drivers/acpi/acpi_pad.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * acpi_pad.c ACPI Processor Aggregator Driver
   4  *
   5  * Copyright (c) 2009, Intel Corporation.
   6  */
   7
   8 #include <linux/kernel.h>
   9 #include <linux/cpumask.h>
  10 #include <linux/module.h>
  11 #include <linux/init.h>
  12 #include <linux/types.h>
  13 #include <linux/kthread.h>
  14 #include <uapi/linux/sched/types.h>
  15 #include <linux/freezer.h>
  16 #include <linux/cpu.h>
  17 #include <linux/tick.h>
  18 #include <linux/slab.h>
  19 #include <linux/acpi.h>
  20 #include <asm/mwait.h>
  21 #include <xen/xen.h>
  22
  23 #define ACPI_PROCESSOR_AGGREGATOR_CLASS "acpi_pad"
  24 #define ACPI_PROCESSOR_AGGREGATOR_DEVICE_NAME "Processor Aggregator"
  25 #define ACPI_PROCESSOR_AGGREGATOR_NOTIFY 0x80
  26 static DEFINE_MUTEX(isolated_cpus_lock);
  27 static DEFINE_MUTEX(round_robin_lock);
  28
  29 static unsigned long power_saving_mwait_eax;
  30
  31 static unsigned char tsc_detected_unstable;
  32 static unsigned char tsc_marked_unstable;
  33
  34 static void power_saving_mwait_init(void)
  35 {
  36         unsigned int eax, ebx, ecx, edx;
  37         unsigned int highest_cstate = 0;
  38         unsigned int highest_subcstate = 0;
  39         int i;
  40
  41         if (!boot_cpu_has(X86_FEATURE_MWAIT))
  42                 return;
  43         if (boot_cpu_data.cpuid_level < CPUID_MWAIT_LEAF)
  44                 return;
  45
  46         cpuid(CPUID_MWAIT_LEAF, &eax, &ebx, &ecx, &edx);
  47
  48         if (!(ecx & CPUID5_ECX_EXTENSIONS_SUPPORTED) ||
  49             !(ecx & CPUID5_ECX_INTERRUPT_BREAK))
  50                 return;
  51
  52         edx >>= MWAIT_SUBSTATE_SIZE;
  53         for (i = 0; i < 7 && edx; i++, edx >>= MWAIT_SUBSTATE_SIZE) {
  54                 if (edx & MWAIT_SUBSTATE_MASK) {
  55                         highest_cstate = i;
  56                         highest_subcstate = edx & MWAIT_SUBSTATE_MASK;
  57                 }
  58         }
  59         power_saving_mwait_eax = (highest_cstate << MWAIT_SUBSTATE_SIZE) |
  60                 (highest_subcstate - 1);
  61
  62 #if defined(CONFIG_X86)
  63         switch (boot_cpu_data.x86_vendor) {
  64         case X86_VENDOR_HYGON:
  65         case X86_VENDOR_AMD:
  66         case X86_VENDOR_INTEL:
  67         case X86_VENDOR_ZHAOXIN:
  68                 /*
  69                  * AMD Fam10h TSC will tick in all
  70                  * C/P/S0/S1 states when this bit is set.
