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
  34 #define ACPI_PROCESSOR_AGGREGATOR_CLASS "acpi_pad"
  35 #define ACPI_PROCESSOR_AGGREGATOR_DEVICE_NAME "Processor Aggregator"
  36 #define ACPI_PROCESSOR_AGGREGATOR_NOTIFY 0x80
  37 static DEFINE_MUTEX(isolated_cpus_lock);
  38
  39 #define MWAIT_SUBSTATE_MASK     (0xf)
  40 #define MWAIT_CSTATE_MASK       (0xf)
  41 #define MWAIT_SUBSTATE_SIZE     (4)
  42 #define CPUID_MWAIT_LEAF (5)
  43 #define CPUID5_ECX_EXTENSIONS_SUPPORTED (0x1)
  44 #define CPUID5_ECX_INTERRUPT_BREAK      (0x2)
  45 static unsigned long power_saving_mwait_eax;
  46
  47 static unsigned char tsc_detected_unstable;
  48 static unsigned char tsc_marked_unstable;
  49 static unsigned char lapic_detected_unstable;
  50 static unsigned char lapic_marked_unstable;
  51
  52 static void power_saving_mwait_init(void)
  53 {
  54         unsigned int eax, ebx, ecx, edx;
  55         unsigned int highest_cstate = 0;
  56         unsigned int highest_subcstate = 0;
  57         int i;
  58
  59         if (!boot_cpu_has(X86_FEATURE_MWAIT))
  60                 return;
  61         if (boot_cpu_data.cpuid_level < CPUID_MWAIT_LEAF)
  62                 return;
  63
  64         cpuid(CPUID_MWAIT_LEAF, &eax, &ebx, &ecx, &edx);
  65
  66         if (!(ecx & CPUID5_ECX_EXTENSIONS_SUPPORTED) ||
  67             !(ecx & CPUID5_ECX_INTERRUPT_BREAK))
  68                 return;
  69
  70         edx >>= MWAIT_SUBSTATE_SIZE;
  71         for (i = 0; i < 7 && edx; i++, edx >>= MWAIT_SUBSTATE_SIZE) {
  72                 if (edx & MWAIT_SUBSTATE_MASK) {
  73                         highest_cstate = i;
  74                         highest_subcstate = edx & MWAIT_SUBSTATE_MASK;
  75                 }
  76         }
  77         power_saving_mwait_eax = (highest_cstate << MWAIT_SUBSTATE_SIZE) |
  78                 (highest_subcstate - 1);
  79
  80 #if defined(CONFIG_GENERIC_TIME) && defined(CONFIG_X86)
  81         switch (boot_cpu_data.x86_vendor) {
  82         case X86_VENDOR_AMD:
  83         case X86_VENDOR_INTEL:
  84                 /*
  85                  * AMD Fam10h TSC will tick in all
  86                  * C/P/S0/S1 states when this bit is set.
