Re: [PATCH] swiotlb: fix out-of-bounds read in swiotlb_release_slots()

[Marek Szyprowski]

On 29.10.2026 18:39, Michael Kelley wrote:
> From: Giorgi Kobakhia <gkobakhi@xxxxxxx> Sent: Sunday, October 25, 2026 8:20 PM
>> When buffer is freed, swiotlb_release_slots() tries to merge it with the
>> following free space by reading the list value of the next slot.
>>
>>     if (index + nslots < ALIGN(index + 1, IO_TLB_SEGSIZE))
>>         count = mem->slots[index + nslots].list;
>>     else
>>         count = 0;
>>
>> If index is in the last segment of the pool and the pool size is not
>> multiple of IO_TLB_SEGSIZE, index + nslots becomes equal to mem->nslabs
>> leading to out of bound read.
>>
>>  BUG: KASAN: slab-out-of-bounds in swiotlb_release_slots
>>  Read of size 2
>>
>>  Call trace:
>>   swiotlb_release_slots (kernel/dma/swiotlb.c:1653)
>>   __swiotlb_tbl_unmap_single (kernel/dma/swiotlb.c:1734)
>>   dma_unmap_phys (./include/linux/swiotlb.h:262 kernel/dma/direct.h:102
>>                     kernel/dma/mapping.c:218)
>>   dma_unmap_page_attrs (kernel/dma/mapping.c:236)
>>
>> The commit 0b84e4f8b793 ("swiotlb: Add restricted DMA pool
>> initialization") makes it possible to initialize a pool with an
>> arbitrary size. Before it nslabs was multiple of IO_TLB_SEGSIZE and no
>> OOB occurred.
>>
>> Add overflow check.
> I can see that this is a legitimate problem. But I'm not sure this is the right
> solution.
>
> All other paths for creating a swiotlb pool ensure that nslabs is a multiple
> of IO_TLB_SEGSIZE. The better solution may be for rmem_swiotlb_device_init()
> to round-up nslabs to be a multiple of IO_TLB_SEGSIZE. My concern is
> that there may be other places in the swiotlb code where there is an
> implicit assumption that there are no partial segments at the end of
> a pool. It would take some effort for me to re-immerse myself in the
> swiotlb code to see if that's true, and I haven't done that. But my
> guess is that there's a good chance of such an assumption.
>
> What are the implications of potentially allocating slightly more memory
> for a restricted pool than is specified? It wouldn't be much additional
> memory since the segment size is 256 KiB.
Frankly speaking it would much easier to round the pool size down to match
the multiple of IO_TLB_SEGSIZE, waste some memory and issue a warning.
It is not that trivial to allocate "slightly more memory" if the size of
the restricted pool is statically specified in device tree.

Best regards
-- 
Marek Szyprowski, PhD
Samsung R&D Institute Poland