df in linux not showing correct free space after file removal

I have file servers which are used to store files. Files might reside there for a week, or for a year. Unfortunately, when I remove files from the server, df command doesn't reflect the freed up space. So eventually, the server gets filled up (df shows 99%), and my script doesn't send any more files there, except there might be a few dozen GB of free space on there.

I got noatime flag on the mounted partitions if that makes any difference.

Solution 1:

Deleting the filename doesn't actually delete the file. Some other process is holding the file open, causing it to not be deleted; restart or kill that process to release the file.

Use

lsof +L1

to find out which process is using a deleted (unlinked) file.

Solution 2:

as Ignacio mentions, deleting the file won't free the space until you delete the processes that have open handles against that file.

Nevertheless, you can reclaim the space without killing the processes. All you need to do is to remove the file descriptors.

First execute lsof | grep deleted to identify the process holding the file

[hudson@opsynxvm0055 log]$ /usr/sbin/lsof |grep deleted
java       8859   hudson    1w      REG              253,0 3662503356    7578206 /crucible/data/current/var/log/fisheye.out (deleted)

Then execute:

cd /proc/PID/fd

then

[hudson@opsynxvm0055 fd]$ ls -l |grep deleted
total 0
l-wx------ 1 hudson devel 64 Feb  7 11:48 1 -> /crucible/data/current/var/log/fisheye.out (deleted)

The "1" will be the file descriptor. Now type "> FD" to reclaim that space

> 1

You might need to repeat the operation if there are other processes holding the file.

Solution 3:

One possibility is that the file(s) you deleted have more references in the filesystem. If you've created hardlinks, several filenames will point to the same data, and the data (the actual contents) won't be marked as free/usable until all references to it has been removed. Before you delete files, either stat them (Entry named Links) or do ls -l on them (should be the second column).

If it does turn out that the files are referenced elsewhere, I guess you'll have to ls -i the file(s) to find the inode-number, and then do a find with -inum <inode-number> to find the other references to that file (you probably also want to use -mount to stay within the same filesystem as well).