drivers/cpufreq/powernv-cpufreq.c

   1 /*
   2  * POWERNV cpufreq driver for the IBM POWER processors
   3  *
   4  * (C) Copyright IBM 2014
   5  *
   6  * Author: Vaidyanathan Srinivasan <svaidy at linux.vnet.ibm.com>
   7  *
   8  * This program is free software; you can redistribute it and/or modify
   9  * it under the terms of the GNU General Public License as published by
  10  * the Free Software Foundation; either version 2, or (at your option)
  11  * any later version.
  12  *
  13  * This program is distributed in the hope that it will be useful,
  14  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  15  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  16  * GNU General Public License for more details.
  17  *
  18  */
  19
  20 #define pr_fmt(fmt)     "powernv-cpufreq: " fmt
  21
  22 #include <linux/kernel.h>
  23 #include <linux/sysfs.h>
  24 #include <linux/cpumask.h>
  25 #include <linux/module.h>
  26 #include <linux/cpufreq.h>
  27 #include <linux/smp.h>
  28 #include <linux/of.h>
  29 #include <linux/reboot.h>
  30 #include <linux/slab.h>
  31
  32 #include <asm/cputhreads.h>
  33 #include <asm/firmware.h>
  34 #include <asm/reg.h>
  35 #include <asm/smp.h> /* Required for cpu_sibling_mask() in UP configs */
  36 #include <asm/opal.h>
  37
  38 #define POWERNV_MAX_PSTATES     256
  39 #define PMSR_PSAFE_ENABLE       (1UL << 30)
  40 #define PMSR_SPR_EM_DISABLE     (1UL << 31)
  41 #define PMSR_MAX(x)             ((x >> 32) & 0xFF)
  42
  43 static struct cpufreq_frequency_table powernv_freqs[POWERNV_MAX_PSTATES+1];
  44 static bool rebooting, throttled, occ_reset;
  45
  46 static struct chip {
  47         unsigned int id;
  48         bool throttled;
  49         cpumask_t mask;
  50         struct work_struct throttle;
  51         bool restore;
  52 } *chips;
  53
  54 static int nr_chips;
  55
  56 /*
  57  * Note: The set of pstates consists of contiguous integers, the
  58  * smallest of which is indicated by powernv_pstate_info.min, the
  59  * largest of which is indicated by powernv_pstate_info.max.
  60  *
  61  * The nominal pstate is the highest non-turbo pstate in this
  62  * platform. This is indicated by powernv_pstate_info.nominal.
  63  */
  64 static struct powernv_pstate_info {
  65         int min;
  66         int max;
  67         int nominal;
  68         int nr_pstates;
  69 } powernv_pstate_info;
  70
  71 /*
  72  * Initialize the freq table based on data obtained
  73  * from the firmware passed via device-tree
  74  */
  75 static int init_powernv_pstates(void)
  76 {
  77         struct device_node *power_mgt;
  78         int i, pstate_min, pstate_max, pstate_nominal, nr_pstates = 0;
  79         const __be32 *pstate_ids, *pstate_freqs;
  80         u32 len_ids, len_freqs;
  81
  82         power_mgt = of_find_node_by_path("/ibm,opal/power-mgt");
  83         if (!power_mgt) {
  84                 pr_warn("power-mgt node not found\n");
  85                 return -ENODEV;
  86         }
  87
  88         if (of_property_read_u32(power_mgt, "ibm,pstate-min", &pstate_min)) {
  89                 pr_warn("ibm,pstate-min node not found\n");
  90                 return -ENODEV;
  91         }
  92
  93         if (of_property_read_u32(power_mgt, "ibm,pstate-max", &pstate_max)) {
  94                 pr_warn("ibm,pstate-max node not found\n");
  95                 return -ENODEV;
  96         }
  97
  98         if (of_property_read_u32(power_mgt, "ibm,pstate-nominal",
  99                                  &pstate_nominal)) {
 100                 pr_warn("ibm,pstate-nominal not found\n");
 101                 return -ENODEV;
 102         }
 103         pr_info("cpufreq pstate min %d nominal %d max %d\n", pstate_min,
 104                 pstate_nominal, pstate_max);
 105
 106         pstate_ids = of_get_property(power_mgt, "ibm,pstate-ids", &len_ids);
 107         if (!pstate_ids) {
