arch/x86/kernel/nmi_32.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  *  Mikael Pettersson   : Pentium 4 support for local APIC NMI watchdog.
  10  *  Pavel Machek and
  11  *  Mikael Pettersson   : PM converted to driver model. Disable/enable API.
  12  */
  13
  14 #include <linux/delay.h>
  15 #include <linux/interrupt.h>
  16 #include <linux/module.h>
  17 #include <linux/nmi.h>
  18 #include <linux/sysdev.h>
  19 #include <linux/sysctl.h>
  20 #include <linux/percpu.h>
  21 #include <linux/kprobes.h>
  22 #include <linux/cpumask.h>
  23 #include <linux/kernel_stat.h>
  24 #include <linux/kdebug.h>
  25
  26 #include <asm/smp.h>
  27 #include <asm/nmi.h>
  28
  29 #include "mach_traps.h"
  30
  31 int unknown_nmi_panic;
  32 int nmi_watchdog_enabled;
  33
  34 static cpumask_t backtrace_mask = CPU_MASK_NONE;
  35
  36 /* nmi_active:
  37  * >0: the lapic NMI watchdog is active, but can be disabled
  38  * <0: the lapic NMI watchdog has not been set up, and cannot
  39  *     be enabled
  40  *  0: the lapic NMI watchdog is disabled, but can be enabled
  41  */
  42 atomic_t nmi_active = ATOMIC_INIT(0);           /* oprofile uses this */
  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_32.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_32.c
  55 static int endflag __initdata = 0;
  56
  57 #ifdef CONFIG_SMP
  58 /* The performance counters used by NMI_LOCAL_APIC don't trigger when
  59  * the CPU is idle. To make sure the NMI watchdog really ticks on all
  60  * CPUs during the test make them busy.
  61  */
  62 static __init void nmi_cpu_busy(void *data)
  63 {
  64         local_irq_enable_in_hardirq();
  65         /* Intentionally don't use cpu_relax here. This is
  66            to make sure that the performance counter really ticks,
  67            even if there is a simulator or similar that catches the
  68            pause instruction. On a real HT machine this is fine because
  69            all other CPUs are busy with "useless" delay loops and don't
  70            care if they get somewhat less cycles. */
  71         while (endflag == 0)
  72                 mb();
  73 }
  74 #endif
  75
  76 static int __init check_nmi_watchdog(void)
  77 {
  78         unsigned int *prev_nmi_count;
  79         int cpu;
  80
  81         if ((nmi_watchdog == NMI_NONE) || (nmi_watchdog == NMI_DISABLED))
  82                 return 0;
  83
  84         if (!atomic_read(&nmi_active))
  85                 return 0;
  86
  87         prev_nmi_count = kmalloc(NR_CPUS * sizeof(int), GFP_KERNEL);
  88         if (!prev_nmi_count)
  89                 return -1;
  90
  91         printk(KERN_INFO "Testing NMI watchdog ... ");
  92
  93 #ifdef CONFIG_SMP
  94         if (nmi_watchdog == NMI_LOCAL_APIC)
  95                 smp_call_function(nmi_cpu_busy, (void *)&endflag, 0, 0);
  96 #endif
  97
  98         for_each_possible_cpu(cpu)
  99                 prev_nmi_count[cpu] = nmi_count(cpu);
 100         local_irq_enable();
 101         mdelay((20*1000)/nmi_hz); // wait 20 ticks
 102
 103         for_each_possible_cpu(cpu) {
 104 #ifdef CONFIG_SMP
 105                 /* Check cpu_callin_map here because that is set
 106                    after the timer is started. */
 107                 if (!cpu_isset(cpu, cpu_callin_map))
 108                         continue;
 109 #endif
 110                 if (!per_cpu(wd_enabled, cpu))
 111                         continue;
 112                 if (nmi_count(cpu) - 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                                 nmi_count(cpu));
 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                 return -1;
 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 }
 138 /* This needs to happen later in boot so counters are working */
 139 late_initcall(check_nmi_watchdog);
 140
 141 static int __init setup_nmi_watchdog(char *str)
 142 {
 143         int nmi;
 144
 145         get_option(&str, &nmi);
 146
 147         if ((nmi >= NMI_INVALID) || (nmi < NMI_NONE))
 148                 return 0;
 149
 150         nmi_watchdog = nmi;
 151         return 1;
 152 }
 153
 154 __setup("nmi_watchdog=", setup_nmi_watchdog);
 155
 156
 157 /* Suspend/resume support */
