Illustrated Redirection Tutorial
This tutorial is not a complete guide to redirection, it will not
cover here docs, here strings, name pipes etc… I just hope it'll help
you to understand what things like 3>&2, 2>&1 or 1>&3- do.
stdin, stdout, stderr
When Bash starts, normally, 3 file descriptors are opened, 0, 1 and 2
also known as standard input (stdin), standard output (stdout) and
standard error (stderr).
For example, with Bash running in a Linux terminal emulator, you'll see:
# lsof +f g -ap $BASHPID -d 0,1,2 COMMAND PID USER FD TYPE FILE-FLAG DEVICE SIZE/OFF NODE NAME bash 12135 root 0u CHR RW,LG 136,13 0t0 16 /dev/pts/5 bash 12135 root 1u CHR RW,LG 136,13 0t0 16 /dev/pts/5 bash 12135 root 2u CHR RW,LG 136,13 0t0 16 /dev/pts/5
This /dev/pts/5 is a pseudo terminal used to emulate a real
terminal. Bash reads (stdin) from this terminal and prints
via stdout and stderr to this terminal.
--- +-----------------------+
standard input ( 0 ) ---->| /dev/pts/5 |
--- +-----------------------+
--- +-----------------------+
standard output ( 1 ) ---->| /dev/pts/5 |
--- +-----------------------+
--- +-----------------------+
standard error ( 2 ) ---->| /dev/pts/5 |
--- +-----------------------+
When a command, a compound command, a subshell etc. is executed, it inherits
these file descriptors. For instance echo foo will send the text foo to the file
descriptor 1 inherited from the shell, which is connected to /dev/pts/5.
Simple Redirections
Output Redirection "n> file"
> is probably the simplest redirection.
echo foo > file
the > file after the command alters the file descriptors belonging to the
command echo. It changes the file descriptor 1 (> file is the same as
1>file) so that it points to the file file. They will look like:
--- +-----------------------+
standard input ( 0 ) ---->| /dev/pts/5 |
--- +-----------------------+
--- +-----------------------+
standard output ( 1 ) ---->| file |
--- +-----------------------+
--- +-----------------------+
standard error ( 2 ) ---->| /dev/pts/5 |
--- +-----------------------+
Now characters written by our command, echo, that are sent to the
standard output, i.e., the file descriptor 1, end up in the file
named file.
In the same way, command 2> file will change the standard error and
will make it point to file. Standard error is used by applications to print errors.
What will command 3> file do? It will open a new file descriptor
pointing to file. The command will then start with:
--- +-----------------------+
standard input ( 0 ) ---->| /dev/pts/5 |
--- +-----------------------+
--- +-----------------------+
standard output ( 1 ) ---->| /dev/pts/5 |
--- +-----------------------+
--- +-----------------------+
standard error ( 2 ) ---->| /dev/pts/5 |
--- +-----------------------+
--- +-----------------------+
new descriptor ( 3 ) ---->| file |
--- +-----------------------+
What will the command do with this descriptor? It depends. Often nothing. We will see later why we might want other file descriptors.
Input Redirection "n< file"
When you run a commandusing command < file, it changes the
file descriptor 0 so that it looks like:
--- +-----------------------+
standard input ( 0 ) <----| file |
--- +-----------------------+
--- +-----------------------+
standard output ( 1 ) ---->| /dev/pts/5 |
--- +-----------------------+
--- +-----------------------+
standard error ( 2 ) ---->| /dev/pts/5 |
--- +-----------------------+
If the command reads from stdin, it now will read from file and not
from the console.
As with >, < can be used to open a new file descriptor for reading,
command 3<file. Later we will see how this can be useful.
Pipes |
What does this | do? Among other things, it connects the standard output of
the command on the left to the standard input of the command on the right.
That is, it creates a special file, a pipe, which is opened as a write destinaton
for the left command, and as a read source for the right command.
echo foo | cat
--- +--------------+ --- +--------------+
( 0 ) ---->| /dev/pts/5 | ------> ( 0 ) ---->|pipe (read) |
--- +--------------+ / --- +--------------+
/
--- +--------------+ / --- +--------------+
( 1 ) ---->| pipe (write) | / ( 1 ) ---->| /dev/pts |
--- +--------------+ --- +--------------+
--- +--------------+ --- +--------------+
( 2 ) ---->| /dev/pts/5 | ( 2 ) ---->| /dev/pts/ |
--- +--------------+ --- +--------------+
This is possible because the redirections are set up by the shell before the commands are executed, and the commands inherit the file descriptors.
