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 #include <linux/slab.h>
  26
  27 #include <asm/smp.h>
  28 #include <asm/nmi.h>
  29 #include <asm/timer.h>
  30
  31 #include "mach_traps.h"
  32
  33 int unknown_nmi_panic;
  34 int nmi_watchdog_enabled;
  35
  36 static cpumask_t backtrace_mask = CPU_MASK_NONE;
  37
  38 /* nmi_active:
  39  * >0: the lapic NMI watchdog is active, but can be disabled
  40  * <0: the lapic NMI watchdog has not been set up, and cannot
  41  *     be enabled
  42  *  0: the lapic NMI watchdog is disabled, but can be enabled
  43  */
  44 atomic_t nmi_active = ATOMIC_INIT(0);           /* oprofile uses this */
  45
  46 unsigned int nmi_watchdog = NMI_DEFAULT;
  47 static unsigned int nmi_hz = HZ;
  48
  49 static DEFINE_PER_CPU(short, wd_enabled);
  50
  51 static int endflag __initdata = 0;
  52
  53 #ifdef CONFIG_SMP
  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         local_irq_enable_in_hardirq();
  61         /* Intentionally don't use cpu_relax here. This is
  62            to make sure that the performance counter really ticks,
  63            even if there is a simulator or similar that catches the
  64            pause instruction. On a real HT machine this is fine because
  65            all other CPUs are busy with "useless" delay loops and don't
  66            care if they get somewhat less cycles. */
  67         while (endflag == 0)
  68                 mb();
  69 }
  70 #endif
  71
  72 int __init check_nmi_watchdog(void)
  73 {
  74         unsigned int *prev_nmi_count;
  75         int cpu;
  76
  77         if ((nmi_watchdog == NMI_NONE) || (nmi_watchdog == NMI_DISABLED))
  78                 return 0;
  79
  80         if (!atomic_read(&nmi_active))
  81                 return 0;
  82
  83         prev_nmi_count = kmalloc(NR_CPUS * sizeof(int), GFP_KERNEL);
  84         if (!prev_nmi_count)
  85                 goto error;
  86
  87         printk(KERN_INFO "Testing NMI watchdog ... ");
  88
  89 #ifdef CONFIG_SMP
  90         if (nmi_watchdog == NMI_LOCAL_APIC)
  91                 smp_call_function(nmi_cpu_busy, (void *)&endflag, 0, 0);
  92 #endif
  93
  94         for_each_possible_cpu(cpu)
  95                 prev_nmi_count[cpu] = nmi_count(cpu);
  96         local_irq_enable();
  97         mdelay((20*1000)/nmi_hz); // wait 20 ticks
  98
  99         for_each_possible_cpu(cpu) {
 100 #ifdef CONFIG_SMP
 101                 /* Check cpu_callin_map here because that is set
 102                    after the timer is started. */
 103                 if (!cpu_isset(cpu, cpu_callin_map))
 104                         continue;
 105 #endif
 106                 if (!per_cpu(wd_enabled, cpu))
 107                         continue;
 108                 if (nmi_count(cpu) - prev_nmi_count[cpu] <= 5) {
 109                         printk(KERN_WARNING "WARNING: CPU#%d: NMI "
 110                                 "appears to be stuck (%d->%d)!\n",
 111                                 cpu,
 112                                 prev_nmi_count[cpu],
 113                                 nmi_count(cpu));
 114                         per_cpu(wd_enabled, cpu) = 0;
 115                         atomic_dec(&nmi_active);
 116                 }
 117         }
 118         endflag = 1;
 119         if (!atomic_read(&nmi_active)) {
 120                 kfree(prev_nmi_count);
 121                 atomic_set(&nmi_active, -1);
 122                 goto error;
 123         }
 124         printk("OK.\n");
 125
 126         /* now that we know it works we can reduce NMI frequency to
 127            something more reasonable; makes a difference in some configs */
 128         if (nmi_watchdog == NMI_LOCAL_APIC)
 129                 nmi_hz = lapic_adjust_nmi_hz(1);
 130
 131         kfree(prev_nmi_count);
