1// SPDX-License-Identifier: GPL-2.0-only 2/* 3 * acpi_pad.c ACPI Processor Aggregator Driver 4 * 5 * Copyright (c) 2009, Intel Corporation. 6 */ 7 8#include <linux/kernel.h> 9#include <linux/cpumask.h> 10#include <linux/module.h> 11#include <linux/init.h> 12#include <linux/types.h> 13#include <linux/kthread.h> 14#include <uapi/linux/sched/types.h> 15#include <linux/freezer.h> 16#include <linux/cpu.h> 17#include <linux/tick.h> 18#include <linux/slab.h> 19#include <linux/acpi.h> 20#include <linux/perf_event.h> 21#include <asm/mwait.h> 22#include <xen/xen.h> 23 24#define ACPI_PROCESSOR_AGGREGATOR_CLASS "acpi_pad" 25#define ACPI_PROCESSOR_AGGREGATOR_DEVICE_NAME "Processor Aggregator" 26#define ACPI_PROCESSOR_AGGREGATOR_NOTIFY 0x80 27static DEFINE_MUTEX(isolated_cpus_lock); 28static DEFINE_MUTEX(round_robin_lock); 29 30static unsigned long power_saving_mwait_eax; 31 32static unsigned char tsc_detected_unstable; 33static unsigned char tsc_marked_unstable; 34 35static void power_saving_mwait_init(void) 36{ 37 unsigned int eax, ebx, ecx, edx; 38 unsigned int highest_cstate = 0; 39 unsigned int highest_subcstate = 0; 40 int i; 41 42 if (!boot_cpu_has(X86_FEATURE_MWAIT)) 43 return; 44 if (boot_cpu_data.cpuid_level < CPUID_MWAIT_LEAF) 45 return; 46 47 cpuid(CPUID_MWAIT_LEAF, &eax, &ebx, &ecx, &edx); 48 49 if (!(ecx & CPUID5_ECX_EXTENSIONS_SUPPORTED) || 50 !(ecx & CPUID5_ECX_INTERRUPT_BREAK)) 51 return; 52 53 edx >>= MWAIT_SUBSTATE_SIZE; 54 for (i = 0; i < 7 && edx; i++, edx >>= MWAIT_SUBSTATE_SIZE) { 55 if (edx & MWAIT_SUBSTATE_MASK) { 56 highest_cstate = i; 57 highest_subcstate = edx & MWAIT_SUBSTATE_MASK; 58 } 59 } 60 power_saving_mwait_eax = (highest_cstate << MWAIT_SUBSTATE_SIZE) | 61 (highest_subcstate - 1); 62 63#if defined(CONFIG_X86) 64 switch (boot_cpu_data.x86_vendor) { 65 case X86_VENDOR_HYGON: 66 case X86_VENDOR_AMD: 67 case X86_VENDOR_INTEL: 68 case X86_VENDOR_ZHAOXIN: 69 case X86_VENDOR_CENTAUR: 70 /* 71 * AMD Fam10h TSC will tick in all 72 * C/P/S0/S1 states when this bit is set. 73 */ 74 if (!boot_cpu_has(X86_FEATURE_NONSTOP_TSC)) 75 tsc_detected_unstable = 1; 76 break; 77 default: 78 /* TSC could halt in idle */ 79 tsc_detected_unstable = 1; 80 } 81#endif 82} 83 84static unsigned long cpu_weight[NR_CPUS]; 85static int tsk_in_cpu[NR_CPUS] = {[0 ... NR_CPUS-1] = -1}; 86static DECLARE_BITMAP(pad_busy_cpus_bits, NR_CPUS); 87static void round_robin_cpu(unsigned int tsk_index) 88{ 89 struct cpumask *pad_busy_cpus = to_cpumask(pad_busy_cpus_bits); 90 cpumask_var_t tmp; 91 int cpu; 92 unsigned long min_weight = -1; 93 unsigned long preferred_cpu; 94 95 if (!alloc_cpumask_var(&tmp, GFP_KERNEL)) 96 return; 97 98 mutex_lock(&round_robin_lock); 99 cpumask_clear(tmp); 100 for_each_cpu(cpu, pad_busy_cpus) 101 cpumask_or(tmp, tmp, topology_sibling_cpumask(cpu)); 102 cpumask_andnot(tmp, cpu_online_mask, tmp); 103 /* avoid HT sibilings if possible */ 104 if (cpumask_empty(tmp)) 105 cpumask_andnot(tmp, cpu_online_mask, pad_busy_cpus); 106 if (cpumask_empty(tmp)) { 107 mutex_unlock(&round_robin_lock); 108 free_cpumask_var(tmp); 109 return; 110 } 111 for_each_cpu(cpu, tmp) { 112 if (cpu_weight[cpu] < min_weight) { 113 min_weight = cpu_weight[cpu]; 114 preferred_cpu = cpu; 115 } 116 } 117 118 if (tsk_in_cpu[tsk_index] != -1) 119 cpumask_clear_cpu(tsk_in_cpu[tsk_index], pad_busy_cpus); 