git.oblomov.eu Git - linux-2.6/blob - 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 static void __acpi_nmi_disable(void *__unused)
 135 {
 136         apic_write(APIC_LVT0, APIC_DM_NMI | APIC_LVT_MASKED);
 137 }
 138
 139 int __init check_nmi_watchdog(void)
 140 {
 141         unsigned int *prev_nmi_count;
 142         int cpu;
 143
 144         if (!nmi_watchdog_active() || !atomic_read(&nmi_active))
 145                 return 0;
 146
 147         prev_nmi_count = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL);
 148         if (!prev_nmi_count)
 149                 goto error;
 150
 151         printk(KERN_INFO "Testing NMI watchdog ... ");
 152
 153 #ifdef CONFIG_SMP
 154         if (nmi_watchdog == NMI_LOCAL_APIC)
 155                 smp_call_function(nmi_cpu_busy, (void *)&endflag, 0);
 156 #endif
 157
 158         for_each_possible_cpu(cpu)
 159                 prev_nmi_count[cpu] = get_nmi_count(cpu);
 160         local_irq_enable();
 161         mdelay((20 * 1000) / nmi_hz); /* wait 20 ticks */
 162
 163         for_each_online_cpu(cpu) {
 164                 if (!per_cpu(wd_enabled, cpu))
 165                         continue;
 166                 if (get_nmi_count(cpu) - prev_nmi_count[cpu] <= 5)
 167                         report_broken_nmi(cpu, prev_nmi_count);
 168         }
 169         endflag = 1;
 170         if (!atomic_read(&nmi_active)) {
 171                 kfree(prev_nmi_count);
 172                 atomic_set(&nmi_active, -1);
 173                 goto error;
 174         }
 175         printk("OK.\n");
 176
 177         /*
 178          * now that we know it works we can reduce NMI frequency to
 179          * something more reasonable; makes a difference in some configs
 180          */
 181         if (nmi_watchdog == NMI_LOCAL_APIC)
 182                 nmi_hz = lapic_adjust_nmi_hz(1);
 183
 184         kfree(prev_nmi_count);
 185         return 0;
 186 error:
 187         if (nmi_watchdog == NMI_IO_APIC) {
 188                 if (!timer_through_8259)
 189                         disable_8259A_irq(0);
 190                 on_each_cpu(__acpi_nmi_disable, NULL, 1);
 191         }
 192
 193 #ifdef CONFIG_X86_32
 194         timer_ack = 0;
 195 #endif
 196         return -1;
 197 }
 198
 199 static int __init setup_nmi_watchdog(char *str)
 200 {
 201         unsigned int nmi;
 202
 203         if (!strncmp(str, "panic", 5)) {
 204                 panic_on_timeout = 1;
 205                 str = strchr(str, ',');
 206                 if (!str)
 207                         return 1;
 208                 ++str;
 209         }
 210
 211         if (!strncmp(str, "lapic", 5))
 212                 nmi_watchdog = NMI_LOCAL_APIC;
 213         else if (!strncmp(str, "ioapic", 6))
 214                 nmi_watchdog = NMI_IO_APIC;
 215         else {
 216                 get_option(&str, &nmi);
 217                 if (nmi >= NMI_INVALID)
 218                         return 0;
 219                 nmi_watchdog = nmi;
 220         }
 221
 222         return 1;
 223 }
 224 __setup("nmi_watchdog=", setup_nmi_watchdog);
 225
 226 /*
 227  * Suspend/resume support
 228  */
 229 #ifdef CONFIG_PM
 230
 231 static int nmi_pm_active; /* nmi_active before suspend */
 232
 233 static int lapic_nmi_suspend(struct sys_device *dev, pm_message_t state)
 234 {
 235         /* only CPU0 goes here, other CPUs should be offline */
 236         nmi_pm_active = atomic_read(&nmi_active);
 237         stop_apic_nmi_watchdog(NULL);
 238         BUG_ON(atomic_read(&nmi_active) != 0);
 239         return 0;
 240 }
 241
 242 static int lapic_nmi_resume(struct sys_device *dev)
 243 {
 244         /* only CPU0 goes here, other CPUs should be offline */
