From: Hugh Dickins <hugh@veritas.com>

Kevin P.  Fleming pointed out that the 2.6 tmpfs does not allow writing huge
sparse files.  This is an unintended side-effect of the strict memory commit
changes: which should make no difference.

The solution is to treat the tmpfs files (of variable size) and the shmem
objects (of fixed size) differently: sounds nasty but works out well.  The
shmem objects follow the VM preallocation convention as before, but the tmpfs
files revert to allocation on demand as a filesystem would.  If there's not
enough memory to write to a tmpfs hole, it is reported as -ENOSPC rather than
-ENOMEM, so the mmap writer gets SIGBUS rather than everyone else getting
OOM-killed.