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
 158         while (!kthread_should_stop()) {
 159                 int cpu;
 160                 u64 expire_time;
 161
 162                 try_to_freeze();
 163
 164                 /* round robin to cpus */
 165                 if (last_jiffies + round_robin_time * HZ < jiffies) {
 166                         last_jiffies = jiffies;
 167                         round_robin_cpu(tsk_index);
 168                 }
 169
 170                 do_sleep = 0;
 171
 172                 expire_time = jiffies + HZ * (100 - idle_pct) / 100;
 173
 174                 while (!need_resched()) {
 175                         if (tsc_detected_unstable && !tsc_marked_unstable) {
 176                                 /* TSC could halt in idle, so notify users */
 177                                 mark_tsc_unstable("TSC halts in idle");
 178                                 tsc_marked_unstable = 1;
 179                         }
 180                         if (lapic_detected_unstable && !lapic_marked_unstable) {
 181                                 int i;
 182                                 /* LAPIC could halt in idle, so notify users */
 183                                 for_each_online_cpu(i)
 184                                         clockevents_notify(
 185                                                 CLOCK_EVT_NOTIFY_BROADCAST_ON,
 186                                                 &i);
 187                                 lapic_marked_unstable = 1;
 188                         }
 189                         local_irq_disable();
 190                         cpu = smp_processor_id();
 191                         if (lapic_marked_unstable)
 192                                 clockevents_notify(
 193                                         CLOCK_EVT_NOTIFY_BROADCAST_ENTER, &cpu);
 194                         stop_critical_timings();
 195
 196                         __monitor((void *)&current_thread_info()->flags, 0, 0);
 197                         smp_mb();
 198                         if (!need_resched())
 199                                 __mwait(power_saving_mwait_eax, 1);
 200
 201                         start_critical_timings();
 202                         if (lapic_marked_unstable)
 203                                 clockevents_notify(
 204                                         CLOCK_EVT_NOTIFY_BROADCAST_EXIT, &cpu);
 205                         local_irq_enable();
 206
 207                         if (jiffies > expire_time) {
 208                                 do_sleep = 1;
 209                                 break;
 210                         }
 211                 }
 212
 213                 /*
 214                  * current sched_rt has threshold for rt task running time.
 215                  * When a rt task uses 95% CPU time, the rt thread will be
 216                  * scheduled out for 5% CPU time to not starve other tasks. But
 217                  * the mechanism only works when all CPUs have RT task running,
 218                  * as if one CPU hasn't RT task, RT task from other CPUs will
 219                  * borrow CPU time from this CPU and cause RT task use > 95%
 220                  * CPU time. To make 'avoid starvation' work, takes a nap here.
 221                  */
 222                 if (do_sleep)
 223                         schedule_timeout_killable(HZ * idle_pct / 100);
 224         }
 225
 226         exit_round_robin(tsk_index);
 227         return 0;
 228 }
 229
 230 static struct task_struct *ps_tsks[NR_CPUS];
 231 static unsigned int ps_tsk_num;
 232 static int create_power_saving_task(void)
 233 {
 234         int rc;
 235
 236         ps_tsks[ps_tsk_num] = kthread_run(power_saving_thread,
 237                 (void *)(unsigned long)ps_tsk_num,
 238                 "acpi_pad/%d", ps_tsk_num);
 239
 240         if (IS_ERR(ps_tsks[ps_tsk_num])) {
 241                 rc = PTR_ERR(ps_tsks[ps_tsk_num]);
 242                 ps_tsks[ps_tsk_num] = NULL;
 243         } else {
 244                 rc = 0;
 245                 ps_tsk_num++;
 246         }
 247
 248         return rc;
 249 }
 250
 251 static void destroy_power_saving_task(void)
 252 {
 253         if (ps_tsk_num > 0) {
 254                 ps_tsk_num--;
 255                 kthread_stop(ps_tsks[ps_tsk_num]);
 256                 ps_tsks[ps_tsk_num] = NULL;
 257         }
 258 }
 259
 260 static void set_power_saving_task_num(unsigned int num)
 261 {
 262         if (num > ps_tsk_num) {
 263                 while (ps_tsk_num < num) {
 264                         if (create_power_saving_task())
