Linux pipes and open3

Three articles on how open3 is implemented in Linux, and how it is properly used

Open3 is a common name for a function that spawns a process and substitutes its standard input, output and error streams with the pipes connected to the original process. This way you may interactively feed the child with data, and read the processed data back.

The problem with open3 is that its power renders it more dangerous, and its usage—more error-prone. As the famous Spider-Man quote shares, "with great power comes great responsibility", and Linux inter-process communication is not an exception. So here comes the effective usage advice number 1.

Avoid open3 unless necessary

For the purpose of clarity, throughout the post, standard input, output, and error file handlers are called "standard filehandlers"; the process that spawns, and the process that is spawned are the "parent", and the "child" respectively.

The first advice about a safe open3 usage is... to avoid its usage. Languages that are popular in Linux world usually come with a handful of tools for process spawning, such as:

Shell redirection instead of pipes

Instead of using open3 in your program, you may try to connect process inputs and outputs with Bash redirection utilities, invoking the shell via standard functions to spawn processes.

For instance, you may write system("find | gzip >archive.gz") in Perl to spawn a process that archives recursive directory listing into a file, instead of passing the data through the parent process. You may use a usual redirection into just a text file as well, if that's what you really need.

However, I wouldn't use intermediate files if they're only to dump information temporarily for an immediate processing by something else, like this:

I consider it a bad stlye, let alone that it requires extra time on disk access and synchronization, while it may be totally unnecessary. The arguments to the commands passed that way will also undergo the interpretation by the shell, so you'll need to escape your arguments, a painful and an error-prone process.

If you need the intermediate result saved in the files for debugging purposes, you stil may dump it in debug mode, saving cycles when your project is compiled for production usage.

  • system — spawn the process, inheriting parent's standard file handlers;
  • fork+exec — ditto, but, possibly, with additional customizations you learned about in the previous post, for instance (such as tuning environment or playing with file handlers);
  • popen—open only one pipe to the process, either for reading or writing (note that popen also suffers from some of the typical misuses of pipes).
  • shell redirection untilities—Bash has a number of measures to take control over how to output and redirect input and output of the process. A section in one of my older posts is devoted to pipes in Bash. What could be helpful is the ability of using shell interpreter when spawning a process with such functions as system in Perl. For instance, you may write system("find | gzip >archive.gz") in Perl to spawn a process that archives recursive directory listing into a file, instead of passing the data through the parent process (see side note). You may use a usual redirection into file as well, if that's what you really need.

If one of these tools fulfills your needs, then go for it! You can replace it with open3 at any time anyway—unlike houses, software is relatively easy to rebuild.

Where open3 should be used

However, if you'll find the alternatives listed above inefficient, or not powerful enough, you may opt our for using open3. I'll try to enumerate the cases where I would certainly advise using open3.

What is select?

The select function (man page) is a common name for a function to wait on file handlers. Assume you develop a task scheduler that connects to several machines via sockets; it should wait for the first socket to have the result in to schedule the next task to that specific machine (since others are obviously busy). How do you accomplish that?

Of course, there is a faster and more easy-to-use alternative than trying a non-blocking read on each socket once per several milliseconds. It is using the select() function (or one of its companions, such as poll() or kpoll) to perform a blocking wait on all of them. This function will return the filehandler list of those that have the data readily available in them—as soon as there will be at least one such descriptor!

You may find a plenty of manuals on how to use select; later in this post you'll see an example.

  1. several consumers of one source—if you are given a source of data that should be redirected to several consumers (such as some data that should be both printed onto the terminal and saved into an archived log file for history keeping reasons), you should connect a pipe to its output and error streams, and when you read some amount of data from there (for instance, when you read a line), send it to all the consumers you have. In some cases you can do it with Bash (not with sh), but the code with open3 should look much more clear.
  2. aggregation of several providers—if you maintain several data sources, which all should be aggregated into a common place (a text file, for instance) that has some concurrency issues, you might benefit from open3. You may spawn the processes, and then select() (see sidenote) their output streams, thus making writing of the data to the aggregator thread-safe.
  3. interactive data exchange with a continuously-running process—you can't avoid open3 if you spawn a process that responds to your input with some text to stdout. A lot of axillary processes that are launched thousands of times, have the interface of reading from stdin, and writing to stdout (such as SAT solvers), and you most likely don't have the time to write or read anything from the disk.

