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 #include <linux/cpu.h>
  32 #include <trace/events/power.h>
  33
  34 #include <asm/cputhreads.h>
  35 #include <asm/firmware.h>
  36 #include <asm/reg.h>
  37 #include <asm/smp.h> /* Required for cpu_sibling_mask() in UP configs */
  38 #include <asm/opal.h>
  39 #include <linux/timer.h>
  40
  41 #define POWERNV_MAX_PSTATES     256
  42 #define PMSR_PSAFE_ENABLE       (1UL << 30)
  43 #define PMSR_SPR_EM_DISABLE     (1UL << 31)
  44 #define PMSR_MAX(x)             ((x >> 32) & 0xFF)
  45
  46 #define MAX_RAMP_DOWN_TIME                              5120
  47 /*
  48  * On an idle system we want the global pstate to ramp-down from max value to
  49  * min over a span of ~5 secs. Also we want it to initially ramp-down slowly and
  50  * then ramp-down rapidly later on.
  51  *
  52  * This gives a percentage rampdown for time elapsed in milliseconds.
  53  * ramp_down_percentage = ((ms * ms) >> 18)
  54  *                      ~= 3.8 * (sec * sec)
  55  *
  56  * At 0 ms      ramp_down_percent = 0
  57  * At 5120 ms   ramp_down_percent = 100
  58  */
  59 #define ramp_down_percent(time)         ((time * time) >> 18)
  60
  61 /* Interval after which the timer is queued to bring down global pstate */
  62 #define GPSTATE_TIMER_INTERVAL                          2000
  63
  64 /**
  65  * struct global_pstate_info -  Per policy data structure to maintain history of
  66  *                              global pstates
  67  * @highest_lpstate_idx:        The local pstate index from which we are
  68  *                              ramping down
  69  * @elapsed_time:               Time in ms spent in ramping down from
  70  *                              highest_lpstate_idx
  71  * @last_sampled_time:          Time from boot in ms when global pstates were
  72  *                              last set
  73  * @last_lpstate_idx,           Last set value of local pstate and global
  74  * last_gpstate_idx             pstate in terms of cpufreq table index
  75  * @timer:                      Is used for ramping down if cpu goes idle for
  76  *                              a long time with global pstate held high
  77  * @gpstate_lock:               A spinlock to maintain synchronization between
  78  *                              routines called by the timer handler and
  79  *                              governer's target_index calls
  80  */
  81 struct global_pstate_info {
  82         int highest_lpstate_idx;
  83         unsigned int elapsed_time;
  84         unsigned int last_sampled_time;
  85         int last_lpstate_idx;
  86         int last_gpstate_idx;
  87         spinlock_t gpstate_lock;
  88         struct timer_list timer;
  89 };
  90
  91 static struct cpufreq_frequency_table powernv_freqs[POWERNV_MAX_PSTATES+1];
  92 static bool rebooting, throttled, occ_reset;
  93
  94 static const char * const throttle_reason[] = {
  95         "No throttling",
  96         "Power Cap",
  97         "Processor Over Temperature",
  98         "Power Supply Failure",
  99         "Over Current",
 100         "OCC Reset"
 101 };
 102
 103 enum throttle_reason_type {
 104         NO_THROTTLE = 0,
 105         POWERCAP,
 106         CPU_OVERTEMP,
 107         POWER_SUPPLY_FAILURE,
 108         OVERCURRENT,
 109         OCC_RESET_THROTTLE,
 110         OCC_MAX_REASON
 111 };
 112
 113 static struct chip {
 114         unsigned int id;
 115         bool throttled;
 116         bool restore;
 117         u8 throttle_reason;
 118         cpumask_t mask;
 119         struct work_struct throttle;
 120         int throttle_turbo;
 121         int throttle_sub_turbo;
 122         int reason[OCC_MAX_REASON];
 123 } *chips;
 124
 125 static int nr_chips;
 126 static DEFINE_PER_CPU(struct chip *, chip_info);
 127
 128 /*
 129  * Note:
 130  * The set of pstates consists of contiguous integers.
 131  * powernv_pstate_info stores the index of the frequency table for
 132  * max, min and nominal frequencies. It also stores number of
 133  * available frequencies.
 134  *
 135  * powernv_pstate_info.nominal indicates the index to the highest
 136  * non-turbo frequency.
