arch/x86/kernel/apic/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/proto.h>
  33 #include <asm/timer.h>
  34
  35 #include <asm/mce.h>
  36
  37 #include <asm/mach_traps.h>
  38
  39 int unknown_nmi_panic;
  40 int nmi_watchdog_enabled;
  41
  42 static cpumask_var_t backtrace_mask;
  43
  44 /* nmi_active:
  45  * >0: the lapic NMI watchdog is active, but can be disabled
  46  * <0: the lapic NMI watchdog has not been set up, and cannot
  47  *     be enabled
  48  *  0: the lapic NMI watchdog is disabled, but can be enabled
  49  */
  50 atomic_t nmi_active = ATOMIC_INIT(0);           /* oprofile uses this */
  51 EXPORT_SYMBOL(nmi_active);
  52
  53 unsigned int nmi_watchdog = NMI_NONE;
  54 EXPORT_SYMBOL(nmi_watchdog);
  55
  56 static int panic_on_timeout;
  57
  58 static unsigned int nmi_hz = HZ;
  59 static DEFINE_PER_CPU(short, wd_enabled);
  60 static int endflag __initdata;
  61
  62 static inline unsigned int get_nmi_count(int cpu)
  63 {
  64         return per_cpu(irq_stat, cpu).__nmi_count;
  65 }
  66
  67 static inline int mce_in_progress(void)
  68 {
  69 #if defined(CONFIG_X86_NEW_MCE)
  70         return atomic_read(&mce_entry) > 0;
  71 #endif
  72         return 0;
  73 }
  74
  75 /*
  76  * Take the local apic timer and PIT/HPET into account. We don't
  77  * know which one is active, when we have highres/dyntick on
  78  */
  79 static inline unsigned int get_timer_irqs(int cpu)
  80 {
  81         return per_cpu(irq_stat, cpu).apic_timer_irqs +
  82                 per_cpu(irq_stat, cpu).irq0_irqs;
  83 }
  84
  85 #ifdef CONFIG_SMP
  86 /*
  87  * The performance counters used by NMI_LOCAL_APIC don't trigger when
  88  * the CPU is idle. To make sure the NMI watchdog really ticks on all
  89  * CPUs during the test make them busy.
  90  */
  91 static __init void nmi_cpu_busy(void *data)
  92 {
  93         local_irq_enable_in_hardirq();
  94         /*
  95          * Intentionally don't use cpu_relax here. This is
  96          * to make sure that the performance counter really ticks,
  97          * even if there is a simulator or similar that catches the
  98          * pause instruction. On a real HT machine this is fine because
  99          * all other CPUs are busy with "useless" delay loops and don't
 100          * care if they get somewhat less cycles.
 101          */
 102         while (endflag == 0)
 103                 mb();
 104 }
 105 #endif
 106
 107 static void report_broken_nmi(int cpu, unsigned int *prev_nmi_count)
 108 {
 109         printk(KERN_CONT "\n");
 110
 111         printk(KERN_WARNING
 112                 "WARNING: CPU#%d: NMI appears to be stuck (%d->%d)!\n",
 113                         cpu, prev_nmi_count[cpu], get_nmi_count(cpu));
 114
 115         printk(KERN_WARNING
 116                 "Please report this to bugzilla.kernel.org,\n");
 117         printk(KERN_WARNING
 118                 "and attach the output of the 'dmesg' command.\n");
 119
 120         per_cpu(wd_enabled, cpu) = 0;
 121         atomic_dec(&nmi_active);
 122 }
 123
 124 static void __acpi_nmi_disable(void *__unused)
 125 {
 126         apic_write(APIC_LVT0, APIC_DM_NMI | APIC_LVT_MASKED);
 127 }
 128
 129 int __init check_nmi_watchdog(void)
 130 {
 131         unsigned int *prev_nmi_count;
 132         int cpu;
 133
 134         if (!nmi_watchdog_active() || !atomic_read(&nmi_active))
 135                 return 0;
 136
 137         prev_nmi_count = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL);
 138         if (!prev_nmi_count)
 139                 goto error;
 140
 141         alloc_cpumask_var(&backtrace_mask, GFP_KERNEL|__GFP_ZERO);
