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         idle_loop_prolog(&in_purr);
  66         local_irq_enable();
  67         set_thread_flag(TIF_POLLING_NRFLAG);
  68         snooze_exit_time = get_tb() + snooze_timeout;
  69
  70         while (!need_resched()) {
  71                 HMT_low();
  72                 HMT_very_low();
  73                 if (snooze_timeout_en && get_tb() > snooze_exit_time)
  74                         break;
  75         }
  76
  77         HMT_medium();
  78         clear_thread_flag(TIF_POLLING_NRFLAG);
  79         smp_mb();
  80
  81         idle_loop_epilog(in_purr);
  82
  83         return index;
  84 }
  85
  86 static void check_and_cede_processor(void)
  87 {
  88         /*
  89          * Ensure our interrupt state is properly tracked,
  90          * also checks if no interrupt has occurred while we
  91          * were soft-disabled
  92          */
  93         if (prep_irq_for_idle()) {
  94                 cede_processor();
  95 #ifdef CONFIG_TRACE_IRQFLAGS
  96                 /* Ensure that H_CEDE returns with IRQs on */
  97                 if (WARN_ON(!(mfmsr() & MSR_EE)))
  98                         __hard_irq_enable();
  99 #endif
 100         }
 101 }
 102
 103 static int dedicated_cede_loop(struct cpuidle_device *dev,
 104                                 struct cpuidle_driver *drv,
 105                                 int index)
 106 {
 107         unsigned long in_purr;
 108
 109         idle_loop_prolog(&in_purr);
 110         get_lppaca()->donate_dedicated_cpu = 1;
 111
 112         HMT_medium();
 113         check_and_cede_processor();
 114
 115         get_lppaca()->donate_dedicated_cpu = 0;
 116
 117         idle_loop_epilog(in_purr);
 118
 119         return index;
 120 }
 121
 122 static int shared_cede_loop(struct cpuidle_device *dev,
 123                         struct cpuidle_driver *drv,
 124                         int index)
 125 {
 126         unsigned long in_purr;
 127
 128         idle_loop_prolog(&in_purr);
 129
 130         /*
 131          * Yield the processor to the hypervisor.  We return if
 132          * an external interrupt occurs (which are driven prior
 133          * to returning here) or if a prod occurs from another
 134          * processor. When returning here, external interrupts
 135          * are enabled.
 136          */
 137         check_and_cede_processor();
 138
 139         idle_loop_epilog(in_purr);
 140
 141         return index;
 142 }
 143
 144 /*
 145  * States for dedicated partition case.
 146  */
 147 static struct cpuidle_state dedicated_states[] = {
 148         { /* Snooze */
 149                 .name = "snooze",
 150                 .desc = "snooze",
 151                 .exit_latency = 0,
 152                 .target_residency = 0,
 153                 .enter = &snooze_loop },
 154         { /* CEDE */
 155                 .name = "CEDE",
 156                 .desc = "CEDE",
 157                 .exit_latency = 10,
 158                 .target_residency = 100,
 159                 .enter = &dedicated_cede_loop },
 160 };
 161
 162 /*
 163  * States for shared partition case.
 164  */
 165 static struct cpuidle_state shared_states[] = {
 166         { /* Shared Cede */
 167                 .name = "Shared Cede",
 168                 .desc = "Shared Cede",
 169                 .exit_latency = 0,
 170                 .target_residency = 0,
 171                 .enter = &shared_cede_loop },
 172 };
 173
 174 static int pseries_cpuidle_cpu_online(unsigned int cpu)
 175 {
 176         struct cpuidle_device *dev = per_cpu(cpuidle_devices, cpu);
 177
 178         if (dev && cpuidle_get_driver()) {
 179                 cpuidle_pause_and_lock();
 180                 cpuidle_enable_device(dev);
 181                 cpuidle_resume_and_unlock();
 182         }
 183         return 0;
 184 }
 185
 186 static int pseries_cpuidle_cpu_dead(unsigned int cpu)
 187 {
 188         struct cpuidle_device *dev = per_cpu(cpuidle_devices, cpu);
 189
 190         if (dev && cpuidle_get_driver()) {
 191                 cpuidle_pause_and_lock();
 192                 cpuidle_disable_device(dev);
 193                 cpuidle_resume_and_unlock();
 194         }
 195         return 0;
 196 }
 197
 198 /*
 199  * pseries_cpuidle_driver_init()
 200  */
 201 static int pseries_cpuidle_driver_init(void)
 202 {
 203         int idle_state;
 204         struct cpuidle_driver *drv = &pseries_idle_driver;
 205
 206         drv->state_count = 0;
 207
 208         for (idle_state = 0; idle_state < max_idle_state; ++idle_state) {
 209                 /* Is the state not enabled? */
 210                 if (cpuidle_state_table[idle_state].enter == NULL)
 211                         continue;
 212
 213                 drv->states[drv->state_count] = /* structure copy */
 214                         cpuidle_state_table[idle_state];
 215
 216                 drv->state_count += 1;
 217         }
 218
 219         return 0;
 220 }
 221
 222 /*
 223  * pseries_idle_probe()
 224  * Choose state table for shared versus dedicated partition
 225  */
 226 static int pseries_idle_probe(void)
 227 {
 228
 229         if (cpuidle_disable != IDLE_NO_OVERRIDE)
 230                 return -ENODEV;
 231
 232         if (firmware_has_feature(FW_FEATURE_SPLPAR)) {
 233                 /*
 234                  * Use local_paca instead of get_lppaca() since
 235                  * preemption is not disabled, and it is not required in
 236                  * fact, since lppaca_ptr does not need to be the value
 237                  * associated to the current CPU, it can be from any CPU.
 238                  */
 239                 if (lppaca_shared_proc(local_paca->lppaca_ptr)) {
 240                         cpuidle_state_table = shared_states;
 241                         max_idle_state = ARRAY_SIZE(shared_states);
 242                 } else {
 243                         cpuidle_state_table = dedicated_states;
 244                         max_idle_state = ARRAY_SIZE(dedicated_states);
 245                 }
 246         } else
 247                 return -ENODEV;
 248
 249         if (max_idle_state > 1) {
 250                 snooze_timeout_en = true;
 251                 snooze_timeout = cpuidle_state_table[1].target_residency *
 252                                  tb_ticks_per_usec;
 253         }
 254         return 0;
 255 }
 256
 257 static int __init pseries_processor_idle_init(void)
 258 {
 259         int retval;
 260
 261         retval = pseries_idle_probe();
 262         if (retval)
 263                 return retval;
 264
 265         pseries_cpuidle_driver_init();
 266         retval = cpuidle_register(&pseries_idle_driver, NULL);
 267         if (retval) {
 268                 printk(KERN_DEBUG "Registration of pseries driver failed.\n");
 269                 return retval;
 270         }
 271
 272         retval = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN,
 273                                            "cpuidle/pseries:online",
 274                                            pseries_cpuidle_cpu_online, NULL);
 275         WARN_ON(retval < 0);
 276         retval = cpuhp_setup_state_nocalls(CPUHP_CPUIDLE_DEAD,
 277                                            "cpuidle/pseries:DEAD", NULL,
 278                                            pseries_cpuidle_cpu_dead);
 279         WARN_ON(retval < 0);
 280         printk(KERN_DEBUG "pseries_idle_driver registered\n");
 281         return 0;
 282 }
 283
 284 device_initcall(pseries_processor_idle_init);