One of the examples of the application of these patterns (more specifically, a combination of the first and the second) is transparent logging—it's a blatant crime to use anything but open3 for this. By transparent logging I mean combining logs from the child processes into a common logging sink in the parent one. Usually it's done automatically: the system call just routes standard output of the child to that of parent. However, assume you spawn several concurrent processes, and unless you prefix each of them with a unique identifier, you'll get lost in the log quickly.

This may be solved by opening these processes with open3, and attaching prefixes to their output lines before printing them. Note also that this way you may control severity of logs: for instance, you might want treat standard error and output streams from the children differently, and that can not be achieved with a simple popen call.

Synchronous vs. asynchronous

In this post we will mostly study issues with a synchronous processing of data exchange with the child process. Being synchronous means that nothing else but the interaction with the process is performed while the child is alive. In other words, there is only one thread in the parent.

Some notes on asynchronous processing will be given at the end of this post. Still, I consider synchronous experience extremely valuable, as it helps to shape the vision what a good asynchronous interaction via open3 should be.

Issues with open3

So far I've been telling that open3 is not straightforward to use, but what causes thee complexity? Why could there be the "Dead Locks" referenced in the title of this post? I had to learn this in the hard way, tacking cumbersome bugs in our project, and here's what it was about.

Following the pattern "aggregation of several providers", I spawned the process that wrote to both stdout and stderr with the intent to combine them both in the archived log file. The code (simplified) looked like this:

I expected to get a file that has all the lines from the stdout of the child, and prefixed lines from the stderr of the child afterwards. However, sometimes the application just deadlocked!

A quick strace demonstrated that the parent hung up on read from child's stdout, and the child hung up... on write to its stderr! How could that be?

Remember that in the first post about pipes, I listed the limited capacity as one of the properties of the pipes. I stressed it on purpose, because it plays its role right here, when you try to use open3. The pipe that connected the parent to the child was full, and the child wanted to write an error message there. While the parent was still reading from the output pipe, because the child was still running, and its pipes were not closed! That's the open3 Dead Lock.

Of course, this could happen with any pair of pipes here. Assume the process just echoes every character it takes on the input to the output (of course, real programs will be doing a useful transformations, but for clarity we may assume it as identical). We want to feed it twice as much characters as the pipe's capacity.

You might think that this won't strike you unless you're dealing with extremely long inputs and outputs. Sorry to disappoint you, but, quoting the man 7 pipe:

In Linux versions before 2.6.11, the capacity of a pipe was the same as the system page size (e.g., 4096 bytes on i386). Since Linux 2.6.11, the pipe capacity is 65536 bytes.

It's not much, though, in certain cases, it's big enough to let badly written programs work. On a larger scale, we definitely need a generic, limit-agnostic solution.

How to prevent the dead lock

It's relatively simple to devise a generic rule of mitigating the effect of such a limitation. To avoid deadlocks with open3, you should clear each of the output pipes (and fill the input pipe) as soon as possible, and do not put a dependency between clearing a pipe and waiting for another pipe. So we need to watch closely to all the file handlers (up to 3) which we plan to read/write data from/to—and it's important that you wait for all of them at once! If you read the sidenote above, you already know that we could use select for this.

Using select to react to input promptly

So, a potential way to fix the program above is to write something like this:

This program, however, only outlines the approach to tackling deadlocks with open3. It is still prone to deadlocks, though less than the original one. Assume that the child started to print a line to its stdout, but haven't finished it, because it got a debugging interrupt. Having not finishing printing the line, it spits 100 Kb of debugging information to stderr with the intent to continue printing to stdout the normal output. However, the parent is still blocked in the out.readline() call, waiting for the line termination character to appear there, and the child gets blocked at the write to stderr, because the err pipe is full, and no one's going to remove data from it. Deadlock again. (You may play with this deadlock by open3-ing this sample program).

