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 <linux/hashtable.h>
  33 #include <trace/events/power.h>
  34
  35 #include <asm/cputhreads.h>
  36 #include <asm/firmware.h>
  37 #include <asm/reg.h>
  38 #include <asm/smp.h> /* Required for cpu_sibling_mask() in UP configs */
  39 #include <asm/opal.h>
  40 #include <linux/timer.h>
  41
  42 #define POWERNV_MAX_PSTATES_ORDER  8
  43 #define POWERNV_MAX_PSTATES     (1UL << (POWERNV_MAX_PSTATES_ORDER))
  44 #define PMSR_PSAFE_ENABLE       (1UL << 30)
  45 #define PMSR_SPR_EM_DISABLE     (1UL << 31)
  46 #define MAX_PSTATE_SHIFT        32
  47 #define LPSTATE_SHIFT           48
  48 #define GPSTATE_SHIFT           56
  49
  50 #define MAX_RAMP_DOWN_TIME                              5120
  51 /*
  52  * On an idle system we want the global pstate to ramp-down from max value to
  53  * min over a span of ~5 secs. Also we want it to initially ramp-down slowly and
  54  * then ramp-down rapidly later on.
  55  *
  56  * This gives a percentage rampdown for time elapsed in milliseconds.
  57  * ramp_down_percentage = ((ms * ms) >> 18)
  58  *                      ~= 3.8 * (sec * sec)
  59  *
  60  * At 0 ms      ramp_down_percent = 0
  61  * At 5120 ms   ramp_down_percent = 100
  62  */
  63 #define ramp_down_percent(time)         ((time * time) >> 18)
  64
  65 /* Interval after which the timer is queued to bring down global pstate */
  66 #define GPSTATE_TIMER_INTERVAL                          2000
  67
  68 /**
  69  * struct global_pstate_info -  Per policy data structure to maintain history of
  70  *                              global pstates
  71  * @highest_lpstate_idx:        The local pstate index from which we are
  72  *                              ramping down
  73  * @elapsed_time:               Time in ms spent in ramping down from
  74  *                              highest_lpstate_idx
  75  * @last_sampled_time:          Time from boot in ms when global pstates were
  76  *                              last set
  77  * @last_lpstate_idx,           Last set value of local pstate and global
  78  * last_gpstate_idx             pstate in terms of cpufreq table index
  79  * @timer:                      Is used for ramping down if cpu goes idle for
  80  *                              a long time with global pstate held high
  81  * @gpstate_lock:               A spinlock to maintain synchronization between
  82  *                              routines called by the timer handler and
  83  *                              governer's target_index calls
  84  */
  85 struct global_pstate_info {
  86         int highest_lpstate_idx;
  87         unsigned int elapsed_time;
  88         unsigned int last_sampled_time;
  89         int last_lpstate_idx;
  90         int last_gpstate_idx;
  91         spinlock_t gpstate_lock;
  92         struct timer_list timer;
  93         struct cpufreq_policy *policy;
  94 };
  95
  96 static struct cpufreq_frequency_table powernv_freqs[POWERNV_MAX_PSTATES+1];
  97
  98 DEFINE_HASHTABLE(pstate_revmap, POWERNV_MAX_PSTATES_ORDER);
  99 /**
 100  * struct pstate_idx_revmap_data: Entry in the hashmap pstate_revmap
 101  *                                indexed by a function of pstate id.
 102  *
 103  * @pstate_id: pstate id for this entry.
 104  *
 105  * @cpufreq_table_idx: Index into the powernv_freqs
 106  *                     cpufreq_frequency_table for frequency
 107  *                     corresponding to pstate_id.
 108  *
 109  * @hentry: hlist_node that hooks this entry into the pstate_revmap
 110  *          hashtable
 111  */
 112 struct pstate_idx_revmap_data {
 113         int pstate_id;
 114         unsigned int cpufreq_table_idx;
 115         struct hlist_node hentry;
 116 };
 117
 118 u32 pstate_sign_prefix;
 119 static bool rebooting, throttled, occ_reset;
 120
 121 static const char * const throttle_reason[] = {
 122         "No throttling",
 123         "Power Cap",
 124         "Processor Over Temperature",
 125         "Power Supply Failure",
 126         "Over Current",
 127         "OCC Reset"
 128 };
 129
 130 enum throttle_reason_type {
 131         NO_THROTTLE = 0,
 132         POWERCAP,
 133         CPU_OVERTEMP,
 134         POWER_SUPPLY_FAILURE,
 135         OVERCURRENT,
 136         OCC_RESET_THROTTLE,
 137         OCC_MAX_REASON
 138 };
 139
 140 static struct chip {
 141         unsigned int id;
 142         bool throttled;
 143         bool restore;
 144         u8 throttle_reason;
 145         cpumask_t mask;
 146         struct work_struct throttle;
 147         int throttle_turbo;
 148         int throttle_sub_turbo;
 149         int reason[OCC_MAX_REASON];
 150 } *chips;
 151
 152 static int nr_chips;
 153 static DEFINE_PER_CPU(struct chip *, chip_info);
 154
 155 /*
 156  * Note:
 157  * The set of pstates consists of contiguous integers.
 158  * powernv_pstate_info stores the index of the frequency table for
 159  * max, min and nominal frequencies. It also stores number of
 160  * available frequencies.
 161  *
 162  * powernv_pstate_info.nominal indicates the index to the highest
 163  * non-turbo frequency.
