[PATCH V1] nvme-pci: Fix NULL pointer dereference in nvme_pci_prp_iter_next

[Pradeep Pragallapati]

On 2/3/2026 7:35 PM, Pradeep Pragallapati wrote:
> 
> 
> On 2/3/2026 10:57 AM, Christoph Hellwig wrote:
>> On Mon, Feb 02, 2026 at 11:59:04AM -0700, Keith Busch wrote:
>>> In the case where this iteration caused dma_need_unmap() to toggle to
>>> true, this is the iteration that allocates the dma_vecs, and it
>>> initializes the first entry to this iter. But the next lines proceed to
>>> the save this iter in the next index, so it's doubly accounted for and
>>> will get unmapped twice in the completion.
>>
>> Yeah.
>>
>>> Also, if the allocation fails, we should set iter->status to
>>> BLK_STS_RESOURCE so the callers know why the iteration can't continue.
>>> Otherwise, the caller will think the request is badly formed if you
>>> return false from here without setting iter->status.
>>>
>>> Here's my quick take. Boot tested with swiotlb enabled, but haven't
>>> tried to test the changing dma_need_unmap() scenario.
>>
>> Looks much better.  Cosmetic nits below.
>>
>> Pradeep, can you test this with your setup?
> Sure, testing has started, and I will share the findings soon.
> Also, I did not pick up the initialization of dma_vecs during testing.

I ran testing for over 20 hours and did not observe the issue on my 
setup. It appears to be helping.
> 
>>
>>> +    if (!dma_use_iova(&iod->dma_state) && dma_need_unmap(dma_dev))
>>> +        return nvme_pci_prp_save_mapping(iter, req);
>>
>>> +    if (!dma_use_iova(&iod->dma_state) && dma_need_unmap(nvmeq->dev- 
>>> >dev))
>>> +        if (!nvme_pci_prp_save_mapping(iter, req))
>>> +            return iter->status;
>>
>> I'd move the dma_use_iova / dma_need_unmap checks into
>> nvme_pci_prp_save_mapping to simplify this a bit more.
>>
>>>       /*
>>>        * PRP1 always points to the start of the DMA transfers.
>>> @@ -1218,6 +1231,8 @@ static blk_status_t nvme_prep_rq(struct request 
>>> *req)
>>>       iod->nr_descriptors = 0;
>>>       iod->total_len = 0;
>>>       iod->meta_total_len = 0;
>>> +    iod->nr_dma_vecs = 0;
>>> +    iod->dma_vecs = NULL;
>>
>> I don't think we need the dma_vecs initialization here, as everything
>> is keyed off nr_dma_vecs.
>