lib/librte_power/power_acpi_cpufreq.c

   1 /* SPDX-License-Identifier: BSD-3-Clause
   2  * Copyright(c) 2010-2014 Intel Corporation
   3  */
   4
   5 #include <stdio.h>
   6 #include <sys/types.h>
   7 #include <sys/stat.h>
   8 #include <fcntl.h>
   9 #include <stdlib.h>
  10 #include <string.h>
  11 #include <unistd.h>
  12 #include <signal.h>
  13 #include <limits.h>
  14
  15 #include <rte_memcpy.h>
  16 #include <rte_atomic.h>
  17
  18 #include "power_acpi_cpufreq.h"
  19 #include "power_common.h"
  20
  21 #ifdef RTE_LIBRTE_POWER_DEBUG
  22 #define POWER_DEBUG_TRACE(fmt, args...) do { \
  23                 RTE_LOG(ERR, POWER, "%s: " fmt, __func__, ## args); \
  24 } while (0)
  25 #else
  26 #define POWER_DEBUG_TRACE(fmt, args...)
  27 #endif
  28
  29 #define FOPEN_OR_ERR_RET(f, retval) do { \
  30                 if ((f) == NULL) { \
  31                         RTE_LOG(ERR, POWER, "File not openned\n"); \
  32                         return retval; \
  33                 } \
  34 } while (0)
  35
  36 #define FOPS_OR_NULL_GOTO(ret, label) do { \
  37                 if ((ret) == NULL) { \
  38                         RTE_LOG(ERR, POWER, "fgets returns nothing\n"); \
  39                         goto label; \
  40                 } \
  41 } while (0)
  42
  43 #define FOPS_OR_ERR_GOTO(ret, label) do { \
  44                 if ((ret) < 0) { \
  45                         RTE_LOG(ERR, POWER, "File operations failed\n"); \
  46                         goto label; \
  47                 } \
  48 } while (0)
  49
  50 #define STR_SIZE     1024
  51 #define POWER_CONVERT_TO_DECIMAL 10
  52
  53 #define POWER_GOVERNOR_USERSPACE "userspace"
  54 #define POWER_SYSFILE_GOVERNOR   \
  55                 "/sys/devices/system/cpu/cpu%u/cpufreq/scaling_governor"
  56 #define POWER_SYSFILE_AVAIL_FREQ \
  57                 "/sys/devices/system/cpu/cpu%u/cpufreq/scaling_available_frequencies"
  58 #define POWER_SYSFILE_SETSPEED   \
  59                 "/sys/devices/system/cpu/cpu%u/cpufreq/scaling_setspeed"
  60
  61 /*
  62  * MSR related
  63  */
  64 #define PLATFORM_INFO     0x0CE
  65 #define TURBO_RATIO_LIMIT 0x1AD
  66 #define IA32_PERF_CTL     0x199
  67 #define CORE_TURBO_DISABLE_BIT ((uint64_t)1<<32)
  68
  69 enum power_state {
  70         POWER_IDLE = 0,
  71         POWER_ONGOING,
  72         POWER_USED,
  73         POWER_UNKNOWN
  74 };
  75
  76 /**
  77  * Power info per lcore.
  78  */
  79 struct rte_power_info {
  80         unsigned int lcore_id;                   /**< Logical core id */
  81         uint32_t freqs[RTE_MAX_LCORE_FREQS]; /**< Frequency array */
  82         uint32_t nb_freqs;                   /**< number of available freqs */
  83         FILE *f;                             /**< FD of scaling_setspeed */
  84         char governor_ori[32];               /**< Original governor name */
  85         uint32_t curr_idx;                   /**< Freq index in freqs array */
  86         volatile uint32_t state;             /**< Power in use state */
  87         uint16_t turbo_available;            /**< Turbo Boost available */
  88         uint16_t turbo_enable;               /**< Turbo Boost enable/disable */
  89 } __rte_cache_aligned;
  90
  91 static struct rte_power_info lcore_power_info[RTE_MAX_LCORE];
  92
  93 /**
  94  * It is to set specific freq for specific logical core, according to the index
  95  * of supported frequencies.
