This is an English translation of a blog item I wrote for AT Computing

While giving a course a student showed me the following:

$ ps -ef > /tmp/file

Where /tmp is 100% filled yields no errors and seems to have worked!

Lets try to see what is going on here.

Firstly, lets fill up a file system. We are going to use an fs mounted under /media/disk:

$ cp /dev/zero /media/disk/HUGE                 
$ cp: writing `/media/disk/HUGE': No space left on device

Further we need a little program to test a few things:

int 
main(void) 
{
    write(1, "hello\n", 6);
}

This writes the 6 bytes hello\n to standard output. Notice how I don’t do any error checking. If this were implemented correctly I should check how many bytes are actually written by the write() system call. Also see man 2 write.

After compilation we have our program output:

$ gcc output.c -o output

The usual I/O redirection works as expected:

$ ./output > tmp_file
$ cat tmp_file
hello

Now try this on our 100% filled file system:

$ ./output > /tmp/tmp_file

No errors - looks like this went OK.

$ ls -l /tmp/tmp_file
-rw-rw-r-- 1 miekg miekg 0 Jun 10 12:47 /tmp/tmp_file

We do have a file in /tmp, but its size is only 0 bytes. We are left with two questions. How is it that you can create a file in a full file system? And, why do we not see any data in /tmp/tmp_file given the fact that we saw no errors?

Directories⌗

If you create a directory in Linux (and Unix), then an ls -ldn will say this (empty) directory has a size of 4.0K:

$ mkdir /tmp/test
$ ls -ldh /tmp/test
drwxrwxr-x 2 miekg miekg 4.0K Jun 10 12:52 /tmp/test

That 4.0K is reserved. When we create files and or subdirectories under /tmp/test the size of the directory will only be enlarged when we cross the 4.0K boundary. When doing a little bit of testing I saw that Linux then sets the new size to 12.0K:

$ cd /tmp/test
$ for i in $(seq 0 260); do echo $i; touch file.$i; done
$ ls -ldh /tmp/test
drwxrwxr-x 2 miekg miekg 12K Jun 10 12:59 /tmp/test

So as long as you don’t cross that 4.0K boundary the system will allow you to create files. Even on otherwise full file systems!

Writing the data

If we do a ./output > /tmp/tmp_file, output does not know it is writing to disk. output just writes to its stdout and in this case output does not check if the writing succeeds. If we add this check we get a different story.

First amend the source:

int 
main(void) 
{
    if (write(1, "hello\n", 6) != 6) {
	    write(2, "ouput: write error\n", 20);
	    return 1;
    }
    return 0;
}

Recompile:

$ gcc output.c -o output

Retest:

$ ./output > /tmp/tmp_file
ouput: write error

Now we have fixed output, but there are still a lot of programs which also have this bug:

The next set of commands all fail without any errors:

$ ps -ef > /tmp/tmp_file  

$ free > /tmp/tmp_file     

$ grep 'as' testfile > /tmp/tmp_file           

$ perl -e 'print "hallo";' > /tmp/tmp_file

Luckily a lot of other programs do the right thing:

$ who > /tmp/tmp_file
who: write error: No space left on device

Also many (if not all) programs made by GNU do the correct thing.

Full file system and I/O redirection

Directories⌗