Another low-level Arcan client: A tray icon handler

This is the third part in the ongoing series exploring the different levels of APIs that are available for Arcan clients. For reference, the previous parts in the series are here:


In this part, we will write a really nifty little tool that will act as a tray icon (and more). When clicked or triggered via a custom input binding or a global hotkey, the tool will launch a second application that attach as a “popup” to the tray icon, with the icon representing whether the application is alive or not. This video shows it running inside of the Durden desktop, first wrapping a terminal emulator, then an X server running windowmaker (tradition by now).

The API features it covers are as follows:

  1. Registering custom key bindings.
  2. Negotiating a connection on behalf of a third party.
  3. Spawning a new process that inherits a negotiated connection.
  4. Server-side triggered coarse-grained timers.

The tool itself turned out useful enough to be added inside the tool collection in the main Arcan repository, so you can find it here:

A neat thing is that this clean and small tool can actually cover a number of use-cases without any modifications to the ‘real’ client, staying entirely in line with the ‘do one thing and well’ philosophy.

From an outline perspective, this is what our tool will do:

  1. Connect and identify as an ‘ICON’.
  2. Render a SVG into the segment from 1.
  3. On activation (click or otherwise), request a ‘HANDOVER’ subsegment.
  4. Inherit this handover connection into a new process and executable.
  5. Keep track of the child and update the icon state accordingly.

Getting the icon (1, 2)

The structure of the code itself will be similar to that of the previous article, so we will use that as a reference and describe the relevant modifications.The first change for this is easy, switch out the MEDIA part of the connection call to an ICON:

struct arg_arr* args;
struct arcan_shmif_cont conn =
arcan_shmif_open(SEGID_ICON, SHMIF_ACQUIRE_FATALFAIL, &args);

As a recap, the primary segment type is a strong hint to the window manager about what kind of a connection this is. This type has some impact on how the connection will be synchronised and prioritised, what decorations will be assigned, which kinds of subsegments will be allowed to be allocated and so on. The rule of thumb is that type + connection-point is the main selector for behaviour.

Capture the initial data like before, this time around we care about the initial size, upper dimensions and output density as we need to follow them:

struct arcan_shmif_initial* cfg;
arcan_shmif_initial(&conn, &cfg);

struct trayicon_state state = {
.density = cfg->density * 0.393700787f,
.bin = argv[3],
.argv = &argv[4],
.client_pid = -1,
.envv = environ,
.icon_normal = argv[1],
.icon_pressed = argv[2]

Only thing of note is that the SVG renderer we will use is stuck in the dark ages and uses the archaic ‘DPI’ for density, so convert to metric. Rendering SVG is out of scope, so we pull that from a 3rd party: nanosvg. The actual drawing and signalling is not different than what was done in the last article, so little point in reiterating that here. This link takes you to the lines in question.

Using the naive event loop from last time would cause some problems , so we will modify it slightly:

static void event_loop(
struct arcan_shmif_cont* C, struct trayicon_state* trayicon)

struct arcan_event ev;
while(arcan_shmif_wait(C, &ev)){
do_event(C, &ev, trayicon);

while(arcan_shmif_poll(C, &ev) > 0)
do_event(C, &ev, trayicon);

if (trayicon->dirty){
arcan_shmif_signal(C, SHMIF_SIGVID);

The innermost while loop is new. For this specific application, it is ok that the first wait is blocking as the program is entirely event driven. Responding to some of the events can be expensive however, particularly those events that would redraw the icon state. To be more responsive to ‘event storms’ where a stall somewhere else in the system could cause multiple events that would counteract each-other, we flush them out and only update after that is done.

Intermission: Durden tool, “external buttons”.

Features like a tray requires cooperation with the window manager. The easiest way to achieve this is through an addressing scheme we call ‘connection points’. Review the previous articles in the series if you do not know about this concept already.

For Durden, there is an explicit script that maintains the connections for external statusbar button, reroutes input and deals with positioning, user configuration and so on. This link points to the source of that tool.

In short, it exposes ways of creating connection points that wait for an icon to connect, and maps actions and data on this connection into the statusbar. It also adds strong restrictions on what this icon connection can do, what other resources it can try and negotiate, and how input gets routed to it.

As with anything else in Durden, the settings and controls for enabling this feature is exposed as a filesystem (one that can be mounted over FUSE). The specific path that needs to be enabled for this demo is:


Then the settings are exposed and discoverable in following path:


This allows the destination and behaviour to be highly tuneable and composes well with all other desktop mechanisms, from input macros to timers.

