arch/x86/kernel/nmi_32.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 <linux/delay.h>
  15 #include <linux/interrupt.h>
  16 #include <linux/module.h>
  17 #include <linux/nmi.h>
  18 #include <linux/sysdev.h>
  19 #include <linux/sysctl.h>
  20 #include <linux/percpu.h>
  21 #include <linux/kprobes.h>
  22 #include <linux/cpumask.h>
  23 #include <linux/kernel_stat.h>
  24 #include <linux/kdebug.h>
  25
  26 #include <asm/smp.h>
  27 #include <asm/nmi.h>
  28
  29 #include "mach_traps.h"
  30
  31 int unknown_nmi_panic;
  32 int nmi_watchdog_enabled;
  33
  34 static cpumask_t backtrace_mask = CPU_MASK_NONE;
  35
  36 /* nmi_active:
  37  * >0: the lapic NMI watchdog is active, but can be disabled
  38  * <0: the lapic NMI watchdog has not been set up, and cannot
  39  *     be enabled
  40  *  0: the lapic NMI watchdog is disabled, but can be enabled
  41  */
  42 atomic_t nmi_active = ATOMIC_INIT(0);           /* oprofile uses this */
  43
  44 unsigned int nmi_watchdog = NMI_DEFAULT;
  45 static unsigned int nmi_hz = HZ;
  46
  47 static DEFINE_PER_CPU(short, wd_enabled);
  48
  49 /* local prototypes */
  50 static int unknown_nmi_panic_callback(struct pt_regs *regs, int cpu);
  51
  52 static int endflag __initdata = 0;
  53
  54 /* The performance counters used by NMI_LOCAL_APIC don't trigger when
  55  * the CPU is idle. To make sure the NMI watchdog really ticks on all
  56  * CPUs during the test make them busy.
  57  */
  58 static __init void nmi_cpu_busy(void *data)
  59 {
  60 #ifdef CONFIG_SMP
  61         local_irq_enable_in_hardirq();
  62         /* Intentionally don't use cpu_relax here. This is
  63            to make sure that the performance counter really ticks,
  64            even if there is a simulator or similar that catches the
  65            pause instruction. On a real HT machine this is fine because
  66            all other CPUs are busy with "useless" delay loops and don't
  67            care if they get somewhat less cycles. */
  68         while (endflag == 0)
  69                 mb();
  70 #endif
  71 }
  72
  73 static int __init check_nmi_watchdog(void)
  74 {
  75         unsigned int *prev_nmi_count;
  76         int cpu;
  77
  78         if ((nmi_watchdog == NMI_NONE) || (nmi_watchdog == NMI_DISABLED))
  79                 return 0;
  80
  81         if (!atomic_read(&nmi_active))
  82                 return 0;
  83
  84         prev_nmi_count = kmalloc(NR_CPUS * sizeof(int), GFP_KERNEL);
  85         if (!prev_nmi_count)
  86                 return -1;
  87
  88         printk(KERN_INFO "Testing NMI watchdog ... ");
  89
  90         if (nmi_watchdog == NMI_LOCAL_APIC)
  91                 smp_call_function(nmi_cpu_busy, (void *)&endflag, 0, 0);
  92
  93         for_each_possible_cpu(cpu)
  94                 prev_nmi_count[cpu] = per_cpu(irq_stat, cpu).__nmi_count;
  95         local_irq_enable();
  96         mdelay((20*1000)/nmi_hz); // wait 20 ticks
  97
  98         for_each_possible_cpu(cpu) {
  99 #ifdef CONFIG_SMP
 100                 /* Check cpu_callin_map here because that is set
 101                    after the timer is started. */
 102                 if (!cpu_isset(cpu, cpu_callin_map))
 103                         continue;
 104 #endif
 105                 if (!per_cpu(wd_enabled, cpu))
 106                         continue;
 107                 if (nmi_count(cpu) - prev_nmi_count[cpu] <= 5) {
 108                         printk(KERN_WARNING "WARNING: CPU#%d: NMI "
 109                                 "appears to be stuck (%d->%d)!\n",
 110                                 cpu,
 111                                 prev_nmi_count[cpu],
 112                                 nmi_count(cpu));
 113                         per_cpu(wd_enabled, cpu) = 0;
 114                         atomic_dec(&nmi_active);
 115                 }
 116         }
 117         endflag = 1;
 118         if (!atomic_read(&nmi_active)) {
 119                 kfree(prev_nmi_count);
 120                 atomic_set(&nmi_active, -1);
