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/irq_work.h>
20 #include <linux/kthread.h>
21 #include <linux/time.h>
22 #include <linux/vmalloc.h>
23 #include <uapi/linux/sched/types.h>
24
25 #include <linux/unaligned.h>
26
27 #include <acpi/cppc_acpi.h>
28
29 /*
30 * This list contains information parsed from per CPU ACPI _CPC and _PSD
31 * structures: e.g. the highest and lowest supported performance, capabilities,
32 * desired performance, level requested etc. Depending on the share_type, not
33 * all CPUs will have an entry in the list.
34 */
35 static LIST_HEAD(cpu_data_list);
36
37 static struct cpufreq_driver cppc_cpufreq_driver;
38
39 #ifdef CONFIG_ACPI_CPPC_CPUFREQ_FIE
40 static enum {
41 FIE_UNSET = -1,
42 FIE_ENABLED,
43 FIE_DISABLED
44 } fie_disabled = FIE_UNSET;
45
46 module_param(fie_disabled, int, 0444);
47 MODULE_PARM_DESC(fie_disabled, "Disable Frequency Invariance Engine (FIE)");
48
49 /* Frequency invariance support */
50 struct cppc_freq_invariance {
51 int cpu;
52 struct irq_work irq_work;
53 struct kthread_work work;
54 struct cppc_perf_fb_ctrs prev_perf_fb_ctrs;
55 struct cppc_cpudata *cpu_data;
56 };
57
58 static DEFINE_PER_CPU(struct cppc_freq_invariance, cppc_freq_inv);
59 static struct kthread_worker *kworker_fie;
60
61 static int cppc_perf_from_fbctrs(struct cppc_cpudata *cpu_data,
62 struct cppc_perf_fb_ctrs *fb_ctrs_t0,
63 struct cppc_perf_fb_ctrs *fb_ctrs_t1);
64
65 /**
66 * cppc_scale_freq_workfn - CPPC arch_freq_scale updater for frequency invariance
67 * @work: The work item.
68 *
69 * The CPPC driver register itself with the topology core to provide its own
70 * implementation (cppc_scale_freq_tick()) of topology_scale_freq_tick() which
71 * gets called by the scheduler on every tick.
72 *
73 * Note that the arch specific counters have higher priority than CPPC counters,
74 * if available, though the CPPC driver doesn't need to have any special
75 * handling for that.
76 *
77 * On an invocation of cppc_scale_freq_tick(), we schedule an irq work (since we
78 * reach here from hard-irq context), which then schedules a normal work item
79 * and cppc_scale_freq_workfn() updates the per_cpu arch_freq_scale variable
80 * based on the counter updates since the last tick.
81 */
cppc_scale_freq_workfn(struct kthread_work * work)82 static void cppc_scale_freq_workfn(struct kthread_work *work)
83 {
84 struct cppc_freq_invariance *cppc_fi;
85 struct cppc_perf_fb_ctrs fb_ctrs = {0};
86 struct cppc_cpudata *cpu_data;
87 unsigned long local_freq_scale;
88 u64 perf;
89
90 cppc_fi = container_of(work, struct cppc_freq_invariance, work);
91 cpu_data = cppc_fi->cpu_data;
92
93 if (cppc_get_perf_ctrs(cppc_fi->cpu, &fb_ctrs)) {
94 pr_warn("%s: failed to read perf counters\n", __func__);
95 return;
96 }
97
98 perf = cppc_perf_from_fbctrs(cpu_data, &cppc_fi->prev_perf_fb_ctrs,
99 &fb_ctrs);
100 if (!perf)
101 return;
102
103 cppc_fi->prev_perf_fb_ctrs = fb_ctrs;
104
105 perf <<= SCHED_CAPACITY_SHIFT;
106 local_freq_scale = div64_u64(perf, cpu_data->perf_caps.highest_perf);
107
108 /* This can happen due to counter's overflow */
109 if (unlikely(local_freq_scale > 1024))
110 local_freq_scale = 1024;
111
112 per_cpu(arch_freq_scale, cppc_fi->cpu) = local_freq_scale;
113 }
114
cppc_irq_work(struct irq_work * irq_work)115 static void cppc_irq_work(struct irq_work *irq_work)
116 {
117 struct cppc_freq_invariance *cppc_fi;
118
119 cppc_fi = container_of(irq_work, struct cppc_freq_invariance, irq_work);
120 kthread_queue_work(kworker_fie, &cppc_fi->work);
121 }
122
cppc_scale_freq_tick(void)123 static void cppc_scale_freq_tick(void)
124 {
125 struct cppc_freq_invariance *cppc_fi = &per_cpu(cppc_freq_inv, smp_processor_id());
126
127 /*
128 * cppc_get_perf_ctrs() can potentially sleep, call that from the right
129 * context.