More On File Descriptors
Duplicating File Descriptor 2>&1
We have seen how to open (or redirect) file descriptors. Let us see
how to duplicate them, starting with the classic 2>&1. What does this
mean? That something written on the file descriptor 2 will go where
file descriptor 1 goes. In a shell command 2>&1 is not a very
interesting example so we will use ls /tmp/ doesnotexist 2>&1 | less
ls /tmp/ doesnotexist 2>&1 | less
--- +--------------+ --- +--------------+
( 0 ) ---->| /dev/pts/5 | ------> ( 0 ) ---->|from the pipe |
--- +--------------+ / ---> --- +--------------+
/ /
--- +--------------+ / / --- +--------------+
( 1 ) ---->| to the pipe | / / ( 1 ) ---->| /dev/pts |
--- +--------------+ / --- +--------------+
/
--- +--------------+ / --- +--------------+
( 2 ) ---->| to the pipe | / ( 2 ) ---->| /dev/pts/ |
--- +--------------+ --- +--------------+
Why is it called duplicating? Because after 2>&1, we have 2 file
descriptors pointing to the same file. Take care not to call this
"File Descriptor Aliasing"; if we redirect stdout after 2>&1 to a
file B, file descriptor 2 will still be opened on the file A where it was. This is
often misunderstood by people wanting to redirect both standard input
and standard output to the file. Continue reading for more on this.
So if you have two file descriptors s and t like:
--- +-----------------------+
a descriptor ( s ) ---->| /some/file |
--- +-----------------------+
--- +-----------------------+
a descriptor ( t ) ---->| /another/file |
--- +-----------------------+
Using a t>&s (where t and s are numbers) it means:
Copy whatever file descriptorscontains into file descriptort
So you got a copy of this descriptor:
--- +-----------------------+
a descriptor ( s ) ---->| /some/file |
--- +-----------------------+
--- +-----------------------+
a descriptor ( t ) ---->| /some/file |
--- +-----------------------+
Internally each of these is represented by a file descriptor opened by the operating system's fopen calls, and is likely just a pointer to the file which has been opened for reading (stdin or file descriptor 0) or writing (stdout /stderr).
Note that the file reading or writing positions are also duplicated. If you have already
read a line of s, then after t>&s if you read a line from t, you will
get the second line of the file.
Similarly for output file descriptors, writing a line to file descriptor s will append a line to a file as will writing a line to file descriptor t.
target>&source effectively.
So, as a simple example (albeit slightly contrived), is the following:
exec 3>&1 # Copy 1 into 3 exec 1> logfile # Make 1 opened to write to logfile lotsa_stdout # Outputs to fd 1, which writes to logfile exec 1>&3 # Copy 3 back into 1 echo Done # Output to original stdout
Order Of Redirection, i.e., "> file 2>&1" vs. "2>&1 >file"
While it doesn't matter where the redirections appears on the command line, their order does matter. They are set up from left to right.
2>&1 >file
A common error, is to do command 2>&1 > file to redirect both stderr
and stdout to file. Let's see what's going on. First we type the
command in our terminal, the descriptors look like this:
--- +-----------------------+
standard input ( 0 ) ---->| /dev/pts/5 |
--- +-----------------------+
--- +-----------------------+
standard output ( 1 ) ---->| /dev/pts/5 |
--- +-----------------------+
--- +-----------------------+
standard error ( 2 ) ---->| /dev/pts/5 |
--- +-----------------------+
Then our shell, Bash sees 2>&1 so it duplicates 1, and the file
descriptor look like this:
--- +-----------------------+
standard input ( 0 ) ---->| /dev/pts/5 |
--- +-----------------------+
--- +-----------------------+
standard output ( 1 ) ---->| /dev/pts/5 |
--- +-----------------------+
--- +-----------------------+
standard error ( 2 ) ---->| /dev/pts/5 |
--- +-----------------------+
That's right, nothing has changed, 2 was already pointing to the same
place as 1. Now Bash sees > file and thus changes stdout:
--- +-----------------------+
standard input ( 0 ) ---->| /dev/pts/5 |
--- +-----------------------+
--- +-----------------------+
standard output ( 1 ) ---->| file |
--- +-----------------------+
--- +-----------------------+
standard error ( 2 ) ---->| /dev/pts/5 |
--- +-----------------------+
And that's not what we want.
