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 static void report_broken_nmi(int cpu, int *prev_nmi_count)
 118 {
 119         printk(KERN_CONT "\n");
 120
 121         printk(KERN_WARNING
 122                 "WARNING: CPU#%d: NMI appears to be stuck (%d->%d)!\n",
 123                         cpu, prev_nmi_count[cpu], get_nmi_count(cpu));
 124
 125         printk(KERN_WARNING
 126                 "Please report this to bugzilla.kernel.org,\n");
 127         printk(KERN_WARNING
 128                 "and attach the output of the 'dmesg' command.\n");
 129
 130         per_cpu(wd_enabled, cpu) = 0;
 131         atomic_dec(&nmi_active);
 132 }
 133
 134 int __init check_nmi_watchdog(void)
 135 {
 136         unsigned int *prev_nmi_count;
 137         int cpu;
 138
 139         if (!nmi_watchdog_active() || !atomic_read(&nmi_active))
 140                 return 0;
 141
 142         prev_nmi_count = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL);
 143         if (!prev_nmi_count)
 144                 goto error;
 145
 146         printk(KERN_INFO "Testing NMI watchdog ... ");
 147
 148 #ifdef CONFIG_SMP
 149         if (nmi_watchdog == NMI_LOCAL_APIC)
 150                 smp_call_function(nmi_cpu_busy, (void *)&endflag, 0);
 151 #endif
 152
 153         for_each_possible_cpu(cpu)
 154                 prev_nmi_count[cpu] = get_nmi_count(cpu);
 155         local_irq_enable();
 156         mdelay((20 * 1000) / nmi_hz); /* wait 20 ticks */
 157
 158         for_each_online_cpu(cpu) {
 159                 if (!per_cpu(wd_enabled, cpu))
 160                         continue;
 161                 if (get_nmi_count(cpu) - prev_nmi_count[cpu] <= 5)
 162                         report_broken_nmi(cpu, prev_nmi_count);
 163         }
 164         endflag = 1;
 165         if (!atomic_read(&nmi_active)) {
 166                 kfree(prev_nmi_count);
 167                 atomic_set(&nmi_active, -1);
 168                 goto error;
 169         }
 170         printk("OK.\n");
 171
 172         /*
 173          * now that we know it works we can reduce NMI frequency to
 174          * something more reasonable; makes a difference in some configs
 175          */
 176         if (nmi_watchdog == NMI_LOCAL_APIC)
 177                 nmi_hz = lapic_adjust_nmi_hz(1);
 178
 179         kfree(prev_nmi_count);
 180         return 0;
 181 error:
 182         if (nmi_watchdog == NMI_IO_APIC && !timer_through_8259)
 183                 disable_8259A_irq(0);
 184 #ifdef CONFIG_X86_32
 185         timer_ack = 0;
 186 #endif
 187         return -1;
 188 }
 189
 190 static int __init setup_nmi_watchdog(char *str)
 191 {
 192         unsigned int nmi;
 193
 194         if (!strncmp(str, "panic", 5)) {
 195                 panic_on_timeout = 1;
 196                 str = strchr(str, ',');
 197                 if (!str)
 198                         return 1;
 199                 ++str;
 200         }
 201
 202         get_option(&str, &nmi);
 203
 204         if (nmi >= NMI_INVALID)
 205                 return 0;
 206
 207         nmi_watchdog = nmi;
 208         return 1;
 209 }
 210 __setup("nmi_watchdog=", setup_nmi_watchdog);
 211
 212 /*
 213  * Suspend/resume support
 214  */
 215 #ifdef CONFIG_PM
 216
 217 static int nmi_pm_active; /* nmi_active before suspend */
 218
 219 static int lapic_nmi_suspend(struct sys_device *dev, pm_message_t state)
 220 {
 221         /* only CPU0 goes here, other CPUs should be offline */
 222         nmi_pm_active = atomic_read(&nmi_active);
 223         stop_apic_nmi_watchdog(NULL);
 224         BUG_ON(atomic_read(&nmi_active) != 0);
 225         return 0;
 226 }
 227
 228 static int lapic_nmi_resume(struct sys_device *dev)
