Loading...
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 <asm/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 */
35static LIST_HEAD(cpu_data_list);
36
37static bool boost_supported;
38
39struct cppc_workaround_oem_info {
40 char oem_id[ACPI_OEM_ID_SIZE + 1];
41 char oem_table_id[ACPI_OEM_TABLE_ID_SIZE + 1];
42 u32 oem_revision;
43};
44
45static struct cppc_workaround_oem_info wa_info[] = {
46 {
47 .oem_id = "HISI ",
48 .oem_table_id = "HIP07 ",
49 .oem_revision = 0,
50 }, {
51 .oem_id = "HISI ",
52 .oem_table_id = "HIP08 ",
53 .oem_revision = 0,
54 }
55};
56
57static struct cpufreq_driver cppc_cpufreq_driver;
58
59static enum {
60 FIE_UNSET = -1,
61 FIE_ENABLED,
62 FIE_DISABLED
63} fie_disabled = FIE_UNSET;
64
65#ifdef CONFIG_ACPI_CPPC_CPUFREQ_FIE
66module_param(fie_disabled, int, 0444);
67MODULE_PARM_DESC(fie_disabled, "Disable Frequency Invariance Engine (FIE)");
68
69/* Frequency invariance support */
70struct cppc_freq_invariance {
71 int cpu;
72 struct irq_work irq_work;
73 struct kthread_work work;
74 struct cppc_perf_fb_ctrs prev_perf_fb_ctrs;
75 struct cppc_cpudata *cpu_data;
76};
77
78static DEFINE_PER_CPU(struct cppc_freq_invariance, cppc_freq_inv);
79static struct kthread_worker *kworker_fie;
80
81static unsigned int hisi_cppc_cpufreq_get_rate(unsigned int cpu);
82static int cppc_perf_from_fbctrs(struct cppc_cpudata *cpu_data,
83 struct cppc_perf_fb_ctrs *fb_ctrs_t0,
84 struct cppc_perf_fb_ctrs *fb_ctrs_t1);
85
86/**
87 * cppc_scale_freq_workfn - CPPC arch_freq_scale updater for frequency invariance
88 * @work: The work item.
89 *
90 * The CPPC driver register itself with the topology core to provide its own
91 * implementation (cppc_scale_freq_tick()) of topology_scale_freq_tick() which
92 * gets called by the scheduler on every tick.
93 *
94 * Note that the arch specific counters have higher priority than CPPC counters,
95 * if available, though the CPPC driver doesn't need to have any special
96 * handling for that.
97 *
98 * On an invocation of cppc_scale_freq_tick(), we schedule an irq work (since we
99 * reach here from hard-irq context), which then schedules a normal work item
100 * and cppc_scale_freq_workfn() updates the per_cpu arch_freq_scale variable
101 * based on the counter updates since the last tick.
102 */
103static void cppc_scale_freq_workfn(struct kthread_work *work)
104{
105 struct cppc_freq_invariance *cppc_fi;
106 struct cppc_perf_fb_ctrs fb_ctrs = {0};
107 struct cppc_cpudata *cpu_data;
108 unsigned long local_freq_scale;
109 u64 perf;
110
111 cppc_fi = container_of(work, struct cppc_freq_invariance, work);
112 cpu_data = cppc_fi->cpu_data;
113
114 if (cppc_get_perf_ctrs(cppc_fi->cpu, &fb_ctrs)) {
115 pr_warn("%s: failed to read perf counters\n", __func__);
116 return;
117 }
118
119 perf = cppc_perf_from_fbctrs(cpu_data, &cppc_fi->prev_perf_fb_ctrs,
120 &fb_ctrs);
121 cppc_fi->prev_perf_fb_ctrs = fb_ctrs;
122
123 perf <<= SCHED_CAPACITY_SHIFT;
124 local_freq_scale = div64_u64(perf, cpu_data->perf_caps.highest_perf);
125
126 /* This can happen due to counter's overflow */
127 if (unlikely(local_freq_scale > 1024))
128 local_freq_scale = 1024;
129
130 per_cpu(arch_freq_scale, cppc_fi->cpu) = local_freq_scale;
131}
132
133static void cppc_irq_work(struct irq_work *irq_work)
134{
135 struct cppc_freq_invariance *cppc_fi;
136
137 cppc_fi = container_of(irq_work, struct cppc_freq_invariance, irq_work);
138 kthread_queue_work(kworker_fie, &cppc_fi->work);
139}
140
141static void cppc_scale_freq_tick(void)
142{
143 struct cppc_freq_invariance *cppc_fi = &per_cpu(cppc_freq_inv, smp_processor_id());
144
145 /*
146 * cppc_get_perf_ctrs() can potentially sleep, call that from the right
147 * context.
148 */
149 irq_work_queue(&cppc_fi->irq_work);
150}
151
152static struct scale_freq_data cppc_sftd = {
153 .source = SCALE_FREQ_SOURCE_CPPC,
154 .set_freq_scale = cppc_scale_freq_tick,
155};
156
157static void cppc_cpufreq_cpu_fie_init(struct cpufreq_policy *policy)
158{
159 struct cppc_freq_invariance *cppc_fi;
160 int cpu, ret;
161
162 if (fie_disabled)
163 return;
164
165 for_each_cpu(cpu, policy->cpus) {
166 cppc_fi = &per_cpu(cppc_freq_inv, cpu);
167 cppc_fi->cpu = cpu;
168 cppc_fi->cpu_data = policy->driver_data;
169 kthread_init_work(&cppc_fi->work, cppc_scale_freq_workfn);
170 init_irq_work(&cppc_fi->irq_work, cppc_irq_work);
171
172 ret = cppc_get_perf_ctrs(cpu, &cppc_fi->prev_perf_fb_ctrs);
173 if (ret) {
174 pr_warn("%s: failed to read perf counters for cpu:%d: %d\n",
175 __func__, cpu, ret);
176
177 /*
178 * Don't abort if the CPU was offline while the driver
179 * was getting registered.
180 */
181 if (cpu_online(cpu))
182 return;
183 }
184 }
185
186 /* Register for freq-invariance */
187 topology_set_scale_freq_source(&cppc_sftd, policy->cpus);
188}
189
190/*
191 * We free all the resources on policy's removal and not on CPU removal as the
192 * irq-work are per-cpu and the hotplug core takes care of flushing the pending
193 * irq-works (hint: smpcfd_dying_cpu()) on CPU hotplug. Even if the kthread-work
194 * fires on another CPU after the concerned CPU is removed, it won't harm.
195 *
196 * We just need to make sure to remove them all on policy->exit().
197 */
198static void cppc_cpufreq_cpu_fie_exit(struct cpufreq_policy *policy)
199{
200 struct cppc_freq_invariance *cppc_fi;
201 int cpu;
202
203 if (fie_disabled)
204 return;
205
206 /* policy->cpus will be empty here, use related_cpus instead */
207 topology_clear_scale_freq_source(SCALE_FREQ_SOURCE_CPPC, policy->related_cpus);
208
209 for_each_cpu(cpu, policy->related_cpus) {
210 cppc_fi = &per_cpu(cppc_freq_inv, cpu);
211 irq_work_sync(&cppc_fi->irq_work);
212 kthread_cancel_work_sync(&cppc_fi->work);
213 }
214}
215
216static void __init cppc_freq_invariance_init(void)
217{
218 struct sched_attr attr = {
219 .size = sizeof(struct sched_attr),
220 .sched_policy = SCHED_DEADLINE,
221 .sched_nice = 0,
222 .sched_priority = 0,
223 /*
224 * Fake (unused) bandwidth; workaround to "fix"
225 * priority inheritance.
