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_atomic.h>
  16 #include <rte_memcpy.h>
  17 #include <rte_memory.h>
  18 #include <rte_string_fns.h>
  19
  20 #include "power_acpi_cpufreq.h"
  21 #include "power_common.h"
  22
  23 #ifdef RTE_LIBRTE_POWER_DEBUG
  24 #define POWER_DEBUG_TRACE(fmt, args...) do { \
  25                 RTE_LOG(ERR, POWER, "%s: " fmt, __func__, ## args); \
  26 } while (0)
  27 #else
  28 #define POWER_DEBUG_TRACE(fmt, args...)
  29 #endif
  30
  31 #define FOPEN_OR_ERR_RET(f, retval) do { \
  32                 if ((f) == NULL) { \
  33                         RTE_LOG(ERR, POWER, "File not openned\n"); \
  34                         return retval; \
  35                 } \
  36 } while (0)
  37
  38 #define FOPS_OR_NULL_GOTO(ret, label) do { \
  39                 if ((ret) == NULL) { \
  40                         RTE_LOG(ERR, POWER, "fgets returns nothing\n"); \
  41                         goto label; \
  42                 } \
  43 } while (0)
  44
  45 #define FOPS_OR_ERR_GOTO(ret, label) do { \
  46                 if ((ret) < 0) { \
  47                         RTE_LOG(ERR, POWER, "File operations failed\n"); \
  48                         goto label; \
  49                 } \
  50 } while (0)
  51
  52 #define STR_SIZE     1024
  53 #define POWER_CONVERT_TO_DECIMAL 10
  54
  55 #define POWER_GOVERNOR_USERSPACE "userspace"
  56 #define POWER_SYSFILE_GOVERNOR   \
  57                 "/sys/devices/system/cpu/cpu%u/cpufreq/scaling_governor"
  58 #define POWER_SYSFILE_AVAIL_FREQ \
  59                 "/sys/devices/system/cpu/cpu%u/cpufreq/scaling_available_frequencies"
  60 #define POWER_SYSFILE_SETSPEED   \
  61                 "/sys/devices/system/cpu/cpu%u/cpufreq/scaling_setspeed"
  62
  63 /*
  64  * MSR related
  65  */
  66 #define PLATFORM_INFO     0x0CE
  67 #define TURBO_RATIO_LIMIT 0x1AD
  68 #define IA32_PERF_CTL     0x199
  69 #define CORE_TURBO_DISABLE_BIT ((uint64_t)1<<32)
  70
  71 enum power_state {
  72         POWER_IDLE = 0,
  73         POWER_ONGOING,
  74         POWER_USED,
  75         POWER_UNKNOWN
  76 };
  77
  78 /**
  79  * Power info per lcore.
  80  */
  81 struct rte_power_info {
  82         unsigned int lcore_id;                   /**< Logical core id */
  83         uint32_t freqs[RTE_MAX_LCORE_FREQS]; /**< Frequency array */
  84         uint32_t nb_freqs;                   /**< number of available freqs */
  85         FILE *f;                             /**< FD of scaling_setspeed */
  86         char governor_ori[32];               /**< Original governor name */
  87         uint32_t curr_idx;                   /**< Freq index in freqs array */
  88         volatile uint32_t state;             /**< Power in use state */
  89         uint16_t turbo_available;            /**< Turbo Boost available */
  90         uint16_t turbo_enable;               /**< Turbo Boost enable/disable */
  91 } __rte_cache_aligned;
  92
  93 static struct rte_power_info lcore_power_info[RTE_MAX_LCORE];
  94
  95 /**
  96  * It is to set specific freq for specific logical core, according to the index
  97  * of supported frequencies.
