Re: [RFC] configfs_{un,}register_group() semantics

[Al Viro]

[now with hopefully correct address of Daniel Baluta]

On Sun, May 12, 2024 at 05:30:12AM +0100, Al Viro wrote:
> 	Folks, could you confirm if the following is correct?
> 
> 1. configfs_unregister_group() callers are supposed to prevent
> having it called when some items/groups had been created under it.
> The original one (in iio) *does* prevent that (the call chains come
> through the module_exit() of modules pinned by ->make_group() for
> the added subdirectory), but I don't see that documented anywhere and
> AFAICS at least in one case (drivers/pci/endpoint/pci-ep-cfs.c) that is
> not guaranteed.  The same goes for symlinks created in or to those.
> 
> 2. rmdir on directory added by configfs_register_group() is supposed to
> fail (is it even supposed to be used inside the stuff created by mkdir?
> Original use was inside a subsystem, AFAICS).
> 
> 3. rmdir that would've taken out the parent group is supposed to take
> the added one out (again, are they even supposed to be used inside the
> stuff created by mkdir?)
> 
> 4. one is *NOT* supposed to use have ->make_group() schedule creation of
> subdirectories via configfs_register_group(); configfs_add_default_group()
> ought to be used instead.
> 
> drivers/pci/endpoint/pci-ep-cfs.c:pci_epf_make() has this:
>         INIT_DELAYED_WORK(&epf_group->cfs_work, pci_epf_cfs_work);
>         queue_delayed_work(system_wq, &epf_group->cfs_work,
>                            msecs_to_jiffies(1));
> 
>         return &epf_group->group;
> 
> with pci_epf_cfs_work() allocating several config_group and calling
> configfs_register_group() to link those in.  I really doubt that this
> kind of "let's hope that configfs_mkdir() will get through directory
> creation in less than 1ms after ->make_group() returns" is the way it
> is supposed to be done; at a guess, configfs_add_default_group()
> should've been used (synchronously), but I might be missing something
> subtle here.
> 
> Comments?
>