Why does `df` lie about free space

[Kyryll A Mirnenko]

  That's an old problem, but yesterday I was working with 1.4G UFS2-slice 
reported by `df` to have 400kb of free space. Alternative calculations (e.g. 
writing a random file until kernel says no inode's free) give a result of 
more that 100M (!) unused. Thats about 7% of the whole size!
  That could a harmless bug, but some applications (PostgreSQL is!) uses 
statfs to get free space info. Posetgres dies saying can't write his *.pid 
cause FS is full (but it has 100M inodes free)...
  So whats wrong with `df` (e.g. statfs/fstatfs)?