120 tsk_in_cpu[tsk_index] = preferred_cpu; 121 cpumask_set_cpu(preferred_cpu, pad_busy_cpus); 122 cpu_weight[preferred_cpu]++; 123 mutex_unlock(&round_robin_lock); 124 125 set_cpus_allowed_ptr(current, cpumask_of(preferred_cpu)); 126 127 free_cpumask_var(tmp); 128} 129 130static void exit_round_robin(unsigned int tsk_index) 131{ 132 struct cpumask *pad_busy_cpus = to_cpumask(pad_busy_cpus_bits); 133 134 cpumask_clear_cpu(tsk_in_cpu[tsk_index], pad_busy_cpus); 135 tsk_in_cpu[tsk_index] = -1; 136} 137 138static unsigned int idle_pct = 5; /* percentage */ 139static unsigned int round_robin_time = 1; /* second */ 140static int power_saving_thread(void *data) 141{ 142 int do_sleep; 143 unsigned int tsk_index = (unsigned long)data; 144 u64 last_jiffies = 0; 145 146 sched_set_fifo_low(current); 147 148 while (!kthread_should_stop()) { 149 unsigned long expire_time; 150 151 /* round robin to cpus */ 152 expire_time = last_jiffies + round_robin_time * HZ; 153 if (time_before(expire_time, jiffies)) { 154 last_jiffies = jiffies; 155 round_robin_cpu(tsk_index); 156 } 157 158 do_sleep = 0; 159 160 expire_time = jiffies + HZ * (100 - idle_pct) / 100; 161 162 while (!need_resched()) { 163 if (tsc_detected_unstable && !tsc_marked_unstable) { 164 /* TSC could halt in idle, so notify users */ 165 mark_tsc_unstable("TSC halts in idle"); 166 tsc_marked_unstable = 1; 167 } 168 local_irq_disable(); 169 170 perf_lopwr_cb(true); 171 172 tick_broadcast_enable(); 173 tick_broadcast_enter(); 174 stop_critical_timings(); 175 176 mwait_idle_with_hints(power_saving_mwait_eax, 1); 177 178 start_critical_timings(); 179 tick_broadcast_exit(); 180 181 perf_lopwr_cb(false); 182 183 local_irq_enable(); 184 185 if (time_before(expire_time, jiffies)) { 186 do_sleep = 1; 187 break; 188 } 189 } 190 191 /* 192 * current sched_rt has threshold for rt task running time. 193 * When a rt task uses 95% CPU time, the rt thread will be 194 * scheduled out for 5% CPU time to not starve other tasks. But 195 * the mechanism only works when all CPUs have RT task running, 196 * as if one CPU hasn't RT task, RT task from other CPUs will 197 * borrow CPU time from this CPU and cause RT task use > 95% 198 * CPU time. To make 'avoid starvation' work, takes a nap here. 199 */ 200 if (unlikely(do_sleep)) 201 schedule_timeout_killable(HZ * idle_pct / 100); 202 203 /* If an external event has set the need_resched flag, then 204 * we need to deal with it, or this loop will continue to 205 * spin without calling __mwait(). 206 */ 207 if (unlikely(need_resched())) 208 schedule(); 209 } 210 211 exit_round_robin(tsk_index); 212 return 0; 213} 214 215static struct task_struct *ps_tsks[NR_CPUS]; 216static unsigned int ps_tsk_num; 217static int create_power_saving_task(void) 218{ 219 int rc; 220 221 ps_tsks[ps_tsk_num] = kthread_run(power_saving_thread, 222 (void *)(unsigned long)ps_tsk_num, 223 "acpi_pad/%d", ps_tsk_num); 224 225 if (IS_ERR(ps_tsks[ps_tsk_num])) { 226 rc = PTR_ERR(ps_tsks[ps_tsk_num]); 227 ps_tsks[ps_tsk_num] = NULL; 228 } else { 229 rc = 0; 230 ps_tsk_num++; 231 } 232 233 return rc; 234} 235 236static void destroy_power_saving_task(void) 237{ 238 if (ps_tsk_num > 0) { 239 ps_tsk_num--; 240 kthread_stop(ps_tsks[ps_tsk_num]); 241 ps_tsks[ps_tsk_num] = NULL; 242 } 243} 244 245static void set_power_saving_task_num(unsigned