On Activation, Request a Subsegment (3)

Back to our tool. Lets modify the event loop to respond to click events, and to register a custom input binding slot. This labelhint can also carry a suggestion as to the default keysymbol (F1, l, UP, …) + modifier (ctrl, alt, …) though it is, as normal, something for the WM to decide what to do with. As soon as we have the connection from arcan_shmif_open:

arcan_shmif_enqueue(&conn, &(struct arcan_event){
.ext.labelhint = {
.label = "ACTIVATE",
.descr = "tooltip description goes here"

Next in the event loop we will check for any digital event matching our label, or a digital press from a mouse type device.

if (ev->category == EVENT_IO){
if (ev->io.kind != EVENT_IO_BUTTON ||

if (strcmp(ev->io.label, "ACTIVATE") == 0 ||
ev->io.devkind == EVENT_IDEVKIND_MOUSE){
toggle_client(C, trayicon);

In toggle_client, we do the normal ‘check if the child is alive, if so, kill it and spawn a TERM to KILL thread’. It is just verbose and not very interesting. See this link for that code. The more interesting is if we want to spawn a child. First, we do this:

arcan_shmif_enqueue(C, &(struct arcan_event){
.ext.kind = ARCAN_EVENT(SEGREQ),
.ext.segreq.kind = SEGID_HANDOVER

This says ‘hey, I would like to create a new segment on behalf of someone else’. This allows the server side to accept the new client in, but also track origin without resorting to fundamentally broken things like asserting that pid == connection which was the X11 dance with _NET_WM_PID. This is an asynchronous request, so we need to do something in the event loop as well.

Handover to new Process (4)

We are reaching the end of the journey. Time to extend the event loop with something. There are two possible events that can come following a segment request. The boring one is TARGET_COMMAND_REQFAILED, which means that the server side said no. This is also the default if the WM does not explicitly opt-in via a call to target_accept as part of the event handling routine on the server side.

If the request is accepted, the proper resources are allocated and pushed to the client. To help dealing with multiple pending requests and so on, the type can be verified and paired against a client chosen ID, though we do not really care about that here for the time being.

trayicon->client_pid =
arcan_shmif_handover_exec(C, *ev,
trayicon->bin, trayicon->argv, trayicon->envv, false);

if (-1 != trayicon->client_pid){
render_to(C, trayicon);

It is worthwhile to note that NEWSEGMENT events can arrive at any time. If they do not match a previous request, it means that the WM force pushed a resource as a feature probe and signalling intent. This will become important in an upcoming article.

If we do not handle this event, the internals of the client side library will drop the resources on the next call into any of the arcan shmif functions that is not directly related to accepting the request. Here we use a specific derivative of the exec- family, arcan_shmif_handover_exec. Internally speaking, it sets up a new process but duplicates the control socket and mapping descriptor, passes it as an environment so the new client will find it on arcan_shmif_open instead of using the connection path mechanism.

Track the Child (5)

The last piece of the puzzle is to keep track of our new pid. Now this pose something of a problem with the event loop we have. The proper options tend to be to fire up a monitoring thread that waitpid()s or a signal handler, then forward that back to the event loop where we, instead of the normal wait / poll I/O multiplexes on the raw descriptors. That is a lot of ugly boring code. For this round, we take something simpler.

arcan_shmif_enqueue(&cont, &(struct arcan_event){
.ext.clock.rate = 5

This will periodically emit STEPFRAME events at a rate of one event every 5 ticks on a 25Hz clock (chosen for my PALs). This is a mechanism that can be used for getting frame delivery notifications, server- initiated variable frame pacing mechanism and for client requested low frequency timers. Here we will just use it to check if the child is still alive and flip the icon if the state changed.

In Closing

If you are curious as to how the client side of the tray icon work in Xorg, look no further than the “System Tray Protocol Specification“. A wrapper similar to this one would of course be possible to make within good old’ X, but in a quite more painful and convoluted way ,with much less security, safety and accountability towards the user. Recall, even the smallest of ‘icon’ in X is no less capable than a full blown application, and this lack of nuance is is kind of the problem.

The main point of this article was mainly to show off some of the basic features, primarily the HANDOVER execution, as it is something we will use later for much more advanced things. The secondary point was to illustrate the balance of power and control, with clients still able to achieve traditional edge case features, but not without the consent and constraints set by the user, and that can only be a good thing.

This entry was posted in Uncategorized. Bookmark the permalink.

Leave a Reply

Fill in your details below or click an icon to log in: Logo

You are commenting using your account. Log Out /  Change )

Google photo

You are commenting using your Google account. Log Out /  Change )

Twitter picture

You are commenting using your Twitter account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )

Connecting to %s