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 #include <linux/nmi.h>
  30
  31 #include <asm/i8259.h>
  32 #include <asm/io_apic.h>
  33 #include <asm/proto.h>
  34 #include <asm/timer.h>
  35
  36 #include <asm/mce.h>
  37
  38 #include <mach_traps.h>
  39
  40 int unknown_nmi_panic;
  41 int nmi_watchdog_enabled;
  42
  43 static cpumask_t backtrace_mask = CPU_MASK_NONE;
  44
  45 /* nmi_active:
  46  * >0: the lapic NMI watchdog is active, but can be disabled
  47  * <0: the lapic NMI watchdog has not been set up, and cannot
  48  *     be enabled
  49  *  0: the lapic NMI watchdog is disabled, but can be enabled
  50  */
  51 atomic_t nmi_active = ATOMIC_INIT(0);           /* oprofile uses this */
  52 EXPORT_SYMBOL(nmi_active);
  53
  54 unsigned int nmi_watchdog = NMI_NONE;
  55 EXPORT_SYMBOL(nmi_watchdog);
  56
  57 static int panic_on_timeout;
  58
  59 static unsigned int nmi_hz = HZ;
  60 static DEFINE_PER_CPU(short, wd_enabled);
  61 static int endflag __initdata;
  62
  63 static inline unsigned int get_nmi_count(int cpu)
  64 {
  65 #ifdef CONFIG_X86_64
  66         return cpu_pda(cpu)->__nmi_count;
  67 #else
  68         return nmi_count(cpu);
  69 #endif
  70 }
  71
  72 static inline int mce_in_progress(void)
  73 {
  74 #if defined(CONFIG_X86_64) && defined(CONFIG_X86_MCE)
  75         return atomic_read(&mce_entry) > 0;
  76 #endif
  77         return 0;
  78 }
  79
  80 /*
  81  * Take the local apic timer and PIT/HPET into account. We don't
  82  * know which one is active, when we have highres/dyntick on
  83  */
  84 static inline unsigned int get_timer_irqs(int cpu)
  85 {
  86 #ifdef CONFIG_X86_64
  87         return read_pda(apic_timer_irqs) + read_pda(irq0_irqs);
  88 #else
  89         return per_cpu(irq_stat, cpu).apic_timer_irqs +
  90                 per_cpu(irq_stat, cpu).irq0_irqs;
  91 #endif
  92 }
  93
  94 #ifdef CONFIG_SMP
  95 /*
  96  * The performance counters used by NMI_LOCAL_APIC don't trigger when
  97  * the CPU is idle. To make sure the NMI watchdog really ticks on all
  98  * CPUs during the test make them busy.
  99  */
 100 static __init void nmi_cpu_busy(void *data)
 101 {
 102         local_irq_enable_in_hardirq();
 103         /*
 104          * Intentionally don't use cpu_relax here. This is
 105          * to make sure that the performance counter really ticks,
 106          * even if there is a simulator or similar that catches the
 107          * pause instruction. On a real HT machine this is fine because
 108          * all other CPUs are busy with "useless" delay loops and don't
 109          * care if they get somewhat less cycles.
