Re: __GFP_HIGH ignored?

[Linus Torvalds]

On Sun, 2 Sep 2001, Daniel Phillips wrote:
>
> __GFP_HIGH is apparently ignored now.  Its intended function is performed
> instead by PF_MEMALLOC.  Should we take it out?

No. We may have a use for it in the future, especially as PF_MEMALLOC
really means something completely different.

I stronly suspect that the current PF_MEMALLOC handling is much too
strict: instead of trating a PF_MEMALLOC as a atomic high-priority thing,
we could treat it as something like

	/*
	 * recursive call? Make sure to strip out anything
	 * that could cause deadlocks and further recursion..
	 */
	if (current->flags & PF_MEMALLOC) {
		gfp_mask &= ~(__GFP_HIGHIO | __GFP_FS);
		page_launder(gfp_mask);
		page = RMQUEUE(..)
		if (page)
			return page;
		goto repeat;
	}

(Yeah, not exactly like the above, but you get the idea - allow a very
limited form of recursion, and allow a PF_MEMALLOC to always use the
reserves - together they should make us better at handling the current
issues with bounce buffers, for example).

It may be that we'll never need __GFP_HIGH, but let's keep it - it does
have semantics that make some amount of sense.

		Linus