> From: Jason Gunthorpe <jgg@xxxxxxxx> > Sent: Friday, November 8, 2019 6:57 PM > > We should be creating 3 different buses, instead of mdev bus being de- > multiplexer of that? > > > > Hence, depending the device flavour specified, create such device on right > bus? > > > > For example, > > $ devlink create subdev pci/0000:05:00.0 flavour virtio name foo > > subdev_id 1 $ devlink create subdev pci/0000:05:00.0 flavour mdev > > <uuid> subdev_id 2 $ devlink create subdev pci/0000:05:00.0 flavour > > mlx5 id 1 subdev_id 3 > > I like the idea of specifying what kind of interface you want at sub device > creation time. It fits the driver model pretty well and doesn't require abusing > the vfio mdev for binding to a netdev driver. > > > $ devlink subdev pci/0000:05:00.0/<subdev_id> config <params> $ echo > > <respective_device_id> <sysfs_path>/bind > > Is explicit binding really needed? No. > If you specify a vfio flavour why shouldn't > the vfio driver autoload and bind to it right away? That is kind of the point > of the driver model... > It some configuration is needed that cannot be passed at device creation time, explicit bind later can be used. > (kind of related, but I don't get while all that GUID and lifecycle stuff in mdev > should apply for something like a SF) > GUID is just the name of the device. But lets park this aside for a moment. > > Implement power management callbacks also on all above 3 buses? > > Abstract out mlx5_bus into more generic virtual bus (vdev bus?) so > > that multiple vendors can reuse? > > In this specific case, why does the SF in mlx5 mode even need a bus? > Is it only because of devlink? That would be unfortunate > Devlink is one part due to identifying using bus/dev. How do we refer to its devlink instance of SF without bus/device? Can we extend devlink_register() to accept optionally have sf_id? If we don't have a bus, creating sub function (a device), without a 'struct device' which will have BAR, resources, etc is odd. Now if we cannot see 'struct device' in sysfs, how do we persistently name them? Are we ok to add /sys/class/net/sf_netdev/subdev_id And /sys/class/infiniband/<rdma_dev>/subdev_id So that systemd/udev can rename them as en<X?><subdev_id> and roce<X><subdev_id> If so, what will be X without a bus type? This route without a bus is certainly helpful to overcome the IOMMU limitation where IOMMU only listens to pci bus type for DMAR setup, dmar_register_bus_notifier(), and in intel_iommu_init()-> bus_set_iommu(&pci_bus_type, &intel_iommu_ops); and other IOMMU doing similar PCI/AMBA binding. This is currently overcome using WA dma_ops.