 110          */
 111         while (endflag == 0)
 112                 mb();
 113 }
 114 #endif
 115
 116 static void report_broken_nmi(int cpu, int *prev_nmi_count)
 117 {
 118         printk(KERN_CONT "\n");
 119
 120         printk(KERN_WARNING
 121                 "WARNING: CPU#%d: NMI appears to be stuck (%d->%d)!\n",
 122                         cpu, prev_nmi_count[cpu], get_nmi_count(cpu));
 123
 124         printk(KERN_WARNING
 125                 "Please report this to bugzilla.kernel.org,\n");
 126         printk(KERN_WARNING
 127                 "and attach the output of the 'dmesg' command.\n");
 128
 129         per_cpu(wd_enabled, cpu) = 0;
 130         atomic_dec(&nmi_active);
 131 }
 132
 133 static void __acpi_nmi_disable(void *__unused)
 134 {
 135         apic_write(APIC_LVT0, APIC_DM_NMI | APIC_LVT_MASKED);
 136 }
 137
 138 int __init check_nmi_watchdog(void)
 139 {
 140         unsigned int *prev_nmi_count;
 141         int cpu;
 142
 143         if (!nmi_watchdog_active() || !atomic_read(&nmi_active))
 144                 return 0;
 145
 146         prev_nmi_count = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL);
 147         if (!prev_nmi_count)
 148                 goto error;
 149
 150         printk(KERN_INFO "Testing NMI watchdog ... ");
 151
 152 #ifdef CONFIG_SMP
 153         if (nmi_watchdog == NMI_LOCAL_APIC)
 154                 smp_call_function(nmi_cpu_busy, (void *)&endflag, 0);
 155 #endif
 156
 157         for_each_possible_cpu(cpu)
 158                 prev_nmi_count[cpu] = get_nmi_count(cpu);
 159         local_irq_enable();
 160         mdelay((20 * 1000) / nmi_hz); /* wait 20 ticks */
 161
 162         for_each_online_cpu(cpu) {
 163                 if (!per_cpu(wd_enabled, cpu))
 164                         continue;
 165                 if (get_nmi_count(cpu) - prev_nmi_count[cpu] <= 5)
 166                         report_broken_nmi(cpu, prev_nmi_count);
 167         }
 168         endflag = 1;
 169         if (!atomic_read(&nmi_active)) {
 170                 kfree(prev_nmi_count);
 171                 atomic_set(&nmi_active, -1);
 172                 goto error;
 173         }
 174         printk("OK.\n");
 175
 176         /*
 177          * now that we know it works we can reduce NMI frequency to
 178          * something more reasonable; makes a difference in some configs
 179          */
 180         if (nmi_watchdog == NMI_LOCAL_APIC)
 181                 nmi_hz = lapic_adjust_nmi_hz(1);
 182
 183         kfree(prev_nmi_count);
 184         return 0;
 185 error:
 186         if (nmi_watchdog == NMI_IO_APIC) {
 187                 if (!timer_through_8259)
 188                         disable_8259A_irq(0);
 189                 on_each_cpu(__acpi_nmi_disable, NULL, 1);
 190         }
 191
 192 #ifdef CONFIG_X86_32
 193         timer_ack = 0;
 194 #endif
 195         return -1;
 196 }
 197
 198 static int __init setup_nmi_watchdog(char *str)
 199 {
 200         unsigned int nmi;
 201
 202         if (!strncmp(str, "panic", 5)) {
 203                 panic_on_timeout = 1;
 204                 str = strchr(str, ',');
 205                 if (!str)
 206                         return 1;
 207                 ++str;
 208         }
 209
 210         if (!strncmp(str, "lapic", 5))
 211                 nmi_watchdog = NMI_LOCAL_APIC;
 212         else if (!strncmp(str, "ioapic", 6))
 213                 nmi_watchdog = NMI_IO_APIC;
 214         else {
 215                 get_option(&str, &nmi);
 216                 if (nmi >= NMI_INVALID)
 217                         return 0;
 218                 nmi_watchdog = nmi;
 219         }
 220
 221         return 1;
 222 }
 223 __setup("nmi_watchdog=", setup_nmi_watchdog);
 224
 225 /*
 226  * Suspend/resume support
 227  */
 228 #ifdef CONFIG_PM
 229
 230 static int nmi_pm_active; /* nmi_active before suspend */
 231
 232 static int lapic_nmi_suspend(struct sys_device *dev, pm_message_t state)
 233 {
 234         /* only CPU0 goes here, other CPUs should be offline */
