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         if (!strncmp(str, "lapic", 5))
 203                 nmi_watchdog = NMI_LOCAL_APIC;
 204         else if (!strncmp(str, "ioapic", 6))
 205                 nmi_watchdog = NMI_IO_APIC;
 206         else {
 207                 get_option(&str, &nmi);
 208                 if (nmi >= NMI_INVALID)
 209                         return 0;
 210                 nmi_watchdog = nmi;
 211         }
 212
 213         return 1;
 214 }
 215 __setup("nmi_watchdog=", setup_nmi_watchdog);
 216
 217 /*
 218  * Suspend/resume support
 219  */
 220 #ifdef CONFIG_PM
 221
 222 static int nmi_pm_active; /* nmi_active before suspend */
 223
 224 static int lapic_nmi_suspend(struct sys_device *dev, pm_message_t state)
 225 {
 226         /* only CPU0 goes here, other CPUs should be offline */
 227         nmi_pm_active = atomic_read(&nmi_active);
 228         stop_apic_nmi_watchdog(NULL);
 229         BUG_ON(atomic_read(&nmi_active) != 0);
 230         return 0;
 231 }
 232
 233 static int lapic_nmi_resume(struct sys_device *dev)
 234 {
 235         /* only CPU0 goes here, other CPUs should be offline */
 236         if (nmi_pm_active > 0) {
 237                 setup_apic_nmi_watchdog(NULL);
 238                 touch_nmi_watchdog();
 239         }
 240         return 0;
 241 }
 242
 243 static struct sysdev_class nmi_sysclass = {
 244         .name           = "lapic_nmi",
 245         .resume         = lapic_nmi_resume,
 246         .suspend        = lapic_nmi_suspend,
 247 };
 248
 249 static struct sys_device device_lapic_nmi = {
 250         .id     = 0,
 251         .cls    = &nmi_sysclass,
 252 };
 253
 254 static int __init init_lapic_nmi_sysfs(void)
 255 {
 256         int error;
 257
 258         /*
 259          * should really be a BUG_ON but b/c this is an
 260          * init call, it just doesn't work.  -dcz
 261          */
 262         if (nmi_watchdog != NMI_LOCAL_APIC)
 263                 return 0;
 264
 265         if (atomic_read(&nmi_active) < 0)
 266                 return 0;
 267
 268         error = sysdev_class_register(&nmi_sysclass);
 269         if (!error)
 270                 error = sysdev_register(&device_lapic_nmi);
 271         return error;
 272 }
 273
 274 /* must come after the local APIC's device_initcall() */
 275 late_initcall(init_lapic_nmi_sysfs);
 276
 277 #endif  /* CONFIG_PM */
 278
 279 static void __acpi_nmi_enable(void *__unused)
 280 {
 281         apic_write(APIC_LVT0, APIC_DM_NMI);
 282 }
 283
 284 /*
 285  * Enable timer based NMIs on all CPUs:
 286  */
 287 void acpi_nmi_enable(void)
 288 {
 289         if (atomic_read(&nmi_active) && nmi_watchdog == NMI_IO_APIC)
 290                 on_each_cpu(__acpi_nmi_enable, NULL, 1);
 291 }
 292
 293 static void __acpi_nmi_disable(void *__unused)
 294 {
 295         apic_write(APIC_LVT0, APIC_DM_NMI | APIC_LVT_MASKED);
 296 }
 297
 298 /*
 299  * Disable timer based NMIs on all CPUs:
 300  */
 301 void acpi_nmi_disable(void)
 302 {
 303         if (atomic_read(&nmi_active) && nmi_watchdog == NMI_IO_APIC)
 304                 on_each_cpu(__acpi_nmi_disable, NULL, 1);
 305 }
 306
 307 /*
 308  * This function is called as soon the LAPIC NMI watchdog driver has everything
 309  * in place and it's ready to check if the NMIs belong to the NMI watchdog
 310  */
 311 void cpu_nmi_set_wd_enabled(void)
 312 {
 313         __get_cpu_var(wd_enabled) = 1;
 314 }
 315
 316 void setup_apic_nmi_watchdog(void *unused)
 317 {
 318         if (__get_cpu_var(wd_enabled))
 319                 return;
 320
 321         /* cheap hack to support suspend/resume */
 322         /* if cpu0 is not active neither should the other cpus */
 323         if (smp_processor_id() != 0 && atomic_read(&nmi_active) <= 0)
 324                 return;
 325
 326         switch (nmi_watchdog) {
 327         case NMI_LOCAL_APIC:
 328                 if (lapic_watchdog_init(nmi_hz) < 0) {
