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  */
  16
  17 #include <linux/kernel.h>
  18 #include <linux/cpumask.h>
  19 #include <linux/module.h>
  20 #include <linux/init.h>
  21 #include <linux/types.h>
  22 #include <linux/kthread.h>
  23 #include <uapi/linux/sched/types.h>
  24 #include <linux/freezer.h>
  25 #include <linux/cpu.h>
  26 #include <linux/tick.h>
  27 #include <linux/slab.h>
  28 #include <linux/acpi.h>
  29 #include <asm/mwait.h>
  30 #include <xen/xen.h>
  31
  32 #define ACPI_PROCESSOR_AGGREGATOR_CLASS "acpi_pad"
  33 #define ACPI_PROCESSOR_AGGREGATOR_DEVICE_NAME "Processor Aggregator"
  34 #define ACPI_PROCESSOR_AGGREGATOR_NOTIFY 0x80
  35 static DEFINE_MUTEX(isolated_cpus_lock);
  36 static DEFINE_MUTEX(round_robin_lock);
  37
  38 static unsigned long power_saving_mwait_eax;
  39
  40 static unsigned char tsc_detected_unstable;
  41 static unsigned char tsc_marked_unstable;
  42
  43 static void power_saving_mwait_init(void)
  44 {
  45         unsigned int eax, ebx, ecx, edx;
  46         unsigned int highest_cstate = 0;
  47         unsigned int highest_subcstate = 0;
  48         int i;
  49
  50         if (!boot_cpu_has(X86_FEATURE_MWAIT))
  51                 return;
  52         if (boot_cpu_data.cpuid_level < CPUID_MWAIT_LEAF)
  53                 return;
  54
  55         cpuid(CPUID_MWAIT_LEAF, &eax, &ebx, &ecx, &edx);
  56
  57         if (!(ecx & CPUID5_ECX_EXTENSIONS_SUPPORTED) ||
  58             !(ecx & CPUID5_ECX_INTERRUPT_BREAK))
  59                 return;
  60
  61         edx >>= MWAIT_SUBSTATE_SIZE;
  62         for (i = 0; i < 7 && edx; i++, edx >>= MWAIT_SUBSTATE_SIZE) {
  63                 if (edx & MWAIT_SUBSTATE_MASK) {
  64                         highest_cstate = i;
  65                         highest_subcstate = edx & MWAIT_SUBSTATE_MASK;
  66                 }
  67         }
  68         power_saving_mwait_eax = (highest_cstate << MWAIT_SUBSTATE_SIZE) |
  69                 (highest_subcstate - 1);
  70
  71 #if defined(CONFIG_X86)
  72         switch (boot_cpu_data.x86_vendor) {
  73         case X86_VENDOR_AMD:
  74         case X86_VENDOR_INTEL:
  75                 /*
  76                  * AMD Fam10h TSC will tick in all
  77                  * C/P/S0/S1 states when this bit is set.
