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, 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         return -1;
 175 }
 176
 177 static int __init setup_nmi_watchdog(char *str)
 178 {
 179         unsigned int nmi;
 180
 181         if (!strncmp(str, "panic", 5)) {
 182                 panic_on_timeout = 1;
 183                 str = strchr(str, ',');
 184                 if (!str)
 185                         return 1;
 186                 ++str;
 187         }
 188
 189         get_option(&str, &nmi);
 190
 191         if (nmi >= NMI_INVALID)
 192                 return 0;
 193
 194         nmi_watchdog = nmi;
 195         return 1;
 196 }
 197 __setup("nmi_watchdog=", setup_nmi_watchdog);
 198
 199 /*
 200  * Suspend/resume support
 201  */
 202 #ifdef CONFIG_PM
 203
 204 static int nmi_pm_active; /* nmi_active before suspend */
 205
 206 static int lapic_nmi_suspend(struct sys_device *dev, pm_message_t state)
 207 {
 208         /* only CPU0 goes here, other CPUs should be offline */
 209         nmi_pm_active = atomic_read(&nmi_active);
 210         stop_apic_nmi_watchdog(NULL);
 211         BUG_ON(atomic_read(&nmi_active) != 0);
 212         return 0;
 213 }
 214
 215 static int lapic_nmi_resume(struct sys_device *dev)
 216 {
 217         /* only CPU0 goes here, other CPUs should be offline */
 218         if (nmi_pm_active > 0) {
 219                 setup_apic_nmi_watchdog(NULL);
 220                 touch_nmi_watchdog();
 221         }
 222         return 0;
 223 }
 224
 225 static struct sysdev_class nmi_sysclass = {
 226         .name           = "lapic_nmi",
 227         .resume         = lapic_nmi_resume,
 228         .suspend        = lapic_nmi_suspend,
 229 };
 230
 231 static struct sys_device device_lapic_nmi = {
 232         .id     = 0,
 233         .cls    = &nmi_sysclass,
 234 };
 235
 236 static int __init init_lapic_nmi_sysfs(void)
 237 {
 238         int error;
 239
 240         /*
 241          * should really be a BUG_ON but b/c this is an
 242          * init call, it just doesn't work.  -dcz
 243          */
 244         if (nmi_watchdog != NMI_LOCAL_APIC)
 245                 return 0;
 246
 247         if (atomic_read(&nmi_active) < 0)
 248                 return 0;
 249
 250         error = sysdev_class_register(&nmi_sysclass);
 251         if (!error)
 252                 error = sysdev_register(&device_lapic_nmi);
 253         return error;
 254 }
 255
 256 /* must come after the local APIC's device_initcall() */
 257 late_initcall(init_lapic_nmi_sysfs);
 258
 259 #endif  /* CONFIG_PM */
 260
 261 static void __acpi_nmi_enable(void *__unused)
 262 {
 263         apic_write_around(APIC_LVT0, APIC_DM_NMI);
 264 }
 265
 266 /*
 267  * Enable timer based NMIs on all CPUs:
 268  */
 269 void acpi_nmi_enable(void)
 270 {
 271         if (atomic_read(&nmi_active) && nmi_watchdog == NMI_IO_APIC)
 272                 on_each_cpu(__acpi_nmi_enable, NULL, 0, 1);
 273 }
 274
 275 static void __acpi_nmi_disable(void *__unused)
 276 {
 277         apic_write_around(APIC_LVT0, APIC_DM_NMI | APIC_LVT_MASKED);
 278 }
 279
 280 /*
 281  * Disable timer based NMIs on all CPUs:
 282  */
 283 void acpi_nmi_disable(void)
 284 {
 285         if (atomic_read(&nmi_active) && nmi_watchdog == NMI_IO_APIC)
 286                 on_each_cpu(__acpi_nmi_disable, NULL, 0, 1);
 287 }
 288
 289 void setup_apic_nmi_watchdog(void *unused)
 290 {
 291         if (__get_cpu_var(wd_enabled))
 292                 return;
 293
 294         /* cheap hack to support suspend/resume */
 295         /* if cpu0 is not active neither should the other cpus */
 296         if (smp_processor_id() != 0 && atomic_read(&nmi_active) <= 0)
 297                 return;
 298
 299         switch (nmi_watchdog) {
 300         case NMI_LOCAL_APIC:
 301                  /* enable it before to avoid race with handler */
 302                 __get_cpu_var(wd_enabled) = 1;
 303                 if (lapic_watchdog_init(nmi_hz) < 0) {