 137  */
 138 static struct powernv_pstate_info {
 139         unsigned int min;
 140         unsigned int max;
 141         unsigned int nominal;
 142         unsigned int nr_pstates;
 143 } powernv_pstate_info;
 144
 145 /* Use following macros for conversions between pstate_id and index */
 146 static inline int idx_to_pstate(unsigned int i)
 147 {
 148         if (unlikely(i >= powernv_pstate_info.nr_pstates)) {
 149                 pr_warn_once("index %u is out of bound\n", i);
 150                 return powernv_freqs[powernv_pstate_info.nominal].driver_data;
 151         }
 152
 153         return powernv_freqs[i].driver_data;
 154 }
 155
 156 static inline unsigned int pstate_to_idx(int pstate)
 157 {
 158         int min = powernv_freqs[powernv_pstate_info.min].driver_data;
 159         int max = powernv_freqs[powernv_pstate_info.max].driver_data;
 160
 161         if (min > 0) {
 162                 if (unlikely((pstate < max) || (pstate > min))) {
 163                         pr_warn_once("pstate %d is out of bound\n", pstate);
 164                         return powernv_pstate_info.nominal;
 165                 }
 166         } else {
 167                 if (unlikely((pstate > max) || (pstate < min))) {
 168                         pr_warn_once("pstate %d is out of bound\n", pstate);
 169                         return powernv_pstate_info.nominal;
 170                 }
 171         }
 172         /*
 173          * abs() is deliberately used so that is works with
 174          * both monotonically increasing and decreasing
 175          * pstate values
 176          */
 177         return abs(pstate - idx_to_pstate(powernv_pstate_info.max));
 178 }
 179
 180 static inline void reset_gpstates(struct cpufreq_policy *policy)
 181 {
 182         struct global_pstate_info *gpstates = policy->driver_data;
 183
 184         gpstates->highest_lpstate_idx = 0;
 185         gpstates->elapsed_time = 0;
 186         gpstates->last_sampled_time = 0;
 187         gpstates->last_lpstate_idx = 0;
 188         gpstates->last_gpstate_idx = 0;
 189 }
 190
 191 /*
 192  * Initialize the freq table based on data obtained
 193  * from the firmware passed via device-tree
 194  */
 195 static int init_powernv_pstates(void)
 196 {
 197         struct device_node *power_mgt;
 198         int i, nr_pstates = 0;
 199         const __be32 *pstate_ids, *pstate_freqs;
 200         u32 len_ids, len_freqs;
 201         u32 pstate_min, pstate_max, pstate_nominal;
 202
 203         power_mgt = of_find_node_by_path("/ibm,opal/power-mgt");
 204         if (!power_mgt) {
 205                 pr_warn("power-mgt node not found\n");
 206                 return -ENODEV;
 207         }
 208
 209         if (of_property_read_u32(power_mgt, "ibm,pstate-min", &pstate_min)) {
 210                 pr_warn("ibm,pstate-min node not found\n");
 211                 return -ENODEV;
 212         }
 213
 214         if (of_property_read_u32(power_mgt, "ibm,pstate-max", &pstate_max)) {
 215                 pr_warn("ibm,pstate-max node not found\n");
 216                 return -ENODEV;
 217         }
 218
 219         if (of_property_read_u32(power_mgt, "ibm,pstate-nominal",
 220                                  &pstate_nominal)) {
 221                 pr_warn("ibm,pstate-nominal not found\n");
 222                 return -ENODEV;
 223         }
 224         pr_info("cpufreq pstate min %d nominal %d max %d\n", pstate_min,
 225                 pstate_nominal, pstate_max);
 226
 227         pstate_ids = of_get_property(power_mgt, "ibm,pstate-ids", &len_ids);
 228         if (!pstate_ids) {
 229                 pr_warn("ibm,pstate-ids not found\n");
 230                 return -ENODEV;
 231         }
 232
 233         pstate_freqs = of_get_property(power_mgt, "ibm,pstate-frequencies-mhz",
 234                                       &len_freqs);
 235         if (!pstate_freqs) {
 236                 pr_warn("ibm,pstate-frequencies-mhz not found\n");
 237                 return -ENODEV;
 238         }
 239
 240         if (len_ids != len_freqs) {
 241                 pr_warn("Entries in ibm,pstate-ids and "
 242                         "ibm,pstate-frequencies-mhz does not match\n");
 243         }
 244
 245         nr_pstates = min(len_ids, len_freqs) / sizeof(u32);
 246         if (!nr_pstates) {
 247                 pr_warn("No PStates found\n");
 248                 return -ENODEV;
 249         }
 250
 251         powernv_pstate_info.nr_pstates = nr_pstates;
 252         pr_debug("NR PStates %d\n", nr_pstates);
 253         for (i = 0; i < nr_pstates; i++) {
 254                 u32 id = be32_to_cpu(pstate_ids[i]);
 255                 u32 freq = be32_to_cpu(pstate_freqs[i]);
 256
 257                 pr_debug("PState id %d freq %d MHz\n", id, freq);
 258                 powernv_freqs[i].frequency = freq * 1000; /* kHz */
 259                 powernv_freqs[i].driver_data = id;
 260
 261                 if (id == pstate_max)
 262                         powernv_pstate_info.max = i;
 263                 if (id == pstate_nominal)
 264                         powernv_pstate_info.nominal = i;
 265                 if (id == pstate_min)
