On Sat, Jan 19, 2019 at 02:02:22PM +0100, Christoph Hellwig wrote:

Interesting.  This suggest it is related to the use of ZONE_DMA by
the FSL SOCs that your board uses.  Let me investigate this a bit more.

As a hack to check that theory I've pushed a new commit to the
powerpc-dma.6-debug branch to use old powerpc GFP_DMA selection
with the new dma direct code:

http://git.infradead.org/users/hch/misc.git/commitdiff/5c532d07c2f3c3972104de505d06b8d85f403f06

And another one that drops the addressability checks that powerpc
never had:

http://git.infradead.org/users/hch/misc.git/commitdiff/18e7629b38465ca98f8e7eed639123a13ac3b669

Can you first test with both patches, and then just with the first
in case that worked?