drivers/idle/intel_idle.c

   1 /*
   2  * intel_idle.c - native hardware idle loop for modern Intel processors
   3  *
   4  * Copyright (c) 2013, Intel Corporation.
   5  * Len Brown <len.brown@intel.com>
   6  *
   7  * This program is free software; you can redistribute it and/or modify it
   8  * under the terms and conditions of the GNU General Public License,
   9  * version 2, as published by the Free Software Foundation.
  10  *
  11  * This program is distributed in the hope it will be useful, but WITHOUT
  12  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  13  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
  14  * more details.
  15  *
  16  * You should have received a copy of the GNU General Public License along with
  17  * this program; if not, write to the Free Software Foundation, Inc.,
  18  * 51 Franklin St - Fifth Floor, Boston, MA 02110-1301 USA.
  19  */
  20
  21 /*
  22  * intel_idle is a cpuidle driver that loads on specific Intel processors
  23  * in lieu of the legacy ACPI processor_idle driver.  The intent is to
  24  * make Linux more efficient on these processors, as intel_idle knows
  25  * more than ACPI, as well as make Linux more immune to ACPI BIOS bugs.
  26  */
  27
  28 /*
  29  * Design Assumptions
  30  *
  31  * All CPUs have same idle states as boot CPU
  32  *
  33  * Chipset BM_STS (bus master status) bit is a NOP
  34  *      for preventing entry into deep C-stats
  35  */
  36
  37 /*
  38  * Known limitations
  39  *
  40  * The driver currently initializes for_each_online_cpu() upon modprobe.
  41  * It it unaware of subsequent processors hot-added to the system.
  42  * This means that if you boot with maxcpus=n and later online
  43  * processors above n, those processors will use C1 only.
  44  *
  45  * ACPI has a .suspend hack to turn off deep c-statees during suspend
  46  * to avoid complications with the lapic timer workaround.
  47  * Have not seen issues with suspend, but may need same workaround here.
  48  *
  49  */
  50
  51 /* un-comment DEBUG to enable pr_debug() statements */
  52 #define DEBUG
  53
  54 #include <linux/kernel.h>
  55 #include <linux/cpuidle.h>
  56 #include <linux/tick.h>
  57 #include <trace/events/power.h>
  58 #include <linux/sched.h>
  59 #include <linux/notifier.h>
  60 #include <linux/cpu.h>
  61 #include <linux/moduleparam.h>
  62 #include <asm/cpu_device_id.h>
  63 #include <asm/intel-family.h>
  64 #include <asm/mwait.h>
  65 #include <asm/msr.h>
  66
  67 #define INTEL_IDLE_VERSION "0.4.1"
  68 #define PREFIX "intel_idle: "
  69
  70 static struct cpuidle_driver intel_idle_driver = {
  71         .name = "intel_idle",
  72         .owner = THIS_MODULE,
  73 };
  74 /* intel_idle.max_cstate=0 disables driver */
  75 static int max_cstate = CPUIDLE_STATE_MAX - 1;
  76
  77 static unsigned int mwait_substates;
  78
  79 #define LAPIC_TIMER_ALWAYS_RELIABLE 0xFFFFFFFF
  80 /* Reliable LAPIC Timer States, bit 1 for C1 etc.  */
  81 static unsigned int lapic_timer_reliable_states = (1 << 1);      /* Default to only C1 */
  82
  83 struct idle_cpu {
  84         struct cpuidle_state *state_table;
  85
  86         /*
  87          * Hardware C-state auto-demotion may not always be optimal.
  88          * Indicate which enable bits to clear here.
  89          */
  90         unsigned long auto_demotion_disable_flags;
  91         bool byt_auto_demotion_disable_flag;
  92         bool disable_promotion_to_c1e;
  93 };
  94
  95 static const struct idle_cpu *icpu;
  96 static struct cpuidle_device __percpu *intel_idle_cpuidle_devices;
  97 static int intel_idle(struct cpuidle_device *dev,
  98                         struct cpuidle_driver *drv, int index);
  99 static void intel_idle_freeze(struct cpuidle_device *dev,
 100                               struct cpuidle_driver *drv, int index);
 101 static int intel_idle_cpu_init(int cpu);
 102
 103 static struct cpuidle_state *cpuidle_state_table;
 104
 105 /*
 106  * Set this flag for states where the HW flushes the TLB for us
 107  * and so we don't need cross-calls to keep it consistent.
 108  * If this flag is set, SW flushes the TLB, so even if the
 109  * HW doesn't do the flushing, this flag is safe to use.
 110  */
 111 #define CPUIDLE_FLAG_TLB_FLUSHED        0x10000
 112
 113 /*
 114  * MWAIT takes an 8-bit "hint" in EAX "suggesting"
 115  * the C-state (top nibble) and sub-state (bottom nibble)
 116  * 0x00 means "MWAIT(C1)", 0x10 means "MWAIT(C2)" etc.
 117  *
 118  * We store the hint at the top of our "flags" for each state.
 119  */
 120 #define flg2MWAIT(flags) (((flags) >> 24) & 0xFF)
 121 #define MWAIT2flg(eax) ((eax & 0xFF) << 24)
 122
 123 /*
 124  * States are indexed by the cstate number,
 125  * which is also the index into the MWAIT hint array.
 126  * Thus C0 is a dummy.
 127  */
 128 static struct cpuidle_state nehalem_cstates[] = {
 129         {
 130                 .name = "C1-NHM",
 131                 .desc = "MWAIT 0x00",
 132                 .flags = MWAIT2flg(0x00),
 133                 .exit_latency = 3,
 134                 .target_residency = 6,
 135                 .enter = &intel_idle,
 136                 .enter_freeze = intel_idle_freeze, },
 137         {
 138                 .name = "C1E-NHM",
 139                 .desc = "MWAIT 0x01",
 140                 .flags = MWAIT2flg(0x01),
 141                 .exit_latency = 10,
 142                 .target_residency = 20,
 143                 .enter = &intel_idle,
 144                 .enter_freeze = intel_idle_freeze, },
 145         {
 146                 .name = "C3-NHM",
 147                 .desc = "MWAIT 0x10",
 148                 .flags = MWAIT2flg(0x10) | CPUIDLE_FLAG_TLB_FLUSHED,
 149                 .exit_latency = 20,
 150                 .target_residency = 80,
 151                 .enter = &intel_idle,
 152                 .enter_freeze = intel_idle_freeze, },
 153         {
 154                 .name = "C6-NHM",
 155                 .desc = "MWAIT 0x20",
 156                 .flags = MWAIT2flg(0x20) | CPUIDLE_FLAG_TLB_FLUSHED,
 157                 .exit_latency = 200,
 158                 .target_residency = 800,
 159                 .enter = &intel_idle,
 160                 .enter_freeze = intel_idle_freeze, },
 161         {
 162                 .enter = NULL }
 163 };
 164
 165 static struct cpuidle_state snb_cstates[] = {
 166         {
 167                 .name = "C1-SNB",
 168                 .desc = "MWAIT 0x00",
 169                 .flags = MWAIT2flg(0x00),
 170                 .exit_latency = 2,
 171                 .target_residency = 2,
 172                 .enter = &intel_idle,
 173                 .enter_freeze = intel_idle_freeze, },
 174         {
 175                 .name = "C1E-SNB",
 176                 .desc = "MWAIT 0x01",
 177                 .flags = MWAIT2flg(0x01),
 178                 .exit_latency = 10,
 179                 .target_residency = 20,
 180                 .enter = &intel_idle,