  87                  */
  88                 if (!boot_cpu_has(X86_FEATURE_NONSTOP_TSC))
  89                         tsc_detected_unstable = 1;
  90                 if (!boot_cpu_has(X86_FEATURE_ARAT))
  91                         lapic_detected_unstable = 1;
  92                 break;
  93         default:
  94                 /* TSC & LAPIC could halt in idle */
  95                 tsc_detected_unstable = 1;
  96                 lapic_detected_unstable = 1;
  97         }
  98 #endif
  99 }
 100
 101 static unsigned long cpu_weight[NR_CPUS];
 102 static int tsk_in_cpu[NR_CPUS] = {[0 ... NR_CPUS-1] = -1};
 103 static DECLARE_BITMAP(pad_busy_cpus_bits, NR_CPUS);
 104 static void round_robin_cpu(unsigned int tsk_index)
 105 {
 106         struct cpumask *pad_busy_cpus = to_cpumask(pad_busy_cpus_bits);
 107         cpumask_var_t tmp;
 108         int cpu;
 109         unsigned long min_weight = -1;
 110         unsigned long uninitialized_var(preferred_cpu);
 111
 112         if (!alloc_cpumask_var(&tmp, GFP_KERNEL))
 113                 return;
 114
 115         mutex_lock(&isolated_cpus_lock);
 116         cpumask_clear(tmp);
 117         for_each_cpu(cpu, pad_busy_cpus)
 118                 cpumask_or(tmp, tmp, topology_thread_cpumask(cpu));
 119         cpumask_andnot(tmp, cpu_online_mask, tmp);
 120         /* avoid HT sibilings if possible */
 121         if (cpumask_empty(tmp))
 122                 cpumask_andnot(tmp, cpu_online_mask, pad_busy_cpus);
 123         if (cpumask_empty(tmp)) {
 124                 mutex_unlock(&isolated_cpus_lock);
 125                 return;
 126         }
 127         for_each_cpu(cpu, tmp) {
 128                 if (cpu_weight[cpu] < min_weight) {
 129                         min_weight = cpu_weight[cpu];
 130                         preferred_cpu = cpu;
 131                 }
 132         }
 133
 134         if (tsk_in_cpu[tsk_index] != -1)
 135                 cpumask_clear_cpu(tsk_in_cpu[tsk_index], pad_busy_cpus);
 136         tsk_in_cpu[tsk_index] = preferred_cpu;
 137         cpumask_set_cpu(preferred_cpu, pad_busy_cpus);
 138         cpu_weight[preferred_cpu]++;
 139         mutex_unlock(&isolated_cpus_lock);
 140
 141         set_cpus_allowed_ptr(current, cpumask_of(preferred_cpu));
 142 }
 143
 144 static void exit_round_robin(unsigned int tsk_index)
 145 {
 146         struct cpumask *pad_busy_cpus = to_cpumask(pad_busy_cpus_bits);
 147         cpumask_clear_cpu(tsk_in_cpu[tsk_index], pad_busy_cpus);
 148         tsk_in_cpu[tsk_index] = -1;
 149 }
 150
 151 static unsigned int idle_pct = 5; /* percentage */
 152 static unsigned int round_robin_time = 10; /* second */
 153 static int power_saving_thread(void *data)
 154 {
 155         struct sched_param param = {.sched_priority = 1};
 156         int do_sleep;
 157         unsigned int tsk_index = (unsigned long)data;
 158         u64 last_jiffies = 0;
 159
 160         sched_setscheduler(current, SCHED_RR, &param);
 161
 162         while (!kthread_should_stop()) {
 163                 int cpu;
 164                 u64 expire_time;
 165
 166                 try_to_freeze();
 167
 168                 /* round robin to cpus */
 169                 if (last_jiffies + round_robin_time * HZ < jiffies) {
 170                         last_jiffies = jiffies;
 171                         round_robin_cpu(tsk_index);
 172                 }
 173
 174                 do_sleep = 0;
 175
 176                 expire_time = jiffies + HZ * (100 - idle_pct) / 100;
 177
 178                 while (!need_resched()) {
 179                         if (tsc_detected_unstable && !tsc_marked_unstable) {
 180                                 /* TSC could halt in idle, so notify users */
 181                                 mark_tsc_unstable("TSC halts in idle");
 182                                 tsc_marked_unstable = 1;
 183                         }
 184                         if (lapic_detected_unstable && !lapic_marked_unstable) {
 185                                 int i;
 186                                 /* LAPIC could halt in idle, so notify users */
 187                                 for_each_online_cpu(i)
 188                                         clockevents_notify(
 189                                                 CLOCK_EVT_NOTIFY_BROADCAST_ON,
 190                                                 &i);
 191                                 lapic_marked_unstable = 1;
 192                         }
 193                         local_irq_disable();
 194                         cpu = smp_processor_id();
 195                         if (lapic_marked_unstable)
 196                                 clockevents_notify(
 197                                         CLOCK_EVT_NOTIFY_BROADCAST_ENTER, &cpu);
 198                         stop_critical_timings();
 199
 200                         __monitor((void *)&current_thread_info()->flags, 0, 0);
 201                         smp_mb();
 202                         if (!need_resched())
 203                                 __mwait(power_saving_mwait_eax, 1);
 204
 205                         start_critical_timings();
 206                         if (lapic_marked_unstable)
 207                                 clockevents_notify(
 208                                         CLOCK_EVT_NOTIFY_BROADCAST_EXIT, &cpu);
 209                         local_irq_enable();
 210
 211                         if (jiffies > expire_time) {
 212                                 do_sleep = 1;
 213                                 break;
 214                         }
 215                 }
 216
 217                 /*
 218                  * current sched_rt has threshold for rt task running time.
