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/i8259.h>
  25 #include <asm/io_apic.h>
  26 #include <asm/smp.h>
  27 #include <asm/nmi.h>
  28 #include <asm/proto.h>
  29 #include <asm/mce.h>
  30
  31 #include <mach_traps.h>
  32
  33 int unknown_nmi_panic;
  34 int nmi_watchdog_enabled;
  35 int panic_on_unrecovered_nmi;
  36
  37 static cpumask_t backtrace_mask = CPU_MASK_NONE;
  38
  39 /* nmi_active:
  40  * >0: the lapic NMI watchdog is active, but can be disabled
  41  * <0: the lapic NMI watchdog has not been set up, and cannot
  42  *     be enabled
  43  *  0: the lapic NMI watchdog is disabled, but can be enabled
  44  */
  45 atomic_t nmi_active = ATOMIC_INIT(0);           /* oprofile uses this */
  46 static int panic_on_timeout;
  47
  48 unsigned int nmi_watchdog = NMI_DEFAULT;
  49 static unsigned int nmi_hz = HZ;
  50
  51 static DEFINE_PER_CPU(short, wd_enabled);
  52
  53 /* Run after command line and cpu_init init, but before all other checks */
  54 void nmi_watchdog_default(void)
  55 {
  56         if (nmi_watchdog != NMI_DEFAULT)
  57                 return;
  58         nmi_watchdog = NMI_NONE;
  59 }
  60
  61 static int endflag __initdata = 0;
  62
  63 #ifdef CONFIG_SMP
  64 /* The performance counters used by NMI_LOCAL_APIC don't trigger when
  65  * the CPU is idle. To make sure the NMI watchdog really ticks on all
  66  * CPUs during the test make them busy.
  67  */
  68 static __init void nmi_cpu_busy(void *data)
  69 {
  70         local_irq_enable_in_hardirq();
  71         /* Intentionally don't use cpu_relax here. This is
  72            to make sure that the performance counter really ticks,
  73            even if there is a simulator or similar that catches the
  74            pause instruction. On a real HT machine this is fine because
  75            all other CPUs are busy with "useless" delay loops and don't
  76            care if they get somewhat less cycles. */
  77         while (endflag == 0)
  78                 mb();
  79 }
  80 #endif
  81
  82 int __init check_nmi_watchdog(void)
  83 {
  84         int *prev_nmi_count;
  85         int cpu;
  86
  87         if ((nmi_watchdog == NMI_NONE) || (nmi_watchdog == NMI_DISABLED))
  88                 return 0;
  89
  90         if (!atomic_read(&nmi_active))
  91                 return 0;
  92
  93         prev_nmi_count = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL);
  94         if (!prev_nmi_count)
  95                 goto error;
  96
  97         printk(KERN_INFO "Testing NMI watchdog ... ");
  98
  99 #ifdef CONFIG_SMP
 100         if (nmi_watchdog == NMI_LOCAL_APIC)
 101                 smp_call_function(nmi_cpu_busy, (void *)&endflag, 0, 0);
 102 #endif
 103
 104         for (cpu = 0; cpu < nr_cpu_ids; cpu++)
 105                 prev_nmi_count[cpu] = cpu_pda(cpu)->__nmi_count;
 106         local_irq_enable();
 107         mdelay((20*1000)/nmi_hz); // wait 20 ticks
 108
 109         for_each_online_cpu(cpu) {
 110                 if (!per_cpu(wd_enabled, cpu))
 111                         continue;
 112                 if (cpu_pda(cpu)->__nmi_count - prev_nmi_count[cpu] <= 5) {
 113                         printk(KERN_WARNING "WARNING: CPU#%d: NMI "
 114                                "appears to be stuck (%d->%d)!\n",
 115                                 cpu,
 116                                 prev_nmi_count[cpu],
 117                                 cpu_pda(cpu)->__nmi_count);
 118                         per_cpu(wd_enabled, cpu) = 0;
 119                         atomic_dec(&nmi_active);
 120                 }
 121         }
 122         endflag = 1;
 123         if (!atomic_read(&nmi_active)) {
 124                 kfree(prev_nmi_count);
 125                 atomic_set(&nmi_active, -1);
 126                 goto error;
 127         }
 128         printk("OK.\n");
 129
 130         /* now that we know it works we can reduce NMI frequency to
 131            something more reasonable; makes a difference in some configs */
 132         if (nmi_watchdog == NMI_LOCAL_APIC)
 133                 nmi_hz = lapic_adjust_nmi_hz(1);
 134
 135         kfree(prev_nmi_count);
 136         return 0;
 137 error:
 138         if (nmi_watchdog == NMI_IO_APIC && !timer_through_8259)
 139                 disable_8259A_irq(0);
 140
 141         return -1;
 142 }
 143
 144 static int __init setup_nmi_watchdog(char *str)
 145 {
 146         int nmi;
 147
 148         if (!strncmp(str,"panic",5)) {
 149                 panic_on_timeout = 1;
