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 #include <asm/plpar_wrappers.h>
  22
  23 struct cpuidle_driver pseries_idle_driver = {
  24         .name             = "pseries_idle",
  25         .owner            = THIS_MODULE,
  26 };
  27
  28 #define MAX_IDLE_STATE_COUNT    2
  29
  30 static int max_idle_state = MAX_IDLE_STATE_COUNT - 1;
  31 static struct cpuidle_device __percpu *pseries_cpuidle_devices;
  32 static struct cpuidle_state *cpuidle_state_table;
  33
  34 static inline void idle_loop_prolog(unsigned long *in_purr)
  35 {
  36         *in_purr = mfspr(SPRN_PURR);
  37         /*
  38          * Indicate to the HV that we are idle. Now would be
  39          * a good time to find other work to dispatch.
  40          */
  41         get_lppaca()->idle = 1;
  42 }
  43
  44 static inline void idle_loop_epilog(unsigned long in_purr)
  45 {
  46         u64 wait_cycles;
  47
  48         wait_cycles = be64_to_cpu(get_lppaca()->wait_state_cycles);
  49         wait_cycles += mfspr(SPRN_PURR) - in_purr;
  50         get_lppaca()->wait_state_cycles = cpu_to_be64(wait_cycles);
  51         get_lppaca()->idle = 0;
  52 }
  53
  54 static int snooze_loop(struct cpuidle_device *dev,
  55                         struct cpuidle_driver *drv,
  56                         int index)
  57 {
  58         unsigned long in_purr;
  59         int cpu = dev->cpu;
  60
  61         idle_loop_prolog(&in_purr);
  62         local_irq_enable();
  63         set_thread_flag(TIF_POLLING_NRFLAG);
  64
  65         while ((!need_resched()) && cpu_online(cpu)) {
  66                 ppc64_runlatch_off();
  67                 HMT_low();
  68                 HMT_very_low();
  69         }
  70
  71         HMT_medium();
  72         clear_thread_flag(TIF_POLLING_NRFLAG);
  73         smp_mb();
  74
  75         idle_loop_epilog(in_purr);
  76
  77         return index;
  78 }
  79
  80 static void check_and_cede_processor(void)
  81 {
  82         /*
  83          * Ensure our interrupt state is properly tracked,
  84          * also checks if no interrupt has occurred while we
  85          * were soft-disabled
  86          */
  87         if (prep_irq_for_idle()) {
  88                 cede_processor();
  89 #ifdef CONFIG_TRACE_IRQFLAGS
  90                 /* Ensure that H_CEDE returns with IRQs on */
  91                 if (WARN_ON(!(mfmsr() & MSR_EE)))
  92                         __hard_irq_enable();
  93 #endif
  94         }
  95 }
  96
  97 static int dedicated_cede_loop(struct cpuidle_device *dev,
  98                                 struct cpuidle_driver *drv,
  99                                 int index)
 100 {
 101         unsigned long in_purr;
 102
 103         idle_loop_prolog(&in_purr);
 104         get_lppaca()->donate_dedicated_cpu = 1;
 105
 106         ppc64_runlatch_off();
 107         HMT_medium();
 108         check_and_cede_processor();
 109
 110         get_lppaca()->donate_dedicated_cpu = 0;
 111
 112         idle_loop_epilog(in_purr);
 113
 114         return index;
 115 }
 116
 117 static int shared_cede_loop(struct cpuidle_device *dev,
 118                         struct cpuidle_driver *drv,
 119                         int index)
 120 {
 121         unsigned long in_purr;
 122
 123         idle_loop_prolog(&in_purr);
 124
 125         /*
 126          * Yield the processor to the hypervisor.  We return if
 127          * an external interrupt occurs (which are driven prior
 128          * to returning here) or if a prod occurs from another
 129          * processor. When returning here, external interrupts
 130          * are enabled.
 131          */
 132         check_and_cede_processor();
 133
 134         idle_loop_epilog(in_purr);
 135
 136         return index;
 137 }
 138
 139 /*
 140  * States for dedicated partition case.
