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/i8259.h>
  31 #include <asm/io_apic.h>
  32 #include <asm/smp.h>
  33 #include <asm/nmi.h>
  34 #include <asm/proto.h>
  35 #include <asm/timer.h>
  36
  37 #include <asm/mce.h>
  38
  39 #include <mach_traps.h>
  40
  41 int unknown_nmi_panic;
  42 int nmi_watchdog_enabled;
  43
  44 static cpumask_t backtrace_mask = CPU_MASK_NONE;
  45
  46 /* nmi_active:
  47  * >0: the lapic NMI watchdog is active, but can be disabled
  48  * <0: the lapic NMI watchdog has not been set up, and cannot
  49  *     be enabled
  50  *  0: the lapic NMI watchdog is disabled, but can be enabled
  51  */
  52 atomic_t nmi_active = ATOMIC_INIT(0);           /* oprofile uses this */
  53 EXPORT_SYMBOL(nmi_active);
  54
  55 unsigned int nmi_watchdog = NMI_NONE;
  56 EXPORT_SYMBOL(nmi_watchdog);
  57
  58 static int panic_on_timeout;
  59
  60 static unsigned int nmi_hz = HZ;
  61 static DEFINE_PER_CPU(short, wd_enabled);
  62 static int endflag __initdata;
  63
  64 static inline unsigned int get_nmi_count(int cpu)
  65 {
  66 #ifdef CONFIG_X86_64
  67         return cpu_pda(cpu)->__nmi_count;
  68 #else
  69         return nmi_count(cpu);
  70 #endif
  71 }
  72
  73 static inline int mce_in_progress(void)
  74 {
  75 #if defined(CONFIG_X86_64) && defined(CONFIG_X86_MCE)
  76         return atomic_read(&mce_entry) > 0;
  77 #endif
  78         return 0;
  79 }
  80
  81 /*
  82  * Take the local apic timer and PIT/HPET into account. We don't
  83  * know which one is active, when we have highres/dyntick on
  84  */
  85 static inline unsigned int get_timer_irqs(int cpu)
  86 {
  87 #ifdef CONFIG_X86_64
  88         return read_pda(apic_timer_irqs) + read_pda(irq0_irqs);
  89 #else
  90         return per_cpu(irq_stat, cpu).apic_timer_irqs +
  91                 per_cpu(irq_stat, cpu).irq0_irqs;
  92 #endif
  93 }
  94
  95 #ifdef CONFIG_SMP
  96 /*
  97  * The performance counters used by NMI_LOCAL_APIC don't trigger when
  98  * the CPU is idle. To make sure the NMI watchdog really ticks on all
  99  * CPUs during the test make them busy.
 100  */
 101 static __init void nmi_cpu_busy(void *data)
 102 {
 103         local_irq_enable_in_hardirq();
 104         /*
 105          * Intentionally don't use cpu_relax here. This is
 106          * to make sure that the performance counter really ticks,
 107          * even if there is a simulator or similar that catches the
 108          * pause instruction. On a real HT machine this is fine because
 109          * all other CPUs are busy with "useless" delay loops and don't
 110          * care if they get somewhat less cycles.
 111          */
 112         while (endflag == 0)
 113                 mb();
 114 }
 115 #endif
 116
 117 int __init check_nmi_watchdog(void)
 118 {
 119         unsigned int *prev_nmi_count;
 120         int cpu;
 121
 122         if (!nmi_watchdog_active() || !atomic_read(&nmi_active))
 123                 return 0;
 124
 125         prev_nmi_count = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL);
 126         if (!prev_nmi_count)
 127                 goto error;
 128
 129         printk(KERN_INFO "Testing NMI watchdog ... ");
 130
 131 #ifdef CONFIG_SMP
 132         if (nmi_watchdog == NMI_LOCAL_APIC)
 133                 smp_call_function(nmi_cpu_busy, (void *)&endflag, 0);
 134 #endif
 135
 136         for_each_possible_cpu(cpu)
 137                 prev_nmi_count[cpu] = get_nmi_count(cpu);
 138         local_irq_enable();
 139         mdelay((20 * 1000) / nmi_hz); /* wait 20 ticks */
 140
 141         for_each_online_cpu(cpu) {
 142                 if (!per_cpu(wd_enabled, cpu))
 143                         continue;
 144                 if (get_nmi_count(cpu) - prev_nmi_count[cpu] <= 5) {
 145                         printk(KERN_WARNING "WARNING: CPU#%d: NMI "
 146                                 "appears to be stuck (%d->%d)!\n",
 147                                 cpu,
 148                                 prev_nmi_count[cpu],
 149                                 get_nmi_count(cpu));
 150                         per_cpu(wd_enabled, cpu) = 0;
 151                         atomic_dec(&nmi_active);
 152                 }
 153         }
 154         endflag = 1;
 155         if (!atomic_read(&nmi_active)) {
