> From: gregkh@xxxxxxxxxxxxxxxxxxx <gregkh@xxxxxxxxxxxxxxxxxxx> > Sent: Thursday, October 1, 2020 6:25 PM > > On Thu, Oct 01, 2020 at 09:49:00AM -0300, Jason Gunthorpe wrote: > > On Thu, Oct 01, 2020 at 12:59:25PM +0200, gregkh@xxxxxxxxxxxxxxxxxxx > wrote: > > > We don't add infrastructure without users. And the normal rule of > > > thumb of "if we have 3 users, then it is a semi-sane api" really applies > here. > > > > Based on recent discussions I'm expecting: > > - Intel SOF > > - New Intel RDMA driver > > - mlx5 RDMA driver conversion > > - mlx4 RDMA driver conversion > > - mlx5 subdevice feature for netdev > > - Intel IDXD vfio-mdev > > - Habana Labs Gaudi netdev driver > > > > Will use this in the short term. > > > > I would like, but don't expect too see, the other RDMA RoCE drivers > > converted - cxgb3/4, i40iw, hns, ocrdma, and qedr. It solves an > > annoying module loading problem we have. > > > > We've seen the New Intel RDMA driver many months ago, if patch 1 is > > going to stay the same we should post some of the mlx items next week. > > > > It is hard to co-ordinate all of this already, having some general > > agreement that there is nothing fundamentally objectionable about > > ancillary bus will help alot. > > I agree, but with just one user (in a very odd way I do have to say, more on > that on the review of that specific patch), it's hard to judge if this is useful are > not, right? > As Jason mentioned above, mlx5 subdevice feature, I like to provide more context before posting the patches. I have rebased and tested mlx5 subfunction devices for netdev to use ancillary device as per the RFC posted at [1]. These subdevices are created dynamically on user request. Typically then are in range of hundreds. Please grep for virtbus to see its intended use in [1]. To refresh the memory, before working on the RFC [1], mlx5 subfunction use is also discussed further with Greg at [2]. Recently I further discussed ancillary bus (virtbus) intended use for mlx5 subfunction with netdev community at [3] and summarized in [4] , jump to last slide 22. mlx5 series is bit long and waiting for mainly ancillary bus to be available apart from some internal reviews to finish. [1] https://lore.kernel.org/netdev/20200519092258.GF4655@nanopsycho/ [2] https://patchwork.kernel.org/patch/11280547/#23056985 [3] https://netdevconf.info/0x14/pub/papers/45/0x14-paper45-talk-paper.pdf [4] https://netdevconf.info/0x14/pub/slides/45/sf_mgmt_using_devlink_netdevconf_0x14.pdf > So, what happened to at least the Intel SOF driver usage? That was the > original user of this bus (before it was renamed), surely that patchset should > be floating around somewhere in Intel, right? > > thanks, > > greg k-h
