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 #include <linux/slab.h>
  26
  27 #include <asm/smp.h>
  28 #include <asm/nmi.h>
  29 #include <asm/timer.h>
  30
  31 #include "mach_traps.h"
  32
  33 int unknown_nmi_panic;
  34 int nmi_watchdog_enabled;
  35
  36 static cpumask_t backtrace_mask = CPU_MASK_NONE;
  37
  38 /* nmi_active:
  39  * >0: the lapic NMI watchdog is active, but can be disabled
  40  * <0: the lapic NMI watchdog has not been set up, and cannot
  41  *     be enabled
  42  *  0: the lapic NMI watchdog is disabled, but can be enabled
  43  */
  44 atomic_t nmi_active = ATOMIC_INIT(0);           /* oprofile uses this */
  45 static int panic_on_timeout;
  46
  47 unsigned int nmi_watchdog = NMI_DEFAULT;
  48 static unsigned int nmi_hz = HZ;
  49
  50 static DEFINE_PER_CPU(short, wd_enabled);
  51
  52 static int endflag __initdata = 0;
  53
  54 /* Run after command line and cpu_init init, but before all other checks */
  55 void nmi_watchdog_default(void)
  56 {
  57         if (nmi_watchdog != NMI_DEFAULT)
  58                 return;
  59         if (lapic_watchdog_ok())
  60                 nmi_watchdog = NMI_LOCAL_APIC;
  61         else
  62                 nmi_watchdog = NMI_IO_APIC;
  63 }
  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         unsigned 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                 goto error;
  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_each_possible_cpu(cpu)
 107                 prev_nmi_count[cpu] = nmi_count(cpu);
 108         local_irq_enable();
 109         mdelay((20*1000)/nmi_hz); // wait 20 ticks
 110
 111         for_each_possible_cpu(cpu) {
 112 #ifdef CONFIG_SMP
 113                 /* Check cpu_callin_map here because that is set
 114                    after the timer is started. */
 115                 if (!cpu_isset(cpu, cpu_callin_map))
 116                         continue;
 117 #endif
 118                 if (!per_cpu(wd_enabled, cpu))
 119                         continue;
 120                 if (nmi_count(cpu) - prev_nmi_count[cpu] <= 5) {
 121                         printk(KERN_WARNING "WARNING: CPU#%d: NMI "
 122                                 "appears to be stuck (%d->%d)!\n",
 123                                 cpu,
 124                                 prev_nmi_count[cpu],
 125                                 nmi_count(cpu));
 126                         per_cpu(wd_enabled, cpu) = 0;
 127                         atomic_dec(&nmi_active);
 128                 }
 129         }
 130         endflag = 1;
 131         if (!atomic_read(&nmi_active)) {
 132                 kfree(prev_nmi_count);
 133                 atomic_set(&nmi_active, -1);
 134                 goto error;
 135         }
 136         printk("OK.\n");
 137
 138         /* now that we know it works we can reduce NMI frequency to
 139            something more reasonable; makes a difference in some configs */
 140         if (nmi_watchdog == NMI_LOCAL_APIC)
 141                 nmi_hz = lapic_adjust_nmi_hz(1);
 142
 143         kfree(prev_nmi_count);
 144         return 0;
 145 error:
 146         timer_ack = !cpu_has_tsc;
 147
 148         return -1;
 149 }
 150
 151 static int __init setup_nmi_watchdog(char *str)
 152 {
 153         int nmi;
 154
 155         if (!strncmp(str, "panic", 5)) {
 156                 panic_on_timeout = 1;
 157                 str = strchr(str, ',');
 158                 if (!str)
 159                         return 1;
 160                 ++str;
 161         }
 162
 163         get_option(&str, &nmi);
 164
 165         if ((nmi >= NMI_INVALID) || (nmi < NMI_NONE))
 166                 return 0;
 167
 168         nmi_watchdog = nmi;
 169         return 1;
 170 }
 171
 172 __setup("nmi_watchdog=", setup_nmi_watchdog);
 173
 174
 175 /* Suspend/resume support */
 176
 177 #ifdef CONFIG_PM
 178
 179 static int nmi_pm_active; /* nmi_active before suspend */
 180
 181 static int lapic_nmi_suspend(struct sys_device *dev, pm_message_t state)
 182 {