130 */
131 irq_work_queue(&cppc_fi->irq_work);
132 }
133
134 static struct scale_freq_data cppc_sftd = {
135 .source = SCALE_FREQ_SOURCE_CPPC,
136 .set_freq_scale = cppc_scale_freq_tick,
137 };
138
cppc_cpufreq_cpu_fie_init(struct cpufreq_policy * policy)139 static void cppc_cpufreq_cpu_fie_init(struct cpufreq_policy *policy)
140 {
141 struct cppc_freq_invariance *cppc_fi;
142 int cpu, ret;
143
144 if (fie_disabled)
145 return;
146
147 for_each_cpu(cpu, policy->cpus) {
148 cppc_fi = &per_cpu(cppc_freq_inv, cpu);
149 cppc_fi->cpu = cpu;
150 cppc_fi->cpu_data = policy->driver_data;
151 kthread_init_work(&cppc_fi->work, cppc_scale_freq_workfn);
152 init_irq_work(&cppc_fi->irq_work, cppc_irq_work);
153
154 ret = cppc_get_perf_ctrs(cpu, &cppc_fi->prev_perf_fb_ctrs);
155 if (ret) {
156 pr_warn("%s: failed to read perf counters for cpu:%d: %d\n",
157 __func__, cpu, ret);
158
159 /*
160 * Don't abort if the CPU was offline while the driver
161 * was getting registered.
162 */
163 if (cpu_online(cpu))
164 return;
165 }
166 }
167
168 /* Register for freq-invariance */
169 topology_set_scale_freq_source(&cppc_sftd, policy->cpus);
170 }
171
172 /*
173 * We free all the resources on policy's removal and not on CPU removal as the
174 * irq-work are per-cpu and the hotplug core takes care of flushing the pending
175 * irq-works (hint: smpcfd_dying_cpu()) on CPU hotplug. Even if the kthread-work
176 * fires on another CPU after the concerned CPU is removed, it won't harm.
177 *
178 * We just need to make sure to remove them all on policy->exit().
179 */
cppc_cpufreq_cpu_fie_exit(struct cpufreq_policy * policy)180 static void cppc_cpufreq_cpu_fie_exit(struct cpufreq_policy *policy)
181 {
182 struct cppc_freq_invariance *cppc_fi;
183 int cpu;
184
185 if (fie_disabled)
186 return;
187
188 /* policy->cpus will be empty here, use related_cpus instead */
189 topology_clear_scale_freq_source(SCALE_FREQ_SOURCE_CPPC, policy->related_cpus);
190
191 for_each_cpu(cpu, policy->related_cpus) {
192 cppc_fi = &per_cpu(cppc_freq_inv, cpu);
193 irq_work_sync(&cppc_fi->irq_work);
194 kthread_cancel_work_sync(&cppc_fi->work);
195 }
196 }
197
cppc_freq_invariance_init(void)198 static void __init cppc_freq_invariance_init(void)
199 {
200 struct sched_attr attr = {
201 .size = sizeof(struct sched_attr),
202 .sched_policy = SCHED_DEADLINE,
203 .sched_nice = 0,
204 .sched_priority = 0,
205 /*
206 * Fake (unused) bandwidth; workaround to "fix"
207 * priority inheritance.
208 */
209 .sched_runtime = NSEC_PER_MSEC,
210 .sched_deadline = 10 * NSEC_PER_MSEC,
211 .sched_period = 10 * NSEC_PER_MSEC,
212 };
213 int ret;
214
215 if (fie_disabled != FIE_ENABLED && fie_disabled != FIE_DISABLED) {
216 fie_disabled = FIE_ENABLED;
217 if (cppc_perf_ctrs_in_pcc()) {
218 pr_info("FIE not enabled on systems with registers in PCC\n");
219 fie_disabled = FIE_DISABLED;
220 }
221 }
222
223 if (fie_disabled)
224 return;
225
226 kworker_fie = kthread_run_worker(0, "cppc_fie");
227 if (IS_ERR(kworker_fie)) {
228 pr_warn("%s: failed to create kworker_fie: %ld\n", __func__,
229 PTR_ERR(kworker_fie));
230 fie_disabled = FIE_DISABLED;
231 return;
232 }
233
234 ret = sched_setattr_nocheck(kworker_fie->task, &attr);
235 if (ret) {
236 pr_warn("%s: failed to set SCHED_DEADLINE: %d\n", __func__,
237 ret);
238 kthread_destroy_worker(kworker_fie);
239 fie_disabled = FIE_DISABLED;
240 }
241 }
242
cppc_freq_invariance_exit(void)243 static void cppc_freq_invariance_exit(void)
244 {
245 if (fie_disabled)
246 return;
247
248 kthread_destroy_worker(kworker_fie);
249 }
250
251 #else
cppc_cpufreq_cpu_fie_init(struct cpufreq_policy * policy)252 static inline void cppc_cpufreq_cpu_fie_init(struct cpufreq_policy *policy)
253 {
254 }
255
cppc_cpufreq_cpu_fie_exit(struct cpufreq_policy * policy)256 static inline void cppc_cpufreq_cpu_fie_exit(struct cpufreq_policy *policy)
257 {
258 }
259
cppc_freq_invariance_init(void)260 static inline void cppc_freq_invariance_init(void)
261 {
262 }
263
cppc_freq_invariance_exit(void)264 static inline void cppc_freq_invariance_exit(void)