 219                  * When a rt task uses 95% CPU time, the rt thread will be
 220                  * scheduled out for 5% CPU time to not starve other tasks. But
 221                  * the mechanism only works when all CPUs have RT task running,
 222                  * as if one CPU hasn't RT task, RT task from other CPUs will
 223                  * borrow CPU time from this CPU and cause RT task use > 95%
 224                  * CPU time. To make 'avoid starvation' work, takes a nap here.
 225                  */
 226                 if (do_sleep)
 227                         schedule_timeout_killable(HZ * idle_pct / 100);
 228         }
 229
 230         exit_round_robin(tsk_index);
 231         return 0;
 232 }
 233
 234 static struct task_struct *ps_tsks[NR_CPUS];
 235 static unsigned int ps_tsk_num;
 236 static int create_power_saving_task(void)
 237 {
 238         int rc = -ENOMEM;
 239
 240         ps_tsks[ps_tsk_num] = kthread_run(power_saving_thread,
 241                 (void *)(unsigned long)ps_tsk_num,
 242                 "power_saving/%d", ps_tsk_num);
 243         rc = IS_ERR(ps_tsks[ps_tsk_num]) ? PTR_ERR(ps_tsks[ps_tsk_num]) : 0;
 244         if (!rc)
 245                 ps_tsk_num++;
 246         else
 247                 ps_tsks[ps_tsk_num] = NULL;
 248
 249         return rc;
 250 }
 251
 252 static void destroy_power_saving_task(void)
 253 {
 254         if (ps_tsk_num > 0) {
 255                 ps_tsk_num--;
 256                 kthread_stop(ps_tsks[ps_tsk_num]);
 257                 ps_tsks[ps_tsk_num] = NULL;
 258         }
 259 }
 260
 261 static void set_power_saving_task_num(unsigned int num)
 262 {
 263         if (num > ps_tsk_num) {
 264                 while (ps_tsk_num < num) {
 265                         if (create_power_saving_task())
 266                                 return;
 267                 }
 268         } else if (num < ps_tsk_num) {
 269                 while (ps_tsk_num > num)
 270                         destroy_power_saving_task();
 271         }
 272 }
 273
 274 static void acpi_pad_idle_cpus(unsigned int num_cpus)
 275 {
 276         get_online_cpus();
 277
 278         num_cpus = min_t(unsigned int, num_cpus, num_online_cpus());
 279         set_power_saving_task_num(num_cpus);
 280
 281         put_online_cpus();
 282 }
 283
 284 static uint32_t acpi_pad_idle_cpus_num(void)
 285 {
 286         return ps_tsk_num;
 287 }
 288
 289 static ssize_t acpi_pad_rrtime_store(struct device *dev,
 290         struct device_attribute *attr, const char *buf, size_t count)
 291 {
 292         unsigned long num;
 293         if (strict_strtoul(buf, 0, &num))
 294                 return -EINVAL;
 295         if (num < 1 || num >= 100)
 296                 return -EINVAL;
 297         mutex_lock(&isolated_cpus_lock);
 298         round_robin_time = num;
 299         mutex_unlock(&isolated_cpus_lock);
 300         return count;
 301 }
 302
 303 static ssize_t acpi_pad_rrtime_show(struct device *dev,
 304         struct device_attribute *attr, char *buf)
 305 {
 306         return scnprintf(buf, PAGE_SIZE, "%d", round_robin_time);
 307 }
 308 static DEVICE_ATTR(rrtime, S_IRUGO|S_IWUSR,
 309         acpi_pad_rrtime_show,
 310         acpi_pad_rrtime_store);
 311
 312 static ssize_t acpi_pad_idlepct_store(struct device *dev,
