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