 265                                 return;
 266                 }
 267         } else if (num < ps_tsk_num) {
 268                 while (ps_tsk_num > num)
 269                         destroy_power_saving_task();
 270         }
 271 }
 272
 273 static void acpi_pad_idle_cpus(unsigned int num_cpus)
 274 {
 275         get_online_cpus();
 276
 277         num_cpus = min_t(unsigned int, num_cpus, num_online_cpus());
 278         set_power_saving_task_num(num_cpus);
 279
 280         put_online_cpus();
 281 }
 282
 283 static uint32_t acpi_pad_idle_cpus_num(void)
 284 {
 285         return ps_tsk_num;
 286 }
 287
 288 static ssize_t acpi_pad_rrtime_store(struct device *dev,
 289         struct device_attribute *attr, const char *buf, size_t count)
 290 {
 291         unsigned long num;
 292         if (kstrtoul(buf, 0, &num))
 293                 return -EINVAL;
 294         if (num < 1 || num >= 100)
 295                 return -EINVAL;
 296         mutex_lock(&isolated_cpus_lock);
 297         round_robin_time = num;
 298         mutex_unlock(&isolated_cpus_lock);
 299         return count;
 300 }
 301
 302 static ssize_t acpi_pad_rrtime_show(struct device *dev,
 303         struct device_attribute *attr, char *buf)
 304 {
 305         return scnprintf(buf, PAGE_SIZE, "%d\n", round_robin_time);
 306 }
 307 static DEVICE_ATTR(rrtime, S_IRUGO|S_IWUSR,
 308         acpi_pad_rrtime_show,
 309         acpi_pad_rrtime_store);
 310
 311 static ssize_t acpi_pad_idlepct_store(struct device *dev,
 312         struct device_attribute *attr, const char *buf, size_t count)
 313 {
 314         unsigned long num;
 315         if (kstrtoul(buf, 0, &num))
 316                 return -EINVAL;
 317         if (num < 1 || num >= 100)
 318                 return -EINVAL;
 319         mutex_lock(&isolated_cpus_lock);
 320         idle_pct = num;
 321         mutex_unlock(&isolated_cpus_lock);
 322         return count;
 323 }
 324
 325 static ssize_t acpi_pad_idlepct_show(struct device *dev,
 326         struct device_attribute *attr, char *buf)
 327 {
 328         return scnprintf(buf, PAGE_SIZE, "%d\n", idle_pct);
 329 }
 330 static DEVICE_ATTR(idlepct, S_IRUGO|S_IWUSR,
 331         acpi_pad_idlepct_show,
 332         acpi_pad_idlepct_store);
 333
 334 static ssize_t acpi_pad_idlecpus_store(struct device *dev,
 335         struct device_attribute *attr, const char *buf, size_t count)
 336 {
 337         unsigned long num;
 338         if (kstrtoul(buf, 0, &num))
 339                 return -EINVAL;
 340         mutex_lock(&isolated_cpus_lock);
 341         acpi_pad_idle_cpus(num);
 342         mutex_unlock(&isolated_cpus_lock);
 343         return count;
 344 }
 345
 346 static ssize_t acpi_pad_idlecpus_show(struct device *dev,
 347         struct device_attribute *attr, char *buf)
 348 {
 349         int n = 0;
 350         n = cpumask_scnprintf(buf, PAGE_SIZE-2, to_cpumask(pad_busy_cpus_bits));
 351         buf[n++] = '\n';
 352         buf[n] = '\0';
 353         return n;
 354 }
 355 static DEVICE_ATTR(idlecpus, S_IRUGO|S_IWUSR,
 356         acpi_pad_idlecpus_show,
 357         acpi_pad_idlecpus_store);
 358
 359 static int acpi_pad_add_sysfs(struct acpi_device *device)
 360 {
 361         int result;
 362
 363         result = device_create_file(&device->dev, &dev_attr_idlecpus);
 364         if (result)
 365                 return -ENODEV;
 366         result = device_create_file(&device->dev, &dev_attr_idlepct);
 367         if (result) {
 368                 device_remove_file(&device->dev, &dev_attr_idlecpus);
 369                 return -ENODEV;
 370         }
 371         result = device_create_file(&device->dev, &dev_attr_rrtime);
 372         if (result) {
 373                 device_remove_file(&device->dev, &dev_attr_idlecpus);
 374                 device_remove_file(&device->dev, &dev_attr_idlepct);
 375                 return -ENODEV;
 376         }
 377         return 0;
 378 }
 379
 380 static void acpi_pad_remove_sysfs(struct acpi_device *device)
 381 {
 382         device_remove_file(&device->dev, &dev_attr_idlecpus);
 383         device_remove_file(&device->dev, &dev_attr_idlepct);
 384         device_remove_file(&device->dev, &dev_attr_rrtime);
 385 }
 386
 387 /*
 388  * Query firmware how many CPUs should be idle
 389  * return -1 on failure
 390  */
 391 static int acpi_pad_pur(acpi_handle handle)
 392 {
 393         struct acpi_buffer buffer = {ACPI_ALLOCATE_BUFFER, NULL};
 394         union acpi_object *package;
 395         int num = -1;
 396