 181                 .enter_freeze = intel_idle_freeze, },
 182         {
 183                 .name = "C3-SNB",
 184                 .desc = "MWAIT 0x10",
 185                 .flags = MWAIT2flg(0x10) | CPUIDLE_FLAG_TLB_FLUSHED,
 186                 .exit_latency = 80,
 187                 .target_residency = 211,
 188                 .enter = &intel_idle,
 189                 .enter_freeze = intel_idle_freeze, },
 190         {
 191                 .name = "C6-SNB",
 192                 .desc = "MWAIT 0x20",
 193                 .flags = MWAIT2flg(0x20) | CPUIDLE_FLAG_TLB_FLUSHED,
 194                 .exit_latency = 104,
 195                 .target_residency = 345,
 196                 .enter = &intel_idle,
 197                 .enter_freeze = intel_idle_freeze, },
 198         {
 199                 .name = "C7-SNB",
 200                 .desc = "MWAIT 0x30",
 201                 .flags = MWAIT2flg(0x30) | CPUIDLE_FLAG_TLB_FLUSHED,
 202                 .exit_latency = 109,
 203                 .target_residency = 345,
 204                 .enter = &intel_idle,
 205                 .enter_freeze = intel_idle_freeze, },
 206         {
 207                 .enter = NULL }
 208 };
 209
 210 static struct cpuidle_state byt_cstates[] = {
 211         {
 212                 .name = "C1-BYT",
 213                 .desc = "MWAIT 0x00",
 214                 .flags = MWAIT2flg(0x00),
 215                 .exit_latency = 1,
 216                 .target_residency = 1,
 217                 .enter = &intel_idle,
 218                 .enter_freeze = intel_idle_freeze, },
 219         {
 220                 .name = "C6N-BYT",
 221                 .desc = "MWAIT 0x58",
 222                 .flags = MWAIT2flg(0x58) | CPUIDLE_FLAG_TLB_FLUSHED,
 223                 .exit_latency = 300,
 224                 .target_residency = 275,
 225                 .enter = &intel_idle,
 226                 .enter_freeze = intel_idle_freeze, },
 227         {
 228                 .name = "C6S-BYT",
 229                 .desc = "MWAIT 0x52",
 230                 .flags = MWAIT2flg(0x52) | CPUIDLE_FLAG_TLB_FLUSHED,
 231                 .exit_latency = 500,
 232                 .target_residency = 560,
 233                 .enter = &intel_idle,
 234                 .enter_freeze = intel_idle_freeze, },
 235         {
 236                 .name = "C7-BYT",
 237                 .desc = "MWAIT 0x60",
 238                 .flags = MWAIT2flg(0x60) | CPUIDLE_FLAG_TLB_FLUSHED,
 239                 .exit_latency = 1200,
 240                 .target_residency = 4000,
 241                 .enter = &intel_idle,
 242                 .enter_freeze = intel_idle_freeze, },
 243         {
 244                 .name = "C7S-BYT",
 245                 .desc = "MWAIT 0x64",
 246                 .flags = MWAIT2flg(0x64) | CPUIDLE_FLAG_TLB_FLUSHED,
 247                 .exit_latency = 10000,
 248                 .target_residency = 20000,
 249                 .enter = &intel_idle,
 250                 .enter_freeze = intel_idle_freeze, },
 251         {
 252                 .enter = NULL }
 253 };
 254
 255 static struct cpuidle_state cht_cstates[] = {
 256         {
 257                 .name = "C1-CHT",
 258                 .desc = "MWAIT 0x00",
 259                 .flags = MWAIT2flg(0x00),
 260                 .exit_latency = 1,
 261                 .target_residency = 1,
 262                 .enter = &intel_idle,
 263                 .enter_freeze = intel_idle_freeze, },
 264         {
 265                 .name = "C6N-CHT",
 266                 .desc = "MWAIT 0x58",
 267                 .flags = MWAIT2flg(0x58) | CPUIDLE_FLAG_TLB_FLUSHED,
 268                 .exit_latency = 80,
 269                 .target_residency = 275,
 270                 .enter = &intel_idle,
 271                 .enter_freeze = intel_idle_freeze, },
 272         {
 273                 .name = "C6S-CHT",
 274                 .desc = "MWAIT 0x52",
 275                 .flags = MWAIT2flg(0x52) | CPUIDLE_FLAG_TLB_FLUSHED,
 276                 .exit_latency = 200,
 277                 .target_residency = 560,
 278                 .enter = &intel_idle,
 279                 .enter_freeze = intel_idle_freeze, },
 280         {
 281                 .name = "C7-CHT",
 282                 .desc = "MWAIT 0x60",
 283                 .flags = MWAIT2flg(0x60) | CPUIDLE_FLAG_TLB_FLUSHED,
 284                 .exit_latency = 1200,
 285                 .target_residency = 4000,
 286                 .enter = &intel_idle,
 287                 .enter_freeze = intel_idle_freeze, },
 288         {
 289                 .name = "C7S-CHT",
 290                 .desc = "MWAIT 0x64",
 291                 .flags = MWAIT2flg(0x64) | CPUIDLE_FLAG_TLB_FLUSHED,
 292                 .exit_latency = 10000,
 293                 .target_residency = 20000,
 294                 .enter = &intel_idle,
 295                 .enter_freeze = intel_idle_freeze, },
 296         {
 297                 .enter = NULL }
 298 };
 299
 300 static struct cpuidle_state ivb_cstates[] = {
 301         {
 302                 .name = "C1-IVB",
 303                 .desc = "MWAIT 0x00",
 304                 .flags = MWAIT2flg(0x00),
 305                 .exit_latency = 1,
 306                 .target_residency = 1,
 307                 .enter = &intel_idle,
 308                 .enter_freeze = intel_idle_freeze, },
 309         {
 310                 .name = "C1E-IVB",
 311                 .desc = "MWAIT 0x01",
 312                 .flags = MWAIT2flg(0x01),
 313                 .exit_latency = 10,
 314                 .target_residency = 20,
 315                 .enter = &intel_idle,
 316                 .enter_freeze = intel_idle_freeze, },
 317         {
 318                 .name = "C3-IVB",
 319                 .desc = "MWAIT 0x10",
 320                 .flags = MWAIT2flg(0x10) | CPUIDLE_FLAG_TLB_FLUSHED,
 321                 .exit_latency = 59,
 322                 .target_residency = 156,
 323                 .enter = &intel_idle,
 324                 .enter_freeze = intel_idle_freeze, },
 325         {
 326                 .name = "C6-IVB",
 327                 .desc = "MWAIT 0x20",
 328                 .flags = MWAIT2flg(0x20) | CPUIDLE_FLAG_TLB_FLUSHED,
 329                 .exit_latency = 80,
 330                 .target_residency = 300,
 331                 .enter = &intel_idle,
 332                 .enter_freeze = intel_idle_freeze, },
 333         {
 334                 .name = "C7-IVB",
 335                 .desc = "MWAIT 0x30",
 336                 .flags = MWAIT2flg(0x30) | CPUIDLE_FLAG_TLB_FLUSHED,
 337                 .exit_latency = 87,
 338                 .target_residency = 300,
 339                 .enter = &intel_idle,
 340                 .enter_freeze = intel_idle_freeze, },
 341         {
 342                 .enter = NULL }
 343 };
 344
 345 static struct cpuidle_state ivt_cstates[] = {
 346         {
 347                 .name = "C1-IVT",
