On Fri, 8 Nov 2019 20:41:18 +0100, Jiri Pirko wrote: > Fri, Nov 08, 2019 at 08:06:40PM CET, jakub.kicinski@xxxxxxxxxxxxx wrote: > >On Fri, 8 Nov 2019 13:12:33 +0100, Jiri Pirko wrote: > >> Thu, Nov 07, 2019 at 09:32:34PM CET, jakub.kicinski@xxxxxxxxxxxxx wrote: > >> >On Thu, 7 Nov 2019 10:04:48 -0600, Parav Pandit wrote: > >> >> Mellanox sub function capability allows users to create several hundreds > >> >> of networking and/or rdma devices without depending on PCI SR-IOV support. > >> > > >> >You call the new port type "sub function" but the devlink port flavour > >> >is mdev. > >> > > >> >As I'm sure you remember you nacked my patches exposing NFP's PCI > >> >sub functions which are just regions of the BAR without any mdev > >> >capability. Am I in the clear to repost those now? Jiri? > >> > >> Well question is, if it makes sense to have SFs without having them as > >> mdev? I mean, we discussed the modelling thoroughtly and eventually we > >> realized that in order to model this correctly, we need SFs on "a bus". > >> Originally we were thinking about custom bus, but mdev is already there > >> to handle this. > > > >But the "main/real" port is not a mdev in your case. NFP is like mlx4. > >It has one PCI PF for multiple ports. > > I don't see how relevant the number of PFs-vs-uplink_ports is. Well. We have a slice per external port, the association between the port and the slice becomes irrelevant once switchdev mode is enabled, but the queues are assigned statically so it'd be a waste of resources to not show all slices as netdevs. > >> Our SFs are also just regions of the BAR, same thing as you have. > >> > >> Can't you do the same for nfp SFs? > >> Then the "mdev" flavour is enough for all. > > > >Absolutely not. > > > >Why not make the main device of mlx5 a mdev, too, if that's acceptable. > >There's (a) long precedence for multiple ports on one PCI PF in > >networking devices, (b) plenty deployed software > >which depend on the main devices hanging off the PCI PF directly. > > > >The point of mdevs is being able to sign them to VFs or run DPDK on > >them (map to user space). > > > >For normal devices existing sysfs hierarchy were one device has > >multiple children of a certain class, without a bus and a separate > >driver is perfectly fine. Do you think we should also slice all serial > >chips into mdevs if they have multiple lines. > > > >Exactly as I predicted much confusion about what's being achieved here, > >heh :) > > Please let me understand how your device is different. > Originally Parav didn't want to have mlx5 subfunctions as mdev. He > wanted to have them tight to the same pci device as the pf. No > difference from what you describe you want. However while we thought > about how to fit things in, how to handle na phys_port_name, how to see > things in sysfs we came up with an idea of a dedicated bus. The difference is that there is naturally a main device and subslices with this new mlx5 code. In mlx4 or nfp all ports are equal and statically allocated when FW initializes based on port breakout. Maybe it's the fact I spent last night at an airport but I'm feeling like I'm arguing about this stronger than I actually care :) > We took it upstream and people suggested to use mdev bus for this. > > Parav, please correct me if I'm wrong but I don't think where is a plan > to push SFs into VM or to userspace as Jakub expects, right? There's definitely a plan to push them to VFs, I believe that was part of the original requirements, otherwise there'd be absolutely no need for a bus to begin with.
