arch/powerpc/platforms/powernv/idle.c

   1 /*
   2  * PowerNV cpuidle code
   3  *
   4  * Copyright 2015 IBM Corp.
   5  *
   6  * This program is free software; you can redistribute it and/or
   7  * modify it under the terms of the GNU General Public License
   8  * as published by the Free Software Foundation; either version
   9  * 2 of the License, or (at your option) any later version.
  10  */
  11
  12 #include <linux/types.h>
  13 #include <linux/mm.h>
  14 #include <linux/slab.h>
  15 #include <linux/of.h>
  16 #include <linux/device.h>
  17 #include <linux/cpu.h>
  18
  19 #include <asm/firmware.h>
  20 #include <asm/machdep.h>
  21 #include <asm/opal.h>
  22 #include <asm/cputhreads.h>
  23 #include <asm/cpuidle.h>
  24 #include <asm/code-patching.h>
  25 #include <asm/smp.h>
  26 #include <asm/runlatch.h>
  27 #include <asm/dbell.h>
  28
  29 #include "powernv.h"
  30 #include "subcore.h"
  31
  32 /* Power ISA 3.0 allows for stop states 0x0 - 0xF */
  33 #define MAX_STOP_STATE  0xF
  34
  35 #define P9_STOP_SPR_MSR 2000
  36 #define P9_STOP_SPR_PSSCR      855
  37
  38 static u32 supported_cpuidle_states;
  39
  40 /*
  41  * The default stop state that will be used by ppc_md.power_save
  42  * function on platforms that support stop instruction.
  43  */
  44 static u64 pnv_default_stop_val;
  45 static u64 pnv_default_stop_mask;
  46 static bool default_stop_found;
  47
  48 /*
  49  * First deep stop state. Used to figure out when to save/restore
  50  * hypervisor context.
  51  */
  52 u64 pnv_first_deep_stop_state = MAX_STOP_STATE;
  53
  54 /*
  55  * psscr value and mask of the deepest stop idle state.
  56  * Used when a cpu is offlined.
  57  */
  58 static u64 pnv_deepest_stop_psscr_val;
  59 static u64 pnv_deepest_stop_psscr_mask;
  60 static u64 pnv_deepest_stop_flag;
  61 static bool deepest_stop_found;
  62
  63 static int pnv_save_sprs_for_deep_states(void)
  64 {
  65         int cpu;
  66         int rc;
  67
  68         /*
  69          * hid0, hid1, hid4, hid5, hmeer and lpcr values are symmetric across
  70          * all cpus at boot. Get these reg values of current cpu and use the
  71          * same across all cpus.
  72          */
  73         uint64_t lpcr_val = mfspr(SPRN_LPCR);
  74         uint64_t hid0_val = mfspr(SPRN_HID0);
  75         uint64_t hid1_val = mfspr(SPRN_HID1);
  76         uint64_t hid4_val = mfspr(SPRN_HID4);
  77         uint64_t hid5_val = mfspr(SPRN_HID5);
  78         uint64_t hmeer_val = mfspr(SPRN_HMEER);
  79         uint64_t msr_val = MSR_IDLE;
  80         uint64_t psscr_val = pnv_deepest_stop_psscr_val;
  81
  82         for_each_possible_cpu(cpu) {
  83                 uint64_t pir = get_hard_smp_processor_id(cpu);
  84                 uint64_t hsprg0_val = (uint64_t)paca_ptrs[cpu];
  85
  86                 rc = opal_slw_set_reg(pir, SPRN_HSPRG0, hsprg0_val);
  87                 if (rc != 0)
  88                         return rc;
  89
  90                 rc = opal_slw_set_reg(pir, SPRN_LPCR, lpcr_val);
  91                 if (rc != 0)
  92                         return rc;
  93
  94                 if (cpu_has_feature(CPU_FTR_ARCH_300)) {
  95                         rc = opal_slw_set_reg(pir, P9_STOP_SPR_MSR, msr_val);
  96                         if (rc)
  97                                 return rc;
  98
  99                         rc = opal_slw_set_reg(pir,
 100                                               P9_STOP_SPR_PSSCR, psscr_val);
 101
 102                         if (rc)
 103                                 return rc;
 104                 }
 105
 106                 /* HIDs are per core registers */
 107                 if (cpu_thread_in_core(cpu) == 0) {
 108
 109                         rc = opal_slw_set_reg(pir, SPRN_HMEER, hmeer_val);
 110                         if (rc != 0)
 111                                 return rc;
 112
 113                         rc = opal_slw_set_reg(pir, SPRN_HID0, hid0_val);
 114                         if (rc != 0)
 115                                 return rc;
 116
 117                         /* Only p8 needs to set extra HID regiters */
 118                         if (!cpu_has_feature(CPU_FTR_ARCH_300)) {