 348                 .desc = "MWAIT 0x00",
 349                 .flags = MWAIT2flg(0x00),
 350                 .exit_latency = 1,
 351                 .target_residency = 1,
 352                 .enter = &intel_idle,
 353                 .enter_freeze = intel_idle_freeze, },
 354         {
 355                 .name = "C1E-IVT",
 356                 .desc = "MWAIT 0x01",
 357                 .flags = MWAIT2flg(0x01),
 358                 .exit_latency = 10,
 359                 .target_residency = 80,
 360                 .enter = &intel_idle,
 361                 .enter_freeze = intel_idle_freeze, },
 362         {
 363                 .name = "C3-IVT",
 364                 .desc = "MWAIT 0x10",
 365                 .flags = MWAIT2flg(0x10) | CPUIDLE_FLAG_TLB_FLUSHED,
 366                 .exit_latency = 59,
 367                 .target_residency = 156,
 368                 .enter = &intel_idle,
 369                 .enter_freeze = intel_idle_freeze, },
 370         {
 371                 .name = "C6-IVT",
 372                 .desc = "MWAIT 0x20",
 373                 .flags = MWAIT2flg(0x20) | CPUIDLE_FLAG_TLB_FLUSHED,
 374                 .exit_latency = 82,
 375                 .target_residency = 300,
 376                 .enter = &intel_idle,
 377                 .enter_freeze = intel_idle_freeze, },
 378         {
 379                 .enter = NULL }
 380 };
 381
 382 static struct cpuidle_state ivt_cstates_4s[] = {
 383         {
 384                 .name = "C1-IVT-4S",
 385                 .desc = "MWAIT 0x00",
 386                 .flags = MWAIT2flg(0x00),
 387                 .exit_latency = 1,
 388                 .target_residency = 1,
 389                 .enter = &intel_idle,
 390                 .enter_freeze = intel_idle_freeze, },
 391         {
 392                 .name = "C1E-IVT-4S",
 393                 .desc = "MWAIT 0x01",
 394                 .flags = MWAIT2flg(0x01),
 395                 .exit_latency = 10,
 396                 .target_residency = 250,
 397                 .enter = &intel_idle,
 398                 .enter_freeze = intel_idle_freeze, },
 399         {
 400                 .name = "C3-IVT-4S",
 401                 .desc = "MWAIT 0x10",
 402                 .flags = MWAIT2flg(0x10) | CPUIDLE_FLAG_TLB_FLUSHED,
 403                 .exit_latency = 59,
 404                 .target_residency = 300,
 405                 .enter = &intel_idle,
 406                 .enter_freeze = intel_idle_freeze, },
 407         {
 408                 .name = "C6-IVT-4S",
 409                 .desc = "MWAIT 0x20",
 410                 .flags = MWAIT2flg(0x20) | CPUIDLE_FLAG_TLB_FLUSHED,
 411                 .exit_latency = 84,
 412                 .target_residency = 400,
 413                 .enter = &intel_idle,
 414                 .enter_freeze = intel_idle_freeze, },
 415         {
 416                 .enter = NULL }
 417 };
 418
 419 static struct cpuidle_state ivt_cstates_8s[] = {
 420         {
 421                 .name = "C1-IVT-8S",
 422                 .desc = "MWAIT 0x00",
 423                 .flags = MWAIT2flg(0x00),
 424                 .exit_latency = 1,
 425                 .target_residency = 1,
 426                 .enter = &intel_idle,
 427                 .enter_freeze = intel_idle_freeze, },
 428         {
 429                 .name = "C1E-IVT-8S",
 430                 .desc = "MWAIT 0x01",
 431                 .flags = MWAIT2flg(0x01),
 432                 .exit_latency = 10,
 433                 .target_residency = 500,
 434                 .enter = &intel_idle,
 435                 .enter_freeze = intel_idle_freeze, },
 436         {
 437                 .name = "C3-IVT-8S",
 438                 .desc = "MWAIT 0x10",
 439                 .flags = MWAIT2flg(0x10) | CPUIDLE_FLAG_TLB_FLUSHED,
 440                 .exit_latency = 59,
 441                 .target_residency = 600,
 442                 .enter = &intel_idle,
 443                 .enter_freeze = intel_idle_freeze, },
 444         {
 445                 .name = "C6-IVT-8S",
 446                 .desc = "MWAIT 0x20",
 447                 .flags = MWAIT2flg(0x20) | CPUIDLE_FLAG_TLB_FLUSHED,
 448                 .exit_latency = 88,
 449                 .target_residency = 700,
 450                 .enter = &intel_idle,
 451                 .enter_freeze = intel_idle_freeze, },
 452         {
 453                 .enter = NULL }
 454 };
 455
 456 static struct cpuidle_state hsw_cstates[] = {
 457         {
 458                 .name = "C1-HSW",
 459                 .desc = "MWAIT 0x00",
 460                 .flags = MWAIT2flg(0x00),
 461                 .exit_latency = 2,
 462                 .target_residency = 2,
 463                 .enter = &intel_idle,
 464                 .enter_freeze = intel_idle_freeze, },
 465         {
 466                 .name = "C1E-HSW",
 467                 .desc = "MWAIT 0x01",
 468                 .flags = MWAIT2flg(0x01),
 469                 .exit_latency = 10,
 470                 .target_residency = 20,
 471                 .enter = &intel_idle,
 472                 .enter_freeze = intel_idle_freeze, },
 473         {
 474                 .name = "C3-HSW",
 475                 .desc = "MWAIT 0x10",
 476                 .flags = MWAIT2flg(0x10) | CPUIDLE_FLAG_TLB_FLUSHED,
 477                 .exit_latency = 33,
 478                 .target_residency = 100,
 479                 .enter = &intel_idle,
 480                 .enter_freeze = intel_idle_freeze, },
 481         {
 482                 .name = "C6-HSW",
 483                 .desc = "MWAIT 0x20",
 484                 .flags = MWAIT2flg(0x20) | CPUIDLE_FLAG_TLB_FLUSHED,
 485                 .exit_latency = 133,
 486                 .target_residency = 400,
 487                 .enter = &intel_idle,
 488                 .enter_freeze = intel_idle_freeze, },
 489         {
 490                 .name = "C7s-HSW",
 491                 .desc = "MWAIT 0x32",
 492                 .flags = MWAIT2flg(0x32) | CPUIDLE_FLAG_TLB_FLUSHED,
 493                 .exit_latency = 166,
 494                 .target_residency = 500,
 495                 .enter = &intel_idle,
 496                 .enter_freeze = intel_idle_freeze, },
 497         {
 498                 .name = "C8-HSW",
 499                 .desc = "MWAIT 0x40",
 500                 .flags = MWAIT2flg(0x40) | CPUIDLE_FLAG_TLB_FLUSHED,
 501                 .exit_latency = 300,
 502                 .target_residency = 900,
 503                 .enter = &intel_idle,
 504                 .enter_freeze = intel_idle_freeze, },
 505         {
 506                 .name = "C9-HSW",
 507                 .desc = "MWAIT 0x50",
 508                 .flags = MWAIT2flg(0x50) | CPUIDLE_FLAG_TLB_FLUSHED,
 509                 .exit_latency = 600,
 510                 .target_residency = 1800,
 511                 .enter = &intel_idle,
 512                 .enter_freeze = intel_idle_freeze, },
 513         {
