I hated systemd when I had to start using it. It tries to do way too much. The antithesis of the the Unix philosophy. Journalctl grew on me though.
Unix is just a collection of random decisions made by various people over the years. The Unix philosophy is really more like "I've always done it that way so don't you dare change it".
fork/exec is garbage. Signals are garbage. You basically can't make any API calls after fork or in a signal handler because threads came after. The interaction of fork and threads is bananas and many hacks have been required over he years to paper over the problems.
The layout of /usr/bin, /usr/sbin, /usr/local/bin, et al isn't a good design. It's dogshit but it was necessary because early Unix file systems couldn't span multiple volumes and early disk systems were small.
The C compilation model of separate header files is not a good design. People have retroactively determined some of the side-effects are not only good but The One True Way. In reality the design was a result of extremely limited RAM and slow CPUs. The preprocessor itself was never designed, just grafted on ad-hoc.
Unix file permissions are shit. Every unique combination of permissions requires a group. Owners are by simple integers. NFS legitimately gives people nightmares.
Let's not even get into everything is a file, except when it's not, and some files are more equal than others.
What about dependency hell? How's that "simple" model working out?
The "Unix philosophy" can piss right off.
"Unix is just a collection of random decisions made by various people over the years."
"Linux has never been about quality. There are so many parts of the system that are just these cheap little hacks, and it happens to run." - Theo de Raadt
"The layout of /usr/bin, /usr/sbin, /usr/local/bin" I don't agree with you. Having base in a directory and having user installed bins in another makes sense. I don't understand why most modern Linux distros install on only one partition by default. You lose the ability to mount with flags ie noexec,nosuid,nodev.
NFS is horrible. Unix/Linux is a multi-user operating system why would you not want to have groups?
KISS
Now that's backwards. He obviously mean a better permission system. Do you seriously think groups is the best way?
Access Control Lists are a better way.
This is probably just your lack of experience not having worked on 50+ million LOC compiling for 12 hours and not having anything else as a better option. There is a reason these things exist.
I agree: C and C++ compilation model is not a good design. It's a patch for not having a decent module system. Heck, even Borland Pascal compiled faster in 90-ies than C++ does now on an orders of magnitude faster machine.
Enlighten me. And while you're at it, explain why this is better than, say, Rust's module system, where we don't need separate header files.
Why wouldn't something like the C++ module system be of benefit in C also?
It would, but by the look of it, C++ evolves faster these days.
I'd love to see a rebuttal of the specific points made as opposed to just "Most of this comment is incorrect".
But I can imagine what xenadu02 might have meant, if you like, and provide some counter arguments.
Signals aren't "garbage" (whatever that means).
Signals can call APIs (the set of async-signal-safe APIs). They can't call non-async-signal-safe APIs not because of threads, but because signals can interrupt a routine at any point (necessary for asynchronous notification of certain events which must be handled before the normal instruction control flow can be resumed) and that interrupted routine may not have been written to be reentrant.
This is true even without threads in the picture.
The fork/exec model is not "garbage". It is actually a fairly nice alternative to the "provide one API to start a child process and give it a large number of parameters for all possible situations". And you can call plenty of APIs between fork and exec in the child safely, just like from signal handlers.
I haven't dealt with dependency hell ever since shared libraries got sonames.
The rest of the comment doesn't list anything of substance. If you want rebuttals for "the file system layout is a bad design" or "the C compilation is a bad design" or anything else, provide some reasons why those are bad designs; some of those reasons may be valid criticism, and some may not be, but one can't just make vacuous statements like that and expect a reasonable discussion to follow.
> So while signal handlers are perfectly workable for some of the early use cases (e.g. SIGSEGV) it seems that they were pushed beyond their competence very early, thus producing a broken design for which there have been repeated attempts at repair. While it may now be possible to write code that handles signal delivery reliably, it is still very easy to get it wrong. The replacement that we find in signalfd() promises to make event handling significantly easier and so more reliable.
Another critic makes the case that "signalfd is [also] useless" [2]:
> "UNIX[] signals are probably one of the worst parts of the UNIX API, and that’s a relatively high bar."
