drivers/cpuidle/cpuidle-pseries.c

   1 /*
   2  *  cpuidle-pseries - 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/plpar_wrappers.h>
  21
  22 struct cpuidle_driver pseries_idle_driver = {
  23         .name             = "pseries_idle",
  24         .owner            = THIS_MODULE,
  25 };
  26
  27 #define MAX_IDLE_STATE_COUNT    2
  28
  29 static int max_idle_state = MAX_IDLE_STATE_COUNT - 1;
  30 static struct cpuidle_device __percpu *pseries_cpuidle_devices;
  31 static struct cpuidle_state *cpuidle_state_table;
  32
  33 static inline void idle_loop_prolog(unsigned long *in_purr)
  34 {
  35         *in_purr = mfspr(SPRN_PURR);
  36         /*
  37          * Indicate to the HV that we are idle. Now would be
  38          * a good time to find other work to dispatch.
  39          */
  40         get_lppaca()->idle = 1;
  41 }
  42
  43 static inline void idle_loop_epilog(unsigned long in_purr)
  44 {
  45         u64 wait_cycles;
  46
  47         wait_cycles = be64_to_cpu(get_lppaca()->wait_state_cycles);
  48         wait_cycles += mfspr(SPRN_PURR) - in_purr;
  49         get_lppaca()->wait_state_cycles = cpu_to_be64(wait_cycles);
  50         get_lppaca()->idle = 0;
  51 }
  52
  53 static int snooze_loop(struct cpuidle_device *dev,
  54                         struct cpuidle_driver *drv,
  55                         int index)
  56 {
  57         unsigned long in_purr;
  58         int cpu = dev->cpu;
  59
  60         idle_loop_prolog(&in_purr);
  61         local_irq_enable();
  62         set_thread_flag(TIF_POLLING_NRFLAG);
  63
  64         while ((!need_resched()) && cpu_online(cpu)) {
  65                 HMT_low();
  66                 HMT_very_low();
  67         }
  68
  69         HMT_medium();
  70         clear_thread_flag(TIF_POLLING_NRFLAG);
  71         smp_mb();
  72
  73         idle_loop_epilog(in_purr);
  74
  75         return index;
  76 }
  77
  78 static void check_and_cede_processor(void)
  79 {
  80         /*
  81          * Ensure our interrupt state is properly tracked,
  82          * also checks if no interrupt has occurred while we
  83          * were soft-disabled
  84          */
  85         if (prep_irq_for_idle()) {
  86                 cede_processor();
  87 #ifdef CONFIG_TRACE_IRQFLAGS
  88                 /* Ensure that H_CEDE returns with IRQs on */
  89                 if (WARN_ON(!(mfmsr() & MSR_EE)))
  90                         __hard_irq_enable();
  91 #endif
  92         }
  93 }
  94
  95 static int dedicated_cede_loop(struct cpuidle_device *dev,
  96                                 struct cpuidle_driver *drv,
  97                                 int index)
  98 {
  99         unsigned long in_purr;
 100
 101         idle_loop_prolog(&in_purr);
 102         get_lppaca()->donate_dedicated_cpu = 1;
 103
 104         HMT_medium();
 105         check_and_cede_processor();
 106
 107         get_lppaca()->donate_dedicated_cpu = 0;
 108
 109         idle_loop_epilog(in_purr);
 110
 111         return index;
 112 }
 113
 114 static int shared_cede_loop(struct cpuidle_device *dev,
 115                         struct cpuidle_driver *drv,
 116                         int index)
 117 {
 118         unsigned long in_purr;
 119
 120         idle_loop_prolog(&in_purr);
 121
 122         /*
 123          * Yield the processor to the hypervisor.  We return if
 124          * an external interrupt occurs (which are driven prior
 125          * to returning here) or if a prod occurs from another
 126          * processor. When returning here, external interrupts
 127          * are enabled.
 128          */
 129         check_and_cede_processor();
 130
 131         idle_loop_epilog(in_purr);
 132
 133         return index;
 134 }
 135
 136 /*
 137  * States for dedicated partition case.