 266                         powernv_pstate_info.min = i;
 267         }
 268
 269         /* End of list marker entry */
 270         powernv_freqs[i].frequency = CPUFREQ_TABLE_END;
 271         return 0;
 272 }
 273
 274 /* Returns the CPU frequency corresponding to the pstate_id. */
 275 static unsigned int pstate_id_to_freq(int pstate_id)
 276 {
 277         int i;
 278
 279         i = pstate_to_idx(pstate_id);
 280         if (i >= powernv_pstate_info.nr_pstates || i < 0) {
 281                 pr_warn("PState id %d outside of PState table, "
 282                         "reporting nominal id %d instead\n",
 283                         pstate_id, idx_to_pstate(powernv_pstate_info.nominal));
 284                 i = powernv_pstate_info.nominal;
 285         }
 286
 287         return powernv_freqs[i].frequency;
 288 }
 289
 290 /*
 291  * cpuinfo_nominal_freq_show - Show the nominal CPU frequency as indicated by
 292  * the firmware
 293  */
 294 static ssize_t cpuinfo_nominal_freq_show(struct cpufreq_policy *policy,
 295                                         char *buf)
 296 {
 297         return sprintf(buf, "%u\n",
 298                 powernv_freqs[powernv_pstate_info.nominal].frequency);
 299 }
 300
 301 struct freq_attr cpufreq_freq_attr_cpuinfo_nominal_freq =
 302         __ATTR_RO(cpuinfo_nominal_freq);
 303
 304 static struct freq_attr *powernv_cpu_freq_attr[] = {
 305         &cpufreq_freq_attr_scaling_available_freqs,
 306         &cpufreq_freq_attr_cpuinfo_nominal_freq,
 307         NULL,
 308 };
 309
 310 #define throttle_attr(name, member)                                     \
 311 static ssize_t name##_show(struct cpufreq_policy *policy, char *buf)    \
 312 {                                                                       \
 313         struct chip *chip = per_cpu(chip_info, policy->cpu);            \
 314                                                                         \
 315         return sprintf(buf, "%u\n", chip->member);                      \
 316 }                                                                       \
 317                                                                         \
 318 static struct freq_attr throttle_attr_##name = __ATTR_RO(name)          \
 319
 320 throttle_attr(unthrottle, reason[NO_THROTTLE]);
 321 throttle_attr(powercap, reason[POWERCAP]);
 322 throttle_attr(overtemp, reason[CPU_OVERTEMP]);
 323 throttle_attr(supply_fault, reason[POWER_SUPPLY_FAILURE]);
 324 throttle_attr(overcurrent, reason[OVERCURRENT]);
 325 throttle_attr(occ_reset, reason[OCC_RESET_THROTTLE]);
 326 throttle_attr(turbo_stat, throttle_turbo);
 327 throttle_attr(sub_turbo_stat, throttle_sub_turbo);
 328
 329 static struct attribute *throttle_attrs[] = {
 330         &throttle_attr_unthrottle.attr,
 331         &throttle_attr_powercap.attr,
 332         &throttle_attr_overtemp.attr,
 333         &throttle_attr_supply_fault.attr,
 334         &throttle_attr_overcurrent.attr,
 335         &throttle_attr_occ_reset.attr,
 336         &throttle_attr_turbo_stat.attr,
 337         &throttle_attr_sub_turbo_stat.attr,
 338         NULL,
 339 };
 340
 341 static const struct attribute_group throttle_attr_grp = {
 342         .name   = "throttle_stats",
 343         .attrs  = throttle_attrs,
 344 };
 345
 346 /* Helper routines */
 347
 348 /* Access helpers to power mgt SPR */
 349
 350 static inline unsigned long get_pmspr(unsigned long sprn)
 351 {
 352         switch (sprn) {
 353         case SPRN_PMCR:
 354                 return mfspr(SPRN_PMCR);
 355
 356         case SPRN_PMICR:
 357                 return mfspr(SPRN_PMICR);
 358
 359         case SPRN_PMSR:
 360                 return mfspr(SPRN_PMSR);
 361         }
 362         BUG();
 363 }
 364
 365 static inline void set_pmspr(unsigned long sprn, unsigned long val)
 366 {
 367         switch (sprn) {
 368         case SPRN_PMCR:
 369                 mtspr(SPRN_PMCR, val);
 370                 return;
 371
 372         case SPRN_PMICR:
 373                 mtspr(SPRN_PMICR, val);
 374                 return;
 375         }
 376         BUG();
 377 }
 378
 379 /*
 380  * Use objects of this type to query/update
 381  * pstates on a remote CPU via smp_call_function.
 382  */
 383 struct powernv_smp_call_data {
 384         unsigned int freq;
 385         int pstate_id;
 386         int gpstate_id;
 387 };
 388
 389 /*
 390  * powernv_read_cpu_freq: Reads the current frequency on this CPU.
 391  *
 392  * Called via smp_call_function.
 393  *
 394  * Note: The caller of the smp_call_function should pass an argument of
 395  * the type 'struct powernv_smp_call_data *' along with this function.
 396  *
 397  * The current frequency on this CPU will be returned via
 398  * ((struct powernv_smp_call_data *)arg)->freq;
 399  */
 400 static void powernv_read_cpu_freq(void *arg)
 401 {
 402         unsigned long pmspr_val;
 403         s8 local_pstate_id;
 404         struct powernv_smp_call_data *freq_data = arg;
 405
 406         pmspr_val = get_pmspr(SPRN_PMSR);
 407
 408         /*
 409          * The local pstate id corresponds bits 48..55 in the PMSR.
