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