Re: [PATCH] mm/slab: align kmalloc to cacheline when DMA API debugging is active

["Harry Yoo (Oracle)" <harry@kernel.org>]

On Fri, Mar 27, 2026 at 11:50:07AM +0500, Mikhail Gavrilov wrote:
> On Fri, Mar 27, 2026 at 11:38 AM Harry Yoo (Oracle) <harry@kernel.org> wrote:
> >
> > On Fri, Mar 27, 2026 at 10:58:46AM +0500, Mikhail Gavrilov wrote:
> > > When CONFIG_DMA_API_DEBUG is enabled, the DMA debug infrastructure
> > > tracks active mappings per cacheline and warns if two different DMA
> > > mappings share the same cacheline ("cacheline tracking EEXIST,
> > > overlapping mappings aren't supported").
> > >
> > > On x86_64, ARCH_KMALLOC_MINALIGN defaults to 8, so small kmalloc
> > > allocations (e.g. the 8-byte hub->buffer and hub->status in the USB
> > > hub driver) frequently land in the same 64-byte cacheline.  When both
> > > are DMA-mapped, this triggers a false positive warning.
> >
> > Is it feasible to suppress the warning if dma_get_cache_alignment() is
> > smaller than L1_CACHE_BYTES?
> 
> Hi Harry,

Hi Mikhail,

Please keep in mind that I have limited understanding of DMA API,
but just wanted to double check if there is (or isn't) a sane way to
fix it on dma-debug side.

> Good question. I considered the dma-debug side, but the issue is
> that the cacheline overlap check in add_dma_entry() is intentionally
> strict -- it catches real bugs on non-coherent architectures where
> two DMA buffers sharing a cacheline can corrupt data.

But dma_get_cache_alignment() < L1_CACHE_BYTES means the architecture
actually allows overlapping cachelines, no?

A non-coherent architecture where two DMA buffers sharing a cacheline
could corrupt data should define ARCH_DMA_MINALIGN >= L1_CACHE_BYTES.

I'm not sure what kind of a real bug this will hide,
or am I missing something?

> Alan Stern discussed this in the bugzilla [1] and concluded that
> the slab alignment approach "seems reasonable" [2],

As long as there's no good alternative way to fix, yeah.

> noting that turning on debugging should not affect the way the kernel
> behaves -- otherwise what you're debugging isn't the same as what normally
> happens.

Yeah, this is why I'm trying to double check if there's no feasible
alternative.

> But given the way the DMA API debugging is set up, I don't
> see any alternative."

I'm trying to say adding the (dma_get_cache_alignment() <
L1_CACHE_BYTES) check might be considered as an alternative ;)
 
> [1] https://bugzilla.kernel.org/show_bug.cgi?id=215740#c31
> [2] https://bugzilla.kernel.org/show_bug.cgi?id=215740#c44

-- 
Cheers,
Harry / Hyeonggon