 410          * Note: Watch out for the sign!
 411          */
 412         local_pstate_id = (pmspr_val >> 48) & 0xFF;
 413         freq_data->pstate_id = local_pstate_id;
 414         freq_data->freq = pstate_id_to_freq(freq_data->pstate_id);
 415
 416         pr_debug("cpu %d pmsr %016lX pstate_id %d frequency %d kHz\n",
 417                 raw_smp_processor_id(), pmspr_val, freq_data->pstate_id,
 418                 freq_data->freq);
 419 }
 420
 421 /*
 422  * powernv_cpufreq_get: Returns the CPU frequency as reported by the
 423  * firmware for CPU 'cpu'. This value is reported through the sysfs
 424  * file cpuinfo_cur_freq.
 425  */
 426 static unsigned int powernv_cpufreq_get(unsigned int cpu)
 427 {
 428         struct powernv_smp_call_data freq_data;
 429
 430         smp_call_function_any(cpu_sibling_mask(cpu), powernv_read_cpu_freq,
 431                         &freq_data, 1);
 432
 433         return freq_data.freq;
 434 }
 435
 436 /*
 437  * set_pstate: Sets the pstate on this CPU.
 438  *
 439  * This is called via an smp_call_function.
 440  *
 441  * The caller must ensure that freq_data is of the type
 442  * (struct powernv_smp_call_data *) and the pstate_id which needs to be set
 443  * on this CPU should be present in freq_data->pstate_id.
 444  */
 445 static void set_pstate(void *data)
 446 {
 447         unsigned long val;
 448         struct powernv_smp_call_data *freq_data = data;
 449         unsigned long pstate_ul = freq_data->pstate_id;
 450         unsigned long gpstate_ul = freq_data->gpstate_id;
 451
 452         val = get_pmspr(SPRN_PMCR);
 453         val = val & 0x0000FFFFFFFFFFFFULL;
 454
 455         pstate_ul = pstate_ul & 0xFF;
 456         gpstate_ul = gpstate_ul & 0xFF;
 457
 458         /* Set both global(bits 56..63) and local(bits 48..55) PStates */
 459         val = val | (gpstate_ul << 56) | (pstate_ul << 48);
 460
 461         pr_debug("Setting cpu %d pmcr to %016lX\n",
 462                         raw_smp_processor_id(), val);
 463         set_pmspr(SPRN_PMCR, val);
 464 }
 465
 466 /*
 467  * get_nominal_index: Returns the index corresponding to the nominal
 468  * pstate in the cpufreq table
 469  */
 470 static inline unsigned int get_nominal_index(void)
 471 {
 472         return powernv_pstate_info.nominal;
 473 }
 474
 475 static void powernv_cpufreq_throttle_check(void *data)
 476 {
 477         struct chip *chip;
 478         unsigned int cpu = smp_processor_id();
 479         unsigned long pmsr;
 480         int pmsr_pmax;
 481         unsigned int pmsr_pmax_idx;
 482
 483         pmsr = get_pmspr(SPRN_PMSR);
 484         chip = this_cpu_read(chip_info);
 485
 486         /* Check for Pmax Capping */
 487         pmsr_pmax = (s8)PMSR_MAX(pmsr);
 488         pmsr_pmax_idx = pstate_to_idx(pmsr_pmax);
 489         if (pmsr_pmax_idx != powernv_pstate_info.max) {
 490                 if (chip->throttled)
 491                         goto next;
 492                 chip->throttled = true;
 493                 if (pmsr_pmax_idx > powernv_pstate_info.nominal) {
 494                         pr_warn_once("CPU %d on Chip %u has Pmax(%d) reduced below nominal frequency(%d)\n",
 495                                      cpu, chip->id, pmsr_pmax,
 496                                      idx_to_pstate(powernv_pstate_info.nominal));
 497                         chip->throttle_sub_turbo++;
 498                 } else {
 499                         chip->throttle_turbo++;
 500                 }
 501                 trace_powernv_throttle(chip->id,
 502                                       throttle_reason[chip->throttle_reason],
 503                                       pmsr_pmax);
 504         } else if (chip->throttled) {
 505                 chip->throttled = false;
 506                 trace_powernv_throttle(chip->id,
 507                                       throttle_reason[chip->throttle_reason],
 508                                       pmsr_pmax);
 509         }
 510
 511         /* Check if Psafe_mode_active is set in PMSR. */
 512 next:
 513         if (pmsr & PMSR_PSAFE_ENABLE) {
 514                 throttled = true;
 515                 pr_info("Pstate set to safe frequency\n");
 516         }
 517
 518         /* Check if SPR_EM_DISABLE is set in PMSR */
 519         if (pmsr & PMSR_SPR_EM_DISABLE) {
 520                 throttled = true;
 521                 pr_info("Frequency Control disabled from OS\n");
 522         }
 523
 524         if (throttled) {
 525                 pr_info("PMSR = %16lx\n", pmsr);
 526                 pr_warn("CPU Frequency could be throttled\n");
 527         }
 528 }
 529
 530 /**
 531  * calc_global_pstate - Calculate global pstate
 532  * @elapsed_time:               Elapsed time in milliseconds
 533  * @local_pstate_idx:           New local pstate
 534  * @highest_lpstate_idx:        pstate from which its ramping down
 535  *
 536  * Finds the appropriate global pstate based on the pstate from which its
 537  * ramping down and the time elapsed in ramping down. It follows a quadratic
 538  * equation which ensures that it reaches ramping down to pmin in 5sec.
