Non Maskable Interrupt

The Non-Maskable Interrupt (NMI) is a hardware-driven interrupt much like the PIC interrupts, but the NMI goes either directly to the CPU, or via another controller (e.g., the ISP)---in which case it can be masked.

About

NMIs occur for RAM errors and unrecoverable hardware problems. For newer computers these things may be handled using machine check exceptions and/or SMI. For the newest chipsets (at least for Intel) there's also a pile of TCO stuff ("total cost of ownership") that is tied into it all (with a special "TCO IRQ" and connections to SMI/SMM, etc). All of the TCO stuff can be connected to an onboard ethernet controller, and (at least part of it) is intended for remote monitoring of the system. Unfortunately, the chipset documentation doesn't say how BIOSes normally configure the chipset, and the chipsets themselves support several different options in each case. For example, for a RAM error it could be handled by the chipset itself, it could generate an SMI (where the BIOS/SMM handler does "RAM scrubbing" in software), it could generate a "TCO interrupt", etc. If you add it all up it's a huge complex mess (TCO + SMI + SMBus + Northbridge + PCI bus/controller/s + PCI-to-LPC-bridge + god-knows-what) that can be completely different between motherboards (even motherboards with the same chipset).

The short version of this story is that there's only really 2 reasons for an NMI. The first reason is a hardware failure. The second reason is a "watchdog timer", which can be used to detect when the kernel itself locks up (and is sometimes also used for more accurate profiling as it allows EIP to be sampled even when IRQs are disabled).

If a hardware failure caused an NMI then there's no way to figure out which piece of hardware caused the NMI. In this case you may want to inform the user that a hardware error has occurred, and then the kernel should shutdown/reset the machine.

For the watchdog timer, it must be setup by the OS first. This can actually be done even when the chipset itself doesn't have a special watchdog timer for it (e.g. setting the PIT, RTC/CMOS IRQ or a HPET IRQ to "NMI, send to all CPUs" in the I/O APIC). In this case you want the watchdog timer to be fast (i.e. no slow hardware task switching and cache flushing) and you'd also want all CPUs to share the same timer, which means all CPUs would receive the same IRQ at the same time.

As an alternative, you could also use the local APIC's timer or the performance monitoring counter overflow for a "per CPU" watchdog timer.

Usage

The NMI is enabled (set high) by the memory module when a memory parity error occurs.

Be careful about disabling the NMI and the PIC for extended periods of time (mind you, watchdog timers typically use NMIs).

On the XT the NMI can be masked by setting bit 7 on I/O port 0xA0. On the AT the NMI can be masked by setting bit 7 on I/O port 0x70. This port is shared with the CMOS RAM index register using bits 0 through 6 of I/O port 0x70. The CMOS RTC expects a read from or write to the data port 0x71 after any write to index port 0x70 or it may go into an undefined state. There may also need to be an I/O delay between accessing the index and data registers. The index port 0x70 may be a write-only port and always return 0xFF on read. Hence the bit masking below to preserve bits 0 through 6 of the CMOS index register may not work, nor may it be possible to retrieve the current state of the NMI mask from port 0x70.

 void NMI_enable() {
    outb(0x70, inb(0x70) & 0x7F);
    inb(0x71);
 }

 void NMI_disable() {
    outb(0x70, inb(0x70) | 0x80);
    inb(0x71);
 }

When an NMI occurs you can check the system control port A and B at I/O addresses 0x92 and 0x61 respectively to get an indication of what caused the error:

System Control Port A (0x92) layout:

BIT	Description
0	Alternate hot reset
1	Alternate gate A20
2	Reserved
3	Security Lock
4*	Watchdog timer status
5	Reserved
6	HDD 2 drive activity
7	HDD 1 drive activity

System Control Port B (0x61)

Bit	Description
0	Timer 2 tied to speaker
1	Speaker data enable
2	Parity check enable
3	Channel check enable
4	Refresh request
5	Timer 2 output
6*	Channel check
7*	Parity check

The important bits are indicated with an '*'. The Channel Check bit indicates a failure on the bus, probably by a peripheral device such as a modem, sound card, NIC, etc, while the Parity check bit indicates a memory read or write failure.