  98  */
  99 static int
 100 set_freq_internal(struct rte_power_info *pi, uint32_t idx)
 101 {
 102         if (idx >= RTE_MAX_LCORE_FREQS || idx >= pi->nb_freqs) {
 103                 RTE_LOG(ERR, POWER, "Invalid frequency index %u, which "
 104                                 "should be less than %u\n", idx, pi->nb_freqs);
 105                 return -1;
 106         }
 107
 108         /* Check if it is the same as current */
 109         if (idx == pi->curr_idx)
 110                 return 0;
 111
 112         POWER_DEBUG_TRACE("Freqency[%u] %u to be set for lcore %u\n",
 113                         idx, pi->freqs[idx], pi->lcore_id);
 114         if (fseek(pi->f, 0, SEEK_SET) < 0) {
 115                 RTE_LOG(ERR, POWER, "Fail to set file position indicator to 0 "
 116                                 "for setting frequency for lcore %u\n", pi->lcore_id);
 117                 return -1;
 118         }
 119         if (fprintf(pi->f, "%u", pi->freqs[idx]) < 0) {
 120                 RTE_LOG(ERR, POWER, "Fail to write new frequency for "
 121                                 "lcore %u\n", pi->lcore_id);
 122                 return -1;
 123         }
 124         fflush(pi->f);
 125         pi->curr_idx = idx;
 126
 127         return 1;
 128 }
 129
 130 /**
 131  * It is to check the current scaling governor by reading sys file, and then
 132  * set it into 'userspace' if it is not by writing the sys file. The original
 133  * governor will be saved for rolling back.
 134  */
 135 static int
 136 power_set_governor_userspace(struct rte_power_info *pi)
 137 {
 138         FILE *f;
 139         int ret = -1;
 140         char buf[BUFSIZ];
 141         char fullpath[PATH_MAX];
 142         char *s;
 143         int val;
 144
 145         snprintf(fullpath, sizeof(fullpath), POWER_SYSFILE_GOVERNOR,
 146                         pi->lcore_id);
 147         f = fopen(fullpath, "rw+");
 148         FOPEN_OR_ERR_RET(f, ret);
 149
 150         s = fgets(buf, sizeof(buf), f);
 151         FOPS_OR_NULL_GOTO(s, out);
 152         /* Strip off terminating '\n' */
 153         strtok(buf, "\n");
 154
 155         /* Check if current governor is userspace */
 156         if (strncmp(buf, POWER_GOVERNOR_USERSPACE,
 157                         sizeof(POWER_GOVERNOR_USERSPACE)) == 0) {
 158                 ret = 0;
 159                 POWER_DEBUG_TRACE("Power management governor of lcore %u is "
 160                                 "already userspace\n", pi->lcore_id);
 161                 goto out;
 162         }
 163         /* Save the original governor */
 164         strlcpy(pi->governor_ori, buf, sizeof(pi->governor_ori));
 165
 166         /* Write 'userspace' to the governor */
 167         val = fseek(f, 0, SEEK_SET);
 168         FOPS_OR_ERR_GOTO(val, out);
 169
 170         val = fputs(POWER_GOVERNOR_USERSPACE, f);
 171         FOPS_OR_ERR_GOTO(val, out);
 172
 173         /* We need to flush to see if the fputs succeeds */
 174         val = fflush(f);
 175         FOPS_OR_ERR_GOTO(val, out);
 176
 177         ret = 0;
 178         RTE_LOG(INFO, POWER, "Power management governor of lcore %u has been "
 179                         "set to user space successfully\n", pi->lcore_id);
 180 out:
 181         fclose(f);
 182
 183         return ret;
 184 }
 185
 186 /**
 187  * It is to get the available frequencies of the specific lcore by reading the
 188  * sys file.