 304                         __get_cpu_var(wd_enabled) = 0;
 305                         return;
 306                 }
 307                 /* FALL THROUGH */
 308         case NMI_IO_APIC:
 309                 __get_cpu_var(wd_enabled) = 1;
 310                 atomic_inc(&nmi_active);
 311         }
 312 }
 313
 314 void stop_apic_nmi_watchdog(void *unused)
 315 {
 316         /* only support LOCAL and IO APICs for now */
 317         if (!nmi_watchdog_active())
 318                 return;
 319         if (__get_cpu_var(wd_enabled) == 0)
 320                 return;
 321         if (nmi_watchdog == NMI_LOCAL_APIC)
 322                 lapic_watchdog_stop();
 323         __get_cpu_var(wd_enabled) = 0;
 324         atomic_dec(&nmi_active);
 325 }
 326
 327 /*
 328  * the best way to detect whether a CPU has a 'hard lockup' problem
 329  * is to check it's local APIC timer IRQ counts. If they are not
 330  * changing then that CPU has some problem.
 331  *
 332  * as these watchdog NMI IRQs are generated on every CPU, we only
 333  * have to check the current processor.
 334  *
 335  * since NMIs don't listen to _any_ locks, we have to be extremely
 336  * careful not to rely on unsafe variables. The printk might lock
 337  * up though, so we have to break up any console locks first ...
 338  * [when there will be more tty-related locks, break them up here too!]
 339  */
 340
 341 static DEFINE_PER_CPU(unsigned, last_irq_sum);
 342 static DEFINE_PER_CPU(local_t, alert_counter);
 343 static DEFINE_PER_CPU(int, nmi_touch);
 344
 345 void touch_nmi_watchdog(void)
 346 {
 347         if (nmi_watchdog_active()) {
 348                 unsigned cpu;
 349
 350                 /*
 351                  * Tell other CPUs to reset their alert counters. We cannot
 352                  * do it ourselves because the alert count increase is not
 353                  * atomic.
 354                  */
 355                 for_each_present_cpu(cpu) {
 356                         if (per_cpu(nmi_touch, cpu) != 1)
 357                                 per_cpu(nmi_touch, cpu) = 1;
 358                 }
 359         }
 360
 361         /*
 362          * Tickle the softlockup detector too:
 363          */
 364         touch_softlockup_watchdog();
 365 }
 366 EXPORT_SYMBOL(touch_nmi_watchdog);
 367
 368 notrace __kprobes int
 369 nmi_watchdog_tick(struct pt_regs *regs, unsigned reason)
 370 {
 371         /*
 372          * Since current_thread_info()-> is always on the stack, and we
 373          * always switch the stack NMI-atomically, it's safe to use
 374          * smp_processor_id().
 375          */
 376         unsigned int sum;
 377         int touched = 0;
 378         int cpu = smp_processor_id();
 379         int rc = 0;
 380
 381         /* check for other users first */
 382         if (notify_die(DIE_NMI, "nmi", regs, reason, 2, SIGINT)
 383                         == NOTIFY_STOP) {
 384                 rc = 1;
 385                 touched = 1;
 386         }
 387
 388         sum = get_timer_irqs(cpu);
 389
 390         if (__get_cpu_var(nmi_touch)) {
 391                 __get_cpu_var(nmi_touch) = 0;
 392                 touched = 1;
 393         }
 394
 395         if (cpu_isset(cpu, backtrace_mask)) {
 396                 static DEFINE_SPINLOCK(lock);   /* Serialise the printks */
 397
 398                 spin_lock(&lock);
 399                 printk(KERN_WARNING "NMI backtrace for cpu %d\n", cpu);
 400                 dump_stack();
 401                 spin_unlock(&lock);
 402                 cpu_clear(cpu, backtrace_mask);
 403         }
 404
 405         /* Could check oops_in_progress here too, but it's safer not to */
 406         if (mce_in_progress())
 407                 touched = 1;
 408
 409         /* if the none of the timers isn't firing, this cpu isn't doing much */
 410         if (!touched && __get_cpu_var(last_irq_sum) == sum) {
 411                 /*
 412                  * Ayiee, looks like this CPU is stuck ...