 156                 kfree(prev_nmi_count);
 157                 atomic_set(&nmi_active, -1);
 158                 goto error;
 159         }
 160         printk("OK.\n");
 161
 162         /*
 163          * now that we know it works we can reduce NMI frequency to
 164          * something more reasonable; makes a difference in some configs
 165          */
 166         if (nmi_watchdog == NMI_LOCAL_APIC)
 167                 nmi_hz = lapic_adjust_nmi_hz(1);
 168
 169         kfree(prev_nmi_count);
 170         return 0;
 171 error:
 172         if (nmi_watchdog == NMI_IO_APIC && !timer_through_8259)
 173                 disable_8259A_irq(0);
 174 #ifdef CONFIG_X86_32
 175         timer_ack = 0;
 176 #endif
 177         return -1;
 178 }
 179
 180 static int __init setup_nmi_watchdog(char *str)
 181 {
 182         unsigned int nmi;
 183
 184         if (!strncmp(str, "panic", 5)) {
 185                 panic_on_timeout = 1;
 186                 str = strchr(str, ',');
 187                 if (!str)
 188                         return 1;
 189                 ++str;
 190         }
 191
 192         get_option(&str, &nmi);
 193
 194         if (nmi >= NMI_INVALID)
 195                 return 0;
 196
 197         nmi_watchdog = nmi;
 198         return 1;
 199 }
 200 __setup("nmi_watchdog=", setup_nmi_watchdog);
 201
 202 /*
 203  * Suspend/resume support
 204  */
 205 #ifdef CONFIG_PM
 206
 207 static int nmi_pm_active; /* nmi_active before suspend */
 208
 209 static int lapic_nmi_suspend(struct sys_device *dev, pm_message_t state)
 210 {
 211         /* only CPU0 goes here, other CPUs should be offline */
 212         nmi_pm_active = atomic_read(&nmi_active);
 213         stop_apic_nmi_watchdog(NULL);
 214         BUG_ON(atomic_read(&nmi_active) != 0);
 215         return 0;
 216 }
 217
 218 static int lapic_nmi_resume(struct sys_device *dev)
 219 {
 220         /* only CPU0 goes here, other CPUs should be offline */
 221         if (nmi_pm_active > 0) {
 222                 setup_apic_nmi_watchdog(NULL);
 223                 touch_nmi_watchdog();
 224         }
 225         return 0;
 226 }
 227
 228 static struct sysdev_class nmi_sysclass = {
 229         .name           = "lapic_nmi",
 230         .resume         = lapic_nmi_resume,
 231         .suspend        = lapic_nmi_suspend,
 232 };
 233
 234 static struct sys_device device_lapic_nmi = {
 235         .id     = 0,
 236         .cls    = &nmi_sysclass,
 237 };
 238
 239 static int __init init_lapic_nmi_sysfs(void)
 240 {
 241         int error;
 242
 243         /*
 244          * should really be a BUG_ON but b/c this is an
 245          * init call, it just doesn't work.  -dcz
 246          */
 247         if (nmi_watchdog != NMI_LOCAL_APIC)
 248                 return 0;
 249
 250         if (atomic_read(&nmi_active) < 0)
 251                 return 0;
 252
 253         error = sysdev_class_register(&nmi_sysclass);
 254         if (!error)
 255                 error = sysdev_register(&device_lapic_nmi);
 256         return error;
 257 }
 258
 259 /* must come after the local APIC's device_initcall() */
 260 late_initcall(init_lapic_nmi_sysfs);
 261
 262 #endif  /* CONFIG_PM */
 263
 264 static void __acpi_nmi_enable(void *__unused)
 265 {
 266         apic_write_around(APIC_LVT0, APIC_DM_NMI);
 267 }
 268
 269 /*
 270  * Enable timer based NMIs on all CPUs:
 271  */
 272 void acpi_nmi_enable(void)
 273 {
 274         if (atomic_read(&nmi_active) && nmi_watchdog == NMI_IO_APIC)
 275                 on_each_cpu(__acpi_nmi_enable, NULL, 1);
 276 }
 277
 278 static void __acpi_nmi_disable(void *__unused)
 279 {
 280         apic_write_around(APIC_LVT0, APIC_DM_NMI | APIC_LVT_MASKED);
 281 }
 282
 283 /*
 284  * Disable timer based NMIs on all CPUs:
 285  */
 286 void acpi_nmi_disable(void)
 287 {
 288         if (atomic_read(&nmi_active) && nmi_watchdog == NMI_IO_APIC)
 289                 on_each_cpu(__acpi_nmi_disable, NULL, 1);
 290 }
 291
 292 void setup_apic_nmi_watchdog(void *unused)
 293 {
 294         if (__get_cpu_var(wd_enabled))
 295                 return;
 296
 297         /* cheap hack to support suspend/resume */
 298         /* if cpu0 is not active neither should the other cpus */
 299         if (smp_processor_id() != 0 && atomic_read(&nmi_active) <= 0)