226 */
227 .sched_runtime = 1000000,
228 .sched_deadline = 10000000,
229 .sched_period = 10000000,
230 };
231 int ret;
232
233 if (fie_disabled != FIE_ENABLED && fie_disabled != FIE_DISABLED) {
234 fie_disabled = FIE_ENABLED;
235 if (cppc_perf_ctrs_in_pcc()) {
236 pr_info("FIE not enabled on systems with registers in PCC\n");
237 fie_disabled = FIE_DISABLED;
238 }
239 }
240
241 if (fie_disabled)
242 return;
243
244 kworker_fie = kthread_create_worker(0, "cppc_fie");
245 if (IS_ERR(kworker_fie)) {
246 pr_warn("%s: failed to create kworker_fie: %ld\n", __func__,
247 PTR_ERR(kworker_fie));
248 fie_disabled = FIE_DISABLED;
249 return;
250 }
251
252 ret = sched_setattr_nocheck(kworker_fie->task, &attr);
253 if (ret) {
254 pr_warn("%s: failed to set SCHED_DEADLINE: %d\n", __func__,
255 ret);
256 kthread_destroy_worker(kworker_fie);
257 fie_disabled = FIE_DISABLED;
258 }
259}
260
261static void cppc_freq_invariance_exit(void)
262{
263 if (fie_disabled)
264 return;
265
266 kthread_destroy_worker(kworker_fie);
267}
268
269#else
270static inline void cppc_cpufreq_cpu_fie_init(struct cpufreq_policy *policy)
271{
272}
273
274static inline void cppc_cpufreq_cpu_fie_exit(struct cpufreq_policy *policy)
275{
276}
277
278static inline void cppc_freq_invariance_init(void)
279{
280}
281
282static inline void cppc_freq_invariance_exit(void)
283{
284}
285#endif /* CONFIG_ACPI_CPPC_CPUFREQ_FIE */
286
287static int cppc_cpufreq_set_target(struct cpufreq_policy *policy,
288 unsigned int target_freq,
289 unsigned int relation)
290{
291 struct cppc_cpudata *cpu_data = policy->driver_data;
292 unsigned int cpu = policy->cpu;
293 struct cpufreq_freqs freqs;
294 u32 desired_perf;
295 int ret = 0;
296
297 desired_perf = cppc_khz_to_perf(&cpu_data->perf_caps, target_freq);
298 /* Return if it is exactly the same perf */
299 if (desired_perf == cpu_data->perf_ctrls.desired_perf)
300 return ret;
301
302 cpu_data->perf_ctrls.desired_perf = desired_perf;
303 freqs.old = policy->cur;
304 freqs.new = target_freq;
305
306 cpufreq_freq_transition_begin(policy, &freqs);
307 ret = cppc_set_perf(cpu, &cpu_data->perf_ctrls);
308 cpufreq_freq_transition_end(policy, &freqs, ret != 0);
309
310 if (ret)
311 pr_debug("Failed to set target on CPU:%d. ret:%d\n",
312 cpu, ret);
313
314 return ret;
315}
316
317static unsigned int cppc_cpufreq_fast_switch(struct cpufreq_policy *policy,
318 unsigned int target_freq)
319{
320 struct cppc_cpudata *cpu_data = policy->driver_data;
321 unsigned int cpu = policy->cpu;
322 u32 desired_perf;
323 int ret;
324
325 desired_perf = cppc_khz_to_perf(&cpu_data->perf_caps, target_freq);
326 cpu_data->perf_ctrls.desired_perf = desired_perf;
327 ret = cppc_set_perf(cpu, &cpu_data->perf_ctrls);
328
329 if (ret) {
330 pr_debug("Failed to set target on CPU:%d. ret:%d\n",
331 cpu, ret);
332 return 0;
333 }
334
335 return target_freq;
336}
337
338static int cppc_verify_policy(struct cpufreq_policy_data *policy)
339{
340 cpufreq_verify_within_cpu_limits(policy);
341 return 0;
342}
343
344/*
345 * The PCC subspace describes the rate at which platform can accept commands
346 * on the shared PCC channel (including READs which do not count towards freq
347 * transition requests), so ideally we need to use the PCC values as a fallback
348 * if we don't have a platform specific transition_delay_us
349 */
350#ifdef CONFIG_ARM64
351#include <asm/cputype.h>
352
353static unsigned int cppc_cpufreq_get_transition_delay_us(unsigned int cpu)
354{
355 unsigned long implementor = read_cpuid_implementor();
356 unsigned long part_num = read_cpuid_part_number();
357
358 switch (implementor) {
359 case ARM_CPU_IMP_QCOM:
360 switch (part_num) {
361 case QCOM_CPU_PART_FALKOR_V1:
362 case QCOM_CPU_PART_FALKOR:
363 return 10000;
364 }
365 }
366 return cppc_get_transition_latency(cpu) / NSEC_PER_USEC;
367}
368#else
369static unsigned int cppc_cpufreq_get_transition_delay_us(unsigned int cpu)
370{
371 return cppc_get_transition_latency(cpu) / NSEC_PER_USEC;
372}
373#endif
374
375#if defined(CONFIG_ARM64) && defined(CONFIG_ENERGY_MODEL)
376
377static DEFINE_PER_CPU(unsigned int, efficiency_class);
378static void cppc_cpufreq_register_em(struct cpufreq_policy *policy);
379
380/* Create an artificial performance state every CPPC_EM_CAP_STEP capacity unit. */
381#define CPPC_EM_CAP_STEP (20)
382/* Increase the cost value by CPPC_EM_COST_STEP every performance state. */
383#define CPPC_EM_COST_STEP (1)
384/* Add a cost gap correspnding to the energy of 4 CPUs. */
385#define CPPC_EM_COST_GAP (4 * SCHED_CAPACITY_SCALE * CPPC_EM_COST_STEP \
386 / CPPC_EM_CAP_STEP)
387
388static unsigned int get_perf_level_count(struct cpufreq_policy *policy)
389{
390 struct cppc_perf_caps *perf_caps;
391 unsigned int min_cap, max_cap;
392 struct cppc_cpudata *cpu_data;
393 int cpu = policy->cpu;
394
395 cpu_data = policy->driver_data;
396 perf_caps = &cpu_data->perf_caps;
397 max_cap = arch_scale_cpu_capacity(cpu);
398 min_cap = div_u64((u64)max_cap * perf_caps->lowest_perf,
399 perf_caps->highest_perf);
400 if ((min_cap == 0) || (max_cap < min_cap))
401 return 0;
402 return 1 + max_cap / CPPC_EM_CAP_STEP - min_cap / CPPC_EM_CAP_STEP;
403}
404
405/*
406 * The cost is defined as:
407 * cost = power * max_frequency / frequency
408 */
409static inline unsigned long compute_cost(int cpu, int step)
410{
411 return CPPC_EM_COST_GAP * per_cpu(efficiency_class, cpu) +
412 step * CPPC_EM_COST_STEP;
413}
414
415static int cppc_get_cpu_power(struct device *cpu_dev,
416 unsigned long *power, unsigned long *KHz)
417{
418 unsigned long perf_step, perf_prev, perf, perf_check;
419 unsigned int min_step, max_step, step, step_check;
420 unsigned long prev_freq = *KHz;
421 unsigned int min_cap, max_cap;
422 struct cpufreq_policy *policy;
423
424 struct cppc_perf_caps *perf_caps;
425 struct cppc_cpudata *cpu_data;
426
427 policy = cpufreq_cpu_get_raw(cpu_dev->id);
428 cpu_data = policy->driver_data;
429 perf_caps = &cpu_data->perf_caps;
430 max_cap = arch_scale_cpu_capacity(cpu_dev->id);
431 min_cap = div_u64((u64)max_cap * perf_caps->lowest_perf,
432 perf_caps->highest_perf);
433 perf_step = div_u64((u64)CPPC_EM_CAP_STEP * perf_caps->highest_perf,
434 max_cap);
435 min_step = min_cap / CPPC_EM_CAP_STEP;
436 max_step = max_cap / CPPC_EM_CAP_STEP;
437
438 perf_prev = cppc_khz_to_perf(perf_caps, *KHz);
439 step = perf_prev / perf_step;
440
441 if (step > max_step)
442 return -EINVAL;
443
444 if (min_step == max_step) {
445 step = max_step;
446 perf = perf_caps->highest_perf;
447 } else if (step < min_step) {
448 step = min_step;
449 perf = perf_caps->lowest_perf;
450 } else {
451 step++;
452 if (step == max_step)
453 perf = perf_caps->highest_perf;
454 else
455 perf = step * perf_step;
456 }
457
458 *KHz = cppc_perf_to_khz(perf_caps, perf);
459 perf_check = cppc_khz_to_perf(perf_caps, *KHz);
460 step_check = perf_check / perf_step;
461
462 /*
463 * To avoid bad integer approximation, check that new frequency value
464 * increased and that the new frequency will be converted to the
465 * desired step value.
