Linux Signals – Internals | Hacker News Reader

Linux Signals – Internals

sklinuxblog.blogspot.com

90 pointsingve30 comments

I really hope when the time comes and someone reinvents the wheel and comes up with a new great OS, the new OS will NOT have signals. Those are one of the worst things ever invented. The idea that your code is randomly interrupted and it starts to execute something else is just mind-blowingly bad. The real fun comes when you want to use different libraries with different signal handling ideas in the same process. It all goes to very quickly. The worst thing is that no one had the courage to say "okay guys, this is bad, let's just forget it ever existed and create a better alternative". /rant

I'm not an operating systems person, so I'm curious - what are the alternatives?

Win32 uses messages for everything. While Windows does include signals[0], only three are actually used (SIGABRT, SIGFPE, SIGTERM), and in practice I don't think I've ever seen a Win32 program implement signal handlers.

[0]: https://msdn.microsoft.com/en-us/library/xdkz3x12.aspx

I don't think message queues and signals serve the same use case. Signal could be used for out-of-band communication.

You may be right, but in practice the syscall is so limited that I think many programs basically process the syscall as a message anyways, that using a messaging model may make more sense, and be easier to work with.

Someone correct me if I'm wrong, but because of the way the syscall callback operates, it could be interrupting threads at almost any stage. This makes executing the signal in a safe manner very difficult, and from the signal handler itself you can only run certain code. If i remember correctly, most signal handler implementations then get reduced to basically set some state and return, where the regular execution will then check for that state and react to it. Well, if your only going to process the signal during you main event loop anyways, then why not just exchange the information over a message queue, and send a message to the process when you want it to do something (ie reload configuration).

Well, from what I know, sometimes you just have to have out of band communication, check the discussion in the comment here: http://250bpm.com/blog:70

It sounds like a necessary use case to me. While most people would probably not use it, that doesn't invalidate its existence. Most people would probably not touch any low level code either way.

Actually, structured exception handling (SEH) is the closest equivalent of UNIX synchronous signals (SIGSEGV, etc., those that are caused by the program itself) on Win32. As for asynchronous signals, there's APC (asynchronous procedure calls). (See https://msdn.microsoft.com/en-us/library/windows/desktop/ms6...)

I find Win32 vastly superior than UNIX in this area.

Signals are the user-space equivalent of a hardware interrupt. It's true that almost any computer that does anything useful in the real world needs to be notified about that world using interrupt service routines (ISRs). But that doesn't mean user programs need interrupts. OSes can and do use interrupts as a cues to notify processes via very different abstractions. Heres one way:

  User Process 99:  
    read socket Foo.
  Kernel:
    socket Foo has no data yet, 
       move 99 from the "running" list to the "sleeping" list.
    run process 86
  User Process 86:
    do stuff
  << HARDWARE INTERRUPT!>>
  Kernel:
    Notice that 99 is waiting for this packet (via socket Foo).
    Copy network data into 99's read buffer.
    Move 99 back to the "running" list.
    Run some processes, soon enough 99 gets a turn.
  User process 99:
    Oh good!  I have data.

This sort of interface is usually much more useful for user applications than having the equivalent of their own ISR becuase it doesn't just send a notification -- it also controls the flow of the main application in a sane way. Simple, non-interactive applications can do this kind of blocking I/O all day long.

More responsive applications need some kind of event loop. I.e. instead of blocking on an concrete I/O resource, they block notification service which tells them what I/O is available. In very different ways, Windows messages and Unix select()/poll() both do this.

The end result is usually a callback driven program. This is slightly similar to signals/ISRs (since signals are a kind of callback) -- but the game-changing difference is that the callbacks are only called when the application has voluntarily gone back to the event loop.

I'm not sure I understand this example. Signals allow the program to receive a notification whether they are blocked on I/O or running on the CPU (and receive it immediately in the latter case).

The application can integrate signals with its event loop with the self-pipe trick, or it can use Linux-specific APIs to have signals delivered over a file descriptor.

http://man7.org/linux/man-pages/man2/signalfd.2.html

I'm not saying signals are fun, but I don't think you've proposed anything different/better.

> Signals allow the program to receive a notification whether they are blocked on I/O or running on the CPU (and receive it immediately in the latter case).

No, they don't. Just because signal handler code gets executed immediately, doesn't mean that the progam "receives" anything, as you cannot really touch anything that's also touched by the rest of the program, as you'd usually produce some form of race condition. The little that you can do safely usually is functionally equivalent to setting a flag for the rest of the program to process, which you can just as well achieve with any other "non-immediate" IPC mechanism, with much lower risk of getting it wrong.

> The application can integrate signals with its event loop with the self-pipe trick, or it can use Linux-specific APIs to have signals delivered over a file descriptor.

Which, for all itents and purposes, transforms them into yet another pipe/socket/event source, in wich case you might as well use one of the numerous other variants of those.