 539  */
 540 static inline int calc_global_pstate(unsigned int elapsed_time,
 541                                      int highest_lpstate_idx,
 542                                      int local_pstate_idx)
 543 {
 544         int index_diff;
 545
 546         /*
 547          * Using ramp_down_percent we get the percentage of rampdown
 548          * that we are expecting to be dropping. Difference between
 549          * highest_lpstate_idx and powernv_pstate_info.min will give a absolute
 550          * number of how many pstates we will drop eventually by the end of
 551          * 5 seconds, then just scale it get the number pstates to be dropped.
 552          */
 553         index_diff =  ((int)ramp_down_percent(elapsed_time) *
 554                         (powernv_pstate_info.min - highest_lpstate_idx)) / 100;
 555
 556         /* Ensure that global pstate is >= to local pstate */
 557         if (highest_lpstate_idx + index_diff >= local_pstate_idx)
 558                 return local_pstate_idx;
 559         else
 560                 return highest_lpstate_idx + index_diff;
 561 }
 562
 563 static inline void  queue_gpstate_timer(struct global_pstate_info *gpstates)
 564 {
 565         unsigned int timer_interval;
 566
 567         /*
 568          * Setting up timer to fire after GPSTATE_TIMER_INTERVAL ms, But
 569          * if it exceeds MAX_RAMP_DOWN_TIME ms for ramp down time.
 570          * Set timer such that it fires exactly at MAX_RAMP_DOWN_TIME
 571          * seconds of ramp down time.
 572          */
 573         if ((gpstates->elapsed_time + GPSTATE_TIMER_INTERVAL)
 574              > MAX_RAMP_DOWN_TIME)
 575                 timer_interval = MAX_RAMP_DOWN_TIME - gpstates->elapsed_time;
 576         else
 577                 timer_interval = GPSTATE_TIMER_INTERVAL;
 578
 579         mod_timer(&gpstates->timer, jiffies + msecs_to_jiffies(timer_interval));
 580 }
 581
 582 /**
 583  * gpstate_timer_handler
 584  *
 585  * @data: pointer to cpufreq_policy on which timer was queued
 586  *
 587  * This handler brings down the global pstate closer to the local pstate
 588  * according quadratic equation. Queues a new timer if it is still not equal
 589  * to local pstate
 590  */
 591 void gpstate_timer_handler(unsigned long data)
 592 {
 593         struct cpufreq_policy *policy = (struct cpufreq_policy *)data;
 594         struct global_pstate_info *gpstates = policy->driver_data;
 595         int gpstate_idx;
 596         unsigned int time_diff = jiffies_to_msecs(jiffies)
 597                                         - gpstates->last_sampled_time;
 598         struct powernv_smp_call_data freq_data;
 599
 600         if (!spin_trylock(&gpstates->gpstate_lock))
 601                 return;
 602         /*
 603          * If the timer has migrated to the different cpu then bring
 604          * it back to one of the policy->cpus
 605          */
 606         if (!cpumask_test_cpu(raw_smp_processor_id(), policy->cpus)) {
 607                 gpstates->timer.expires = jiffies + msecs_to_jiffies(1);
 608                 add_timer_on(&gpstates->timer, cpumask_first(policy->cpus));
 609                 spin_unlock(&gpstates->gpstate_lock);
 610                 return;
 611         }
 612
 613         gpstates->last_sampled_time += time_diff;
 614         gpstates->elapsed_time += time_diff;
 615         freq_data.pstate_id = idx_to_pstate(gpstates->last_lpstate_idx);
 616
 617         if ((gpstates->last_gpstate_idx == gpstates->last_lpstate_idx) ||
 618             (gpstates->elapsed_time > MAX_RAMP_DOWN_TIME)) {
 619                 gpstate_idx = pstate_to_idx(freq_data.pstate_id);
 620                 reset_gpstates(policy);
 621                 gpstates->highest_lpstate_idx = gpstate_idx;
 622         } else {
 623                 gpstate_idx = calc_global_pstate(gpstates->elapsed_time,
 624                                                  gpstates->highest_lpstate_idx,
 625                                                  gpstates->last_lpstate_idx);
 626         }
 627
 628         /*
 629          * If local pstate is equal to global pstate, rampdown is over
 630          * So timer is not required to be queued.
