\input texinfo @c -*-texinfo-*-
@c %**start of header
@setfilename mds.info
@settitle mds
@afourpaper
@documentencoding UTF-8
@documentlanguage en
@finalout
@c %**end of header
@dircategory Graphics environment
@direntry
* mds: (mds). The micro-display server
@end direntry
@copying
Copyright @copyright{} 2014 Mattias Andrée
@quotation
Permission is granted to copy, distribute and/or modify this document
under the terms of the GNU Free Documentation License, Version 1.3 or
any later version published by the Free Software Foundation; with no
Invariant Sections, with no Front-Cover Texts, and with no Back-Cover
Texts. A copy of the license is included in the section entitled
``GNU Free Documentation License''.
@end quotation
@end copying
@ifnottex
@node Top
@top mds -- The micro-display server
@insertcopying
@end ifnottex
@titlepage
@title mds
@subtitle The micro-display server
@author by Mattias Andrée (maandree)
@page
@c @center `'
@vskip 0pt plus 1filll
@insertcopying
@end titlepage
@contents
@menu
* Overview:: Brief overview of @command{mds}.
* Architecture:: Architectural overview of @command{mds}.
* Protocol:: The @command{mds} procotol.
* GNU Free Documentation License:: Copying and sharing this manual.
@end menu
@node Overview
@chapter Overview
@command{mds}@footnote{mds stands for micro-display server}
is a display server protocol and an implementation of said
protocol. What makes @command{mds} stand out is its core
design choice: it is desigend just like a microkernel.
Rather than one, possibly modular, process --- a monolithic
process --- mds is comprised of many small servers, each
exchangable and responsible for one thing.
@command{mds} goal is neither security, performance nor
a perfect graphical experience. @command{mds} is all
about flexibility and freedom 0@footnote{The freedom to run
the program as you wish, for any purpose}.
The reason for having a display server architectured as a
microkernel is so that components can be added, remove
and replaced online. Additionally, the message passing
between the servers makes it easy to design a system that
lets you make clients that can listen on messages between
the servers and perhaps modify them. This enables you to
do so much more with your display server. Moreover, if
a single part of the system crashes it does not bring down
the whole system, and the crashed server can be respawned
with minor side effects. @command{mds} is architectured
in three layers: a microkernel, a master server and a
collection of servers. And clients are actually located
on the same layer as the servers, because there is no
actual difference, the only thing that separates a server
from a client is for what purpose you run it. @command{mds}'s
kernel is a minimal program that do initialisation of the
display, such as giving it an index and create runtime
files and directories for servers and other programs
to use. Then the kernel creates a domain socket for the
master server and spawns the master server and respawns
it if it crashes. Because of this, if the master server
crashes it will not lose its socket when it is respawned.
The master server than, on its initial spawn, starts
the all servers and other programs that the user have
choosen and then starts accepting connections to it and
coordinates messages between servers and clients. Further,
separating all components into separate processes enables
us to only give the servers the privileges they actually
need, rather than having one program with root privileges
that takes care of everything even things that do not do
require any privileges.
All @command{mds}'s servers, that is all running parts of
@command{mds} except the kernel, are designed so that they
can re-exec themself so that they can be updated online
without any side effects. Servers serialises their state,
saves it to RAM (in a directory created by the kernel),
re-exec themself and loads their serialised state. The
kernel cannot do this because when it has spawned the
master server it has no reason to re-exec, its only mission
is to respawn the master server it if would happen to crash.
It would technically be possible to enable the kernel to
re-exec but it is not worth it as it as no reason to
re-exec, and doing so puts the display server at risk
of crashing.
@node Architecture
@chapter Architecture
@menu
* Layers:: The layers of the display server.
* Interprocess Communication:: How servers and clients communicate.
@end menu
@node Layers
@section Layers
The @command{mds} display server in architectured in
three layers. The first layer is called the kernel.
The kernel is responsible for acquiring a display
server index@footnote{As with any display server,
the system can have multiple instances of
@command{mds} running at the same time.}, set up
environment variables to indicate which display
server and display server instance is being used,
create a domain socket for the display server and
start the master server and restart it if it crashes,
and then clean up the system when the display server
closes. The kernel only responsible for creating
the domain socket for communication with the display
server, it is not responsible for using it, that
mission falls to the master server.
The second layer is the master server. The master
server has two responsibilities: coordinating
message passing between other servers and clients
@footnote{In @command{mds} their is no functional
distinction between servers and clients, the
distinction is purely semantic.} and starting
other servers.
The third layer is the other servers and clients.
protocolwise there is no specification on how
they are started. But in the reference
implementation of the master server, this is
done by starting a shell script with the
pathname @file{$@{XDG_CONFIG_HOME@}/mdsinitrc}
and the user is responsible for providing the
logic in that shell script.@footnote{Moonstruck
users are allowed to implement this in C
or any other language of their choosing.}
@c Which is better: cray-cray users, lunatic users,
@c moonstruck users, insane users, ballers, madmen,
@c loony tunes?
These servers implements the actual functionality
of the display server.
@node Interprocess Communication
@section Interprocess Communication
Intrinsic to @command{mds} is a powerful
interprocess communication mechanism. Servers
and clients connect to the display server by
connecting to a domain socket served by the
master server. A server or client that has
connected to the display server can do three
things:
@itemize
@item
Request assignment of a unique ID.
@item
Multicast a message.
@item
Join or leave a multicast groups.
@end itemize
Upon assignment of an ID the master server
will automatically place the client in a
multicast group for that specific client.
This automatically multicast group assignment
is done by the master server simply so you
as a debugger do not forget to do so. When
a client is disconnected it will and out a
message to a specific multicast group that
the client, refered to by it's ID, have closed.
