drivers/acpi/acpi_pad.c

   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 <linux/slab.h>
  31 #include <acpi/acpi_bus.h>
  32 #include <acpi/acpi_drivers.h>
  33 #include <asm/mwait.h>
  34
  35 #define ACPI_PROCESSOR_AGGREGATOR_CLASS "acpi_pad"
  36 #define ACPI_PROCESSOR_AGGREGATOR_DEVICE_NAME "Processor Aggregator"
  37 #define ACPI_PROCESSOR_AGGREGATOR_NOTIFY 0x80
  38 static DEFINE_MUTEX(isolated_cpus_lock);
  39 static DEFINE_MUTEX(round_robin_lock);
  40
  41 static unsigned long power_saving_mwait_eax;
  42
  43 static unsigned char tsc_detected_unstable;
  44 static unsigned char tsc_marked_unstable;
  45 static unsigned char lapic_detected_unstable;
  46 static unsigned char lapic_marked_unstable;
  47
  48 static void power_saving_mwait_init(void)
  49 {
  50         unsigned int eax, ebx, ecx, edx;
  51         unsigned int highest_cstate = 0;
  52         unsigned int highest_subcstate = 0;
  53         int i;
  54
  55         if (!boot_cpu_has(X86_FEATURE_MWAIT))
  56                 return;
  57         if (boot_cpu_data.cpuid_level < CPUID_MWAIT_LEAF)
  58                 return;
  59
  60         cpuid(CPUID_MWAIT_LEAF, &eax, &ebx, &ecx, &edx);
  61
  62         if (!(ecx & CPUID5_ECX_EXTENSIONS_SUPPORTED) ||
  63             !(ecx & CPUID5_ECX_INTERRUPT_BREAK))
  64                 return;
  65
  66         edx >>= MWAIT_SUBSTATE_SIZE;
  67         for (i = 0; i < 7 && edx; i++, edx >>= MWAIT_SUBSTATE_SIZE) {
  68                 if (edx & MWAIT_SUBSTATE_MASK) {
  69                         highest_cstate = i;
  70                         highest_subcstate = edx & MWAIT_SUBSTATE_MASK;
  71                 }
  72         }
  73         power_saving_mwait_eax = (highest_cstate << MWAIT_SUBSTATE_SIZE) |
  74                 (highest_subcstate - 1);
  75
  76 #if 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                         tsc_detected_unstable = 1;
  86                 if (!boot_cpu_has(X86_FEATURE_ARAT))
  87                         lapic_detected_unstable = 1;
  88                 break;
  89         default:
  90                 /* TSC & LAPIC could halt in idle */
  91                 tsc_detected_unstable = 1;
  92                 lapic_detected_unstable = 1;
  93         }
  94 #endif
  95 }
  96
  97 static unsigned long cpu_weight[NR_CPUS];
  98 static int tsk_in_cpu[NR_CPUS] = {[0 ... NR_CPUS-1] = -1};
  99 static DECLARE_BITMAP(pad_busy_cpus_bits, NR_CPUS);
 100 static void round_robin_cpu(unsigned int tsk_index)
 101 {
 102         struct cpumask *pad_busy_cpus = to_cpumask(pad_busy_cpus_bits);
 103         cpumask_var_t tmp;
 104         int cpu;
 105         unsigned long min_weight = -1;
 106         unsigned long uninitialized_var(preferred_cpu);
 107
 108         if (!alloc_cpumask_var(&tmp, GFP_KERNEL))
 109                 return;
 110
 111         mutex_lock(&round_robin_lock);
 112         cpumask_clear(tmp);
 113         for_each_cpu(cpu, pad_busy_cpus)
 114                 cpumask_or(tmp, tmp, topology_thread_cpumask(cpu));
 115         cpumask_andnot(tmp, cpu_online_mask, tmp);
 116         /* avoid HT sibilings if possible */
 117         if (cpumask_empty(tmp))
 118                 cpumask_andnot(tmp, cpu_online_mask, pad_busy_cpus);
 119         if (cpumask_empty(tmp)) {
 120                 mutex_unlock(&round_robin_lock);
 121                 return;
 122         }
 123         for_each_cpu(cpu, tmp) {
 124                 if (cpu_weight[cpu] < min_weight) {
 125                         min_weight = cpu_weight[cpu];
 126                         preferred_cpu = cpu;
 127                 }
 128         }
 129
 130         if (tsk_in_cpu[tsk_index] != -1)
 131                 cpumask_clear_cpu(tsk_in_cpu[tsk_index], pad_busy_cpus);
 132         tsk_in_cpu[tsk_index] = preferred_cpu;
 133         cpumask_set_cpu(preferred_cpu, pad_busy_cpus);
 134         cpu_weight[preferred_cpu]++;
 135         mutex_unlock(&round_robin_lock);
 136
