Oh gawd, that's terrible. Never, ever duplicate code like that.

What the patch does is:
- formally shift the loop one level down in the call graph, adding instances of __tmp_remove_tables() 
exactly to locations where instances of __tmp_remove_table() already exist,
- on architectures where __tmp_remove_tables() resulted into calling free_page_and_swap_cache() in loop, 
call batched free_page_and_swap_cache_nolru() instead,
- on other places, keep the loop as is - perhaps as a possible target for future optimizations.

The extra duplication added by this patch just highlights already existing duplication of 
__tlb_remove_table() implementations.

Ok let's follow your suggestion instead. AFAIU, that is:
- remove the free_page_and_swap_cache() based implementation from archs,
- instead, add it into mm/mmu_gather.c, ifdef-ed by a new Kconfig key, and define that Kconfig key into 
the archs that use it,
- then, keep the optimization inside mm/mmu_gather.c.

Indeed, the overall change will become smaller then. Thanks for the idea. Will post patches doing that soon.

Nikita