Re: [PATCH v2] dma-debug: suppress cacheline overlap warning when arch has no DMA alignment requirement

[Robin Murphy <robin.murphy@arm.com>]

On 2026-03-30 8:44 am, Marek Szyprowski wrote:
> On 27.03.2026 13:41, 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.
>>
>> This has been reported repeatedly since v5.14 (when the EEXIST check
>> was added) across various USB host controllers and devices including
>> xhci_hcd with USB hubs, USB audio devices, and USB ethernet adapters.
>>
>> The cacheline overlap is only a real concern on architectures that
>> require DMA buffer alignment to cacheline boundaries (i.e. where
>> ARCH_DMA_MINALIGN >= L1_CACHE_BYTES).  On architectures like x86_64
>> where dma_get_cache_alignment() returns 1, the hardware is
>> cache-coherent and overlapping cacheline mappings are harmless.
>>
>> Suppress the EEXIST warning when dma_get_cache_alignment() is less
>> than L1_CACHE_BYTES, indicating the architecture does not require
>> cacheline-aligned DMA buffers.

Really the value of this check is for mappings of structure members or 
array elements which have no inherent guarantee of being individually 
aligned to ARCH_DMA_MINALIGN, and while x86 can also get away with that, 
it represents a genuine issue for non-coherent architectures. USB's 
mapping of a dedicated tiny allocation is, if anything, rather the 
special case.

TBH I'd be inclined to have CONFIG_DMA_DEBUG raise ARCH_DMA_MINALIGN as 
appropriate such that genuine false-positives can't happen, rather than 
effectively defeat the whole check, but I suppose that might carry a 
risk of tripping up arch code that doesn't expect it... Oh well.

Thanks,
Robin.

>> Verified with a kernel module reproducer that performs two kmalloc(8)
>> allocations back-to-back and DMA-maps both:
>>
>>    Before: allocations share a cacheline, EEXIST fires within ~50 pairs
>>    After:  same cacheline pair found, but no warning emitted
>>
>> Fixes: 2b4bbc6231d7 ("dma-debug: report -EEXIST errors in add_dma_entry")
>> Link: https://bugzilla.kernel.org/show_bug.cgi?id=215740
>> Suggested-by: Harry Yoo <harry@kernel.org>
>> Tested-by: Mikhail Gavrilov <mikhail.v.gavrilov@gmail.com>
>> Signed-off-by: Mikhail Gavrilov <mikhail.v.gavrilov@gmail.com>
> 
> Applied to dma-mapping-fixes. Thanks!
> 
>> ---
>>
>> v1 -> v2:
>>    - Moved fix from include/linux/slab.h (ARCH_KMALLOC_MINALIGN)
>>      to kernel/dma/debug.c per Harry Yoo's suggestion.
>>    - Instead of forcing cacheline-aligned allocations, suppress
>>      the warning when the architecture has no DMA alignment
>>      requirement (dma_get_cache_alignment() < L1_CACHE_BYTES).
>>
>> v1: https://lore.kernel.org/all/20260327055846.248829-1-mikhail.v.gavrilov@gmail.com/
>>
>> Reproducer module that triggers the bug reliably:
>>    https://bugzilla.kernel.org/attachment.cgi?id=309769
>>
>>   kernel/dma/debug.c | 1 +
>>   1 file changed, 1 insertion(+)
>>
>> diff --git a/kernel/dma/debug.c b/kernel/dma/debug.c
>> index 0677918f06a8..1a725edbbbf6 100644
>> --- a/kernel/dma/debug.c
>> +++ b/kernel/dma/debug.c
>> @@ -615,6 +615,7 @@ static void add_dma_entry(struct dma_debug_entry *entry, unsigned long attrs)
>>   	} else if (rc == -EEXIST &&
>>   		   !(attrs & DMA_ATTR_SKIP_CPU_SYNC) &&
>>   		   !(entry->is_cache_clean && overlap_cache_clean) &&
>> +		   dma_get_cache_alignment() >= L1_CACHE_BYTES &&
>>   		   !(IS_ENABLED(CONFIG_DMA_BOUNCE_UNALIGNED_KMALLOC) &&
>>   		     is_swiotlb_active(entry->dev))) {
>>   		err_printk(entry->dev, entry,
> 
> Best regards