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