466 */
467 while ((*KHz == prev_freq) || (step_check != step)) {
468 perf++;
469 *KHz = cppc_perf_to_khz(perf_caps, perf);
470 perf_check = cppc_khz_to_perf(perf_caps, *KHz);
471 step_check = perf_check / perf_step;
472 }
473
474 /*
475 * With an artificial EM, only the cost value is used. Still the power
476 * is populated such as 0 < power < EM_MAX_POWER. This allows to add
477 * more sense to the artificial performance states.
478 */
479 *power = compute_cost(cpu_dev->id, step);
480
481 return 0;
482}
483
484static int cppc_get_cpu_cost(struct device *cpu_dev, unsigned long KHz,
485 unsigned long *cost)
486{
487 unsigned long perf_step, perf_prev;
488 struct cppc_perf_caps *perf_caps;
489 struct cpufreq_policy *policy;
490 struct cppc_cpudata *cpu_data;
491 unsigned int max_cap;
492 int step;
493
494 policy = cpufreq_cpu_get_raw(cpu_dev->id);
495 cpu_data = policy->driver_data;
496 perf_caps = &cpu_data->perf_caps;
497 max_cap = arch_scale_cpu_capacity(cpu_dev->id);
498
499 perf_prev = cppc_khz_to_perf(perf_caps, KHz);
500 perf_step = CPPC_EM_CAP_STEP * perf_caps->highest_perf / max_cap;
501 step = perf_prev / perf_step;
502
503 *cost = compute_cost(cpu_dev->id, step);
504
505 return 0;
506}
507
508static int populate_efficiency_class(void)
509{
510 struct acpi_madt_generic_interrupt *gicc;
511 DECLARE_BITMAP(used_classes, 256) = {};
512 int class, cpu, index;
513
514 for_each_possible_cpu(cpu) {
515 gicc = acpi_cpu_get_madt_gicc(cpu);
516 class = gicc->efficiency_class;
517 bitmap_set(used_classes, class, 1);
518 }
519
520 if (bitmap_weight(used_classes, 256) <= 1) {
521 pr_debug("Efficiency classes are all equal (=%d). "
522 "No EM registered", class);
523 return -EINVAL;
524 }
525
526 /*
527 * Squeeze efficiency class values on [0:#efficiency_class-1].
528 * Values are per spec in [0:255].
529 */
530 index = 0;
531 for_each_set_bit(class, used_classes, 256) {
532 for_each_possible_cpu(cpu) {
533 gicc = acpi_cpu_get_madt_gicc(cpu);
534 if (gicc->efficiency_class == class)
535 per_cpu(efficiency_class, cpu) = index;
536 }
537 index++;
538 }
539 cppc_cpufreq_driver.register_em = cppc_cpufreq_register_em;
540
541 return 0;
542}
543
544static void cppc_cpufreq_register_em(struct cpufreq_policy *policy)
545{
546 struct cppc_cpudata *cpu_data;
547 struct em_data_callback em_cb =
548 EM_ADV_DATA_CB(cppc_get_cpu_power, cppc_get_cpu_cost);
549
550 cpu_data = policy->driver_data;
551 em_dev_register_perf_domain(get_cpu_device(policy->cpu),
552 get_perf_level_count(policy), &em_cb,
553 cpu_data->shared_cpu_map, 0);
554}
555
556#else
557static int populate_efficiency_class(void)
558{
559 return 0;
560}
561#endif
562
563static struct cppc_cpudata *cppc_cpufreq_get_cpu_data(unsigned int cpu)
564{
565 struct cppc_cpudata *cpu_data;
566 int ret;
567
568 cpu_data = kzalloc(sizeof(struct cppc_cpudata), GFP_KERNEL);
569 if (!cpu_data)
570 goto out;
571
572 if (!zalloc_cpumask_var(&cpu_data->shared_cpu_map, GFP_KERNEL))
573 goto free_cpu;
574
575 ret = acpi_get_psd_map(cpu, cpu_data);
576 if (ret) {
577 pr_debug("Err parsing CPU%d PSD data: ret:%d\n", cpu, ret);
578 goto free_mask;
579 }
580
581 ret = cppc_get_perf_caps(cpu, &cpu_data->perf_caps);
582 if (ret) {
583 pr_debug("Err reading CPU%d perf caps: ret:%d\n", cpu, ret);
584 goto free_mask;
585 }
586
587 list_add(&cpu_data->node, &cpu_data_list);
588
589 return cpu_data;
590
591free_mask:
592 free_cpumask_var(cpu_data->shared_cpu_map);
593free_cpu:
594 kfree(cpu_data);
595out:
596 return NULL;
597}
598
599static void cppc_cpufreq_put_cpu_data(struct cpufreq_policy *policy)
600{
601 struct cppc_cpudata *cpu_data = policy->driver_data;
602
603 list_del(&cpu_data->node);
604 free_cpumask_var(cpu_data->shared_cpu_map);
605 kfree(cpu_data);
606 policy->driver_data = NULL;
607}
608
609static int cppc_cpufreq_cpu_init(struct cpufreq_policy *policy)
610{
611 unsigned int cpu = policy->cpu;
612 struct cppc_cpudata *cpu_data;
613 struct cppc_perf_caps *caps;
614 int ret;
615
616 cpu_data = cppc_cpufreq_get_cpu_data(cpu);
617 if (!cpu_data) {
618 pr_err("Error in acquiring _CPC/_PSD data for CPU%d.\n", cpu);
619 return -ENODEV;
620 }
621 caps = &cpu_data->perf_caps;
622 policy->driver_data = cpu_data;
623
624 /*
625 * Set min to lowest nonlinear perf to avoid any efficiency penalty (see
626 * Section 8.4.7.1.1.5 of ACPI 6.1 spec)
627 */
628 policy->min = cppc_perf_to_khz(caps, caps->lowest_nonlinear_perf);
629 policy->max = cppc_perf_to_khz(caps, caps->nominal_perf);
630
631 /*
632 * Set cpuinfo.min_freq to Lowest to make the full range of performance
633 * available if userspace wants to use any perf between lowest & lowest
634 * nonlinear perf
635 */
636 policy->cpuinfo.min_freq = cppc_perf_to_khz(caps, caps->lowest_perf);
637 policy->cpuinfo.max_freq = cppc_perf_to_khz(caps, caps->nominal_perf);
638
639 policy->transition_delay_us = cppc_cpufreq_get_transition_delay_us(cpu);
640 policy->shared_type = cpu_data->shared_type;
641
642 switch (policy->shared_type) {
643 case CPUFREQ_SHARED_TYPE_HW:
644 case CPUFREQ_SHARED_TYPE_NONE:
645 /* Nothing to be done - we'll have a policy for each CPU */
646 break;
647 case CPUFREQ_SHARED_TYPE_ANY:
648 /*
649 * All CPUs in the domain will share a policy and all cpufreq
650 * operations will use a single cppc_cpudata structure stored
651 * in policy->driver_data.
