ZFS Compression, when du and ls appear to disagree.

I have recently been migrated to a new ZFS based home directory at work, I’ve got hourly snapshots of my data and (as I’ve just found it) automatic on disk compression is enabled. Recently, when dealing with a corrupt .tar archive the on disk compression managed to suitably confuse me - enough to write this blog entry.

After saving my .tar file into my home directory I decided to verify the size on disk with the size on the ftp, the tar file on the ftp was roughly 20mb.

# du -hs download.tar
6.2M download.tar

That didn’t look right so I verified the file size with ls

# ls -ahl download.tar
-rw-r–r– 1 iain staff 20M May 9 17:41 download.tar

So, ls thinks my file is 20mb in size and du thinks its only 6mb, turns out that this is the expected behavior when you have compression enabled on the file system.
Both du and ls get their information about a file by performing the stat() system call against the file, /bin/stat under Solaris Nevada shows us the information returned from the stat() system call.

#/bin/stat download.tar
File: `download.tar’
Size: 20616912 Blocks: 12769 IO Block: 8192 regular file
Device: 5102d9ah/84946330d Inode: 76465 Links: 1
Access: (0644/-rw-r–r–) Uid: (50000/iain) Gid: ( 10/ staff)
Access: 2008-05-09 17:53:17.653543700 +0100
Modify: 2008-05-09 17:41:18.136951400 +0100
Change: 2008-05-09 17:41:18.136951400 +0100

The file attributes returned contain the files size in bytes, this is what ls returns to us when performing a directory listing. However, the same is not true for du. Based on the source of du.c one can see that du calculates the file size on disk by multiplying the number of blocks the file takes on disk up by 512.

12769 * 512 = 6537728 Bytes or 6.23Mb

Although there is a clear explanation for this behavior I think that it may cause some confusion. I am sure that many people rely on du returning the actual size of the file, this could cause problems if you are preparing to copy the files to cdrom or to any other file system that does not perform compression. It is certainly something that anyone administrating ZFS file systems should be aware of.