 138  */
 139 static struct cpuidle_state dedicated_states[MAX_IDLE_STATE_COUNT] = {
 140         { /* Snooze */
 141                 .name = "snooze",
 142                 .desc = "snooze",
 143                 .flags = CPUIDLE_FLAG_TIME_VALID,
 144                 .exit_latency = 0,
 145                 .target_residency = 0,
 146                 .enter = &snooze_loop },
 147         { /* CEDE */
 148                 .name = "CEDE",
 149                 .desc = "CEDE",
 150                 .flags = CPUIDLE_FLAG_TIME_VALID,
 151                 .exit_latency = 10,
 152                 .target_residency = 100,
 153                 .enter = &dedicated_cede_loop },
 154 };
 155
 156 /*
 157  * States for shared partition case.
 158  */
 159 static struct cpuidle_state shared_states[MAX_IDLE_STATE_COUNT] = {
 160         { /* Shared Cede */
 161                 .name = "Shared Cede",
 162                 .desc = "Shared Cede",
 163                 .flags = CPUIDLE_FLAG_TIME_VALID,
 164                 .exit_latency = 0,
 165                 .target_residency = 0,
 166                 .enter = &shared_cede_loop },
 167 };
 168
 169 void update_smt_snooze_delay(int cpu, int residency)
 170 {
 171         struct cpuidle_driver *drv = cpuidle_get_driver();
 172         struct cpuidle_device *dev = per_cpu(cpuidle_devices, cpu);
 173
 174         if (cpuidle_state_table != dedicated_states)
 175                 return;
 176
 177         if (residency < 0) {
 178                 /* Disable the Nap state on that cpu */
 179                 if (dev)
 180                         dev->states_usage[1].disable = 1;
 181         } else
 182                 if (drv)
 183                         drv->states[1].target_residency = residency;
 184 }
 185
 186 static int pseries_cpuidle_add_cpu_notifier(struct notifier_block *n,
 187                         unsigned long action, void *hcpu)
 188 {
 189         int hotcpu = (unsigned long)hcpu;
 190         struct cpuidle_device *dev =
 191                         per_cpu_ptr(pseries_cpuidle_devices, hotcpu);
 192
 193         if (dev && cpuidle_get_driver()) {
 194                 switch (action) {
 195                 case CPU_ONLINE:
 196                 case CPU_ONLINE_FROZEN:
 197                         cpuidle_pause_and_lock();
 198                         cpuidle_enable_device(dev);
 199                         cpuidle_resume_and_unlock();
 200                         break;
 201
 202                 case CPU_DEAD:
 203                 case CPU_DEAD_FROZEN:
 204                         cpuidle_pause_and_lock();
 205                         cpuidle_disable_device(dev);
 206                         cpuidle_resume_and_unlock();
 207                         break;
 208
 209                 default:
 210                         return NOTIFY_DONE;
 211                 }
 212         }
 213         return NOTIFY_OK;
 214 }
 215
 216 static struct notifier_block setup_hotplug_notifier = {
 217         .notifier_call = pseries_cpuidle_add_cpu_notifier,
 218 };
 219
 220 /*
 221  * pseries_cpuidle_driver_init()
 222  */
 223 static int pseries_cpuidle_driver_init(void)
 224 {
 225         int idle_state;
 226         struct cpuidle_driver *drv = &pseries_idle_driver;
 227
 228         drv->state_count = 0;
 229
 230         for (idle_state = 0; idle_state < MAX_IDLE_STATE_COUNT; ++idle_state) {
 231
 232                 if (idle_state > max_idle_state)
 233                         break;
 234
 235                 /* is the state not enabled? */
 236                 if (cpuidle_state_table[idle_state].enter == NULL)
 237                         continue;
 238
 239                 drv->states[drv->state_count] = /* structure copy */
 240                         cpuidle_state_table[idle_state];
