Re: [PATCH] fs: Handle register_shrinker failure

[Al Viro]

On Fri, Apr 28, 2017 at 05:30:46AM +0100, Al Viro wrote:

> I really don't like that.  Your "remove it from all lists and pray that
> nobody has picked a reference of any kind" at the very least needs a careful
> written proof of correctness.  AFAICS, somebody might've found it on the
> list and attempted to grab ->s_umount (grab_super() from another thread
> calling sget()).  Then they'd block until your up_write() in there and
> bugger the system up trying to play with ->s_umount in the object you've
> freed.
> 
> NAK.  Yes, the bug is real, but this is not a solution.

Why do we register it that early, anyway?  super_cache_scan() won't do
anything until we are done with setting the sucker up and dropped ->s_umount.
How about we initialize ->s_shrink.list in alloc_super(), have
deactivate_locked_super() call unregister_shrinker() only if list_empty(...)
and have mount_fs() do
	error = register_shrinker(&sb->s_shrink);
	if (error)
		goto out_sb;
        sb->s_flags |= MS_BORN;
        error = security_sb_kern_mount(sb, flags, secdata);
        if (error)
                goto out_sb;

Folks?  Am I missing something subtle here?