arch/x86/kernel/nmi_64.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  *  Pavel Machek and
  10  *  Mikael Pettersson   : PM converted to driver model. Disable/enable API.
  11  */
  12
  13 #include <linux/nmi.h>
  14 #include <linux/mm.h>
  15 #include <linux/delay.h>
  16 #include <linux/interrupt.h>
  17 #include <linux/module.h>
  18 #include <linux/sysdev.h>
  19 #include <linux/sysctl.h>
  20 #include <linux/kprobes.h>
  21 #include <linux/cpumask.h>
  22 #include <linux/kdebug.h>
  23
  24 #include <asm/smp.h>
  25 #include <asm/nmi.h>
  26 #include <asm/proto.h>
  27 #include <asm/mce.h>
  28
  29 #include <mach_traps.h>
  30
  31 int unknown_nmi_panic;
  32 int nmi_watchdog_enabled;
  33 int panic_on_unrecovered_nmi;
  34
  35 static cpumask_t backtrace_mask = CPU_MASK_NONE;
  36
  37 /* nmi_active:
  38  * >0: the lapic NMI watchdog is active, but can be disabled
  39  * <0: the lapic NMI watchdog has not been set up, and cannot
  40  *     be enabled
  41  *  0: the lapic NMI watchdog is disabled, but can be enabled
  42  */
  43 atomic_t nmi_active = ATOMIC_INIT(0);           /* oprofile uses this */
  44 static int panic_on_timeout;
  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 /* Run after command line and cpu_init init, but before all other checks */
  52 void nmi_watchdog_default(void)
  53 {
  54         if (nmi_watchdog != NMI_DEFAULT)
  55                 return;
  56         nmi_watchdog = NMI_NONE;
  57 }
  58
  59 static int endflag __initdata = 0;
  60
  61 #ifdef CONFIG_SMP
  62 /* The performance counters used by NMI_LOCAL_APIC don't trigger when
  63  * the CPU is idle. To make sure the NMI watchdog really ticks on all
  64  * CPUs during the test make them busy.
  65  */
  66 static __init void nmi_cpu_busy(void *data)
  67 {
  68         local_irq_enable_in_hardirq();
  69         /* Intentionally don't use cpu_relax here. This is
  70            to make sure that the performance counter really ticks,
  71            even if there is a simulator or similar that catches the
  72            pause instruction. On a real HT machine this is fine because
  73            all other CPUs are busy with "useless" delay loops and don't
  74            care if they get somewhat less cycles. */
  75         while (endflag == 0)
  76                 mb();
  77 }
  78 #endif
  79
  80 int __init check_nmi_watchdog(void)
  81 {
  82         int *prev_nmi_count;
  83         int cpu;
  84
  85         if ((nmi_watchdog == NMI_NONE) || (nmi_watchdog == NMI_DISABLED))
  86                 return 0;
  87
  88         if (!atomic_read(&nmi_active))
  89                 return 0;
  90
  91         prev_nmi_count = kmalloc(NR_CPUS * sizeof(int), GFP_KERNEL);
  92         if (!prev_nmi_count)
  93                 return -1;
  94
  95         printk(KERN_INFO "Testing NMI watchdog ... ");
  96
  97 #ifdef CONFIG_SMP
  98         if (nmi_watchdog == NMI_LOCAL_APIC)
  99                 smp_call_function(nmi_cpu_busy, (void *)&endflag, 0, 0);
 100 #endif
 101
 102         for (cpu = 0; cpu < NR_CPUS; cpu++)
 103                 prev_nmi_count[cpu] = cpu_pda(cpu)->__nmi_count;
 104         local_irq_enable();
 105         mdelay((20*1000)/nmi_hz); // wait 20 ticks
 106
 107         for_each_online_cpu(cpu) {
 108                 if (!per_cpu(wd_enabled, cpu))
 109                         continue;
 110                 if (cpu_pda(cpu)->__nmi_count - prev_nmi_count[cpu] <= 5) {
 111                         printk(KERN_WARNING "WARNING: CPU#%d: NMI "
 112                                "appears to be stuck (%d->%d)!\n",
 113                                 cpu,
 114                                 prev_nmi_count[cpu],
 115                                 cpu_pda(cpu)->__nmi_count);
 116                         per_cpu(wd_enabled, cpu) = 0;
 117                         atomic_dec(&nmi_active);
 118                 }
 119         }
 120         endflag = 1;
