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