 164  */
 165 static struct powernv_pstate_info {
 166         unsigned int min;
 167         unsigned int max;
 168         unsigned int nominal;
 169         unsigned int nr_pstates;
 170         bool wof_enabled;
 171 } powernv_pstate_info;
 172
 173 static inline int extract_pstate(u64 pmsr_val, unsigned int shift)
 174 {
 175         int ret = ((pmsr_val >> shift) & 0xFF);
 176
 177         if (!ret)
 178                 return ret;
 179
 180         return (pstate_sign_prefix | ret);
 181 }
 182
 183 #define extract_local_pstate(x) extract_pstate(x, LPSTATE_SHIFT)
 184 #define extract_global_pstate(x) extract_pstate(x, GPSTATE_SHIFT)
 185 #define extract_max_pstate(x)  extract_pstate(x, MAX_PSTATE_SHIFT)
 186
 187 /* Use following functions for conversions between pstate_id and index */
 188
 189 /**
 190  * idx_to_pstate : Returns the pstate id corresponding to the
 191  *                 frequency in the cpufreq frequency table
 192  *                 powernv_freqs indexed by @i.
 193  *
 194  *                 If @i is out of bound, this will return the pstate
 195  *                 corresponding to the nominal frequency.
 196  */
 197 static inline int idx_to_pstate(unsigned int i)
 198 {
 199         if (unlikely(i >= powernv_pstate_info.nr_pstates)) {
 200                 pr_warn_once("idx_to_pstate: index %u is out of bound\n", i);
 201                 return powernv_freqs[powernv_pstate_info.nominal].driver_data;
 202         }
 203
 204         return powernv_freqs[i].driver_data;
 205 }
 206
 207 /**
 208  * pstate_to_idx : Returns the index in the cpufreq frequencytable
 209  *                 powernv_freqs for the frequency whose corresponding
 210  *                 pstate id is @pstate.
 211  *
 212  *                 If no frequency corresponding to @pstate is found,
 213  *                 this will return the index of the nominal
 214  *                 frequency.
 215  */
 216 static unsigned int pstate_to_idx(int pstate)
 217 {
 218         unsigned int key = pstate % POWERNV_MAX_PSTATES;
 219         struct pstate_idx_revmap_data *revmap_data;
 220
 221         hash_for_each_possible(pstate_revmap, revmap_data, hentry, key) {
 222                 if (revmap_data->pstate_id == pstate)
 223                         return revmap_data->cpufreq_table_idx;
 224         }
 225
 226         pr_warn_once("pstate_to_idx: pstate %d not found\n", pstate);
 227         return powernv_pstate_info.nominal;
 228 }
 229
 230 static inline void reset_gpstates(struct cpufreq_policy *policy)
 231 {
 232         struct global_pstate_info *gpstates = policy->driver_data;
 233
 234         gpstates->highest_lpstate_idx = 0;
 235         gpstates->elapsed_time = 0;
 236         gpstates->last_sampled_time = 0;
 237         gpstates->last_lpstate_idx = 0;
 238         gpstates->last_gpstate_idx = 0;
 239 }
 240
 241 /*
 242  * Initialize the freq table based on data obtained
 243  * from the firmware passed via device-tree
 244  */
 245 static int init_powernv_pstates(void)
 246 {
 247         struct device_node *power_mgt;
 248         int i, nr_pstates = 0;
 249         const __be32 *pstate_ids, *pstate_freqs;
 250         u32 len_ids, len_freqs;
 251         u32 pstate_min, pstate_max, pstate_nominal;
 252         u32 pstate_turbo, pstate_ultra_turbo;
 253
 254         power_mgt = of_find_node_by_path("/ibm,opal/power-mgt");
 255         if (!power_mgt) {
 256                 pr_warn("power-mgt node not found\n");
 257                 return -ENODEV;
 258         }
 259
 260         if (of_property_read_u32(power_mgt, "ibm,pstate-min", &pstate_min)) {
 261                 pr_warn("ibm,pstate-min node not found\n");
 262                 return -ENODEV;
 263         }
 264
 265         if (of_property_read_u32(power_mgt, "ibm,pstate-max", &pstate_max)) {
 266                 pr_warn("ibm,pstate-max node not found\n");
 267                 return -ENODEV;
 268         }
 269
 270         if (of_property_read_u32(power_mgt, "ibm,pstate-nominal",
 271                                  &pstate_nominal)) {
 272                 pr_warn("ibm,pstate-nominal not found\n");
 273                 return -ENODEV;
 274         }
 275
 276         if (of_property_read_u32(power_mgt, "ibm,pstate-ultra-turbo",
 277                                  &pstate_ultra_turbo)) {
 278                 powernv_pstate_info.wof_enabled = false;
 279                 goto next;
 280         }
 281
 282         if (of_property_read_u32(power_mgt, "ibm,pstate-turbo",
 283                                  &pstate_turbo)) {
 284                 powernv_pstate_info.wof_enabled = false;
 285                 goto next;
 286         }
 287
 288         if (pstate_turbo == pstate_ultra_turbo)
 289                 powernv_pstate_info.wof_enabled = false;
 290         else
 291                 powernv_pstate_info.wof_enabled = true;
 292
 293 next:
 294         pr_info("cpufreq pstate min %d nominal %d max %d\n", pstate_min,
 295                 pstate_nominal, pstate_max);
 296         pr_info("Workload Optimized Frequency is %s in the platform\n",
 297                 (powernv_pstate_info.wof_enabled) ? "enabled" : "disabled");
 298
 299         pstate_ids = of_get_property(power_mgt, "ibm,pstate-ids", &len_ids);
 300         if (!pstate_ids) {
 301                 pr_warn("ibm,pstate-ids not found\n");
 302                 return -ENODEV;
 303         }
 304
 305         pstate_freqs = of_get_property(power_mgt, "ibm,pstate-frequencies-mhz",
 306                                       &len_freqs);
 307         if (!pstate_freqs) {
 308                 pr_warn("ibm,pstate-frequencies-mhz not found\n");
 309                 return -ENODEV;
 310         }
 311
 312         if (len_ids != len_freqs) {
 313                 pr_warn("Entries in ibm,pstate-ids and "
 314                         "ibm,pstate-frequencies-mhz does not match\n");
 315         }
 316
 317         nr_pstates = min(len_ids, len_freqs) / sizeof(u32);
 318         if (!nr_pstates) {
 319                 pr_warn("No PStates found\n");
