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  */
  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 <linux/freezer.h>
  24 #include <linux/cpu.h>
  25 #include <linux/tick.h>
  26 #include <linux/slab.h>
  27 #include <linux/acpi.h>
  28 #include <asm/mwait.h>
  29
  30 #define ACPI_PROCESSOR_AGGREGATOR_CLASS "acpi_pad"
  31 #define ACPI_PROCESSOR_AGGREGATOR_DEVICE_NAME "Processor Aggregator"
  32 #define ACPI_PROCESSOR_AGGREGATOR_NOTIFY 0x80
  33 static DEFINE_MUTEX(isolated_cpus_lock);
  34 static DEFINE_MUTEX(round_robin_lock);
  35
  36 static unsigned long power_saving_mwait_eax;
  37
  38 static unsigned char tsc_detected_unstable;
  39 static unsigned char tsc_marked_unstable;
  40
  41 static void power_saving_mwait_init(void)
  42 {
  43         unsigned int eax, ebx, ecx, edx;
  44         unsigned int highest_cstate = 0;
  45         unsigned int highest_subcstate = 0;
  46         int i;
  47
  48         if (!boot_cpu_has(X86_FEATURE_MWAIT))
  49                 return;
  50         if (boot_cpu_data.cpuid_level < CPUID_MWAIT_LEAF)
  51                 return;
  52
  53         cpuid(CPUID_MWAIT_LEAF, &eax, &ebx, &ecx, &edx);
  54
  55         if (!(ecx & CPUID5_ECX_EXTENSIONS_SUPPORTED) ||
  56             !(ecx & CPUID5_ECX_INTERRUPT_BREAK))
  57                 return;
  58
  59         edx >>= MWAIT_SUBSTATE_SIZE;
  60         for (i = 0; i < 7 && edx; i++, edx >>= MWAIT_SUBSTATE_SIZE) {
  61                 if (edx & MWAIT_SUBSTATE_MASK) {
  62                         highest_cstate = i;
  63                         highest_subcstate = edx & MWAIT_SUBSTATE_MASK;
  64                 }
  65         }
  66         power_saving_mwait_eax = (highest_cstate << MWAIT_SUBSTATE_SIZE) |
  67                 (highest_subcstate - 1);
  68
  69 #if defined(CONFIG_X86)
  70         switch (boot_cpu_data.x86_vendor) {
  71         case X86_VENDOR_AMD:
  72         case X86_VENDOR_INTEL:
  73                 /*
  74                  * AMD Fam10h TSC will tick in all
  75                  * C/P/S0/S1 states when this bit is set.
  76                  */
  77                 if (!boot_cpu_has(X86_FEATURE_NONSTOP_TSC))
  78                         tsc_detected_unstable = 1;
  79                 break;
  80         default:
  81                 /* TSC could halt in idle */
  82                 tsc_detected_unstable = 1;
  83         }
  84 #endif
  85 }
  86
  87 static unsigned long cpu_weight[NR_CPUS];
  88 static int tsk_in_cpu[NR_CPUS] = {[0 ... NR_CPUS-1] = -1};
  89 static DECLARE_BITMAP(pad_busy_cpus_bits, NR_CPUS);
  90 static void round_robin_cpu(unsigned int tsk_index)
  91 {
  92         struct cpumask *pad_busy_cpus = to_cpumask(pad_busy_cpus_bits);
  93         cpumask_var_t tmp;
  94         int cpu;
  95         unsigned long min_weight = -1;
  96         unsigned long uninitialized_var(preferred_cpu);
  97
  98         if (!alloc_cpumask_var(&tmp, GFP_KERNEL))
  99                 return;
 100
 101         mutex_lock(&round_robin_lock);
 102         cpumask_clear(tmp);
 103         for_each_cpu(cpu, pad_busy_cpus)
 104                 cpumask_or(tmp, tmp, topology_sibling_cpumask(cpu));
 105         cpumask_andnot(tmp, cpu_online_mask, tmp);
 106         /* avoid HT sibilings if possible */
 107         if (cpumask_empty(tmp))
 108                 cpumask_andnot(tmp, cpu_online_mask, pad_busy_cpus);
 109         if (cpumask_empty(tmp)) {
 110                 mutex_unlock(&round_robin_lock);
 111                 free_cpumask_var(tmp);
 112                 return;
 113         }