 397         if (ACPI_FAILURE(acpi_evaluate_object(handle, "_PUR", NULL, &buffer)))
 398                 return num;
 399
 400         if (!buffer.length || !buffer.pointer)
 401                 return num;
 402
 403         package = buffer.pointer;
 404
 405         if (package->type == ACPI_TYPE_PACKAGE &&
 406                 package->package.count == 2 &&
 407                 package->package.elements[0].integer.value == 1) /* rev 1 */
 408
 409                 num = package->package.elements[1].integer.value;
 410
 411         kfree(buffer.pointer);
 412         return num;
 413 }
 414
 415 /* Notify firmware how many CPUs are idle */
 416 static void acpi_pad_ost(acpi_handle handle, int stat,
 417         uint32_t idle_cpus)
 418 {
 419         union acpi_object params[3] = {
 420                 {.type = ACPI_TYPE_INTEGER,},
 421                 {.type = ACPI_TYPE_INTEGER,},
 422                 {.type = ACPI_TYPE_BUFFER,},
 423         };
 424         struct acpi_object_list arg_list = {3, params};
 425
 426         params[0].integer.value = ACPI_PROCESSOR_AGGREGATOR_NOTIFY;
 427         params[1].integer.value =  stat;
 428         params[2].buffer.length = 4;
 429         params[2].buffer.pointer = (void *)&idle_cpus;
 430         acpi_evaluate_object(handle, "_OST", &arg_list, NULL);
 431 }
 432
 433 static void acpi_pad_handle_notify(acpi_handle handle)
 434 {
 435         int num_cpus;
 436         uint32_t idle_cpus;
 437
 438         mutex_lock(&isolated_cpus_lock);
 439         num_cpus = acpi_pad_pur(handle);
 440         if (num_cpus < 0) {
 441                 mutex_unlock(&isolated_cpus_lock);
 442                 return;
 443         }
 444         acpi_pad_idle_cpus(num_cpus);
 445         idle_cpus = acpi_pad_idle_cpus_num();
 446         acpi_pad_ost(handle, 0, idle_cpus);
 447         mutex_unlock(&isolated_cpus_lock);
 448 }
 449
 450 static void acpi_pad_notify(acpi_handle handle, u32 event,
 451         void *data)
 452 {
 453         struct acpi_device *device = data;
 454
 455         switch (event) {
 456         case ACPI_PROCESSOR_AGGREGATOR_NOTIFY:
 457                 acpi_pad_handle_notify(handle);
 458                 acpi_bus_generate_netlink_event(device->pnp.device_class,
 459                         dev_name(&device->dev), event, 0);
 460                 break;
 461         default:
 462                 pr_warn("Unsupported event [0x%x]\n", event);
 463                 break;
 464         }
 465 }
 466
 467 static int acpi_pad_add(struct acpi_device *device)
 468 {
 469         acpi_status status;
 470
 471         strcpy(acpi_device_name(device), ACPI_PROCESSOR_AGGREGATOR_DEVICE_NAME);
 472         strcpy(acpi_device_class(device), ACPI_PROCESSOR_AGGREGATOR_CLASS);
 473
 474         if (acpi_pad_add_sysfs(device))
 475                 return -ENODEV;
 476
 477         status = acpi_install_notify_handler(device->handle,
 478                 ACPI_DEVICE_NOTIFY, acpi_pad_notify, device);
 479         if (ACPI_FAILURE(status)) {
 480                 acpi_pad_remove_sysfs(device);
 481                 return -ENODEV;
 482         }
 483
 484         return 0;
 485 }
 486
 487 static int acpi_pad_remove(struct acpi_device *device)
 488 {
 489         mutex_lock(&isolated_cpus_lock);
 490         acpi_pad_idle_cpus(0);
 491         mutex_unlock(&isolated_cpus_lock);
 492
 493         acpi_remove_notify_handler(device->handle,
 494                 ACPI_DEVICE_NOTIFY, acpi_pad_notify);
 495         acpi_pad_remove_sysfs(device);
 496         return 0;
 497 }
 498
 499 static const struct acpi_device_id pad_device_ids[] = {
 500         {"ACPI000C", 0},
 501         {"", 0},
 502 };
 503 MODULE_DEVICE_TABLE(acpi, pad_device_ids);
 504
 505 static struct acpi_driver acpi_pad_driver = {
 506         .name = "processor_aggregator",
 507         .class = ACPI_PROCESSOR_AGGREGATOR_CLASS,
 508         .ids = pad_device_ids,
 509         .ops = {
 510                 .add = acpi_pad_add,
 511                 .remove = acpi_pad_remove,
 512         },
 513 };
 514
 515 static int __init acpi_pad_init(void)
 516 {
 517         power_saving_mwait_init();
 518         if (power_saving_mwait_eax == 0)
 519                 return -EINVAL;
 520
 521         return acpi_bus_register_driver(&acpi_pad_driver);
 522 }
 523
 524 static void __exit acpi_pad_exit(void)
 525 {
 526         acpi_bus_unregister_driver(&acpi_pad_driver);
 527 }
 528
 529 module_init(acpi_pad_init);
 530 module_exit(acpi_pad_exit);
 531 MODULE_AUTHOR("Shaohua Li<shaohua.li@intel.com>");
 532 MODULE_DESCRIPTION("ACPI Processor Aggregator Driver");
 533 MODULE_LICENSE("GPL");