1 // SPDX-License-Identifier: GPL-2.0 2 /* 3 * cpuidle-powernv - idle state cpuidle driver. 4 * Adapted from drivers/cpuidle/cpuidle-pseries 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 #include <linux/clockchips.h> 16 #include <linux/of.h> 17 #include <linux/slab.h> 18 19 #include <asm/machdep.h> 20 #include <asm/firmware.h> 21 #include <asm/opal.h> 22 #include <asm/runlatch.h> 23 #include <asm/cpuidle.h> 24 25 /* 26 * Expose only those Hardware idle states via the cpuidle framework 27 * that have latency value below POWERNV_THRESHOLD_LATENCY_NS. 28 */ 29 #define POWERNV_THRESHOLD_LATENCY_NS 200000 30 31 static struct cpuidle_driver powernv_idle_driver = { 32 .name = "powernv_idle", 33 .owner = THIS_MODULE, 34 }; 35 36 static int max_idle_state __read_mostly; 37 static struct cpuidle_state *cpuidle_state_table __read_mostly; 38 39 struct stop_psscr_table { 40 u64 val; 41 u64 mask; 42 }; 43 44 static struct stop_psscr_table stop_psscr_table[CPUIDLE_STATE_MAX] __read_mostly; 45 46 static u64 default_snooze_timeout __read_mostly; 47 static bool snooze_timeout_en __read_mostly; 48 49 static u64 get_snooze_timeout(struct cpuidle_device *dev, 50 struct cpuidle_driver *drv, 51 int index) 52 { 53 int i; 54 55 if (unlikely(!snooze_timeout_en)) 56 return default_snooze_timeout; 57 58 for (i = index + 1; i < drv->state_count; i++) { 59 if (dev->states_usage[i].disable) 60 continue; 61 62 return drv->states[i].target_residency * tb_ticks_per_usec; 63 } 64 65 return default_snooze_timeout; 66 } 67 68 static int snooze_loop(struct cpuidle_device *dev, 69 struct cpuidle_driver *drv, 70 int index) 71 { 72 u64 snooze_exit_time; 73 74 set_thread_flag(TIF_POLLING_NRFLAG); 75 76 local_irq_enable(); 77 78 snooze_exit_time = get_tb() + get_snooze_timeout(dev, drv, index); 79 dev->poll_time_limit = false; 80 ppc64_runlatch_off(); 81 HMT_very_low(); 82 while (!need_resched()) { 83 if (likely(snooze_timeout_en) && get_tb() > snooze_exit_time) { 84 /* 85 * Task has not woken up but we are exiting the polling 86 * loop anyway. Require a barrier after polling is 87 * cleared to order subsequent test of need_resched(). 88 */ 89 clear_thread_flag(TIF_POLLING_NRFLAG); 90 dev->poll_time_limit = true; 91 smp_mb(); 92 break; 93 } 94 } 95 96 HMT_medium(); 97 ppc64_runlatch_on(); 98 99 /* Avoid double clear when breaking */ 100 if (!dev->poll_time_limit) 101 clear_thread_flag(TIF_POLLING_NRFLAG); 102 103 local_irq_disable(); 104 105 return index; 106 } 107 108 static int nap_loop(struct cpuidle_device *dev, 109 struct cpuidle_driver *drv, 110 int index) 111 { 112 power7_idle_type(PNV_THREAD_NAP); 113 114 return index; 115 } 116 117 /* Register for fastsleep only in oneshot mode of broadcast */ 118 #ifdef CONFIG_TICK_ONESHOT 119 static int fastsleep_loop(struct cpuidle_device *dev, 120 struct cpuidle_driver *drv, 121 int index) 122 { 123 unsigned long old_lpcr = mfspr(SPRN_LPCR); 124 unsigned long new_lpcr; 125 126 if (unlikely(system_state < SYSTEM_RUNNING)) 127 return index; 128 129 new_lpcr = old_lpcr; 130 /* Do not exit powersave upon decrementer as we've setup the timer 131 * offload. 132 */ 133 new_lpcr &= ~LPCR_PECE1; 134 135 mtspr(SPRN_LPCR, new_lpcr); 136 137 power7_idle_type(PNV_THREAD_SLEEP); 138 139 mtspr(SPRN_LPCR, old_lpcr); 140 141 return index; 142 } 143 #endif 144 145 static int stop_loop(struct cpuidle_device *dev, 146 struct cpuidle_driver *drv, 147 int index) 148 { 149 arch300_idle_type(stop_psscr_table[index].val, 150 stop_psscr_table[index].mask); 151 return index; 152 } 153 154 /* 155 * States for dedicated partition case. 156 */ 157 static struct cpuidle_state powernv_states[CPUIDLE_STATE_MAX] = { 158 { /* Snooze */ 159 .name = "snooze", 160 .desc = "snooze", 161 .exit_latency = 0, 162 .target_residency = 0, 163 .enter = snooze_loop, 164 .flags = CPUIDLE_FLAG_POLLING }, 165 }; 166 167 static int powernv_cpuidle_cpu_online(unsigned int cpu) 168 { 169 struct cpuidle_device *dev = per_cpu(cpuidle_devices, cpu); 170 171 if (dev && cpuidle_get_driver()) { 172 cpuidle_pause_and_lock(); 173 cpuidle_enable_device(dev); 174 cpuidle_resume_and_unlock(); 175 } 176 return 0; 177 } 178 179 static int powernv_cpuidle_cpu_dead(unsigned int cpu) 180 { 181 struct cpuidle_device *dev = per_cpu(cpuidle_devices, cpu); 182 183 if (dev && cpuidle_get_driver()) { 184 cpuidle_pause_and_lock(); 185 cpuidle_disable_device(dev); 186 cpuidle_resume_and_unlock(); 187 } 188 return 0; 189 } 190 191 /* 192 * powernv_cpuidle_driver_init() 193 */ 194 static int powernv_cpuidle_driver_init(void) 195 { 196 int idle_state; 197 struct cpuidle_driver *drv = &powernv_idle_driver; 198 199 drv->state_count = 0; 200 201 for (idle_state = 0; idle_state < max_idle_state; ++idle_state) { 202 /* Is the state not enabled? */ 203 if (cpuidle_state_table[idle_state].enter == NULL) 204 continue; 205 206 drv->states[drv->state_count] = /* structure copy */ 207 cpuidle_state_table[idle_state]; 208 209 drv->state_count += 1; 210 } 211 212 /* 213 * On the PowerNV platform cpu_present may be less than cpu_possible in 214 * cases when firmware detects the CPU, but it is not available to the 215 * OS. If CONFIG_HOTPLUG_CPU=n, then such CPUs are not hotplugable at 216 * run time and hence cpu_devices are not created for those CPUs by the 217 * generic topology_init(). 218 * 219 * drv->cpumask defaults to cpu_possible_mask in 220 * __cpuidle_driver_init(). This breaks cpuidle on PowerNV where 221 * cpu_devices are not created for CPUs in cpu_possible_mask that 222 * cannot be hot-added later at run time. 223 * 224 * Trying cpuidle_register_device() on a CPU without a cpu_device is 225 * incorrect, so pass a correct CPU mask to the generic cpuidle driver. 226 */ 227 228 drv->cpumask = (struct cpumask *)cpu_present_mask; 229 230 return 0; 231 } 232 233 static inline void add_powernv_state(int index, const char *name, 234 unsigned int flags, 235 int (*idle_fn)(struct cpuidle_device *, 236 struct cpuidle_driver *, 237 int), 238 unsigned int target_residency, 239 unsigned int exit_latency, 240 u64 psscr_val, u64 psscr_mask) 241 { 242 strscpy(powernv_states[index].name, name, CPUIDLE_NAME_LEN); 243 strscpy(powernv_states[index].desc, name, CPUIDLE_NAME_LEN); 244 powernv_states[index].flags = flags; 245 powernv_states[index].target_residency = target_residency; 246 powernv_states[index].exit_latency = exit_latency; 247 powernv_states[index].enter = idle_fn; 248 /* For power8 and below psscr_* will be 0 */ 249 stop_psscr_table[index].val = psscr_val; 250 stop_psscr_table[index].mask = psscr_mask; 251 } 252 253 extern u32 pnv_get_supported_cpuidle_states(void); 254 static int powernv_add_idle_states(void) 255 { 256 int nr_idle_states = 1; /* Snooze */ 257 int dt_idle_states; 258 u32 has_stop_states = 0; 259 int i; 260 u32 supported_flags = pnv_get_supported_cpuidle_states(); 261 262 263 /* Currently we have snooze statically defined */ 264 if (nr_pnv_idle_states <= 0) { 265 pr_warn("cpuidle-powernv : Only Snooze is available\n"); 266 goto out; 267 } 268 269 /* TODO: Count only states which are eligible for cpuidle */ 270 dt_idle_states = nr_pnv_idle_states; 271 272 /* 273 * Since snooze is used as first idle state, max idle states allowed is 274 * CPUIDLE_STATE_MAX -1 275 */ 276 if (nr_pnv_idle_states > CPUIDLE_STATE_MAX - 1) { 277 pr_warn("cpuidle-powernv: discovered idle states more than allowed"); 278 dt_idle_states = CPUIDLE_STATE_MAX - 1; 279 } 280 281 /* 282 * If the idle states use stop instruction, probe for psscr values 283 * and psscr mask which are necessary to specify required stop level. 