drivers/cpuidle/cpuidle-powernv.c

   1 /*
   2  *  cpuidle-powernv - idle state cpuidle driver.
   3  *  Adapted from drivers/cpuidle/cpuidle-pseries
   4  *
   5  */
   6
   7 #include <linux/kernel.h>
   8 #include <linux/module.h>
   9 #include <linux/init.h>
  10 #include <linux/moduleparam.h>
  11 #include <linux/cpuidle.h>
  12 #include <linux/cpu.h>
  13 #include <linux/notifier.h>
  14 #include <linux/clockchips.h>
  15 #include <linux/of.h>
  16 #include <linux/slab.h>
  17
  18 #include <asm/machdep.h>
  19 #include <asm/firmware.h>
  20 #include <asm/opal.h>
  21 #include <asm/runlatch.h>
  22
  23 #define POWERNV_THRESHOLD_LATENCY_NS 200000
  24
  25 struct cpuidle_driver powernv_idle_driver = {
  26         .name             = "powernv_idle",
  27         .owner            = THIS_MODULE,
  28 };
  29
  30 static int max_idle_state;
  31 static struct cpuidle_state *cpuidle_state_table;
  32
  33 static u64 stop_psscr_table[CPUIDLE_STATE_MAX];
  34
  35 static u64 default_snooze_timeout;
  36 static bool snooze_timeout_en;
  37
  38 static u64 get_snooze_timeout(struct cpuidle_device *dev,
  39                               struct cpuidle_driver *drv,
  40                               int index)
  41 {
  42         int i;
  43
  44         if (unlikely(!snooze_timeout_en))
  45                 return default_snooze_timeout;
  46
  47         for (i = index + 1; i < drv->state_count; i++) {
  48                 struct cpuidle_state *s = &drv->states[i];
  49                 struct cpuidle_state_usage *su = &dev->states_usage[i];
  50
  51                 if (s->disabled || su->disable)
  52                         continue;
  53
  54                 return s->target_residency * tb_ticks_per_usec;
  55         }
  56
  57         return default_snooze_timeout;
  58 }
  59
  60 static int snooze_loop(struct cpuidle_device *dev,
  61                         struct cpuidle_driver *drv,
  62                         int index)
  63 {
  64         u64 snooze_exit_time;
  65
  66         local_irq_enable();
  67         set_thread_flag(TIF_POLLING_NRFLAG);
  68
  69         snooze_exit_time = get_tb() + get_snooze_timeout(dev, drv, index);
  70         ppc64_runlatch_off();
  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         ppc64_runlatch_on();
  80         clear_thread_flag(TIF_POLLING_NRFLAG);
  81         smp_mb();
  82         return index;
  83 }
  84
  85 static int nap_loop(struct cpuidle_device *dev,
  86                         struct cpuidle_driver *drv,
  87                         int index)
  88 {
  89         ppc64_runlatch_off();
  90         power7_idle();
  91         ppc64_runlatch_on();
  92         return index;
  93 }
  94
  95 /* Register for fastsleep only in oneshot mode of broadcast */
  96 #ifdef CONFIG_TICK_ONESHOT
  97 static int fastsleep_loop(struct cpuidle_device *dev,
  98                                 struct cpuidle_driver *drv,
  99                                 int index)
 100 {
 101         unsigned long old_lpcr = mfspr(SPRN_LPCR);
 102         unsigned long new_lpcr;
 103
 104         if (unlikely(system_state < SYSTEM_RUNNING))
 105                 return index;
 106
 107         new_lpcr = old_lpcr;
 108         /* Do not exit powersave upon decrementer as we've setup the timer
 109          * offload.
 110          */
 111         new_lpcr &= ~LPCR_PECE1;
 112
 113         mtspr(SPRN_LPCR, new_lpcr);
 114         power7_sleep();
 115
 116         mtspr(SPRN_LPCR, old_lpcr);
 117
 118         return index;
 119 }
 120 #endif
 121
 122 static int stop_loop(struct cpuidle_device *dev,
 123                      struct cpuidle_driver *drv,
 124                      int index)
 125 {
 126         ppc64_runlatch_off();
 127         power9_idle_stop(stop_psscr_table[index]);
 128         ppc64_runlatch_on();
 129         return index;
 130 }
 131
 132 /*
 133  * States for dedicated partition case.
