On Thu, Jun 06, 2019 at 08:24:41AM +0200, Christoph Hellwig wrote: > On Wed, Jun 05, 2019 at 05:22:35PM -0300, Jason Gunthorpe wrote: > > On Wed, Jun 05, 2019 at 09:08:31PM +0200, Christoph Hellwig wrote: > > > This ensures all proper DMA layer handling is taken care of by the > > > SCSI midlayer. > > > > Maybe not entirely related to this series, but it looks like the SCSI > > layer is changing the device global dma_set_max_seg_size() - at least > > in RDMA the dma device is being shared between many users, so we > > really don't want SCSI to make this value smaller. > > > > Can we do something about this? > > We could do something about it as outlined in my mail - pass the > dma_params explicitly to the dma_map_sg call. But that isn't really > suitable for a short term fix and will take a little more time.
Sounds good to me, having every dma mapping specify its restrictions makes a lot more sense than a device global setting, IMHO. In RDMA the restrictions to build a SGL, create a device queue or build a MR are all a little different. ie for MRs alignment of the post-IOMMU DMA address is very important for performance as the MR logic can only build device huge pages out of properly aligned DMA addresses. While for SGLs we don't care about this, instead SGLs usually have the 32 bit per-element length limit in the HW that MRs do not. > Until we've sorted that out the device paramter needs to be set to > the smallest value supported. smallest? largest? We've been setting it to the largest value the device can handle (ie 2G) Jason
