Re: [PATCH V1] nvme-pci: Fix NULL pointer dereference in nvme_pci_prp_iter_next
From: Robin Murphy
Date: Mon Feb 02 2026 - 10:20:48 EST
On 2026-02-02 2:35 pm, Christoph Hellwig wrote:
On Mon, Feb 02, 2026 at 06:27:38PM +0530, Pradeep P V K wrote:
Fix a NULL pointer dereference that occurs in nvme_pci_prp_iter_next()
when SWIOTLB bounce buffering becomes active during runtime.
The issue occurs when SWIOTLB activation changes the device's DMA
mapping requirements at runtime,
creating a mismatch between
iod->dma_vecs allocation and access logic.
The problem manifests when:
1. Device initially operates with dma_skip_sync=true
(coherent DMA assumed)
2. First SWIOTLB mapping occurs due to DMA address limitations,
memory encryption, or IOMMU bounce buffering requirements
3. SWIOTLB calls dma_reset_need_sync(), permanently setting
dma_skip_sync=false
4. Subsequent I/Os now have dma_need_unmap()=true, requiring
iod->dma_vecs
I think this patch just papers over the bug. If dma_need_unmap
can't be trusted before the dma_map_* call, we've not saved
the unmap information and the unmap won't work properly.
The dma_need_unmap() kerneldoc says:
"This function must be called after all mappings that might
need to be unmapped have been performed."
Trying to infer anything from it beforehand is definitely a bug in the caller.
So we'll need to extend the core code to tell if a mapping
will set dma_skip_sync=false before doing the mapping.
I don't see that being possible - at best we could reasonably infer that a fully-coherent system with no sync ops, no SWIOTLB and no DMA_DEBUG shouldn't ever set it to true, but as for the other way round, by the time you've run through all the SWIOTLB logic to guess whether a particular mapping would be bounced or not, you've basically performed the mapping anyway. Thus at best, such an API to potentially do a whole dry-run mapping before every actual mapping would seem like a pretty pointless anti-optimisation.
Thanks,
Robin.