 132         return 0;
 133 error:
 134         timer_ack = !cpu_has_tsc;
 135
 136         return -1;
 137 }
 138
 139 static int __init setup_nmi_watchdog(char *str)
 140 {
 141         int nmi;
 142
 143         get_option(&str, &nmi);
 144
 145         if ((nmi >= NMI_INVALID) || (nmi < NMI_NONE))
 146                 return 0;
 147
 148         nmi_watchdog = nmi;
 149         return 1;
 150 }
 151
 152 __setup("nmi_watchdog=", setup_nmi_watchdog);
 153
 154
 155 /* Suspend/resume support */
 156
 157 #ifdef CONFIG_PM
 158
 159 static int nmi_pm_active; /* nmi_active before suspend */
 160
 161 static int lapic_nmi_suspend(struct sys_device *dev, pm_message_t state)
 162 {
 163         /* only CPU0 goes here, other CPUs should be offline */
 164         nmi_pm_active = atomic_read(&nmi_active);
 165         stop_apic_nmi_watchdog(NULL);
 166         BUG_ON(atomic_read(&nmi_active) != 0);
 167         return 0;
 168 }
 169
 170 static int lapic_nmi_resume(struct sys_device *dev)
 171 {
 172         /* only CPU0 goes here, other CPUs should be offline */
 173         if (nmi_pm_active > 0) {
 174                 setup_apic_nmi_watchdog(NULL);
 175                 touch_nmi_watchdog();
 176         }
 177         return 0;
 178 }
 179
 180
 181 static struct sysdev_class nmi_sysclass = {
 182         .name           = "lapic_nmi",
 183         .resume         = lapic_nmi_resume,
 184         .suspend        = lapic_nmi_suspend,
 185 };
 186
 187 static struct sys_device device_lapic_nmi = {
 188         .id     = 0,
 189         .cls    = &nmi_sysclass,
 190 };
 191
 192 static int __init init_lapic_nmi_sysfs(void)
 193 {
 194         int error;
 195
 196         /* should really be a BUG_ON but b/c this is an
 197          * init call, it just doesn't work.  -dcz
 198          */
 199         if (nmi_watchdog != NMI_LOCAL_APIC)
 200                 return 0;
 201
 202         if (atomic_read(&nmi_active) < 0)
 203                 return 0;
 204
 205         error = sysdev_class_register(&nmi_sysclass);
 206         if (!error)
 207                 error = sysdev_register(&device_lapic_nmi);
 208         return error;
 209 }
 210 /* must come after the local APIC's device_initcall() */
 211 late_initcall(init_lapic_nmi_sysfs);
 212
 213 #endif  /* CONFIG_PM */
 214
 215 static void __acpi_nmi_enable(void *__unused)
 216 {
 217         apic_write_around(APIC_LVT0, APIC_DM_NMI);
 218 }
 219
 220 /*
 221  * Enable timer based NMIs on all CPUs:
 222  */
 223 void acpi_nmi_enable(void)
 224 {
 225         if (atomic_read(&nmi_active) && nmi_watchdog == NMI_IO_APIC)
 226                 on_each_cpu(__acpi_nmi_enable, NULL, 0, 1);
 227 }
 228
 229 static void __acpi_nmi_disable(void *__unused)
 230 {
 231         apic_write(APIC_LVT0, APIC_DM_NMI | APIC_LVT_MASKED);
 232 }
 233
 234 /*
 235  * Disable timer based NMIs on all CPUs:
 236  */
 237 void acpi_nmi_disable(void)
 238 {
 239         if (atomic_read(&nmi_active) && nmi_watchdog == NMI_IO_APIC)
 240                 on_each_cpu(__acpi_nmi_disable, NULL, 0, 1);
 241 }
 242
 243 void setup_apic_nmi_watchdog(void *unused)
 244 {
 245         if (__get_cpu_var(wd_enabled))
 246                 return;
 247
 248         /* cheap hack to support suspend/resume */
 249         /* if cpu0 is not active neither should the other cpus */
 250         if ((smp_processor_id() != 0) && (atomic_read(&nmi_active) <= 0))
 251                 return;
 252
 253         switch (nmi_watchdog) {
 254         case NMI_LOCAL_APIC:
 255                 __get_cpu_var(wd_enabled) = 1; /* enable it before to avoid race with handler */
 256                 if (lapic_watchdog_init(nmi_hz) < 0) {
 257                         __get_cpu_var(wd_enabled) = 0;
 258                         return;
 259                 }