 235         nmi_pm_active = atomic_read(&nmi_active);
 236         stop_apic_nmi_watchdog(NULL);
 237         BUG_ON(atomic_read(&nmi_active) != 0);
 238         return 0;
 239 }
 240
 241 static int lapic_nmi_resume(struct sys_device *dev)
 242 {
 243         /* only CPU0 goes here, other CPUs should be offline */
 244         if (nmi_pm_active > 0) {
 245                 setup_apic_nmi_watchdog(NULL);
 246                 touch_nmi_watchdog();
 247         }
 248         return 0;
 249 }
 250
 251 static struct sysdev_class nmi_sysclass = {
 252         .name           = "lapic_nmi",
 253         .resume         = lapic_nmi_resume,
 254         .suspend        = lapic_nmi_suspend,
 255 };
 256
 257 static struct sys_device device_lapic_nmi = {
 258         .id     = 0,
 259         .cls    = &nmi_sysclass,
 260 };
 261
 262 static int __init init_lapic_nmi_sysfs(void)
 263 {
 264         int error;
 265
 266         /*
 267          * should really be a BUG_ON but b/c this is an
 268          * init call, it just doesn't work.  -dcz
 269          */
 270         if (nmi_watchdog != NMI_LOCAL_APIC)
 271                 return 0;
 272
 273         if (atomic_read(&nmi_active) < 0)
 274                 return 0;
 275
 276         error = sysdev_class_register(&nmi_sysclass);
 277         if (!error)
 278                 error = sysdev_register(&device_lapic_nmi);
 279         return error;
 280 }
 281
 282 /* must come after the local APIC's device_initcall() */
 283 late_initcall(init_lapic_nmi_sysfs);
 284
 285 #endif  /* CONFIG_PM */
 286
 287 static void __acpi_nmi_enable(void *__unused)
 288 {
 289         apic_write(APIC_LVT0, APIC_DM_NMI);
 290 }
 291
 292 /*
 293  * Enable timer based NMIs on all CPUs:
 294  */
 295 void acpi_nmi_enable(void)
 296 {
 297         if (atomic_read(&nmi_active) && nmi_watchdog == NMI_IO_APIC)
 298                 on_each_cpu(__acpi_nmi_enable, NULL, 1);
 299 }
 300
 301 /*
 302  * Disable timer based NMIs on all CPUs:
 303  */
 304 void acpi_nmi_disable(void)
 305 {
 306         if (atomic_read(&nmi_active) && nmi_watchdog == NMI_IO_APIC)
 307                 on_each_cpu(__acpi_nmi_disable, NULL, 1);
 308 }
 309
 310 /*
 311  * This function is called as soon the LAPIC NMI watchdog driver has everything
 312  * in place and it's ready to check if the NMIs belong to the NMI watchdog
 313  */
 314 void cpu_nmi_set_wd_enabled(void)
 315 {
 316         __get_cpu_var(wd_enabled) = 1;
 317 }
 318
 319 void setup_apic_nmi_watchdog(void *unused)
 320 {
 321         if (__get_cpu_var(wd_enabled))
 322                 return;
 323
 324         /* cheap hack to support suspend/resume */
 325         /* if cpu0 is not active neither should the other cpus */
 326         if (smp_processor_id() != 0 && atomic_read(&nmi_active) <= 0)
 327                 return;
 328
 329         switch (nmi_watchdog) {
 330         case NMI_LOCAL_APIC:
 331                 if (lapic_watchdog_init(nmi_hz) < 0) {
 332                         __get_cpu_var(wd_enabled) = 0;
 333                         return;
 334                 }
 335                 /* FALL THROUGH */
 336         case NMI_IO_APIC:
 337                 __get_cpu_var(wd_enabled) = 1;
 338                 atomic_inc(&nmi_active);
 339         }
 340 }
 341
 342 void stop_apic_nmi_watchdog(void *unused)
 343 {
 344         /* only support LOCAL and IO APICs for now */
 345         if (!nmi_watchdog_active())
 346                 return;
 347         if (__get_cpu_var(wd_enabled) == 0)
 348                 return;
 349         if (nmi_watchdog == NMI_LOCAL_APIC)
 350                 lapic_watchdog_stop();
 351         else
 352                 __acpi_nmi_disable(NULL);
 353         __get_cpu_var(wd_enabled) = 0;
 354         atomic_dec(&nmi_active);
 355 }
 356
 357 /*
 358  * the best way to detect whether a CPU has a 'hard lockup' problem
 359  * is to check it's local APIC timer IRQ counts. If they are not
 360  * changing then that CPU has some problem.
