Re: Pagecache: find_or_create_page does not call a proper pageallocator function

[Andrew Morton]

On Tue, 24 Apr 2007 13:58:35 -0700 (PDT) Christoph Lameter <clameter@xxxxxxx> wrote:

> On Tue, 24 Apr 2007, Andrew Morton wrote:
> 
> > > Then I think we should disable page migration for allocations that do not 
> > > allow access to the policy zone. That would fix it.
> > 
> > Can't we use mapping_gfp_mask() when allocating the destination page?
> 
> There is no point in migrating something if you cannot reach the 
> destination. If the policy zone is not allowed by an allocation then the 
> page cannot be migrated because (on 32 bit NUMA) there is only a single 
> ZONE_NORMAL on the system. A different node requires a HIGHMEM allocation!

don't care really.  If we're allocating a page for an address_space, we
should be using its allocation mask.  As I said, we might add more things in
the future.

If in so doing the user pointlessly wastes some CPU cycles in one weird
case, it doesn't matter a lot.

> > It would be better to do so, really.  Who knows, mapping_gfp_mask() might
> > be extented in the future to say "I want GFP_NOIO" or something.  Or a
> > filesystem might specify GFP_KERNEL for regular pagecache pages or
> > whatever.
> 
> Hmmm..... How about a VM_DONTMIGRATE flag instead? That would be easy to 
> check and could be set by a device that must have all pages of the address
> space conforming to the gfp mask.
> 
> Or more general
> 
> VM_STRICT_ALLOC?

ug.  That's special-casing the blockdev pagecache peculiarity when someone
mmaps it.  It'd be better to stick with the existing interfaces and
conventions: the address_space gets to dictate how is pages are allocated.

> > Generally, the interface is "address_space tells core kernel how to
> > allocate its pages", and to be nice we should honour that in all places
> > where we allocate a page for an address_space.
> > 
> > If we'd had any brains we would have implemented this function as an
> > address_space_operations callback, but we don't so we didn't.
> 
> We have enough flags I think.

mapping_gfp_mask if a pretty foul thing.  Adding

	struct page (*alloc_page)(struct address_space *mapping);

to address_space_operations would be a quite nice cleanup.

But I ain't taking cleanups at present.

<goes back to working on a critical 2.6.21 bug>
-
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/