Re: [PATCH 10/10] mm/migrate.c: call detach_page_private to cleanup code

[Guoqing Jiang]

On 5/19/20 12:06 PM, Gao Xiang wrote:
> On Tue, May 19, 2020 at 09:35:59AM +0200, Guoqing Jiang wrote:
> > On 5/19/20 7:12 AM, Andrew Morton wrote:
> > > On Sun, 17 May 2020 23:47:18 +0200 Guoqing Jiang <guoqing.jiang@xxxxxxxxxxxxxxx> wrote:
> > >
> > > > We can cleanup code a little by call detach_page_private here.
> > > >
> > > > ...
> > > >
> > > > --- a/mm/migrate.c
> > > > +++ b/mm/migrate.c
> > > > @@ -804,10 +804,7 @@ static int __buffer_migrate_page(struct address_space *mapping,
> > > >    	if (rc != MIGRATEPAGE_SUCCESS)
> > > >    		goto unlock_buffers;
> > > > -	ClearPagePrivate(page);
> > > > -	set_page_private(newpage, page_private(page));
> > > > -	set_page_private(page, 0);
> > > > -	put_page(page);
> > > > +	set_page_private(newpage, detach_page_private(page));
> > > >    	get_page(newpage);
> > > >    	bh = head;
> > > mm/migrate.c: In function '__buffer_migrate_page':
> > > ./include/linux/mm_types.h:243:52: warning: assignment makes integer from pointer without a cast [-Wint-conversion]
> > >    #define set_page_private(page, v) ((page)->private = (v))
> > >                                                       ^
> > > mm/migrate.c:800:2: note: in expansion of macro 'set_page_private'
> > >     set_page_private(newpage, detach_page_private(page));
> > >     ^~~~~~~~~~~~~~~~
> > >
> > > The fact that set_page_private(detach_page_private()) generates a type
> > > mismatch warning seems deeply wrong, surely.
> > >
> > > Please let's get the types sorted out - either unsigned long or void *,
> > > not half-one and half-the other.  Whatever needs the least typecasting
> > > at callsites, I suggest.
> > Sorry about that, I should notice the warning before. I will double check if
> > other
> > places need the typecast or not, then send a new version.
> >
> > > And can we please implement set_page_private() and page_private() with
> > > inlined C code?  There is no need for these to be macros.
> > Just did a quick change.
> >
> > -#define page_private(page)Ã Ã Ã Ã Ã Ã Ã Ã Ã Ã Ã Ã  ((page)->private)
> > -#define set_page_private(page, v)Ã Ã Ã Ã Ã  ((page)->private = (v))
> > +static inline unsigned long page_private(struct page *page)
> > +{
> > +Ã Ã Ã Ã Ã Ã  return page->private;
> > +}
> > +
> > +static inline void set_page_private(struct page *page, unsigned long
> > priv_data)
> > +{
> > +Ã Ã Ã Ã Ã Ã  page->private = priv_data;
> > +}
> >
> > Then I get error like.
> >
> > fs/erofs/zdata.h: In function ÃâËz_erofs_onlinepage_indexÃââ:
> > fs/erofs/zdata.h:126:8: error: lvalue required as unary ÃâË&Ãââ operand
> > Ã  u.v = &page_private(page);
> > Ã Ã Ã Ã Ã Ã Ã  ^
> >
> > I guess it is better to keep page_private as macro, please correct me in
> > case I
> > missed something.
> I guess that you could Cc me in the reply.

Sorry for not do that ...

> In that case, EROFS uses page->private as an atomic integer to
> trace 2 partial subpages in one page.
>
> I think that you could also use &page->private instead directly to
> replace &page_private(page) here since I didn't find some hint to
> pick &page_private(page) or &page->private.

Thanks for the input, I just did a quick test, so need to investigate more.
And I think it is better to have another thread to change those macros to
inline function, then fix related issues due to the change.

> In addition, I found some limitation of new {attach,detach}_page_private
> helper (that is why I was interested in this series at that time [1] [2],
> but I gave up finally) since many patterns (not all) in EROFS are
>
> io_submit (origin, page locked):
> attach_page_private(page);
> ...
> put_page(page);
>
> end_io (page locked):
> SetPageUptodate(page);
> unlock_page(page);
>
> since the page is always locked, so io_submit could be simplified as
> set_page_private(page, ...);
> SetPagePrivate(page);
> , which can save both one temporary get_page(page) and one
> put_page(page) since it could be regarded as safe with page locked.

The SetPageUptodate is not called inside {attach,detach}_page_private,
I could probably misunderstand your point, maybe you want the new pairs
can handle the locked page, care to elaborate more?

> btw, I noticed the patchset versions are PATCH [3], RFC PATCH [4],
> RFC PATCH v2 [5], RFC PATCH v3 [6], PATCH [7]. Although I also
> noticed the patchset title was once changed, but it could be some
> harder to trace the whole history discussion.
>
> [1] https://lore.kernel.org/linux-fsdevel/20200419051404.GA30986@hsiangkao-HP-ZHAN-66-Pro-G1/
> [2] https://lore.kernel.org/linux-fsdevel/20200427025752.GA3979@hsiangkao-HP-ZHAN-66-Pro-G1/
> [3] https://lore.kernel.org/linux-fsdevel/20200418225123.31850-1-guoqing.jiang@xxxxxxxxxxxxxxx/
> [4] https://lore.kernel.org/linux-fsdevel/20200426214925.10970-1-guoqing.jiang@xxxxxxxxxxxxxxx/
> [5] https://lore.kernel.org/linux-fsdevel/20200430214450.10662-1-guoqing.jiang@xxxxxxxxxxxxxxx/
> [6] https://lore.kernel.org/linux-fsdevel/20200507214400.15785-1-guoqing.jiang@xxxxxxxxxxxxxxx/
> [7] https://lore.kernel.org/linux-fsdevel/20200517214718.468-1-guoqing.jiang@xxxxxxxxxxxxxxx/

All the cover letter of those series are here.

RFC V3:https://lore.kernel.org/lkml/20200507214400.15785-1-guoqing.jiang@xxxxxxxxxxxxxxx/
RFC V2:https://lore.kernel.org/lkml/20200430214450.10662-1-guoqing.jiang@xxxxxxxxxxxxxxx/
RFC:https://lore.kernel.org/lkml/20200426214925.10970-1-guoqing.jiang@xxxxxxxxxxxxxxx/

And the latest one:

https://lore.kernel.org/lkml/20200430214450.10662-1-guoqing.jiang@xxxxxxxxxxxxxxx/


Thanks,
Guoqing