linux/drivers/cpufreq/cppc_cpufreq.c
<<
>>
Prefs
   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);
 789