 134  */
 135 static struct cpuidle_state powernv_states[CPUIDLE_STATE_MAX] = {
 136         { /* Snooze */
 137                 .name = "snooze",
 138                 .desc = "snooze",
 139                 .exit_latency = 0,
 140                 .target_residency = 0,
 141                 .enter = snooze_loop },
 142 };
 143
 144 static int powernv_cpuidle_cpu_online(unsigned int cpu)
 145 {
 146         struct cpuidle_device *dev = per_cpu(cpuidle_devices, cpu);
 147
 148         if (dev && cpuidle_get_driver()) {
 149                 cpuidle_pause_and_lock();
 150                 cpuidle_enable_device(dev);
 151                 cpuidle_resume_and_unlock();
 152         }
 153         return 0;
 154 }
 155
 156 static int powernv_cpuidle_cpu_dead(unsigned int cpu)
 157 {
 158         struct cpuidle_device *dev = per_cpu(cpuidle_devices, cpu);
 159
 160         if (dev && cpuidle_get_driver()) {
 161                 cpuidle_pause_and_lock();
 162                 cpuidle_disable_device(dev);
 163                 cpuidle_resume_and_unlock();
 164         }
 165         return 0;
 166 }
 167
 168 /*
 169  * powernv_cpuidle_driver_init()
 170  */
 171 static int powernv_cpuidle_driver_init(void)
 172 {
 173         int idle_state;
 174         struct cpuidle_driver *drv = &powernv_idle_driver;
 175
 176         drv->state_count = 0;
 177
 178         for (idle_state = 0; idle_state < max_idle_state; ++idle_state) {
 179                 /* Is the state not enabled? */
 180                 if (cpuidle_state_table[idle_state].enter == NULL)
 181                         continue;
 182
 183                 drv->states[drv->state_count] = /* structure copy */
 184                         cpuidle_state_table[idle_state];
 185
 186                 drv->state_count += 1;
 187         }
 188
 189         /*
 190          * On the PowerNV platform cpu_present may be less than cpu_possible in
 191          * cases when firmware detects the CPU, but it is not available to the
 192          * OS.  If CONFIG_HOTPLUG_CPU=n, then such CPUs are not hotplugable at
 193          * run time and hence cpu_devices are not created for those CPUs by the
 194          * generic topology_init().
 195          *
 196          * drv->cpumask defaults to cpu_possible_mask in
 197          * __cpuidle_driver_init().  This breaks cpuidle on PowerNV where
 198          * cpu_devices are not created for CPUs in cpu_possible_mask that
 199          * cannot be hot-added later at run time.
 200          *
 201          * Trying cpuidle_register_device() on a CPU without a cpu_device is
 202          * incorrect, so pass a correct CPU mask to the generic cpuidle driver.