652 */
653 cpumask_copy(policy->cpus, cpu_data->shared_cpu_map);
654 break;
655 default:
656 pr_debug("Unsupported CPU co-ord type: %d\n",
657 policy->shared_type);
658 ret = -EFAULT;
659 goto out;
660 }
661
662 policy->fast_switch_possible = cppc_allow_fast_switch();
663 policy->dvfs_possible_from_any_cpu = true;
664
665 /*
666 * If 'highest_perf' is greater than 'nominal_perf', we assume CPU Boost
667 * is supported.
668 */
669 if (caps->highest_perf > caps->nominal_perf)
670 boost_supported = true;
671
672 /* Set policy->cur to max now. The governors will adjust later. */
673 policy->cur = cppc_perf_to_khz(caps, caps->highest_perf);
674 cpu_data->perf_ctrls.desired_perf = caps->highest_perf;
675
676 ret = cppc_set_perf(cpu, &cpu_data->perf_ctrls);
677 if (ret) {
678 pr_debug("Err setting perf value:%d on CPU:%d. ret:%d\n",
679 caps->highest_perf, cpu, ret);
680 goto out;
681 }
682
683 cppc_cpufreq_cpu_fie_init(policy);
684 return 0;
685
686out:
687 cppc_cpufreq_put_cpu_data(policy);
688 return ret;
689}
690
691static int cppc_cpufreq_cpu_exit(struct cpufreq_policy *policy)
692{
693 struct cppc_cpudata *cpu_data = policy->driver_data;
694 struct cppc_perf_caps *caps = &cpu_data->perf_caps;
695 unsigned int cpu = policy->cpu;
696 int ret;
697
698 cppc_cpufreq_cpu_fie_exit(policy);
699
700 cpu_data->perf_ctrls.desired_perf = caps->lowest_perf;
701
702 ret = cppc_set_perf(cpu, &cpu_data->perf_ctrls);
703 if (ret)
704 pr_debug("Err setting perf value:%d on CPU:%d. ret:%d\n",
705 caps->lowest_perf, cpu, ret);
706
707 cppc_cpufreq_put_cpu_data(policy);
708 return 0;
709}
710
711static inline u64 get_delta(u64 t1, u64 t0)
712{
713 if (t1 > t0 || t0 > ~(u32)0)
714 return t1 - t0;
715
716 return (u32)t1 - (u32)t0;
717}
718
719static int cppc_perf_from_fbctrs(struct cppc_cpudata *cpu_data,
720 struct cppc_perf_fb_ctrs *fb_ctrs_t0,
721 struct cppc_perf_fb_ctrs *fb_ctrs_t1)
722{
723 u64 delta_reference, delta_delivered;
724 u64 reference_perf;
725
726 reference_perf = fb_ctrs_t0->reference_perf;
727
728 delta_reference = get_delta(fb_ctrs_t1->reference,
729 fb_ctrs_t0->reference);
730 delta_delivered = get_delta(fb_ctrs_t1->delivered,
731 fb_ctrs_t0->delivered);
732
733 /* Check to avoid divide-by zero and invalid delivered_perf */
734 if (!delta_reference || !delta_delivered)
735 return cpu_data->perf_ctrls.desired_perf;
736
737 return (reference_perf * delta_delivered) / delta_reference;
738}
739
740static unsigned int cppc_cpufreq_get_rate(unsigned int cpu)
741{
742 struct cppc_perf_fb_ctrs fb_ctrs_t0 = {0}, fb_ctrs_t1 = {0};
743 struct cpufreq_policy *policy = cpufreq_cpu_get(cpu);
744 struct cppc_cpudata *cpu_data = policy->driver_data;
745 u64 delivered_perf;
746 int ret;
747
748 cpufreq_cpu_put(policy);
749
750 ret = cppc_get_perf_ctrs(cpu, &fb_ctrs_t0);
751 if (ret)
752 return 0;
753
754 udelay(2); /* 2usec delay between sampling */
755
756 ret = cppc_get_perf_ctrs(cpu, &fb_ctrs_t1);
757 if (ret)
758 return 0;
759
760 delivered_perf = cppc_perf_from_fbctrs(cpu_data, &fb_ctrs_t0,
761 &fb_ctrs_t1);
762
763 return cppc_perf_to_khz(&cpu_data->perf_caps, delivered_perf);
764}
765
766static int cppc_cpufreq_set_boost(struct cpufreq_policy *policy, int state)
767{
768 struct cppc_cpudata *cpu_data = policy->driver_data;
769 struct cppc_perf_caps *caps = &cpu_data->perf_caps;
770 int ret;
771
772 if (!boost_supported) {
773 pr_err("BOOST not supported by CPU or firmware\n");
774 return -EINVAL;
775 }
776
777 if (state)
778 policy->max = cppc_perf_to_khz(caps, caps->highest_perf);
779 else
780 policy->max = cppc_perf_to_khz(caps, caps->nominal_perf);
781 policy->cpuinfo.max_freq = policy->max;
782
783 ret = freq_qos_update_request(policy->max_freq_req, policy->max);
784 if (ret < 0)
785 return ret;
786
787 return 0;
788}
789
790static ssize_t show_freqdomain_cpus(struct cpufreq_policy *policy, char *buf)
791{
792 struct cppc_cpudata *cpu_data = policy->driver_data;
793
794 return cpufreq_show_cpus(cpu_data->shared_cpu_map, buf);
795}
796cpufreq_freq_attr_ro(freqdomain_cpus);
797
798static struct freq_attr *cppc_cpufreq_attr[] = {
799 &freqdomain_cpus,
800 NULL,
801};
802
803static struct cpufreq_driver cppc_cpufreq_driver = {
804 .flags = CPUFREQ_CONST_LOOPS,
805 .verify = cppc_verify_policy,
806 .target = cppc_cpufreq_set_target,
807 .get = cppc_cpufreq_get_rate,
808 .fast_switch = cppc_cpufreq_fast_switch,
809 .init = cppc_cpufreq_cpu_init,
810 .exit = cppc_cpufreq_cpu_exit,
811 .set_boost = cppc_cpufreq_set_boost,
812 .attr = cppc_cpufreq_attr,
813 .name = "cppc_cpufreq",
814};
815
816/*
817 * HISI platform does not support delivered performance counter and
818 * reference performance counter. It can calculate the performance using the
819 * platform specific mechanism. We reuse the desired performance register to
820 * store the real performance calculated by the platform.