265 {
266 }
267 #endif /* CONFIG_ACPI_CPPC_CPUFREQ_FIE */
268
cppc_cpufreq_set_target(struct cpufreq_policy * policy,unsigned int target_freq,unsigned int relation)269 static int cppc_cpufreq_set_target(struct cpufreq_policy *policy,
270 unsigned int target_freq,
271 unsigned int relation)
272 {
273 struct cppc_cpudata *cpu_data = policy->driver_data;
274 unsigned int cpu = policy->cpu;
275 struct cpufreq_freqs freqs;
276 int ret = 0;
277
278 cpu_data->perf_ctrls.desired_perf =
279 cppc_khz_to_perf(&cpu_data->perf_caps, target_freq);
280 freqs.old = policy->cur;
281 freqs.new = target_freq;
282
283 cpufreq_freq_transition_begin(policy, &freqs);
284 ret = cppc_set_perf(cpu, &cpu_data->perf_ctrls);
285 cpufreq_freq_transition_end(policy, &freqs, ret != 0);
286
287 if (ret)
288 pr_debug("Failed to set target on CPU:%d. ret:%d\n",
289 cpu, ret);
290
291 return ret;
292 }
293
cppc_cpufreq_fast_switch(struct cpufreq_policy * policy,unsigned int target_freq)294 static unsigned int cppc_cpufreq_fast_switch(struct cpufreq_policy *policy,
295 unsigned int target_freq)
296 {
297 struct cppc_cpudata *cpu_data = policy->driver_data;
298 unsigned int cpu = policy->cpu;
299 u32 desired_perf;
300 int ret;
301
302 desired_perf = cppc_khz_to_perf(&cpu_data->perf_caps, target_freq);
303 cpu_data->perf_ctrls.desired_perf = desired_perf;
304 ret = cppc_set_perf(cpu, &cpu_data->perf_ctrls);
305
306 if (ret) {
307 pr_debug("Failed to set target on CPU:%d. ret:%d\n",
308 cpu, ret);
309 return 0;
310 }
311
312 return target_freq;
313 }
314
cppc_verify_policy(struct cpufreq_policy_data * policy)315 static int cppc_verify_policy(struct cpufreq_policy_data *policy)
316 {
317 cpufreq_verify_within_cpu_limits(policy);
318 return 0;
319 }
320
321 /*
322 * The PCC subspace describes the rate at which platform can accept commands
323 * on the shared PCC channel (including READs which do not count towards freq
324 * transition requests), so ideally we need to use the PCC values as a fallback
325 * if we don't have a platform specific transition_delay_us
326 */
327 #ifdef CONFIG_ARM64
328 #include <asm/cputype.h>
329
cppc_cpufreq_get_transition_delay_us(unsigned int cpu)330 static unsigned int cppc_cpufreq_get_transition_delay_us(unsigned int cpu)
331 {
332 unsigned long implementor = read_cpuid_implementor();
333 unsigned long part_num = read_cpuid_part_number();
334
335 switch (implementor) {
336 case ARM_CPU_IMP_QCOM:
337 switch (part_num) {
338 case QCOM_CPU_PART_FALKOR_V1:
339 case QCOM_CPU_PART_FALKOR:
340 return 10000;
341 }
342 }
343 return cppc_get_transition_latency(cpu) / NSEC_PER_USEC;
344 }
345 #else
cppc_cpufreq_get_transition_delay_us(unsigned int cpu)346 static unsigned int cppc_cpufreq_get_transition_delay_us(unsigned int cpu)
347 {
348 return cppc_get_transition_latency(cpu) / NSEC_PER_USEC;
349 }
350 #endif
351
352 #if defined(CONFIG_ARM64) && defined(CONFIG_ENERGY_MODEL)
353
354 static DEFINE_PER_CPU(unsigned int, efficiency_class);
355 static void cppc_cpufreq_register_em(struct cpufreq_policy *policy);
356
357 /* Create an artificial performance state every CPPC_EM_CAP_STEP capacity unit. */
358 #define CPPC_EM_CAP_STEP (20)
359 /* Increase the cost value by CPPC_EM_COST_STEP every performance state. */
360 #define CPPC_EM_COST_STEP (1)
361 /* Add a cost gap correspnding to the energy of 4 CPUs. */
362 #define CPPC_EM_COST_GAP (4 * SCHED_CAPACITY_SCALE * CPPC_EM_COST_STEP \
363 / CPPC_EM_CAP_STEP)
364
get_perf_level_count(struct cpufreq_policy * policy)365 static unsigned int get_perf_level_count(struct cpufreq_policy *policy)
366 {
367 struct cppc_perf_caps *perf_caps;
368 unsigned int min_cap, max_cap;
369 struct cppc_cpudata *cpu_data;
370 int cpu = policy->cpu;
371
372 cpu_data = policy->driver_data;
373 perf_caps = &cpu_data->perf_caps;
374 max_cap = arch_scale_cpu_capacity(cpu);
375 min_cap = div_u64((u64)max_cap * perf_caps->lowest_perf,
376 perf_caps->highest_perf);
377 if ((min_cap == 0) || (max_cap < min_cap))
378 return 0;
