// Patrick Louis

Daemons on Unix

Daemons on Unix

(Transcript of the podcast)

Intro

You’ve certainly heard of daemons, those processes that lurk in the background and do what they’re supposed to do.

You might even have written and run programs that are daemons.

Today we’ll talk about them, those daemons ({day,dee}mon), what there is to know about their mechanism and details.

A big generic overview of daemons on Unix.

What does Daemon mean, where does it come from

Let’s start with the history, etymology, and terminology.

The where does it come from, what does it mean, why it’s written this way.

By now, if you’ve been using Unix for quite a while you should already know that daemons are not truly demons. You may even have made the distinction in your mind separating the two concepts.

Some even use different pronunciations not to mix up the terms, {day,dee}mons

However, this distinction doesn’t really exist when it comes to the terms and history, daemon truly means demon, it’s just an older alternative spelling of the word demon.

Daemon are spirits/supernatural beings of the Greek mythology that go around doing their deeds in the shadows, overlooking the world or individuals, sometimes affecting them in good ways, sometimes in bad ways. They aren’t biased, what they do is more or less objective in that regard.

That’s rather different than today’s perception of demons as evil mean satanic spirits.

Those daemons are always there, in the background, symbolizing the not visible, yet always present and working its will.

But where did the metaphor come from?

To put it bluntly, it was coined by the programmers of the MIT’s project MAC, which ran the compatible time-sharing systems, which was one of the first time-sharing OS. It was the ancestor that inspired Multics, Unix, and ITS.

The term was first used to refer to what was called a dragon, and then more generically used to refer to anything that acted like that dragon.

A dragon being similar to what we call today daemon but in the difference that it is not invoked at all but instead is used by the system to perform secondary tasks.

I couldn’t find much info about dragon, so if anyone got more on that please add up resources when this gets released.

Daemon was then used to refer to the more generic concept.

The name daemon itself was inspired by Maxwell’s demon, which is a physics philosophical experiment where an imaginary being interferes constantly in the background with the particles to make them behave in certain ways.

This describes well today’s use on Unix, a background process that does system chores.

The first prototype of the daemon concept was a program called DAEMON that automatically made tape backups of the file system.

After that some started rationalizing the name choice and built the backcronym Disk And Execution MONitor, a constructed phrase claimed to be the source but that really isn’t.

Today the concept of daemon is present in many areas under other names such as services, the concept isn’t necessarily unique to Unix but it’s most often related to it when you hear that name.

Still it’s hard not to equate demon with the satanic underworld definition when even the BSD mascot, beastie, makes fun of it.

process tree

Generic overview

So that’s where the name comes from but what’s a generic overview of daemons, what do they really do, what’s their job, what’s their purpose, what’s special about them.

After much reading, I’ve compacted the definition of a daemon into the following:

A daemon is a process in a multitasking OS that offers services/functionalities or does autonomous tasks, sometimes repetitive/periodic, sometimes in response to the occurrence of specific system events or conditions aka supervising the system, most of the time it’s a long running process but not always, most of the time without interventions and independent of a user, and most of the time in an unnoticed unobtrusive way, aka in the background.

So mix those and it gives you the recipe for a daemon.

But what about the terms services and servers.

They are subset of what we call daemons.

Daemons are just a bit broader, they don’t have to be services.

A daemon can be a long-running background process that answers requests for services.

So, a service, like the name implies, is a software that provides capabilities to other softwares.

Services are typically automatically started at boot up, are long-lived, they have states that can be altered (running/not running), they have relationships and dependencies with other services, and they are usually critical components of a system and so are started during boot.

Those services are managed by a daemon, the init process, the first process started on a system, the PID 1. It’s job is to manage what runs at what time, to handle the dependencies, and regenerates them after crashing so that the system keeps running correctly. More on that topic later on.

Some examples of such services can be the logging system, a database, or a network manager.

Daemon can also act as middle man, as message passing systems to the kernel api/system calls which are services providing access to some lower level layer.

You can see how services fit as subset of the daemon definition.

A server is also a subset of the daemon definition.

Simply said, servers are processes that are long-running and perform services based on requests that can be launched remotely from another machine.

The obvious example here are a webserver, a mail server, or a secure shell server, sshd.

