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 <<<<<<< HEAD:arch/x86/kernel/nmi_64.c
  50 /* local prototypes */
  51 static int unknown_nmi_panic_callback(struct pt_regs *regs, int cpu);
  52
  53 =======
  54 >>>>>>> 264e3e889d86e552b4191d69bb60f4f3b383135a:arch/x86/kernel/nmi_64.c
  55 /* Run after command line and cpu_init init, but before all other checks */
  56 void nmi_watchdog_default(void)
  57 {
  58         if (nmi_watchdog != NMI_DEFAULT)
  59                 return;
  60         nmi_watchdog = NMI_NONE;
  61 }
  62
  63 static int endflag __initdata = 0;
  64
  65 #ifdef CONFIG_SMP
  66 /* The performance counters used by NMI_LOCAL_APIC don't trigger when
  67  * the CPU is idle. To make sure the NMI watchdog really ticks on all
  68  * CPUs during the test make them busy.
  69  */
  70 static __init void nmi_cpu_busy(void *data)
  71 {
  72         local_irq_enable_in_hardirq();
  73         /* Intentionally don't use cpu_relax here. This is
  74            to make sure that the performance counter really ticks,
  75            even if there is a simulator or similar that catches the
  76            pause instruction. On a real HT machine this is fine because
  77            all other CPUs are busy with "useless" delay loops and don't
  78            care if they get somewhat less cycles. */
  79         while (endflag == 0)
  80                 mb();
  81 }
  82 #endif
  83
  84 int __init check_nmi_watchdog(void)
  85 {
  86         int *prev_nmi_count;
  87         int cpu;
  88
  89         if ((nmi_watchdog == NMI_NONE) || (nmi_watchdog == NMI_DISABLED))
  90                 return 0;
  91
  92         if (!atomic_read(&nmi_active))
  93                 return 0;
  94
  95         prev_nmi_count = kmalloc(NR_CPUS * sizeof(int), GFP_KERNEL);
  96         if (!prev_nmi_count)
  97                 return -1;
  98
  99         printk(KERN_INFO "Testing NMI watchdog ... ");
 100
 101 #ifdef CONFIG_SMP
 102         if (nmi_watchdog == NMI_LOCAL_APIC)
 103                 smp_call_function(nmi_cpu_busy, (void *)&endflag, 0, 0);
 104 #endif
 105
 106         for (cpu = 0; cpu < NR_CPUS; cpu++)
 107                 prev_nmi_count[cpu] = cpu_pda(cpu)->__nmi_count;
 108         local_irq_enable();
 109         mdelay((20*1000)/nmi_hz); // wait 20 ticks
 110
 111         for_each_online_cpu(cpu) {
 112                 if (!per_cpu(wd_enabled, cpu))
 113                         continue;
 114                 if (cpu_pda(cpu)->__nmi_count - prev_nmi_count[cpu] <= 5) {
 115                         printk(KERN_WARNING "WARNING: CPU#%d: NMI "
 116                                "appears to be stuck (%d->%d)!\n",
 117                                 cpu,
 118                                 prev_nmi_count[cpu],
 119                                 cpu_pda(cpu)->__nmi_count);
 120                         per_cpu(wd_enabled, cpu) = 0;
 121                         atomic_dec(&nmi_active);
 122                 }
 123         }
 124         endflag = 1;
 125         if (!atomic_read(&nmi_active)) {
 126                 kfree(prev_nmi_count);
 127                 atomic_set(&nmi_active, -1);
 128                 return -1;
 129         }
 130         printk("OK.\n");
 131
 132         /* now that we know it works we can reduce NMI frequency to
 133            something more reasonable; makes a difference in some configs */
 134         if (nmi_watchdog == NMI_LOCAL_APIC)
 135                 nmi_hz = lapic_adjust_nmi_hz(1);
 136
 137         kfree(prev_nmi_count);
 138         return 0;
 139 }
 140
 141 static int __init setup_nmi_watchdog(char *str)
 142 {
 143         int nmi;
 144
 145         if (!strncmp(str,"panic",5)) {
 146                 panic_on_timeout = 1;
 147                 str = strchr(str, ',');
 148                 if (!str)
 149                         return 1;
 150                 ++str;
 151         }
 152
 153         get_option(&str, &nmi);
 154
 155         if ((nmi >= NMI_INVALID) || (nmi < NMI_NONE))
 156                 return 0;
 157
 158         nmi_watchdog = nmi;
 159         return 1;
 160 }
 161
 162 __setup("nmi_watchdog=", setup_nmi_watchdog);