Signals came up recently on HN when someone remarked that not even memset() is signal-safe! [3]
All in all, working with signals correctly requires mastering a tremendous degree of complexity. Other platforms have provided simpler APIs, such as Structure Event Handling (SEH) [4].
[1] https://lwn.net/Articles/414618/
[2] article link from https://news.ycombinator.com/item?id=9564975
[3] https://news.ycombinator.com/item?id=13313563
[4] An HN comment describing how it's simpler: https://news.ycombinator.com/item?id=13323870
P.S. Please note that the views quoted above are not necessarily my views.
> Why does ls do sorting? Why does grep do -R recursive searching? How is that "Do one thing and do it well"?
I think these are valid examples of how Unix itself fails to follow the "Unix philosophy" of "Do One Thing and Do It Well".
> The fork/exec model is not "garbage". It is actually a fairly nice alternative to the "provide one API to start a child process and give it a large number of parameters for all possible situations". And you can call plenty of APIs between fork and exec in the child safely, just like from signal handlers.
fork-exec complicates the implementation of threads (see atfork handlers). Rather than "a large number of parameters for all possible situations", another alternative would be to have (1) a call which given executable name and arguments returns an opaque handle (or file descriptor) representing the process to be started (2) a bunch of further calls to set attributes on that handle – new features could add new APIs acting on the handle, or an extensible API like ioctl could be used – if there is a handle to represent the current process, then you only need one API call to set it for the current process or a child to be started (3) finally, a start call which turns the process-to-be-started handle into a running process handle.
> Unix file permissions are shit
The user-group-other model is arguably too limiting. ACLs are a better idea, but then should you use POSIX ACLs or NFSv4 ACLs?
The distinction between primary group ID and supplementary group IDs is silly.
Why must every file have both a UID and a GID? For files owned by a single user, you end up creating a dummy group like "staff" or so on just to obey the rule that every file must have a GID. For shared files, e.g. project files, files generally end up owned by their creator, even though in a business sense they really belong to the project not to whoever created them. It would make more sense if the owner could be either a user or a group, and then also have zero or more non-owning groups associated with it.
In most cases permissions should only exist on the directory, and then automatically apply to any files in the directory. (In most cases every file in the same directory should have the same permission; Unix bases its design on the exception rather than the rule.) Of course, hard links make this impossible, but I think hard links were a mistake.
The executable permission bits actually do double duty as a file type indicator. That's rather ugly. If Unix had explicit file types (rather than just a naming convention of file extensions), then certain file types could be declared to be executable. Executable permission would then mean "you are allowed to execute this if it is an executable" instead of "this is an executable". Stuff like the +x vs +X distinction in chmod would never have been necessary.
> Let's not even get into everything is a file
Unix would have been much better if everything were a file descriptor, rather than having stuff like pid_t. Linux at least is evolving in this direction. Plan9 does it better. Even the WindowsNT philosophy of "everything is a handle" is better than the traditional Unix approach.
The rationale for this is that POSIX ACLs are firstly too simple to model what we need. And they are also non-standard (POSIX .1e ACLs are a DRAFT specification which was never ratified).
NFSv4 ACLs are vastly more featureful, already implemented to support NFSv4 in kernel, though not available in userspace AFAICT. On FreeBSD and other platforms using ZFS, they are also used by ZFS and are directly exposed to userspace, making rich ACLs usable as the default permissions model system-wide when running on ZFS. Linux, unfortunately, doesn't yet do any of this, even when using ZFS.
Fork-and-exec isn't complicated by threads. Only fork-and-keep-executing is.
UNIX doesn't have a naming convention using file extensions.
Some of your points are valid opinions that are shared by others, but I don't know how much they have to do with the UNIX philosophy.
Some APIs can be improved, sure. And some are being improved. It takes time because of unix's success and most systems' desire to remain backward compatible (especially in source form).
