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 ppc64_runlatch_off(); 80 HMT_very_low(); 81 while (!need_resched()) { 82 if (likely(snooze_timeout_en) && get_tb() > snooze_exit_time) { 83 /* 84 * Task has not woken up but we are exiting the polling 85 * loop anyway. Require a barrier after polling is 86 * cleared to order subsequent test of need_resched(). 87 */ 88 clear_thread_flag(TIF_POLLING_NRFLAG); 89 smp_mb(); 90 break; 91 } 92 } 93 94 HMT_medium(); 95 ppc64_runlatch_on(); 96 clear_thread_flag(TIF_POLLING_NRFLAG); 97 98 local_irq_disable(); 99 100 return index; 101 } 102 103 static int nap_loop(struct cpuidle_device *dev, 104 struct cpuidle_driver *drv, 105 int index) 106 { 107 power7_idle_type(PNV_THREAD_NAP); 108 109 return index; 110 } 111 112 /* Register for fastsleep only in oneshot mode of broadcast */ 113 #ifdef CONFIG_TICK_ONESHOT 114 static int fastsleep_loop(struct cpuidle_device *dev, 115 struct cpuidle_driver *drv, 116 int index) 117 { 118 unsigned long old_lpcr = mfspr(SPRN_LPCR); 119 unsigned long new_lpcr; 120 121 if (unlikely(system_state < SYSTEM_RUNNING)) 122 return index; 123 124 new_lpcr = old_lpcr; 125 /* Do not exit powersave upon decrementer as we've setup the timer 126 * offload. 127 */ 128 new_lpcr &= ~LPCR_PECE1; 129 130 mtspr(SPRN_LPCR, new_lpcr); 131 132 power7_idle_type(PNV_THREAD_SLEEP); 133 134 mtspr(SPRN_LPCR, old_lpcr); 135 136 return index; 137 } 138 #endif 139 140 static int stop_loop(struct cpuidle_device *dev, 141 struct cpuidle_driver *drv, 142 int index) 143 { 144 power9_idle_type(stop_psscr_table[index].val, 145 stop_psscr_table[index].mask); 146 return index; 147 } 148 149 /* 150 * States for dedicated partition case. 151 */ 152 static struct cpuidle_state powernv_states[CPUIDLE_STATE_MAX] = { 153 { /* Snooze */ 154 .name = "snooze", 155 .desc = "snooze", 156 .exit_latency = 0, 157 .target_residency = 0, 158 .enter = snooze_loop }, 159 }; 160 161 static int powernv_cpuidle_cpu_online(unsigned int cpu) 162 { 163 struct cpuidle_device *dev = per_cpu(cpuidle_devices, cpu); 164 165 if (dev && cpuidle_get_driver()) { 166 cpuidle_pause_and_lock(); 167 cpuidle_enable_device(dev); 168 cpuidle_resume_and_unlock(); 169 } 170 return 0; 171 } 172 173 static int powernv_cpuidle_cpu_dead(unsigned int cpu) 174 { 175 struct cpuidle_device *dev = per_cpu(cpuidle_devices, cpu); 176 177 if (dev && cpuidle_get_driver()) { 178 cpuidle_pause_and_lock(); 179 cpuidle_disable_device(dev); 180 cpuidle_resume_and_unlock(); 181 } 182 return 0; 183 } 184 185 /* 186 * powernv_cpuidle_driver_init() 187 */ 188 static int powernv_cpuidle_driver_init(void) 189 { 190 int idle_state; 191 struct cpuidle_driver *drv = &powernv_idle_driver; 192 193 drv->state_count = 0; 194 195 for (idle_state = 0; idle_state < max_idle_state; ++idle_state) { 196 /* Is the state not enabled? */ 197 if (cpuidle_state_table[idle_state].enter == NULL) 198 continue; 199 200 drv->states[drv->state_count] = /* structure copy */ 201 cpuidle_state_table[idle_state]; 202 203 drv->state_count += 1; 204 } 205 206 /* 207 * On the PowerNV platform cpu_present may be less than cpu_possible in 208 * cases when firmware detects the CPU, but it is not available to the 209 * OS. If CONFIG_HOTPLUG_CPU=n, then such CPUs are not hotplugable at 210 * run time and hence cpu_devices are not created for those CPUs by the 211 * generic topology_init(). 