Re: is dma_mapping_error() check necessary for dma_alloc_noncoherent()?

On 26.08.2025 11:44, Baochen Qiang wrote:
> Hi guys,
>
> I have a driver which allocate noncoherent DMA buffer and get the returned CPU addr tested:
>
> 	vaddr_unaligned = dma_alloc_noncoherent(ab->dev, rx_tid->unaligned_size, &paddr,
> 						DMA_BIDIRECTIONAL, GFP_ATOMIC);
> 	if (!vaddr_unaligned) {
> 		spin_unlock_bh(&ab->base_lock);
> 		return -ENOMEM;
> 	}
>
> while free the buffer
>
> 	dma_free_noncoherent(ab->dev, rx_tid->unaligned_size,
> 			     rx_tid->vaddr_unaligned,
> 			     rx_tid->paddr_unaligned, DMA_BIDIRECTIONAL);
>
> I get below warnings:
>
>   DMA-API: ath11k_pci 0000:03:00.0: device driver failed to check map error[device
> address=0x00000000f3ad7000] [size=639 bytes] [mapped as single]
>   WARNING: CPU: 15 PID: 64303 at kernel/dma/debug.c:1036 check_unmap+0x7e2/0x950
>   RIP: 0010:check_unmap+0x7e2/0x950
>   Call Trace:
>    <TASK>
>    ? free_to_partial_list+0x9d/0x350
>    debug_dma_unmap_page+0xac/0xc0
>    ? debug_smp_processor_id+0x17/0x20
>    ? rcu_is_watching+0x13/0x70
>    dma_free_pages+0x56/0x180
>    [...]
>    </TASK>
>   ---[ end trace 0000000000000000 ]---
>   DMA-API: Mapped at:
>    debug_dma_map_page+0x7c/0x140
>    dma_alloc_pages+0x74/0x220
>    [...]
>
> Checking code gives me the impression that I should do dma_mapping_error() check as well.
> And indeed with below diff the warning is gone:
>
> +       dma_mapping_error(ab->dev, paddr);
>
> However this does not make sense to me since IMO testing the CPU address is good enough, I
> can not imagine a valid case where DMA alloc/map fails while returning a valid CPU
> address, no?
>
> If I was right, should we remove invocation to debug_dma_map_page() in dma_alloc_pages()?

Simply replace "if (!vaddr_unaligned)" with "if 
dma_mapping_error(ab->dev, paddr)" and the debug code will be happy.

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