 631          */
 632         if (gpstate_idx != gpstates->last_lpstate_idx)
 633                 queue_gpstate_timer(gpstates);
 634
 635         freq_data.gpstate_id = idx_to_pstate(gpstate_idx);
 636         gpstates->last_gpstate_idx = pstate_to_idx(freq_data.gpstate_id);
 637         gpstates->last_lpstate_idx = pstate_to_idx(freq_data.pstate_id);
 638
 639         set_pstate(&freq_data);
 640         spin_unlock(&gpstates->gpstate_lock);
 641 }
 642
 643 /*
 644  * powernv_cpufreq_target_index: Sets the frequency corresponding to
 645  * the cpufreq table entry indexed by new_index on the cpus in the
 646  * mask policy->cpus
 647  */
 648 static int powernv_cpufreq_target_index(struct cpufreq_policy *policy,
 649                                         unsigned int new_index)
 650 {
 651         struct powernv_smp_call_data freq_data;
 652         unsigned int cur_msec, gpstate_idx;
 653         struct global_pstate_info *gpstates = policy->driver_data;
 654
 655         if (unlikely(rebooting) && new_index != get_nominal_index())
 656                 return 0;
 657
 658         if (!throttled) {
 659                 /* we don't want to be preempted while
 660                  * checking if the CPU frequency has been throttled
 661                  */
 662                 preempt_disable();
 663                 powernv_cpufreq_throttle_check(NULL);
 664                 preempt_enable();
 665         }
 666
 667         cur_msec = jiffies_to_msecs(get_jiffies_64());
 668
 669         spin_lock(&gpstates->gpstate_lock);
 670         freq_data.pstate_id = idx_to_pstate(new_index);
 671
 672         if (!gpstates->last_sampled_time) {
 673                 gpstate_idx = new_index;
 674                 gpstates->highest_lpstate_idx = new_index;
 675                 goto gpstates_done;
 676         }
 677
 678         if (gpstates->last_gpstate_idx < new_index) {
 679                 gpstates->elapsed_time += cur_msec -
 680                                                  gpstates->last_sampled_time;
 681
 682                 /*
 683                  * If its has been ramping down for more than MAX_RAMP_DOWN_TIME
 684                  * we should be resetting all global pstate related data. Set it
 685                  * equal to local pstate to start fresh.
 686                  */
 687                 if (gpstates->elapsed_time > MAX_RAMP_DOWN_TIME) {
 688                         reset_gpstates(policy);
 689                         gpstates->highest_lpstate_idx = new_index;
 690                         gpstate_idx = new_index;
 691                 } else {
 692                 /* Elaspsed_time is less than 5 seconds, continue to rampdown */
 693                         gpstate_idx = calc_global_pstate(gpstates->elapsed_time,
 694                                                          gpstates->highest_lpstate_idx,
 695                                                          new_index);
 696                 }
 697         } else {
 698                 reset_gpstates(policy);
 699                 gpstates->highest_lpstate_idx = new_index;
 700                 gpstate_idx = new_index;
 701         }
 702
 703         /*
 704          * If local pstate is equal to global pstate, rampdown is over
 705          * So timer is not required to be queued.
 706          */
 707         if (gpstate_idx != new_index)
 708                 queue_gpstate_timer(gpstates);
 709         else
 710                 del_timer_sync(&gpstates->timer);
 711
 712 gpstates_done:
 713         freq_data.gpstate_id = idx_to_pstate(gpstate_idx);
 714         gpstates->last_sampled_time = cur_msec;
 715         gpstates->last_gpstate_idx = gpstate_idx;
 716         gpstates->last_lpstate_idx = new_index;
 717
 718         spin_unlock(&gpstates->gpstate_lock);
 719
 720         /*
 721          * Use smp_call_function to send IPI and execute the
 722          * mtspr on target CPU.  We could do that without IPI
 723          * if current CPU is within policy->cpus (core)
 724          */
 725         smp_call_function_any(policy->cpus, set_pstate, &freq_data, 1);
 726         return 0;
 727 }
 728
 729 static int powernv_cpufreq_cpu_init(struct cpufreq_policy *policy)
 730 {
 731         int base, i, ret;
 732         struct kernfs_node *kn;
 733         struct global_pstate_info *gpstates;
 734
 735         base = cpu_first_thread_sibling(policy->cpu);
 736
 737         for (i = 0; i < threads_per_core; i++)
