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