arch/x86/kernel/nmi.c

   1 /*
   2  *  NMI watchdog support on APIC systems
   3  *
   4  *  Started by Ingo Molnar <mingo@redhat.com>
   5  *
   6  *  Fixes:
   7  *  Mikael Pettersson   : AMD K7 support for local APIC NMI watchdog.
   8  *  Mikael Pettersson   : Power Management for local APIC NMI watchdog.
   9  *  Mikael Pettersson   : Pentium 4 support for local APIC NMI watchdog.
  10  *  Pavel Machek and
  11  *  Mikael Pettersson   : PM converted to driver model. Disable/enable API.
  12  */
  13
  14 #include <asm/apic.h>
  15
  16 #include <linux/nmi.h>
  17 #include <linux/mm.h>
  18 #include <linux/delay.h>
  19 #include <linux/interrupt.h>
  20 #include <linux/module.h>
  21 #include <linux/sysdev.h>
  22 #include <linux/sysctl.h>
  23 #include <linux/percpu.h>
  24 #include <linux/kprobes.h>
  25 #include <linux/cpumask.h>
  26 #include <linux/kernel_stat.h>
  27 #include <linux/kdebug.h>
  28 #include <linux/smp.h>
  29
  30 #include <asm/proto.h>
  31 #include <asm/timer.h>
  32
  33 #include <asm/mce.h>
  34
  35 #include <mach_traps.h>
  36
  37 int unknown_nmi_panic;
  38 int nmi_watchdog_enabled;
  39
  40 static cpumask_t backtrace_mask = CPU_MASK_NONE;
  41
  42 /* nmi_active:
  43  * >0: the lapic NMI watchdog is active, but can be disabled
  44  * <0: the lapic NMI watchdog has not been set up, and cannot
  45  *     be enabled
  46  *  0: the lapic NMI watchdog is disabled, but can be enabled
  47  */
  48 atomic_t nmi_active = ATOMIC_INIT(0);           /* oprofile uses this */
  49 EXPORT_SYMBOL(nmi_active);
  50
  51 unsigned int nmi_watchdog = NMI_DEFAULT;
  52 EXPORT_SYMBOL(nmi_watchdog);
  53
  54 static int panic_on_timeout;
  55
  56 static unsigned int nmi_hz = HZ;
  57 static DEFINE_PER_CPU(short, wd_enabled);
  58 static int endflag __initdata;
  59
  60 static inline unsigned int get_nmi_count(int cpu)
  61 {
  62 #ifdef CONFIG_X86_64
  63         return cpu_pda(cpu)->__nmi_count;
  64 #else
  65         return nmi_count(cpu);
  66 #endif
  67 }
  68
  69 static inline int mce_in_progress(void)
  70 {
  71 #if defined(CONFIX_X86_64) && defined(CONFIG_X86_MCE)
  72         return atomic_read(&mce_entry) > 0;
  73 #endif
  74         return 0;
  75 }
  76
  77 /*
  78  * Take the local apic timer and PIT/HPET into account. We don't
  79  * know which one is active, when we have highres/dyntick on
  80  */
  81 static inline unsigned int get_timer_irqs(int cpu)
  82 {
  83 #ifdef CONFIG_X86_64
  84         return read_pda(apic_timer_irqs) + read_pda(irq0_irqs);
  85 #else
  86         return per_cpu(irq_stat, cpu).apic_timer_irqs +
  87                 per_cpu(irq_stat, cpu).irq0_irqs;
  88 #endif
  89 }
  90
  91 #ifdef CONFIG_X86_64
  92 /* Run after command line and cpu_init init, but before all other checks */
  93 void nmi_watchdog_default(void)
  94 {
  95         if (nmi_watchdog != NMI_DEFAULT)
  96                 return;
  97         nmi_watchdog = NMI_NONE;
  98 }
  99 #endif
 100
 101 #ifdef CONFIG_SMP
 102 /*
 103  * The performance counters used by NMI_LOCAL_APIC don't trigger when
 104  * the CPU is idle. To make sure the NMI watchdog really ticks on all
 105  * CPUs during the test make them busy.