 189  */
 190 static int
 191 power_get_available_freqs(struct rte_power_info *pi)
 192 {
 193         FILE *f;
 194         int ret = -1, i, count;
 195         char *p;
 196         char buf[BUFSIZ];
 197         char fullpath[PATH_MAX];
 198         char *freqs[RTE_MAX_LCORE_FREQS];
 199         char *s;
 200
 201         snprintf(fullpath, sizeof(fullpath), POWER_SYSFILE_AVAIL_FREQ,
 202                         pi->lcore_id);
 203         f = fopen(fullpath, "r");
 204         FOPEN_OR_ERR_RET(f, ret);
 205
 206         s = fgets(buf, sizeof(buf), f);
 207         FOPS_OR_NULL_GOTO(s, out);
 208
 209         /* Strip the line break if there is */
 210         p = strchr(buf, '\n');
 211         if (p != NULL)
 212                 *p = 0;
 213
 214         /* Split string into at most RTE_MAX_LCORE_FREQS frequencies */
 215         count = rte_strsplit(buf, sizeof(buf), freqs,
 216                         RTE_MAX_LCORE_FREQS, ' ');
 217         if (count <= 0) {
 218                 RTE_LOG(ERR, POWER, "No available frequency in "
 219                                 ""POWER_SYSFILE_AVAIL_FREQ"\n", pi->lcore_id);
 220                 goto out;
 221         }
 222         if (count >= RTE_MAX_LCORE_FREQS) {
 223                 RTE_LOG(ERR, POWER, "Too many available frequencies : %d\n",
 224                                 count);
 225                 goto out;
 226         }
 227
 228         /* Store the available frequncies into power context */
 229         for (i = 0, pi->nb_freqs = 0; i < count; i++) {
 230                 POWER_DEBUG_TRACE("Lcore %u frequency[%d]: %s\n", pi->lcore_id,
 231                                 i, freqs[i]);
 232                 pi->freqs[pi->nb_freqs++] = strtoul(freqs[i], &p,
 233                                 POWER_CONVERT_TO_DECIMAL);
 234         }
 235
 236         if ((pi->freqs[0]-1000) == pi->freqs[1]) {
 237                 pi->turbo_available = 1;
 238                 pi->turbo_enable = 1;
 239                 POWER_DEBUG_TRACE("Lcore %u Can do Turbo Boost\n",
 240                                 pi->lcore_id);
 241         } else {
 242                 pi->turbo_available = 0;
 243                 pi->turbo_enable = 0;
 244                 POWER_DEBUG_TRACE("Turbo Boost not available on Lcore %u\n",
 245                                 pi->lcore_id);
 246         }
 247
 248         ret = 0;
 249         POWER_DEBUG_TRACE("%d frequency(s) of lcore %u are available\n",
 250                         count, pi->lcore_id);
 251 out:
 252         fclose(f);
 253
 254         return ret;
 255 }
 256
 257 /**
 258  * It is to fopen the sys file for the future setting the lcore frequency.
 259  */
 260 static int
 261 power_init_for_setting_freq(struct rte_power_info *pi)
 262 {
 263         FILE *f;
 264         char fullpath[PATH_MAX];
 265         char buf[BUFSIZ];
 266         uint32_t i, freq;
 267         char *s;
 268
 269         snprintf(fullpath, sizeof(fullpath), POWER_SYSFILE_SETSPEED,
 270                         pi->lcore_id);
 271         f = fopen(fullpath, "rw+");
 272         FOPEN_OR_ERR_RET(f, -1);
 273
 274         s = fgets(buf, sizeof(buf), f);
 275         FOPS_OR_NULL_GOTO(s, out);
 276
 277         freq = strtoul(buf, NULL, POWER_CONVERT_TO_DECIMAL);
 278         for (i = 0; i < pi->nb_freqs; i++) {
 279                 if (freq == pi->freqs[i]) {
 280                         pi->curr_idx = i;
 281                         pi->f = f;
 282                         return 0;
 283                 }
 284         }
 285
 286 out:
 287         fclose(f);
 288
 289         return -1;
 290 }
 291
 292 int
 293 power_acpi_cpufreq_init(unsigned int lcore_id)
 294 {
 295         struct rte_power_info *pi;
 296
 297         if (lcore_id >= RTE_MAX_LCORE) {
 298                 RTE_LOG(ERR, POWER, "Lcore id %u can not exceeds %u\n",
 299                                 lcore_id, RTE_MAX_LCORE - 1U);
 300                 return -1;
 301         }
 302
 303         pi = &lcore_power_info[lcore_id];
