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