[PATCH v5 3/3] nvme/pci: make PRP list DMA pools per-NUMA-node

[Keith Busch]

On Fri, Apr 25, 2025 at 03:21:11PM +0200, Christoph Hellwig wrote:
> On Thu, Apr 24, 2025 at 09:40:18AM -0600, Keith Busch wrote:
> > The dmapool allocates dma coherent memory, and it's mapped for the
> > remainder of lifetime of the pool. Allocating slab memory and dma
> > mapping per-io would be pretty costly in comparison, I think.
> 
> True.  Although we don't even need dma coherent memory, a single
> cache writeback after writing the PRPs/SGLs would probably be more
> efficient on not cache coherent platforms.  But no one really cares
> about performance on those anyway..

Sure, but it's not just about non-coherent platform performance
concerns. Allocations out of the dma pool are iommu mapped if necessary
too. We frequently allocate and free these lists, and the dmapool makes
it quick and easy to reuse previously mapped memory.