Another issue is that fork-and-exec doesn't work well with languages with complicated runtimes, e.g. multithreaded garbage collection. It forces you to use a lower level language (such as C) to write all the code between fork and exec. An API based on process handles with a separate "start" call to convert a not-yet-started handle into a running process wouldn't have that deficiency.
Another issue is that it is very hard to implement robust error handling without race conditions in the fork-exec model. What if the child process encounters an error between the fork and the exec? How does it notify the parent process of exactly what error it got (e.g. "setsid failed"?) You need some sort of IPC mechanism between the child and the parent. And such an IPC mechanism is prone to race conditions. By contrast, the process handle-based API I suggested doesn't have this problem since it doesn't introduce more concurrency into the system than is absolutely necessary.
> UNIX doesn't have a naming convention using file extensions.
Yes it does. The average Unix system is full of file extensions like .c, .h, .so, .html, etc. Even in Unix V1 file extensions were used as a convention - http://minnie.tuhs.org/cgi-bin/utree.pl?file=V1
> Some of your points are valid opinions that are shared by others, but I don't know how much they have to do with the UNIX philosophy.
Is there a clear definition of what the "UNIX philosophy" is? Is any criticism of Unix systems as actually implemented a valid criticism of the "Unix philosophy"? Or do you want to define the "Unix philosophy" so vaguely as to put it beyond any possibility of criticism?
How are you doing fork-and-exec in a language with a large runtime? You are either using the language-provided APIs to do it, in which case they should document the restrictions on what you can call (and you should follow those), or you are dipping down into the C or system call layer to do your own fork-and-exec, in which case yeah, you still need to keep to the safe list of routines you can call between fork and exec, and you may have extra limitations since you are mucking around underneath your language's runtime (like you may have to unignore signals on your own, close file descriptors, etc). No surprises there.
> Another issue is that it is very hard to implement robust error handling without race conditions in the fork-exec model.
I don't think it is. You just print an error to stderr (write() is safe to call), and you return a bad error code (fork has built-in IPC for error codes via wait() in the parent).
> Is there a clear definition of what the "UNIX philosophy" is?
I don't know, ask the person who first invoked that phrase in this thread. They claimed it meant "do one thing and do it well" to them, and then they complained about things that didn't seem related to me (like file extensions, what does that have to do with programs "doing one thing"?).
> How are you doing fork-and-exec in a language with a large runtime? You are either using the language-provided APIs to do it, in which case they should document the restrictions on what you can call (and you should follow those), or you are dipping down into the C or system call layer to do your own fork-and-exec, in which case yeah, you still need to keep to the safe list of routines you can call between fork and exec, and you may have extra limitations since you are mucking around underneath your language's runtime (like you may have to unignore signals on your own, close file descriptors, etc). No surprises there.
Let's say I am using JNA – https://github.com/java-native-access/jna – under Java. It is safe to call posix_spawn from Java code using JNA. It is safe to call the Windows API equivalent (CreateProcess). It would be safe to call the handle/descriptor-based API I proposed. It is not safe to call fork. This is an undeniable deficiency of the fork-exec approach which competing approaches don't have. Furthermore, whatever compensating advantages fork-exec may have, the handle/descriptor-based API I proposed has the same advantages without this disadvantage.
> > Another issue is that it is very hard to implement robust error handling without race conditions in the fork-exec model.
> I don't think it is. You just print an error to stderr (write() is safe to call), and you return a bad error code (fork has built-in IPC for error codes via wait() in the parent).
But that isn't robust. How can the parent process reliably distinguish output sent by the child process prior to the exec from output sent by the child process post the exec? Likewise, how can the parent process reliably distinguish an error return value from the child process prior to the exec from an error return value from the exec'd program? It can't.
For truly robust error handling, you'd actually need to do something like this: (1) have a pipe between parent and child process with FD_CLOEXEC set on the child side; (2) the child sends the parent a message "I'm about to exec" before calling exec; (3) the child sends the parent a message saying "exec failed with errno=.." if the exec call fails; (4) if the exec call succeeds, the child process will close its end of the pipe without sending any message post "I'm about to exec". This is my point, actually robustly handling errors in the fork-exec model is quite complex. In a handle/descriptor based API it would be much simpler.
