Re: [PATCH] libfs: fix accidental overflow in offset calculation

[Al Viro <viro@zeniv.linux.org.uk>]

On Fri, May 10, 2024 at 01:49:06AM +0100, Al Viro wrote:
> On Fri, May 10, 2024 at 12:35:51AM +0000, Justin Stitt wrote:
> > @@ -147,7 +147,9 @@ loff_t dcache_dir_lseek(struct file *file, loff_t offset, int whence)
> >  	struct dentry *dentry = file->f_path.dentry;
> >  	switch (whence) {
> >  		case 1:
> > -			offset += file->f_pos;
> > +			/* cannot represent offset with loff_t */
> > +			if (check_add_overflow(offset, file->f_pos, &offset))
> > +				return -EOVERFLOW;
> 
> Instead of -EINVAL it correctly returns in such cases?  Why?

We have file->f_pos in range 0..LLONG_MAX.  We are adding a value in
range LLONG_MIN..LLONG_MAX.  The sum (in $\Bbb Z$) cannot be below
LLONG_MIN or above 2 * LLONG_MAX, so if it can't be represented by loff_t,
it must have been in range LLONG_MAX + 1 .. 2 * LLONG_MAX.  Result of
wraparound would be equal to that sum - 2 * LLONG_MAX - 2, which is going
to be in no greater than -2.  We will run
			fallthrough;
		case 0:
			if (offset >= 0)
				break;
			fallthrough;
		default:
			return -EINVAL;
and fail with -EINVAL.

Could you explain why would -EOVERFLOW be preferable and why should we
engage in that bit of cargo cult?