212 * 213 * drv->cpumask defaults to cpu_possible_mask in 214 * __cpuidle_driver_init(). This breaks cpuidle on PowerNV where 215 * cpu_devices are not created for CPUs in cpu_possible_mask that 216 * cannot be hot-added later at run time. 217 * 218 * Trying cpuidle_register_device() on a CPU without a cpu_device is 219 * incorrect, so pass a correct CPU mask to the generic cpuidle driver. 220 */ 221 222 drv->cpumask = (struct cpumask *)cpu_present_mask; 223 224 return 0; 225 } 226 227 static inline void add_powernv_state(int index, const char *name, 228 unsigned int flags, 229 int (*idle_fn)(struct cpuidle_device *, 230 struct cpuidle_driver *, 231 int), 232 unsigned int target_residency, 233 unsigned int exit_latency, 234 u64 psscr_val, u64 psscr_mask) 235 { 236 strlcpy(powernv_states[index].name, name, CPUIDLE_NAME_LEN); 237 strlcpy(powernv_states[index].desc, name, CPUIDLE_NAME_LEN); 238 powernv_states[index].flags = flags; 239 powernv_states[index].target_residency = target_residency; 240 powernv_states[index].exit_latency = exit_latency; 241 powernv_states[index].enter = idle_fn; 242 /* For power8 and below psscr_* will be 0 */ 243 stop_psscr_table[index].val = psscr_val; 244 stop_psscr_table[index].mask = psscr_mask; 245 } 246 247 extern u32 pnv_get_supported_cpuidle_states(void); 248 static int powernv_add_idle_states(void) 249 { 250 int nr_idle_states = 1; /* Snooze */ 251 int dt_idle_states; 252 u32 has_stop_states = 0; 253 int i; 254 u32 supported_flags = pnv_get_supported_cpuidle_states(); 255 256 257 /* Currently we have snooze statically defined */ 258 if (nr_pnv_idle_states <= 0) { 259 pr_warn("cpuidle-powernv : Only Snooze is available\n"); 260 goto out; 261 } 262 263 /* TODO: Count only states which are eligible for cpuidle */ 264 dt_idle_states = nr_pnv_idle_states; 265 266 /* 267 * Since snooze is used as first idle state, max idle states allowed is 268 * CPUIDLE_STATE_MAX -1 269 */ 270 if (nr_pnv_idle_states > CPUIDLE_STATE_MAX - 1) { 271 pr_warn("cpuidle-powernv: discovered idle states more than allowed"); 272 dt_idle_states = CPUIDLE_STATE_MAX - 1; 273 } 274 275 /* 276 * If the idle states use stop instruction, probe for psscr values 277 * and psscr mask which are necessary to specify required stop level. 278 */ 279 has_stop_states = (pnv_idle_states[0].flags & 280 (OPAL_PM_STOP_INST_FAST | OPAL_PM_STOP_INST_DEEP)); 281 282 for (i = 0; i < dt_idle_states; i++) { 283 unsigned int exit_latency, target_residency; 284 bool stops_timebase = false; 285 struct pnv_idle_states_t *state = &pnv_idle_states[i]; 286 287 /* 288 * Skip the platform idle state whose flag isn't in 289 * the supported_cpuidle_states flag mask. 290 */ 291 if ((state->flags & supported_flags) != state->flags) 292 continue; 293 /* 294 * If an idle state has exit latency beyond 295 * POWERNV_THRESHOLD_LATENCY_NS then don't use it 296 * in cpu-idle. 297 */ 298 if (state->latency_ns > POWERNV_THRESHOLD_LATENCY_NS) 299 continue; 300 /* 301 * Firmware passes residency and latency values in ns. 302 * cpuidle expects it in us. 303 */ 304 exit_latency = DIV_ROUND_UP(state->latency_ns, 1000); 305 target_residency = DIV_ROUND_UP(state->residency_ns, 1000); 306 307 if (has_stop_states && !