 304         if (rte_atomic32_cmpset(&(pi->state), POWER_IDLE, POWER_ONGOING)
 305                         == 0) {
 306                 RTE_LOG(INFO, POWER, "Power management of lcore %u is "
 307                                 "in use\n", lcore_id);
 308                 return -1;
 309         }
 310
 311         pi->lcore_id = lcore_id;
 312         /* Check and set the governor */
 313         if (power_set_governor_userspace(pi) < 0) {
 314                 RTE_LOG(ERR, POWER, "Cannot set governor of lcore %u to "
 315                                 "userspace\n", lcore_id);
 316                 goto fail;
 317         }
 318
 319         /* Get the available frequencies */
 320         if (power_get_available_freqs(pi) < 0) {
 321                 RTE_LOG(ERR, POWER, "Cannot get available frequencies of "
 322                                 "lcore %u\n", lcore_id);
 323                 goto fail;
 324         }
 325
 326         /* Init for setting lcore frequency */
 327         if (power_init_for_setting_freq(pi) < 0) {
 328                 RTE_LOG(ERR, POWER, "Cannot init for setting frequency for "
 329                                 "lcore %u\n", lcore_id);
 330                 goto fail;
 331         }
 332
 333         /* Set freq to max by default */
 334         if (power_acpi_cpufreq_freq_max(lcore_id) < 0) {
 335                 RTE_LOG(ERR, POWER, "Cannot set frequency of lcore %u "
 336                                 "to max\n", lcore_id);
 337                 goto fail;
 338         }
 339
 340         RTE_LOG(INFO, POWER, "Initialized successfully for lcore %u "
 341                         "power management\n", lcore_id);
 342         rte_atomic32_cmpset(&(pi->state), POWER_ONGOING, POWER_USED);
 343
 344         return 0;
 345
 346 fail:
 347         rte_atomic32_cmpset(&(pi->state), POWER_ONGOING, POWER_UNKNOWN);
 348
 349         return -1;
 350 }
 351
 352 /**
 353  * It is to check the governor and then set the original governor back if
 354  * needed by writing the sys file.
 355  */
 356 static int
 357 power_set_governor_original(struct rte_power_info *pi)
 358 {
 359         FILE *f;
 360         int ret = -1;
 361         char buf[BUFSIZ];
 362         char fullpath[PATH_MAX];
 363         char *s;
 364         int val;
 365
 366         snprintf(fullpath, sizeof(fullpath), POWER_SYSFILE_GOVERNOR,
 367                         pi->lcore_id);
 368         f = fopen(fullpath, "rw+");
 369         FOPEN_OR_ERR_RET(f, ret);
 370
 371         s = fgets(buf, sizeof(buf), f);
 372         FOPS_OR_NULL_GOTO(s, out);
 373
 374         /* Check if the governor to be set is the same as current */
 375         if (strncmp(buf, pi->governor_ori, sizeof(pi->governor_ori)) == 0) {
 376                 ret = 0;
 377                 POWER_DEBUG_TRACE("Power management governor of lcore %u "
 378                                 "has already been set to %s\n",
 379                                 pi->lcore_id, pi->governor_ori);
 380                 goto out;
 381         }
 382
 383         /* Write back the original governor */
 384         val = fseek(f, 0, SEEK_SET);
 385         FOPS_OR_ERR_GOTO(val, out);
 386
 387         val = fputs(pi->governor_ori, f);
 388         FOPS_OR_ERR_GOTO(val, out);
 389
 390         ret = 0;
 391         RTE_LOG(INFO, POWER, "Power management governor of lcore %u "
 392                         "has been set back to %s successfully\n",
 393                         pi->lcore_id, pi->governor_ori);
 394 out:
 395         fclose(f);
 396
 397         return ret;
 398 }
 399
 400 int
 401 power_acpi_cpufreq_exit(unsigned int lcore_id)
 402 {
 403         struct rte_power_info *pi;
 404
 405         if (lcore_id >= RTE_MAX_LCORE) {
 406                 RTE_LOG(ERR, POWER, "Lcore id %u can not exceeds %u\n",
 407                                 lcore_id, RTE_MAX_LCORE - 1U);
 408                 return -1;
 409         }
 410         pi = &lcore_power_info[lcore_id];
 411         if (rte_atomic32_cmpset(&(pi->state), POWER_USED, POWER_ONGOING)
