arch/powerpc/platforms/pseries/processor_idle.c

   1 /*
   2  *  processor_idle - idle state cpuidle driver.
   3  *  Adapted from drivers/idle/intel_idle.c and
   4  *  drivers/acpi/processor_idle.c
   5  *
   6  */
   7
   8 #include <linux/kernel.h>
   9 #include <linux/module.h>
  10 #include <linux/init.h>
  11 #include <linux/moduleparam.h>
  12 #include <linux/cpuidle.h>
  13 #include <linux/cpu.h>
  14 #include <linux/notifier.h>
  15
  16 #include <asm/paca.h>
  17 #include <asm/reg.h>
  18 #include <asm/machdep.h>
  19 #include <asm/firmware.h>
  20 #include <asm/runlatch.h>
  21
  22 #include "plpar_wrappers.h"
  23 #include "pseries.h"
  24
  25 struct cpuidle_driver pseries_idle_driver = {
  26         .name =         "pseries_idle",
  27         .owner =        THIS_MODULE,
  28 };
  29
  30 #define MAX_IDLE_STATE_COUNT    2
  31
  32 static int max_idle_state = MAX_IDLE_STATE_COUNT - 1;
  33 static struct cpuidle_device __percpu *pseries_cpuidle_devices;
  34 static struct cpuidle_state *cpuidle_state_table;
  35
  36 static inline void idle_loop_prolog(unsigned long *in_purr, ktime_t *kt_before)
  37 {
  38
  39         *kt_before = ktime_get_real();
  40         *in_purr = mfspr(SPRN_PURR);
  41         /*
  42          * Indicate to the HV that we are idle. Now would be
  43          * a good time to find other work to dispatch.
  44          */
  45         get_lppaca()->idle = 1;
  46 }
  47
  48 static inline  s64 idle_loop_epilog(unsigned long in_purr, ktime_t kt_before)
  49 {
  50         get_lppaca()->wait_state_cycles += mfspr(SPRN_PURR) - in_purr;
  51         get_lppaca()->idle = 0;
  52
  53         return ktime_to_us(ktime_sub(ktime_get_real(), kt_before));
  54 }
  55
  56 static int snooze_loop(struct cpuidle_device *dev,
  57                         struct cpuidle_driver *drv,
  58                         int index)
  59 {
  60         unsigned long in_purr;
  61         ktime_t kt_before;
  62         int cpu = dev->cpu;
  63
  64         idle_loop_prolog(&in_purr, &kt_before);
  65         local_irq_enable();
  66         set_thread_flag(TIF_POLLING_NRFLAG);
  67
  68         while ((!need_resched()) && cpu_online(cpu)) {
  69                 ppc64_runlatch_off();
  70                 HMT_low();
  71                 HMT_very_low();
  72         }
  73
  74         HMT_medium();
  75         clear_thread_flag(TIF_POLLING_NRFLAG);
  76         smp_mb();
  77
  78         dev->last_residency =
  79                 (int)idle_loop_epilog(in_purr, kt_before);
  80         return index;
  81 }
  82
  83 static void check_and_cede_processor(void)
  84 {
  85         /*
  86          * Ensure our interrupt state is properly tracked,
  87          * also checks if no interrupt has occurred while we
  88          * were soft-disabled
  89          */
  90         if (prep_irq_for_idle()) {
  91                 cede_processor();
  92 #ifdef CONFIG_TRACE_IRQFLAGS
  93                 /* Ensure that H_CEDE returns with IRQs on */
  94                 if (WARN_ON(!(mfmsr() & MSR_EE)))
  95                         __hard_irq_enable();
  96 #endif
  97         }
  98 }
  99
 100 static int dedicated_cede_loop(struct cpuidle_device *dev,
 101                                 struct cpuidle_driver *drv,
 102                                 int index)
 103 {
 104         unsigned long in_purr;
 105         ktime_t kt_before;
 106
 107         idle_loop_prolog(&in_purr, &kt_before);
 108         get_lppaca()->donate_dedicated_cpu = 1;
 109
 110         ppc64_runlatch_off();
 111         HMT_medium();
 112         check_and_cede_processor();
 113
 114         get_lppaca()->donate_dedicated_cpu = 0;
 115         dev->last_residency =
 116                 (int)idle_loop_epilog(in_purr, kt_before);
 117         return index;
 118 }
 119
 120 static int shared_cede_loop(struct cpuidle_device *dev,
 121                         struct cpuidle_driver *drv,
 122                         int index)
 123 {
 124         unsigned long in_purr;
 125         ktime_t kt_before;
 126
 127         idle_loop_prolog(&in_purr, &kt_before);
 128
 129         /*
 130          * Yield the processor to the hypervisor.  We return if
 131          * an external interrupt occurs (which are driven prior
 132          * to returning here) or if a prod occurs from another
 133          * processor. When returning here, external interrupts
 134          * are enabled.