 150                 str = strchr(str, ',');
 151                 if (!str)
 152                         return 1;
 153                 ++str;
 154         }
 155
 156         get_option(&str, &nmi);
 157
 158         if ((nmi >= NMI_INVALID) || (nmi < NMI_NONE))
 159                 return 0;
 160
 161         nmi_watchdog = nmi;
 162         return 1;
 163 }
 164
 165 __setup("nmi_watchdog=", setup_nmi_watchdog);
 166
 167 #ifdef CONFIG_PM
 168
 169 static int nmi_pm_active; /* nmi_active before suspend */
 170
 171 static int lapic_nmi_suspend(struct sys_device *dev, pm_message_t state)
 172 {
 173         /* only CPU0 goes here, other CPUs should be offline */
 174         nmi_pm_active = atomic_read(&nmi_active);
 175         stop_apic_nmi_watchdog(NULL);
 176         BUG_ON(atomic_read(&nmi_active) != 0);
 177         return 0;
 178 }
 179
 180 static int lapic_nmi_resume(struct sys_device *dev)
 181 {
 182         /* only CPU0 goes here, other CPUs should be offline */
 183         if (nmi_pm_active > 0) {
 184                 setup_apic_nmi_watchdog(NULL);
 185                 touch_nmi_watchdog();
 186         }
 187         return 0;
 188 }
 189
 190 static struct sysdev_class nmi_sysclass = {
 191         .name           = "lapic_nmi",
 192         .resume         = lapic_nmi_resume,
 193         .suspend        = lapic_nmi_suspend,
 194 };
 195
 196 static struct sys_device device_lapic_nmi = {
 197         .id     = 0,
 198         .cls    = &nmi_sysclass,
 199 };
 200
 201 static int __init init_lapic_nmi_sysfs(void)
 202 {
 203         int error;
 204
 205         /* should really be a BUG_ON but b/c this is an
 206          * init call, it just doesn't work.  -dcz
 207          */
 208         if (nmi_watchdog != NMI_LOCAL_APIC)
 209                 return 0;
 210
 211         if (atomic_read(&nmi_active) < 0)
 212                 return 0;
 213
 214         error = sysdev_class_register(&nmi_sysclass);
 215         if (!error)
 216                 error = sysdev_register(&device_lapic_nmi);
 217         return error;
 218 }
 219 /* must come after the local APIC's device_initcall() */
 220 late_initcall(init_lapic_nmi_sysfs);
 221
 222 #endif  /* CONFIG_PM */
 223
 224 static void __acpi_nmi_enable(void *__unused)
 225 {
 226         apic_write(APIC_LVT0, APIC_DM_NMI);
 227 }
 228
 229 /*
 230  * Enable timer based NMIs on all CPUs:
 231  */
 232 void acpi_nmi_enable(void)
 233 {
 234         if (atomic_read(&nmi_active) && nmi_watchdog == NMI_IO_APIC)
 235                 on_each_cpu(__acpi_nmi_enable, NULL, 0, 1);
 236 }
 237
 238 static void __acpi_nmi_disable(void *__unused)
 239 {
 240         apic_write(APIC_LVT0, APIC_DM_NMI | APIC_LVT_MASKED);
 241 }
 242
 243 /*
 244  * Disable timer based NMIs on all CPUs:
 245  */
 246 void acpi_nmi_disable(void)
 247 {
 248         if (atomic_read(&nmi_active) && nmi_watchdog == NMI_IO_APIC)
 249                 on_each_cpu(__acpi_nmi_disable, NULL, 0, 1);
 250 }
 251
 252 void setup_apic_nmi_watchdog(void *unused)
 253 {
 254         if (__get_cpu_var(wd_enabled))
 255                 return;
 256
 257         /* cheap hack to support suspend/resume */
 258         /* if cpu0 is not active neither should the other cpus */
 259         if ((smp_processor_id() != 0) && (atomic_read(&nmi_active) <= 0))
 260                 return;
 261
 262         switch (nmi_watchdog) {
 263         case NMI_LOCAL_APIC:
 264                 __get_cpu_var(wd_enabled) = 1;
 265                 if (lapic_watchdog_init(nmi_hz) < 0) {
 266                         __get_cpu_var(wd_enabled) = 0;
 267                         return;
 268                 }
 269                 /* FALL THROUGH */
 270         case NMI_IO_APIC:
 271                 __get_cpu_var(wd_enabled) = 1;
 272                 atomic_inc(&nmi_active);
 273         }
 274 }
 275
 276 void stop_apic_nmi_watchdog(void *unused)
 277 {
 278         /* only support LOCAL and IO APICs for now */
 279         if ((nmi_watchdog != NMI_LOCAL_APIC) &&
 280             (nmi_watchdog != NMI_IO_APIC))
 281                 return;
 282         if (__get_cpu_var(wd_enabled) == 0)
 283                 return;
 284         if (nmi_watchdog == NMI_LOCAL_APIC)
 285                 lapic_watchdog_stop();
 286         __get_cpu_var(wd_enabled) = 0;
 287         atomic_dec(&nmi_active);
 288 }
 289
 290 /*
 291  * the best way to detect whether a CPU has a 'hard lockup' problem
 292  * is to check it's local APIC timer IRQ counts. If they are not
 293  * changing then that CPU has some problem.
