Deadlock detection algorithm distinguish processes by pair pid+export. It is not correct when one process access files from different mount points.
Proper entity should be pid+NID

testcase patch attached:

/local/work/lustre/lustre/tests/flocks_test 4 /mnt/lustre/file1 /mnt/lustre2/file2
6646: get lock2
6647: get lock1
6646: done
6647: done
6647: get lock2
6646: get lock1

it waits forever, while works with files from one mount point

/local/work/lustre/lustre/tests/flocks_test 4 /mnt/lustre/file1 /mnt/lustre/file2
6639: get lock2
6640: get lock1
6639: done
6640: done
6639: get lock1
6640: get lock2
fcntl cmd 7 failed: Resource deadlock avoided
6640: done
6640: exit rc=0
6639: done
6639: exit rc=0