 119
 120                                 rc = opal_slw_set_reg(pir, SPRN_HID1, hid1_val);
 121                                 if (rc != 0)
 122                                         return rc;
 123
 124                                 rc = opal_slw_set_reg(pir, SPRN_HID4, hid4_val);
 125                                 if (rc != 0)
 126                                         return rc;
 127
 128                                 rc = opal_slw_set_reg(pir, SPRN_HID5, hid5_val);
 129                                 if (rc != 0)
 130                                         return rc;
 131                         }
 132                 }
 133         }
 134
 135         return 0;
 136 }
 137
 138 static void pnv_alloc_idle_core_states(void)
 139 {
 140         int i, j;
 141         int nr_cores = cpu_nr_cores();
 142         u32 *core_idle_state;
 143
 144         /*
 145          * core_idle_state - The lower 8 bits track the idle state of
 146          * each thread of the core.
 147          *
 148          * The most significant bit is the lock bit.
 149          *
 150          * Initially all the bits corresponding to threads_per_core
 151          * are set. They are cleared when the thread enters deep idle
 152          * state like sleep and winkle/stop.
 153          *
 154          * Initially the lock bit is cleared.  The lock bit has 2
 155          * purposes:
 156          *      a. While the first thread in the core waking up from
 157          *         idle is restoring core state, it prevents other
 158          *         threads in the core from switching to process
 159          *         context.
 160          *      b. While the last thread in the core is saving the
 161          *         core state, it prevents a different thread from
 162          *         waking up.
 163          */
 164         for (i = 0; i < nr_cores; i++) {
 165                 int first_cpu = i * threads_per_core;
 166                 int node = cpu_to_node(first_cpu);
 167                 size_t paca_ptr_array_size;
 168
 169                 core_idle_state = kmalloc_node(sizeof(u32), GFP_KERNEL, node);
 170                 *core_idle_state = (1 << threads_per_core) - 1;
 171                 paca_ptr_array_size = (threads_per_core *
 172                                        sizeof(struct paca_struct *));
 173
 174                 for (j = 0; j < threads_per_core; j++) {
 175                         int cpu = first_cpu + j;
 176
 177                         paca_ptrs[cpu]->core_idle_state_ptr = core_idle_state;
 178                         paca_ptrs[cpu]->thread_idle_state = PNV_THREAD_RUNNING;
 179                         paca_ptrs[cpu]->thread_mask = 1 << j;
 180                         if (!cpu_has_feature(CPU_FTR_POWER9_DD1))
 181                                 continue;
 182                         paca_ptrs[cpu]->thread_sibling_pacas =
 183                                 kmalloc_node(paca_ptr_array_size,
 184                                              GFP_KERNEL, node);
 185                 }
 186         }
 187
 188         update_subcore_sibling_mask();
 189
 190         if (supported_cpuidle_states & OPAL_PM_LOSE_FULL_CONTEXT) {
 191                 int rc = pnv_save_sprs_for_deep_states();
 192
 193                 if (likely(!rc))
 194                         return;
 195
 196                 /*
 197                  * The stop-api is unable to restore hypervisor
 198                  * resources on wakeup from platform idle states which
 199                  * lose full context. So disable such states.
 200                  */
 201                 supported_cpuidle_states &= ~OPAL_PM_LOSE_FULL_CONTEXT;
 202                 pr_warn("cpuidle-powernv: Disabling idle states that lose full context\n");
 203                 pr_warn("cpuidle-powernv: Idle power-savings, CPU-Hotplug affected\n");
 204
 205                 if (cpu_has_feature(CPU_FTR_ARCH_300) &&
 206                     (pnv_deepest_stop_flag & OPAL_PM_LOSE_FULL_CONTEXT)) {
 207                         /*
 208                          * Use the default stop state for CPU-Hotplug
 209                          * if available.