Other examples of daemons include:

  • crond - time-based job scheduler
  • BIND - the Berkeley Internet Name Daemon
  • dhcpd - dynamically configure tcp/ip information for clients
  • syslogd - system logger collector
  • ntpd - network time protocol daemon
  • inetd - listend for network connection requests and launches daemons to handle them

As you might have noticed a lot of those daemons end in ‘d’.

It’s a tradition and convention, taken for the sake of clarity, that daemon names, when checked in the process tree, all have a trailing ‘d’.

This is just so that when you give processes a glance you can denote which is a daemon and which is not.

Foreground/Background process

We said that daemons run as a background processes rather than being under the direct control of an interactive user, what does this imply?

What does it mean to run in the background?

Let’s do a bit of explanations.

There are, arguably, three cases where a process is considered a background daemon.

The first one is when the process is started during boot time, usually by the init process, and thus is not attached to any terminal.

The second one is when the process is invoked explicitly, on a terminal, and then detached from all terminals and ran in the background.

The third case is more or less commonly said to be about “transforming a process into a daemon”, even though it’s not really. It consists of simply having the process in the process group attached to the terminal but running in the background.

We’ll dive into the details of every one of these categories later but first let’s talk about processes in general.

A process is an instance of a running program, it is managed by the kernel of the OS and is referred to by it’s process ID.

When looking from an interaction point of view there are only three types of processes: interactive - from CLI interaction, batch - from a pipeline, daemon which are the rest.

Another way to perceive daemons is to see them as those repetitive administrative tasks that are fired by a scheduler from time to time. But maybe in that case it’s the scheduler, such as cron that is the daemon and not the tasks themselves.

Yet again, sometimes even the automated background tasks that are fired upon login can be said to be daemons.

This depends on your perception.

Now from a terminal process management point of view processes are either in the foreground or in the background.

The foreground processes are the ones we interact with on a terminal, they get the stdin/stdout/sterr, they are the session leader of a terminal, and there can only be one session leader, one controlling process of a terminal at a time.

When one process is the session leader all the other processes in the group are in the background and run by themselves instead of running from user inputs.

The signals sent to the terminal are always sent to the process group which has it as a controlling terminal. The session leader catches it.

With today’s job control, that is the BSD style of job control, terminals send a STOP signal to background processes that try to write to a terminal, the SIGTTIN and SIGTTOU signals.

SIGTTIN   21,21,26    Stop    Terminal input for background process
SIGTTOU   22,22,27    Stop    Terminal output for background process

This makes it a pain if you want to have long running processes in the background that might inadvertently write to a terminal and be stopped by mistake.

For that reason, the background processes that want to “act” like daemons have to become immune/ignore those signals. Or you could possibly disable the feature in the terminal itself to not send that signal when a background process tries to write to its terminal.

But that’s not the start of the issue with having a “daemon” attached to a terminal like that. The biggest issue is that if the terminal process exits the wannabe daemon will be killed because it’s still associated with it.

In fact it’s killed because it will receive the SIGHUP, hangup signal from the terminal, which the default action is to terminate the process.

So one way to alleviate the pain, again, is to ignore that signal.

The nohup command does just that, it’s a wrapper to ignore SIGHUP.

If stdout and stderr are connected to a terminal, it also redirects them to nohup.out

But then the question is: What happens when the terminal process ends, when the parent of the wannabe daemon dies?

Well, that’s simple, when the parent of a process terminates it is assigned/adopted by the init process, PID 1, as parent by default.

Now in this case it will have no controlling terminal and will run in the background as intended.

That is one reason the daemons are started during the boot process, but not the only one, as we’ll see later.

So we’re starting to understand the idea behind running daemons.

It doesn’t really lie in the association or disassociation from a terminal but more in the other criteria we talked about in the earlier section.

However, as we’ve seen, it’s not as simple to run processes with the intention of having them as daemons.

It’s not as simple when they’re not session leaders, nor when they don’t have a controlling terminal.

How do you even communicate with daemons?

This poses many problems, how does this all fall into place?

This makes it hard to write a daemon and to manage them without missing edge cases and that is why some standards and good practices got built around daemons over the years.

And that is what we’re going to explore next.

Programming Wise

Because of everything we’ve discussed in the previous section it is tricky to write a daemon. Fortunately, all you have to do is follow a set of rules and norms to get your daemon running correctly.

You could possibly create a daemon using the technique we talked about earlier, namely forking a child process in the background and immediately exiting so that it is adopted by the init process.

