arch/x86/kernel/cpu/mcheck/therm_throt.c

   1 /*
   2  * Thermal throttle event support code (such as syslog messaging and rate
   3  * limiting) that was factored out from x86_64 (mce_intel.c) and i386 (p4.c).
   4  *
   5  * This allows consistent reporting of CPU thermal throttle events.
   6  *
   7  * Maintains a counter in /sys that keeps track of the number of thermal
   8  * events, such that the user knows how bad the thermal problem might be
   9  * (since the logging to syslog and mcelog is rate limited).
  10  *
  11  * Author: Dmitriy Zavin (dmitriyz@google.com)
  12  *
  13  * Credits: Adapted from Zwane Mwaikambo's original code in mce_intel.c.
  14  *          Inspired by Ross Biro's and Al Borchers' counter code.
  15  */
  16 #include <linux/interrupt.h>
  17 #include <linux/notifier.h>
  18 #include <linux/jiffies.h>
  19 #include <linux/kernel.h>
  20 #include <linux/percpu.h>
  21 #include <linux/sysdev.h>
  22 #include <linux/types.h>
  23 #include <linux/init.h>
  24 #include <linux/smp.h>
  25 #include <linux/cpu.h>
  26
  27 #include <asm/processor.h>
  28 #include <asm/system.h>
  29 #include <asm/apic.h>
  30 #include <asm/idle.h>
  31 #include <asm/mce.h>
  32 #include <asm/msr.h>
  33
  34 /* How long to wait between reporting thermal events */
  35 #define CHECK_INTERVAL          (300 * HZ)
  36
  37 #define THERMAL_THROTTLING_EVENT        0
  38 #define POWER_LIMIT_EVENT               1
  39
  40 /*
  41  * Current thermal event state:
  42  */
  43 struct _thermal_state {
  44         bool                    new_event;
  45         int                     event;
  46         u64                     next_check;
  47         unsigned long           count;
  48         unsigned long           last_count;
  49 };
  50
  51 struct thermal_state {
  52         struct _thermal_state core_throttle;
  53         struct _thermal_state core_power_limit;
  54         struct _thermal_state package_throttle;
  55         struct _thermal_state package_power_limit;
  56         struct _thermal_state core_thresh0;
  57         struct _thermal_state core_thresh1;
  58 };
  59
  60 /* Callback to handle core threshold interrupts */
  61 int (*platform_thermal_notify)(__u64 msr_val);
  62
  63 static DEFINE_PER_CPU(struct thermal_state, thermal_state);
  64
  65 static atomic_t therm_throt_en  = ATOMIC_INIT(0);
  66
  67 static u32 lvtthmr_init __read_mostly;
  68
  69 #ifdef CONFIG_SYSFS
  70 #define define_therm_throt_sysdev_one_ro(_name)                         \
  71         static SYSDEV_ATTR(_name, 0444,                                 \
  72                            therm_throt_sysdev_show_##_name,             \
  73                                    NULL)                                \
  74
  75 #define define_therm_throt_sysdev_show_func(event, name)                \
  76                                                                         \
  77 static ssize_t therm_throt_sysdev_show_##event##_##name(                \
  78                         struct sys_device *dev,                         \
  79                         struct sysdev_attribute *attr,                  \
  80                         char *buf)                                      \
  81 {                                                                       \
  82         unsigned int cpu = dev->id;                                     \
  83         ssize_t ret;                                                    \
  84                                                                         \
  85         preempt_disable();      /* CPU hotplug */                       \
  86         if (cpu_online(cpu)) {                                          \
  87                 ret = sprintf(buf, "%lu\n",                             \
  88                               per_cpu(thermal_state, cpu).event.name);  \
  89         } else                                                          \
  90                 ret = 0;                                                \
  91         preempt_enable();                                               \
  92                                                                         \
  93         return ret;                                                     \
  94 }
  95
  96 define_therm_throt_sysdev_show_func(core_throttle, count);
  97 define_therm_throt_sysdev_one_ro(core_throttle_count);
  98
  99 define_therm_throt_sysdev_show_func(core_power_limit, count);
 100 define_therm_throt_sysdev_one_ro(core_power_limit_count);
 101
 102 define_therm_throt_sysdev_show_func(package_throttle, count);
 103 define_therm_throt_sysdev_one_ro(package_throttle_count);
 104
 105 define_therm_throt_sysdev_show_func(package_power_limit, count);
 106 define_therm_throt_sysdev_one_ro(package_power_limit_count);
 107
 108 static struct attribute *thermal_throttle_attrs[] = {
 109         &attr_core_throttle_count.attr,
 110         NULL
 111 };
 112
 113 static struct attribute_group thermal_attr_group = {
 114         .attrs  = thermal_throttle_attrs,
 115         .name   = "thermal_throttle"
 116 };
 117 #endif /* CONFIG_SYSFS */
 118
 119 #define CORE_LEVEL      0
 120 #define PACKAGE_LEVEL   1
 121
 122 /***
 123  * therm_throt_process - Process thermal throttling event from interrupt
 124  * @curr: Whether the condition is current or not (boolean), since the
 125  *        thermal interrupt normally gets called both when the thermal
 126  *        event begins and once the event has ended.
