1 // SPDX-License-Identifier: GPL-2.0 2 /* 3 * cpuidle-pseries - idle state cpuidle driver. 4 * Adapted from drivers/idle/intel_idle.c and 5 * drivers/acpi/processor_idle.c 6 * 7 */ 8 9 #include <linux/kernel.h> 10 #include <linux/module.h> 11 #include <linux/init.h> 12 #include <linux/moduleparam.h> 13 #include <linux/cpuidle.h> 14 #include <linux/cpu.h> 15 #include <linux/notifier.h> 16 17 #include <asm/paca.h> 18 #include <asm/reg.h> 19 #include <asm/machdep.h> 20 #include <asm/firmware.h> 21 #include <asm/runlatch.h> 22 #include <asm/idle.h> 23 #include <asm/plpar_wrappers.h> 24 #include <asm/rtas.h> 25 26 static struct cpuidle_driver pseries_idle_driver = { 27 .name = "pseries_idle", 28 .owner = THIS_MODULE, 29 }; 30 31 static int max_idle_state __read_mostly; 32 static struct cpuidle_state *cpuidle_state_table __read_mostly; 33 static u64 snooze_timeout __read_mostly; 34 static bool snooze_timeout_en __read_mostly; 35 36 static int snooze_loop(struct cpuidle_device *dev, 37 struct cpuidle_driver *drv, 38 int index) 39 { 40 u64 snooze_exit_time; 41 42 set_thread_flag(TIF_POLLING_NRFLAG); 43 44 pseries_idle_prolog(); 45 local_irq_enable(); 46 snooze_exit_time = get_tb() + snooze_timeout; 47 48 while (!need_resched()) { 49 HMT_low(); 50 HMT_very_low(); 51 if (likely(snooze_timeout_en) && get_tb() > snooze_exit_time) { 52 /* 53 * Task has not woken up but we are exiting the polling 54 * loop anyway. Require a barrier after polling is 55 * cleared to order subsequent test of need_resched(). 56 */ 57 clear_thread_flag(TIF_POLLING_NRFLAG); 58 smp_mb(); 59 break; 60 } 61 } 62 63 HMT_medium(); 64 clear_thread_flag(TIF_POLLING_NRFLAG); 65 66 local_irq_disable(); 67 68 pseries_idle_epilog(); 69 70 return index; 71 } 72 73 static void check_and_cede_processor(void) 74 { 75 /* 76 * Ensure our interrupt state is properly tracked, 77 * also checks if no interrupt has occurred while we 78 * were soft-disabled 79 */ 80 if (prep_irq_for_idle()) { 81 cede_processor(); 82 #ifdef CONFIG_TRACE_IRQFLAGS 83 /* Ensure that H_CEDE returns with IRQs on */ 84 if (WARN_ON(!(mfmsr() & MSR_EE))) 85 __hard_irq_enable(); 86 #endif 87 } 88 } 89 90 /* 91 * XCEDE: Extended CEDE states discovered through the 92 * "ibm,get-systems-parameter" RTAS call with the token 93 * CEDE_LATENCY_TOKEN 94 */ 95 96 /* 97 * Section 7.3.16 System Parameters Option of PAPR version 2.8.1 has a 98 * table with all the parameters to ibm,get-system-parameters. 99 * CEDE_LATENCY_TOKEN corresponds to the token value for Cede Latency 100 * Settings Information. 101 */ 102 #define CEDE_LATENCY_TOKEN 45 103 104 /* 105 * If the platform supports the cede latency settings information system 106 * parameter it must provide the following information in the NULL terminated 107 * parameter string: 108 * 109 * a. The first byte is the length “N” of each cede latency setting record minus 110 * one (zero indicates a length of 1 byte). 111 * 112 * b. For each supported cede latency setting a cede latency setting record 113 * consisting of the first “N” bytes as per the following table. 114 * 115 * ----------------------------- 116 * | Field | Field | 117 * | Name | Length | 118 * ----------------------------- 119 * | Cede Latency | 1 Byte | 120 * | Specifier Value | | 121 * ----------------------------- 122 * | Maximum wakeup | | 123 * | latency in | 8 Bytes | 124 * | tb-ticks | | 125 * ----------------------------- 126 * | Responsive to | | 127 * | external | 1 Byte | 128 * | interrupts | | 129 * ----------------------------- 130 * 131 * This version has cede latency record size = 10. 