  78                  */
  79                 if (!boot_cpu_has(X86_FEATURE_NONSTOP_TSC))
  80                         tsc_detected_unstable = 1;
  81                 break;
  82         default:
  83                 /* TSC could halt in idle */
  84                 tsc_detected_unstable = 1;
  85         }
  86 #endif
  87 }
  88
  89 static unsigned long cpu_weight[NR_CPUS];
  90 static int tsk_in_cpu[NR_CPUS] = {[0 ... NR_CPUS-1] = -1};
  91 static DECLARE_BITMAP(pad_busy_cpus_bits, NR_CPUS);
  92 static void round_robin_cpu(unsigned int tsk_index)
  93 {
  94         struct cpumask *pad_busy_cpus = to_cpumask(pad_busy_cpus_bits);
  95         cpumask_var_t tmp;
  96         int cpu;
  97         unsigned long min_weight = -1;
  98         unsigned long uninitialized_var(preferred_cpu);
  99
 100         if (!alloc_cpumask_var(&tmp, GFP_KERNEL))
 101                 return;
 102
 103         mutex_lock(&round_robin_lock);
 104         cpumask_clear(tmp);
 105         for_each_cpu(cpu, pad_busy_cpus)
 106                 cpumask_or(tmp, tmp, topology_sibling_cpumask(cpu));
 107         cpumask_andnot(tmp, cpu_online_mask, tmp);
 108         /* avoid HT sibilings if possible */
 109         if (cpumask_empty(tmp))
 110                 cpumask_andnot(tmp, cpu_online_mask, pad_busy_cpus);
 111         if (cpumask_empty(tmp)) {
 112                 mutex_unlock(&round_robin_lock);
 113                 return;
 114         }
 115         for_each_cpu(cpu, tmp) {
 116                 if (cpu_weight[cpu] < min_weight) {
 117                         min_weight = cpu_weight[cpu];
 118                         preferred_cpu = cpu;
 119                 }
 120         }
 121
 122         if (tsk_in_cpu[tsk_index] != -1)
 123                 cpumask_clear_cpu(tsk_in_cpu[tsk_index], pad_busy_cpus);
 124         tsk_in_cpu[tsk_index] = preferred_cpu;
 125         cpumask_set_cpu(preferred_cpu, pad_busy_cpus);
 126         cpu_weight[preferred_cpu]++;
 127         mutex_unlock(&round_robin_lock);
 128
 129         set_cpus_allowed_ptr(current, cpumask_of(preferred_cpu));
 130 }
 131
 132 static void exit_round_robin(unsigned int tsk_index)
 133 {
 134         struct cpumask *pad_busy_cpus = to_cpumask(pad_busy_cpus_bits);
 135         cpumask_clear_cpu(tsk_in_cpu[tsk_index], pad_busy_cpus);
 136         tsk_in_cpu[tsk_index] = -1;
 137 }
 138
 139 static unsigned int idle_pct = 5; /* percentage */
 140 static unsigned int round_robin_time = 1; /* second */
 141 static int power_saving_thread(void *data)
 142 {
 143         struct sched_param param = {.sched_priority = 1};
 144         int do_sleep;
 145         unsigned int tsk_index = (unsigned long)data;
 146         u64 last_jiffies = 0;
 147
 148         sched_setscheduler(current, SCHED_RR, &param);
 149
 150         while (!kthread_should_stop()) {
 151                 unsigned long expire_time;
 152
 153                 /* round robin to cpus */
 154                 expire_time = last_jiffies + round_robin_time * HZ;
 155                 if (time_before(expire_time, jiffies)) {
 156                         last_jiffies = jiffies;
 157                         round_robin_cpu(tsk_index);
 158                 }
 159
 160                 do_sleep = 0;
 161
 162                 expire_time = jiffies + HZ * (100 - idle_pct) / 100;
 163
 164                 while (!need_resched()) {
 165                         if (tsc_detected_unstable && !tsc_marked_unstable) {
 166                                 /* TSC could halt in idle, so notify users */
 167                                 mark_tsc_unstable("TSC halts in idle");
 168                                 tsc_marked_unstable = 1;
 169                         }
 170                         local_irq_disable();
 171                         tick_broadcast_enable();
 172                         tick_broadcast_enter();
 173                         stop_critical_timings();
 174
 175                         mwait_idle_with_hints(power_saving_mwait_eax, 1);
 176
 177                         start_critical_timings();
 178                         tick_broadcast_exit();
 179                         local_irq_enable();
 180
 181                         if (time_before(expire_time, jiffies)) {
 182                                 do_sleep = 1;
 183                                 break;
 184                         }
 185                 }
 186
 187                 /*
 188                  * current sched_rt has threshold for rt task running time.
 189                  * When a rt task uses 95% CPU time, the rt thread will be
 190                  * scheduled out for 5% CPU time to not starve other tasks. But
 191                  * the mechanism only works when all CPUs have RT task running,
 192                  * as if one CPU hasn't RT task, RT task from other CPUs will
 193                  * borrow CPU time from this CPU and cause RT task use > 95%
 194                  * CPU time. To make 'avoid starvation' work, takes a nap here.
