drivers/cpufreq/cppc_cpufreq.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * CPPC (Collaborative Processor Performance Control) driver for
   4  * interfacing with the CPUfreq layer and governors. See
   5  * cppc_acpi.c for CPPC specific methods.
   6  *
   7  * (C) Copyright 2014, 2015 Linaro Ltd.
   8  * Author: Ashwin Chaugule <ashwin.chaugule@linaro.org>
   9  */
  10
  11 #define pr_fmt(fmt)     "CPPC Cpufreq:" fmt
  12
  13 #include <linux/kernel.h>
  14 #include <linux/module.h>
  15 #include <linux/delay.h>
  16 #include <linux/cpu.h>
  17 #include <linux/cpufreq.h>
  18 #include <linux/dmi.h>
  19 #include <linux/time.h>
  20 #include <linux/vmalloc.h>
  21
  22 #include <asm/unaligned.h>
  23
  24 #include <acpi/cppc_acpi.h>
  25
  26 /* Minimum struct length needed for the DMI processor entry we want */
  27 #define DMI_ENTRY_PROCESSOR_MIN_LENGTH  48
  28
  29 /* Offset in the DMI processor structure for the max frequency */
  30 #define DMI_PROCESSOR_MAX_SPEED         0x14
  31
  32 /*
  33  * These structs contain information parsed from per CPU
  34  * ACPI _CPC structures.
  35  * e.g. For each CPU the highest, lowest supported
  36  * performance capabilities, desired performance level
  37  * requested etc.
  38  */
  39 static struct cppc_cpudata **all_cpu_data;
  40 static bool boost_supported;
  41
  42 struct cppc_workaround_oem_info {
  43         char oem_id[ACPI_OEM_ID_SIZE + 1];
  44         char oem_table_id[ACPI_OEM_TABLE_ID_SIZE + 1];
  45         u32 oem_revision;
  46 };
  47
  48 static struct cppc_workaround_oem_info wa_info[] = {
  49         {
  50                 .oem_id         = "HISI  ",
  51                 .oem_table_id   = "HIP07   ",
  52                 .oem_revision   = 0,
  53         }, {
  54                 .oem_id         = "HISI  ",
  55                 .oem_table_id   = "HIP08   ",
  56                 .oem_revision   = 0,
  57         }
  58 };
  59
  60 /* Callback function used to retrieve the max frequency from DMI */
  61 static void cppc_find_dmi_mhz(const struct dmi_header *dm, void *private)
  62 {
  63         const u8 *dmi_data = (const u8 *)dm;
  64         u16 *mhz = (u16 *)private;
  65
  66         if (dm->type == DMI_ENTRY_PROCESSOR &&
  67             dm->length >= DMI_ENTRY_PROCESSOR_MIN_LENGTH) {
  68                 u16 val = (u16)get_unaligned((const u16 *)
  69                                 (dmi_data + DMI_PROCESSOR_MAX_SPEED));
  70                 *mhz = val > *mhz ? val : *mhz;
  71         }
  72 }
  73
  74 /* Look up the max frequency in DMI */
  75 static u64 cppc_get_dmi_max_khz(void)
  76 {
  77         u16 mhz = 0;
  78
  79         dmi_walk(cppc_find_dmi_mhz, &mhz);
  80
  81         /*
  82          * Real stupid fallback value, just in case there is no
  83          * actual value set.