  96  */
  97 static int
  98 set_freq_internal(struct rte_power_info *pi, uint32_t idx)
  99 {
 100         if (idx >= RTE_MAX_LCORE_FREQS || idx >= pi->nb_freqs) {
 101                 RTE_LOG(ERR, POWER, "Invalid frequency index %u, which "
 102                                 "should be less than %u\n", idx, pi->nb_freqs);
 103                 return -1;
 104         }
 105
 106         /* Check if it is the same as current */
 107         if (idx == pi->curr_idx)
 108                 return 0;
 109
 110         POWER_DEBUG_TRACE("Freqency[%u] %u to be set for lcore %u\n",
 111                         idx, pi->freqs[idx], pi->lcore_id);
 112         if (fseek(pi->f, 0, SEEK_SET) < 0) {
 113                 RTE_LOG(ERR, POWER, "Fail to set file position indicator to 0 "
 114                                 "for setting frequency for lcore %u\n", pi->lcore_id);
 115                 return -1;
 116         }
 117         if (fprintf(pi->f, "%u", pi->freqs[idx]) < 0) {
 118                 RTE_LOG(ERR, POWER, "Fail to write new frequency for "
 119                                 "lcore %u\n", pi->lcore_id);
 120                 return -1;
 121         }
 122         fflush(pi->f);
 123         pi->curr_idx = idx;
 124
 125         return 1;
 126 }
 127
 128 /**
 129  * It is to check the current scaling governor by reading sys file, and then
 130  * set it into 'userspace' if it is not by writing the sys file. The original
 131  * governor will be saved for rolling back.
 132  */
 133 static int
 134 power_set_governor_userspace(struct rte_power_info *pi)
 135 {
 136         FILE *f;
 137         int ret = -1;
 138         char buf[BUFSIZ];
 139         char fullpath[PATH_MAX];
 140         char *s;
 141         int val;
 142
 143         snprintf(fullpath, sizeof(fullpath), POWER_SYSFILE_GOVERNOR,
 144                         pi->lcore_id);
 145         f = fopen(fullpath, "rw+");
 146         FOPEN_OR_ERR_RET(f, ret);
 147
 148         s = fgets(buf, sizeof(buf), f);
 149         FOPS_OR_NULL_GOTO(s, out);
 150
 151         /* Check if current governor is userspace */
 152         if (strncmp(buf, POWER_GOVERNOR_USERSPACE,
 153                         sizeof(POWER_GOVERNOR_USERSPACE)) == 0) {
 154                 ret = 0;
 155                 POWER_DEBUG_TRACE("Power management governor of lcore %u is "
 156                                 "already userspace\n", pi->lcore_id);
 157                 goto out;
 158         }
 159         /* Save the original governor */
 160         snprintf(pi->governor_ori, sizeof(pi->governor_ori), "%s", buf);
 161
 162         /* Write 'userspace' to the governor */
 163         val = fseek(f, 0, SEEK_SET);
 164         FOPS_OR_ERR_GOTO(val, out);
 165
 166         val = fputs(POWER_GOVERNOR_USERSPACE, f);
 167         FOPS_OR_ERR_GOTO(val, out);
 168
 169         ret = 0;
 170         RTE_LOG(INFO, POWER, "Power management governor of lcore %u has been "
 171                         "set to user space successfully\n", pi->lcore_id);
 172 out:
 173         fclose(f);
 174
 175         return ret;
 176 }
 177
 178 /**
 179  * It is to get the available frequencies of the specific lcore by reading the
 180  * sys file.