The issue here is that we still do not "remove data from pipes as soon as possible". For that, we need nonblocking reads, more low-level than those of the readline()- and scanf-like functions.

Using nonblocking reads and your own buffers

The problem with nonblocking low-level reads, as Capt. Obvious notes, is that they are low-level. We can't read more or less structured data from them. Assume that we want to read a number (a number of seconds to wait before launching the starship, for instance) from the child's stdout. If that debugging interrupt described above is triggered just in the middle of printing 1000, our nonblocking read will read 10 (before turning to reading from stderr), and act accordingly, launching the multi-billion-dollar ship prematurely. From the child's viewpoint, however, doing so is totally legitimate, since it printed 1000 and debugging information to the different output channels (stdout and stderr), and if the reader confuses these numbers, it's its problem.

Do not use strings as buffers (like I do here)

In the samples below we used "strings" as buffers. However, strings in the modern scripting languages (including those used here, as it's Ruby) consist of multi-byte characters with variable length (see Joel's post on Unicode), and not with one-byte symbols. On the other hand, in some less modern and "less scripting" languages, strings can not contain zero bytes, as they would be treated as the end of the string.

Therefore, "strings" are going to misbehave if chosen as a buffer for an abstract byte stream. I used them for simplicity, and for the sake of demonstration of open3 usage; in real programs, however, you should not use them.

Therefore, we need to handle these situations accordingly, adding another level of indirection between the child and the parent. We will store the data we read from pipes in the intermediate storage; we could name it a buffer. In fact, we are going to re-implement buffered read.

In the next example we will implement a linewise-open3, a subroutine that invokes user-defined callbacks at reading a complete line from stdin or stderr. You could play with it more, introducing scanf-like behavior (instead of these regexps you'll see). However, we have two more issues to discuss before getting to the code.

Reactive interaction with both input and output data

The select with callbacks works well for reading output with the input pipe closed at all. What should we do to if we want to write something to the input pipe of the child based what's being read from it?

To talk interactively with the child, you'll most likely need an asynchronous processing. However, there is one pattern which allows the exchange of data through both the input and the output in the synchronous mode. We already agreed that we will invoke user-defined callbacks after reading lines from stdin and stdout. However, we didn't use the return value of these callbacks in any way! The idea about this arises immediately:

If the user-defined callbacks to stdout and stderr lines return a non-empty string, we feed this string to stdin of the child as soon as possible. To get more control over the child, we treat NULL return value from a callback as a sign to close the standard input. We will also make our wrapper get a string as an input and feed it at the very beginning, as it is what could trigger the outputting of the data in the first place.

This still sounds not powerful enough, but you still have aplenty of asynchronous options, such as interrupting pselect with a specific signal, or setting up a separate thread for feeding data into the input pipe. We will omit these options in this blog post.

Watching for the child termination

As we're implementing a synchronous open3 wrapper, the natural assumption would be that at return from the wrapper the child should be dead and reaped. This way we can also return its exit status to the parent.

As we discussed in the previous post, process termination does not depend on the status of the pipes, so we should watch for it independently.

What we actually want is a version of select that watches for filehandlers and for the child's termination. If you know such a version of select (which I don't), make a comment, please. For now, let's search for another solution.

Such functionality could be simulated with a special wrapper thread that only waits for the child (with wait), and sends signal at its termination. This signal would interrupt select, and we would handle that.

In our project I implemented a simpler solution. It is based on the assumption that the more time the child is running, the less it's likely to terminate during the next second. So we can use timely wakeups to check for a process status (implemented as a non-null timeout to select), and increase the wait period with the course of time. Having the upper boundary for that period is a good idea as well.

Note that if a process has terminated, and the pipes are still open, we assume that the last nonblocking read will fetch us all the data we're interested in, and we may close the pipes. This may not be true, but in such specific cases you'll need specific solutions anyway.

