Re: [PATCH v7 1/4] spinlock: A new lockref structure for locklessupdate of refcount

[Linus Torvalds]

On Sun, Sep 8, 2013 at 5:03 PM, Al Viro <viro@xxxxxxxxxxxxxxxxxx> wrote:
>
> Well...  unlazy_walk() is always followed by terminate_walk() very shortly,
> but there's a minor problem - terminate_walk() uses "are we in RCU
> mode?" for two things:
>         a) do we need to do path_put() here?
>         b) do we need to unlock?
> If you introduce the third case ("no need to do unlock and no need to
> do path_put()"), we'd better decide how to check for that case...

Actually, I decided to take advantage of those two cases instead, and
I have a patch that I think does the right thing. Basically, I start
off unlazy_walk() with just doing that lockref_get_not_dead() on the
parent dentry, and if that fails I just return an error in RCU mode
(so terminate_walk() does what it always used to do, and we haven't
done anything else to any refcounts).

Now, if the lockref_get_not_dead() succeeded, that means that we have
a reference on the nd->path.dentry, and we can now just do
"mntget(nd->path.mnt);". Ta-Daa! We now have everything done for the
non-RCU case for terminate_walk().

So after that point, we clear LOOKUP_RCU, and make the rule be that
any return (error or success) has to do unlock_rcu_walk(). And then
all the other refcounts are easy, we can just "dput(dentry);" after
that.

I haven't tested it yet, I was going to reboot into it just now. But
I'm attaching the patch here. Maybe I missed some detail, but it all
seems simpler.

Note that this patch requires the "lockref_get_not_dead()" cleanup at
the top of my current -git.

        Linus

Attachment: patch.diff
Description: Binary data