Re: [PATCH net-next -v5 3/4] mm: introduce __get_page() and __put_page()

[Yunsheng Lin <linyunsheng@huawei.com>]

On 2021/10/10 4:15, Matthew Wilcox wrote:
> On Sat, Oct 09, 2021 at 12:49:29PM -0700, John Hubbard wrote:
>> On 10/9/21 02:37, Yunsheng Lin wrote:
>>> Introduce __get_page() and __put_page() to operate on the
>>> base page or head of a compound page for the cases when a
>>> page is known to be a base page or head of a compound page.
>>
>> Hi,
>>
>> I wonder if you are aware of a much larger, 137-patch seriesto do that:
>> folio/pageset [1]?
>>
>> The naming you are proposing here does not really improve clarity. There
>> is nothing about __get_page() that makes it clear that it's meant only
>> for head/base pages, while get_page() tail pages as well. And the
>> well-known and widely used get_page() and put_page() get their meaning
>> shifted.
>>
>> This area is hard to get right, and that's why there have been 15
>> versions, and a lot of contention associated with [1]. If you have an
>> alternate approach, I think it would be better in its own separate
>> series, with a cover letter that, at a minimum, explains how it compares
>> to folios/pagesets.

As I was not familiar enough with mm, so I tried the semantic of
__page_frag_cache_drain(), which expects a base page or the head
page of a compound page too.

I suppose we may need to put a BUG_ON() to catch the case of
user passing a tail page accidentally, which is a run time error
and has run time overhead?
And adding a new type like folio will allow the compiler to
catch the error without any overhead?

> 
> I wasn't initially sure whether network pagepools should be part of
> struct folio or should be their own separate type.  At this point, I

Actually only a few driver are using page pool now, and others are mostly
using page allocator directly, see page_frag_alloc_align() and
skb_page_frag_refill(), only changing the page pool does not seems helpful
here, maybe the whole network stack should be using a new type like folio,
as the netstask does not need to deal with tail page directly? And it seems
virt_to_page() is one of the things need handling when netstack is changed
to use a new type like folio?

>