 514                 .name = "C10-HSW",
 515                 .desc = "MWAIT 0x60",
 516                 .flags = MWAIT2flg(0x60) | CPUIDLE_FLAG_TLB_FLUSHED,
 517                 .exit_latency = 2600,
 518                 .target_residency = 7700,
 519                 .enter = &intel_idle,
 520                 .enter_freeze = intel_idle_freeze, },
 521         {
 522                 .enter = NULL }
 523 };
 524 static struct cpuidle_state bdw_cstates[] = {
 525         {
 526                 .name = "C1-BDW",
 527                 .desc = "MWAIT 0x00",
 528                 .flags = MWAIT2flg(0x00),
 529                 .exit_latency = 2,
 530                 .target_residency = 2,
 531                 .enter = &intel_idle,
 532                 .enter_freeze = intel_idle_freeze, },
 533         {
 534                 .name = "C1E-BDW",
 535                 .desc = "MWAIT 0x01",
 536                 .flags = MWAIT2flg(0x01),
 537                 .exit_latency = 10,
 538                 .target_residency = 20,
 539                 .enter = &intel_idle,
 540                 .enter_freeze = intel_idle_freeze, },
 541         {
 542                 .name = "C3-BDW",
 543                 .desc = "MWAIT 0x10",
 544                 .flags = MWAIT2flg(0x10) | CPUIDLE_FLAG_TLB_FLUSHED,
 545                 .exit_latency = 40,
 546                 .target_residency = 100,
 547                 .enter = &intel_idle,
 548                 .enter_freeze = intel_idle_freeze, },
 549         {
 550                 .name = "C6-BDW",
 551                 .desc = "MWAIT 0x20",
 552                 .flags = MWAIT2flg(0x20) | CPUIDLE_FLAG_TLB_FLUSHED,
 553                 .exit_latency = 133,
 554                 .target_residency = 400,
 555                 .enter = &intel_idle,
 556                 .enter_freeze = intel_idle_freeze, },
 557         {
 558                 .name = "C7s-BDW",
 559                 .desc = "MWAIT 0x32",
 560                 .flags = MWAIT2flg(0x32) | CPUIDLE_FLAG_TLB_FLUSHED,
 561                 .exit_latency = 166,
 562                 .target_residency = 500,
 563                 .enter = &intel_idle,
 564                 .enter_freeze = intel_idle_freeze, },
 565         {
 566                 .name = "C8-BDW",
 567                 .desc = "MWAIT 0x40",
 568                 .flags = MWAIT2flg(0x40) | CPUIDLE_FLAG_TLB_FLUSHED,
 569                 .exit_latency = 300,
 570                 .target_residency = 900,
 571                 .enter = &intel_idle,
 572                 .enter_freeze = intel_idle_freeze, },
 573         {
 574                 .name = "C9-BDW",
 575                 .desc = "MWAIT 0x50",
 576                 .flags = MWAIT2flg(0x50) | CPUIDLE_FLAG_TLB_FLUSHED,
 577                 .exit_latency = 600,
 578                 .target_residency = 1800,
 579                 .enter = &intel_idle,
 580                 .enter_freeze = intel_idle_freeze, },
 581         {
 582                 .name = "C10-BDW",
 583                 .desc = "MWAIT 0x60",
 584                 .flags = MWAIT2flg(0x60) | CPUIDLE_FLAG_TLB_FLUSHED,
 585                 .exit_latency = 2600,
 586                 .target_residency = 7700,
 587                 .enter = &intel_idle,
 588                 .enter_freeze = intel_idle_freeze, },
 589         {
 590                 .enter = NULL }
 591 };
 592
 593 static struct cpuidle_state skl_cstates[] = {
 594         {
 595                 .name = "C1-SKL",
 596                 .desc = "MWAIT 0x00",
 597                 .flags = MWAIT2flg(0x00),
 598                 .exit_latency = 2,
 599                 .target_residency = 2,
 600                 .enter = &intel_idle,
 601                 .enter_freeze = intel_idle_freeze, },
 602         {
 603                 .name = "C1E-SKL",
 604                 .desc = "MWAIT 0x01",
 605                 .flags = MWAIT2flg(0x01),
 606                 .exit_latency = 10,
 607                 .target_residency = 20,
 608                 .enter = &intel_idle,
 609                 .enter_freeze = intel_idle_freeze, },
 610         {
 611                 .name = "C3-SKL",
 612                 .desc = "MWAIT 0x10",
 613                 .flags = MWAIT2flg(0x10) | CPUIDLE_FLAG_TLB_FLUSHED,
 614                 .exit_latency = 70,
 615                 .target_residency = 100,
 616                 .enter = &intel_idle,
 617                 .enter_freeze = intel_idle_freeze, },
 618         {
 619                 .name = "C6-SKL",
 620                 .desc = "MWAIT 0x20",
 621                 .flags = MWAIT2flg(0x20) | CPUIDLE_FLAG_TLB_FLUSHED,
 622                 .exit_latency = 85,
 623                 .target_residency = 200,
 624                 .enter = &intel_idle,
 625                 .enter_freeze = intel_idle_freeze, },
 626         {
 627                 .name = "C7s-SKL",
 628                 .desc = "MWAIT 0x33",
 629                 .flags = MWAIT2flg(0x33) | CPUIDLE_FLAG_TLB_FLUSHED,
 630                 .exit_latency = 124,
 631                 .target_residency = 800,
 632                 .enter = &intel_idle,
 633                 .enter_freeze = intel_idle_freeze, },
 634         {
 635                 .name = "C8-SKL",
 636                 .desc = "MWAIT 0x40",
 637                 .flags = MWAIT2flg(0x40) | CPUIDLE_FLAG_TLB_FLUSHED,
 638                 .exit_latency = 200,
 639                 .target_residency = 800,
 640                 .enter = &intel_idle,
 641                 .enter_freeze = intel_idle_freeze, },
 642         {
 643                 .name = "C9-SKL",
 644                 .desc = "MWAIT 0x50",
 645                 .flags = MWAIT2flg(0x50) | CPUIDLE_FLAG_TLB_FLUSHED,
 646                 .exit_latency = 480,
 647                 .target_residency = 5000,
 648                 .enter = &intel_idle,
 649                 .enter_freeze = intel_idle_freeze, },
 650         {
 651                 .name = "C10-SKL",
 652                 .desc = "MWAIT 0x60",
 653                 .flags = MWAIT2flg(0x60) | CPUIDLE_FLAG_TLB_FLUSHED,
 654                 .exit_latency = 890,
 655                 .target_residency = 5000,
 656                 .enter = &intel_idle,
 657                 .enter_freeze = intel_idle_freeze, },
 658         {
 659                 .enter = NULL }
 660 };
 661
 662 static struct cpuidle_state skx_cstates[] = {
 663         {
 664                 .name = "C1-SKX",
 665                 .desc = "MWAIT 0x00",
 666                 .flags = MWAIT2flg(0x00),
 667                 .exit_latency = 2,
 668                 .target_residency = 2,
 669                 .enter = &intel_idle,
 670                 .enter_freeze = intel_idle_freeze, },
 671         {
 672                 .name = "C1E-SKX",
 673                 .desc = "MWAIT 0x01",
 674                 .flags = MWAIT2flg(0x01),
 675                 .exit_latency = 10,
 676                 .target_residency = 20,
 677                 .enter = &intel_idle,
 678                 .enter_freeze = intel_idle_freeze, },
