xref: /linux/drivers/cpufreq/cppc_cpufreq.c (revision 48dea9a700c8728cc31a1dd44588b97578de86ee)
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 /* Offest in the DMI processor structure for the max frequency */
30 #define DMI_PROCESSOR_MAX_SPEED  0x14
31 
32 /*
33  * These structs contain information parsed from per CPU
34  * ACPI _CPC structures.
35  * e.g. For each CPU the highest, lowest supported
36  * performance capabilities, desired performance level
37  * requested etc.
38  */
39 static struct cppc_cpudata **all_cpu_data;
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,
100 					unsigned int perf)
101 {
102 	static u64 max_khz;
103 	struct cppc_perf_caps *caps = &cpu->perf_caps;
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,
124 					unsigned int freq)
125 {
126 	static u64 max_khz;
127 	struct cppc_perf_caps *caps = &cpu->perf_caps;
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 	struct cppc_cpudata *cpu;
153 	struct cpufreq_freqs freqs;
154 	u32 desired_perf;
155 	int ret = 0;
156 
157 	cpu = all_cpu_data[policy->cpu];
158 
159 	desired_perf = cppc_cpufreq_khz_to_perf(cpu, target_freq);
160 	/* Return if it is exactly the same perf */
161 	if (desired_perf == cpu->perf_ctrls.desired_perf)
162 		return ret;
163 
164 	cpu->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, &cpu->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->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 	int cpu_num = policy->cpu;
188 	struct cppc_cpudata *cpu = all_cpu_data[cpu_num];
189 	int ret;
190 
191 	cpu->perf_ctrls.desired_perf = cpu->perf_caps.lowest_perf;
192 
193 	ret = cppc_set_perf(cpu_num, &cpu->perf_ctrls);
194 	if (ret)
195 		pr_debug("Err setting perf value:%d on CPU:%d. ret:%d\n",
196 				cpu->perf_caps.lowest_perf, cpu_num, ret);
197 }
198 
199 /*
200  * The PCC subspace describes the rate at which platform can accept commands
201  * on the shared PCC channel (including READs which do not count towards freq
202  * trasition requests), so ideally we need to use the PCC values as a fallback
203  * if we don't have a platform specific transition_delay_us
204  */
205 #ifdef CONFIG_ARM64
206 #include <asm/cputype.h>
207 
208 static unsigned int cppc_cpufreq_get_transition_delay_us(int cpu)
209 {
210 	unsigned long implementor = read_cpuid_implementor();
211 	unsigned long part_num = read_cpuid_part_number();
212 	unsigned int delay_us = 0;
213 
214 	switch (implementor) {
215 	case ARM_CPU_IMP_QCOM:
216 		switch (part_num) {
217 		case QCOM_CPU_PART_FALKOR_V1:
218 		case QCOM_CPU_PART_FALKOR:
219 			delay_us = 10000;
220 			break;
221 		default:
222 			delay_us = cppc_get_transition_latency(cpu) / NSEC_PER_USEC;
223 			break;
224 		}
225 		break;
226 	default:
227 		delay_us = cppc_get_transition_latency(cpu) / NSEC_PER_USEC;
228 		break;
229 	}
230 
231 	return delay_us;
232 }
233 
234 #else
235 
236 static unsigned int cppc_cpufreq_get_transition_delay_us(int cpu)
237 {
238 	return cppc_get_transition_latency(cpu) / NSEC_PER_USEC;
239 }
240 #endif
241 
242 static int cppc_cpufreq_cpu_init(struct cpufreq_policy *policy)
243 {
244 	struct cppc_cpudata *cpu;
245 	unsigned int cpu_num = policy->cpu;
246 	int ret = 0;
247 
248 	cpu = all_cpu_data[policy->cpu];
249 
250 	cpu->cpu = cpu_num;
251 	ret = cppc_get_perf_caps(policy->cpu, &cpu->perf_caps);
252 
253 	if (ret) {
254 		pr_debug("Err reading CPU%d perf capabilities. ret:%d\n",
255 				cpu_num, ret);
256 		return ret;
257 	}
258 
259 	/* Convert the lowest and nominal freq from MHz to KHz */
260 	cpu->perf_caps.lowest_freq *= 1000;
261 	cpu->perf_caps.nominal_freq *= 1000;
262 
263 	/*
264 	 * Set min to lowest nonlinear perf to avoid any efficiency penalty (see
265 	 * Section 8.4.7.1.1.5 of ACPI 6.1 spec)
266 	 */
267 	policy->min = cppc_cpufreq_perf_to_khz(cpu, cpu->perf_caps.lowest_nonlinear_perf);
268 	policy->max = cppc_cpufreq_perf_to_khz(cpu, cpu->perf_caps.nominal_perf);
269 
270 	/*
271 	 * Set cpuinfo.min_freq to Lowest to make the full range of performance
272 	 * available if userspace wants to use any perf between lowest & lowest
273 	 * nonlinear perf
274 	 */
275 	policy->cpuinfo.min_freq = cppc_cpufreq_perf_to_khz(cpu, cpu->perf_caps.lowest_perf);
276 	policy->cpuinfo.max_freq = cppc_cpufreq_perf_to_khz(cpu, cpu->perf_caps.nominal_perf);
277 
278 	policy->transition_delay_us = cppc_cpufreq_get_transition_delay_us(cpu_num);
279 	policy->shared_type = cpu->shared_type;
280 
281 	if (policy->shared_type == CPUFREQ_SHARED_TYPE_ANY) {
282 		int i;
283 
284 		cpumask_copy(policy->cpus, cpu->shared_cpu_map);
285 
286 		for_each_cpu(i, policy->cpus) {
287 			if (unlikely(i == policy->cpu))
288 				continue;
289 
290 			memcpy(&all_cpu_data[i]->perf_caps, &cpu->perf_caps,
291 			       sizeof(cpu->perf_caps));
292 		}
293 	} else if (policy->shared_type == CPUFREQ_SHARED_TYPE_ALL) {
294 		/* Support only SW_ANY for now. */
295 		pr_debug("Unsupported CPU co-ord type\n");
296 		return -EFAULT;
297 	}
298 
299 	cpu->cur_policy = policy;
300 
301 	/*
302 	 * If 'highest_perf' is greater than 'nominal_perf', we assume CPU Boost
303 	 * is supported.
