arch/x86/kernel/nmi_32.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 <linux/delay.h>
  15 #include <linux/interrupt.h>
  16 #include <linux/module.h>
  17 #include <linux/nmi.h>
  18 #include <linux/sysdev.h>
  19 #include <linux/sysctl.h>
  20 #include <linux/percpu.h>
  21 #include <linux/kprobes.h>
  22 #include <linux/cpumask.h>
  23 #include <linux/kernel_stat.h>
  24 #include <linux/kdebug.h>
  25
  26 #include <asm/smp.h>
  27 #include <asm/nmi.h>
  28
  29 #include "mach_traps.h"
  30
  31 int unknown_nmi_panic;
  32 int nmi_watchdog_enabled;
  33
  34 static cpumask_t backtrace_mask = CPU_MASK_NONE;
  35
  36 /* nmi_active:
  37  * >0: the lapic NMI watchdog is active, but can be disabled
  38  * <0: the lapic NMI watchdog has not been set up, and cannot
  39  *     be enabled
  40  *  0: the lapic NMI watchdog is disabled, but can be enabled
  41  */
  42 atomic_t nmi_active = ATOMIC_INIT(0);           /* oprofile uses this */
  43
  44 unsigned int nmi_watchdog = NMI_DEFAULT;
  45 static unsigned int nmi_hz = HZ;
  46
  47 static DEFINE_PER_CPU(short, wd_enabled);
  48
  49 static int endflag __initdata = 0;
  50
  51 #ifdef CONFIG_SMP
  52 /* The performance counters used by NMI_LOCAL_APIC don't trigger when
  53  * the CPU is idle. To make sure the NMI watchdog really ticks on all
  54  * CPUs during the test make them busy.
  55  */
  56 static __init void nmi_cpu_busy(void *data)
  57 {
  58         local_irq_enable_in_hardirq();
  59         /* Intentionally don't use cpu_relax here. This is
  60            to make sure that the performance counter really ticks,
  61            even if there is a simulator or similar that catches the
  62            pause instruction. On a real HT machine this is fine because
  63            all other CPUs are busy with "useless" delay loops and don't
  64            care if they get somewhat less cycles. */
  65         while (endflag == 0)
  66                 mb();
  67 }
  68 #endif
  69
  70 static int __init check_nmi_watchdog(void)
  71 {
  72         unsigned int *prev_nmi_count;
  73         int cpu;
  74
  75         if ((nmi_watchdog == NMI_NONE) || (nmi_watchdog == NMI_DISABLED))
  76                 return 0;
  77
  78         if (!atomic_read(&nmi_active))
  79                 return 0;
  80
  81         prev_nmi_count = kmalloc(NR_CPUS * sizeof(int), GFP_KERNEL);
  82         if (!prev_nmi_count)
  83                 return -1;
  84
  85         printk(KERN_INFO "Testing NMI watchdog ... ");
  86
  87 #ifdef CONFIG_SMP
  88         if (nmi_watchdog == NMI_LOCAL_APIC)
  89                 smp_call_function(nmi_cpu_busy, (void *)&endflag, 0, 0);
  90 #endif
  91
  92         for_each_possible_cpu(cpu)
  93                 prev_nmi_count[cpu] = nmi_count(cpu);
  94         local_irq_enable();
  95         mdelay((20*1000)/nmi_hz); // wait 20 ticks
  96
  97         for_each_possible_cpu(cpu) {
  98 #ifdef CONFIG_SMP
  99                 /* Check cpu_callin_map here because that is set
 100                    after the timer is started. */
 101                 if (!cpu_isset(cpu, cpu_callin_map))
 102                         continue;
 103 #endif
 104                 if (!per_cpu(wd_enabled, cpu))
 105                         continue;
 106                 if (nmi_count(cpu) - prev_nmi_count[cpu] <= 5) {
 107                         printk(KERN_WARNING "WARNING: CPU#%d: NMI "
 108                                 "appears to be stuck (%d->%d)!\n",
 109                                 cpu,
 110                                 prev_nmi_count[cpu],
 111                                 nmi_count(cpu));
 112                         per_cpu(wd_enabled, cpu) = 0;
 113                         atomic_dec(&nmi_active);
 114                 }
 115         }
 116         endflag = 1;
 117         if (!atomic_read(&nmi_active)) {
 118                 kfree(prev_nmi_count);
 119                 atomic_set(&nmi_active, -1);
 120                 return -1;
 121         }