 229 {
 230         /* only CPU0 goes here, other CPUs should be offline */
 231         if (nmi_pm_active > 0) {
 232                 setup_apic_nmi_watchdog(NULL);
 233                 touch_nmi_watchdog();
 234         }
 235         return 0;
 236 }
 237
 238 static struct sysdev_class nmi_sysclass = {
 239         .name           = "lapic_nmi",
 240         .resume         = lapic_nmi_resume,
 241         .suspend        = lapic_nmi_suspend,
 242 };
 243
 244 static struct sys_device device_lapic_nmi = {
 245         .id     = 0,
 246         .cls    = &nmi_sysclass,
 247 };
 248
 249 static int __init init_lapic_nmi_sysfs(void)
 250 {
 251         int error;
 252
 253         /*
 254          * should really be a BUG_ON but b/c this is an
 255          * init call, it just doesn't work.  -dcz
 256          */
 257         if (nmi_watchdog != NMI_LOCAL_APIC)
 258                 return 0;
 259
 260         if (atomic_read(&nmi_active) < 0)
 261                 return 0;
 262
 263         error = sysdev_class_register(&nmi_sysclass);
 264         if (!error)
 265                 error = sysdev_register(&device_lapic_nmi);
 266         return error;
 267 }
 268
 269 /* must come after the local APIC's device_initcall() */
 270 late_initcall(init_lapic_nmi_sysfs);
 271
 272 #endif  /* CONFIG_PM */
 273
 274 static void __acpi_nmi_enable(void *__unused)
 275 {
 276         apic_write(APIC_LVT0, APIC_DM_NMI);
 277 }
 278
 279 /*
 280  * Enable timer based NMIs on all CPUs:
 281  */
 282 void acpi_nmi_enable(void)
 283 {
 284         if (atomic_read(&nmi_active) && nmi_watchdog == NMI_IO_APIC)
 285                 on_each_cpu(__acpi_nmi_enable, NULL, 1);
 286 }
 287
 288 static void __acpi_nmi_disable(void *__unused)
 289 {
 290         apic_write(APIC_LVT0, APIC_DM_NMI | APIC_LVT_MASKED);
 291 }
 292
 293 /*
 294  * Disable timer based NMIs on all CPUs:
 295  */
 296 void acpi_nmi_disable(void)
 297 {
 298         if (atomic_read(&nmi_active) && nmi_watchdog == NMI_IO_APIC)
 299                 on_each_cpu(__acpi_nmi_disable, NULL, 1);
 300 }
 301
 302 void setup_apic_nmi_watchdog(void *unused)
 303 {
 304         if (__get_cpu_var(wd_enabled))
 305                 return;
 306
 307         /* cheap hack to support suspend/resume */
 308         /* if cpu0 is not active neither should the other cpus */
 309         if (smp_processor_id() != 0 && atomic_read(&nmi_active) <= 0)
 310                 return;
 311
 312         switch (nmi_watchdog) {
 313         case NMI_LOCAL_APIC:
 314                  /* enable it before to avoid race with handler */
 315                 __get_cpu_var(wd_enabled) = 1;
 316                 if (lapic_watchdog_init(nmi_hz) < 0) {
 317                         __get_cpu_var(wd_enabled) = 0;
 318                         return;
 319                 }
 320                 /* FALL THROUGH */
 321         case NMI_IO_APIC:
 322                 __get_cpu_var(wd_enabled) = 1;
 323                 atomic_inc(&nmi_active);
 324         }
 325 }
 326
 327 void stop_apic_nmi_watchdog(void *unused)
 328 {
 329         /* only support LOCAL and IO APICs for now */
 330         if (!nmi_watchdog_active())
 331                 return;
 332         if (__get_cpu_var(wd_enabled) == 0)
 333                 return;
 334         if (nmi_watchdog == NMI_LOCAL_APIC)
 335                 lapic_watchdog_stop();
 336         __get_cpu_var(wd_enabled) = 0;
 337         atomic_dec(&nmi_active);
 338 }
 339
 340 /*
 341  * the best way to detect whether a CPU has a 'hard lockup' problem
 342  * is to check it's local APIC timer IRQ counts. If they are not
 343  * changing then that CPU has some problem.