 241
 242                 drv->state_count += 1;
 243         }
 244
 245         return 0;
 246 }
 247
 248 /* pseries_idle_devices_uninit(void)
 249  * unregister cpuidle devices and de-allocate memory
 250  */
 251 static void pseries_idle_devices_uninit(void)
 252 {
 253         int i;
 254         struct cpuidle_device *dev;
 255
 256         for_each_possible_cpu(i) {
 257                 dev = per_cpu_ptr(pseries_cpuidle_devices, i);
 258                 cpuidle_unregister_device(dev);
 259         }
 260
 261         free_percpu(pseries_cpuidle_devices);
 262         return;
 263 }
 264
 265 /* pseries_idle_devices_init()
 266  * allocate, initialize and register cpuidle device
 267  */
 268 static int pseries_idle_devices_init(void)
 269 {
 270         int i;
 271         struct cpuidle_driver *drv = &pseries_idle_driver;
 272         struct cpuidle_device *dev;
 273
 274         pseries_cpuidle_devices = alloc_percpu(struct cpuidle_device);
 275         if (pseries_cpuidle_devices == NULL)
 276                 return -ENOMEM;
 277
 278         for_each_possible_cpu(i) {
 279                 dev = per_cpu_ptr(pseries_cpuidle_devices, i);
 280                 dev->state_count = drv->state_count;
 281                 dev->cpu = i;
 282                 if (cpuidle_register_device(dev)) {
 283                         printk(KERN_DEBUG \
 284                                 "cpuidle_register_device %d failed!\n", i);
 285                         return -EIO;
 286                 }
 287         }
 288
 289         return 0;
 290 }
 291
 292 /*
 293  * pseries_idle_probe()
 294  * Choose state table for shared versus dedicated partition
 295  */
 296 static int pseries_idle_probe(void)
 297 {
 298
 299         if (!firmware_has_feature(FW_FEATURE_SPLPAR))
 300                 return -ENODEV;
 301
 302         if (cpuidle_disable != IDLE_NO_OVERRIDE)
 303                 return -ENODEV;
 304
 305         if (max_idle_state == 0) {
 306                 printk(KERN_DEBUG "pseries processor idle disabled.\n");
 307                 return -EPERM;
 308         }
 309
 310         if (lppaca_shared_proc(get_lppaca()))
 311                 cpuidle_state_table = shared_states;
 312         else
 313                 cpuidle_state_table = dedicated_states;
 314
 315         return 0;
 316 }
 317
 318 static int __init pseries_processor_idle_init(void)
 319 {
 320         int retval;
 321
 322         retval = pseries_idle_probe();
 323         if (retval)
 324                 return retval;
 325
 326         pseries_cpuidle_driver_init();
 327         retval = cpuidle_register_driver(&pseries_idle_driver);
 328         if (retval) {
 329                 printk(KERN_DEBUG "Registration of pseries driver failed.\n");
 330                 return retval;
 331         }
 332
 333         retval = pseries_idle_devices_init();
 334         if (retval) {
 335                 pseries_idle_devices_uninit();
 336                 cpuidle_unregister_driver(&pseries_idle_driver);
 337                 return retval;
 338         }
 339
 340         register_cpu_notifier(&setup_hotplug_notifier);
 341         printk(KERN_DEBUG "pseries_idle_driver registered\n");
 342
 343         return 0;
 344 }
 345
 346 static void __exit pseries_processor_idle_exit(void)
 347 {
 348
 349         unregister_cpu_notifier(&setup_hotplug_notifier);
 350         pseries_idle_devices_uninit();
 351         cpuidle_unregister_driver(&pseries_idle_driver);
 352
 353         return;
 354 }
 355
 356 module_init(pseries_processor_idle_init);
 357 module_exit(pseries_processor_idle_exit);
 358
 359 MODULE_AUTHOR("Deepthi Dharwar <deepthi@linux.vnet.ibm.com>");
 360 MODULE_DESCRIPTION("Cpuidle driver for POWER");
 361 MODULE_LICENSE("GPL");