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 * You should have received a copy of the GNU General Public License along with 16 * this program; if not, write to the Free Software Foundation, Inc., 17 * 51 Franklin St - Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 */ 20 21 #include <linux/kernel.h> 22 #include <linux/cpumask.h> 23 #include <linux/module.h> 24 #include <linux/init.h> 25 #include <linux/types.h> 26 #include <linux/kthread.h> 27 #include <linux/freezer.h> 28 #include <linux/cpu.h> 29 #include <linux/clockchips.h> 30 #include <acpi/acpi_bus.h> 31 #include <acpi/acpi_drivers.h> 32 33 #define ACPI_PROCESSOR_AGGREGATOR_CLASS "processor_aggregator" 34 #define ACPI_PROCESSOR_AGGREGATOR_DEVICE_NAME "Processor Aggregator" 35 #define ACPI_PROCESSOR_AGGREGATOR_NOTIFY 0x80 36 static DEFINE_MUTEX(isolated_cpus_lock); 37 38 #define MWAIT_SUBSTATE_MASK (0xf) 39 #define MWAIT_CSTATE_MASK (0xf) 40 #define MWAIT_SUBSTATE_SIZE (4) 41 #define CPUID_MWAIT_LEAF (5) 42 #define CPUID5_ECX_EXTENSIONS_SUPPORTED (0x1) 43 #define CPUID5_ECX_INTERRUPT_BREAK (0x2) 44 static unsigned long power_saving_mwait_eax; 45 static void power_saving_mwait_init(void) 46 { 47 unsigned int eax, ebx, ecx, edx; 48 unsigned int highest_cstate = 0; 49 unsigned int highest_subcstate = 0; 50 int i; 51 52 if (!boot_cpu_has(X86_FEATURE_MWAIT)) 53 return; 54 if (boot_cpu_data.cpuid_level < CPUID_MWAIT_LEAF) 55 return; 56 57 cpuid(CPUID_MWAIT_LEAF, &eax, &ebx, &ecx, &edx); 58 59 if (!(ecx & CPUID5_ECX_EXTENSIONS_SUPPORTED) || 60 !(ecx & CPUID5_ECX_INTERRUPT_BREAK)) 61 return; 62 63 edx >>= MWAIT_SUBSTATE_SIZE; 64 for (i = 0; i < 7 && edx; i++, edx >>= MWAIT_SUBSTATE_SIZE) { 65 if (edx & MWAIT_SUBSTATE_MASK) { 66 highest_cstate = i; 67 highest_subcstate = edx & MWAIT_SUBSTATE_MASK; 68 } 69 } 70 power_saving_mwait_eax = (highest_cstate << MWAIT_SUBSTATE_SIZE) | 71 (highest_subcstate - 1); 72 73 for_each_online_cpu(i) 74 clockevents_notify(CLOCK_EVT_NOTIFY_BROADCAST_ON, &i); 75 76 #if defined(CONFIG_GENERIC_TIME) && defined(CONFIG_X86) 77 switch (boot_cpu_data.x86_vendor) { 78 case X86_VENDOR_AMD: 79 case X86_VENDOR_INTEL: 80 /* 81 * AMD Fam10h TSC will tick in all 82 * C/P/S0/S1 states when this bit is set. 83 */ 84 if (boot_cpu_has(X86_FEATURE_NONSTOP_TSC)) 85 return; 86 87 /*FALL THROUGH*/ 88 default: 89 /* TSC could halt in idle, so notify users */ 90 mark_tsc_unstable("TSC halts in idle"); 91 } 92 #endif 93 } 94 95 static unsigned long cpu_weight[NR_CPUS]; 96 static int tsk_in_cpu[NR_CPUS] = {[0 ... NR_CPUS-1] = -1}; 97 static DECLARE_BITMAP(pad_busy_cpus_bits, NR_CPUS); 98 static void round_robin_cpu(unsigned int tsk_index) 99 { 100 struct cpumask *pad_busy_cpus = to_cpumask(pad_busy_cpus_bits); 101 cpumask_var_t tmp; 102 int cpu; 103 unsigned long min_weight = -1; 104 unsigned long uninitialized_var(preferred_cpu); 105 106 if (!alloc_cpumask_var(&tmp, GFP_KERNEL)) 107 return; 108 109 mutex_lock(&isolated_cpus_lock); 110 cpumask_clear(tmp); 111 for_each_cpu(cpu, pad_busy_cpus) 112 cpumask_or(tmp, tmp, topology_thread_cpumask(cpu)); 113 cpumask_andnot(tmp, cpu_online_mask, tmp); 114 /* avoid HT sibilings if