On Fri, 2018-06-29 at 06:56 -0700, Linus Torvalds wrote: > On Thu, Jun 28, 2018 at 7:22 PM Benjamin Herrenschmidt > <b...@kernel.crashing.org> wrote: > > > > This fixes it by instead doing an explicit kobject_del() when > > the glue dir is empty, by keeping track of the number of > > child devices of the gluedir. > > Ugh. I was hoping that you'd just do the "only check duplicate names > if actually in use" instead.
I had a look (see my other email). It's non-trivial. We can still look into it, but from what I gathered of how sysfs works, it's based on kernfs which doesn't have the kobjects nor access to the reference count, and is the holder of the names rbtree. So we'd need to find a way to convey that "in-use" information down to kernfs (I thought maybe adding an optional pointer to a kref ? but then, it's still somewhat racy ...) Additionally, we would need a few more pairs of eyes to make sure that sticking duplicates in that rbtree isn't going to break some corner case in there. It's a code base I'm not at all familiar with. So while I agree with the overall idea that a kobject whose reference has gone down to 0 should probably be ignored by sysfs, actually doing it doesn't seem simple. > This patch seems to make patch 1/2 pointless. No? Somewhat. It's stil the "right thing" to do in case a hole shows up elsewhere, and it's an easier stable backport. But yes. I wrote it before I figured out what to do with 2/2 (I was still trying to do what you wanted and just skip ref=0 in sysfs name lookups). But yes, with 2/2, there shouldn't be a kobject with a 0 ref in the list anymore... hopefully. I would very much appreciate the opinion of someone more familiar with sysfs and the device core on all this though. Cheers, Ben.