G'day Peter, The problem is that the physical memory you are I/Oing from is fragmented. In X it is very likely that you will always get a maximum physically contiguous I/O of only 4k because that is the page size of the VM system. You have a couple of solutions. 1> Limit your I/Os to 1/2 Meg 2> Or if your DMA engine is capable then my solution is to setup multiple DMA programs then at 'Filter' time (i.e. at primary interrupt time) reprogram the DMA engine with the new DMA. I personally have used approach 2 with really dumb hardware, i.e. no page list at all just a single physical address and length. Using Filter dma chaining I managed to get 112Mb/s throughput with a not bad 5% CPU load. Remember if you go for this solution totally prepare the ENTIRE I/O before starting the first I/O running. There is very little you are allowed to do at primary interrupt time, you certainly CAN NOT call genPhysicalSegments. Godfrey At 16:43 +0100 02-6-13, Peter Young wrote: Hi, I am creating an IOMemoryCursor using: CursP = IOMemoryCursor::withSpecification ( SegFn, UINT_MAX, UINT_MAX, 1 ); which gives me a valid MemoryCursor. I then call CursP->genPhysicalSegments( MemDescP, start, Segments, NumSegments, MaxTransferSize, &BytesToTransfer ); to create the DMA address/len pairs I require in 'Segments'. When I debug SegFn, the Segment.length field is ALWAYS 4096 regardless of the values passed to the withSpecification function above. This gives me a max transfer size of 0.5Mb (128 entries), which is way too small for my requirements. Questions: 1. Why does my segment function ONLY get 4k segments? What am I doing wrong? 2. When 'wiring' the User buffer, do I just need to 'prepare' the MemoryDescriptor, or do I need to 'map' it as well? _______________________________________________ darwin-kernel mailing list | darwin-kernel@lists.apple.com Help/Unsubscribe/Archives: http://www.lists.apple.com/mailman/listinfo/darwin-kernel Do not post admin requests to the list. They will be ignored.