 135          */
 136         check_and_cede_processor();
 137
 138         dev->last_residency =
 139                 (int)idle_loop_epilog(in_purr, kt_before);
 140         return index;
 141 }
 142
 143 /*
 144  * States for dedicated partition case.
 145  */
 146 static struct cpuidle_state dedicated_states[MAX_IDLE_STATE_COUNT] = {
 147         { /* Snooze */
 148                 .name = "snooze",
 149                 .desc = "snooze",
 150                 .flags = CPUIDLE_FLAG_TIME_VALID,
 151                 .exit_latency = 0,
 152                 .target_residency = 0,
 153                 .enter = &snooze_loop },
 154         { /* CEDE */
 155                 .name = "CEDE",
 156                 .desc = "CEDE",
 157                 .flags = CPUIDLE_FLAG_TIME_VALID,
 158                 .exit_latency = 10,
 159                 .target_residency = 100,
 160                 .enter = &dedicated_cede_loop },
 161 };
 162
 163 /*
 164  * States for shared partition case.
 165  */
 166 static struct cpuidle_state shared_states[MAX_IDLE_STATE_COUNT] = {
 167         { /* Shared Cede */
 168                 .name = "Shared Cede",
 169                 .desc = "Shared Cede",
 170                 .flags = CPUIDLE_FLAG_TIME_VALID,
 171                 .exit_latency = 0,
 172                 .target_residency = 0,
 173                 .enter = &shared_cede_loop },
 174 };
 175
 176 void update_smt_snooze_delay(int cpu, int residency)
 177 {
 178         struct cpuidle_driver *drv = cpuidle_get_driver();
 179         struct cpuidle_device *dev = per_cpu(cpuidle_devices, cpu);
 180
 181         if (cpuidle_state_table != dedicated_states)
 182                 return;
 183
 184         if (residency < 0) {
 185                 /* Disable the Nap state on that cpu */
 186                 if (dev)
 187                         dev->states_usage[1].disable = 1;
 188         } else
 189                 if (drv)
 190                         drv->states[1].target_residency = residency;
 191 }
 192
 193 static int pseries_cpuidle_add_cpu_notifier(struct notifier_block *n,
 194                         unsigned long action, void *hcpu)
 195 {
 196         int hotcpu = (unsigned long)hcpu;
 197         struct cpuidle_device *dev =
 198                         per_cpu_ptr(pseries_cpuidle_devices, hotcpu);
 199
 200         if (dev && cpuidle_get_driver()) {
 201                 switch (action) {
 202                 case CPU_ONLINE:
 203                 case CPU_ONLINE_FROZEN:
 204                         cpuidle_pause_and_lock();
 205                         cpuidle_enable_device(dev);
 206                         cpuidle_resume_and_unlock();
 207                         break;
 208
 209                 case CPU_DEAD:
 210                 case CPU_DEAD_FROZEN:
 211                         cpuidle_pause_and_lock();
 212                         cpuidle_disable_device(dev);
 213                         cpuidle_resume_and_unlock();
 214                         break;
 215
 216                 default:
 217                         return NOTIFY_DONE;
 218                 }
 219         }
 220         return NOTIFY_OK;
 221 }
 222
 223 static struct notifier_block setup_hotplug_notifier = {
 224         .notifier_call = pseries_cpuidle_add_cpu_notifier,
 225 };
 226
 227 /*
 228  * pseries_cpuidle_driver_init()
 229  */
 230 static int pseries_cpuidle_driver_init(void)
 231 {
 232         int idle_state;
 233         struct cpuidle_driver *drv = &pseries_idle_driver;
 234
 235         drv->state_count = 0;
 236
 237         for (idle_state = 0; idle_state < MAX_IDLE_STATE_COUNT; ++idle_state) {
 238
 239                 if (idle_state > max_idle_state)
 240                         break;
 241
 242                 /* is the state not enabled? */
 243                 if (cpuidle_state_table[idle_state].enter == NULL)
 244                         continue;
