On 2019-07-25 11:58 a.m., Matthew Wilcox wrote: > On Thu, Jul 25, 2019 at 11:53:20AM -0600, Logan Gunthorpe wrote: >> >> >> On 2019-07-25 11:40 a.m., Greg Kroah-Hartman wrote: >>> On Thu, Jul 25, 2019 at 11:23:21AM -0600, Logan Gunthorpe wrote: >>>> cdev_get_by_path() attempts to retrieve a struct cdev from >>>> a path name. It is analagous to blkdev_get_by_path(). >>>> >>>> This will be necessary to create a nvme_ctrl_get_by_path()to >>>> support NVMe-OF passthru. >>> >>> Ick, why? Why would a cdev have a "pathname"? >> >> So we can go from "/dev/nvme0" (which points to a char device) to its >> struct cdev and eventually it's struct nvme_ctrl. Doing it this way also >> allows supporting symlinks that might be created by udev rules. > > But you're not really trying to go from a string to a chardev. You're > trying to go from a nvmet_subsys to a chardev. Isn't there a better > way to link the two somewhere else? > > (I must confess that once I would have known the answer to this, but > the NVMe subsystem has grown ridiculously complex and I can no longer > fit it in my head) Well the nvmet_subsys isn't related to the nvme_ctrl (and thus char dev) at all. An nvmet_subsys is created via configfs and the user has to specify an NVMe controller for it to use (by writting a string to a config attribute). The best handle the user has is a path to the controller's cdev (/dev/nvmeX) so the fabrics code has to be able to lookup the corresponding struct nvme_ctrl from the path. This is directly analogous to the way NVMe-of works today: it uses blkdev_get_by_path() to translate a user provided path to a struct block_device. The only difference here is that, for passthru, we need a nvme_ctrl, not a block device. Logan