 114         for_each_cpu(cpu, tmp) {
 115                 if (cpu_weight[cpu] < min_weight) {
 116                         min_weight = cpu_weight[cpu];
 117                         preferred_cpu = cpu;
 118                 }
 119         }
 120
 121         if (tsk_in_cpu[tsk_index] != -1)
 122                 cpumask_clear_cpu(tsk_in_cpu[tsk_index], pad_busy_cpus);
 123         tsk_in_cpu[tsk_index] = preferred_cpu;
 124         cpumask_set_cpu(preferred_cpu, pad_busy_cpus);
 125         cpu_weight[preferred_cpu]++;
 126         mutex_unlock(&round_robin_lock);
 127
 128         set_cpus_allowed_ptr(current, cpumask_of(preferred_cpu));
 129
 130         free_cpumask_var(tmp);
 131 }
 132
 133 static void exit_round_robin(unsigned int tsk_index)
 134 {
 135         struct cpumask *pad_busy_cpus = to_cpumask(pad_busy_cpus_bits);
 136         cpumask_clear_cpu(tsk_in_cpu[tsk_index], pad_busy_cpus);
 137         tsk_in_cpu[tsk_index] = -1;
 138 }
 139
 140 static unsigned int idle_pct = 5; /* percentage */
 141 static unsigned int round_robin_time = 1; /* second */
 142 static int power_saving_thread(void *data)
 143 {
 144         struct sched_param param = {.sched_priority = 1};
 145         int do_sleep;
 146         unsigned int tsk_index = (unsigned long)data;
 147         u64 last_jiffies = 0;
 148
 149         sched_setscheduler(current, SCHED_RR, &param);
 150
 151         while (!kthread_should_stop()) {
 152                 unsigned long expire_time;
 153
 154                 /* round robin to cpus */
 155                 expire_time = last_jiffies + round_robin_time * HZ;
 156                 if (time_before(expire_time, jiffies)) {
 157                         last_jiffies = jiffies;
 158                         round_robin_cpu(tsk_index);
 159                 }
 160
 161                 do_sleep = 0;
 162
 163                 expire_time = jiffies + HZ * (100 - idle_pct) / 100;
 164
 165                 while (!need_resched()) {
 166                         if (tsc_detected_unstable && !tsc_marked_unstable) {
 167                                 /* TSC could halt in idle, so notify users */
 168                                 mark_tsc_unstable("TSC halts in idle");
 169                                 tsc_marked_unstable = 1;
 170                         }
 171                         local_irq_disable();
 172                         tick_broadcast_enable();
 173                         tick_broadcast_enter();
 174                         stop_critical_timings();
 175
 176                         mwait_idle_with_hints(power_saving_mwait_eax, 1);
 177
 178                         start_critical_timings();
 179                         tick_broadcast_exit();
 180                         local_irq_enable();
 181
 182                         if (time_before(expire_time, jiffies)) {
 183                                 do_sleep = 1;
 184                                 break;
 185                         }
 186                 }
 187
 188                 /*
 189                  * current sched_rt has threshold for rt task running time.
 190                  * When a rt task uses 95% CPU time, the rt thread will be
 191                  * scheduled out for 5% CPU time to not starve other tasks. But
 192                  * the mechanism only works when all CPUs have RT task running,
 193                  * as if one CPU hasn't RT task, RT task from other CPUs will
 194                  * borrow CPU time from this CPU and cause RT task use > 95%
 195                  * CPU time. To make 'avoid starvation' work, takes a nap here.