304 	 */
305 	if (cpu->perf_caps.highest_perf > cpu->perf_caps.nominal_perf)
306 		boost_supported = true;
307 
308 	/* Set policy->cur to max now. The governors will adjust later. */
309 	policy->cur = cppc_cpufreq_perf_to_khz(cpu,
310 					cpu->perf_caps.highest_perf);
311 	cpu->perf_ctrls.desired_perf = cpu->perf_caps.highest_perf;
312 
313 	ret = cppc_set_perf(cpu_num, &cpu->perf_ctrls);
314 	if (ret)
315 		pr_debug("Err setting perf value:%d on CPU:%d. ret:%d\n",
316 				cpu->perf_caps.highest_perf, cpu_num, ret);
317 
318 	return ret;
319 }
320 
321 static inline u64 get_delta(u64 t1, u64 t0)
322 {
323 	if (t1 > t0 || t0 > ~(u32)0)
324 		return t1 - t0;
325 
326 	return (u32)t1 - (u32)t0;
327 }
328 
329 static int cppc_get_rate_from_fbctrs(struct cppc_cpudata *cpu,
330 				     struct cppc_perf_fb_ctrs fb_ctrs_t0,
331 				     struct cppc_perf_fb_ctrs fb_ctrs_t1)
332 {
333 	u64 delta_reference, delta_delivered;
334 	u64 reference_perf, delivered_perf;
335 
336 	reference_perf = fb_ctrs_t0.reference_perf;
337 
338 	delta_reference = get_delta(fb_ctrs_t1.reference,
339 				    fb_ctrs_t0.reference);
340 	delta_delivered = get_delta(fb_ctrs_t1.delivered,
341 				    fb_ctrs_t0.delivered);
342 
343 	/* Check to avoid divide-by zero */
344 	if (delta_reference || delta_delivered)
345 		delivered_perf = (reference_perf * delta_delivered) /
346 					delta_reference;
347 	else
348 		delivered_perf = cpu->perf_ctrls.desired_perf;
349 
350 	return cppc_cpufreq_perf_to_khz(cpu, delivered_perf);
351 }
352 
353 static unsigned int cppc_cpufreq_get_rate(unsigned int cpunum)
354 {
355 	struct cppc_perf_fb_ctrs fb_ctrs_t0 = {0}, fb_ctrs_t1 = {0};
356 	struct cppc_cpudata *cpu = all_cpu_data[cpunum];
357 	int ret;
358 
359 	ret = cppc_get_perf_ctrs(cpunum, &fb_ctrs_t0);
360 	if (ret)
361 		return ret;
362 
363 	udelay(2); /* 2usec delay between sampling */
364 
365 	ret = cppc_get_perf_ctrs(cpunum, &fb_ctrs_t1);
366 	if (ret)
367 		return ret;
368 
369 	return cppc_get_rate_from_fbctrs(cpu, fb_ctrs_t0, fb_ctrs_t1);
370 }
371 
372 static int cppc_cpufreq_set_boost(struct cpufreq_policy *policy, int state)
373 {
374 	struct cppc_cpudata *cpudata;
375 	int ret;
376 
377 	if (!boost_supported) {
378 		pr_err("BOOST not supported by CPU or firmware\n");
379 		return -EINVAL;
380 	}
381 
382 	cpudata = all_cpu_data[policy->cpu];
383 	if (state)
384 		policy->max = cppc_cpufreq_perf_to_khz(cpudata,
385 					cpudata->perf_caps.highest_perf);
386 	else
387 		policy->max = cppc_cpufreq_perf_to_khz(cpudata,
388 					cpudata->perf_caps.nominal_perf);
389 	policy->cpuinfo.max_freq = policy->max;
390 
391 	ret = freq_qos_update_request(policy->max_freq_req, policy->max);
392 	if (ret < 0)
393 		return ret;
394 
395 	return 0;
396 }
397 
398 static struct cpufreq_driver cppc_cpufreq_driver = {
399 	.flags = CPUFREQ_CONST_LOOPS,
400 	.verify = cppc_verify_policy,
401 	.target = cppc_cpufreq_set_target,
402 	.get = cppc_cpufreq_get_rate,
403 	.init = cppc_cpufreq_cpu_init,
404 	.stop_cpu = cppc_cpufreq_stop_cpu,
405 	.set_boost = cppc_cpufreq_set_boost,
406 	.name = "cppc_cpufreq",
407 };
408 
409 /*
410  * HISI platform does not support delivered performance counter and
411  * reference performance counter. It can calculate the performance using the
412  * platform specific mechanism. We reuse the desired performance register to
413  * store the real performance calculated by the platform.