 679         {
 680                 .name = "C6-SKX",
 681                 .desc = "MWAIT 0x20",
 682                 .flags = MWAIT2flg(0x20) | CPUIDLE_FLAG_TLB_FLUSHED,
 683                 .exit_latency = 133,
 684                 .target_residency = 600,
 685                 .enter = &intel_idle,
 686                 .enter_freeze = intel_idle_freeze, },
 687         {
 688                 .enter = NULL }
 689 };
 690
 691 static struct cpuidle_state atom_cstates[] = {
 692         {
 693                 .name = "C1E-ATM",
 694                 .desc = "MWAIT 0x00",
 695                 .flags = MWAIT2flg(0x00),
 696                 .exit_latency = 10,
 697                 .target_residency = 20,
 698                 .enter = &intel_idle,
 699                 .enter_freeze = intel_idle_freeze, },
 700         {
 701                 .name = "C2-ATM",
 702                 .desc = "MWAIT 0x10",
 703                 .flags = MWAIT2flg(0x10),
 704                 .exit_latency = 20,
 705                 .target_residency = 80,
 706                 .enter = &intel_idle,
 707                 .enter_freeze = intel_idle_freeze, },
 708         {
 709                 .name = "C4-ATM",
 710                 .desc = "MWAIT 0x30",
 711                 .flags = MWAIT2flg(0x30) | CPUIDLE_FLAG_TLB_FLUSHED,
 712                 .exit_latency = 100,
 713                 .target_residency = 400,
 714                 .enter = &intel_idle,
 715                 .enter_freeze = intel_idle_freeze, },
 716         {
 717                 .name = "C6-ATM",
 718                 .desc = "MWAIT 0x52",
 719                 .flags = MWAIT2flg(0x52) | CPUIDLE_FLAG_TLB_FLUSHED,
 720                 .exit_latency = 140,
 721                 .target_residency = 560,
 722                 .enter = &intel_idle,
 723                 .enter_freeze = intel_idle_freeze, },
 724         {
 725                 .enter = NULL }
 726 };
 727 static struct cpuidle_state avn_cstates[] = {
 728         {
 729                 .name = "C1-AVN",
 730                 .desc = "MWAIT 0x00",
 731                 .flags = MWAIT2flg(0x00),
 732                 .exit_latency = 2,
 733                 .target_residency = 2,
 734                 .enter = &intel_idle,
 735                 .enter_freeze = intel_idle_freeze, },
 736         {
 737                 .name = "C6-AVN",
 738                 .desc = "MWAIT 0x51",
 739                 .flags = MWAIT2flg(0x51) | CPUIDLE_FLAG_TLB_FLUSHED,
 740                 .exit_latency = 15,
 741                 .target_residency = 45,
 742                 .enter = &intel_idle,
 743                 .enter_freeze = intel_idle_freeze, },
 744         {
 745                 .enter = NULL }
 746 };
 747 static struct cpuidle_state knl_cstates[] = {
 748         {
 749                 .name = "C1-KNL",
 750                 .desc = "MWAIT 0x00",
 751                 .flags = MWAIT2flg(0x00),
 752                 .exit_latency = 1,
 753                 .target_residency = 2,
 754                 .enter = &intel_idle,
 755                 .enter_freeze = intel_idle_freeze },
 756         {
 757                 .name = "C6-KNL",
 758                 .desc = "MWAIT 0x10",
 759                 .flags = MWAIT2flg(0x10) | CPUIDLE_FLAG_TLB_FLUSHED,
 760                 .exit_latency = 120,
 761                 .target_residency = 500,
 762                 .enter = &intel_idle,
 763                 .enter_freeze = intel_idle_freeze },
 764         {
 765                 .enter = NULL }
 766 };
 767
 768 static struct cpuidle_state bxt_cstates[] = {
 769         {
 770                 .name = "C1-BXT",
 771                 .desc = "MWAIT 0x00",
 772                 .flags = MWAIT2flg(0x00),
 773                 .exit_latency = 2,
 774                 .target_residency = 2,
 775                 .enter = &intel_idle,
 776                 .enter_freeze = intel_idle_freeze, },
 777         {
 778                 .name = "C1E-BXT",
 779                 .desc = "MWAIT 0x01",
 780                 .flags = MWAIT2flg(0x01),
 781                 .exit_latency = 10,
 782                 .target_residency = 20,
 783                 .enter = &intel_idle,
 784                 .enter_freeze = intel_idle_freeze, },
 785         {
 786                 .name = "C6-BXT",
 787                 .desc = "MWAIT 0x20",
 788                 .flags = MWAIT2flg(0x20) | CPUIDLE_FLAG_TLB_FLUSHED,
 789                 .exit_latency = 133,
 790                 .target_residency = 133,
 791                 .enter = &intel_idle,
 792                 .enter_freeze = intel_idle_freeze, },
 793         {
 794                 .name = "C7s-BXT",
 795                 .desc = "MWAIT 0x31",
 796                 .flags = MWAIT2flg(0x31) | CPUIDLE_FLAG_TLB_FLUSHED,
 797                 .exit_latency = 155,
 798                 .target_residency = 155,
 799                 .enter = &intel_idle,
 800                 .enter_freeze = intel_idle_freeze, },
 801         {
 802                 .name = "C8-BXT",
 803                 .desc = "MWAIT 0x40",
 804                 .flags = MWAIT2flg(0x40) | CPUIDLE_FLAG_TLB_FLUSHED,
 805                 .exit_latency = 1000,
 806                 .target_residency = 1000,
 807                 .enter = &intel_idle,
 808                 .enter_freeze = intel_idle_freeze, },
 809         {
 810                 .name = "C9-BXT",
 811                 .desc = "MWAIT 0x50",
 812                 .flags = MWAIT2flg(0x50) | CPUIDLE_FLAG_TLB_FLUSHED,
 813                 .exit_latency = 2000,
 814                 .target_residency = 2000,
 815                 .enter = &intel_idle,
 816                 .enter_freeze = intel_idle_freeze, },
 817         {
 818                 .name = "C10-BXT",
 819                 .desc = "MWAIT 0x60",
 820                 .flags = MWAIT2flg(0x60) | CPUIDLE_FLAG_TLB_FLUSHED,
 821                 .exit_latency = 10000,
 822                 .target_residency = 10000,
 823                 .enter = &intel_idle,
 824                 .enter_freeze = intel_idle_freeze, },
 825         {
 826                 .enter = NULL }
 827 };
 828
 829 /**
 830  * intel_idle
 831  * @dev: cpuidle_device
 832  * @drv: cpuidle driver
 833  * @index: index of cpuidle state
 834  *
 835  * Must be called under local_irq_disable().
 836  */
 837 static int intel_idle(struct cpuidle_device *dev,
 838                 struct cpuidle_driver *drv, int index)
 839 {
 840         unsigned long ecx = 1; /* break on interrupt flag */
 841         struct cpuidle_state *state = &drv->states[index];
 842         unsigned long eax = flg2MWAIT(state->flags);
 843         unsigned int cstate;
 844         int cpu = smp_processor_id();
 845
 846         cstate = (((eax) >> MWAIT_SUBSTATE_SIZE) & MWAIT_CSTATE_MASK) + 1;
 847
 848         /*
 849          * leave_mm() to avoid costly and often unnecessary wakeups
 850          * for flushing the user TLB's associated with the active mm.