 137         set_cpus_allowed_ptr(current, cpumask_of(preferred_cpu));
 138 }
 139
 140 static void exit_round_robin(unsigned int tsk_index)
 141 {
 142         struct cpumask *pad_busy_cpus = to_cpumask(pad_busy_cpus_bits);
 143         cpumask_clear_cpu(tsk_in_cpu[tsk_index], pad_busy_cpus);
 144         tsk_in_cpu[tsk_index] = -1;
 145 }
 146
 147 static unsigned int idle_pct = 5; /* percentage */
 148 static unsigned int round_robin_time = 1; /* second */
 149 static int power_saving_thread(void *data)
 150 {
 151         struct sched_param param = {.sched_priority = 1};
 152         int do_sleep;
 153         unsigned int tsk_index = (unsigned long)data;
 154         u64 last_jiffies = 0;
 155
 156         sched_setscheduler(current, SCHED_RR, &param);
 157         set_freezable();
 158
 159         while (!kthread_should_stop()) {
 160                 int cpu;
 161                 u64 expire_time;
 162
 163                 try_to_freeze();
 164
 165                 /* round robin to cpus */
 166                 if (last_jiffies + round_robin_time * HZ < jiffies) {
 167                         last_jiffies = jiffies;
 168                         round_robin_cpu(tsk_index);
 169                 }
 170
 171                 do_sleep = 0;
 172
 173                 expire_time = jiffies + HZ * (100 - idle_pct) / 100;
 174
 175                 while (!need_resched()) {
 176                         if (tsc_detected_unstable && !tsc_marked_unstable) {
 177                                 /* TSC could halt in idle, so notify users */
 178                                 mark_tsc_unstable("TSC halts in idle");
 179                                 tsc_marked_unstable = 1;
 180                         }
 181                         if (lapic_detected_unstable && !lapic_marked_unstable) {
 182                                 int i;
 183                                 /* LAPIC could halt in idle, so notify users */
 184                                 for_each_online_cpu(i)
 185                                         clockevents_notify(
 186                                                 CLOCK_EVT_NOTIFY_BROADCAST_ON,
 187                                                 &i);
 188                                 lapic_marked_unstable = 1;
 189                         }
 190                         local_irq_disable();
 191                         cpu = smp_processor_id();
 192                         if (lapic_marked_unstable)
 193                                 clockevents_notify(
 194                                         CLOCK_EVT_NOTIFY_BROADCAST_ENTER, &cpu);
 195                         stop_critical_timings();
 196
 197                         __monitor((void *)&current_thread_info()->flags, 0, 0);
 198                         smp_mb();
 199                         if (!need_resched())
 200                                 __mwait(power_saving_mwait_eax, 1);
 201
 202                         start_critical_timings();
 203                         if (lapic_marked_unstable)
 204                                 clockevents_notify(
 205                                         CLOCK_EVT_NOTIFY_BROADCAST_EXIT, &cpu);
 206                         local_irq_enable();
 207
 208                         if (jiffies > expire_time) {
 209                                 do_sleep = 1;
 210                                 break;
 211                         }
 212                 }
 213
 214                 /*
 215                  * current sched_rt has threshold for rt task running time.
 216                  * When a rt task uses 95% CPU time, the rt thread will be
 217                  * scheduled out for 5% CPU time to not starve other tasks. But
 218                  * the mechanism only works when all CPUs have RT task running,
 219                  * as if one CPU hasn't RT task, RT task from other CPUs will
 220                  * borrow CPU time from this CPU and cause RT task use > 95%
 221                  * CPU time. To make 'avoid starvation' work, takes a nap here.
