Re: [ANNOUNCE] 3.2.9-rt16

[Steven Rostedt <rostedt@goodmis.org>]

On Tue, 2012-03-06 at 11:20 +0100, Thomas Gleixner wrote:
> Dear RT Folks,
> 
> I'm pleased to announce the 3.2.9-rt16 release.
> 
> Changes vs. 3.2.9-rt15:
> 
>   * cpu hotplug lock init fix [ Steven ]
> 
>   * seqlock fix CONFIG typo
> 

Note, yesterday while running some stress tests I hit a live lock here:

static inline struct dentry *dentry_kill(struct dentry *dentry, int ref)
        __releases(dentry->d_lock)
{
        struct inode *inode;
        struct dentry *parent;

        inode = dentry->d_inode;
        if (inode && !spin_trylock(&inode->i_lock)) {
relock:
                seq_spin_unlock(&dentry->d_lock);
                cpu_relax();
                return dentry; /* try again with same dentry */
        }
        if (IS_ROOT(dentry))
                parent = NULL;
        else
                parent = dentry->d_parent;
        if (parent && !seq_spin_trylock(&parent->d_lock)) {
                if (inode)
                        spin_unlock(&inode->i_lock);
                goto relock;
        }


When it fails to grab either the inode->i_lock or the parent->d_lock it
returns back to dput() and dput() will retry. We get into another one of
these cases where we can spin blocking the holder of the locks.

I experimented with adding a grab lock of the inode->i_lock or
parent->d_lock if they existed (required initializing parent to NULL),
which seemed to help a lot, but then eventually it locked up. As I'm not
sure its safe to grab them straight here even after we release the
dentry->d_lock. I'll have to enable full lockdep to see if this breaks
the ordering.

I haven't looked too deeply into this code yet, but I'm assuming that
dput() can be called where we can't just take the inode or parent lock?

-- Steve