 158
 159 #ifdef CONFIG_PM
 160
 161 static int nmi_pm_active; /* nmi_active before suspend */
 162
 163 static int lapic_nmi_suspend(struct sys_device *dev, pm_message_t state)
 164 {
 165         /* only CPU0 goes here, other CPUs should be offline */
 166         nmi_pm_active = atomic_read(&nmi_active);
 167         stop_apic_nmi_watchdog(NULL);
 168         BUG_ON(atomic_read(&nmi_active) != 0);
 169         return 0;
 170 }
 171
 172 static int lapic_nmi_resume(struct sys_device *dev)
 173 {
 174         /* only CPU0 goes here, other CPUs should be offline */
 175         if (nmi_pm_active > 0) {
 176                 setup_apic_nmi_watchdog(NULL);
 177                 touch_nmi_watchdog();
 178         }
 179         return 0;
 180 }
 181
 182
 183 static struct sysdev_class nmi_sysclass = {
 184         .name           = "lapic_nmi",
 185         .resume         = lapic_nmi_resume,
 186         .suspend        = lapic_nmi_suspend,
 187 };
 188
 189 static struct sys_device device_lapic_nmi = {
 190         .id     = 0,
 191         .cls    = &nmi_sysclass,
 192 };
 193
 194 static int __init init_lapic_nmi_sysfs(void)
 195 {
 196         int error;
 197
 198         /* should really be a BUG_ON but b/c this is an
 199          * init call, it just doesn't work.  -dcz
 200          */
 201         if (nmi_watchdog != NMI_LOCAL_APIC)
 202                 return 0;
 203
 204         if (atomic_read(&nmi_active) < 0)
 205                 return 0;
 206
 207         error = sysdev_class_register(&nmi_sysclass);
 208         if (!error)
 209                 error = sysdev_register(&device_lapic_nmi);
 210         return error;
 211 }
 212 /* must come after the local APIC's device_initcall() */
 213 late_initcall(init_lapic_nmi_sysfs);
 214
 215 #endif  /* CONFIG_PM */
 216
 217 static void __acpi_nmi_enable(void *__unused)
 218 {
 219         apic_write_around(APIC_LVT0, APIC_DM_NMI);
 220 }
 221
 222 /*
 223  * Enable timer based NMIs on all CPUs:
 224  */
 225 void acpi_nmi_enable(void)
 226 {
 227         if (atomic_read(&nmi_active) && nmi_watchdog == NMI_IO_APIC)
 228                 on_each_cpu(__acpi_nmi_enable, NULL, 0, 1);
 229 }
 230
 231 static void __acpi_nmi_disable(void *__unused)
 232 {
 233         apic_write(APIC_LVT0, APIC_DM_NMI | APIC_LVT_MASKED);
 234 }
 235
 236 /*
 237  * Disable timer based NMIs on all CPUs:
 238  */
 239 void acpi_nmi_disable(void)
 240 {
 241         if (atomic_read(&nmi_active) && nmi_watchdog == NMI_IO_APIC)
 242                 on_each_cpu(__acpi_nmi_disable, NULL, 0, 1);
 243 }
 244
 245 void setup_apic_nmi_watchdog(void *unused)
 246 {
 247         if (__get_cpu_var(wd_enabled))
 248                 return;
 249
 250         /* cheap hack to support suspend/resume */
 251         /* if cpu0 is not active neither should the other cpus */
 252         if ((smp_processor_id() != 0) && (atomic_read(&nmi_active) <= 0))
 253                 return;
 254
 255         switch (nmi_watchdog) {
 256         case NMI_LOCAL_APIC:
 257                 __get_cpu_var(wd_enabled) = 1; /* enable it before to avoid race with handler */
 258                 if (lapic_watchdog_init(nmi_hz) < 0) {
 259                         __get_cpu_var(wd_enabled) = 0;
 260                         return;
 261                 }
 262                 /* FALL THROUGH */
 263         case NMI_IO_APIC:
 264                 __get_cpu_var(wd_enabled) = 1;
 265                 atomic_inc(&nmi_active);
 266         }
 267 }
 268
 269 void stop_apic_nmi_watchdog(void *unused)
 270 {
 271         /* only support LOCAL and IO APICs for now */
 272         if ((nmi_watchdog != NMI_LOCAL_APIC) &&
 273             (nmi_watchdog != NMI_IO_APIC))
 274                 return;
 275         if (__get_cpu_var(wd_enabled) == 0)
 276                 return;
 277         if (nmi_watchdog == NMI_LOCAL_APIC)
 278                 lapic_watchdog_stop();
 279         __get_cpu_var(wd_enabled) = 0;
 280         atomic_dec(&nmi_active);
 281 }
 282
 283 /*
 284  * the best way to detect whether a CPU has a 'hard lockup' problem
 285  * is to check it's local APIC timer IRQ counts. If they are not
 286  * changing then that CPU has some problem.