 222                  */
 223                 if (do_sleep)
 224                         schedule_timeout_killable(HZ * idle_pct / 100);
 225         }
 226
 227         exit_round_robin(tsk_index);
 228         return 0;
 229 }
 230
 231 static struct task_struct *ps_tsks[NR_CPUS];
 232 static unsigned int ps_tsk_num;
 233 static int create_power_saving_task(void)
 234 {
 235         int rc = -ENOMEM;
 236
 237         ps_tsks[ps_tsk_num] = kthread_run(power_saving_thread,
 238                 (void *)(unsigned long)ps_tsk_num,
 239                 "acpi_pad/%d", ps_tsk_num);
 240         rc = PTR_RET(ps_tsks[ps_tsk_num]);
 241         if (!rc)
 242                 ps_tsk_num++;
 243         else
 244                 ps_tsks[ps_tsk_num] = NULL;
 245
 246         return rc;
 247 }
 248
 249 static void destroy_power_saving_task(void)
 250 {
 251         if (ps_tsk_num > 0) {
 252                 ps_tsk_num--;
 253                 kthread_stop(ps_tsks[ps_tsk_num]);
 254                 ps_tsks[ps_tsk_num] = NULL;
 255         }
 256 }
 257
 258 static void set_power_saving_task_num(unsigned int num)
 259 {
 260         if (num > ps_tsk_num) {
 261                 while (ps_tsk_num < num) {
 262                         if (create_power_saving_task())
 263                                 return;
 264                 }
 265         } else if (num < ps_tsk_num) {
 266                 while (ps_tsk_num > num)
 267                         destroy_power_saving_task();
 268         }
 269 }
 270
 271 static void acpi_pad_idle_cpus(unsigned int num_cpus)
 272 {
 273         get_online_cpus();
 274
 275         num_cpus = min_t(unsigned int, num_cpus, num_online_cpus());
 276         set_power_saving_task_num(num_cpus);
 277
 278         put_online_cpus();
 279 }
 280
 281 static uint32_t acpi_pad_idle_cpus_num(void)
 282 {
 283         return ps_tsk_num;
 284 }
 285
 286 static ssize_t acpi_pad_rrtime_store(struct device *dev,
 287         struct device_attribute *attr, const char *buf, size_t count)
 288 {
 289         unsigned long num;
 290         if (kstrtoul(buf, 0, &num))
 291                 return -EINVAL;
 292         if (num < 1 || num >= 100)
 293                 return -EINVAL;
 294         mutex_lock(&isolated_cpus_lock);
 295         round_robin_time = num;
 296         mutex_unlock(&isolated_cpus_lock);
 297         return count;
 298 }
 299
 300 static ssize_t acpi_pad_rrtime_show(struct device *dev,
 301         struct device_attribute *attr, char *buf)
 302 {
 303         return scnprintf(buf, PAGE_SIZE, "%d\n", round_robin_time);
 304 }
 305 static DEVICE_ATTR(rrtime, S_IRUGO|S_IWUSR,
 306         acpi_pad_rrtime_show,
 307         acpi_pad_rrtime_store);
 308
 309 static ssize_t acpi_pad_idlepct_store(struct device *dev,
 310         struct device_attribute *attr, const char *buf, size_t count)
 311 {
 312         unsigned long num;
 313         if (kstrtoul(buf, 0, &num))
 314                 return -EINVAL;
 315         if (num < 1 || num >= 100)
 316                 return -EINVAL;
 317         mutex_lock(&isolated_cpus_lock);
 318         idle_pct = num;
 319         mutex_unlock(&isolated_cpus_lock);
 320         return count;
 321 }
 322
 323 static ssize_t acpi_pad_idlepct_show(struct device *dev,
 324         struct device_attribute *attr, char *buf)
 325 {
 326         return scnprintf(buf, PAGE_SIZE, "%d\n", idle_pct);
 327 }
 328 static DEVICE_ATTR(idlepct, S_IRUGO|S_IWUSR,
 329         acpi_pad_idlepct_show,
