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