>file 2>&1
Now let's look at the correct command >file 2>&1. We start as in the
previous example, and Bash sees > file:
--- +-----------------------+
standard input ( 0 ) ---->| /dev/pts/5 |
--- +-----------------------+
--- +-----------------------+
standard output ( 1 ) ---->| file |
--- +-----------------------+
--- +-----------------------+
standard error ( 2 ) ---->| /dev/pts/5 |
--- +-----------------------+
Then it sees our duplication 2>&1:
--- +-----------------------+
standard input ( 0 ) ---->| /dev/pts/5 |
--- +-----------------------+
--- +-----------------------+
standard output ( 1 ) ---->| file |
--- +-----------------------+
--- +-----------------------+
standard error ( 2 ) ---->| file |
--- +-----------------------+
And voila, both 1 and 2 are redirected to file.
Why sed 's/foo/bar/' file >file Doesn't Work
This is a common error, we want to modify a file using something that
reads from a file and writes the result to stdout. To do this, we redirect stdout
to the file we want to modify. The problem here is that, as we
have seen, the redirections are setup before the command is actually
executed.
So BEFORE sed starts, standard output has already been redirected, with
the additional side effect that, because we used >, "file" gets truncated. When sed starts to
read the file, it contains nothing.
exec
In Bash the exec built-in replaces the shell with the specified
program. So what does this have to do with redirection?
exec also allow us to manipulate the file descriptors.
If you don't specify a program, the redirection after exec modifies the
file descriptors of the current shell.
For example, all the commands after exec 2>file will have file
descriptors like:
--- +-----------------------+
standard input ( 0 ) ---->| /dev/pts/5 |
--- +-----------------------+
--- +-----------------------+
standard output ( 1 ) ---->| /dev/pts/5 |
--- +-----------------------+
--- +-----------------------+
standard error ( 2 ) ---->| file |
--- +-----------------------+
All the the errors sent to stderr by the commands after the exec 2>file
will go to the file, just as if you had the command in a script and
ran myscript 2>file.
exec can be used, if, for instance, you want to log the errors the
commands in your script produce, just add exec 2>myscript.errors at
the beginning of your script.
Let's see another use case. We want to read a file line by line, this is easy, we just do:
while read -r line;do echo "$line";done < file
Now, we want, after printing each line, to do a pause, waiting for the user to press a key:
while read -r line;do echo "$line"; read -p "Press any key" -n 1;done < file
And, surprise this doesn't work. Why? because the shell descriptor of the while loop looks like:
--- +-----------------------+
standard input ( 0 ) ---->| file |
--- +-----------------------+
--- +-----------------------+
standard output ( 1 ) ---->| /dev/pts/5 |
--- +-----------------------+
--- +-----------------------+
standard error ( 2 ) ---->| /dev/pts/5 |
--- +-----------------------+
and our read inherits these descriptors, and our command (read -p "Press any key" -n 1)
inherits them, and thus reads from file and not from our terminal.
A quick look at help read tells us that we can specify a file
descriptor from which read should read. Cool. Now let's use exec to
get another descriptor:
exec 3<file while read -u 3 line;do echo "$line"; read -p "Press any key" -n 1;done
Now the file descriptors look like:
--- +-----------------------+
standard input ( 0 ) ---->| /dev/pts/5 |
--- +-----------------------+
--- +-----------------------+
standard output ( 1 ) ---->| /dev/pts/5 |
--- +-----------------------+
--- +-----------------------+
standard error ( 2 ) ---->| /dev/pts/5 |
--- +-----------------------+
--- +-----------------------+
new descriptor ( 3 ) ---->| file |
--- +-----------------------+
and it works.
Closing The File Descriptors
Closing a file through a file descriptor is easy, just make it a duplicate of -.
For instance, let's close stdin <&- and stderr 2>&-:
bash -c '{ lsof -a -p $$ -d0,1,2 ;} <&- 2>&-'
COMMAND PID USER FD TYPE DEVICE SIZE NODE NAME
bash 10668 pgas 1u CHR 136,2 4 /dev/pts/2
we see that inside the {} that only 1 is still here.