possible */ 115 if (cpumask_empty(tmp)) 116 cpumask_andnot(tmp, cpu_online_mask, pad_busy_cpus); 117 if (cpumask_empty(tmp)) { 118 mutex_unlock(&isolated_cpus_lock); 119 return; 120 } 121 for_each_cpu(cpu, tmp) { 122 if (cpu_weight[cpu] < min_weight) { 123 min_weight = cpu_weight[cpu]; 124 preferred_cpu = cpu; 125 } 126 } 127 128 if (tsk_in_cpu[tsk_index] != -1) 129 cpumask_clear_cpu(tsk_in_cpu[tsk_index], pad_busy_cpus); 130 tsk_in_cpu[tsk_index] = preferred_cpu; 131 cpumask_set_cpu(preferred_cpu, pad_busy_cpus); 132 cpu_weight[preferred_cpu]++; 133 mutex_unlock(&isolated_cpus_lock); 134 135 set_cpus_allowed_ptr(current, cpumask_of(preferred_cpu)); 136 } 137 138 static void exit_round_robin(unsigned int tsk_index) 139 { 140 struct cpumask *pad_busy_cpus = to_cpumask(pad_busy_cpus_bits); 141 cpumask_clear_cpu(tsk_in_cpu[tsk_index], pad_busy_cpus); 142 tsk_in_cpu[tsk_index] = -1; 143 } 144 145 static unsigned int idle_pct = 5; /* percentage */ 146 static unsigned int round_robin_time = 10; /* second */ 147 static int power_saving_thread(void *data) 148 { 149 struct sched_param param = {.sched_priority = 1}; 150 int do_sleep; 151 unsigned int tsk_index = (unsigned long)data; 152 u64 last_jiffies = 0; 153 154 sched_setscheduler(current, SCHED_RR, ¶m); 155 156 while (!kthread_should_stop()) { 157 int cpu; 158 u64 expire_time; 159 160 try_to_freeze(); 161 162 /* round robin to cpus */ 163 if (last_jiffies + round_robin_time * HZ < jiffies) { 164 last_jiffies = jiffies; 165 round_robin_cpu(tsk_index); 166 } 167 168 do_sleep = 0; 169 170 current_thread_info()->status &= ~TS_POLLING; 171 /* 172 * TS_POLLING-cleared state must be visible before we test 173 * NEED_RESCHED: 174 */ 175 smp_mb(); 176 177 expire_time = jiffies + HZ * (100 - idle_pct) / 100; 178 179 while (!need_resched()) { 180 local_irq_disable(); 181 cpu = smp_processor_id(); 182 clockevents_notify(CLOCK_EVT_NOTIFY_BROADCAST_ENTER, 183 &cpu); 184 stop_critical_timings(); 185 186 __monitor((void *)¤t_thread_info()->flags, 0, 0); 187 smp_mb(); 188 if (!need_resched()) 189 __mwait(power_saving_mwait_eax, 1); 190 191 start_critical_timings(); 192 clockevents_notify(CLOCK_EVT_NOTIFY_BROADCAST_EXIT, 193 &cpu); 194 local_irq_enable(); 195 196 if (jiffies > expire_time) { 197 do_sleep = 1; 198 break; 199 } 200 } 201 202 current_thread_info()->status |= TS_POLLING; 203 204 /* 205 * current sched_rt has threshold for rt task running time. 206 * When a rt task uses 95% CPU time, the rt thread will be 207 * scheduled out for 5% CPU time to not starve other tasks. But 208 * the mechanism only works when all CPUs have RT task running, 209 * as if one CPU hasn't RT task, RT task from other CPUs will 210 * borrow CPU time from this CPU and cause RT task use > 95% 211 * CPU time. To make 'avoid staration' work, takes a nap here. 212 */ 213 if (do_sleep) 214 schedule_timeout_killable(HZ * idle_pct / 100); 215 } 216 217 exit_round_robin(tsk_index); 218 return 0; 219 } 220 221 static struct task_struct *ps_tsks[NR_CPUS]; 222 static unsigned int ps_tsk_num; 223 static int create_power_saving_task(void) 224 { 225 ps_tsks[ps_tsk_num] = kthread_run(power_saving_thread, 