 330         acpi_pad_idlepct_store);
 331
 332 static ssize_t acpi_pad_idlecpus_store(struct device *dev,
 333         struct device_attribute *attr, const char *buf, size_t count)
 334 {
 335         unsigned long num;
 336         if (kstrtoul(buf, 0, &num))
 337                 return -EINVAL;
 338         mutex_lock(&isolated_cpus_lock);
 339         acpi_pad_idle_cpus(num);
 340         mutex_unlock(&isolated_cpus_lock);
 341         return count;
 342 }
 343
 344 static ssize_t acpi_pad_idlecpus_show(struct device *dev,
 345         struct device_attribute *attr, char *buf)
 346 {
 347         int n = 0;
 348         n = cpumask_scnprintf(buf, PAGE_SIZE-2, to_cpumask(pad_busy_cpus_bits));
 349         buf[n++] = '\n';
 350         buf[n] = '\0';
 351         return n;
 352 }
 353 static DEVICE_ATTR(idlecpus, S_IRUGO|S_IWUSR,
 354         acpi_pad_idlecpus_show,
 355         acpi_pad_idlecpus_store);
 356
 357 static int acpi_pad_add_sysfs(struct acpi_device *device)
 358 {
 359         int result;
 360
 361         result = device_create_file(&device->dev, &dev_attr_idlecpus);
 362         if (result)
 363                 return -ENODEV;
 364         result = device_create_file(&device->dev, &dev_attr_idlepct);
 365         if (result) {
 366                 device_remove_file(&device->dev, &dev_attr_idlecpus);
 367                 return -ENODEV;
 368         }
 369         result = device_create_file(&device->dev, &dev_attr_rrtime);
 370         if (result) {
 371                 device_remove_file(&device->dev, &dev_attr_idlecpus);
 372                 device_remove_file(&device->dev, &dev_attr_idlepct);
 373                 return -ENODEV;
 374         }
 375         return 0;
 376 }
 377
 378 static void acpi_pad_remove_sysfs(struct acpi_device *device)
 379 {
 380         device_remove_file(&device->dev, &dev_attr_idlecpus);
 381         device_remove_file(&device->dev, &dev_attr_idlepct);
 382         device_remove_file(&device->dev, &dev_attr_rrtime);
 383 }
 384
 385 /*
 386  * Query firmware how many CPUs should be idle
 387  * return -1 on failure
 388  */
 389 static int acpi_pad_pur(acpi_handle handle)
 390 {
 391         struct acpi_buffer buffer = {ACPI_ALLOCATE_BUFFER, NULL};
 392         union acpi_object *package;
 393         int num = -1;
 394
 395         if (ACPI_FAILURE(acpi_evaluate_object(handle, "_PUR", NULL, &buffer)))
 396                 return num;
 397
 398         if (!buffer.length || !buffer.pointer)
 399                 return num;
 400
 401         package = buffer.pointer;
 402
 403         if (package->type == ACPI_TYPE_PACKAGE &&
 404                 package->package.count == 2 &&
 405                 package->package.elements[0].integer.value == 1) /* rev 1 */
 406
 407                 num = package->package.elements[1].integer.value;
 408
 409         kfree(buffer.pointer);
 410         return num;
 411 }
 412
 413 /* Notify firmware how many CPUs are idle */
 414 static void acpi_pad_ost(acpi_handle handle, int stat,
 415         uint32_t idle_cpus)
 416 {
 417         union acpi_object params[3] = {
 418                 {.type = ACPI_TYPE_INTEGER,},
 419                 {.type = ACPI_TYPE_INTEGER,},
 420                 {.type = ACPI_TYPE_BUFFER,},
 421         };
 422         struct acpi_object_list arg_list = {3, params};
 423
 424         params[0].integer.value = ACPI_PROCESSOR_AGGREGATOR_NOTIFY;
 425         params[1].integer.value =  stat;
 426         params[2].buffer.length = 4;
