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