 196                  */
 197                 if (unlikely(do_sleep))
 198                         schedule_timeout_killable(HZ * idle_pct / 100);
 199
 200                 /* If an external event has set the need_resched flag, then
 201                  * we need to deal with it, or this loop will continue to
 202                  * spin without calling __mwait().
 203                  */
 204                 if (unlikely(need_resched()))
 205                         schedule();
 206         }
 207
 208         exit_round_robin(tsk_index);
 209         return 0;
 210 }
 211
 212 static struct task_struct *ps_tsks[NR_CPUS];
 213 static unsigned int ps_tsk_num;
 214 static int create_power_saving_task(void)
 215 {
 216         int rc;
 217
 218         ps_tsks[ps_tsk_num] = kthread_run(power_saving_thread,
 219                 (void *)(unsigned long)ps_tsk_num,
 220                 "acpi_pad/%d", ps_tsk_num);
 221
 222         if (IS_ERR(ps_tsks[ps_tsk_num])) {
 223                 rc = PTR_ERR(ps_tsks[ps_tsk_num]);
 224                 ps_tsks[ps_tsk_num] = NULL;
 225         } else {
 226                 rc = 0;
 227                 ps_tsk_num++;
 228         }
 229
 230         return rc;
 231 }
 232
 233 static void destroy_power_saving_task(void)
 234 {
 235         if (ps_tsk_num > 0) {
 236                 ps_tsk_num--;
 237                 kthread_stop(ps_tsks[ps_tsk_num]);
 238                 ps_tsks[ps_tsk_num] = NULL;
 239         }
 240 }
 241
 242 static void set_power_saving_task_num(unsigned int num)
 243 {
 244         if (num > ps_tsk_num) {
 245                 while (ps_tsk_num < num) {
 246                         if (create_power_saving_task())
 247                                 return;
 248                 }
 249         } else if (num < ps_tsk_num) {
 250                 while (ps_tsk_num > num)
 251                         destroy_power_saving_task();
 252         }
 253 }
 254
 255 static void acpi_pad_idle_cpus(unsigned int num_cpus)
 256 {
 257         get_online_cpus();
 258
 259         num_cpus = min_t(unsigned int, num_cpus, num_online_cpus());
 260         set_power_saving_task_num(num_cpus);
 261
 262         put_online_cpus();
 263 }
 264
 265 static uint32_t acpi_pad_idle_cpus_num(void)
 266 {
 267         return ps_tsk_num;
 268 }
 269
 270 static ssize_t acpi_pad_rrtime_store(struct device *dev,
 271         struct device_attribute *attr, const char *buf, size_t count)
 272 {
 273         unsigned long num;
 274         if (kstrtoul(buf, 0, &num))
 275                 return -EINVAL;
 276         if (num < 1 || num >= 100)
 277                 return -EINVAL;
 278         mutex_lock(&isolated_cpus_lock);
 279         round_robin_time = num;
 280         mutex_unlock(&isolated_cpus_lock);
 281         return count;
 282 }
 283
 284 static ssize_t acpi_pad_rrtime_show(struct device *dev,
 285         struct device_attribute *attr, char *buf)
 286 {
 287         return scnprintf(buf, PAGE_SIZE, "%d\n", round_robin_time);
 288 }
 289 static DEVICE_ATTR(rrtime, S_IRUGO|S_IWUSR,
 290         acpi_pad_rrtime_show,
 291         acpi_pad_rrtime_store);
 292
 293 static ssize_t acpi_pad_idlepct_store(struct device *dev,
 294         struct device_attribute *attr, const char *buf, size_t count)
 295 {
 296         unsigned long num;
 297         if (kstrtoul(buf, 0, &num))
 298                 return -EINVAL;
 299         if (num < 1 || num >= 100)
