data,
320 							    caps->lowest_perf);
321 	policy->cpuinfo.max_freq = cppc_cpufreq_perf_to_khz(cpu_data,
322 							    caps->nominal_perf);
323 
324 	policy->transition_delay_us = cppc_cpufreq_get_transition_delay_us(cpu);
325 	policy->shared_type = cpu_data->shared_type;
326 
327 	switch (policy->shared_type) {
328 	case CPUFREQ_SHARED_TYPE_HW:
329 	case CPUFREQ_SHARED_TYPE_NONE:
330 		/* Nothing to be done - we'll have a policy for each CPU */
331 		break;
332 	case CPUFREQ_SHARED_TYPE_ANY:
333 		/*
334 		 * All CPUs in the domain will share a policy and all cpufreq
335 		 * operations will use a single cppc_cpudata structure stored
336 		 * in policy->driver_data.
337 		 */
338 		cpumask_copy(policy->cpus, cpu_data->shared_cpu_map);
339 		break;
340 	default:
341 		pr_debug("Unsupported CPU co-ord type: %d\n",
342 			 policy->shared_type);
343 		return -EFAULT;
344 	}
345 
346 	/*
347 	 * If 'highest_perf' is greater than 'nominal_perf', we assume CPU Boost
348 	 * is supported.
349 	 */
350 	if (caps->highest_perf > caps->nominal_perf)
351 		boost_supported = true;
352 
353 	/* Set policy->cur to max now. The governors will adjust later. */
354 	policy->cur = cppc_cpufreq_perf_to_khz(cpu_data, caps->highest_perf);
355 	cpu_data->perf_ctrls.desired_perf =  caps->highest_perf;
356 
357 	ret = cppc_set_perf(cpu, &cpu_data->perf_ctrls);
358 	if (ret)
359 		pr_debug("Err setting perf value:%d on CPU:%d. ret:%d\n",
360 			 caps->highest_perf, cpu, ret);
361 
362 	return ret;
363 }
364 
365 static inline u64 get_delta(u64 t1, u64 t0)
366 {
367 	if (t1 > t0 || t0 > ~(u32)0)
368 		return t1 - t0;
369 
370 	return (u32)t1 - (u32)t0;
371 }
372 
373 static int cppc_get_rate_from_fbctrs(struct cppc_cpudata *cpu_data,
374 				     struct cppc_perf_fb_ctrs fb_ctrs_t0,
375 				     struct cppc_perf_fb_ctrs fb_ctrs_t1)
376 {
377 	u64 delta_reference, delta_delivered;
378 	u64 reference_perf, delivered_perf;
379 
380 	reference_perf = fb_ctrs_t0.reference_perf;
381 
382 	delta_reference = get_delta(fb_ctrs_t1.reference,
383 				    fb_ctrs_t0.reference);
384 	delta_delivered = get_delta(fb_ctrs_t1.delivered,
385 				    fb_ctrs_t0.delivered);
386 
387 	/* Check to avoid divide-by zero */
388 	if (delta_reference || delta_delivered)
389 		delivered_perf = (reference_perf * delta_delivered) /
390 					delta_reference;
391 	else
392 		delivered_perf = cpu_data->perf_ctrls.desired_perf;
393 
394 	return cppc_cpufreq_perf_to_khz(cpu_data, delivered_perf);
395 }
396 
397 static unsigned int cppc_cpufreq_get_rate(unsigned int cpu)
398 {
399 	struct cppc_perf_fb_ctrs fb_ctrs_t0 = {0}, fb_ctrs_t1 = {0};
400 	struct cpufreq_policy *policy = cpufreq_cpu_get(cpu);
401 	struct cppc_cpudata *cpu_data = policy->driver_data;
402 	int ret;
403 
404 	cpufreq_cpu_put(policy);
405 
406 	ret = cppc_get_perf_ctrs(cpu, &fb_ctrs_t0);
407 	if (ret)
408 		return ret;
409 
410 	udelay(2); /* 2usec delay between sampling */
411 
412 	ret = cppc_get_perf_ctrs(cpu, &fb_ctrs_t1);
413 	if (ret)
414 		return ret;
415 
416 	return cppc_get_rate_from_fbctrs(cpu_data, fb_ctrs_t0, fb_ctrs_t1);
417 }
418 
419 static int cppc_cpufreq_set_boost(struct cpufreq_policy *policy, int state)
420 {
421 	struct cppc_cpudata *cpu_data = policy->driver_data;
422 	struct cppc_perf_caps *caps = &cpu_data->perf_caps;
423 	int ret;
424 
425 	if (!boost_supported) {
426 		pr_err("BOOST not supported by CPU or firmware\n");
427 		return -EINVAL;
428 	}
429 
430 	if (state)
431 		policy->max = cppc_cpufreq_perf_to_khz(cpu_data,
432 						       caps->highest_perf);
433 	else