  84          */
  85         mhz = mhz ? mhz : 1;
  86
  87         return (1000 * mhz);
  88 }
  89
  90 /*
  91  * If CPPC lowest_freq and nominal_freq registers are exposed then we can
  92  * use them to convert perf to freq and vice versa
  93  *
  94  * If the perf/freq point lies between Nominal and Lowest, we can treat
  95  * (Low perf, Low freq) and (Nom Perf, Nom freq) as 2D co-ordinates of a line
  96  * and extrapolate the rest
  97  * For perf/freq > Nominal, we use the ratio perf:freq at Nominal for conversion
  98  */
  99 static unsigned int cppc_cpufreq_perf_to_khz(struct cppc_cpudata *cpu_data,
 100                                              unsigned int perf)
 101 {
 102         struct cppc_perf_caps *caps = &cpu_data->perf_caps;
 103         static u64 max_khz;
 104         u64 mul, div;
 105
 106         if (caps->lowest_freq && caps->nominal_freq) {
 107                 if (perf >= caps->nominal_perf) {
 108                         mul = caps->nominal_freq;
 109                         div = caps->nominal_perf;
 110                 } else {
 111                         mul = caps->nominal_freq - caps->lowest_freq;
 112                         div = caps->nominal_perf - caps->lowest_perf;
 113                 }
 114         } else {
 115                 if (!max_khz)
 116                         max_khz = cppc_get_dmi_max_khz();
 117                 mul = max_khz;
 118                 div = caps->highest_perf;
 119         }
 120         return (u64)perf * mul / div;
 121 }
 122
 123 static unsigned int cppc_cpufreq_khz_to_perf(struct cppc_cpudata *cpu_data,
 124                                              unsigned int freq)
 125 {
 126         struct cppc_perf_caps *caps = &cpu_data->perf_caps;
 127         static u64 max_khz;
 128         u64  mul, div;
 129
 130         if (caps->lowest_freq && caps->nominal_freq) {
 131                 if (freq >= caps->nominal_freq) {
 132                         mul = caps->nominal_perf;
 133                         div = caps->nominal_freq;
 134                 } else {
 135                         mul = caps->lowest_perf;
 136                         div = caps->lowest_freq;
 137                 }
 138         } else {
 139                 if (!max_khz)
 140                         max_khz = cppc_get_dmi_max_khz();
 141                 mul = caps->highest_perf;
 142                 div = max_khz;
 143         }
 144
 145         return (u64)freq * mul / div;
 146 }
 147
 148 static int cppc_cpufreq_set_target(struct cpufreq_policy *policy,
 149                                    unsigned int target_freq,
 150                                    unsigned int relation)
 151 {
 152         struct cppc_cpudata *cpu_data = all_cpu_data[policy->cpu];
 153         struct cpufreq_freqs freqs;
 154         u32 desired_perf;
 155         int ret = 0;
 156
 157         desired_perf = cppc_cpufreq_khz_to_perf(cpu_data, target_freq);
 158         /* Return if it is exactly the same perf */
 159         if (desired_perf == cpu_data->perf_ctrls.desired_perf)
 160                 return ret;
 161
 162         cpu_data->perf_ctrls.desired_perf = desired_perf;
 163         freqs.old = policy->cur;
 164         freqs.new = target_freq;
 165
 166         cpufreq_freq_transition_begin(policy, &freqs);
 167         ret = cppc_set_perf(cpu_data->cpu, &cpu_data->perf_ctrls);
 168         cpufreq_freq_transition_end(policy, &freqs, ret != 0);
 169
 170         if (ret)
 171                 pr_debug("Failed to set target on CPU:%d. ret:%d\n",
 172                          cpu_data->cpu, ret);
 173
 174         return ret;
 175 }
 176
 177 static int cppc_verify_policy(struct cpufreq_policy_data *policy)
 178 {
 179         cpufreq_verify_within_cpu_limits(policy);
 180         return 0;
 181 }
 182
 183 static void cppc_cpufreq_stop_cpu(struct cpufreq_policy *policy)
 184 {
 185         struct cppc_cpudata *cpu_data = all_cpu_data[policy->cpu];
 186         struct cppc_perf_caps *caps = &cpu_data->perf_caps;
 187         unsigned int cpu = policy->cpu;
 188         int ret;
 189
 190         cpu_data->perf_ctrls.desired_perf = caps->lowest_perf;
 191
 192         ret = cppc_set_perf(cpu, &cpu_data->perf_ctrls);
 193         if (ret)
 194                 pr_debug("Err setting perf value:%d on CPU:%d. ret:%d\n",
 195                          caps->lowest_perf, cpu, ret);