 320                 return -ENODEV;
 321         }
 322
 323         powernv_pstate_info.nr_pstates = nr_pstates;
 324         pr_debug("NR PStates %d\n", nr_pstates);
 325
 326         pstate_sign_prefix = pstate_min & ~0xFF;
 327
 328         for (i = 0; i < nr_pstates; i++) {
 329                 u32 id = be32_to_cpu(pstate_ids[i]);
 330                 u32 freq = be32_to_cpu(pstate_freqs[i]);
 331                 struct pstate_idx_revmap_data *revmap_data;
 332                 unsigned int key;
 333
 334                 pr_debug("PState id %d freq %d MHz\n", id, freq);
 335                 powernv_freqs[i].frequency = freq * 1000; /* kHz */
 336                 powernv_freqs[i].driver_data = id;
 337
 338                 revmap_data = (struct pstate_idx_revmap_data *)
 339                               kmalloc(sizeof(*revmap_data), GFP_KERNEL);
 340
 341                 revmap_data->pstate_id = id;
 342                 revmap_data->cpufreq_table_idx = i;
 343                 key = id % POWERNV_MAX_PSTATES;
 344                 hash_add(pstate_revmap, &revmap_data->hentry, key);
 345
 346                 if (id == pstate_max)
 347                         powernv_pstate_info.max = i;
 348                 else if (id == pstate_nominal)
 349                         powernv_pstate_info.nominal = i;
 350                 else if (id == pstate_min)
 351                         powernv_pstate_info.min = i;
 352
 353                 if (powernv_pstate_info.wof_enabled && id == pstate_turbo) {
 354                         int j;
 355
 356                         for (j = i - 1; j >= (int)powernv_pstate_info.max; j--)
 357                                 powernv_freqs[j].flags = CPUFREQ_BOOST_FREQ;
 358                 }
 359         }
 360
 361         /* End of list marker entry */
 362         powernv_freqs[i].frequency = CPUFREQ_TABLE_END;
 363         return 0;
 364 }
 365
 366 /* Returns the CPU frequency corresponding to the pstate_id. */
 367 static unsigned int pstate_id_to_freq(int pstate_id)
 368 {
 369         int i;
 370
 371         i = pstate_to_idx(pstate_id);
 372         if (i >= powernv_pstate_info.nr_pstates || i < 0) {
 373                 pr_warn("PState id %d outside of PState table, "
 374                         "reporting nominal id %d instead\n",
 375                         pstate_id, idx_to_pstate(powernv_pstate_info.nominal));
 376                 i = powernv_pstate_info.nominal;
 377         }
 378
 379         return powernv_freqs[i].frequency;
 380 }
 381
 382 /*
 383  * cpuinfo_nominal_freq_show - Show the nominal CPU frequency as indicated by
 384  * the firmware
 385  */
 386 static ssize_t cpuinfo_nominal_freq_show(struct cpufreq_policy *policy,
 387                                         char *buf)
 388 {
 389         return sprintf(buf, "%u\n",
 390                 powernv_freqs[powernv_pstate_info.nominal].frequency);
 391 }
 392
 393 struct freq_attr cpufreq_freq_attr_cpuinfo_nominal_freq =
 394         __ATTR_RO(cpuinfo_nominal_freq);
 395
 396 #define SCALING_BOOST_FREQS_ATTR_INDEX          2
 397
 398 static struct freq_attr *powernv_cpu_freq_attr[] = {
 399         &cpufreq_freq_attr_scaling_available_freqs,
 400         &cpufreq_freq_attr_cpuinfo_nominal_freq,
 401         &cpufreq_freq_attr_scaling_boost_freqs,
 402         NULL,
 403 };
 404
 405 #define throttle_attr(name, member)                                     \
 406 static ssize_t name##_show(struct cpufreq_policy *policy, char *buf)    \
 407 {                                                                       \
 408         struct chip *chip = per_cpu(chip_info, policy->cpu);            \
 409                                                                         \
 410         return sprintf(buf, "%u\n", chip->member);                      \
 411 }                                                                       \
 412                                                                         \
 413 static struct freq_attr throttle_attr_##name = __ATTR_RO(name)          \
 414
 415 throttle_attr(unthrottle, reason[NO_THROTTLE]);
 416 throttle_attr(powercap, reason[POWERCAP]);
 417 throttle_attr(overtemp, reason[CPU_OVERTEMP]);
 418 throttle_attr(supply_fault, reason[POWER_SUPPLY_FAILURE]);
 419 throttle_attr(overcurrent, reason[OVERCURRENT]);
 420 throttle_attr(occ_reset, reason[OCC_RESET_THROTTLE]);
 421 throttle_attr(turbo_stat, throttle_turbo);
 422 throttle_attr(sub_turbo_stat, throttle_sub_turbo);
 423
 424 static struct attribute *throttle_attrs[] = {
 425         &throttle_attr_unthrottle.attr,
 426         &throttle_attr_powercap.attr,
 427         &throttle_attr_overtemp.attr,
 428         &throttle_attr_supply_fault.attr,
 429         &throttle_attr_overcurrent.attr,
 430         &throttle_attr_occ_reset.attr,
 431         &throttle_attr_turbo_stat.attr,
 432         &throttle_attr_sub_turbo_stat.attr,
 433         NULL,
 434 };
 435
 436 static const struct attribute_group throttle_attr_grp = {
 437         .name   = "throttle_stats",
 438         .attrs  = throttle_attrs,
 439 };
 440
 441 /* Helper routines */
 442
 443 /* Access helpers to power mgt SPR */
 444
 445 static inline unsigned long get_pmspr(unsigned long sprn)
 446 {
 447         switch (sprn) {
 448         case SPRN_PMCR:
 449                 return mfspr(SPRN_PMCR);
 450
 451         case SPRN_PMICR:
 452                 return mfspr(SPRN_PMICR);
 453
 454         case SPRN_PMSR:
 455                 return mfspr(SPRN_PMSR);
 456         }
 457         BUG();
 458 }
 459
 460 static inline void set_pmspr(unsigned long sprn, unsigned long val)
 461 {
 462         switch (sprn) {
 463         case SPRN_PMCR:
 464                 mtspr(SPRN_PMCR, val);
 465                 return;
 466
 467         case SPRN_PMICR:
 468                 mtspr(SPRN_PMICR, val);
 469                 return;
 470         }
 471         BUG();
 472 }
 473
 474 /*
 475  * Use objects of this type to query/update
 476  * pstates on a remote CPU via smp_call_function.