 181  */
 182 static int
 183 power_get_available_freqs(struct rte_power_info *pi)
 184 {
 185         FILE *f;
 186         int ret = -1, i, count;
 187         char *p;
 188         char buf[BUFSIZ];
 189         char fullpath[PATH_MAX];
 190         char *freqs[RTE_MAX_LCORE_FREQS];
 191         char *s;
 192
 193         snprintf(fullpath, sizeof(fullpath), POWER_SYSFILE_AVAIL_FREQ,
 194                         pi->lcore_id);
 195         f = fopen(fullpath, "r");
 196         FOPEN_OR_ERR_RET(f, ret);
 197
 198         s = fgets(buf, sizeof(buf), f);
 199         FOPS_OR_NULL_GOTO(s, out);
 200
 201         /* Strip the line break if there is */
 202         p = strchr(buf, '\n');
 203         if (p != NULL)
 204                 *p = 0;
 205
 206         /* Split string into at most RTE_MAX_LCORE_FREQS frequencies */
 207         count = rte_strsplit(buf, sizeof(buf), freqs,
 208                         RTE_MAX_LCORE_FREQS, ' ');
 209         if (count <= 0) {
 210                 RTE_LOG(ERR, POWER, "No available frequency in "
 211                                 ""POWER_SYSFILE_AVAIL_FREQ"\n", pi->lcore_id);
 212                 goto out;
 213         }
 214         if (count >= RTE_MAX_LCORE_FREQS) {
 215                 RTE_LOG(ERR, POWER, "Too many available frequencies : %d\n",
 216                                 count);
 217                 goto out;
 218         }
 219
 220         /* Store the available frequncies into power context */
 221         for (i = 0, pi->nb_freqs = 0; i < count; i++) {
 222                 POWER_DEBUG_TRACE("Lcore %u frequency[%d]: %s\n", pi->lcore_id,
 223                                 i, freqs[i]);
 224                 pi->freqs[pi->nb_freqs++] = strtoul(freqs[i], &p,
 225                                 POWER_CONVERT_TO_DECIMAL);
 226         }
 227
 228         if ((pi->freqs[0]-1000) == pi->freqs[1]) {
 229                 pi->turbo_available = 1;
 230                 pi->turbo_enable = 1;
 231                 POWER_DEBUG_TRACE("Lcore %u Can do Turbo Boost\n",
 232                                 pi->lcore_id);
 233         } else {
 234                 pi->turbo_available = 0;
 235                 pi->turbo_enable = 0;
 236                 POWER_DEBUG_TRACE("Turbo Boost not available on Lcore %u\n",
 237                                 pi->lcore_id);
 238         }
 239
 240         ret = 0;
 241         POWER_DEBUG_TRACE("%d frequency(s) of lcore %u are available\n",
 242                         count, pi->lcore_id);
 243 out:
 244         fclose(f);
 245
 246         return ret;
 247 }
 248
 249 /**
 250  * It is to fopen the sys file for the future setting the lcore frequency.
 251  */
 252 static int
 253 power_init_for_setting_freq(struct rte_power_info *pi)
 254 {
 255         FILE *f;
 256         char fullpath[PATH_MAX];
 257         char buf[BUFSIZ];
 258         uint32_t i, freq;
 259         char *s;
 260
 261         snprintf(fullpath, sizeof(fullpath), POWER_SYSFILE_SETSPEED,
 262                         pi->lcore_id);
 263         f = fopen(fullpath, "rw+");
 264         FOPEN_OR_ERR_RET(f, -1);
 265
 266         s = fgets(buf, sizeof(buf), f);
 267         FOPS_OR_NULL_GOTO(s, out);
 268
 269         freq = strtoul(buf, NULL, POWER_CONVERT_TO_DECIMAL);
 270         for (i = 0; i < pi->nb_freqs; i++) {
 271                 if (freq == pi->freqs[i]) {
 272                         pi->curr_idx = i;
 273                         pi->f = f;
 274                         return 0;
 275                 }
 276         }
 277
 278 out:
 279         fclose(f);
 280
 281         return -1;
 282 }
 283
 284 int
 285 power_acpi_cpufreq_init(unsigned int lcore_id)
 286 {
 287         struct rte_power_info *pi;
 288
 289         if (lcore_id >= RTE_MAX_LCORE) {
 290                 RTE_LOG(ERR, POWER, "Lcore id %u can not exceeds %u\n",