 851          */
 852         if (state->flags & CPUIDLE_FLAG_TLB_FLUSHED)
 853                 leave_mm(cpu);
 854
 855         if (!(lapic_timer_reliable_states & (1 << (cstate))))
 856                 tick_broadcast_enter();
 857
 858         mwait_idle_with_hints(eax, ecx);
 859
 860         if (!(lapic_timer_reliable_states & (1 << (cstate))))
 861                 tick_broadcast_exit();
 862
 863         return index;
 864 }
 865
 866 /**
 867  * intel_idle_freeze - simplified "enter" callback routine for suspend-to-idle
 868  * @dev: cpuidle_device
 869  * @drv: cpuidle driver
 870  * @index: state index
 871  */
 872 static void intel_idle_freeze(struct cpuidle_device *dev,
 873                              struct cpuidle_driver *drv, int index)
 874 {
 875         unsigned long ecx = 1; /* break on interrupt flag */
 876         unsigned long eax = flg2MWAIT(drv->states[index].flags);
 877
 878         mwait_idle_with_hints(eax, ecx);
 879 }
 880
 881 static void __setup_broadcast_timer(void *arg)
 882 {
 883         unsigned long on = (unsigned long)arg;
 884
 885         if (on)
 886                 tick_broadcast_enable();
 887         else
 888                 tick_broadcast_disable();
 889 }
 890
 891 static int cpu_hotplug_notify(struct notifier_block *n,
 892                               unsigned long action, void *hcpu)
 893 {
 894         int hotcpu = (unsigned long)hcpu;
 895         struct cpuidle_device *dev;
 896
 897         switch (action & ~CPU_TASKS_FROZEN) {
 898         case CPU_ONLINE:
 899
 900                 if (lapic_timer_reliable_states != LAPIC_TIMER_ALWAYS_RELIABLE)
 901                         smp_call_function_single(hotcpu, __setup_broadcast_timer,
 902                                                  (void *)true, 1);
 903
 904                 /*
 905                  * Some systems can hotplug a cpu at runtime after
 906                  * the kernel has booted, we have to initialize the
 907                  * driver in this case
 908                  */
 909                 dev = per_cpu_ptr(intel_idle_cpuidle_devices, hotcpu);
 910                 if (dev->registered)
 911                         break;
 912
 913                 if (intel_idle_cpu_init(hotcpu))
 914                         return NOTIFY_BAD;
 915
 916                 break;
 917         }
 918         return NOTIFY_OK;
 919 }
 920
 921 static struct notifier_block cpu_hotplug_notifier = {
 922         .notifier_call = cpu_hotplug_notify,
 923 };
 924
 925 static void auto_demotion_disable(void *dummy)
 926 {
 927         unsigned long long msr_bits;
 928
 929         rdmsrl(MSR_NHM_SNB_PKG_CST_CFG_CTL, msr_bits);
 930         msr_bits &= ~(icpu->auto_demotion_disable_flags);
 931         wrmsrl(MSR_NHM_SNB_PKG_CST_CFG_CTL, msr_bits);
 932 }
 933 static void c1e_promotion_disable(void *dummy)
 934 {
 935         unsigned long long msr_bits;
 936
 937         rdmsrl(MSR_IA32_POWER_CTL, msr_bits);
 938         msr_bits &= ~0x2;
 939         wrmsrl(MSR_IA32_POWER_CTL, msr_bits);
 940 }
 941
 942 static const struct idle_cpu idle_cpu_nehalem = {
 943         .state_table = nehalem_cstates,
 944         .auto_demotion_disable_flags = NHM_C1_AUTO_DEMOTE | NHM_C3_AUTO_DEMOTE,
 945         .disable_promotion_to_c1e = true,
 946 };
 947
 948 static const struct idle_cpu idle_cpu_atom = {
 949         .state_table = atom_cstates,
 950 };
 951
 952 static const struct idle_cpu idle_cpu_lincroft = {
 953         .state_table = atom_cstates,
 954         .auto_demotion_disable_flags = ATM_LNC_C6_AUTO_DEMOTE,
 955 };
 956
 957 static const struct idle_cpu idle_cpu_snb = {
 958         .state_table = snb_cstates,
 959         .disable_promotion_to_c1e = true,
 960 };
 961
 962 static const struct idle_cpu idle_cpu_byt = {
 963         .state_table = byt_cstates,
 964         .disable_promotion_to_c1e = true,
 965         .byt_auto_demotion_disable_flag = true,
 966 };
 967
 968 static const struct idle_cpu idle_cpu_cht = {
 969         .state_table = cht_cstates,
 970         .disable_promotion_to_c1e = true,
 971         .byt_auto_demotion_disable_flag = true,
 972 };
 973
 974 static const struct idle_cpu idle_cpu_ivb = {
 975         .state_table = ivb_cstates,
 976         .disable_promotion_to_c1e = true,
 977 };
 978
 979 static const struct idle_cpu idle_cpu_ivt = {
 980         .state_table = ivt_cstates,
 981         .disable_promotion_to_c1e = true,
 982 };
 983
 984 static const struct idle_cpu idle_cpu_hsw = {
 985         .state_table = hsw_cstates,
 986         .disable_promotion_to_c1e = true,
 987 };
 988
 989 static const struct idle_cpu idle_cpu_bdw = {
 990         .state_table = bdw_cstates,
 991         .disable_promotion_to_c1e = true,
 992 };
 993
 994 static const struct idle_cpu idle_cpu_skl = {
 995         .state_table = skl_cstates,
 996         .disable_promotion_to_c1e = true,
 997 };
 998
 999 static const struct idle_cpu idle_cpu_skx = {
1000         .state_table = skx_cstates,
1001         .disable_promotion_to_c1e = true,
1002 };
1003
1004 static const struct idle_cpu idle_cpu_avn = {
1005         .state_table = avn_cstates,
1006         .disable_promotion_to_c1e = true,
1007 };
1008
1009 static const struct idle_cpu idle_cpu_knl = {
1010         .state_table = knl_cstates,
1011 };
1012
1013 static const struct idle_cpu idle_cpu_bxt = {
1014         .state_table = bxt_cstates,
1015         .disable_promotion_to_c1e = true,
1016 };
1017
1018 #define ICPU(model, cpu) \
1019         { X86_VENDOR_INTEL, 6, model, X86_FEATURE_MWAIT, (unsigned long)&cpu }
1020
1021 static const struct x86_cpu_id intel_idle_ids[] __initconst = {
1022         ICPU(INTEL_FAM6_NEHALEM_EP,             idle_cpu_nehalem),