821 */
822static unsigned int hisi_cppc_cpufreq_get_rate(unsigned int cpu)
823{
824 struct cpufreq_policy *policy = cpufreq_cpu_get(cpu);
825 struct cppc_cpudata *cpu_data = policy->driver_data;
826 u64 desired_perf;
827 int ret;
828
829 cpufreq_cpu_put(policy);
830
831 ret = cppc_get_desired_perf(cpu, &desired_perf);
832 if (ret < 0)
833 return -EIO;
834
835 return cppc_perf_to_khz(&cpu_data->perf_caps, desired_perf);
836}
837
838static void cppc_check_hisi_workaround(void)
839{
840 struct acpi_table_header *tbl;
841 acpi_status status = AE_OK;
842 int i;
843
844 status = acpi_get_table(ACPI_SIG_PCCT, 0, &tbl);
845 if (ACPI_FAILURE(status) || !tbl)
846 return;
847
848 for (i = 0; i < ARRAY_SIZE(wa_info); i++) {
849 if (!memcmp(wa_info[i].oem_id, tbl->oem_id, ACPI_OEM_ID_SIZE) &&
850 !memcmp(wa_info[i].oem_table_id, tbl->oem_table_id, ACPI_OEM_TABLE_ID_SIZE) &&
851 wa_info[i].oem_revision == tbl->oem_revision) {
852 /* Overwrite the get() callback */
853 cppc_cpufreq_driver.get = hisi_cppc_cpufreq_get_rate;
854 fie_disabled = FIE_DISABLED;
855 break;
856 }
857 }
858
859 acpi_put_table(tbl);
860}
861
862static int __init cppc_cpufreq_init(void)
863{
864 int ret;
865
866 if (!acpi_cpc_valid())
867 return -ENODEV;
868
869 cppc_check_hisi_workaround();
870 cppc_freq_invariance_init();
871 populate_efficiency_class();
872
873 ret = cpufreq_register_driver(&cppc_cpufreq_driver);
874 if (ret)
875 cppc_freq_invariance_exit();
876
877 return ret;
878}
879
880static inline void free_cpu_data(void)
881{
882 struct cppc_cpudata *iter, *tmp;
883
884 list_for_each_entry_safe(iter, tmp, &cpu_data_list, node) {
885 free_cpumask_var(iter->shared_cpu_map);
886 list_del(&iter->node);
887 kfree(iter);
888 }
889
890}
891
892static void __exit cppc_cpufreq_exit(void)
893{
894 cpufreq_unregister_driver(&cppc_cpufreq_driver);
895 cppc_freq_invariance_exit();
896
897 free_cpu_data();
898}
899
900module_exit(cppc_cpufreq_exit);
901MODULE_AUTHOR("Ashwin Chaugule");
902MODULE_DESCRIPTION("CPUFreq driver based on the ACPI CPPC v5.0+ spec");
903MODULE_LICENSE("GPL");
904
905late_initcall(cppc_cpufreq_init);
906
907static const struct acpi_device_id cppc_acpi_ids[] __used = {
908 {ACPI_PROCESSOR_DEVICE_HID, },
909 {}
910};
911
912MODULE_DEVICE_TABLE(acpi, cppc_acpi_ids);
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 */
42static LIST_HEAD(cpu_data_list);
43
44static bool boost_supported;
45
46struct 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
52static 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#ifdef CONFIG_ACPI_CPPC_CPUFREQ_FIE
65
66/* Frequency invariance support */
67struct cppc_freq_invariance {
68 int cpu;
69 struct irq_work irq_work;
70 struct kthread_work work;
71 struct cppc_perf_fb_ctrs prev_perf_fb_ctrs;
72 struct cppc_cpudata *cpu_data;
73};
74
75static DEFINE_PER_CPU(struct cppc_freq_invariance, cppc_freq_inv);
76static struct kthread_worker *kworker_fie;
77
78static struct cpufreq_driver cppc_cpufreq_driver;
79static unsigned int hisi_cppc_cpufreq_get_rate(unsigned int cpu);
80static int cppc_perf_from_fbctrs(struct cppc_cpudata *cpu_data,
81 struct cppc_perf_fb_ctrs *fb_ctrs_t0,
82 struct cppc_perf_fb_ctrs *fb_ctrs_t1);
83
84/**
85 * cppc_scale_freq_workfn - CPPC arch_freq_scale updater for frequency invariance
86 * @work: The work item.
87 *
88 * The CPPC driver register itself with the topology core to provide its own
89 * implementation (cppc_scale_freq_tick()) of topology_scale_freq_tick() which
90 * gets called by the scheduler on every tick.
91 *
92 * Note that the arch specific counters have higher priority than CPPC counters,
93 * if available, though the CPPC driver doesn't need to have any special
94 * handling for that.
95 *
96 * On an invocation of cppc_scale_freq_tick(), we schedule an irq work (since we
97 * reach here from hard-irq context), which then schedules a normal work item
98 * and cppc_scale_freq_workfn() updates the per_cpu arch_freq_scale variable
99 * based on the counter updates since the last tick.
100 */
101static void cppc_scale_freq_workfn(struct kthread_work *work)
102{
103 struct cppc_freq_invariance *cppc_fi;
104 struct cppc_perf_fb_ctrs fb_ctrs = {0};
105 struct cppc_cpudata *cpu_data;
106 unsigned long local_freq_scale;
107 u64 perf;
108
109 cppc_fi = container_of(work, struct cppc_freq_invariance, work);
110 cpu_data = cppc_fi->cpu_data;
111
112 if (cppc_get_perf_ctrs(cppc_fi->cpu, &fb_ctrs)) {
113 pr_warn("%s: failed to read perf counters\n", __func__);
114 return;
115 }
116
117 perf = cppc_perf_from_fbctrs(cpu_data, &cppc_fi->prev_perf_fb_ctrs,
118 &fb_ctrs);
119 cppc_fi->prev_perf_fb_ctrs = fb_ctrs;
120
121 perf <<= SCHED_CAPACITY_SHIFT;
122 local_freq_scale = div64_u64(perf, cpu_data->perf_caps.highest_perf);
123
124 /* This can happen due to counter's overflow */
125 if (unlikely(local_freq_scale > 1024))
126 local_freq_scale = 1024;
127
128 per_cpu(arch_freq_scale, cppc_fi->cpu) = local_freq_scale;
129}
130
131static void cppc_irq_work(struct irq_work *irq_work)
132{
133 struct cppc_freq_invariance *cppc_fi;
134
135 cppc_fi = container_of(irq_work, struct cppc_freq_invariance, irq_work);
136 kthread_queue_work(kworker_fie, &cppc_fi->work);
137}
138
139static void cppc_scale_freq_tick(void)
140{
141 struct cppc_freq_invariance *cppc_fi = &per_cpu(cppc_freq_inv, smp_processor_id());
142
143 /*
144 * cppc_get_perf_ctrs() can potentially sleep, call that from the right
145 * context.
146 */
147 irq_work_queue(&cppc_fi->irq_work);
148}
149
150static struct scale_freq_data cppc_sftd = {
151 .source = SCALE_FREQ_SOURCE_CPPC,
152 .set_freq_scale = cppc_scale_freq_tick,
153};
154
155static void cppc_cpufreq_cpu_fie_init(struct cpufreq_policy *policy)
156{
157 struct cppc_freq_invariance *cppc_fi;
158 int cpu, ret;
159
160 if (cppc_cpufreq_driver.get == hisi_cppc_cpufreq_get_rate)
161 return;
162
163 for_each_cpu(cpu, policy->cpus) {
164 cppc_fi = &per_cpu(cppc_freq_inv, cpu);
165 cppc_fi->cpu = cpu;
166 cppc_fi->cpu_data = policy->driver_data;
167 kthread_init_work(&cppc_fi->work, cppc_scale_freq_workfn);
168 init_irq_work(&cppc_fi->irq_work, cppc_irq_work);
169
170 ret = cppc_get_perf_ctrs(cpu, &cppc_fi->prev_perf_fb_ctrs);
171 if (ret) {
172 pr_warn("%s: failed to read perf counters for cpu:%d: %d\n",
173 __func__, cpu, ret);
174
175 /*
176 * Don't abort if the CPU was offline while the driver
177 * was getting registered.