 477  */
 478 struct powernv_smp_call_data {
 479         unsigned int freq;
 480         int pstate_id;
 481         int gpstate_id;
 482 };
 483
 484 /*
 485  * powernv_read_cpu_freq: Reads the current frequency on this CPU.
 486  *
 487  * Called via smp_call_function.
 488  *
 489  * Note: The caller of the smp_call_function should pass an argument of
 490  * the type 'struct powernv_smp_call_data *' along with this function.
 491  *
 492  * The current frequency on this CPU will be returned via
 493  * ((struct powernv_smp_call_data *)arg)->freq;
 494  */
 495 static void powernv_read_cpu_freq(void *arg)
 496 {
 497         unsigned long pmspr_val;
 498         struct powernv_smp_call_data *freq_data = arg;
 499
 500         pmspr_val = get_pmspr(SPRN_PMSR);
 501         freq_data->pstate_id = extract_local_pstate(pmspr_val);
 502         freq_data->freq = pstate_id_to_freq(freq_data->pstate_id);
 503
 504         pr_debug("cpu %d pmsr %016lX pstate_id %d frequency %d kHz\n",
 505                 raw_smp_processor_id(), pmspr_val, freq_data->pstate_id,
 506                 freq_data->freq);
 507 }
 508
 509 /*
 510  * powernv_cpufreq_get: Returns the CPU frequency as reported by the
 511  * firmware for CPU 'cpu'. This value is reported through the sysfs
 512  * file cpuinfo_cur_freq.
 513  */
 514 static unsigned int powernv_cpufreq_get(unsigned int cpu)
 515 {
 516         struct powernv_smp_call_data freq_data;
 517
 518         smp_call_function_any(cpu_sibling_mask(cpu), powernv_read_cpu_freq,
 519                         &freq_data, 1);
 520
 521         return freq_data.freq;
 522 }
 523
 524 /*
 525  * set_pstate: Sets the pstate on this CPU.
 526  *
 527  * This is called via an smp_call_function.
 528  *
 529  * The caller must ensure that freq_data is of the type
 530  * (struct powernv_smp_call_data *) and the pstate_id which needs to be set
 531  * on this CPU should be present in freq_data->pstate_id.
 532  */
 533 static void set_pstate(void *data)
 534 {
 535         unsigned long val;
 536         struct powernv_smp_call_data *freq_data = data;
 537         unsigned long pstate_ul = freq_data->pstate_id;
 538         unsigned long gpstate_ul = freq_data->gpstate_id;
 539
 540         val = get_pmspr(SPRN_PMCR);
 541         val = val & 0x0000FFFFFFFFFFFFULL;
 542
 543         pstate_ul = pstate_ul & 0xFF;
 544         gpstate_ul = gpstate_ul & 0xFF;
 545
 546         /* Set both global(bits 56..63) and local(bits 48..55) PStates */
 547         val = val | (gpstate_ul << 56) | (pstate_ul << 48);
 548
 549         pr_debug("Setting cpu %d pmcr to %016lX\n",
 550                         raw_smp_processor_id(), val);
 551         set_pmspr(SPRN_PMCR, val);
 552 }
 553
 554 /*
 555  * get_nominal_index: Returns the index corresponding to the nominal
 556  * pstate in the cpufreq table
 557  */
 558 static inline unsigned int get_nominal_index(void)
 559 {
 560         return powernv_pstate_info.nominal;
 561 }
 562
 563 static void powernv_cpufreq_throttle_check(void *data)
 564 {
 565         struct chip *chip;
 566         unsigned int cpu = smp_processor_id();
 567         unsigned long pmsr;
 568         int pmsr_pmax;
 569         unsigned int pmsr_pmax_idx;
 570
 571         pmsr = get_pmspr(SPRN_PMSR);
 572         chip = this_cpu_read(chip_info);
 573
 574         /* Check for Pmax Capping */
 575         pmsr_pmax = extract_max_pstate(pmsr);
 576         pmsr_pmax_idx = pstate_to_idx(pmsr_pmax);
 577         if (pmsr_pmax_idx != powernv_pstate_info.max) {
 578                 if (chip->throttled)
 579                         goto next;
 580                 chip->throttled = true;
 581                 if (pmsr_pmax_idx > powernv_pstate_info.nominal) {
 582                         pr_warn_once("CPU %d on Chip %u has Pmax(%d) reduced below nominal frequency(%d)\n",
 583                                      cpu, chip->id, pmsr_pmax,
 584                                      idx_to_pstate(powernv_pstate_info.nominal));
 585                         chip->throttle_sub_turbo++;
 586                 } else {
 587                         chip->throttle_turbo++;
 588                 }
 589                 trace_powernv_throttle(chip->id,
 590                                       throttle_reason[chip->throttle_reason],
 591                                       pmsr_pmax);
 592         } else if (chip->throttled) {