 260                 /* FALL THROUGH */
 261         case NMI_IO_APIC:
 262                 __get_cpu_var(wd_enabled) = 1;
 263                 atomic_inc(&nmi_active);
 264         }
 265 }
 266
 267 void stop_apic_nmi_watchdog(void *unused)
 268 {
 269         /* only support LOCAL and IO APICs for now */
 270         if ((nmi_watchdog != NMI_LOCAL_APIC) &&
 271             (nmi_watchdog != NMI_IO_APIC))
 272                 return;
 273         if (__get_cpu_var(wd_enabled) == 0)
 274                 return;
 275         if (nmi_watchdog == NMI_LOCAL_APIC)
 276                 lapic_watchdog_stop();
 277         __get_cpu_var(wd_enabled) = 0;
 278         atomic_dec(&nmi_active);
 279 }
 280
 281 /*
 282  * the best way to detect whether a CPU has a 'hard lockup' problem
 283  * is to check it's local APIC timer IRQ counts. If they are not
 284  * changing then that CPU has some problem.
 285  *
 286  * as these watchdog NMI IRQs are generated on every CPU, we only
 287  * have to check the current processor.
 288  *
 289  * since NMIs don't listen to _any_ locks, we have to be extremely
 290  * careful not to rely on unsafe variables. The printk might lock
 291  * up though, so we have to break up any console locks first ...
 292  * [when there will be more tty-related locks, break them up
 293  *  here too!]
 294  */
 295
 296 static unsigned int
 297         last_irq_sums [NR_CPUS],
 298         alert_counter [NR_CPUS];
 299
 300 void touch_nmi_watchdog(void)
 301 {
 302         if (nmi_watchdog > 0) {
 303                 unsigned cpu;
 304
 305                 /*
 306                  * Just reset the alert counters, (other CPUs might be
 307                  * spinning on locks we hold):
 308                  */
 309                 for_each_present_cpu(cpu) {
 310                         if (alert_counter[cpu])
 311                                 alert_counter[cpu] = 0;
 312                 }
 313         }
 314
 315         /*
 316          * Tickle the softlockup detector too:
 317          */
 318         touch_softlockup_watchdog();
 319 }
 320 EXPORT_SYMBOL(touch_nmi_watchdog);
 321
 322 extern void die_nmi(struct pt_regs *, const char *msg);
 323
 324 __kprobes int nmi_watchdog_tick(struct pt_regs * regs, unsigned reason)
 325 {
 326
 327         /*
 328          * Since current_thread_info()-> is always on the stack, and we
 329          * always switch the stack NMI-atomically, it's safe to use
 330          * smp_processor_id().
 331          */
 332         unsigned int sum;
 333         int touched = 0;
 334         int cpu = smp_processor_id();
 335         int rc = 0;
 336
 337         /* check for other users first */
 338         if (notify_die(DIE_NMI, "nmi", regs, reason, 2, SIGINT)
 339                         == NOTIFY_STOP) {
 340                 rc = 1;
 341                 touched = 1;
 342         }
 343
 344         if (cpu_isset(cpu, backtrace_mask)) {
 345                 static DEFINE_SPINLOCK(lock);   /* Serialise the printks */
 346
 347                 spin_lock(&lock);
 348                 printk("NMI backtrace for cpu %d\n", cpu);
 349                 dump_stack();
 350                 spin_unlock(&lock);
 351                 cpu_clear(cpu, backtrace_mask);
 352         }
 353
 354         /*
 355          * Take the local apic timer and PIT/HPET into account. We don't
 356          * know which one is active, when we have highres/dyntick on
 357          */
 358         sum = per_cpu(irq_stat, cpu).apic_timer_irqs +
 359                 per_cpu(irq_stat, cpu).irq0_irqs;
 360
 361         /* if the none of the timers isn't firing, this cpu isn't doing much */
 362         if (!touched && last_irq_sums[cpu] == sum) {
 363                 /*
 364                  * Ayiee, looks like this CPU is stuck ...