379 return 1 + max_cap / CPPC_EM_CAP_STEP - min_cap / CPPC_EM_CAP_STEP;
380 }
381
382 /*
383 * The cost is defined as:
384 * cost = power * max_frequency / frequency
385 */
compute_cost(int cpu,int step)386 static inline unsigned long compute_cost(int cpu, int step)
387 {
388 return CPPC_EM_COST_GAP * per_cpu(efficiency_class, cpu) +
389 step * CPPC_EM_COST_STEP;
390 }
391
cppc_get_cpu_power(struct device * cpu_dev,unsigned long * power,unsigned long * KHz)392 static int cppc_get_cpu_power(struct device *cpu_dev,
393 unsigned long *power, unsigned long *KHz)
394 {
395 unsigned long perf_step, perf_prev, perf, perf_check;
396 unsigned int min_step, max_step, step, step_check;
397 unsigned long prev_freq = *KHz;
398 unsigned int min_cap, max_cap;
399 struct cpufreq_policy *policy;
400
401 struct cppc_perf_caps *perf_caps;
402 struct cppc_cpudata *cpu_data;
403
404 policy = cpufreq_cpu_get_raw(cpu_dev->id);
405 if (!policy)
406 return -EINVAL;
407
408 cpu_data = policy->driver_data;
409 perf_caps = &cpu_data->perf_caps;
410 max_cap = arch_scale_cpu_capacity(cpu_dev->id);
411 min_cap = div_u64((u64)max_cap * perf_caps->lowest_perf,
412 perf_caps->highest_perf);
413 perf_step = div_u64((u64)CPPC_EM_CAP_STEP * perf_caps->highest_perf,
414 max_cap);
415 min_step = min_cap / CPPC_EM_CAP_STEP;
416 max_step = max_cap / CPPC_EM_CAP_STEP;
417
418 perf_prev = cppc_khz_to_perf(perf_caps, *KHz);
419 step = perf_prev / perf_step;
420
421 if (step > max_step)
422 return -EINVAL;
423
424 if (min_step == max_step) {
425 step = max_step;
426 perf = perf_caps->highest_perf;
427 } else if (step < min_step) {
428 step = min_step;
429 perf = perf_caps->lowest_perf;
430 } else {
431 step++;
432 if (step == max_step)
433 perf = perf_caps->highest_perf;
434 else
435 perf = step * perf_step;
436 }
437
438 *KHz = cppc_perf_to_khz(perf_caps, perf);
439 perf_check = cppc_khz_to_perf(perf_caps, *KHz);
440 step_check = perf_check / perf_step;
441
442 /*
443 * To avoid bad integer approximation, check that new frequency value
444 * increased and that the new frequency will be converted to the
445 * desired step value.
446 */
447 while ((*KHz == prev_freq) || (step_check != step)) {
448 perf++;
449 *KHz = cppc_perf_to_khz(perf_caps, perf);
450 perf_check = cppc_khz_to_perf(perf_caps, *KHz);
451 step_check = perf_check / perf_step;
452 }
453
454 /*
455 * With an artificial EM, only the cost value is used. Still the power
456 * is populated such as 0 < power < EM_MAX_POWER. This allows to add
457 * more sense to the artificial performance states.
458 */
459 *power = compute_cost(cpu_dev->id, step);
460
461 return 0;
462 }
463
cppc_get_cpu_cost(struct device * cpu_dev,unsigned long KHz,unsigned long * cost)464 static int cppc_get_cpu_cost(struct device *cpu_dev, unsigned long KHz,
465 unsigned long *cost)
466 {
467 unsigned long perf_step, perf_prev;
468 struct cppc_perf_caps *perf_caps;
469 struct cpufreq_policy *policy;
470 struct cppc_cpudata *cpu_data;
471 unsigned int max_cap;
472 int step;
473
474 policy = cpufreq_cpu_get_raw(cpu_dev->id);
475 if (!policy)
476 return -EINVAL;
477
478 cpu_data = policy->driver_data;
479 perf_caps = &cpu_data->perf_caps;
480 max_cap = arch_scale_cpu_capacity(cpu_dev->id);
481
482 perf_prev = cppc_khz_to_perf(perf_caps, KHz);
483 perf_step = CPPC_EM_CAP_STEP * perf_caps->highest_perf / max_cap;
484 step = perf_prev / perf_step;
485
486 *cost = compute_cost(cpu_dev->id, step);
487
488 return 0;
489 }
490
populate_efficiency_class(void)491 static int populate_efficiency_class(void)
492 {
493 struct acpi_madt_generic_interrupt *gicc;
494 DECLARE_BITMAP(used_classes, 256) = {};
495 int class, cpu, index;
496
497 for_each_possible_cpu(cpu) {
498 gicc = acpi_cpu_get_madt_gicc(cpu);
499 class = gicc->efficiency_class;
500 bitmap_set(used_classes, class, 1);
501 }
502
503 if (bitmap_weight(used_classes, 256) <= 1) {
504 pr_debug("Efficiency classes are all equal (=%d). "
505 "No EM registered", class);
506 return -EINVAL;
507 }
508
509 /*
510 * Squeeze efficiency class values on [0:#efficiency_class-1].