(And the above approach using a pipe isn't perfectly robust – what if the child process crashes for some reason between sending the "I'm about to exec" message and actually calling exec()? It is very difficult for the parent process to reliably distinguish that scenario from some failure in the program being exec()'d.)
Are you calling fork() from Java, from C, or using the system call number?
Because I'd agree calling it from Java might be unsafe (depends on how Java and JNA interact), but I believe calling it from C or the system call is perfectly fine. And this is in line with what I've written previously.
> But that isn't robust.
It's not supposed to be robust in the way you are describing.
The fork-exec model is low level. It is supposed to be low level. Doing high level things with it is supposed to take some work by the application. That's not a deficiency.
If you build too many things into the low level code, you run into trouble because now you've got 10x as many ways to fail (building your pipes, writing your error messages, marshalling error state, cleaning up, you name it).
Also, some programs will want to do some of those higher level things differently, so instead of baking them into the API and having tons of parameters and paying for some of that overhead (like creating a pipe and writing error messages to the parent for every single fork and exec) you only do that when you want it.
However, FreeBSD has had process descriptors since roughly 2010. They have the slightly odd semantics of terminating processes when all descriptors to them are closed. But they can be used as descriptors with kqueue() and the like.
* Integrated daemon status. This one caught me by surprise, since the
systemd journal was functionality that I expected to dislike. But I was
surprised at how well-implemented it is, and systemctl status blew me
away. I think any systems administrator who has tried to debug a
running service will be immediately struck by the differences between
upstart:
lbcd start/running, process 32294
and systemd:
lbcd.service - responder for load balancing
Loaded: loaded (/lib/systemd/system/lbcd.service; enabled)
Active: active (running) since Sun 2013-12-29 13:01:24 PST; 1h 11min ago
Docs: man:lbcd(8)
http://www.eyrie.org/~eagle/software/lbcd/
Main PID: 25290 (lbcd)
CGroup: name=systemd:/system/lbcd.service
└─25290 /usr/sbin/lbcd -f -l
Dec 29 13:01:24 wanderer systemd[1]: Starting responder for load balancing...
Dec 29 13:01:24 wanderer systemd[1]: Started responder for load balancing.
Dec 29 13:01:24 wanderer lbcd[25290]: ready to accept requests
Dec 29 13:01:43 wanderer lbcd[25290]: request from ::1 (version 3)
Both are clearly superior to sysvinit, which bails on the problem
entirely and forces reimplementation in every init script, but the
systemd approach takes this to another level. And this is not an easy
change for upstart. While some more data could be added, like the
command line taken from ps, the most useful addition in systemd is the
log summary. And that relies on the journal, which is a fundamental
design decision of systemd.
And yes, all of those log messages are also in the syslog files where
one would expect to find them. And systemd can also capture standard
output and standard error from daemons and drop that in the journal and
from there into syslog, which makes it much easier to uncover daemon
startup problems that resulted in complaints to standard error instead
of syslog. This cannot even be easily replaced with something that
might parse the syslog files, even given output forwarding to syslog
(something upstart currently doesn't have), since the journal will
continue to work properly even if all syslog messages are forwarded off
the host, stored in some other format, or stored in some other file.
systemd is agnostic to the underlying syslog implementation.
[Unit]
Description=Demo service
[Service]
Type=forking
ExecStart=/usr/sbin/my-daemon
PrivateTmp=yes
PrivateDevices=yes
PrivateNetwork=yes
CapabilityBoundingSet=CAP_NET_BIND_SERVICE
I can drop capabilities with CapabilityBoundingSet=, or limit resource usage with CPUSchedulingPriority=, IOSchedulingPriority=, etc. I could even tell systemd to open the listening socket for me so the service doesn't need CAP_NET_BIND_SERVICE! Moving these options into the init system makes a ton of sense, because it gives administrators the ability to employ these features from outside applications, not just by enabling them within applications that bother to explicitly support them via command line arguments. Systemd better encourages the principle of least privilege: if a system daemon does not need the ability to "ptrace" other processes, or bind to ports <1024, then as the administrator I can take those away with CapabilityBoundingSet= in the unit file. Chrooting the service is as easy as RootDirectory=. This is a huge step forward compared to the world where every service must be relied upon to expose these settings, and must be trusted to implement them correctly.
