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 /* Run after command line and cpu_init init, but before all other checks */
  50 void nmi_watchdog_default(void)
  51 {
  52         if (nmi_watchdog != NMI_DEFAULT)
  53                 return;
  54         nmi_watchdog = NMI_NONE;
  55 }
  56
  57 static int endflag __initdata = 0;
  58
  59 #ifdef CONFIG_SMP
  60 /* The performance counters used by NMI_LOCAL_APIC don't trigger when
  61  * the CPU is idle. To make sure the NMI watchdog really ticks on all
  62  * CPUs during the test make them busy.
  63  */
  64 static __init void nmi_cpu_busy(void *data)
  65 {
  66         local_irq_enable_in_hardirq();
  67         /* Intentionally don't use cpu_relax here. This is
  68            to make sure that the performance counter really ticks,
  69            even if there is a simulator or similar that catches the
  70            pause instruction. On a real HT machine this is fine because
  71            all other CPUs are busy with "useless" delay loops and don't
  72            care if they get somewhat less cycles. */
  73         while (endflag == 0)
  74                 mb();
  75 }
  76 #endif
  77
  78 int __init check_nmi_watchdog(void)
  79 {
  80         int *prev_nmi_count;
  81         int cpu;
  82
  83         if ((nmi_watchdog == NMI_NONE) || (nmi_watchdog == NMI_DISABLED))
  84                 return 0;
  85
  86         if (!atomic_read(&nmi_active))
  87                 return 0;
  88
  89         prev_nmi_count = kmalloc(NR_CPUS * sizeof(int), GFP_KERNEL);
  90         if (!prev_nmi_count)
  91                 return -1;
  92
  93         printk(KERN_INFO "Testing NMI watchdog ... ");
  94
  95 #ifdef CONFIG_SMP
  96         if (nmi_watchdog == NMI_LOCAL_APIC)
  97                 smp_call_function(nmi_cpu_busy, (void *)&endflag, 0, 0);
  98 #endif
  99
 100         for (cpu = 0; cpu < NR_CPUS; cpu++)
 101                 prev_nmi_count[cpu] = cpu_pda(cpu)->__nmi_count;
 102         local_irq_enable();
 103         mdelay((20*1000)/nmi_hz); // wait 20 ticks
 104
 105         for_each_online_cpu(cpu) {
 106                 if (!per_cpu(wd_enabled, cpu))
 107                         continue;
 108                 if (cpu_pda(cpu)->__nmi_count - 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                                 cpu_pda(cpu)->__nmi_count);
 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                 return -1;
 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 }
 134
 135 static int __init setup_nmi_watchdog(char *str)
 136 {
 137         int nmi;
 138
 139         if (!strncmp(str,"panic",5)) {
 140                 panic_on_timeout = 1;
 141                 str = strchr(str, ',');
 142                 if (!str)
 143                         return 1;
 144                 ++str;
 145         }
 146
 147         get_option(&str, &nmi);
 148
 149         if ((nmi >= NMI_INVALID) || (nmi < NMI_NONE))
 150                 return 0;
 151
 152         nmi_watchdog = nmi;
 153         return 1;
 154 }
 155
 156 __setup("nmi_watchdog=", setup_nmi_watchdog);
 157
 158 #ifdef CONFIG_PM
 159
 160 static int nmi_pm_active; /* nmi_active before suspend */
 161
 162 static int lapic_nmi_suspend(struct sys_device *dev, pm_message_t state)
 163 {
 164         /* only CPU0 goes here, other CPUs should be offline */
 165         nmi_pm_active = atomic_read(&nmi_active);
 166         stop_apic_nmi_watchdog(NULL);
 167         BUG_ON(atomic_read(&nmi_active) != 0);
 168         return 0;
 169 }
 170
 171 static int lapic_nmi_resume(struct sys_device *dev)
 172 {
 173         /* only CPU0 goes here, other CPUs should be offline */
 174         if (nmi_pm_active > 0) {
 175                 setup_apic_nmi_watchdog(NULL);
 176                 touch_nmi_watchdog();
 177         }
 178         return 0;
 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(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;
 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
 290 static DEFINE_PER_CPU(unsigned, last_irq_sum);
 291 static DEFINE_PER_CPU(local_t, alert_counter);
 292 static DEFINE_PER_CPU(int, nmi_touch);
 293
 294 void touch_nmi_watchdog(void)
 295 {
 296         if (nmi_watchdog > 0) {
 297                 unsigned cpu;
 298
 299                 /*
 300                  * Tell other CPUs to reset their alert counters. We cannot
 301                  * do it ourselves because the alert count increase is not
 302                  * atomic.
