Re: [PATCH 1/5] mm: remove page_is_file_lru function

[Matthew Wilcox]

On Fri, Jan 21, 2022 at 03:02:12PM +0800, Alex Shi wrote:
> On Thu, Jan 20, 2022 at 9:28 PM Matthew Wilcox <willy@xxxxxxxxxxxxx> wrote:
> >
> > On Thu, Jan 20, 2022 at 09:10:20PM +0800, alexs@xxxxxxxxxx wrote:
> > > From: Alex Shi <alexs@xxxxxxxxxx>
> > >
> > > This function could be full replaced by folio_is_file_lru, so no reason
> > > to keep a duplicate function.
> >
> > This is not a helpful way to do this kind of replacement.
> >
> > Instead of choosing a function to remove and doing a blind replacement,
> > choose a call site and convert the whole calling function to use folios.
> > Once you've removed all callers, you can remove the wrapper function.
> >
> > Also, a number of changes here will conflict with patches I've already
> > posted.  Try doing change_pte_range() in mprotect.c to get a feel for
> > how to convert a function entirely to folios.
> 
> Hi Willy,
> 
> Thanks for your comments!
> 
> The patchset did the thing as you required "convert the whole calling
> function to use folios. then remove the wrapper function" on yesterday's
> Linus and next tree, that included your patchset "Page cache/iomap for 5.17".

That's not what I meant.  What I meant is you're currently doing:

 - Find folio wrapper function
 - Inline it into all callers
 - Delete wrapper function

That creates a lot of churn and not a lot of improvement.

What would be helpful is doing:

 - Find folio wrapper function
 - Find a caller, convert it from using pages to using folios

That's harder, but it actually accomplishes something (ie auditing
a function to make it work with folios).  These wrapper functions are
signals that the callers need to be converted to use folios.

> Is the conflicting patch "Enabling large folios for 5.17" or others? Sorry
> for can't check everyone, your patches are many. If just the former, I see
> you mentioned: "I'd be uncomfortable seeing it merged before 5.18".
> Would you point out which of your patches was interfered or blocked?

The GUP series was the specific series that this conflicted with.
And yes, I have a lot of patches outstanding in this area.  That's a
sign that small cleanup patches aren't going to be welcomed because
they're going to conflict with meaningful patches.

> And yes, replacing page functions in change_pte_range is a bit harder,
> but it seems it has no much relation with this trival patchset.

That is, indeed, the point.