  71                  */
  72                 if (!boot_cpu_has(X86_FEATURE_NONSTOP_TSC))
  73                         tsc_detected_unstable = 1;
  74                 break;
  75         default:
  76                 /* TSC could halt in idle */
  77                 tsc_detected_unstable = 1;
  78         }
  79 #endif
  80 }
  81
  82 static unsigned long cpu_weight[NR_CPUS];
  83 static int tsk_in_cpu[NR_CPUS] = {[0 ... NR_CPUS-1] = -1};
  84 static DECLARE_BITMAP(pad_busy_cpus_bits, NR_CPUS);
  85 static void round_robin_cpu(unsigned int tsk_index)
  86 {
  87         struct cpumask *pad_busy_cpus = to_cpumask(pad_busy_cpus_bits);
  88         cpumask_var_t tmp;
  89         int cpu;
  90         unsigned long min_weight = -1;
  91         unsigned long uninitialized_var(preferred_cpu);
  92
  93         if (!alloc_cpumask_var(&tmp, GFP_KERNEL))
  94                 return;
  95
  96         mutex_lock(&round_robin_lock);
  97         cpumask_clear(tmp);
  98         for_each_cpu(cpu, pad_busy_cpus)
  99                 cpumask_or(tmp, tmp, topology_sibling_cpumask(cpu));
 100         cpumask_andnot(tmp, cpu_online_mask, tmp);
 101         /* avoid HT sibilings if possible */
 102         if (cpumask_empty(tmp))
 103                 cpumask_andnot(tmp, cpu_online_mask, pad_busy_cpus);
 104         if (cpumask_empty(tmp)) {
 105                 mutex_unlock(&round_robin_lock);
 106                 free_cpumask_var(tmp);
 107                 return;
 108         }
 109         for_each_cpu(cpu, tmp) {
 110                 if (cpu_weight[cpu] < min_weight) {
 111                         min_weight = cpu_weight[cpu];
 112                         preferred_cpu = cpu;
 113                 }
 114         }
 115
 116         if (tsk_in_cpu[tsk_index] != -1)
 117                 cpumask_clear_cpu(tsk_in_cpu[tsk_index], pad_busy_cpus);
 118         tsk_in_cpu[tsk_index] = preferred_cpu;
 119         cpumask_set_cpu(preferred_cpu, pad_busy_cpus);
 120         cpu_weight[preferred_cpu]++;
 121         mutex_unlock(&round_robin_lock);
 122
 123         set_cpus_allowed_ptr(current, cpumask_of(preferred_cpu));
 124
 125         free_cpumask_var(tmp);
 126 }
 127
 128 static void exit_round_robin(unsigned int tsk_index)
 129 {
 130         struct cpumask *pad_busy_cpus = to_cpumask(pad_busy_cpus_bits);
 131         cpumask_clear_cpu(tsk_in_cpu[tsk_index], pad_busy_cpus);
 132         tsk_in_cpu[tsk_index] = -1;
 133 }
 134
 135 static unsigned int idle_pct = 5; /* percentage */
 136 static unsigned int round_robin_time = 1; /* second */
 137 static int power_saving_thread(void *data)
 138 {
 139         struct sched_param param = {.sched_priority = 1};
 140         int do_sleep;
 141         unsigned int tsk_index = (unsigned long)data;
 142         u64 last_jiffies = 0;
 143
 144         sched_setscheduler(current, SCHED_RR, &param);
 145
 146         while (!kthread_should_stop()) {
 147                 unsigned long expire_time;
 148
 149                 /* round robin to cpus */
 150                 expire_time = last_jiffies + round_robin_time * HZ;
 151                 if (time_before(expire_time, jiffies)) {
 152                         last_jiffies = jiffies;
 153                         round_robin_cpu(tsk_index);
 154                 }
 155
 156                 do_sleep = 0;
 157
 158                 expire_time = jiffies + HZ * (100 - idle_pct) / 100;
 159
 160                 while (!need_resched()) {
 161                         if (tsc_detected_unstable && !tsc_marked_unstable) {
 162                                 /* TSC could halt in idle, so notify users */
 163                                 mark_tsc_unstable("TSC halts in idle");
 164                                 tsc_marked_unstable = 1;
 165                         }
 166                         local_irq_disable();
 167                         tick_broadcast_enable();
 168                         tick_broadcast_enter();
 169                         stop_critical_timings();
 170
 171                         mwait_idle_with_hints(power_saving_mwait_eax, 1);
 172
 173                         start_critical_timings();
 174                         tick_broadcast_exit();
 175                         local_irq_enable();
 176
 177                         if (time_before(expire_time, jiffies)) {
 178                                 do_sleep = 1;
 179                                 break;
 180                         }
 181                 }
 182
 183                 /*
 184                  * current sched_rt has threshold for rt task running time.