 183         /* only CPU0 goes here, other CPUs should be offline */
 184         nmi_pm_active = atomic_read(&nmi_active);
 185         stop_apic_nmi_watchdog(NULL);
 186         BUG_ON(atomic_read(&nmi_active) != 0);
 187         return 0;
 188 }
 189
 190 static int lapic_nmi_resume(struct sys_device *dev)
 191 {
 192         /* only CPU0 goes here, other CPUs should be offline */
 193         if (nmi_pm_active > 0) {
 194                 setup_apic_nmi_watchdog(NULL);
 195                 touch_nmi_watchdog();
 196         }
 197         return 0;
 198 }
 199
 200
 201 static struct sysdev_class nmi_sysclass = {
 202         .name           = "lapic_nmi",
 203         .resume         = lapic_nmi_resume,
 204         .suspend        = lapic_nmi_suspend,
 205 };
 206
 207 static struct sys_device device_lapic_nmi = {
 208         .id     = 0,
 209         .cls    = &nmi_sysclass,
 210 };
 211
 212 static int __init init_lapic_nmi_sysfs(void)
 213 {
 214         int error;
 215
 216         /* should really be a BUG_ON but b/c this is an
 217          * init call, it just doesn't work.  -dcz
 218          */
 219         if (nmi_watchdog != NMI_LOCAL_APIC)
 220                 return 0;
 221
 222         if (atomic_read(&nmi_active) < 0)
 223                 return 0;
 224
 225         error = sysdev_class_register(&nmi_sysclass);
 226         if (!error)
 227                 error = sysdev_register(&device_lapic_nmi);
 228         return error;
 229 }
 230 /* must come after the local APIC's device_initcall() */
 231 late_initcall(init_lapic_nmi_sysfs);
 232
 233 #endif  /* CONFIG_PM */
 234
 235 static void __acpi_nmi_enable(void *__unused)
 236 {
 237         apic_write_around(APIC_LVT0, APIC_DM_NMI);
 238 }
 239
 240 /*
 241  * Enable timer based NMIs on all CPUs:
 242  */
 243 void acpi_nmi_enable(void)
 244 {
 245         if (atomic_read(&nmi_active) && nmi_watchdog == NMI_IO_APIC)
 246                 on_each_cpu(__acpi_nmi_enable, NULL, 0, 1);
 247 }
 248
 249 static void __acpi_nmi_disable(void *__unused)
 250 {
 251         apic_write_around(APIC_LVT0, APIC_DM_NMI | APIC_LVT_MASKED);
 252 }
 253
 254 /*
 255  * Disable timer based NMIs on all CPUs:
 256  */
 257 void acpi_nmi_disable(void)
 258 {
 259         if (atomic_read(&nmi_active) && nmi_watchdog == NMI_IO_APIC)
 260                 on_each_cpu(__acpi_nmi_disable, NULL, 0, 1);
 261 }
 262
 263 void setup_apic_nmi_watchdog(void *unused)
 264 {
 265         if (__get_cpu_var(wd_enabled))
 266                 return;
 267
 268         /* cheap hack to support suspend/resume */
 269         /* if cpu0 is not active neither should the other cpus */
 270         if ((smp_processor_id() != 0) && (atomic_read(&nmi_active) <= 0))
 271                 return;
 272
 273         switch (nmi_watchdog) {
 274         case NMI_LOCAL_APIC:
 275                 __get_cpu_var(wd_enabled) = 1; /* enable it before to avoid race with handler */
 276                 if (lapic_watchdog_init(nmi_hz) < 0) {
 277                         __get_cpu_var(wd_enabled) = 0;
 278                         return;
 279                 }
 280                 /* FALL THROUGH */
 281         case NMI_IO_APIC:
 282                 __get_cpu_var(wd_enabled) = 1;
 283                 atomic_inc(&nmi_active);
 284         }
 285 }
 286
 287 void stop_apic_nmi_watchdog(void *unused)
 288 {
 289         /* only support LOCAL and IO APICs for now */
 290         if ((nmi_watchdog != NMI_LOCAL_APIC) &&
 291             (nmi_watchdog != NMI_IO_APIC))
 292                 return;
 293         if (__get_cpu_var(wd_enabled) == 0)
 294                 return;
 295         if (nmi_watchdog == NMI_LOCAL_APIC)
 296                 lapic_watchdog_stop();
 297         __get_cpu_var(wd_enabled) = 0;
 298         atomic_dec(&nmi_active);
 299 }
 300
 301 /*
 302  * the best way to detect whether a CPU has a 'hard lockup' problem
 303  * is to check it's local APIC timer IRQ counts. If they are not
 304  * changing then that CPU has some problem.
