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 __cpuidle 37 int snooze_loop(struct cpuidle_device *dev, 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 raw_local_irq_enable(); 46 snooze_exit_time = get_tb() + snooze_timeout; 47 dev->poll_time_limit = false; 48 49 while (!need_resched()) { 50 HMT_low(); 51 HMT_very_low(); 52 if (likely(snooze_timeout_en) && get_tb() > snooze_exit_time) { 53 /* 54 * Task has not woken up but we are exiting the polling 55 * loop anyway. Require a barrier after polling is 56 * cleared to order subsequent test of need_resched(). 57 */ 58 dev->poll_time_limit = true; 59 clear_thread_flag(TIF_POLLING_NRFLAG); 60 smp_mb(); 61 break; 62 } 63 } 64 65 HMT_medium(); 66 clear_thread_flag(TIF_POLLING_NRFLAG); 67 68 raw_local_irq_disable(); 69 70 pseries_idle_epilog(); 71 72 return index; 73 } 74 75 static __cpuidle void check_and_cede_processor(void) 76 { 77 /* 78 * Ensure our interrupt state is properly tracked, 79 * also checks if no interrupt has occurred while we 80 * were soft-disabled 81 */ 82 if (prep_irq_for_idle()) { 83 cede_processor(); 84 #ifdef CONFIG_TRACE_IRQFLAGS 85 /* Ensure that H_CEDE returns with IRQs on */ 86 if (WARN_ON(!(mfmsr() & MSR_EE))) 87 __hard_irq_enable(); 88 #endif 89 } 90 } 91 92 /* 93 * XCEDE: Extended CEDE states discovered through the 94 * "ibm,get-systems-parameter" RTAS call with the token 95 * CEDE_LATENCY_TOKEN 96 */ 97 98 /* 99 * Section 7.3.16 System Parameters Option of PAPR version 2.8.1 has a 100 * table with all the parameters to ibm,get-system-parameters. 101 * CEDE_LATENCY_TOKEN corresponds to the token value for Cede Latency 102 * Settings Information. 103 */ 104 #define CEDE_LATENCY_TOKEN 45 105 106 /* 107 * If the platform supports the cede latency settings information system 108 * parameter it must provide the following information in the NULL terminated 109 * parameter string: 110 * 111 * a. The first byte is the length “N” of each cede latency setting record minus 112 * one (zero indicates a length of 1 byte). 113 * 114 * b. For each supported cede latency setting a cede latency setting record 115 * consisting of the first “N” bytes as per the following table. 116 * 117 * ----------------------------- 118 * | Field | Field | 119 * | Name | Length | 120 * ----------------------------- 121 * | Cede Latency | 1 Byte | 122 * | Specifier Value | | 123 * ----------------------------- 124 * | Maximum wakeup | | 125 * | latency in | 8 Bytes | 126 * | tb-ticks | | 127 * ----------------------------- 128 * | Responsive to | | 129 * | external | 1 Byte | 130 * | interrupts | | 131 * ----------------------------- 132 * 133 * This version has cede latency record size = 10. 134 * 135 * The structure xcede_latency_payload represents a) and b) with 136 * xcede_latency_record representing the table in b). 137 * 138 * xcede_latency_parameter is what gets returned by 139 * ibm,get-systems-parameter RTAS call when made with 140 * CEDE_LATENCY_TOKEN. 141 * 142 * These structures are only used to represent the data obtained by the RTAS 143 * call. The data is in big-endian. 144 */ 145 struct xcede_latency_record { 146 u8 hint; 147 __be64 latency_ticks; 148 u8 wake_on_irqs; 149 } __packed; 150 151 // Make space for 16 records, which "should be enough". 152 struct xcede_latency_payload { 153 u8 record_size; 154 struct xcede_latency_record records[16]; 155 } __packed; 156 157 struct xcede_latency_parameter { 158 __be16 payload_size; 159 struct xcede_latency_payload payload; 160 u8 null_char; 161 } __packed; 162 163 static unsigned int nr_xcede_records; 164 static struct xcede_latency_parameter xcede_latency_parameter __initdata; 165 166 static int __init parse_cede_parameters(void) 167 { 168 struct xcede_latency_payload *payload; 169 u32 total_xcede_records_size; 170 u8 xcede_record_size; 171 u16 payload_size; 172 int ret, i; 173 174 ret = rtas_call(rtas_token("ibm,get-system-parameter"), 3, 1, 175 NULL, CEDE_LATENCY_TOKEN, __pa(&xcede_latency_parameter), 176 sizeof(xcede_latency_parameter)); 177 if (ret) { 178 pr_err("xcede: Error parsing CEDE_LATENCY_TOKEN\n"); 179 return ret; 180 } 181 182 payload_size = be16_to_cpu(xcede_latency_parameter.payload_size); 183 payload = &xcede_latency_parameter.payload; 184 185 xcede_record_size = payload->record_size + 1; 186 187 if (xcede_record_size != sizeof(struct xcede_latency_record)) { 188 pr_err("xcede: Expected record-size %lu. Observed size %u.