SAFE public projects git trees. - safe/jmp/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/proto.h>
  33 #include <asm/timer.h>
  34
  35 #include <asm/mce.h>
  36
  37 #include <mach_traps.h>
  38
  39 int unknown_nmi_panic;
  40 int nmi_watchdog_enabled;
  41
  42 static cpumask_t backtrace_mask = CPU_MASK_NONE;
  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 #ifdef CONFIG_X86_64
  65         return cpu_pda(cpu)->__nmi_count;
  66 #else
  67         return nmi_count(cpu);
  68 #endif
  69 }
  70
  71 static inline int mce_in_progress(void)
  72 {
  73 #if defined(CONFIG_X86_64) && defined(CONFIG_X86_MCE)
  74         return atomic_read(&mce_entry) > 0;
  75 #endif
  76         return 0;
  77 }
  78
  79 /*
  80  * Take the local apic timer and PIT/HPET into account. We don't
  81  * know which one is active, when we have highres/dyntick on
  82  */
  83 static inline unsigned int get_timer_irqs(int cpu)
  84 {
  85 #ifdef CONFIG_X86_64
  86         return read_pda(apic_timer_irqs) + read_pda(irq0_irqs);
  87 #else
  88         return per_cpu(irq_stat, cpu).apic_timer_irqs +
  89                 per_cpu(irq_stat, cpu).irq0_irqs;
  90 #endif
  91 }
  92
  93 #ifdef CONFIG_SMP
  94 /*
  95  * The performance counters used by NMI_LOCAL_APIC don't trigger when
  96  * the CPU is idle. To make sure the NMI watchdog really ticks on all
  97  * CPUs during the test make them busy.
  98  */
  99 static __init void nmi_cpu_busy(void *data)
 100 {
 101         local_irq_enable_in_hardirq();
 102         /*
 103          * Intentionally don't use cpu_relax here. This is
 104          * to make sure that the performance counter really ticks,
 105          * even if there is a simulator or similar that catches the
 106          * pause instruction. On a real HT machine this is fine because
 107          * all other CPUs are busy with "useless" delay loops and don't
 108          * care if they get somewhat less cycles.
 109          */
 110         while (endflag == 0)
 111                 mb();
 112 }
 113 #endif
 114
 115 static void report_broken_nmi(int cpu, int *prev_nmi_count)
 116 {
 117         printk(KERN_CONT "\n");
 118
 119         printk(KERN_WARNING
 120                 "WARNING: CPU#%d: NMI appears to be stuck (%d->%d)!\n",
 121                         cpu, prev_nmi_count[cpu], get_nmi_count(cpu));
 122
 123         printk(KERN_WARNING
 124                 "Please report this to bugzilla.kernel.org,\n");
 125         printk(KERN_WARNING
 126                 "and attach the output of the 'dmesg' command.\n");
 127
 128         per_cpu(wd_enabled, cpu) = 0;
 129         atomic_dec(&nmi_active);
 130 }
 131
 132 static void __acpi_nmi_disable(void *__unused)
 133 {
 134         apic_write(APIC_LVT0, APIC_DM_NMI | APIC_LVT_MASKED);
 135 }
 136
 137 int __init check_nmi_watchdog(void)
 138 {
 139         unsigned int *prev_nmi_count;
 140         int cpu;
 141
 142         if (!nmi_watchdog_active() || !atomic_read(&nmi_active))
 143                 return 0;
 144
 145         prev_nmi_count = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL);
 146         if (!prev_nmi_count)
 147                 goto error;
 148
 149         printk(KERN_INFO "Testing NMI watchdog ... ");
 150
 151 #ifdef CONFIG_SMP
 152         if (nmi_watchdog == NMI_LOCAL_APIC)
 153                 smp_call_function(nmi_cpu_busy, (void *)&endflag, 0);
 154 #endif
 155
 156         for_each_possible_cpu(cpu)