 203          */
 204
 205         drv->cpumask = (struct cpumask *)cpu_present_mask;
 206
 207         return 0;
 208 }
 209
 210 static int powernv_add_idle_states(void)
 211 {
 212         struct device_node *power_mgt;
 213         int nr_idle_states = 1; /* Snooze */
 214         int dt_idle_states;
 215         u32 latency_ns[CPUIDLE_STATE_MAX];
 216         u32 residency_ns[CPUIDLE_STATE_MAX];
 217         u32 flags[CPUIDLE_STATE_MAX];
 218         u64 psscr_val[CPUIDLE_STATE_MAX];
 219         const char *names[CPUIDLE_STATE_MAX];
 220         int i, rc;
 221
 222         /* Currently we have snooze statically defined */
 223
 224         power_mgt = of_find_node_by_path("/ibm,opal/power-mgt");
 225         if (!power_mgt) {
 226                 pr_warn("opal: PowerMgmt Node not found\n");
 227                 goto out;
 228         }
 229
 230         /* Read values of any property to determine the num of idle states */
 231         dt_idle_states = of_property_count_u32_elems(power_mgt, "ibm,cpu-idle-state-flags");
 232         if (dt_idle_states < 0) {
 233                 pr_warn("cpuidle-powernv: no idle states found in the DT\n");
 234                 goto out;
 235         }
 236
 237         /*
 238          * Since snooze is used as first idle state, max idle states allowed is
 239          * CPUIDLE_STATE_MAX -1
 240          */
 241         if (dt_idle_states > CPUIDLE_STATE_MAX - 1) {
 242                 pr_warn("cpuidle-powernv: discovered idle states more than allowed");
 243                 dt_idle_states = CPUIDLE_STATE_MAX - 1;
 244         }
 245
 246         if (of_property_read_u32_array(power_mgt,
 247                         "ibm,cpu-idle-state-flags", flags, dt_idle_states)) {
 248                 pr_warn("cpuidle-powernv : missing ibm,cpu-idle-state-flags in DT\n");
 249                 goto out;
 250         }
 251
 252         if (of_property_read_u32_array(power_mgt,
 253                 "ibm,cpu-idle-state-latencies-ns", latency_ns,
 254                 dt_idle_states)) {
 255                 pr_warn("cpuidle-powernv: missing ibm,cpu-idle-state-latencies-ns in DT\n");
 256                 goto out;
 257         }
 258         if (of_property_read_string_array(power_mgt,
 259                 "ibm,cpu-idle-state-names", names, dt_idle_states) < 0) {
 260                 pr_warn("cpuidle-powernv: missing ibm,cpu-idle-state-names in DT\n");
 261                 goto out;
 262         }
 263
 264         /*
 265          * If the idle states use stop instruction, probe for psscr values
 266          * which are necessary to specify required stop level.
 267          */
 268         if (flags[0] & (OPAL_PM_STOP_INST_FAST | OPAL_PM_STOP_INST_DEEP))
 269                 if (of_property_read_u64_array(power_mgt,
 270                     "ibm,cpu-idle-state-psscr", psscr_val, dt_idle_states)) {
 271                         pr_warn("cpuidle-powernv: missing ibm,cpu-idle-states-psscr in DT\n");
 272                         goto out;
 273                 }
 274
 275         rc = of_property_read_u32_array(power_mgt,
 276                 "ibm,cpu-idle-state-residency-ns", residency_ns, dt_idle_states);
 277
 278         for (i = 0; i < dt_idle_states; i++) {
 279                 /*
 280                  * If an idle state has exit latency beyond
 281                  * POWERNV_THRESHOLD_LATENCY_NS then don't use it
 282                  * in cpu-idle.
 283                  */
 284                 if (latency_ns[i] > POWERNV_THRESHOLD_LATENCY_NS)
 285                         continue;
 286
 287                 /*
 288                  * Cpuidle accepts exit_latency and target_residency in us.
 289                  * Use default target_residency values if f/w does not expose it.
 290                  */
 291                 if (flags[i] & OPAL_PM_NAP_ENABLED) {
 292                         /* Add NAP state */
 293                         strcpy(powernv_states[nr_idle_states].name, "Nap");
 294                         strcpy(powernv_states[nr_idle_states].desc, "Nap");
 295                         powernv_states[nr_idle_states].flags = 0;
 296                         powernv_states[nr_idle_states].target_residency = 100;
 297                         powernv_states[nr_idle_states].enter = nap_loop;
 298                 } else if ((flags[i] & OPAL_PM_STOP_INST_FAST) &&
 299                                 !(flags[i] & OPAL_PM_TIMEBASE_STOP)) {
 300                         strncpy(powernv_states[nr_idle_states].name,
 301                                 names[i], CPUIDLE_NAME_LEN);
 302                         strncpy(powernv_states[nr_idle_states].desc,
 303                                 names[i], CPUIDLE_NAME_LEN);
 304                         powernv_states[nr_idle_states].flags = 0;
 305
 306                         powernv_states[nr_idle_states].enter = stop_loop;
 307                         stop_psscr_table[nr_idle_states] = psscr_val[i];
 308                 }
 309
 310                 /*
 311                  * All cpuidle states with CPUIDLE_FLAG_TIMER_STOP set must come
 312                  * within this config dependency check.