 106  */
 107 static __init void nmi_cpu_busy(void *data)
 108 {
 109         local_irq_enable_in_hardirq();
 110         /*
 111          * Intentionally don't use cpu_relax here. This is
 112          * to make sure that the performance counter really ticks,
 113          * even if there is a simulator or similar that catches the
 114          * pause instruction. On a real HT machine this is fine because
 115          * all other CPUs are busy with "useless" delay loops and don't
 116          * care if they get somewhat less cycles.
 117          */
 118         while (endflag == 0)
 119                 mb();
 120 }
 121 #endif
 122
 123 int __init check_nmi_watchdog(void)
 124 {
 125         unsigned int *prev_nmi_count;
 126         int cpu;
 127
 128         if (nmi_watchdog == NMI_NONE || nmi_watchdog == NMI_DISABLED)
 129                 return 0;
 130
 131         if (!atomic_read(&nmi_active))
 132                 return 0;
 133
 134         prev_nmi_count = kmalloc(NR_CPUS * sizeof(int), GFP_KERNEL);
 135         if (!prev_nmi_count)
 136                 goto error;
 137
 138         printk(KERN_INFO "Testing NMI watchdog ... ");
 139
 140 #ifdef CONFIG_SMP
 141         if (nmi_watchdog == NMI_LOCAL_APIC)
 142                 smp_call_function(nmi_cpu_busy, (void *)&endflag, 0, 0);
 143 #endif
 144
 145         for_each_possible_cpu(cpu)
 146                 prev_nmi_count[cpu] = get_nmi_count(cpu);
 147         local_irq_enable();
 148         mdelay((20 * 1000) / nmi_hz); /* wait 20 ticks */
 149
 150         for_each_online_cpu(cpu) {
 151                 if (!per_cpu(wd_enabled, cpu))
 152                         continue;
 153                 if (get_nmi_count(cpu) - prev_nmi_count[cpu] <= 5) {
 154                         printk(KERN_WARNING "WARNING: CPU#%d: NMI "
 155                                 "appears to be stuck (%d->%d)!\n",
 156                                 cpu,
 157                                 prev_nmi_count[cpu],
 158                                 get_nmi_count(cpu));
 159                         per_cpu(wd_enabled, cpu) = 0;
 160                         atomic_dec(&nmi_active);
 161                 }
 162         }
 163         endflag = 1;
 164         if (!atomic_read(&nmi_active)) {
 165                 kfree(prev_nmi_count);
 166                 atomic_set(&nmi_active, -1);
 167                 goto error;
 168         }
 169         printk("OK.\n");
 170
 171         /*
 172          * now that we know it works we can reduce NMI frequency to
 173          * something more reasonable; makes a difference in some configs
 174          */
 175         if (nmi_watchdog == NMI_LOCAL_APIC)
 176                 nmi_hz = lapic_adjust_nmi_hz(1);
 177
 178         kfree(prev_nmi_count);
 179         return 0;
 180
 181 error:
 182 #ifdef CONFIG_X86_32
 183         timer_ack = !cpu_has_tsc;
 184 #endif
 185         return -1;
 186 }
 187
 188 static int __init setup_nmi_watchdog(char *str)
 189 {
 190         int nmi;
 191
 192         if (!strncmp(str, "panic", 5)) {
 193                 panic_on_timeout = 1;
 194                 str = strchr(str, ',');
 195                 if (!str)
 196                         return 1;
 197                 ++str;
 198         }
 199
 200         get_option(&str, &nmi);
 201
 202         if (nmi >= NMI_INVALID || nmi < NMI_NONE)
 203                 return 0;
 204
 205         nmi_watchdog = nmi;
 206         return 1;
 207 }
 208 __setup("nmi_watchdog=", setup_nmi_watchdog);
 209
 210 /*
 211  * Suspend/resume support
 212  */
 213 #ifdef CONFIG_PM
 214
 215 static int nmi_pm_active; /* nmi_active before suspend */
 216
 217 static int lapic_nmi_suspend(struct sys_device *dev, pm_message_t state)
 218 {
 219         /* only CPU0 goes here, other CPUs should be offline */
 220         nmi_pm_active = atomic_read(&nmi_active);
 221         stop_apic_nmi_watchdog(NULL);
