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