 108                 pr_warn("ibm,pstate-ids not found\n");
 109                 return -ENODEV;
 110         }
 111
 112         pstate_freqs = of_get_property(power_mgt, "ibm,pstate-frequencies-mhz",
 113                                       &len_freqs);
 114         if (!pstate_freqs) {
 115                 pr_warn("ibm,pstate-frequencies-mhz not found\n");
 116                 return -ENODEV;
 117         }
 118
 119         if (len_ids != len_freqs) {
 120                 pr_warn("Entries in ibm,pstate-ids and "
 121                         "ibm,pstate-frequencies-mhz does not match\n");
 122         }
 123
 124         nr_pstates = min(len_ids, len_freqs) / sizeof(u32);
 125         if (!nr_pstates) {
 126                 pr_warn("No PStates found\n");
 127                 return -ENODEV;
 128         }
 129
 130         pr_debug("NR PStates %d\n", nr_pstates);
 131         for (i = 0; i < nr_pstates; i++) {
 132                 u32 id = be32_to_cpu(pstate_ids[i]);
 133                 u32 freq = be32_to_cpu(pstate_freqs[i]);
 134
 135                 pr_debug("PState id %d freq %d MHz\n", id, freq);
 136                 powernv_freqs[i].frequency = freq * 1000; /* kHz */
 137                 powernv_freqs[i].driver_data = id;
 138         }
 139         /* End of list marker entry */
 140         powernv_freqs[i].frequency = CPUFREQ_TABLE_END;
 141
 142         powernv_pstate_info.min = pstate_min;
 143         powernv_pstate_info.max = pstate_max;
 144         powernv_pstate_info.nominal = pstate_nominal;
 145         powernv_pstate_info.nr_pstates = nr_pstates;
 146
 147         return 0;
 148 }
 149
 150 /* Returns the CPU frequency corresponding to the pstate_id. */
 151 static unsigned int pstate_id_to_freq(int pstate_id)
 152 {
 153         int i;
 154
 155         i = powernv_pstate_info.max - pstate_id;
 156         if (i >= powernv_pstate_info.nr_pstates || i < 0) {
 157                 pr_warn("PState id %d outside of PState table, "
 158                         "reporting nominal id %d instead\n",
 159                         pstate_id, powernv_pstate_info.nominal);
 160                 i = powernv_pstate_info.max - powernv_pstate_info.nominal;
 161         }
 162
 163         return powernv_freqs[i].frequency;
 164 }
 165
 166 /*
 167  * cpuinfo_nominal_freq_show - Show the nominal CPU frequency as indicated by
 168  * the firmware
 169  */
 170 static ssize_t cpuinfo_nominal_freq_show(struct cpufreq_policy *policy,
 171                                         char *buf)
 172 {
 173         return sprintf(buf, "%u\n",
 174                 pstate_id_to_freq(powernv_pstate_info.nominal));
 175 }
 176
 177 struct freq_attr cpufreq_freq_attr_cpuinfo_nominal_freq =
 178         __ATTR_RO(cpuinfo_nominal_freq);
 179
 180 static struct freq_attr *powernv_cpu_freq_attr[] = {
 181         &cpufreq_freq_attr_scaling_available_freqs,
 182         &cpufreq_freq_attr_cpuinfo_nominal_freq,
 183         NULL,
 184 };
 185
 186 /* Helper routines */
 187
 188 /* Access helpers to power mgt SPR */
 189
 190 static inline unsigned long get_pmspr(unsigned long sprn)
 191 {
 192         switch (sprn) {
 193         case SPRN_PMCR:
 194                 return mfspr(SPRN_PMCR);
 195
 196         case SPRN_PMICR:
 197                 return mfspr(SPRN_PMICR);
 198
 199         case SPRN_PMSR:
 200                 return mfspr(SPRN_PMSR);
 201         }
 202         BUG();
 203 }
 204
 205 static inline void set_pmspr(unsigned long sprn, unsigned long val)
 206 {
 207         switch (sprn) {
 208         case SPRN_PMCR:
 209                 mtspr(SPRN_PMCR, val);
 210                 return;
 211
 212         case SPRN_PMICR:
 213                 mtspr(SPRN_PMICR, val);
 214                 return;
 215         }
 216         BUG();
 217 }
 218
 219 /*
 220  * Use objects of this type to query/update
 221  * pstates on a remote CPU via smp_call_function.
