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