next-20250627: IOMMU DMA warning during NVMe I/O completion after 06cae0e3f61c

[Keith Busch]

On Thu, Jul 03, 2025 at 11:30:42AM +0200, Christoph Hellwig wrote:
> I think the idea to reconstruct the dma addresses from PRPs should
> be considered a failure by now.  It works fine for SGLs, but for
> PRPs we're better off just stashing them away.  Bob, can you try

s/Bob/Ben

> something like the patch below?  To be fully safe it needs a mempool,
> and it could use some cleanups, but it does pass testing on my setup
> here, so I'd love to see if if fixes your issue.

Thanks for confirming.

While this is starting to look a bit messy, I believe it's still an
overall win: you've cut down the vector walking in the setup path from 3
to 1, which reduces a non-trivial amount of overhead for even moderately
sized IO.