 195                  */
 196                 if (unlikely(do_sleep))
 197                         schedule_timeout_killable(HZ * idle_pct / 100);
 198
 199                 /* If an external event has set the need_resched flag, then
 200                  * we need to deal with it, or this loop will continue to
 201                  * spin without calling __mwait().
 202                  */
 203                 if (unlikely(need_resched()))
 204                         schedule();
 205         }
 206
 207         exit_round_robin(tsk_index);
 208         return 0;
 209 }
 210
 211 static struct task_struct *ps_tsks[NR_CPUS];
 212 static unsigned int ps_tsk_num;
 213 static int create_power_saving_task(void)
 214 {
 215         int rc;
 216
 217         ps_tsks[ps_tsk_num] = kthread_run(power_saving_thread,
 218                 (void *)(unsigned long)ps_tsk_num,
 219                 "acpi_pad/%d", ps_tsk_num);
 220
 221         if (IS_ERR(ps_tsks[ps_tsk_num])) {
 222                 rc = PTR_ERR(ps_tsks[ps_tsk_num]);
 223                 ps_tsks[ps_tsk_num] = NULL;
 224         } else {
 225                 rc = 0;
 226                 ps_tsk_num++;
 227         }
 228
 229         return rc;
 230 }
 231
 232 static void destroy_power_saving_task(void)
 233 {
 234         if (ps_tsk_num > 0) {
 235                 ps_tsk_num--;
 236                 kthread_stop(ps_tsks[ps_tsk_num]);
 237                 ps_tsks[ps_tsk_num] = NULL;
 238         }
 239 }
 240
 241 static void set_power_saving_task_num(unsigned int num)
 242 {
 243         if (num > ps_tsk_num) {
 244                 while (ps_tsk_num < num) {
 245                         if (create_power_saving_task())
 246                                 return;
 247                 }
 248         } else if (num < ps_tsk_num) {
 249                 while (ps_tsk_num > num)
 250                         destroy_power_saving_task();
 251         }
 252 }
 253
 254 static void acpi_pad_idle_cpus(unsigned int num_cpus)
 255 {
 256         get_online_cpus();
 257
 258         num_cpus = min_t(unsigned int, num_cpus, num_online_cpus());
 259         set_power_saving_task_num(num_cpus);
 260
 261         put_online_cpus();
 262 }
 263
 264 static uint32_t acpi_pad_idle_cpus_num(void)
 265 {
 266         return ps_tsk_num;
 267 }
 268
 269 static ssize_t acpi_pad_rrtime_store(struct device *dev,
 270         struct device_attribute *attr, const char *buf, size_t count)
 271 {
 272         unsigned long num;
 273         if (kstrtoul(buf, 0, &num))
 274                 return -EINVAL;
 275         if (num < 1 || num >= 100)
 276                 return -EINVAL;
 277         mutex_lock(&isolated_cpus_lock);
 278         round_robin_time = num;
 279         mutex_unlock(&isolated_cpus_lock);
 280         return count;
 281 }
 282
 283 static ssize_t acpi_pad_rrtime_show(struct device *dev,
 284         struct device_attribute *attr, char *buf)
 285 {
 286         return scnprintf(buf, PAGE_SIZE, "%d\n", round_robin_time);
 287 }
 288 static DEVICE_ATTR(rrtime, S_IRUGO|S_IWUSR,
 289         acpi_pad_rrtime_show,
 290         acpi_pad_rrtime_store);
 291
 292 static ssize_t acpi_pad_idlepct_store(struct device *dev,
 293         struct device_attribute *attr, const char *buf, size_t count)
 294 {
 295         unsigned long num;
 296         if (kstrtoul(buf, 0, &num))
 297                 return -EINVAL;
 298         if (num < 1 || num >= 100)
 299                 return -EINVAL;
