xesite/blog/fun-with-redirection-2021-0...

12 KiB

title date author tags
Fun with Redirection 2021-09-22 Twi
shell
redirection
osdev

When you're hacking in the shell or in a script, sometimes you want to change how the output of a command is routed. Today I'm gonna cover common shell redirection tips and tricks that I use every day at work and how it all works under the hood.

Let's say you're trying to capture the output of a command to a file, such as uname -av:

$ uname -av
Linux shachi 5.13.15 #1-NixOS SMP Wed Sep 8 06:50:21 UTC 2021 x86_64 GNU/Linux

You could copy that to the clipboard and paste it into a file, but there is a better way thanks to the > operator:

$ uname -av > uname.txt
$ cat uname.txt
Linux shachi 5.13.15 #1-NixOS SMP Wed Sep 8 06:50:21 UTC 2021 x86_64 GNU/Linux

Let's say you want to run this on a few machines and put all of the output into uname.txt. You could write a shell script loop like this:

# make sure the file doesn't already exist
rm -f uname.txt

for host in shachi chrysalis kos-mos ontos pneuma
do
  ssh $host -- uname -av >> uname.txt
done

Then uname.txt should look like this:

Linux shachi 5.13.15 #1-NixOS SMP Wed Sep 8 06:50:21 UTC 2021 x86_64 GNU/Linux
Linux chrysalis 5.10.63 #1-NixOS SMP Wed Sep 8 06:49:02 UTC 2021 x86_64 GNU/Linux
Linux kos-mos 5.10.45 #1-NixOS SMP Fri Jun 18 08:00:06 UTC 2021 x86_64 GNU/Linux
Linux ontos 5.10.52 #1-NixOS SMP Tue Jul 20 14:05:59 UTC 2021 x86_64 GNU/Linux
Linux pneuma 5.10.57 #1-NixOS SMP Sun Aug 8 07:05:24 UTC 2021 x86_64 GNU/Linux

Now let's say you want to extract all of the hostnames from that uname.txt. The pattern of the file seems to specify that fields are separated by spaces and the hostname seems to be the second space-separated field in each line. You can use the cut command to select that small subset from each line, and you can feed the cut command's standard input using the < operator:

$ cut -d' ' -f2 < uname.txt
shachi
chrysalis
kos-mos
ontos
pneuma

It's worth noting that a lot of these core CLI utilities are built on the idea that they are filters, or things that take one infinite stream of text in on one end and then return another stream of text out the other end. This is done through a channel called "standard input/output", where standard input refers to input to the command and standard output refers to the output of the command.

That's a great metaphor, let's build onto it using the | (pipe) operator. The pipe operator lets you pipe the standard output of one command to the standard input of another.

You mentioned that you can pass files as input and output for commands, does this mean that standard input and standard output are files?

Precisely! They are just files that are automatically open for every process. Usually commands will output to standard out and some will also accept input via standard in.

Doesn't that have some level of overhead though? Isn't it expensive to spin up a whole heckin' cat process for that?

Not on any decent system made in the last 20 years. This may have some impact on Windows (because they have core architectural mistakes that make processes take up to 100 milliseconds to spin up), but this is about Unix/Linux. I think these should work on Windows too if you use Cygwin, but if you're using WSL you shouldn't have any real issues there

Let's say we want to rewrite that cut command above to use pipes. You could write it like this:

cat uname.txt | cut -d' ' -f2

The mnemonic we use for remembering the cut command is that fields are separated by the delimiter and you cut out the nth field/s. You can use

This will get you the exact same output:

$ cat uname.txt | cut -d' ' -f2
shachi
chrysalis
kos-mos
ontos
pneuma

Personally I prefer writing shell pipelines like that as it makes it a bit easier to tack on more specific selectors or operations as you go along. For example, if you wanted to sort them you could pipe the result to sort:

$ cat uname.txt | cut -d' ' -f2 | sort
chrysalis
kos-mos
ontos
pneuma
shachi

This lets you gradually build up a shell pipeline as you drill down to the data you want in the format you want.

I wanted to save this compiler error to a file but it didn't work. I tried doing this:

$ rustc foo.rs > foo.log

But the output printed to the screen instead of the file:

$ rustc foo.rs > foo.log
error: expected one of `!` or `::`, found `main`
 --> foo.rs:1:5
  |
1 | fun main() {}
  |     ^^^^ expected one of `!` or `::`

error: aborting due to previous error

$ cat foo.log
$

This happens because there are actually two output streams per program. There is the standard out stream and there is also a standard error stream. The reason that standard error exists is so that you can see if any errors have happened if you redirect standard out.

Sometimes standard out may not be a stream of text, say you have a compressed file you want to analyze and there's an issue with the decompression. If the decompressor wrote its errors to the standard output stream, it could confuse or corrupt your analysis.

However, we can redirect standard error in particular by modifying how we redirect to the file:

$ rustc foo.rs 2> foo.log
$ cat foo.log
error: expected one of `!` or `::`, found `main`
 --> foo.rs:1:5
  |
1 | fun main() {}
  |     ^^^^ expected one of `!` or `::`

error: aborting due to previous error

Where did the 2 come from?

So I mentioned earlier that redirection modifies the standard input and output of programs. This is not entirely true, but it was a convenient half-truth to help build this part of the explanation.

For every process on a Unix-like system (such as Linux and macOS), the kernel stores a list of active file-like objects. This includes real files on the filesystem, pipes between processes, network sockets, and more. When a program reads or writes a file, they tell the kernel which file they want to use by giving it a number index into that list, starting at zero. Standard in/out/error are just the conventional names for the first three open files in the list, like this:

File Descriptor Purpose
0 Standard input
1 Standard output
2 Standard error

Shell redirection simply changes which files are in that list of open files when the program starts running.

