1 /* 2 * acpi_pad.c ACPI Processor Aggregator Driver 3 * 4 * Copyright (c) 2009, Intel Corporation. 5 * 6 * This program is free software; you can redistribute it and/or modify it 7 * under the terms and conditions of the GNU General Public License, 8 * version 2, as published by the Free Software Foundation. 9 * 10 * This program is distributed in the hope it will be useful, but WITHOUT 11 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 12 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for 13 * more details. 14 * 15 */ 16 17 #include <linux/kernel.h> 18 #include <linux/cpumask.h> 19 #include <linux/module.h> 20 #include <linux/init.h> 21 #include <linux/types.h> 22 #include <linux/kthread.h> 23 #include <uapi/linux/sched/types.h> 24 #include <linux/freezer.h> 25 #include <linux/cpu.h> 26 #include <linux/tick.h> 27 #include <linux/slab.h> 28 #include <linux/acpi.h> 29 #include <asm/mwait.h> 30 #include <xen/xen.h> 31 32 #define ACPI_PROCESSOR_AGGREGATOR_CLASS "acpi_pad" 33 #define ACPI_PROCESSOR_AGGREGATOR_DEVICE_NAME "Processor Aggregator" 34 #define ACPI_PROCESSOR_AGGREGATOR_NOTIFY 0x80 35 static DEFINE_MUTEX(isolated_cpus_lock); 36 static DEFINE_MUTEX(round_robin_lock); 37 38 static unsigned long power_saving_mwait_eax; 39 40 static unsigned char tsc_detected_unstable; 41 static unsigned char tsc_marked_unstable; 42 43 static void power_saving_mwait_init(void) 44 { 45 unsigned int eax, ebx, ecx, edx; 46 unsigned int highest_cstate = 0; 47 unsigned int highest_subcstate = 0; 48 int i; 49 50 if (!boot_cpu_has(X86_FEATURE_MWAIT)) 51 return; 52 if (boot_cpu_data.cpuid_level < CPUID_MWAIT_LEAF) 53 return; 54 55 cpuid(CPUID_MWAIT_LEAF, &eax, &ebx, &ecx, &edx); 56 57 if (!(ecx & CPUID5_ECX_EXTENSIONS_SUPPORTED) || 58 !(ecx & CPUID5_ECX_INTERRUPT_BREAK)) 59 return; 60 61 edx >>= MWAIT_SUBSTATE_SIZE; 62 for (i = 0; i < 7 && edx; i++, edx >>= MWAIT_SUBSTATE_SIZE) { 63 if (edx & MWAIT_SUBSTATE_MASK) { 64 highest_cstate = i; 65 highest_subcstate = edx & MWAIT_SUBSTATE_MASK; 66 } 67 } 68 power_saving_mwait_eax = (highest_cstate << MWAIT_SUBSTATE_SIZE) | 69 (highest_subcstate - 1); 70 71 #if defined(CONFIG_X86) 72 switch (boot_cpu_data.x86_vendor) { 73 case X86_VENDOR_AMD: 74 case X86_VENDOR_INTEL: 75 /* 76 * AMD Fam10h TSC will tick in all 77 * C/P/S0/S1 states when this bit is set. 78 */ 79 if (!boot_cpu_has(X86_FEATURE_NONSTOP_TSC)) 80 tsc_detected_unstable = 1; 81 break; 82 default: 83 /* TSC could halt in idle */ 84 tsc_detected_unstable = 1; 85 } 86 #endif 87 } 88 89 static unsigned long cpu_weight[NR_CPUS]; 90 static int tsk_in_cpu[NR_CPUS] = {[0 ... NR_CPUS-1] = -1}; 91 static DECLARE_BITMAP(pad_busy_cpus_bits, NR_CPUS); 92 static void round_robin_cpu(unsigned int tsk_index) 93 { 94 struct cpumask *pad_busy_cpus = to_cpumask(pad_busy_cpus_bits); 95 cpumask_var_t tmp; 96 int cpu; 97 unsigned long min_weight = -1; 98 unsigned long uninitialized_var(preferred_cpu); 