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 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 *counts;
  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         counts = kmalloc(NR_CPUS * sizeof(int), GFP_KERNEL);
  93         if (!counts)
  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                 counts[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 - counts[cpu] <= 5) {
 112                         printk("CPU#%d: NMI appears to be stuck (%d->%d)!\n",
 113                                cpu,
 114                                counts[cpu],
 115                                cpu_pda(cpu)->__nmi_count);
 116                         per_cpu(wd_enabled, cpu) = 0;
 117                         atomic_dec(&nmi_active);
 118                 }
 119         }
 120         if (!atomic_read(&nmi_active)) {
 121                 kfree(counts);
 122                 atomic_set(&nmi_active, -1);
 123                 endflag = 1;
 124                 return -1;
 125         }
 126         endflag = 1;
 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(counts);
 135         return 0;
 136 }
 137
 138 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
 162 static void __acpi_nmi_disable(void *__unused)
 163 {
 164         apic_write(APIC_LVT0, APIC_DM_NMI | APIC_LVT_MASKED);
 165 }
 166
 167 /*
 168  * Disable timer based NMIs on all CPUs:
 169  */
 170 void acpi_nmi_disable(void)
 171 {
 172         if (atomic_read(&nmi_active) && nmi_watchdog == NMI_IO_APIC)
 173                 on_each_cpu(__acpi_nmi_disable, NULL, 0, 1);
 174 }
 175
 176 static void __acpi_nmi_enable(void *__unused)
 177 {
 178         apic_write(APIC_LVT0, APIC_DM_NMI);
 179 }
 180
 181 /*
 182  * Enable timer based NMIs on all CPUs:
 183  */
 184 void acpi_nmi_enable(void)
 185 {
 186         if (atomic_read(&nmi_active) && nmi_watchdog == NMI_IO_APIC)
 187                 on_each_cpu(__acpi_nmi_enable, NULL, 0, 1);
 188 }
 189 #ifdef CONFIG_PM
 190
 191 static int nmi_pm_active; /* nmi_active before suspend */
 192
 193 static int lapic_nmi_suspend(struct sys_device *dev, pm_message_t state)
 194 {
 195         /* only CPU0 goes here, other CPUs should be offline */
 196         nmi_pm_active = atomic_read(&nmi_active);
 197         stop_apic_nmi_watchdog(NULL);
 198         BUG_ON(atomic_read(&nmi_active) != 0);
 199         return 0;
 200 }
 201
 202 static int lapic_nmi_resume(struct sys_device *dev)
 203 {
 204         /* only CPU0 goes here, other CPUs should be offline */
 205         if (nmi_pm_active > 0) {
 206                 setup_apic_nmi_watchdog(NULL);
 207                 touch_nmi_watchdog();
 208         }
 209         return 0;
 210 }
 211
 212 static struct sysdev_class nmi_sysclass = {
 213         set_kset_name("lapic_nmi"),
 214         .resume         = lapic_nmi_resume,
 215         .suspend        = lapic_nmi_suspend,
 216 };
 217
 218 static struct sys_device device_lapic_nmi = {
 219         .id             = 0,
 220         .cls    = &nmi_sysclass,
 221 };
 222
 223 static int __init init_lapic_nmi_sysfs(void)
 224 {
 225         int error;
 226
 227         /* should really be a BUG_ON but b/c this is an
 228          * init call, it just doesn't work.  -dcz
 229          */
 230         if (nmi_watchdog != NMI_LOCAL_APIC)
 231                 return 0;
 232
 233         if ( atomic_read(&nmi_active) < 0 )
 234                 return 0;
 235
 236         error = sysdev_class_register(&nmi_sysclass);
 237         if (!error)
 238                 error = sysdev_register(&device_lapic_nmi);
 239         return error;
 240 }
 241 /* must come after the local APIC's device_initcall() */
 242 late_initcall(init_lapic_nmi_sysfs);
 243
 244 #endif  /* CONFIG_PM */
 245
 246 void setup_apic_nmi_watchdog(void *unused)
 247 {
 248         if (__get_cpu_var(wd_enabled) == 1)
 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
 316 int __kprobes nmi_watchdog_tick(struct pt_regs * regs, unsigned reason)
 317 {
 318         int sum;
 319         int touched = 0;
 320         int cpu = smp_processor_id();
 321         int rc = 0;
 322
 323         /* check for other users first */
 324         if (notify_die(DIE_NMI, "nmi", regs, reason, 2, SIGINT)
 325                         == NOTIFY_STOP) {
 326                 rc = 1;
 327                 touched = 1;
 328         }
 329
 330         sum = read_pda(apic_timer_irqs) + read_pda(irq0_irqs);
 331         if (__get_cpu_var(nmi_touch)) {
 332                 __get_cpu_var(nmi_touch) = 0;
 333                 touched = 1;
 334         }
 335
 336         if (cpu_isset(cpu, backtrace_mask)) {
 337                 static DEFINE_SPINLOCK(lock);   /* Serialise the printks */
 338
 339                 spin_lock(&lock);
 340                 printk("NMI backtrace for cpu %d\n", cpu);
 341                 dump_stack();
 342                 spin_unlock(&lock);
 343                 cpu_clear(cpu, backtrace_mask);
 344         }
 345
 346 #ifdef CONFIG_X86_MCE
 347         /* Could check oops_in_progress here too, but it's safer
 348            not too */
 349         if (atomic_read(&mce_entry) > 0)
 350                 touched = 1;
 351 #endif
 352         /* if the apic timer isn't firing, this cpu isn't doing much */
 353         if (!touched && __get_cpu_var(last_irq_sum) == sum) {
 354                 /*
 355                  * Ayiee, looks like this CPU is stuck ...