But that’s not good enough you have to do other things so that you’ll be sure it won’t break.

You can find a list of the series of steps all over the internet.

Some “tips” can even be found in your man pages, checkout the man 7 daemon page.

Most of those recommendations are similar in some ways, and emerge from the requirements/demands of the service manager of that system, be it the init system or not, be it Apple MacOS X Daemon, SysV style init systems, “new-style” systemd, inetd, launchd, etc..

So again, daemons are not really hard to write, on the contrary, they’re easy to put together, what’s tricky is to think of the edge cases, quirks, and side effects.

Things that are usually taken care of when processes are running in an

interactive session on a terminal but not when detached.

For example, what we’ve talked about earlier, namely the process group, the user ID running the process, the input/output, etc.

If the daemon is started at boot it has to manage those itself explicitly and if it’s launched from within a login session then it needs to undo a lot of what has been set during the login sequence to ensure that the association with the calling environment is destroyed.

In the best of the worlds the service manager or sometimes called super-server daemon, the process that manages the daemons, which usually launches them at boot and is also usually the init process, would handle those steps and perform the functions you need.

That’s just a nota-bene to say that the actual procedure doesn’t have to be coded inside the daemon itself but can be external. Let’s also note that for old style SysV daemons those steps are required during the initialization. Other init systems/service-manager might handle that for you, so be sure to checkout their documentation.

So now, let’s go through the usual criteria that are recommended to turn a process into a daemon, the order might vary according to different specs but the ideas are the same.

Criteria

  • Create background process

The first step that needs to be taken care of is the one we talked about before, forking off the parent process of the daemon and let it terminate gracefully so that the soon-to-be daemon process run autonomously in the background.

For those who don’t know, forking is a system call that creates a child process that is a near copy of its parent. It’s the only system call that returns two distinct values, the PID of the child for the parent that keeps running, and 0 for the child, in the case of errors it returns -1.

In sum this step is about separating the daemon from its parent.

  • Becoming a session leader & group leader.

On Unix every process has multiple identifiers attached to it, and each of those identifiers mean something different, usually a way to regroup related processes.

You certainly know the PID, the process ID, a unique number associated to the process so that you can refer to it.

There’s also the parent ID, the ID of the process that started it. There’s the group id which is the PID of the process group leader. There’s the session ID, which is again, just the ID of the session leader.

Processes in the same session have the same controlling terminal, processes in the same group also belong to the same session. But a session can have multiple process groups.

To picture that, the shell is usually the session leader, it controls the terminal, and every pipeline executed is a process group, they run grouped with each other.

Ok so we have a background process without a parent, or actually an orphan process inherited by the init process, but that process is still attached to the terminal itself and the process group associated to it because it has inherited it from the dead parent.

In this case the daemon is still subject to receiving terminal generated signals and signals sent to the process group, remember for example SIGHUP which we talked about earlier. So it would be pretty bad. We also don’t want to ignore those signals, as we said.

It might also be a good idea to reset all the signal handlers to their default behavior because you might not know how the parent has changed them.

We need to detach/disassociate ourselves, to create our own process group and become our own process leader, that means disassociating the daemon from any controlling terminals and having its own session.

Under some systems the previous step, the fork step, is not mandatory and the detachment can be done directly, however on other systems it’s mandatory because a process that is already a group leader cannot change its group and forking is an easy way to solve that. Forking is also a way to prevent the terminal from locking up making the shell believe the daemon has terminated.

There are different system calls to do what we talked about, ioctl on some older system or setsid on newer that is for the associated session ID, and for the group we have setpgrp amongst others, which is not used along with setsid.

So at this point if you looked at the process tree you should see that your daemon has no controlling terminal, that it’s parent process ID is 1, that its session ID is the same as its group id and the same as its own ID. It’s its own session leader and group leader.

  • Fork again

If you know a thing or two about process management, which we might discuss soon in another episode, you may know that being a session leader is the only way to potentially reacquire a terminal.

We don’t want that, we don’t want the daemons to acquire any terminal, not because its bad but because we have a limited set of terminals and we’re not gonna spare one for every daemon we have. Also, even in systems where we don’t have that problem, where process gets attached to already acquired terminals, we don’t want this because we don’t even need a terminal for the daemons, it’ll only create complications.

