nut-debian/docs/design.txt

NUT design document
===================

This software is designed around a layered scheme with drivers, a
server and clients.   These layers communicate with text-based
protocols for easier maintenance and diagnostics.

The layering
------------

image:images/nut_layering.png[NUT layering]


How information gets around
---------------------------

From the equipment
~~~~~~~~~~~~~~~~~~

DRIVERS talk to the EQUIPMENT and receive updates.  For most hardware this
is polled (DRIVER asks EQUIPMENT about a variable), but forced updates are
also possible.  The exact method is not important, as it is abstracted
by the driver.

From the driver
~~~~~~~~~~~~~~~

The core of all DRIVERS maintains internal storage for every variable
that is known along with the auxiliary data for those variables.  It
sends updates to this data to any process which connects to the Unix
domain socket.

The DRIVERS will also provide a full atomic copy of their internal
knowledge upon receiving the "DUMPALL" command on the socket.  The dump
is in the same format as updates, and is followed by "DUMPDONE".  When
"DUMPDONE" has been received, the view is complete.

The SERVER will connect to the socket of each DRIVER and will request a
dump at that time.  It retains this data in local storage for later use.
It continues to listen on the socket for additional updates.

This protocol is documented in link:sock-protocol.txt[].

From the server
~~~~~~~~~~~~~~~

The SERVER's internal storage maintains a complete copy of the data
which is in the DRIVER, so it is capable of answering any request
immediately.  When a request for data arrives from a CLIENT, the SERVER
looks through the internal storage for that UPS and returns the
requested data if it is available.

The format for requests from the CLIENT is documented in link:protocol.txt[].

Instant commands
----------------

"Instant commands" is the term given to a set of actions that result in
something happening to the UPS.  Some of the common ones are
`test.battery.start` to initiate a battery test and `test.panel.start` to
test the front panel of the UPS.

They are passed to the SERVER from a CLIENT using an authenticated
network connection.  The SERVER first checks to make sure that the instant
command is valid for the DRIVER.  If it's supported, a message is sent
via a socket to the DRIVER containing the command and any auxiliary
information.

At this point, there is no confirmation to the SERVER of the command's
execution.  This is (still) planned for a future release.  This has been
delayed since returning a response involves some potentially interesting
timing issues.  Remember that `upsd` services clients in a round-robin
fashion, so all queries must be lightweight and speedy.

NOTE: FIXME: Wasn't "TRACKING" mechanism for "INSTCMD/SET VAR" introduced
to address just this? See https://github.com/networkupstools/nut/pull/659

Setting variables
-----------------

Some variables in the DRIVER or EQUIPMENT can be changed, and carry the
FLAG_RW flag.  Upon receiving a SET command from the CLIENT, the SERVER
first verifies that it is valid for that DRIVER in terms of writability
and data type.  If those checks pass, it then sends the SET command
through the socket, much like the instant command design.

The DRIVER is expected to commit the value to the EQUIPMENT and update
its internal representation of that variable.

Like the instant commands, there is currently no acknowledgement of the
command's completion from the DRIVER.  This, too, is planned for a future
release.

NOTE: FIXME: Wasn't "TRACKING" mechanism for "INSTCMD/SET VAR" introduced
to address just this? See https://github.com/networkupstools/nut/pull/659

Example data path
-----------------

Here's the path a piece of data might take through this architecture.
The event is a UPS going on battery, and the final result is a pager
delivering the alpha message to the admin.

1. EQUIPMENT reports on battery by setting flag in status register

2. DRIVER notices this flag and stores it in the `ups.status` variable as
   OB.  This update gets pushed out to any listeners via the sockets.

3. SERVER `upsd` sees activity on the socket, reads it, parses it, and
   commits the new data to its local version of the status variable.

4. CLIENT `upsmon` does a routine poll of SERVER for `ups.status` and
   gets `OB`.

5. CLIENT `upsmon` then invokes its `NOTIFYCMD` which is `upssched`.

6. `upssched` starts up a daemon to handle a timer which will expire about
   30 seconds into the future.

7. 30 seconds later, the timer expires since the UPS is still on battery,
   and so `upssched` calls the `CMDSCRIPT` which is `upssched-cmd`.

8. `upssched-cmd` parses the args and calls `sendmail`.

9. Avian carriers, smoke signals, SMTP, and some magic result in the
   message getting from the pager company's gateway to a transmitter
   and then to the admin's pager.

This scenario requires some configuration, obviously:

1. There's an UPS driver running.
   (Whatever applies for the hardware)

2. `upsd` has a valid UPS entry in 'ups.conf' for this UPS.

	[myups]
		driver = nutupsdrv
		port = /dev/ttySx

3. `upsd` has a valid user for `upsmon` in 'upsd.users' file.

	[monuser]
		password = somepass
		upsmon primary

4. `upsmon` is set to monitor this UPS with this user in 'upsmon.conf' file.

	MONITOR myups@localhost 1 monuser somepass primary

5. `upsmon` is set to `EXEC` the `NOTIFYCMD` for the `ONBATT` condition in
   'upsmon.conf' file.

	NOTIFYFLAG ONBATT EXEC

6. `upsmon` calls `upssched` as the `NOTIFYCMD` in 'upsmon.conf' file.

	NOTIFYCMD /path/to/upssched

7. `upssched` has a 30 second timer for `ONBATT` in 'upssched.conf' file.

	AT ONBATT * START-TIMER upsonbatt 30

8. `upssched` calls `upssched-cmd` as the `CMDSCRIPT` in 'upssched.conf'.

	CMDSCRIPT /path/to/upssched-cmd

9. `upssched-cmd` knows what to do with `upsonbatt` keyword as its first
   argument (a quick `case..esac` construct, see the examples)


History
-------

The oldest versions of this software (1998) had no separation between
the driver and the network server, and only supported the latest APC
Smart-UPS hardware as a result.  The network protocol used brittle
binary structs.  This had numerous bad implications for compatibility
and portability.

After the driver and server were separated, data was shared through the
state file concept.  Status was written into a static array (the "info
array") by drivers, and that array was stored on disk.  The `upsd` would
periodically read that file into a local copy of that array.

Shared memory mode was added a bit later, and that removed some of the
lag from the status updates.  Unfortunately, it didn't have any locking
originally, and the possibility for corruption due to races existed.

`mmap()` support was added at some point after that, and became the
default.  The drivers and `upsd` would `mmap()` the file into memory and
read or write from it.  Locking was done using the state file as the
token, so contention problems were avoided.  This method was relatively
quick, but it involved at least 3 copies of the data (driver, disk/mmap,
server) and a whole lot of locking and unlocking.  It could occasionally
delay the driver or server when waiting for a lock.

In April 2003, the entire state management subsystem was removed and
replaced with a single local socket.  The drivers listen for
connections and push updates asynchronously to any listeners.  They also
recognize a few commands.  Drivers also dampen updates, and only push
them out when something actually changes.

As a result, `upsd` no longer has to poll any files on the disk, and can
just `select()` all of its file descriptors (fds) and wait for activity.
When one of them is active, it reads the fd and parses the results.
Updates from the hardware now get to `upsd` about as fast as they possibly
can.

Drivers used to call `setinfo()` to change the local array, and then would
call `writeinfo()` to push the array onto the disk, or into the
mmap/shared memory space.  This introduced a lag since many drivers poll
quite a few variables during an update.