178 */
179 if (cpu_online(cpu))
180 return;
181 }
182 }
183
184 /* Register for freq-invariance */
185 topology_set_scale_freq_source(&cppc_sftd, policy->cpus);
186}
187
188/*
189 * We free all the resources on policy's removal and not on CPU removal as the
190 * irq-work are per-cpu and the hotplug core takes care of flushing the pending
191 * irq-works (hint: smpcfd_dying_cpu()) on CPU hotplug. Even if the kthread-work
192 * fires on another CPU after the concerned CPU is removed, it won't harm.
193 *
194 * We just need to make sure to remove them all on policy->exit().
195 */
196static void cppc_cpufreq_cpu_fie_exit(struct cpufreq_policy *policy)
197{
198 struct cppc_freq_invariance *cppc_fi;
199 int cpu;
200
201 if (cppc_cpufreq_driver.get == hisi_cppc_cpufreq_get_rate)
202 return;
203
204 /* policy->cpus will be empty here, use related_cpus instead */
205 topology_clear_scale_freq_source(SCALE_FREQ_SOURCE_CPPC, policy->related_cpus);
206
207 for_each_cpu(cpu, policy->related_cpus) {
208 cppc_fi = &per_cpu(cppc_freq_inv, cpu);
209 irq_work_sync(&cppc_fi->irq_work);
210 kthread_cancel_work_sync(&cppc_fi->work);
211 }
212}
213
214static void __init cppc_freq_invariance_init(void)
215{
216 struct sched_attr attr = {
217 .size = sizeof(struct sched_attr),
218 .sched_policy = SCHED_DEADLINE,
219 .sched_nice = 0,
220 .sched_priority = 0,
221 /*
222 * Fake (unused) bandwidth; workaround to "fix"
223 * priority inheritance.
224 */
225 .sched_runtime = 1000000,
226 .sched_deadline = 10000000,
227 .sched_period = 10000000,
228 };
229 int ret;
230
231 if (cppc_cpufreq_driver.get == hisi_cppc_cpufreq_get_rate)
232 return;
233
234 kworker_fie = kthread_create_worker(0, "cppc_fie");
235 if (IS_ERR(kworker_fie))
236 return;
237
238 ret = sched_setattr_nocheck(kworker_fie->task, &attr);
239 if (ret) {
240 pr_warn("%s: failed to set SCHED_DEADLINE: %d\n", __func__,
241 ret);
242 kthread_destroy_worker(kworker_fie);
243 return;
244 }
245}
246
247static void cppc_freq_invariance_exit(void)
248{
249 if (cppc_cpufreq_driver.get == hisi_cppc_cpufreq_get_rate)
250 return;
251
252 kthread_destroy_worker(kworker_fie);
253 kworker_fie = NULL;
254}
255
256#else
257static inline void cppc_cpufreq_cpu_fie_init(struct cpufreq_policy *policy)
258{
259}
260
261static inline void cppc_cpufreq_cpu_fie_exit(struct cpufreq_policy *policy)
262{
263}
264
265static inline void cppc_freq_invariance_init(void)
266{
267}
268
269static inline void cppc_freq_invariance_exit(void)
270{
271}
272#endif /* CONFIG_ACPI_CPPC_CPUFREQ_FIE */
273
274/* Callback function used to retrieve the max frequency from DMI */
275static void cppc_find_dmi_mhz(const struct dmi_header *dm, void *private)
276{
277 const u8 *dmi_data = (const u8 *)dm;
278 u16 *mhz = (u16 *)private;
279
280 if (dm->type == DMI_ENTRY_PROCESSOR &&
281 dm->length >= DMI_ENTRY_PROCESSOR_MIN_LENGTH) {
282 u16 val = (u16)get_unaligned((const u16 *)
283 (dmi_data + DMI_PROCESSOR_MAX_SPEED));
284 *mhz = val > *mhz ? val : *mhz;
285 }
286}
287
288/* Look up the max frequency in DMI */
289static u64 cppc_get_dmi_max_khz(void)
290{
291 u16 mhz = 0;
292
293 dmi_walk(cppc_find_dmi_mhz, &mhz);
294
295 /*
296 * Real stupid fallback value, just in case there is no
297 * actual value set.
298 */
299 mhz = mhz ? mhz : 1;
300
301 return (1000 * mhz);
302}
303
304/*
305 * If CPPC lowest_freq and nominal_freq registers are exposed then we can
306 * use them to convert perf to freq and vice versa
307 *
308 * If the perf/freq point lies between Nominal and Lowest, we can treat
309 * (Low perf, Low freq) and (Nom Perf, Nom freq) as 2D co-ordinates of a line
310 * and extrapolate the rest
311 * For perf/freq > Nominal, we use the ratio perf:freq at Nominal for conversion
312 */
313static unsigned int cppc_cpufreq_perf_to_khz(struct cppc_cpudata *cpu_data,
314 unsigned int perf)
315{
316 struct cppc_perf_caps *caps = &cpu_data->perf_caps;
317 static u64 max_khz;
318 u64 mul, div;
319
320 if (caps->lowest_freq && caps->nominal_freq) {
321 if (perf >= caps->nominal_perf) {
322 mul = caps->nominal_freq;
323 div = caps->nominal_perf;
324 } else {
325 mul = caps->nominal_freq - caps->lowest_freq;
326 div = caps->nominal_perf - caps->lowest_perf;
327 }
328 } else {
329 if (!max_khz)
330 max_khz = cppc_get_dmi_max_khz();
331 mul = max_khz;
332 div = caps->highest_perf;
333 }
334 return (u64)perf * mul / div;
335}
336
337static unsigned int cppc_cpufreq_khz_to_perf(struct cppc_cpudata *cpu_data,
338 unsigned int freq)
339{
340 struct cppc_perf_caps *caps = &cpu_data->perf_caps;
341 static u64 max_khz;
342 u64 mul, div;
343
344 if (caps->lowest_freq && caps->nominal_freq) {
345 if (freq >= caps->nominal_freq) {
346 mul = caps->nominal_perf;
347 div = caps->nominal_freq;
348 } else {
349 mul = caps->lowest_perf;
350 div = caps->lowest_freq;
351 }
352 } else {
353 if (!max_khz)
354 max_khz = cppc_get_dmi_max_khz();
355 mul = caps->highest_perf;
356 div = max_khz;
357 }
358
359 return (u64)freq * mul / div;
360}
361
362static int cppc_cpufreq_set_target(struct cpufreq_policy *policy,
363 unsigned int target_freq,
364 unsigned int relation)
