drivers/thermal/intel_powerclamp.c

   1 /*
   2  * intel_powerclamp.c - package c-state idle injection
   3  *
   4  * Copyright (c) 2012, Intel Corporation.
   5  *
   6  * Authors:
   7  *     Arjan van de Ven <arjan@linux.intel.com>
   8  *     Jacob Pan <jacob.jun.pan@linux.intel.com>
   9  *
  10  * This program is free software; you can redistribute it and/or modify it
  11  * under the terms and conditions of the GNU General Public License,
  12  * version 2, as published by the Free Software Foundation.
  13  *
  14  * This program is distributed in the hope it will be useful, but WITHOUT
  15  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  16  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
  17  * more details.
  18  *
  19  * You should have received a copy of the GNU General Public License along with
  20  * this program; if not, write to the Free Software Foundation, Inc.,
  21  * 51 Franklin St - Fifth Floor, Boston, MA 02110-1301 USA.
  22  *
  23  *
  24  *      TODO:
  25  *           1. better handle wakeup from external interrupts, currently a fixed
  26  *              compensation is added to clamping duration when excessive amount
  27  *              of wakeups are observed during idle time. the reason is that in
  28  *              case of external interrupts without need for ack, clamping down
  29  *              cpu in non-irq context does not reduce irq. for majority of the
  30  *              cases, clamping down cpu does help reduce irq as well, we should
  31  *              be able to differenciate the two cases and give a quantitative
  32  *              solution for the irqs that we can control. perhaps based on
  33  *              get_cpu_iowait_time_us()
  34  *
  35  *           2. synchronization with other hw blocks
  36  *
  37  *
  38  */
  39
  40 #define pr_fmt(fmt)     KBUILD_MODNAME ": " fmt
  41
  42 #include <linux/module.h>
  43 #include <linux/kernel.h>
  44 #include <linux/delay.h>
  45 #include <linux/kthread.h>
  46 #include <linux/freezer.h>
  47 #include <linux/cpu.h>
  48 #include <linux/thermal.h>
  49 #include <linux/slab.h>
  50 #include <linux/tick.h>
  51 #include <linux/debugfs.h>
  52 #include <linux/seq_file.h>
  53 #include <linux/sched/rt.h>
  54
  55 #include <asm/nmi.h>
  56 #include <asm/msr.h>
  57 #include <asm/mwait.h>
  58 #include <asm/cpu_device_id.h>
  59 #include <asm/idle.h>
  60 #include <asm/hardirq.h>
  61
  62 #define MAX_TARGET_RATIO (50U)
  63 /* For each undisturbed clamping period (no extra wake ups during idle time),
  64  * we increment the confidence counter for the given target ratio.
  65  * CONFIDENCE_OK defines the level where runtime calibration results are
  66  * valid.
  67  */
  68 #define CONFIDENCE_OK (3)
  69 /* Default idle injection duration, driver adjust sleep time to meet target
  70  * idle ratio. Similar to frequency modulation.
  71  */
  72 #define DEFAULT_DURATION_JIFFIES (6)
  73
  74 static unsigned int target_mwait;
  75 static struct dentry *debug_dir;
  76
  77 /* user selected target */
  78 static unsigned int set_target_ratio;
  79 static unsigned int current_ratio;
  80 static bool should_skip;
  81 static bool reduce_irq;
  82 static atomic_t idle_wakeup_counter;
  83 static unsigned int control_cpu; /* The cpu assigned to collect stat and update
  84                                   * control parameters. default to BSP but BSP
  85                                   * can be offlined.