 210                          */
 211                         if (default_stop_found) {
 212                                 pnv_deepest_stop_psscr_val =
 213                                         pnv_default_stop_val;
 214                                 pnv_deepest_stop_psscr_mask =
 215                                         pnv_default_stop_mask;
 216                                 pr_warn("cpuidle-powernv: Offlined CPUs will stop with psscr = 0x%016llx\n",
 217                                         pnv_deepest_stop_psscr_val);
 218                         } else { /* Fallback to snooze loop for CPU-Hotplug */
 219                                 deepest_stop_found = false;
 220                                 pr_warn("cpuidle-powernv: Offlined CPUs will busy wait\n");
 221                         }
 222                 }
 223         }
 224 }
 225
 226 u32 pnv_get_supported_cpuidle_states(void)
 227 {
 228         return supported_cpuidle_states;
 229 }
 230 EXPORT_SYMBOL_GPL(pnv_get_supported_cpuidle_states);
 231
 232 static void pnv_fastsleep_workaround_apply(void *info)
 233
 234 {
 235         int rc;
 236         int *err = info;
 237
 238         rc = opal_config_cpu_idle_state(OPAL_CONFIG_IDLE_FASTSLEEP,
 239                                         OPAL_CONFIG_IDLE_APPLY);
 240         if (rc)
 241                 *err = 1;
 242 }
 243
 244 /*
 245  * Used to store fastsleep workaround state
 246  * 0 - Workaround applied/undone at fastsleep entry/exit path (Default)
 247  * 1 - Workaround applied once, never undone.
 248  */
 249 static u8 fastsleep_workaround_applyonce;
 250
 251 static ssize_t show_fastsleep_workaround_applyonce(struct device *dev,
 252                 struct device_attribute *attr, char *buf)
 253 {
 254         return sprintf(buf, "%u\n", fastsleep_workaround_applyonce);
 255 }
 256
 257 static ssize_t store_fastsleep_workaround_applyonce(struct device *dev,
 258                 struct device_attribute *attr, const char *buf,
 259                 size_t count)
 260 {
 261         cpumask_t primary_thread_mask;
 262         int err;
 263         u8 val;
 264
 265         if (kstrtou8(buf, 0, &val) || val != 1)
 266                 return -EINVAL;
 267
 268         if (fastsleep_workaround_applyonce == 1)
 269                 return count;
 270
 271         /*
 272          * fastsleep_workaround_applyonce = 1 implies
 273          * fastsleep workaround needs to be left in 'applied' state on all
 274          * the cores. Do this by-
 275          * 1. Patching out the call to 'undo' workaround in fastsleep exit path
 276          * 2. Sending ipi to all the cores which have at least one online thread
 277          * 3. Patching out the call to 'apply' workaround in fastsleep entry
 278          * path
 279          * There is no need to send ipi to cores which have all threads
 280          * offlined, as last thread of the core entering fastsleep or deeper
 281          * state would have applied workaround.
 282          */
 283         err = patch_instruction(
 284                 (unsigned int *)pnv_fastsleep_workaround_at_exit,
 285                 PPC_INST_NOP);
 286         if (err) {
 287                 pr_err("fastsleep_workaround_applyonce change failed while patching pnv_fastsleep_workaround_at_exit");
 288                 goto fail;
 289         }
 290
 291         get_online_cpus();
 292         primary_thread_mask = cpu_online_cores_map();
 293         on_each_cpu_mask(&primary_thread_mask,
 294                                 pnv_fastsleep_workaround_apply,
 295                                 &err, 1);
 296         put_online_cpus();
 297         if (err) {
 298                 pr_err("fastsleep_workaround_applyonce change failed while running pnv_fastsleep_workaround_apply");
 299                 goto fail;
 300         }
 301
 302         err = patch_instruction(
 303                 (unsigned int *)pnv_fastsleep_workaround_at_entry,
 304                 PPC_INST_NOP);
 305         if (err) {
 306                 pr_err("fastsleep_workaround_applyonce change failed while patching pnv_fastsleep_workaround_at_entry");
 307                 goto fail;
 308         }
