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 MAX_PSTATE_SHIFT        32
  45 #define LPSTATE_SHIFT           48
  46 #define GPSTATE_SHIFT           56
  47 #define MAX_NR_CHIPS            32
  48
  49 #define MAX_RAMP_DOWN_TIME                              5120
  50 /*
  51  * On an idle system we want the global pstate to ramp-down from max value to
  52  * min over a span of ~5 secs. Also we want it to initially ramp-down slowly and
  53  * then ramp-down rapidly later on.
  54  *
  55  * This gives a percentage rampdown for time elapsed in milliseconds.
  56  * ramp_down_percentage = ((ms * ms) >> 18)
  57  *                      ~= 3.8 * (sec * sec)
  58  *
  59  * At 0 ms      ramp_down_percent = 0
  60  * At 5120 ms   ramp_down_percent = 100
  61  */
  62 #define ramp_down_percent(time)         ((time * time) >> 18)
  63
  64 /* Interval after which the timer is queued to bring down global pstate */
  65 #define GPSTATE_TIMER_INTERVAL                          2000
  66
  67 /**
  68  * struct global_pstate_info -  Per policy data structure to maintain history of
  69  *                              global pstates
  70  * @highest_lpstate_idx:        The local pstate index from which we are
  71  *                              ramping down
  72  * @elapsed_time:               Time in ms spent in ramping down from
  73  *                              highest_lpstate_idx
  74  * @last_sampled_time:          Time from boot in ms when global pstates were
  75  *                              last set
  76  * @last_lpstate_idx,           Last set value of local pstate and global
  77  * last_gpstate_idx             pstate in terms of cpufreq table index
  78  * @timer:                      Is used for ramping down if cpu goes idle for
  79  *                              a long time with global pstate held high
  80  * @gpstate_lock:               A spinlock to maintain synchronization between
  81  *                              routines called by the timer handler and
  82  *                              governer's target_index calls
  83  */
  84 struct global_pstate_info {
  85         int highest_lpstate_idx;
  86         unsigned int elapsed_time;
  87         unsigned int last_sampled_time;
  88         int last_lpstate_idx;
  89         int last_gpstate_idx;
  90         spinlock_t gpstate_lock;
  91         struct timer_list timer;
  92 };
  93
  94 static struct cpufreq_frequency_table powernv_freqs[POWERNV_MAX_PSTATES+1];
  95 u32 pstate_sign_prefix;
  96 static bool rebooting, throttled, occ_reset;
  97
  98 static const char * const throttle_reason[] = {
  99         "No throttling",
 100         "Power Cap",
 101         "Processor Over Temperature",
 102         "Power Supply Failure",
 103         "Over Current",
 104         "OCC Reset"
 105 };
 106
 107 enum throttle_reason_type {
 108         NO_THROTTLE = 0,
 109         POWERCAP,
 110         CPU_OVERTEMP,
 111         POWER_SUPPLY_FAILURE,
 112         OVERCURRENT,
 113         OCC_RESET_THROTTLE,
 114         OCC_MAX_REASON
 115 };
 116
 117 static struct chip {
 118         unsigned int id;
 119         bool throttled;
 120         bool restore;
 121         u8 throttle_reason;
 122         cpumask_t mask;
 123         struct work_struct throttle;
 124         int throttle_turbo;
 125         int throttle_sub_turbo;
 126         int reason[OCC_MAX_REASON];
 127 } *chips;
 128
 129 static int nr_chips;
 130 static DEFINE_PER_CPU(struct chip *, chip_info);
 131
 132 /*
 133  * Note:
 134  * The set of pstates consists of contiguous integers.
 135  * powernv_pstate_info stores the index of the frequency table for
 136  * max, min and nominal frequencies. It also stores number of
 137  * available frequencies.
 138  *
 139  * powernv_pstate_info.nominal indicates the index to the highest
 140  * non-turbo frequency.
 141  */
 142 static struct powernv_pstate_info {
 143         unsigned int min;
 144         unsigned int max;
 145         unsigned int nominal;
 146         unsigned int nr_pstates;
 147         bool wof_enabled;
 148 } powernv_pstate_info;
 149
 150 static inline int extract_pstate(u64 pmsr_val, unsigned int shift)
 151 {
 152         int ret = ((pmsr_val >> shift) & 0xFF);
 153
 154         if (!ret)
 155                 return ret;
 156
 157         return (pstate_sign_prefix | ret);
 158 }
 159
 160 #define extract_local_pstate(x) extract_pstate(x, LPSTATE_SHIFT)
 161 #define extract_global_pstate(x) extract_pstate(x, GPSTATE_SHIFT)
 162 #define extract_max_pstate(x)  extract_pstate(x, MAX_PSTATE_SHIFT)
 163
 164 /* Use following macros for conversions between pstate_id and index */
 165 static inline int idx_to_pstate(unsigned int i)
 166 {
 167         if (unlikely(i >= powernv_pstate_info.nr_pstates)) {
 168                 pr_warn_once("index %u is out of bound\n", i);
 169                 return powernv_freqs[powernv_pstate_info.nominal].driver_data;
 170         }
 171
 172         return powernv_freqs[i].driver_data;
 173 }
 174
 175 static inline unsigned int pstate_to_idx(int pstate)
 176 {
 177         int min = powernv_freqs[powernv_pstate_info.min].driver_data;
 178         int max = powernv_freqs[powernv_pstate_info.max].driver_data;
