The Linux IPMI Driver
                          ---------------------
                              Corey Minyard
                          <minyard@mvista.com>
                            <minyard@acm.org>

The Intelligent Platform Management Interface, or IPMI, is a
standard for controlling intelligent devices that monitor a system.
It provides for dynamic discovery of sensors in the system and the
ability to monitor the sensors and be informed when the sensor's
values change or go outside certain boundaries.  It also has a
standardized database for field-replacable units (FRUs) and a watchdog
timer.

To use this, you need an interface to an IPMI controller in your
system (called a Baseboard Management Controller, or BMC) and
management software that can use the IPMI system.

This document describes how to use the IPMI driver for Linux.  If you
are not familiar with IPMI itself, see the web site at
http://www.intel.com/design/servers/ipmi/index.htm.  IPMI is a big
subject and I can't cover it all here!

Basic Design
------------

The Linux IPMI driver is designed to be very modular and flexible, you
only need to take the pieces you need and you can use it in many
different ways.  Because of that, it's broken into many chunks of
code.  These chunks are:

ipmi_msghandler - This is the central piece of software for the IPMI
system.  It handles all messages, message timing, and responses.  The
IPMI users tie into this, and the IPMI physical interfaces (called
System Management Interfaces, or SMIs) also tie in here.  This
provides the kernelland interface for IPMI, but does not provide an
interface for use by application processes.

ipmi_devintf - This provides a userland IOCTL interface for the IPMI
driver, each open file for this device ties in to the message handler
as an IPMI user.

ipmi_kcs_drv - A driver for the KCS SMI.  Most system have a KCS
interface for IPMI.


Much documentation for the interface is in the include files.  The
IPMI include files are:

ipmi.h - Contains the user interface and IOCTL interface for IPMI.

ipmi_smi.h - Contains the interface for SMI drivers to use.

ipmi_msgdefs.h - General definitions for base IPMI messaging.


Addressing
----------

The IPMI addressing works much like IP addresses, you have an overlay
to handle the different address types.  The overlay is:

  struct ipmi_addr
  {
        int   addr_type;
        short channel;
        char  data[IPMI_MAX_ADDR_SIZE];
  };

The addr_type determines what the address really is.  The driver
currently understands two different types of addresses.

"System Interface" addresses are defined as:

  struct ipmi_system_interface_addr
  {
        int   addr_type;
        short channel;
  };

and the type is IPMI_SYSTEM_INTERFACE_ADDR_TYPE.  This is used for talking
straight to the BMC on the current card.  The channel must be
IPMI_BMC_CHANNEL.

Messages that are destined to go out on the IPMB bus use the
IPMI_IPMB_ADDR_TYPE address type.  The format is

  struct ipmi_ipmb_addr
  {
        int           addr_type;
        short         channel;
        unsigned char slave_addr;
        unsigned char lun;
  };

The "channel" here is generally zero, but some devices support more
than one channel, it corresponds to the channel as defined in the IPMI
spec.


Messages
--------

Messages are defined as:

struct ipmi_msg
{
        unsigned char netfn;
        unsigned char lun;
        unsigned char cmd;
        unsigned char *data;
        int           data_len;
};

The driver takes care of adding/stripping the header information.  The
data portion is just the data to be send (do NOT put addressing info
here) or the response.  Note that the completion code of a response is
the first item in "data", it is not stripped out because that is how
all the messages are defined in the spec (and thus makes counting the
offsets a little easier :-).

When using the IOCTL interface from userland, you must provide a block
of data for "data", fill it, and set data_len to the length of the
block of data, even when receiving messages.  Otherwise the driver
will have no place to put the message.

Messages coming up from the message handler in kernelland will come in
as:

  struct ipmi_recv_msg
  {
        struct list_head link;

        /* The type of message as defined in the "Receive Types"
           defines above. */
        int         recv_type;

        ipmi_user_t      *user;
        struct ipmi_addr addr;
        long             msgid;
        struct ipmi_msg  msg;

        /* Call this when done with the message.  It will presumably free
           the message and do any other necessary cleanup. */
        void (*done)(struct ipmi_recv_msg *msg);

        /* Place-holder for the data, don't make any assumptions about
           the size or existence of this, since it may change. */
        unsigned char   msg_data[IPMI_MAX_MSG_LENGTH];
  };

You should look at the receive type and handle the message
appropriately.


