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