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  * This list contains information parsed from per CPU ACPI _CPC and _PSD
  34  * structures: e.g. the highest and lowest supported performance, capabilities,
  35  * desired performance, level requested etc. Depending on the share_type, not
  36  * all CPUs will have an entry in the list.
  37  */
  38 static LIST_HEAD(cpu_data_list);
  39
  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 {
 153         struct cppc_cpudata *cpu_data = policy->driver_data;
 154         unsigned int cpu = policy->cpu;
 155         struct cpufreq_freqs freqs;
 156         u32 desired_perf;
 157         int ret = 0;
 158
 159         desired_perf = cppc_cpufreq_khz_to_perf(cpu_data, target_freq);
 160         /* Return if it is exactly the same perf */
 161         if (desired_perf == cpu_data->perf_ctrls.desired_perf)
 162                 return ret;
 163
 164         cpu_data->perf_ctrls.desired_perf = desired_perf;
 165         freqs.old = policy->cur;
 166         freqs.new = target_freq;
 167
 168         cpufreq_freq_transition_begin(policy, &freqs);
 169         ret = cppc_set_perf(cpu, &cpu_data->perf_ctrls);
 170         cpufreq_freq_transition_end(policy, &freqs, ret != 0);
 171
 172         if (ret)
 173                 pr_debug("Failed to set target on CPU:%d. ret:%d\n",
 174                          cpu, ret);
 175
 176         return ret;
 177 }
 178
 179 static int cppc_verify_policy(struct cpufreq_policy_data *policy)
 180 {
 181         cpufreq_verify_within_cpu_limits(policy);
 182         return 0;
 183 }
 184
 185 static void cppc_cpufreq_stop_cpu(struct cpufreq_policy *policy)
 186 {
 187         struct cppc_cpudata *cpu_data = policy->driver_data;
 188         struct cppc_perf_caps *caps = &cpu_data->perf_caps;
 189         unsigned int cpu = policy->cpu;
 190         int ret;
 191
 192         cpu_data->perf_ctrls.desired_perf = caps->lowest_perf;
 193
 194         ret = cppc_set_perf(cpu, &cpu_data->perf_ctrls);
 195         if (ret)
 196                 pr_debug("Err setting perf value:%d on CPU:%d. ret:%d\n",
 197                          caps->lowest_perf, cpu, ret);
 198
 199         /* Remove CPU node from list and free driver data for policy */
 200         free_cpumask_var(cpu_data->shared_cpu_map);
 201         list_del(&cpu_data->node);
 202         kfree(policy->driver_data);
 203         policy->driver_data = NULL;
 204 }
 205
 206 /*
 207  * The PCC subspace describes the rate at which platform can accept commands
 208  * on the shared PCC channel (including READs which do not count towards freq
 209  * transition requests), so ideally we need to use the PCC values as a fallback
 210  * if we don't have a platform specific transition_delay_us
 211  */
 212 #ifdef CONFIG_ARM64
 213 #include <asm/cputype.h>
 214
 215 static unsigned int cppc_cpufreq_get_transition_delay_us(unsigned int cpu)
 216 {
 217         unsigned long implementor = read_cpuid_implementor();
 218         unsigned long part_num = read_cpuid_part_number();
 219
 220         switch (implementor) {
 221         case ARM_CPU_IMP_QCOM:
 222                 switch (part_num) {
 223                 case QCOM_CPU_PART_FALKOR_V1:
 224                 case QCOM_CPU_PART_FALKOR:
 225                         return 10000;
 226                 }
 227         }
 228         return cppc_get_transition_latency(cpu) / NSEC_PER_USEC;
 229 }
 230
 231 #else
 232
 233 static unsigned int cppc_cpufreq_get_transition_delay_us(unsigned int cpu)
 234 {
 235         return cppc_get_transition_latency(cpu) / NSEC_PER_USEC;
 236 }
 237 #endif
 238
 239
 240 static struct cppc_cpudata *cppc_cpufreq_get_cpu_data(unsigned int cpu)
 241 {
 242         struct cppc_cpudata *cpu_data;
 243         int ret;
 244