 179
 180         if (min > 0) {
 181                 if (unlikely((pstate < max) || (pstate > min))) {
 182                         pr_warn_once("pstate %d is out of bound\n", pstate);
 183                         return powernv_pstate_info.nominal;
 184                 }
 185         } else {
 186                 if (unlikely((pstate > max) || (pstate < min))) {
 187                         pr_warn_once("pstate %d is out of bound\n", pstate);
 188                         return powernv_pstate_info.nominal;
 189                 }
 190         }
 191         /*
 192          * abs() is deliberately used so that is works with
 193          * both monotonically increasing and decreasing
 194          * pstate values
 195          */
 196         return abs(pstate - idx_to_pstate(powernv_pstate_info.max));
 197 }
 198
 199 static inline void reset_gpstates(struct cpufreq_policy *policy)
 200 {
 201         struct global_pstate_info *gpstates = policy->driver_data;
 202
 203         gpstates->highest_lpstate_idx = 0;
 204         gpstates->elapsed_time = 0;
 205         gpstates->last_sampled_time = 0;
 206         gpstates->last_lpstate_idx = 0;
 207         gpstates->last_gpstate_idx = 0;
 208 }
 209
 210 /*
 211  * Initialize the freq table based on data obtained
 212  * from the firmware passed via device-tree
 213  */
 214 static int init_powernv_pstates(void)
 215 {
 216         struct device_node *power_mgt;
 217         int i, nr_pstates = 0;
 218         const __be32 *pstate_ids, *pstate_freqs;
 219         u32 len_ids, len_freqs;
 220         u32 pstate_min, pstate_max, pstate_nominal;
 221         u32 pstate_turbo, pstate_ultra_turbo;
 222
 223         power_mgt = of_find_node_by_path("/ibm,opal/power-mgt");
 224         if (!power_mgt) {
 225                 pr_warn("power-mgt node not found\n");
 226                 return -ENODEV;
 227         }
 228
 229         if (of_property_read_u32(power_mgt, "ibm,pstate-min", &pstate_min)) {
 230                 pr_warn("ibm,pstate-min node not found\n");
 231                 return -ENODEV;
 232         }
 233
 234         if (of_property_read_u32(power_mgt, "ibm,pstate-max", &pstate_max)) {
 235                 pr_warn("ibm,pstate-max node not found\n");
 236                 return -ENODEV;
 237         }
 238
 239         if (of_property_read_u32(power_mgt, "ibm,pstate-nominal",
 240                                  &pstate_nominal)) {
 241                 pr_warn("ibm,pstate-nominal not found\n");
 242                 return -ENODEV;
 243         }
 244
 245         if (of_property_read_u32(power_mgt, "ibm,pstate-ultra-turbo",
 246                                  &pstate_ultra_turbo)) {
 247                 powernv_pstate_info.wof_enabled = false;
 248                 goto next;
 249         }
 250
 251         if (of_property_read_u32(power_mgt, "ibm,pstate-turbo",
 252                                  &pstate_turbo)) {
 253                 powernv_pstate_info.wof_enabled = false;
 254                 goto next;
 255         }
 256
 257         if (pstate_turbo == pstate_ultra_turbo)
 258                 powernv_pstate_info.wof_enabled = false;
 259         else
 260                 powernv_pstate_info.wof_enabled = true;
 261
 262 next:
 263         pr_info("cpufreq pstate min %d nominal %d max %d\n", pstate_min,
 264                 pstate_nominal, pstate_max);
 265         pr_info("Workload Optimized Frequency is %s in the platform\n",
 266                 (powernv_pstate_info.wof_enabled) ? "enabled" : "disabled");
 267
 268         pstate_ids = of_get_property(power_mgt, "ibm,pstate-ids", &len_ids);
 269         if (!pstate_ids) {
 270                 pr_warn("ibm,pstate-ids not found\n");
 271                 return -ENODEV;
 272         }
 273
 274         pstate_freqs = of_get_property(power_mgt, "ibm,pstate-frequencies-mhz",
 275                                       &len_freqs);
 276         if (!pstate_freqs) {
 277                 pr_warn("ibm,pstate-frequencies-mhz not found\n");
 278                 return -ENODEV;
 279         }
 280
 281         if (len_ids != len_freqs) {
 282                 pr_warn("Entries in ibm,pstate-ids and "
 283                         "ibm,pstate-frequencies-mhz does not match\n");
 284         }
 285
 286         nr_pstates = min(len_ids, len_freqs) / sizeof(u32);
 287         if (!nr_pstates) {
 288                 pr_warn("No PStates found\n");
 289                 return -ENODEV;
 290         }
 291
 292         powernv_pstate_info.nr_pstates = nr_pstates;
 293         pr_debug("NR PStates %d\n", nr_pstates);
 294
 295         pstate_sign_prefix = pstate_min & ~0xFF;
 296
 297         for (i = 0; i < nr_pstates; i++) {
 298                 u32 id = be32_to_cpu(pstate_ids[i]);
 299                 u32 freq = be32_to_cpu(pstate_freqs[i]);
 300
 301                 pr_debug("PState id %d freq %d MHz\n", id, freq);
 302                 powernv_freqs[i].frequency = freq * 1000; /* kHz */
 303                 powernv_freqs[i].driver_data = id;
 304
 305                 if (id == pstate_max)
 306                         powernv_pstate_info.max = i;
 307                 if (id == pstate_nominal)