 122         printk("OK.\n");
 123
 124         /* now that we know it works we can reduce NMI frequency to
 125            something more reasonable; makes a difference in some configs */
 126         if (nmi_watchdog == NMI_LOCAL_APIC)
 127                 nmi_hz = lapic_adjust_nmi_hz(1);
 128
 129         kfree(prev_nmi_count);
 130         return 0;
 131 }
 132 /* This needs to happen later in boot so counters are working */
 133 late_initcall(check_nmi_watchdog);
 134
 135 static int __init setup_nmi_watchdog(char *str)
 136 {
 137         int nmi;
 138
 139         get_option(&str, &nmi);
 140
 141         if ((nmi >= NMI_INVALID) || (nmi < NMI_NONE))
 142                 return 0;
 143
 144         nmi_watchdog = nmi;
 145         return 1;
 146 }
 147
 148 __setup("nmi_watchdog=", setup_nmi_watchdog);
 149
 150
 151 /* Suspend/resume support */
 152
 153 #ifdef CONFIG_PM
 154
 155 static int nmi_pm_active; /* nmi_active before suspend */
 156
 157 static int lapic_nmi_suspend(struct sys_device *dev, pm_message_t state)
 158 {
 159         /* only CPU0 goes here, other CPUs should be offline */
 160         nmi_pm_active = atomic_read(&nmi_active);
 161         stop_apic_nmi_watchdog(NULL);
 162         BUG_ON(atomic_read(&nmi_active) != 0);
 163         return 0;
 164 }
 165
 166 static int lapic_nmi_resume(struct sys_device *dev)
 167 {
 168         /* only CPU0 goes here, other CPUs should be offline */
 169         if (nmi_pm_active > 0) {
 170                 setup_apic_nmi_watchdog(NULL);
 171                 touch_nmi_watchdog();
 172         }
 173         return 0;
 174 }
 175
 176
 177 static struct sysdev_class nmi_sysclass = {
 178         .name           = "lapic_nmi",
 179         .resume         = lapic_nmi_resume,
 180         .suspend        = lapic_nmi_suspend,
 181 };
 182
 183 static struct sys_device device_lapic_nmi = {
 184         .id     = 0,
 185         .cls    = &nmi_sysclass,
 186 };
 187
 188 static int __init init_lapic_nmi_sysfs(void)
 189 {
 190         int error;
 191
 192         /* should really be a BUG_ON but b/c this is an
 193          * init call, it just doesn't work.  -dcz
 194          */
 195         if (nmi_watchdog != NMI_LOCAL_APIC)
 196                 return 0;
 197
 198         if (atomic_read(&nmi_active) < 0)
 199                 return 0;
 200
 201         error = sysdev_class_register(&nmi_sysclass);
 202         if (!error)
 203                 error = sysdev_register(&device_lapic_nmi);
 204         return error;
 205 }
 206 /* must come after the local APIC's device_initcall() */
 207 late_initcall(init_lapic_nmi_sysfs);
 208
 209 #endif  /* CONFIG_PM */
 210
 211 static void __acpi_nmi_enable(void *__unused)
 212 {
 213         apic_write_around(APIC_LVT0, APIC_DM_NMI);
 214 }
 215
 216 /*
 217  * Enable timer based NMIs on all CPUs:
 218  */
 219 void acpi_nmi_enable(void)
 220 {
 221         if (atomic_read(&nmi_active) && nmi_watchdog == NMI_IO_APIC)
 222                 on_each_cpu(__acpi_nmi_enable, NULL, 0, 1);
 223 }
 224
 225 static void __acpi_nmi_disable(void *__unused)
 226 {
 227         apic_write(APIC_LVT0, APIC_DM_NMI | APIC_LVT_MASKED);
 228 }
 229
 230 /*
 231  * Disable timer based NMIs on all CPUs:
 232  */
 233 void acpi_nmi_disable(void)
 234 {
 235         if (atomic_read(&nmi_active) && nmi_watchdog == NMI_IO_APIC)
 236                 on_each_cpu(__acpi_nmi_disable, NULL, 0, 1);
 237 }
 238
 239 void setup_apic_nmi_watchdog(void *unused)
 240 {
 241         if (__get_cpu_var(wd_enabled))
 242                 return;
 243
 244         /* cheap hack to support suspend/resume */
 245         /* if cpu0 is not active neither should the other cpus */
 246         if ((smp_processor_id() != 0) && (atomic_read(&nmi_active) <= 0))
 247                 return;
 248
 249         switch (nmi_watchdog) {
 250         case NMI_LOCAL_APIC:
 251                 __get_cpu_var(wd_enabled) = 1; /* enable it before to avoid race with handler */