 157                 prev_nmi_count[cpu] = get_nmi_count(cpu);
 158         local_irq_enable();
 159         mdelay((20 * 1000) / nmi_hz); /* wait 20 ticks */
 160
 161         for_each_online_cpu(cpu) {
 162                 if (!per_cpu(wd_enabled, cpu))
 163                         continue;
 164                 if (get_nmi_count(cpu) - prev_nmi_count[cpu] <= 5)
 165                         report_broken_nmi(cpu, prev_nmi_count);
 166         }
 167         endflag = 1;
 168         if (!atomic_read(&nmi_active)) {
 169                 kfree(prev_nmi_count);
 170                 atomic_set(&nmi_active, -1);
 171                 goto error;
 172         }
 173         printk("OK.\n");
 174
 175         /*
 176          * now that we know it works we can reduce NMI frequency to
 177          * something more reasonable; makes a difference in some configs
 178          */
 179         if (nmi_watchdog == NMI_LOCAL_APIC)
 180                 nmi_hz = lapic_adjust_nmi_hz(1);
 181
 182         kfree(prev_nmi_count);
 183         return 0;
 184 error:
 185         if (nmi_watchdog == NMI_IO_APIC) {
 186                 if (!timer_through_8259)
 187                         disable_8259A_irq(0);
 188                 on_each_cpu(__acpi_nmi_disable, NULL, 1);
 189         }
 190
 191 #ifdef CONFIG_X86_32
 192         timer_ack = 0;
 193 #endif
 194         return -1;
 195 }
 196
 197 static int __init setup_nmi_watchdog(char *str)
 198 {
 199         unsigned int nmi;
 200
 201         if (!strncmp(str, "panic", 5)) {
 202                 panic_on_timeout = 1;
 203                 str = strchr(str, ',');
 204                 if (!str)
 205                         return 1;
 206                 ++str;
 207         }
 208
 209         if (!strncmp(str, "lapic", 5))
 210                 nmi_watchdog = NMI_LOCAL_APIC;
 211         else if (!strncmp(str, "ioapic", 6))
 212                 nmi_watchdog = NMI_IO_APIC;
 213         else {
 214                 get_option(&str, &nmi);
 215                 if (nmi >= NMI_INVALID)
 216                         return 0;
 217                 nmi_watchdog = nmi;
 218         }
 219
 220         return 1;
 221 }
 222 __setup("nmi_watchdog=", setup_nmi_watchdog);
 223
 224 /*
 225  * Suspend/resume support
 226  */
 227 #ifdef CONFIG_PM
 228
 229 static int nmi_pm_active; /* nmi_active before suspend */
 230
 231 static int lapic_nmi_suspend(struct sys_device *dev, pm_message_t state)
 232 {
 233         /* only CPU0 goes here, other CPUs should be offline */
 234         nmi_pm_active = atomic_read(&nmi_active);
 235         stop_apic_nmi_watchdog(NULL);
 236         BUG_ON(atomic_read(&nmi_active) != 0);
 237         return 0;
 238 }
 239
 240 static int lapic_nmi_resume(struct sys_device *dev)
 241 {
 242         /* only CPU0 goes here, other CPUs should be offline */
 243         if (nmi_pm_active > 0) {
 244                 setup_apic_nmi_watchdog(NULL);
 245                 touch_nmi_watchdog();
 246         }
 247         return 0;
 248 }
 249
 250 static struct sysdev_class nmi_sysclass = {
 251         .name           = "lapic_nmi",
 252         .resume         = lapic_nmi_resume,
 253         .suspend        = lapic_nmi_suspend,
 254 };
 255
 256 static struct sys_device device_lapic_nmi = {
 257         .id     = 0,
 258         .cls    = &nmi_sysclass,
 259 };
 260
 261 static int __init init_lapic_nmi_sysfs(void)
 262 {
 263         int error;
 264
 265         /*