Though the OS will probably clean up the mess, it is perhaps a
good idea to close the file descriptors you open. For instance, if you
open a file descriptor with exec 3>file, all the commands afterwards
will inherit it. It's probably better to do something like:
exec 3>file ..... #commands that uses 3 ..... exec 3>&- #we don't need 3 any more
I've seen some people using this as a way to discard, say stderr, using something like: command 2>&-. Though it might work, I'm not sure if you can expect all applications to behave correctly with a closed stderr.
When in doubt, I use 2>/dev/null.
An Example
This example comes from this post (ffe4c2e382034ed9) on the comp.unix.shell group:
{
{
cmd1 3>&- |
cmd2 2>&3 3>&-
} 2>&1 >&4 4>&- |
cmd3 3>&- 4>&-
} 3>&2 4>&1
The redirections are processed from left to right, but as the file
descriptors are inherited we will also have to work from the outer to
the inner contexts. We will assume that we run this command in a
terminal. Let's start with the outer { } 3>&2 4>&1.
--- +-------------+ --- +-------------+ ( 0 ) ---->| /dev/pts/5 | ( 3 ) ---->| /dev/pts/5 | --- +-------------+ --- +-------------+ --- +-------------+ --- +-------------+ ( 1 ) ---->| /dev/pts/5 | ( 4 ) ---->| /dev/pts/5 | --- +-------------+ --- +-------------+ --- +-------------+ ( 2 ) ---->| /dev/pts/5 | --- +-------------+
We only made 2 copies of stderr and stdout. 3>&1 4>&1 would have
produced the same result here because we ran the command in a terminal
and thus 1 and 2 go to the terminal. As an exercise, you can start
with 1 pointing to file.stdout and 2 pointing to file.stderr, you will
see why these redirections are very nice.
Let's continue with the right part of the second pipe: | cmd3 3>&- 4>&-
--- +-------------+ ( 0 ) ---->| 2nd pipe | --- +-------------+ --- +-------------+ ( 1 ) ---->| /dev/pts/5 | --- +-------------+ --- +-------------+ ( 2 ) ---->| /dev/pts/5 | --- +-------------+
It inherits the previous file descriptors, closes 3 and 4 and sets up a pipe
for reading. Now for the left part of the second pipe {…} 2>&1 >&4 4>&- |
--- +-------------+ --- +-------------+ ( 0 ) ---->| /dev/pts/5 | ( 3 ) ---->| /dev/pts/5 | --- +-------------+ --- +-------------+ --- +-------------+ ( 1 ) ---->| /dev/pts/5 | --- +-------------+ --- +-------------+ ( 2 ) ---->| 2nd pipe | --- +-------------+
First, The file descriptor 1 is connected to the pipe (|), then 2 is
made a copy of 1 and thus is made an fd to the pipe (2>&1), then 1 is
made a copy of 4 (>&4), then 4 is closed. These are the file
descriptors of the inner {}. Lcet's go inside and have a look at the
right part of the first pipe: | cmd2 2>&3 3>&-
--- +-------------+ ( 0 ) ---->| 1st pipe | --- +-------------+ --- +-------------+ ( 1 ) ---->| /dev/pts/5 | --- +-------------+ --- +-------------+ ( 2 ) ---->| /dev/pts/5 | --- +-------------+
It inherits the previous file descriptors, connects 0 to the 1st pipe, the file descriptor 2 is made a copy of 3, and 3 is closed. Finally, for the left part of the pipe:
--- +-------------+ ( 0 ) ---->| /dev/pts/5 | --- +-------------+ --- +-------------+ ( 1 ) ---->| 1st pipe | --- +-------------+ --- +-------------+ ( 2 ) ---->| 2nd pipe | --- +-------------+
It also inherits the file descriptor of the left part of the 2nd
pipe, file descriptor 1 is connected to the first pipe, 3 is closed.