 738                 cpumask_set_cpu(base + i, policy->cpus);
 739
 740         kn = kernfs_find_and_get(policy->kobj.sd, throttle_attr_grp.name);
 741         if (!kn) {
 742                 int ret;
 743
 744                 ret = sysfs_create_group(&policy->kobj, &throttle_attr_grp);
 745                 if (ret) {
 746                         pr_info("Failed to create throttle stats directory for cpu %d\n",
 747                                 policy->cpu);
 748                         return ret;
 749                 }
 750         } else {
 751                 kernfs_put(kn);
 752         }
 753
 754         gpstates =  kzalloc(sizeof(*gpstates), GFP_KERNEL);
 755         if (!gpstates)
 756                 return -ENOMEM;
 757
 758         policy->driver_data = gpstates;
 759
 760         /* initialize timer */
 761         init_timer_pinned_deferrable(&gpstates->timer);
 762         gpstates->timer.data = (unsigned long)policy;
 763         gpstates->timer.function = gpstate_timer_handler;
 764         gpstates->timer.expires = jiffies +
 765                                 msecs_to_jiffies(GPSTATE_TIMER_INTERVAL);
 766         spin_lock_init(&gpstates->gpstate_lock);
 767         ret = cpufreq_table_validate_and_show(policy, powernv_freqs);
 768
 769         if (ret < 0)
 770                 kfree(policy->driver_data);
 771
 772         return ret;
 773 }
 774
 775 static int powernv_cpufreq_cpu_exit(struct cpufreq_policy *policy)
 776 {
 777         /* timer is deleted in cpufreq_cpu_stop() */
 778         kfree(policy->driver_data);
 779
 780         return 0;
 781 }
 782
 783 static int powernv_cpufreq_reboot_notifier(struct notifier_block *nb,
 784                                 unsigned long action, void *unused)
 785 {
 786         int cpu;
 787         struct cpufreq_policy *cpu_policy;
 788
 789         rebooting = true;
 790         for_each_online_cpu(cpu) {
 791                 cpu_policy = cpufreq_cpu_get(cpu);
 792                 if (!cpu_policy)
 793                         continue;
 794                 powernv_cpufreq_target_index(cpu_policy, get_nominal_index());
 795                 cpufreq_cpu_put(cpu_policy);
 796         }
 797
 798         return NOTIFY_DONE;
 799 }
 800
 801 static struct notifier_block powernv_cpufreq_reboot_nb = {
 802         .notifier_call = powernv_cpufreq_reboot_notifier,
 803 };
 804
 805 void powernv_cpufreq_work_fn(struct work_struct *work)
 806 {
 807         struct chip *chip = container_of(work, struct chip, throttle);
 808         struct cpufreq_policy *policy;
 809         unsigned int cpu;
 810         cpumask_t mask;
 811
 812         get_online_cpus();
 813         cpumask_and(&mask, &chip->mask, cpu_online_mask);
 814         smp_call_function_any(&mask,
 815                               powernv_cpufreq_throttle_check, NULL, 0);
 816
 817         if (!chip->restore)
 818                 goto out;
 819
 820         chip->restore = false;
 821         for_each_cpu(cpu, &mask) {
 822                 int index;
 823
 824                 policy = cpufreq_cpu_get(cpu);
 825                 if (!policy)
 826                         continue;
 827                 index = cpufreq_table_find_index_c(policy, policy->cur);
 828                 powernv_cpufreq_target_index(policy, index);
 829                 cpumask_andnot(&mask, &mask, policy->cpus);
 830                 cpufreq_cpu_put(policy);
 831         }
 832 out:
 833         put_online_cpus();
 834 }
 835
 836 static int powernv_cpufreq_occ_msg(struct notifier_block *nb,
 837                                    unsigned long msg_type, void *_msg)
 838 {
 839         struct opal_msg *msg = _msg;
 840         struct opal_occ_msg omsg;
 841         int i;
 842
 843         if (msg_type != OPAL_MSG_OCC)
 844                 return 0;
 845
 846         omsg.type = be64_to_cpu(msg->params[0]);
 847
 848         switch (omsg.type) {
 849         case OCC_RESET:
 850                 occ_reset = true;
 851                 pr_info("OCC (On Chip Controller - enforces hard thermal/power limits) Resetting\n");
 852                 /*
 853                  * powernv_cpufreq_throttle_check() is called in
 854                  * target() callback which can detect the throttle state
 855                  * for governors like ondemand.
 856                  * But static governors will not call target() often thus
 857                  * report throttling here.
