SAFE public projects git trees. - safe/jmp/linux-2.6/blob - 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
  50 static void power_saving_mwait_init(void)
  51 {
  52         unsigned int eax, ebx, ecx, edx;
  53         unsigned int highest_cstate = 0;
  54         unsigned int highest_subcstate = 0;
  55         int i;
  56
  57         if (!boot_cpu_has(X86_FEATURE_MWAIT))
  58                 return;
  59         if (boot_cpu_data.cpuid_level < CPUID_MWAIT_LEAF)
  60                 return;
  61
  62         cpuid(CPUID_MWAIT_LEAF, &eax, &ebx, &ecx, &edx);
  63
  64         if (!(ecx & CPUID5_ECX_EXTENSIONS_SUPPORTED) ||
  65             !(ecx & CPUID5_ECX_INTERRUPT_BREAK))
  66                 return;
  67
  68         edx >>= MWAIT_SUBSTATE_SIZE;
  69         for (i = 0; i < 7 && edx; i++, edx >>= MWAIT_SUBSTATE_SIZE) {
  70                 if (edx & MWAIT_SUBSTATE_MASK) {
  71                         highest_cstate = i;
  72                         highest_subcstate = edx & MWAIT_SUBSTATE_MASK;
  73                 }
  74         }
  75         power_saving_mwait_eax = (highest_cstate << MWAIT_SUBSTATE_SIZE) |
  76                 (highest_subcstate - 1);
  77
  78         for_each_online_cpu(i)
  79                 clockevents_notify(CLOCK_EVT_NOTIFY_BROADCAST_ON, &i);
  80
  81 #if defined(CONFIG_GENERIC_TIME) && defined(CONFIG_X86)
  82         switch (boot_cpu_data.x86_vendor) {
  83         case X86_VENDOR_AMD:
  84         case X86_VENDOR_INTEL:
  85                 /*
  86                  * AMD Fam10h TSC will tick in all
  87                  * C/P/S0/S1 states when this bit is set.
  88                  */
  89                 if (boot_cpu_has(X86_FEATURE_NONSTOP_TSC))
  90                         return;
  91
  92                 /*FALL THROUGH*/
  93         default:
  94                 /* TSC could halt in idle */
  95                 tsc_detected_unstable = 1;
  96         }
  97 #endif
  98 }
  99
 100 static unsigned long cpu_weight[NR_CPUS];
 101 static int tsk_in_cpu[NR_CPUS] = {[0 ... NR_CPUS-1] = -1};
 102 static DECLARE_BITMAP(pad_busy_cpus_bits, NR_CPUS);
 103 static void round_robin_cpu(unsigned int tsk_index)
 104 {
 105         struct cpumask *pad_busy_cpus = to_cpumask(pad_busy_cpus_bits);
 106         cpumask_var_t tmp;
 107         int cpu;
 108         unsigned long min_weight = -1;
 109         unsigned long uninitialized_var(preferred_cpu);
 110
 111         if (!alloc_cpumask_var(&tmp, GFP_KERNEL))
 112                 return;
 113
 114         mutex_lock(&isolated_cpus_lock);
 115         cpumask_clear(tmp);
 116         for_each_cpu(cpu, pad_busy_cpus)
 117                 cpumask_or(tmp, tmp, topology_thread_cpumask(cpu));
 118         cpumask_andnot(tmp, cpu_online_mask, tmp);
 119         /* avoid HT sibilings if possible */
 120         if (cpumask_empty(tmp))
 121                 cpumask_andnot(tmp, cpu_online_mask, pad_busy_cpus);
 122         if (cpumask_empty(tmp)) {
 123                 mutex_unlock(&isolated_cpus_lock);
 124                 return;
 125         }
 126         for_each_cpu(cpu, tmp) {
 127                 if (cpu_weight[cpu] < min_weight) {
 128                         min_weight = cpu_weight[cpu];
 129                         preferred_cpu = cpu;
 130                 }
 131         }
 132
 133         if (tsk_in_cpu[tsk_index] != -1)
 134                 cpumask_clear_cpu(tsk_in_cpu[tsk_index], pad_busy_cpus);
 135         tsk_in_cpu[tsk_index] = preferred_cpu;
 136         cpumask_set_cpu(preferred_cpu, pad_busy_cpus);
 137         cpu_weight[preferred_cpu]++;
 138         mutex_unlock(&isolated_cpus_lock);
 139
 140         set_cpus_allowed_ptr(current, cpumask_of(preferred_cpu));
 141 }
 142
 143 static void exit_round_robin(unsigned int tsk_index)
 144 {
 145         struct cpumask *pad_busy_cpus = to_cpumask(pad_busy_cpus_bits);
 146         cpumask_clear_cpu(tsk_in_cpu[tsk_index], pad_busy_cpus);
 147         tsk_in_cpu[tsk_index] = -1;
 148 }
 149
 150 static unsigned int idle_pct = 5; /* percentage */
 151 static unsigned int round_robin_time = 10; /* second */
 152 static int power_saving_thread(void *data)
 153 {
 154         struct sched_param param = {.sched_priority = 1};
 155         int do_sleep;
 156         unsigned int tsk_index = (unsigned long)data;
 157         u64 last_jiffies = 0;
 158
 159         sched_setscheduler(current, SCHED_RR, &param);
 160
 161         while (!kthread_should_stop()) {