 593                 chip->throttled = false;
 594                 trace_powernv_throttle(chip->id,
 595                                       throttle_reason[chip->throttle_reason],
 596                                       pmsr_pmax);
 597         }
 598
 599         /* Check if Psafe_mode_active is set in PMSR. */
 600 next:
 601         if (pmsr & PMSR_PSAFE_ENABLE) {
 602                 throttled = true;
 603                 pr_info("Pstate set to safe frequency\n");
 604         }
 605
 606         /* Check if SPR_EM_DISABLE is set in PMSR */
 607         if (pmsr & PMSR_SPR_EM_DISABLE) {
 608                 throttled = true;
 609                 pr_info("Frequency Control disabled from OS\n");
 610         }
 611
 612         if (throttled) {
 613                 pr_info("PMSR = %16lx\n", pmsr);
 614                 pr_warn("CPU Frequency could be throttled\n");
 615         }
 616 }
 617
 618 /**
 619  * calc_global_pstate - Calculate global pstate
 620  * @elapsed_time:               Elapsed time in milliseconds
 621  * @local_pstate_idx:           New local pstate
 622  * @highest_lpstate_idx:        pstate from which its ramping down
 623  *
 624  * Finds the appropriate global pstate based on the pstate from which its
 625  * ramping down and the time elapsed in ramping down. It follows a quadratic
 626  * equation which ensures that it reaches ramping down to pmin in 5sec.
 627  */
 628 static inline int calc_global_pstate(unsigned int elapsed_time,
 629                                      int highest_lpstate_idx,
 630                                      int local_pstate_idx)
 631 {
 632         int index_diff;
 633
 634         /*
 635          * Using ramp_down_percent we get the percentage of rampdown
 636          * that we are expecting to be dropping. Difference between
 637          * highest_lpstate_idx and powernv_pstate_info.min will give a absolute
 638          * number of how many pstates we will drop eventually by the end of
 639          * 5 seconds, then just scale it get the number pstates to be dropped.
 640          */
 641         index_diff =  ((int)ramp_down_percent(elapsed_time) *
 642                         (powernv_pstate_info.min - highest_lpstate_idx)) / 100;
 643
 644         /* Ensure that global pstate is >= to local pstate */
 645         if (highest_lpstate_idx + index_diff >= local_pstate_idx)
 646                 return local_pstate_idx;
 647         else
 648                 return highest_lpstate_idx + index_diff;
 649 }
 650
 651 static inline void  queue_gpstate_timer(struct global_pstate_info *gpstates)
 652 {
 653         unsigned int timer_interval;
 654
 655         /*
 656          * Setting up timer to fire after GPSTATE_TIMER_INTERVAL ms, But
 657          * if it exceeds MAX_RAMP_DOWN_TIME ms for ramp down time.
 658          * Set timer such that it fires exactly at MAX_RAMP_DOWN_TIME
 659          * seconds of ramp down time.
 660          */
 661         if ((gpstates->elapsed_time + GPSTATE_TIMER_INTERVAL)
 662              > MAX_RAMP_DOWN_TIME)
 663                 timer_interval = MAX_RAMP_DOWN_TIME - gpstates->elapsed_time;
 664         else
 665                 timer_interval = GPSTATE_TIMER_INTERVAL;
 666
 667         mod_timer(&gpstates->timer, jiffies + msecs_to_jiffies(timer_interval));
 668 }
 669
 670 /**
 671  * gpstate_timer_handler
 672  *
 673  * @data: pointer to cpufreq_policy on which timer was queued
 674  *
 675  * This handler brings down the global pstate closer to the local pstate
 676  * according quadratic equation. Queues a new timer if it is still not equal
 677  * to local pstate
 678  */
 679 void gpstate_timer_handler(struct timer_list *t)
 680 {
 681         struct global_pstate_info *gpstates = from_timer(gpstates, t, timer);
 682         struct cpufreq_policy *policy = gpstates->policy;
 683         int gpstate_idx, lpstate_idx;
 684         unsigned long val;
 685         unsigned int time_diff = jiffies_to_msecs(jiffies)
 686                                         - gpstates->last_sampled_time;
 687         struct powernv_smp_call_data freq_data;
 688
 689         if (!spin_trylock(&gpstates->gpstate_lock))
 690                 return;
 691
 692         /*
 693          * If PMCR was last updated was using fast_swtich then
 694          * We may have wrong in gpstate->last_lpstate_idx
 695          * value. Hence, read from PMCR to get correct data.
