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