 308                         powernv_pstate_info.nominal = i;
 309                 if (id == pstate_min)
 310                         powernv_pstate_info.min = i;
 311
 312                 if (powernv_pstate_info.wof_enabled && id == pstate_turbo) {
 313                         int j;
 314
 315                         for (j = i - 1; j >= (int)powernv_pstate_info.max; j--)
 316                                 powernv_freqs[j].flags = CPUFREQ_BOOST_FREQ;
 317                 }
 318         }
 319
 320         /* End of list marker entry */
 321         powernv_freqs[i].frequency = CPUFREQ_TABLE_END;
 322         return 0;
 323 }
 324
 325 /* Returns the CPU frequency corresponding to the pstate_id. */
 326 static unsigned int pstate_id_to_freq(int pstate_id)
 327 {
 328         int i;
 329
 330         i = pstate_to_idx(pstate_id);
 331         if (i >= powernv_pstate_info.nr_pstates || i < 0) {
 332                 pr_warn("PState id %d outside of PState table, "
 333                         "reporting nominal id %d instead\n",
 334                         pstate_id, idx_to_pstate(powernv_pstate_info.nominal));
 335                 i = powernv_pstate_info.nominal;
 336         }
 337
 338         return powernv_freqs[i].frequency;
 339 }
 340
 341 /*
 342  * cpuinfo_nominal_freq_show - Show the nominal CPU frequency as indicated by
 343  * the firmware
 344  */
 345 static ssize_t cpuinfo_nominal_freq_show(struct cpufreq_policy *policy,
 346                                         char *buf)
 347 {
 348         return sprintf(buf, "%u\n",
 349                 powernv_freqs[powernv_pstate_info.nominal].frequency);
 350 }
 351
 352 struct freq_attr cpufreq_freq_attr_cpuinfo_nominal_freq =
 353         __ATTR_RO(cpuinfo_nominal_freq);
 354
 355 #define SCALING_BOOST_FREQS_ATTR_INDEX          2
 356
 357 static struct freq_attr *powernv_cpu_freq_attr[] = {
 358         &cpufreq_freq_attr_scaling_available_freqs,
 359         &cpufreq_freq_attr_cpuinfo_nominal_freq,
 360         &cpufreq_freq_attr_scaling_boost_freqs,
 361         NULL,
 362 };
 363
 364 #define throttle_attr(name, member)                                     \
 365 static ssize_t name##_show(struct cpufreq_policy *policy, char *buf)    \
 366 {                                                                       \
 367         struct chip *chip = per_cpu(chip_info, policy->cpu);            \
 368                                                                         \
 369         return sprintf(buf, "%u\n", chip->member);                      \
 370 }                                                                       \
 371                                                                         \
 372 static struct freq_attr throttle_attr_##name = __ATTR_RO(name)          \
 373
 374 throttle_attr(unthrottle, reason[NO_THROTTLE]);
 375 throttle_attr(powercap, reason[POWERCAP]);
 376 throttle_attr(overtemp, reason[CPU_OVERTEMP]);
 377 throttle_attr(supply_fault, reason[POWER_SUPPLY_FAILURE]);
 378 throttle_attr(overcurrent, reason[OVERCURRENT]);
 379 throttle_attr(occ_reset, reason[OCC_RESET_THROTTLE]);
 380 throttle_attr(turbo_stat, throttle_turbo);
 381 throttle_attr(sub_turbo_stat, throttle_sub_turbo);
 382
 383 static struct attribute *throttle_attrs[] = {
 384         &throttle_attr_unthrottle.attr,
 385         &throttle_attr_powercap.attr,
 386         &throttle_attr_overtemp.attr,
 387         &throttle_attr_supply_fault.attr,
 388         &throttle_attr_overcurrent.attr,
 389         &throttle_attr_occ_reset.attr,
 390         &throttle_attr_turbo_stat.attr,
 391         &throttle_attr_sub_turbo_stat.attr,
 392         NULL,
 393 };
 394
 395 static const struct attribute_group throttle_attr_grp = {
 396         .name   = "throttle_stats",
 397         .attrs  = throttle_attrs,
 398 };
 399
 400 /* Helper routines */
 401
 402 /* Access helpers to power mgt SPR */
 403
 404 static inline unsigned long get_pmspr(unsigned long sprn)
 405 {
 406         switch (sprn) {
 407         case SPRN_PMCR:
 408                 return mfspr(SPRN_PMCR);
 409
 410         case SPRN_PMICR:
 411                 return mfspr(SPRN_PMICR);
 412
 413         case SPRN_PMSR:
 414                 return mfspr(SPRN_PMSR);
 415         }
 416         BUG();
 417 }
 418
 419 static inline void set_pmspr(unsigned long sprn, unsigned long val)
 420 {
 421         switch (sprn) {
 422         case SPRN_PMCR:
 423                 mtspr(SPRN_PMCR, val);
 424                 return;
 425
 426         case SPRN_PMICR:
 427                 mtspr(SPRN_PMICR, val);
 428                 return;
 429         }
 430         BUG();
 431 }
 432
 433 /*
 434  * Use objects of this type to query/update
 435  * pstates on a remote CPU via smp_call_function.
 436  */
 437 struct powernv_smp_call_data {
 438         unsigned int freq;
 439         int pstate_id;
 440         int gpstate_id;
 441 };
 442
 443 /*
 444  * powernv_read_cpu_freq: Reads the current frequency on this CPU.
 445  *
 446  * Called via smp_call_function.
 447  *
 448  * Note: The caller of the smp_call_function should pass an argument of
 449  * the type 'struct powernv_smp_call_data *' along with this function.
