arch/ppc64/kernel/idle.c

   1 /*
   2  * Idle daemon for PowerPC.  Idle daemon will handle any action
   3  * that needs to be taken when the system becomes idle.
   4  *
   5  * Originally Written by Cort Dougan (cort@cs.nmt.edu)
   6  *
   7  * iSeries supported added by Mike Corrigan <mikejc@us.ibm.com>
   8  *
   9  * Additional shared processor, SMT, and firmware support
  10  *    Copyright (c) 2003 Dave Engebretsen <engebret@us.ibm.com>
  11  *
  12  * This program is free software; you can redistribute it and/or
  13  * modify it under the terms of the GNU General Public License
  14  * as published by the Free Software Foundation; either version
  15  * 2 of the License, or (at your option) any later version.
  16  */
  17
  18 #include <linux/config.h>
  19 #include <linux/sched.h>
  20 #include <linux/kernel.h>
  21 #include <linux/smp.h>
  22 #include <linux/cpu.h>
  23 #include <linux/module.h>
  24 #include <linux/sysctl.h>
  25 #include <linux/smp.h>
  26
  27 #include <asm/system.h>
  28 #include <asm/processor.h>
  29 #include <asm/mmu.h>
  30 #include <asm/cputable.h>
  31 #include <asm/time.h>
  32 #include <asm/iSeries/HvCall.h>
  33 #include <asm/iSeries/ItLpQueue.h>
  34 #include <asm/plpar_wrappers.h>
  35 #include <asm/systemcfg.h>
  36
  37 extern void power4_idle(void);
  38
  39 static int (*idle_loop)(void);
  40
  41 #ifdef CONFIG_PPC_ISERIES
  42 static unsigned long maxYieldTime = 0;
  43 static unsigned long minYieldTime = 0xffffffffffffffffUL;
  44
  45 static inline void process_iSeries_events(void)
  46 {
  47         asm volatile ("li 0,0x5555; sc" : : : "r0", "r3");
  48 }
  49
  50 static void yield_shared_processor(void)
  51 {
  52         unsigned long tb;
  53         unsigned long yieldTime;
  54
  55         HvCall_setEnabledInterrupts(HvCall_MaskIPI |
  56                                     HvCall_MaskLpEvent |
  57                                     HvCall_MaskLpProd |
  58                                     HvCall_MaskTimeout);
  59
  60         tb = get_tb();
  61         /* Compute future tb value when yield should expire */
  62         HvCall_yieldProcessor(HvCall_YieldTimed, tb+tb_ticks_per_jiffy);
  63
  64         yieldTime = get_tb() - tb;
  65         if (yieldTime > maxYieldTime)
  66                 maxYieldTime = yieldTime;
  67
  68         if (yieldTime < minYieldTime)
  69                 minYieldTime = yieldTime;
  70
  71         /*
  72          * The decrementer stops during the yield.  Force a fake decrementer
  73          * here and let the timer_interrupt code sort out the actual time.
  74          */
  75         get_paca()->lppaca.int_dword.fields.decr_int = 1;
  76         process_iSeries_events();
  77 }
  78
  79 static int iSeries_idle(void)
  80 {
  81         struct paca_struct *lpaca;
  82         long oldval;
  83
  84         /* ensure iSeries run light will be out when idle */
  85         ppc64_runlatch_off();
  86
  87         lpaca = get_paca();
  88
  89         while (1) {
  90                 if (lpaca->lppaca.shared_proc) {
  91                         if (hvlpevent_is_pending())
  92                                 process_iSeries_events();
  93                         if (!need_resched())
  94                                 yield_shared_processor();
  95                 } else {
  96                         oldval = test_and_clear_thread_flag(TIF_NEED_RESCHED);
  97
  98                         if (!oldval) {
  99                                 set_thread_flag(TIF_POLLING_NRFLAG);
 100
 101                                 while (!need_resched()) {
 102                                         HMT_medium();
 103                                         if (hvlpevent_is_pending())
 104                                                 process_iSeries_events();
 105                                         HMT_low();
 106                                 }
 107
 108                                 HMT_medium();
 109                                 clear_thread_flag(TIF_POLLING_NRFLAG);
 110                         } else {
 111                                 set_need_resched();
 112                         }
 113                 }
 114
 115                 ppc64_runlatch_on();
 116                 schedule();
 117                 ppc64_runlatch_off();
 118         }
 119
 120         return 0;
 121 }
 122
 123 #else
 124
 125 static int default_idle(void)
 126 {
 127         long oldval;
 128         unsigned int cpu = smp_processor_id();
 129
 130         while (1) {
 131                 oldval = test_and_clear_thread_flag(TIF_NEED_RESCHED);
 132
 133                 if (!oldval) {
 134                         set_thread_flag(TIF_POLLING_NRFLAG);
 135
 136                         while (!need_resched() && !cpu_is_offline(cpu)) {
 137                                 barrier();
 138                                 /*
 139                                  * Go into low thread priority and possibly
 140                                  * low power mode.