 141  */
 142 static struct cpuidle_state dedicated_states[MAX_IDLE_STATE_COUNT] = {
 143         { /* Snooze */
 144                 .name = "snooze",
 145                 .desc = "snooze",
 146                 .flags = CPUIDLE_FLAG_TIME_VALID,
 147                 .exit_latency = 0,
 148                 .target_residency = 0,
 149                 .enter = &snooze_loop },
 150         { /* CEDE */
 151                 .name = "CEDE",
 152                 .desc = "CEDE",
 153                 .flags = CPUIDLE_FLAG_TIME_VALID,
 154                 .exit_latency = 10,
 155                 .target_residency = 100,
 156                 .enter = &dedicated_cede_loop },
 157 };
 158
 159 /*
 160  * States for shared partition case.
 161  */
 162 static struct cpuidle_state shared_states[MAX_IDLE_STATE_COUNT] = {
 163         { /* Shared Cede */
 164                 .name = "Shared Cede",
 165                 .desc = "Shared Cede",
 166                 .flags = CPUIDLE_FLAG_TIME_VALID,
 167                 .exit_latency = 0,
 168                 .target_residency = 0,
 169                 .enter = &shared_cede_loop },
 170 };
 171
 172 void update_smt_snooze_delay(int cpu, int residency)
 173 {
 174         struct cpuidle_driver *drv = cpuidle_get_driver();
 175         struct cpuidle_device *dev = per_cpu(cpuidle_devices, cpu);
 176
 177         if (cpuidle_state_table != dedicated_states)
 178                 return;
 179
 180         if (residency < 0) {
 181                 /* Disable the Nap state on that cpu */
 182                 if (dev)
 183                         dev->states_usage[1].disable = 1;
 184         } else
 185                 if (drv)
 186                         drv->states[1].target_residency = residency;
 187 }
 188
 189 static int pseries_cpuidle_add_cpu_notifier(struct notifier_block *n,
 190                         unsigned long action, void *hcpu)
 191 {
 192         int hotcpu = (unsigned long)hcpu;
 193         struct cpuidle_device *dev =
 194                         per_cpu_ptr(pseries_cpuidle_devices, hotcpu);
 195
 196         if (dev && cpuidle_get_driver()) {
 197                 switch (action) {
 198                 case CPU_ONLINE:
 199                 case CPU_ONLINE_FROZEN:
 200                         cpuidle_pause_and_lock();
 201                         cpuidle_enable_device(dev);
 202                         cpuidle_resume_and_unlock();
 203                         break;
 204
 205                 case CPU_DEAD:
 206                 case CPU_DEAD_FROZEN:
 207                         cpuidle_pause_and_lock();
 208                         cpuidle_disable_device(dev);
 209                         cpuidle_resume_and_unlock();
 210                         break;
 211
 212                 default:
 213                         return NOTIFY_DONE;
 214                 }
 215         }
 216         return NOTIFY_OK;
 217 }
 218
 219 static struct notifier_block setup_hotplug_notifier = {
 220         .notifier_call = pseries_cpuidle_add_cpu_notifier,
 221 };
 222
 223 /*
 224  * pseries_cpuidle_driver_init()
 225  */
 226 static int pseries_cpuidle_driver_init(void)
 227 {
 228         int idle_state;
 229         struct cpuidle_driver *drv = &pseries_idle_driver;
 230
 231         drv->state_count = 0;
 232
 233         for (idle_state = 0; idle_state < MAX_IDLE_STATE_COUNT; ++idle_state) {
 234
 235                 if (idle_state > max_idle_state)
 236                         break;
 237
 238                 /* is the state not enabled? */
 239                 if (cpuidle_state_table[idle_state].enter == NULL)
 240                         continue;
 241
 242                 drv->states[drv->state_count] = /* structure copy */
 243                         cpuidle_state_table[idle_state];