[1] https://lists.debian.org/debian-ctte/2013/12/msg00234.html
Also, I have never seen a systemd box emit log lines like that for a failed service. It invariably points at some useless logfile with obscure systemd messages in it instead of the stderr of the failed process. This is on clean ubuntu and debian installs. Maybe it is user error, but I doubt it. (Though there is no command line in the examples...)
Anyway, I'm happy to cleanse with fire instead of RTFM at this point. On a related note, I just learned the solaris init system and started using openbsd's again.
I prefer them both to systemd. They are at opposite ends of a spectrum. The openbsd approach is well curated shell scripts. I think systemd was heavily inspired by the solaris thing.
It doesn't. Systemd has its own resolver (systemd-resolved), which has other issues, but it does not run in PID 1. It's a completely separate process.
// I'm enjoying watching the points bouncing up and down, but if you disagree in some way, please comment :)
What exactly does 'PrivateUsers' do? What uid do I have? When I write that uidin a db, what value does it keep? Between invocations, does the uid change or is it per unit? If a file is owned by a private uid, what do other processes on the system see? Is PrivateUsers for this unit file only, for the unit files in this group of unit files, across the entire system, across the entire cluster? If I want two different programs to share this PrivateUsers concept, how do I do that?
It turns out that gluing random shit to the side of a monolith gives you the illusion of convenience, but since the monolith will not do that thing well -- for example, identity management -- you will end up with some programs that adopt the half-assed solution, and some programs that are forced to do things a different way because their use case is complex. Now you have two problems.
https://www.freedesktop.org/software/systemd/man/systemd.exe...
Systemd's job and goal is to provide a simple configuration file format that makes it easy to enable these features with installed system daemons.
You may overlooking the parameters supported by systemd and their benefits. Admins get a single way to manage their services and dependencies, and a way to do that that works consistently across all services. These features can be enabled even if the services were not designed for it (e.g., chroot). With systemd you can employ these settings from the outside with any service, and that's a big advance. Difficult to achieve otherwise.
> The point was that it's a leaky abstraction composed of half implemented concepts
I am unclear what part you want to criticize. The Linux kernel is what provides the User Namespace feature. Systemd helps users take advantage of it. What part do you consider half-implemented or a leaky abstraction?
Please also consider whether you might have the wrong mental model for the feature or its employment. In particular, the documentation for PrivateUsers= says: "This is useful to securely detach the user and group databases used by the unit from the rest of the system, and thus to create an effective sandbox environment." The usage you had in mind when you wrote your comment may not be compatible with the purpose of PrivateUsers. I recommend reading up systemd.exec params [1] before criticism. PrivateUsers= is intended for scenarios like transient sandboxed environments, so I'd suggest we discuss a simpler example like PrivateDevices=
[1] http://man7.org/linux/man-pages/man7/user_namespaces.7.html
I'm aware of what systemd claims to be. I'm aware of the benefits that its fans claim it, and it alone has.
My criticism is not with the Linux kernel (!?!) or user namespacing as a concept. My criticism is that systemd takes all of the rich complexity of user namespacing and, in response, adds the flag 'PrivateUsers=yes' -- a boolean. That's not what user namespacing is for and now we have two problems: systemd, which has no business making that decision and has done it the wrong way, and the continuing need, which has not been solved by a boolean flag, for daemons to have competent, complex user namespaces. Now devs have to know both ways: the half-assed way, and the real way, instead of just having a tool that gives them the real way.
That's what we in the software design business call a shitty design that would make Guy Fieri blush. But I guess we're all in Flavortown now.
As opposed to a multitude of implementations of varying quality and functionality for the same concept in each init script that needs it?.