 309
 310         fastsleep_workaround_applyonce = 1;
 311
 312         return count;
 313 fail:
 314         return -EIO;
 315 }
 316
 317 static DEVICE_ATTR(fastsleep_workaround_applyonce, 0600,
 318                         show_fastsleep_workaround_applyonce,
 319                         store_fastsleep_workaround_applyonce);
 320
 321 static unsigned long __power7_idle_type(unsigned long type)
 322 {
 323         unsigned long srr1;
 324
 325         if (!prep_irq_for_idle_irqsoff())
 326                 return 0;
 327
 328         __ppc64_runlatch_off();
 329         srr1 = power7_idle_insn(type);
 330         __ppc64_runlatch_on();
 331
 332         fini_irq_for_idle_irqsoff();
 333
 334         return srr1;
 335 }
 336
 337 void power7_idle_type(unsigned long type)
 338 {
 339         unsigned long srr1;
 340
 341         srr1 = __power7_idle_type(type);
 342         irq_set_pending_from_srr1(srr1);
 343 }
 344
 345 void power7_idle(void)
 346 {
 347         if (!powersave_nap)
 348                 return;
 349
 350         power7_idle_type(PNV_THREAD_NAP);
 351 }
 352
 353 static unsigned long __power9_idle_type(unsigned long stop_psscr_val,
 354                                       unsigned long stop_psscr_mask)
 355 {
 356         unsigned long psscr;
 357         unsigned long srr1;
 358
 359         if (!prep_irq_for_idle_irqsoff())
 360                 return 0;
 361
 362         psscr = mfspr(SPRN_PSSCR);
 363         psscr = (psscr & ~stop_psscr_mask) | stop_psscr_val;
 364
 365         __ppc64_runlatch_off();
 366         srr1 = power9_idle_stop(psscr);
 367         __ppc64_runlatch_on();
 368
 369         fini_irq_for_idle_irqsoff();
 370
 371         return srr1;
 372 }
 373
 374 void power9_idle_type(unsigned long stop_psscr_val,
 375                                       unsigned long stop_psscr_mask)
 376 {
 377         unsigned long srr1;
 378
 379         srr1 = __power9_idle_type(stop_psscr_val, stop_psscr_mask);
 380         irq_set_pending_from_srr1(srr1);
 381 }
 382
 383 /*
 384  * Used for ppc_md.power_save which needs a function with no parameters
 385  */
 386 void power9_idle(void)
 387 {
 388         power9_idle_type(pnv_default_stop_val, pnv_default_stop_mask);
 389 }
 390
 391 #ifdef CONFIG_KVM_BOOK3S_HV_POSSIBLE
 392 /*
 393  * This is used in working around bugs in thread reconfiguration
 394  * on POWER9 (at least up to Nimbus DD2.2) relating to transactional
 395  * memory and the way that XER[SO] is checkpointed.
 396  * This function forces the core into SMT4 in order by asking
 397  * all other threads not to stop, and sending a message to any
 398  * that are in a stop state.
 399  * Must be called with preemption disabled.
 400  */
 401 void pnv_power9_force_smt4_catch(void)
 402 {
 403         int cpu, cpu0, thr;
 404         int awake_threads = 1;          /* this thread is awake */
 405         int poke_threads = 0;
 406         int need_awake = threads_per_core;
 407
 408         cpu = smp_processor_id();
 409         cpu0 = cpu & ~(threads_per_core - 1);
 410         for (thr = 0; thr < threads_per_core; ++thr) {
 411                 if (cpu != cpu0 + thr)
 412                         atomic_inc(&paca_ptrs[cpu0+thr]->dont_stop);
 413         }
 414         /* order setting dont_stop vs testing requested_psscr */
 415         mb();
 416         for (thr = 0; thr < threads_per_core; ++thr) {
 417                 if (!paca_ptrs[cpu0+thr]->requested_psscr)
 418                         ++awake_threads;
 419                 else
 420                         poke_threads |= (1 << thr);
 421         }
 422
 423         /* If at least 3 threads are awake, the core is in SMT4 already */
 424         if (awake_threads < need_awake) {
 425                 /* We have to wake some threads; we'll use msgsnd */
 426                 for (thr = 0; thr < threads_per_core; ++thr) {
 427                         if (poke_threads & (1 << thr)) {
 428                                 ppc_msgsnd_sync();
 429                                 ppc_msgsnd(PPC_DBELL_MSGTYPE, 0,
 430                                            paca_ptrs[cpu0+thr]->hw_cpu_id);
 431                         }
