Re: [PATCH 10/10] mm: Use folio_next_pos()

[Christian Brauner <brauner@kernel.org>]

On Sat, Nov 08, 2025 at 05:00:28AM +0000, Matthew Wilcox wrote:
> On Fri, Oct 24, 2025 at 06:08:18PM +0100, Matthew Wilcox (Oracle) wrote:
> > +++ b/mm/truncate.c
> > @@ -387,7 +387,7 @@ void truncate_inode_pages_range(struct address_space *mapping,
> >  	same_folio = (lstart >> PAGE_SHIFT) == (lend >> PAGE_SHIFT);
> >  	folio = __filemap_get_folio(mapping, lstart >> PAGE_SHIFT, FGP_LOCK, 0);
> >  	if (!IS_ERR(folio)) {
> > -		same_folio = lend < folio_pos(folio) + folio_size(folio);
> > +		same_folio = lend < folio_next_pos(folio);
> 
> This causes an intermittent failure with XFS.  Two reports here:
> https://lore.kernel.org/linux-xfs/aQohjfEFmU8lef6M@casper.infradead.org/
> 
> This is a fun one.  The "fix" I'm running with right now is:
> 
> -               same_folio = lend < folio_next_pos(folio);
> +               same_folio = lend < (u64)folio_next_pos(folio);
> 
> folio_pos() and folio_next_pos() return an loff_t.  folio_size() returns
> a size_t.  So folio_pos() + folio_size() is unsigned (by the usual C
> promotion rules).  Before this patch, this was an unsigned comparison
> against lend, and with the patch it's now a signed comparison.  Since
> lend can be -1 (to mean 'end of file'), same_folio will now be 'true'
> when it used to be 'false'.
> 
> Funnily, on 32-bit systems, size_t is u32 and loff_t is s64, so their
> addition is also s64.  That means this has been wrong on 32-bit systems
> for ... a while.  And nobody noticed, so I guess nobody's testing 32-bit
> all that hard.

death by slow degradation...

> Anyway, what's the *right* way to fix this?  Cast to (unsigned long

Naively, I would think it should be an unsigned comparison for u64.

> long)?  There's an ssize_t, but I really want the opposite, a uloff_t.

Fwiw, I don't think anything is stopping us from defining uloff_t in the
kernel if it's helpful.

> (now i'm going to go back through all the other patches in this series
> and see if I've made the same mistake anywhere else)