From mboxrd@z Thu Jan 1 00:00:00 1970 From: Alex Williamson Subject: Re: [v7 0/3] plumb kvm/vfio to notify kvm:group attach/detach Date: Wed, 30 Nov 2016 10:06:57 -0700 Message-ID: <20161130100657.04ba599f@t450s.home> References: <1479794973-11910-1-git-send-email-jike.song@intel.com> <583CEFAA.3070600@intel.com> Mime-Version: 1.0 Content-Type: text/plain; charset=US-ASCII Content-Transfer-Encoding: 7bit Cc: pbonzini@redhat.com, guangrong.xiao@linux.intel.com, kevin.tian@intel.com, kwankhede@nvidia.com, cjia@nvidia.com, kvm@vger.kernel.org To: Jike Song Return-path: Received: from mx1.redhat.com ([209.132.183.28]:53246 "EHLO mx1.redhat.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1758264AbcK3RHn (ORCPT ); Wed, 30 Nov 2016 12:07:43 -0500 In-Reply-To: <583CEFAA.3070600@intel.com> Sender: kvm-owner@vger.kernel.org List-ID: On Tue, 29 Nov 2016 11:02:02 +0800 Jike Song wrote: > On 11/22/2016 02:09 PM, Jike Song wrote: > > So far KVM and VFIO are mostly transparent to each other. > > However, there are users who would rely on them both. For > > example, KVMGT relies on VFIO to mediate device operations, > > and it also relies on KVM for features such as guest page > > tracking. To do that, it needs to know which KVM instance > > a vfio_group is attached to, when it will be detached. > > > > There is already a kvm_vfio device serving for similar > > purpose, this series extends it to allow users register > > notifiers on a vfio_group, and notify kvm:group attaching > > and detaching events to that. > > > > > > Changes v7: > > - replace vfio->opened with container user like iommu notifier registration > > - fix a typo > > Changes v6: > > - only allow registration on already opened vfio_group > > - moved WARN_ON from vfio_group_release to vfio_group_fops_release > > - remove unnecessary check > > Changes v5: > > - separate vfio_group and vfio_iommu events from each other > > - force user to specify which type it is interested in, and which events. > > Register it IFF all required events are supported > > Changes v4: > > - use a notifier-based implementation in vfio_group > > - combine the notifier along with iommu notifier by Kirti > > Changes v3: > > - don't touch kvm_put_kvm, vfio won't need it > > - fix a race condition > > Changes v2: > > - As Paolo suggested, drop the type and get/put ops from > > vfio, assuming that the data will always be kvm. > > > > > > > > Jike Song (3): > > vfio: vfio_register_notifier: classify iommu notifier > > vfio: support notifier chain in vfio_group > > kvm: set/clear kvm to/from vfio_group when group add/delete > > > > drivers/vfio/vfio.c | 167 ++++++++++++++++++++++++++++++++++++++++++++------- > > include/linux/vfio.h | 22 ++++++- > > virt/kvm/vfio.c | 31 ++++++++++ > > 3 files changed, 195 insertions(+), 25 deletions(-) > > Hi Alex, > > Since Paolo has ACKed a previous version of PATCH [3/3] before, if you have > no more comments on this series, could you pick them up? Please be aware that > 1/3 was updated, or I can send a v8 if you prefer :) Sorry for the delay, I was on holiday. Please send a v8, particularly since I noted updates in patch 3/3 as well. I would prefer to commit this with a R-b from Kirti since the NVIDIA driver is also affected by this, but I believe this is the right approach for us long term. I'm also having trouble finding an ack from Paolo that's particularly relevant for the kvm-vfio portion, but I don't think this is a controversial issue, I feel comfortable that he's onboard with the idea. Thanks, Alex