 300                 return -EINVAL;
 301         mutex_lock(&isolated_cpus_lock);
 302         idle_pct = num;
 303         mutex_unlock(&isolated_cpus_lock);
 304         return count;
 305 }
 306
 307 static ssize_t acpi_pad_idlepct_show(struct device *dev,
 308         struct device_attribute *attr, char *buf)
 309 {
 310         return scnprintf(buf, PAGE_SIZE, "%d\n", idle_pct);
 311 }
 312 static DEVICE_ATTR(idlepct, S_IRUGO|S_IWUSR,
 313         acpi_pad_idlepct_show,
 314         acpi_pad_idlepct_store);
 315
 316 static ssize_t acpi_pad_idlecpus_store(struct device *dev,
 317         struct device_attribute *attr, const char *buf, size_t count)
 318 {
 319         unsigned long num;
 320         if (kstrtoul(buf, 0, &num))
 321                 return -EINVAL;
 322         mutex_lock(&isolated_cpus_lock);
 323         acpi_pad_idle_cpus(num);
 324         mutex_unlock(&isolated_cpus_lock);
 325         return count;
 326 }
 327
 328 static ssize_t acpi_pad_idlecpus_show(struct device *dev,
 329         struct device_attribute *attr, char *buf)
 330 {
 331         return cpumap_print_to_pagebuf(false, buf,
 332                                        to_cpumask(pad_busy_cpus_bits));
 333 }
 334
 335 static DEVICE_ATTR(idlecpus, S_IRUGO|S_IWUSR,
 336         acpi_pad_idlecpus_show,
 337         acpi_pad_idlecpus_store);
 338
 339 static int acpi_pad_add_sysfs(struct acpi_device *device)
 340 {
 341         int result;
 342
 343         result = device_create_file(&device->dev, &dev_attr_idlecpus);
 344         if (result)
 345                 return -ENODEV;
 346         result = device_create_file(&device->dev, &dev_attr_idlepct);
 347         if (result) {
 348                 device_remove_file(&device->dev, &dev_attr_idlecpus);
 349                 return -ENODEV;
 350         }
 351         result = device_create_file(&device->dev, &dev_attr_rrtime);
 352         if (result) {
 353                 device_remove_file(&device->dev, &dev_attr_idlecpus);
 354                 device_remove_file(&device->dev, &dev_attr_idlepct);
 355                 return -ENODEV;
 356         }
 357         return 0;
 358 }
 359
 360 static void acpi_pad_remove_sysfs(struct acpi_device *device)
 361 {
 362         device_remove_file(&device->dev, &dev_attr_idlecpus);
 363         device_remove_file(&device->dev, &dev_attr_idlepct);
 364         device_remove_file(&device->dev, &dev_attr_rrtime);
 365 }
 366
 367 /*
 368  * Query firmware how many CPUs should be idle
 369  * return -1 on failure
 370  */
 371 static int acpi_pad_pur(acpi_handle handle)
 372 {
 373         struct acpi_buffer buffer = {ACPI_ALLOCATE_BUFFER, NULL};
 374         union acpi_object *package;
 375         int num = -1;
 376
 377         if (ACPI_FAILURE(acpi_evaluate_object(handle, "_PUR", NULL, &buffer)))
 378                 return num;
 379
 380         if (!buffer.length || !buffer.pointer)
 381                 return num;
 382
 383         package = buffer.pointer;
 384
 385         if (package->type == ACPI_TYPE_PACKAGE &&
 386                 package->package.count == 2 &&
 387                 package->package.elements[0].integer.value == 1) /* rev 1 */
 388
 389                 num = package->package.elements[1].integer.value;
 390
 391         kfree(buffer.pointer);
 392         return num;
 393 }
 394
 395 static void acpi_pad_handle_notify(acpi_handle handle)
 396 {