 300                 return;
 301
 302         switch (nmi_watchdog) {
 303         case NMI_LOCAL_APIC:
 304                  /* enable it before to avoid race with handler */
 305                 __get_cpu_var(wd_enabled) = 1;
 306                 if (lapic_watchdog_init(nmi_hz) < 0) {
 307                         __get_cpu_var(wd_enabled) = 0;
 308                         return;
 309                 }
 310                 /* FALL THROUGH */
 311         case NMI_IO_APIC:
 312                 __get_cpu_var(wd_enabled) = 1;
 313                 atomic_inc(&nmi_active);
 314         }
 315 }
 316
 317 void stop_apic_nmi_watchdog(void *unused)
 318 {
 319         /* only support LOCAL and IO APICs for now */
 320         if (!nmi_watchdog_active())
 321                 return;
 322         if (__get_cpu_var(wd_enabled) == 0)
 323                 return;
 324         if (nmi_watchdog == NMI_LOCAL_APIC)
 325                 lapic_watchdog_stop();
 326         __get_cpu_var(wd_enabled) = 0;
 327         atomic_dec(&nmi_active);
 328 }
 329
 330 /*
 331  * the best way to detect whether a CPU has a 'hard lockup' problem
 332  * is to check it's local APIC timer IRQ counts. If they are not
 333  * changing then that CPU has some problem.
 334  *
 335  * as these watchdog NMI IRQs are generated on every CPU, we only
 336  * have to check the current processor.
 337  *
 338  * since NMIs don't listen to _any_ locks, we have to be extremely
 339  * careful not to rely on unsafe variables. The printk might lock
 340  * up though, so we have to break up any console locks first ...
 341  * [when there will be more tty-related locks, break them up here too!]
 342  */
 343
 344 static DEFINE_PER_CPU(unsigned, last_irq_sum);
 345 static DEFINE_PER_CPU(local_t, alert_counter);
 346 static DEFINE_PER_CPU(int, nmi_touch);
 347
 348 void touch_nmi_watchdog(void)
 349 {
 350         if (nmi_watchdog_active()) {
 351                 unsigned cpu;
 352
 353                 /*
 354                  * Tell other CPUs to reset their alert counters. We cannot
 355                  * do it ourselves because the alert count increase is not
 356                  * atomic.
 357                  */
 358                 for_each_present_cpu(cpu) {
 359                         if (per_cpu(nmi_touch, cpu) != 1)
 360                                 per_cpu(nmi_touch, cpu) = 1;
 361                 }
 362         }
 363
 364         /*
 365          * Tickle the softlockup detector too:
 366          */
 367         touch_softlockup_watchdog();
 368 }
 369 EXPORT_SYMBOL(touch_nmi_watchdog);
 370
 371 notrace __kprobes int
 372 nmi_watchdog_tick(struct pt_regs *regs, unsigned reason)
 373 {
 374         /*
 375          * Since current_thread_info()-> is always on the stack, and we
 376          * always switch the stack NMI-atomically, it's safe to use
 377          * smp_processor_id().
 378          */
 379         unsigned int sum;
 380         int touched = 0;
 381         int cpu = smp_processor_id();
 382         int rc = 0;
 383
 384         /* check for other users first */
 385         if (notify_die(DIE_NMI, "nmi", regs, reason, 2, SIGINT)
 386                         == NOTIFY_STOP) {
 387                 rc = 1;
 388                 touched = 1;
 389         }
 390
 391         sum = get_timer_irqs(cpu);
 392
 393         if (__get_cpu_var(nmi_touch)) {
 394                 __get_cpu_var(nmi_touch) = 0;
 395                 touched = 1;
 396         }
 397
 398         if (cpu_isset(cpu, backtrace_mask)) {
 399                 static DEFINE_SPINLOCK(lock);   /* Serialise the printks */
 400
 401                 spin_lock(&lock);
 402                 printk(KERN_WARNING "NMI backtrace for cpu %d\n", cpu);
 403                 dump_stack();
 404                 spin_unlock(&lock);
 405                 cpu_clear(cpu, backtrace_mask);
 406         }
 407
 408         /* Could check oops_in_progress here too, but it's safer not to */
 409         if (mce_in_progress())
 410                 touched = 1;
 411
 412         /* if the none of the timers isn't firing, this cpu isn't doing much */
 413         if (!touched && __get_cpu_var(last_irq_sum) == sum) {
 414                 /*
 415                  * Ayiee, looks like this CPU is stuck ...