 291                                 lcore_id, RTE_MAX_LCORE - 1U);
 292                 return -1;
 293         }
 294
 295         pi = &lcore_power_info[lcore_id];
 296         if (rte_atomic32_cmpset(&(pi->state), POWER_IDLE, POWER_ONGOING)
 297                         == 0) {
 298                 RTE_LOG(INFO, POWER, "Power management of lcore %u is "
 299                                 "in use\n", lcore_id);
 300                 return -1;
 301         }
 302
 303         pi->lcore_id = lcore_id;
 304         /* Check and set the governor */
 305         if (power_set_governor_userspace(pi) < 0) {
 306                 RTE_LOG(ERR, POWER, "Cannot set governor of lcore %u to "
 307                                 "userspace\n", lcore_id);
 308                 goto fail;
 309         }
 310
 311         /* Get the available frequencies */
 312         if (power_get_available_freqs(pi) < 0) {
 313                 RTE_LOG(ERR, POWER, "Cannot get available frequencies of "
 314                                 "lcore %u\n", lcore_id);
 315                 goto fail;
 316         }
 317
 318         /* Init for setting lcore frequency */
 319         if (power_init_for_setting_freq(pi) < 0) {
 320                 RTE_LOG(ERR, POWER, "Cannot init for setting frequency for "
 321                                 "lcore %u\n", lcore_id);
 322                 goto fail;
 323         }
 324
 325         /* Set freq to max by default */
 326         if (power_acpi_cpufreq_freq_max(lcore_id) < 0) {
 327                 RTE_LOG(ERR, POWER, "Cannot set frequency of lcore %u "
 328                                 "to max\n", lcore_id);
 329                 goto fail;
 330         }
 331
 332         RTE_LOG(INFO, POWER, "Initialized successfully for lcore %u "
 333                         "power management\n", lcore_id);
 334         rte_atomic32_cmpset(&(pi->state), POWER_ONGOING, POWER_USED);
 335
 336         return 0;
 337
 338 fail:
 339         rte_atomic32_cmpset(&(pi->state), POWER_ONGOING, POWER_UNKNOWN);
 340
 341         return -1;
 342 }
 343
 344 /**
 345  * It is to check the governor and then set the original governor back if
 346  * needed by writing the sys file.
 347  */
 348 static int
 349 power_set_governor_original(struct rte_power_info *pi)
 350 {
 351         FILE *f;
 352         int ret = -1;
 353         char buf[BUFSIZ];
 354         char fullpath[PATH_MAX];
 355         char *s;
 356         int val;
 357
 358         snprintf(fullpath, sizeof(fullpath), POWER_SYSFILE_GOVERNOR,
 359                         pi->lcore_id);
 360         f = fopen(fullpath, "rw+");
 361         FOPEN_OR_ERR_RET(f, ret);
 362
 363         s = fgets(buf, sizeof(buf), f);
 364         FOPS_OR_NULL_GOTO(s, out);
 365
 366         /* Check if the governor to be set is the same as current */
 367         if (strncmp(buf, pi->governor_ori, sizeof(pi->governor_ori)) == 0) {
 368                 ret = 0;
 369                 POWER_DEBUG_TRACE("Power management governor of lcore %u "
 370                                 "has already been set to %s\n",
 371                                 pi->lcore_id, pi->governor_ori);
 372                 goto out;
 373         }
 374
 375         /* Write back the original governor */
 376         val = fseek(f, 0, SEEK_SET);
 377         FOPS_OR_ERR_GOTO(val, out);
 378
 379         val = fputs(pi->governor_ori, f);
 380         FOPS_OR_ERR_GOTO(val, out);
 381
 382         ret = 0;
 383         RTE_LOG(INFO, POWER, "Power management governor of lcore %u "
 384                         "has been set back to %s successfully\n",
 385                         pi->lcore_id, pi->governor_ori);
 386 out:
 387         fclose(f);
 388
 389         return ret;
 390 }
 391
 392 int
 393 power_acpi_cpufreq_exit(unsigned int lcore_id)