 244
 245                 drv->state_count += 1;
 246         }
 247
 248         return 0;
 249 }
 250
 251 /* pseries_idle_devices_uninit(void)
 252  * unregister cpuidle devices and de-allocate memory
 253  */
 254 static void pseries_idle_devices_uninit(void)
 255 {
 256         int i;
 257         struct cpuidle_device *dev;
 258
 259         for_each_possible_cpu(i) {
 260                 dev = per_cpu_ptr(pseries_cpuidle_devices, i);
 261                 cpuidle_unregister_device(dev);
 262         }
 263
 264         free_percpu(pseries_cpuidle_devices);
 265         return;
 266 }
 267
 268 /* pseries_idle_devices_init()
 269  * allocate, initialize and register cpuidle device
 270  */
 271 static int pseries_idle_devices_init(void)
 272 {
 273         int i;
 274         struct cpuidle_driver *drv = &pseries_idle_driver;
 275         struct cpuidle_device *dev;
 276
 277         pseries_cpuidle_devices = alloc_percpu(struct cpuidle_device);
 278         if (pseries_cpuidle_devices == NULL)
 279                 return -ENOMEM;
 280
 281         for_each_possible_cpu(i) {
 282                 dev = per_cpu_ptr(pseries_cpuidle_devices, i);
 283                 dev->state_count = drv->state_count;
 284                 dev->cpu = i;
 285                 if (cpuidle_register_device(dev)) {
 286                         printk(KERN_DEBUG \
 287                                 "cpuidle_register_device %d failed!\n", i);
 288                         return -EIO;
 289                 }
 290         }
 291
 292         return 0;
 293 }
 294
 295 /*
 296  * pseries_idle_probe()
 297  * Choose state table for shared versus dedicated partition
 298  */
 299 static int pseries_idle_probe(void)
 300 {
 301
 302         if (!firmware_has_feature(FW_FEATURE_SPLPAR))
 303                 return -ENODEV;
 304
 305         if (cpuidle_disable != IDLE_NO_OVERRIDE)
 306                 return -ENODEV;
 307
 308         if (max_idle_state == 0) {
 309                 printk(KERN_DEBUG "pseries processor idle disabled.\n");
 310                 return -EPERM;
 311         }
 312
 313         if (lppaca_shared_proc(get_lppaca()))
 314                 cpuidle_state_table = shared_states;
 315         else
 316                 cpuidle_state_table = dedicated_states;
 317
 318         return 0;
 319 }
 320
 321 static int __init pseries_processor_idle_init(void)
 322 {
 323         int retval;
 324
 325         retval = pseries_idle_probe();
 326         if (retval)
 327                 return retval;
 328
 329         pseries_cpuidle_driver_init();
 330         retval = cpuidle_register_driver(&pseries_idle_driver);
 331         if (retval) {
 332                 printk(KERN_DEBUG "Registration of pseries driver failed.\n");
 333                 return retval;
 334         }
 335
 336         retval = pseries_idle_devices_init();
 337         if (retval) {
 338                 pseries_idle_devices_uninit();
 339                 cpuidle_unregister_driver(&pseries_idle_driver);
 340                 return retval;
 341         }
 342
 343         register_cpu_notifier(&setup_hotplug_notifier);
 344         printk(KERN_DEBUG "pseries_idle_driver registered\n");
 345
 346         return 0;
 347 }
 348
 349 static void __exit pseries_processor_idle_exit(void)
 350 {
 351
 352         unregister_cpu_notifier(&setup_hotplug_notifier);
 353         pseries_idle_devices_uninit();
 354         cpuidle_unregister_driver(&pseries_idle_driver);
 355
 356         return;
 357 }
 358
 359 module_init(pseries_processor_idle_init);
 360 module_exit(pseries_processor_idle_exit);
 361
 362 MODULE_AUTHOR("Deepthi Dharwar <deepthi@linux.vnet.ibm.com>");
 363 MODULE_DESCRIPTION("Cpuidle driver for POWER");
 364 MODULE_LICENSE("GPL");