 287  *
 288  * as these watchdog NMI IRQs are generated on every CPU, we only
 289  * have to check the current processor.
 290  *
 291  * since NMIs don't listen to _any_ locks, we have to be extremely
 292  * careful not to rely on unsafe variables. The printk might lock
 293  * up though, so we have to break up any console locks first ...
 294  * [when there will be more tty-related locks, break them up
 295  *  here too!]
 296  */
 297
 298 static unsigned int
 299         last_irq_sums [NR_CPUS],
 300         alert_counter [NR_CPUS];
 301
 302 void touch_nmi_watchdog(void)
 303 {
 304         if (nmi_watchdog > 0) {
 305                 unsigned cpu;
 306
 307                 /*
 308                  * Just reset the alert counters, (other CPUs might be
 309                  * spinning on locks we hold):
 310                  */
 311                 for_each_present_cpu(cpu) {
 312                         if (alert_counter[cpu])
 313                                 alert_counter[cpu] = 0;
 314                 }
 315         }
 316
 317         /*
 318          * Tickle the softlockup detector too:
 319          */
 320         touch_softlockup_watchdog();
 321 }
 322 EXPORT_SYMBOL(touch_nmi_watchdog);
 323
 324 extern void die_nmi(struct pt_regs *, const char *msg);
 325
 326 __kprobes int nmi_watchdog_tick(struct pt_regs * regs, unsigned reason)
 327 {
 328
 329         /*
 330          * Since current_thread_info()-> is always on the stack, and we
 331          * always switch the stack NMI-atomically, it's safe to use
 332          * smp_processor_id().
 333          */
 334         unsigned int sum;
 335         int touched = 0;
 336         int cpu = smp_processor_id();
 337         int rc = 0;
 338
 339         /* check for other users first */
 340         if (notify_die(DIE_NMI, "nmi", regs, reason, 2, SIGINT)
 341                         == NOTIFY_STOP) {
 342                 rc = 1;
 343                 touched = 1;
 344         }
 345
 346         if (cpu_isset(cpu, backtrace_mask)) {
 347                 static DEFINE_SPINLOCK(lock);   /* Serialise the printks */
 348
 349                 spin_lock(&lock);
 350                 printk("NMI backtrace for cpu %d\n", cpu);
 351                 dump_stack();
 352                 spin_unlock(&lock);
 353                 cpu_clear(cpu, backtrace_mask);
 354         }
 355
 356         /*
 357          * Take the local apic timer and PIT/HPET into account. We don't
 358          * know which one is active, when we have highres/dyntick on
 359          */
 360         sum = per_cpu(irq_stat, cpu).apic_timer_irqs +
 361                 per_cpu(irq_stat, cpu).irq0_irqs;
 362
 363         /* if the none of the timers isn't firing, this cpu isn't doing much */
 364         if (!touched && last_irq_sums[cpu] == sum) {
 365                 /*
 366                  * Ayiee, looks like this CPU is stuck ...
 367                  * wait a few IRQs (5 seconds) before doing the oops ...
 368                  */
 369                 alert_counter[cpu]++;
 370                 if (alert_counter[cpu] == 5*nmi_hz)
 371                         /*
 372                          * die_nmi will return ONLY if NOTIFY_STOP happens..