  86                                   */
  87 static bool clamping;
  88
  89
  90 static struct task_struct * __percpu *powerclamp_thread;
  91 static struct thermal_cooling_device *cooling_dev;
  92 static unsigned long *cpu_clamping_mask;  /* bit map for tracking per cpu
  93                                            * clamping thread
  94                                            */
  95
  96 static unsigned int duration;
  97 static unsigned int pkg_cstate_ratio_cur;
  98 static unsigned int window_size;
  99
 100 static int duration_set(const char *arg, const struct kernel_param *kp)
 101 {
 102         int ret = 0;
 103         unsigned long new_duration;
 104
 105         ret = kstrtoul(arg, 10, &new_duration);
 106         if (ret)
 107                 goto exit;
 108         if (new_duration > 25 || new_duration < 6) {
 109                 pr_err("Out of recommended range %lu, between 6-25ms\n",
 110                         new_duration);
 111                 ret = -EINVAL;
 112         }
 113
 114         duration = clamp(new_duration, 6ul, 25ul);
 115         smp_mb();
 116
 117 exit:
 118
 119         return ret;
 120 }
 121
 122 static const struct kernel_param_ops duration_ops = {
 123         .set = duration_set,
 124         .get = param_get_int,
 125 };
 126
 127
 128 module_param_cb(duration, &duration_ops, &duration, 0644);
 129 MODULE_PARM_DESC(duration, "forced idle time for each attempt in msec.");
 130
 131 struct powerclamp_calibration_data {
 132         unsigned long confidence;  /* used for calibration, basically a counter
 133                                     * gets incremented each time a clamping
 134                                     * period is completed without extra wakeups
 135                                     * once that counter is reached given level,
 136                                     * compensation is deemed usable.
 137                                     */
 138         unsigned long steady_comp; /* steady state compensation used when
 139                                     * no extra wakeups occurred.
 140                                     */
 141         unsigned long dynamic_comp; /* compensate excessive wakeup from idle
 142                                      * mostly from external interrupts.
 143                                      */
 144 };
 145
 146 static struct powerclamp_calibration_data cal_data[MAX_TARGET_RATIO];
 147
 148 static int window_size_set(const char *arg, const struct kernel_param *kp)
 149 {
 150         int ret = 0;
 151         unsigned long new_window_size;
 152
 153         ret = kstrtoul(arg, 10, &new_window_size);
 154         if (ret)
 155                 goto exit_win;
 156         if (new_window_size > 10 || new_window_size < 2) {
 157                 pr_err("Out of recommended window size %lu, between 2-10\n",
 158                         new_window_size);
 159                 ret = -EINVAL;
 160         }
 161
 162         window_size = clamp(new_window_size, 2ul, 10ul);
 163         smp_mb();
 164
 165 exit_win:
 166
 167         return ret;
 168 }
 169
 170 static const struct kernel_param_ops window_size_ops = {
 171         .set = window_size_set,
 172         .get = param_get_int,
 173 };
 174
 175 module_param_cb(window_size, &window_size_ops, &window_size, 0644);
 176 MODULE_PARM_DESC(window_size, "sliding window in number of clamping cycles\n"
 177         "\tpowerclamp controls idle ratio within this window. larger\n"
 178         "\twindow size results in slower response time but more smooth\n"
 179         "\tclamping results. default to 2.");
 180
 181 static void find_target_mwait(void)
 182 {
 183         unsigned int eax, ebx, ecx, edx;
 184         unsigned int highest_cstate = 0;
 185         unsigned int highest_subcstate = 0;
 186         int i;
 187
 188         if (boot_cpu_data.cpuid_level < CPUID_MWAIT_LEAF)
 189                 return;
 190
 191         cpuid(CPUID_MWAIT_LEAF, &eax, &ebx, &ecx, &edx);
 192
 193         if (!(ecx & CPUID5_ECX_EXTENSIONS_SUPPORTED) ||
 194             !(ecx & CPUID5_ECX_INTERRUPT_BREAK))
 195                 return;
 196
 197         edx >>= MWAIT_SUBSTATE_SIZE;
 198         for (i = 0; i < 7 && edx; i++, edx >>= MWAIT_SUBSTATE_SIZE) {
 199                 if (edx & MWAIT_SUBSTATE_MASK) {