 245         cpu_data = kzalloc(sizeof(struct cppc_cpudata), GFP_KERNEL);
 246         if (!cpu_data)
 247                 goto out;
 248
 249         if (!zalloc_cpumask_var(&cpu_data->shared_cpu_map, GFP_KERNEL))
 250                 goto free_cpu;
 251
 252         ret = acpi_get_psd_map(cpu, cpu_data);
 253         if (ret) {
 254                 pr_debug("Err parsing CPU%d PSD data: ret:%d\n", cpu, ret);
 255                 goto free_mask;
 256         }
 257
 258         ret = cppc_get_perf_caps(cpu, &cpu_data->perf_caps);
 259         if (ret) {
 260                 pr_debug("Err reading CPU%d perf caps: ret:%d\n", cpu, ret);
 261                 goto free_mask;
 262         }
 263
 264         /* Convert the lowest and nominal freq from MHz to KHz */
 265         cpu_data->perf_caps.lowest_freq *= 1000;
 266         cpu_data->perf_caps.nominal_freq *= 1000;
 267
 268         list_add(&cpu_data->node, &cpu_data_list);
 269
 270         return cpu_data;
 271
 272 free_mask:
 273         free_cpumask_var(cpu_data->shared_cpu_map);
 274 free_cpu:
 275         kfree(cpu_data);
 276 out:
 277         return NULL;
 278 }
 279
 280 static int cppc_cpufreq_cpu_init(struct cpufreq_policy *policy)
 281 {
 282         unsigned int cpu = policy->cpu;
 283         struct cppc_cpudata *cpu_data;
 284         struct cppc_perf_caps *caps;
 285         int ret;
 286
 287         cpu_data = cppc_cpufreq_get_cpu_data(cpu);
 288         if (!cpu_data) {
 289                 pr_err("Error in acquiring _CPC/_PSD data for CPU%d.\n", cpu);
 290                 return -ENODEV;
 291         }
 292         caps = &cpu_data->perf_caps;
 293         policy->driver_data = cpu_data;
 294
 295         /*
 296          * Set min to lowest nonlinear perf to avoid any efficiency penalty (see
 297          * Section 8.4.7.1.1.5 of ACPI 6.1 spec)
 298          */
 299         policy->min = cppc_cpufreq_perf_to_khz(cpu_data,
 300                                                caps->lowest_nonlinear_perf);
 301         policy->max = cppc_cpufreq_perf_to_khz(cpu_data,
 302                                                caps->nominal_perf);
 303
 304         /*
 305          * Set cpuinfo.min_freq to Lowest to make the full range of performance
 306          * available if userspace wants to use any perf between lowest & lowest
 307          * nonlinear perf
 308          */
 309         policy->cpuinfo.min_freq = cppc_cpufreq_perf_to_khz(cpu_data,
 310                                                             caps->lowest_perf);
 311         policy->cpuinfo.max_freq = cppc_cpufreq_perf_to_khz(cpu_data,
 312                                                             caps->nominal_perf);
 313
 314         policy->transition_delay_us = cppc_cpufreq_get_transition_delay_us(cpu);
 315         policy->shared_type = cpu_data->shared_type;
 316
 317         switch (policy->shared_type) {
 318         case CPUFREQ_SHARED_TYPE_HW:
 319         case CPUFREQ_SHARED_TYPE_NONE:
 320                 /* Nothing to be done - we'll have a policy for each CPU */
 321                 break;
 322         case CPUFREQ_SHARED_TYPE_ANY:
 323                 /*
 324                  * All CPUs in the domain will share a policy and all cpufreq
 325                  * operations will use a single cppc_cpudata structure stored
 326                  * in policy->driver_data.
 327                  */
 328                 cpumask_copy(policy->cpus, cpu_data->shared_cpu_map);
 329                 break;
 330         default:
 331                 pr_debug("Unsupported CPU co-ord type: %d\n",
 332                          policy->shared_type);
 333                 return -EFAULT;
 334         }
 335
 336         /*
 337          * If 'highest_perf' is greater than 'nominal_perf', we assume CPU Boost
 338          * is supported.
