On Tue, Dec 18, 2018 at 10:22:36AM -0500, Doug Ledford wrote: > On Tue, 2018-12-18 at 13:30 +0000, Ruhl, Michael J wrote: > > Hmm, taking a little close look at the debugfs.c module for HFI I am wondering if > > there is some confusion WRT how the HFI driver uses debugfs. > > > > The init function (hfi1_dbg_ibdev_init() occurs after hfi1_ib_register(), with a > > comment of "create debugfs files after init and ib register". This uses the > > hfi1_ibdev data structure as the "private data" for the debugfs information. > > > > However, the root debufs system is based off of "DRIVER_NAME" (hfi1), and > > the actual filenames within that directory is based off of hfi1_class_name (hfi1) > > and the "port" value. Well, it is 'unit' and unit comes from an internal global IDR and is thus guaranteed globally unique. > > For each HFI we will increment port, so for two instance of the driver, there will > > always be a uniquely defined name. > > > > From that perspective, as far as I can tell, the HFI driver is meeting the > > "globally uniqueness" requirement. > > You are correct. This satisfies the needs for device rename and can be > left just as it is. > > The original problem stemmed from drivers using their ib device name > (such as mlx4_0), which presents a problem on rename in that when the > first mlx4 device is registered it would be mlx4_0, then udev renames > the device to <whatever> but the debugfs files are still in a directory > named mlx4_0, and then you register another mlx4 device, and because > there is no mlx4_0 any more, the new device gets the name mlx4_0 again, > and now the creation of the debugfs files for the second mlx4 device > fail because the name mlx4_0 is already in use. Your naming scheme > avoids that problem, and so it is fine just as it is and there is no > need to break any existing user space or switch to PCI address naming > schemes. Yah, I agree, it still breaks userspace in the sense that the hfi_1 debugfs name doesn't correlate to any ibdev name, but it doesn't cause any kernel level problem. You should probably still use the PCI name for debugfs and add a symlink to it from from the unit based name. Jason
