Re: fs/dcache.c - BUG: soft lockup - CPU#5 stuck for 22s! [systemd-udevd:1667]

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

On Thu, May 29, 2014 at 05:53:51PM +0100, Al Viro wrote:
> On Thu, May 29, 2014 at 09:29:42AM -0700, Linus Torvalds wrote:
> > On Thu, May 29, 2014 at 9:23 AM, Al Viro <viro@zeniv.linux.org.uk> wrote:
> > >
> > > BTW, lock_parent() might be better off if in contended case it would not
> > > bother with rename_lock and did something like this:
> > > again:
> > 
> > Ack. I think that's much better.
> 
> Pushed to #for-linus (with dumb braino fixed - it's if (parent != dentry),
> not if (parent)).  I'll wait with folding it back into the commit that
> introduces lock_parent() until we get testing results...

Grrr...  Sadly, that's not good enough.  Leaking rcu_read_lock() on
success is trivial, but there's more serious problem: suppose dentries
involved get moved before we get to locking what we thought was parent.
We end up taking ->d_lock on two dentries that might be nowhere near each
other in the tree, with obvious nasty implications.  Would be _very_ hard
to reproduce ;-/

AFAICS, the following would be safe, but I'd really appreciate any extra
eyes on that sucker:

static inline struct dentry *lock_parent(struct dentry *dentry)
{
        struct dentry *parent = dentry->d_parent;
        if (IS_ROOT(dentry))
                return NULL;
        if (likely(spin_trylock(&parent->d_lock)))
                return parent;
        spin_unlock(&dentry->d_lock);
        rcu_read_lock();
again:
        parent = ACCESS_ONCE(dentry->d_parent);
        spin_lock(&parent->d_lock);
        /*
         * We can't blindly lock dentry until we are sure
         * that we won't violate the locking order.
         * While parent->d_lock is not enough to stabilize
	 * dentry->d_parent, it *is* enough to stabilize
	 * dentry->d_parent == parent.
         */
        if (unlikely(parent != dentry->d_parent)) {
                spin_unlock(&parent->d_lock);
                goto again;
        }
        rcu_read_unlock();
        if (parent != dentry)
                spin_lock(&dentry->d_lock);
        else
                parent = NULL;
        return parent;
}

That variant got force-pushed in place of the previous one, again at the
head of #for-linus.  And I'm definitely not folding it in until it gets
more review and testing.