The purpose of all this becomes clear if we take only the commands:
cmd2
--- +-------------+
-->( 0 ) ---->| 1st pipe |
/ --- +-------------+
/
/ --- +-------------+
cmd 1 / ( 1 ) ---->| /dev/pts/5 |
/ --- +-------------+
/
--- +-------------+ / --- +-------------+
( 0 ) ---->| /dev/pts/5 | / ( 2 ) ---->| /dev/pts/5 |
--- +-------------+ / --- +-------------+
/
--- +-------------+ / cmd3
( 1 ) ---->| 1st pipe | /
--- +-------------+ --- +-------------+
------------>( 0 ) ---->| 2nd pipe |
--- +-------------+ / --- +-------------+
( 2 ) ---->| 2nd pipe |/
--- +-------------+ --- +-------------+
( 1 ) ---->| /dev/pts/5 |
--- +-------------+
--- +-------------+
( 2 ) ---->| /dev/pts/5 |
--- +-------------+
As said previously, as an exercise, you can start with 1 open on a file
and 2 open on another file to see how the stdin from cmd2 and cmd3 goes
to the original stdin and how the stderr goes to the original stderr.
Syntax
I used to have trouble choosing between 0&<3 3&>1 3>&1 ->2 -<&0 &-<0
0<&- etc… (I think probably because the syntax is more representative of the result, i.e., the redirection, than
what is done, i.e., opening, closing, or duplicating file descriptors).
If this fits your situation, then maybe the following "rules" will help you, a redirection is always like the following:
lhs op rhs
lhsis always a file description, i.e., a number:- Either we want to open, duplicate, move or we want to close. If the op is
<then there is an implicit 0, if it's>or>>, there is an implicit 1.
opis<,>,>>,>|, or<>:<if the file decriptor inlhswill be read,>if it will be written,>>if data is to be appended to the file,>|to overwrite an existing file or<>if it will be both read and written.
rhsis the thing that the file descriptor will describe:- It can be the name of a file, the place where another descriptor goes (
&1), or,&-, which will close the file descriptor.
You might not like this description, and find it a bit incomplete or
inexact, but I think it really helps to easily find that, say &->0 is
incorrect.
A note on style
The shell is pretty loose about what it considers a valid redirect. While opinions probably differ, this author has some (strong) recommendations:
- Always keep redirections "tightly grouped" – that is, do not include whitespace anywhere within the redirection syntax except within quotes if required on the RHS (e.g. a filename that contains a space). Since shells fundamentally use whitespace to delimit fields in general, it is visually much clearer for each redirection to be separated by whitespace, but grouped in chunks that contain no unnecessary whitespace.
- Do always put a space between each redirection, and between the argument list and the first redirect.
- Always place redirections together at the very end of a command after all arguments. Never precede a command with a redirect. Never put a redirect in the middle of the arguments.
# Good! This is clearly a simple commmand with two arguments and 4 redirections
cmd arg1 arg2 <myFile 3<&1 2>/dev/null >&2
# Good!
{ cmd1 <<<'my input'; cmd2; } >someFile
# Bad. Is the "1" a file descriptor or an argument to cmd? (answer: it's the FD). Is the space after the herestring part of the input data? (answer: No).
# The redirects are also not delimited in any obvious way.
cmd 2>& 1 <<< stuff
# Hideously Bad. It's difficult to tell where the redirects are and whether they're even valid redirects.
# This is in fact one command with one argument, an assignment, and three redirects.
foo=bar<baz bork<<< blarg>bleh
Conclusion
I hope this tutorial worked for you.
I lied, I did not explain 1>&3-, go check the manual 
Thanks to Stéphane Chazelas from whom I stole both the intro and the example….
The intro is inspired by this introduction, you'll find a nice exercise there too:
The last example comes from this post:
See also
- Internal: Redirection syntax overview
Discussion
Hello,
thank you for that comprehensive explanation.
I was looking for a solution for the following problem: I want to execute a shell script (both remotely via RSH and locally). The output from stdout and stderr should go to a file, to see the scripts progress at the terminal I wanted to redirect the output of some echo commands to the same file and to the terminal. Based on this tutorial I implemented the following solution
(I don't know how to produce an ampersand, therefore I use "amp;" instead):As soon as output to stderr happens it doesn't work as I expected. E.g.
results in the following content of file logfile:
I.e. the texts "my message" and "Hello again" have been overwritten by the stderr output of the ls commands.