 266          * should really be a BUG_ON but b/c this is an
 267          * init call, it just doesn't work.  -dcz
 268          */
 269         if (nmi_watchdog != NMI_LOCAL_APIC)
 270                 return 0;
 271
 272         if (atomic_read(&nmi_active) < 0)
 273                 return 0;
 274
 275         error = sysdev_class_register(&nmi_sysclass);
 276         if (!error)
 277                 error = sysdev_register(&device_lapic_nmi);
 278         return error;
 279 }
 280
 281 /* must come after the local APIC's device_initcall() */
 282 late_initcall(init_lapic_nmi_sysfs);
 283
 284 #endif  /* CONFIG_PM */
 285
 286 static void __acpi_nmi_enable(void *__unused)
 287 {
 288         apic_write(APIC_LVT0, APIC_DM_NMI);
 289 }
 290
 291 /*
 292  * Enable timer based NMIs on all CPUs:
 293  */
 294 void acpi_nmi_enable(void)
 295 {
 296         if (atomic_read(&nmi_active) && nmi_watchdog == NMI_IO_APIC)
 297                 on_each_cpu(__acpi_nmi_enable, NULL, 1);
 298 }
 299
 300 /*
 301  * Disable timer based NMIs on all CPUs:
 302  */
 303 void acpi_nmi_disable(void)
 304 {
 305         if (atomic_read(&nmi_active) && nmi_watchdog == NMI_IO_APIC)
 306                 on_each_cpu(__acpi_nmi_disable, NULL, 1);
 307 }
 308
 309 /*
 310  * This function is called as soon the LAPIC NMI watchdog driver has everything
 311  * in place and it's ready to check if the NMIs belong to the NMI watchdog
 312  */
 313 void cpu_nmi_set_wd_enabled(void)
 314 {
 315         __get_cpu_var(wd_enabled) = 1;
 316 }
 317
 318 void setup_apic_nmi_watchdog(void *unused)
 319 {
 320         if (__get_cpu_var(wd_enabled))
 321                 return;
 322
 323         /* cheap hack to support suspend/resume */
 324         /* if cpu0 is not active neither should the other cpus */
 325         if (smp_processor_id() != 0 && atomic_read(&nmi_active) <= 0)
 326                 return;
 327
 328         switch (nmi_watchdog) {
 329         case NMI_LOCAL_APIC:
 330                 if (lapic_watchdog_init(nmi_hz) < 0) {
 331                         __get_cpu_var(wd_enabled) = 0;
 332                         return;
 333                 }
 334                 /* FALL THROUGH */
 335         case NMI_IO_APIC:
 336                 __get_cpu_var(wd_enabled) = 1;
 337                 atomic_inc(&nmi_active);
 338         }
 339 }
 340
 341 void stop_apic_nmi_watchdog(void *unused)
 342 {
 343         /* only support LOCAL and IO APICs for now */
 344         if (!nmi_watchdog_active())
 345                 return;
 346         if (__get_cpu_var(wd_enabled) == 0)
 347                 return;
 348         if (nmi_watchdog == NMI_LOCAL_APIC)
 349                 lapic_watchdog_stop();
 350         else
 351                 __acpi_nmi_disable(NULL);
 352         __get_cpu_var(wd_enabled) = 0;
 353         atomic_dec(&nmi_active);
 354 }
 355
 356 /*
 357  * the best way to detect whether a CPU has a 'hard lockup' problem
 358  * is to check it's local APIC timer IRQ counts. If they are not
 359  * changing then that CPU has some problem.
 360  *
 361  * as these watchdog NMI IRQs are generated on every CPU, we only
 362  * have to check the current processor.
 363  *
 364  * since NMIs don't listen to _any_ locks, we have to be extremely
 365  * careful not to rely on unsafe variables. The printk might lock
 366  * up though, so we have to break up any console locks first ...
 367  * [when there will be more tty-related locks, break them up here too!]