 373                          */
 374                         die_nmi(regs, "BUG: NMI Watchdog detected LOCKUP");
 375         } else {
 376                 last_irq_sums[cpu] = sum;
 377                 alert_counter[cpu] = 0;
 378         }
 379         /* see if the nmi watchdog went off */
 380         if (!__get_cpu_var(wd_enabled))
 381                 return rc;
 382         switch (nmi_watchdog) {
 383         case NMI_LOCAL_APIC:
 384                 rc |= lapic_wd_event(nmi_hz);
 385                 break;
 386         case NMI_IO_APIC:
 387                 /* don't know how to accurately check for this.
 388                  * just assume it was a watchdog timer interrupt
 389                  * This matches the old behaviour.
 390                  */
 391                 rc = 1;
 392                 break;
 393         }
 394         return rc;
 395 }
 396
 397 <<<<<<< HEAD:arch/x86/kernel/nmi_32.c
 398 int do_nmi_callback(struct pt_regs * regs, int cpu)
 399 {
 400 #ifdef CONFIG_SYSCTL
 401         if (unknown_nmi_panic)
 402                 return unknown_nmi_panic_callback(regs, cpu);
 403 #endif
 404         return 0;
 405 }
 406
 407 =======
 408 >>>>>>> 264e3e889d86e552b4191d69bb60f4f3b383135a:arch/x86/kernel/nmi_32.c
 409 #ifdef CONFIG_SYSCTL
 410
 411 static int unknown_nmi_panic_callback(struct pt_regs *regs, int cpu)
 412 {
 413         unsigned char reason = get_nmi_reason();
 414         char buf[64];
 415
 416         sprintf(buf, "NMI received for unknown reason %02x\n", reason);
 417         die_nmi(regs, buf);
 418         return 0;
 419 }
 420
 421 /*
 422  * proc handler for /proc/sys/kernel/nmi
 423  */
 424 int proc_nmi_enabled(struct ctl_table *table, int write, struct file *file,
 425                         void __user *buffer, size_t *length, loff_t *ppos)
 426 {
 427         int old_state;
 428
 429         nmi_watchdog_enabled = (atomic_read(&nmi_active) > 0) ? 1 : 0;
 430         old_state = nmi_watchdog_enabled;
 431         proc_dointvec(table, write, file, buffer, length, ppos);
 432         if (!!old_state == !!nmi_watchdog_enabled)
 433                 return 0;
 434
 435         if (atomic_read(&nmi_active) < 0 || nmi_watchdog == NMI_DISABLED) {
 436                 printk( KERN_WARNING "NMI watchdog is permanently disabled\n");
 437                 return -EIO;
 438         }
 439
 440         if (nmi_watchdog == NMI_DEFAULT) {
 441                 if (lapic_watchdog_ok())
 442                         nmi_watchdog = NMI_LOCAL_APIC;
 443                 else
 444                         nmi_watchdog = NMI_IO_APIC;
 445         }
 446
 447         if (nmi_watchdog == NMI_LOCAL_APIC) {
 448                 if (nmi_watchdog_enabled)
 449                         enable_lapic_nmi_watchdog();
 450                 else
 451                         disable_lapic_nmi_watchdog();
 452         } else {
 453                 printk( KERN_WARNING
 454                         "NMI watchdog doesn't know what hardware to touch\n");
 455                 return -EIO;
 456         }
 457         return 0;
 458 }
 459
 460 #endif
 461
 462 <<<<<<< HEAD:arch/x86/kernel/nmi_32.c
 463 =======
 464 int do_nmi_callback(struct pt_regs *regs, int cpu)
 465 {
 466 #ifdef CONFIG_SYSCTL
 467         if (unknown_nmi_panic)
 468                 return unknown_nmi_panic_callback(regs, cpu);
 469 #endif
 470         return 0;
 471 }
 472
 473 >>>>>>> 264e3e889d86e552b4191d69bb60f4f3b383135a:arch/x86/kernel/nmi_32.c
 474 void __trigger_all_cpu_backtrace(void)
 475 {
 476         int i;
 477
 478         backtrace_mask = cpu_online_map;
 479         /* Wait for up to 10 seconds for all CPUs to do the backtrace */
 480         for (i = 0; i < 10 * 1000; i++) {
 481                 if (cpus_empty(backtrace_mask))
 482                         break;
 483                 mdelay(1);
 484         }
 485 }
 486
 487 EXPORT_SYMBOL(nmi_active);
 488 EXPORT_SYMBOL(nmi_watchdog);