 344  *
 345  * as these watchdog NMI IRQs are generated on every CPU, we only
 346  * have to check the current processor.
 347  *
 348  * since NMIs don't listen to _any_ locks, we have to be extremely
 349  * careful not to rely on unsafe variables. The printk might lock
 350  * up though, so we have to break up any console locks first ...
 351  * [when there will be more tty-related locks, break them up here too!]
 352  */
 353
 354 static DEFINE_PER_CPU(unsigned, last_irq_sum);
 355 static DEFINE_PER_CPU(local_t, alert_counter);
 356 static DEFINE_PER_CPU(int, nmi_touch);
 357
 358 void touch_nmi_watchdog(void)
 359 {
 360         if (nmi_watchdog_active()) {
 361                 unsigned cpu;
 362
 363                 /*
 364                  * Tell other CPUs to reset their alert counters. We cannot
 365                  * do it ourselves because the alert count increase is not
 366                  * atomic.
 367                  */
 368                 for_each_present_cpu(cpu) {
 369                         if (per_cpu(nmi_touch, cpu) != 1)
 370                                 per_cpu(nmi_touch, cpu) = 1;
 371                 }
 372         }
 373
 374         /*
 375          * Tickle the softlockup detector too:
 376          */
 377         touch_softlockup_watchdog();
 378 }
 379 EXPORT_SYMBOL(touch_nmi_watchdog);
 380
 381 notrace __kprobes int
 382 nmi_watchdog_tick(struct pt_regs *regs, unsigned reason)
 383 {
 384         /*
 385          * Since current_thread_info()-> is always on the stack, and we
 386          * always switch the stack NMI-atomically, it's safe to use
 387          * smp_processor_id().
 388          */
 389         unsigned int sum;
 390         int touched = 0;
 391         int cpu = smp_processor_id();
 392         int rc = 0;
 393
 394         /* check for other users first */
 395         if (notify_die(DIE_NMI, "nmi", regs, reason, 2, SIGINT)
 396                         == NOTIFY_STOP) {
 397                 rc = 1;
 398                 touched = 1;
 399         }
 400
 401         sum = get_timer_irqs(cpu);
 402
 403         if (__get_cpu_var(nmi_touch)) {
 404                 __get_cpu_var(nmi_touch) = 0;
 405                 touched = 1;
 406         }
 407
 408         if (cpu_isset(cpu, backtrace_mask)) {
 409                 static DEFINE_SPINLOCK(lock);   /* Serialise the printks */
 410
 411                 spin_lock(&lock);
 412                 printk(KERN_WARNING "NMI backtrace for cpu %d\n", cpu);
 413                 dump_stack();
 414                 spin_unlock(&lock);
 415                 cpu_clear(cpu, backtrace_mask);
 416         }
 417
 418         /* Could check oops_in_progress here too, but it's safer not to */
 419         if (mce_in_progress())
 420                 touched = 1;
 421
 422         /* if the none of the timers isn't firing, this cpu isn't doing much */
 423         if (!touched && __get_cpu_var(last_irq_sum) == sum) {
 424                 /*
 425                  * Ayiee, looks like this CPU is stuck ...
 426                  * wait a few IRQs (5 seconds) before doing the oops ...