 368  */
 369
 370 static DEFINE_PER_CPU(unsigned, last_irq_sum);
 371 static DEFINE_PER_CPU(local_t, alert_counter);
 372 static DEFINE_PER_CPU(int, nmi_touch);
 373
 374 void touch_nmi_watchdog(void)
 375 {
 376         if (nmi_watchdog_active()) {
 377                 unsigned cpu;
 378
 379                 /*
 380                  * Tell other CPUs to reset their alert counters. We cannot
 381                  * do it ourselves because the alert count increase is not
 382                  * atomic.
 383                  */
 384                 for_each_present_cpu(cpu) {
 385                         if (per_cpu(nmi_touch, cpu) != 1)
 386                                 per_cpu(nmi_touch, cpu) = 1;
 387                 }
 388         }
 389
 390         /*
 391          * Tickle the softlockup detector too:
 392          */
 393         touch_softlockup_watchdog();
 394 }
 395 EXPORT_SYMBOL(touch_nmi_watchdog);
 396
 397 notrace __kprobes int
 398 nmi_watchdog_tick(struct pt_regs *regs, unsigned reason)
 399 {
 400         /*
 401          * Since current_thread_info()-> is always on the stack, and we
 402          * always switch the stack NMI-atomically, it's safe to use
 403          * smp_processor_id().
 404          */
 405         unsigned int sum;
 406         int touched = 0;
 407         int cpu = smp_processor_id();
 408         int rc = 0;
 409
 410         /* check for other users first */
 411         if (notify_die(DIE_NMI, "nmi", regs, reason, 2, SIGINT)
 412                         == NOTIFY_STOP) {
 413                 rc = 1;
 414                 touched = 1;
 415         }
 416
 417         sum = get_timer_irqs(cpu);
 418
 419         if (__get_cpu_var(nmi_touch)) {
 420                 __get_cpu_var(nmi_touch) = 0;
 421                 touched = 1;
 422         }
 423
 424         if (cpu_isset(cpu, backtrace_mask)) {
 425                 static DEFINE_SPINLOCK(lock);   /* Serialise the printks */
 426
 427                 spin_lock(&lock);
 428                 printk(KERN_WARNING "NMI backtrace for cpu %d\n", cpu);
 429                 dump_stack();
 430                 spin_unlock(&lock);
 431                 cpu_clear(cpu, backtrace_mask);
 432         }
 433
 434         /* Could check oops_in_progress here too, but it's safer not to */
 435         if (mce_in_progress())
 436                 touched = 1;
 437
 438         /* if the none of the timers isn't firing, this cpu isn't doing much */
 439         if (!touched && __get_cpu_var(last_irq_sum) == sum) {
 440                 /*
 441                  * Ayiee, looks like this CPU is stuck ...
 442                  * wait a few IRQs (5 seconds) before doing the oops ...
 443                  */
 444                 local_inc(&__get_cpu_var(alert_counter));
 445                 if (local_read(&__get_cpu_var(alert_counter)) == 5 * nmi_hz)
 446                         /*
 447                          * die_nmi will return ONLY if NOTIFY_STOP happens..
 448                          */
 449                         die_nmi("BUG: NMI Watchdog detected LOCKUP",
 450                                 regs, panic_on_timeout);
 451         } else {
 452                 __get_cpu_var(last_irq_sum) = sum;
 453                 local_set(&__get_cpu_var(alert_counter), 0);
 454         }
 455
 456         /* see if the nmi watchdog went off */
 457         if (!__get_cpu_var(wd_enabled))
 458                 return rc;
 459         switch (nmi_watchdog) {
 460         case NMI_LOCAL_APIC:
 461                 rc |= lapic_wd_event(nmi_hz);
 462                 break;
 463         case NMI_IO_APIC:
 464                 /*
 465                  * don't know how to accurately check for this.
 466                  * just assume it was a watchdog timer interrupt
 467                  * This matches the old behaviour.