The Upper Layer Interface (Message Handler)
-------------------------------------------

The upper layer of the interface provides the users with a consistent
view of the IPMI interfaces.  It allows multiple SMI interfaces to be
addressed (because some boards actually have multiple BMCs on them)
and the user should not have to care what type of SMI is below them.


Creating the User

To user the message handler, you must first create a user using
ipmi_create_user.  The interface number specifies which SMI you want
to connect to, and you must supply callback functions to be called
when data comes in.  The callback function can run at interrupt level,
so be careful using the callbacks.  This also allows to you pass in a
piece of data, the handler_data, that will be passed back to you on
all calls.

Once you are done, call ipmi_destroy_user() to get rid of the user.

From userland, opening the device automatically creates a user, and
closing the device automatically destroys the user.


Messaging

To send a message from kernel-land, the ipmi_request() call does
pretty much all message handling.  Most of the parameter are
self-explanatory.  However, it takes a "msgid" parameter.  This is NOT
the sequence number of messages.  It is simply a long value that is
passed back when the response for the message is returned.  You may
use it for anything you like.

Responses come back in the function pointed to by the ipmi_recv_hndl
field of the "handler" that you passed in to ipmi_create_user().
Remember again, these may be running at interrupt level.  Remember to
look at the receive type, too.

From userland, you fill out an ipmi_req_t structure and use the
IPMICTL_SEND_COMMAND ioctl.  For incoming stuff, you can use select()
or poll() to wait for messages to come in.  However, you cannot use
read() to get them, you must call the IPMICTL_RECEIVE_MSG with the
ipmi_recv_t structure to actually get the message.  Remember that you
must supply a pointer to a block of data in the msg.data field, and
you must fill in the msg.data_len field with the size of the data.
This gives the receiver a place to actually put the message.

If the message cannot fit into the data you provide, you will get an
EMSGSIZE error and the driver will leave the data in the receive
queue.  If you want to get it and have it truncate the message, us
the IPMICTL_RECEIVE_MSG_TRUNC ioctl.

When you send a command (which is defined by the lowest-order bit of
the netfn per the IPMI spec) on the IPMB bus, the driver will
automatically assign the sequence number to the command and save the
command.  If the response is not receive in the IPMI-specified 5
seconds, it will generate a response automatically saying the command
timed out.  If an unsolicited response comes in (if it was after 5
seconds, for instance), that response will be ignored.

In kernelland, after you receive a message and are done with it, you
MUST call ipmi_free_recv_msg() on it, or you will leak messages.  Note
that you should NEVER mess with the "done" field of a message, that is
required to properly clean up the message.

Note that when sending, there is an ipmi_request_supply_msgs() call
that lets you supply the smi and receive message.  This is useful for
pieces of code that need to work even if the system is out of buffers
(the watchdog timer uses this, for instance).  You supply your own
buffer and own free routines.  This is not recommended for normal use,
though, since it is tricky to manage your own buffers.


Events and Incoming Commands

The driver takes care of polling for IPMI events and receiving
commands (commands are messages that are not responses, they are
commands that other things on the IPMB bus have sent you).  To receive
these, you must register for them, they will not automatically be sent
to you.

To receive events, you must call ipmi_set_gets_events() and set the
"val" to non-zero.  Any events that have been received by the driver
since startup will immediately be delivered to the first user that
registers for events.  After that, if multiple users are registered
for events, they will all receive all events that come in.

For receiving commands, you have to individually register commands you
want to receive.  Call ipmi_register_for_cmd() and supply the netfn
and command name for each command you want to receive.  Only one user
may be registered for each netfn/cmd, but different users may register
for different commands.

From userland, equivalent IOCTLs are provided to do these functions.


The Lower Layer (SMI) Interface
-------------------------------

As mentioned before, multiple SMI interfaces may be registered to the
message handler, each of these is assigned an interface number when
they register with the message handler.  They are generally assigned
in the order they register, although if an SMI unregisters and then
another one registers, all bets are off.

