Omair Majid

Omair Majid

Announcements, thoughts, code and fun

10 Jun 2020

The magic of file

Ever wonder how file(1)1 works?

After reading this post, you should understand all about file and its underpinnings.

file

Using the file tool is quite simple. It’s just file file-name-1. You can call it with multiple arguments too. Here’s a few examples of running it on my system.

$ file document.pdf
document.pdf: PDF document, version 1.4
$ file image.jpg
image.jpg: JPEG image data, Exif standard: [TIFF image data, big-endian, direntries=7 ...orientation=upper-left, datetime=2019:10:30 21:44:30, width=1740], baseline, precision 8, 2017x961, components 3
$ file video.mp4
video.mp4: ISO Media, MP4 v2 [ISO 14496-14]
$ file notes.org
notes.org: UTF-8 Unicode text, with very long lines
$ file /dev/video0
/dev/video0: character special (81/0)

You can already see a few things:

  1. It works on any file.

  2. It tries to examine the contents of the file (not just the extensions) to figure out the file type.

  3. It understands file formats in some detail. It can extract extra information from the file itself. You can see that it understands versions for mp4 and jpeg files.

  4. It knows about special files like devices.

There’s a few gotchas, though.

  1. file may not do what you want in shell scripts.

    $ file no-such-file
    no-such-file: cannot open `no-such-file' (No such file or directory)
    $ echo $?
    0
    

    This version of file always seems to exit with a code of 0. Even when file can’t figure it out. Even if the file doesn’t exist.

  2. For many formats that are not common on Unix/Linux, file is particularly fragile:

    $ file Vagrantfile
    Vagrantfile: HTML document, ASCII text
    $ file microsoft.netcore.app.2.1.19.nupkg
    microsoft.netcore.app.2.1.19.nupkg: Microsoft OOXML
    

    Yes, it really thinks a Vagrantfile is an HTML document and a NuGet package file is an Office document!

But how does file actually figure out the file type?

Like other classic Unix tools, file has excellent documentation. man file points out that file looks for things in this order.

  1. It checks to see if a file is a special type of file, like devices.

  2. It checks if the magic(5) database contains a “magic pattern” that can identify this file type.

  3. It checks to see if it’s a text-file. In this case, file will try and find out additional properties, such as language, encoding, line-length and the line terminator type (CR, LF, or CRLF).

file uses the result from the first check that works. If none of these checks are successful, file simply reports the file type as ‘data’.

The most interesting one of these is the magic(5) database option. While the other checks are fairly static, you can examine the magic database see how file works. You can even modify the database manually to fix bugs2 in file!

magic

The magic database is a plain text file. It is often located at /usr/share/magic. You can view it in your favourite editor.

/usr/share/magic contains a series of test-and-output lines. A test describes something to look for. If the test matches, the output is printed.

Here’s what a line might look like:

0   string      Draw        RISC OS Draw file data

The first 3 columns are the test and the final column is the output.

This line specifies that at if the string at position 0 in the file matches the constant string Draw, then show RISC OS Draw file data as the file type.

To make things easier for the people maintaining the database, the database supports nesting tests. That is, a test can say that it only applies if the previous test matched.

Here’s an example of a test and a nested test:

0   belong      0xcafebabe
>4  belong      >30     compiled Java class data,

The first line is a test that can always be performed. The second line contains a test that should only be performed if the first test is successful, indicated by a “>” in the beginning of the line.

The first test here says: does the big-endian long at file position 0 match the constant 0xcafebabe. There’s no output column, so it doesn’t do anything directly if it’s successful.

However, if the test is successful, the nested test runs. The nested test in this example checks if the big-endian long at position 4 in the file is greater than 30. If that’s true, print compiled Java class data3.

man 5 magic contains extensive documentation that covers all the syntax and options available if you want to learn about this in more detail.

a real life bug

Now that you know have a good understanding of file and magic, lets see how you can use that to figure out a real life “bug”.

As part of releasing .NET Core this week, one of the tools in our release pipeline flagged an error saying the type of a particular file had changed. The error looked something like this:

 file type changed for Microsoft.NETCore.App.Ref/3.1.0/data/PlatformManifest.txt:
-ASCII text
+Message Sequence Chart (chart)

The error is telling us that the file type changed from ASCII text to Message Sequence Chart (chart).

That error looks bogus. It’s a text file. The extension is .txt.

The contents look like plain text too:

$ head Microsoft.NETCore.App.Ref/3.1.0/data/PlatformManifest.txt
mscorlib.dll|Microsoft.NETCore.App.Ref|4.0.0.0|4.700.19.56404
System.IO.Compression.Native.a|Microsoft.NETCore.App.Ref||0.0.0.0
System.IO.Compression.Native.so|Microsoft.NETCore.App.Ref||0.0.0.0

So why is file saying this is a Message Sequence Chart? Let’s see if we can find an answer.

We know what algorithm file uses. So we will walk through it.

First, file will check if it’s a special file. We can use stat to find out what if that’s true.

$ stat -c '%F' Microsoft.NETCore.App.Ref/3.1.0/data/PlatformManifest.txt
regular file

Here, stat confirms that this is regular file.

Next, file will use magic to guess the file type.

Does the magic database know about this file type at all?

$ grep 'Message Sequence Chart' /usr/share/magic
0       string          mscdocument     Message Sequence Chart (document)
0       string          msc             Message Sequence Chart (chart)
0       string          submsc          Message Sequence Chart (subchart)

Aha! It does.

For this set of tests, magic looks at the string at position 0 and if it matches with msc (or mscdocument, or submsc), it thinks this is a Message Sequence Chart.

If you recall, the first line of the text file was:

$ head Microsoft.NETCore.App.Ref/3.1.0/data/PlatformManifest.txt
mscorlib.dll|Microsoft.NETCore.App.Ref|4.0.0.0|4.700.19.56404

The first few bytes of the file incorrectly match with msc. magic and file see the match and think this file is a Message Sequence Chart (chart).

This was clearly a false alarm from file (and our tooling). We ignored it.

closing thoughts

Hopefully, you now have a good idea of how file and magic work. You also have a good starting point to use and debug these tools if you never need to.


  1. The thing(number) convention is pretty common in Unix/Linux documentation. The thing part indicates the name of the concept. The number indicates the type of the thing as well as the section number of man pages where this thing is documented. For example, someone can be talking about exec(1p) and exec(3). By exec(1p) they mean the one that is documented in section 1: Executable programs. So exec(1p) is the exec program. By exec(3) they mean the one that is documented in section 3: Library calls. So exec(3) is a library call. You can read the manual pages of a particular version by running man 1p exec or man 3 exec. ↩︎

  2. If you do patch the database to fix a bug, please contribute the change back upstream. ↩︎

  3. It turns out that Java class files and MACH executables use the same first sequence of bytes in the file (0xcafebabe). The next set of bytes can be used to guess whether this is Java class file or a MACH executable. Java class files have the class file version here, which is closer to 50. MACH executables use a much lower value. That’s why the first test here for 0xcafebabe doesn’t print anything: at this point, magic doesn’t have enough information to claim if this is Java or not. ↩︎