 361  *
 362  * as these watchdog NMI IRQs are generated on every CPU, we only
 363  * have to check the current processor.
 364  *
 365  * since NMIs don't listen to _any_ locks, we have to be extremely
 366  * careful not to rely on unsafe variables. The printk might lock
 367  * up though, so we have to break up any console locks first ...
 368  * [when there will be more tty-related locks, break them up here too!]
 369  */
 370
 371 static DEFINE_PER_CPU(unsigned, last_irq_sum);
 372 static DEFINE_PER_CPU(local_t, alert_counter);
 373 static DEFINE_PER_CPU(int, nmi_touch);
 374
 375 void touch_nmi_watchdog(void)
 376 {
 377         if (nmi_watchdog_active()) {
 378                 unsigned cpu;
 379
 380                 /*
 381                  * Tell other CPUs to reset their alert counters. We cannot
 382                  * do it ourselves because the alert count increase is not
 383                  * atomic.
 384                  */
 385                 for_each_present_cpu(cpu) {
 386                         if (per_cpu(nmi_touch, cpu) != 1)
 387                                 per_cpu(nmi_touch, cpu) = 1;
 388                 }
 389         }
 390
 391         /*
 392          * Tickle the softlockup detector too:
 393          */
 394         touch_softlockup_watchdog();
 395 }
 396 EXPORT_SYMBOL(touch_nmi_watchdog);
 397
 398 notrace __kprobes int
 399 nmi_watchdog_tick(struct pt_regs *regs, unsigned reason)
 400 {
 401         /*
 402          * Since current_thread_info()-> is always on the stack, and we
 403          * always switch the stack NMI-atomically, it's safe to use
 404          * smp_processor_id().
 405          */
 406         unsigned int sum;
 407         int touched = 0;
 408         int cpu = smp_processor_id();
 409         int rc = 0;
 410
 411         /* check for other users first */
 412         if (notify_die(DIE_NMI, "nmi", regs, reason, 2, SIGINT)
 413                         == NOTIFY_STOP) {
 414                 rc = 1;
 415                 touched = 1;
 416         }
 417
 418         sum = get_timer_irqs(cpu);
 419
 420         if (__get_cpu_var(nmi_touch)) {
 421                 __get_cpu_var(nmi_touch) = 0;
 422                 touched = 1;
 423         }
 424
 425         if (cpu_isset(cpu, backtrace_mask)) {
 426                 static DEFINE_SPINLOCK(lock);   /* Serialise the printks */
 427
 428                 spin_lock(&lock);
 429                 printk(KERN_WARNING "NMI backtrace for cpu %d\n", cpu);
 430                 dump_stack();
 431                 spin_unlock(&lock);
 432                 cpu_clear(cpu, backtrace_mask);
 433         }
 434
 435         /* Could check oops_in_progress here too, but it's safer not to */
 436         if (mce_in_progress())
 437                 touched = 1;
 438
 439         /* if the none of the timers isn't firing, this cpu isn't doing much */
 440         if (!touched && __get_cpu_var(last_irq_sum) == sum) {
 441                 /*
 442                  * Ayiee, looks like this CPU is stuck ...
 443                  * wait a few IRQs (5 seconds) before doing the oops ...
 444                  */
 445                 local_inc(&__get_cpu_var(alert_counter));
 446                 if (local_read(&__get_cpu_var(alert_counter)) == 5 * nmi_hz)
 447                         /*
 448                          * die_nmi will return ONLY if NOTIFY_STOP happens..
 449                          */
 450                         die_nmi("BUG: NMI Watchdog detected LOCKUP",
 451                                 regs, panic_on_timeout);
 452         } else {
 453                 __get_cpu_var(last_irq_sum) = sum;
 454                 local_set(&__get_cpu_var(alert_counter), 0);
 455         }
 456
 457         /* see if the nmi watchdog went off */
 458         if (!__get_cpu_var(wd_enabled))
 459                 return rc;
 460         switch (nmi_watchdog) {
 461         case NMI_LOCAL_APIC:
 462                 rc |= lapic_wd_event(nmi_hz);
 463                 break;
 464         case NMI_IO_APIC:
 465                 /*
 466                  * don't know how to accurately check for this.