 141                                  */
 142                                 HMT_low();
 143                                 HMT_very_low();
 144                         }
 145
 146                         HMT_medium();
 147                         clear_thread_flag(TIF_POLLING_NRFLAG);
 148                 } else {
 149                         set_need_resched();
 150                 }
 151
 152                 schedule();
 153                 if (cpu_is_offline(cpu) && system_state == SYSTEM_RUNNING)
 154                         cpu_die();
 155         }
 156
 157         return 0;
 158 }
 159
 160 #ifdef CONFIG_PPC_PSERIES
 161
 162 DECLARE_PER_CPU(unsigned long, smt_snooze_delay);
 163
 164 int dedicated_idle(void)
 165 {
 166         long oldval;
 167         struct paca_struct *lpaca = get_paca(), *ppaca;
 168         unsigned long start_snooze;
 169         unsigned long *smt_snooze_delay = &__get_cpu_var(smt_snooze_delay);
 170         unsigned int cpu = smp_processor_id();
 171
 172         ppaca = &paca[cpu ^ 1];
 173
 174         while (1) {
 175                 /*
 176                  * Indicate to the HV that we are idle. Now would be
 177                  * a good time to find other work to dispatch.
 178                  */
 179                 lpaca->lppaca.idle = 1;
 180
 181                 oldval = test_and_clear_thread_flag(TIF_NEED_RESCHED);
 182                 if (!oldval) {
 183                         set_thread_flag(TIF_POLLING_NRFLAG);
 184                         start_snooze = __get_tb() +
 185                                 *smt_snooze_delay * tb_ticks_per_usec;
 186                         while (!need_resched() && !cpu_is_offline(cpu)) {
 187                                 /*
 188                                  * Go into low thread priority and possibly
 189                                  * low power mode.
 190                                  */
 191                                 HMT_low();
 192                                 HMT_very_low();
 193
 194                                 if (*smt_snooze_delay == 0 ||
 195                                     __get_tb() < start_snooze)
 196                                         continue;
 197
 198                                 HMT_medium();
 199
 200                                 if (!(ppaca->lppaca.idle)) {
 201                                         local_irq_disable();
 202
 203                                         /*
 204                                          * We are about to sleep the thread
 205                                          * and so wont be polling any
 206                                          * more.
 207                                          */
 208                                         clear_thread_flag(TIF_POLLING_NRFLAG);
 209
 210                                         /*
 211                                          * SMT dynamic mode. Cede will result
 212                                          * in this thread going dormant, if the
 213                                          * partner thread is still doing work.
 214                                          * Thread wakes up if partner goes idle,
 215                                          * an interrupt is presented, or a prod
 216                                          * occurs.  Returning from the cede
 217                                          * enables external interrupts.
 218                                          */
 219                                         if (!need_resched())
 220                                                 cede_processor();
 221                                         else
 222                                                 local_irq_enable();
 223                                 } else {
 224                                         /*
 225                                          * Give the HV an opportunity at the
 226                                          * processor, since we are not doing
 227                                          * any work.
 228                                          */
 229                                         poll_pending();
 230                                 }
 231                         }
 232
 233                         clear_thread_flag(TIF_POLLING_NRFLAG);
 234                 } else {
 235                         set_need_resched();
 236                 }
 237
 238                 HMT_medium();
 239                 lpaca->lppaca.idle = 0;
 240                 schedule();
 241                 if (cpu_is_offline(cpu) && system_state == SYSTEM_RUNNING)
 242                         cpu_die();
 243         }
 244         return 0;
 245 }
 246
 247 static int shared_idle(void)
 248 {
 249         struct paca_struct *lpaca = get_paca();
 250         unsigned int cpu = smp_processor_id();
 251
 252         while (1) {
 253                 /*
 254                  * Indicate to the HV that we are idle. Now would be
 255                  * a good time to find other work to dispatch.