 305  *
 306  * as these watchdog NMI IRQs are generated on every CPU, we only
 307  * have to check the current processor.
 308  *
 309  * since NMIs don't listen to _any_ locks, we have to be extremely
 310  * careful not to rely on unsafe variables. The printk might lock
 311  * up though, so we have to break up any console locks first ...
 312  * [when there will be more tty-related locks, break them up
 313  *  here too!]
 314  */
 315
 316 static DEFINE_PER_CPU(unsigned, last_irq_sum);
 317 static DEFINE_PER_CPU(local_t, alert_counter);
 318 static DEFINE_PER_CPU(int, nmi_touch);
 319
 320 void touch_nmi_watchdog(void)
 321 {
 322         if (nmi_watchdog > 0) {
 323                 unsigned cpu;
 324
 325                 /*
 326                  * Tell other CPUs to reset their alert counters. We cannot
 327                  * do it ourselves because the alert count increase is not
 328                  * atomic.
 329                  */
 330                 for_each_present_cpu(cpu) {
 331                         if (per_cpu(nmi_touch, cpu) != 1)
 332                                 per_cpu(nmi_touch, cpu) = 1;
 333                 }
 334         }
 335
 336         /*
 337          * Tickle the softlockup detector too:
 338          */
 339         touch_softlockup_watchdog();
 340 }
 341 EXPORT_SYMBOL(touch_nmi_watchdog);
 342
 343 notrace __kprobes int
 344 nmi_watchdog_tick(struct pt_regs *regs, unsigned reason)
 345 {
 346
 347         /*
 348          * Since current_thread_info()-> is always on the stack, and we
 349          * always switch the stack NMI-atomically, it's safe to use
 350          * smp_processor_id().
 351          */
 352         unsigned int sum;
 353         int touched = 0;
 354         int cpu = smp_processor_id();
 355         int rc = 0;
 356
 357         /* check for other users first */
 358         if (notify_die(DIE_NMI, "nmi", regs, reason, 2, SIGINT)
 359                         == NOTIFY_STOP) {
 360                 rc = 1;
 361                 touched = 1;
 362         }
 363
 364         if (cpu_isset(cpu, backtrace_mask)) {
 365                 static DEFINE_SPINLOCK(lock);   /* Serialise the printks */
 366
 367                 spin_lock(&lock);
 368                 printk("NMI backtrace for cpu %d\n", cpu);
 369                 dump_stack();
 370                 spin_unlock(&lock);
 371                 cpu_clear(cpu, backtrace_mask);
 372         }
 373
 374         /*
 375          * Take the local apic timer and PIT/HPET into account. We don't
 376          * know which one is active, when we have highres/dyntick on
 377          */
 378         sum = per_cpu(irq_stat, cpu).apic_timer_irqs +
 379                 per_cpu(irq_stat, cpu).irq0_irqs;
 380         if (__get_cpu_var(nmi_touch)) {
 381                 __get_cpu_var(nmi_touch) = 0;
 382                 touched = 1;
 383         }
 384
 385         /* if the none of the timers isn't firing, this cpu isn't doing much */
 386         if (!touched && __get_cpu_var(last_irq_sum) == sum) {
 387                 /*
 388                  * Ayiee, looks like this CPU is stuck ...