 397         int num_cpus;
 398         uint32_t idle_cpus;
 399         struct acpi_buffer param = {
 400                 .length = 4,
 401                 .pointer = (void *)&idle_cpus,
 402         };
 403
 404         mutex_lock(&isolated_cpus_lock);
 405         num_cpus = acpi_pad_pur(handle);
 406         if (num_cpus < 0) {
 407                 mutex_unlock(&isolated_cpus_lock);
 408                 return;
 409         }
 410         acpi_pad_idle_cpus(num_cpus);
 411         idle_cpus = acpi_pad_idle_cpus_num();
 412         acpi_evaluate_ost(handle, ACPI_PROCESSOR_AGGREGATOR_NOTIFY, 0, &param);
 413         mutex_unlock(&isolated_cpus_lock);
 414 }
 415
 416 static void acpi_pad_notify(acpi_handle handle, u32 event,
 417         void *data)
 418 {
 419         struct acpi_device *device = data;
 420
 421         switch (event) {
 422         case ACPI_PROCESSOR_AGGREGATOR_NOTIFY:
 423                 acpi_pad_handle_notify(handle);
 424                 acpi_bus_generate_netlink_event(device->pnp.device_class,
 425                         dev_name(&device->dev), event, 0);
 426                 break;
 427         default:
 428                 pr_warn("Unsupported event [0x%x]\n", event);
 429                 break;
 430         }
 431 }
 432
 433 static int acpi_pad_add(struct acpi_device *device)
 434 {
 435         acpi_status status;
 436
 437         strcpy(acpi_device_name(device), ACPI_PROCESSOR_AGGREGATOR_DEVICE_NAME);
 438         strcpy(acpi_device_class(device), ACPI_PROCESSOR_AGGREGATOR_CLASS);
 439
 440         if (acpi_pad_add_sysfs(device))
 441                 return -ENODEV;
 442
 443         status = acpi_install_notify_handler(device->handle,
 444                 ACPI_DEVICE_NOTIFY, acpi_pad_notify, device);
 445         if (ACPI_FAILURE(status)) {
 446                 acpi_pad_remove_sysfs(device);
 447                 return -ENODEV;
 448         }
 449
 450         return 0;
 451 }
 452
 453 static int acpi_pad_remove(struct acpi_device *device)
 454 {
 455         mutex_lock(&isolated_cpus_lock);
 456         acpi_pad_idle_cpus(0);
 457         mutex_unlock(&isolated_cpus_lock);
 458
 459         acpi_remove_notify_handler(device->handle,
 460                 ACPI_DEVICE_NOTIFY, acpi_pad_notify);
 461         acpi_pad_remove_sysfs(device);
 462         return 0;
 463 }
 464
 465 static const struct acpi_device_id pad_device_ids[] = {
 466         {"ACPI000C", 0},
 467         {"", 0},
 468 };
 469 MODULE_DEVICE_TABLE(acpi, pad_device_ids);
 470
 471 static struct acpi_driver acpi_pad_driver = {
 472         .name = "processor_aggregator",
 473         .class = ACPI_PROCESSOR_AGGREGATOR_CLASS,
 474         .ids = pad_device_ids,
 475         .ops = {
 476                 .add = acpi_pad_add,
 477                 .remove = acpi_pad_remove,
 478         },
 479 };
 480
 481 static int __init acpi_pad_init(void)
 482 {
 483         power_saving_mwait_init();
 484         if (power_saving_mwait_eax == 0)
 485                 return -EINVAL;
 486
 487         return acpi_bus_register_driver(&acpi_pad_driver);
 488 }
 489
 490 static void __exit acpi_pad_exit(void)
 491 {
 492         acpi_bus_unregister_driver(&acpi_pad_driver);
 493 }
 494
 495 module_init(acpi_pad_init);
 496 module_exit(acpi_pad_exit);
 497 MODULE_AUTHOR("Shaohua Li<shaohua.li@intel.com>");
 498 MODULE_DESCRIPTION("ACPI Processor Aggregator Driver");
 499 MODULE_LICENSE("GPL");