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 static int max_idle_state;
  28 static struct cpuidle_state *cpuidle_state_table;
  29
  30 static inline void idle_loop_prolog(unsigned long *in_purr)
  31 {
  32         *in_purr = mfspr(SPRN_PURR);
  33         /*
  34          * Indicate to the HV that we are idle. Now would be
  35          * a good time to find other work to dispatch.
  36          */
  37         get_lppaca()->idle = 1;
  38 }
  39
  40 static inline void idle_loop_epilog(unsigned long in_purr)
  41 {
  42         u64 wait_cycles;
  43
  44         wait_cycles = be64_to_cpu(get_lppaca()->wait_state_cycles);
  45         wait_cycles += mfspr(SPRN_PURR) - in_purr;
  46         get_lppaca()->wait_state_cycles = cpu_to_be64(wait_cycles);
  47         get_lppaca()->idle = 0;
  48 }
  49
  50 static int snooze_loop(struct cpuidle_device *dev,
  51                         struct cpuidle_driver *drv,
  52                         int index)
  53 {
  54         unsigned long in_purr;
  55
  56         idle_loop_prolog(&in_purr);
  57         local_irq_enable();
  58         set_thread_flag(TIF_POLLING_NRFLAG);
  59
  60         while (!need_resched()) {
  61                 HMT_low();
  62                 HMT_very_low();
  63         }
  64
  65         HMT_medium();
  66         clear_thread_flag(TIF_POLLING_NRFLAG);
  67         smp_mb();
  68
  69         idle_loop_epilog(in_purr);
  70
  71         return index;
  72 }
  73
  74 static void check_and_cede_processor(void)
  75 {
  76         /*
  77          * Ensure our interrupt state is properly tracked,
  78          * also checks if no interrupt has occurred while we
  79          * were soft-disabled
  80          */
  81         if (prep_irq_for_idle()) {
  82                 cede_processor();
  83 #ifdef CONFIG_TRACE_IRQFLAGS
  84                 /* Ensure that H_CEDE returns with IRQs on */
  85                 if (WARN_ON(!(mfmsr() & MSR_EE)))
  86                         __hard_irq_enable();
  87 #endif
  88         }
  89 }
  90
  91 static int dedicated_cede_loop(struct cpuidle_device *dev,
  92                                 struct cpuidle_driver *drv,
  93                                 int index)
  94 {
  95         unsigned long in_purr;
  96
  97         idle_loop_prolog(&in_purr);
  98         get_lppaca()->donate_dedicated_cpu = 1;
  99
 100         HMT_medium();
 101         check_and_cede_processor();
 102
 103         get_lppaca()->donate_dedicated_cpu = 0;
 104
 105         idle_loop_epilog(in_purr);
 106
 107         return index;
 108 }
 109
 110 static int shared_cede_loop(struct cpuidle_device *dev,
 111                         struct cpuidle_driver *drv,
 112                         int index)
 113 {
 114         unsigned long in_purr;
 115
 116         idle_loop_prolog(&in_purr);
 117
 118         /*
 119          * Yield the processor to the hypervisor.  We return if
 120          * an external interrupt occurs (which are driven prior
 121          * to returning here) or if a prod occurs from another
 122          * processor. When returning here, external interrupts
 123          * are enabled.
 124          */
 125         check_and_cede_processor();
 126
 127         idle_loop_epilog(in_purr);
 128
 129         return index;
 130 }
 131
 132 /*
 133  * States for dedicated partition case.
 134  */
 135 static struct cpuidle_state dedicated_states[] = {
 136         { /* Snooze */
 137                 .name = "snooze",
 138                 .desc = "snooze",
 139                 .flags = CPUIDLE_FLAG_TIME_VALID,
 140                 .exit_latency = 0,
 141                 .target_residency = 0,
 142                 .enter = &snooze_loop },
 143         { /* CEDE */
 144                 .name = "CEDE",
 145                 .desc = "CEDE",
 146                 .flags = CPUIDLE_FLAG_TIME_VALID,
 147                 .exit_latency = 10,
 148                 .target_residency = 100,
 149                 .enter = &dedicated_cede_loop },
 150 };
 151
 152 /*
 153  * States for shared partition case.