 450  *
 451  * The current frequency on this CPU will be returned via
 452  * ((struct powernv_smp_call_data *)arg)->freq;
 453  */
 454 static void powernv_read_cpu_freq(void *arg)
 455 {
 456         unsigned long pmspr_val;
 457         struct powernv_smp_call_data *freq_data = arg;
 458
 459         pmspr_val = get_pmspr(SPRN_PMSR);
 460         freq_data->pstate_id = extract_local_pstate(pmspr_val);
 461         freq_data->freq = pstate_id_to_freq(freq_data->pstate_id);
 462
 463         pr_debug("cpu %d pmsr %016lX pstate_id %d frequency %d kHz\n",
 464                 raw_smp_processor_id(), pmspr_val, freq_data->pstate_id,
 465                 freq_data->freq);
 466 }
 467
 468 /*
 469  * powernv_cpufreq_get: Returns the CPU frequency as reported by the
 470  * firmware for CPU 'cpu'. This value is reported through the sysfs
 471  * file cpuinfo_cur_freq.
 472  */
 473 static unsigned int powernv_cpufreq_get(unsigned int cpu)
 474 {
 475         struct powernv_smp_call_data freq_data;
 476
 477         smp_call_function_any(cpu_sibling_mask(cpu), powernv_read_cpu_freq,
 478                         &freq_data, 1);
 479
 480         return freq_data.freq;
 481 }
 482
 483 /*
 484  * set_pstate: Sets the pstate on this CPU.
 485  *
 486  * This is called via an smp_call_function.
 487  *
 488  * The caller must ensure that freq_data is of the type
 489  * (struct powernv_smp_call_data *) and the pstate_id which needs to be set
 490  * on this CPU should be present in freq_data->pstate_id.
 491  */
 492 static void set_pstate(void *data)
 493 {
 494         unsigned long val;
 495         struct powernv_smp_call_data *freq_data = data;
 496         unsigned long pstate_ul = freq_data->pstate_id;
 497         unsigned long gpstate_ul = freq_data->gpstate_id;
 498
 499         val = get_pmspr(SPRN_PMCR);
 500         val = val & 0x0000FFFFFFFFFFFFULL;
 501
 502         pstate_ul = pstate_ul & 0xFF;
 503         gpstate_ul = gpstate_ul & 0xFF;
 504
 505         /* Set both global(bits 56..63) and local(bits 48..55) PStates */
 506         val = val | (gpstate_ul << 56) | (pstate_ul << 48);
 507
 508         pr_debug("Setting cpu %d pmcr to %016lX\n",
 509                         raw_smp_processor_id(), val);
 510         set_pmspr(SPRN_PMCR, val);
 511 }
 512
 513 /*
 514  * get_nominal_index: Returns the index corresponding to the nominal
 515  * pstate in the cpufreq table
 516  */
 517 static inline unsigned int get_nominal_index(void)
 518 {
 519         return powernv_pstate_info.nominal;
 520 }
 521
 522 static void powernv_cpufreq_throttle_check(void *data)
 523 {
 524         struct chip *chip;
 525         unsigned int cpu = smp_processor_id();
 526         unsigned long pmsr;
 527         int pmsr_pmax;
 528         unsigned int pmsr_pmax_idx;
 529
 530         pmsr = get_pmspr(SPRN_PMSR);
 531         chip = this_cpu_read(chip_info);
 532
 533         /* Check for Pmax Capping */
 534         pmsr_pmax = extract_max_pstate(pmsr);
 535         pmsr_pmax_idx = pstate_to_idx(pmsr_pmax);
 536         if (pmsr_pmax_idx != powernv_pstate_info.max) {
 537                 if (chip->throttled)
 538                         goto next;
 539                 chip->throttled = true;
 540                 if (pmsr_pmax_idx > powernv_pstate_info.nominal) {
 541                         pr_warn_once("CPU %d on Chip %u has Pmax(%d) reduced below nominal frequency(%d)\n",
 542                                      cpu, chip->id, pmsr_pmax,
 543                                      idx_to_pstate(powernv_pstate_info.nominal));
 544                         chip->throttle_sub_turbo++;
 545                 } else {
 546                         chip->throttle_turbo++;
 547                 }
 548                 trace_powernv_throttle(chip->id,
 549                                       throttle_reason[chip->throttle_reason],
 550                                       pmsr_pmax);
 551         } else if (chip->throttled) {
 552                 chip->throttled = false;
 553                 trace_powernv_throttle(chip->id,
 554                                       throttle_reason[chip->throttle_reason],
 555                                       pmsr_pmax);
 556         }
 557
 558         /* Check if Psafe_mode_active is set in PMSR. */
 559 next:
 560         if (pmsr & PMSR_PSAFE_ENABLE) {
 561                 throttled = true;
 562                 pr_info("Pstate set to safe frequency\n");
 563         }
 564
 565         /* Check if SPR_EM_DISABLE is set in PMSR */
 566         if (pmsr & PMSR_SPR_EM_DISABLE) {
 567                 throttled = true;
 568                 pr_info("Frequency Control disabled from OS\n");
 569         }
 570
 571         if (throttled) {
 572                 pr_info("PMSR = %16lx\n", pmsr);
 573                 pr_warn("CPU Frequency could be throttled\n");
 574         }
 575 }
 576
 577 /**
 578  * calc_global_pstate - Calculate global pstate
 579  * @elapsed_time:               Elapsed time in milliseconds
 580  * @local_pstate_idx:           New local pstate
 581  * @highest_lpstate_idx:        pstate from which its ramping down
 582  *
 583  * Finds the appropriate global pstate based on the pstate from which its
 584  * ramping down and the time elapsed in ramping down. It follows a quadratic
 585  * equation which ensures that it reaches ramping down to pmin in 5sec.
