Re: swapper_space export

[Hugh Dickins <hugh@veritas.com>]

On Wed, 17 May 2006, Arjan van de Ven wrote:
> On Tue, 2006-05-16 at 19:24 -0400, Josef Sipek wrote:
> > I was trying to compile the Unionfs[1] to get it up to sync it up with
> > the kernel developments from the past few months. Anyway, long story
> > short...swapper_space (defined in mm/swap_state.c) is not exported
> > anymore (git commit: 4936967374c1ad0eb3b734f24875e2484c3786cc). This
> > apparently is not an issue for most modules. Troubles arise when the
> > modules include mm.h or any of its relatives.
> > 
> > One simply gets a linker error about swapper_space not being defined.
> > The problem is that it is used in mm.h.
> 
> don't you think it's really suspect that no other filesystem, in fact no
> other driver in the kernel needs this? Could it just be that unionfs is 
> using a wrong API ? Because if that's the case you're patch is just the
> wrong thing. Maybe the unionfs people should try to submit their code
> for review etc......

Much as I'd love to side with Jeff against you and Adrian ;)
I think you're right.

I see no reference to page_mapping() in the unionfs source (and at
present there's no other justifiable modular use of swapper_space);
but my guess would be that Jeff is being more conscientious than is
called for in getting it to sync up with the kernel -

The unionfs source does contain its own inline "sync_page" which
comments that it "is copied verbatim from mm/filemap.c".  I'm
guessing Jeff has noticed that it's no longer a verbatim copy,
has made it so, and is thereby involving page_mapping().

No need for that here (nor for the smp_mb nor for the io_schedule):
unionfs's sync_page is working on a locked pagecache page of the
lower-level filesystem, that's not going to be a PageSwapCache page
nor a PageAnon page (nor even a truncated page with NULL mapping:
page lock is held).  Just use page->mapping as before.

(But I notice that unionfs better not have a tmpfs in its union:
the unionfs use of grab_cache_page is not strictly compatible with
the way tmpfs pages are swapped out under memory pressure.)

Hugh