 432                 }
 433                 /* now spin until at least 3 threads are awake */
 434                 do {
 435                         for (thr = 0; thr < threads_per_core; ++thr) {
 436                                 if ((poke_threads & (1 << thr)) &&
 437                                     !paca_ptrs[cpu0+thr]->requested_psscr) {
 438                                         ++awake_threads;
 439                                         poke_threads &= ~(1 << thr);
 440                                 }
 441                         }
 442                 } while (awake_threads < need_awake);
 443         }
 444 }
 445 EXPORT_SYMBOL_GPL(pnv_power9_force_smt4_catch);
 446
 447 void pnv_power9_force_smt4_release(void)
 448 {
 449         int cpu, cpu0, thr;
 450
 451         cpu = smp_processor_id();
 452         cpu0 = cpu & ~(threads_per_core - 1);
 453
 454         /* clear all the dont_stop flags */
 455         for (thr = 0; thr < threads_per_core; ++thr) {
 456                 if (cpu != cpu0 + thr)
 457                         atomic_dec(&paca_ptrs[cpu0+thr]->dont_stop);
 458         }
 459 }
 460 EXPORT_SYMBOL_GPL(pnv_power9_force_smt4_release);
 461 #endif /* CONFIG_KVM_BOOK3S_HV_POSSIBLE */
 462
 463 #ifdef CONFIG_HOTPLUG_CPU
 464 static void pnv_program_cpu_hotplug_lpcr(unsigned int cpu, u64 lpcr_val)
 465 {
 466         u64 pir = get_hard_smp_processor_id(cpu);
 467
 468         mtspr(SPRN_LPCR, lpcr_val);
 469
 470         /*
 471          * Program the LPCR via stop-api only if the deepest stop state
 472          * can lose hypervisor context.
 473          */
 474         if (supported_cpuidle_states & OPAL_PM_LOSE_FULL_CONTEXT)
 475                 opal_slw_set_reg(pir, SPRN_LPCR, lpcr_val);
 476 }
 477
 478 /*
 479  * pnv_cpu_offline: A function that puts the CPU into the deepest
 480  * available platform idle state on a CPU-Offline.
 481  * interrupts hard disabled and no lazy irq pending.
 482  */
 483 unsigned long pnv_cpu_offline(unsigned int cpu)
 484 {
 485         unsigned long srr1;
 486         u32 idle_states = pnv_get_supported_cpuidle_states();
 487         u64 lpcr_val;
 488
 489         /*
 490          * We don't want to take decrementer interrupts while we are
 491          * offline, so clear LPCR:PECE1. We keep PECE2 (and
 492          * LPCR_PECE_HVEE on P9) enabled as to let IPIs in.
 493          *
 494          * If the CPU gets woken up by a special wakeup, ensure that
 495          * the SLW engine sets LPCR with decrementer bit cleared, else
 496          * the CPU will come back to the kernel due to a spurious
 497          * wakeup.
 498          */
 499         lpcr_val = mfspr(SPRN_LPCR) & ~(u64)LPCR_PECE1;
 500         pnv_program_cpu_hotplug_lpcr(cpu, lpcr_val);
 501
 502         __ppc64_runlatch_off();
 503
 504         if (cpu_has_feature(CPU_FTR_ARCH_300) && deepest_stop_found) {
 505                 unsigned long psscr;
 506
 507                 psscr = mfspr(SPRN_PSSCR);
 508                 psscr = (psscr & ~pnv_deepest_stop_psscr_mask) |
 509                                                 pnv_deepest_stop_psscr_val;
 510                 srr1 = power9_offline_stop(psscr);
 511
 512         } else if ((idle_states & OPAL_PM_WINKLE_ENABLED) &&
 513                    (idle_states & OPAL_PM_LOSE_FULL_CONTEXT)) {
 514                 srr1 = power7_idle_insn(PNV_THREAD_WINKLE);
 515         } else if ((idle_states & OPAL_PM_SLEEP_ENABLED) ||
 516                    (idle_states & OPAL_PM_SLEEP_ENABLED_ER1)) {
 517                 srr1 = power7_idle_insn(PNV_THREAD_SLEEP);
 518         } else if (idle_states & OPAL_PM_NAP_ENABLED) {
 519                 srr1 = power7_idle_insn(PNV_THREAD_NAP);
 520         } else {
 521                 /* This is the fallback method. We emulate snooze */
 522                 while (!generic_check_cpu_restart(cpu)) {
 523                         HMT_low();
 524                         HMT_very_low();
 525                 }
 526                 srr1 = 0;
 527                 HMT_medium();
 528         }
 529
 530         __ppc64_runlatch_on();
 531
 532         /*
 533          * Re-enable decrementer interrupts in LPCR.