 356                  * wait a few IRQs (5 seconds) before doing the oops ...
 357                  */
 358                 local_inc(&__get_cpu_var(alert_counter));
 359                 if (local_read(&__get_cpu_var(alert_counter)) == 5*nmi_hz)
 360                         die_nmi("NMI Watchdog detected LOCKUP on CPU %d\n", regs,
 361                                 panic_on_timeout);
 362         } else {
 363                 __get_cpu_var(last_irq_sum) = sum;
 364                 local_set(&__get_cpu_var(alert_counter), 0);
 365         }
 366
 367         /* see if the nmi watchdog went off */
 368         if (!__get_cpu_var(wd_enabled))
 369                 return rc;
 370         switch (nmi_watchdog) {
 371         case NMI_LOCAL_APIC:
 372                 rc |= lapic_wd_event(nmi_hz);
 373                 break;
 374         case NMI_IO_APIC:
 375                 /* don't know how to accurately check for this.
 376                  * just assume it was a watchdog timer interrupt
 377                  * This matches the old behaviour.
 378                  */
 379                 rc = 1;
 380                 break;
 381         }
 382         return rc;
 383 }
 384
 385 static unsigned ignore_nmis;
 386
 387 asmlinkage __kprobes void do_nmi(struct pt_regs * regs, long error_code)
 388 {
 389         nmi_enter();
 390         add_pda(__nmi_count,1);
 391         if (!ignore_nmis)
 392                 default_do_nmi(regs);
 393         nmi_exit();
 394 }
 395
 396 int do_nmi_callback(struct pt_regs * regs, int cpu)
 397 {
 398 #ifdef CONFIG_SYSCTL
 399         if (unknown_nmi_panic)
 400                 return unknown_nmi_panic_callback(regs, cpu);
 401 #endif
 402         return 0;
 403 }
 404
 405 void stop_nmi(void)
 406 {
 407         acpi_nmi_disable();
 408         ignore_nmis++;
 409 }
 410
 411 void restart_nmi(void)
 412 {
 413         ignore_nmis--;
 414         acpi_nmi_enable();
 415 }
 416
 417 #ifdef CONFIG_SYSCTL
 418
 419 static int unknown_nmi_panic_callback(struct pt_regs *regs, int cpu)
 420 {
 421         unsigned char reason = get_nmi_reason();
 422         char buf[64];
 423
 424         sprintf(buf, "NMI received for unknown reason %02x\n", reason);
 425         die_nmi(buf, regs, 1);  /* Always panic here */
 426         return 0;
 427 }
 428
 429 /*
 430  * proc handler for /proc/sys/kernel/nmi
 431  */
 432 int proc_nmi_enabled(struct ctl_table *table, int write, struct file *file,
 433                         void __user *buffer, size_t *length, loff_t *ppos)
 434 {
 435         int old_state;
 436
 437         nmi_watchdog_enabled = (atomic_read(&nmi_active) > 0) ? 1 : 0;
 438         old_state = nmi_watchdog_enabled;
 439         proc_dointvec(table, write, file, buffer, length, ppos);
 440         if (!!old_state == !!nmi_watchdog_enabled)
 441                 return 0;
 442
 443         if (atomic_read(&nmi_active) < 0 || nmi_watchdog == NMI_DISABLED) {
 444                 printk( KERN_WARNING "NMI watchdog is permanently disabled\n");
 445                 return -EIO;
 446         }
 447
 448         /* if nmi_watchdog is not set yet, then set it */
 449         nmi_watchdog_default();
 450
 451         if (nmi_watchdog == NMI_LOCAL_APIC) {
 452                 if (nmi_watchdog_enabled)
 453                         enable_lapic_nmi_watchdog();
 454                 else
 455                         disable_lapic_nmi_watchdog();
 456         } else {
 457                 printk( KERN_WARNING
 458                         "NMI watchdog doesn't know what hardware to touch\n");
 459                 return -EIO;
 460         }
 461         return 0;
 462 }
 463
 464 #endif
 465
 466 void __trigger_all_cpu_backtrace(void)
 467 {
 468         int i;
 469
 470         backtrace_mask = cpu_online_map;
 471         /* Wait for up to 10 seconds for all CPUs to do the backtrace */
 472         for (i = 0; i < 10 * 1000; i++) {
 473                 if (cpus_empty(backtrace_mask))
 474                         break;
 475                 mdelay(1);
 476         }
 477 }
 478
 479 EXPORT_SYMBOL(nmi_active);
 480 EXPORT_SYMBOL(nmi_watchdog);
 481 EXPORT_SYMBOL(touch_nmi_watchdog);