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