 222  */
 223 struct powernv_smp_call_data {
 224         unsigned int freq;
 225         int pstate_id;
 226 };
 227
 228 /*
 229  * powernv_read_cpu_freq: Reads the current frequency on this CPU.
 230  *
 231  * Called via smp_call_function.
 232  *
 233  * Note: The caller of the smp_call_function should pass an argument of
 234  * the type 'struct powernv_smp_call_data *' along with this function.
 235  *
 236  * The current frequency on this CPU will be returned via
 237  * ((struct powernv_smp_call_data *)arg)->freq;
 238  */
 239 static void powernv_read_cpu_freq(void *arg)
 240 {
 241         unsigned long pmspr_val;
 242         s8 local_pstate_id;
 243         struct powernv_smp_call_data *freq_data = arg;
 244
 245         pmspr_val = get_pmspr(SPRN_PMSR);
 246
 247         /*
 248          * The local pstate id corresponds bits 48..55 in the PMSR.
 249          * Note: Watch out for the sign!
 250          */
 251         local_pstate_id = (pmspr_val >> 48) & 0xFF;
 252         freq_data->pstate_id = local_pstate_id;
 253         freq_data->freq = pstate_id_to_freq(freq_data->pstate_id);
 254
 255         pr_debug("cpu %d pmsr %016lX pstate_id %d frequency %d kHz\n",
 256                 raw_smp_processor_id(), pmspr_val, freq_data->pstate_id,
 257                 freq_data->freq);
 258 }
 259
 260 /*
 261  * powernv_cpufreq_get: Returns the CPU frequency as reported by the
 262  * firmware for CPU 'cpu'. This value is reported through the sysfs
 263  * file cpuinfo_cur_freq.
 264  */
 265 static unsigned int powernv_cpufreq_get(unsigned int cpu)
 266 {
 267         struct powernv_smp_call_data freq_data;
 268
 269         smp_call_function_any(cpu_sibling_mask(cpu), powernv_read_cpu_freq,
 270                         &freq_data, 1);
 271
 272         return freq_data.freq;
 273 }
 274
 275 /*
 276  * set_pstate: Sets the pstate on this CPU.
 277  *
 278  * This is called via an smp_call_function.
 279  *
 280  * The caller must ensure that freq_data is of the type
 281  * (struct powernv_smp_call_data *) and the pstate_id which needs to be set
 282  * on this CPU should be present in freq_data->pstate_id.
 283  */
 284 static void set_pstate(void *freq_data)
 285 {
 286         unsigned long val;
 287         unsigned long pstate_ul =
 288                 ((struct powernv_smp_call_data *) freq_data)->pstate_id;
 289
 290         val = get_pmspr(SPRN_PMCR);
 291         val = val & 0x0000FFFFFFFFFFFFULL;
 292
 293         pstate_ul = pstate_ul & 0xFF;
 294
 295         /* Set both global(bits 56..63) and local(bits 48..55) PStates */
 296         val = val | (pstate_ul << 56) | (pstate_ul << 48);
 297
 298         pr_debug("Setting cpu %d pmcr to %016lX\n",
 299                         raw_smp_processor_id(), val);
 300         set_pmspr(SPRN_PMCR, val);
 301 }
 302
 303 /*
 304  * get_nominal_index: Returns the index corresponding to the nominal
 305  * pstate in the cpufreq table
 306  */
 307 static inline unsigned int get_nominal_index(void)
 308 {
 309         return powernv_pstate_info.max - powernv_pstate_info.nominal;
 310 }
 311
 312 static void powernv_cpufreq_throttle_check(void *data)
 313 {
 314         unsigned int cpu = smp_processor_id();
 315         unsigned long pmsr;
 316         int pmsr_pmax, i;
 317
 318         pmsr = get_pmspr(SPRN_PMSR);
 319
 320         for (i = 0; i < nr_chips; i++)
 321                 if (chips[i].id == cpu_to_chip_id(cpu))
 322                         break;
 323
 324         /* Check for Pmax Capping */
 325         pmsr_pmax = (s8)PMSR_MAX(pmsr);
 326         if (pmsr_pmax != powernv_pstate_info.max) {
 327                 if (chips[i].throttled)
 328                         goto next;
 329                 chips[i].throttled = true;
 330                 if (pmsr_pmax < powernv_pstate_info.nominal)
 331                         pr_crit("CPU %d on Chip %u has Pmax reduced below nominal frequency (%d < %d)\n",