1023         ICPU(INTEL_FAM6_NEHALEM,                idle_cpu_nehalem),
1024         ICPU(INTEL_FAM6_WESTMERE2,              idle_cpu_nehalem),
1025         ICPU(INTEL_FAM6_WESTMERE,               idle_cpu_nehalem),
1026         ICPU(INTEL_FAM6_WESTMERE_EP,            idle_cpu_nehalem),
1027         ICPU(INTEL_FAM6_NEHALEM_EX,             idle_cpu_nehalem),
1028         ICPU(INTEL_FAM6_ATOM_PINEVIEW,          idle_cpu_atom),
1029         ICPU(INTEL_FAM6_ATOM_LINCROFT,          idle_cpu_lincroft),
1030         ICPU(INTEL_FAM6_WESTMERE_EX,            idle_cpu_nehalem),
1031         ICPU(INTEL_FAM6_SANDYBRIDGE,            idle_cpu_snb),
1032         ICPU(INTEL_FAM6_SANDYBRIDGE_X,          idle_cpu_snb),
1033         ICPU(INTEL_FAM6_ATOM_CEDARVIEW,         idle_cpu_atom),
1034         ICPU(INTEL_FAM6_ATOM_SILVERMONT1,       idle_cpu_byt),
1035         ICPU(INTEL_FAM6_ATOM_AIRMONT,           idle_cpu_cht),
1036         ICPU(INTEL_FAM6_IVYBRIDGE,              idle_cpu_ivb),
1037         ICPU(INTEL_FAM6_IVYBRIDGE_X,            idle_cpu_ivt),
1038         ICPU(INTEL_FAM6_HASWELL_CORE,           idle_cpu_hsw),
1039         ICPU(INTEL_FAM6_HASWELL_X,              idle_cpu_hsw),
1040         ICPU(INTEL_FAM6_HASWELL_ULT,            idle_cpu_hsw),
1041         ICPU(INTEL_FAM6_HASWELL_GT3E,           idle_cpu_hsw),
1042         ICPU(INTEL_FAM6_ATOM_SILVERMONT2,       idle_cpu_avn),
1043         ICPU(INTEL_FAM6_BROADWELL_CORE,         idle_cpu_bdw),
1044         ICPU(INTEL_FAM6_BROADWELL_GT3E,         idle_cpu_bdw),
1045         ICPU(INTEL_FAM6_BROADWELL_X,            idle_cpu_bdw),
1046         ICPU(INTEL_FAM6_BROADWELL_XEON_D,       idle_cpu_bdw),
1047         ICPU(INTEL_FAM6_SKYLAKE_MOBILE,         idle_cpu_skl),
1048         ICPU(INTEL_FAM6_SKYLAKE_DESKTOP,        idle_cpu_skl),
1049         ICPU(INTEL_FAM6_KABYLAKE_MOBILE,        idle_cpu_skl),
1050         ICPU(INTEL_FAM6_KABYLAKE_DESKTOP,       idle_cpu_skl),
1051         ICPU(INTEL_FAM6_SKYLAKE_X,              idle_cpu_skx),
1052         ICPU(INTEL_FAM6_XEON_PHI_KNL,           idle_cpu_knl),
1053         ICPU(INTEL_FAM6_ATOM_GOLDMONT,          idle_cpu_bxt),
1054         {}
1055 };
1056
1057 /*
1058  * intel_idle_probe()
1059  */
1060 static int __init intel_idle_probe(void)
1061 {
1062         unsigned int eax, ebx, ecx;
1063         const struct x86_cpu_id *id;
1064
1065         if (max_cstate == 0) {
1066                 pr_debug(PREFIX "disabled\n");
1067                 return -EPERM;
1068         }
1069
1070         id = x86_match_cpu(intel_idle_ids);
1071         if (!id) {
1072                 if (boot_cpu_data.x86_vendor == X86_VENDOR_INTEL &&
1073                     boot_cpu_data.x86 == 6)
1074                         pr_debug(PREFIX "does not run on family %d model %d\n",
1075                                 boot_cpu_data.x86, boot_cpu_data.x86_model);
1076                 return -ENODEV;
1077         }
1078
1079         if (boot_cpu_data.cpuid_level < CPUID_MWAIT_LEAF)
1080                 return -ENODEV;
1081
1082         cpuid(CPUID_MWAIT_LEAF, &eax, &ebx, &ecx, &mwait_substates);
1083
1084         if (!(ecx & CPUID5_ECX_EXTENSIONS_SUPPORTED) ||
1085             !(ecx & CPUID5_ECX_INTERRUPT_BREAK) ||
1086             !mwait_substates)
1087                         return -ENODEV;
1088
1089         pr_debug(PREFIX "MWAIT substates: 0x%x\n", mwait_substates);
1090
1091         icpu = (const struct idle_cpu *)id->driver_data;
1092         cpuidle_state_table = icpu->state_table;
1093
1094         pr_debug(PREFIX "v" INTEL_IDLE_VERSION
1095                 " model 0x%X\n", boot_cpu_data.x86_model);
1096
1097         return 0;
1098 }
1099
1100 /*
1101  * intel_idle_cpuidle_devices_uninit()
1102  * Unregisters the cpuidle devices.
1103  */
1104 static void intel_idle_cpuidle_devices_uninit(void)
1105 {
1106         int i;
1107         struct cpuidle_device *dev;
1108
1109         for_each_online_cpu(i) {
1110                 dev = per_cpu_ptr(intel_idle_cpuidle_devices, i);
1111                 cpuidle_unregister_device(dev);
1112         }
1113 }
1114
1115 /*
1116  * ivt_idle_state_table_update(void)
1117  *
1118  * Tune IVT multi-socket targets
1119  * Assumption: num_sockets == (max_package_num + 1)
1120  */
1121 static void ivt_idle_state_table_update(void)
1122 {
1123         /* IVT uses a different table for 1-2, 3-4, and > 4 sockets */
1124         int cpu, package_num, num_sockets = 1;
1125
1126         for_each_online_cpu(cpu) {
1127                 package_num = topology_physical_package_id(cpu);
1128                 if (package_num + 1 > num_sockets) {
1129                         num_sockets = package_num + 1;
1130
1131                         if (num_sockets > 4) {
1132                                 cpuidle_state_table = ivt_cstates_8s;
1133                                 return;
1134                         }
1135                 }
1136         }
1137
1138         if (num_sockets > 2)
1139                 cpuidle_state_table = ivt_cstates_4s;
1140
1141         /* else, 1 and 2 socket systems use default ivt_cstates */
1142 }
1143
1144 /*
1145  * Translate IRTL (Interrupt Response Time Limit) MSR to usec
1146  */
1147
1148 static unsigned int irtl_ns_units[] = {
1149         1, 32, 1024, 32768, 1048576, 33554432, 0, 0 };
1150
1151 static unsigned long long irtl_2_usec(unsigned long long irtl)
1152 {
1153         unsigned long long ns;
1154
1155         ns = irtl_ns_units[(irtl >> 10) & 0x3];
1156
1157         return div64_u64((irtl & 0x3FF) * ns, 1000);
1158 }
1159 /*
1160  * bxt_idle_state_table_update(void)
1161  *
1162  * On BXT, we trust the IRTL to show the definitive maximum latency
1163  * We use the same value for target_residency.