 413                  * wait a few IRQs (5 seconds) before doing the oops ...
 414                  */
 415                 local_inc(&__get_cpu_var(alert_counter));
 416                 if (local_read(&__get_cpu_var(alert_counter)) == 5 * nmi_hz)
 417                         /*
 418                          * die_nmi will return ONLY if NOTIFY_STOP happens..
 419                          */
 420                         die_nmi("BUG: NMI Watchdog detected LOCKUP",
 421                                 regs, panic_on_timeout);
 422         } else {
 423                 __get_cpu_var(last_irq_sum) = sum;
 424                 local_set(&__get_cpu_var(alert_counter), 0);
 425         }
 426
 427         /* see if the nmi watchdog went off */
 428         if (!__get_cpu_var(wd_enabled))
 429                 return rc;
 430         switch (nmi_watchdog) {
 431         case NMI_LOCAL_APIC:
 432                 rc |= lapic_wd_event(nmi_hz);
 433                 break;
 434         case NMI_IO_APIC:
 435                 /*
 436                  * don't know how to accurately check for this.
 437                  * just assume it was a watchdog timer interrupt
 438                  * This matches the old behaviour.
 439                  */
 440                 rc = 1;
 441                 break;
 442         }
 443         return rc;
 444 }
 445
 446 #ifdef CONFIG_SYSCTL
 447
 448 static int unknown_nmi_panic_callback(struct pt_regs *regs, int cpu)
 449 {
 450         unsigned char reason = get_nmi_reason();
 451         char buf[64];
 452
 453         sprintf(buf, "NMI received for unknown reason %02x\n", reason);
 454         die_nmi(buf, regs, 1); /* Always panic here */
 455         return 0;
 456 }
 457
 458 /*
 459  * proc handler for /proc/sys/kernel/nmi
 460  */
 461 int proc_nmi_enabled(struct ctl_table *table, int write, struct file *file,
 462                         void __user *buffer, size_t *length, loff_t *ppos)
 463 {
 464         int old_state;
 465
 466         nmi_watchdog_enabled = (atomic_read(&nmi_active) > 0) ? 1 : 0;
 467         old_state = nmi_watchdog_enabled;
 468         proc_dointvec(table, write, file, buffer, length, ppos);
 469         if (!!old_state == !!nmi_watchdog_enabled)
 470                 return 0;
 471
 472         if (atomic_read(&nmi_active) < 0 || !nmi_watchdog_active()) {
 473                 printk(KERN_WARNING
 474                         "NMI watchdog is permanently disabled\n");
 475                 return -EIO;
 476         }
 477
 478         if (nmi_watchdog == NMI_LOCAL_APIC) {
 479                 if (nmi_watchdog_enabled)
 480                         enable_lapic_nmi_watchdog();
 481                 else
 482                         disable_lapic_nmi_watchdog();
 483         } else {
 484                 printk(KERN_WARNING
 485                         "NMI watchdog doesn't know what hardware to touch\n");
 486                 return -EIO;
 487         }
 488         return 0;
 489 }
 490
 491 #endif /* CONFIG_SYSCTL */
 492
 493 int do_nmi_callback(struct pt_regs *regs, int cpu)
 494 {
 495 #ifdef CONFIG_SYSCTL
 496         if (unknown_nmi_panic)
 497                 return unknown_nmi_panic_callback(regs, cpu);
 498 #endif
 499         return 0;
 500 }
 501
 502 void __trigger_all_cpu_backtrace(void)
 503 {
 504         int i;
 505
 506         backtrace_mask = cpu_online_map;
 507         /* Wait for up to 10 seconds for all CPUs to do the backtrace */
 508         for (i = 0; i < 10 * 1000; i++) {
 509                 if (cpus_empty(backtrace_mask))
 510                         break;
 511                 mdelay(1);
 512         }
 513 }