 332                                 cpu, chips[i].id, pmsr_pmax,
 333                                 powernv_pstate_info.nominal);
 334                 else
 335                         pr_info("CPU %d on Chip %u has Pmax reduced below turbo frequency (%d < %d)\n",
 336                                 cpu, chips[i].id, pmsr_pmax,
 337                                 powernv_pstate_info.max);
 338         } else if (chips[i].throttled) {
 339                 chips[i].throttled = false;
 340                 pr_info("CPU %d on Chip %u has Pmax restored to %d\n", cpu,
 341                         chips[i].id, pmsr_pmax);
 342         }
 343
 344         /* Check if Psafe_mode_active is set in PMSR. */
 345 next:
 346         if (pmsr & PMSR_PSAFE_ENABLE) {
 347                 throttled = true;
 348                 pr_info("Pstate set to safe frequency\n");
 349         }
 350
 351         /* Check if SPR_EM_DISABLE is set in PMSR */
 352         if (pmsr & PMSR_SPR_EM_DISABLE) {
 353                 throttled = true;
 354                 pr_info("Frequency Control disabled from OS\n");
 355         }
 356
 357         if (throttled) {
 358                 pr_info("PMSR = %16lx\n", pmsr);
 359                 pr_crit("CPU Frequency could be throttled\n");
 360         }
 361 }
 362
 363 /*
 364  * powernv_cpufreq_target_index: Sets the frequency corresponding to
 365  * the cpufreq table entry indexed by new_index on the cpus in the
 366  * mask policy->cpus
 367  */
 368 static int powernv_cpufreq_target_index(struct cpufreq_policy *policy,
 369                                         unsigned int new_index)
 370 {
 371         struct powernv_smp_call_data freq_data;
 372
 373         if (unlikely(rebooting) && new_index != get_nominal_index())
 374                 return 0;
 375
 376         if (!throttled) {
 377                 /* we don't want to be preempted while
 378                  * checking if the CPU frequency has been throttled
 379                  */
 380                 preempt_disable();
 381                 powernv_cpufreq_throttle_check(NULL);
 382                 preempt_enable();
 383         }
 384
 385         freq_data.pstate_id = powernv_freqs[new_index].driver_data;
 386
 387         /*
 388          * Use smp_call_function to send IPI and execute the
 389          * mtspr on target CPU.  We could do that without IPI
 390          * if current CPU is within policy->cpus (core)
 391          */
 392         smp_call_function_any(policy->cpus, set_pstate, &freq_data, 1);
 393
 394         return 0;
 395 }
 396
 397 static int powernv_cpufreq_cpu_init(struct cpufreq_policy *policy)
 398 {
 399         int base, i;
 400
 401         base = cpu_first_thread_sibling(policy->cpu);
 402
 403         for (i = 0; i < threads_per_core; i++)
 404                 cpumask_set_cpu(base + i, policy->cpus);
 405
 406         return cpufreq_table_validate_and_show(policy, powernv_freqs);
 407 }
 408
 409 static int powernv_cpufreq_reboot_notifier(struct notifier_block *nb,
 410                                 unsigned long action, void *unused)
 411 {
 412         int cpu;
 413         struct cpufreq_policy *cpu_policy;
 414
 415         rebooting = true;
 416         for_each_online_cpu(cpu) {
 417                 cpu_policy = cpufreq_cpu_get(cpu);
 418                 if (!cpu_policy)
 419                         continue;
 420                 powernv_cpufreq_target_index(cpu_policy, get_nominal_index());
 421                 cpufreq_cpu_put(cpu_policy);
 422         }
 423
 424         return NOTIFY_DONE;
 425 }
 426
 427 static struct notifier_block powernv_cpufreq_reboot_nb = {
 428         .notifier_call = powernv_cpufreq_reboot_notifier,
 429 };
 430
 431 void powernv_cpufreq_work_fn(struct work_struct *work)
 432 {
 433         struct chip *chip = container_of(work, struct chip, throttle);
 434         unsigned int cpu;
 435         cpumask_var_t mask;
 436