 394 {
 395         struct rte_power_info *pi;
 396
 397         if (lcore_id >= RTE_MAX_LCORE) {
 398                 RTE_LOG(ERR, POWER, "Lcore id %u can not exceeds %u\n",
 399                                 lcore_id, RTE_MAX_LCORE - 1U);
 400                 return -1;
 401         }
 402         pi = &lcore_power_info[lcore_id];
 403         if (rte_atomic32_cmpset(&(pi->state), POWER_USED, POWER_ONGOING)
 404                         == 0) {
 405                 RTE_LOG(INFO, POWER, "Power management of lcore %u is "
 406                                 "not used\n", lcore_id);
 407                 return -1;
 408         }
 409
 410         /* Close FD of setting freq */
 411         fclose(pi->f);
 412         pi->f = NULL;
 413
 414         /* Set the governor back to the original */
 415         if (power_set_governor_original(pi) < 0) {
 416                 RTE_LOG(ERR, POWER, "Cannot set the governor of %u back "
 417                                 "to the original\n", lcore_id);
 418                 goto fail;
 419         }
 420
 421         RTE_LOG(INFO, POWER, "Power management of lcore %u has exited from "
 422                         "'userspace' mode and been set back to the "
 423                         "original\n", lcore_id);
 424         rte_atomic32_cmpset(&(pi->state), POWER_ONGOING, POWER_IDLE);
 425
 426         return 0;
 427
 428 fail:
 429         rte_atomic32_cmpset(&(pi->state), POWER_ONGOING, POWER_UNKNOWN);
 430
 431         return -1;
 432 }
 433
 434 uint32_t
 435 power_acpi_cpufreq_freqs(unsigned int lcore_id, uint32_t *freqs, uint32_t num)
 436 {
 437         struct rte_power_info *pi;
 438
 439         if (lcore_id >= RTE_MAX_LCORE || !freqs) {
 440                 RTE_LOG(ERR, POWER, "Invalid input parameter\n");
 441                 return 0;
 442         }
 443
 444         pi = &lcore_power_info[lcore_id];
 445         if (num < pi->nb_freqs) {
 446                 RTE_LOG(ERR, POWER, "Buffer size is not enough\n");
 447                 return 0;
 448         }
 449         rte_memcpy(freqs, pi->freqs, pi->nb_freqs * sizeof(uint32_t));
 450
 451         return pi->nb_freqs;
 452 }
 453
 454 uint32_t
 455 power_acpi_cpufreq_get_freq(unsigned int lcore_id)
 456 {
 457         if (lcore_id >= RTE_MAX_LCORE) {
 458                 RTE_LOG(ERR, POWER, "Invalid lcore ID\n");
 459                 return RTE_POWER_INVALID_FREQ_INDEX;
 460         }
 461
 462         return lcore_power_info[lcore_id].curr_idx;
 463 }
 464
 465 int
 466 power_acpi_cpufreq_set_freq(unsigned int lcore_id, uint32_t index)
 467 {
 468         if (lcore_id >= RTE_MAX_LCORE) {
 469                 RTE_LOG(ERR, POWER, "Invalid lcore ID\n");
 470                 return -1;
 471         }
 472
 473         return set_freq_internal(&(lcore_power_info[lcore_id]), index);
 474 }
 475
 476 int
 477 power_acpi_cpufreq_freq_down(unsigned int lcore_id)
 478 {
 479         struct rte_power_info *pi;
 480
 481         if (lcore_id >= RTE_MAX_LCORE) {
 482                 RTE_LOG(ERR, POWER, "Invalid lcore ID\n");
 483                 return -1;
 484         }
 485
 486         pi = &lcore_power_info[lcore_id];
 487         if (pi->curr_idx + 1 == pi->nb_freqs)
 488                 return 0;
 489
 490         /* Frequencies in the array are from high to low. */
 491         return set_freq_internal(pi, pi->curr_idx + 1);
 492 }
 493
 494 int
 495 power_acpi_cpufreq_freq_up(unsigned int lcore_id)
 496 {
 497         struct rte_power_info *pi;
 498
 499         if (lcore_id >= RTE_MAX_LCORE) {
 500                 RTE_LOG(ERR, POWER, "Invalid lcore ID\n");
 501                 return -1;
 502         }
 503
 504         pi = &lcore_power_info[lcore_id];
 505         if (pi->curr_idx == 0)
 506                 return 0;
 507
 508         /* Frequencies in the array are from high to low. */