 858                  */
 859                 if (!throttled) {
 860                         throttled = true;
 861                         pr_warn("CPU frequency is throttled for duration\n");
 862                 }
 863
 864                 break;
 865         case OCC_LOAD:
 866                 pr_info("OCC Loading, CPU frequency is throttled until OCC is started\n");
 867                 break;
 868         case OCC_THROTTLE:
 869                 omsg.chip = be64_to_cpu(msg->params[1]);
 870                 omsg.throttle_status = be64_to_cpu(msg->params[2]);
 871
 872                 if (occ_reset) {
 873                         occ_reset = false;
 874                         throttled = false;
 875                         pr_info("OCC Active, CPU frequency is no longer throttled\n");
 876
 877                         for (i = 0; i < nr_chips; i++) {
 878                                 chips[i].restore = true;
 879                                 schedule_work(&chips[i].throttle);
 880                         }
 881
 882                         return 0;
 883                 }
 884
 885                 for (i = 0; i < nr_chips; i++)
 886                         if (chips[i].id == omsg.chip)
 887                                 break;
 888
 889                 if (omsg.throttle_status >= 0 &&
 890                     omsg.throttle_status <= OCC_MAX_THROTTLE_STATUS) {
 891                         chips[i].throttle_reason = omsg.throttle_status;
 892                         chips[i].reason[omsg.throttle_status]++;
 893                 }
 894
 895                 if (!omsg.throttle_status)
 896                         chips[i].restore = true;
 897
 898                 schedule_work(&chips[i].throttle);
 899         }
 900         return 0;
 901 }
 902
 903 static struct notifier_block powernv_cpufreq_opal_nb = {
 904         .notifier_call  = powernv_cpufreq_occ_msg,
 905         .next           = NULL,
 906         .priority       = 0,
 907 };
 908
 909 static void powernv_cpufreq_stop_cpu(struct cpufreq_policy *policy)
 910 {
 911         struct powernv_smp_call_data freq_data;
 912         struct global_pstate_info *gpstates = policy->driver_data;
 913
 914         freq_data.pstate_id = idx_to_pstate(powernv_pstate_info.min);
 915         freq_data.gpstate_id = idx_to_pstate(powernv_pstate_info.min);
 916         smp_call_function_single(policy->cpu, set_pstate, &freq_data, 1);
 917         del_timer_sync(&gpstates->timer);
 918 }
 919
 920 static struct cpufreq_driver powernv_cpufreq_driver = {
 921         .name           = "powernv-cpufreq",
 922         .flags          = CPUFREQ_CONST_LOOPS,
 923         .init           = powernv_cpufreq_cpu_init,
 924         .exit           = powernv_cpufreq_cpu_exit,
 925         .verify         = cpufreq_generic_frequency_table_verify,
 926         .target_index   = powernv_cpufreq_target_index,
 927         .get            = powernv_cpufreq_get,
 928         .stop_cpu       = powernv_cpufreq_stop_cpu,
 929         .attr           = powernv_cpu_freq_attr,
 930 };
 931
 932 static int init_chip_info(void)
 933 {
 934         unsigned int chip[256];
 935         unsigned int cpu, i;
 936         unsigned int prev_chip_id = UINT_MAX;
 937
 938         for_each_possible_cpu(cpu) {
 939                 unsigned int id = cpu_to_chip_id(cpu);
 940
 941                 if (prev_chip_id != id) {
 942                         prev_chip_id = id;
 943                         chip[nr_chips++] = id;
 944                 }
 945         }
 946
 947         chips = kcalloc(nr_chips, sizeof(struct chip), GFP_KERNEL);
 948         if (!chips)
 949                 return -ENOMEM;
 950
 951         for (i = 0; i < nr_chips; i++) {
 952                 chips[i].id = chip[i];
 953                 cpumask_copy(&chips[i].mask, cpumask_of_node(chip[i]));
 954                 INIT_WORK(&chips[i].throttle, powernv_cpufreq_work_fn);
 955                 for_each_cpu(cpu, &chips[i].mask)
 956                         per_cpu(chip_info, cpu) =  &chips[i];
 957         }
 958
 959         return 0;
 960 }
 961
 962 static inline void clean_chip_info(void)
 963 {
 964         int i;
 965
 966         /* flush any pending work items */
 967         if (chips)
 968                 for (i = 0; i < nr_chips; i++)
 969                         cancel_work_sync(&chips[i].throttle);
 970         kfree(chips);
 971 }
 972
 973 static inline void unregister_all_notifiers(void)
 974 {
 975         opal_message_notifier_unregister(OPAL_MSG_OCC,
 976                                          &powernv_cpufreq_opal_nb);
 977         unregister_reboot_notifier(&powernv_cpufreq_reboot_nb);
 978 }
 979
 980 static int __init powernv_cpufreq_init(void)
 981 {
 982         int rc = 0;
 983
 984         /* Don't probe on pseries (guest) platforms */
 985         if (!firmware_has_feature(FW_FEATURE_OPAL))
 986                 return -ENODEV;
 987
 988         /* Discover pstates from device tree and init */
 989         rc = init_powernv_pstates();
 990         if (rc)
 991                 goto out;
 992
 993         /* Populate chip info */
 994         rc = init_chip_info();
 995         if (rc)
 996                 goto out;
 997
 998         register_reboot_notifier(&powernv_cpufreq_reboot_nb);
 999         opal_message_notifier_register(OPAL_MSG_OCC, &powernv_cpufreq_opal_nb);
1000
1001         rc = cpufreq_register_driver(&powernv_cpufreq_driver);
1002         if (!rc)
1003                 return 0;
1004
1005         pr_info("Failed to register the cpufreq driver (%d)\n", rc);
1006         unregister_all_notifiers();
1007         clean_chip_info();
1008 out:
1009         pr_info("Platform driver disabled. System does not support PState control\n");
1010         return rc;
1011 }
1012 module_init(powernv_cpufreq_init);
1013
1014 static void __exit powernv_cpufreq_exit(void)
1015 {
1016         cpufreq_unregister_driver(&powernv_cpufreq_driver);
1017         unregister_all_notifiers();
1018         clean_chip_info();
1019 }
1020 module_exit(powernv_cpufreq_exit);
1021
1022 MODULE_LICENSE("GPL");
1023 MODULE_AUTHOR("Vaidyanathan Srinivasan <svaidy at linux.vnet.ibm.com>");