On Fri, 26 Jan 2024 14:26:08 -0800 Linus Torvalds <torvalds@xxxxxxxxxxxxxxxxxxxx> wrote: > nother thing that worries me is that odd locking that releases the > lock in the middle. I don't understand why you release the > tracefs_mutex() over create_file(), for example. There's a lot of > "take, drop, re-take, re-drop" of that mutex that seems strange. This was because the create_file/dir() would call into the VFS which would grab locks, and on a final dput() on a ei dentry that is to be freed, calls back into eventfs_set_ei_status_free() which also grabs the eventfs_mutex. But it gets called with the same VFS locks that are taken by create_file/dir() VFS calls. This was caught by lockdep. Hence the dropping of those locks. The eventfs_mutex is just protecting the ei list and also assigning and clearing the ei->dentry. Now that dentry is used to synchronize the last close, and also to know if the ei was ever referenced. If ei->dentry is NULL it can be freed immediately (after SRCU) when the directory is deleted. But if ei->dentry is set, it means that something may still have a reference to it and must be freed after the last dput() and SRCU. Now some of this was needed due to the way the dir wrapper worked so I may be able to revisit this and possibly just use an ei->ref counter. But I wasted enough time on this and I'm way behind in my other responsibilities, so this is not something I can work on now. -- Steve