 252                 if (lapic_watchdog_init(nmi_hz) < 0) {
 253                         __get_cpu_var(wd_enabled) = 0;
 254                         return;
 255                 }
 256                 /* FALL THROUGH */
 257         case NMI_IO_APIC:
 258                 __get_cpu_var(wd_enabled) = 1;
 259                 atomic_inc(&nmi_active);
 260         }
 261 }
 262
 263 void stop_apic_nmi_watchdog(void *unused)
 264 {
 265         /* only support LOCAL and IO APICs for now */
 266         if ((nmi_watchdog != NMI_LOCAL_APIC) &&
 267             (nmi_watchdog != NMI_IO_APIC))
 268                 return;
 269         if (__get_cpu_var(wd_enabled) == 0)
 270                 return;
 271         if (nmi_watchdog == NMI_LOCAL_APIC)
 272                 lapic_watchdog_stop();
 273         __get_cpu_var(wd_enabled) = 0;
 274         atomic_dec(&nmi_active);
 275 }
 276
 277 /*
 278  * the best way to detect whether a CPU has a 'hard lockup' problem
 279  * is to check it's local APIC timer IRQ counts. If they are not
 280  * changing then that CPU has some problem.
 281  *
 282  * as these watchdog NMI IRQs are generated on every CPU, we only
 283  * have to check the current processor.
 284  *
 285  * since NMIs don't listen to _any_ locks, we have to be extremely
 286  * careful not to rely on unsafe variables. The printk might lock
 287  * up though, so we have to break up any console locks first ...
 288  * [when there will be more tty-related locks, break them up
 289  *  here too!]
 290  */
 291
 292 static unsigned int
 293         last_irq_sums [NR_CPUS],
 294         alert_counter [NR_CPUS];
 295
 296 void touch_nmi_watchdog(void)
 297 {
 298         if (nmi_watchdog > 0) {
 299                 unsigned cpu;
 300
 301                 /*
 302                  * Just reset the alert counters, (other CPUs might be
 303                  * spinning on locks we hold):
 304                  */
 305                 for_each_present_cpu(cpu) {
 306                         if (alert_counter[cpu])
 307                                 alert_counter[cpu] = 0;
 308                 }
 309         }
 310
 311         /*
 312          * Tickle the softlockup detector too:
 313          */
 314         touch_softlockup_watchdog();
 315 }
 316 EXPORT_SYMBOL(touch_nmi_watchdog);
 317
 318 extern void die_nmi(struct pt_regs *, const char *msg);
 319
 320 __kprobes int nmi_watchdog_tick(struct pt_regs * regs, unsigned reason)
 321 {
 322
 323         /*
 324          * Since current_thread_info()-> is always on the stack, and we
 325          * always switch the stack NMI-atomically, it's safe to use
 326          * smp_processor_id().
 327          */
 328         unsigned int sum;
 329         int touched = 0;
 330         int cpu = smp_processor_id();
 331         int rc = 0;
 332
 333         /* check for other users first */
 334         if (notify_die(DIE_NMI, "nmi", regs, reason, 2, SIGINT)
 335                         == NOTIFY_STOP) {
 336                 rc = 1;
 337                 touched = 1;
 338         }
 339
 340         if (cpu_isset(cpu, backtrace_mask)) {
 341                 static DEFINE_SPINLOCK(lock);   /* Serialise the printks */
 342
 343                 spin_lock(&lock);
 344                 printk("NMI backtrace for cpu %d\n", cpu);
 345                 dump_stack();
 346                 spin_unlock(&lock);
 347                 cpu_clear(cpu, backtrace_mask);
 348         }
 349
 350         /*
 351          * Take the local apic timer and PIT/HPET into account. We don't
 352          * know which one is active, when we have highres/dyntick on
 353          */
 354         sum = per_cpu(irq_stat, cpu).apic_timer_irqs +
 355                 per_cpu(irq_stat, cpu).irq0_irqs;
 356
 357         /* if the none of the timers isn't firing, this cpu isn't doing much */
 358         if (!touched && last_irq_sums[cpu] == sum) {
 359                 /*
 360                  * Ayiee, looks like this CPU is stuck ...