 467                  * just assume it was a watchdog timer interrupt
 468                  * This matches the old behaviour.
 469                  */
 470                 rc = 1;
 471                 break;
 472         }
 473         return rc;
 474 }
 475
 476 #ifdef CONFIG_SYSCTL
 477
 478 static void enable_ioapic_nmi_watchdog_single(void *unused)
 479 {
 480         __get_cpu_var(wd_enabled) = 1;
 481         atomic_inc(&nmi_active);
 482         __acpi_nmi_enable(NULL);
 483 }
 484
 485 static void enable_ioapic_nmi_watchdog(void)
 486 {
 487         on_each_cpu(enable_ioapic_nmi_watchdog_single, NULL, 1);
 488         touch_nmi_watchdog();
 489 }
 490
 491 static void disable_ioapic_nmi_watchdog(void)
 492 {
 493         on_each_cpu(stop_apic_nmi_watchdog, NULL, 1);
 494 }
 495
 496 static int __init setup_unknown_nmi_panic(char *str)
 497 {
 498         unknown_nmi_panic = 1;
 499         return 1;
 500 }
 501 __setup("unknown_nmi_panic", setup_unknown_nmi_panic);
 502
 503 static int unknown_nmi_panic_callback(struct pt_regs *regs, int cpu)
 504 {
 505         unsigned char reason = get_nmi_reason();
 506         char buf[64];
 507
 508         sprintf(buf, "NMI received for unknown reason %02x\n", reason);
 509         die_nmi(buf, regs, 1); /* Always panic here */
 510         return 0;
 511 }
 512
 513 /*
 514  * proc handler for /proc/sys/kernel/nmi
 515  */
 516 int proc_nmi_enabled(struct ctl_table *table, int write, struct file *file,
 517                         void __user *buffer, size_t *length, loff_t *ppos)
 518 {
 519         int old_state;
 520
 521         nmi_watchdog_enabled = (atomic_read(&nmi_active) > 0) ? 1 : 0;
 522         old_state = nmi_watchdog_enabled;
 523         proc_dointvec(table, write, file, buffer, length, ppos);
 524         if (!!old_state == !!nmi_watchdog_enabled)
 525                 return 0;
 526
 527         if (atomic_read(&nmi_active) < 0 || !nmi_watchdog_active()) {
 528                 printk(KERN_WARNING
 529                         "NMI watchdog is permanently disabled\n");
 530                 return -EIO;
 531         }
 532
 533         if (nmi_watchdog == NMI_LOCAL_APIC) {
 534                 if (nmi_watchdog_enabled)
 535                         enable_lapic_nmi_watchdog();
 536                 else
 537                         disable_lapic_nmi_watchdog();
 538         } else if (nmi_watchdog == NMI_IO_APIC) {
 539                 if (nmi_watchdog_enabled)
 540                         enable_ioapic_nmi_watchdog();
 541                 else
 542                         disable_ioapic_nmi_watchdog();
 543         } else {
 544                 printk(KERN_WARNING
 545                         "NMI watchdog doesn't know what hardware to touch\n");
 546                 return -EIO;
 547         }
 548         return 0;
 549 }
 550
 551 #endif /* CONFIG_SYSCTL */
 552
 553 int do_nmi_callback(struct pt_regs *regs, int cpu)
 554 {
 555 #ifdef CONFIG_SYSCTL
 556         if (unknown_nmi_panic)
 557                 return unknown_nmi_panic_callback(regs, cpu);
 558 #endif
 559         return 0;
 560 }
 561
 562 void __trigger_all_cpu_backtrace(void)
 563 {
 564         int i;
 565
 566         backtrace_mask = cpu_online_map;
 567         /* Wait for up to 10 seconds for all CPUs to do the backtrace */
 568         for (i = 0; i < 10 * 1000; i++) {
 569                 if (cpus_empty(backtrace_mask))
 570                         break;
 571                 mdelay(1);
 572         }
 573 }