On Wed, 2020-07-08 at 17:35 +0200, Christoph Hellwig wrote:
> On Tue, Jul 07, 2020 at 02:28:04PM +0200, Nicolas Saenz Julienne wrote:
> > When allocating atomic DMA memory for a device, the dma-pool core
> > queries __dma_direct_optimal_gfp_mask() to check which atomic pool to
> > use. It turns out the GFP flag returned is only an optimistic guess.
> > The pool selected might sometimes live in a zone higher than the
> > device's view of memory.
> > 
> > As there isn't a way to grantee a mapping between a device's DMA
> > constraints and correct GFP flags this unifies both DMA atomic pools.
> > The resulting pool is allocated in the lower DMA zone available, if any,
> > so as for devices to always get accessible memory while having the
> > flexibility of using dma_pool_kernel for the non constrained ones.
> > 
> > Fixes: c84dc6e68a1d ("dma-pool: add additional coherent pools to map to gfp
> > mask")
> > Reported-by: Jeremy Linton <jeremy.linton@arm.com>
> > Suggested-by: Robin Murphy <robin.murphy@arm.com>
> > Signed-off-by: Nicolas Saenz Julienne <nsaenzjulienne@suse.de>
> 
> Hmm, this is not what I expected from the previous thread.  I thought
> we'd just use one dma pool based on runtime available of the zones..

I may be misunderstanding you, but isn't that going back to how things used to
be before pulling in David Rientjes' work? The benefit of having a GFP_KERNEL
pool is that non-address-constrained devices can get their atomic memory there,
instead of consuming somewhat scarcer low memory.