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/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 static int max_idle_state;
  29 static struct cpuidle_state *cpuidle_state_table;
  30 static u64 snooze_timeout;
  31 static bool snooze_timeout_en;
  32
  33 static inline void idle_loop_prolog(unsigned long *in_purr)
  34 {
  35         ppc64_runlatch_off();
  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         if (irqs_disabled())
  54                 local_irq_enable();
  55         ppc64_runlatch_on();
  56 }
  57
  58 static int snooze_loop(struct cpuidle_device *dev,
  59                         struct cpuidle_driver *drv,
  60                         int index)
  61 {
  62         unsigned long in_purr;
  63         u64 snooze_exit_time;
  64
  65         set_thread_flag(TIF_POLLING_NRFLAG);
  66
  67         idle_loop_prolog(&in_purr);
  68         local_irq_enable();
  69         snooze_exit_time = get_tb() + snooze_timeout;
  70
  71         while (!need_resched()) {
  72                 HMT_low();
  73                 HMT_very_low();
  74                 if (snooze_timeout_en && get_tb() > snooze_exit_time)
  75                         break;
  76         }
  77
  78         HMT_medium();
  79         clear_thread_flag(TIF_POLLING_NRFLAG);
  80         smp_mb();
  81
  82         idle_loop_epilog(in_purr);
  83
  84         return index;
  85 }
  86
  87 static void check_and_cede_processor(void)
  88 {
  89         /*
  90          * Ensure our interrupt state is properly tracked,
  91          * also checks if no interrupt has occurred while we
  92          * were soft-disabled
  93          */
  94         if (prep_irq_for_idle()) {
  95                 cede_processor();
  96 #ifdef CONFIG_TRACE_IRQFLAGS
  97                 /* Ensure that H_CEDE returns with IRQs on */
  98                 if (WARN_ON(!(mfmsr() & MSR_EE)))
  99                         __hard_irq_enable();
 100 #endif
 101         }
 102 }
 103
 104 static int dedicated_cede_loop(struct cpuidle_device *dev,
 105                                 struct cpuidle_driver *drv,
 106                                 int index)
 107 {
 108         unsigned long in_purr;
 109
 110         idle_loop_prolog(&in_purr);
 111         get_lppaca()->donate_dedicated_cpu = 1;
 112
 113         HMT_medium();
 114         check_and_cede_processor();
 115
 116         get_lppaca()->donate_dedicated_cpu = 0;
 117
 118         idle_loop_epilog(in_purr);
 119
 120         return index;
 121 }
 122
 123 static int shared_cede_loop(struct cpuidle_device *dev,
 124                         struct cpuidle_driver *drv,
 125                         int index)
 126 {
 127         unsigned long in_purr;
 128
 129         idle_loop_prolog(&in_purr);
 130
 131         /*
 132          * Yield the processor to the hypervisor.  We return if
 133          * an external interrupt occurs (which are driven prior
 134          * to returning here) or if a prod occurs from another
 135          * processor. When returning here, external interrupts
 136          * are enabled.
 137          */
 138         check_and_cede_processor();
 139
 140         idle_loop_epilog(in_purr);
 141
 142         return index;
 143 }
 144
 145 /*
 146  * States for dedicated partition case.
 147  */
 148 static struct cpuidle_state dedicated_states[] = {
 149         { /* Snooze */
 150                 .name = "snooze",
 151                 .desc = "snooze",
 152                 .exit_latency = 0,
 153                 .target_residency = 0,
 154                 .enter = &snooze_loop },
 155         { /* CEDE */
 156                 .name = "CEDE",
 157                 .desc = "CEDE",
 158                 .exit_latency = 10,
 159                 .target_residency = 100,
 160                 .enter = &dedicated_cede_loop },
 161 };
 162
 163 /*
 164  * States for shared partition case.