 696          */
 697         val = get_pmspr(SPRN_PMCR);
 698         freq_data.gpstate_id = extract_global_pstate(val);
 699         freq_data.pstate_id = extract_local_pstate(val);
 700         if (freq_data.gpstate_id  == freq_data.pstate_id) {
 701                 reset_gpstates(policy);
 702                 spin_unlock(&gpstates->gpstate_lock);
 703                 return;
 704         }
 705
 706         gpstates->last_sampled_time += time_diff;
 707         gpstates->elapsed_time += time_diff;
 708
 709         if (gpstates->elapsed_time > MAX_RAMP_DOWN_TIME) {
 710                 gpstate_idx = pstate_to_idx(freq_data.pstate_id);
 711                 lpstate_idx = gpstate_idx;
 712                 reset_gpstates(policy);
 713                 gpstates->highest_lpstate_idx = gpstate_idx;
 714         } else {
 715                 lpstate_idx = pstate_to_idx(freq_data.pstate_id);
 716                 gpstate_idx = calc_global_pstate(gpstates->elapsed_time,
 717                                                  gpstates->highest_lpstate_idx,
 718                                                  lpstate_idx);
 719         }
 720         freq_data.gpstate_id = idx_to_pstate(gpstate_idx);
 721         gpstates->last_gpstate_idx = gpstate_idx;
 722         gpstates->last_lpstate_idx = lpstate_idx;
 723         /*
 724          * If local pstate is equal to global pstate, rampdown is over
 725          * So timer is not required to be queued.
 726          */
 727         if (gpstate_idx != gpstates->last_lpstate_idx)
 728                 queue_gpstate_timer(gpstates);
 729
 730         spin_unlock(&gpstates->gpstate_lock);
 731
 732         /* Timer may get migrated to a different cpu on cpu hot unplug */
 733         smp_call_function_any(policy->cpus, set_pstate, &freq_data, 1);
 734 }
 735
 736 /*
 737  * powernv_cpufreq_target_index: Sets the frequency corresponding to
 738  * the cpufreq table entry indexed by new_index on the cpus in the
 739  * mask policy->cpus
 740  */
 741 static int powernv_cpufreq_target_index(struct cpufreq_policy *policy,
 742                                         unsigned int new_index)
 743 {
 744         struct powernv_smp_call_data freq_data;
 745         unsigned int cur_msec, gpstate_idx;
 746         struct global_pstate_info *gpstates = policy->driver_data;
 747
 748         if (unlikely(rebooting) && new_index != get_nominal_index())
 749                 return 0;
 750
 751         if (!throttled) {
 752                 /* we don't want to be preempted while
 753                  * checking if the CPU frequency has been throttled
 754                  */
 755                 preempt_disable();
 756                 powernv_cpufreq_throttle_check(NULL);
 757                 preempt_enable();
 758         }
 759
 760         cur_msec = jiffies_to_msecs(get_jiffies_64());
 761
 762         spin_lock(&gpstates->gpstate_lock);
 763         freq_data.pstate_id = idx_to_pstate(new_index);
 764
 765         if (!gpstates->last_sampled_time) {
 766                 gpstate_idx = new_index;
 767                 gpstates->highest_lpstate_idx = new_index;
 768                 goto gpstates_done;
 769         }
 770
 771         if (gpstates->last_gpstate_idx < new_index) {
 772                 gpstates->elapsed_time += cur_msec -
 773                                                  gpstates->last_sampled_time;
 774
 775                 /*
 776                  * If its has been ramping down for more than MAX_RAMP_DOWN_TIME
 777                  * we should be resetting all global pstate related data. Set it
 778                  * equal to local pstate to start fresh.
 779                  */
 780                 if (gpstates->elapsed_time > MAX_RAMP_DOWN_TIME) {
 781                         reset_gpstates(policy);
 782                         gpstates->highest_lpstate_idx = new_index;
 783                         gpstate_idx = new_index;
 784                 } else {
 785                 /* Elaspsed_time is less than 5 seconds, continue to rampdown */
 786                         gpstate_idx = calc_global_pstate(gpstates->elapsed_time,
 787                                                          gpstates->highest_lpstate_idx,
 788                                                          new_index);
 789                 }
 790         } else {
 791                 reset_gpstates(policy);
 792                 gpstates->highest_lpstate_idx = new_index;
 793                 gpstate_idx = new_index;
 794         }
 795
 796         /*
 797          * If local pstate is equal to global pstate, rampdown is over
 798          * So timer is not required to be queued.
