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