284 */ 285 has_stop_states = (pnv_idle_states[0].flags & 286 (OPAL_PM_STOP_INST_FAST | OPAL_PM_STOP_INST_DEEP)); 287 288 for (i = 0; i < dt_idle_states; i++) { 289 unsigned int exit_latency, target_residency; 290 bool stops_timebase = false; 291 struct pnv_idle_states_t *state = &pnv_idle_states[i]; 292 293 /* 294 * Skip the platform idle state whose flag isn't in 295 * the supported_cpuidle_states flag mask. 296 */ 297 if ((state->flags & supported_flags) != state->flags) 298 continue; 299 /* 300 * If an idle state has exit latency beyond 301 * POWERNV_THRESHOLD_LATENCY_NS then don't use it 302 * in cpu-idle. 303 */ 304 if (state->latency_ns > POWERNV_THRESHOLD_LATENCY_NS) 305 continue; 306 /* 307 * Firmware passes residency and latency values in ns. 308 * cpuidle expects it in us. 309 */ 310 exit_latency = DIV_ROUND_UP(state->latency_ns, 1000); 311 target_residency = DIV_ROUND_UP(state->residency_ns, 1000); 312 313 if (has_stop_states && !(state->valid)) 314 continue; 315 316 if (state->flags & OPAL_PM_TIMEBASE_STOP) 317 stops_timebase = true; 318 319 if (state->flags & OPAL_PM_NAP_ENABLED) { 320 /* Add NAP state */ 321 add_powernv_state(nr_idle_states, "Nap", 322 CPUIDLE_FLAG_NONE, nap_loop, 323 target_residency, exit_latency, 0, 0); 324 } else if (has_stop_states && !stops_timebase) { 325 add_powernv_state(nr_idle_states, state->name, 326 CPUIDLE_FLAG_NONE, stop_loop, 327 target_residency, exit_latency, 328 state->psscr_val, 329 state->psscr_mask); 330 } 331 332 /* 333 * All cpuidle states with CPUIDLE_FLAG_TIMER_STOP set must come 334 * within this config dependency check. 335 */ 336 #ifdef CONFIG_TICK_ONESHOT 337 else if (state->flags & OPAL_PM_SLEEP_ENABLED || 338 state->flags & OPAL_PM_SLEEP_ENABLED_ER1) { 339 /* Add FASTSLEEP state */ 340 add_powernv_state(nr_idle_states, "FastSleep", 341 CPUIDLE_FLAG_TIMER_STOP, 342 fastsleep_loop, 343 target_residency, exit_latency, 0, 0); 344 } else if (has_stop_states && stops_timebase) { 345 add_powernv_state(nr_idle_states, state->name, 346 CPUIDLE_FLAG_TIMER_STOP, stop_loop, 347 target_residency, exit_latency, 348 state->psscr_val, 349 state->psscr_mask); 350 } 351 #endif 352 else 353 continue; 354 nr_idle_states++; 355 } 356 out: 357 return nr_idle_states; 358 } 359 360 /* 361 * powernv_idle_probe() 362 * Choose state table for shared versus dedicated partition 363 */ 364 static int powernv_idle_probe(void) 365 { 366 if (cpuidle_disable != IDLE_NO_OVERRIDE) 367 return -ENODEV; 368 369 if (firmware_has_feature(FW_FEATURE_OPAL)) { 370 cpuidle_state_table = powernv_states; 371 /* Device tree can indicate more idle states */ 372 max_idle_state = powernv_add_idle_states(); 373 default_snooze_timeout = TICK_USEC * tb_ticks_per_usec; 374 if (max_idle_state > 1) 375 snooze_timeout_en = true; 376 } else 377 return -ENODEV; 378 379 return 0; 380 } 381 382 static int __init powernv_processor_idle_init(void) 383 { 384 int retval; 385 386 retval = powernv_idle_probe(); 387 if (retval) 388 return retval; 389 390 powernv_cpuidle_driver_init(); 391 retval = cpuidle_register(&powernv_idle_driver, NULL); 392 if (retval) { 393 printk(KERN_DEBUG "Registration of powernv driver failed.\n"); 394 return retval; 395 } 396 397 retval = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN, 398 "cpuidle/powernv:online", 399 powernv_cpuidle_cpu_online, NULL); 400 WARN_ON(retval < 0); 401 retval = cpuhp_setup_state_nocalls(CPUHP_CPUIDLE_DEAD, 402 "cpuidle/powernv:dead", NULL, 403 powernv_cpuidle_cpu_dead); 404 WARN_ON(retval < 0); 405 printk(KERN_DEBUG "powernv_idle_driver registered\n"); 406 return 0; 407 } 408 409 device_initcall(powernv_processor_idle_init); 410