226 (void *)(unsigned long)ps_tsk_num, 227 "power_saving/%d", ps_tsk_num); 228 if (ps_tsks[ps_tsk_num]) { 229 ps_tsk_num++; 230 return 0; 231 } 232 return -EINVAL; 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 } 241 } 242 243 static void set_power_saving_task_num(unsigned int num) 244 { 245 if (num > ps_tsk_num) { 246 while (ps_tsk_num < num) { 247 if (create_power_saving_task()) 248 return; 249 } 250 } else if (num < ps_tsk_num) { 251 while (ps_tsk_num > num) 252 destroy_power_saving_task(); 253 } 254 } 255 256 static int acpi_pad_idle_cpus(unsigned int num_cpus) 257 { 258 get_online_cpus(); 259 260 num_cpus = min_t(unsigned int, num_cpus, num_online_cpus()); 261 set_power_saving_task_num(num_cpus); 262 263 put_online_cpus(); 264 return 0; 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 (strict_strtoul(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", 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 (strict_strtoul(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", 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 (strict_strtoul(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 cpumask_scnprintf(buf, PAGE_SIZE, 334 to_cpumask(pad_busy_cpus_bits)); 335 } 336 static DEVICE_ATTR(idlecpus, S_IRUGO|S_IWUSR, 337 acpi_pad_idlecpus_show, 338 acpi_pad_idlecpus_store); 339 340 static int acpi_pad_add_sysfs(struct acpi_device *device) 341 { 342 int result; 343 344 result = device_create_file(&device->dev, &dev_attr_idlecpus); 345 if (result) 346 return -ENODEV; 347 result = device_create_file(&device->dev, &dev_attr_idlepct); 348 if (result) { 349 device_remove_file(&device->dev, &dev_attr_idlecpus); 350 return -ENODEV; 351 } 352 result = device_create_file(&device->dev, &dev_attr_rrtime); 353 if (result) { 354 device_remove_file(&device->dev, &dev_attr_idlecpus); 355 device_remove_file(&device->dev, &dev_attr_idlepct); 356 return -ENODEV; 357 } 358 return 0; 359 } 360 361 static void acpi_pad_remove_sysfs(struct acpi_device *device) 362 { 363 device_remove_file(&device->dev, &dev_attr_idlecpus); 364 device_remove_file(&device->dev, &dev_attr_idlepct); 365 device_remove_file(&device->dev, &dev_attr_rrtime); 366 } 367 368 /* Query firmware how many CPUs should be idle */ 369 static int acpi_pad_pur(acpi_handle handle, int *num_cpus) 370 { 371 struct acpi_buffer buffer = {ACPI_ALLOCATE_BUFFER, NULL}; 372 acpi_status status; 373 union acpi_object *package; 374 int rev, num, ret = -EINVAL; 375 376 status = acpi_evaluate_object(handle, "_PUR", NULL, &buffer); 377 if (ACPI_FAILURE(status)) 378 return -EINVAL; 379 package = buffer.pointer; 380 if (package->type != ACPI_TYPE_PACKAGE || package->package.count != 2) 381 goto out; 382 rev = package->package.elements[0].integer.value; 383 num = package->package.elements[1].integer.value; 384 if (rev != 1) 385 goto out; 386 *num_cpus = num; 387 ret = 0; 388 out: 389 kfree(buffer.pointer); 390 return ret; 391 } 392 393 /* Notify firmware how many CPUs are idle */ 394 static void acpi_pad_ost(acpi_handle handle, int stat, 395 uint32_t idle_cpus) 396 { 397 union acpi_object params[3] = { 398 {.type = ACPI_TYPE_INTEGER,}, 