132 * 133 * The structure xcede_latency_payload represents a) and b) with 134 * xcede_latency_record representing the table in b). 135 * 136 * xcede_latency_parameter is what gets returned by 137 * ibm,get-systems-parameter RTAS call when made with 138 * CEDE_LATENCY_TOKEN. 139 * 140 * These structures are only used to represent the data obtained by the RTAS 141 * call. The data is in big-endian. 142 */ 143 struct xcede_latency_record { 144 u8 hint; 145 __be64 latency_ticks; 146 u8 wake_on_irqs; 147 } __packed; 148 149 // Make space for 16 records, which "should be enough". 150 struct xcede_latency_payload { 151 u8 record_size; 152 struct xcede_latency_record records[16]; 153 } __packed; 154 155 struct xcede_latency_parameter { 156 __be16 payload_size; 157 struct xcede_latency_payload payload; 158 u8 null_char; 159 } __packed; 160 161 static unsigned int nr_xcede_records; 162 static struct xcede_latency_parameter xcede_latency_parameter __initdata; 163 164 static int __init parse_cede_parameters(void) 165 { 166 struct xcede_latency_payload *payload; 167 u32 total_xcede_records_size; 168 u8 xcede_record_size; 169 u16 payload_size; 170 int ret, i; 171 172 ret = rtas_call(rtas_token("ibm,get-system-parameter"), 3, 1, 173 NULL, CEDE_LATENCY_TOKEN, __pa(&xcede_latency_parameter), 174 sizeof(xcede_latency_parameter)); 175 if (ret) { 176 pr_err("xcede: Error parsing CEDE_LATENCY_TOKEN\n"); 177 return ret; 178 } 179 180 payload_size = be16_to_cpu(xcede_latency_parameter.payload_size); 181 payload = &xcede_latency_parameter.payload; 182 183 xcede_record_size = payload->record_size + 1; 184 185 if (xcede_record_size != sizeof(struct xcede_latency_record)) { 186 pr_err("xcede: Expected record-size %lu. Observed size %u.\n", 187 sizeof(struct xcede_latency_record), xcede_record_size); 188 return -EINVAL; 189 } 190 191 pr_info("xcede: xcede_record_size = %d\n", xcede_record_size); 192 193 /* 194 * Since the payload_size includes the last NULL byte and the 195 * xcede_record_size, the remaining bytes correspond to array of all 196 * cede_latency settings. 197 */ 198 total_xcede_records_size = payload_size - 2; 199 nr_xcede_records = total_xcede_records_size / xcede_record_size; 200 201 for (i = 0; i < nr_xcede_records; i++) { 202 struct xcede_latency_record *record = &payload->records[i]; 203 u64 latency_ticks = be64_to_cpu(record->latency_ticks); 204 u8 wake_on_irqs = record->wake_on_irqs; 205 u8 hint = record->hint; 206 207 pr_info("xcede: Record %d : hint = %u, latency = 0x%llx tb ticks, Wake-on-irq = %u\n", 208 i, hint, latency_ticks, wake_on_irqs); 209 } 210 211 return 0; 212 } 213 214 #define NR_DEDICATED_STATES 2 /* snooze, CEDE */ 215 static u8 cede_latency_hint[NR_DEDICATED_STATES]; 216 217 static int dedicated_cede_loop(struct cpuidle_device *dev, 218 struct cpuidle_driver *drv, 219 int index) 220 { 221 u8 old_latency_hint; 222 223 pseries_idle_prolog(); 224 get_lppaca()->donate_dedicated_cpu = 1; 225 old_latency_hint = get_lppaca()->cede_latency_hint; 226 get_lppaca()->cede_latency_hint = cede_latency_hint[index]; 227 228 HMT_medium(); 229 check_and_cede_processor(); 230 231 local_irq_disable(); 232 get_lppaca()->donate_dedicated_cpu = 0; 233 get_lppaca()->cede_latency_hint = old_latency_hint; 234 235 pseries_idle_epilog(); 236 237 return index; 238 } 239 240 static int shared_cede_loop(struct cpuidle_device *dev, 241 struct cpuidle_driver *drv, 242 int index) 243 { 244 245 pseries_idle_prolog(); 246 247 /* 248 * Yield the processor to the hypervisor. We return if 249 * an external interrupt occurs (which are driven prior 250 * to returning here) or if a prod occurs from another 251 * processor. When returning here, external interrupts 252 * are enabled. 