 121                 return -1;
 122         }
 123         printk("OK.\n");
 124
 125         /* now that we know it works we can reduce NMI frequency to
 126            something more reasonable; makes a difference in some configs */
 127         if (nmi_watchdog == NMI_LOCAL_APIC)
 128                 nmi_hz = lapic_adjust_nmi_hz(1);
 129
 130         kfree(prev_nmi_count);
 131         return 0;
 132 }
 133 /* This needs to happen later in boot so counters are working */
 134 late_initcall(check_nmi_watchdog);
 135
 136 static int __init setup_nmi_watchdog(char *str)
 137 {
 138         int nmi;
 139
 140         get_option(&str, &nmi);
 141
 142         if ((nmi >= NMI_INVALID) || (nmi < NMI_NONE))
 143                 return 0;
 144
 145         nmi_watchdog = nmi;
 146         return 1;
 147 }
 148
 149 __setup("nmi_watchdog=", setup_nmi_watchdog);
 150
 151
 152 /* Suspend/resume support */
 153
 154 #ifdef CONFIG_PM
 155
 156 static int nmi_pm_active; /* nmi_active before suspend */
 157
 158 static int lapic_nmi_suspend(struct sys_device *dev, pm_message_t state)
 159 {
 160         /* only CPU0 goes here, other CPUs should be offline */
 161         nmi_pm_active = atomic_read(&nmi_active);
 162         stop_apic_nmi_watchdog(NULL);
 163         BUG_ON(atomic_read(&nmi_active) != 0);
 164         return 0;
 165 }
 166
 167 static int lapic_nmi_resume(struct sys_device *dev)
 168 {
 169         /* only CPU0 goes here, other CPUs should be offline */
 170         if (nmi_pm_active > 0) {
 171                 setup_apic_nmi_watchdog(NULL);
 172                 touch_nmi_watchdog();
 173         }
 174         return 0;
 175 }
 176
 177
 178 static struct sysdev_class nmi_sysclass = {
 179         set_kset_name("lapic_nmi"),
 180         .resume         = lapic_nmi_resume,
 181         .suspend        = lapic_nmi_suspend,
 182 };
 183
 184 static struct sys_device device_lapic_nmi = {
 185         .id     = 0,
 186         .cls    = &nmi_sysclass,
 187 };
 188
 189 static int __init init_lapic_nmi_sysfs(void)
 190 {
 191         int error;
 192
 193         /* should really be a BUG_ON but b/c this is an
 194          * init call, it just doesn't work.  -dcz
 195          */
 196         if (nmi_watchdog != NMI_LOCAL_APIC)
 197                 return 0;
 198
 199         if (atomic_read(&nmi_active) < 0)
 200                 return 0;
 201
 202         error = sysdev_class_register(&nmi_sysclass);
 203         if (!error)
 204                 error = sysdev_register(&device_lapic_nmi);
 205         return error;
 206 }
 207 /* must come after the local APIC's device_initcall() */
 208 late_initcall(init_lapic_nmi_sysfs);
 209
 210 #endif  /* CONFIG_PM */
 211
 212 static void __acpi_nmi_enable(void *__unused)
 213 {
 214         apic_write_around(APIC_LVT0, APIC_DM_NMI);
 215 }
 216
 217 /*
 218  * Enable timer based NMIs on all CPUs:
 219  */
 220 void acpi_nmi_enable(void)
 221 {
 222         if (atomic_read(&nmi_active) && nmi_watchdog == NMI_IO_APIC)
 223                 on_each_cpu(__acpi_nmi_enable, NULL, 0, 1);
 224 }
 225
 226 static void __acpi_nmi_disable(void *__unused)
 227 {
 228         apic_write(APIC_LVT0, APIC_DM_NMI | APIC_LVT_MASKED);
 229 }
 230
 231 /*
 232  * Disable timer based NMIs on all CPUs:
 233  */
 234 void acpi_nmi_disable(void)
 235 {
 236         if (atomic_read(&nmi_active) && nmi_watchdog == NMI_IO_APIC)
 237                 on_each_cpu(__acpi_nmi_disable, NULL, 0, 1);
 238 }
 239
 240 void setup_apic_nmi_watchdog (void *unused)
 241 {
 242         if (__get_cpu_var(wd_enabled))
 243                 return;
 244
 245         /* cheap hack to support suspend/resume */
 246         /* if cpu0 is not active neither should the other cpus */
 247         if ((smp_processor_id() != 0) && (atomic_read(&nmi_active) <= 0))
 248                 return;
 249
 250         switch (nmi_watchdog) {
 251         case NMI_LOCAL_APIC:
 252                 __get_cpu_var(wd_enabled) = 1; /* enable it before to avoid race with handler */