 185                  * When a rt task uses 95% CPU time, the rt thread will be
 186                  * scheduled out for 5% CPU time to not starve other tasks. But
 187                  * the mechanism only works when all CPUs have RT task running,
 188                  * as if one CPU hasn't RT task, RT task from other CPUs will
 189                  * borrow CPU time from this CPU and cause RT task use > 95%
 190                  * CPU time. To make 'avoid starvation' work, takes a nap here.
 191                  */
 192                 if (unlikely(do_sleep))
 193                         schedule_timeout_killable(HZ * idle_pct / 100);
 194
 195                 /* If an external event has set the need_resched flag, then
 196                  * we need to deal with it, or this loop will continue to
 197                  * spin without calling __mwait().
 198                  */
 199                 if (unlikely(need_resched()))
 200                         schedule();
 201         }
 202
 203         exit_round_robin(tsk_index);
 204         return 0;
 205 }
 206
 207 static struct task_struct *ps_tsks[NR_CPUS];
 208 static unsigned int ps_tsk_num;
 209 static int create_power_saving_task(void)
 210 {
 211         int rc;
 212
 213         ps_tsks[ps_tsk_num] = kthread_run(power_saving_thread,
 214                 (void *)(unsigned long)ps_tsk_num,
 215                 "acpi_pad/%d", ps_tsk_num);
 216
 217         if (IS_ERR(ps_tsks[ps_tsk_num])) {
 218                 rc = PTR_ERR(ps_tsks[ps_tsk_num]);
 219                 ps_tsks[ps_tsk_num] = NULL;
 220         } else {
 221                 rc = 0;
 222                 ps_tsk_num++;
 223         }
 224
 225         return rc;
 226 }
 227
 228 static void destroy_power_saving_task(void)
 229 {
 230         if (ps_tsk_num > 0) {
 231                 ps_tsk_num--;
 232                 kthread_stop(ps_tsks[ps_tsk_num]);
 233                 ps_tsks[ps_tsk_num] = NULL;
 234         }
 235 }
 236
 237 static void set_power_saving_task_num(unsigned int num)
 238 {
 239         if (num > ps_tsk_num) {
 240                 while (ps_tsk_num < num) {
 241                         if (create_power_saving_task())
 242                                 return;
 243                 }
 244         } else if (num < ps_tsk_num) {
 245                 while (ps_tsk_num > num)
 246                         destroy_power_saving_task();
 247         }
 248 }
 249
 250 static void acpi_pad_idle_cpus(unsigned int num_cpus)
 251 {
 252         get_online_cpus();
 253
 254         num_cpus = min_t(unsigned int, num_cpus, num_online_cpus());
 255         set_power_saving_task_num(num_cpus);
 256
 257         put_online_cpus();
 258 }
 259
 260 static uint32_t acpi_pad_idle_cpus_num(void)
 261 {
 262         return ps_tsk_num;
 263 }
 264
 265 static ssize_t acpi_pad_rrtime_store(struct device *dev,
 266         struct device_attribute *attr, const char *buf, size_t count)
 267 {
 268         unsigned long num;
 269         if (kstrtoul(buf, 0, &num))
 270                 return -EINVAL;
 271         if (num < 1 || num >= 100)
 272                 return -EINVAL;
 273         mutex_lock(&isolated_cpus_lock);
 274         round_robin_time = num;
 275         mutex_unlock(&isolated_cpus_lock);
 276         return count;
 277 }
 278
 279 static ssize_t acpi_pad_rrtime_show(struct device *dev,
 280         struct device_attribute *attr, char *buf)
 281 {
 282         return scnprintf(buf, PAGE_SIZE, "%d\n", round_robin_time);
 283 }
 284 static DEVICE_ATTR(rrtime, S_IRUGO|S_IWUSR,
 285         acpi_pad_rrtime_show,
 286         acpi_pad_rrtime_store);
 287