 200                         highest_cstate = i;
 201                         highest_subcstate = edx & MWAIT_SUBSTATE_MASK;
 202                 }
 203         }
 204         target_mwait = (highest_cstate << MWAIT_SUBSTATE_SIZE) |
 205                 (highest_subcstate - 1);
 206
 207 }
 208
 209 struct pkg_cstate_info {
 210         bool skip;
 211         int msr_index;
 212         int cstate_id;
 213 };
 214
 215 #define PKG_CSTATE_INIT(id) {                           \
 216                 .msr_index = MSR_PKG_C##id##_RESIDENCY, \
 217                 .cstate_id = id                         \
 218                         }
 219
 220 static struct pkg_cstate_info pkg_cstates[] = {
 221         PKG_CSTATE_INIT(2),
 222         PKG_CSTATE_INIT(3),
 223         PKG_CSTATE_INIT(6),
 224         PKG_CSTATE_INIT(7),
 225         PKG_CSTATE_INIT(8),
 226         PKG_CSTATE_INIT(9),
 227         PKG_CSTATE_INIT(10),
 228         {NULL},
 229 };
 230
 231 static bool has_pkg_state_counter(void)
 232 {
 233         u64 val;
 234         struct pkg_cstate_info *info = pkg_cstates;
 235
 236         /* check if any one of the counter msrs exists */
 237         while (info->msr_index) {
 238                 if (!rdmsrl_safe(info->msr_index, &val))
 239                         return true;
 240                 info++;
 241         }
 242
 243         return false;
 244 }
 245
 246 static u64 pkg_state_counter(void)
 247 {
 248         u64 val;
 249         u64 count = 0;
 250         struct pkg_cstate_info *info = pkg_cstates;
 251
 252         while (info->msr_index) {
 253                 if (!info->skip) {
 254                         if (!rdmsrl_safe(info->msr_index, &val))
 255                                 count += val;
 256                         else
 257                                 info->skip = true;
 258                 }
 259                 info++;
 260         }
 261
 262         return count;
 263 }
 264
 265 static void noop_timer(unsigned long foo)
 266 {
 267         /* empty... just the fact that we get the interrupt wakes us up */
 268 }
 269
 270 static unsigned int get_compensation(int ratio)
 271 {
 272         unsigned int comp = 0;
 273
 274         /* we only use compensation if all adjacent ones are good */
 275         if (ratio == 1 &&
 276                 cal_data[ratio].confidence >= CONFIDENCE_OK &&
 277                 cal_data[ratio + 1].confidence >= CONFIDENCE_OK &&
 278                 cal_data[ratio + 2].confidence >= CONFIDENCE_OK) {
 279                 comp = (cal_data[ratio].steady_comp +
 280                         cal_data[ratio + 1].steady_comp +
 281                         cal_data[ratio + 2].steady_comp) / 3;
 282         } else if (ratio == MAX_TARGET_RATIO - 1 &&
 283                 cal_data[ratio].confidence >= CONFIDENCE_OK &&
 284                 cal_data[ratio - 1].confidence >= CONFIDENCE_OK &&
 285                 cal_data[ratio - 2].confidence >= CONFIDENCE_OK) {
 286                 comp = (cal_data[ratio].steady_comp +
 287                         cal_data[ratio - 1].steady_comp +
 288                         cal_data[ratio - 2].steady_comp) / 3;
 289         } else if (cal_data[ratio].confidence >= CONFIDENCE_OK &&
 290                 cal_data[ratio - 1].confidence >= CONFIDENCE_OK &&
 291                 cal_data[ratio + 1].confidence >= CONFIDENCE_OK) {
 292                 comp = (cal_data[ratio].steady_comp +
 293                         cal_data[ratio - 1].steady_comp +
 294                         cal_data[ratio + 1].steady_comp) / 3;
 295         }
 296
 297         /* REVISIT: simple penalty of double idle injection */
 298         if (reduce_irq)
 299                 comp = ratio;
 300         /* do not exceed limit */
 301         if (comp + ratio >= MAX_TARGET_RATIO)
 302                 comp = MAX_TARGET_RATIO - ratio - 1;
 303
 304         return comp;
 305 }
 306
 307 static void adjust_compensation(int target_ratio, unsigned int win)
 308 {
 309         int delta;
 310         struct powerclamp_calibration_data *d = &cal_data[target_ratio];
 311
 312         /*
 313          * adjust compensations if confidence level has not been reached or
 314          * there are too many wakeups during the last idle injection period, we
 315          * cannot trust the data for compensation.
