Re: [PATCH] add __GFP_ZERO to GFP_LEVEL_MASK

[Christoph Lameter <clameter@sgi.com>]

GFP_LEVEL_MASK is used to allow the pass through of page allocator
flags. Currently these are

#define GFP_LEVEL_MASK (__GFP_WAIT|__GFP_HIGH|__GFP_IO|__GFP_FS| \
                        __GFP_COLD|__GFP_NOWARN|__GFP_REPEAT| \
                        __GFP_NOFAIL|__GFP_NORETRY|__GFP_COMP| \
                        __GFP_NOMEMALLOC|__GFP_HARDWALL|__GFP_THISNODE|
			__GFP_MOVABLE)

Some of these flags control page allocator reclaim and fallback
behavior. If they are specified for a slab alloc operation then they
are effective if a new slab has to be allocated. These are

1. Reclaim control

__GFP_WAIT
__GFP_IO
__GFP_FS
__GFP_NOWARN
__GFP_REPEAT
__GFP_NOFAIL
__GFP_NORETRY

2. Reserve control

__GFP_HIGH
__GFP_NOMEMALLOC

2. Fallback control

__GFP_HARDWALL	(cpuset contraints)
__GFP_THISNODE (handled by SLAB on its own, SLUB/SLOB pass through)

AFAIK these make sense.

Then there are some other flags. I am wondering why they are in
GFP_LEVEL_MASK?

__GFP_COLD	Does not make sense for slab allocators since we have
		to touch the page immediately.

__GFP_COMP	No effect. Added by the page allocator on their own
		if a higher order allocs are used for a slab.

__GFP_MOVABLE	The movability of a slab is determined by the
		options specified at kmem_cache_create time. If this is
		specified at kmalloc time then we will have some random
		slabs movable and others not.