Re: [PATCH] pagemap: Add alert to mapping_set_release_always() for mapping with no release_folio

[Jan Kara]

On Fri 12-12-25 12:18:11, Matthew Wilcox wrote:
> On Thu, Dec 11, 2025 at 10:23:13AM +0100, Jan Kara wrote:
> > On Wed 10-12-25 21:36:43, Matthew Wilcox wrote:
> > > On Thu, Dec 11, 2025 at 01:31:04AM +0530, Deepakkumar Karn wrote:
> > > >  static inline void mapping_set_release_always(struct address_space *mapping)
> > > >  {
> > > > +	/* Alert while setting the flag with no release_folio callback */
> > > 
> > > The comment is superfluous.
> > 
> > Agreed.
> > 
> > > > +	VM_WARN_ONCE(!mapping->a_ops->release_folio,
> > > > +		     "Setting AS_RELEASE_ALWAYS with no release_folio");
> > > 
> > > But you haven't said why we need to do this.  Surely the NULL pointer
> > > splat is enough to tell you that you did something stupid?
> > 
> > Well, but this will tell it much earlier and it will directly point to the
> > place were you've done the mistake (instead of having to figure out why
> > drop_buffers() is crashing on you). So I think this assert makes sense to
> > ease debugging and as kind of self-reminding documentation :).
> 
> Oh.  So the real problem here is this:
> 
>         if (mapping && mapping->a_ops->release_folio)
>                 return mapping->a_ops->release_folio(folio, gfp);
>         return try_to_free_buffers(folio);
> 
> We should have a block_release_folio(), change all the BH-based
> filesystems to add it to their aops, and then change
> filemap_release_folio() to do:
> 
> 	if (mapping)
> 		return mapping->a_ops->release_folio(folio, gfp);
> 	return true;

OK, yes, this would work for me and I agree it looks like a nice cleanup.

> (actually, can the !mapping case be hit?  surely this can't be called
> for folios which have already been truncated?)

You'd think ;). There's some usage of try_to_free_buffers() from the dark
ages predating git era in jbd2 (back then jbd) which is specifically run
when we are done with journalling buffers on a page that was truncated -
see fs/jbd2/commit.c:release_buffer_page(). Also there's an interesting
case in migrate_folio_unmap() which calls try_to_free_buffers() for a
truncated page. All the other users seem to have a valid mapping.

								Honza
-- 
Jan Kara <jack@xxxxxxxx>
SUSE Labs, CR