 586  */
 587 static inline int calc_global_pstate(unsigned int elapsed_time,
 588                                      int highest_lpstate_idx,
 589                                      int local_pstate_idx)
 590 {
 591         int index_diff;
 592
 593         /*
 594          * Using ramp_down_percent we get the percentage of rampdown
 595          * that we are expecting to be dropping. Difference between
 596          * highest_lpstate_idx and powernv_pstate_info.min will give a absolute
 597          * number of how many pstates we will drop eventually by the end of
 598          * 5 seconds, then just scale it get the number pstates to be dropped.
 599          */
 600         index_diff =  ((int)ramp_down_percent(elapsed_time) *
 601                         (powernv_pstate_info.min - highest_lpstate_idx)) / 100;
 602
 603         /* Ensure that global pstate is >= to local pstate */
 604         if (highest_lpstate_idx + index_diff >= local_pstate_idx)
 605                 return local_pstate_idx;
 606         else
 607                 return highest_lpstate_idx + index_diff;
 608 }
 609
 610 static inline void  queue_gpstate_timer(struct global_pstate_info *gpstates)
 611 {
 612         unsigned int timer_interval;
 613
 614         /*
 615          * Setting up timer to fire after GPSTATE_TIMER_INTERVAL ms, But
 616          * if it exceeds MAX_RAMP_DOWN_TIME ms for ramp down time.
 617          * Set timer such that it fires exactly at MAX_RAMP_DOWN_TIME
 618          * seconds of ramp down time.
 619          */
 620         if ((gpstates->elapsed_time + GPSTATE_TIMER_INTERVAL)
 621              > MAX_RAMP_DOWN_TIME)
 622                 timer_interval = MAX_RAMP_DOWN_TIME - gpstates->elapsed_time;
 623         else
 624                 timer_interval = GPSTATE_TIMER_INTERVAL;
 625
 626         mod_timer(&gpstates->timer, jiffies + msecs_to_jiffies(timer_interval));
 627 }
 628
 629 /**
 630  * gpstate_timer_handler
 631  *
 632  * @data: pointer to cpufreq_policy on which timer was queued
 633  *
 634  * This handler brings down the global pstate closer to the local pstate
 635  * according quadratic equation. Queues a new timer if it is still not equal
 636  * to local pstate
 637  */
 638 void gpstate_timer_handler(unsigned long data)
 639 {
 640         struct cpufreq_policy *policy = (struct cpufreq_policy *)data;
 641         struct global_pstate_info *gpstates = policy->driver_data;
 642         int gpstate_idx, lpstate_idx;
 643         unsigned long val;
 644         unsigned int time_diff = jiffies_to_msecs(jiffies)
 645                                         - gpstates->last_sampled_time;
 646         struct powernv_smp_call_data freq_data;
 647
 648         if (!spin_trylock(&gpstates->gpstate_lock))
 649                 return;
 650         /*
 651          * If the timer has migrated to the different cpu then bring
 652          * it back to one of the policy->cpus
 653          */
 654         if (!cpumask_test_cpu(raw_smp_processor_id(), policy->cpus)) {
 655                 gpstates->timer.expires = jiffies + msecs_to_jiffies(1);
 656                 add_timer_on(&gpstates->timer, cpumask_first(policy->cpus));
 657                 spin_unlock(&gpstates->gpstate_lock);
 658                 return;
 659         }
 660
 661         /*
 662          * If PMCR was last updated was using fast_swtich then
 663          * We may have wrong in gpstate->last_lpstate_idx
 664          * value. Hence, read from PMCR to get correct data.
 665          */
 666         val = get_pmspr(SPRN_PMCR);
 667         freq_data.gpstate_id = extract_global_pstate(val);
 668         freq_data.pstate_id = extract_local_pstate(val);
 669         if (freq_data.gpstate_id  == freq_data.pstate_id) {
 670                 reset_gpstates(policy);
 671                 spin_unlock(&gpstates->gpstate_lock);
 672                 return;
 673         }
 674
 675         gpstates->last_sampled_time += time_diff;
 676         gpstates->elapsed_time += time_diff;
 677
 678         if (gpstates->elapsed_time > MAX_RAMP_DOWN_TIME) {
 679                 gpstate_idx = pstate_to_idx(freq_data.pstate_id);
 680                 lpstate_idx = gpstate_idx;
 681                 reset_gpstates(policy);
 682                 gpstates->highest_lpstate_idx = gpstate_idx;
 683         } else {
 684                 lpstate_idx = pstate_to_idx(freq_data.pstate_id);
 685                 gpstate_idx = calc_global_pstate(gpstates->elapsed_time,
 686                                                  gpstates->highest_lpstate_idx,
 687                                                  lpstate_idx);
 688         }
 689         freq_data.gpstate_id = idx_to_pstate(gpstate_idx);
 690         gpstates->last_gpstate_idx = gpstate_idx;
 691         gpstates->last_lpstate_idx = lpstate_idx;
 692         /*
 693          * If local pstate is equal to global pstate, rampdown is over
 694          * So timer is not required to be queued.
