Re: [PATCH] mm/slab: improve kmem_cache_alloc_bulk

On 6/1/26 13:38, Christoph Hellwig wrote:
> On Mon, Jun 01, 2026 at 10:16:30AM +0200, Vlastimil Babka (SUSE) wrote:
>> > kmem_cache_alloc_bulk() returning 0 was considered a success in that case.
>> > 
>> > Either fixing kmem_cache_alloc_bulk() (and the comment) or fixing the
>> > user sounds fine to me.
>> 
>> Would it be wrong if we just returned true for size of 0? Would something
>> else break?
> 
> I don't think it is wrong per se, but it feels like the wrong kind of
> API.  I.e. I don't think the MSM caller actually wants this, as they'd
> also do a zero-sized kvmalloc.

If p->count is 0 then indeed there's a zero-sized kvmalloc so p->pages ==
ZERO_SIZE_PTR but then nothing breaks because nothing tries to dereference it?

msm_iommu_pagetable_prealloc_cleanup() has a "if (p->count > 0)" branch so
it seems it's considered possible. But then the rest of the functions also
seems working fine, i.e. kmem_cache_free_bulk() of zero size does nothing,
kvfree() of ZERO_SIZE_PTR does nothing.

It seems to me kmem_cache_alloc_bulk() returning true for size == 0 fits
naturally in this world and is less likely to result in a gotcha?