Why does tar appear to skip file contents when output file is /dev/null?

Question

I have a directory with over 400 GiB of data in it. I wanted to check that all the files can be read without errors, so a simple way I thought of was to tar it into /dev/null. But instead I see the following behavior:

$ time tar cf /dev/null .

real    0m4.387s
user    0m3.462s
sys     0m0.185s
$ time tar cf - . > /dev/null

real    0m3.130s
user    0m3.091s
sys     0m0.035s
$ time tar cf - . | cat > /dev/null
^C

real    10m32.985s
user    0m1.942s
sys     0m33.764s

The third command above was forcibly stopped by Ctrl+C after having run for quite long already. Moreover, while the first two commands were working, activity indicator of the storage device containing . was nearly always idle. With the third command the indicator is constantly lit up, meaning extreme busyness.

So it seems that, when tar is able to find out that its output file is /dev/null, i.e. when /dev/null is directly opened to have the file handle which tar writes to, file body appears skipped. (Adding v option to tar does print all the files in the directory being tar'red.)

So I wonder, why is this so? Is it some kind of optimization? If yes, then why would tar even want to do such a dubious optimization for such a special case?

I'm using GNU tar 1.26 with glibc 2.27 on Linux 4.14.105 amd64.

Answer 1

It is a documented optimization:

When the archive is being created to /dev/null, GNU tar tries to minimize input and output operations. The Amanda backup system, when used with GNU tar, has an initial sizing pass which uses this feature.

Answer 2

This can happen with a variety of programs, for example, I had that behavior once when just using cp file /dev/null; instead of getting an estimate of my disk read speed, the command returned after a few milliseconds.

As far as I remember, that was on Solaris or AIX, but the principle applies to all kinds of unix-y systems.

In the old times, when a program copied a file to somewhere, it'd alternate between read calls that get some data from disk (or whatever the file descriptor is referring to) to memory (with a guarantee everything is there when read returns) and write calls (which take the chunk of memory and send the content to the destination).

However, there are at least two newer ways to achieve the same:

• Linux has system calls copy_file_range (not portable to other unixes at all) and sendfile (somewhat portable; originally intended to send a file to the network, but can use any destination now). They're intended to optimize transfers; if the program uses one of those, it's easily conceivable the kernel recognizes the target is /dev/null and turns the system call into a no-op

• Programs can use mmap to get the file contents instead of read, this basically means "make sure the data is there when I try to access that chunk of memory" instead of "make sure the data is there when the system call returns". So a program can mmap the source file, then call write on that chunk of mapped memory. However, as writing /dev/null doesn't need to access the written data, the "make sure it's there" condition isn't ever triggered, resulting in the file not being read either.

Not sure if gnu tar uses any, and which, of these two mechanisms when it detects it's writing to /dev/null, but they're the reason why any program, when used to check read-speeds, should be run with | cat > /dev/null instead of > /dev/null - and why | cat > /dev/null should be avoided in all other cases.