 329                         __get_cpu_var(wd_enabled) = 0;
 330                         return;
 331                 }
 332                 /* FALL THROUGH */
 333         case NMI_IO_APIC:
 334                 __get_cpu_var(wd_enabled) = 1;
 335                 atomic_inc(&nmi_active);
 336         }
 337 }
 338
 339 void stop_apic_nmi_watchdog(void *unused)
 340 {
 341         /* only support LOCAL and IO APICs for now */
 342         if (!nmi_watchdog_active())
 343                 return;
 344         if (__get_cpu_var(wd_enabled) == 0)
 345                 return;
 346         if (nmi_watchdog == NMI_LOCAL_APIC)
 347                 lapic_watchdog_stop();
 348         __get_cpu_var(wd_enabled) = 0;
 349         atomic_dec(&nmi_active);
 350 }
 351
 352 /*
 353  * the best way to detect whether a CPU has a 'hard lockup' problem
 354  * is to check it's local APIC timer IRQ counts. If they are not
 355  * changing then that CPU has some problem.
 356  *
 357  * as these watchdog NMI IRQs are generated on every CPU, we only
 358  * have to check the current processor.
 359  *
 360  * since NMIs don't listen to _any_ locks, we have to be extremely
 361  * careful not to rely on unsafe variables. The printk might lock
 362  * up though, so we have to break up any console locks first ...
 363  * [when there will be more tty-related locks, break them up here too!]
 364  */
 365
 366 static DEFINE_PER_CPU(unsigned, last_irq_sum);
 367 static DEFINE_PER_CPU(local_t, alert_counter);
 368 static DEFINE_PER_CPU(int, nmi_touch);
 369
 370 void touch_nmi_watchdog(void)
 371 {
 372         if (nmi_watchdog_active()) {
 373                 unsigned cpu;
 374
 375                 /*
 376                  * Tell other CPUs to reset their alert counters. We cannot
 377                  * do it ourselves because the alert count increase is not
 378                  * atomic.
 379                  */
 380                 for_each_present_cpu(cpu) {
 381                         if (per_cpu(nmi_touch, cpu) != 1)
 382                                 per_cpu(nmi_touch, cpu) = 1;
 383                 }
 384         }
 385
 386         /*
 387          * Tickle the softlockup detector too:
 388          */
 389         touch_softlockup_watchdog();
 390 }
 391 EXPORT_SYMBOL(touch_nmi_watchdog);
 392
 393 notrace __kprobes int
 394 nmi_watchdog_tick(struct pt_regs *regs, unsigned reason)
 395 {
 396         /*
 397          * Since current_thread_info()-> is always on the stack, and we
 398          * always switch the stack NMI-atomically, it's safe to use
 399          * smp_processor_id().
 400          */
 401         unsigned int sum;
 402         int touched = 0;
 403         int cpu = smp_processor_id();
 404         int rc = 0;
 405
 406         /* check for other users first */
 407         if (notify_die(DIE_NMI, "nmi", regs, reason, 2, SIGINT)
 408                         == NOTIFY_STOP) {
 409                 rc = 1;
 410                 touched = 1;
 411         }
 412
 413         sum = get_timer_irqs(cpu);
 414
 415         if (__get_cpu_var(nmi_touch)) {
 416                 __get_cpu_var(nmi_touch) = 0;
 417                 touched = 1;
 418         }
 419
 420         if (cpu_isset(cpu, backtrace_mask)) {
 421                 static DEFINE_SPINLOCK(lock);   /* Serialise the printks */
 422
 423                 spin_lock(&lock);
 424                 printk(KERN_WARNING "NMI backtrace for cpu %d\n", cpu);
 425                 dump_stack();
 426                 spin_unlock(&lock);
 427                 cpu_clear(cpu, backtrace_mask);
 428         }
 429
 430         /* Could check oops_in_progress here too, but it's safer not to */
 431         if (mce_in_progress())
 432                 touched = 1;
 433
 434         /* if the none of the timers isn't firing, this cpu isn't doing much */
 435         if (!touched && __get_cpu_var(last_irq_sum) == sum) {
 436                 /*
 437                  * Ayiee, looks like this CPU is stuck ...