 416                  * wait a few IRQs (5 seconds) before doing the oops ...
 417                  */
 418                 local_inc(&__get_cpu_var(alert_counter));
 419                 if (local_read(&__get_cpu_var(alert_counter)) == 5 * nmi_hz)
 420                         /*
 421                          * die_nmi will return ONLY if NOTIFY_STOP happens..
 422                          */
 423                         die_nmi("BUG: NMI Watchdog detected LOCKUP",
 424                                 regs, panic_on_timeout);
 425         } else {
 426                 __get_cpu_var(last_irq_sum) = sum;
 427                 local_set(&__get_cpu_var(alert_counter), 0);
 428         }
 429
 430         /* see if the nmi watchdog went off */
 431         if (!__get_cpu_var(wd_enabled))
 432                 return rc;
 433         switch (nmi_watchdog) {
 434         case NMI_LOCAL_APIC:
 435                 rc |= lapic_wd_event(nmi_hz);
 436                 break;
 437         case NMI_IO_APIC:
 438                 /*
 439                  * don't know how to accurately check for this.
 440                  * just assume it was a watchdog timer interrupt
 441                  * This matches the old behaviour.
 442                  */
 443                 rc = 1;
 444                 break;
 445         }
 446         return rc;
 447 }
 448
 449 #ifdef CONFIG_SYSCTL
 450
 451 static int unknown_nmi_panic_callback(struct pt_regs *regs, int cpu)
 452 {
 453         unsigned char reason = get_nmi_reason();
 454         char buf[64];
 455
 456         sprintf(buf, "NMI received for unknown reason %02x\n", reason);
 457         die_nmi(buf, regs, 1); /* Always panic here */
 458         return 0;
 459 }
 460
 461 /*
 462  * proc handler for /proc/sys/kernel/nmi
 463  */
 464 int proc_nmi_enabled(struct ctl_table *table, int write, struct file *file,
 465                         void __user *buffer, size_t *length, loff_t *ppos)
 466 {
 467         int old_state;
 468
 469         nmi_watchdog_enabled = (atomic_read(&nmi_active) > 0) ? 1 : 0;
 470         old_state = nmi_watchdog_enabled;
 471         proc_dointvec(table, write, file, buffer, length, ppos);
 472         if (!!old_state == !!nmi_watchdog_enabled)
 473                 return 0;
 474
 475         if (atomic_read(&nmi_active) < 0 || !nmi_watchdog_active()) {
 476                 printk(KERN_WARNING
 477                         "NMI watchdog is permanently disabled\n");
 478                 return -EIO;
 479         }
 480
 481         if (nmi_watchdog == NMI_LOCAL_APIC) {
 482                 if (nmi_watchdog_enabled)
 483                         enable_lapic_nmi_watchdog();
 484                 else
 485                         disable_lapic_nmi_watchdog();
 486         } else {
 487                 printk(KERN_WARNING
 488                         "NMI watchdog doesn't know what hardware to touch\n");
 489                 return -EIO;
 490         }
 491         return 0;
 492 }
 493
 494 #endif /* CONFIG_SYSCTL */
 495
 496 int do_nmi_callback(struct pt_regs *regs, int cpu)
 497 {
 498 #ifdef CONFIG_SYSCTL
 499         if (unknown_nmi_panic)
 500                 return unknown_nmi_panic_callback(regs, cpu);
 501 #endif
 502         return 0;
 503 }
 504
 505 void __trigger_all_cpu_backtrace(void)
 506 {
 507         int i;
 508
 509         backtrace_mask = cpu_online_map;
 510         /* Wait for up to 10 seconds for all CPUs to do the backtrace */
 511         for (i = 0; i < 10 * 1000; i++) {
 512                 if (cpus_empty(backtrace_mask))
 513                         break;
 514                 mdelay(1);
 515         }
 516 }