Re: hugetlb BUILD REGRESSION in linux-next20221121

On 11/28/22 11:40 PM, Rasmus Villemoes wrote:
> On 29/11/2022 05.57, Dan Carpenter wrote:
>> On Mon, Nov 28, 2022 at 02:34:54PM -0800, Sidhartha Kumar wrote:
>>> Hello,
>>>
>>> One of my patches in linux-next was flagged as a Unverified Error/Warning
>>> with the following warning[1]:
>>>
>>> mm/hugetlb.c:2073 alloc_pool_huge_page() error: uninitialized symbol
>>> 'folio'.
>>>
>>> The relevant change is:
>>>
>>> -    struct page *page;
>>> +    struct folio *folio;
>>>       int nr_nodes, node;
>>>       gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE;
>>>
>>>       for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
>>> -        page = alloc_fresh_huge_page(h, gfp_mask, node, nodes_allowed,
>>> -                        node_alloc_noretry);
>>> -        if (page)
>>> +        folio = alloc_fresh_hugetlb_folio(h, gfp_mask, node,
>>> +                    nodes_allowed, node_alloc_noretry);
>>> +        if (folio)
>>>               break;
>>>       }
>>>
>>> -    if (!page)
>>> +    if (!folio)
>>>
>>> It looks like I can initialize folio to NULL to avoid this error but I'm not
>>> sure how this would cause a regression as previously the page variable was
>>> unitialized as well. Please let me know if I am missing something in my
>>> patch or if this should be ignored.
>> Both the original and the new code trigger a Smatch warning.  I don't
>> know why the kbuild-bot marks this as a new warning.  Possibly that's
>> because the variable name has changed?  The kbuild-bot is run by Intel.
>>
>> The problem is obviously that Smatch doesn't know that we always enter
>> the loop.  There are hack arounds that I could do for this, but
>> sometimes we dont' actually enter the loop so changing this will hide
>> bugs...  I am conflicted on this.
> It seems that a reasonably clean way to deal with this, since there's
> not a lot of code following the loop, is to simply move that code into
> the "if (page/folio)" branch. I.e. change the tail to
>
>
>     if (page) {
>       free_huge_page(page); /* free it into the hugepage allocator */
>       return 1;
>     }
>   } // this is the end of loop
>   return 0;
>
> or however it looks after the above change.

Thanks for the suggestion, this looks like a good way to avoid the 
smatch warning and I'll add it to the next version of my patch.

Thanks.

Sidhartha Kumar


> Rasmus
>