(state->valid)) 308 continue; 309 310 if (state->flags & OPAL_PM_TIMEBASE_STOP) 311 stops_timebase = true; 312 313 if (state->flags & OPAL_PM_NAP_ENABLED) { 314 /* Add NAP state */ 315 add_powernv_state(nr_idle_states, "Nap", 316 CPUIDLE_FLAG_NONE, nap_loop, 317 target_residency, exit_latency, 0, 0); 318 } else if (has_stop_states && !stops_timebase) { 319 add_powernv_state(nr_idle_states, state->name, 320 CPUIDLE_FLAG_NONE, stop_loop, 321 target_residency, exit_latency, 322 state->psscr_val, 323 state->psscr_mask); 324 } 325 326 /* 327 * All cpuidle states with CPUIDLE_FLAG_TIMER_STOP set must come 328 * within this config dependency check. 329 */ 330 #ifdef CONFIG_TICK_ONESHOT 331 else if (state->flags & OPAL_PM_SLEEP_ENABLED || 332 state->flags & OPAL_PM_SLEEP_ENABLED_ER1) { 333 /* Add FASTSLEEP state */ 334 add_powernv_state(nr_idle_states, "FastSleep", 335 CPUIDLE_FLAG_TIMER_STOP, 336 fastsleep_loop, 337 target_residency, exit_latency, 0, 0); 338 } else if (has_stop_states && stops_timebase) { 339 add_powernv_state(nr_idle_states, state->name, 340 CPUIDLE_FLAG_TIMER_STOP, stop_loop, 341 target_residency, exit_latency, 342 state->psscr_val, 343 state->psscr_mask); 344 } 345 #endif 346 else 347 continue; 348 nr_idle_states++; 349 } 350 out: 351 return nr_idle_states; 352 } 353 354 /* 355 * powernv_idle_probe() 356 * Choose state table for shared versus dedicated partition 357 */ 358 static int powernv_idle_probe(void) 359 { 360 if (cpuidle_disable != IDLE_NO_OVERRIDE) 361 return -ENODEV; 362 363 if (firmware_has_feature(FW_FEATURE_OPAL)) { 364 cpuidle_state_table = powernv_states; 365 /* Device tree can indicate more idle states */ 366 max_idle_state = powernv_add_idle_states(); 367 default_snooze_timeout = TICK_USEC * tb_ticks_per_usec; 368 if (max_idle_state > 1) 369 snooze_timeout_en = true; 370 } else 371 return -ENODEV; 372 373 return 0; 374 } 375 376 static int __init powernv_processor_idle_init(void) 377 { 378 int retval; 379 380 retval = powernv_idle_probe(); 381 if (retval) 382 return retval; 383 384 powernv_cpuidle_driver_init(); 385 retval = cpuidle_register(&powernv_idle_driver, NULL); 386 if (retval) { 387 printk(KERN_DEBUG "Registration of powernv driver failed.\n"); 388 return retval; 389 } 390 391 retval = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN, 392 "cpuidle/powernv:online", 393 powernv_cpuidle_cpu_online, NULL); 394 WARN_ON(retval < 0); 395 retval = cpuhp_setup_state_nocalls(CPUHP_CPUIDLE_DEAD, 396 "cpuidle/powernv:dead", NULL, 397 powernv_cpuidle_cpu_dead); 398 WARN_ON(retval < 0); 399 printk(KERN_DEBUG "powernv_idle_driver registered\n"); 400 return 0; 401 } 402 403 device_initcall(powernv_processor_idle_init); 404