 389                  * wait a few IRQs (5 seconds) before doing the oops ...
 390                  */
 391                 local_inc(&__get_cpu_var(alert_counter));
 392                 if (local_read(&__get_cpu_var(alert_counter)) == 5*nmi_hz)
 393                         /*
 394                          * die_nmi will return ONLY if NOTIFY_STOP happens..
 395                          */
 396                         die_nmi("BUG: NMI Watchdog detected LOCKUP",
 397                                 regs, panic_on_timeout);
 398         } else {
 399                 __get_cpu_var(last_irq_sum) = sum;
 400                 local_set(&__get_cpu_var(alert_counter), 0);
 401         }
 402         /* see if the nmi watchdog went off */
 403         if (!__get_cpu_var(wd_enabled))
 404                 return rc;
 405         switch (nmi_watchdog) {
 406         case NMI_LOCAL_APIC:
 407                 rc |= lapic_wd_event(nmi_hz);
 408                 break;
 409         case NMI_IO_APIC:
 410                 /* don't know how to accurately check for this.
 411                  * just assume it was a watchdog timer interrupt
 412                  * This matches the old behaviour.
 413                  */
 414                 rc = 1;
 415                 break;
 416         }
 417         return rc;
 418 }
 419
 420 #ifdef CONFIG_SYSCTL
 421
 422 static int unknown_nmi_panic_callback(struct pt_regs *regs, int cpu)
 423 {
 424         unsigned char reason = get_nmi_reason();
 425         char buf[64];
 426
 427         sprintf(buf, "NMI received for unknown reason %02x\n", reason);
 428         die_nmi(buf, regs, 0);
 429         return 0;
 430 }
 431
 432 /*
 433  * proc handler for /proc/sys/kernel/nmi
 434  */
 435 int proc_nmi_enabled(struct ctl_table *table, int write, struct file *file,
 436                         void __user *buffer, size_t *length, loff_t *ppos)
 437 {
 438         int old_state;
 439
 440         nmi_watchdog_enabled = (atomic_read(&nmi_active) > 0) ? 1 : 0;
 441         old_state = nmi_watchdog_enabled;
 442         proc_dointvec(table, write, file, buffer, length, ppos);
 443         if (!!old_state == !!nmi_watchdog_enabled)
 444                 return 0;
 445
 446         if (atomic_read(&nmi_active) < 0 || nmi_watchdog == NMI_DISABLED) {
 447                 printk( KERN_WARNING "NMI watchdog is permanently disabled\n");
 448                 return -EIO;
 449         }
 450
 451         /* if nmi_watchdog is not set yet, then set it */
 452         nmi_watchdog_default();
 453
 454         if (nmi_watchdog == NMI_LOCAL_APIC) {
 455                 if (nmi_watchdog_enabled)
 456                         enable_lapic_nmi_watchdog();
 457                 else
 458                         disable_lapic_nmi_watchdog();
 459         } else {
 460                 printk( KERN_WARNING
 461                         "NMI watchdog doesn't know what hardware to touch\n");
 462                 return -EIO;
 463         }
 464         return 0;
 465 }
 466
 467 #endif
 468
 469 int do_nmi_callback(struct pt_regs *regs, int cpu)
 470 {
 471 #ifdef CONFIG_SYSCTL
 472         if (unknown_nmi_panic)
 473                 return unknown_nmi_panic_callback(regs, cpu);
 474 #endif
 475         return 0;
 476 }
 477
 478 void __trigger_all_cpu_backtrace(void)
 479 {
 480         int i;
 481
 482         backtrace_mask = cpu_online_map;
 483         /* Wait for up to 10 seconds for all CPUs to do the backtrace */
 484         for (i = 0; i < 10 * 1000; i++) {
 485                 if (cpus_empty(backtrace_mask))
 486                         break;
 487                 mdelay(1);
 488         }
 489 }
 490
 491 EXPORT_SYMBOL(nmi_active);
 492 EXPORT_SYMBOL(nmi_watchdog);