99 100 if (!alloc_cpumask_var(&tmp, GFP_KERNEL)) 101 return; 102 103 mutex_lock(&round_robin_lock); 104 cpumask_clear(tmp); 105 for_each_cpu(cpu, pad_busy_cpus) 106 cpumask_or(tmp, tmp, topology_sibling_cpumask(cpu)); 107 cpumask_andnot(tmp, cpu_online_mask, tmp); 108 /* avoid HT sibilings if possible */ 109 if (cpumask_empty(tmp)) 110 cpumask_andnot(tmp, cpu_online_mask, pad_busy_cpus); 111 if (cpumask_empty(tmp)) { 112 mutex_unlock(&round_robin_lock); 113 free_cpumask_var(tmp); 114 return; 115 } 116 for_each_cpu(cpu, tmp) { 117 if (cpu_weight[cpu] < min_weight) { 118 min_weight = cpu_weight[cpu]; 119 preferred_cpu = cpu; 120 } 121 } 122 123 if (tsk_in_cpu[tsk_index] != -1) 124 cpumask_clear_cpu(tsk_in_cpu[tsk_index], pad_busy_cpus); 125 tsk_in_cpu[tsk_index] = preferred_cpu; 126 cpumask_set_cpu(preferred_cpu, pad_busy_cpus); 127 cpu_weight[preferred_cpu]++; 128 mutex_unlock(&round_robin_lock); 129 130 set_cpus_allowed_ptr(current, cpumask_of(preferred_cpu)); 131 132 free_cpumask_var(tmp); 133 } 134 135 static void exit_round_robin(unsigned int tsk_index) 136 { 137 struct cpumask *pad_busy_cpus = to_cpumask(pad_busy_cpus_bits); 138 cpumask_clear_cpu(tsk_in_cpu[tsk_index], pad_busy_cpus); 139 tsk_in_cpu[tsk_index] = -1; 140 } 141 142 static unsigned int idle_pct = 5; /* percentage */ 143 static unsigned int round_robin_time = 1; /* second */ 144 static int power_saving_thread(void *data) 145 { 146 struct sched_param param = {.sched_priority = 1}; 147 int do_sleep; 148 unsigned int tsk_index = (unsigned long)data; 149 u64 last_jiffies = 0; 150 151 sched_setscheduler(current, SCHED_RR, ¶m); 152 153 while (!kthread_should_stop()) { 154 unsigned long expire_time; 155 156 /* round robin to cpus */ 157 expire_time = last_jiffies + round_robin_time * HZ; 158 if (time_before(expire_time, jiffies)) { 159 last_jiffies = jiffies; 160 round_robin_cpu(tsk_index); 161 } 162 163 do_sleep = 0; 164 165 expire_time = jiffies + HZ * (100 - idle_pct) / 100; 166 167 while (!need_resched()) { 168 if (tsc_detected_unstable && !tsc_marked_unstable) { 169 /* TSC could halt in idle, so notify users */ 170 mark_tsc_unstable("TSC halts in idle"); 171 tsc_marked_unstable = 1; 172 } 173 local_irq_disable(); 174 tick_broadcast_enable(); 175 tick_broadcast_enter(); 176 stop_critical_timings(); 177 178 mwait_idle_with_hints(power_saving_mwait_eax, 1); 179 180 start_critical_timings(); 181 tick_broadcast_exit(); 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 get_online_cpus(); 260 261 num_cpus = min_t(unsigned int, num_cpus, num_online_cpus()); 262 set_power_saving_task_num(num_cpus); 263 264 put_online_cpus(); 265 } 266 267 static uint32_t acpi_pad_idle_cpus_num(void) 268 { 269 return ps_tsk_num; 270 } 271 272 static ssize_t acpi_pad_rrtime_store(struct device *dev, 273 struct device_attribute *attr, const char *buf, size_t count) 274 { 275 unsigned long num; 276 if (kstrtoul(buf, 