 412                         == 0) {
 413                 RTE_LOG(INFO, POWER, "Power management of lcore %u is "
 414                                 "not used\n", lcore_id);
 415                 return -1;
 416         }
 417
 418         /* Close FD of setting freq */
 419         fclose(pi->f);
 420         pi->f = NULL;
 421
 422         /* Set the governor back to the original */
 423         if (power_set_governor_original(pi) < 0) {
 424                 RTE_LOG(ERR, POWER, "Cannot set the governor of %u back "
 425                                 "to the original\n", lcore_id);
 426                 goto fail;
 427         }
 428
 429         RTE_LOG(INFO, POWER, "Power management of lcore %u has exited from "
 430                         "'userspace' mode and been set back to the "
 431                         "original\n", lcore_id);
 432         rte_atomic32_cmpset(&(pi->state), POWER_ONGOING, POWER_IDLE);
 433
 434         return 0;
 435
 436 fail:
 437         rte_atomic32_cmpset(&(pi->state), POWER_ONGOING, POWER_UNKNOWN);
 438
 439         return -1;
 440 }
 441
 442 uint32_t
 443 power_acpi_cpufreq_freqs(unsigned int lcore_id, uint32_t *freqs, uint32_t num)
 444 {
 445         struct rte_power_info *pi;
 446
 447         if (lcore_id >= RTE_MAX_LCORE) {
 448                 RTE_LOG(ERR, POWER, "Invalid lcore ID\n");
 449                 return 0;
 450         }
 451
 452         if (freqs == NULL) {
 453                 RTE_LOG(ERR, POWER, "NULL buffer supplied\n");
 454                 return 0;
 455         }
 456
 457         pi = &lcore_power_info[lcore_id];
 458         if (num < pi->nb_freqs) {
 459                 RTE_LOG(ERR, POWER, "Buffer size is not enough\n");
 460                 return 0;
 461         }
 462         rte_memcpy(freqs, pi->freqs, pi->nb_freqs * sizeof(uint32_t));
 463
 464         return pi->nb_freqs;
 465 }
 466
 467 uint32_t
 468 power_acpi_cpufreq_get_freq(unsigned int lcore_id)
 469 {
 470         if (lcore_id >= RTE_MAX_LCORE) {
 471                 RTE_LOG(ERR, POWER, "Invalid lcore ID\n");
 472                 return RTE_POWER_INVALID_FREQ_INDEX;
 473         }
 474
 475         return lcore_power_info[lcore_id].curr_idx;
 476 }
 477
 478 int
 479 power_acpi_cpufreq_set_freq(unsigned int lcore_id, uint32_t index)
 480 {
 481         if (lcore_id >= RTE_MAX_LCORE) {
 482                 RTE_LOG(ERR, POWER, "Invalid lcore ID\n");
 483                 return -1;
 484         }
 485
 486         return set_freq_internal(&(lcore_power_info[lcore_id]), index);
 487 }
 488
 489 int
 490 power_acpi_cpufreq_freq_down(unsigned int lcore_id)
 491 {
 492         struct rte_power_info *pi;
 493
 494         if (lcore_id >= RTE_MAX_LCORE) {
 495                 RTE_LOG(ERR, POWER, "Invalid lcore ID\n");
 496                 return -1;
 497         }
 498
 499         pi = &lcore_power_info[lcore_id];
 500         if (pi->curr_idx + 1 == pi->nb_freqs)
 501                 return 0;
 502
 503         /* Frequencies in the array are from high to low. */
 504         return set_freq_internal(pi, pi->curr_idx + 1);
 505 }
 506
 507 int
 508 power_acpi_cpufreq_freq_up(unsigned int lcore_id)
 509 {
 510         struct rte_power_info *pi;
 511
 512         if (lcore_id >= RTE_MAX_LCORE) {
 513                 RTE_LOG(ERR, POWER, "Invalid lcore ID\n");
 514                 return -1;
 515         }
 516
 517         pi = &lcore_power_info[lcore_id];
 518         if (pi->curr_idx == 0)
 519                 return 0;
 520
 521         /* Frequencies in the array are from high to low. */
 522         return set_freq_internal(pi, pi->curr_idx - 1);
 523 }
 524
 525 int
 526 power_acpi_cpufreq_freq_max(unsigned int lcore_id)
 527 {
 528         if (lcore_id >= RTE_MAX_LCORE) {
 529                 RTE_LOG(ERR, POWER, "Invalid lcore ID\n");
 530                 return -1;
 531         }
 532
 533         /* Frequencies in the array are from high to low. */
 534         if (lcore_power_info[lcore_id].turbo_available) {