399 {.type = ACPI_TYPE_INTEGER,}, 400 {.type = ACPI_TYPE_BUFFER,}, 401 }; 402 struct acpi_object_list arg_list = {3, params}; 403 404 params[0].integer.value = ACPI_PROCESSOR_AGGREGATOR_NOTIFY; 405 params[1].integer.value = stat; 406 params[2].buffer.length = 4; 407 params[2].buffer.pointer = (void *)&idle_cpus; 408 acpi_evaluate_object(handle, "_OST", &arg_list, NULL); 409 } 410 411 static void acpi_pad_handle_notify(acpi_handle handle) 412 { 413 int num_cpus, ret; 414 uint32_t idle_cpus; 415 416 mutex_lock(&isolated_cpus_lock); 417 if (acpi_pad_pur(handle, &num_cpus)) { 418 mutex_unlock(&isolated_cpus_lock); 419 return; 420 } 421 ret = acpi_pad_idle_cpus(num_cpus); 422 idle_cpus = acpi_pad_idle_cpus_num(); 423 if (!ret) 424 acpi_pad_ost(handle, 0, idle_cpus); 425 else 426 acpi_pad_ost(handle, 1, 0); 427 mutex_unlock(&isolated_cpus_lock); 428 } 429 430 static void acpi_pad_notify(acpi_handle handle, u32 event, 431 void *data) 432 { 433 struct acpi_device *device = data; 434 435 switch (event) { 436 case ACPI_PROCESSOR_AGGREGATOR_NOTIFY: 437 acpi_pad_handle_notify(handle); 438 acpi_bus_generate_proc_event(device, event, 0); 439 acpi_bus_generate_netlink_event(device->pnp.device_class, 440 dev_name(&device->dev), event, 0); 441 break; 442 default: 443 printk(KERN_WARNING"Unsupported event [0x%x]\n", event); 444 break; 445 } 446 } 447 448 static int acpi_pad_add(struct acpi_device *device) 449 { 450 acpi_status status; 451 452 strcpy(acpi_device_name(device), ACPI_PROCESSOR_AGGREGATOR_DEVICE_NAME); 453 strcpy(acpi_device_class(device), ACPI_PROCESSOR_AGGREGATOR_CLASS); 454 455 if (acpi_pad_add_sysfs(device)) 456 return -ENODEV; 457 458 status = acpi_install_notify_handler(device->handle, 459 ACPI_DEVICE_NOTIFY, acpi_pad_notify, device); 460 if (ACPI_FAILURE(status)) { 461 acpi_pad_remove_sysfs(device); 462 return -ENODEV; 463 } 464 465 return 0; 466 } 467 468 static int acpi_pad_remove(struct acpi_device *device, 469 int type) 470 { 471 mutex_lock(&isolated_cpus_lock); 472 acpi_pad_idle_cpus(0); 473 mutex_unlock(&isolated_cpus_lock); 474 475 acpi_remove_notify_handler(device->handle, 476 ACPI_DEVICE_NOTIFY, acpi_pad_notify); 477 acpi_pad_remove_sysfs(device); 478 return 0; 479 } 480 481 static const struct acpi_device_id pad_device_ids[] = { 482 {"ACPI000C", 0}, 483 {"", 0}, 484 }; 485 MODULE_DEVICE_TABLE(acpi, pad_device_ids); 486 487 static struct acpi_driver acpi_pad_driver = { 488 .name = "processor_aggregator", 489 .class = ACPI_PROCESSOR_AGGREGATOR_CLASS, 490 .ids = pad_device_ids, 491 .ops = { 492 .add = acpi_pad_add, 493 .remove = acpi_pad_remove, 494 }, 495 }; 496 497 static int __init acpi_pad_init(void) 498 { 499 power_saving_mwait_init(); 500 if (power_saving_mwait_eax == 0) 501 return -EINVAL; 502 503 return acpi_bus_register_driver(&acpi_pad_driver); 504 } 505 506 static void __exit acpi_pad_exit(void) 507 { 508 acpi_bus_unregister_driver(&acpi_pad_driver); 509 } 510 511 module_init(acpi_pad_init); 512 module_exit(acpi_pad_exit); 513 MODULE_AUTHOR("Shaohua Li<shaohua.li@intel.com>"); 514 MODULE_DESCRIPTION("ACPI Processor Aggregator Driver"); 515 MODULE_LICENSE("GPL"); 516