 256                  */
 257                 lpaca->lppaca.idle = 1;
 258
 259                 while (!need_resched() && !cpu_is_offline(cpu)) {
 260                         local_irq_disable();
 261
 262                         /*
 263                          * Yield the processor to the hypervisor.  We return if
 264                          * an external interrupt occurs (which are driven prior
 265                          * to returning here) or if a prod occurs from another
 266                          * processor. When returning here, external interrupts
 267                          * are enabled.
 268                          *
 269                          * Check need_resched() again with interrupts disabled
 270                          * to avoid a race.
 271                          */
 272                         if (!need_resched())
 273                                 cede_processor();
 274                         else
 275                                 local_irq_enable();
 276                 }
 277
 278                 HMT_medium();
 279                 lpaca->lppaca.idle = 0;
 280                 schedule();
 281                 if (cpu_is_offline(smp_processor_id()) &&
 282                     system_state == SYSTEM_RUNNING)
 283                         cpu_die();
 284         }
 285
 286         return 0;
 287 }
 288
 289 #endif /* CONFIG_PPC_PSERIES */
 290
 291 static int native_idle(void)
 292 {
 293         while(1) {
 294                 /* check CPU type here */
 295                 if (!need_resched())
 296                         power4_idle();
 297                 if (need_resched())
 298                         schedule();
 299
 300                 if (cpu_is_offline(raw_smp_processor_id()) &&
 301                     system_state == SYSTEM_RUNNING)
 302                         cpu_die();
 303         }
 304         return 0;
 305 }
 306
 307 #endif /* CONFIG_PPC_ISERIES */
 308
 309 void cpu_idle(void)
 310 {
 311         idle_loop();
 312 }
 313
 314 int powersave_nap;
 315
 316 #ifdef CONFIG_SYSCTL
 317 /*
 318  * Register the sysctl to set/clear powersave_nap.
 319  */
 320 static ctl_table powersave_nap_ctl_table[]={
 321         {
 322                 .ctl_name       = KERN_PPC_POWERSAVE_NAP,
 323                 .procname       = "powersave-nap",
 324                 .data           = &powersave_nap,
 325                 .maxlen         = sizeof(int),
 326                 .mode           = 0644,
 327                 .proc_handler   = &proc_dointvec,
 328         },
 329         { 0, },
 330 };
 331 static ctl_table powersave_nap_sysctl_root[] = {
 332         { 1, "kernel", NULL, 0, 0755, powersave_nap_ctl_table, },
 333         { 0,},
 334 };
 335
 336 static int __init
 337 register_powersave_nap_sysctl(void)
 338 {
 339         register_sysctl_table(powersave_nap_sysctl_root, 0);
 340
 341         return 0;
 342 }
 343 __initcall(register_powersave_nap_sysctl);
 344 #endif
 345
 346 int idle_setup(void)
 347 {
 348         /*
 349          * Move that junk to each platform specific file, eventually define
 350          * a pSeries_idle for shared processor stuff
 351          */
 352 #ifdef CONFIG_PPC_ISERIES
 353         idle_loop = iSeries_idle;
 354         return 1;
 355 #else
 356         idle_loop = default_idle;
 357 #endif
 358 #ifdef CONFIG_PPC_PSERIES
 359         if (systemcfg->platform & PLATFORM_PSERIES) {
 360                 if (cur_cpu_spec->firmware_features & FW_FEATURE_SPLPAR) {
 361                         if (get_paca()->lppaca.shared_proc) {
 362                                 printk(KERN_INFO "Using shared processor idle loop\n");
 363                                 idle_loop = shared_idle;
 364                         } else {
 365                                 printk(KERN_INFO "Using dedicated idle loop\n");
 366                                 idle_loop = dedicated_idle;
 367                         }
 368                 } else {
 369                         printk(KERN_INFO "Using default idle loop\n");
 370                         idle_loop = default_idle;
 371                 }
 372         }
 373 #endif /* CONFIG_PPC_PSERIES */
 374 #ifndef CONFIG_PPC_ISERIES
 375         if (systemcfg->platform == PLATFORM_POWERMAC ||
 376             systemcfg->platform == PLATFORM_MAPLE) {
 377                 printk(KERN_INFO "Using native/NAP idle loop\n");
 378                 idle_loop = native_idle;
 379         }
 380 #endif /* CONFIG_PPC_ISERIES */
 381
 382         return 1;
 383 }