int num) 246{ 247 if (num > ps_tsk_num) { 248 while (ps_tsk_num < num) { 249 if (create_power_saving_task()) 250 return; 251 } 252 } else if (num < ps_tsk_num) { 253 while (ps_tsk_num > num) 254 destroy_power_saving_task(); 255 } 256} 257 258static void acpi_pad_idle_cpus(unsigned int num_cpus) 259{ 260 cpus_read_lock(); 261 262 num_cpus = min_t(unsigned int, num_cpus, num_online_cpus()); 263 set_power_saving_task_num(num_cpus); 264 265 cpus_read_unlock(); 266} 267 268static uint32_t acpi_pad_idle_cpus_num(void) 269{ 270 return ps_tsk_num; 271} 272 273static ssize_t rrtime_store(struct device *dev, 274 struct device_attribute *attr, const char *buf, size_t count) 275{ 276 unsigned long num; 277 278 if (kstrtoul(buf, 0, &num)) 279 return -EINVAL; 280 if (num < 1 || num >= 100) 281 return -EINVAL; 282 mutex_lock(&isolated_cpus_lock); 283 round_robin_time = num; 284 mutex_unlock(&isolated_cpus_lock); 285 return count; 286} 287 288static ssize_t rrtime_show(struct device *dev, 289 struct device_attribute *attr, char *buf) 290{ 291 return sysfs_emit(buf, "%d\n", round_robin_time); 292} 293static DEVICE_ATTR_RW(rrtime); 294 295static ssize_t idlepct_store(struct device *dev, 296 struct device_attribute *attr, const char *buf, size_t count) 297{ 298 unsigned long num; 299 300 if (kstrtoul(buf, 0, &num)) 301 return -EINVAL; 302 if (num < 1 || num >= 100) 303 return -EINVAL; 304 mutex_lock(&isolated_cpus_lock); 305 idle_pct = num; 306 mutex_unlock(&isolated_cpus_lock); 307 return count; 308} 309 310static ssize_t idlepct_show(struct device *dev, 311 struct device_attribute *attr, char *buf) 312{ 313 return sysfs_emit(buf, "%d\n", idle_pct); 314} 315static DEVICE_ATTR_RW(idlepct); 316 317static ssize_t idlecpus_store(struct device *dev, 318 struct device_attribute *attr, const char *buf, size_t count) 319{ 320 unsigned long num; 321 322 if (kstrtoul(buf, 0, &num)) 323 return -EINVAL; 324 mutex_lock(&isolated_cpus_lock); 325 acpi_pad_idle_cpus(num); 326 mutex_unlock(&isolated_cpus_lock); 327 return count; 328} 329 330static ssize_t idlecpus_show(struct device *dev, 331 struct device_attribute *attr, char *buf) 332{ 333 return cpumap_print_to_pagebuf(false, buf, 334 to_cpumask(pad_busy_cpus_bits)); 335} 336 337static DEVICE_ATTR_RW(idlecpus); 338 339static int acpi_pad_add_sysfs(struct acpi_device *device) 340{ 341 int result; 342 343 result = device_create_file(&device->dev, &dev_attr_idlecpus); 344 if (result) 345 return -ENODEV; 346 result = device_create_file(&device->dev, &dev_attr_idlepct); 347 if (result) { 348 device_remove_file(&device->dev, &dev_attr_idlecpus); 349 return -ENODEV; 350 } 351 result = device_create_file(&device->dev, &dev_attr_rrtime); 352 if (result) { 353 device_remove_file(&device->dev, &dev_attr_idlecpus); 354 device_remove_file(&device->dev, &dev_attr_idlepct); 355 return -ENODEV; 356 } 357 return 0; 358} 359 360static void acpi_pad_remove_sysfs(struct acpi_device *device) 361{ 362 device_remove_file(&device->dev, &dev_attr_idlecpus); 363 device_remove_file(&device->dev, &dev_attr_idlepct); 364 device_remove_file(&device->dev, &dev_attr_rrtime); 365} 366 367/* 368 * Query firmware how many CPUs should be idle 369 * return -1 on failure 370 */ 371static int acpi_pad_pur(acpi_handle handle) 372{ 373 struct acpi_buffer buffer = {ACPI_ALLOCATE_BUFFER, NULL}; 