253 */ 254 check_and_cede_processor(); 255 256 local_irq_disable(); 257 pseries_idle_epilog(); 258 259 return index; 260 } 261 262 /* 263 * States for dedicated partition case. 264 */ 265 static struct cpuidle_state dedicated_states[NR_DEDICATED_STATES] = { 266 { /* Snooze */ 267 .name = "snooze", 268 .desc = "snooze", 269 .exit_latency = 0, 270 .target_residency = 0, 271 .enter = &snooze_loop }, 272 { /* CEDE */ 273 .name = "CEDE", 274 .desc = "CEDE", 275 .exit_latency = 10, 276 .target_residency = 100, 277 .enter = &dedicated_cede_loop }, 278 }; 279 280 /* 281 * States for shared partition case. 282 */ 283 static struct cpuidle_state shared_states[] = { 284 { /* Snooze */ 285 .name = "snooze", 286 .desc = "snooze", 287 .exit_latency = 0, 288 .target_residency = 0, 289 .enter = &snooze_loop }, 290 { /* Shared Cede */ 291 .name = "Shared Cede", 292 .desc = "Shared Cede", 293 .exit_latency = 10, 294 .target_residency = 100, 295 .enter = &shared_cede_loop }, 296 }; 297 298 static int pseries_cpuidle_cpu_online(unsigned int cpu) 299 { 300 struct cpuidle_device *dev = per_cpu(cpuidle_devices, cpu); 301 302 if (dev && cpuidle_get_driver()) { 303 cpuidle_pause_and_lock(); 304 cpuidle_enable_device(dev); 305 cpuidle_resume_and_unlock(); 306 } 307 return 0; 308 } 309 310 static int pseries_cpuidle_cpu_dead(unsigned int cpu) 311 { 312 struct cpuidle_device *dev = per_cpu(cpuidle_devices, cpu); 313 314 if (dev && cpuidle_get_driver()) { 315 cpuidle_pause_and_lock(); 316 cpuidle_disable_device(dev); 317 cpuidle_resume_and_unlock(); 318 } 319 return 0; 320 } 321 322 /* 323 * pseries_cpuidle_driver_init() 324 */ 325 static int pseries_cpuidle_driver_init(void) 326 { 327 int idle_state; 328 struct cpuidle_driver *drv = &pseries_idle_driver; 329 330 drv->state_count = 0; 331 332 for (idle_state = 0; idle_state < max_idle_state; ++idle_state) { 333 /* Is the state not enabled? */ 334 if (cpuidle_state_table[idle_state].enter == NULL) 335 continue; 336 337 drv->states[drv->state_count] = /* structure copy */ 338 cpuidle_state_table[idle_state]; 339 340 drv->state_count += 1; 341 } 342 343 return 0; 344 } 345 346 static void __init fixup_cede0_latency(void) 347 { 348 struct xcede_latency_payload *payload; 349 u64 min_xcede_latency_us = UINT_MAX; 350 int i; 351 352 if (parse_cede_parameters()) 353 return; 354 355 pr_info("cpuidle: Skipping the %d Extended CEDE idle states\n", 356 nr_xcede_records); 357 358 payload = &xcede_latency_parameter.payload; 359 360 /* 361 * The CEDE idle state maps to CEDE(0). While the hypervisor 362 * does not advertise CEDE(0) exit latency values, it does 363 * advertise the latency values of the extended CEDE states. 364 * We use the lowest advertised exit latency value as a proxy 365 * for the exit latency of CEDE(0). 