 468                  */
 469                 rc = 1;
 470                 break;
 471         }
 472         return rc;
 473 }
 474
 475 #ifdef CONFIG_SYSCTL
 476
 477 static void enable_ioapic_nmi_watchdog_single(void *unused)
 478 {
 479         __get_cpu_var(wd_enabled) = 1;
 480         atomic_inc(&nmi_active);
 481         __acpi_nmi_enable(NULL);
 482 }
 483
 484 static void enable_ioapic_nmi_watchdog(void)
 485 {
 486         on_each_cpu(enable_ioapic_nmi_watchdog_single, NULL, 1);
 487         touch_nmi_watchdog();
 488 }
 489
 490 static void disable_ioapic_nmi_watchdog(void)
 491 {
 492         on_each_cpu(stop_apic_nmi_watchdog, NULL, 1);
 493 }
 494
 495 static int __init setup_unknown_nmi_panic(char *str)
 496 {
 497         unknown_nmi_panic = 1;
 498         return 1;
 499 }
 500 __setup("unknown_nmi_panic", setup_unknown_nmi_panic);
 501
 502 static int unknown_nmi_panic_callback(struct pt_regs *regs, int cpu)
 503 {
 504         unsigned char reason = get_nmi_reason();
 505         char buf[64];
 506
 507         sprintf(buf, "NMI received for unknown reason %02x\n", reason);
 508         die_nmi(buf, regs, 1); /* Always panic here */
 509         return 0;
 510 }
 511
 512 /*
 513  * proc handler for /proc/sys/kernel/nmi
 514  */
 515 int proc_nmi_enabled(struct ctl_table *table, int write, struct file *file,
 516                         void __user *buffer, size_t *length, loff_t *ppos)
 517 {
 518         int old_state;
 519
 520         nmi_watchdog_enabled = (atomic_read(&nmi_active) > 0) ? 1 : 0;
 521         old_state = nmi_watchdog_enabled;
 522         proc_dointvec(table, write, file, buffer, length, ppos);
 523         if (!!old_state == !!nmi_watchdog_enabled)
 524                 return 0;
 525
 526         if (atomic_read(&nmi_active) < 0 || !nmi_watchdog_active()) {
 527                 printk(KERN_WARNING
 528                         "NMI watchdog is permanently disabled\n");
 529                 return -EIO;
 530         }
 531
 532         if (nmi_watchdog == NMI_LOCAL_APIC) {
 533                 if (nmi_watchdog_enabled)
 534                         enable_lapic_nmi_watchdog();
 535                 else
 536                         disable_lapic_nmi_watchdog();
 537         } else if (nmi_watchdog == NMI_IO_APIC) {
 538                 if (nmi_watchdog_enabled)
 539                         enable_ioapic_nmi_watchdog();
 540                 else
 541                         disable_ioapic_nmi_watchdog();
 542         } else {
 543                 printk(KERN_WARNING
 544                         "NMI watchdog doesn't know what hardware to touch\n");
 545                 return -EIO;
 546         }
 547         return 0;
 548 }
 549
 550 #endif /* CONFIG_SYSCTL */
 551
 552 int do_nmi_callback(struct pt_regs *regs, int cpu)
 553 {
 554 #ifdef CONFIG_SYSCTL
 555         if (unknown_nmi_panic)
 556                 return unknown_nmi_panic_callback(regs, cpu);
 557 #endif
 558         return 0;
 559 }
 560
 561 void __trigger_all_cpu_backtrace(void)
 562 {
 563         int i;
 564
 565         backtrace_mask = cpu_online_map;
 566         /* Wait for up to 10 seconds for all CPUs to do the backtrace */
 567         for (i = 0; i < 10 * 1000; i++) {
 568                 if (cpus_empty(backtrace_mask))
 569                         break;
 570                 mdelay(1);
 571         }
 572 }