 799          */
 800         if (gpstate_idx != new_index)
 801                 queue_gpstate_timer(gpstates);
 802         else
 803                 del_timer_sync(&gpstates->timer);
 804
 805 gpstates_done:
 806         freq_data.gpstate_id = idx_to_pstate(gpstate_idx);
 807         gpstates->last_sampled_time = cur_msec;
 808         gpstates->last_gpstate_idx = gpstate_idx;
 809         gpstates->last_lpstate_idx = new_index;
 810
 811         spin_unlock(&gpstates->gpstate_lock);
 812
 813         /*
 814          * Use smp_call_function to send IPI and execute the
 815          * mtspr on target CPU.  We could do that without IPI
 816          * if current CPU is within policy->cpus (core)
 817          */
 818         smp_call_function_any(policy->cpus, set_pstate, &freq_data, 1);
 819         return 0;
 820 }
 821
 822 static int powernv_cpufreq_cpu_init(struct cpufreq_policy *policy)
 823 {
 824         int base, i, ret;
 825         struct kernfs_node *kn;
 826         struct global_pstate_info *gpstates;
 827
 828         base = cpu_first_thread_sibling(policy->cpu);
 829
 830         for (i = 0; i < threads_per_core; i++)
 831                 cpumask_set_cpu(base + i, policy->cpus);
 832
 833         kn = kernfs_find_and_get(policy->kobj.sd, throttle_attr_grp.name);
 834         if (!kn) {
 835                 int ret;
 836
 837                 ret = sysfs_create_group(&policy->kobj, &throttle_attr_grp);
 838                 if (ret) {
 839                         pr_info("Failed to create throttle stats directory for cpu %d\n",
 840                                 policy->cpu);
 841                         return ret;
 842                 }
 843         } else {
 844                 kernfs_put(kn);
 845         }
 846
 847         gpstates =  kzalloc(sizeof(*gpstates), GFP_KERNEL);
 848         if (!gpstates)
 849                 return -ENOMEM;
 850
 851         policy->driver_data = gpstates;
 852
 853         /* initialize timer */
 854         gpstates->policy = policy;
 855         timer_setup(&gpstates->timer, gpstate_timer_handler,
 856                     TIMER_PINNED | TIMER_DEFERRABLE);
 857         gpstates->timer.expires = jiffies +
 858                                 msecs_to_jiffies(GPSTATE_TIMER_INTERVAL);
 859         spin_lock_init(&gpstates->gpstate_lock);
 860         ret = cpufreq_table_validate_and_show(policy, powernv_freqs);
 861
 862         if (ret < 0) {
 863                 kfree(policy->driver_data);
 864                 return ret;
 865         }
 866
 867         policy->fast_switch_possible = true;
 868         return ret;
 869 }
 870
 871 static int powernv_cpufreq_cpu_exit(struct cpufreq_policy *policy)
 872 {
 873         /* timer is deleted in cpufreq_cpu_stop() */
 874         kfree(policy->driver_data);
 875
 876         return 0;
 877 }
 878
 879 static int powernv_cpufreq_reboot_notifier(struct notifier_block *nb,
 880                                 unsigned long action, void *unused)
 881 {
 882         int cpu;
 883         struct cpufreq_policy cpu_policy;
 884
 885         rebooting = true;
 886         for_each_online_cpu(cpu) {
 887                 cpufreq_get_policy(&cpu_policy, cpu);
 888                 powernv_cpufreq_target_index(&cpu_policy, get_nominal_index());
 889         }
 890
 891         return NOTIFY_DONE;
 892 }
 893
 894 static struct notifier_block powernv_cpufreq_reboot_nb = {
 895         .notifier_call = powernv_cpufreq_reboot_notifier,
 896 };
 897
 898 void powernv_cpufreq_work_fn(struct work_struct *work)
 899 {
 900         struct chip *chip = container_of(work, struct chip, throttle);
 901         unsigned int cpu;
 902         cpumask_t mask;
 903
 904         get_online_cpus();
 905         cpumask_and(&mask, &chip->mask, cpu_online_mask);
 906         smp_call_function_any(&mask,
 907                               powernv_cpufreq_throttle_check, NULL, 0);
 908
 909         if (!chip->restore)
 910                 goto out;
 911
 912         chip->restore = false;
 913         for_each_cpu(cpu, &mask) {
 914                 int index;
 915                 struct cpufreq_policy policy;
 916
 917                 cpufreq_get_policy(&policy, cpu);
 918                 index = cpufreq_table_find_index_c(&policy, policy.cur);
 919                 powernv_cpufreq_target_index(&policy, index);
 920                 cpumask_andnot(&mask, &mask, policy.cpus);
 921         }
 922 out:
 923         put_online_cpus();
 924 }
 925
 926 static int powernv_cpufreq_occ_msg(struct notifier_block *nb,
 927                                    unsigned long msg_type, void *_msg)
 928 {
 929         struct opal_msg *msg = _msg;
 930         struct opal_occ_msg omsg;
 931         int i;
 932
 933         if (msg_type != OPAL_MSG_OCC)
 934                 return 0;
 935
 936         omsg.type = be64_to_cpu(msg->params[0]);
 937
 938         switch (omsg.type) {
 939         case OCC_RESET:
 940                 occ_reset = true;
 941                 pr_info("OCC (On Chip Controller - enforces hard thermal/power limits) Resetting\n");
 942                 /*
 943                  * powernv_cpufreq_throttle_check() is called in
 944                  * target() callback which can detect the throttle state
 945                  * for governors like ondemand.
 946                  * But static governors will not call target() often thus
 947                  * report throttling here.