 313                  */
 314 #ifdef CONFIG_TICK_ONESHOT
 315                 if (flags[i] & OPAL_PM_SLEEP_ENABLED ||
 316                         flags[i] & OPAL_PM_SLEEP_ENABLED_ER1) {
 317                         /* Add FASTSLEEP state */
 318                         strcpy(powernv_states[nr_idle_states].name, "FastSleep");
 319                         strcpy(powernv_states[nr_idle_states].desc, "FastSleep");
 320                         powernv_states[nr_idle_states].flags = CPUIDLE_FLAG_TIMER_STOP;
 321                         powernv_states[nr_idle_states].target_residency = 300000;
 322                         powernv_states[nr_idle_states].enter = fastsleep_loop;
 323                 } else if ((flags[i] & OPAL_PM_STOP_INST_DEEP) &&
 324                                 (flags[i] & OPAL_PM_TIMEBASE_STOP)) {
 325                         strncpy(powernv_states[nr_idle_states].name,
 326                                 names[i], CPUIDLE_NAME_LEN);
 327                         strncpy(powernv_states[nr_idle_states].desc,
 328                                 names[i], CPUIDLE_NAME_LEN);
 329
 330                         powernv_states[nr_idle_states].flags = CPUIDLE_FLAG_TIMER_STOP;
 331                         powernv_states[nr_idle_states].enter = stop_loop;
 332                         stop_psscr_table[nr_idle_states] = psscr_val[i];
 333                 }
 334 #endif
 335                 powernv_states[nr_idle_states].exit_latency =
 336                                 ((unsigned int)latency_ns[i]) / 1000;
 337
 338                 if (!rc) {
 339                         powernv_states[nr_idle_states].target_residency =
 340                                 ((unsigned int)residency_ns[i]) / 1000;
 341                 }
 342
 343                 nr_idle_states++;
 344         }
 345 out:
 346         return nr_idle_states;
 347 }
 348
 349 /*
 350  * powernv_idle_probe()
 351  * Choose state table for shared versus dedicated partition
 352  */
 353 static int powernv_idle_probe(void)
 354 {
 355         if (cpuidle_disable != IDLE_NO_OVERRIDE)
 356                 return -ENODEV;
 357
 358         if (firmware_has_feature(FW_FEATURE_OPAL)) {
 359                 cpuidle_state_table = powernv_states;
 360                 /* Device tree can indicate more idle states */
 361                 max_idle_state = powernv_add_idle_states();
 362                 default_snooze_timeout = TICK_USEC * tb_ticks_per_usec;
 363                 if (max_idle_state > 1)
 364                         snooze_timeout_en = true;
 365         } else
 366                 return -ENODEV;
 367
 368         return 0;
 369 }
 370
 371 static int __init powernv_processor_idle_init(void)
 372 {
 373         int retval;
 374
 375         retval = powernv_idle_probe();
 376         if (retval)
 377                 return retval;
 378
 379         powernv_cpuidle_driver_init();
 380         retval = cpuidle_register(&powernv_idle_driver, NULL);
 381         if (retval) {
 382                 printk(KERN_DEBUG "Registration of powernv driver failed.\n");
 383                 return retval;
 384         }
 385
 386         retval = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN,
 387                                            "cpuidle/powernv:online",
 388                                            powernv_cpuidle_cpu_online, NULL);
 389         WARN_ON(retval < 0);
 390         retval = cpuhp_setup_state_nocalls(CPUHP_CPUIDLE_DEAD,
 391                                            "cpuidle/powernv:dead", NULL,
 392                                            powernv_cpuidle_cpu_dead);
 393         WARN_ON(retval < 0);
 394         printk(KERN_DEBUG "powernv_idle_driver registered\n");
 395         return 0;
 396 }
 397
 398 device_initcall(powernv_processor_idle_init);