 196 }
 197
 198 /*
 199  * The PCC subspace describes the rate at which platform can accept commands
 200  * on the shared PCC channel (including READs which do not count towards freq
 201  * transition requests), so ideally we need to use the PCC values as a fallback
 202  * if we don't have a platform specific transition_delay_us
 203  */
 204 #ifdef CONFIG_ARM64
 205 #include <asm/cputype.h>
 206
 207 static unsigned int cppc_cpufreq_get_transition_delay_us(unsigned int cpu)
 208 {
 209         unsigned long implementor = read_cpuid_implementor();
 210         unsigned long part_num = read_cpuid_part_number();
 211         unsigned int delay_us = 0;
 212
 213         switch (implementor) {
 214         case ARM_CPU_IMP_QCOM:
 215                 switch (part_num) {
 216                 case QCOM_CPU_PART_FALKOR_V1:
 217                 case QCOM_CPU_PART_FALKOR:
 218                         delay_us = 10000;
 219                         break;
 220                 default:
 221                         delay_us = cppc_get_transition_latency(cpu) / NSEC_PER_USEC;
 222                         break;
 223                 }
 224                 break;
 225         default:
 226                 delay_us = cppc_get_transition_latency(cpu) / NSEC_PER_USEC;
 227                 break;
 228         }
 229
 230         return delay_us;
 231 }
 232
 233 #else
 234
 235 static unsigned int cppc_cpufreq_get_transition_delay_us(unsigned int cpu)
 236 {
 237         return cppc_get_transition_latency(cpu) / NSEC_PER_USEC;
 238 }
 239 #endif
 240
 241 static int cppc_cpufreq_cpu_init(struct cpufreq_policy *policy)
 242 {
 243         struct cppc_cpudata *cpu_data = all_cpu_data[policy->cpu];
 244         struct cppc_perf_caps *caps = &cpu_data->perf_caps;
 245         unsigned int cpu = policy->cpu;
 246         int ret = 0;
 247
 248         cpu_data->cpu = cpu;
 249         ret = cppc_get_perf_caps(cpu, caps);
 250
 251         if (ret) {
 252                 pr_debug("Err reading CPU%d perf capabilities. ret:%d\n",
 253                          cpu, ret);
 254                 return ret;
 255         }
 256
 257         /* Convert the lowest and nominal freq from MHz to KHz */
 258         caps->lowest_freq *= 1000;
 259         caps->nominal_freq *= 1000;
 260
 261         /*
 262          * Set min to lowest nonlinear perf to avoid any efficiency penalty (see
 263          * Section 8.4.7.1.1.5 of ACPI 6.1 spec)
 264          */
 265         policy->min = cppc_cpufreq_perf_to_khz(cpu_data,
 266                                                caps->lowest_nonlinear_perf);
 267         policy->max = cppc_cpufreq_perf_to_khz(cpu_data,
 268                                                caps->nominal_perf);
 269
 270         /*
 271          * Set cpuinfo.min_freq to Lowest to make the full range of performance
 272          * available if userspace wants to use any perf between lowest & lowest
 273          * nonlinear perf
 274          */
 275         policy->cpuinfo.min_freq = cppc_cpufreq_perf_to_khz(cpu_data,
 276                                                             caps->lowest_perf);
 277         policy->cpuinfo.max_freq = cppc_cpufreq_perf_to_khz(cpu_data,
 278                                                             caps->nominal_perf);
 279
 280         policy->transition_delay_us = cppc_cpufreq_get_transition_delay_us(cpu);
 281         policy->shared_type = cpu_data->shared_type;
 282
 283         if (policy->shared_type == CPUFREQ_SHARED_TYPE_ANY) {
 284                 int i;
 285
 286                 cpumask_copy(policy->cpus, cpu_data->shared_cpu_map);
 287
 288                 for_each_cpu(i, policy->cpus) {
 289                         if (unlikely(i == cpu))
 290                                 continue;
 291
 292                         memcpy(&all_cpu_data[i]->perf_caps, caps,
 293                                sizeof(cpu_data->perf_caps));
 294                 }
 295         } else if (policy->shared_type == CPUFREQ_SHARED_TYPE_ALL) {
 296                 /* Support only SW_ANY for now. */
 297                 pr_debug("Unsupported CPU co-ord type\n");
 298                 return -EFAULT;
 299         }
 300
 301         cpu_data->cur_policy = policy;
 302
 303         /*
 304          * If 'highest_perf' is greater than 'nominal_perf', we assume CPU Boost
 305          * is supported.