 316          */
 317         if (d->confidence >= CONFIDENCE_OK ||
 318                 atomic_read(&idle_wakeup_counter) >
 319                 win * num_online_cpus())
 320                 return;
 321
 322         delta = set_target_ratio - current_ratio;
 323         /* filter out bad data */
 324         if (delta >= 0 && delta <= (1+target_ratio/10)) {
 325                 if (d->steady_comp)
 326                         d->steady_comp =
 327                                 roundup(delta+d->steady_comp, 2)/2;
 328                 else
 329                         d->steady_comp = delta;
 330                 d->confidence++;
 331         }
 332 }
 333
 334 static bool powerclamp_adjust_controls(unsigned int target_ratio,
 335                                 unsigned int guard, unsigned int win)
 336 {
 337         static u64 msr_last, tsc_last;
 338         u64 msr_now, tsc_now;
 339         u64 val64;
 340
 341         /* check result for the last window */
 342         msr_now = pkg_state_counter();
 343         tsc_now = rdtsc();
 344
 345         /* calculate pkg cstate vs tsc ratio */
 346         if (!msr_last || !tsc_last)
 347                 current_ratio = 1;
 348         else if (tsc_now-tsc_last) {
 349                 val64 = 100*(msr_now-msr_last);
 350                 do_div(val64, (tsc_now-tsc_last));
 351                 current_ratio = val64;
 352         }
 353
 354         /* update record */
 355         msr_last = msr_now;
 356         tsc_last = tsc_now;
 357
 358         adjust_compensation(target_ratio, win);
 359         /*
 360          * too many external interrupts, set flag such
 361          * that we can take measure later.
 362          */
 363         reduce_irq = atomic_read(&idle_wakeup_counter) >=
 364                 2 * win * num_online_cpus();
 365
 366         atomic_set(&idle_wakeup_counter, 0);
 367         /* if we are above target+guard, skip */
 368         return set_target_ratio + guard <= current_ratio;
 369 }
 370
 371 static int clamp_thread(void *arg)
 372 {
 373         int cpunr = (unsigned long)arg;
 374         DEFINE_TIMER(wakeup_timer, noop_timer, 0, 0);
 375         static const struct sched_param param = {
 376                 .sched_priority = MAX_USER_RT_PRIO/2,
 377         };
 378         unsigned int count = 0;
 379         unsigned int target_ratio;
 380
 381         set_bit(cpunr, cpu_clamping_mask);
 382         set_freezable();
 383         init_timer_on_stack(&wakeup_timer);
 384         sched_setscheduler(current, SCHED_FIFO, &param);
 385
 386         while (true == clamping && !kthread_should_stop() &&
 387                 cpu_online(cpunr)) {
 388                 int sleeptime;
 389                 unsigned long target_jiffies;
 390                 unsigned int guard;
 391                 unsigned int compensation = 0;
 392                 int interval; /* jiffies to sleep for each attempt */
 393                 unsigned int duration_jiffies = msecs_to_jiffies(duration);
 394                 unsigned int window_size_now;
 395
 396                 try_to_freeze();
 397                 /*
 398                  * make sure user selected ratio does not take effect until
 399                  * the next round. adjust target_ratio if user has changed
 400                  * target such that we can converge quickly.
 401                  */
 402                 target_ratio = set_target_ratio;
 403                 guard = 1 + target_ratio/20;
 404                 window_size_now = window_size;
 405                 count++;
 406
 407                 /*
 408                  * systems may have different ability to enter package level
 409                  * c-states, thus we need to compensate the injected idle ratio
 410                  * to achieve the actual target reported by the HW.
 411                  */
 412                 compensation = get_compensation(target_ratio);
 413                 interval = duration_jiffies*100/(target_ratio+compensation);
 414
 415                 /* align idle time */
 416                 target_jiffies = roundup(jiffies, interval);
 417                 sleeptime = target_jiffies - jiffies;
 418                 if (sleeptime <= 0)
 419                         sleeptime = 1;
 420                 schedule_timeout_interruptible(sleeptime);
 421                 /*
 422                  * only elected controlling cpu can collect stats and update
 423                  * control parameters.
