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 #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 static int panic_on_timeout;
  44
  45 unsigned int nmi_watchdog = NMI_DEFAULT;
  46 static unsigned int nmi_hz = HZ;
  47
  48 static DEFINE_PER_CPU(short, wd_enabled);
  49
  50 /* Run after command line and cpu_init init, but before all other checks */
  51 void nmi_watchdog_default(void)
  52 {
  53         if (nmi_watchdog != NMI_DEFAULT)
  54                 return;
  55         nmi_watchdog = NMI_NONE;
  56 }
  57
  58 static int endflag __initdata = 0;
  59
  60 #ifdef CONFIG_SMP
  61 /* The performance counters used by NMI_LOCAL_APIC don't trigger when
  62  * the CPU is idle. To make sure the NMI watchdog really ticks on all
  63  * CPUs during the test make them busy.
  64  */
  65 static __init void nmi_cpu_busy(void *data)
  66 {
  67         local_irq_enable_in_hardirq();
  68         /* Intentionally don't use cpu_relax here. This is
  69            to make sure that the performance counter really ticks,
  70            even if there is a simulator or similar that catches the
  71            pause instruction. On a real HT machine this is fine because
  72            all other CPUs are busy with "useless" delay loops and don't
  73            care if they get somewhat less cycles. */
  74         while (endflag == 0)
  75                 mb();
  76 }
  77 #endif
  78
  79 int __init check_nmi_watchdog(void)
  80 {
  81         unsigned int *prev_nmi_count;
  82         int cpu;
  83
  84         if ((nmi_watchdog == NMI_NONE) || (nmi_watchdog == NMI_DISABLED))
  85                 return 0;
  86
  87         if (!atomic_read(&nmi_active))
  88                 return 0;
  89
  90         prev_nmi_count = kmalloc(NR_CPUS * sizeof(int), GFP_KERNEL);
  91         if (!prev_nmi_count)
  92                 return -1;
  93
  94         printk(KERN_INFO "Testing NMI watchdog ... ");
  95
  96 #ifdef CONFIG_SMP
  97         if (nmi_watchdog == NMI_LOCAL_APIC)
  98                 smp_call_function(nmi_cpu_busy, (void *)&endflag, 0, 0);
  99 #endif
 100
 101         for_each_possible_cpu(cpu)
 102                 prev_nmi_count[cpu] = cpu_pda(cpu)->__nmi_count;
 103         local_irq_enable();
 104         mdelay((20*1000)/nmi_hz); // wait 20 ticks
 105
 106         for_each_online_cpu(cpu) {
 107                 if (!per_cpu(wd_enabled, cpu))
 108                         continue;
 109                 if (cpu_pda(cpu)->__nmi_count - prev_nmi_count[cpu] <= 5) {
 110                         printk(KERN_WARNING "WARNING: CPU#%d: NMI "
 111                                "appears to be stuck (%d->%d)!\n",
 112                                 cpu,
 113                                 prev_nmi_count[cpu],
 114                                 cpu_pda(cpu)->__nmi_count);
 115                         per_cpu(wd_enabled, cpu) = 0;
 116                         atomic_dec(&nmi_active);
 117                 }
 118         }
 119         endflag = 1;
 120         if (!atomic_read(&nmi_active)) {
 121                 kfree(prev_nmi_count);
 122                 atomic_set(&nmi_active, -1);
 123                 return -1;
 124         }
 125         printk("OK.\n");
 126
 127         /* now that we know it works we can reduce NMI frequency to
 128            something more reasonable; makes a difference in some configs */
 129         if (nmi_watchdog == NMI_LOCAL_APIC)
 130                 nmi_hz = lapic_adjust_nmi_hz(1);
 131
 132         kfree(prev_nmi_count);
 133         return 0;
 134 }
 135
 136 static int __init setup_nmi_watchdog(char *str)
 137 {
 138         int nmi;
 139
 140         if (!strncmp(str,"panic",5)) {
 141                 panic_on_timeout = 1;
 142                 str = strchr(str, ',');
 143                 if (!str)
 144                         return 1;
 145                 ++str;
 146         }
 147
 148         get_option(&str, &nmi);
 149
 150         if ((nmi >= NMI_INVALID) || (nmi < NMI_NONE))
 151                 return 0;
 152
 153         nmi_watchdog = nmi;
 154         return 1;
 155 }
 156