 339          */
 340         if (caps->highest_perf > caps->nominal_perf)
 341                 boost_supported = true;
 342
 343         /* Set policy->cur to max now. The governors will adjust later. */
 344         policy->cur = cppc_cpufreq_perf_to_khz(cpu_data, caps->highest_perf);
 345         cpu_data->perf_ctrls.desired_perf =  caps->highest_perf;
 346
 347         ret = cppc_set_perf(cpu, &cpu_data->perf_ctrls);
 348         if (ret)
 349                 pr_debug("Err setting perf value:%d on CPU:%d. ret:%d\n",
 350                          caps->highest_perf, cpu, ret);
 351
 352         return ret;
 353 }
 354
 355 static inline u64 get_delta(u64 t1, u64 t0)
 356 {
 357         if (t1 > t0 || t0 > ~(u32)0)
 358                 return t1 - t0;
 359
 360         return (u32)t1 - (u32)t0;
 361 }
 362
 363 static int cppc_get_rate_from_fbctrs(struct cppc_cpudata *cpu_data,
 364                                      struct cppc_perf_fb_ctrs fb_ctrs_t0,
 365                                      struct cppc_perf_fb_ctrs fb_ctrs_t1)
 366 {
 367         u64 delta_reference, delta_delivered;
 368         u64 reference_perf, delivered_perf;
 369
 370         reference_perf = fb_ctrs_t0.reference_perf;
 371
 372         delta_reference = get_delta(fb_ctrs_t1.reference,
 373                                     fb_ctrs_t0.reference);
 374         delta_delivered = get_delta(fb_ctrs_t1.delivered,
 375                                     fb_ctrs_t0.delivered);
 376
 377         /* Check to avoid divide-by zero */
 378         if (delta_reference || delta_delivered)
 379                 delivered_perf = (reference_perf * delta_delivered) /
 380                                         delta_reference;
 381         else
 382                 delivered_perf = cpu_data->perf_ctrls.desired_perf;
 383
 384         return cppc_cpufreq_perf_to_khz(cpu_data, delivered_perf);
 385 }
 386
 387 static unsigned int cppc_cpufreq_get_rate(unsigned int cpu)
 388 {
 389         struct cppc_perf_fb_ctrs fb_ctrs_t0 = {0}, fb_ctrs_t1 = {0};
 390         struct cpufreq_policy *policy = cpufreq_cpu_get(cpu);
 391         struct cppc_cpudata *cpu_data = policy->driver_data;
 392         int ret;
 393
 394         cpufreq_cpu_put(policy);
 395
 396         ret = cppc_get_perf_ctrs(cpu, &fb_ctrs_t0);
 397         if (ret)
 398                 return ret;
 399
 400         udelay(2); /* 2usec delay between sampling */
 401
 402         ret = cppc_get_perf_ctrs(cpu, &fb_ctrs_t1);
 403         if (ret)
 404                 return ret;
 405
 406         return cppc_get_rate_from_fbctrs(cpu_data, fb_ctrs_t0, fb_ctrs_t1);
 407 }
 408
 409 static int cppc_cpufreq_set_boost(struct cpufreq_policy *policy, int state)
 410 {
 411         struct cppc_cpudata *cpu_data = policy->driver_data;
 412         struct cppc_perf_caps *caps = &cpu_data->perf_caps;
 413         int ret;
 414
 415         if (!boost_supported) {
 416                 pr_err("BOOST not supported by CPU or firmware\n");
 417                 return -EINVAL;
 418         }
 419
 420         if (state)
 421                 policy->max = cppc_cpufreq_perf_to_khz(cpu_data,
 422                                                        caps->highest_perf);
 423         else
 424                 policy->max = cppc_cpufreq_perf_to_khz(cpu_data,
 425                                                        caps->nominal_perf);
 426         policy->cpuinfo.max_freq = policy->max;
 427
 428         ret = freq_qos_update_request(policy->max_freq_req, policy->max);
 429         if (ret < 0)
 430                 return ret;
 431
 432         return 0;
 433 }
 434
 435 static ssize_t show_freqdomain_cpus(struct cpufreq_policy *policy, char *buf)
 436 {
 437         struct cppc_cpudata *cpu_data = policy->driver_data;
 438
 439         return cpufreq_show_cpus(cpu_data->shared_cpu_map, buf);
 440 }
 441 cpufreq_freq_attr_ro(freqdomain_cpus);
 442
 443 static struct freq_attr *cppc_cpufreq_attr[] = {
 444         &freqdomain_cpus,
 445         NULL,
 446 };
 447
 448 static struct cpufreq_driver cppc_cpufreq_driver = {
 449         .flags = CPUFREQ_CONST_LOOPS,
 450         .verify = cppc_verify_policy,
 451         .target = cppc_cpufreq_set_target,
 452         .get = cppc_cpufreq_get_rate,
 453         .init = cppc_cpufreq_cpu_init,
 454         .stop_cpu = cppc_cpufreq_stop_cpu,
 455         .set_boost = cppc_cpufreq_set_boost,
 456         .attr = cppc_cpufreq_attr,
 457         .name = "cppc_cpufreq",
 458 };
 459
 460 /*
 461  * HISI platform does not support delivered performance counter and
 462  * reference performance counter. It can calculate the performance using the
 463  * platform specific mechanism. We reuse the desired performance register to
 464  * store the real performance calculated by the platform.
