On Thu, 9 Jan 2014 14:42:39 -0700
Matthew Wilcox [off-list ref] wrote:

On Thu, Jan 09, 2014 at 04:27:31PM -0500, Steven Rostedt wrote:

quoted

Note, the crash came from stressing the deletion and reading of debugfs
files. I was not able to recreate this via normal files. But I'm not
sure they are safe. It may just be that the race window is much harder
to hit.

But "normal" files have a 'destroy_inode' method.  So you've basically
only fixed it for debugfs (and maybe a few other unusual filesystems).
Why doesn't the code look like this:

Because I thought of that after I sent the email ;-)

Well, that's not really true. I don't know the semantics of the
destroy_inode() call. But I should have asked that in my change log.

static void i_callback(struct rcu_head *head)
{
	struct inode *inode = container_of(head, struct inode, i_rcu);
	__destroy_inode(inode);
	if (inode->i_sb->s_op->destroy_inode)
		inode->i_sb->s_op->destroy_inode(inode);
	else
		kmem_cache_free(inode_cachep, inode);
}

static void destroy_inode(struct inode *inode)
{
	BUG_ON(!list_empty(&inode->i_lru));
	call_rcu(&inode->i_rcu, i_callback);
}

We'd then have to get rid of all the call_rcu() invocations in individual
filesystems' destroy_inode methods, but that doesn't sound like a bad
thing to me.

Which is another reason that I didn't do it, as I didn't know all the
happenings inside the ->destroy_inode() calls. But yeah, I agree with
this.

Also, can iput() sleep? If not then we are OK. Otherwise, we need to be
careful about any mutex being grabbed in those call backs, as the
rcu_callback can't sleep either.

-- Steve