 427         params[2].buffer.pointer = (void *)&idle_cpus;
 428         acpi_evaluate_object(handle, "_OST", &arg_list, NULL);
 429 }
 430
 431 static void acpi_pad_handle_notify(acpi_handle handle)
 432 {
 433         int num_cpus;
 434         uint32_t idle_cpus;
 435
 436         mutex_lock(&isolated_cpus_lock);
 437         num_cpus = acpi_pad_pur(handle);
 438         if (num_cpus < 0) {
 439                 mutex_unlock(&isolated_cpus_lock);
 440                 return;
 441         }
 442         acpi_pad_idle_cpus(num_cpus);
 443         idle_cpus = acpi_pad_idle_cpus_num();
 444         acpi_pad_ost(handle, 0, idle_cpus);
 445         mutex_unlock(&isolated_cpus_lock);
 446 }
 447
 448 static void acpi_pad_notify(acpi_handle handle, u32 event,
 449         void *data)
 450 {
 451         struct acpi_device *device = data;
 452
 453         switch (event) {
 454         case ACPI_PROCESSOR_AGGREGATOR_NOTIFY:
 455                 acpi_pad_handle_notify(handle);
 456                 acpi_bus_generate_proc_event(device, event, 0);
 457                 acpi_bus_generate_netlink_event(device->pnp.device_class,
 458                         dev_name(&device->dev), event, 0);
 459                 break;
 460         default:
 461                 pr_warn("Unsupported event [0x%x]\n", event);
 462                 break;
 463         }
 464 }
 465
 466 static int acpi_pad_add(struct acpi_device *device)
 467 {
 468         acpi_status status;
 469
 470         strcpy(acpi_device_name(device), ACPI_PROCESSOR_AGGREGATOR_DEVICE_NAME);
 471         strcpy(acpi_device_class(device), ACPI_PROCESSOR_AGGREGATOR_CLASS);
 472
 473         if (acpi_pad_add_sysfs(device))
 474                 return -ENODEV;
 475
 476         status = acpi_install_notify_handler(device->handle,
 477                 ACPI_DEVICE_NOTIFY, acpi_pad_notify, device);
 478         if (ACPI_FAILURE(status)) {
 479                 acpi_pad_remove_sysfs(device);
 480                 return -ENODEV;
 481         }
 482
 483         return 0;
 484 }
 485
 486 static int acpi_pad_remove(struct acpi_device *device)
 487 {
 488         mutex_lock(&isolated_cpus_lock);
 489         acpi_pad_idle_cpus(0);
 490         mutex_unlock(&isolated_cpus_lock);
 491
 492         acpi_remove_notify_handler(device->handle,
 493                 ACPI_DEVICE_NOTIFY, acpi_pad_notify);
 494         acpi_pad_remove_sysfs(device);
 495         return 0;
 496 }
 497
 498 static const struct acpi_device_id pad_device_ids[] = {
 499         {"ACPI000C", 0},
 500         {"", 0},
 501 };
 502 MODULE_DEVICE_TABLE(acpi, pad_device_ids);
 503
 504 static struct acpi_driver acpi_pad_driver = {
 505         .name = "processor_aggregator",
 506         .class = ACPI_PROCESSOR_AGGREGATOR_CLASS,
 507         .ids = pad_device_ids,
 508         .ops = {
 509                 .add = acpi_pad_add,
 510                 .remove = acpi_pad_remove,
 511         },
 512 };
 513
 514 static int __init acpi_pad_init(void)
 515 {
 516         power_saving_mwait_init();
 517         if (power_saving_mwait_eax == 0)
 518                 return -EINVAL;
 519
 520         return acpi_bus_register_driver(&acpi_pad_driver);
 521 }
 522
 523 static void __exit acpi_pad_exit(void)
 524 {
 525         acpi_bus_unregister_driver(&acpi_pad_driver);
 526 }
 527
 528 module_init(acpi_pad_init);
 529 module_exit(acpi_pad_exit);
 530 MODULE_AUTHOR("Shaohua Li<shaohua.li@intel.com>");
 531 MODULE_DESCRIPTION("ACPI Processor Aggregator Driver");
 532 MODULE_LICENSE("GPL");