 303                  */
 304                 for_each_present_cpu(cpu) {
 305                         if (per_cpu(nmi_touch, cpu) != 1)
 306                                 per_cpu(nmi_touch, cpu) = 1;
 307                 }
 308         }
 309
 310         touch_softlockup_watchdog();
 311 }
 312 EXPORT_SYMBOL(touch_nmi_watchdog);
 313
 314 int __kprobes nmi_watchdog_tick(struct pt_regs * regs, unsigned reason)
 315 {
 316         int sum;
 317         int touched = 0;
 318         int cpu = smp_processor_id();
 319         int rc = 0;
 320
 321         /* check for other users first */
 322         if (notify_die(DIE_NMI, "nmi", regs, reason, 2, SIGINT)
 323                         == NOTIFY_STOP) {
 324                 rc = 1;
 325                 touched = 1;
 326         }
 327
 328         sum = read_pda(apic_timer_irqs) + read_pda(irq0_irqs);
 329         if (__get_cpu_var(nmi_touch)) {
 330                 __get_cpu_var(nmi_touch) = 0;
 331                 touched = 1;
 332         }
 333
 334         if (cpu_isset(cpu, backtrace_mask)) {
 335                 static DEFINE_SPINLOCK(lock);   /* Serialise the printks */
 336
 337                 spin_lock(&lock);
 338                 printk("NMI backtrace for cpu %d\n", cpu);
 339                 dump_stack();
 340                 spin_unlock(&lock);
 341                 cpu_clear(cpu, backtrace_mask);
 342         }
 343
 344 #ifdef CONFIG_X86_MCE
 345         /* Could check oops_in_progress here too, but it's safer
 346            not too */
 347         if (atomic_read(&mce_entry) > 0)
 348                 touched = 1;
 349 #endif
 350         /* if the apic timer isn't firing, this cpu isn't doing much */
 351         if (!touched && __get_cpu_var(last_irq_sum) == sum) {
 352                 /*
 353                  * Ayiee, looks like this CPU is stuck ...
 354                  * wait a few IRQs (5 seconds) before doing the oops ...
 355                  */
 356                 local_inc(&__get_cpu_var(alert_counter));
 357                 if (local_read(&__get_cpu_var(alert_counter)) == 5*nmi_hz)
 358                         die_nmi("NMI Watchdog detected LOCKUP on CPU %d\n", regs,
 359                                 panic_on_timeout);
 360         } else {
 361                 __get_cpu_var(last_irq_sum) = sum;
 362                 local_set(&__get_cpu_var(alert_counter), 0);
 363         }
 364
 365         /* see if the nmi watchdog went off */
 366         if (!__get_cpu_var(wd_enabled))
 367                 return rc;
 368         switch (nmi_watchdog) {
 369         case NMI_LOCAL_APIC:
 370                 rc |= lapic_wd_event(nmi_hz);
 371                 break;
 372         case NMI_IO_APIC:
 373                 /* don't know how to accurately check for this.
 374                  * just assume it was a watchdog timer interrupt
 375                  * This matches the old behaviour.
 376                  */
 377                 rc = 1;
 378                 break;
 379         }
 380         return rc;
 381 }
 382
 383 static unsigned ignore_nmis;
 384
 385 asmlinkage __kprobes void do_nmi(struct pt_regs * regs, long error_code)
 386 {
 387         nmi_enter();
 388         add_pda(__nmi_count,1);
 389         if (!ignore_nmis)
 390                 default_do_nmi(regs);
 391         nmi_exit();
 392 }
 393
 394 void stop_nmi(void)
 395 {
 396         acpi_nmi_disable();
 397         ignore_nmis++;
 398 }
 399
 400 void restart_nmi(void)
 401 {
 402         ignore_nmis--;
 403         acpi_nmi_enable();
 404 }
 405
 406 #ifdef CONFIG_SYSCTL
 407
 408 static int unknown_nmi_panic_callback(struct pt_regs *regs, int cpu)
 409 {
 410         unsigned char reason = get_nmi_reason();
 411         char buf[64];
 412
 413         sprintf(buf, "NMI received for unknown reason %02x\n", reason);
 414         die_nmi(buf, regs, 1);  /* Always panic here */
 415         return 0;
 416 }
 417
 418 /*
 419  * proc handler for /proc/sys/kernel/nmi
 420  */
 421 int proc_nmi_enabled(struct ctl_table *table, int write, struct file *file,
 422                         void __user *buffer, size_t *length, loff_t *ppos)
 423 {
 424         int old_state;
 425
 426         nmi_watchdog_enabled = (atomic_read(&nmi_active) > 0) ? 1 : 0;
 427         old_state = nmi_watchdog_enabled;
 428         proc_dointvec(table, write, file, buffer, length, ppos);
 429         if (!!old_state == !!nmi_watchdog_enabled)
 430                 return 0;
 431
 432         if (atomic_read(&nmi_active) < 0 || nmi_watchdog == NMI_DISABLED) {
 433                 printk( KERN_WARNING "NMI watchdog is permanently disabled\n");
 434                 return -EIO;
 435         }
 436
 437         /* if nmi_watchdog is not set yet, then set it */
 438         nmi_watchdog_default();
 439
 440         if (nmi_watchdog == NMI_LOCAL_APIC) {
 441                 if (nmi_watchdog_enabled)
 442                         enable_lapic_nmi_watchdog();
 443                 else
 444                         disable_lapic_nmi_watchdog();
 445         } else {
 446                 printk( KERN_WARNING
 447                         "NMI watchdog doesn't know what hardware to touch\n");
 448                 return -EIO;
 449         }
 450         return 0;
 451 }
 452
 453 #endif
 454
 455 int do_nmi_callback(struct pt_regs *regs, int cpu)
 456 {
 457 #ifdef CONFIG_SYSCTL
 458         if (unknown_nmi_panic)
 459                 return unknown_nmi_panic_callback(regs, cpu);
 460 #endif
 461         return 0;
 462 }
 463
 464 void __trigger_all_cpu_backtrace(void)
 465 {
 466         int i;
 467
 468         backtrace_mask = cpu_online_map;
 469         /* Wait for up to 10 seconds for all CPUs to do the backtrace */
 470         for (i = 0; i < 10 * 1000; i++) {
 471                 if (cpus_empty(backtrace_mask))
 472                         break;
 473                 mdelay(1);
 474         }
 475 }
 476
 477 EXPORT_SYMBOL(nmi_active);
 478 EXPORT_SYMBOL(nmi_watchdog);