These patches implement a new set of DMA operations designed to allow devices that cannot address bit 59 to use more than 32-bit DMA.
The previous implementation for PHB3 assumed contiguous memory, which is no longer the case on POWER9 and possibly later. As a result, a new approach was necessary, which is what this new implementation does. These patches aren't perfect yet - most notably, there is no implementation of unmap() - once a TCE is created it is never freed. For the most part this isn't an issue, but could potentially be a problem for a device that can't address many bits and maps *a lot* of memory. I have a much more complex implementation in the back that I will post in future, but until that's ready and fully tested, I believe the drastic performance improvement is worth getting this upstream sooner rather than later. Thanks to Benjamin Herrenschmidt, Alistair Popple and Timothy Pearson for their help. Russell Currey (3): powerpc/powernv/pci: Track largest available TCE order per PHB powerpc/powernv: DMA operations for discontiguous allocation powerpc/powernv/pci: Track TCE tables in debugfs arch/powerpc/include/asm/dma-mapping.h | 1 + arch/powerpc/platforms/powernv/Makefile | 2 +- arch/powerpc/platforms/powernv/pci-dma.c | 243 ++++++++++++++++++++++ arch/powerpc/platforms/powernv/pci-ioda.c | 132 +++++++----- arch/powerpc/platforms/powernv/pci.h | 10 + 5 files changed, 334 insertions(+), 54 deletions(-) create mode 100644 arch/powerpc/platforms/powernv/pci-dma.c -- 2.17.1