Sorry if I gve the impression I was prpoposing something new. I was trying to explain, in concrete terms, to a someone who is "not an operating systems person" what existing OSes do.

The self-pipe trick just shows that what most applications need is an event polling mechanism, not a preemptive callback. As for signalfd, I'd say that is exactly the kind of interface nemaar wants INSTEAD of traditional signals.

Message passing, mainly. I.e. if your program is an event-based loop, then an interrupt is just another kind of message that you can receive on your main thread.

Signals are almost the app version of interrupts which is a fundamental way the peripherals (and your programs!) communicate with the kernel.

Sure handling them can be a pain in the ass, but I'd argue that signals are beautiful in their simplicity and extremely powerful. They are used for far more than signaling hang up and termination.

Signals are an incredibly brittle interface. Not only can you lose signals all the time (if a signal is sent to a process while it is handling another signal then you can lose that signal), but you lose a lot of semantic information because the model of signal is "random callback that you jump to and you lose your old stack state while in the signal" is incredibly lossy. There's a lot of other problems (signals not being queueable, as well as the fact that signals expose a bunch of kernel race conditions by design) but you get the gist.

Unix did a lot of things right, but signals was certainly not one of them. Sometimes the simple ideas aren't the best ones.

> if a signal is sent to a process while it is handling another signal then you can lose that signal

> signals not being queueable

Are you sure? I'm not an expert on signal handling, but I believe these statements are not true in the case of signalfd(2), which allows you to read signals from a file descriptor and does not require a signal handler.

It is necessarily true because sending signals cannot block and queueing an unbounded number of signals would require unbounded amounts of memory.

Signal sending cannot block, so you'd need to have an unlimited amount of memory in order to guarantee that all signals are delivered. Not to mention that some signals allow you to block a process in the middle of a syscall (resulting in the lovely restart_syscall and EINTR hacks).

Also, signalfd(2) has its own lovely host of problems (caused by the fact that a file descriptor and signals don't match in abstractions). On fork(), sendmsg() or exec() or any other interesting event, the file descriptor starts to lean towards breaking POSIX. In addition, because of signals' interactions with threads, you get some even odder interactions when you read from the fd in a multithreaded program (they share the file descriptor table but will read different data).

> signals not being queueable

real-time signals are queueable.

CPU interrupts aren't too bad in a kernel, because a kernel is built around the idea of frequently switching task, so one more kind of switching isn't a big deal. Also, when the kernel asks the CPU to do something (like write to memory), the CPU will generally wait until the task is complete, or reaches some convenient checkpoint, before stopping and jumping to the interrupt vector.

On the other hand, most programs aren't designed for context switching, so the design overhead of dealing with signals is significant. Furthermore, signals can arrive at any time, even during a syscall, and while it's possible to do reliable I/O in the face of random interruptions, it's quite tricky.

Signals seem like the kind of thing one would inflict on userspace after having had to deal with interrupts in the kernel, yes. "Share in my pain, userspace programmers!"

A standard mechanism needs to exist to send a message to an arbitrary process for things like "please gracefully terminate", but I think the ideal interface looks like a combination of 1) signalfd (minus some of the legacy signal-specific bits), for any signals the process can handle, and 2) an in-kernel mechanism to SIGKILL, SIGSTOP, or SIGCONT processes, which the process can't do anything about anyway.

(And things like SIGWINCH should never have been signals; those should just be a standard message associated with the terminal device itself.)

> but I'd argue that signals are beautiful in their simplicity and extremely powerful.

There's nothing beautiful when you have a MT program which also has to handle signals for various reasons. Signals are an abomination, not the least because a signal handler is global for the whole process.

It should be noted that the example code is defective: printf() is being called inside the while loop and in the signal handler, which could lead to the printf() in the loop being interrupted by the signal handler, which would call printf() again, but printf() is not reentrant.

Doesn't ctrl-z actually send SIGTSTP rather than SIGSTOP? I think the difference is that SIGTSTP can be caught, while SIGSTOP is always an unconditional stop.

That's correct.

I like the use of jprobes to show the call stack within the kernel. I wish all of the tracing functionality within Linux would have one API (like DTrace). I considered writing a module to provide such an API, but there's so many complications when trying to create conditional probes (you don't have enough power with eBPF, and the default string-based globbing is useless).

I keep meaning to learn Linux from the ground up, by following the bootloader assembly, kernel initialisation etc and be able to track it in the kernel code, but its so vast I forget half of it, as I try to make my way (sporadically) through all there is.

There are so many hours in the day.

I get that feeling; I've been meaning to do that for OS X myself.

Appendix A of Understanding the Linux Kernel (3rd edition) has a good overview of this.

I am trying to read this book from cover to cover while I am in hospital. It is vastly more entertaining than the cable TV.

Or you can read Linux Insides… :-)

https://github.com/0xAX/linux-insides

Lots of good HN discussions of Linux signals:

https://hn.algolia.com/?query=linux%20signals&sort=byPopular...

Reply on news.ycombinator.com