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 __read_mostly;
  29 static struct cpuidle_state *cpuidle_state_table __read_mostly;
  30 static u64 snooze_timeout __read_mostly;
  31 static bool snooze_timeout_en __read_mostly;
  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 (likely(snooze_timeout_en) && get_tb() > snooze_exit_time) {
  75                         /*
  76                          * Task has not woken up but we are exiting the polling
  77                          * loop anyway. Require a barrier after polling is
  78                          * cleared to order subsequent test of need_resched().
  79                          */
  80                         clear_thread_flag(TIF_POLLING_NRFLAG);
  81                         smp_mb();
  82                         break;
  83                 }
  84         }
  85
  86         HMT_medium();
  87         clear_thread_flag(TIF_POLLING_NRFLAG);
  88
  89         idle_loop_epilog(in_purr);
  90
  91         return index;
  92 }
  93
  94 static void check_and_cede_processor(void)
  95 {
  96         /*
  97          * Ensure our interrupt state is properly tracked,
  98          * also checks if no interrupt has occurred while we
  99          * were soft-disabled
 100          */
 101         if (prep_irq_for_idle()) {
 102                 cede_processor();
 103 #ifdef CONFIG_TRACE_IRQFLAGS
 104                 /* Ensure that H_CEDE returns with IRQs on */
 105                 if (WARN_ON(!(mfmsr() & MSR_EE)))
 106                         __hard_irq_enable();
 107 #endif
 108         }
 109 }
 110
 111 static int dedicated_cede_loop(struct cpuidle_device *dev,
 112                                 struct cpuidle_driver *drv,
 113                                 int index)
 114 {
 115         unsigned long in_purr;
 116
 117         idle_loop_prolog(&in_purr);
 118         get_lppaca()->donate_dedicated_cpu = 1;
 119
 120         HMT_medium();
 121         check_and_cede_processor();
 122
 123         get_lppaca()->donate_dedicated_cpu = 0;
 124
 125         idle_loop_epilog(in_purr);
 126
 127         return index;
 128 }
 129
 130 static int shared_cede_loop(struct cpuidle_device *dev,
 131                         struct cpuidle_driver *drv,
 132                         int index)
 133 {
 134         unsigned long in_purr;
 135
 136         idle_loop_prolog(&in_purr);
 137
 138         /*
 139          * Yield the processor to the hypervisor.  We return if
 140          * an external interrupt occurs (which are driven prior
 141          * to returning here) or if a prod occurs from another
 142          * processor. When returning here, external interrupts
 143          * are enabled.
 144          */
 145         check_and_cede_processor();
 146
 147         idle_loop_epilog(in_purr);
 148
 149         return index;
 150 }
 151
 152 /*
 153  * States for dedicated partition case.
 154  */
 155 static struct cpuidle_state dedicated_states[] = {
 156         { /* Snooze */
 157                 .name = "snooze",
 158                 .desc = "snooze",
 159                 .exit_latency = 0,
 160                 .target_residency = 0,
 161                 .enter = &snooze_loop },
 162         { /* CEDE */
 163                 .name = "CEDE",
 164                 .desc = "CEDE",
 165                 .exit_latency = 10,
 166                 .target_residency = 100,
 167                 .enter = &dedicated_cede_loop },
 168 };
 169
 170 /*
 171  * States for shared partition case.
 172  */
 173 static struct cpuidle_state shared_states[] = {
 174         { /* Shared Cede */
 175                 .name = "Shared Cede",
 176                 .desc = "Shared Cede",
 177                 .exit_latency = 0,
 178                 .target_residency = 0,
 179                 .enter = &shared_cede_loop },
 180 };
 181
 182 static int pseries_cpuidle_cpu_online(unsigned int cpu)
 183 {
 184         struct cpuidle_device *dev = per_cpu(cpuidle_devices, cpu);
 185
 186         if (dev && cpuidle_get_driver()) {
 187                 cpuidle_pause_and_lock();
 188                 cpuidle_enable_device(dev);
 189                 cpuidle_resume_and_unlock();
 190         }
 191         return 0;
 192 }
 193
 194 static int pseries_cpuidle_cpu_dead(unsigned int cpu)
 195 {
 196         struct cpuidle_device *dev = per_cpu(cpuidle_devices, cpu);
 197
 198         if (dev && cpuidle_get_driver()) {
 199                 cpuidle_pause_and_lock();
 200                 cpuidle_disable_device(dev);
 201                 cpuidle_resume_and_unlock();
 202         }
 203         return 0;
 204 }
 205
 206 /*
 207  * pseries_cpuidle_driver_init()
 208  */
 209 static int pseries_cpuidle_driver_init(void)
 210 {
 211         int idle_state;
 212         struct cpuidle_driver *drv = &pseries_idle_driver;
 213
 214         drv->state_count = 0;
 215
 216         for (idle_state = 0; idle_state < max_idle_state; ++idle_state) {
 217                 /* Is the state not enabled? */
 218                 if (cpuidle_state_table[idle_state].enter == NULL)
 219                         continue;
 220
 221                 drv->states[drv->state_count] = /* structure copy */
 222                         cpuidle_state_table[idle_state];
 223
 224                 drv->state_count += 1;
 225         }
 226
 227         return 0;
 228 }
 229
 230 /*
 231  * pseries_idle_probe()
 232  * Choose state table for shared versus dedicated partition
 233  */
 234 static int pseries_idle_probe(void)
 235 {
 236
 237         if (cpuidle_disable != IDLE_NO_OVERRIDE)
 238                 return -ENODEV;
 239
 240         if (firmware_has_feature(FW_FEATURE_SPLPAR)) {
 241                 if (lppaca_shared_proc(get_lppaca())) {
 242                         cpuidle_state_table = shared_states;
 243                         max_idle_state = ARRAY_SIZE(shared_states);
 244                 } else {
 245                         cpuidle_state_table = dedicated_states;
 246                         max_idle_state = ARRAY_SIZE(dedicated_states);
 247                 }
 248         } else
 249                 return -ENODEV;
 250
 251         if (max_idle_state > 1) {
 252                 snooze_timeout_en = true;
 253                 snooze_timeout = cpuidle_state_table[1].target_residency *
 254                                  tb_ticks_per_usec;
 255         }
 256         return 0;
 257 }
 258
 259 static int __init pseries_processor_idle_init(void)
 260 {
 261         int retval;
 262
 263         retval = pseries_idle_probe();
 264         if (retval)
 265                 return retval;
 266
 267         pseries_cpuidle_driver_init();
 268         retval = cpuidle_register(&pseries_idle_driver, NULL);
 269         if (retval) {
 270                 printk(KERN_DEBUG "Registration of pseries driver failed.\n");
 271                 return retval;
 272         }
 273
 274         retval = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN,
 275                                            "cpuidle/pseries:online",
 276                                            pseries_cpuidle_cpu_online, NULL);
 277         WARN_ON(retval < 0);
 278         retval = cpuhp_setup_state_nocalls(CPUHP_CPUIDLE_DEAD,
 279                                            "cpuidle/pseries:DEAD", NULL,
 280                                            pseries_cpuidle_cpu_dead);
 281         WARN_ON(retval < 0);
 282         printk(KERN_DEBUG "pseries_idle_driver registered\n");
 283         return 0;
 284 }
 285
 286 device_initcall(pseries_processor_idle_init);