Re: [linux-next:master] BUILD REGRESSION 8cb8311e95e3bb58bd84d6350365f14a718faa6d

[Matthew Wilcox <willy@infradead.org>]

On Wed, May 25, 2022 at 03:20:06PM -0700, Andrew Morton wrote:
> On Wed, 25 May 2022 23:07:35 +0100 Jessica Clarke <jrtc27@jrtc27.com> wrote:
> 
> > This is i386, so an unsigned long is 32-bit, but i_blocks is a blkcnt_t
> > i.e. a u64, which makes the shift without a cast of the LHS fishy.
> 
> Ah, of course, thanks.  I remember 32 bits ;)
> 
> --- a/mm/shmem.c~mm-shmemc-suppress-shift-warning
> +++ a/mm/shmem.c
> @@ -1945,7 +1945,7 @@ alloc_nohuge:
>  
>  	spin_lock_irq(&info->lock);
>  	info->alloced += folio_nr_pages(folio);
> -	inode->i_blocks += BLOCKS_PER_PAGE << folio_order(folio);
> +	inode->i_blocks += (blkcnt_t)BLOCKS_PER_PAGE << folio_order(folio);

Bizarre this started showing up now.  The recent patch was:

-       info->alloced += compound_nr(page);
-       inode->i_blocks += BLOCKS_PER_PAGE << compound_order(page);
+       info->alloced += folio_nr_pages(folio);
+       inode->i_blocks += BLOCKS_PER_PAGE << folio_order(folio);

so it could tell that compound_order() was small, but folio_order()
might be large?

Silencing the warning is a good thing, but folio_order() can (at the
moment) be at most 9 on i386, so it isn't actually going to be
larger than 4096.