\n", 189 sizeof(struct xcede_latency_record), xcede_record_size); 190 return -EINVAL; 191 } 192 193 pr_info("xcede: xcede_record_size = %d\n", xcede_record_size); 194 195 /* 196 * Since the payload_size includes the last NULL byte and the 197 * xcede_record_size, the remaining bytes correspond to array of all 198 * cede_latency settings. 199 */ 200 total_xcede_records_size = payload_size - 2; 201 nr_xcede_records = total_xcede_records_size / xcede_record_size; 202 203 for (i = 0; i < nr_xcede_records; i++) { 204 struct xcede_latency_record *record = &payload->records[i]; 205 u64 latency_ticks = be64_to_cpu(record->latency_ticks); 206 u8 wake_on_irqs = record->wake_on_irqs; 207 u8 hint = record->hint; 208 209 pr_info("xcede: Record %d : hint = %u, latency = 0x%llx tb ticks, Wake-on-irq = %u\n", 210 i, hint, latency_ticks, wake_on_irqs); 211 } 212 213 return 0; 214 } 215 216 #define NR_DEDICATED_STATES 2 /* snooze, CEDE */ 217 static u8 cede_latency_hint[NR_DEDICATED_STATES]; 218 219 static __cpuidle 220 int dedicated_cede_loop(struct cpuidle_device *dev, struct cpuidle_driver *drv, 221 int index) 222 { 223 u8 old_latency_hint; 224 225 pseries_idle_prolog(); 226 get_lppaca()->donate_dedicated_cpu = 1; 227 old_latency_hint = get_lppaca()->cede_latency_hint; 228 get_lppaca()->cede_latency_hint = cede_latency_hint[index]; 229 230 HMT_medium(); 231 check_and_cede_processor(); 232 233 raw_local_irq_disable(); 234 get_lppaca()->donate_dedicated_cpu = 0; 235 get_lppaca()->cede_latency_hint = old_latency_hint; 236 237 pseries_idle_epilog(); 238 239 return index; 240 } 241 242 static __cpuidle 243 int shared_cede_loop(struct cpuidle_device *dev, struct cpuidle_driver *drv, 244 int index) 245 { 246 247 pseries_idle_prolog(); 248 249 /* 250 * Yield the processor to the hypervisor. We return if 251 * an external interrupt occurs (which are driven prior 252 * to returning here) or if a prod occurs from another 253 * processor. When returning here, external interrupts 254 * are enabled. 255 */ 256 check_and_cede_processor(); 257 258 raw_local_irq_disable(); 259 pseries_idle_epilog(); 260 261 return index; 262 } 263 264 /* 265 * States for dedicated partition case. 266 */ 267 static struct cpuidle_state dedicated_states[NR_DEDICATED_STATES] = { 268 { /* Snooze */ 269 .name = "snooze", 270 .desc = "snooze", 271 .exit_latency = 0, 272 .target_residency = 0, 273 .enter = &snooze_loop, 274 .flags = CPUIDLE_FLAG_POLLING }, 275 { /* CEDE */ 276 .name = "CEDE", 277 .desc = "CEDE", 278 .exit_latency = 10, 279 .target_residency = 100, 280 .enter = &dedicated_cede_loop }, 281 }; 282 283 /* 284 * States for shared partition case. 285 */ 286 static struct cpuidle_state shared_states[] = { 287 { /* Snooze */ 288 .name = "snooze", 289 .desc = "snooze", 290 .exit_latency = 0, 291 .target_residency = 0, 292 .enter = &snooze_loop, 293 .flags = CPUIDLE_FLAG_POLLING }, 294 { /* Shared Cede */ 295 .name = "Shared Cede", 296 .desc = "Shared Cede", 297 .exit_latency = 10, 298 .target_residency = 100, 299 .enter = &shared_cede_loop }, 300 }; 301 302 static int pseries_cpuidle_cpu_online(unsigned int cpu) 303 { 304 struct cpuidle_device *dev = per_cpu(cpuidle_devices, cpu); 305 306 if (dev && cpuidle_get_driver()) { 307 cpuidle_pause_and_lock(); 308 cpuidle_enable_device(dev); 309 cpuidle_resume_and_unlock(); 310 } 311 return 0; 312 } 313 314 static int pseries_cpuidle_cpu_dead(unsigned int cpu) 315 { 316 struct cpuidle_device *dev = per_cpu(cpuidle_devices, cpu); 317 318 if (dev && cpuidle_get_driver()) { 319 cpuidle_pause_and_lock(); 320 cpuidle_disable_device(dev); 321 cpuidle_resume_and_unlock(); 322 } 323 return 0; 324 } 325 326 /* 327 * pseries_cpuidle_driver_init() 328 */ 329 static int pseries_cpuidle_driver_init(void) 330 { 331 int idle_state; 332 struct cpuidle_driver *drv = &pseries_idle_driver; 333 334 drv->state_count = 0; 335 336 for (idle_state = 0; idle_state < max_idle_state; ++idle_state) { 337 /* Is the state not enabled? */ 338 if (cpuidle_state_table[idle_state].enter == NULL) 339 continue; 340 341 drv->states[drv->state_count] = /* structure