Re: [PATCH v2 3/4] mm: hugetlb: fix GFP mask propagation to CMA

[Qi Zheng]

On 11/6/26 1:50 PM, Oscar Salvador (SUSE) wrote:
> On Fri, Nov 06, 2026 at 10:43:53AM +0800, Qi Zheng wrote:
> > I'd still prefer to keep __GFP_COMP so that
> > cma_alloc_frozen_compound() itself guarantees compound allocation,
> > without requiring callers to supply the flag.
>
> Well, but aren't all callers hugetlb? And we already do supply __GFP_COMP there
> for evident reasons.
> And should that not be the case, and you want to catch someone who went rogue,
> I would just place a WARN_ON_ONCE to catch it here, so we know who is not following
> the rules.
>
> Makes sense?

Ah, make sense.

How about this:

diff --git a/mm/cma.c b/mm/cma.c
index 876635a56180e..45017ed876f76 100644
--- a/mm/cma.c
+++ b/mm/cma.c
@@ -935,7 +935,8 @@ struct page *cma_alloc_frozen(struct cma *cma, 
unsigned long count,

 struct page *cma_alloc_frozen_compound(struct cma *cma, unsigned int 
order, gfp_t gfp)
 {
-       gfp |= __GFP_COMP | __GFP_NOWARN;
+       if (WARN_ON_ONCE(!(gfp & __GFP_COMP)))
+               return NULL;

        return __cma_alloc_frozen(cma, 1 << order, order, gfp);
 }

>