365
366{
367 struct cppc_cpudata *cpu_data = policy->driver_data;
368 unsigned int cpu = policy->cpu;
369 struct cpufreq_freqs freqs;
370 u32 desired_perf;
371 int ret = 0;
372
373 desired_perf = cppc_cpufreq_khz_to_perf(cpu_data, target_freq);
374 /* Return if it is exactly the same perf */
375 if (desired_perf == cpu_data->perf_ctrls.desired_perf)
376 return ret;
377
378 cpu_data->perf_ctrls.desired_perf = desired_perf;
379 freqs.old = policy->cur;
380 freqs.new = target_freq;
381
382 cpufreq_freq_transition_begin(policy, &freqs);
383 ret = cppc_set_perf(cpu, &cpu_data->perf_ctrls);
384 cpufreq_freq_transition_end(policy, &freqs, ret != 0);
385
386 if (ret)
387 pr_debug("Failed to set target on CPU:%d. ret:%d\n",
388 cpu, ret);
389
390 return ret;
391}
392
393static int cppc_verify_policy(struct cpufreq_policy_data *policy)
394{
395 cpufreq_verify_within_cpu_limits(policy);
396 return 0;
397}
398
399/*
400 * The PCC subspace describes the rate at which platform can accept commands
401 * on the shared PCC channel (including READs which do not count towards freq
402 * transition requests), so ideally we need to use the PCC values as a fallback
403 * if we don't have a platform specific transition_delay_us
404 */
405#ifdef CONFIG_ARM64
406#include <asm/cputype.h>
407
408static unsigned int cppc_cpufreq_get_transition_delay_us(unsigned int cpu)
409{
410 unsigned long implementor = read_cpuid_implementor();
411 unsigned long part_num = read_cpuid_part_number();
412
413 switch (implementor) {
414 case ARM_CPU_IMP_QCOM:
415 switch (part_num) {
416 case QCOM_CPU_PART_FALKOR_V1:
417 case QCOM_CPU_PART_FALKOR:
418 return 10000;
419 }
420 }
421 return cppc_get_transition_latency(cpu) / NSEC_PER_USEC;
422}
423
424#else
425
426static unsigned int cppc_cpufreq_get_transition_delay_us(unsigned int cpu)
427{
428 return cppc_get_transition_latency(cpu) / NSEC_PER_USEC;
429}
430#endif
431
432
433static struct cppc_cpudata *cppc_cpufreq_get_cpu_data(unsigned int cpu)
434{
435 struct cppc_cpudata *cpu_data;
436 int ret;
437
438 cpu_data = kzalloc(sizeof(struct cppc_cpudata), GFP_KERNEL);
439 if (!cpu_data)
440 goto out;
441
442 if (!zalloc_cpumask_var(&cpu_data->shared_cpu_map, GFP_KERNEL))
443 goto free_cpu;
444
445 ret = acpi_get_psd_map(cpu, cpu_data);
446 if (ret) {
447 pr_debug("Err parsing CPU%d PSD data: ret:%d\n", cpu, ret);
448 goto free_mask;
449 }
450
451 ret = cppc_get_perf_caps(cpu, &cpu_data->perf_caps);
452 if (ret) {
453 pr_debug("Err reading CPU%d perf caps: ret:%d\n", cpu, ret);
454 goto free_mask;
455 }
456
457 /* Convert the lowest and nominal freq from MHz to KHz */
458 cpu_data->perf_caps.lowest_freq *= 1000;
459 cpu_data->perf_caps.nominal_freq *= 1000;
460
461 list_add(&cpu_data->node, &cpu_data_list);
462
463 return cpu_data;
464
465free_mask:
466 free_cpumask_var(cpu_data->shared_cpu_map);
467free_cpu:
468 kfree(cpu_data);
469out:
470 return NULL;
471}
472
473static void cppc_cpufreq_put_cpu_data(struct cpufreq_policy *policy)
474{
475 struct cppc_cpudata *cpu_data = policy->driver_data;
476
477 list_del(&cpu_data->node);
478 free_cpumask_var(cpu_data->shared_cpu_map);
479 kfree(cpu_data);
480 policy->driver_data = NULL;
481}
482
483static int cppc_cpufreq_cpu_init(struct cpufreq_policy *policy)
484{
485 unsigned int cpu = policy->cpu;
486 struct cppc_cpudata *cpu_data;
487 struct cppc_perf_caps *caps;
488 int ret;
489
490 cpu_data = cppc_cpufreq_get_cpu_data(cpu);
491 if (!cpu_data) {
492 pr_err("Error in acquiring _CPC/_PSD data for CPU%d.\n", cpu);
493 return -ENODEV;
494 }
495 caps = &cpu_data->perf_caps;
496 policy->driver_data = cpu_data;
497
498 /*
499 * Set min to lowest nonlinear perf to avoid any efficiency penalty (see
500 * Section 8.4.7.1.1.5 of ACPI 6.1 spec)
501 */
502 policy->min = cppc_cpufreq_perf_to_khz(cpu_data,
503 caps->lowest_nonlinear_perf);
504 policy->max = cppc_cpufreq_perf_to_khz(cpu_data,
505 caps->nominal_perf);
506
507 /*
508 * Set cpuinfo.min_freq to Lowest to make the full range of performance
509 * available if userspace wants to use any perf between lowest & lowest
510 * nonlinear perf
511 */
512 policy->cpuinfo.min_freq = cppc_cpufreq_perf_to_khz(cpu_data,
513 caps->lowest_perf);
514 policy->cpuinfo.max_freq = cppc_cpufreq_perf_to_khz(cpu_data,
515 caps->nominal_perf);
516
517 policy->transition_delay_us = cppc_cpufreq_get_transition_delay_us(cpu);
518 policy->shared_type = cpu_data->shared_type;
519
520 switch (policy->shared_type) {
521 case CPUFREQ_SHARED_TYPE_HW:
522 case CPUFREQ_SHARED_TYPE_NONE:
523 /* Nothing to be done - we'll have a policy for each CPU */
524 break;
525 case CPUFREQ_SHARED_TYPE_ANY:
526 /*
527 * All CPUs in the domain will share a policy and all cpufreq
528 * operations will use a single cppc_cpudata structure stored
529 * in policy->driver_data.
530 */
531 cpumask_copy(policy->cpus, cpu_data->shared_cpu_map);
532 break;
533 default:
534 pr_debug("Unsupported CPU co-ord type: %d\n",
535 policy->shared_type);
536 ret = -EFAULT;
537 goto out;
538 }
539
540 /*
541 * If 'highest_perf' is greater than 'nominal_perf', we assume CPU Boost
542 * is supported.