Under some systems, like 4.2BSD, this is already taken care of because only processes with the group ID of zero can acquire a controlling terminal, and our daemon has a group ID equal to its process ID, as you may remember.

But we have to take care of all systems and all cases.

The easiest way to prevent the reacquisition of a terminal is to fork again and run the actual daemon in that child while, again, killing the parent. And because the parent is the process group leader it has to ignore SIGHUP because otherwise it’ll forward SIGHUP to the child and kill it, since the parent is the session leader. This signal handling will also be inherited by the child.

This has the final effect of having a daemon that has no controlling terminal, that is immune to signals from the tty driver, and cannot acquire a new terminal since it’s not a process group leader, its PID is different than its PGID and SID.

  • Resetting the environment

Alright, so those steps already should give us a nice daemon but there are other small things to take into consideration too.

For example, it’s a good idea to set the current working directory of the daemon to the root directory instead of the one inherited from the parent. Like that the process doesn’t keep hold of any directory or resources under any mounted filesystem, allowing them to be unmounted at any time, and the root directory is always guaranteed to be there.

A daemon also doesn’t have any use for file descriptors and so it’s also a good idea to close the ones that have been inherited by the parent or to redirect them to /dev/null, be them file descriptors to files or the standard ones stdin/stdout/stderr, anyway the daemons don’t have a terminal. It’ll prevent security issues too.

It’s also good to do that because sometimes the files related to the terminal settings are still opened, and when they are the terminal settings are not reset until the last process releases them.

Listen to the episode about terminals to know more about terminal settings.

Another attribute a daemon inherit from the parent is the umask, the file creation mask, it’s also a good idea to set it to 0 to allow the permission masks to depend only on the daemon and not what the parent wanted it to be.

You can listen to the podcast about bits and bytes, I’ve discussed about masks in it.

You may also need to reset environment variables to sane settings that don’t interfere with the daemon instead of using the ones from the parent.

  • Main loop

Nice, and sweet, we should be settled now, we got our daemon running in its own environment.

Now we can add whatever extra we want to add, be them related to the way we want to communicate with the daemon like logs and whatever else.

Usually the main part, the payload, the core, of a daemon is a big loop, an infinite, while 1, loop that repeats the same tasks over and over again in some cases checking for conditions. It would’ve been better to call them “the myth of Sisyphus” instead of daemons.

So let’s summarize the steps you need to take care of before the actual daemon code:

  1. Create a background process, the parent of the daemon is PID 1.
  2. Become a session leader & group leader, GID == SID == PID.
  3. Fork again to detach and force it not to reacquire a terminal. (the parent should ignore SIGHUP) PID != GID and PID != SID
  4. Reset the environment: * Close all open file descriptors. * Connect /dev/null to standard input, output, and error. * Reset all signal handlers to their default. * Reset the signal mask. * Sanitize the environment variables. * Reset the umask to 0. * Change the current directory to the root directory.
  5. Actual daemon loop

Sample


/*
 * daemonize.c
 * This example daemonizes a process, writes a few log messages,
 * sleeps 20 seconds and terminates afterwards.
 */

#include <stdio.h>
#include <stdlib.h>
#include <unistd.h>
#include <signal.h>
#include <sys/types.h>
#include <sys/stat.h>
#include <syslog.h>

static void skeleton_daemon()
{
    pid_t pid;

    /* Fork off the parent process */
    pid = fork();

    /* An error occurred */
    if (pid < 0)
        exit(EXIT_FAILURE);

    /* Success: Let the parent terminate */
    if (pid > 0)
        exit(EXIT_SUCCESS);

    /* On success: The child process becomes session leader */
    if (setsid() < 0)
        exit(EXIT_FAILURE);

    /* Catch, ignore and handle signals */
    //TODO: Implement a working signal handler */
    signal(SIGCHLD, SIG_IGN);
    signal(SIGHUP, SIG_IGN);

    /* Fork off for the second time*/
    pid = fork();

    /* An error occurred */
    if (pid < 0)
        exit(EXIT_FAILURE);

    /* Success: Let the parent terminate */
    if (pid > 0)
        exit(EXIT_SUCCESS);

    /* Set new file permissions */
    umask(0);

    /* Change the working directory to the root directory */
    /* or another appropriated directory */
    chdir("/");