 437         smp_call_function_any(&chip->mask,
 438                               powernv_cpufreq_throttle_check, NULL, 0);
 439
 440         if (!chip->restore)
 441                 return;
 442
 443         chip->restore = false;
 444         cpumask_copy(mask, &chip->mask);
 445         for_each_cpu_and(cpu, mask, cpu_online_mask) {
 446                 int index, tcpu;
 447                 struct cpufreq_policy policy;
 448
 449                 cpufreq_get_policy(&policy, cpu);
 450                 cpufreq_frequency_table_target(&policy, policy.freq_table,
 451                                                policy.cur,
 452                                                CPUFREQ_RELATION_C, &index);
 453                 powernv_cpufreq_target_index(&policy, index);
 454                 for_each_cpu(tcpu, policy.cpus)
 455                         cpumask_clear_cpu(tcpu, mask);
 456         }
 457 }
 458
 459 static char throttle_reason[][30] = {
 460                                         "No throttling",
 461                                         "Power Cap",
 462                                         "Processor Over Temperature",
 463                                         "Power Supply Failure",
 464                                         "Over Current",
 465                                         "OCC Reset"
 466                                      };
 467
 468 static int powernv_cpufreq_occ_msg(struct notifier_block *nb,
 469                                    unsigned long msg_type, void *_msg)
 470 {
 471         struct opal_msg *msg = _msg;
 472         struct opal_occ_msg omsg;
 473         int i;
 474
 475         if (msg_type != OPAL_MSG_OCC)
 476                 return 0;
 477
 478         omsg.type = be64_to_cpu(msg->params[0]);
 479
 480         switch (omsg.type) {
 481         case OCC_RESET:
 482                 occ_reset = true;
 483                 pr_info("OCC (On Chip Controller - enforces hard thermal/power limits) Resetting\n");
 484                 /*
 485                  * powernv_cpufreq_throttle_check() is called in
 486                  * target() callback which can detect the throttle state
 487                  * for governors like ondemand.
 488                  * But static governors will not call target() often thus
 489                  * report throttling here.
 490                  */
 491                 if (!throttled) {
 492                         throttled = true;
 493                         pr_crit("CPU frequency is throttled for duration\n");
 494                 }
 495
 496                 break;
 497         case OCC_LOAD:
 498                 pr_info("OCC Loading, CPU frequency is throttled until OCC is started\n");
 499                 break;
 500         case OCC_THROTTLE:
 501                 omsg.chip = be64_to_cpu(msg->params[1]);
 502                 omsg.throttle_status = be64_to_cpu(msg->params[2]);
 503
 504                 if (occ_reset) {
 505                         occ_reset = false;
 506                         throttled = false;
 507                         pr_info("OCC Active, CPU frequency is no longer throttled\n");
 508
 509                         for (i = 0; i < nr_chips; i++) {
 510                                 chips[i].restore = true;
 511                                 schedule_work(&chips[i].throttle);
 512                         }
 513
 514                         return 0;
 515                 }
 516
 517                 if (omsg.throttle_status &&
 518                     omsg.throttle_status <= OCC_MAX_THROTTLE_STATUS)
 519                         pr_info("OCC: Chip %u Pmax reduced due to %s\n",
 520                                 (unsigned int)omsg.chip,
 521                                 throttle_reason[omsg.throttle_status]);
 522                 else if (!omsg.throttle_status)
 523                         pr_info("OCC: Chip %u %s\n", (unsigned int)omsg.chip,