 313         struct device_attribute *attr, const char *buf, size_t count)
 314 {
 315         unsigned long num;
 316         if (strict_strtoul(buf, 0, &num))
 317                 return -EINVAL;
 318         if (num < 1 || num >= 100)
 319                 return -EINVAL;
 320         mutex_lock(&isolated_cpus_lock);
 321         idle_pct = num;
 322         mutex_unlock(&isolated_cpus_lock);
 323         return count;
 324 }
 325
 326 static ssize_t acpi_pad_idlepct_show(struct device *dev,
 327         struct device_attribute *attr, char *buf)
 328 {
 329         return scnprintf(buf, PAGE_SIZE, "%d", idle_pct);
 330 }
 331 static DEVICE_ATTR(idlepct, S_IRUGO|S_IWUSR,
 332         acpi_pad_idlepct_show,
 333         acpi_pad_idlepct_store);
 334
 335 static ssize_t acpi_pad_idlecpus_store(struct device *dev,
 336         struct device_attribute *attr, const char *buf, size_t count)
 337 {
 338         unsigned long num;
 339         if (strict_strtoul(buf, 0, &num))
 340                 return -EINVAL;
 341         mutex_lock(&isolated_cpus_lock);
 342         acpi_pad_idle_cpus(num);
 343         mutex_unlock(&isolated_cpus_lock);
 344         return count;
 345 }
 346
 347 static ssize_t acpi_pad_idlecpus_show(struct device *dev,
 348         struct device_attribute *attr, char *buf)
 349 {
 350         return cpumask_scnprintf(buf, PAGE_SIZE,
 351                 to_cpumask(pad_busy_cpus_bits));
 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 /* Query firmware how many CPUs should be idle */
 386 static int acpi_pad_pur(acpi_handle handle, int *num_cpus)
 387 {
 388         struct acpi_buffer buffer = {ACPI_ALLOCATE_BUFFER, NULL};
 389         union acpi_object *package;
 390         int rev, num, ret = -EINVAL;
 391
 392         if (ACPI_FAILURE(acpi_evaluate_object(handle, "_PUR", NULL, &buffer)))
 393                 return -EINVAL;
 394
 395         if (!buffer.length || !buffer.pointer)
 396                 return -EINVAL;
 397
 398         package = buffer.pointer;
 399         if (package->type != ACPI_TYPE_PACKAGE || package->package.count != 2)
 400                 goto out;
 401         rev = package->package.elements[0].integer.value;
 402         num = package->package.elements[1].integer.value;
 403         if (rev != 1 || num < 0)
 404                 goto out;
 405         *num_cpus = num;
 406         ret = 0;
 407 out:
 408         kfree(buffer.pointer);
 409         return ret;
 410 }
 411
 412 /* Notify firmware how many CPUs are idle */
 413 static void acpi_pad_ost(acpi_handle handle, int stat,
 414         uint32_t idle_cpus)
 415 {
 416         union acpi_object params[3] = {
 417                 {.type = ACPI_TYPE_INTEGER,},
 418                 {.type = ACPI_TYPE_INTEGER,},
 419                 {.type = ACPI_TYPE_BUFFER,},
 420         };
 421         struct acpi_object_list arg_list = {3, params};
 422
 423         params[0].integer.value = ACPI_PROCESSOR_AGGREGATOR_NOTIFY;
 424         params[1].integer.value =  stat;
 425         params[2].buffer.length = 4;
 426         params[2].buffer.pointer = (void *)&idle_cpus;
 427         acpi_evaluate_object(handle, "_OST", &arg_list, NULL);
 428 }
 429
 430 static void acpi_pad_handle_notify(acpi_handle handle)
 431 {
 432         int num_cpus;
