On Fri, Jan 13, 2023 at 11:57:33PM +0000, Saleem, Shiraz wrote: > > Subject: Re: [PATCH v2 rdma-core] irdma: Add ice and irdma to kernel-boot rules > > > > On Wed, Nov 02, 2022 at 04:40:20PM +0000, Nikolova, Tatyana E wrote: > > > Hi Jason, > > > > > > I know it has been a while since we discussed this. Based on your feedback, we > > are proposing another solution for the irdma kernel-boot rules. Could you please > > review it? > > > > > > > > udevadm info --attribute-walk /sys/class/infiniband/rocep47s0f0 > > > > > > > > > > looking at device > > > > '/devices/pci0000:2e/0000:2e:00.0/0000:2f:00.0/infiniband/rocep47s0f0': > > > > > > > > This looks like the problem. For any of this to work the infiniband > > > > device needs to be parented to the aux device, not the PCI device. > > > > > > > > mlx5 did not due this for backwards compat reasons, but this is a > > > > new driver so it could do it properly. > > > > > > Hi Jason - This also impacts us in terms of backwards compatibility. This is a new driver, it by definition does not have backwards compatability. What you are talking about is drop in compatability against mlx5 which is a very different thing. > There are current applications/customers who are looking at > "'/sys/class/infiniband/<ib_device>/device/" for sysfs attributes > like numa_node, local_cpus etc. under the PCI device. That is just wrong assumptions for how sysfs is structured, sadly. > i.e. assuming the PCI device is ib device parent. > > With parent change to auxiliary device, they won't be able find > these attributes and potentially resort to default configurations > that are sub-optimal. Introducing new stuff often comes with an ecosystem upgrade as well.. Jason
