Re: [PATCH 1/11] Add __GFP_MOVABLE flag and update callers

[Mel Gorman]

On Sun, 26 Nov 2006, Hugh Dickins wrote:

> On Sat, 25 Nov 2006, Linus Torvalds wrote:
> > On Fri, 24 Nov 2006, Hugh Dickins wrote:
> > > You need to add in something like the patch below (mutatis mutandis
> > > for whichever approach you end up taking): tmpfs uses highmem pages
> > > for its swap vector blocks, noting where on swap the data pages are,
> > > and allocates them with mapping_gfp_mask(inode->i_mapping); but we
> > > don't have any mechanism in place for reclaiming or migrating those.
> >
> > I think this really just points out that you should _not_ put MOVABLE into
> > the "mapping_gfp_mask()" at all.
> >
> > The mapping_gfp_mask() should really just contain the "constraints" on
> > the allocation, not the "how the allocation is used". So things like "I
> > need all my pages to be in the 32bit DMA'able region" is a constraint on
> > the allocator, as is something like "I need the allocation to be atomic".
> >
> > But MOVABLE is really not a constraint on the allocator, it's a guarantee
> > by the code _calling_ the allocator that it will then make sure that it
> > _uses_ the allocation in a way that means that it is movable.

Later, MOVABLE might be a constraint. For example, hotpluggable nodes may 
only allow MOVABLE allocations to be allocated.

> > So it shouldn't be a property of the mapping itself, it should always be a
> > property of the code that actually does the allocation.
> >
> > Hmm?
>
> Not anything I feel strongly about, but I don't see it that way.
>
> mapping_gfp_mask() seems to me nothing more than a pragmatic way
> of getting the appropriate gfp_mask down to page_cache_alloc().

And that is important for any filesystem that uses generic_file_read(). As 
page_cache_alloc() has no knowledge of the filesystem, it depends on the 
mapping_gfp_mask() to determine if the pages are movable or not.

> alloc_inode() initializes it to whatever suits most filesystems
> (currently GFP_HIGHUSER), and those who differ adjust it (e.g.
> block_dev has good reason to avoid highmem so sets it to GFP_USER
> instead).  It used to be the case that several filesystems lacked
> kmap() where needed, and those too would set GFP_USER: what you call
> a constraint seems to me equally a property of the surrounding code.
>
> If __GFP_MOVABLE is coming in, and most fs's are indeed allocating
> movable pages, then I don't see why MOVABLE shouldn't be in the
> mapping_gfp_mask.  Specifying MOVABLE constrains both the caller's
> use of the pages, and the way they are allocated; as does HIGHMEM.

> From what I've seen, the majority of filesystems are suitable for using 
__GFP_MOVABLE and it would be clearer to have GFP_HIGH_MOVABLE as the 
default and setting GFP_HIGHUSER in filesystems like ramfs.

> And we shouldn't be guided by the way tmpfs (ab?)uses that gfp_mask
> for its metadata allocations as well as its page_cache_alloc()s:
> that's just a special case.  Though the ramfs case is more telling
> (its pagecache pages being not at present movable).
>
> Hugh

--
Mel Gorman
Part-time Phd Student                          Linux Technology Center
University of Limerick                         IBM Dublin Software Lab
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/