Re: [PATCH mm-unstable] mm: clarify folio_set_compound_order() zero support

[John Hubbard <jhubbard@nvidia.com>]

On 12/9/22 06:27, Muchun Song wrote:
>  From you advise, I think we can remove VM_BUG_ON and handle non-zero
> order page, something like:

Yes, and thanks for summarizing all the individual feedback into a
proposed solution.

If we go this route, then I'd suggest a little note above the function,
such as:

/*
  * For non-large folios, this will have no effect, other than possibly
  * generating a warning, if the caller attempts to set a non-zero folio order
  * for a non-large folio.
  */

> static inline void folio_set_order(struct folio *folio,
> 		                   unsigned int order)
> {
> 	if (!folio_test_large(folio)) {
> 		WARN_ON(order);

Better make that a WARN_ON_ONCE(), to avoid taking the machine down
with excessive warnings in the log.

> 		return;
> 	}
> 
> 	folio->_folio_order = order;
> #ifdef CONFIG_64BIT
> 	folio->_folio_nr_pages = order ? 1U << order : 0;
> #endif
> }
> 
> In this case,
> 
>    1) we can handle both non-zero and zero (folio_order() works as well
>       for this case) order page.
>    2) it can prevent OOB for non-large folio and warn unexpected users.
>    3) Do not BUG.
>    4) No need to rename folio_set_order.
> 
> What do you think?

If the new behavior is OK with everyone, it seems good to me.

thanks,
-- 
John Hubbard
NVIDIA