Re: put_page() tries to handle hugepages but fails

[Andrew Morton <akpm@osdl.org>]

David Gibson <david@gibson.dropbear.id.au> wrote:
>
> Andrew, please apply.
> 
> The code of put_page() is misleading, in that it appears to have code
> handling PageCompound pages (i.e. hugepages).  However it won't
> actually handle them correctly - __page_cache_release() will not work
> properly on a compound.  Instead, hugepages should be and are released
> with huge_page_release() from mm/hugetlb.c.  This patch removes the
> broken PageCompound path from put_page(), replacing it with a
> BUG_ON().  This also removes the initialization of page[1].mapping
> from compoound pages, which was only ever used in this broken code
> path.

We could certainly remove the test for a null destructor in there and
require that compound pages have a destructor installed.

But the main reason why that code is in there is for transparently handling
direct-io into hugepage regions.  That code does perform put_page against
4k pageframes within the huge page and it does follow the pointer to the
head page.

With your patch applied get_user_pages() and bio_release_pages() will
manipulate the refcounts of the inner 4k pages rather than the head pages
and things will explode.

We could change follow_hugetlb_page() to always take a ref against the head
page and we could teach bio_release_pages() to perform appropriate pfn
masking to locate the head page, and perform similar tricks for
futexes-in-large-pages.  But with the code as-is the refcounting works
transparently.


If it's "broken" I wanna know why.