wli suffered OOMs because tmpfs was allocating GFP_USER, for its
metadata pages.  This patch allocates them GFP_HIGHUSER (default
mapping->gfp_mask) and uses atomic kmaps to access (KM_USER0 for upper
levels, KM_USER1 for lowest level).  shmem_unuse_inode and
shmem_truncate rewritten alike to avoid repeated maps and unmaps of the
same page: cr's truncate was much more elegant, but I couldn't quite
see how to convert it.

I do wonder whether this patch is a bloat too far for tmpfs, and even
non-highmem configs will be penalised by page_address overhead (perhaps
a further patch could get over that).  There is an attractive
alternative (keep swp_entry_ts in the existing radix-tree, no metadata
pages at all), but we haven't worked out an unhacky interface to that.
For now at least, let's give tmpfs highmem metadata a spin.