374 union acpi_object *package; 375 int num = -1; 376 377 if (ACPI_FAILURE(acpi_evaluate_object(handle, "_PUR", NULL, &buffer))) 378 return num; 379 380 if (!buffer.length || !buffer.pointer) 381 return num; 382 383 package = buffer.pointer; 384 385 if (package->type == ACPI_TYPE_PACKAGE && 386 package->package.count == 2 && 387 package->package.elements[0].integer.value == 1) /* rev 1 */ 388 389 num = package->package.elements[1].integer.value; 390 391 kfree(buffer.pointer); 392 return num; 393} 394 395static void acpi_pad_handle_notify(acpi_handle handle) 396{ 397 int num_cpus; 398 uint32_t idle_cpus; 399 struct acpi_buffer param = { 400 .length = 4, 401 .pointer = (void *)&idle_cpus, 402 }; 403 404 mutex_lock(&isolated_cpus_lock); 405 num_cpus = acpi_pad_pur(handle); 406 if (num_cpus < 0) { 407 mutex_unlock(&isolated_cpus_lock); 408 return; 409 } 410 acpi_pad_idle_cpus(num_cpus); 411 idle_cpus = acpi_pad_idle_cpus_num(); 412 acpi_evaluate_ost(handle, ACPI_PROCESSOR_AGGREGATOR_NOTIFY, 0, ¶m); 413 mutex_unlock(&isolated_cpus_lock); 414} 415 416static void acpi_pad_notify(acpi_handle handle, u32 event, 417 void *data) 418{ 419 struct acpi_device *device = data; 420 421 switch (event) { 422 case ACPI_PROCESSOR_AGGREGATOR_NOTIFY: 423 acpi_pad_handle_notify(handle); 424 acpi_bus_generate_netlink_event(device->pnp.device_class, 425 dev_name(&device->dev), event, 0); 426 break; 427 default: 428 pr_warn("Unsupported event [0x%x]\n", event); 429 break; 430 } 431} 432 433static int acpi_pad_add(struct acpi_device *device) 434{ 435 acpi_status status; 436 437 strcpy(acpi_device_name(device), ACPI_PROCESSOR_AGGREGATOR_DEVICE_NAME); 438 strcpy(acpi_device_class(device), ACPI_PROCESSOR_AGGREGATOR_CLASS); 439 440 if (acpi_pad_add_sysfs(device)) 441 return -ENODEV; 442 443 status = acpi_install_notify_handler(device->handle, 444 ACPI_DEVICE_NOTIFY, acpi_pad_notify, device); 445 if (ACPI_FAILURE(status)) { 446 acpi_pad_remove_sysfs(device); 447 return -ENODEV; 448 } 449 450 return 0; 451} 452 453static void acpi_pad_remove(struct acpi_device *device) 454{ 455 mutex_lock(&isolated_cpus_lock); 456 acpi_pad_idle_cpus(0); 457 mutex_unlock(&isolated_cpus_lock); 458 459 acpi_remove_notify_handler(device->handle, 460 ACPI_DEVICE_NOTIFY, acpi_pad_notify); 461 acpi_pad_remove_sysfs(device); 462} 463 464static const struct acpi_device_id pad_device_ids[] = { 465 {"ACPI000C", 0}, 466 {"", 0}, 467}; 468MODULE_DEVICE_TABLE(acpi, pad_device_ids); 469 470static struct acpi_driver acpi_pad_driver = { 471 .name = "processor_aggregator", 472 .class = ACPI_PROCESSOR_AGGREGATOR_CLASS, 473 .ids = pad_device_ids, 474 .ops = { 475 .add = acpi_pad_add, 476 .remove = acpi_pad_remove, 477 }, 478}; 479 480static int __init acpi_pad_init(void) 481{ 482 /* Xen ACPI PAD is used when running as Xen Dom0. */ 483 if (xen_initial_domain()) 484 return -ENODEV; 485 486 power_saving_mwait_init(); 487 if (power_saving_mwait_eax == 0) 488 return -EINVAL; 489 490 return acpi_bus_register_driver(&acpi_pad_driver); 491} 492 493static void __exit acpi_pad_exit(void) 494{ 495 acpi_bus_unregister_driver(&acpi_pad_driver); 496} 497 498module_init(acpi_pad_init); 499module_exit(acpi_pad_exit); 500MODULE_AUTHOR("Shaohua Li<shaohua.li@intel.com>"); 501MODULE_DESCRIPTION("ACPI Processor Aggregator Driver"); 502MODULE_LICENSE("GPL"); 503