 162                 int cpu;
 163                 u64 expire_time;
 164
 165                 try_to_freeze();
 166
 167                 /* round robin to cpus */
 168                 if (last_jiffies + round_robin_time * HZ < jiffies) {
 169                         last_jiffies = jiffies;
 170                         round_robin_cpu(tsk_index);
 171                 }
 172
 173                 do_sleep = 0;
 174
 175                 current_thread_info()->status &= ~TS_POLLING;
 176                 /*
 177                  * TS_POLLING-cleared state must be visible before we test
 178                  * NEED_RESCHED:
 179                  */
 180                 smp_mb();
 181
 182                 expire_time = jiffies + HZ * (100 - idle_pct) / 100;
 183
 184                 while (!need_resched()) {
 185                         if (tsc_detected_unstable && !tsc_marked_unstable) {
 186                                 /* TSC could halt in idle, so notify users */
 187                                 mark_tsc_unstable("TSC halts in idle");
 188                                 tsc_marked_unstable = 1;
 189                         }
 190                         local_irq_disable();
 191                         cpu = smp_processor_id();
 192                         clockevents_notify(CLOCK_EVT_NOTIFY_BROADCAST_ENTER,
 193                                 &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                         clockevents_notify(CLOCK_EVT_NOTIFY_BROADCAST_EXIT,
 203                                 &cpu);
 204                         local_irq_enable();
 205
 206                         if (jiffies > expire_time) {
 207                                 do_sleep = 1;
 208                                 break;
 209                         }
 210                 }
 211
 212                 current_thread_info()->status |= TS_POLLING;
 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                 "power_saving/%d", ps_tsk_num);
 240         rc = IS_ERR(ps_tsks[ps_tsk_num]) ? PTR_ERR(ps_tsks[ps_tsk_num]) : 0;
 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 (strict_strtoul(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", 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 (strict_strtoul(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", 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 (strict_strtoul(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         return cpumask_scnprintf(buf, PAGE_SIZE,
 348                 to_cpumask(pad_busy_cpus_bits));
 349 }
 350 static DEVICE_ATTR(idlecpus, S_IRUGO|S_IWUSR,
 351         acpi_pad_idlecpus_show,
 352         acpi_pad_idlecpus_store);
 353
 354 static int acpi_pad_add_sysfs(struct acpi_device *device)
 355 {
 356         int result;
 357
 358         result = device_create_file(&device->dev, &dev_attr_idlecpus);
 359         if (result)
 360                 return -ENODEV;
 361         result = device_create_file(&device->dev, &dev_attr_idlepct);
 362         if (result) {
 363                 device_remove_file(&device->dev, &dev_attr_idlecpus);
 364                 return -ENODEV;
 365         }
 366         result = device_create_file(&device->dev, &dev_attr_rrtime);
 367         if (result) {
 368                 device_remove_file(&device->dev, &dev_attr_idlecpus);
 369                 device_remove_file(&device->dev, &dev_attr_idlepct);
 370                 return -ENODEV;
 371         }
 372         return 0;
 373 }
 374
 375 static void acpi_pad_remove_sysfs(struct acpi_device *device)
 376 {
 377         device_remove_file(&device->dev, &dev_attr_idlecpus);
 378         device_remove_file(&device->dev, &dev_attr_idlepct);
 379         device_remove_file(&device->dev, &dev_attr_rrtime);
 380 }
 381
 382 /* Query firmware how many CPUs should be idle */
 383 static int acpi_pad_pur(acpi_handle handle, int *num_cpus)
 384 {
 385         struct acpi_buffer buffer = {ACPI_ALLOCATE_BUFFER, NULL};
 386         union acpi_object *package;
 387         int rev, num, ret = -EINVAL;
 388
 389         if (ACPI_FAILURE(acpi_evaluate_object(handle, "_PUR", NULL, &buffer)))
 390                 return -EINVAL;
 391
 392         if (!buffer.length || !buffer.pointer)
 393                 return -EINVAL;
 394
 395         package = buffer.pointer;
 396         if (package->type != ACPI_TYPE_PACKAGE || package->package.count != 2)
 397                 goto out;