 534          *
 535          * Further, we want stop states to be woken up by decrementer
 536          * for non-hotplug cases. So program the LPCR via stop api as
 537          * well.
 538          */
 539         lpcr_val = mfspr(SPRN_LPCR) | (u64)LPCR_PECE1;
 540         pnv_program_cpu_hotplug_lpcr(cpu, lpcr_val);
 541
 542         return srr1;
 543 }
 544 #endif
 545
 546 /*
 547  * Power ISA 3.0 idle initialization.
 548  *
 549  * POWER ISA 3.0 defines a new SPR Processor stop Status and Control
 550  * Register (PSSCR) to control idle behavior.
 551  *
 552  * PSSCR layout:
 553  * ----------------------------------------------------------
 554  * | PLS | /// | SD | ESL | EC | PSLL | /// | TR | MTL | RL |
 555  * ----------------------------------------------------------
 556  * 0      4     41   42    43   44     48    54   56    60
 557  *
 558  * PSSCR key fields:
 559  *      Bits 0:3  - Power-Saving Level Status (PLS). This field indicates the
 560  *      lowest power-saving state the thread entered since stop instruction was
 561  *      last executed.
 562  *
 563  *      Bit 41 - Status Disable(SD)
 564  *      0 - Shows PLS entries
 565  *      1 - PLS entries are all 0
 566  *
 567  *      Bit 42 - Enable State Loss
 568  *      0 - No state is lost irrespective of other fields
 569  *      1 - Allows state loss
 570  *
 571  *      Bit 43 - Exit Criterion
 572  *      0 - Exit from power-save mode on any interrupt
 573  *      1 - Exit from power-save mode controlled by LPCR's PECE bits
 574  *
 575  *      Bits 44:47 - Power-Saving Level Limit
 576  *      This limits the power-saving level that can be entered into.
 577  *
 578  *      Bits 60:63 - Requested Level
 579  *      Used to specify which power-saving level must be entered on executing
 580  *      stop instruction
 581  */
 582
 583 int validate_psscr_val_mask(u64 *psscr_val, u64 *psscr_mask, u32 flags)
 584 {
 585         int err = 0;
 586
 587         /*
 588          * psscr_mask == 0xf indicates an older firmware.
 589          * Set remaining fields of psscr to the default values.
 590          * See NOTE above definition of PSSCR_HV_DEFAULT_VAL
 591          */
 592         if (*psscr_mask == 0xf) {
 593                 *psscr_val = *psscr_val | PSSCR_HV_DEFAULT_VAL;
 594                 *psscr_mask = PSSCR_HV_DEFAULT_MASK;
 595                 return err;
 596         }
 597
 598         /*
 599          * New firmware is expected to set the psscr_val bits correctly.
 600          * Validate that the following invariants are correctly maintained by
 601          * the new firmware.
 602          * - ESL bit value matches the EC bit value.
 603          * - ESL bit is set for all the deep stop states.
 604          */
 605         if (GET_PSSCR_ESL(*psscr_val) != GET_PSSCR_EC(*psscr_val)) {
 606                 err = ERR_EC_ESL_MISMATCH;
 607         } else if ((flags & OPAL_PM_LOSE_FULL_CONTEXT) &&
 608                 GET_PSSCR_ESL(*psscr_val) == 0) {
 609                 err = ERR_DEEP_STATE_ESL_MISMATCH;
 610         }
 611
 612         return err;
 613 }
 614
 615 /*
 616  * pnv_arch300_idle_init: Initializes the default idle state, first
 617  *                        deep idle state and deepest idle state on
 618  *                        ISA 3.0 CPUs.