 222         BUG_ON(atomic_read(&nmi_active) != 0);
 223         return 0;
 224 }
 225
 226 static int lapic_nmi_resume(struct sys_device *dev)
 227 {
 228         /* only CPU0 goes here, other CPUs should be offline */
 229         if (nmi_pm_active > 0) {
 230                 setup_apic_nmi_watchdog(NULL);
 231                 touch_nmi_watchdog();
 232         }
 233         return 0;
 234 }
 235
 236 static struct sysdev_class nmi_sysclass = {
 237         .name           = "lapic_nmi",
 238         .resume         = lapic_nmi_resume,
 239         .suspend        = lapic_nmi_suspend,
 240 };
 241
 242 static struct sys_device device_lapic_nmi = {
 243         .id     = 0,
 244         .cls    = &nmi_sysclass,
 245 };
 246
 247 static int __init init_lapic_nmi_sysfs(void)
 248 {
 249         int error;
 250
 251         /*
 252          * should really be a BUG_ON but b/c this is an
 253          * init call, it just doesn't work.  -dcz
 254          */
 255         if (nmi_watchdog != NMI_LOCAL_APIC)
 256                 return 0;
 257
 258         if (atomic_read(&nmi_active) < 0)
 259                 return 0;
 260
 261         error = sysdev_class_register(&nmi_sysclass);
 262         if (!error)
 263                 error = sysdev_register(&device_lapic_nmi);
 264         return error;
 265 }
 266
 267 /* must come after the local APIC's device_initcall() */
 268 late_initcall(init_lapic_nmi_sysfs);
 269
 270 #endif  /* CONFIG_PM */
 271
 272 static void __acpi_nmi_enable(void *__unused)
 273 {
 274         apic_write_around(APIC_LVT0, APIC_DM_NMI);
 275 }
 276
 277 /*
 278  * Enable timer based NMIs on all CPUs:
 279  */
 280 void acpi_nmi_enable(void)
 281 {
 282         if (atomic_read(&nmi_active) && nmi_watchdog == NMI_IO_APIC)
 283                 on_each_cpu(__acpi_nmi_enable, NULL, 0, 1);
 284 }
 285
 286 static void __acpi_nmi_disable(void *__unused)
 287 {
 288         apic_write_around(APIC_LVT0, APIC_DM_NMI | APIC_LVT_MASKED);
 289 }
 290
 291 /*
 292  * Disable timer based NMIs on all CPUs:
 293  */
 294 void acpi_nmi_disable(void)
 295 {
 296         if (atomic_read(&nmi_active) && nmi_watchdog == NMI_IO_APIC)
 297                 on_each_cpu(__acpi_nmi_disable, NULL, 0, 1);
 298 }
 299
 300 void setup_apic_nmi_watchdog(void *unused)
 301 {
 302         if (__get_cpu_var(wd_enabled))
 303                 return;
 304
 305         /* cheap hack to support suspend/resume */
 306         /* if cpu0 is not active neither should the other cpus */
 307         if (smp_processor_id() != 0 && atomic_read(&nmi_active) <= 0)
 308                 return;
 309
 310         switch (nmi_watchdog) {
 311         case NMI_LOCAL_APIC:
 312                  /* enable it before to avoid race with handler */
 313                 __get_cpu_var(wd_enabled) = 1;
 314                 if (lapic_watchdog_init(nmi_hz) < 0) {
 315                         __get_cpu_var(wd_enabled) = 0;
 316                         return;
 317                 }
 318                 /* FALL THROUGH */
 319         case NMI_IO_APIC:
 320                 __get_cpu_var(wd_enabled) = 1;
 321                 atomic_inc(&nmi_active);
 322         }
 323 }
 324
 325 void stop_apic_nmi_watchdog(void *unused)
 326 {
 327         /* only support LOCAL and IO APICs for now */
 328         if (nmi_watchdog != NMI_LOCAL_APIC &&
 329             nmi_watchdog != NMI_IO_APIC)
 330                 return;
 331         if (__get_cpu_var(wd_enabled) == 0)
 332                 return;
 333         if (nmi_watchdog == NMI_LOCAL_APIC)
 334                 lapic_watchdog_stop();
 335         __get_cpu_var(wd_enabled) = 0;
 336         atomic_dec(&nmi_active);
 337 }
 338
 339 /*
 340  * the best way to detect whether a CPU has a 'hard lockup' problem
 341  * is to check it's local APIC timer IRQ counts. If they are not
 342  * changing then that CPU has some problem.