 142         printk(KERN_INFO "Testing NMI watchdog ... ");
 143
 144 #ifdef CONFIG_SMP
 145         if (nmi_watchdog == NMI_LOCAL_APIC)
 146                 smp_call_function(nmi_cpu_busy, (void *)&endflag, 0);
 147 #endif
 148
 149         for_each_possible_cpu(cpu)
 150                 prev_nmi_count[cpu] = get_nmi_count(cpu);
 151         local_irq_enable();
 152         mdelay((20 * 1000) / nmi_hz); /* wait 20 ticks */
 153
 154         for_each_online_cpu(cpu) {
 155                 if (!per_cpu(wd_enabled, cpu))
 156                         continue;
 157                 if (get_nmi_count(cpu) - prev_nmi_count[cpu] <= 5)
 158                         report_broken_nmi(cpu, prev_nmi_count);
 159         }
 160         endflag = 1;
 161         if (!atomic_read(&nmi_active)) {
 162                 kfree(prev_nmi_count);
 163                 atomic_set(&nmi_active, -1);
 164                 goto error;
 165         }
 166         printk("OK.\n");
 167
 168         /*
 169          * now that we know it works we can reduce NMI frequency to
 170          * something more reasonable; makes a difference in some configs
 171          */
 172         if (nmi_watchdog == NMI_LOCAL_APIC)
 173                 nmi_hz = lapic_adjust_nmi_hz(1);
 174
 175         kfree(prev_nmi_count);
 176         return 0;
 177 error:
 178         if (nmi_watchdog == NMI_IO_APIC) {
 179                 if (!timer_through_8259)
 180                         disable_8259A_irq(0);
 181                 on_each_cpu(__acpi_nmi_disable, NULL, 1);
 182         }
 183
 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 /*
 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 /*
 303  * This function is called as soon the LAPIC NMI watchdog driver has everything
 304  * in place and it's ready to check if the NMIs belong to the NMI watchdog
 305  */
 306 void cpu_nmi_set_wd_enabled(void)
 307 {
 308         __get_cpu_var(wd_enabled) = 1;
 309 }
 310
 311 void setup_apic_nmi_watchdog(void *unused)
 312 {
 313         if (__get_cpu_var(wd_enabled))
 314                 return;
 315
 316         /* cheap hack to support suspend/resume */
 317         /* if cpu0 is not active neither should the other cpus */
 318         if (smp_processor_id() != 0 && atomic_read(&nmi_active) <= 0)
 319                 return;
 320
 321         switch (nmi_watchdog) {
 322         case NMI_LOCAL_APIC:
 323                 if (lapic_watchdog_init(nmi_hz) < 0) {
 324                         __get_cpu_var(wd_enabled) = 0;
 325                         return;
 326                 }
 327                 /* FALL THROUGH */
 328         case NMI_IO_APIC:
 329                 __get_cpu_var(wd_enabled) = 1;
 330                 atomic_inc(&nmi_active);
 331         }
 332 }
 333
 334 void stop_apic_nmi_watchdog(void *unused)
 335 {
 336         /* only support LOCAL and IO APICs for now */
 337         if (!nmi_watchdog_active())
 338                 return;
 339         if (__get_cpu_var(wd_enabled) == 0)
 340                 return;
 341         if (nmi_watchdog == NMI_LOCAL_APIC)
 342                 lapic_watchdog_stop();
 343         else
 344                 __acpi_nmi_disable(NULL);
 345         __get_cpu_var(wd_enabled) = 0;
 346         atomic_dec(&nmi_active);
 347 }
 348
 349 /*
 350  * the best way to detect whether a CPU has a 'hard lockup' problem
 351  * is to check it's local APIC timer IRQ counts. If they are not
 352  * changing then that CPU has some problem.
 353  *
 354  * as these watchdog NMI IRQs are generated on every CPU, we only
 355  * have to check the current processor.