 253                 if (lapic_watchdog_init(nmi_hz) < 0) {
 254                         __get_cpu_var(wd_enabled) = 0;
 255                         return;
 256                 }
 257                 /* FALL THROUGH */
 258         case NMI_IO_APIC:
 259                 __get_cpu_var(wd_enabled) = 1;
 260                 atomic_inc(&nmi_active);
 261         }
 262 }
 263
 264 void stop_apic_nmi_watchdog(void *unused)
 265 {
 266         /* only support LOCAL and IO APICs for now */
 267         if ((nmi_watchdog != NMI_LOCAL_APIC) &&
 268             (nmi_watchdog != NMI_IO_APIC))
 269                 return;
 270         if (__get_cpu_var(wd_enabled) == 0)
 271                 return;
 272         if (nmi_watchdog == NMI_LOCAL_APIC)
 273                 lapic_watchdog_stop();
 274         __get_cpu_var(wd_enabled) = 0;
 275         atomic_dec(&nmi_active);
 276 }
 277
 278 /*
 279  * the best way to detect whether a CPU has a 'hard lockup' problem
 280  * is to check it's local APIC timer IRQ counts. If they are not
 281  * changing then that CPU has some problem.
 282  *
 283  * as these watchdog NMI IRQs are generated on every CPU, we only
 284  * have to check the current processor.
 285  *
 286  * since NMIs don't listen to _any_ locks, we have to be extremely
 287  * careful not to rely on unsafe variables. The printk might lock
 288  * up though, so we have to break up any console locks first ...
 289  * [when there will be more tty-related locks, break them up
 290  *  here too!]
 291  */
 292
 293 static unsigned int
 294         last_irq_sums [NR_CPUS],
 295         alert_counter [NR_CPUS];
 296
 297 void touch_nmi_watchdog(void)
 298 {
 299         if (nmi_watchdog > 0) {
 300                 unsigned cpu;
 301
 302                 /*
 303                  * Just reset the alert counters, (other CPUs might be
 304                  * spinning on locks we hold):
 305                  */
 306                 for_each_present_cpu(cpu) {
 307                         if (alert_counter[cpu])
 308                                 alert_counter[cpu] = 0;
 309                 }
 310         }
 311
 312         /*
 313          * Tickle the softlockup detector too:
 314          */
 315         touch_softlockup_watchdog();
 316 }
 317 EXPORT_SYMBOL(touch_nmi_watchdog);
 318
 319 extern void die_nmi(struct pt_regs *, const char *msg);
 320
 321 __kprobes int nmi_watchdog_tick(struct pt_regs * regs, unsigned reason)
 322 {
 323
 324         /*
 325          * Since current_thread_info()-> is always on the stack, and we
 326          * always switch the stack NMI-atomically, it's safe to use
 327          * smp_processor_id().
 328          */
 329         unsigned int sum;
 330         int touched = 0;
 331         int cpu = smp_processor_id();
 332         int rc=0;
 333
 334         /* check for other users first */
 335         if (notify_die(DIE_NMI, "nmi", regs, reason, 2, SIGINT)
 336                         == NOTIFY_STOP) {
 337                 rc = 1;
 338                 touched = 1;
 339         }
 340
 341         if (cpu_isset(cpu, backtrace_mask)) {
 342                 static DEFINE_SPINLOCK(lock);   /* Serialise the printks */
 343
 344                 spin_lock(&lock);
 345                 printk("NMI backtrace for cpu %d\n", cpu);
 346                 dump_stack();
 347                 spin_unlock(&lock);
 348                 cpu_clear(cpu, backtrace_mask);
 349         }
 350
 351         /*
 352          * Take the local apic timer and PIT/HPET into account. We don't
 353          * know which one is active, when we have highres/dyntick on
 354          */
 355         sum = per_cpu(irq_stat, cpu).apic_timer_irqs +
 356                 per_cpu(irq_stat, cpu).irq0_irqs;
 357
 358         /* if the none of the timers isn't firing, this cpu isn't doing much */
 359         if (!touched && last_irq_sums[cpu] == sum) {
 360                 /*
 361                  * Ayiee, looks like this CPU is stuck ...
