Comment 16 for bug 1823753

The hisi_sas maintainer is experimenting with a patch that will change the driver's 33 page allocations to single page allocations. If there is no significant performance impact, that could be a way forward to deal with the fragmentation issue costing us up to 31M of CMA.

In addition, there is DMA/CMA optimization making its way upstream, that uses non-CMA memory for single page requests:
  https://lkml.org/lkml/2019/5/6/1219

This avoids the impact of the CMA memory usage increase in the RDMA/hns driver (see comment #11) and - with the aforementioned hisi_sas patch - those allocations as well.

After those optimizations, I am able to fulfill all cma allocation requests from our D06 CS board[*] with cma=64M. However, Ubuntu ships with only 16M of CMA. Note that upstream's defconfig allocates 32M of CMA,
 which is apparently required for the RPi VC4:
  https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=ebf089248dab2ef569e5e26a607f0977a71182b7

So there maybe other reasons we want to bump to at least 32M - I'm not sure that we'd want to go all the way to 64M for general purpose kernel though.

[*] Keeping in mind that this is just tuning for a specific system/config. Throw in a bunch of mlx5 cards, and I suspect we'll still see cma_alloc log spew on this platform.