 509         return set_freq_internal(pi, pi->curr_idx - 1);
 510 }
 511
 512 int
 513 power_acpi_cpufreq_freq_max(unsigned int lcore_id)
 514 {
 515         if (lcore_id >= RTE_MAX_LCORE) {
 516                 RTE_LOG(ERR, POWER, "Invalid lcore ID\n");
 517                 return -1;
 518         }
 519
 520         /* Frequencies in the array are from high to low. */
 521         if (lcore_power_info[lcore_id].turbo_available) {
 522                 if (lcore_power_info[lcore_id].turbo_enable)
 523                         /* Set to Turbo */
 524                         return set_freq_internal(
 525                                         &lcore_power_info[lcore_id], 0);
 526                 else
 527                         /* Set to max non-turbo */
 528                         return set_freq_internal(
 529                                         &lcore_power_info[lcore_id], 1);
 530         } else
 531                 return set_freq_internal(&lcore_power_info[lcore_id], 0);
 532 }
 533
 534 int
 535 power_acpi_cpufreq_freq_min(unsigned int lcore_id)
 536 {
 537         struct rte_power_info *pi;
 538
 539         if (lcore_id >= RTE_MAX_LCORE) {
 540                 RTE_LOG(ERR, POWER, "Invalid lcore ID\n");
 541                 return -1;
 542         }
 543
 544         pi = &lcore_power_info[lcore_id];
 545
 546         /* Frequencies in the array are from high to low. */
 547         return set_freq_internal(pi, pi->nb_freqs - 1);
 548 }
 549
 550
 551 int
 552 power_acpi_turbo_status(unsigned int lcore_id)
 553 {
 554         struct rte_power_info *pi;
 555
 556         if (lcore_id >= RTE_MAX_LCORE) {
 557                 RTE_LOG(ERR, POWER, "Invalid lcore ID\n");
 558                 return -1;
 559         }
 560
 561         pi = &lcore_power_info[lcore_id];
 562
 563         return pi->turbo_enable;
 564 }
 565
 566
 567 int
 568 power_acpi_enable_turbo(unsigned int lcore_id)
 569 {
 570         struct rte_power_info *pi;
 571
 572         if (lcore_id >= RTE_MAX_LCORE) {
 573                 RTE_LOG(ERR, POWER, "Invalid lcore ID\n");
 574                 return -1;
 575         }
 576
 577         pi = &lcore_power_info[lcore_id];
 578
 579         if (pi->turbo_available)
 580                 pi->turbo_enable = 1;
 581         else {
 582                 pi->turbo_enable = 0;
 583                 RTE_LOG(ERR, POWER,
 584                         "Failed to enable turbo on lcore %u\n",
 585                         lcore_id);
 586                         return -1;
 587         }
 588
 589         /* Max may have changed, so call to max function */
 590         if (power_acpi_cpufreq_freq_max(lcore_id) < 0) {
 591                 RTE_LOG(ERR, POWER,
 592                         "Failed to set frequency of lcore %u to max\n",
 593                         lcore_id);
 594                         return -1;
 595         }
 596
 597         return 0;
 598 }
 599
 600 int
 601 power_acpi_disable_turbo(unsigned int lcore_id)
 602 {
 603         struct rte_power_info *pi;
 604
 605         if (lcore_id >= RTE_MAX_LCORE) {
 606                 RTE_LOG(ERR, POWER, "Invalid lcore ID\n");
 607                 return -1;
 608         }
 609
 610         pi = &lcore_power_info[lcore_id];
 611
 612          pi->turbo_enable = 0;
 613
 614         if ((pi->turbo_available) && (pi->curr_idx <= 1)) {
 615                 /* Try to set freq to max by default coming out of turbo */
 616                 if (power_acpi_cpufreq_freq_max(lcore_id) < 0) {
 617                         RTE_LOG(ERR, POWER,
 618                                 "Failed to set frequency of lcore %u to max\n",
 619                                 lcore_id);
 620                         return -1;
 621                 }
 622         }
 623
 624         return 0;
 625 }