0, &num)) 277 return -EINVAL; 278 if (num < 1 || num >= 100) 279 return -EINVAL; 280 mutex_lock(&isolated_cpus_lock); 281 round_robin_time = num; 282 mutex_unlock(&isolated_cpus_lock); 283 return count; 284 } 285 286 static ssize_t acpi_pad_rrtime_show(struct device *dev, 287 struct device_attribute *attr, char *buf) 288 { 289 return scnprintf(buf, PAGE_SIZE, "%d\n", round_robin_time); 290 } 291 static DEVICE_ATTR(rrtime, S_IRUGO|S_IWUSR, 292 acpi_pad_rrtime_show, 293 acpi_pad_rrtime_store); 294 295 static ssize_t acpi_pad_idlepct_store(struct device *dev, 296 struct device_attribute *attr, const char *buf, size_t count) 297 { 298 unsigned long num; 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 acpi_pad_idlepct_show(struct device *dev, 310 struct device_attribute *attr, char *buf) 311 { 312 return scnprintf(buf, PAGE_SIZE, "%d\n", idle_pct); 313 } 314 static DEVICE_ATTR(idlepct, S_IRUGO|S_IWUSR, 315 acpi_pad_idlepct_show, 316 acpi_pad_idlepct_store); 317 318 static ssize_t acpi_pad_idlecpus_store(struct device *dev, 319 struct device_attribute *attr, const char *buf, size_t count) 320 { 321 unsigned long num; 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 } 329 330 static ssize_t acpi_pad_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 } 336 337 static DEVICE_ATTR(idlecpus, S_IRUGO|S_IWUSR, 338 acpi_pad_idlecpus_show, 339 acpi_pad_idlecpus_store); 340 341 static int acpi_pad_add_sysfs(struct acpi_device *device) 342 { 343 int result; 344 345 result = device_create_file(&device->dev, &dev_attr_idlecpus); 346 if (result) 347 return -ENODEV; 348 result = device_create_file(&device->dev, &dev_attr_idlepct); 349 if (result) { 350 device_remove_file(&device->dev, &dev_attr_idlecpus); 351 return -ENODEV; 352 } 353 result = device_create_file(&device->dev, &dev_attr_rrtime); 354 if (result) { 355 device_remove_file(&device->dev, &dev_attr_idlecpus); 356 device_remove_file(&device->dev, &dev_attr_idlepct); 357 return -ENODEV; 358 } 359 return 0; 360 } 361 362 static void acpi_pad_remove_sysfs(struct acpi_device *device) 363 { 364 device_remove_file(&device->dev, &dev_attr_idlecpus); 365 device_remove_file(&device->dev, &dev_attr_idlepct); 366 device_remove_file(&device->dev, &dev_attr_rrtime); 367 } 368 369 /* 370 * Query firmware how many CPUs should be idle 371 * return -1 on failure 372 */ 373 static int acpi_pad_pur(acpi_handle handle) 374 { 375 struct acpi_buffer buffer = {ACPI_ALLOCATE_BUFFER, NULL}; 376 union acpi_object *package; 377 int num = -1; 378 379 if (ACPI_FAILURE(acpi_evaluate_object(handle, "_PUR", NULL, &buffer))) 380 return num; 381 382 if (!buffer.length || !buffer.pointer) 383 return num; 384 385 package = buffer.pointer; 386 387 if (package->type == ACPI_TYPE_PACKAGE && 388 package->package.count == 2 && 389 package->package.elements[0].integer.value == 1) /* rev 1 */ 390 391 num = package->package.elements[1].integer.value; 392 393 