Re: page_order private bit causes problems with dma_alloc_coherent?

[KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>]

Jim Duchek wrote:
> I'm attempting to debug a problem where pci_free_consistent causes a spewage
> of bad_page's due to the LG_private bit being set.  The remove goes through
> free_pages_ok, which does the free_page_check (which errors on the private
> bit) _BEFORE_ free_pages_bulk (which removes the private bit)
> 

When page is free , page_count(page) == 0. But this just means page is 'free'.
To show page is a head of buddy (contiguous pages of 2^order pages), order is
stored into page->private. To show page->private has a valid 'order', PG_private
is used.

See page_is_buddy() function.

> I'm not entirely sure why the pages are getting this bit -- this is a DRI
> driver and the bits don't get added during the whole time I'm running an
> application -- they get added shortly before I kill the GL client.  Any
> hints?  Is there a way to force pages never to use the page_order stuff?
> Should free_pages_check not consider the private bit such a bad thing?
> Should I just hack in something where I remove all the bits right before I
> do my free? :)
> 
I think removing PG_private before freeing is sane.
The page is free, so it's needless to use PG_private bit when free_pages() is called.

Thanks,
-- Kame