On Wed, Jan 09, 2019 at 12:22:25PM -0700, Logan Gunthorpe wrote: > Hey, > > I'm resending this because I've recently found out that the change we > made to use the NTB struct device in DMA allocations is wrong and > needs to be reverted. Turns out that, when running with an IOMMU, > dma_alloc_coherent() will always fail if you pass it the NTB struct > device. This is because the device has not been assigned an IOMMU > group and the Intel IOMMU at least expect the devices to be on the PCI > bus and be able to find a proper bus-dev-fn number through a struct > pci device. Therefore, we must revert the change and I've changed > patch 2 to do this and remove the no longer necessary DMA mask > adjustments. > > I'm not sure if we can get past the impass in getting this series merged: > I still maintain every patch in this series is necessary to fix a > regression and there's no way to add port numbers to switchtec in the > crosslink configuration so it can't be fixed in the other way that was > suggested.
Given the need for this to get in (as it does fix actual problems) and the absence of comments on v3, I'm adding this to the ntb-next branch. If there is anyone that has anything more to say on this, please do so immediately. Otherwise, I think this probably needs to be moved to the ntb branch and go into 5.1-rc1. Thoughts? Thanks, Jon > > Logan > > -- > > Changes since v2: > - Rebased on v5.0-rc1 > - Modify Patch 2 to revert back to using the PCI struct device > instead of the NTB struct device in DMA calls > - Collected Allen's Acks > - Collected Alexander's Tested-By > > Changes since v1: > - Rebased onto ntb-next (there was a minor conflict in a recent change > to the intel driver) > - Collected Dave's Ack > > -- > > Logan Gunthorpe (8): > NTB: ntb_tool: reading the link file should not end in a NULL byte > NTB: Revert the change to use the NTB device dev for DMA allocations > NTB: Fix the default port and peer numbers for legacy drivers > NTB: ntb_pingpong: Choose doorbells based on port number > NTB: perf: Don't require one more memory window than number of peers > NTB: perf: Fix support for hardware that doesn't have port numbers > NTB: perf: Fix race condition when run with ntb_test > NTB: ntb_test: Fix bug when counting remote files > > drivers/ntb/hw/amd/ntb_hw_amd.c | 4 ---- > drivers/ntb/hw/idt/ntb_hw_idt.c | 6 ----- > drivers/ntb/hw/intel/ntb_hw_gen1.c | 4 ---- > drivers/ntb/ntb.c | 9 ++------ > drivers/ntb/test/ntb_perf.c | 29 +++++++++++++++++++------ > drivers/ntb/test/ntb_pingpong.c | 14 +++++------- > drivers/ntb/test/ntb_tool.c | 9 ++++---- > tools/testing/selftests/ntb/ntb_test.sh | 2 +- > 8 files changed, 35 insertions(+), 42 deletions(-) > > -- > 2.19.0