 121         if (!atomic_read(&nmi_active)) {
 122                 kfree(prev_nmi_count);
 123                 atomic_set(&nmi_active, -1);
 124                 return -1;
 125         }
 126         printk("OK.\n");
 127
 128         /* now that we know it works we can reduce NMI frequency to
 129            something more reasonable; makes a difference in some configs */
 130         if (nmi_watchdog == NMI_LOCAL_APIC)
 131                 nmi_hz = lapic_adjust_nmi_hz(1);
 132
 133         kfree(prev_nmi_count);
 134         return 0;
 135 }
 136
 137 static int __init setup_nmi_watchdog(char *str)
 138 {
 139         int nmi;
 140
 141         if (!strncmp(str,"panic",5)) {
 142                 panic_on_timeout = 1;
 143                 str = strchr(str, ',');
 144                 if (!str)
 145                         return 1;
 146                 ++str;
 147         }
 148
 149         get_option(&str, &nmi);
 150
 151         if ((nmi >= NMI_INVALID) || (nmi < NMI_NONE))
 152                 return 0;
 153
 154         nmi_watchdog = nmi;
 155         return 1;
 156 }
 157
 158 __setup("nmi_watchdog=", setup_nmi_watchdog);
 159
 160 #ifdef CONFIG_PM
 161
 162 static int nmi_pm_active; /* nmi_active before suspend */
 163
 164 static int lapic_nmi_suspend(struct sys_device *dev, pm_message_t state)
 165 {
 166         /* only CPU0 goes here, other CPUs should be offline */
 167         nmi_pm_active = atomic_read(&nmi_active);
 168         stop_apic_nmi_watchdog(NULL);
 169         BUG_ON(atomic_read(&nmi_active) != 0);
 170         return 0;
 171 }
 172
 173 static int lapic_nmi_resume(struct sys_device *dev)
 174 {
 175         /* only CPU0 goes here, other CPUs should be offline */
 176         if (nmi_pm_active > 0) {
 177                 setup_apic_nmi_watchdog(NULL);
 178                 touch_nmi_watchdog();
 179         }
 180         return 0;
 181 }
 182
 183 static struct sysdev_class nmi_sysclass = {
 184         .name           = "lapic_nmi",
 185         .resume         = lapic_nmi_resume,
 186         .suspend        = lapic_nmi_suspend,
 187 };
 188
 189 static struct sys_device device_lapic_nmi = {
 190         .id     = 0,
 191         .cls    = &nmi_sysclass,
 192 };
 193
 194 static int __init init_lapic_nmi_sysfs(void)
 195 {
 196         int error;
 197
 198         /* should really be a BUG_ON but b/c this is an
 199          * init call, it just doesn't work.  -dcz
 200          */
 201         if (nmi_watchdog != NMI_LOCAL_APIC)
 202                 return 0;
 203
 204         if (atomic_read(&nmi_active) < 0)
 205                 return 0;
 206
 207         error = sysdev_class_register(&nmi_sysclass);
 208         if (!error)
 209                 error = sysdev_register(&device_lapic_nmi);
 210         return error;
 211 }
 212 /* must come after the local APIC's device_initcall() */
 213 late_initcall(init_lapic_nmi_sysfs);
 214
 215 #endif  /* CONFIG_PM */
 216
 217 static void __acpi_nmi_enable(void *__unused)
 218 {
 219         apic_write(APIC_LVT0, APIC_DM_NMI);
 220 }
 221
 222 /*
 223  * Enable timer based NMIs on all CPUs:
 224  */
 225 void acpi_nmi_enable(void)
 226 {
 227         if (atomic_read(&nmi_active) && nmi_watchdog == NMI_IO_APIC)
 228                 on_each_cpu(__acpi_nmi_enable, NULL, 0, 1);
 229 }
 230
 231 static void __acpi_nmi_disable(void *__unused)
 232 {
 233         apic_write(APIC_LVT0, APIC_DM_NMI | APIC_LVT_MASKED);
 234 }
 235
 236 /*
 237  * Disable timer based NMIs on all CPUs:
 238  */
 239 void acpi_nmi_disable(void)
 240 {
 241         if (atomic_read(&nmi_active) && nmi_watchdog == NMI_IO_APIC)
 242                 on_each_cpu(__acpi_nmi_disable, NULL, 0, 1);
 243 }
 244
 245 void setup_apic_nmi_watchdog(void *unused)
 246 {
 247         if (__get_cpu_var(wd_enabled))
 248                 return;
 249
 250         /* cheap hack to support suspend/resume */
