Re: [PATCH] dma: swiotlb: Round restricted DMA pool size down to IO_TLB_SEGSIZE

[Marek Szyprowski]

On 03.11.2026 22:18, Michael Kelley wrote:
> From: Marek Szyprowski <m.szyprowski@xxxxxxxxxxx> Sent: Tuesday, November 3, 2026 7:50 AM
>> All other paths creating a swiotlb pool ensure that nslabs is a multiple
>> of IO_TLB_SEGSIZE, and the swiotlb code relies on that. Since commit
>> 0b84e4f8b793 ("swiotlb: Add restricted DMA pool initialization") a
>> restricted pool can have an arbitrary size taken from the device tree.
>>
>> If the pool size is not a multiple of IO_TLB_SEGSIZE, the last segment is
>> partial and swiotlb_release_slots() reads mem->slots[index + nslabs],
>> which is one past the end of the slots array when the freed buffer ends
>> at the end of the pool:
>>
>>  BUG: KASAN: slab-out-of-bounds in swiotlb_release_slots
>>  Read of size 2
>>
>> Round the pool size down to a multiple of IO_TLB_SEGSIZE, wasting a bit
>> of memory, and warn about it.
> I'm not familiar with the macro scenarios where a device has a
> restricted pool specified in DT. But I wonder if it would be better to
> round up instead of down. DT has specified that the device should
> have a restricted DMA pool of a certain size, and rounding down gives
> the device a smaller pool than was requested. Could DMA operations
> then fail because of not having a large enough pool? Rounding down
> also allows the pool to end up being size zero, which this patch
> detects and results in a definite failure at device initialization time.
>
> What's the tradeoff in rounding up instead?The device restricted
> pool at least as large as requested. Presumably, being larger than
> requested doesn't cause any problems for the device or its driver.
> The downside is using up to 256 KiB more system memory than
> expected. In an embedded system with only a few MiB of system
> memory, maybe that's a relatively big problem. I don't know the
> scenarios well enough to judge whether using a bit more system
> memory is better or worse than the device not having all the
> restricted pool space it said it needed for DMA.

The main problem with enlarging the restricted pool is that the memory
has been already allocated (reserved) by the generic of_reserved_mem
code before the rmem_swiotlb_device_init() is called. There might be
even another reserved region just after the one being processed by
rmem_swiotlb_device_init(). The reserved region might be specified in
the device tree exactly, by pointing its base address and size. In
such case the rmem_swiotlb_device_init() really cannot do anything
else than using the provided already allocated region partially.
The only thing that might be improved is freeing the 'wasted' part
of the region back to the memory allocator if You think that
printing a warning is not enough.



>
> Your commit message also mentions "wasting a bit of memory".
> What is wasted in the round down case? The rounding down is
> done before the pool memory is allocated. I would have thought
> the wasting would occur in the round up case, where the restricted
> pool is given more memory than the device said it needed.
>
> Michael
>
>> Reported-by: Giorgi Kobakhia <gkobakhi@xxxxxxx>
>> Closes: https://lore.kernel.org/all/20261026032007.1957949-1-gkobakhi@xxxxxxx/
>> Fixes: 0b84e4f8b793 ("swiotlb: Add restricted DMA pool initialization")
>> CC: stable@xxxxxxxxxxxxxxx
>> Signed-off-by: Marek Szyprowski <m.szyprowski@xxxxxxxxxxx>
>> ---
>>  kernel/dma/swiotlb.c | 15 ++++++++++++++-
>>  1 file changed, 14 insertions(+), 1 deletion(-)
>>
>> diff --git a/kernel/dma/swiotlb.c b/kernel/dma/swiotlb.c
>> index 2c57b49ac7c0..a68f35192582 100644
>> --- a/kernel/dma/swiotlb.c
>> +++ b/kernel/dma/swiotlb.c
>> @@ -2007,6 +2007,19 @@ static int rmem_swiotlb_device_init(struct reserved_mem *rmem,
>>  	 */
>>  	if (!mem) {
>>  		struct io_tlb_pool *pool;
>> +		unsigned long aligned_nslabs = ALIGN_DOWN(nslabs, IO_TLB_SEGSIZE);
>> +
>> +		if (aligned_nslabs != nslabs) {
>> +			dev_warn(dev, "Restricted DMA pool size %zu is not a multiple of %lu, rounding down to %lu bytes\n",
>> +				 (size_t)rmem->size,
>> +				 (unsigned long)IO_TLB_SEGSIZE << IO_TLB_SHIFT,
>> +				 aligned_nslabs << IO_TLB_SHIFT);
>> +			nslabs = aligned_nslabs;
>> +		}
>> +		if (!nslabs) {
>> +			dev_err(dev, "Restricted DMA pool is too small\n");
>> +			return -EINVAL;
>> +		}
>>
>>  		mem = kzalloc_obj(*mem);
>>  		if (!mem)
>> @@ -2034,7 +2047,7 @@ static int rmem_swiotlb_device_init(struct reserved_mem *rmem,
>>
>>  			mem->cc_shared = true;
>>  			ret = set_memory_decrypted((unsigned long)phys_to_virt(rmem->base),
>> -						   rmem->size >> PAGE_SHIFT);
>> +						   (nslabs << IO_TLB_SHIFT) >> PAGE_SHIFT);
>>  			if (ret) {
>>  				dev_err(dev, "Failed to decrypt restricted DMA pool\n");
>>  				kfree(pool->areas);
>> --
>> 2.43.0
>
Best regards
-- 
Marek Szyprowski, PhD
Samsung R&D Institute Poland