 163
 164 #ifdef CONFIG_PM
 165
 166 static int nmi_pm_active; /* nmi_active before suspend */
 167
 168 static int lapic_nmi_suspend(struct sys_device *dev, pm_message_t state)
 169 {
 170         /* only CPU0 goes here, other CPUs should be offline */
 171         nmi_pm_active = atomic_read(&nmi_active);
 172         stop_apic_nmi_watchdog(NULL);
 173         BUG_ON(atomic_read(&nmi_active) != 0);
 174         return 0;
 175 }
 176
 177 static int lapic_nmi_resume(struct sys_device *dev)
 178 {
 179         /* only CPU0 goes here, other CPUs should be offline */
 180         if (nmi_pm_active > 0) {
 181                 setup_apic_nmi_watchdog(NULL);
 182                 touch_nmi_watchdog();
 183         }
 184         return 0;
 185 }
 186
 187 static struct sysdev_class nmi_sysclass = {
 188         .name           = "lapic_nmi",
 189         .resume         = lapic_nmi_resume,
 190         .suspend        = lapic_nmi_suspend,
 191 };
 192
 193 static struct sys_device device_lapic_nmi = {
 194         .id     = 0,
 195         .cls    = &nmi_sysclass,
 196 };
 197
 198 static int __init init_lapic_nmi_sysfs(void)
 199 {
 200         int error;
 201
 202         /* should really be a BUG_ON but b/c this is an
 203          * init call, it just doesn't work.  -dcz
 204          */
 205         if (nmi_watchdog != NMI_LOCAL_APIC)
 206                 return 0;
 207
 208         if (atomic_read(&nmi_active) < 0)
 209                 return 0;
 210
 211         error = sysdev_class_register(&nmi_sysclass);
 212         if (!error)
 213                 error = sysdev_register(&device_lapic_nmi);
 214         return error;
 215 }
 216 /* must come after the local APIC's device_initcall() */
 217 late_initcall(init_lapic_nmi_sysfs);
 218
 219 #endif  /* CONFIG_PM */
 220
 221 static void __acpi_nmi_enable(void *__unused)
 222 {
 223         apic_write(APIC_LVT0, APIC_DM_NMI);
 224 }
 225
 226 /*
 227  * Enable timer based NMIs on all CPUs:
 228  */
 229 void acpi_nmi_enable(void)
 230 {
 231         if (atomic_read(&nmi_active) && nmi_watchdog == NMI_IO_APIC)
 232                 on_each_cpu(__acpi_nmi_enable, NULL, 0, 1);
 233 }
 234
 235 static void __acpi_nmi_disable(void *__unused)
 236 {
 237         apic_write(APIC_LVT0, APIC_DM_NMI | APIC_LVT_MASKED);
 238 }
 239
 240 /*
 241  * Disable timer based NMIs on all CPUs:
 242  */
 243 void acpi_nmi_disable(void)
 244 {
 245         if (atomic_read(&nmi_active) && nmi_watchdog == NMI_IO_APIC)
 246                 on_each_cpu(__acpi_nmi_disable, NULL, 0, 1);
 247 }
 248
 249 void setup_apic_nmi_watchdog(void *unused)
 250 {
 251         if (__get_cpu_var(wd_enabled))
 252                 return;
 253
 254         /* cheap hack to support suspend/resume */
 255         /* if cpu0 is not active neither should the other cpus */
 256         if ((smp_processor_id() != 0) && (atomic_read(&nmi_active) <= 0))
 257                 return;
 258
 259         switch (nmi_watchdog) {
 260         case NMI_LOCAL_APIC:
 261                 __get_cpu_var(wd_enabled) = 1;
 262                 if (lapic_watchdog_init(nmi_hz) < 0) {
 263                         __get_cpu_var(wd_enabled) = 0;
 264                         return;
 265                 }
 266                 /* FALL THROUGH */
 267         case NMI_IO_APIC:
 268                 __get_cpu_var(wd_enabled) = 1;
 269                 atomic_inc(&nmi_active);
 270         }
 271 }
 272
 273 void stop_apic_nmi_watchdog(void *unused)
 274 {
 275         /* only support LOCAL and IO APICs for now */
 276         if ((nmi_watchdog != NMI_LOCAL_APIC) &&
 277             (nmi_watchdog != NMI_IO_APIC))
 278                 return;
 279         if (__get_cpu_var(wd_enabled) == 0)
 280                 return;
 281         if (nmi_watchdog == NMI_LOCAL_APIC)
 282                 lapic_watchdog_stop();
 283         __get_cpu_var(wd_enabled) = 0;
 284         atomic_dec(&nmi_active);
 285 }
 286
 287 /*
 288  * the best way to detect whether a CPU has a 'hard lockup' problem
 289  * is to check it's local APIC timer IRQ counts. If they are not
 290  * changing then that CPU has some problem.