 424                  */
 425                 if (cpunr == control_cpu && !(count%window_size_now)) {
 426                         should_skip =
 427                                 powerclamp_adjust_controls(target_ratio,
 428                                                         guard, window_size_now);
 429                         smp_mb();
 430                 }
 431
 432                 if (should_skip)
 433                         continue;
 434
 435                 target_jiffies = jiffies + duration_jiffies;
 436                 mod_timer(&wakeup_timer, target_jiffies);
 437                 if (unlikely(local_softirq_pending()))
 438                         continue;
 439                 /*
 440                  * stop tick sched during idle time, interrupts are still
 441                  * allowed. thus jiffies are updated properly.
 442                  */
 443                 preempt_disable();
 444                 /* mwait until target jiffies is reached */
 445                 while (time_before(jiffies, target_jiffies)) {
 446                         unsigned long ecx = 1;
 447                         unsigned long eax = target_mwait;
 448
 449                         /*
 450                          * REVISIT: may call enter_idle() to notify drivers who
 451                          * can save power during cpu idle. same for exit_idle()
 452                          */
 453                         local_touch_nmi();
 454                         stop_critical_timings();
 455                         mwait_idle_with_hints(eax, ecx);
 456                         start_critical_timings();
 457                         atomic_inc(&idle_wakeup_counter);
 458                 }
 459                 preempt_enable();
 460         }
 461         del_timer_sync(&wakeup_timer);
 462         clear_bit(cpunr, cpu_clamping_mask);
 463
 464         return 0;
 465 }
 466
 467 /*
 468  * 1 HZ polling while clamping is active, useful for userspace
 469  * to monitor actual idle ratio.
 470  */
 471 static void poll_pkg_cstate(struct work_struct *dummy);
 472 static DECLARE_DELAYED_WORK(poll_pkg_cstate_work, poll_pkg_cstate);
 473 static void poll_pkg_cstate(struct work_struct *dummy)
 474 {
 475         static u64 msr_last;
 476         static u64 tsc_last;
 477         static unsigned long jiffies_last;
 478
 479         u64 msr_now;
 480         unsigned long jiffies_now;
 481         u64 tsc_now;
 482         u64 val64;
 483
 484         msr_now = pkg_state_counter();
 485         tsc_now = rdtsc();
 486         jiffies_now = jiffies;
 487
 488         /* calculate pkg cstate vs tsc ratio */
 489         if (!msr_last || !tsc_last)
 490                 pkg_cstate_ratio_cur = 1;
 491         else {
 492                 if (tsc_now - tsc_last) {
 493                         val64 = 100 * (msr_now - msr_last);
 494                         do_div(val64, (tsc_now - tsc_last));
 495                         pkg_cstate_ratio_cur = val64;
 496                 }
 497         }
 498
 499         /* update record */
 500         msr_last = msr_now;
 501         jiffies_last = jiffies_now;
 502         tsc_last = tsc_now;
 503
 504         if (true == clamping)
 505                 schedule_delayed_work(&poll_pkg_cstate_work, HZ);
 506 }
 507
 508 static int start_power_clamp(void)
 509 {
 510         unsigned long cpu;
 511         struct task_struct *thread;
 512
 513         /* check if pkg cstate counter is completely 0, abort in this case */
 514         if (!has_pkg_state_counter()) {
 515                 pr_err("pkg cstate counter not functional, abort\n");
 516                 return -EINVAL;
 517         }
 518
 519         set_target_ratio = clamp(set_target_ratio, 0U, MAX_TARGET_RATIO - 1);
 520         /* prevent cpu hotplug */
 521         get_online_cpus();
 522
 523         /* prefer BSP */
 524         control_cpu = 0;
 525         if (!cpu_online(control_cpu))
 526                 control_cpu = smp_processor_id();
 527
 528         clamping = true;
 529         schedule_delayed_work(&poll_pkg_cstate_work, 0);
 530
 531         /* start one thread per online cpu */
 532         for_each_online_cpu(cpu) {
 533                 struct task_struct **p =
 534                         per_cpu_ptr(powerclamp_thread, cpu);
 535
 536                 thread = kthread_create_on_node(clamp_thread,
 537                                                 (void *) cpu,
 538                                                 cpu_to_node(cpu),
 539                                                 "kidle_inject/%ld", cpu);
 540                 /* bind to cpu here */
 541                 if (likely(!IS_ERR(thread))) {
 542                         kthread_bind(thread, cpu);
 543                         wake_up_process(thread);
 544                         *p = thread;
 545                 }
 546
 547         }
 548         put_online_cpus();
 549
 550         return 0;
 551 }
 552
 553 static void end_power_clamp(void)
 554 {
 555         int i;
 556         struct task_struct *thread;
 557
 558         clamping = false;
 559         /*
 560          * make clamping visible to other cpus and give per cpu clamping threads
 561          * sometime to exit, or gets killed later.