 356  *
 357  * since NMIs don't listen to _any_ locks, we have to be extremely
 358  * careful not to rely on unsafe variables. The printk might lock
 359  * up though, so we have to break up any console locks first ...
 360  * [when there will be more tty-related locks, break them up here too!]
 361  */
 362
 363 static DEFINE_PER_CPU(unsigned, last_irq_sum);
 364 static DEFINE_PER_CPU(local_t, alert_counter);
 365 static DEFINE_PER_CPU(int, nmi_touch);
 366
 367 void touch_nmi_watchdog(void)
 368 {
 369         if (nmi_watchdog_active()) {
 370                 unsigned cpu;
 371
 372                 /*
 373                  * Tell other CPUs to reset their alert counters. We cannot
 374                  * do it ourselves because the alert count increase is not
 375                  * atomic.
 376                  */
 377                 for_each_present_cpu(cpu) {
 378                         if (per_cpu(nmi_touch, cpu) != 1)
 379                                 per_cpu(nmi_touch, cpu) = 1;
 380                 }
 381         }
 382
 383         /*
 384          * Tickle the softlockup detector too:
 385          */
 386         touch_softlockup_watchdog();
 387 }
 388 EXPORT_SYMBOL(touch_nmi_watchdog);
 389
 390 notrace __kprobes int
 391 nmi_watchdog_tick(struct pt_regs *regs, unsigned reason)
 392 {
 393         /*
 394          * Since current_thread_info()-> is always on the stack, and we
 395          * always switch the stack NMI-atomically, it's safe to use
 396          * smp_processor_id().
 397          */
 398         unsigned int sum;
 399         int touched = 0;
 400         int cpu = smp_processor_id();
 401         int rc = 0;
 402
 403         /* check for other users first */
 404         if (notify_die(DIE_NMI, "nmi", regs, reason, 2, SIGINT)
 405                         == NOTIFY_STOP) {
 406                 rc = 1;
 407                 touched = 1;
 408         }
 409
 410         sum = get_timer_irqs(cpu);
 411
 412         if (__get_cpu_var(nmi_touch)) {
 413                 __get_cpu_var(nmi_touch) = 0;
 414                 touched = 1;
 415         }
 416
 417         /* We can be called before check_nmi_watchdog, hence NULL check. */
 418         if (backtrace_mask != NULL && cpumask_test_cpu(cpu, backtrace_mask)) {
 419                 static DEFINE_SPINLOCK(lock);   /* Serialise the printks */
 420
 421                 spin_lock(&lock);
 422                 printk(KERN_WARNING "NMI backtrace for cpu %d\n", cpu);
 423                 dump_stack();
 424                 spin_unlock(&lock);
 425                 cpumask_clear_cpu(cpu, backtrace_mask);
 426         }
 427
 428         /* Could check oops_in_progress here too, but it's safer not to */
 429         if (mce_in_progress())
 430                 touched = 1;
 431
 432         /* if the none of the timers isn't firing, this cpu isn't doing much */
 433         if (!touched && __get_cpu_var(last_irq_sum) == sum) {
 434                 /*
 435                  * Ayiee, looks like this CPU is stuck ...
 436                  * wait a few IRQs (5 seconds) before doing the oops ...
 437                  */
 438                 local_inc(&__get_cpu_var(alert_counter));
 439                 if (local_read(&__get_cpu_var(alert_counter)) == 5 * nmi_hz)
 440                         /*
 441                          * die_nmi will return ONLY if NOTIFY_STOP happens..
 442                          */
 443                         die_nmi("BUG: NMI Watchdog detected LOCKUP",
 444                                 regs, panic_on_timeout);
 445         } else {
 446                 __get_cpu_var(last_irq_sum) = sum;
 447                 local_set(&__get_cpu_var(alert_counter), 0);
 448         }
 449
 450         /* see if the nmi watchdog went off */
 451         if (!__get_cpu_var(wd_enabled))
 452                 return rc;
 453         switch (nmi_watchdog) {
 454         case NMI_LOCAL_APIC:
 455                 rc |= lapic_wd_event(nmi_hz);
 456                 break;
 457         case NMI_IO_APIC:
 458                 /*
 459                  * don't know how to accurately check for this.