 695          */
 696         if (gpstate_idx != gpstates->last_lpstate_idx)
 697                 queue_gpstate_timer(gpstates);
 698
 699         set_pstate(&freq_data);
 700         spin_unlock(&gpstates->gpstate_lock);
 701 }
 702
 703 /*
 704  * powernv_cpufreq_target_index: Sets the frequency corresponding to
 705  * the cpufreq table entry indexed by new_index on the cpus in the
 706  * mask policy->cpus
 707  */
 708 static int powernv_cpufreq_target_index(struct cpufreq_policy *policy,
 709                                         unsigned int new_index)
 710 {
 711         struct powernv_smp_call_data freq_data;
 712         unsigned int cur_msec, gpstate_idx;
 713         struct global_pstate_info *gpstates = policy->driver_data;
 714
 715         if (unlikely(rebooting) && new_index != get_nominal_index())
 716                 return 0;
 717
 718         if (!throttled) {
 719                 /* we don't want to be preempted while
 720                  * checking if the CPU frequency has been throttled
 721                  */
 722                 preempt_disable();
 723                 powernv_cpufreq_throttle_check(NULL);
 724                 preempt_enable();
 725         }
 726
 727         cur_msec = jiffies_to_msecs(get_jiffies_64());
 728
 729         spin_lock(&gpstates->gpstate_lock);
 730         freq_data.pstate_id = idx_to_pstate(new_index);
 731
 732         if (!gpstates->last_sampled_time) {
 733                 gpstate_idx = new_index;
 734                 gpstates->highest_lpstate_idx = new_index;
 735                 goto gpstates_done;
 736         }
 737
 738         if (gpstates->last_gpstate_idx < new_index) {
 739                 gpstates->elapsed_time += cur_msec -
 740                                                  gpstates->last_sampled_time;
 741
 742                 /*
 743                  * If its has been ramping down for more than MAX_RAMP_DOWN_TIME
 744                  * we should be resetting all global pstate related data. Set it
 745                  * equal to local pstate to start fresh.
 746                  */
 747                 if (gpstates->elapsed_time > MAX_RAMP_DOWN_TIME) {
 748                         reset_gpstates(policy);
 749                         gpstates->highest_lpstate_idx = new_index;
 750                         gpstate_idx = new_index;
 751                 } else {
 752                 /* Elaspsed_time is less than 5 seconds, continue to rampdown */
 753                         gpstate_idx = calc_global_pstate(gpstates->elapsed_time,
 754                                                          gpstates->highest_lpstate_idx,
 755                                                          new_index);
 756                 }
 757         } else {
 758                 reset_gpstates(policy);
 759                 gpstates->highest_lpstate_idx = new_index;
 760                 gpstate_idx = new_index;
 761         }
 762
 763         /*
 764          * If local pstate is equal to global pstate, rampdown is over
 765          * So timer is not required to be queued.
 766          */
 767         if (gpstate_idx != new_index)
 768                 queue_gpstate_timer(gpstates);
 769         else
 770                 del_timer_sync(&gpstates->timer);
 771
 772 gpstates_done:
 773         freq_data.gpstate_id = idx_to_pstate(gpstate_idx);
 774         gpstates->last_sampled_time = cur_msec;
 775         gpstates->last_gpstate_idx = gpstate_idx;
 776         gpstates->last_lpstate_idx = new_index;
 777
 778         spin_unlock(&gpstates->gpstate_lock);
 779
 780         /*
 781          * Use smp_call_function to send IPI and execute the
 782          * mtspr on target CPU.  We could do that without IPI
 783          * if current CPU is within policy->cpus (core)
 784          */
 785         smp_call_function_any(policy->cpus, set_pstate, &freq_data, 1);
 786         return 0;
 787 }
 788
 789 static int powernv_cpufreq_cpu_init(struct cpufreq_policy *policy)
 790 {
 791         int base, i, ret;
 792         struct kernfs_node *kn;
 793         struct global_pstate_info *gpstates;
 794
 795         base = cpu_first_thread_sibling(policy->cpu);
 796
 797         for (i = 0; i < threads_per_core; i++)
 798                 cpumask_set_cpu(base + i, policy->cpus);
 799
 800         kn = kernfs_find_and_get(policy->kobj.sd, throttle_attr_grp.name);
 801         if (!kn) {
 802                 int ret;
 803
 804                 ret = sysfs_create_group(&policy->kobj, &throttle_attr_grp);
 805                 if (ret) {
 806                         pr_info("Failed to create throttle stats directory for cpu %d\n",
 807                                 policy->cpu);
 808                         return ret;
 809                 }
 810         } else {
 811                 kernfs_put(kn);
 812         }
 813
 814         gpstates =  kzalloc(sizeof(*gpstates), GFP_KERNEL);
 815         if (!gpstates)
 816                 return -ENOMEM;
 817
 818         policy->driver_data = gpstates;
 819
 820         /* initialize timer */
 821         init_timer_pinned_deferrable(&gpstates->timer);
 822         gpstates->timer.data = (unsigned long)policy;
 823         gpstates->timer.function = gpstate_timer_handler;
 824         gpstates->timer.expires = jiffies +
 825                                 msecs_to_jiffies(GPSTATE_TIMER_INTERVAL);
 826         spin_lock_init(&gpstates->gpstate_lock);
 827         ret = cpufreq_table_validate_and_show(policy, powernv_freqs);
 828
 829         if (ret < 0) {
 830                 kfree(policy->driver_data);