 300         mutex_lock(&isolated_cpus_lock);
 301         idle_pct = num;
 302         mutex_unlock(&isolated_cpus_lock);
 303         return count;
 304 }
 305
 306 static ssize_t acpi_pad_idlepct_show(struct device *dev,
 307         struct device_attribute *attr, char *buf)
 308 {
 309         return scnprintf(buf, PAGE_SIZE, "%d\n", idle_pct);
 310 }
 311 static DEVICE_ATTR(idlepct, S_IRUGO|S_IWUSR,
 312         acpi_pad_idlepct_show,
 313         acpi_pad_idlepct_store);
 314
 315 static ssize_t acpi_pad_idlecpus_store(struct device *dev,
 316         struct device_attribute *attr, const char *buf, size_t count)
 317 {
 318         unsigned long num;
 319         if (kstrtoul(buf, 0, &num))
 320                 return -EINVAL;
 321         mutex_lock(&isolated_cpus_lock);
 322         acpi_pad_idle_cpus(num);
 323         mutex_unlock(&isolated_cpus_lock);
 324         return count;
 325 }
 326
 327 static ssize_t acpi_pad_idlecpus_show(struct device *dev,
 328         struct device_attribute *attr, char *buf)
 329 {
 330         return cpumap_print_to_pagebuf(false, buf,
 331                                        to_cpumask(pad_busy_cpus_bits));
 332 }
 333
 334 static DEVICE_ATTR(idlecpus, S_IRUGO|S_IWUSR,
 335         acpi_pad_idlecpus_show,
 336         acpi_pad_idlecpus_store);
 337
 338 static int acpi_pad_add_sysfs(struct acpi_device *device)
 339 {
 340         int result;
 341
 342         result = device_create_file(&device->dev, &dev_attr_idlecpus);
 343         if (result)
 344                 return -ENODEV;
 345         result = device_create_file(&device->dev, &dev_attr_idlepct);
 346         if (result) {
 347                 device_remove_file(&device->dev, &dev_attr_idlecpus);
 348                 return -ENODEV;
 349         }
 350         result = device_create_file(&device->dev, &dev_attr_rrtime);
 351         if (result) {
 352                 device_remove_file(&device->dev, &dev_attr_idlecpus);
 353                 device_remove_file(&device->dev, &dev_attr_idlepct);
 354                 return -ENODEV;
 355         }
 356         return 0;
 357 }
 358
 359 static void acpi_pad_remove_sysfs(struct acpi_device *device)
 360 {
 361         device_remove_file(&device->dev, &dev_attr_idlecpus);
 362         device_remove_file(&device->dev, &dev_attr_idlepct);
 363         device_remove_file(&device->dev, &dev_attr_rrtime);
 364 }
 365
 366 /*
 367  * Query firmware how many CPUs should be idle
 368  * return -1 on failure
 369  */
 370 static int acpi_pad_pur(acpi_handle handle)
 371 {
 372         struct acpi_buffer buffer = {ACPI_ALLOCATE_BUFFER, NULL};
 373         union acpi_object *package;
 374         int num = -1;
 375
 376         if (ACPI_FAILURE(acpi_evaluate_object(handle, "_PUR", NULL, &buffer)))
 377                 return num;
 378
 379         if (!buffer.length || !buffer.pointer)
 380                 return num;
 381
 382         package = buffer.pointer;
 383
 384         if (package->type == ACPI_TYPE_PACKAGE &&
 385                 package->package.count == 2 &&
 386                 package->package.elements[0].integer.value == 1) /* rev 1 */
 387
 388                 num = package->package.elements[1].integer.value;
 389
 390         kfree(buffer.pointer);
 391         return num;
 392 }
 393
 394 static void acpi_pad_handle_notify(acpi_handle handle)
 395 {
 396         int num_cpus;
 397         uint32_t idle_cpus;
 398         struct acpi_buffer param = {
