Re: page refcount race between prep_compound_gigantic_page() and __page_cache_add_speculative()?

[Matthew Wilcox <willy@infradead.org>]

On Tue, Jun 15, 2021 at 01:03:53PM +0200, Jann Horn wrote:
> The messier path, as the original commit describes, is "gigantic" page
> allocation. In that case, we'll go through the following path (if we
> ignore CMA):
> 
>   alloc_fresh_huge_page():
>     alloc_gigantic_page()
>       alloc_contig_pages()
>         __alloc_contig_pages()
>           alloc_contig_range()
>             isolate_freepages_range()
>               split_map_pages()
>                 post_alloc_hook() [FOR EVERY PAGE]
>                   set_page_refcounted()
>                     set_page_count(page, 1)
>     prep_compound_gigantic_page()
>       set_page_count(p, 0) [FOR EVERY TAIL PAGE]
> 
> so all the tail pages are initially allocated with refcount 1 by the
> page allocator, and then we overwrite those refcounts with zeroes.
> 
> 
> Luckily, the only non-__init codepath that can get here is
> __nr_hugepages_store_common(), which is only invoked from privileged
> writes to sysfs/sysctls.

Argh.  What if we passed __GFP_COMP into alloc_contig_pages()?
The current callers of alloc_contig_range() do not pass __GFP_COMP,
so it's no behaviour change for them, and __GFP_COMP implies this
kind of behaviour.  I think that would imply _not_ calling
split_map_pages(), which implies not calling post_alloc_hook(),
which means we probably need to do a lot of the parts of
post_alloc_hook() in alloc_gigantic_page().  Yuck.