Re: [PATCH] btrfs: fix signedness issue in min()

[David Laight]

On Mon, 17 Mar 2025 20:26:39 +0100
David Sterba <dsterba@xxxxxxx> wrote:

> On Mon, Mar 17, 2025 at 02:16:37PM +0000, David Laight wrote:
> > On Fri, 14 Mar 2025 16:54:41 +0100
> > Arnd Bergmann <arnd@xxxxxxxxxx> wrote:
> >   
> > > From: Arnd Bergmann <arnd@xxxxxxxx>
> > > 
> > > Comparing a u64 to an loff_t causes a warning in min()
> > > 
> > > fs/btrfs/extent_io.c: In function 'extent_write_locked_range':
> > > include/linux/compiler_types.h:557:45: error: call to '__compiletime_assert_588' declared with attribute error: min(folio_pos(folio) + folio_size(folio) - 1, end) signedness error
> > > fs/btrfs/extent_io.c:2472:27: note: in expansion of macro 'min'
> > >  2472 |                 cur_end = min(folio_pos(folio) + folio_size(folio) - 1, end);
> > >       |                           ^~~
> > > 
> > > Use min_t() instead.  
> > 
> > It would be slightly better to use min_unsigned() since, regardless of the types
> > involved, it can't discard significant bits.
> > 
> > OTOH the real problem here is that both folio_pos() and folio_size() return signed types.  
> 
> folio_size() returns size_t:
> 
> static inline size_t folio_size(const struct folio *folio)
> {
> 	return PAGE_SIZE << folio_order(folio);
> }
> 
> Otherwise the min_t with force u64 is ok and lots of min() use (in
> btrfs) was converted to the typed variant in case the types don't match.

That is just broken.
min_t(u64, x, y) is just min((u64)x, (u64)y) and you wouldn't do the
same casts anywhere else unless you really had to.
So you really shouldn't use min_t() unless there is no other way around the problem.

Ok (u64) are unlikely to be a problem, but there are plenty of places where
(u8) get used and can (and actually has) discard significant bits and cause bugs.

	David