    /* Close all open file descriptors */
    int x;
    for (x = sysconf(_SC_OPEN_MAX); x>=0; x--)
    {
        close (x);
    }

    /* Open the log file */
    openlog ("firstdaemon", LOG_PID, LOG_DAEMON);
}
int main()
{
    skeleton_daemon();

    while (1)
    {
        //TODO: Insert daemon code here.
        syslog (LOG_NOTICE, "First daemon started.");
        sleep (20);
        break;
    }

    syslog (LOG_NOTICE, "First daemon terminated.");
    closelog();

    return EXIT_SUCCESS;
}

When it is someone else that does the job

As we said earlier, all those steps setting up the environment for a daemon can be taken care of by a third party, for example the init process.

Still in those cases there are certain requirements that need to be met by the daemon, not only to be able to launch them but to supervise and control them also.

Executing the steps we mentioned before may even interfere with how well all this functionality happens.

Those new requirements are quite straight forward, they are usually the following but not limited to them.

Two signals should be handled SIGTERM to exit cleanly the daemon and SIGHUP to reload the configuration files if it applies.

The exit code from the main daemon process should reflect if there are errors and problems while executing the daemon.

The daemon should be integrate correctly within the system, be it an init file of some sort or anything else. It also should rely on the system’s functionality as much as possible instead of re-inventing them.

If applicable the daemon should notify that system upon startup completion or status update via an interface provided by that system.

And sometimes using the communication mechanism offered by that system, be it for inter-process communication or for logging.

All and all, you’ll have to refer to the system providing the daemon handling.

daemon (3)

So those were a lot of steps we talked about earlier, isn’t that a repetitive task, why not wrap it up?

There’s a function called daemon() that first appeared in 4.4BSD and that does the core of what we talked about earlier.

Some disavow using it as it doesn’t implement every single one of the steps, for example the ones about cleaning up the environment.

However, that a nifty convenience routine that let you do in one line the whole detachment, running in the background, closing files, and changing directory parts. This is all really good if you want to throw in 2 min a daemon and don’t want to bother re-implementing everything while not relying on a third party system.

Management

There’s the daemon, and it’s running.

Now how do you keep it running, how do you manage, supervise, and track it, and how do you interact with it.

There are programs that we call service manager or super-server daemon, the ones we mentioned before, for instance launchd, systemd, Upstart, Service Management Facility, Android init, sysvinit, runit, OpenRC, supervisord, daemontools, etc..

Most of those are also init systems, but regardless of that and the other features and functions offered by the init, what do those service managers have in common, what are some commonalities they offer.

Let’s check different features that could be present in them so that we can get the big idea of their role:

  • Start the daemons in a synchronized and deterministic manner according to a configuration file or in an asynchronous and non-deterministic manner according to cross-dependencies between daemons set in configuration files or activated dynamically according to the needs.

  • With the above comes what we talked about earlier about daemon creation but sometimes not necessarily all the steps. All of this also might be configurable, for example you can control the environment variables passed to the daemons from the service manager.

  • Control and monitor all the processes and services in the system.

  • Recover the system in case of errors or crashes, make it more reliable. For example it can implement a configurable restart and retry mechanisms depending on the daemon configuration.

  • Provide useful and detailed debug information in case of a crash: Process details, registers dump and memory map.

  • Create a graphical representation of the system’s processes.

  • An internal and centralized log mechanism implemented in it that can be relied on by the daemon.

  • A centralized way to control the state of daemons, start, stop, and monitor them.

  • An integrated inter-process communication protocol integrated inside of it to control and notify daemons.

This all sounds amazing, right? And it certainly is!

Now how does the service manager do all those things, how does it know which daemon is where and its state.

There are many approach to this problem.

First of all, let’s assume the daemon has implemented the criteria we mentioned earlier when we said it’s a third party entity that starts it.

Apart and additionally from those criteria, which if we want to sum them up are about handling SIGHUP and SIGTERM, returning the right exit code, and use the functionalities provided by that system instead of re-inventing its own.

A way to approach it is to have the daemon not started completely in the background but to leave it as a child of the supervisor so that it will receive SIGCHLD when the daemon stops. The fork/exec approach, which is a rather direct management approach.

Another way is to write the PID of the daemon in a file in a specific location, for example in /run/ or /var/run, and then have the supervisor poll or use something such as libnotify to assess the changes happening to the daemon. It’s also a nice way to have only one instance of a daemon that can’t allow to have more than one.