 365                  * wait a few IRQs (5 seconds) before doing the oops ...
 366                  */
 367                 alert_counter[cpu]++;
 368                 if (alert_counter[cpu] == 5*nmi_hz)
 369                         /*
 370                          * die_nmi will return ONLY if NOTIFY_STOP happens..
 371                          */
 372                         die_nmi(regs, "BUG: NMI Watchdog detected LOCKUP");
 373         } else {
 374                 last_irq_sums[cpu] = sum;
 375                 alert_counter[cpu] = 0;
 376         }
 377         /* see if the nmi watchdog went off */
 378         if (!__get_cpu_var(wd_enabled))
 379                 return rc;
 380         switch (nmi_watchdog) {
 381         case NMI_LOCAL_APIC:
 382                 rc |= lapic_wd_event(nmi_hz);
 383                 break;
 384         case NMI_IO_APIC:
 385                 /* don't know how to accurately check for this.
 386                  * just assume it was a watchdog timer interrupt
 387                  * This matches the old behaviour.
 388                  */
 389                 rc = 1;
 390                 break;
 391         }
 392         return rc;
 393 }
 394
 395 #ifdef CONFIG_SYSCTL
 396
 397 static int unknown_nmi_panic_callback(struct pt_regs *regs, int cpu)
 398 {
 399         unsigned char reason = get_nmi_reason();
 400         char buf[64];
 401
 402         sprintf(buf, "NMI received for unknown reason %02x\n", reason);
 403         die_nmi(regs, buf);
 404         return 0;
 405 }
 406
 407 /*
 408  * proc handler for /proc/sys/kernel/nmi
 409  */
 410 int proc_nmi_enabled(struct ctl_table *table, int write, struct file *file,
 411                         void __user *buffer, size_t *length, loff_t *ppos)
 412 {
 413         int old_state;
 414
 415         nmi_watchdog_enabled = (atomic_read(&nmi_active) > 0) ? 1 : 0;
 416         old_state = nmi_watchdog_enabled;
 417         proc_dointvec(table, write, file, buffer, length, ppos);
 418         if (!!old_state == !!nmi_watchdog_enabled)
 419                 return 0;
 420
 421         if (atomic_read(&nmi_active) < 0 || nmi_watchdog == NMI_DISABLED) {
 422                 printk( KERN_WARNING "NMI watchdog is permanently disabled\n");
 423                 return -EIO;
 424         }
 425
 426         if (nmi_watchdog == NMI_DEFAULT) {
 427                 if (lapic_watchdog_ok())
 428                         nmi_watchdog = NMI_LOCAL_APIC;
 429                 else
 430                         nmi_watchdog = NMI_IO_APIC;
 431         }
 432
 433         if (nmi_watchdog == NMI_LOCAL_APIC) {
 434                 if (nmi_watchdog_enabled)
 435                         enable_lapic_nmi_watchdog();
 436                 else
 437                         disable_lapic_nmi_watchdog();
 438         } else {
 439                 printk( KERN_WARNING
 440                         "NMI watchdog doesn't know what hardware to touch\n");
 441                 return -EIO;
 442         }
 443         return 0;
 444 }
 445
 446 #endif
 447
 448 int do_nmi_callback(struct pt_regs *regs, int cpu)
 449 {
 450 #ifdef CONFIG_SYSCTL
 451         if (unknown_nmi_panic)
 452                 return unknown_nmi_panic_callback(regs, cpu);
 453 #endif
 454         return 0;
 455 }
 456
 457 void __trigger_all_cpu_backtrace(void)
 458 {
 459         int i;
 460
 461         backtrace_mask = cpu_online_map;
 462         /* Wait for up to 10 seconds for all CPUs to do the backtrace */
 463         for (i = 0; i < 10 * 1000; i++) {
 464                 if (cpus_empty(backtrace_mask))
 465                         break;
 466                 mdelay(1);
 467         }
 468 }
 469
 470 EXPORT_SYMBOL(nmi_active);
 471 EXPORT_SYMBOL(nmi_watchdog);