On Mon, May 26, 2014 at 04:27:03PM +0100, Al Viro wrote:

> Do these livelocks keep happening indefinitely, once triggered?  IOW,
> is that a buggered state of dcache and/or kernfs, or is it a transient pileup
> that happens when we invalidate a subtree there?

Could you slap
	if (dentry->d_sb->s_magic == SYSFS_MAGIC) {
		printk(KERN_INFO "killing %p4d", dentry);
		WARN_ON(1);
	}
in the very beginning of dentry_kill(),
	if (dentry->d_sb->s_magic == SYSFS_MAGIC) {
		printk(KERN_INFO "invalidate %p4d", dentry);
		WARN_ON(1);
	}
right after the
        if (!dentry->d_inode) {
                d_drop(dentry);
                goto out;
        }
in check_submounts_and_drop(), reproduce that shite and see what
gets into the log between USB disconnect and soft lockup?  Warning:
it will produce an obscene amount of output.  If it gets _really_ excessive
(as in "can't even get through the boot without drowning in syslog traffic"),
we could add a sysctl turning those on, but let's try and see if it's
survivable in that form first...