 165  */
 166 static struct cpuidle_state shared_states[] = {
 167         { /* Shared Cede */
 168                 .name = "Shared Cede",
 169                 .desc = "Shared Cede",
 170                 .exit_latency = 0,
 171                 .target_residency = 0,
 172                 .enter = &shared_cede_loop },
 173 };
 174
 175 static int pseries_cpuidle_cpu_online(unsigned int cpu)
 176 {
 177         struct cpuidle_device *dev = per_cpu(cpuidle_devices, cpu);
 178
 179         if (dev && cpuidle_get_driver()) {
 180                 cpuidle_pause_and_lock();
 181                 cpuidle_enable_device(dev);
 182                 cpuidle_resume_and_unlock();
 183         }
 184         return 0;
 185 }
 186
 187 static int pseries_cpuidle_cpu_dead(unsigned int cpu)
 188 {
 189         struct cpuidle_device *dev = per_cpu(cpuidle_devices, cpu);
 190
 191         if (dev && cpuidle_get_driver()) {
 192                 cpuidle_pause_and_lock();
 193                 cpuidle_disable_device(dev);
 194                 cpuidle_resume_and_unlock();
 195         }
 196         return 0;
 197 }
 198
 199 /*
 200  * pseries_cpuidle_driver_init()
 201  */
 202 static int pseries_cpuidle_driver_init(void)
 203 {
 204         int idle_state;
 205         struct cpuidle_driver *drv = &pseries_idle_driver;
 206
 207         drv->state_count = 0;
 208
 209         for (idle_state = 0; idle_state < max_idle_state; ++idle_state) {
 210                 /* Is the state not enabled? */
 211                 if (cpuidle_state_table[idle_state].enter == NULL)
 212                         continue;
 213
 214                 drv->states[drv->state_count] = /* structure copy */
 215                         cpuidle_state_table[idle_state];
 216
 217                 drv->state_count += 1;
 218         }
 219
 220         return 0;
 221 }
 222
 223 /*
 224  * pseries_idle_probe()
 225  * Choose state table for shared versus dedicated partition
 226  */
 227 static int pseries_idle_probe(void)
 228 {
 229
 230         if (cpuidle_disable != IDLE_NO_OVERRIDE)
 231                 return -ENODEV;
 232
 233         if (firmware_has_feature(FW_FEATURE_SPLPAR)) {
 234                 if (lppaca_shared_proc(get_lppaca())) {
 235                         cpuidle_state_table = shared_states;
 236                         max_idle_state = ARRAY_SIZE(shared_states);
 237                 } else {
 238                         cpuidle_state_table = dedicated_states;
 239                         max_idle_state = ARRAY_SIZE(dedicated_states);
 240                 }
 241         } else
 242                 return -ENODEV;
 243
 244         if (max_idle_state > 1) {
 245                 snooze_timeout_en = true;
 246                 snooze_timeout = cpuidle_state_table[1].target_residency *
 247                                  tb_ticks_per_usec;
 248         }
 249         return 0;
 250 }
 251
 252 static int __init pseries_processor_idle_init(void)
 253 {
 254         int retval;
 255
 256         retval = pseries_idle_probe();
 257         if (retval)
 258                 return retval;
 259
 260         pseries_cpuidle_driver_init();
 261         retval = cpuidle_register(&pseries_idle_driver, NULL);
 262         if (retval) {
 263                 printk(KERN_DEBUG "Registration of pseries driver failed.\n");
 264                 return retval;
 265         }
 266
 267         retval = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN,
 268                                            "cpuidle/pseries:online",
 269                                            pseries_cpuidle_cpu_online, NULL);
 270         WARN_ON(retval < 0);
 271         retval = cpuhp_setup_state_nocalls(CPUHP_CPUIDLE_DEAD,
 272                                            "cpuidle/pseries:DEAD", NULL,
 273                                            pseries_cpuidle_cpu_dead);
 274         WARN_ON(retval < 0);
 275         printk(KERN_DEBUG "pseries_idle_driver registered\n");
 276         return 0;
 277 }
 278
 279 device_initcall(pseries_processor_idle_init);