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

[Dan Carpenter <dan.carpenter@oracle.com>]

On Thu, May 26, 2022 at 03:28:25PM +0100, Matthew Wilcox wrote:
> On Thu, May 26, 2022 at 11:48:32AM +0300, Dan Carpenter wrote:
> > On Thu, May 26, 2022 at 02:16:34AM +0100, Matthew Wilcox wrote:
> > > 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?
> > 
> > The old code also generates a warning on my test system.  Smatch thinks
> > both compound_order() and folio_order() are 0-255.  I guess because of
> > the "unsigned char compound_order;" in the struct page.
> 
> It'd be nice if we could annotate that as "contains a value between
> 1 and BITS_PER_LONG - PAGE_SHIFT".  Then be able to optionally enable
> a checker that ensures that's true on loads/stores.  Maybe we need a
> language that isn't C :-P  Ada can do this ... I don't think Rust can.

Machine Parsable Comments.  It's a matter of figuring out the best
format and writing the code.

In Smatch, I have table of hard coded return values in the format:
<function> <old return> <new hard coded return>
https://github.com/error27/smatch/blob/master/smatch_data/db/kernel.return_fixes
I don't have code to handle something like BITS_PER_LONG or PAGE_SHIFT.
To be honest, Smatch code always assumes that PAGE_SIZE is 4096 but I
should actually look it up...  It's not impossible to do.  The GFP_KERNEL
values changed enough so that I eventually made that look up the actual
defines.

I also have a table in the database where I could edit the values of
(struct page)->compound_order.

regards,
dan carpenter