 562          */
 563         smp_mb();
 564         msleep(20);
 565         if (bitmap_weight(cpu_clamping_mask, num_possible_cpus())) {
 566                 for_each_set_bit(i, cpu_clamping_mask, num_possible_cpus()) {
 567                         pr_debug("clamping thread for cpu %d alive, kill\n", i);
 568                         thread = *per_cpu_ptr(powerclamp_thread, i);
 569                         kthread_stop(thread);
 570                 }
 571         }
 572 }
 573
 574 static int powerclamp_cpu_callback(struct notifier_block *nfb,
 575                                 unsigned long action, void *hcpu)
 576 {
 577         unsigned long cpu = (unsigned long)hcpu;
 578         struct task_struct *thread;
 579         struct task_struct **percpu_thread =
 580                 per_cpu_ptr(powerclamp_thread, cpu);
 581
 582         if (false == clamping)
 583                 goto exit_ok;
 584
 585         switch (action) {
 586         case CPU_ONLINE:
 587                 thread = kthread_create_on_node(clamp_thread,
 588                                                 (void *) cpu,
 589                                                 cpu_to_node(cpu),
 590                                                 "kidle_inject/%lu", cpu);
 591                 if (likely(!IS_ERR(thread))) {
 592                         kthread_bind(thread, cpu);
 593                         wake_up_process(thread);
 594                         *percpu_thread = thread;
 595                 }
 596                 /* prefer BSP as controlling CPU */
 597                 if (cpu == 0) {
 598                         control_cpu = 0;
 599                         smp_mb();
 600                 }
 601                 break;
 602         case CPU_DEAD:
 603                 if (test_bit(cpu, cpu_clamping_mask)) {
 604                         pr_err("cpu %lu dead but powerclamping thread is not\n",
 605                                 cpu);
 606                         kthread_stop(*percpu_thread);
 607                 }
 608                 if (cpu == control_cpu) {
 609                         control_cpu = smp_processor_id();
 610                         smp_mb();
 611                 }
 612         }
 613
 614 exit_ok:
 615         return NOTIFY_OK;
 616 }
 617
 618 static struct notifier_block powerclamp_cpu_notifier = {
 619         .notifier_call = powerclamp_cpu_callback,
 620 };
 621
 622 static int powerclamp_get_max_state(struct thermal_cooling_device *cdev,
 623                                  unsigned long *state)
 624 {
 625         *state = MAX_TARGET_RATIO;
 626
 627         return 0;
 628 }
 629
 630 static int powerclamp_get_cur_state(struct thermal_cooling_device *cdev,
 631                                  unsigned long *state)
 632 {
 633         if (true == clamping)
 634                 *state = pkg_cstate_ratio_cur;
 635         else
 636                 /* to save power, do not poll idle ratio while not clamping */
 637                 *state = -1; /* indicates invalid state */
 638
 639         return 0;
 640 }
 641
 642 static int powerclamp_set_cur_state(struct thermal_cooling_device *cdev,
 643                                  unsigned long new_target_ratio)
 644 {
 645         int ret = 0;
 646
 647         new_target_ratio = clamp(new_target_ratio, 0UL,
 648                                 (unsigned long) (MAX_TARGET_RATIO-1));
 649         if (set_target_ratio == 0 && new_target_ratio > 0) {
 650                 pr_info("Start idle injection to reduce power\n");