 251         /* if cpu0 is not active neither should the other cpus */
 252         if ((smp_processor_id() != 0) && (atomic_read(&nmi_active) <= 0))
 253                 return;
 254
 255         switch (nmi_watchdog) {
 256         case NMI_LOCAL_APIC:
 257                 __get_cpu_var(wd_enabled) = 1;
 258                 if (lapic_watchdog_init(nmi_hz) < 0) {
 259                         __get_cpu_var(wd_enabled) = 0;
 260                         return;
 261                 }
 262                 /* FALL THROUGH */
 263         case NMI_IO_APIC:
 264                 __get_cpu_var(wd_enabled) = 1;
 265                 atomic_inc(&nmi_active);
 266         }
 267 }
 268
 269 void stop_apic_nmi_watchdog(void *unused)
 270 {
 271         /* only support LOCAL and IO APICs for now */
 272         if ((nmi_watchdog != NMI_LOCAL_APIC) &&
 273             (nmi_watchdog != NMI_IO_APIC))
 274                 return;
 275         if (__get_cpu_var(wd_enabled) == 0)
 276                 return;
 277         if (nmi_watchdog == NMI_LOCAL_APIC)
 278                 lapic_watchdog_stop();
 279         __get_cpu_var(wd_enabled) = 0;
 280         atomic_dec(&nmi_active);
 281 }
 282
 283 /*
 284  * the best way to detect whether a CPU has a 'hard lockup' problem
 285  * is to check it's local APIC timer IRQ counts. If they are not
 286  * changing then that CPU has some problem.
 287  *
 288  * as these watchdog NMI IRQs are generated on every CPU, we only
 289  * have to check the current processor.
 290  */
 291
 292 static DEFINE_PER_CPU(unsigned, last_irq_sum);
 293 static DEFINE_PER_CPU(local_t, alert_counter);
 294 static DEFINE_PER_CPU(int, nmi_touch);
 295
 296 void touch_nmi_watchdog(void)
 297 {
 298         if (nmi_watchdog > 0) {
 299                 unsigned cpu;
 300
 301                 /*
 302                  * Tell other CPUs to reset their alert counters. We cannot
 303                  * do it ourselves because the alert count increase is not
 304                  * atomic.
 305                  */
 306                 for_each_present_cpu(cpu) {
 307                         if (per_cpu(nmi_touch, cpu) != 1)
 308                                 per_cpu(nmi_touch, cpu) = 1;
 309                 }
 310         }
 311
 312         touch_softlockup_watchdog();
 313 }
 314 EXPORT_SYMBOL(touch_nmi_watchdog);
 315
 316 notrace __kprobes int
 317 nmi_watchdog_tick(struct pt_regs *regs, unsigned reason)
 318 {
 319         int sum;
 320         int touched = 0;
 321         int cpu = smp_processor_id();
 322         int rc = 0;
 323
 324         /* check for other users first */
 325         if (notify_die(DIE_NMI, "nmi", regs, reason, 2, SIGINT)
 326                         == NOTIFY_STOP) {
 327                 rc = 1;
 328                 touched = 1;
 329         }
 330
 331         sum = read_pda(apic_timer_irqs) + read_pda(irq0_irqs);
 332         if (__get_cpu_var(nmi_touch)) {
 333                 __get_cpu_var(nmi_touch) = 0;
 334                 touched = 1;
 335         }
 336
 337         if (cpu_isset(cpu, backtrace_mask)) {
 338                 static DEFINE_SPINLOCK(lock);   /* Serialise the printks */
 339
 340                 spin_lock(&lock);
 341                 printk("NMI backtrace for cpu %d\n", cpu);
 342                 dump_stack();
 343                 spin_unlock(&lock);
 344                 cpu_clear(cpu, backtrace_mask);
 345         }
 346
 347 #ifdef CONFIG_X86_MCE
 348         /* Could check oops_in_progress here too, but it's safer
 349            not too */
 350         if (atomic_read(&mce_entry) > 0)
 351                 touched = 1;
 352 #endif
 353         /* if the apic timer isn't firing, this cpu isn't doing much */
 354         if (!touched && __get_cpu_var(last_irq_sum) == sum) {
 355                 /*
 356                  * Ayiee, looks like this CPU is stuck ...
 357                  * wait a few IRQs (5 seconds) before doing the oops ...