 245         if (nmi_pm_active > 0) {
 246                 setup_apic_nmi_watchdog(NULL);
 247                 touch_nmi_watchdog();
 248         }
 249         return 0;
 250 }
 251
 252 static struct sysdev_class nmi_sysclass = {
 253         .name           = "lapic_nmi",
 254         .resume         = lapic_nmi_resume,
 255         .suspend        = lapic_nmi_suspend,
 256 };
 257
 258 static struct sys_device device_lapic_nmi = {
 259         .id     = 0,
 260         .cls    = &nmi_sysclass,
 261 };
 262
 263 static int __init init_lapic_nmi_sysfs(void)
 264 {
 265         int error;
 266
 267         /*
 268          * should really be a BUG_ON but b/c this is an
 269          * init call, it just doesn't work.  -dcz
 270          */
 271         if (nmi_watchdog != NMI_LOCAL_APIC)
 272                 return 0;
 273
 274         if (atomic_read(&nmi_active) < 0)
 275                 return 0;
 276
 277         error = sysdev_class_register(&nmi_sysclass);
 278         if (!error)
 279                 error = sysdev_register(&device_lapic_nmi);
 280         return error;
 281 }
 282
 283 /* must come after the local APIC's device_initcall() */
 284 late_initcall(init_lapic_nmi_sysfs);
 285
 286 #endif  /* CONFIG_PM */
 287
 288 static void __acpi_nmi_enable(void *__unused)
 289 {
 290         apic_write(APIC_LVT0, APIC_DM_NMI);
 291 }
 292
 293 /*
 294  * Enable timer based NMIs on all CPUs:
 295  */
 296 void acpi_nmi_enable(void)
 297 {
 298         if (atomic_read(&nmi_active) && nmi_watchdog == NMI_IO_APIC)
 299                 on_each_cpu(__acpi_nmi_enable, NULL, 1);
 300 }
 301
 302 /*
 303  * Disable timer based NMIs on all CPUs:
 304  */
 305 void acpi_nmi_disable(void)
 306 {
 307         if (atomic_read(&nmi_active) && nmi_watchdog == NMI_IO_APIC)
 308                 on_each_cpu(__acpi_nmi_disable, NULL, 1);
 309 }
 310
 311 /*
 312  * This function is called as soon the LAPIC NMI watchdog driver has everything
 313  * in place and it's ready to check if the NMIs belong to the NMI watchdog
 314  */
 315 void cpu_nmi_set_wd_enabled(void)
 316 {
 317         __get_cpu_var(wd_enabled) = 1;
 318 }
 319
 320 void setup_apic_nmi_watchdog(void *unused)
 321 {
 322         if (__get_cpu_var(wd_enabled))
 323                 return;
 324
 325         /* cheap hack to support suspend/resume */
 326         /* if cpu0 is not active neither should the other cpus */
 327         if (smp_processor_id() != 0 && atomic_read(&nmi_active) <= 0)
 328                 return;
 329
 330         switch (nmi_watchdog) {
 331         case NMI_LOCAL_APIC:
 332                 if (lapic_watchdog_init(nmi_hz) < 0) {
 333                         __get_cpu_var(wd_enabled) = 0;
 334                         return;
 335                 }
 336                 /* FALL THROUGH */
 337         case NMI_IO_APIC:
 338                 __get_cpu_var(wd_enabled) = 1;
 339                 atomic_inc(&nmi_active);
 340         }
 341 }
 342
 343 void stop_apic_nmi_watchdog(void *unused)
 344 {
 345         /* only support LOCAL and IO APICs for now */
 346         if (!nmi_watchdog_active())
 347                 return;
 348         if (__get_cpu_var(wd_enabled) == 0)
 349                 return;
 350         if (nmi_watchdog == NMI_LOCAL_APIC)
 351                 lapic_watchdog_stop();
 352         else
 353                 __acpi_nmi_disable(NULL);
 354         __get_cpu_var(wd_enabled) = 0;
 355         atomic_dec(&nmi_active);
 356 }
 357
 358 /*
 359  * the best way to detect whether a CPU has a 'hard lockup' problem
 360  * is to check it's local APIC timer IRQ counts. If they are not
 361  * changing then that CPU has some problem.