 343  *
 344  * as these watchdog NMI IRQs are generated on every CPU, we only
 345  * have to check the current processor.
 346  *
 347  * since NMIs don't listen to _any_ locks, we have to be extremely
 348  * careful not to rely on unsafe variables. The printk might lock
 349  * up though, so we have to break up any console locks first ...
 350  * [when there will be more tty-related locks, break them up here too!]
 351  */
 352
 353 static DEFINE_PER_CPU(unsigned, last_irq_sum);
 354 static DEFINE_PER_CPU(local_t, alert_counter);
 355 static DEFINE_PER_CPU(int, nmi_touch);
 356
 357 void touch_nmi_watchdog(void)
 358 {
 359         if (nmi_watchdog > 0) {
 360                 unsigned cpu;
 361
 362                 /*
 363                  * Tell other CPUs to reset their alert counters. We cannot
 364                  * do it ourselves because the alert count increase is not
 365                  * atomic.
 366                  */
 367                 for_each_present_cpu(cpu) {
 368                         if (per_cpu(nmi_touch, cpu) != 1)
 369                                 per_cpu(nmi_touch, cpu) = 1;
 370                 }
 371         }
 372
 373         /*
 374          * Tickle the softlockup detector too:
 375          */
 376         touch_softlockup_watchdog();
 377 }
 378 EXPORT_SYMBOL(touch_nmi_watchdog);
 379
 380 notrace __kprobes int
 381 nmi_watchdog_tick(struct pt_regs *regs, unsigned reason)
 382 {
 383         /*
 384          * Since current_thread_info()-> is always on the stack, and we
 385          * always switch the stack NMI-atomically, it's safe to use
 386          * smp_processor_id().
 387          */
 388         unsigned int sum;
 389         int touched = 0;
 390         int cpu = smp_processor_id();
 391         int rc = 0;
 392
 393         /* check for other users first */
 394         if (notify_die(DIE_NMI, "nmi", regs, reason, 2, SIGINT)
 395                         == NOTIFY_STOP) {
 396                 rc = 1;
 397                 touched = 1;
 398         }
 399
 400         sum = get_timer_irqs(cpu);
 401
 402         if (__get_cpu_var(nmi_touch)) {
 403                 __get_cpu_var(nmi_touch) = 0;
 404                 touched = 1;
 405         }
 406
 407         if (cpu_isset(cpu, backtrace_mask)) {
 408                 static DEFINE_SPINLOCK(lock);   /* Serialise the printks */
 409
 410                 spin_lock(&lock);
 411                 printk(KERN_WARNING "NMI backtrace for cpu %d\n", cpu);
 412                 dump_stack();
 413                 spin_unlock(&lock);
 414                 cpu_clear(cpu, backtrace_mask);
 415         }
 416
 417         /* Could check oops_in_progress here too, but it's safer not to */
 418         if (mce_in_progress())
 419                 touched = 1;
 420
 421         /* if the none of the timers isn't firing, this cpu isn't doing much */
 422         if (!touched && __get_cpu_var(last_irq_sum) == sum) {
 423                 /*
 424                  * Ayiee, looks like this CPU is stuck ...
 425                  * wait a few IRQs (5 seconds) before doing the oops ...
 426                  */
 427                 local_inc(&__get_cpu_var(alert_counter));
 428                 if (local_read(&__get_cpu_var(alert_counter)) == 5 * nmi_hz)
 429                         /*
 430                          * die_nmi will return ONLY if NOTIFY_STOP happens..