 127  *
 128  * This function is called by the thermal interrupt after the
 129  * IRQ has been acknowledged.
 130  *
 131  * It will take care of rate limiting and printing messages to the syslog.
 132  *
 133  * Returns: 0 : Event should NOT be further logged, i.e. still in
 134  *              "timeout" from previous log message.
 135  *          1 : Event should be logged further, and a message has been
 136  *              printed to the syslog.
 137  */
 138 static int therm_throt_process(bool new_event, int event, int level)
 139 {
 140         struct _thermal_state *state;
 141         unsigned int this_cpu = smp_processor_id();
 142         bool old_event;
 143         u64 now;
 144         struct thermal_state *pstate = &per_cpu(thermal_state, this_cpu);
 145
 146         now = get_jiffies_64();
 147         if (level == CORE_LEVEL) {
 148                 if (event == THERMAL_THROTTLING_EVENT)
 149                         state = &pstate->core_throttle;
 150                 else if (event == POWER_LIMIT_EVENT)
 151                         state = &pstate->core_power_limit;
 152                 else
 153                          return 0;
 154         } else if (level == PACKAGE_LEVEL) {
 155                 if (event == THERMAL_THROTTLING_EVENT)
 156                         state = &pstate->package_throttle;
 157                 else if (event == POWER_LIMIT_EVENT)
 158                         state = &pstate->package_power_limit;
 159                 else
 160                         return 0;
 161         } else
 162                 return 0;
 163
 164         old_event = state->new_event;
 165         state->new_event = new_event;
 166
 167         if (new_event)
 168                 state->count++;
 169
 170         if (time_before64(now, state->next_check) &&
 171                         state->count != state->last_count)
 172                 return 0;
 173
 174         state->next_check = now + CHECK_INTERVAL;
 175         state->last_count = state->count;
 176
 177         /* if we just entered the thermal event */
 178         if (new_event) {
 179                 if (event == THERMAL_THROTTLING_EVENT)
 180                         printk(KERN_CRIT "CPU%d: %s temperature above threshold, cpu clock throttled (total events = %lu)\n",
 181                                 this_cpu,
 182                                 level == CORE_LEVEL ? "Core" : "Package",
 183                                 state->count);
 184                 else
 185                         printk(KERN_CRIT "CPU%d: %s power limit notification (total events = %lu)\n",
 186                                 this_cpu,
 187                                 level == CORE_LEVEL ? "Core" : "Package",
 188                                 state->count);
 189
 190                 add_taint(TAINT_MACHINE_CHECK);
 191                 return 1;
 192         }
 193         if (old_event) {
 194                 if (event == THERMAL_THROTTLING_EVENT)
 195                         printk(KERN_INFO "CPU%d: %s temperature/speed normal\n",
 196                                 this_cpu,
 197                                 level == CORE_LEVEL ? "Core" : "Package");
 198                 else
 199                         printk(KERN_INFO "CPU%d: %s power limit normal\n",
 200                                 this_cpu,
 201                                 level == CORE_LEVEL ? "Core" : "Package");
 202                 return 1;
 203         }
 204
 205         return 0;
 206 }
 207
 208 static int thresh_event_valid(int event)
 209 {
 210         struct _thermal_state *state;
 211         unsigned int this_cpu = smp_processor_id();
 212         struct thermal_state *pstate = &per_cpu(thermal_state, this_cpu);
 213         u64 now = get_jiffies_64();
 214
 215         state = (event == 0) ? &pstate->core_thresh0 : &pstate->core_thresh1;
 216
 217         if (time_before64(now, state->next_check))
 218                 return 0;
 219
 220         state->next_check = now + CHECK_INTERVAL;