 306          */
 307         if (caps->highest_perf > caps->nominal_perf)
 308                 boost_supported = true;
 309
 310         /* Set policy->cur to max now. The governors will adjust later. */
 311         policy->cur = cppc_cpufreq_perf_to_khz(cpu_data, caps->highest_perf);
 312         cpu_data->perf_ctrls.desired_perf =  caps->highest_perf;
 313
 314         ret = cppc_set_perf(cpu, &cpu_data->perf_ctrls);
 315         if (ret)
 316                 pr_debug("Err setting perf value:%d on CPU:%d. ret:%d\n",
 317                          caps->highest_perf, cpu, ret);
 318
 319         return ret;
 320 }
 321
 322 static inline u64 get_delta(u64 t1, u64 t0)
 323 {
 324         if (t1 > t0 || t0 > ~(u32)0)
 325                 return t1 - t0;
 326
 327         return (u32)t1 - (u32)t0;
 328 }
 329
 330 static int cppc_get_rate_from_fbctrs(struct cppc_cpudata *cpu_data,
 331                                      struct cppc_perf_fb_ctrs fb_ctrs_t0,
 332                                      struct cppc_perf_fb_ctrs fb_ctrs_t1)
 333 {
 334         u64 delta_reference, delta_delivered;
 335         u64 reference_perf, delivered_perf;
 336
 337         reference_perf = fb_ctrs_t0.reference_perf;
 338
 339         delta_reference = get_delta(fb_ctrs_t1.reference,
 340                                     fb_ctrs_t0.reference);
 341         delta_delivered = get_delta(fb_ctrs_t1.delivered,
 342                                     fb_ctrs_t0.delivered);
 343
 344         /* Check to avoid divide-by zero */
 345         if (delta_reference || delta_delivered)
 346                 delivered_perf = (reference_perf * delta_delivered) /
 347                                         delta_reference;
 348         else
 349                 delivered_perf = cpu_data->perf_ctrls.desired_perf;
 350
 351         return cppc_cpufreq_perf_to_khz(cpu_data, delivered_perf);
 352 }
 353
 354 static unsigned int cppc_cpufreq_get_rate(unsigned int cpu)
 355 {
 356         struct cppc_perf_fb_ctrs fb_ctrs_t0 = {0}, fb_ctrs_t1 = {0};
 357         struct cppc_cpudata *cpu_data = all_cpu_data[cpu];
 358         int ret;
 359
 360         ret = cppc_get_perf_ctrs(cpu, &fb_ctrs_t0);
 361         if (ret)
 362                 return ret;
 363
 364         udelay(2); /* 2usec delay between sampling */
 365
 366         ret = cppc_get_perf_ctrs(cpu, &fb_ctrs_t1);
 367         if (ret)
 368                 return ret;
 369
 370         return cppc_get_rate_from_fbctrs(cpu_data, fb_ctrs_t0, fb_ctrs_t1);
 371 }
 372
 373 static int cppc_cpufreq_set_boost(struct cpufreq_policy *policy, int state)
 374 {
 375         struct cppc_cpudata *cpu_data = all_cpu_data[policy->cpu];
 376         struct cppc_perf_caps *caps = &cpu_data->perf_caps;
 377         int ret;
 378
 379         if (!boost_supported) {
 380                 pr_err("BOOST not supported by CPU or firmware\n");
 381                 return -EINVAL;
 382         }
 383
 384         if (state)
 385                 policy->max = cppc_cpufreq_perf_to_khz(cpu_data,
 386                                                        caps->highest_perf);
 387         else
 388                 policy->max = cppc_cpufreq_perf_to_khz(cpu_data,
 389                                                        caps->nominal_perf);
 390         policy->cpuinfo.max_freq = policy->max;
 391
 392         ret = freq_qos_update_request(policy->max_freq_req, policy->max);
 393         if (ret < 0)
 394                 return ret;
 395
 396         return 0;
 397 }
 398
 399 static struct cpufreq_driver cppc_cpufreq_driver = {
 400         .flags = CPUFREQ_CONST_LOOPS,
 401         .verify = cppc_verify_policy,
 402         .target = cppc_cpufreq_set_target,
 403         .get = cppc_cpufreq_get_rate,
 404         .init = cppc_cpufreq_cpu_init,
 405         .stop_cpu = cppc_cpufreq_stop_cpu,
 406         .set_boost = cppc_cpufreq_set_boost,
 407         .name = "cppc_cpufreq",
 408 };
 409
 410 /*
 411  * HISI platform does not support delivered performance counter and
 412  * reference performance counter. It can calculate the performance using the
 413  * platform specific mechanism. We reuse the desired performance register to
 414  * store the real performance calculated by the platform.
