On Mon, Sep 12, 2022 at 04:32:09PM +0530, Kashyap Desai wrote: > I noted your response. Issue is possible to any RDMA h/w and issue is not > completely Rejected. It is just that you need another way to fix it > (preferred to handle it in iommu) > We can reopen discussion if we see another instance from other vendors. > Quick workaround is - use 63 bit DMA mask in rdma low level driver if > someone really wants to work around this issue until something more robust > fix committed in upstream kernel (either ib stack or iommu stack). That is not quite right, I said it should also be fixed in RDMA, just none of the patches you are proposing fix it correctly. You must avoid the mathematical overflow by reworking the logic. Jason