A message in the @command{mds} protocol is
comprised of two parts: headers and a payload.
When a client joins a multicast group it is
actually say that it is interested and receiving
broadcasts containing a specific header or a
specific header--value pair, or that it is
interesting in all messages@footnote{This
could be used for logging, possibly spying and
networking.}. Thus a message is automatically
multicasted to groups indicated by its headers.
The multicast groups and receiving of groups
is called interceptions. The interesting
property of interceptions is that they may
be modifying. When a server registers for
message interception it can say that it wants
to be able to modify messages. If this is done
and the server receives a message for which it
has said it want to be able to modify it,
the master server will wait for that server
to respond before it send the message to
the next server in the interception list.
The server can choose to do three things
with a message that it has opted in for
modification of: leave the message as-is,
modify the message, or consume the message.
A message consumption is done by modify
the message to make it empty. A consumed
message will not be send to any further
clients or servers in the interception list.
To make this mechanism sensible, a server or
client can set a priority when it registers
for interception (does not need to be
modifying.) When a message is broadcasted it
will be received by all servers in the
interception except the original sender,
unless it gets consumes. The order in which
the master server sends the message to the
recipients is determined by priority the
servers registed with. The message first sent
to the recipients with highest priority and
last to the recipients with lowestr priority,
and orderd by the priority between those
priorities. Of two or more servers have the
same priority the order in which they will
receive the message, of those recipients,
is arbitrary.
An interesting property of this machanism
is demonstrated in the @command{mds-vt}
server. Unlike most servers @command{mds-vt}
maintains two concurrent connections to
the display. Once @command{mds-vt} receives
a signal from the OS kernel requesting to
switch virtual terminal, @command{mds-vt}
will from one of its connections send
out a message and wait for it to be
received in its other connection and the
let the OS kernel switch virtual terminal.
The secondary connection to the display
has registered interception with lower
priority of the message that the primary
connection broadcasts. This message will
be received by other servers that will
let the message continue to the next
server in the interception list once that
server is ready for the OS kernel to switch
virtual terminal. All of these server has
registered modifying interception of the
message but none will actually modify or
consume the message; it is only used a
mechanism for letting @command{mds-vt} know
when all servers are ready for the switch
without having to know how many they are
and wait for a reply from all of them.
@node Protocol
@chapter Protocol
@menu
* Environment Variables:: Identifying the active display server
* Signals:: Signalling individual servers
* Filesystem:: The display server's footprint on the filesystem
* Message Passing:: Sending messages between servers and clients
@end menu
@node Environment Variables
@section Environment Variables
A crucial of any display server is letting child
processes know which display server they should
connect to. @command{X.org} does by setting the
environment variable @env{DISPLAY} to
@code{<host>:<display index>}, where @code{<host>}
is empty if the display is one the local machine.
In this tradition @command{mds} does the same thing
with the environment variable @env{MDS_DISPLAY}.
@command{mds} also creates a new process group and
export the new process group ID to the environment
variable @command{MDS_PGROUP}. This process group
can be used to send signals to all @command{mds}
servers collectively.
@node Signals
@section Signals
@command{mds} servers can re-execute into an
updated version of their binary. This can be
used to update display server online after
a new version has been installed. To do this
send the signal @command{SIGUSR1} to the server
you want update. If a server does not support
online updating it will ignore this signal.
If the operating system defines a signal named
@command{SIGUPDATE}, this signal is used
instead of @command{SIGUSR1}.
If you need servers to free up allocated
memory that they do not use, send the signal
@command{SIGDANGER}, or if not defined
@command{SIGRTMAX}. Unimportant servers may
choose to die on @command{SIGDANGER}.
@node Filesystem
@section Filesystem
The @command{mds} kernel creates two directories
for the @command{mds} servers to use: one for
runtime data and one for temporary data.
These directories are named by
@code{MDS_RUNTIME_ROOT_DIRECTORY} and
@code{MDS_STORAGE_ROOT_DIRECTORY}, respectively,
by the header file @file{<libmdsserver/config.h>}.
If the systems runtime data directory is @file{/run}
and transient temporary data directory is @file{/tmp},
and the package name of @command{mds} is @command{mds},
these directories will be @file{/run/mds} and
@file{/tmp/.@{system-directory@}.mds}, respectively.
In @file{/tmp/.@{system-directory@}.mds} the kernel
will create a directory for the display server instance
named @file{.data} prefixed by the display server index.
For example if the display server index is zero,
temporary data may be stored in
@file{/tmp/.@{system-directory@}.mds/0.data}
As defined by @code{SHM_PATH_PATTERN} by
@file{<libmdsserver/config.h>}, when a server
re-executes itself it will marshal its state to
the POSIX shared memory unit named
@file{/.proc-pid-%ji}, where @file{%ji}
@footnote{@code{%ji} is the pattern in @code{*printf}
functions for the data type @code{intmax_t}.} is
replaced with the process ID of the server. This
file will be bound to the pathname
@file{/dev/shm/.proc-pid-%ji} if POSIX shared
memory is stored in @file{/dev/shm} by the
operating system.
In @code{MDS_RUNTIME_ROOT_DIRECTORY} the kernel
will create two files. @file{.pid} and @file{.socket},
both prefixed with the display server index
@footnote{@file{0.pid} and @file{0.socket} if
the display server index is 0.}. The @file{.pid}
file contains the process ID of the display server
and is used by the kernel to figure out whether
an display server index is still in use or just
not properly cleaned up. Of course it can be used
by any program to find the process ID of the
kernel process of a display server instance.
The @file{.socket} is the domain socket used
for communication with the display server and
its servers and clients.
@node Message Passing
@section Message Passing
@node GNU Free Documentation License
@appendix GNU Free Documentation License
@include fdl.texinfo
@bye