copy */ 342 cpuidle_state_table[idle_state]; 343 344 drv->state_count += 1; 345 } 346 347 return 0; 348 } 349 350 static void __init fixup_cede0_latency(void) 351 { 352 struct xcede_latency_payload *payload; 353 u64 min_xcede_latency_us = UINT_MAX; 354 int i; 355 356 if (parse_cede_parameters()) 357 return; 358 359 pr_info("cpuidle: Skipping the %d Extended CEDE idle states\n", 360 nr_xcede_records); 361 362 payload = &xcede_latency_parameter.payload; 363 364 /* 365 * The CEDE idle state maps to CEDE(0). While the hypervisor 366 * does not advertise CEDE(0) exit latency values, it does 367 * advertise the latency values of the extended CEDE states. 368 * We use the lowest advertised exit latency value as a proxy 369 * for the exit latency of CEDE(0). 370 */ 371 for (i = 0; i < nr_xcede_records; i++) { 372 struct xcede_latency_record *record = &payload->records[i]; 373 u8 hint = record->hint; 374 u64 latency_tb = be64_to_cpu(record->latency_ticks); 375 u64 latency_us = DIV_ROUND_UP_ULL(tb_to_ns(latency_tb), NSEC_PER_USEC); 376 377 /* 378 * We expect the exit latency of an extended CEDE 379 * state to be non-zero, it to since it takes at least 380 * a few nanoseconds to wakeup the idle CPU and 381 * dispatch the virtual processor into the Linux 382 * Guest. 383 * 384 * So we consider only non-zero value for performing 385 * the fixup of CEDE(0) latency. 386 */ 387 if (latency_us == 0) { 388 pr_warn("cpuidle: Skipping xcede record %d [hint=%d]. Exit latency = 0us\n", 389 i, hint); 390 continue; 391 } 392 393 if (latency_us < min_xcede_latency_us) 394 min_xcede_latency_us = latency_us; 395 } 396 397 if (min_xcede_latency_us != UINT_MAX) { 398 dedicated_states[1].exit_latency = min_xcede_latency_us; 399 dedicated_states[1].target_residency = 10 * (min_xcede_latency_us); 400 pr_info("cpuidle: Fixed up CEDE exit latency to %llu us\n", 401 min_xcede_latency_us); 402 } 403 404 } 405 406 /* 407 * pseries_idle_probe() 408 * Choose state table for shared versus dedicated partition 409 */ 410 static int __init pseries_idle_probe(void) 411 { 412 413 if (cpuidle_disable != IDLE_NO_OVERRIDE) 414 return -ENODEV; 415 416 if (firmware_has_feature(FW_FEATURE_SPLPAR)) { 417 /* 418 * Use local_paca instead of get_lppaca() since 419 * preemption is not disabled, and it is not required in 420 * fact, since lppaca_ptr does not need to be the value 421 * associated to the current CPU, it can be from any CPU. 422 */ 423 if (lppaca_shared_proc(local_paca->lppaca_ptr)) { 424 cpuidle_state_table = shared_states; 425 max_idle_state = ARRAY_SIZE(shared_states); 426 } else { 427 /* 428 * Use firmware provided latency values 429 * starting with POWER10 platforms. In the 430 * case that we are running on a POWER10 431 * platform but in an earlier compat mode, we 432 * can still use the firmware provided values. 433 * 434 * However, on platforms prior to POWER10, we 435 * cannot rely on the accuracy of the firmware 436 * provided latency values. On such platforms, 437 * go with the conservative default estimate 438 * of 10us. 439 */ 440 if (cpu_has_feature(CPU_FTR_ARCH_31) || pvr_version_is(PVR_POWER10)) 441 fixup_cede0_latency(); 442 cpuidle_state_table = dedicated_states; 443 max_idle_state = NR_DEDICATED_STATES; 444 } 445 } else 446 return -ENODEV; 447 448 if (max_idle_state > 1) { 449 snooze_timeout_en = true; 450 snooze_timeout = cpuidle_state_table[1].target_residency * 451 tb_ticks_per_usec; 452 } 453 return 0; 454 } 455 456 static int __init pseries_processor_idle_init(void) 457 { 458 int retval; 459 460 retval = pseries_idle_probe(); 461 if (retval) 462 return retval; 463 464 pseries_cpuidle_driver_init(); 465 retval = cpuidle_register(&pseries_idle_driver, NULL); 466 if (retval) { 467 printk(KERN_DEBUG "Registration of pseries driver failed.\n"); 468 return retval; 469 } 470 471 retval = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN, 472 "cpuidle/pseries:online", 473 pseries_cpuidle_cpu_online, NULL); 474 WARN_ON(retval < 0); 475 retval = cpuhp_setup_state_nocalls(CPUHP_CPUIDLE_DEAD, 476 "cpuidle/pseries:DEAD", NULL, 477 pseries_cpuidle_cpu_dead); 478 WARN_ON(retval < 0); 479 printk(KERN_DEBUG "pseries_idle_driver registered\n"); 480 return 0; 481 } 482 483 device_initcall(pseries_processor_idle_init); 484