 221         return 1;
 222 }
 223
 224 #ifdef CONFIG_SYSFS
 225 /* Add/Remove thermal_throttle interface for CPU device: */
 226 static __cpuinit int thermal_throttle_add_dev(struct sys_device *sys_dev,
 227                                 unsigned int cpu)
 228 {
 229         int err;
 230         struct cpuinfo_x86 *c = &cpu_data(cpu);
 231
 232         err = sysfs_create_group(&sys_dev->kobj, &thermal_attr_group);
 233         if (err)
 234                 return err;
 235
 236         if (cpu_has(c, X86_FEATURE_PLN))
 237                 err = sysfs_add_file_to_group(&sys_dev->kobj,
 238                                               &attr_core_power_limit_count.attr,
 239                                               thermal_attr_group.name);
 240         if (cpu_has(c, X86_FEATURE_PTS)) {
 241                 err = sysfs_add_file_to_group(&sys_dev->kobj,
 242                                               &attr_package_throttle_count.attr,
 243                                               thermal_attr_group.name);
 244                 if (cpu_has(c, X86_FEATURE_PLN))
 245                         err = sysfs_add_file_to_group(&sys_dev->kobj,
 246                                         &attr_package_power_limit_count.attr,
 247                                         thermal_attr_group.name);
 248         }
 249
 250         return err;
 251 }
 252
 253 static __cpuinit void thermal_throttle_remove_dev(struct sys_device *sys_dev)
 254 {
 255         sysfs_remove_group(&sys_dev->kobj, &thermal_attr_group);
 256 }
 257
 258 /* Mutex protecting device creation against CPU hotplug: */
 259 static DEFINE_MUTEX(therm_cpu_lock);
 260
 261 /* Get notified when a cpu comes on/off. Be hotplug friendly. */
 262 static __cpuinit int
 263 thermal_throttle_cpu_callback(struct notifier_block *nfb,
 264                               unsigned long action,
 265                               void *hcpu)
 266 {
 267         unsigned int cpu = (unsigned long)hcpu;
 268         struct sys_device *sys_dev;
 269         int err = 0;
 270
 271         sys_dev = get_cpu_sysdev(cpu);
 272
 273         switch (action) {
 274         case CPU_UP_PREPARE:
 275         case CPU_UP_PREPARE_FROZEN:
 276                 mutex_lock(&therm_cpu_lock);
 277                 err = thermal_throttle_add_dev(sys_dev, cpu);
 278                 mutex_unlock(&therm_cpu_lock);
 279                 WARN_ON(err);
 280                 break;
 281         case CPU_UP_CANCELED:
 282         case CPU_UP_CANCELED_FROZEN:
 283         case CPU_DEAD:
 284         case CPU_DEAD_FROZEN:
 285                 mutex_lock(&therm_cpu_lock);
 286                 thermal_throttle_remove_dev(sys_dev);
 287                 mutex_unlock(&therm_cpu_lock);
 288                 break;
 289         }
 290         return notifier_from_errno(err);
 291 }
 292
 293 static struct notifier_block thermal_throttle_cpu_notifier __cpuinitdata =
 294 {
 295         .notifier_call = thermal_throttle_cpu_callback,
 296 };
 297
 298 static __init int thermal_throttle_init_device(void)
 299 {
 300         unsigned int cpu = 0;
 301         int err;
 302
 303         if (!atomic_read(&therm_throt_en))
 304                 return 0;
 305
 306         register_hotcpu_notifier(&thermal_throttle_cpu_notifier);
 307
 308 #ifdef CONFIG_HOTPLUG_CPU
 309         mutex_lock(&therm_cpu_lock);
 310 #endif
 311         /* connect live CPUs to sysfs */
 312         for_each_online_cpu(cpu) {
 313                 err = thermal_throttle_add_dev(get_cpu_sysdev(cpu), cpu);
 314                 WARN_ON(err);
 315         }
 316 #ifdef CONFIG_HOTPLUG_CPU
 317         mutex_unlock(&therm_cpu_lock);
 318 #endif
 319
 320         return 0;
 321 }
 322 device_initcall(thermal_throttle_init_device);
 323
 324 #endif /* CONFIG_SYSFS */
 325
 326 /*
 327  * Set up the most two significant bit to notify mce log that this thermal
 328  * event type.
 329  * This is a temp solution. May be changed in the future with mce log
 330  * infrasture.