 291  *
 292  * as these watchdog NMI IRQs are generated on every CPU, we only
 293  * have to check the current processor.
 294  */
 295
 296 static DEFINE_PER_CPU(unsigned, last_irq_sum);
 297 static DEFINE_PER_CPU(local_t, alert_counter);
 298 static DEFINE_PER_CPU(int, nmi_touch);
 299
 300 void touch_nmi_watchdog(void)
 301 {
 302         if (nmi_watchdog > 0) {
 303                 unsigned cpu;
 304
 305                 /*
 306                  * Tell other CPUs to reset their alert counters. We cannot
 307                  * do it ourselves because the alert count increase is not
 308                  * atomic.
 309                  */
 310                 for_each_present_cpu(cpu) {
 311                         if (per_cpu(nmi_touch, cpu) != 1)
 312                                 per_cpu(nmi_touch, cpu) = 1;
 313                 }
 314         }
 315
 316         touch_softlockup_watchdog();
 317 }
 318 EXPORT_SYMBOL(touch_nmi_watchdog);
 319
 320 int __kprobes nmi_watchdog_tick(struct pt_regs * regs, unsigned reason)
 321 {
 322         int sum;
 323         int touched = 0;
 324         int cpu = smp_processor_id();
 325         int rc = 0;
 326
 327         /* check for other users first */
 328         if (notify_die(DIE_NMI, "nmi", regs, reason, 2, SIGINT)
 329                         == NOTIFY_STOP) {
 330                 rc = 1;
 331                 touched = 1;
 332         }
 333
 334         sum = read_pda(apic_timer_irqs) + read_pda(irq0_irqs);
 335         if (__get_cpu_var(nmi_touch)) {
 336                 __get_cpu_var(nmi_touch) = 0;
 337                 touched = 1;
 338         }
 339
 340         if (cpu_isset(cpu, backtrace_mask)) {
 341                 static DEFINE_SPINLOCK(lock);   /* Serialise the printks */
 342
 343                 spin_lock(&lock);
 344                 printk("NMI backtrace for cpu %d\n", cpu);
 345                 dump_stack();
 346                 spin_unlock(&lock);
 347                 cpu_clear(cpu, backtrace_mask);
 348         }
 349
 350 #ifdef CONFIG_X86_MCE
 351         /* Could check oops_in_progress here too, but it's safer
 352            not too */
 353         if (atomic_read(&mce_entry) > 0)
 354                 touched = 1;
 355 #endif
 356         /* if the apic timer isn't firing, this cpu isn't doing much */
 357         if (!touched && __get_cpu_var(last_irq_sum) == sum) {
 358                 /*
 359                  * Ayiee, looks like this CPU is stuck ...
 360                  * wait a few IRQs (5 seconds) before doing the oops ...
 361                  */
 362                 local_inc(&__get_cpu_var(alert_counter));
 363                 if (local_read(&__get_cpu_var(alert_counter)) == 5*nmi_hz)
 364                         die_nmi("NMI Watchdog detected LOCKUP on CPU %d\n", regs,
 365                                 panic_on_timeout);
 366         } else {
 367                 __get_cpu_var(last_irq_sum) = sum;
 368                 local_set(&__get_cpu_var(alert_counter), 0);
 369         }
 370
 371         /* see if the nmi watchdog went off */
 372         if (!__get_cpu_var(wd_enabled))
 373                 return rc;
 374         switch (nmi_watchdog) {
 375         case NMI_LOCAL_APIC:
 376                 rc |= lapic_wd_event(nmi_hz);
 377                 break;
 378         case NMI_IO_APIC:
 379                 /* don't know how to accurately check for this.