 427                  */
 428                 local_inc(&__get_cpu_var(alert_counter));
 429                 if (local_read(&__get_cpu_var(alert_counter)) == 5 * nmi_hz)
 430                         /*
 431                          * die_nmi will return ONLY if NOTIFY_STOP happens..
 432                          */
 433                         die_nmi("BUG: NMI Watchdog detected LOCKUP",
 434                                 regs, panic_on_timeout);
 435         } else {
 436                 __get_cpu_var(last_irq_sum) = sum;
 437                 local_set(&__get_cpu_var(alert_counter), 0);
 438         }
 439
 440         /* see if the nmi watchdog went off */
 441         if (!__get_cpu_var(wd_enabled))
 442                 return rc;
 443         switch (nmi_watchdog) {
 444         case NMI_LOCAL_APIC:
 445                 rc |= lapic_wd_event(nmi_hz);
 446                 break;
 447         case NMI_IO_APIC:
 448                 /*
 449                  * don't know how to accurately check for this.
 450                  * just assume it was a watchdog timer interrupt
 451                  * This matches the old behaviour.
 452                  */
 453                 rc = 1;
 454                 break;
 455         }
 456         return rc;
 457 }
 458
 459 #ifdef CONFIG_SYSCTL
 460
 461 static int __init setup_unknown_nmi_panic(char *str)
 462 {
 463         unknown_nmi_panic = 1;
 464         return 1;
 465 }
 466 __setup("unknown_nmi_panic", setup_unknown_nmi_panic);
 467
 468 static int unknown_nmi_panic_callback(struct pt_regs *regs, int cpu)
 469 {
 470         unsigned char reason = get_nmi_reason();
 471         char buf[64];
 472
 473         sprintf(buf, "NMI received for unknown reason %02x\n", reason);
 474         die_nmi(buf, regs, 1); /* Always panic here */
 475         return 0;
 476 }
 477
 478 /*
 479  * proc handler for /proc/sys/kernel/nmi
 480  */
 481 int proc_nmi_enabled(struct ctl_table *table, int write, struct file *file,
 482                         void __user *buffer, size_t *length, loff_t *ppos)
 483 {
 484         int old_state;
 485
 486         nmi_watchdog_enabled = (atomic_read(&nmi_active) > 0) ? 1 : 0;
 487         old_state = nmi_watchdog_enabled;
 488         proc_dointvec(table, write, file, buffer, length, ppos);
 489         if (!!old_state == !!nmi_watchdog_enabled)
 490                 return 0;
 491
 492         if (atomic_read(&nmi_active) < 0 || !nmi_watchdog_active()) {
 493                 printk(KERN_WARNING
 494                         "NMI watchdog is permanently disabled\n");
 495                 return -EIO;
 496         }
 497
 498         if (nmi_watchdog == NMI_LOCAL_APIC) {
 499                 if (nmi_watchdog_enabled)
 500                         enable_lapic_nmi_watchdog();
 501                 else
 502                         disable_lapic_nmi_watchdog();
 503         } else {
 504                 printk(KERN_WARNING
 505                         "NMI watchdog doesn't know what hardware to touch\n");
 506                 return -EIO;
 507         }
 508         return 0;
 509 }
 510
 511 #endif /* CONFIG_SYSCTL */
 512
 513 int do_nmi_callback(struct pt_regs *regs, int cpu)
 514 {
 515 #ifdef CONFIG_SYSCTL
 516         if (unknown_nmi_panic)
 517                 return unknown_nmi_panic_callback(regs, cpu);
 518 #endif
 519         return 0;
 520 }
 521
 522 void __trigger_all_cpu_backtrace(void)
 523 {
 524         int i;
 525
 526         backtrace_mask = cpu_online_map;
 527         /* Wait for up to 10 seconds for all CPUs to do the backtrace */
 528         for (i = 0; i < 10 * 1000; i++) {
 529                 if (cpus_empty(backtrace_mask))
 530                         break;
 531                 mdelay(1);
 532         }
 533 }