Re: [PATCH] swap 13/13 may_enter_fs?

[Andrew Morton <akpm@digeo.com>]

Hugh Dickins <hugh@veritas.com> wrote:
>
> shrink_list's may_enter_fs (may_write_page would be a better name)
> currently reflects that swapcache page I/O won't suffer from FS
> complications, so can be done if __GFP_IO without __GFP_FS; but
> the same is true of a tmpfs page (which is just this stage away
> from being a swapcache page), so check bdi->memory_backed instead.
> 
> ...
>
> +			if (!(gfp_mask & (bdi->memory_backed?
> +					__GFP_IO: __GFP_FS)))
>  				goto keep_locked;

Barf.  I haven't used a question mark operator in ten years, and this is a
fine demonstration of why ;)

I think a feasibly comprehensible transformation would be:

	/*
	 * A comment goes here
	 */
	if (bdi->memory_backed)
		may_enter_fs = gfp_mask & __GFP_IO;
	else
		may_enter_fs = gfp_mask & __GFP_FS;


That being said, this is a bit presumptuous.  PF_MEMALLOC will protect us
from infinite recursion but there are other reasons for GFP_NOFS.

For example, a memory-backed filesystem may be trying to allocate GFP_NOFS
memory while holding filesystem locks which are taken by its writepage.

How about adding a new field to backing_dev_info for this case?  Damned if I
can think of a name for it though.