 380                  * just assume it was a watchdog timer interrupt
 381                  * This matches the old behaviour.
 382                  */
 383                 rc = 1;
 384                 break;
 385         }
 386         return rc;
 387 }
 388
 389 static unsigned ignore_nmis;
 390
 391 asmlinkage __kprobes void do_nmi(struct pt_regs * regs, long error_code)
 392 {
 393         nmi_enter();
 394         add_pda(__nmi_count,1);
 395         if (!ignore_nmis)
 396                 default_do_nmi(regs);
 397         nmi_exit();
 398 }
 399
 400 <<<<<<< HEAD:arch/x86/kernel/nmi_64.c
 401 int do_nmi_callback(struct pt_regs * regs, int cpu)
 402 {
 403 #ifdef CONFIG_SYSCTL
 404         if (unknown_nmi_panic)
 405                 return unknown_nmi_panic_callback(regs, cpu);
 406 #endif
 407         return 0;
 408 }
 409
 410 =======
 411 >>>>>>> 264e3e889d86e552b4191d69bb60f4f3b383135a:arch/x86/kernel/nmi_64.c
 412 void stop_nmi(void)
 413 {
 414         acpi_nmi_disable();
 415         ignore_nmis++;
 416 }
 417
 418 void restart_nmi(void)
 419 {
 420         ignore_nmis--;
 421         acpi_nmi_enable();
 422 }
 423
 424 #ifdef CONFIG_SYSCTL
 425
 426 static int unknown_nmi_panic_callback(struct pt_regs *regs, int cpu)
 427 {
 428         unsigned char reason = get_nmi_reason();
 429         char buf[64];
 430
 431         sprintf(buf, "NMI received for unknown reason %02x\n", reason);
 432         die_nmi(buf, regs, 1);  /* Always panic here */
 433         return 0;
 434 }
 435
 436 /*
 437  * proc handler for /proc/sys/kernel/nmi
 438  */
 439 int proc_nmi_enabled(struct ctl_table *table, int write, struct file *file,
 440                         void __user *buffer, size_t *length, loff_t *ppos)
 441 {
 442         int old_state;
 443
 444         nmi_watchdog_enabled = (atomic_read(&nmi_active) > 0) ? 1 : 0;
 445         old_state = nmi_watchdog_enabled;
 446         proc_dointvec(table, write, file, buffer, length, ppos);
 447         if (!!old_state == !!nmi_watchdog_enabled)
 448                 return 0;
 449
 450         if (atomic_read(&nmi_active) < 0 || nmi_watchdog == NMI_DISABLED) {
 451                 printk( KERN_WARNING "NMI watchdog is permanently disabled\n");
 452                 return -EIO;
 453         }
 454
 455         /* if nmi_watchdog is not set yet, then set it */
 456         nmi_watchdog_default();
 457
 458         if (nmi_watchdog == NMI_LOCAL_APIC) {
 459                 if (nmi_watchdog_enabled)
 460                         enable_lapic_nmi_watchdog();
 461                 else
 462                         disable_lapic_nmi_watchdog();
 463         } else {
 464                 printk( KERN_WARNING
 465                         "NMI watchdog doesn't know what hardware to touch\n");
 466                 return -EIO;
 467         }
 468         return 0;
 469 }
 470
 471 #endif
 472
 473 <<<<<<< HEAD:arch/x86/kernel/nmi_64.c
 474 =======
 475 int do_nmi_callback(struct pt_regs *regs, int cpu)
 476 {
 477 #ifdef CONFIG_SYSCTL
 478         if (unknown_nmi_panic)
 479                 return unknown_nmi_panic_callback(regs, cpu);
 480 #endif
 481         return 0;
 482 }
 483
 484 >>>>>>> 264e3e889d86e552b4191d69bb60f4f3b383135a:arch/x86/kernel/nmi_64.c
 485 void __trigger_all_cpu_backtrace(void)
 486 {
 487         int i;
 488
 489         backtrace_mask = cpu_online_map;
 490         /* Wait for up to 10 seconds for all CPUs to do the backtrace */
 491         for (i = 0; i < 10 * 1000; i++) {
 492                 if (cpus_empty(backtrace_mask))
 493                         break;
 494                 mdelay(1);
 495         }
 496 }
 497
 498 EXPORT_SYMBOL(nmi_active);
 499 EXPORT_SYMBOL(nmi_watchdog);