 331  */
 332 #define CORE_THROTTLED          (0)
 333 #define CORE_POWER_LIMIT        ((__u64)1 << 62)
 334 #define PACKAGE_THROTTLED       ((__u64)2 << 62)
 335 #define PACKAGE_POWER_LIMIT     ((__u64)3 << 62)
 336
 337 static void notify_thresholds(__u64 msr_val)
 338 {
 339         /* check whether the interrupt handler is defined;
 340          * otherwise simply return
 341          */
 342         if (!platform_thermal_notify)
 343                 return;
 344
 345         /* lower threshold reached */
 346         if ((msr_val & THERM_LOG_THRESHOLD0) && thresh_event_valid(0))
 347                 platform_thermal_notify(msr_val);
 348         /* higher threshold reached */
 349         if ((msr_val & THERM_LOG_THRESHOLD1) && thresh_event_valid(1))
 350                 platform_thermal_notify(msr_val);
 351 }
 352
 353 /* Thermal transition interrupt handler */
 354 static void intel_thermal_interrupt(void)
 355 {
 356         __u64 msr_val;
 357         struct cpuinfo_x86 *c = &cpu_data(smp_processor_id());
 358
 359         rdmsrl(MSR_IA32_THERM_STATUS, msr_val);
 360
 361         /* Check for violation of core thermal thresholds*/
 362         notify_thresholds(msr_val);
 363
 364         if (therm_throt_process(msr_val & THERM_STATUS_PROCHOT,
 365                                 THERMAL_THROTTLING_EVENT,
 366                                 CORE_LEVEL) != 0)
 367                 mce_log_therm_throt_event(CORE_THROTTLED | msr_val);
 368
 369         if (cpu_has(c, X86_FEATURE_PLN))
 370                 if (therm_throt_process(msr_val & THERM_STATUS_POWER_LIMIT,
 371                                         POWER_LIMIT_EVENT,
 372                                         CORE_LEVEL) != 0)
 373                         mce_log_therm_throt_event(CORE_POWER_LIMIT | msr_val);
 374
 375         if (cpu_has(c, X86_FEATURE_PTS)) {
 376                 rdmsrl(MSR_IA32_PACKAGE_THERM_STATUS, msr_val);
 377                 if (therm_throt_process(msr_val & PACKAGE_THERM_STATUS_PROCHOT,
 378                                         THERMAL_THROTTLING_EVENT,
 379                                         PACKAGE_LEVEL) != 0)
 380                         mce_log_therm_throt_event(PACKAGE_THROTTLED | msr_val);
 381                 if (cpu_has(c, X86_FEATURE_PLN))
 382                         if (therm_throt_process(msr_val &
 383                                         PACKAGE_THERM_STATUS_POWER_LIMIT,
 384                                         POWER_LIMIT_EVENT,
 385                                         PACKAGE_LEVEL) != 0)
 386                                 mce_log_therm_throt_event(PACKAGE_POWER_LIMIT
 387                                                           | msr_val);
 388         }
 389 }
 390
 391 static void unexpected_thermal_interrupt(void)
 392 {
 393         printk(KERN_ERR "CPU%d: Unexpected LVT thermal interrupt!\n",
 394                         smp_processor_id());
 395         add_taint(TAINT_MACHINE_CHECK);
 396 }
 397
 398 static void (*smp_thermal_vector)(void) = unexpected_thermal_interrupt;
 399
 400 asmlinkage void smp_thermal_interrupt(struct pt_regs *regs)
 401 {
 402         exit_idle();
 403         irq_enter();
 404         inc_irq_stat(irq_thermal_count);
 405         smp_thermal_vector();
 406         irq_exit();
 407         /* Ack only at the end to avoid potential reentry */
 408         ack_APIC_irq();
 409 }
 410
 411 /* Thermal monitoring depends on APIC, ACPI and clock modulation */
 412 static int intel_thermal_supported(struct cpuinfo_x86 *c)
 413 {
 414         if (!cpu_has_apic)
 415                 return 0;
 416         if (!cpu_has(c, X86_FEATURE_ACPI) || !cpu_has(c, X86_FEATURE_ACC))
 417                 return 0;
 418         return 1;
 419 }
 420
 421 void __init mcheck_intel_therm_init(void)
 422 {
 423         /*
 424          * This function is only called on boot CPU. Save the init thermal
 425          * LVT value on BSP and use that value to restore APs' thermal LVT
 426          * entry BIOS programmed later
 427          */
 428         if (intel_thermal_supported(&boot_cpu_data))
 429                 lvtthmr_init = apic_read(APIC_LVTTHMR);
 430 }
 431
 432 void intel_init_thermal(struct cpuinfo_x86 *c)
 433 {
 434         unsigned int cpu = smp_processor_id();
 435         int tm2 = 0;
 436         u32 l, h;
 437
 438         if (!intel_thermal_supported(c))
 439                 return;
 440
 441         /*
 442          * First check if its enabled already, in which case there might
 443          * be some SMM goo which handles it, so we can't even put a handler
 444          * since it might be delivered via SMI already:
 445          */
 446         rdmsr(MSR_IA32_MISC_ENABLE, l, h);
 447
 448         /*
 449          * The initial value of thermal LVT entries on all APs always reads
 450          * 0x10000 because APs are woken up by BSP issuing INIT-SIPI-SIPI
 451          * sequence to them and LVT registers are reset to 0s except for
 452          * the mask bits which are set to 1s when APs receive INIT IPI.