 358                  */
 359                 local_inc(&__get_cpu_var(alert_counter));
 360                 if (local_read(&__get_cpu_var(alert_counter)) == 5*nmi_hz)
 361                         die_nmi("NMI Watchdog detected LOCKUP on CPU %d\n", regs,
 362                                 panic_on_timeout);
 363         } else {
 364                 __get_cpu_var(last_irq_sum) = sum;
 365                 local_set(&__get_cpu_var(alert_counter), 0);
 366         }
 367
 368         /* see if the nmi watchdog went off */
 369         if (!__get_cpu_var(wd_enabled))
 370                 return rc;
 371         switch (nmi_watchdog) {
 372         case NMI_LOCAL_APIC:
 373                 rc |= lapic_wd_event(nmi_hz);
 374                 break;
 375         case NMI_IO_APIC:
 376                 /* don't know how to accurately check for this.
 377                  * just assume it was a watchdog timer interrupt
 378                  * This matches the old behaviour.
 379                  */
 380                 rc = 1;
 381                 break;
 382         }
 383         return rc;
 384 }
 385
 386 static unsigned ignore_nmis;
 387
 388 asmlinkage notrace __kprobes void
 389 do_nmi(struct pt_regs *regs, long error_code)
 390 {
 391         nmi_enter();
 392         add_pda(__nmi_count,1);
 393         if (!ignore_nmis)
 394                 default_do_nmi(regs);
 395         nmi_exit();
 396 }
 397
 398 void stop_nmi(void)
 399 {
 400         acpi_nmi_disable();
 401         ignore_nmis++;
 402 }
 403
 404 void restart_nmi(void)
 405 {
 406         ignore_nmis--;
 407         acpi_nmi_enable();
 408 }
 409
 410 #ifdef CONFIG_SYSCTL
 411
 412 static int unknown_nmi_panic_callback(struct pt_regs *regs, int cpu)
 413 {
 414         unsigned char reason = get_nmi_reason();
 415         char buf[64];
 416
 417         sprintf(buf, "NMI received for unknown reason %02x\n", reason);
 418         die_nmi(buf, regs, 1);  /* Always panic here */
 419         return 0;
 420 }
 421
 422 /*
 423  * proc handler for /proc/sys/kernel/nmi
 424  */
 425 int proc_nmi_enabled(struct ctl_table *table, int write, struct file *file,
 426                         void __user *buffer, size_t *length, loff_t *ppos)
 427 {
 428         int old_state;
 429
 430         nmi_watchdog_enabled = (atomic_read(&nmi_active) > 0) ? 1 : 0;
 431         old_state = nmi_watchdog_enabled;
 432         proc_dointvec(table, write, file, buffer, length, ppos);
 433         if (!!old_state == !!nmi_watchdog_enabled)
 434                 return 0;
 435
 436         if (atomic_read(&nmi_active) < 0 || nmi_watchdog == NMI_DISABLED) {
 437                 printk( KERN_WARNING "NMI watchdog is permanently disabled\n");
 438                 return -EIO;
 439         }
 440
 441         /* if nmi_watchdog is not set yet, then set it */
 442         nmi_watchdog_default();
 443
 444         if (nmi_watchdog == NMI_LOCAL_APIC) {
 445                 if (nmi_watchdog_enabled)
 446                         enable_lapic_nmi_watchdog();
 447                 else
 448                         disable_lapic_nmi_watchdog();
 449         } else {
 450                 printk( KERN_WARNING
 451                         "NMI watchdog doesn't know what hardware to touch\n");
 452                 return -EIO;
 453         }
 454         return 0;
 455 }
 456
 457 #endif
 458
 459 int do_nmi_callback(struct pt_regs *regs, int cpu)
 460 {
 461 #ifdef CONFIG_SYSCTL
 462         if (unknown_nmi_panic)
 463                 return unknown_nmi_panic_callback(regs, cpu);
 464 #endif
 465         return 0;
 466 }
 467
 468 void __trigger_all_cpu_backtrace(void)
 469 {
 470         int i;
 471
 472         backtrace_mask = cpu_online_map;
 473         /* Wait for up to 10 seconds for all CPUs to do the backtrace */
 474         for (i = 0; i < 10 * 1000; i++) {
 475                 if (cpus_empty(backtrace_mask))
 476                         break;
 477                 mdelay(1);
 478         }
 479 }
 480
 481 EXPORT_SYMBOL(nmi_active);
 482 EXPORT_SYMBOL(nmi_watchdog);