 157 __setup("nmi_watchdog=", setup_nmi_watchdog);
 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 static struct sysdev_class nmi_sysclass = {
 183         .name           = "lapic_nmi",
 184         .resume         = lapic_nmi_resume,
 185         .suspend        = lapic_nmi_suspend,
 186 };
 187
 188 static struct sys_device device_lapic_nmi = {
 189         .id     = 0,
 190         .cls    = &nmi_sysclass,
 191 };
 192
 193 static int __init init_lapic_nmi_sysfs(void)
 194 {
 195         int error;
 196
 197         /* should really be a BUG_ON but b/c this is an
 198          * init call, it just doesn't work.  -dcz
 199          */
 200         if (nmi_watchdog != NMI_LOCAL_APIC)
 201                 return 0;
 202
 203         if (atomic_read(&nmi_active) < 0)
 204                 return 0;
 205
 206         error = sysdev_class_register(&nmi_sysclass);
 207         if (!error)
 208                 error = sysdev_register(&device_lapic_nmi);
 209         return error;
 210 }
 211 /* must come after the local APIC's device_initcall() */
 212 late_initcall(init_lapic_nmi_sysfs);
 213
 214 #endif  /* CONFIG_PM */
 215
 216 static void __acpi_nmi_enable(void *__unused)
 217 {
 218         apic_write_around(APIC_LVT0, APIC_DM_NMI);
 219 }
 220
 221 /*
 222  * Enable timer based NMIs on all CPUs:
 223  */
 224 void acpi_nmi_enable(void)
 225 {
 226         if (atomic_read(&nmi_active) && nmi_watchdog == NMI_IO_APIC)
 227                 on_each_cpu(__acpi_nmi_enable, NULL, 0, 1);
 228 }
 229
 230 static void __acpi_nmi_disable(void *__unused)
 231 {
 232         apic_write_around(APIC_LVT0, APIC_DM_NMI | APIC_LVT_MASKED);
 233 }
 234
 235 /*
 236  * Disable timer based NMIs on all CPUs:
 237  */
 238 void acpi_nmi_disable(void)
 239 {
 240         if (atomic_read(&nmi_active) && nmi_watchdog == NMI_IO_APIC)
 241                 on_each_cpu(__acpi_nmi_disable, NULL, 0, 1);
 242 }
 243
 244 void setup_apic_nmi_watchdog(void *unused)
 245 {
 246         if (__get_cpu_var(wd_enabled))
 247                 return;
 248
 249         /* cheap hack to support suspend/resume */
 250         /* if cpu0 is not active neither should the other cpus */
 251         if ((smp_processor_id() != 0) && (atomic_read(&nmi_active) <= 0))
 252                 return;
 253
 254         switch (nmi_watchdog) {
 255         case NMI_LOCAL_APIC:
 256                 __get_cpu_var(wd_enabled) = 1;
 257                 if (lapic_watchdog_init(nmi_hz) < 0) {
 258                         __get_cpu_var(wd_enabled) = 0;
 259                         return;
 260                 }
 261                 /* FALL THROUGH */
 262         case NMI_IO_APIC:
 263                 __get_cpu_var(wd_enabled) = 1;
 264                 atomic_inc(&nmi_active);
 265         }
 266 }
 267
 268 void stop_apic_nmi_watchdog(void *unused)
 269 {
 270         /* only support LOCAL and IO APICs for now */
 271         if ((nmi_watchdog != NMI_LOCAL_APIC) &&
 272             (nmi_watchdog != NMI_IO_APIC))
 273                 return;
 274         if (__get_cpu_var(wd_enabled) == 0)
 275                 return;
 276         if (nmi_watchdog == NMI_LOCAL_APIC)
 277                 lapic_watchdog_stop();
 278         __get_cpu_var(wd_enabled) = 0;
 279         atomic_dec(&nmi_active);
 280 }
 281
 282 /*
 283  * the best way to detect whether a CPU has a 'hard lockup' problem
 284  * is to check it's local APIC timer IRQ counts. If they are not
 285  * changing then that CPU has some problem.
 286  *
 287  * as these watchdog NMI IRQs are generated on every CPU, we only
 288  * have to check the current processor.
 289  */
 290
 291 static DEFINE_PER_CPU(unsigned, last_irq_sum);
 292 static DEFINE_PER_CPU(local_t, alert_counter);
 293 static DEFINE_PER_CPU(int, nmi_touch);
 294
 295 void touch_nmi_watchdog(void)
 296 {
 297         if (nmi_watchdog > 0) {
 298                 unsigned cpu;
 299
 300                 /*
 301                  * Tell other CPUs to reset their alert counters. We cannot
 302                  * do it ourselves because the alert count increase is not
 303                  * atomic.