 619  *
 620  * @np: /ibm,opal/power-mgt device node
 621  * @flags: cpu-idle-state-flags array
 622  * @dt_idle_states: Number of idle state entries
 623  * Returns 0 on success
 624  */
 625 static int __init pnv_power9_idle_init(struct device_node *np, u32 *flags,
 626                                         int dt_idle_states)
 627 {
 628         u64 *psscr_val = NULL;
 629         u64 *psscr_mask = NULL;
 630         u32 *residency_ns = NULL;
 631         u64 max_residency_ns = 0;
 632         int rc = 0, i;
 633
 634         psscr_val = kcalloc(dt_idle_states, sizeof(*psscr_val), GFP_KERNEL);
 635         psscr_mask = kcalloc(dt_idle_states, sizeof(*psscr_mask), GFP_KERNEL);
 636         residency_ns = kcalloc(dt_idle_states, sizeof(*residency_ns),
 637                                GFP_KERNEL);
 638
 639         if (!psscr_val || !psscr_mask || !residency_ns) {
 640                 rc = -1;
 641                 goto out;
 642         }
 643
 644         if (of_property_read_u64_array(np,
 645                 "ibm,cpu-idle-state-psscr",
 646                 psscr_val, dt_idle_states)) {
 647                 pr_warn("cpuidle-powernv: missing ibm,cpu-idle-state-psscr in DT\n");
 648                 rc = -1;
 649                 goto out;
 650         }
 651
 652         if (of_property_read_u64_array(np,
 653                                        "ibm,cpu-idle-state-psscr-mask",
 654                                        psscr_mask, dt_idle_states)) {
 655                 pr_warn("cpuidle-powernv: missing ibm,cpu-idle-state-psscr-mask in DT\n");
 656                 rc = -1;
 657                 goto out;
 658         }
 659
 660         if (of_property_read_u32_array(np,
 661                                        "ibm,cpu-idle-state-residency-ns",
 662                                         residency_ns, dt_idle_states)) {
 663                 pr_warn("cpuidle-powernv: missing ibm,cpu-idle-state-residency-ns in DT\n");
 664                 rc = -1;
 665                 goto out;
 666         }
 667
 668         /*
 669          * Set pnv_first_deep_stop_state, pnv_deepest_stop_psscr_{val,mask},
 670          * and the pnv_default_stop_{val,mask}.
 671          *
 672          * pnv_first_deep_stop_state should be set to the first stop
 673          * level to cause hypervisor state loss.
 674          *
 675          * pnv_deepest_stop_{val,mask} should be set to values corresponding to
 676          * the deepest stop state.
 677          *
 678          * pnv_default_stop_{val,mask} should be set to values corresponding to
 679          * the shallowest (OPAL_PM_STOP_INST_FAST) loss-less stop state.
 680          */
 681         pnv_first_deep_stop_state = MAX_STOP_STATE;
 682         for (i = 0; i < dt_idle_states; i++) {
 683                 int err;
 684                 u64 psscr_rl = psscr_val[i] & PSSCR_RL_MASK;
 685
 686                 if ((flags[i] & OPAL_PM_LOSE_FULL_CONTEXT) &&
 687                      (pnv_first_deep_stop_state > psscr_rl))
 688                         pnv_first_deep_stop_state = psscr_rl;
 689
 690                 err = validate_psscr_val_mask(&psscr_val[i], &psscr_mask[i],
 691                                               flags[i]);
 692                 if (err) {
 693                         report_invalid_psscr_val(psscr_val[i], err);
 694                         continue;
 695                 }
 696
 697                 if (max_residency_ns < residency_ns[i]) {
 698                         max_residency_ns = residency_ns[i];
 699                         pnv_deepest_stop_psscr_val = psscr_val[i];
 700                         pnv_deepest_stop_psscr_mask = psscr_mask[i];
 701                         pnv_deepest_stop_flag = flags[i];
 702                         deepest_stop_found = true;
 703                 }
 704
 705                 if (!default_stop_found &&
 706                     (flags[i] & OPAL_PM_STOP_INST_FAST)) {
 707                         pnv_default_stop_val = psscr_val[i];
 708                         pnv_default_stop_mask = psscr_mask[i];
 709                         default_stop_found = true;
 710                 }
 711         }
 712
 713         if (unlikely(!default_stop_found)) {
 714                 pr_warn("cpuidle-powernv: No suitable default stop state found. Disabling platform idle.\n");
 715         } else {
 716                 ppc_md.power_save = power9_idle;
 717                 pr_info("cpuidle-powernv: Default stop: psscr = 0x%016llx,mask=0x%016llx\n",
 718                         pnv_default_stop_val, pnv_default_stop_mask);
