Re: linux-next: cgroup_mount() falls asleep forever

[Al Viro <viro@ZenIV.linux.org.uk>]

On Wed, Sep 24, 2014 at 06:29:27PM +0400, Andrey Wagin wrote:
> 2014-09-24 14:31 GMT+04:00 Andrey Wagin <avagin@gmail.com>:
> > Hi All,
> 
> The problem is in a following commit:
> 
> commit 0c7bf3e8cab7900e17ce7f97104c39927d835469
> Author: Zefan Li <lizefan@huawei.com>
> Date:   Sat Sep 20 14:49:10 2014 +0800
> 
>     cgroup: remove redundant variable in cgroup_mount()
> 
>     Both pinned_sb and new_sb indicate if a new superblock is needed,
>     so we can just remove new_sb.
> 
>     Note now we must check if kernfs_tryget_sb() returns NULL, because
>     when it returns NULL, kernfs_mount() may still re-use an existing
>     superblock, which is just allocated by another concurent mount.
> 
>     Suggested-by: Tejun Heo <tj@kernel.org>
>     Signed-off-by: Zefan Li <lizefan@huawei.com>
>     Signed-off-by: Tejun Heo <tj@kernel.org>

Lovely...  First of all, that thing is obviously racy - there's nothing
to prevent another mount happening between these two places.  Moreover,
kernfs_mount() calling conventions are really atrocious - pointer to
bool just to indicate that superblock is new?

Could somebody explain WTF is the whole construction trying to do?  Not
to mention anything else, what *does* this pinning a superblock protect
from?  Suppose we have a superblock for the same root with non-NULL ns
and _that_ gets killed.  We get hit by the same
	percpu_ref_kill(&root->cgrp.self.refcnt);
so what's the point of pinned_sb?  Might as well have just bumped the
refcount, superblock or no superblock.  And no, delaying that kernfs_kill_sb()
does you no good whatsoever - again, pinned_sb might have nothing to do with
the superblock we are after.