 304                  */
 305                 for_each_present_cpu(cpu) {
 306                         if (per_cpu(nmi_touch, cpu) != 1)
 307                                 per_cpu(nmi_touch, cpu) = 1;
 308                 }
 309         }
 310
 311         touch_softlockup_watchdog();
 312 }
 313 EXPORT_SYMBOL(touch_nmi_watchdog);
 314
 315 notrace __kprobes int
 316 nmi_watchdog_tick(struct pt_regs *regs, unsigned reason)
 317 {
 318         unsigned int sum;
 319         int touched = 0;
 320         int cpu = smp_processor_id();
 321         int rc = 0;
 322
 323         /* check for other users first */
 324         if (notify_die(DIE_NMI, "nmi", regs, reason, 2, SIGINT)
 325                         == NOTIFY_STOP) {
 326                 rc = 1;
 327                 touched = 1;
 328         }
 329
 330         sum = read_pda(apic_timer_irqs) + read_pda(irq0_irqs);
 331         if (__get_cpu_var(nmi_touch)) {
 332                 __get_cpu_var(nmi_touch) = 0;
 333                 touched = 1;
 334         }
 335
 336         if (cpu_isset(cpu, backtrace_mask)) {
 337                 static DEFINE_SPINLOCK(lock);   /* Serialise the printks */
 338
 339                 spin_lock(&lock);
 340                 printk("NMI backtrace for cpu %d\n", cpu);
 341                 dump_stack();
 342                 spin_unlock(&lock);
 343                 cpu_clear(cpu, backtrace_mask);
 344         }
 345
 346 #ifdef CONFIG_X86_MCE
 347         /* Could check oops_in_progress here too, but it's safer
 348            not too */
 349         if (atomic_read(&mce_entry) > 0)
 350                 touched = 1;
 351 #endif
 352         /* if the apic timer isn't firing, this cpu isn't doing much */
 353         if (!touched && __get_cpu_var(last_irq_sum) == sum) {
 354                 /*
 355                  * Ayiee, looks like this CPU is stuck ...
 356                  * wait a few IRQs (5 seconds) before doing the oops ...
 357                  */
 358                 local_inc(&__get_cpu_var(alert_counter));
 359                 if (local_read(&__get_cpu_var(alert_counter)) == 5*nmi_hz)
 360                         die_nmi("NMI Watchdog detected LOCKUP",
 361                                 regs, panic_on_timeout);
 362         } else {
 363                 __get_cpu_var(last_irq_sum) = sum;
 364                 local_set(&__get_cpu_var(alert_counter), 0);
 365         }
 366
 367         /* see if the nmi watchdog went off */
 368         if (!__get_cpu_var(wd_enabled))
 369                 return rc;
 370         switch (nmi_watchdog) {
 371         case NMI_LOCAL_APIC:
 372                 rc |= lapic_wd_event(nmi_hz);
 373                 break;
 374         case NMI_IO_APIC:
 375                 /* don't know how to accurately check for this.
 376                  * just assume it was a watchdog timer interrupt
 377                  * This matches the old behaviour.
 378                  */
 379                 rc = 1;
 380                 break;
 381         }
 382         return rc;
 383 }
 384
 385 #ifdef CONFIG_SYSCTL
 386
 387 static int unknown_nmi_panic_callback(struct pt_regs *regs, int cpu)
 388 {
 389         unsigned char reason = get_nmi_reason();
 390         char buf[64];
 391
 392         sprintf(buf, "NMI received for unknown reason %02x\n", reason);
 393         die_nmi(buf, regs, 1);  /* Always panic here */
 394         return 0;
 395 }
 396
 397 /*
 398  * proc handler for /proc/sys/kernel/nmi
 399  */
 400 int proc_nmi_enabled(struct ctl_table *table, int write, struct file *file,
 401                         void __user *buffer, size_t *length, loff_t *ppos)
 402 {
 403         int old_state;
 404
 405         nmi_watchdog_enabled = (atomic_read(&nmi_active) > 0) ? 1 : 0;
 406         old_state = nmi_watchdog_enabled;
 407         proc_dointvec(table, write, file, buffer, length, ppos);
 408         if (!!old_state == !!nmi_watchdog_enabled)
 409                 return 0;
 410
 411         if (atomic_read(&nmi_active) < 0 || nmi_watchdog == NMI_DISABLED) {
 412                 printk( KERN_WARNING "NMI watchdog is permanently disabled\n");
 413                 return -EIO;
 414         }
 415
 416         /* if nmi_watchdog is not set yet, then set it */
 417         nmi_watchdog_default();
 418
 419         if (nmi_watchdog == NMI_LOCAL_APIC) {
 420                 if (nmi_watchdog_enabled)
 421                         enable_lapic_nmi_watchdog();
 422                 else
 423                         disable_lapic_nmi_watchdog();
 424         } else {
 425                 printk( KERN_WARNING
 426                         "NMI watchdog doesn't know what hardware to touch\n");
 427                 return -EIO;
 428         }
 429         return 0;
 430 }
 431
 432 #endif
 433
 434 int do_nmi_callback(struct pt_regs *regs, int cpu)
 435 {
 436 #ifdef CONFIG_SYSCTL
 437         if (unknown_nmi_panic)
 438                 return unknown_nmi_panic_callback(regs, cpu);
 439 #endif
 440         return 0;
 441 }
 442
 443 void __trigger_all_cpu_backtrace(void)
 444 {
 445         int i;
 446
 447         backtrace_mask = cpu_online_map;
 448         /* Wait for up to 10 seconds for all CPUs to do the backtrace */
 449         for (i = 0; i < 10 * 1000; i++) {
 450                 if (cpus_empty(backtrace_mask))
 451                         break;
 452                 mdelay(1);
 453         }
 454 }
 455
 456 EXPORT_SYMBOL(nmi_active);
 457 EXPORT_SYMBOL(nmi_watchdog);