511 * Values are per spec in [0:255].
512 */
513 index = 0;
514 for_each_set_bit(class, used_classes, 256) {
515 for_each_possible_cpu(cpu) {
516 gicc = acpi_cpu_get_madt_gicc(cpu);
517 if (gicc->efficiency_class == class)
518 per_cpu(efficiency_class, cpu) = index;
519 }
520 index++;
521 }
522 cppc_cpufreq_driver.register_em = cppc_cpufreq_register_em;
523
524 return 0;
525 }
526
cppc_cpufreq_register_em(struct cpufreq_policy * policy)527 static void cppc_cpufreq_register_em(struct cpufreq_policy *policy)
528 {
529 struct cppc_cpudata *cpu_data;
530 struct em_data_callback em_cb =
531 EM_ADV_DATA_CB(cppc_get_cpu_power, cppc_get_cpu_cost);
532
533 cpu_data = policy->driver_data;
534 em_dev_register_perf_domain(get_cpu_device(policy->cpu),
535 get_perf_level_count(policy), &em_cb,
536 cpu_data->shared_cpu_map, 0);
537 }
538
539 #else
populate_efficiency_class(void)540 static int populate_efficiency_class(void)
541 {
542 return 0;
543 }
544 #endif
545
cppc_cpufreq_get_cpu_data(unsigned int cpu)546 static struct cppc_cpudata *cppc_cpufreq_get_cpu_data(unsigned int cpu)
547 {
548 struct cppc_cpudata *cpu_data;
549 int ret;
550
551 cpu_data = kzalloc(sizeof(struct cppc_cpudata), GFP_KERNEL);
552 if (!cpu_data)
553 goto out;
554
555 if (!zalloc_cpumask_var(&cpu_data->shared_cpu_map, GFP_KERNEL))
556 goto free_cpu;
557
558 ret = acpi_get_psd_map(cpu, cpu_data);
559 if (ret) {
560 pr_debug("Err parsing CPU%d PSD data: ret:%d\n", cpu, ret);
561 goto free_mask;
562 }
563
564 ret = cppc_get_perf_caps(cpu, &cpu_data->perf_caps);
565 if (ret) {
566 pr_debug("Err reading CPU%d perf caps: ret:%d\n", cpu, ret);
567 goto free_mask;
568 }
569
570 list_add(&cpu_data->node, &cpu_data_list);
571
572 return cpu_data;
573
574 free_mask:
575 free_cpumask_var(cpu_data->shared_cpu_map);
576 free_cpu:
577 kfree(cpu_data);
578 out:
579 return NULL;
580 }
581
cppc_cpufreq_put_cpu_data(struct cpufreq_policy * policy)582 static void cppc_cpufreq_put_cpu_data(struct cpufreq_policy *policy)
583 {
584 struct cppc_cpudata *cpu_data = policy->driver_data;
585
586 list_del(&cpu_data->node);
587 free_cpumask_var(cpu_data->shared_cpu_map);
588 kfree(cpu_data);
589 policy->driver_data = NULL;
590 }
591
cppc_cpufreq_cpu_init(struct cpufreq_policy * policy)592 static int cppc_cpufreq_cpu_init(struct cpufreq_policy *policy)
593 {
594 unsigned int cpu = policy->cpu;
595 struct cppc_cpudata *cpu_data;
596 struct cppc_perf_caps *caps;
597 int ret;
598
599 cpu_data = cppc_cpufreq_get_cpu_data(cpu);
600 if (!cpu_data) {
601 pr_err("Error in acquiring _CPC/_PSD data for CPU%d.\n", cpu);
602 return -ENODEV;
603 }
604 caps = &cpu_data->perf_caps;
605 policy->driver_data = cpu_data;
606
607 /*
608 * Set min to lowest nonlinear perf to avoid any efficiency penalty (see
609 * Section 8.4.7.1.1.5 of ACPI 6.1 spec)
610 */
611 policy->min = cppc_perf_to_khz(caps, caps->lowest_nonlinear_perf);
612 policy->max = cppc_perf_to_khz(caps, policy->boost_enabled ?
