The notion of a "shrinker" in vmscan.c is interesting.  I ran across it while
working out how inodes are destroyed (one way is using prune_icache, which
ends up being called in such a way).

struct shrinker {
	shrinker_t              shrinker;
	struct list_head        list;
	int                     seeks;  /* seeks to recreate an obj */
	long                    nr;     /* objs pending delete */
};

typedef int (*shrinker_t)(int nr_to_scan, unsigned int gfp_mask);
	Scans nr_to_scan objects, freeing them if possible.  Returns number of
	objects left in the cache after shrinking.  VM may call with
	nr_to_scan == 0 just to find out number of objects.

"list" field links shrinkers into global shrinker_list.

set_shrinker(int seeks, shrinker_t theshrinker);
	install a "shrinker" on the shrinker list.
	The number of seeks is an estimate of the number of disk seeks that
	will be required to free each object.

Currently is used for dcache, inodes, a few other things.

Could we use something like this e.g. for deciding when to free things like
nfsv4 server-side state or gss contexts, or would that be overkill?