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