 453          * Always restore the value that BIOS has programmed on AP based on
 454          * BSP's info we saved since BIOS is always setting the same value
 455          * for all threads/cores
 456          */
 457         apic_write(APIC_LVTTHMR, lvtthmr_init);
 458
 459         h = lvtthmr_init;
 460
 461         if ((l & MSR_IA32_MISC_ENABLE_TM1) && (h & APIC_DM_SMI)) {
 462                 printk(KERN_DEBUG
 463                        "CPU%d: Thermal monitoring handled by SMI\n", cpu);
 464                 return;
 465         }
 466
 467         /* Check whether a vector already exists */
 468         if (h & APIC_VECTOR_MASK) {
 469                 printk(KERN_DEBUG
 470                        "CPU%d: Thermal LVT vector (%#x) already installed\n",
 471                        cpu, (h & APIC_VECTOR_MASK));
 472                 return;
 473         }
 474
 475         /* early Pentium M models use different method for enabling TM2 */
 476         if (cpu_has(c, X86_FEATURE_TM2)) {
 477                 if (c->x86 == 6 && (c->x86_model == 9 || c->x86_model == 13)) {
 478                         rdmsr(MSR_THERM2_CTL, l, h);
 479                         if (l & MSR_THERM2_CTL_TM_SELECT)
 480                                 tm2 = 1;
 481                 } else if (l & MSR_IA32_MISC_ENABLE_TM2)
 482                         tm2 = 1;
 483         }
 484
 485         /* We'll mask the thermal vector in the lapic till we're ready: */
 486         h = THERMAL_APIC_VECTOR | APIC_DM_FIXED | APIC_LVT_MASKED;
 487         apic_write(APIC_LVTTHMR, h);
 488
 489         rdmsr(MSR_IA32_THERM_INTERRUPT, l, h);
 490         if (cpu_has(c, X86_FEATURE_PLN))
 491                 wrmsr(MSR_IA32_THERM_INTERRUPT,
 492                       l | (THERM_INT_LOW_ENABLE
 493                         | THERM_INT_HIGH_ENABLE | THERM_INT_PLN_ENABLE), h);
 494         else
 495                 wrmsr(MSR_IA32_THERM_INTERRUPT,
 496                       l | (THERM_INT_LOW_ENABLE | THERM_INT_HIGH_ENABLE), h);
 497
 498         if (cpu_has(c, X86_FEATURE_PTS)) {
 499                 rdmsr(MSR_IA32_PACKAGE_THERM_INTERRUPT, l, h);
 500                 if (cpu_has(c, X86_FEATURE_PLN))
 501                         wrmsr(MSR_IA32_PACKAGE_THERM_INTERRUPT,
 502                               l | (PACKAGE_THERM_INT_LOW_ENABLE
 503                                 | PACKAGE_THERM_INT_HIGH_ENABLE
 504                                 | PACKAGE_THERM_INT_PLN_ENABLE), h);
 505                 else
 506                         wrmsr(MSR_IA32_PACKAGE_THERM_INTERRUPT,
 507                               l | (PACKAGE_THERM_INT_LOW_ENABLE
 508                                 | PACKAGE_THERM_INT_HIGH_ENABLE), h);
 509         }
 510
 511         smp_thermal_vector = intel_thermal_interrupt;
 512
 513         rdmsr(MSR_IA32_MISC_ENABLE, l, h);
 514         wrmsr(MSR_IA32_MISC_ENABLE, l | MSR_IA32_MISC_ENABLE_TM1, h);
 515
 516         /* Unmask the thermal vector: */
 517         l = apic_read(APIC_LVTTHMR);
 518         apic_write(APIC_LVTTHMR, l & ~APIC_LVT_MASKED);
 519
 520         printk_once(KERN_INFO "CPU0: Thermal monitoring enabled (%s)\n",
 521                        tm2 ? "TM2" : "TM1");
 522
 523         /* enable thermal throttle processing */
 524         atomic_set(&therm_throt_en, 1);
 525 }