 651                 set_target_ratio = new_target_ratio;
 652                 ret = start_power_clamp();
 653                 goto exit_set;
 654         } else  if (set_target_ratio > 0 && new_target_ratio == 0) {
 655                 pr_info("Stop forced idle injection\n");
 656                 set_target_ratio = 0;
 657                 end_power_clamp();
 658         } else  /* adjust currently running */ {
 659                 set_target_ratio = new_target_ratio;
 660                 /* make new set_target_ratio visible to other cpus */
 661                 smp_mb();
 662         }
 663
 664 exit_set:
 665         return ret;
 666 }
 667
 668 /* bind to generic thermal layer as cooling device*/
 669 static struct thermal_cooling_device_ops powerclamp_cooling_ops = {
 670         .get_max_state = powerclamp_get_max_state,
 671         .get_cur_state = powerclamp_get_cur_state,
 672         .set_cur_state = powerclamp_set_cur_state,
 673 };
 674
 675 /* runs on Nehalem and later */
 676 static const struct x86_cpu_id intel_powerclamp_ids[] __initconst = {
 677         { X86_VENDOR_INTEL, 6, 0x1a},
 678         { X86_VENDOR_INTEL, 6, 0x1c},
 679         { X86_VENDOR_INTEL, 6, 0x1e},
 680         { X86_VENDOR_INTEL, 6, 0x1f},
 681         { X86_VENDOR_INTEL, 6, 0x25},
 682         { X86_VENDOR_INTEL, 6, 0x26},
 683         { X86_VENDOR_INTEL, 6, 0x2a},
 684         { X86_VENDOR_INTEL, 6, 0x2c},
 685         { X86_VENDOR_INTEL, 6, 0x2d},
 686         { X86_VENDOR_INTEL, 6, 0x2e},
 687         { X86_VENDOR_INTEL, 6, 0x2f},
 688         { X86_VENDOR_INTEL, 6, 0x37},
 689         { X86_VENDOR_INTEL, 6, 0x3a},
 690         { X86_VENDOR_INTEL, 6, 0x3c},
 691         { X86_VENDOR_INTEL, 6, 0x3d},
 692         { X86_VENDOR_INTEL, 6, 0x3e},
 693         { X86_VENDOR_INTEL, 6, 0x3f},
 694         { X86_VENDOR_INTEL, 6, 0x45},
 695         { X86_VENDOR_INTEL, 6, 0x46},
 696         { X86_VENDOR_INTEL, 6, 0x47},
 697         { X86_VENDOR_INTEL, 6, 0x4c},
 698         { X86_VENDOR_INTEL, 6, 0x4d},
 699         { X86_VENDOR_INTEL, 6, 0x4e},
 700         { X86_VENDOR_INTEL, 6, 0x4f},
 701         { X86_VENDOR_INTEL, 6, 0x56},
 702         { X86_VENDOR_INTEL, 6, 0x57},
 703         { X86_VENDOR_INTEL, 6, 0x5e},
 704         {}
 705 };
 706 MODULE_DEVICE_TABLE(x86cpu, intel_powerclamp_ids);
 707
 708 static int __init powerclamp_probe(void)
 709 {
 710         if (!x86_match_cpu(intel_powerclamp_ids)) {
 711                 pr_err("Intel powerclamp does not run on family %d model %d\n",
 712                                 boot_cpu_data.x86, boot_cpu_data.x86_model);
 713                 return -ENODEV;
 714         }
 715         if (!boot_cpu_has(X86_FEATURE_NONSTOP_TSC) ||
 716                 !boot_cpu_has(X86_FEATURE_CONSTANT_TSC) ||
 717                 !boot_cpu_has(X86_FEATURE_MWAIT) ||
 718                 !boot_cpu_has(X86_FEATURE_ARAT))
 719                 return -ENODEV;
 720
 721         /* find the deepest mwait value */
 722         find_target_mwait();
 723
 724         return 0;
 725 }
 726
 727 static int powerclamp_debug_show(struct seq_file *m, void *unused)
 728 {
 729         int i = 0;
 730
 731         seq_printf(m, "controlling cpu: %d\n", control_cpu);
 732         seq_printf(m, "pct confidence steady dynamic (compensation)\n");
 733         for (i = 0; i < MAX_TARGET_RATIO; i++) {
 734                 seq_printf(m, "%d\t%lu\t%lu\t%lu\n",
 735                         i,
 736                         cal_data[i].confidence,
 737                         cal_data[i].steady_comp,