 524                                 throttle_reason[omsg.throttle_status]);
 525                 else
 526                         return 0;
 527
 528                 for (i = 0; i < nr_chips; i++)
 529                         if (chips[i].id == omsg.chip) {
 530                                 if (!omsg.throttle_status)
 531                                         chips[i].restore = true;
 532                                 schedule_work(&chips[i].throttle);
 533                         }
 534         }
 535         return 0;
 536 }
 537
 538 static struct notifier_block powernv_cpufreq_opal_nb = {
 539         .notifier_call  = powernv_cpufreq_occ_msg,
 540         .next           = NULL,
 541         .priority       = 0,
 542 };
 543
 544 static void powernv_cpufreq_stop_cpu(struct cpufreq_policy *policy)
 545 {
 546         struct powernv_smp_call_data freq_data;
 547
 548         freq_data.pstate_id = powernv_pstate_info.min;
 549         smp_call_function_single(policy->cpu, set_pstate, &freq_data, 1);
 550 }
 551
 552 static struct cpufreq_driver powernv_cpufreq_driver = {
 553         .name           = "powernv-cpufreq",
 554         .flags          = CPUFREQ_CONST_LOOPS,
 555         .init           = powernv_cpufreq_cpu_init,
 556         .verify         = cpufreq_generic_frequency_table_verify,
 557         .target_index   = powernv_cpufreq_target_index,
 558         .get            = powernv_cpufreq_get,
 559         .stop_cpu       = powernv_cpufreq_stop_cpu,
 560         .attr           = powernv_cpu_freq_attr,
 561 };
 562
 563 static int init_chip_info(void)
 564 {
 565         unsigned int chip[256];
 566         unsigned int cpu, i;
 567         unsigned int prev_chip_id = UINT_MAX;
 568
 569         for_each_possible_cpu(cpu) {
 570                 unsigned int id = cpu_to_chip_id(cpu);
 571
 572                 if (prev_chip_id != id) {
 573                         prev_chip_id = id;
 574                         chip[nr_chips++] = id;
 575                 }
 576         }
 577
 578         chips = kmalloc_array(nr_chips, sizeof(struct chip), GFP_KERNEL);
 579         if (!chips)
 580                 return -ENOMEM;
 581
 582         for (i = 0; i < nr_chips; i++) {
 583                 chips[i].id = chip[i];
 584                 chips[i].throttled = false;
 585                 cpumask_copy(&chips[i].mask, cpumask_of_node(chip[i]));
 586                 INIT_WORK(&chips[i].throttle, powernv_cpufreq_work_fn);
 587                 chips[i].restore = false;
 588         }
 589
 590         return 0;
 591 }
 592
 593 static int __init powernv_cpufreq_init(void)
 594 {
 595         int rc = 0;
 596
 597         /* Don't probe on pseries (guest) platforms */
 598         if (!firmware_has_feature(FW_FEATURE_OPAL))
 599                 return -ENODEV;
 600
 601         /* Discover pstates from device tree and init */
 602         rc = init_powernv_pstates();
 603         if (rc) {
 604                 pr_info("powernv-cpufreq disabled. System does not support PState control\n");
 605                 return rc;
 606         }
 607
 608         /* Populate chip info */
 609         rc = init_chip_info();
 610         if (rc)
 611                 return rc;
 612
 613         register_reboot_notifier(&powernv_cpufreq_reboot_nb);
 614         opal_message_notifier_register(OPAL_MSG_OCC, &powernv_cpufreq_opal_nb);
 615         return cpufreq_register_driver(&powernv_cpufreq_driver);
 616 }
 617 module_init(powernv_cpufreq_init);
 618
 619 static void __exit powernv_cpufreq_exit(void)
 620 {
 621         unregister_reboot_notifier(&powernv_cpufreq_reboot_nb);
 622         opal_message_notifier_unregister(OPAL_MSG_OCC,
 623                                          &powernv_cpufreq_opal_nb);
 624         cpufreq_unregister_driver(&powernv_cpufreq_driver);
 625 }
 626 module_exit(powernv_cpufreq_exit);
 627
 628 MODULE_LICENSE("GPL");
 629 MODULE_AUTHOR("Vaidyanathan Srinivasan <svaidy at linux.vnet.ibm.com>");