On Sun, Jan 07, 2024 at 01:32:28PM -0500, Steven Rostedt wrote: > On Sun, 7 Jan 2024 13:29:12 -0500 > Steven Rostedt <rostedt@xxxxxxxxxxx> wrote: > > > > > > > IOW, the inode_permission() in lookup_one_len() that eventfs does is > > > redundant and just wrong. > > > > I don't think so. > > Just to make it clear. eventfs has nothing to do with mkdir instance/foo. > It exists without that. Although one rationale to do eventfs was so Every instance/foo/ tracefs instances also contains an events directory and thus a eventfs portion. Eventfs is just a subtree of tracefs. It's not a separate filesystem. Both eventfs and tracefs are on the same single, system wide superblock. > that the instance directories wouldn't recreate the same 10thousands > event inodes and dentries for every mkdir done. I know but that's irrelevant to what I'm trying to tell you. A mkdir /sys/kernel/tracing/instances/foo creates a new tracefs instance. With or without the on-demand dentry and inode creation for the eventfs portion that tracefs "instance" has now been created in its entirety including all the required information for someone to later come along and perform a lookup on /sys/kernel/tracing/instances/foo/events. All you've done is to defer the addition of the dentries and inodes when someone does actually look at the events directory of the tracefs instance. Whether you choose to splice in the dentries and inodes for the eventfs portion during lookup and readdir or if you had chosen to not do the on-demand thing at all and the entries were created at the same time as the mkdir call are equivalent from the perspective of permission checking. If you have the required permissions to look at the events directory then there's no reason why listing the directory entries in there should fail. This can't even happen right now.