 361                  * wait a few IRQs (5 seconds) before doing the oops ...
 362                  */
 363                 alert_counter[cpu]++;
 364                 if (alert_counter[cpu] == 5*nmi_hz)
 365                         /*
 366                          * die_nmi will return ONLY if NOTIFY_STOP happens..
 367                          */
 368                         die_nmi(regs, "BUG: NMI Watchdog detected LOCKUP");
 369         } else {
 370                 last_irq_sums[cpu] = sum;
 371                 alert_counter[cpu] = 0;
 372         }
 373         /* see if the nmi watchdog went off */
 374         if (!__get_cpu_var(wd_enabled))
 375                 return rc;
 376         switch (nmi_watchdog) {
 377         case NMI_LOCAL_APIC:
 378                 rc |= lapic_wd_event(nmi_hz);
 379                 break;
 380         case NMI_IO_APIC:
 381                 /* don't know how to accurately check for this.
 382                  * just assume it was a watchdog timer interrupt
 383                  * This matches the old behaviour.
 384                  */
 385                 rc = 1;
 386                 break;
 387         }
 388         return rc;
 389 }
 390
 391 #ifdef CONFIG_SYSCTL
 392
 393 static int unknown_nmi_panic_callback(struct pt_regs *regs, int cpu)
 394 {
 395         unsigned char reason = get_nmi_reason();
 396         char buf[64];
 397
 398         sprintf(buf, "NMI received for unknown reason %02x\n", reason);
 399         die_nmi(regs, buf);
 400         return 0;
 401 }
 402
 403 /*
 404  * proc handler for /proc/sys/kernel/nmi
 405  */
 406 int proc_nmi_enabled(struct ctl_table *table, int write, struct file *file,
 407                         void __user *buffer, size_t *length, loff_t *ppos)
 408 {
 409         int old_state;
 410
 411         nmi_watchdog_enabled = (atomic_read(&nmi_active) > 0) ? 1 : 0;
 412         old_state = nmi_watchdog_enabled;
 413         proc_dointvec(table, write, file, buffer, length, ppos);
 414         if (!!old_state == !!nmi_watchdog_enabled)
 415                 return 0;
 416
 417         if (atomic_read(&nmi_active) < 0 || nmi_watchdog == NMI_DISABLED) {
 418                 printk( KERN_WARNING "NMI watchdog is permanently disabled\n");
 419                 return -EIO;
 420         }
 421
 422         if (nmi_watchdog == NMI_DEFAULT) {
 423                 if (lapic_watchdog_ok())
 424                         nmi_watchdog = NMI_LOCAL_APIC;
 425                 else
 426                         nmi_watchdog = NMI_IO_APIC;
 427         }
 428
 429         if (nmi_watchdog == NMI_LOCAL_APIC) {
 430                 if (nmi_watchdog_enabled)
 431                         enable_lapic_nmi_watchdog();
 432                 else
 433                         disable_lapic_nmi_watchdog();
 434         } else {
 435                 printk( KERN_WARNING
 436                         "NMI watchdog doesn't know what hardware to touch\n");
 437                 return -EIO;
 438         }
 439         return 0;
 440 }
 441
 442 #endif
 443
 444 int do_nmi_callback(struct pt_regs *regs, int cpu)
 445 {
 446 #ifdef CONFIG_SYSCTL
 447         if (unknown_nmi_panic)
 448                 return unknown_nmi_panic_callback(regs, cpu);
 449 #endif
 450         return 0;
 451 }
 452
 453 void __trigger_all_cpu_backtrace(void)
 454 {
 455         int i;
 456
 457         backtrace_mask = cpu_online_map;
 458         /* Wait for up to 10 seconds for all CPUs to do the backtrace */
 459         for (i = 0; i < 10 * 1000; i++) {
 460                 if (cpus_empty(backtrace_mask))
 461                         break;
 462                 mdelay(1);
 463         }
 464 }
 465
 466 EXPORT_SYMBOL(nmi_active);
 467 EXPORT_SYMBOL(nmi_watchdog);