434 		policy->max = cppc_cpufreq_perf_to_khz(cpu_data,
435 						       caps->nominal_perf);
436 	policy->cpuinfo.max_freq = policy->max;
437 
438 	ret = freq_qos_update_request(policy->max_freq_req, policy->max);
439 	if (ret < 0)
440 		return ret;
441 
442 	return 0;
443 }
444 
445 static ssize_t show_freqdomain_cpus(struct cpufreq_policy *policy, char *buf)
446 {
447 	struct cppc_cpudata *cpu_data = policy->driver_data;
448 
449 	return cpufreq_show_cpus(cpu_data->shared_cpu_map, buf);
450 }
451 cpufreq_freq_attr_ro(freqdomain_cpus);
452 
453 static struct freq_attr *cppc_cpufreq_attr[] = {
454 	&freqdomain_cpus,
455 	NULL,
456 };
457 
458 static struct cpufreq_driver cppc_cpufreq_driver = {
459 	.flags = CPUFREQ_CONST_LOOPS,
460 	.verify = cppc_verify_policy,
461 	.target = cppc_cpufreq_set_target,
462 	.get = cppc_cpufreq_get_rate,
463 	.init = cppc_cpufreq_cpu_init,
464 	.stop_cpu = cppc_cpufreq_stop_cpu,
465 	.set_boost = cppc_cpufreq_set_boost,
466 	.attr = cppc_cpufreq_attr,
467 	.name = "cppc_cpufreq",
468 };
469 
470 /*
471  * HISI platform does not support delivered performance counter and
472  * reference performance counter. It can calculate the performance using the
473  * platform specific mechanism. We reuse the desired performance register to
474  * store the real performance calculated by the platform.
475  */
476 static unsigned int hisi_cppc_cpufreq_get_rate(unsigned int cpu)
477 {
478 	struct cpufreq_policy *policy = cpufreq_cpu_get(cpu);
479 	struct cppc_cpudata *cpu_data = policy->driver_data;
480 	u64 desired_perf;
481 	int ret;
482 
483 	cpufreq_cpu_put(policy);
484 
485 	ret = cppc_get_desired_perf(cpu, &desired_perf);
486 	if (ret < 0)
487 		return -EIO;
488 
489 	return cppc_cpufreq_perf_to_khz(cpu_data, desired_perf);
490 }
491 
492 static void cppc_check_hisi_workaround(void)
493 {
494 	struct acpi_table_header *tbl;
495 	acpi_status status = AE_OK;
496 	int i;
497 
498 	status = acpi_get_table(ACPI_SIG_PCCT, 0, &tbl);
499 	if (ACPI_FAILURE(status) || !tbl)
500 		return;
501 
502 	for (i = 0; i < ARRAY_SIZE(wa_info); i++) {
503 		if (!memcmp(wa_info[i].oem_id, tbl->oem_id, ACPI_OEM_ID_SIZE) &&
504 		    !memcmp(wa_info[i].oem_table_id, tbl->oem_table_id, ACPI_OEM_TABLE_ID_SIZE) &&
505 		    wa_info[i].oem_revision == tbl->oem_revision) {
506 			/* Overwrite the get() callback */
507 			cppc_cpufreq_driver.get = hisi_cppc_cpufreq_get_rate;
508 			break;
509 		}
510 	}
511 
512 	acpi_put_table(tbl);
513 }
514 
515 static int __init cppc_cpufreq_init(void)
516 {
517 	if ((acpi_disabled) || !acpi_cpc_valid())
518 		return -ENODEV;
519 
520 	INIT_LIST_HEAD(&cpu_data_list);
521 
522 	cppc_check_hisi_workaround();
523 
524 	return cpufreq_register_driver(&cppc_cpufreq_driver);
525 }
526 
527 static inline void free_cpu_data(void)
528 {
529 	struct cppc_cpudata *iter, *tmp;
530 
531 	list_for_each_entry_safe(iter, tmp, &cpu_data_list, node) {
532 		free_cpumask_var(iter->shared_cpu_map);
533 		list_del(&iter->node);
534 		kfree(iter);
535 	}
536 
537 }
538 
539 static void __exit cppc_cpufreq_exit(void)
540 {
541 	cpufreq_unregister_driver(&cppc_cpufreq_driver);
542 
543 	free_cpu_data();
544 }
545 
546 module_exit(cppc_cpufreq_exit);
547 MODULE_AUTHOR("Ashwin Chaugule");
548 MODULE_DESCRIPTION("CPUFreq driver based on the ACPI CPPC v5.0+ spec");
549 MODULE_LICENSE("GPL");
550 
551 late_initcall(cppc_cpufreq_init);
552 
553 static const struct acpi_device_id cppc_acpi_ids[] __used = {
554 	{ACPI_PROCESSOR_DEVICE_HID, },
555 	{}
556 };
557 
558 MODULE_DEVICE_TABLE(acpi, cppc_acpi_ids);
559