That is why you use a 2 there, because you are telling the shell to change file descriptor number 2 of the rustc process to point to the filesystem file foo.log, which in turn makes the standard error of rustc get written to that file for you.

In turn, this also means that cat foo.txt > foo2.txt is actually a shortcut for saying cat foo.txt 1> foo2.txt, but the 1 can be omitted there because standard out is usually the "default" output that most of these kind of pipelines cares about.

How would I get both standard output and standard error in the same file?

The cool part about the > operator is that it doesn't just stop with output to files on the desk, you can actually have one file descriptor get pointed to another. Let's say you have a need for both standard out and standard error to go to the same file. You can do this with a command like this:

$ rustc foo.rs 2>&1 > foo.log

This tells the shell to point standard error to standard out and then the combined output to foo.log. There's a short form of this too:

$ rustc foo.rs &> foo.log

Where can I expect to use that?

It's a bourne shell extension, but I've tested it in zsh and fish. You can also do &| to pipe both standard out and standard error at the same time in the same way you'd do 2>&1 | whatever.

That will put standard out and standard error to foo.log the same way that 2>&1 > foo.log will. You can also use this with >>:

$ rustc foo.rs &>> foo.log
$ cat foo.log
error: expected one of `!` or `::`, found `main`
 --> foo.rs:1:5 
  | 
1 | fun main() {}
  |     ^^^^ expected one of `!` or `::

error: aborting due to previous error

error: expected one of `!` or `::`, found `main`
 --> foo.rs:1:5
  |
1 | fun main() {}
  |     ^^^^ expected one of `!` or `::`

error: aborting due to previous error

How do I redirect standard in to a file?

The answer there is not directly! There is a workaround in the form of a tool called tee which outputs its standard in to both standard out and a file. For example:

$ dmesg | tee dmesg.txt | grep 'msedge'
[   70.585463] traps: msedge[4715] trap invalid opcode ip:5630ddcedc4c sp:7ffd41f67700 error:0 in msedge[5630d8fc2000+952d000]
[   70.702544] traps: msedge[4745] trap invalid opcode ip:5630ddcedc4c sp:7ffd41f67700 error:0 in msedge[5630d8fc2000+952d000]
[   70.806296] traps: msedge[4781] trap invalid opcode ip:5630ddcedc4c sp:7ffd41f67700 error:0 in msedge[5630d8fc2000+952d000]
[   70.918095] traps: msedge[4889] trap invalid opcode ip:5630ddcedc4c sp:7ffd41f67700 error:0 in msedge[5630d8fc2000+952d000]
[   71.031938] traps: msedge[4926] trap invalid opcode ip:5630ddcedc4c sp:7ffd41f67700 error:0 in msedge[5630d8fc2000+952d000]
[   71.138974] traps: msedge[4935] trap invalid opcode ip:5630ddcedc4c sp:7ffd41f67700 error:0 in msedge[5630d8fc2000+952d000]
[ 1169.163603] traps: msedge[35719] trap invalid opcode ip:556a93951c4c sp:7ffc533f35c0 error:0 in msedge[556a8ec26000+952d000]
[ 1213.301722] traps: msedge[36054] trap invalid opcode ip:55a245960c4c sp:7ffe6d169b40 error:0 in msedge[55a240c35000+952d000]
[10963.234459] traps: msedge[104732] trap invalid opcode ip:55fdb864fc4c sp:7ffc996dfee0 error:0 in msedge[55fdb3924000+952d000]

This would put the output of the dmesg command (read from kernel logs) into dmesg.txt, as well as sending it into the grep command. You might want to do this when debugging long command pipelines to see exactly what is going into a program that isn't doing what you expect.

Redirections also work in scripts too. You can also set "default" redirects for every command in a script using the exec command:

exec > out.log 2> error.log

ls
rustc foo.rs

This will have the file listing from ls written to out.log and any errors from rustc written to error.log.

A lot of other shell tricks and fun is built on top of these fundamentals. For example you can take a folder, zip it up and then unzip it over on another machine using a command like this:

$ tar cz ./blog | ssh pneuma tar xz -C ~/code/christine.website/blog

This will run tar to create a compressed copy of the ./blog folder and then pipe that to tar on another computer to extract that into ~/code/christine.website/blog. It's just pipes and redirection all the way down! Deep inside ssh it's really just piping output of commands back and forth over an encrypted network socket. Connecting to an IRC server is just piping in and out data to the chat server, even more so if you use TLS to connect there. In a way you can model just about everything in Unix with pipes and file descriptors because that is the cornerstone of its design: Everything is a file.

This doesn't mean it's literally a file on the disk, it means you can interact with just about everything using the same system interface as you do with files. Even things like hard disks and video cards.

Here's a fun thing to do. Using curl to read the contents of a URL and jq to select out bits from a JSON stream, you can make a script that lets you read the most recent title from my blog's JSONFeed:

#!/usr/bin/env bash
# xeblog-post.sh

curl -s https://christine.website/blog.json | jq -r '.items[0] | "\(.title) \(.url)"'

At the time of writing this post, here is the output I get from this command:

$ ./xeblog-post.sh
Anbernic RG280M Review https://christine.website/blog/rg280m-review

What else could you do with pipes and redirection? The cloud's the limit!


Thanks to violet spark for looking over this post and fact-checking as well as helping mend some of the brain dump and awkward wording into more polished sentences.