366 */ 367 for (i = 0; i < nr_xcede_records; i++) { 368 struct xcede_latency_record *record = &payload->records[i]; 369 u8 hint = record->hint; 370 u64 latency_tb = be64_to_cpu(record->latency_ticks); 371 u64 latency_us = DIV_ROUND_UP_ULL(tb_to_ns(latency_tb), NSEC_PER_USEC); 372 373 /* 374 * We expect the exit latency of an extended CEDE 375 * state to be non-zero, it to since it takes at least 376 * a few nanoseconds to wakeup the idle CPU and 377 * dispatch the virtual processor into the Linux 378 * Guest. 379 * 380 * So we consider only non-zero value for performing 381 * the fixup of CEDE(0) latency. 382 */ 383 if (latency_us == 0) { 384 pr_warn("cpuidle: Skipping xcede record %d [hint=%d]. Exit latency = 0us\n", 385 i, hint); 386 continue; 387 } 388 389 if (latency_us < min_xcede_latency_us) 390 min_xcede_latency_us = latency_us; 391 } 392 393 if (min_xcede_latency_us != UINT_MAX) { 394 dedicated_states[1].exit_latency = min_xcede_latency_us; 395 dedicated_states[1].target_residency = 10 * (min_xcede_latency_us); 396 pr_info("cpuidle: Fixed up CEDE exit latency to %llu us\n", 397 min_xcede_latency_us); 398 } 399 400 } 401 402 /* 403 * pseries_idle_probe() 404 * Choose state table for shared versus dedicated partition 405 */ 406 static int __init pseries_idle_probe(void) 407 { 408 409 if (cpuidle_disable != IDLE_NO_OVERRIDE) 410 return -ENODEV; 411 412 if (firmware_has_feature(FW_FEATURE_SPLPAR)) { 413 /* 414 * Use local_paca instead of get_lppaca() since 415 * preemption is not disabled, and it is not required in 416 * fact, since lppaca_ptr does not need to be the value 417 * associated to the current CPU, it can be from any CPU. 418 */ 419 if (lppaca_shared_proc(local_paca->lppaca_ptr)) { 420 cpuidle_state_table = shared_states; 421 max_idle_state = ARRAY_SIZE(shared_states); 422 } else { 423 /* 424 * Use firmware provided latency values 425 * starting with POWER10 platforms. In the 426 * case that we are running on a POWER10 427 * platform but in an earlier compat mode, we 428 * can still use the firmware provided values. 429 * 430 * However, on platforms prior to POWER10, we 431 * cannot rely on the accuracy of the firmware 432 * provided latency values. On such platforms, 433 * go with the conservative default estimate 434 * of 10us. 435 */ 436 if (cpu_has_feature(CPU_FTR_ARCH_31) || pvr_version_is(PVR_POWER10)) 437 fixup_cede0_latency(); 438 cpuidle_state_table = dedicated_states; 439 max_idle_state = NR_DEDICATED_STATES; 440 } 441 } else 442 return -ENODEV; 443 444 if (max_idle_state > 1) { 445 snooze_timeout_en = true; 446 snooze_timeout = cpuidle_state_table[1].target_residency * 447 tb_ticks_per_usec; 448 } 449 return 0; 450 } 451 452 static int __init pseries_processor_idle_init(void) 453 { 454 int retval; 455 456 retval = pseries_idle_probe(); 457 if (retval) 458 return retval; 459 460 pseries_cpuidle_driver_init(); 461 retval = cpuidle_register(&pseries_idle_driver, NULL); 462 if (retval) { 463 printk(KERN_DEBUG "Registration of pseries driver failed.\n"); 464 return retval; 465 } 466 467 retval = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN, 468 "cpuidle/pseries:online", 469 pseries_cpuidle_cpu_online, NULL); 470 WARN_ON(retval < 0); 471 retval = cpuhp_setup_state_nocalls(CPUHP_CPUIDLE_DEAD, 472 "cpuidle/pseries:DEAD", NULL, 473 pseries_cpuidle_cpu_dead); 474 WARN_ON(retval < 0); 475 printk(KERN_DEBUG "pseries_idle_driver registered\n"); 476 return 0; 477 } 478 479 device_initcall(pseries_processor_idle_init); 480