 460                  * just assume it was a watchdog timer interrupt
 461                  * This matches the old behaviour.
 462                  */
 463                 rc = 1;
 464                 break;
 465         }
 466         return rc;
 467 }
 468
 469 #ifdef CONFIG_SYSCTL
 470
 471 static void enable_ioapic_nmi_watchdog_single(void *unused)
 472 {
 473         __get_cpu_var(wd_enabled) = 1;
 474         atomic_inc(&nmi_active);
 475         __acpi_nmi_enable(NULL);
 476 }
 477
 478 static void enable_ioapic_nmi_watchdog(void)
 479 {
 480         on_each_cpu(enable_ioapic_nmi_watchdog_single, NULL, 1);
 481         touch_nmi_watchdog();
 482 }
 483
 484 static void disable_ioapic_nmi_watchdog(void)
 485 {
 486         on_each_cpu(stop_apic_nmi_watchdog, NULL, 1);
 487 }
 488
 489 static int __init setup_unknown_nmi_panic(char *str)
 490 {
 491         unknown_nmi_panic = 1;
 492         return 1;
 493 }
 494 __setup("unknown_nmi_panic", setup_unknown_nmi_panic);
 495
 496 static int unknown_nmi_panic_callback(struct pt_regs *regs, int cpu)
 497 {
 498         unsigned char reason = get_nmi_reason();
 499         char buf[64];
 500
 501         sprintf(buf, "NMI received for unknown reason %02x\n", reason);
 502         die_nmi(buf, regs, 1); /* Always panic here */
 503         return 0;
 504 }
 505
 506 /*
 507  * proc handler for /proc/sys/kernel/nmi
 508  */
 509 int proc_nmi_enabled(struct ctl_table *table, int write, struct file *file,
 510                         void __user *buffer, size_t *length, loff_t *ppos)
 511 {
 512         int old_state;
 513
 514         nmi_watchdog_enabled = (atomic_read(&nmi_active) > 0) ? 1 : 0;
 515         old_state = nmi_watchdog_enabled;
 516         proc_dointvec(table, write, file, buffer, length, ppos);
 517         if (!!old_state == !!nmi_watchdog_enabled)
 518                 return 0;
 519
 520         if (atomic_read(&nmi_active) < 0 || !nmi_watchdog_active()) {
 521                 printk(KERN_WARNING
 522                         "NMI watchdog is permanently disabled\n");
 523                 return -EIO;
 524         }
 525
 526         if (nmi_watchdog == NMI_LOCAL_APIC) {
 527                 if (nmi_watchdog_enabled)
 528                         enable_lapic_nmi_watchdog();
 529                 else
 530                         disable_lapic_nmi_watchdog();
 531         } else if (nmi_watchdog == NMI_IO_APIC) {
 532                 if (nmi_watchdog_enabled)
 533                         enable_ioapic_nmi_watchdog();
 534                 else
 535                         disable_ioapic_nmi_watchdog();
 536         } else {
 537                 printk(KERN_WARNING
 538                         "NMI watchdog doesn't know what hardware to touch\n");
 539                 return -EIO;
 540         }
 541         return 0;
 542 }
 543
 544 #endif /* CONFIG_SYSCTL */
 545
 546 int do_nmi_callback(struct pt_regs *regs, int cpu)
 547 {
 548 #ifdef CONFIG_SYSCTL
 549         if (unknown_nmi_panic)
 550                 return unknown_nmi_panic_callback(regs, cpu);
 551 #endif
 552         return 0;
 553 }
 554
 555 void __trigger_all_cpu_backtrace(void)
 556 {
 557         int i;
 558
 559         cpumask_copy(backtrace_mask, cpu_online_mask);
 560         /* Wait for up to 10 seconds for all CPUs to do the backtrace */
 561         for (i = 0; i < 10 * 1000; i++) {
 562                 if (cpumask_empty(backtrace_mask))
 563                         break;
 564                 mdelay(1);
 565         }
 566 }