 415  */
 416 static unsigned int hisi_cppc_cpufreq_get_rate(unsigned int cpu)
 417 {
 418         struct cppc_cpudata *cpu_data = all_cpu_data[cpu];
 419         u64 desired_perf;
 420         int ret;
 421
 422         ret = cppc_get_desired_perf(cpu, &desired_perf);
 423         if (ret < 0)
 424                 return -EIO;
 425
 426         return cppc_cpufreq_perf_to_khz(cpu_data, desired_perf);
 427 }
 428
 429 static void cppc_check_hisi_workaround(void)
 430 {
 431         struct acpi_table_header *tbl;
 432         acpi_status status = AE_OK;
 433         int i;
 434
 435         status = acpi_get_table(ACPI_SIG_PCCT, 0, &tbl);
 436         if (ACPI_FAILURE(status) || !tbl)
 437                 return;
 438
 439         for (i = 0; i < ARRAY_SIZE(wa_info); i++) {
 440                 if (!memcmp(wa_info[i].oem_id, tbl->oem_id, ACPI_OEM_ID_SIZE) &&
 441                     !memcmp(wa_info[i].oem_table_id, tbl->oem_table_id, ACPI_OEM_TABLE_ID_SIZE) &&
 442                     wa_info[i].oem_revision == tbl->oem_revision) {
 443                         /* Overwrite the get() callback */
 444                         cppc_cpufreq_driver.get = hisi_cppc_cpufreq_get_rate;
 445                         break;
 446                 }
 447         }
 448
 449         acpi_put_table(tbl);
 450 }
 451
 452 static int __init cppc_cpufreq_init(void)
 453 {
 454         struct cppc_cpudata *cpu_data;
 455         int i, ret = 0;
 456
 457         if (acpi_disabled)
 458                 return -ENODEV;
 459
 460         all_cpu_data = kcalloc(num_possible_cpus(), sizeof(void *),
 461                                GFP_KERNEL);
 462         if (!all_cpu_data)
 463                 return -ENOMEM;
 464
 465         for_each_possible_cpu(i) {
 466                 all_cpu_data[i] = kzalloc(sizeof(struct cppc_cpudata), GFP_KERNEL);
 467                 if (!all_cpu_data[i])
 468                         goto out;
 469
 470                 cpu_data = all_cpu_data[i];
 471                 if (!zalloc_cpumask_var(&cpu_data->shared_cpu_map, GFP_KERNEL))
 472                         goto out;
 473         }
 474
 475         ret = acpi_get_psd_map(all_cpu_data);
 476         if (ret) {
 477                 pr_debug("Error parsing PSD data. Aborting cpufreq registration.\n");
 478                 goto out;
 479         }
 480
 481         cppc_check_hisi_workaround();
 482
 483         ret = cpufreq_register_driver(&cppc_cpufreq_driver);
 484         if (ret)
 485                 goto out;
 486
 487         return ret;
 488
 489 out:
 490         for_each_possible_cpu(i) {
 491                 cpu_data = all_cpu_data[i];
 492                 if (!cpu_data)
 493                         break;
 494                 free_cpumask_var(cpu_data->shared_cpu_map);
 495                 kfree(cpu_data);
 496         }
 497
 498         kfree(all_cpu_data);
 499         return -ENODEV;
 500 }
 501
 502 static void __exit cppc_cpufreq_exit(void)
 503 {
 504         struct cppc_cpudata *cpu_data;
 505         int i;
 506
 507         cpufreq_unregister_driver(&cppc_cpufreq_driver);
 508
 509         for_each_possible_cpu(i) {
 510                 cpu_data = all_cpu_data[i];
 511                 free_cpumask_var(cpu_data->shared_cpu_map);
 512                 kfree(cpu_data);
 513         }
 514
 515         kfree(all_cpu_data);
 516 }
 517
 518 module_exit(cppc_cpufreq_exit);
 519 MODULE_AUTHOR("Ashwin Chaugule");
 520 MODULE_DESCRIPTION("CPUFreq driver based on the ACPI CPPC v5.0+ spec");
 521 MODULE_LICENSE("GPL");
 522
 523 late_initcall(cppc_cpufreq_init);
 524
 525 static const struct acpi_device_id cppc_acpi_ids[] __used = {
 526         {ACPI_PROCESSOR_DEVICE_HID, },
 527         {}
 528 };
 529
 530 MODULE_DEVICE_TABLE(acpi, cppc_acpi_ids);