Hi Joerg, On 09/09/16 at 10:41am, Joerg Roedel wrote: > > Hi Baoquan, > > On Fri, Sep 09, 2016 at 04:22:25PM +0800, Baoquan He wrote: > > Recently I tried to fix the kdump failure in amd iommu system again, and > > now the latest code works, IO_PAGE_FAULT can't be seen any more. But on > > several amd iommu system with bnx2 NIC, always IO_PAGE_FAULT will be > > printed out. After investegating I found out bnx2 driver doesn't reset > > hardware/reg like other pci device, it does the reset job in bnx2_open > > which is the net device up stage. So with this patch the IO_PAGE_FAULT > > is away too on the system with bnx2 NIC. I will > > > > However when I got a intel system with vt-d and bnx2 NIC, kdump works > > well, and no any error message can be seen. From code it clearly shows > > the domain assignment is done in __intel_map_single, at this time bnx2 > > driver hasn't reset device, the on-flight DMA should still exist. Do you > > have any idea on this? Or I missed anything? I also deferred the > > set_dte_entry calling to __map_single calling, the principal should be > > similar. > > Did you make sure that all unity-mappings are in place in the newly > assigned domain for the bnx2 device before domains are switched? What I am doing is that in iommu driver init stage anything will keep going forware as it does in normal kernel except for set_dte_entry calling. If in kdump kernel and in iommu init stage, just return directly at the beginning of set_dte_entry. For identity mapping, the pass through handling will do everything but return from the beginning of set_dte_entry too. Since it need install pte_root into dev table entry too though its pte_root is NULL. The unity mapping range only does iova reservation. It doesn't do dev table entry handling before device driver init. So unity mappings should be OK. Thanks a lot! Baoquan