 831                 return ret;
 832         }
 833
 834         policy->fast_switch_possible = true;
 835         return ret;
 836 }
 837
 838 static int powernv_cpufreq_cpu_exit(struct cpufreq_policy *policy)
 839 {
 840         /* timer is deleted in cpufreq_cpu_stop() */
 841         kfree(policy->driver_data);
 842
 843         return 0;
 844 }
 845
 846 static int powernv_cpufreq_reboot_notifier(struct notifier_block *nb,
 847                                 unsigned long action, void *unused)
 848 {
 849         int cpu;
 850         struct cpufreq_policy *cpu_policy;
 851
 852         rebooting = true;
 853         for_each_online_cpu(cpu) {
 854                 cpu_policy = cpufreq_cpu_get(cpu);
 855                 if (!cpu_policy)
 856                         continue;
 857                 powernv_cpufreq_target_index(cpu_policy, get_nominal_index());
 858                 cpufreq_cpu_put(cpu_policy);
 859         }
 860
 861         return NOTIFY_DONE;
 862 }
 863
 864 static struct notifier_block powernv_cpufreq_reboot_nb = {
 865         .notifier_call = powernv_cpufreq_reboot_notifier,
 866 };
 867
 868 void powernv_cpufreq_work_fn(struct work_struct *work)
 869 {
 870         struct chip *chip = container_of(work, struct chip, throttle);
 871         struct cpufreq_policy *policy;
 872         unsigned int cpu;
 873         cpumask_t mask;
 874
 875         get_online_cpus();
 876         cpumask_and(&mask, &chip->mask, cpu_online_mask);
 877         smp_call_function_any(&mask,
 878                               powernv_cpufreq_throttle_check, NULL, 0);
 879
 880         if (!chip->restore)
 881                 goto out;
 882
 883         chip->restore = false;
 884         for_each_cpu(cpu, &mask) {
 885                 int index;
 886
 887                 policy = cpufreq_cpu_get(cpu);
 888                 if (!policy)
 889                         continue;
 890                 index = cpufreq_table_find_index_c(policy, policy->cur);
 891                 powernv_cpufreq_target_index(policy, index);
 892                 cpumask_andnot(&mask, &mask, policy->cpus);
 893                 cpufreq_cpu_put(policy);
 894         }
 895 out:
 896         put_online_cpus();
 897 }
 898
 899 static int powernv_cpufreq_occ_msg(struct notifier_block *nb,
 900                                    unsigned long msg_type, void *_msg)
 901 {
 902         struct opal_msg *msg = _msg;
 903         struct opal_occ_msg omsg;
 904         int i;
 905
 906         if (msg_type != OPAL_MSG_OCC)
 907                 return 0;
 908
 909         omsg.type = be64_to_cpu(msg->params[0]);
 910
 911         switch (omsg.type) {
 912         case OCC_RESET:
 913                 occ_reset = true;
 914                 pr_info("OCC (On Chip Controller - enforces hard thermal/power limits) Resetting\n");
 915                 /*
 916                  * powernv_cpufreq_throttle_check() is called in
 917                  * target() callback which can detect the throttle state
 918                  * for governors like ondemand.
 919                  * But static governors will not call target() often thus
 920                  * report throttling here.
 921                  */
 922                 if (!throttled) {
 923                         throttled = true;
 924                         pr_warn("CPU frequency is throttled for duration\n");
 925                 }
 926
 927                 break;
 928         case OCC_LOAD:
 929                 pr_info("OCC Loading, CPU frequency is throttled until OCC is started\n");
 930                 break;
 931         case OCC_THROTTLE:
 932                 omsg.chip = be64_to_cpu(msg->params[1]);
 933                 omsg.throttle_status = be64_to_cpu(msg->params[2]);
 934
 935                 if (occ_reset) {
 936                         occ_reset = false;
 937                         throttled = false;
 938                         pr_info("OCC Active, CPU frequency is no longer throttled\n");
 939
 940                         for (i = 0; i < nr_chips; i++) {
 941                                 chips[i].restore = true;
 942                                 schedule_work(&chips[i].throttle);
 943                         }
 944
 945                         return 0;
 946                 }
 947
 948                 for (i = 0; i < nr_chips; i++)
 949                         if (chips[i].id == omsg.chip)
 950                                 break;
 951
 952                 if (omsg.throttle_status >= 0 &&
 953                     omsg.throttle_status <= OCC_MAX_THROTTLE_STATUS) {
 954                         chips[i].throttle_reason = omsg.throttle_status;
 955                         chips[i].reason[omsg.throttle_status]++;
 956                 }
 957
 958                 if (!omsg.throttle_status)
 959                         chips[i].restore = true;
 960
 961                 schedule_work(&chips[i].throttle);
 962         }
 963         return 0;
 964 }
 965
 966 static struct notifier_block powernv_cpufreq_opal_nb = {
 967         .notifier_call  = powernv_cpufreq_occ_msg,
 968         .next           = NULL,
 969         .priority       = 0,
 970 };
 971
 972 static void powernv_cpufreq_stop_cpu(struct cpufreq_policy *policy)
 973 {
 974         struct powernv_smp_call_data freq_data;
 975         struct global_pstate_info *gpstates = policy->driver_data;
 976
 977         freq_data.pstate_id = idx_to_pstate(powernv_pstate_info.min);
 978         freq_data.gpstate_id = idx_to_pstate(powernv_pstate_info.min);