 399                 .length = 4,
 400                 .pointer = (void *)&idle_cpus,
 401         };
 402
 403         mutex_lock(&isolated_cpus_lock);
 404         num_cpus = acpi_pad_pur(handle);
 405         if (num_cpus < 0) {
 406                 mutex_unlock(&isolated_cpus_lock);
 407                 return;
 408         }
 409         acpi_pad_idle_cpus(num_cpus);
 410         idle_cpus = acpi_pad_idle_cpus_num();
 411         acpi_evaluate_ost(handle, ACPI_PROCESSOR_AGGREGATOR_NOTIFY, 0, &param);
 412         mutex_unlock(&isolated_cpus_lock);
 413 }
 414
 415 static void acpi_pad_notify(acpi_handle handle, u32 event,
 416         void *data)
 417 {
 418         struct acpi_device *device = data;
 419
 420         switch (event) {
 421         case ACPI_PROCESSOR_AGGREGATOR_NOTIFY:
 422                 acpi_pad_handle_notify(handle);
 423                 acpi_bus_generate_netlink_event(device->pnp.device_class,
 424                         dev_name(&device->dev), event, 0);
 425                 break;
 426         default:
 427                 pr_warn("Unsupported event [0x%x]\n", event);
 428                 break;
 429         }
 430 }
 431
 432 static int acpi_pad_add(struct acpi_device *device)
 433 {
 434         acpi_status status;
 435
 436         strcpy(acpi_device_name(device), ACPI_PROCESSOR_AGGREGATOR_DEVICE_NAME);
 437         strcpy(acpi_device_class(device), ACPI_PROCESSOR_AGGREGATOR_CLASS);
 438
 439         if (acpi_pad_add_sysfs(device))
 440                 return -ENODEV;
 441
 442         status = acpi_install_notify_handler(device->handle,
 443                 ACPI_DEVICE_NOTIFY, acpi_pad_notify, device);
 444         if (ACPI_FAILURE(status)) {
 445                 acpi_pad_remove_sysfs(device);
 446                 return -ENODEV;
 447         }
 448
 449         return 0;
 450 }
 451
 452 static int acpi_pad_remove(struct acpi_device *device)
 453 {
 454         mutex_lock(&isolated_cpus_lock);
 455         acpi_pad_idle_cpus(0);
 456         mutex_unlock(&isolated_cpus_lock);
 457
 458         acpi_remove_notify_handler(device->handle,
 459                 ACPI_DEVICE_NOTIFY, acpi_pad_notify);
 460         acpi_pad_remove_sysfs(device);
 461         return 0;
 462 }
 463
 464 static const struct acpi_device_id pad_device_ids[] = {
 465         {"ACPI000C", 0},
 466         {"", 0},
 467 };
 468 MODULE_DEVICE_TABLE(acpi, pad_device_ids);
 469
 470 static struct acpi_driver acpi_pad_driver = {
 471         .name = "processor_aggregator",
 472         .class = ACPI_PROCESSOR_AGGREGATOR_CLASS,
 473         .ids = pad_device_ids,
 474         .ops = {
 475                 .add = acpi_pad_add,
 476                 .remove = acpi_pad_remove,
 477         },
 478 };
 479
 480 static int __init acpi_pad_init(void)
 481 {
 482         /* Xen ACPI PAD is used when running as Xen Dom0. */
 483         if (xen_initial_domain())
 484                 return -ENODEV;
 485
 486         power_saving_mwait_init();
 487         if (power_saving_mwait_eax == 0)
 488                 return -EINVAL;
 489
 490         return acpi_bus_register_driver(&acpi_pad_driver);
 491 }
 492
 493 static void __exit acpi_pad_exit(void)
 494 {
 495         acpi_bus_unregister_driver(&acpi_pad_driver);
 496 }
 497
 498 module_init(acpi_pad_init);
 499 module_exit(acpi_pad_exit);
 500 MODULE_AUTHOR("Shaohua Li<shaohua.li@intel.com>");
 501 MODULE_DESCRIPTION("ACPI Processor Aggregator Driver");
 502 MODULE_LICENSE("GPL");