 362  *
 363  * as these watchdog NMI IRQs are generated on every CPU, we only
 364  * have to check the current processor.
 365  *
 366  * since NMIs don't listen to _any_ locks, we have to be extremely
 367  * careful not to rely on unsafe variables. The printk might lock
 368  * up though, so we have to break up any console locks first ...
 369  * [when there will be more tty-related locks, break them up here too!]
 370  */
 371
 372 static DEFINE_PER_CPU(unsigned, last_irq_sum);
 373 static DEFINE_PER_CPU(local_t, alert_counter);
 374 static DEFINE_PER_CPU(int, nmi_touch);
 375
 376 void touch_nmi_watchdog(void)
 377 {
 378         if (nmi_watchdog_active()) {
 379                 unsigned cpu;
 380
 381                 /*
 382                  * Tell other CPUs to reset their alert counters. We cannot
 383                  * do it ourselves because the alert count increase is not
 384                  * atomic.
 385                  */
 386                 for_each_present_cpu(cpu) {
 387                         if (per_cpu(nmi_touch, cpu) != 1)
 388                                 per_cpu(nmi_touch, cpu) = 1;
 389                 }
 390         }
 391
 392         /*
 393          * Tickle the softlockup detector too:
 394          */
 395         touch_softlockup_watchdog();
 396 }
 397 EXPORT_SYMBOL(touch_nmi_watchdog);
 398
 399 notrace __kprobes int
 400 nmi_watchdog_tick(struct pt_regs *regs, unsigned reason)
 401 {
 402         /*
 403          * Since current_thread_info()-> is always on the stack, and we
 404          * always switch the stack NMI-atomically, it's safe to use
 405          * smp_processor_id().
 406          */
 407         unsigned int sum;
 408         int touched = 0;
 409         int cpu = smp_processor_id();
 410         int rc = 0;
 411
 412         /* check for other users first */
 413         if (notify_die(DIE_NMI, "nmi", regs, reason, 2, SIGINT)
 414                         == NOTIFY_STOP) {
 415                 rc = 1;
 416                 touched = 1;
 417         }
 418
 419         sum = get_timer_irqs(cpu);
 420
 421         if (__get_cpu_var(nmi_touch)) {
 422                 __get_cpu_var(nmi_touch) = 0;
 423                 touched = 1;
 424         }
 425
 426         if (cpu_isset(cpu, backtrace_mask)) {
 427                 static DEFINE_SPINLOCK(lock);   /* Serialise the printks */
 428
 429                 spin_lock(&lock);
 430                 printk(KERN_WARNING "NMI backtrace for cpu %d\n", cpu);
 431                 dump_stack();
 432                 spin_unlock(&lock);
 433                 cpu_clear(cpu, backtrace_mask);
 434         }
 435
 436         /* Could check oops_in_progress here too, but it's safer not to */
 437         if (mce_in_progress())
 438                 touched = 1;
 439
 440         /* if the none of the timers isn't firing, this cpu isn't doing much */
 441         if (!touched && __get_cpu_var(last_irq_sum) == sum) {
 442                 /*
 443                  * Ayiee, looks like this CPU is stuck ...
 444                  * wait a few IRQs (5 seconds) before doing the oops ...
 445                  */
 446                 local_inc(&__get_cpu_var(alert_counter));
 447                 if (local_read(&__get_cpu_var(alert_counter)) == 5 * nmi_hz)
 448                         /*
 449                          * die_nmi will return ONLY if NOTIFY_STOP happens..
 450                          */
 451                         die_nmi("BUG: NMI Watchdog detected LOCKUP",
 452                                 regs, panic_on_timeout);
 453         } else {
 454                 __get_cpu_var(last_irq_sum) = sum;
 455                 local_set(&__get_cpu_var(alert_counter), 0);
 456         }
 457
 458         /* see if the nmi watchdog went off */
 459         if (!__get_cpu_var(wd_enabled))
 460                 return rc;
 461         switch (nmi_watchdog) {
 462         case NMI_LOCAL_APIC:
 463                 rc |= lapic_wd_event(nmi_hz);
 464                 break;
 465         case NMI_IO_APIC:
 466                 /*
 467                  * don't know how to accurately check for this.