 535                 if (lcore_power_info[lcore_id].turbo_enable)
 536                         /* Set to Turbo */
 537                         return set_freq_internal(
 538                                         &lcore_power_info[lcore_id], 0);
 539                 else
 540                         /* Set to max non-turbo */
 541                         return set_freq_internal(
 542                                         &lcore_power_info[lcore_id], 1);
 543         } else
 544                 return set_freq_internal(&lcore_power_info[lcore_id], 0);
 545 }
 546
 547 int
 548 power_acpi_cpufreq_freq_min(unsigned int lcore_id)
 549 {
 550         struct rte_power_info *pi;
 551
 552         if (lcore_id >= RTE_MAX_LCORE) {
 553                 RTE_LOG(ERR, POWER, "Invalid lcore ID\n");
 554                 return -1;
 555         }
 556
 557         pi = &lcore_power_info[lcore_id];
 558
 559         /* Frequencies in the array are from high to low. */
 560         return set_freq_internal(pi, pi->nb_freqs - 1);
 561 }
 562
 563
 564 int
 565 power_acpi_turbo_status(unsigned int lcore_id)
 566 {
 567         struct rte_power_info *pi;
 568
 569         if (lcore_id >= RTE_MAX_LCORE) {
 570                 RTE_LOG(ERR, POWER, "Invalid lcore ID\n");
 571                 return -1;
 572         }
 573
 574         pi = &lcore_power_info[lcore_id];
 575
 576         return pi->turbo_enable;
 577 }
 578
 579
 580 int
 581 power_acpi_enable_turbo(unsigned int lcore_id)
 582 {
 583         struct rte_power_info *pi;
 584
 585         if (lcore_id >= RTE_MAX_LCORE) {
 586                 RTE_LOG(ERR, POWER, "Invalid lcore ID\n");
 587                 return -1;
 588         }
 589
 590         pi = &lcore_power_info[lcore_id];
 591
 592         if (pi->turbo_available)
 593                 pi->turbo_enable = 1;
 594         else {
 595                 pi->turbo_enable = 0;
 596                 RTE_LOG(ERR, POWER,
 597                         "Failed to enable turbo on lcore %u\n",
 598                         lcore_id);
 599                         return -1;
 600         }
 601
 602         /* Max may have changed, so call to max function */
 603         if (power_acpi_cpufreq_freq_max(lcore_id) < 0) {
 604                 RTE_LOG(ERR, POWER,
 605                         "Failed to set frequency of lcore %u to max\n",
 606                         lcore_id);
 607                         return -1;
 608         }
 609
 610         return 0;
 611 }
 612
 613 int
 614 power_acpi_disable_turbo(unsigned int lcore_id)
 615 {
 616         struct rte_power_info *pi;
 617
 618         if (lcore_id >= RTE_MAX_LCORE) {
 619                 RTE_LOG(ERR, POWER, "Invalid lcore ID\n");
 620                 return -1;
 621         }
 622
 623         pi = &lcore_power_info[lcore_id];
 624
 625          pi->turbo_enable = 0;
 626
 627         if ((pi->turbo_available) && (pi->curr_idx <= 1)) {
 628                 /* Try to set freq to max by default coming out of turbo */
 629                 if (power_acpi_cpufreq_freq_max(lcore_id) < 0) {
 630                         RTE_LOG(ERR, POWER,
 631                                 "Failed to set frequency of lcore %u to max\n",
 632                                 lcore_id);
 633                         return -1;
 634                 }
 635         }
 636
 637         return 0;
 638 }
 639
 640 int power_acpi_get_capabilities(unsigned int lcore_id,
 641                 struct rte_power_core_capabilities *caps)
 642 {
 643         struct rte_power_info *pi;
 644
 645         if (lcore_id >= RTE_MAX_LCORE) {
 646                 RTE_LOG(ERR, POWER, "Invalid lcore ID\n");
 647                 return -1;
 648         }
 649         if (caps == NULL) {
 650                 RTE_LOG(ERR, POWER, "Invalid argument\n");
 651                 return -1;
 652         }
 653
 654         pi = &lcore_power_info[lcore_id];
 655         caps->capabilities = 0;
 656         caps->turbo = !!(pi->turbo_available);
 657
 658         return 0;
 659 }