 948                  */
 949                 if (!throttled) {
 950                         throttled = true;
 951                         pr_warn("CPU frequency is throttled for duration\n");
 952                 }
 953
 954                 break;
 955         case OCC_LOAD:
 956                 pr_info("OCC Loading, CPU frequency is throttled until OCC is started\n");
 957                 break;
 958         case OCC_THROTTLE:
 959                 omsg.chip = be64_to_cpu(msg->params[1]);
 960                 omsg.throttle_status = be64_to_cpu(msg->params[2]);
 961
 962                 if (occ_reset) {
 963                         occ_reset = false;
 964                         throttled = false;
 965                         pr_info("OCC Active, CPU frequency is no longer throttled\n");
 966
 967                         for (i = 0; i < nr_chips; i++) {
 968                                 chips[i].restore = true;
 969                                 schedule_work(&chips[i].throttle);
 970                         }
 971
 972                         return 0;
 973                 }
 974
 975                 for (i = 0; i < nr_chips; i++)
 976                         if (chips[i].id == omsg.chip)
 977                                 break;
 978
 979                 if (omsg.throttle_status >= 0 &&
 980                     omsg.throttle_status <= OCC_MAX_THROTTLE_STATUS) {
 981                         chips[i].throttle_reason = omsg.throttle_status;
 982                         chips[i].reason[omsg.throttle_status]++;
 983                 }
 984
 985                 if (!omsg.throttle_status)
 986                         chips[i].restore = true;
 987
 988                 schedule_work(&chips[i].throttle);
 989         }
 990         return 0;
 991 }
 992
 993 static struct notifier_block powernv_cpufreq_opal_nb = {
 994         .notifier_call  = powernv_cpufreq_occ_msg,
 995         .next           = NULL,
 996         .priority       = 0,
 997 };
 998
 999 static void powernv_cpufreq_stop_cpu(struct cpufreq_policy *policy)
1000 {
1001         struct powernv_smp_call_data freq_data;
1002         struct global_pstate_info *gpstates = policy->driver_data;
1003
1004         freq_data.pstate_id = idx_to_pstate(powernv_pstate_info.min);
1005         freq_data.gpstate_id = idx_to_pstate(powernv_pstate_info.min);
1006         smp_call_function_single(policy->cpu, set_pstate, &freq_data, 1);
1007         del_timer_sync(&gpstates->timer);
1008 }
1009
1010 static unsigned int powernv_fast_switch(struct cpufreq_policy *policy,
1011                                         unsigned int target_freq)
1012 {
1013         int index;
1014         struct powernv_smp_call_data freq_data;
1015
1016         index = cpufreq_table_find_index_dl(policy, target_freq);
1017         freq_data.pstate_id = powernv_freqs[index].driver_data;
1018         freq_data.gpstate_id = powernv_freqs[index].driver_data;
1019         set_pstate(&freq_data);
1020
1021         return powernv_freqs[index].frequency;
1022 }
1023
1024 static struct cpufreq_driver powernv_cpufreq_driver = {
1025         .name           = "powernv-cpufreq",
1026         .flags          = CPUFREQ_CONST_LOOPS,
1027         .init           = powernv_cpufreq_cpu_init,
1028         .exit           = powernv_cpufreq_cpu_exit,
1029         .verify         = cpufreq_generic_frequency_table_verify,
1030         .target_index   = powernv_cpufreq_target_index,
1031         .fast_switch    = powernv_fast_switch,
1032         .get            = powernv_cpufreq_get,
1033         .stop_cpu       = powernv_cpufreq_stop_cpu,
1034         .attr           = powernv_cpu_freq_attr,
1035 };
1036
1037 static int init_chip_info(void)
1038 {
1039         unsigned int chip[256];
1040         unsigned int cpu, i;
1041         unsigned int prev_chip_id = UINT_MAX;
1042
1043         for_each_possible_cpu(cpu) {
1044                 unsigned int id = cpu_to_chip_id(cpu);
1045
1046                 if (prev_chip_id != id) {
1047                         prev_chip_id = id;
1048                         chip[nr_chips++] = id;
1049                 }
1050         }
1051
1052         chips = kcalloc(nr_chips, sizeof(struct chip), GFP_KERNEL);
1053         if (!chips)
1054                 return -ENOMEM;
1055
1056         for (i = 0; i < nr_chips; i++) {
1057                 chips[i].id = chip[i];
1058                 cpumask_copy(&chips[i].mask, cpumask_of_node(chip[i]));
1059                 INIT_WORK(&chips[i].throttle, powernv_cpufreq_work_fn);
1060                 for_each_cpu(cpu, &chips[i].mask)
1061                         per_cpu(chip_info, cpu) =  &chips[i];
1062         }
1063
1064         return 0;
1065 }
1066
1067 static inline void clean_chip_info(void)
1068 {
1069         kfree(chips);
1070 }
1071
1072 static inline void unregister_all_notifiers(void)
1073 {
1074         opal_message_notifier_unregister(OPAL_MSG_OCC,
1075                                          &powernv_cpufreq_opal_nb);
1076         unregister_reboot_notifier(&powernv_cpufreq_reboot_nb);
1077 }
1078
1079 static int __init powernv_cpufreq_init(void)
1080 {
1081         int rc = 0;
1082
1083         /* Don't probe on pseries (guest) platforms */
1084         if (!firmware_has_feature(FW_FEATURE_OPAL))
1085                 return -ENODEV;
1086
1087         /* Discover pstates from device tree and init */
1088         rc = init_powernv_pstates();
1089         if (rc)
1090                 goto out;
1091
1092         /* Populate chip info */
1093         rc = init_chip_info();
1094         if (rc)
1095                 goto out;
1096
1097         register_reboot_notifier(&powernv_cpufreq_reboot_nb);
1098         opal_message_notifier_register(OPAL_MSG_OCC, &powernv_cpufreq_opal_nb);
1099
1100         if (powernv_pstate_info.wof_enabled)
1101                 powernv_cpufreq_driver.boost_enabled = true;
1102         else
1103                 powernv_cpu_freq_attr[SCALING_BOOST_FREQS_ATTR_INDEX] = NULL;
1104
1105         rc = cpufreq_register_driver(&powernv_cpufreq_driver);
1106         if (rc) {
1107                 pr_info("Failed to register the cpufreq driver (%d)\n", rc);
1108                 goto cleanup_notifiers;
1109         }
1110
1111         if (powernv_pstate_info.wof_enabled)
1112                 cpufreq_enable_boost_support();
1113
1114         return 0;
1115 cleanup_notifiers:
1116         unregister_all_notifiers();
1117         clean_chip_info();
1118 out:
1119         pr_info("Platform driver disabled. System does not support PState control\n");
1120         return rc;
1121 }
1122 module_init(powernv_cpufreq_init);
1123
1124 static void __exit powernv_cpufreq_exit(void)
1125 {
1126         cpufreq_unregister_driver(&powernv_cpufreq_driver);
1127         unregister_all_notifiers();
1128         clean_chip_info();
1129 }
1130 module_exit(powernv_cpufreq_exit);
1131
1132 MODULE_LICENSE("GPL");
1133 MODULE_AUTHOR("Vaidyanathan Srinivasan <svaidy at linux.vnet.ibm.com>");