414  */
415 static unsigned int hisi_cppc_cpufreq_get_rate(unsigned int cpunum)
416 {
417 	struct cppc_cpudata *cpudata = all_cpu_data[cpunum];
418 	u64 desired_perf;
419 	int ret;
420 
421 	ret = cppc_get_desired_perf(cpunum, &desired_perf);
422 	if (ret < 0)
423 		return -EIO;
424 
425 	return cppc_cpufreq_perf_to_khz(cpudata, desired_perf);
426 }
427 
428 static void cppc_check_hisi_workaround(void)
429 {
430 	struct acpi_table_header *tbl;
431 	acpi_status status = AE_OK;
432 	int i;
433 
434 	status = acpi_get_table(ACPI_SIG_PCCT, 0, &tbl);
435 	if (ACPI_FAILURE(status) || !tbl)
436 		return;
437 
438 	for (i = 0; i < ARRAY_SIZE(wa_info); i++) {
439 		if (!memcmp(wa_info[i].oem_id, tbl->oem_id, ACPI_OEM_ID_SIZE) &&
440 		    !memcmp(wa_info[i].oem_table_id, tbl->oem_table_id, ACPI_OEM_TABLE_ID_SIZE) &&
441 		    wa_info[i].oem_revision == tbl->oem_revision) {
442 			/* Overwrite the get() callback */
443 			cppc_cpufreq_driver.get = hisi_cppc_cpufreq_get_rate;
444 			break;
445 		}
446 	}
447 
448 	acpi_put_table(tbl);
449 }
450 
451 static int __init cppc_cpufreq_init(void)
452 {
453 	int i, ret = 0;
454 	struct cppc_cpudata *cpu;
455 
456 	if (acpi_disabled)
457 		return -ENODEV;
458 
459 	all_cpu_data = kcalloc(num_possible_cpus(), sizeof(void *),
460 			       GFP_KERNEL);
461 	if (!all_cpu_data)
462 		return -ENOMEM;
463 
464 	for_each_possible_cpu(i) {
465 		all_cpu_data[i] = kzalloc(sizeof(struct cppc_cpudata), GFP_KERNEL);
466 		if (!all_cpu_data[i])
467 			goto out;
468 
469 		cpu = all_cpu_data[i];
470 		if (!zalloc_cpumask_var(&cpu->shared_cpu_map, GFP_KERNEL))
471 			goto out;
472 	}
473 
474 	ret = acpi_get_psd_map(all_cpu_data);
475 	if (ret) {
476 		pr_debug("Error parsing PSD data. Aborting cpufreq registration.\n");
477 		goto out;
478 	}
479 
480 	cppc_check_hisi_workaround();
481 
482 	ret = cpufreq_register_driver(&cppc_cpufreq_driver);
483 	if (ret)
484 		goto out;
485 
486 	return ret;
487 
488 out:
489 	for_each_possible_cpu(i) {
490 		cpu = all_cpu_data[i];
491 		if (!cpu)
492 			break;
493 		free_cpumask_var(cpu->shared_cpu_map);
494 		kfree(cpu);
495 	}
496 
497 	kfree(all_cpu_data);
498 	return -ENODEV;
499 }
500 
501 static void __exit cppc_cpufreq_exit(void)
502 {
503 	struct cppc_cpudata *cpu;
504 	int i;
505 
506 	cpufreq_unregister_driver(&cppc_cpufreq_driver);
507 
508 	for_each_possible_cpu(i) {
509 		cpu = all_cpu_data[i];
510 		free_cpumask_var(cpu->shared_cpu_map);
511 		kfree(cpu);
512 	}
513 
514 	kfree(all_cpu_data);
515 }
516 
517 module_exit(cppc_cpufreq_exit);
518 MODULE_AUTHOR("Ashwin Chaugule");
519 MODULE_DESCRIPTION("CPUFreq driver based on the ACPI CPPC v5.0+ spec");
520 MODULE_LICENSE("GPL");
521 
522 late_initcall(cppc_cpufreq_init);
523 
524 static const struct acpi_device_id cppc_acpi_ids[] __used = {
525 	{ACPI_PROCESSOR_DEVICE_HID, },
526 	{}
527 };
528 
529 MODULE_DEVICE_TABLE(acpi, cppc_acpi_ids);
530