 154  */
 155 static struct cpuidle_state shared_states[] = {
 156         { /* Shared Cede */
 157                 .name = "Shared Cede",
 158                 .desc = "Shared Cede",
 159                 .flags = CPUIDLE_FLAG_TIME_VALID,
 160                 .exit_latency = 0,
 161                 .target_residency = 0,
 162                 .enter = &shared_cede_loop },
 163 };
 164
 165 void update_smt_snooze_delay(int cpu, int residency)
 166 {
 167         struct cpuidle_driver *drv = cpuidle_get_driver();
 168         struct cpuidle_device *dev = per_cpu(cpuidle_devices, cpu);
 169
 170         if (cpuidle_state_table != dedicated_states)
 171                 return;
 172
 173         if (residency < 0) {
 174                 /* Disable the Nap state on that cpu */
 175                 if (dev)
 176                         dev->states_usage[1].disable = 1;
 177         } else
 178                 if (drv)
 179                         drv->states[1].target_residency = residency;
 180 }
 181
 182 static int pseries_cpuidle_add_cpu_notifier(struct notifier_block *n,
 183                         unsigned long action, void *hcpu)
 184 {
 185         int hotcpu = (unsigned long)hcpu;
 186         struct cpuidle_device *dev =
 187                                 per_cpu(cpuidle_devices, hotcpu);
 188
 189         if (dev && cpuidle_get_driver()) {
 190                 switch (action) {
 191                 case CPU_ONLINE:
 192                 case CPU_ONLINE_FROZEN:
 193                         cpuidle_pause_and_lock();
 194                         cpuidle_enable_device(dev);
 195                         cpuidle_resume_and_unlock();
 196                         break;
 197
 198                 case CPU_DEAD:
 199                 case CPU_DEAD_FROZEN:
 200                         cpuidle_pause_and_lock();
 201                         cpuidle_disable_device(dev);
 202                         cpuidle_resume_and_unlock();
 203                         break;
 204
 205                 default:
 206                         return NOTIFY_DONE;
 207                 }
 208         }
 209         return NOTIFY_OK;
 210 }
 211
 212 static struct notifier_block setup_hotplug_notifier = {
 213         .notifier_call = pseries_cpuidle_add_cpu_notifier,
 214 };
 215
 216 /*
 217  * pseries_cpuidle_driver_init()
 218  */
 219 static int pseries_cpuidle_driver_init(void)
 220 {
 221         int idle_state;
 222         struct cpuidle_driver *drv = &pseries_idle_driver;
 223
 224         drv->state_count = 0;
 225
 226         for (idle_state = 0; idle_state < max_idle_state; ++idle_state) {
 227                 /* Is the state not enabled? */
 228                 if (cpuidle_state_table[idle_state].enter == NULL)
 229                         continue;
 230
 231                 drv->states[drv->state_count] = /* structure copy */
 232                         cpuidle_state_table[idle_state];
 233
 234                 drv->state_count += 1;
 235         }
 236
 237         return 0;
 238 }
 239
 240 /*
 241  * pseries_idle_probe()
 242  * Choose state table for shared versus dedicated partition
 243  */
 244 static int pseries_idle_probe(void)
 245 {
 246
 247         if (cpuidle_disable != IDLE_NO_OVERRIDE)
 248                 return -ENODEV;
 249
 250         if (firmware_has_feature(FW_FEATURE_SPLPAR)) {
 251                 if (lppaca_shared_proc(get_lppaca())) {
 252                         cpuidle_state_table = shared_states;
 253                         max_idle_state = ARRAY_SIZE(shared_states);
 254                 } else {
 255                         cpuidle_state_table = dedicated_states;
 256                         max_idle_state = ARRAY_SIZE(dedicated_states);
 257                 }
 258         } else
 259                 return -ENODEV;
 260
 261         return 0;
 262 }
 263
 264 static int __init pseries_processor_idle_init(void)
 265 {
 266         int retval;
 267
 268         retval = pseries_idle_probe();
 269         if (retval)
 270                 return retval;
 271
 272         pseries_cpuidle_driver_init();
 273         retval = cpuidle_register(&pseries_idle_driver, NULL);
 274         if (retval) {
 275                 printk(KERN_DEBUG "Registration of pseries driver failed.\n");
 276                 return retval;
 277         }
 278
 279         register_cpu_notifier(&setup_hotplug_notifier);
 280         printk(KERN_DEBUG "pseries_idle_driver registered\n");
 281         return 0;
 282 }
 283
 284 device_initcall(pseries_processor_idle_init);