| NMI(9) | Kernel Developer's Manual (x86) | NMI(9) | 
nmi, nmi_establish,
  nmi_disestablish, —
#include <x86/nmi.h>
nmi_handler_t *
  
  nmi_establish(int
    (*func)(const struct trapframe *, void *),
    void *arg);
void
  
  nmi_disestablish(nmi_handler_t
    *handle);
nmi interface lets the kernel establish handlers for
  x86 Non-Maskable Interrupts (NMIs). An NMI signals to the processor an
  exception on a processor, memory controller, or I/O bus that is irrecoverable
  or else needs attention at a high priority. A “debug switch” or
  a performance/watchdog timer may also trigger an NMI.
An NMI handler will run to completion on the same processor where it began without being preempted by any thread or interrupt except for another NMI. An NMI handler must prepare for re-entry. An NMI handler may run simultaneously on more than one CPU.
Synchronizing access to a shared data structure from an NMI handler is a different challenge than synchronizing access from hardware/software interrupt routines or from kernel threads. An NMI handler may not perform any operation that may sleep, acquire a mutex, or schedule a software interrupt. An NMI handler may use atomic_ops(3). An NMI handler may reference per-CPU storage (percpu(9)).
An NMI handler may not write to the kernel message buffer.
nmi_establish(func,
    arg)nmi_establish(), the kernel will call
      (*func)(tf,
      arg); every time an NMI occurs until the handler is
      removed with nmi_disestablish().
      func should return non-zero if it handled a
      condition that causes NMI, or zero if it did not. If, for a given NMI, all
      handlers return zero, the system will panic or enter the kernel debugger,
      ddb(4).
      nmi_establish() returns
      NULL on failure, and a handle for the NMI handler
      on success.nmi_disestablish(handle)nmi_establish().nmi interface is implemented within the file
  sys/arch/x86/x86/nmi.c.
nmi interface first appeared in
  NetBSD 6.0.
| March 17, 2011 | NetBSD 9.4 |