1164  */
1165 static void bxt_idle_state_table_update(void)
1166 {
1167         unsigned long long msr;
1168
1169         rdmsrl(MSR_PKGC6_IRTL, msr);
1170         if (msr) {
1171                 unsigned int usec = irtl_2_usec(msr);
1172
1173                 bxt_cstates[2].exit_latency = usec;
1174                 bxt_cstates[2].target_residency = usec;
1175         }
1176
1177         rdmsrl(MSR_PKGC7_IRTL, msr);
1178         if (msr) {
1179                 unsigned int usec = irtl_2_usec(msr);
1180
1181                 bxt_cstates[3].exit_latency = usec;
1182                 bxt_cstates[3].target_residency = usec;
1183         }
1184
1185         rdmsrl(MSR_PKGC8_IRTL, msr);
1186         if (msr) {
1187                 unsigned int usec = irtl_2_usec(msr);
1188
1189                 bxt_cstates[4].exit_latency = usec;
1190                 bxt_cstates[4].target_residency = usec;
1191         }
1192
1193         rdmsrl(MSR_PKGC9_IRTL, msr);
1194         if (msr) {
1195                 unsigned int usec = irtl_2_usec(msr);
1196
1197                 bxt_cstates[5].exit_latency = usec;
1198                 bxt_cstates[5].target_residency = usec;
1199         }
1200
1201         rdmsrl(MSR_PKGC10_IRTL, msr);
1202         if (msr) {
1203                 unsigned int usec = irtl_2_usec(msr);
1204
1205                 bxt_cstates[6].exit_latency = usec;
1206                 bxt_cstates[6].target_residency = usec;
1207         }
1208
1209 }
1210 /*
1211  * sklh_idle_state_table_update(void)
1212  *
1213  * On SKL-H (model 0x5e) disable C8 and C9 if:
1214  * C10 is enabled and SGX disabled
1215  */
1216 static void sklh_idle_state_table_update(void)
1217 {
1218         unsigned long long msr;
1219         unsigned int eax, ebx, ecx, edx;
1220
1221
1222         /* if PC10 disabled via cmdline intel_idle.max_cstate=7 or shallower */
1223         if (max_cstate <= 7)
1224                 return;
1225
1226         /* if PC10 not present in CPUID.MWAIT.EDX */
1227         if ((mwait_substates & (0xF << 28)) == 0)
1228                 return;
1229
1230         rdmsrl(MSR_NHM_SNB_PKG_CST_CFG_CTL, msr);
1231
1232         /* PC10 is not enabled in PKG C-state limit */
1233         if ((msr & 0xF) != 8)
1234                 return;
1235
1236         ecx = 0;
1237         cpuid(7, &eax, &ebx, &ecx, &edx);
1238
1239         /* if SGX is present */
1240         if (ebx & (1 << 2)) {
1241
1242                 rdmsrl(MSR_IA32_FEATURE_CONTROL, msr);
1243
1244                 /* if SGX is enabled */
1245                 if (msr & (1 << 18))
1246                         return;
1247         }
1248
1249         skl_cstates[5].disabled = 1;    /* C8-SKL */
1250         skl_cstates[6].disabled = 1;    /* C9-SKL */
1251 }
1252 /*
1253  * intel_idle_state_table_update()
1254  *
1255  * Update the default state_table for this CPU-id
1256  */
1257
1258 static void intel_idle_state_table_update(void)
1259 {
1260         switch (boot_cpu_data.x86_model) {
1261
1262         case INTEL_FAM6_IVYBRIDGE_X:
1263                 ivt_idle_state_table_update();
1264                 break;
1265         case INTEL_FAM6_ATOM_GOLDMONT:
1266                 bxt_idle_state_table_update();
1267                 break;
1268         case INTEL_FAM6_SKYLAKE_DESKTOP:
1269                 sklh_idle_state_table_update();
1270                 break;
1271         }
1272 }
1273
1274 /*
1275  * intel_idle_cpuidle_driver_init()
1276  * allocate, initialize cpuidle_states
1277  */
1278 static void __init intel_idle_cpuidle_driver_init(void)
1279 {
1280         int cstate;
1281         struct cpuidle_driver *drv = &intel_idle_driver;
1282
1283         intel_idle_state_table_update();
1284
1285         drv->state_count = 1;
1286
1287         for (cstate = 0; cstate < CPUIDLE_STATE_MAX; ++cstate) {
1288                 int num_substates, mwait_hint, mwait_cstate;
1289
1290                 if ((cpuidle_state_table[cstate].enter == NULL) &&
1291                     (cpuidle_state_table[cstate].enter_freeze == NULL))
1292                         break;
1293
1294                 if (cstate + 1 > max_cstate) {
1295                         printk(PREFIX "max_cstate %d reached\n",
1296                                 max_cstate);
1297                         break;
1298                 }
1299
1300                 mwait_hint = flg2MWAIT(cpuidle_state_table[cstate].flags);
1301                 mwait_cstate = MWAIT_HINT2CSTATE(mwait_hint);
1302
1303                 /* number of sub-states for this state in CPUID.MWAIT */
1304                 num_substates = (mwait_substates >> ((mwait_cstate + 1) * 4))
1305                                         & MWAIT_SUBSTATE_MASK;
1306
1307                 /* if NO sub-states for this state in CPUID, skip it */
1308                 if (num_substates == 0)
1309                         continue;
1310
1311                 /* if state marked as disabled, skip it */
1312                 if (cpuidle_state_table[cstate].disabled != 0) {
1313                         pr_debug(PREFIX "state %s is disabled",
1314                                 cpuidle_state_table[cstate].name);
1315                         continue;
1316                 }
1317
1318
1319                 if (((mwait_cstate + 1) > 2) &&
1320                         !boot_cpu_has(X86_FEATURE_NONSTOP_TSC))
1321                         mark_tsc_unstable("TSC halts in idle"
1322                                         " states deeper than C2");
1323
1324                 drv->states[drv->state_count] = /* structure copy */
1325                         cpuidle_state_table[cstate];
1326
1327                 drv->state_count += 1;
1328         }
1329
1330         if (icpu->byt_auto_demotion_disable_flag) {
1331                 wrmsrl(MSR_CC6_DEMOTION_POLICY_CONFIG, 0);
1332                 wrmsrl(MSR_MC6_DEMOTION_POLICY_CONFIG, 0);
1333         }
1334 }
1335
1336
1337 /*
1338  * intel_idle_cpu_init()
1339  * allocate, initialize, register cpuidle_devices
1340  * @cpu: cpu/core to initialize
1341  */
1342 static int intel_idle_cpu_init(int cpu)
1343 {
1344         struct cpuidle_device *dev;
1345
1346         dev = per_cpu_ptr(intel_idle_cpuidle_devices, cpu);
1347
1348         dev->cpu = cpu;
1349
1350         if (cpuidle_register_device(dev)) {
1351                 pr_debug(PREFIX "cpuidle_register_device %d failed!\n", cpu);
1352                 return -EIO;
1353         }
1354
1355         if (icpu->auto_demotion_disable_flags)
1356                 smp_call_function_single(cpu, auto_demotion_disable, NULL, 1);
1357
1358         if (icpu->disable_promotion_to_c1e)
1359                 smp_call_function_single(cpu, c1e_promotion_disable, NULL, 1);
1360
1361         return 0;
1362 }
1363
1364 static int __init intel_idle_init(void)
1365 {
1366         int retval, i;
1367
1368         /* Do not load intel_idle at all for now if idle= is passed */
1369         if (boot_option_idle_override != IDLE_NO_OVERRIDE)
1370                 return -ENODEV;
1371
1372         retval = intel_idle_probe();
1373         if (retval)
1374                 return retval;
1375
1376         intel_idle_cpuidle_devices = alloc_percpu(struct cpuidle_device);
1377         if (intel_idle_cpuidle_devices == NULL)
1378                 return -ENOMEM;
1379
1380         intel_idle_cpuidle_driver_init();
1381         retval = cpuidle_register_driver(&intel_idle_driver);
1382         if (retval) {
1383                 struct cpuidle_driver *drv = cpuidle_get_driver();
1384                 printk(KERN_DEBUG PREFIX "intel_idle yielding to %s",
1385                         drv ? drv->name : "none");
1386                 free_percpu(intel_idle_cpuidle_devices);
1387                 return retval;
1388         }
1389
1390         cpu_notifier_register_begin();
1391
1392         for_each_online_cpu(i) {
1393                 retval = intel_idle_cpu_init(i);
1394                 if (retval) {
1395                         intel_idle_cpuidle_devices_uninit();
1396                         cpu_notifier_register_done();
1397                         cpuidle_unregister_driver(&intel_idle_driver);
1398                         free_percpu(intel_idle_cpuidle_devices);
1399                         return retval;
1400                 }
1401         }
1402         __register_cpu_notifier(&cpu_hotplug_notifier);
1403
1404         if (boot_cpu_has(X86_FEATURE_ARAT))     /* Always Reliable APIC Timer */
1405                 lapic_timer_reliable_states = LAPIC_TIMER_ALWAYS_RELIABLE;
1406         else
1407                 on_each_cpu(__setup_broadcast_timer, (void *)true, 1);
1408
1409         cpu_notifier_register_done();
1410
1411         pr_debug(PREFIX "lapic_timer_reliable_states 0x%x\n",
1412                 lapic_timer_reliable_states);
1413
1414         return 0;
1415 }
1416 device_initcall(intel_idle_init);
1417
1418 /*
1419  * We are not really modular, but we used to support that.  Meaning we also
1420  * support "intel_idle.max_cstate=..." at boot and also a read-only export of
1421  * it at /sys/module/intel_idle/parameters/max_cstate -- so using module_param
1422  * is the easiest way (currently) to continue doing that.
1423  */
1424 module_param(max_cstate, int, 0444);