 431                          */
 432                         die_nmi("BUG: NMI Watchdog detected LOCKUP",
 433                                 regs, panic_on_timeout);
 434         } else {
 435                 __get_cpu_var(last_irq_sum) = sum;
 436                 local_set(&__get_cpu_var(alert_counter), 0);
 437         }
 438
 439         /* see if the nmi watchdog went off */
 440         if (!__get_cpu_var(wd_enabled))
 441                 return rc;
 442         switch (nmi_watchdog) {
 443         case NMI_LOCAL_APIC:
 444                 rc |= lapic_wd_event(nmi_hz);
 445                 break;
 446         case NMI_IO_APIC:
 447                 /*
 448                  * don't know how to accurately check for this.
 449                  * just assume it was a watchdog timer interrupt
 450                  * This matches the old behaviour.
 451                  */
 452                 rc = 1;
 453                 break;
 454         }
 455         return rc;
 456 }
 457
 458 #ifdef CONFIG_SYSCTL
 459
 460 static int unknown_nmi_panic_callback(struct pt_regs *regs, int cpu)
 461 {
 462         unsigned char reason = get_nmi_reason();
 463         char buf[64];
 464
 465         sprintf(buf, "NMI received for unknown reason %02x\n", reason);
 466         die_nmi(buf, regs, 1); /* Always panic here */
 467         return 0;
 468 }
 469
 470 /*
 471  * proc handler for /proc/sys/kernel/nmi
 472  */
 473 int proc_nmi_enabled(struct ctl_table *table, int write, struct file *file,
 474                         void __user *buffer, size_t *length, loff_t *ppos)
 475 {
 476         int old_state;
 477
 478         nmi_watchdog_enabled = (atomic_read(&nmi_active) > 0) ? 1 : 0;
 479         old_state = nmi_watchdog_enabled;
 480         proc_dointvec(table, write, file, buffer, length, ppos);
 481         if (!!old_state == !!nmi_watchdog_enabled)
 482                 return 0;
 483
 484         if (atomic_read(&nmi_active) < 0 || nmi_watchdog == NMI_DISABLED) {
 485                 printk(KERN_WARNING
 486                         "NMI watchdog is permanently disabled\n");
 487                 return -EIO;
 488         }
 489
 490         /* if nmi_watchdog is not set yet, then set it */
 491         nmi_watchdog_default();
 492
 493 #ifdef CONFIG_X86_32
 494         if (nmi_watchdog == NMI_NONE) {
 495                 if (lapic_watchdog_ok())
 496                         nmi_watchdog = NMI_LOCAL_APIC;
 497                 else
 498                         nmi_watchdog = NMI_IO_APIC;
 499         }
 500 #endif
 501
 502         if (nmi_watchdog == NMI_LOCAL_APIC) {
 503                 if (nmi_watchdog_enabled)
 504                         enable_lapic_nmi_watchdog();
 505                 else
 506                         disable_lapic_nmi_watchdog();
 507         } else {
 508                 printk(KERN_WARNING
 509                         "NMI watchdog doesn't know what hardware to touch\n");
 510                 return -EIO;
 511         }
 512         return 0;
 513 }
 514
 515 #endif /* CONFIG_SYSCTL */
 516
 517 int do_nmi_callback(struct pt_regs *regs, int cpu)
 518 {
 519 #ifdef CONFIG_SYSCTL
 520         if (unknown_nmi_panic)
 521                 return unknown_nmi_panic_callback(regs, cpu);
 522 #endif
 523         return 0;
 524 }
 525
 526 void __trigger_all_cpu_backtrace(void)
 527 {
 528         int i;
 529
 530         backtrace_mask = cpu_online_map;
 531         /* Wait for up to 10 seconds for all CPUs to do the backtrace */
 532         for (i = 0; i < 10 * 1000; i++) {
 533                 if (cpus_empty(backtrace_mask))
 534                         break;
 535                 mdelay(1);
 536         }
 537 }