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