613 caps->highest_perf : caps->nominal_perf);
614
615 /*
616 * Set cpuinfo.min_freq to Lowest to make the full range of performance
617 * available if userspace wants to use any perf between lowest & lowest
618 * nonlinear perf
619 */
620 policy->cpuinfo.min_freq = cppc_perf_to_khz(caps, caps->lowest_perf);
621 policy->cpuinfo.max_freq = policy->max;
622
623 policy->transition_delay_us = cppc_cpufreq_get_transition_delay_us(cpu);
624 policy->shared_type = cpu_data->shared_type;
625
626 switch (policy->shared_type) {
627 case CPUFREQ_SHARED_TYPE_HW:
628 case CPUFREQ_SHARED_TYPE_NONE:
629 /* Nothing to be done - we'll have a policy for each CPU */
630 break;
631 case CPUFREQ_SHARED_TYPE_ANY:
632 /*
633 * All CPUs in the domain will share a policy and all cpufreq
634 * operations will use a single cppc_cpudata structure stored
635 * in policy->driver_data.
636 */
637 cpumask_copy(policy->cpus, cpu_data->shared_cpu_map);
638 break;
639 default:
640 pr_debug("Unsupported CPU co-ord type: %d\n",
641 policy->shared_type);
642 ret = -EFAULT;
643 goto out;
644 }
645
646 policy->fast_switch_possible = cppc_allow_fast_switch();
647 policy->dvfs_possible_from_any_cpu = true;
648
649 /*
650 * If 'highest_perf' is greater than 'nominal_perf', we assume CPU Boost
651 * is supported.
652 */
653 if (caps->highest_perf > caps->nominal_perf)
654 policy->boost_supported = true;
655
656 /* Set policy->cur to max now. The governors will adjust later. */
657 policy->cur = cppc_perf_to_khz(caps, caps->highest_perf);
658 cpu_data->perf_ctrls.desired_perf = caps->highest_perf;
659
660 ret = cppc_set_perf(cpu, &cpu_data->perf_ctrls);
661 if (ret) {
662 pr_debug("Err setting perf value:%d on CPU:%d. ret:%d\n",
663 caps->highest_perf, cpu, ret);
664 goto out;
665 }
666
667 cppc_cpufreq_cpu_fie_init(policy);
668 return 0;
669
670 out:
671 cppc_cpufreq_put_cpu_data(policy);
672 return ret;
673 }
674
cppc_cpufreq_cpu_exit(struct cpufreq_policy * policy)675 static void cppc_cpufreq_cpu_exit(struct cpufreq_policy *policy)
676 {
677 struct cppc_cpudata *cpu_data = policy->driver_data;
678 struct cppc_perf_caps *caps = &cpu_data->perf_caps;
679 unsigned int cpu = policy->cpu;
680 int ret;
681
682 cppc_cpufreq_cpu_fie_exit(policy);
683
684 cpu_data->perf_ctrls.desired_perf = caps->lowest_perf;
685
686 ret = cppc_set_perf(cpu, &cpu_data->perf_ctrls);
687 if (ret)
688 pr_debug("Err setting perf value:%d on CPU:%d. ret:%d\n",
689 caps->lowest_perf, cpu, ret);
690
691 cppc_cpufreq_put_cpu_data(policy);
692 }
693
get_delta(u64 t1,u64 t0)694 static inline u64 get_delta(u64 t1, u64 t0)
695 {
696 if (t1 > t0 || t0 > ~(u32)0)
697 return t1 - t0;
698
699 return (u32)t1 - (u32)t0;
700 }
701
cppc_perf_from_fbctrs(struct cppc_cpudata * cpu_data,struct cppc_perf_fb_ctrs * fb_ctrs_t0,struct cppc_perf_fb_ctrs * fb_ctrs_t1)702 static int cppc_perf_from_fbctrs(struct cppc_cpudata *cpu_data,
703 struct cppc_perf_fb_ctrs *fb_ctrs_t0,
704 struct cppc_perf_fb_ctrs *fb_ctrs_t1)
705 {
706 u64 delta_reference, delta_delivered;
707 u64 reference_perf;
708
709 reference_perf = fb_ctrs_t0->reference_perf;
710
711 delta_reference = get_delta(fb_ctrs_t1->reference,
712 fb_ctrs_t0->reference);
713 delta_delivered = get_delta(fb_ctrs_t1->delivered,
714 fb_ctrs_t0->delivered);
715
716 /*
717 * Avoid divide-by zero and unchanged feedback counters.
718 * Leave it for callers to handle.
