Re: [PATCH] add __GFP_ZERO to GFP_LEVEL_MASK

[Andrew Morton <akpm@linux-foundation.org>]

On Tue, 24 Jul 2007 16:58:51 -0700 (PDT) Christoph Lameter <clameter@sgi.com> wrote:

> On Tue, 24 Jul 2007, Andrew Morton wrote:
> 
> > __GFP_COMP I'm not so sure about. 
> > drivers/char/drm/drm_pci.c:drm_pci_alloc() (and other places like infiniband)
> > pass it into dma_alloc_coherent() which some architectures implement via slab.  umm,
> > arch/arm/mm/consistent.c is one such.
> 
> Should  drm_pci_alloc really aright in setting __GFP_COMP? 

I don't see what's special about that dma_alloc_coherent() call.

> dma_alloc_coherent does not set __GFP_COMP for other higher order allocs 
> and expects to be able to operate on the page structs indepedently. That 
> is not the case for a compound page.
> 
> Creates a really interesting case for SLAB. Slab did not use __GFP_COMP in 
> order to be able to allow the use page->private (No longer an issue since 
> the 2.6.22 cleanups and avoiding the use of page->private for the compound 
> head).
> 
> Now the __GFP_COMP flag is passed through for any higher order page alloc 
> (such as a kmalloc allocation > PAGE_SIZE). Then we may have allocated one 
> slab that is a compound page amoung others higher order pages allocated 
> without __GFP_COMP. May have caused rare and strange failures in 2.6.21 
> and earlier because of the concurrent page->private use in compound head 
> pages and arch pages.
> 
> SLUB will always use __GFP_COMP so the pages are consistent regardless if 
> __GFP_COMP is passed in or not.
> 
> The strange scenarios come about by expecting a page allocation when 
> sometimes we just substitute a slab alloc.
> 
> We could filter __GFP_COMP out to avoid the BUG()? Or deal with it on a 
> case by case basis?

Fix callers, I'd suggest.  There are a number of fishy-looking open-coded
usages of __GFP_COMP around the place.

It's a bit sad that some architectures are using slab for dma_alloc_coherent()
while others go to alloc_pages().