 979         smp_call_function_single(policy->cpu, set_pstate, &freq_data, 1);
 980         del_timer_sync(&gpstates->timer);
 981 }
 982
 983 static unsigned int powernv_fast_switch(struct cpufreq_policy *policy,
 984                                         unsigned int target_freq)
 985 {
 986         int index;
 987         struct powernv_smp_call_data freq_data;
 988
 989         index = cpufreq_table_find_index_dl(policy, target_freq);
 990         freq_data.pstate_id = powernv_freqs[index].driver_data;
 991         freq_data.gpstate_id = powernv_freqs[index].driver_data;
 992         set_pstate(&freq_data);
 993
 994         return powernv_freqs[index].frequency;
 995 }
 996
 997 static struct cpufreq_driver powernv_cpufreq_driver = {
 998         .name           = "powernv-cpufreq",
 999         .flags          = CPUFREQ_CONST_LOOPS,
1000         .init           = powernv_cpufreq_cpu_init,
1001         .exit           = powernv_cpufreq_cpu_exit,
1002         .verify         = cpufreq_generic_frequency_table_verify,
1003         .target_index   = powernv_cpufreq_target_index,
1004         .fast_switch    = powernv_fast_switch,
1005         .get            = powernv_cpufreq_get,
1006         .stop_cpu       = powernv_cpufreq_stop_cpu,
1007         .attr           = powernv_cpu_freq_attr,
1008 };
1009
1010 static int init_chip_info(void)
1011 {
1012         unsigned int *chip;
1013         unsigned int cpu, i;
1014         unsigned int prev_chip_id = UINT_MAX;
1015         cpumask_t *chip_cpu_mask;
1016         int ret = 0;
1017
1018         chip = kcalloc(num_possible_cpus(), sizeof(*chip), GFP_KERNEL);
1019         if (!chip)
1020                 return -ENOMEM;
1021
1022         /* Allocate a chip cpu mask large enough to fit mask for all chips */
1023         chip_cpu_mask = kcalloc(MAX_NR_CHIPS, sizeof(cpumask_t), GFP_KERNEL);
1024         if (!chip_cpu_mask) {
1025                 ret = -ENOMEM;
1026                 goto free_and_return;
1027         }
1028
1029         for_each_possible_cpu(cpu) {
1030                 unsigned int id = cpu_to_chip_id(cpu);
1031
1032                 if (prev_chip_id != id) {
1033                         prev_chip_id = id;
1034                         chip[nr_chips++] = id;
1035                 }
1036                 cpumask_set_cpu(cpu, &chip_cpu_mask[nr_chips-1]);
1037         }
1038
1039         chips = kcalloc(nr_chips, sizeof(struct chip), GFP_KERNEL);
1040         if (!chips) {
1041                 ret = -ENOMEM;
1042                 goto out_free_chip_cpu_mask;
1043         }
1044
1045         for (i = 0; i < nr_chips; i++) {
1046                 chips[i].id = chip[i];
1047                 cpumask_copy(&chips[i].mask, &chip_cpu_mask[i]);
1048                 INIT_WORK(&chips[i].throttle, powernv_cpufreq_work_fn);
1049                 for_each_cpu(cpu, &chips[i].mask)
1050                         per_cpu(chip_info, cpu) =  &chips[i];
1051         }
1052
1053 out_free_chip_cpu_mask:
1054         kfree(chip_cpu_mask);
1055 free_and_return:
1056         kfree(chip);
1057         return ret;
1058 }
1059
1060 static inline void clean_chip_info(void)
1061 {
1062         int i;
1063
1064         /* flush any pending work items */
1065         if (chips)
1066                 for (i = 0; i < nr_chips; i++)
1067                         cancel_work_sync(&chips[i].throttle);
1068         kfree(chips);
1069 }
1070
1071 static inline void unregister_all_notifiers(void)
1072 {
1073         opal_message_notifier_unregister(OPAL_MSG_OCC,
1074                                          &powernv_cpufreq_opal_nb);
1075         unregister_reboot_notifier(&powernv_cpufreq_reboot_nb);
1076 }
1077
1078 static int __init powernv_cpufreq_init(void)
1079 {
1080         int rc = 0;
1081
1082         /* Don't probe on pseries (guest) platforms */
1083         if (!firmware_has_feature(FW_FEATURE_OPAL))
1084                 return -ENODEV;
1085
1086         /* Discover pstates from device tree and init */
1087         rc = init_powernv_pstates();
1088         if (rc)
1089                 goto out;
1090
1091         /* Populate chip info */
1092         rc = init_chip_info();
1093         if (rc)
1094                 goto out;
1095
1096         register_reboot_notifier(&powernv_cpufreq_reboot_nb);
1097         opal_message_notifier_register(OPAL_MSG_OCC, &powernv_cpufreq_opal_nb);
1098
1099         if (powernv_pstate_info.wof_enabled)
1100                 powernv_cpufreq_driver.boost_enabled = true;
1101         else
1102                 powernv_cpu_freq_attr[SCALING_BOOST_FREQS_ATTR_INDEX] = NULL;
1103
1104         rc = cpufreq_register_driver(&powernv_cpufreq_driver);
1105         if (rc) {
1106                 pr_info("Failed to register the cpufreq driver (%d)\n", rc);
1107                 goto cleanup_notifiers;
1108         }
1109
1110         if (powernv_pstate_info.wof_enabled)
1111                 cpufreq_enable_boost_support();
1112
1113         return 0;
1114 cleanup_notifiers:
1115         unregister_all_notifiers();
1116         clean_chip_info();
1117 out:
1118         pr_info("Platform driver disabled. System does not support PState control\n");
1119         return rc;
1120 }
1121 module_init(powernv_cpufreq_init);
1122
1123 static void __exit powernv_cpufreq_exit(void)
1124 {
1125         cpufreq_unregister_driver(&powernv_cpufreq_driver);
1126         unregister_all_notifiers();
1127         clean_chip_info();
1128 }
1129 module_exit(powernv_cpufreq_exit);
1130
1131 MODULE_LICENSE("GPL");
1132 MODULE_AUTHOR("Vaidyanathan Srinivasan <svaidy at linux.vnet.ibm.com>");