 288 static ssize_t acpi_pad_idlepct_store(struct device *dev,
 289         struct device_attribute *attr, const char *buf, size_t count)
 290 {
 291         unsigned long num;
 292         if (kstrtoul(buf, 0, &num))
 293                 return -EINVAL;
 294         if (num < 1 || num >= 100)
 295                 return -EINVAL;
 296         mutex_lock(&isolated_cpus_lock);
 297         idle_pct = num;
 298         mutex_unlock(&isolated_cpus_lock);
 299         return count;
 300 }
 301
 302 static ssize_t acpi_pad_idlepct_show(struct device *dev,
 303         struct device_attribute *attr, char *buf)
 304 {
 305         return scnprintf(buf, PAGE_SIZE, "%d\n", idle_pct);
 306 }
 307 static DEVICE_ATTR(idlepct, S_IRUGO|S_IWUSR,
 308         acpi_pad_idlepct_show,
 309         acpi_pad_idlepct_store);
 310
 311 static ssize_t acpi_pad_idlecpus_store(struct device *dev,
 312         struct device_attribute *attr, const char *buf, size_t count)
 313 {
 314         unsigned long num;
 315         if (kstrtoul(buf, 0, &num))
 316                 return -EINVAL;
 317         mutex_lock(&isolated_cpus_lock);
 318         acpi_pad_idle_cpus(num);
 319         mutex_unlock(&isolated_cpus_lock);
 320         return count;
 321 }
 322
 323 static ssize_t acpi_pad_idlecpus_show(struct device *dev,
 324         struct device_attribute *attr, char *buf)
 325 {
 326         return cpumap_print_to_pagebuf(false, buf,
 327                                        to_cpumask(pad_busy_cpus_bits));
 328 }
 329
 330 static DEVICE_ATTR(idlecpus, S_IRUGO|S_IWUSR,
 331         acpi_pad_idlecpus_show,
 332         acpi_pad_idlecpus_store);
 333
 334 static int acpi_pad_add_sysfs(struct acpi_device *device)
 335 {
 336         int result;
 337
 338         result = device_create_file(&device->dev, &dev_attr_idlecpus);
 339         if (result)
 340                 return -ENODEV;
 341         result = device_create_file(&device->dev, &dev_attr_idlepct);
 342         if (result) {
 343                 device_remove_file(&device->dev, &dev_attr_idlecpus);
 344                 return -ENODEV;
 345         }
 346         result = device_create_file(&device->dev, &dev_attr_rrtime);
 347         if (result) {
 348                 device_remove_file(&device->dev, &dev_attr_idlecpus);
 349                 device_remove_file(&device->dev, &dev_attr_idlepct);
 350                 return -ENODEV;
 351         }
 352         return 0;
 353 }
 354
 355 static void acpi_pad_remove_sysfs(struct acpi_device *device)
 356 {
 357         device_remove_file(&device->dev, &dev_attr_idlecpus);
 358         device_remove_file(&device->dev, &dev_attr_idlepct);
 359         device_remove_file(&device->dev, &dev_attr_rrtime);
 360 }
 361
 362 /*
 363  * Query firmware how many CPUs should be idle
 364  * return -1 on failure
 365  */
 366 static int acpi_pad_pur(acpi_handle handle)
 367 {
 368         struct acpi_buffer buffer = {ACPI_ALLOCATE_BUFFER, NULL};
 369         union acpi_object *package;
 370         int num = -1;
 371
 372         if (ACPI_FAILURE(acpi_evaluate_object(handle, "_PUR", NULL, &buffer)))
 373                 return num;
 374
 375         if (!buffer.length || !buffer.pointer)
 376                 return num;
 377
 378         package = buffer.pointer;
 379
 380         if (package->type == ACPI_TYPE_PACKAGE &&
 381                 package->package.count == 2 &&
 382                 package->package.elements[0].integer.value == 1) /* rev 1 */
 383
 384                 num = package->package.elements[1].integer.value;
 385
 386         kfree(buffer.pointer);
 387         return num;
 388 }
 389