543 */
544 if (caps->highest_perf > caps->nominal_perf)
545 boost_supported = true;
546
547 /* Set policy->cur to max now. The governors will adjust later. */
548 policy->cur = cppc_cpufreq_perf_to_khz(cpu_data, caps->highest_perf);
549 cpu_data->perf_ctrls.desired_perf = caps->highest_perf;
550
551 ret = cppc_set_perf(cpu, &cpu_data->perf_ctrls);
552 if (ret) {
553 pr_debug("Err setting perf value:%d on CPU:%d. ret:%d\n",
554 caps->highest_perf, cpu, ret);
555 goto out;
556 }
557
558 cppc_cpufreq_cpu_fie_init(policy);
559 return 0;
560
561out:
562 cppc_cpufreq_put_cpu_data(policy);
563 return ret;
564}
565
566static int cppc_cpufreq_cpu_exit(struct cpufreq_policy *policy)
567{
568 struct cppc_cpudata *cpu_data = policy->driver_data;
569 struct cppc_perf_caps *caps = &cpu_data->perf_caps;
570 unsigned int cpu = policy->cpu;
571 int ret;
572
573 cppc_cpufreq_cpu_fie_exit(policy);
574
575 cpu_data->perf_ctrls.desired_perf = caps->lowest_perf;
576
577 ret = cppc_set_perf(cpu, &cpu_data->perf_ctrls);
578 if (ret)
579 pr_debug("Err setting perf value:%d on CPU:%d. ret:%d\n",
580 caps->lowest_perf, cpu, ret);
581
582 cppc_cpufreq_put_cpu_data(policy);
583 return 0;
584}
585
586static inline u64 get_delta(u64 t1, u64 t0)
587{
588 if (t1 > t0 || t0 > ~(u32)0)
589 return t1 - t0;
590
591 return (u32)t1 - (u32)t0;
592}
593
594static int cppc_perf_from_fbctrs(struct cppc_cpudata *cpu_data,
595 struct cppc_perf_fb_ctrs *fb_ctrs_t0,
596 struct cppc_perf_fb_ctrs *fb_ctrs_t1)
597{
598 u64 delta_reference, delta_delivered;
599 u64 reference_perf;
600
601 reference_perf = fb_ctrs_t0->reference_perf;
602
603 delta_reference = get_delta(fb_ctrs_t1->reference,
604 fb_ctrs_t0->reference);
605 delta_delivered = get_delta(fb_ctrs_t1->delivered,
606 fb_ctrs_t0->delivered);
607
608 /* Check to avoid divide-by zero and invalid delivered_perf */
609 if (!delta_reference || !delta_delivered)
610 return cpu_data->perf_ctrls.desired_perf;
611
612 return (reference_perf * delta_delivered) / delta_reference;
613}
614
615static unsigned int cppc_cpufreq_get_rate(unsigned int cpu)
616{
617 struct cppc_perf_fb_ctrs fb_ctrs_t0 = {0}, fb_ctrs_t1 = {0};
618 struct cpufreq_policy *policy = cpufreq_cpu_get(cpu);
619 struct cppc_cpudata *cpu_data = policy->driver_data;
620 u64 delivered_perf;
621 int ret;
622
623 cpufreq_cpu_put(policy);
624
625 ret = cppc_get_perf_ctrs(cpu, &fb_ctrs_t0);
626 if (ret)
627 return ret;
628
629 udelay(2); /* 2usec delay between sampling */
630
631 ret = cppc_get_perf_ctrs(cpu, &fb_ctrs_t1);
632 if (ret)
633 return ret;
634
635 delivered_perf = cppc_perf_from_fbctrs(cpu_data, &fb_ctrs_t0,
636 &fb_ctrs_t1);
637
638 return cppc_cpufreq_perf_to_khz(cpu_data, delivered_perf);
639}
640
641static int cppc_cpufreq_set_boost(struct cpufreq_policy *policy, int state)
642{
643 struct cppc_cpudata *cpu_data = policy->driver_data;
644 struct cppc_perf_caps *caps = &cpu_data->perf_caps;
645 int ret;
646
647 if (!boost_supported) {
648 pr_err("BOOST not supported by CPU or firmware\n");
649 return -EINVAL;
650 }
651
652 if (state)
653 policy->max = cppc_cpufreq_perf_to_khz(cpu_data,
654 caps->highest_perf);
655 else
656 policy->max = cppc_cpufreq_perf_to_khz(cpu_data,
657 caps->nominal_perf);
658 policy->cpuinfo.max_freq = policy->max;
659
660 ret = freq_qos_update_request(policy->max_freq_req, policy->max);
661 if (ret < 0)
662 return ret;
663
664 return 0;
665}
666
667static ssize_t show_freqdomain_cpus(struct cpufreq_policy *policy, char *buf)
668{
669 struct cppc_cpudata *cpu_data = policy->driver_data;
670
671 return cpufreq_show_cpus(cpu_data->shared_cpu_map, buf);
672}
673cpufreq_freq_attr_ro(freqdomain_cpus);
674
675static struct freq_attr *cppc_cpufreq_attr[] = {
676 &freqdomain_cpus,
677 NULL,
678};
679
680static struct cpufreq_driver cppc_cpufreq_driver = {
681 .flags = CPUFREQ_CONST_LOOPS,
682 .verify = cppc_verify_policy,
683 .target = cppc_cpufreq_set_target,
684 .get = cppc_cpufreq_get_rate,
685 .init = cppc_cpufreq_cpu_init,
686 .exit = cppc_cpufreq_cpu_exit,
687 .set_boost = cppc_cpufreq_set_boost,
688 .attr = cppc_cpufreq_attr,
689 .name = "cppc_cpufreq",
690};
691
692/*
693 * HISI platform does not support delivered performance counter and
694 * reference performance counter. It can calculate the performance using the
695 * platform specific mechanism. We reuse the desired performance register to
696 * store the real performance calculated by the platform.
697 */
698static unsigned int hisi_cppc_cpufreq_get_rate(unsigned int cpu)
699{
700 struct cpufreq_policy *policy = cpufreq_cpu_get(cpu);
701 struct cppc_cpudata *cpu_data = policy->driver_data;
702 u64 desired_perf;
703 int ret;
704
705 cpufreq_cpu_put(policy);
706
707 ret = cppc_get_desired_perf(cpu, &desired_perf);
708 if (ret < 0)
709 return -EIO;
710
711 return cppc_cpufreq_perf_to_khz(cpu_data, desired_perf);
712}
713
714static void cppc_check_hisi_workaround(void)
715{
716 struct acpi_table_header *tbl;
717 acpi_status status = AE_OK;
718 int i;
719
720 status = acpi_get_table(ACPI_SIG_PCCT, 0, &tbl);
721 if (ACPI_FAILURE(status) || !tbl)
722 return;
723
724 for (i = 0; i < ARRAY_SIZE(wa_info); i++) {
725 if (!memcmp(wa_info[i].oem_id, tbl->oem_id, ACPI_OEM_ID_SIZE) &&
726 !memcmp(wa_info[i].oem_table_id, tbl->oem_table_id, ACPI_OEM_TABLE_ID_SIZE) &&
727 wa_info[i].oem_revision == tbl->oem_revision) {
728 /* Overwrite the get() callback */
729 cppc_cpufreq_driver.get = hisi_cppc_cpufreq_get_rate;
730 break;
731 }
732 }
733
734 acpi_put_table(tbl);
735}
736
737static int __init cppc_cpufreq_init(void)
738{
739 int ret;
740
741 if ((acpi_disabled) || !acpi_cpc_valid())
742 return -ENODEV;
743
744 INIT_LIST_HEAD(&cpu_data_list);
745
746 cppc_check_hisi_workaround();
747 cppc_freq_invariance_init();
748
749 ret = cpufreq_register_driver(&cppc_cpufreq_driver);
750 if (ret)
751 cppc_freq_invariance_exit();
752
753 return ret;
754}
755
756static inline void free_cpu_data(void)
757{
758 struct cppc_cpudata *iter, *tmp;
759
760 list_for_each_entry_safe(iter, tmp, &cpu_data_list, node) {
761 free_cpumask_var(iter->shared_cpu_map);
762 list_del(&iter->node);
763 kfree(iter);
764 }
765
766}
767
768static void __exit cppc_cpufreq_exit(void)
769{
770 cpufreq_unregister_driver(&cppc_cpufreq_driver);
771 cppc_freq_invariance_exit();
772
773 free_cpu_data();
774}
775
776module_exit(cppc_cpufreq_exit);
777MODULE_AUTHOR("Ashwin Chaugule");
778MODULE_DESCRIPTION("CPUFreq driver based on the ACPI CPPC v5.0+ spec");
779MODULE_LICENSE("GPL");
780
781late_initcall(cppc_cpufreq_init);
782
783static const struct acpi_device_id cppc_acpi_ids[] __used = {
784 {ACPI_PROCESSOR_DEVICE_HID, },
785 {}
786};
787
788MODULE_DEVICE_TABLE(acpi, cppc_acpi_ids);