Yet another way could be to notify the supervisor from the daemon process itself when its initialization is complete. Or when the parent of the sub-second-child exits during the daemon creation.

The blocks are starting to fit into each others, now let’s see what’s up with the init system, why is it so special compared to the supervisor’s roles we just mentioned.

Init & Runlevels

The init is not so different than service managers, actually that’s one of the functionality that an init system can have. The init system itself being a daemon that is automatically started by the kernel during the boot process /sbin/init, it’s the first process.

This isn’t an episode about init systems so we’ll keep it short.

The real difference an init system has as a service manager over others lies in the fact that it’s the first process to start on a Unix system, the PID 1.

Whatever process it starts, then, will have it as parent, and as we said having the init system as parent is one of the criteria to run a daemon, so it can skip the forking step altogether. It’s also nice to manage daemons from the init process because it’s the process that will receive the SIGCHLD when one of it’s children dies, namely the daemon, and thus is in a nice position to manage them.

Because it is the first process it gives the leasure of being able to manipulate, start and stop daemons early on boot and on shutdown, because it’s a process that cannot be killed you can rely on it to efficiently manage daemons.

However that poses one problem we just vaguely mentioned before, the dependency and order in which those daemons are started.

There are many schools of thought that take this approach differently.

For instance, the original research UNIX init simply ran an initialization shell script in /etc/rc.

Today we have two main categories, the synchronous ones and the asynchronous ones.

Let’s take the traditional and prevalent System V as an example of a synchronous one.

A synchronous init system means that daemons are started sequentially, waiting for the previous one to finish before starting the next one, it follows a predetermined simple order, which may lead to delay during the boot.

But daemons are not only started on boot, they can be started or stopped at different stages in time, at different machine states.

To automate that, the SysV init style uses what is called a runlevel, which is a number representing the state the machine is currently in after boot. So for example there could be one for normal user mode, another for single-user mode, another for shutdown, another for booting maybe, another for graphical environment, etc..

When switching from one runlevel to another the init process executes or stops the appropriate daemons that is tagged with the appropriate runlevel number.

Those numbers usually go from 0 to 6 and their meanings vary depending on the implementation, there are only 3 of them that are considered “standard”, otherwise the init refers to the /etc/inittab file which defines what each configured runlevel does in a given system.

  1. Halt
  2. Single user mode (also known as S or s)
  3. Reboot

You can also manually change the runlevel using telinit.

That’s the generic idea behind a the synchronous service manager as init, it’s quite simple.

On the other side of the spectrum we have the asynchronous init systems that manages daemons in parallel, automatically handling dependencies between them.

In those kind of init systems the daemons have to be configured to say what they depend upon so that the init can resolve that dependency. It might even choose to store the way it “rightly” handled that dependency so that it’ll be able to use it without recomputing it on the next boot.

This makes the booting process relatively faster but it doesn’t stop there.

When asynchronously handling daemons you may not even have to rely on things such as state of the machine or runlevels, you can do the daemon activation dynamically depending on multiple conditions and mechanisms at the same time.

For example the bluetooth daemon can be automatically started when a bluetooth capable hardware is plugged in and stopped when plugged out.

Those mechanisms range from the usual activation on boot, to socket-based trigger, to dbus-based activation, to device-based activation which rely on sysfs or udev or whatever, to path-based activation listening to filesystem changes, to timer-based activation, etc..

And don’t forget that you can mix those up for the same daemon. However this all comes at the cost of the init process being more complex and having more dependencies.

Things to care about when its a Daemon

So the daemon is running…

But it’s not connected to a terminal, how does it communicate with the rest of the system and how does it know when to do things?

The three common ways are event based notifications, inter-process communication, and logging.

Let’s start with notifications and events.

There are many things that can be checked and looked at for notifications.

For example you could poll, keep checking once in a while, the filesystem to see if some specific files status has changed and trigger actions in the daemon accordingly.

Or instead of keeping redoing this maneuver over and over again you could rely on instant notification of changing using your operating system facilities such as epoll and inotify on Linux, or Kqueue on BSDs, or more generically a multi-platform abstraction such as libevent.

Another example of notification of system changes could be about the hardwares connected to the machine. You could implement in your daemon a way to do actions based on device events. For that you’ll have to rely on your operating system device manager, devd on FreeBSD, polling the sys/devfs maybe for new devices, or udev or uedev, or again your could use a multi-platform abstraction.