kfree(buffer.pointer); 394 return num; 395 } 396 397 static void acpi_pad_handle_notify(acpi_handle handle) 398 { 399 int num_cpus; 400 uint32_t idle_cpus; 401 struct acpi_buffer param = { 402 .length = 4, 403 .pointer = (void *)&idle_cpus, 404 }; 405 406 mutex_lock(&isolated_cpus_lock); 407 num_cpus = acpi_pad_pur(handle); 408 if (num_cpus < 0) { 409 mutex_unlock(&isolated_cpus_lock); 410 return; 411 } 412 acpi_pad_idle_cpus(num_cpus); 413 idle_cpus = acpi_pad_idle_cpus_num(); 414 acpi_evaluate_ost(handle, ACPI_PROCESSOR_AGGREGATOR_NOTIFY, 0, ¶m); 415 mutex_unlock(&isolated_cpus_lock); 416 } 417 418 static void acpi_pad_notify(acpi_handle handle, u32 event, 419 void *data) 420 { 421 struct acpi_device *device = data; 422 423 switch (event) { 424 case ACPI_PROCESSOR_AGGREGATOR_NOTIFY: 425 acpi_pad_handle_notify(handle); 426 acpi_bus_generate_netlink_event(device->pnp.device_class, 427 dev_name(&device->dev), event, 0); 428 break; 429 default: 430 pr_warn("Unsupported event [0x%x]\n", event); 431 break; 432 } 433 } 434 435 static int acpi_pad_add(struct acpi_device *device) 436 { 437 acpi_status status; 438 439 strcpy(acpi_device_name(device), ACPI_PROCESSOR_AGGREGATOR_DEVICE_NAME); 440 strcpy(acpi_device_class(device), ACPI_PROCESSOR_AGGREGATOR_CLASS); 441 442 if (acpi_pad_add_sysfs(device)) 443 return -ENODEV; 444 445 status = acpi_install_notify_handler(device->handle, 446 ACPI_DEVICE_NOTIFY, acpi_pad_notify, device); 447 if (ACPI_FAILURE(status)) { 448 acpi_pad_remove_sysfs(device); 449 return -ENODEV; 450 } 451 452 return 0; 453 } 454 455 static int acpi_pad_remove(struct acpi_device *device) 456 { 457 mutex_lock(&isolated_cpus_lock); 458 acpi_pad_idle_cpus(0); 459 mutex_unlock(&isolated_cpus_lock); 460 461 acpi_remove_notify_handler(device->handle, 462 ACPI_DEVICE_NOTIFY, acpi_pad_notify); 463 acpi_pad_remove_sysfs(device); 464 return 0; 465 } 466 467 static const struct acpi_device_id pad_device_ids[] = { 468 {"ACPI000C", 0}, 469 {"", 0}, 470 }; 471 MODULE_DEVICE_TABLE(acpi, pad_device_ids); 472 473 static struct acpi_driver acpi_pad_driver = { 474 .name = "processor_aggregator", 475 .class = ACPI_PROCESSOR_AGGREGATOR_CLASS, 476 .ids = pad_device_ids, 477 .ops = { 478 .add = acpi_pad_add, 479 .remove = acpi_pad_remove, 480 }, 481 }; 482 483 static int __init acpi_pad_init(void) 484 { 485 /* Xen ACPI PAD is used when running as Xen Dom0. */ 486 if (xen_initial_domain()) 487 return -ENODEV; 488 489 power_saving_mwait_init(); 490 if (power_saving_mwait_eax == 0) 491 return -EINVAL; 492 493 return acpi_bus_register_driver(&acpi_pad_driver); 494 } 495 496 static void __exit acpi_pad_exit(void) 497 { 498 acpi_bus_unregister_driver(&acpi_pad_driver); 499 } 500 501 module_init(acpi_pad_init); 502 module_exit(acpi_pad_exit); 503 MODULE_AUTHOR("Shaohua Li<shaohua.li@intel.com>"); 504 MODULE_DESCRIPTION("ACPI Processor Aggregator Driver"); 505 MODULE_LICENSE("GPL"); 506