719 */
720 if (!delta_reference || !delta_delivered)
721 return 0;
722
723 return (reference_perf * delta_delivered) / delta_reference;
724 }
725
cppc_get_perf_ctrs_sample(int cpu,struct cppc_perf_fb_ctrs * fb_ctrs_t0,struct cppc_perf_fb_ctrs * fb_ctrs_t1)726 static int cppc_get_perf_ctrs_sample(int cpu,
727 struct cppc_perf_fb_ctrs *fb_ctrs_t0,
728 struct cppc_perf_fb_ctrs *fb_ctrs_t1)
729 {
730 int ret;
731
732 ret = cppc_get_perf_ctrs(cpu, fb_ctrs_t0);
733 if (ret)
734 return ret;
735
736 udelay(2); /* 2usec delay between sampling */
737
738 return cppc_get_perf_ctrs(cpu, fb_ctrs_t1);
739 }
740
cppc_cpufreq_get_rate(unsigned int cpu)741 static unsigned int cppc_cpufreq_get_rate(unsigned int cpu)
742 {
743 struct cppc_perf_fb_ctrs fb_ctrs_t0 = {0}, fb_ctrs_t1 = {0};
744 struct cpufreq_policy *policy = cpufreq_cpu_get(cpu);
745 struct cppc_cpudata *cpu_data;
746 u64 delivered_perf;
747 int ret;
748
749 if (!policy)
750 return 0;
751
752 cpu_data = policy->driver_data;
753
754 cpufreq_cpu_put(policy);
755
756 ret = cppc_get_perf_ctrs_sample(cpu, &fb_ctrs_t0, &fb_ctrs_t1);
757 if (ret) {
758 if (ret == -EFAULT)
759 /* Any of the associated CPPC regs is 0. */
760 goto out_invalid_counters;
761 else
762 return 0;
763 }
764
765 delivered_perf = cppc_perf_from_fbctrs(cpu_data, &fb_ctrs_t0,
766 &fb_ctrs_t1);
767 if (!delivered_perf)
768 goto out_invalid_counters;
769
770 return cppc_perf_to_khz(&cpu_data->perf_caps, delivered_perf);
771
772 out_invalid_counters:
773 /*
774 * Feedback counters could be unchanged or 0 when a cpu enters a
775 * low-power idle state, e.g. clock-gated or power-gated.
776 * Use desired perf for reflecting frequency. Get the latest register
777 * value first as some platforms may update the actual delivered perf
778 * there; if failed, resort to the cached desired perf.
779 */
780 if (cppc_get_desired_perf(cpu, &delivered_perf))
781 delivered_perf = cpu_data->perf_ctrls.desired_perf;
782
783 return cppc_perf_to_khz(&cpu_data->perf_caps, delivered_perf);
784 }
785
cppc_cpufreq_set_boost(struct cpufreq_policy * policy,int state)786 static int cppc_cpufreq_set_boost(struct cpufreq_policy *policy, int state)
787 {
788 struct cppc_cpudata *cpu_data = policy->driver_data;
789 struct cppc_perf_caps *caps = &cpu_data->perf_caps;
790 int ret;
791
792 if (state)
793 policy->max = cppc_perf_to_khz(caps, caps->highest_perf);
794 else
795 policy->max = cppc_perf_to_khz(caps, caps->nominal_perf);
796 policy->cpuinfo.max_freq = policy->max;
797
798 ret = freq_qos_update_request(policy->max_freq_req, policy->max);
799 if (ret < 0)
800 return ret;
801
802 return 0;
803 }
804
show_freqdomain_cpus(struct cpufreq_policy * policy,char * buf)805 static ssize_t show_freqdomain_cpus(struct cpufreq_policy *policy, char *buf)
806 {
807 struct cppc_cpudata *cpu_data = policy->driver_data;
808
809 return cpufreq_show_cpus(cpu_data->shared_cpu_map, buf);
810 }
811
show_auto_select(struct cpufreq_policy * policy,char * buf)812 static ssize_t show_auto_select(struct cpufreq_policy *policy, char *buf)
813 {
814 bool val;
815 int ret;
816
817 ret = cppc_get_auto_sel(policy->cpu, &val);
818
819 /* show "<unsupported>" when this register is not supported by cpc */
820 if (ret == -EOPNOTSUPP)
821 return sysfs_emit(buf, "<unsupported>\n");
822
823 if (ret)
824 return ret;
825
826 return sysfs_emit(buf, "%d\n", val);
827 }
828
store_auto_select(struct cpufreq_policy * policy,const char * buf,size_t count)829 static ssize_t store_auto_select(struct cpufreq_policy *policy,
830 const char *buf, size_t count)
831 {
832 bool val;
833 int ret;
834
835 ret = kstrtobool(buf, &val);
836 if (ret)
837 return ret;
838
839 ret = cppc_set_auto_sel(policy->cpu, val);
840 if (ret)
841 return ret;
842
843 return count;
844 }
845