You can listen to the podcast about input devices for more info on that.

Let’s move to the inter-process communication, the IPC techniques available for a daemon.

Well, there are many, actually it could be anything that doesn’t rely on a terminal, so forget about pipes, redirections, or filters.

Let’s mention a few.

The simplest way but not necessarily the safest is to use temporary files. It works using the techniques we just mentioned about notification of changes to that file. The drawback is that it can lead to deadlock, you cannot know which process can access the file at what time.

Another simple way but limited and also not the safest is about using signals. You probably only want to handle the SIGUSR1 and SIGUSR2 signals, so that only gives you two possible actions that you can do after receiving a signal, there’s no data transfer and real communication. If you implement signal handling in a daemon be sure to document what it does.

You can go back to the podcast about signals to know more about them.

A more complex way but again limited, this time limited to the same machine, is the memory map, which maps a file into memory so that it can be shared across processes. It poses the same problems as a file on the filesystem though.

The last IPC method I want to share is one that has many super categories above it, it’s socket based IPC.

Sockets are a bidirectional stream that works on the same machine (AF_UNIX) and over the network too (AF_INET).

So you can simply receive any request from clients that would connect to your daemon, or your daemon could act as a client to a server somewhere else.

There are layers above that, for instane, HTTP and XML-RPC.

XML-RPC is an xml protocol that does remote procedure call, the most prevalent instance of this today is dbus. A remote procedure call simply means that you send a request so that a procedure will be executed remotely and the result will be returned to you.

Many init system wrap daemons with a dbus interface around them, Upstart and systemd do that for instance. It sort of have become a defacto protocol.

It was used as a replacement to similar systems such as CORBA and DCOP, which are truly horrible, and so dbus is clean compared to them.

A simple example you can relate to would be the way you communicate with your notification daemon aka popups daemon aka poptarts daemon. You use the notify-send command which is simply a wrapper over dbus that will send an XML-RPC query to the dbus-daemon which will then notify whoever listens to that type of requests and get the result back. In our case it will be a notification request and our notification daemon will receive it.

Example of something going over the wire:

Method call sender=:1.1257 -> dest=:1.30 serial=7 path=/org/freedesktop/Notifications; interface=org.freedesktop.Notifications; member=Notify
   string "notify-send"
   uint32 0
   string ""
   string "hello"
   string ""
   array [
   ]
   array [
      dict entry(
         string "urgency"
         variant             byte 1
      )
   ]
   int32 -1
method return sender=:1.30 -> dest=:1.1257 reply_serial=7
   uint32 221

Always keep in mind that the simplest organization is the best when it does what it’s supposed to do.

What about logging?

Logging is more generically a user nudge, a way to tell the user what is going on in the daemon, to be able to debug it as there’s no other way to tell.

This could be implemented, yet again, in any possible way.

From sending and email to a user, to a POST on a remote server, to the usual syslog or any logging system and mechanisms.

You can listen back to the episode about logs to get an idea of the possible options.

Conclusion

So that’s about it, that’s everything I have to say about daemons.

Maybe one last thing would be to make sure the daemon code is secure and fails gracefully, there’s nothing more annoying than a daemon that keeps crashing for no apparent reasons.

Your daemon should also do all the blahblahblah of the Unix philosophy, offer one service and do it well.

Yeah, that’s it!

I hope it gave you a great deal of info about the ins and outs of daemons!

As usual, if you have more info or things to mention or rectify about this podcast subject you can do this on the podcast extension thread.


References:

What does Daemon mean, where does it come from

Programming Wise - Foreground/Background process - forking - dissociating

Management - process tracking - Keep running - crash recovery

Init & Runlevels & States

Things to care about when its a Daemon

Notifications & Event based

Permission and files - security

IPC - signals

Logging

  • man syslog
  • or anything else
  • Check the podcast about logs

Music:
Internet People (2017) - SEAMLESS MUSIC
https://www.jamendo.com/track/1436050/internet-people




If you want to have a more in depth discussion I'm always available by email or irc. We can discuss and argue about what you like and dislike, about new ideas to consider, opinions, etc..
If you don't feel like "having a discussion" or are intimidated by emails then you can simply say something small in the comment sections below and/or share it with your friends.