Just build and try this small example.  There are two pieces, some source
in "mapit.c", and the script that uses it to demonstrate the problem, in
"doit".  If you place both of those in the same directory on an NFS
partition, and run the script, you'll see output like:

rm: cannot remove directory `dir': File exists
rm: cannot unlink `dir/.nfs0001081d0000042f': Device or resource busy
rm: cannot remove directory `dir': File exists

It's the nfs_sillyrename() code that's creating that .nfs file.  I
understand its reason for existing.

The only trouble is that the kernel doesn't realize that it can delete it
in any kind of timely fashion.  So, when a subsequent rmdir() is called on
the parent directory, it fails.  And to make matters worse, a subsequent rm
-rf command fails (with EBUSY) trying to delete the .nfs file itself.

=====mapit.c=====
#include <sys/types.h>
#include <sys/stat.h>
#include <fcntl.h>
#include <unistd.h>
#include <sys/mman.h>

main()
{
   int    fd;
   char*  addr;

   fd = open("dir/file", O_RDWR|O_CREAT|O_TRUNC, 0666);
   ftruncate(fd, 4096);
   addr = mmap((void*)0, 4096, PROT_READ|PROT_WRITE, MAP_SHARED, fd, 0);
   memset(addr, ' ', 4096);
   close(fd);
}

=====doit=====
#!/bin/ksh
#

make -s mapit

rm -rf dir 2>/dev/null
if [ -d dir ]; then
   echo "... waiting for directory \"dir\" to clear out to re-run test ..."
   while [ -d dir ]; do
      rm -rf dir 2>/dev/null
   done
fi
mkdir -p dir

./mapit

rm -rf dir || rm -rf dir