 719         }
 720
 721         if (unlikely(!deepest_stop_found)) {
 722                 pr_warn("cpuidle-powernv: No suitable stop state for CPU-Hotplug. Offlined CPUs will busy wait");
 723         } else {
 724                 pr_info("cpuidle-powernv: Deepest stop: psscr = 0x%016llx,mask=0x%016llx\n",
 725                         pnv_deepest_stop_psscr_val,
 726                         pnv_deepest_stop_psscr_mask);
 727         }
 728
 729         pr_info("cpuidle-powernv: Requested Level (RL) value of first deep stop = 0x%llx\n",
 730                 pnv_first_deep_stop_state);
 731 out:
 732         kfree(psscr_val);
 733         kfree(psscr_mask);
 734         kfree(residency_ns);
 735         return rc;
 736 }
 737
 738 /*
 739  * Probe device tree for supported idle states
 740  */
 741 static void __init pnv_probe_idle_states(void)
 742 {
 743         struct device_node *np;
 744         int dt_idle_states;
 745         u32 *flags = NULL;
 746         int i;
 747
 748         np = of_find_node_by_path("/ibm,opal/power-mgt");
 749         if (!np) {
 750                 pr_warn("opal: PowerMgmt Node not found\n");
 751                 goto out;
 752         }
 753         dt_idle_states = of_property_count_u32_elems(np,
 754                         "ibm,cpu-idle-state-flags");
 755         if (dt_idle_states < 0) {
 756                 pr_warn("cpuidle-powernv: no idle states found in the DT\n");
 757                 goto out;
 758         }
 759
 760         flags = kcalloc(dt_idle_states, sizeof(*flags),  GFP_KERNEL);
 761
 762         if (of_property_read_u32_array(np,
 763                         "ibm,cpu-idle-state-flags", flags, dt_idle_states)) {
 764                 pr_warn("cpuidle-powernv: missing ibm,cpu-idle-state-flags in DT\n");
 765                 goto out;
 766         }
 767
 768         if (cpu_has_feature(CPU_FTR_ARCH_300)) {
 769                 if (pnv_power9_idle_init(np, flags, dt_idle_states))
 770                         goto out;
 771         }
 772
 773         for (i = 0; i < dt_idle_states; i++)
 774                 supported_cpuidle_states |= flags[i];
 775
 776 out:
 777         kfree(flags);
 778 }
 779 static int __init pnv_init_idle_states(void)
 780 {
 781
 782         supported_cpuidle_states = 0;
 783
 784         if (cpuidle_disable != IDLE_NO_OVERRIDE)
 785                 goto out;
 786
 787         pnv_probe_idle_states();
 788
 789         if (!(supported_cpuidle_states & OPAL_PM_SLEEP_ENABLED_ER1)) {
 790                 patch_instruction(
 791                         (unsigned int *)pnv_fastsleep_workaround_at_entry,
 792                         PPC_INST_NOP);
 793                 patch_instruction(
 794                         (unsigned int *)pnv_fastsleep_workaround_at_exit,
 795                         PPC_INST_NOP);
 796         } else {
 797                 /*
 798                  * OPAL_PM_SLEEP_ENABLED_ER1 is set. It indicates that
 799                  * workaround is needed to use fastsleep. Provide sysfs
 800                  * control to choose how this workaround has to be applied.
 801                  */
 802                 device_create_file(cpu_subsys.dev_root,
 803                                 &dev_attr_fastsleep_workaround_applyonce);
 804         }
 805
 806         pnv_alloc_idle_core_states();
 807
 808         /*
 809          * For each CPU, record its PACA address in each of it's
 810          * sibling thread's PACA at the slot corresponding to this
 811          * CPU's index in the core.
 812          */
 813         if (cpu_has_feature(CPU_FTR_POWER9_DD1)) {
 814                 int cpu;
 815
 816                 pr_info("powernv: idle: Saving PACA pointers of all CPUs in their thread sibling PACA\n");
 817                 for_each_possible_cpu(cpu) {
 818                         int base_cpu = cpu_first_thread_sibling(cpu);
 819                         int idx = cpu_thread_in_core(cpu);
 820                         int i;
 821
 822                         for (i = 0; i < threads_per_core; i++) {
 823                                 int j = base_cpu + i;
 824
 825                                 paca_ptrs[j]->thread_sibling_pacas[idx] =
 826                                         paca_ptrs[cpu];
 827                         }
 828                 }
 829         }
 830
 831         if (supported_cpuidle_states & OPAL_PM_NAP_ENABLED)
 832                 ppc_md.power_save = power7_idle;
 833
 834 out:
 835         return 0;
 836 }
 837 machine_subsys_initcall(powernv, pnv_init_idle_states);