 438                  * wait a few IRQs (5 seconds) before doing the oops ...
 439                  */
 440                 local_inc(&__get_cpu_var(alert_counter));
 441                 if (local_read(&__get_cpu_var(alert_counter)) == 5 * nmi_hz)
 442                         /*
 443                          * die_nmi will return ONLY if NOTIFY_STOP happens..
 444                          */
 445                         die_nmi("BUG: NMI Watchdog detected LOCKUP",
 446                                 regs, panic_on_timeout);
 447         } else {
 448                 __get_cpu_var(last_irq_sum) = sum;
 449                 local_set(&__get_cpu_var(alert_counter), 0);
 450         }
 451
 452         /* see if the nmi watchdog went off */
 453         if (!__get_cpu_var(wd_enabled))
 454                 return rc;
 455         switch (nmi_watchdog) {
 456         case NMI_LOCAL_APIC:
 457                 rc |= lapic_wd_event(nmi_hz);
 458                 break;
 459         case NMI_IO_APIC:
 460                 /*
 461                  * don't know how to accurately check for this.
 462                  * just assume it was a watchdog timer interrupt
 463                  * This matches the old behaviour.
 464                  */
 465                 rc = 1;
 466                 break;
 467         }
 468         return rc;
 469 }
 470
 471 #ifdef CONFIG_SYSCTL
 472
 473 static int __init setup_unknown_nmi_panic(char *str)
 474 {
 475         unknown_nmi_panic = 1;
 476         return 1;
 477 }
 478 __setup("unknown_nmi_panic", setup_unknown_nmi_panic);
 479
 480 static int unknown_nmi_panic_callback(struct pt_regs *regs, int cpu)
 481 {
 482         unsigned char reason = get_nmi_reason();
 483         char buf[64];
 484
 485         sprintf(buf, "NMI received for unknown reason %02x\n", reason);
 486         die_nmi(buf, regs, 1); /* Always panic here */
 487         return 0;
 488 }
 489
 490 /*
 491  * proc handler for /proc/sys/kernel/nmi
 492  */
 493 int proc_nmi_enabled(struct ctl_table *table, int write, struct file *file,
 494                         void __user *buffer, size_t *length, loff_t *ppos)
 495 {
 496         int old_state;
 497
 498         nmi_watchdog_enabled = (atomic_read(&nmi_active) > 0) ? 1 : 0;
 499         old_state = nmi_watchdog_enabled;
 500         proc_dointvec(table, write, file, buffer, length, ppos);
 501         if (!!old_state == !!nmi_watchdog_enabled)
 502                 return 0;
 503
 504         if (atomic_read(&nmi_active) < 0 || !nmi_watchdog_active()) {
 505                 printk(KERN_WARNING
 506                         "NMI watchdog is permanently disabled\n");
 507                 return -EIO;
 508         }
 509
 510         if (nmi_watchdog == NMI_LOCAL_APIC) {
 511                 if (nmi_watchdog_enabled)
 512                         enable_lapic_nmi_watchdog();
 513                 else
 514                         disable_lapic_nmi_watchdog();
 515         } else {
 516                 printk(KERN_WARNING
 517                         "NMI watchdog doesn't know what hardware to touch\n");
 518                 return -EIO;
 519         }
 520         return 0;
 521 }
 522
 523 #endif /* CONFIG_SYSCTL */
 524
 525 int do_nmi_callback(struct pt_regs *regs, int cpu)
 526 {
 527 #ifdef CONFIG_SYSCTL
 528         if (unknown_nmi_panic)
 529                 return unknown_nmi_panic_callback(regs, cpu);
 530 #endif
 531         return 0;
 532 }
 533
 534 void __trigger_all_cpu_backtrace(void)
 535 {
 536         int i;
 537
 538         backtrace_mask = cpu_online_map;
 539         /* Wait for up to 10 seconds for all CPUs to do the backtrace */
 540         for (i = 0; i < 10 * 1000; i++) {
 541                 if (cpus_empty(backtrace_mask))
 542                         break;
 543                 mdelay(1);
 544         }
 545 }