 245
 246                 drv->states[drv->state_count] = /* structure copy */
 247                         cpuidle_state_table[idle_state];
 248
 249                 drv->state_count += 1;
 250         }
 251
 252         return 0;
 253 }
 254
 255 /* pseries_idle_devices_uninit(void)
 256  * unregister cpuidle devices and de-allocate memory
 257  */
 258 static void pseries_idle_devices_uninit(void)
 259 {
 260         int i;
 261         struct cpuidle_device *dev;
 262
 263         for_each_possible_cpu(i) {
 264                 dev = per_cpu_ptr(pseries_cpuidle_devices, i);
 265                 cpuidle_unregister_device(dev);
 266         }
 267
 268         free_percpu(pseries_cpuidle_devices);
 269         return;
 270 }
 271
 272 /* pseries_idle_devices_init()
 273  * allocate, initialize and register cpuidle device
 274  */
 275 static int pseries_idle_devices_init(void)
 276 {
 277         int i;
 278         struct cpuidle_driver *drv = &pseries_idle_driver;
 279         struct cpuidle_device *dev;
 280
 281         pseries_cpuidle_devices = alloc_percpu(struct cpuidle_device);
 282         if (pseries_cpuidle_devices == NULL)
 283                 return -ENOMEM;
 284
 285         for_each_possible_cpu(i) {
 286                 dev = per_cpu_ptr(pseries_cpuidle_devices, i);
 287                 dev->state_count = drv->state_count;
 288                 dev->cpu = i;
 289                 if (cpuidle_register_device(dev)) {
 290                         printk(KERN_DEBUG \
 291                                 "cpuidle_register_device %d failed!\n", i);
 292                         return -EIO;
 293                 }
 294         }
 295
 296         return 0;
 297 }
 298
 299 /*
 300  * pseries_idle_probe()
 301  * Choose state table for shared versus dedicated partition
 302  */
 303 static int pseries_idle_probe(void)
 304 {
 305
 306         if (!firmware_has_feature(FW_FEATURE_SPLPAR))
 307                 return -ENODEV;
 308
 309         if (cpuidle_disable != IDLE_NO_OVERRIDE)
 310                 return -ENODEV;
 311
 312         if (max_idle_state == 0) {
 313                 printk(KERN_DEBUG "pseries processor idle disabled.\n");
 314                 return -EPERM;
 315         }
 316
 317         if (get_lppaca()->shared_proc)
 318                 cpuidle_state_table = shared_states;
 319         else
 320                 cpuidle_state_table = dedicated_states;
 321
 322         return 0;
 323 }
 324
 325 static int __init pseries_processor_idle_init(void)
 326 {
 327         int retval;
 328
 329         retval = pseries_idle_probe();
 330         if (retval)
 331                 return retval;
 332
 333         pseries_cpuidle_driver_init();
 334         retval = cpuidle_register_driver(&pseries_idle_driver);
 335         if (retval) {
 336                 printk(KERN_DEBUG "Registration of pseries driver failed.\n");
 337                 return retval;
 338         }
 339
 340         retval = pseries_idle_devices_init();
 341         if (retval) {
 342                 pseries_idle_devices_uninit();
 343                 cpuidle_unregister_driver(&pseries_idle_driver);
 344                 return retval;
 345         }
 346
 347         register_cpu_notifier(&setup_hotplug_notifier);
 348         printk(KERN_DEBUG "pseries_idle_driver registered\n");
 349
 350         return 0;
 351 }
 352
 353 static void __exit pseries_processor_idle_exit(void)
 354 {
 355
 356         unregister_cpu_notifier(&setup_hotplug_notifier);
 357         pseries_idle_devices_uninit();
 358         cpuidle_unregister_driver(&pseries_idle_driver);
 359
 360         return;
 361 }
 362
 363 module_init(pseries_processor_idle_init);
 364 module_exit(pseries_processor_idle_exit);
 365
 366 MODULE_AUTHOR("Deepthi Dharwar <deepthi@linux.vnet.ibm.com>");
 367 MODULE_DESCRIPTION("Cpuidle driver for POWER");
 368 MODULE_LICENSE("GPL");