16

I have two simple programs: A and B. A would run first, then B gets the “stdout” of A and uses it as its “stdin”. Assume I am using a GNU/Linux operating system and the simplest possible way to do this would be:

./A | ./B

If I had to describe this command, I would say that it is a command that takes input (i.e., reads) from a producer (A) and writes to a consumer (B). Is that a correct description? Am I missing anything?

  • Related: In what order do piped commands run? – G-Man Apr 21 at 21:35
  • It's not command, it's an kenerl object created by bash process, which is used as stdout of process A and stdin as B. Two processes are started nearly at the same time. – 炸鱼薯条德里克 Apr 22 at 0:45
  • 1
    @炸鱼 You're correct - for kernel pipeline is an object in pipefs filesystem, but as far as shell itself is concerned - technically that's a pipeline command – Sergiy Kolodyazhnyy Apr 22 at 1:30
26

The only thing about your question that stands out as wrong is that you say

A would run first, then B gets the stdout of A

In fact, both programs would be started at pretty much the same time. If there's no input for B when it tries to read, it will block until there is input to read. Likewise, if there's nobody reading the output from A, its writes will block until its output is read (some of it will be buffered by the pipe).

The only thing synchronising the processes that take part in a pipeline is the I/O, i.e. the reading and writing. If no writing or reading happens, then the two processes will run totally independent of each other. If one ignores the reading or writing of the other, the ignored process will block and eventually be killed by a SIGPIPE signal (if writing) or get an end-of-file condition on its standard input stream (if reading) when the other process terminates.

The idiomatic way to describe A | B is that it's a pipeline containing two programs. The output produced on standard output from the first program is available to be read on the standard input by the second ("[the output of] A is piped into [the input of] B"). The shell does the required plumbing to allow this to happen.

If you want to use the words "consumer" and "producer", I suppose that's ok too.

The fact that these are programs written in C is not relevant. The fact that this is Linux, macOS, OpenBSD or AIX is not relevant.

  • 2
    Writing to a temporary file was used in DOS, as that didn't support multiple processes. – CSM Apr 21 at 17:56
  • 2
    @AlexVong Note though that your example with a temporary file is not exactly equivalent. A program may choose to seek though the contents of a file, but data coming off a pipe is not seekable. A better examlp would be to use mkfifo to create a named pipe, then start B in the background reading from the pipe, and then A writing to it. This is nit-picking though, as the effect would be the same. – Kusalananda Apr 21 at 17:58
  • 2
    @AlexVong The simplifications made in that article divorce it from real pipelines; the parallel execution is truly semantic, not an optimisation. It's a reasonable lies-to-children explanation of monadic evaluation or composition for someone who's seen shell pipelines, but it's not valid in the other direction. Kusalananda's fifo version is closer, but the error propagation parts of the model are genuinely important and not replicable. (all of which I say as someone who is very on the "shell pipelines are just function composition" train) – Michael Homer Apr 21 at 21:41
  • 6
    @AlexVong No, that's completely off track. That isn't able to explain even something simple like yes | sed 10q – Uncle Billy Apr 21 at 22:29
  • 1
    @UncleBilly I agree with your example. This shows that parallel execution is really required also noted by Michael. Otherwise, we'll get non-termination. – Alex Vong Apr 21 at 23:53
2

The term usually used in documentation is "pipeline" , which consists of one or more commands, see POSIX definition So technically speaking, that's two commands you have there, two subprocesses for the shell (either fork()+exec()'ed external commands or subshells )

As for producer-consumer part, the pipeline can be described by that pattern, since:

  • Producer and Consumer share fixed-size buffer, and at least on Linux and MacOS X, there's fixed size for pipeline buffer
  • Producer and Consumer are loosely-coupled, commands in pipeline don't know of each other's existence ( unless they are actively checking /proc/<pid>/fd directory ).
  • Producers write to stdout and consumers read stdin as if they were a single command being executed, aka they can exist without each other.

The difference I see here is that unlike Producer-Consumer in other languges, shell commands use buffering and they write stdout once buffer is filled, but there's no mention that Producer-Consumer has to follow that rule - only wait when queue is filled or discard data (which is something else that pipeline doesn't do).

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.