The ipmi_smi.h defines the interface for SMIs, see that for more
details.


The KCS Driver
--------------

The KCS driver allows up to 4 KCS interfaces to be configured in the
system.  By default, the driver will register one KCS interface at the
spec-specified I/O port 0xca2 without interrupts.  You can change this
at module load time (for a module) with:

  insmod ipmi_kcs_drv.o kcs_ports=<port1>,<port2>... kcs_addrs=<addr1>,<addr2>
       kcs_irqs=<irq1>,<irq2>... kcs_trydefaults=[0|1]

The KCS driver supports two types of interfaces, ports (for I/O port
based KCS interfaces) and memory addresses (for KCS interfaces in
memory).  The driver will support both of them simultaneously, setting
the port to zero (or just not specifying it) will allow the memory
address to be used.  The port will override the memory address if it
is specified and non-zero.  kcs_trydefaults sets whether the standard
IPMI interface at 0xca2 and any interfaces specified by ACPE are
tried.  By default, the driver tries it, set this value to zero to
turn this off.

When compiled into the kernel, the addresses can be specified on the
kernel command line as:

  ipmi_kcs=<bmc1>:<irq1>,<bmc2>:<irq2>....,[nodefault]

The <bmcx> values is either "p<port>" or "m<addr>" for port or memory
addresses.  So for instance, a KCS interface at port 0xca2 using
interrupt 9 and a memory interface at address 0xf9827341 with no
interrupt would be specified "ipmi_kcs=p0xca2:9,m0xf9827341".
If you specify zero for in irq or don't specify it, the driver will
run polled unless the software can detect the interrupt to use in the
ACPI tables.

By default, the driver will attempt to detect a KCS device at the
spec-specified 0xca2 address and any address specified by ACPI.  If
you want to turn this off, use the "nodefault" option.

If you have high-res timers compiled into the kernel, the driver will
use them to provide much better performance.  Note that if you do not
have high-res timers enabled in the kernel and you don't have
interrupts enabled, the driver will run VERY slowly.  Don't blame me,
the KCS interface sucks.


Other Pieces
------------

Watchdog

A watchdog timer is provided that implements the Linux-standard
watchdog timer interface.  It has three module parameters that can be
used to control it:

  insmod ipmi_watchdog timeout=<t> pretimeout=<t> action=<action type>
      preaction=<preaction type> preop=<preop type>

The timeout is the number of seconds to the action, and the pretimeout
is the amount of seconds before the reset that the pre-timeout panic will
occur (if pretimeout is zero, then pretimeout will not be enabled).

The action may be "reset", "power_cycle", or "power_off", and
specifies what to do when the timer times out, and defaults to
"reset".

The preaction may be "pre_smi" for an indication through the SMI
interface, "pre_int" for an indication through the SMI with an
interrupts, and "pre_nmi" for a NMI on a preaction.  This is how
the driver is informed of the pretimeout.

The preop may be set to "preop_none" for no operation on a pretimeout,
"preop_panic" to set the preoperation to panic, or "preop_give_data"
to provide data to read from the watchdog device when the pretimeout
occurs.  A "pre_nmi" setting CANNOT be used with "preop_give_data"
because you can't do data operations from an NMI.

When preop is set to "preop_give_data", one byte comes ready to read
on the device when the pretimeout occurs.  Select and fasync work on
the device, as well.

When compiled into the kernel, the kernel command line is available
for configuring the watchdog:

  ipmi_wdog=<timeout>[,<pretimeout>[,<option>[,<options>....]]]

The options are the actions and preaction above (if an option
controlling the same thing is specified twice, the last is taken).  An
options "start_now" is also there, if included, the watchdog will
start running immediately when all the drivers are ready, it doesn't
have to have a user hooked up to start it.

The watchdog will panic and start a 120 second reset timeout if it
gets a pre-action.  During a panic or a reboot, the watchdog will
start a 120 timer if it is running to make sure the reboot occurs.

Note that if you use the NMI preaction for the watchdog, you MUST
NOT use nmi watchdog mode 1.  If you use the NMI watchdog, you
must use mode 2.