show_auto_act_window(struct cpufreq_policy * policy,char * buf)846 static ssize_t show_auto_act_window(struct cpufreq_policy *policy, char *buf)
847 {
848 u64 val;
849 int ret;
850
851 ret = cppc_get_auto_act_window(policy->cpu, &val);
852
853 /* show "<unsupported>" when this register is not supported by cpc */
854 if (ret == -EOPNOTSUPP)
855 return sysfs_emit(buf, "<unsupported>\n");
856
857 if (ret)
858 return ret;
859
860 return sysfs_emit(buf, "%llu\n", val);
861 }
862
store_auto_act_window(struct cpufreq_policy * policy,const char * buf,size_t count)863 static ssize_t store_auto_act_window(struct cpufreq_policy *policy,
864 const char *buf, size_t count)
865 {
866 u64 usec;
867 int ret;
868
869 ret = kstrtou64(buf, 0, &usec);
870 if (ret)
871 return ret;
872
873 ret = cppc_set_auto_act_window(policy->cpu, usec);
874 if (ret)
875 return ret;
876
877 return count;
878 }
879
show_energy_performance_preference_val(struct cpufreq_policy * policy,char * buf)880 static ssize_t show_energy_performance_preference_val(struct cpufreq_policy *policy, char *buf)
881 {
882 u64 val;
883 int ret;
884
885 ret = cppc_get_epp_perf(policy->cpu, &val);
886
887 /* show "<unsupported>" when this register is not supported by cpc */
888 if (ret == -EOPNOTSUPP)
889 return sysfs_emit(buf, "<unsupported>\n");
890
891 if (ret)
892 return ret;
893
894 return sysfs_emit(buf, "%llu\n", val);
895 }
896
store_energy_performance_preference_val(struct cpufreq_policy * policy,const char * buf,size_t count)897 static ssize_t store_energy_performance_preference_val(struct cpufreq_policy *policy,
898 const char *buf, size_t count)
899 {
900 u64 val;
901 int ret;
902
903 ret = kstrtou64(buf, 0, &val);
904 if (ret)
905 return ret;
906
907 ret = cppc_set_epp(policy->cpu, val);
908 if (ret)
909 return ret;
910
911 return count;
912 }
913
914 cpufreq_freq_attr_ro(freqdomain_cpus);
915 cpufreq_freq_attr_rw(auto_select);
916 cpufreq_freq_attr_rw(auto_act_window);
917 cpufreq_freq_attr_rw(energy_performance_preference_val);
918
919 static struct freq_attr *cppc_cpufreq_attr[] = {
920 &freqdomain_cpus,
921 &auto_select,
922 &auto_act_window,
923 &energy_performance_preference_val,
924 NULL,
925 };
926
927 static struct cpufreq_driver cppc_cpufreq_driver = {
928 .flags = CPUFREQ_CONST_LOOPS,
929 .verify = cppc_verify_policy,
930 .target = cppc_cpufreq_set_target,
931 .get = cppc_cpufreq_get_rate,
932 .fast_switch = cppc_cpufreq_fast_switch,
933 .init = cppc_cpufreq_cpu_init,
934 .exit = cppc_cpufreq_cpu_exit,
935 .set_boost = cppc_cpufreq_set_boost,
936 .attr = cppc_cpufreq_attr,
937 .name = "cppc_cpufreq",
938 };
939
cppc_cpufreq_init(void)940 static int __init cppc_cpufreq_init(void)
941 {
942 int ret;
943
944 if (!acpi_cpc_valid())
945 return -ENODEV;
946
947 cppc_freq_invariance_init();
948 populate_efficiency_class();
949
950 ret = cpufreq_register_driver(&cppc_cpufreq_driver);
951 if (ret)
952 cppc_freq_invariance_exit();
953
954 return ret;
955 }
956
free_cpu_data(void)957 static inline void free_cpu_data(void)
958 {
959 struct cppc_cpudata *iter, *tmp;
960
961 list_for_each_entry_safe(iter, tmp, &cpu_data_list, node) {
962 free_cpumask_var(iter->shared_cpu_map);
963 list_del(&iter->node);
964 kfree(iter);
965 }
966
967 }
968
cppc_cpufreq_exit(void)969 static void __exit cppc_cpufreq_exit(void)
970 {
971 cpufreq_unregister_driver(&cppc_cpufreq_driver);
972 cppc_freq_invariance_exit();
973
974 free_cpu_data();
975 }
976
977 module_exit(cppc_cpufreq_exit);
978 MODULE_AUTHOR("Ashwin Chaugule");
979 MODULE_DESCRIPTION("CPUFreq driver based on the ACPI CPPC v5.0+ spec");
980 MODULE_LICENSE("GPL");
981
982 late_initcall(cppc_cpufreq_init);
983
984 static const struct acpi_device_id cppc_acpi_ids[] __used = {
985 {ACPI_PROCESSOR_DEVICE_HID, },
986 {}
987 };
988
989 MODULE_DEVICE_TABLE(acpi, cppc_acpi_ids);
990