Re: [PATCH v2] vfs: Don't exchange "short" filenames unconditionally.

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

On Sun, Sep 28, 2014 at 07:05:56PM +0100, Al Viro wrote:

> One thing that worries me is the barriers that might be needed on assignments
> to ->d_name.name.  We should be no worse than we are right now - either RCU
> accessors are careful enough with ACCESS_ONCE() and everything's fine,
> or they are not, in which case we already have a bug in mainline - swapping
> ->d_name followed by dput() and freeing of target is no better than
> copying ->d_name from target to source followed by kfree_rcu() of what
> used to be ->d_name.name of source.  IOW, if RCU lookup could pick
> a value of ->d_name.name that got obsolete by d_move() happening
> before our read_lock_rcu(), we would be in trouble anyway - it might
> already have had its freeing RCU-scheduled and thus not delayed
> by read_lock_rcu() done afterwards.  So I think the patch below doesn't
> introduce new problems of that sort, but I'd really appreciate if RCU
> people would take a look at the situation with barriers in that area.
> Are those ACCESS_ONCE() in dentry_cmp() and prepend_name() enough, or
> do we need some barriers in switch_names() as well?

Hmm...  OK, dentry_cmp() is doing something similar to open-coded
rcu_dereference().  prepend_name() does not, and I really wonder if
that's correct...

I'm afraid that the answer is "should've been more careful when switching
d_path() to RCU", but maybe there's something subtle I'm missing there...

I really hate memory ordering rules on alpha ;-/