If I change in the 1st exec to append stdout to logfile (
exec 3>&1 1>>logfile 2>&1) the result is correct:If I change the 7th line to
exec 1>>logfile 2>&1the result looks better, but is not correct (still the line with "my message" is missing):I finally implemented the following solution, which is maybe not so elegant but works:
This has the effect that the error output of the 2nd ls goes to the terminal and not to the file which is absolutely ok for my case.
Can anybody explain what exactly happens? I assume it has something to with file pointers. What is the preferred solution of my problem?
Regards
Armin
P.S.: I have some problems with formatting, esp. with line feeds and empty lines.
Try this. In short:
teegets a process substitution as output file, inside acatand a redirection to FD1 (logfile)tees standard output is redirected to FD3 (terminal/original stdout)This is a Bash specific thing.
For the wiki quirks: I surrounded your code with
<code>...</code>tags. For the ampersand issue I have no solution, sorry. Seems to be a bug in this plugin. It only happens on "preview", but it works for the real view.I am looking to implement the features of Expect, with bash. That is, to design a wrapper program that will assign the called program to redirect its 0-2 to named pipes. The wrapper will then open the other end of the named pipes. Something like this:
exec 3<>pipe.out exec 4<>pipe.in ( PS3="enter choice:"; select choice in one two three; do echo "you choose \"$choice\""; done )0<&4 1>&3 2>&1 while read -u pipe.out line do case $line in *choice:) echo "$line" read i; echo i >&4 ;; EOF) echo "caught \"EOF\", exiting"; break;; *) echo "$line" ;; esac doneOf course, this doesn't work at all. The first problem is, when using a pipe, the process hangs until both ends of the pipe are established. The second part of the problem is that the bash built-in "read" returns on a newline or the option of N chars or delimiter X –neither of which would be useful in this case. So the input of the while loop never "sees" the "enter choice:" prompt, since there is no newline. There are other problems as well.
Is it possible to get Bash to do this? Any suggestions? Here is something that does work.
terminal 1:
terminal 2:
In this case, the program continues when both ends connect to the pipe, but since we arent redirecting stdin from a pipe for the select, you have to enter the choice in terminal 2. You will also notice that even in this scenario, terminal 1 does not see the PS3 prompt since it does not return a newline.Hello,
Many thanks for the comprehensive tutorial. I have learned a great deal about redirection.
In my script, I want to redirect stderr to a file and both stderr and stdout to another file. I found this construction works but I don't quite understand how. Could you explain ?
Also, if I want to do the same in the script using exec to avoid this kind of redirection in every command in the script, what should I do ?
thanks
Tony
You pump
STDERRof the command to descriptor 1, so that it can be transported by the pipe and seen as input by theteecommand. At the same time you redirect the originalSTDOUTto descriptor 3. Theteecommand writes your original standard error output to the file plus outputs it to itsSTDOUT.Outside the whole construct you collect your original standard output (descriptor 3) and your original standard error output (descriptor 1 - through
tee) to the normal descriptors (1 and 2), the rest is a simple file redirection for both descriptors.In short, you use a third descriptor to switch a bypass throuch
tee.I don't know a global method (
execor thelike) off my head. I can imagine that you can hack something with process substitution, but I'm not sure.Many thanks for these explanations! Just one point which confused me. In the example from comp.unix.shell, you wrote: "Now for the left part of the second pipe…" The illustration for the result confused me because I was assuming the fds where coming from the previous illustration, but I then understood that they come from their parent process and hence from the previous previous illustration. I think it would be a little bit clearer if you would put a label on each of your illustrations and make more explicit the transition from one illustration to another. Anyway, many thanks again. )jack(
Pipes seem to introduce an extraneous line at EOF. Is this true? Try this:
When three lines go into the pipe, four lines come out. The problem is not present in the here-string facility.
What a helpful and badly needed site! Thank you!
I see those additional line coming from the previous echo:
It is the additional newline echo adds itself to finalize the output. In your first echo, this is the newline after the closing bracket.
You can verify it when you use
echo -n(suppresses the newline echo itself generates)Thank you for comprehensive manual.
Plase add this example, http://stackoverflow.com/questions/3141738/duplicating-stdout-to-stderr.
echo foo |tee /dev/stderr
Are there better/cleaner solutions? It seems that /dev/stderr can have problem in cron.
It's a functionality of the shell itself, the shell duplicates the relevant file descriptors when it sees those filenames. So it may depend on the shell (or shell compatibility level) you use in cron.