 468                  * just assume it was a watchdog timer interrupt
 469                  * This matches the old behaviour.
 470                  */
 471                 rc = 1;
 472                 break;
 473         }
 474         return rc;
 475 }
 476
 477 #ifdef CONFIG_SYSCTL
 478
 479 static void enable_ioapic_nmi_watchdog_single(void *unused)
 480 {
 481         __get_cpu_var(wd_enabled) = 1;
 482         atomic_inc(&nmi_active);
 483         __acpi_nmi_enable(NULL);
 484 }
 485
 486 static void enable_ioapic_nmi_watchdog(void)
 487 {
 488         on_each_cpu(enable_ioapic_nmi_watchdog_single, NULL, 1);
 489         touch_nmi_watchdog();
 490 }
 491
 492 static void disable_ioapic_nmi_watchdog(void)
 493 {
 494         on_each_cpu(stop_apic_nmi_watchdog, NULL, 1);
 495 }
 496
 497 static int __init setup_unknown_nmi_panic(char *str)
 498 {
 499         unknown_nmi_panic = 1;
 500         return 1;
 501 }
 502 __setup("unknown_nmi_panic", setup_unknown_nmi_panic);
 503
 504 static int unknown_nmi_panic_callback(struct pt_regs *regs, int cpu)
 505 {
 506         unsigned char reason = get_nmi_reason();
 507         char buf[64];
 508
 509         sprintf(buf, "NMI received for unknown reason %02x\n", reason);
 510         die_nmi(buf, regs, 1); /* Always panic here */
 511         return 0;
 512 }
 513
 514 /*
 515  * proc handler for /proc/sys/kernel/nmi
 516  */
 517 int proc_nmi_enabled(struct ctl_table *table, int write, struct file *file,
 518                         void __user *buffer, size_t *length, loff_t *ppos)
 519 {
 520         int old_state;
 521
 522         nmi_watchdog_enabled = (atomic_read(&nmi_active) > 0) ? 1 : 0;
 523         old_state = nmi_watchdog_enabled;
 524         proc_dointvec(table, write, file, buffer, length, ppos);
 525         if (!!old_state == !!nmi_watchdog_enabled)
 526                 return 0;
 527
 528         if (atomic_read(&nmi_active) < 0 || !nmi_watchdog_active()) {
 529                 printk(KERN_WARNING
 530                         "NMI watchdog is permanently disabled\n");
 531                 return -EIO;
 532         }
 533
 534         if (nmi_watchdog == NMI_LOCAL_APIC) {
 535                 if (nmi_watchdog_enabled)
 536                         enable_lapic_nmi_watchdog();
 537                 else
 538                         disable_lapic_nmi_watchdog();
 539         } else if (nmi_watchdog == NMI_IO_APIC) {
 540                 if (nmi_watchdog_enabled)
 541                         enable_ioapic_nmi_watchdog();
 542                 else
 543                         disable_ioapic_nmi_watchdog();
 544         } else {
 545                 printk(KERN_WARNING
 546                         "NMI watchdog doesn't know what hardware to touch\n");
 547                 return -EIO;
 548         }
 549         return 0;
 550 }
 551
 552 #endif /* CONFIG_SYSCTL */
 553
 554 int do_nmi_callback(struct pt_regs *regs, int cpu)
 555 {
 556 #ifdef CONFIG_SYSCTL
 557         if (unknown_nmi_panic)
 558                 return unknown_nmi_panic_callback(regs, cpu);
 559 #endif
 560         return 0;
 561 }
 562
 563 void __trigger_all_cpu_backtrace(void)
 564 {
 565         int i;
 566
 567         backtrace_mask = cpu_online_map;
 568         /* Wait for up to 10 seconds for all CPUs to do the backtrace */
 569         for (i = 0; i < 10 * 1000; i++) {
 570                 if (cpus_empty(backtrace_mask))
 571                         break;
 572                 mdelay(1);
 573         }
 574 }