From mboxrd@z Thu Jan 1 00:00:00 1970 From: Alex Williamson Subject: Re: Device is ineligible for IOMMU domain attach due to platform RMRR requirement Date: Sat, 07 Mar 2015 07:41:12 -0700 Message-ID: <1425739272.4675.83.camel@redhat.com> References: <54F9392B.3060102@hp.com> <1425622242.5200.368.camel@redhat.com> <54FA6C13.1000002@hp.com> <1425703389.4675.49.camel@redhat.com> <54FACF5A.9080504@hp.com> Mime-Version: 1.0 Content-Type: text/plain; charset="UTF-8" Content-Transfer-Encoding: 7bit Cc: kvm@vger.kernel.org To: Steven DuChene Return-path: Received: from mx1.redhat.com ([209.132.183.28]:55165 "EHLO mx1.redhat.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1752907AbbCGOlU (ORCPT ); Sat, 7 Mar 2015 09:41:20 -0500 In-Reply-To: <54FACF5A.9080504@hp.com> Sender: kvm-owner@vger.kernel.org List-ID: On Sat, 2015-03-07 at 05:13 -0500, Steven DuChene wrote: > Alex: > What would be the result of running an earlier kernel that did not have > your RMRR patch on a system that was known to have these problems with > RMRR issues? Would there possibly be some instability when trying to do > PCI passthrough of these same NVidia devices? > > We have a debian install on one of these same systems and it is running > a 3.14.23-2 kernel and we are seeing some issues with PCI passthrough. The potential problems depend on whether the RMRR memory region is actually used, whether the device or the platform depend on that ongoing access, and the address space consumed by the RMRR. In the case of a VM, the fear is that the RMRR is used as a data reporting location by the device, that use continues after the device is assigned to the VM, and that the RMRR memory region overlaps guest RAM. If all of those conditions hold true, then we have a memory integrity issue in the VM. On the platform side, whatever data the platform was depending on the device to report is now disconnected from specified data reporting range and lost. In the case of an Nvidia device, I'd speculate that the more likely cause of issues for that kernel would be that the VT-d hardware may not support snoop-control and the kvm-vfio interface that manages whether KVM emulates cache coherence instructions wasn't added until kernel v3.15. Thanks, Alex