 294  *
 295  * as these watchdog NMI IRQs are generated on every CPU, we only
 296  * have to check the current processor.
 297  */
 298
 299 static DEFINE_PER_CPU(unsigned, last_irq_sum);
 300 static DEFINE_PER_CPU(local_t, alert_counter);
 301 static DEFINE_PER_CPU(int, nmi_touch);
 302
 303 void touch_nmi_watchdog(void)
 304 {
 305         if (nmi_watchdog > 0) {
 306                 unsigned cpu;
 307
 308                 /*
 309                  * Tell other CPUs to reset their alert counters. We cannot
 310                  * do it ourselves because the alert count increase is not
 311                  * atomic.
 312                  */
 313                 for_each_present_cpu(cpu) {
 314                         if (per_cpu(nmi_touch, cpu) != 1)
 315                                 per_cpu(nmi_touch, cpu) = 1;
 316                 }
 317         }
 318
 319         touch_softlockup_watchdog();
 320 }
 321 EXPORT_SYMBOL(touch_nmi_watchdog);
 322
 323 notrace __kprobes int
 324 nmi_watchdog_tick(struct pt_regs *regs, unsigned reason)
 325 {
 326         int sum;
 327         int touched = 0;
 328         int cpu = smp_processor_id();
 329         int rc = 0;
 330
 331         /* check for other users first */
 332         if (notify_die(DIE_NMI, "nmi", regs, reason, 2, SIGINT)
 333                         == NOTIFY_STOP) {
 334                 rc = 1;
 335                 touched = 1;
 336         }
 337
 338         sum = read_pda(apic_timer_irqs) + read_pda(irq0_irqs);
 339         if (__get_cpu_var(nmi_touch)) {
 340                 __get_cpu_var(nmi_touch) = 0;
 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 #ifdef CONFIG_X86_MCE
 355         /* Could check oops_in_progress here too, but it's safer
 356            not too */
 357         if (atomic_read(&mce_entry) > 0)
 358                 touched = 1;
 359 #endif
 360         /* if the apic timer isn't firing, this cpu isn't doing much */
 361         if (!touched && __get_cpu_var(last_irq_sum) == sum) {
 362                 /*
 363                  * Ayiee, looks like this CPU is stuck ...
 364                  * wait a few IRQs (5 seconds) before doing the oops ...
 365                  */
 366                 local_inc(&__get_cpu_var(alert_counter));
 367                 if (local_read(&__get_cpu_var(alert_counter)) == 5*nmi_hz)
 368                         die_nmi("NMI Watchdog detected LOCKUP on CPU %d\n", regs,
 369                                 panic_on_timeout);
 370         } else {
 371                 __get_cpu_var(last_irq_sum) = sum;
 372                 local_set(&__get_cpu_var(alert_counter), 0);
 373         }
 374
 375         /* see if the nmi watchdog went off */
 376         if (!__get_cpu_var(wd_enabled))
 377                 return rc;
 378         switch (nmi_watchdog) {
 379         case NMI_LOCAL_APIC:
 380                 rc |= lapic_wd_event(nmi_hz);
 381                 break;
 382         case NMI_IO_APIC:
 383                 /* don't know how to accurately check for this.
 384                  * just assume it was a watchdog timer interrupt
 385                  * This matches the old behaviour.
 386                  */
 387                 rc = 1;
 388                 break;
 389         }
 390         return rc;
 391 }
 392
 393 static unsigned ignore_nmis;
 394
 395 asmlinkage notrace __kprobes void
 396 do_nmi(struct pt_regs *regs, long error_code)
 397 {
 398         nmi_enter();
 399         add_pda(__nmi_count,1);
 400         if (!ignore_nmis)
 401                 default_do_nmi(regs);
 402         nmi_exit();
 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 int do_nmi_callback(struct pt_regs *regs, int cpu)
 467 {
 468 #ifdef CONFIG_SYSCTL
 469         if (unknown_nmi_panic)
 470                 return unknown_nmi_panic_callback(regs, cpu);
 471 #endif
 472         return 0;
 473 }
 474
 475 void __trigger_all_cpu_backtrace(void)
 476 {
 477         int i;
 478
 479         backtrace_mask = cpu_online_map;
 480         /* Wait for up to 10 seconds for all CPUs to do the backtrace */
 481         for (i = 0; i < 10 * 1000; i++) {
 482                 if (cpus_empty(backtrace_mask))
 483                         break;
 484                 mdelay(1);
 485         }
 486 }
 487
 488 EXPORT_SYMBOL(nmi_active);
 489 EXPORT_SYMBOL(nmi_watchdog);