 738                         cal_data[i].dynamic_comp);
 739         }
 740
 741         return 0;
 742 }
 743
 744 static int powerclamp_debug_open(struct inode *inode,
 745                         struct file *file)
 746 {
 747         return single_open(file, powerclamp_debug_show, inode->i_private);
 748 }
 749
 750 static const struct file_operations powerclamp_debug_fops = {
 751         .open           = powerclamp_debug_open,
 752         .read           = seq_read,
 753         .llseek         = seq_lseek,
 754         .release        = single_release,
 755         .owner          = THIS_MODULE,
 756 };
 757
 758 static inline void powerclamp_create_debug_files(void)
 759 {
 760         debug_dir = debugfs_create_dir("intel_powerclamp", NULL);
 761         if (!debug_dir)
 762                 return;
 763
 764         if (!debugfs_create_file("powerclamp_calib", S_IRUGO, debug_dir,
 765                                         cal_data, &powerclamp_debug_fops))
 766                 goto file_error;
 767
 768         return;
 769
 770 file_error:
 771         debugfs_remove_recursive(debug_dir);
 772 }
 773
 774 static int __init powerclamp_init(void)
 775 {
 776         int retval;
 777         int bitmap_size;
 778
 779         bitmap_size = BITS_TO_LONGS(num_possible_cpus()) * sizeof(long);
 780         cpu_clamping_mask = kzalloc(bitmap_size, GFP_KERNEL);
 781         if (!cpu_clamping_mask)
 782                 return -ENOMEM;
 783
 784         /* probe cpu features and ids here */
 785         retval = powerclamp_probe();
 786         if (retval)
 787                 goto exit_free;
 788
 789         /* set default limit, maybe adjusted during runtime based on feedback */
 790         window_size = 2;
 791         register_hotcpu_notifier(&powerclamp_cpu_notifier);
 792
 793         powerclamp_thread = alloc_percpu(struct task_struct *);
 794         if (!powerclamp_thread) {
 795                 retval = -ENOMEM;
 796                 goto exit_unregister;
 797         }
 798
 799         cooling_dev = thermal_cooling_device_register("intel_powerclamp", NULL,
 800                                                 &powerclamp_cooling_ops);
 801         if (IS_ERR(cooling_dev)) {
 802                 retval = -ENODEV;
 803                 goto exit_free_thread;
 804         }
 805
 806         if (!duration)
 807                 duration = jiffies_to_msecs(DEFAULT_DURATION_JIFFIES);
 808
 809         powerclamp_create_debug_files();
 810
 811         return 0;
 812
 813 exit_free_thread:
 814         free_percpu(powerclamp_thread);
 815 exit_unregister:
 816         unregister_hotcpu_notifier(&powerclamp_cpu_notifier);
 817 exit_free:
 818         kfree(cpu_clamping_mask);
 819         return retval;
 820 }
 821 module_init(powerclamp_init);
 822
 823 static void __exit powerclamp_exit(void)
 824 {
 825         unregister_hotcpu_notifier(&powerclamp_cpu_notifier);
 826         end_power_clamp();
 827         free_percpu(powerclamp_thread);
 828         thermal_cooling_device_unregister(cooling_dev);
 829         kfree(cpu_clamping_mask);
 830
 831         cancel_delayed_work_sync(&poll_pkg_cstate_work);
 832         debugfs_remove_recursive(debug_dir);
 833 }
 834 module_exit(powerclamp_exit);
 835
 836 MODULE_LICENSE("GPL");
 837 MODULE_AUTHOR("Arjan van de Ven <arjan@linux.intel.com>");
 838 MODULE_AUTHOR("Jacob Pan <jacob.jun.pan@linux.intel.com>");
 839 MODULE_DESCRIPTION("Package Level C-state Idle Injection for Intel CPUs");