Linewise-open3 code

Let's sum up what we're up to. Here's a listing of an open3 wrapper that prints the string supplied into stdin of the child, and then invokes one of two user-specified callbacks when a complete, \n-terminated line is read from stdin or stdout respectively. These callbacks may return more data to put into the stdin of the process. The execution terminates when the child is terminated. The wrapper returns the return code of the process (everything else is assumed to be done by callbacks).

I tested this program on random echo. You may also view the complete listing for open3 usage, and test it either with echo or with sshfs installed on your system.

Note also, that in our project we have also developed an open3 wrapper for Perl; you can view it here. It is less capable (without interactiveness), but it's live and working.

Notes on asynchronous open3 usage

In some cases you may trade the complexity for efficiency. The problem with the dead lock above was that we had a single thread designated to read from several pipes. Instead of introducing a buffering system, we may just spawn several threads, and attach each of them to its own endpoint. The burden of waking up the thread that actually has something to process doesn't vanish, it just gets imposed on the OS scheduler, which should contain fewer bugs. Or, to a language runtime scheduler (if it employs green threads, as in Ruby).

This may sound like an easier solution, especially in the multicore era, where developers cheer upon every possibility to develop a multithreaded program that makes the high-end hardware stall less. To me, it's a bit of an overkill for simple tasks (if the can be expressed in the terms of the interface discussed above). And in the context we used open3 in our project, too many competing threads had already started to become a problem.

Perhaps, I'll study the asynchronous option in other posts.


This post concludes what I initially planned for the series of blog posts of how to work with pipes in Linux. Another topic emerged, on the asynchronous interaction with an open3-driven process, but I don't know if I will write about it.

We have observed what are pipes, how they are used for inter-process communication, what options we have for spawning processes with pipes attached to them, and how it may be achieved in modern Linux scripting languages. However, we mainly focused on open3, studying details of its implementation, and the use-cases it's the most efficient in. We studied its quirks and how to avoid the traps set up among the joint of pipes we have to deal with.

I have learned this all spending a couple of days messing with processes that magically deadlocked without any obvious reasons, and with nontrivial multithreaded debugging. I hope these posts will help you when you will be working with pipes, so that you'll avoid my mistakes.

Comments imported from the old website

Paul A. Jungwirth on 24 October 2011 commented:

Thank you for writing up such a detailed description of the problem and your solution! Another approach is to use Ruby's non-blocking IO#ready? method, which I describe on my blog.

It's amazing to me that none of these languages can give a simple interface to safely handle common use-cases. If you can do it in the Bourne shell, why not Perl, Python, and Ruby? For Ruby, I wish you could just pass two procs, one to feed the filter and one to consume it, and their reading/writing would just work. Alas!

Brilliant article, and excellent blog in general. Cheers man!

I used an adaptation of this to solve my problems with hangs I was getting when capturing the output of doxygen.exe, which evidently interleaves stdout and stderr without newline termination. I did find that the readpartial/new line splitting code didn't seem to work properly though and failed a couple of my unit tests. I changed it to eg:

stillOpen = [stdout, stderr]
while !stillOpen.empty?
  fhs = select(stillOpen, nil, nil, nil)
  handleIO(stillOpen, fhs[0], stdout, @outputBuffer)
  handleIO(stillOpen, fhs[0], stderr, @stderrBuffer)

def handleIO(stillOpen, ioArray, io, buffer)
  if ioArray.include?(io)
      buffer << io.readpartial(4096)
      postNewLine = buffer =~ /([^\n]+)\z/ ? $1 : nil

      buffer.scan(/.*\n/) do |line|
      if postNewLine
    rescue EOFError
      stillOpen.delete_if{|s| s == io}

Which enabled my unit tests to pass, and I think is more efficient too.

I tried Paul A. Jungwirth's suggestion but io/wait appeared broken on windows. In any case I couldn't get it to work.

Damn, my attempt at making the code look nice didn't work :(

Pavel Shved on 22 November 2012 commented:

Thank you, James. It's always nice to know that somebody likes my blog! :-)

I fixed your comment to make code look nicer; you should use Markdown to format it; i.e. the code is 4 spaces at the beginning of the line.