 433         uint32_t idle_cpus;
 434
 435         mutex_lock(&isolated_cpus_lock);
 436         if (acpi_pad_pur(handle, &num_cpus)) {
 437                 mutex_unlock(&isolated_cpus_lock);
 438                 return;
 439         }
 440         acpi_pad_idle_cpus(num_cpus);
 441         idle_cpus = acpi_pad_idle_cpus_num();
 442         acpi_pad_ost(handle, 0, idle_cpus);
 443         mutex_unlock(&isolated_cpus_lock);
 444 }
 445
 446 static void acpi_pad_notify(acpi_handle handle, u32 event,
 447         void *data)
 448 {
 449         struct acpi_device *device = data;
 450
 451         switch (event) {
 452         case ACPI_PROCESSOR_AGGREGATOR_NOTIFY:
 453                 acpi_pad_handle_notify(handle);
 454                 acpi_bus_generate_proc_event(device, event, 0);
 455                 acpi_bus_generate_netlink_event(device->pnp.device_class,
 456                         dev_name(&device->dev), event, 0);
 457                 break;
 458         default:
 459                 printk(KERN_WARNING"Unsupported event [0x%x]\n", event);
 460                 break;
 461         }
 462 }
 463
 464 static int acpi_pad_add(struct acpi_device *device)
 465 {
 466         acpi_status status;
 467
 468         strcpy(acpi_device_name(device), ACPI_PROCESSOR_AGGREGATOR_DEVICE_NAME);
 469         strcpy(acpi_device_class(device), ACPI_PROCESSOR_AGGREGATOR_CLASS);
 470
 471         if (acpi_pad_add_sysfs(device))
 472                 return -ENODEV;
 473
 474         status = acpi_install_notify_handler(device->handle,
 475                 ACPI_DEVICE_NOTIFY, acpi_pad_notify, device);
 476         if (ACPI_FAILURE(status)) {
 477                 acpi_pad_remove_sysfs(device);
 478                 return -ENODEV;
 479         }
 480
 481         return 0;
 482 }
 483
 484 static int acpi_pad_remove(struct acpi_device *device,
 485         int type)
 486 {
 487         mutex_lock(&isolated_cpus_lock);
 488         acpi_pad_idle_cpus(0);
 489         mutex_unlock(&isolated_cpus_lock);
 490
 491         acpi_remove_notify_handler(device->handle,
 492                 ACPI_DEVICE_NOTIFY, acpi_pad_notify);
 493         acpi_pad_remove_sysfs(device);
 494         return 0;
 495 }
 496
 497 static const struct acpi_device_id pad_device_ids[] = {
 498         {"ACPI000C", 0},
 499         {"", 0},
 500 };
 501 MODULE_DEVICE_TABLE(acpi, pad_device_ids);
 502
 503 static struct acpi_driver acpi_pad_driver = {
 504         .name = "processor_aggregator",
 505         .class = ACPI_PROCESSOR_AGGREGATOR_CLASS,
 506         .ids = pad_device_ids,
 507         .ops = {
 508                 .add = acpi_pad_add,
 509                 .remove = acpi_pad_remove,
 510         },
 511 };
 512
 513 static int __init acpi_pad_init(void)
 514 {
 515         power_saving_mwait_init();
 516         if (power_saving_mwait_eax == 0)
 517                 return -EINVAL;
 518
 519         return acpi_bus_register_driver(&acpi_pad_driver);
 520 }
 521
 522 static void __exit acpi_pad_exit(void)
 523 {
 524         acpi_bus_unregister_driver(&acpi_pad_driver);
 525 }
 526
 527 module_init(acpi_pad_init);
 528 module_exit(acpi_pad_exit);
 529 MODULE_AUTHOR("Shaohua Li<shaohua.li@intel.com>");
 530 MODULE_DESCRIPTION("ACPI Processor Aggregator Driver");
 531 MODULE_LICENSE("GPL");