 390 static void acpi_pad_handle_notify(acpi_handle handle)
 391 {
 392         int num_cpus;
 393         uint32_t idle_cpus;
 394         struct acpi_buffer param = {
 395                 .length = 4,
 396                 .pointer = (void *)&idle_cpus,
 397         };
 398
 399         mutex_lock(&isolated_cpus_lock);
 400         num_cpus = acpi_pad_pur(handle);
 401         if (num_cpus < 0) {
 402                 mutex_unlock(&isolated_cpus_lock);
 403                 return;
 404         }
 405         acpi_pad_idle_cpus(num_cpus);
 406         idle_cpus = acpi_pad_idle_cpus_num();
 407         acpi_evaluate_ost(handle, ACPI_PROCESSOR_AGGREGATOR_NOTIFY, 0, &param);
 408         mutex_unlock(&isolated_cpus_lock);
 409 }
 410
 411 static void acpi_pad_notify(acpi_handle handle, u32 event,
 412         void *data)
 413 {
 414         struct acpi_device *device = data;
 415
 416         switch (event) {
 417         case ACPI_PROCESSOR_AGGREGATOR_NOTIFY:
 418                 acpi_pad_handle_notify(handle);
 419                 acpi_bus_generate_netlink_event(device->pnp.device_class,
 420                         dev_name(&device->dev), event, 0);
 421                 break;
 422         default:
 423                 pr_warn("Unsupported event [0x%x]\n", event);
 424                 break;
 425         }
 426 }
 427
 428 static int acpi_pad_add(struct acpi_device *device)
 429 {
 430         acpi_status status;
 431
 432         strcpy(acpi_device_name(device), ACPI_PROCESSOR_AGGREGATOR_DEVICE_NAME);
 433         strcpy(acpi_device_class(device), ACPI_PROCESSOR_AGGREGATOR_CLASS);
 434
 435         if (acpi_pad_add_sysfs(device))
 436                 return -ENODEV;
 437
 438         status = acpi_install_notify_handler(device->handle,
 439                 ACPI_DEVICE_NOTIFY, acpi_pad_notify, device);
 440         if (ACPI_FAILURE(status)) {
 441                 acpi_pad_remove_sysfs(device);
 442                 return -ENODEV;
 443         }
 444
 445         return 0;
 446 }
 447
 448 static int acpi_pad_remove(struct acpi_device *device)
 449 {
 450         mutex_lock(&isolated_cpus_lock);
 451         acpi_pad_idle_cpus(0);
 452         mutex_unlock(&isolated_cpus_lock);
 453
 454         acpi_remove_notify_handler(device->handle,
 455                 ACPI_DEVICE_NOTIFY, acpi_pad_notify);
 456         acpi_pad_remove_sysfs(device);
 457         return 0;
 458 }
 459
 460 static const struct acpi_device_id pad_device_ids[] = {
 461         {"ACPI000C", 0},
 462         {"", 0},
 463 };
 464 MODULE_DEVICE_TABLE(acpi, pad_device_ids);
 465
 466 static struct acpi_driver acpi_pad_driver = {
 467         .name = "processor_aggregator",
 468         .class = ACPI_PROCESSOR_AGGREGATOR_CLASS,
 469         .ids = pad_device_ids,
 470         .ops = {
 471                 .add = acpi_pad_add,
 472                 .remove = acpi_pad_remove,
 473         },
 474 };
 475
 476 static int __init acpi_pad_init(void)
 477 {
 478         /* Xen ACPI PAD is used when running as Xen Dom0. */
 479         if (xen_initial_domain())
 480                 return -ENODEV;
 481
 482         power_saving_mwait_init();
 483         if (power_saving_mwait_eax == 0)
 484                 return -EINVAL;
 485
 486         return acpi_bus_register_driver(&acpi_pad_driver);
 487 }
 488
 489 static void __exit acpi_pad_exit(void)
 490 {
 491         acpi_bus_unregister_driver(&acpi_pad_driver);
 492 }
 493
 494 module_init(acpi_pad_init);
 495 module_exit(acpi_pad_exit);
 496 MODULE_AUTHOR("Shaohua Li<shaohua.li@intel.com>");
 497 MODULE_DESCRIPTION("ACPI Processor Aggregator Driver");
 498 MODULE_LICENSE("GPL");