 398         rev = package->package.elements[0].integer.value;
 399         num = package->package.elements[1].integer.value;
 400         if (rev != 1 || num < 0)
 401                 goto out;
 402         *num_cpus = num;
 403         ret = 0;
 404 out:
 405         kfree(buffer.pointer);
 406         return ret;
 407 }
 408
 409 /* Notify firmware how many CPUs are idle */
 410 static void acpi_pad_ost(acpi_handle handle, int stat,
 411         uint32_t idle_cpus)
 412 {
 413         union acpi_object params[3] = {
 414                 {.type = ACPI_TYPE_INTEGER,},
 415                 {.type = ACPI_TYPE_INTEGER,},
 416                 {.type = ACPI_TYPE_BUFFER,},
 417         };
 418         struct acpi_object_list arg_list = {3, params};
 419
 420         params[0].integer.value = ACPI_PROCESSOR_AGGREGATOR_NOTIFY;
 421         params[1].integer.value =  stat;
 422         params[2].buffer.length = 4;
 423         params[2].buffer.pointer = (void *)&idle_cpus;
 424         acpi_evaluate_object(handle, "_OST", &arg_list, NULL);
 425 }
 426
 427 static void acpi_pad_handle_notify(acpi_handle handle)
 428 {
 429         int num_cpus;
 430         uint32_t idle_cpus;
 431
 432         mutex_lock(&isolated_cpus_lock);
 433         if (acpi_pad_pur(handle, &num_cpus)) {
 434                 mutex_unlock(&isolated_cpus_lock);
 435                 return;
 436         }
 437         acpi_pad_idle_cpus(num_cpus);
 438         idle_cpus = acpi_pad_idle_cpus_num();
 439         acpi_pad_ost(handle, 0, idle_cpus);
 440         mutex_unlock(&isolated_cpus_lock);
 441 }
 442
 443 static void acpi_pad_notify(acpi_handle handle, u32 event,
 444         void *data)
 445 {
 446         struct acpi_device *device = data;
 447
 448         switch (event) {
 449         case ACPI_PROCESSOR_AGGREGATOR_NOTIFY:
 450                 acpi_pad_handle_notify(handle);
 451                 acpi_bus_generate_proc_event(device, event, 0);
 452                 acpi_bus_generate_netlink_event(device->pnp.device_class,
 453                         dev_name(&device->dev), event, 0);
 454                 break;
 455         default:
 456                 printk(KERN_WARNING"Unsupported event [0x%x]\n", event);
 457                 break;
 458         }
 459 }
 460
 461 static int acpi_pad_add(struct acpi_device *device)
 462 {
 463         acpi_status status;
 464
 465         strcpy(acpi_device_name(device), ACPI_PROCESSOR_AGGREGATOR_DEVICE_NAME);
 466         strcpy(acpi_device_class(device), ACPI_PROCESSOR_AGGREGATOR_CLASS);
 467
 468         if (acpi_pad_add_sysfs(device))
 469                 return -ENODEV;
 470
 471         status = acpi_install_notify_handler(device->handle,
 472                 ACPI_DEVICE_NOTIFY, acpi_pad_notify, device);
 473         if (ACPI_FAILURE(status)) {
 474                 acpi_pad_remove_sysfs(device);
 475                 return -ENODEV;
 476         }
 477
 478         return 0;
 479 }
 480
 481 static int acpi_pad_remove(struct acpi_device *device,
 482         int type)
 483 {
 484         mutex_lock(&isolated_cpus_lock);
 485         acpi_pad_idle_cpus(0);
 486         mutex_unlock(&isolated_cpus_lock);
 487
 488         acpi_remove_notify_handler(device->handle,
 489                 ACPI_DEVICE_NOTIFY, acpi_pad_notify);
 490         acpi_pad_remove_sysfs(device);
 491         return 0;
 492 }
 493
 494 static const struct acpi_device_id pad_device_ids[] = {
 495         {"ACPI000C", 0},
 496         {"", 0},
 497 };
 498 MODULE_DEVICE_TABLE(acpi, pad_device_ids);
 499
 500 static struct acpi_driver acpi_pad_driver = {
 501         .name = "processor_aggregator",
 502         .class = ACPI_PROCESSOR_AGGREGATOR_CLASS,
 503         .ids = pad_device_ids,
 504         .ops = {
 505                 .add = acpi_pad_add,
 506                 .remove = acpi_pad_remove,
 507         },
 508 };
 509
 510 static int __init acpi_pad_init(void)
 511 {
 512         power_saving_mwait_init();
 513         if (power_saving_mwait_eax == 0)
 514                 return -EINVAL;
 515
 516         return acpi_bus_register_driver(&acpi_pad_driver);
 517 }
 518
 519 static void __exit acpi_pad_exit(void)
 520 {
 521         acpi_bus_unregister_driver(&acpi_pad_driver);
 522 }
 523
 524 module_init(acpi_pad_init);
 525 module_exit(acpi_pad_exit);
 526 MODULE_AUTHOR("Shaohua Li<shaohua.li@intel.com>");
 527 MODULE_DESCRIPTION("ACPI Processor Aggregator Driver");
 528 MODULE_LICENSE("GPL");