 465  */
 466 static unsigned int hisi_cppc_cpufreq_get_rate(unsigned int cpu)
 467 {
 468         struct cpufreq_policy *policy = cpufreq_cpu_get(cpu);
 469         struct cppc_cpudata *cpu_data = policy->driver_data;
 470         u64 desired_perf;
 471         int ret;
 472
 473         cpufreq_cpu_put(policy);
 474
 475         ret = cppc_get_desired_perf(cpu, &desired_perf);
 476         if (ret < 0)
 477                 return -EIO;
 478
 479         return cppc_cpufreq_perf_to_khz(cpu_data, desired_perf);
 480 }
 481
 482 static void cppc_check_hisi_workaround(void)
 483 {
 484         struct acpi_table_header *tbl;
 485         acpi_status status = AE_OK;
 486         int i;
 487
 488         status = acpi_get_table(ACPI_SIG_PCCT, 0, &tbl);
 489         if (ACPI_FAILURE(status) || !tbl)
 490                 return;
 491
 492         for (i = 0; i < ARRAY_SIZE(wa_info); i++) {
 493                 if (!memcmp(wa_info[i].oem_id, tbl->oem_id, ACPI_OEM_ID_SIZE) &&
 494                     !memcmp(wa_info[i].oem_table_id, tbl->oem_table_id, ACPI_OEM_TABLE_ID_SIZE) &&
 495                     wa_info[i].oem_revision == tbl->oem_revision) {
 496                         /* Overwrite the get() callback */
 497                         cppc_cpufreq_driver.get = hisi_cppc_cpufreq_get_rate;
 498                         break;
 499                 }
 500         }
 501
 502         acpi_put_table(tbl);
 503 }
 504
 505 static int __init cppc_cpufreq_init(void)
 506 {
 507         if ((acpi_disabled) || !acpi_cpc_valid())
 508                 return -ENODEV;
 509
 510         INIT_LIST_HEAD(&cpu_data_list);
 511
 512         cppc_check_hisi_workaround();
 513
 514         return cpufreq_register_driver(&cppc_cpufreq_driver);
 515 }
 516
 517 static inline void free_cpu_data(void)
 518 {
 519         struct cppc_cpudata *iter, *tmp;
 520
 521         list_for_each_entry_safe(iter, tmp, &cpu_data_list, node) {
 522                 free_cpumask_var(iter->shared_cpu_map);
 523                 list_del(&iter->node);
 524                 kfree(iter);
 525         }
 526
 527 }
 528
 529 static void __exit cppc_cpufreq_exit(void)
 530 {
 531         cpufreq_unregister_driver(&cppc_cpufreq_driver);
 532
 533         free_cpu_data();
 534 }
 535
 536 module_exit(cppc_cpufreq_exit);
 537 MODULE_AUTHOR("Ashwin Chaugule");
 538 MODULE_DESCRIPTION("CPUFreq driver based on the ACPI CPPC v5.0+ spec");
 539 MODULE_LICENSE("GPL");
 540
 541 late_initcall(cppc_cpufreq_init);
 542
 543 static const struct acpi_device_id cppc_acpi_ids[] __used = {
 544         {ACPI_PROCESSOR_DEVICE_HID, },
 545         {}
 546 };
 547
 548 MODULE_DEVICE_TABLE(acpi, cppc_acpi_ids);