 362                  * wait a few IRQs (5 seconds) before doing the oops ...
 363                  */
 364                 alert_counter[cpu]++;
 365                 if (alert_counter[cpu] == 5*nmi_hz)
 366                         /*
 367                          * die_nmi will return ONLY if NOTIFY_STOP happens..
 368                          */
 369                         die_nmi(regs, "BUG: NMI Watchdog detected LOCKUP");
 370         } else {
 371                 last_irq_sums[cpu] = sum;
 372                 alert_counter[cpu] = 0;
 373         }
 374         /* see if the nmi watchdog went off */
 375         if (!__get_cpu_var(wd_enabled))
 376                 return rc;
 377         switch (nmi_watchdog) {
 378         case NMI_LOCAL_APIC:
 379                 rc |= lapic_wd_event(nmi_hz);
 380                 break;
 381         case NMI_IO_APIC:
 382                 /* don't know how to accurately check for this.
 383                  * just assume it was a watchdog timer interrupt
 384                  * This matches the old behaviour.
 385                  */
 386                 rc = 1;
 387                 break;
 388         }
 389         return rc;
 390 }
 391
 392 int do_nmi_callback(struct pt_regs * regs, int cpu)
 393 {
 394 #ifdef CONFIG_SYSCTL
 395         if (unknown_nmi_panic)
 396                 return unknown_nmi_panic_callback(regs, cpu);
 397 #endif
 398         return 0;
 399 }
 400
 401 #ifdef CONFIG_SYSCTL
 402
 403 static int unknown_nmi_panic_callback(struct pt_regs *regs, int cpu)
 404 {
 405         unsigned char reason = get_nmi_reason();
 406         char buf[64];
 407
 408         sprintf(buf, "NMI received for unknown reason %02x\n", reason);
 409         die_nmi(regs, buf);
 410         return 0;
 411 }
 412
 413 /*
 414  * proc handler for /proc/sys/kernel/nmi
 415  */
 416 int proc_nmi_enabled(struct ctl_table *table, int write, struct file *file,
 417                         void __user *buffer, size_t *length, loff_t *ppos)
 418 {
 419         int old_state;
 420
 421         nmi_watchdog_enabled = (atomic_read(&nmi_active) > 0) ? 1 : 0;
 422         old_state = nmi_watchdog_enabled;
 423         proc_dointvec(table, write, file, buffer, length, ppos);
 424         if (!!old_state == !!nmi_watchdog_enabled)
 425                 return 0;
 426
 427         if (atomic_read(&nmi_active) < 0 || nmi_watchdog == NMI_DISABLED) {
 428                 printk( KERN_WARNING "NMI watchdog is permanently disabled\n");
 429                 return -EIO;
 430         }
 431
 432         if (nmi_watchdog == NMI_DEFAULT) {
 433                 if (lapic_watchdog_ok())
 434                         nmi_watchdog = NMI_LOCAL_APIC;
 435                 else
 436                         nmi_watchdog = NMI_IO_APIC;
 437         }
 438
 439         if (nmi_watchdog == NMI_LOCAL_APIC) {
 440                 if (nmi_watchdog_enabled)
 441                         enable_lapic_nmi_watchdog();
 442                 else
 443                         disable_lapic_nmi_watchdog();
 444         } else {
 445                 printk( KERN_WARNING
 446                         "NMI watchdog doesn't know what hardware to touch\n");
 447                 return -EIO;
 448         }
 449         return 0;
 450 }
 451
 452 #endif
 453
 454 void __trigger_all_cpu_backtrace(void)
 455 {
 456         int i;
 457
 458         backtrace_mask = cpu_online_map;
 459         /* Wait for up to 10 seconds for all CPUs to do the backtrace */
 460         for (i = 0; i < 10 * 1000; i++) {
 461                 if (cpus_empty(backtrace_mask))
 462                         break;
 463                 mdelay(1);
 464         }
 465 }
 466
 467 EXPORT_SYMBOL(nmi_active);
 468 EXPORT_SYMBOL(nmi_watchdog);