Re: [PATCH 6/6] mm/page_alloc.c: avoid allocating highmem pages via alloc_pages_exact_nid()

[Vlastimil Babka]

On 8/31/21 03:56, Miaohe Lin wrote:
> On 2021/8/30 22:24, Matthew Wilcox wrote:
>> On Mon, Aug 30, 2021 at 10:10:51PM +0800, Miaohe Lin wrote:
>>> Don't use with __GFP_HIGHMEM because page_address() cannot represent
>>> highmem pages without kmap(). Newly allocated pages would leak as
>>> page_address() will return NULL for highmem pages here. But It works
>>> now because the only caller does not specify __GFP_HIGHMEM now.
>> 
>> This is a misunderstanding of how alloc_pages_exact() /
>> alloc_pages_exact_nid() work.  You simply can't call them with
>> GFP_HIGHMEM.
>> 
> 
> Yep, they can't work with GFP_HIGHMEM. So IMO it might be better to
> get rid of GFP_HIGHMEM explicitly or add a comment to clarify this
> situation to avoid future misbehavior. But this may be a unnecessary
> worry... Do you prefer to not change anything here?

I agree with the suggestion below...

> Many thanks.
> 
>> If you really must change anything here,
>> s/__GFP_COMP/(__GFP_COMP|__GFP_HIGHMEM)/g throughout both
>> alloc_pages_exact() and alloc_pages_exact_nid().

... which means __GFP_HIGHMEM would be stripped and additionally there would
be a warning.